volII Hal

Finite Elements II: Galerkin Approximation, Elliptic
and Mixed PDEs

Alexandre Ern, Jean-Luc Guermond
To cite this version:

Alexandre Ern, Jean-Luc Guermond. Finite Elements II: Galerkin Approximation, Elliptic and Mixed
PDEs. Springer, 2021, �10.1007/978-3-030-56923-5�. �hal-03226050�
HAL Id: hal-03226050

https://hal.science/hal-03226050
Submitted on 18 May 2021
HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents
entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non,
lished or not. The documents may come from émanant des établissements d’enseignement et de
teaching and research institutions in France or recherche français ou étrangers, des laboratoires
abroad, or from public or private research centers. publics ou privés.
Finite Elements II:
Galerkin approximation, elliptic and mixed PDEs
Alexandre Ern Jean-Luc Guermond
May 13, 2021

Contents
Part V. Weak formulations and well-posedness
24 Weak formulation of model problems 1

24.1 A second-order PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
24.2 A first-order PDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
24.3 A complex-valued model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
24.4 Toward an abstract model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
25 Main results on well-posedness 11

25.1 Mathematical setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
25.2 Lax–Milgram lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
25.3 Banach–Nečas–Babuška (BNB) theorem . . . . . . . . . . . . . . . . . . . . . . . . 14
25.4 Two examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Part VI. Galerkin approximation
26 Basic error analysis 21

26.1 The Galerkin method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
26.2 Discrete well-posedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
26.3 Basic error estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
27 Error analysis with variational crimes 33

27.1 Setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
27.2 Main results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
27.3 Two simple examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
27.4 Strang’s lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
28 Linear algebra 45
28.1 Stiffness and mass matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
28.2 Bounds on the stiffness and mass matrices . . . . . . . . . . . . . . . . . . . . . . . 47
28.3 Solution methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Contents
29 Sparse matrices 57
29.1 Origin of sparsity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
29.2 Storage and assembling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
29.3 Reordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
30 Quadratures 69
30.1 Definition and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
30.2 Quadrature error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
30.3 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Part VII. Elliptic PDEs: conforming approximation
31 Scalar second-order elliptic PDEs 79

31.1 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
31.2 Dirichlet boundary condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
31.3 Robin/Neumann conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
31.4 Elliptic regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
32 H 1 -conforming approximation (I) 93

32.1 Continuous and discrete problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
32.2 Error analysis and best approximation in H 1 . . . . . . . . . . . . . . . . . . . . . 94
32.3 L2 -error analysis: the duality argument . . . . . . . . . . . . . . . . . . . . . . . . 97
32.4 Elliptic projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
33 H 1 -conforming approximation (II) 103

33.1 Non-homogeneous Dirichlet conditions . . . . . . . . . . . . . . . . . . . . . . . . . 103
33.2 Discrete maximum principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
33.3 Discrete problem with quadratures . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
34 A posteriori error analysis 117

34.1 The residual and its dual norm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
34.2 Global upper bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
34.3 Local lower bound . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
34.4 Adaptivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
35 The Helmholtz problem 131

35.1 Robin boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
35.2 Mixed boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
35.3 Dirichlet boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
35.4 H 1 -conforming approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Part VIII. Elliptic PDEs: nonconforming approximation

Contents
36 Crouzeix–Raviart approximation 145

36.1 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
36.2 Crouzeix–Raviart discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
36.3 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
37 Nitsche’s boundary penalty method 159

37.1 Main ideas and discrete problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
37.2 Stability and well-posedness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
37.3 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
38 Discontinuous Galerkin 167

38.1 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
38.2 Symmetric interior penalty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
38.3 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
38.4 Discrete gradient and fluxes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
39 Hybrid high-order method 179

39.1 Local operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
39.2 Discrete problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
39.3 Error analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
40 Contrasted diffusivity (I) 193

40.1 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
40.2 Discrete setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
40.3 The bilinear form n♯ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
41 Contrasted diffusivity (II) 201

41.1 Continuous and discrete settings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
41.2 Crouzeix–Raviart approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
41.3 Nitsche’s boundary penalty method . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
41.4 Discontinuous Galerkin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 206
41.5 The hybrid high-order method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
Part IX. Vector-valued elliptic PDEs
42 Linear elasticity 211

42.1 Continuum mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
42.2 Weak formulation and well-posedness . . . . . . . . . . . . . . . . . . . . . . . . . . 213
42.3 H 1 -conforming approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
42.4 Further topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
43 Maxwell’s equations: H(curl)-approximation 225

43.1 Maxwell’s equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
43.2 Weak formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
43.3 Approximation using edge elements . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Contents
44 Maxwell’s equations: control on the divergence 233

44.1 Functional setting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
44.2 Coercivity revisited for edge elements . . . . . . . . . . . . . . . . . . . . . . . . . 235
44.3 The duality argument for edge elements . . . . . . . . . . . . . . . . . . . . . . . . 239
45 Maxwell’s equations: further topics 243

45.1 Model problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
45.2 Boundary penalty method in H(curl) . . . . . . . . . . . . . . . . . . . . . . . . . 244
45.3 Boundary penalty method in H 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249
45.4 H 1 -approximation with divergence control . . . . . . . . . . . . . . . . . . . . . . . 250
Part X. Eigenvalue problems
46 Symmetric elliptic eigenvalue problems 253

46.1 Spectral theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
46.2 Introductory examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
47 Symmetric operators, conforming approximation 265

47.1 Symmetric and coercive eigenvalue problems . . . . . . . . . . . . . . . . . . . . . . 265
47.2 H 1 -conforming approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
48 Nonsymmetric problems 277

48.1 Abstract theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
48.2 Conforming approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
48.3 Nonconforming approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Part XI. PDEs in mixed form
49 Well-posedness for PDEs in mixed form 287

49.1 Model problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
49.2 Well-posedness in Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
49.3 Saddle point problems in Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . 293
49.4 Babuška–Brezzi theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
50 Mixed finite element approximation 301

50.1 Conforming Galerkin approximation . . . . . . . . . . . . . . . . . . . . . . . . . . 301
50.2 Algebraic viewpoint . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
50.3 Iterative solvers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
51 Darcy’s equations 315

51.1 Weak mixed formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
51.2 Primal, dual, and dual mixed formulations . . . . . . . . . . . . . . . . . . . . . . . 319
51.3 Approximation of the mixed formulation . . . . . . . . . . . . . . . . . . . . . . . . 321
Contents
52 Potential and flux recovery 327

52.1 Hybridization of mixed finite elements . . . . . . . . . . . . . . . . . . . . . . . . . 327
52.2 Flux recovery for H 1 -conforming elements . . . . . . . . . . . . . . . . . . . . . . . 331
53 Stokes equations: Basic ideas 337

53.1 Incompressible fluid mechanics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
53.2 Weak formulation and well-posedness . . . . . . . . . . . . . . . . . . . . . . . . . . 338
53.3 Conforming approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
53.4 Classical examples of unstable pairs . . . . . . . . . . . . . . . . . . . . . . . . . . 346
54 Stokes equations: Stable pairs (I) 351

54.1 Proving the inf-sup condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351
54.2 Mini element: the (PP1 -bubble, P1 ) pair . . . . . . . . . . . . . . . . . . . . . . . . . 353
54.3 Taylor–Hood element: the (P P2 , P1 ) pair . . . . . . . . . . . . . . . . . . . . . . . . 356
54.4 Generalizations of the Taylor–Hood element . . . . . . . . . . . . . . . . . . . . . . 358
55 Stokes equations: Stable pairs (II) 363

55.1 Macroelement techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
55.2 Discontinuous pressures and bubbles . . . . . . . . . . . . . . . . . . . . . . . . . . 366
55.3 Scott–Vogelius elements and generalizations . . . . . . . . . . . . . . . . . . . . . . 368
55.4 Nonconforming and hybrid methods . . . . . . . . . . . . . . . . . . . . . . . . . . 371
55.5 Stable pairs with Q k -based velocities . . . . . . . . . . . . . . . . . . . . . . . . . . 374
Appendix
C Bijective operators in Banach spaces 377

C.1 Injection, surjection, bijection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
C.2 Banach spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
C.3 Hilbert spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
C.4 Duality, reflexivity, and adjoint operators . . . . . . . . . . . . . . . . . . . . . . . 380
C.5 Open mapping and closed range theorems . . . . . . . . . . . . . . . . . . . . . . . 383
C.6 Characterization of surjectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
C.7 Characterization of bijectivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 388
C.8 Coercive operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390
Chapter 24
Weak formulation of model

problems
In Part V, composed of Chapters 24 and 25, we introduce the notion of weak formulations and
state two well-posedness results: the Lax–Milgram lemma and the more fundamental Banach–
Nečas–Babuška theorem. Weak formulations are useful for building finite element approximations
to partial differential equations (PDEs). This chapter presents a step-by-step derivation of weak
formulations. We start by considering a few simple PDEs posed over a bounded subset D of Rd .
Our goal is to reformulate these problems in weak form using the important notion of test functions.
We show by examples that there are many ways to write weak formulations. Choosing one can be
guided, e.g., by the smoothness of the data and the quantities of interest (e.g., the solution or its
gradient). The reader who is not familiar with functional analysis arguments is invited to review
the four chapters composing Part I before reading Part V.
24.1 A second-order PDE

Let D be a Lipschitz domain in Rd (see §3.1) and consider a function f : D → R. The problem we
want to solve consists of seeking a function u : D → R with some appropriate smoothness yet to
be clearly defined such that
−∆u = f in D u = 0 on ∂D, (24.1)
where the Laplace operator is defined by ∆u := ∇·(∇u). In Cartesian coordinates, we have

P 2
∆u := i∈{1: d} ∂∂xu2 .
i
The PDE −∆u = f in D is called Poisson equation (and Laplace equation when f = 0). The
Laplace operator is ubiquitous in physics since it is the prototypical operator modelling diffusion
processes. Applications include heat transfer (where u is the temperature and f the heat source),
mass transfer (where u is the concentration of a species and f the mass source), porous media flow
(where u is the hydraulic head and f the mass source), electrostatics (where u is the electrostatic
potential and f the charge density), and static equilibria of membranes (where u is the transverse
membrane displacement and f the transverse load).
The condition enforced on ∂D in (24.1) is called boundary condition. A condition prescribing
the value of the solution at the boundary is called Dirichlet condition, and when the prescribed
2 Chapter 24. Weak formulation of model problems
value is zero, the condition is called homogeneous Dirichlet condition. In the context of the above
models, the Dirichlet condition means that the temperature (the concentration, the hydraulic
head, the electrostatic potential, or the transverse membrane displacement) is prescribed at the
boundary. Other boundary conditions can be prescribed for the Poisson equation, as reviewed in
Chapter 31 in the more general context of second-order elliptic PDEs.
To sum up, (24.1) is the Poisson equation (or problem) with a homogeneous Dirichlet condition.
We now present three weak formulations of (24.1).
24.1.1 First weak formulation

We derive a weak formulation of (24.1) by proceeding informally. Consider an arbitrary test
function ϕ ∈ C0∞ (D), where C0∞ (D) is the space of infinitely differentiable functions compactly
supported in D. As a first step, we multiply the PDE in (24.1) by ϕ and integrate over D to obtain
Z Z
− (∆u)ϕ dx = f ϕ dx. (24.2)
D D
Equation (24.2) is equivalent to the PDE in (24.1)

R if ∆u is smooth enough (e.g., integrable over
D). Indeed, if an integrable function g satisfies D gϕ dx = 0 for all ϕ ∈ C0∞ (D), Theorem 1.32
implies that g = 0 a.e. in D.
As a second step, we use the divergence formula stating that for any smooth vector-valued
function Φ, Z Z
∇·Φ dx = Φ·n ds, (24.3)
D ∂D
where n is the outward unit normal to D. We apply this formula to the function Φ := w∇v, where
v and w are two scalar-valued smooth functions. Since ∇·Φ = ∇w·∇v + w∆v, we infer that
Z Z Z
− (∆v)w dx = ∇v·∇w dx − (n·∇v)w ds. (24.4)
D D ∂D
This is Green’s formula, which is a very useful tool to derive weak formulations of PDEs involving
the Laplace operator. This formula is valid for instance if v ∈ C 2 (D) ∩ C 1 (D) and w ∈ C 1 (D) ∩
C 0 (D), and it can be extended to functions in the usual Sobolev spaces. In particular, it remains
valid for all v ∈ H 2 (D) and all w ∈ H 1 (D). We apply Green’s formula to the functions v := u and
w := ϕ, assuming enough smoothness for u. Since ϕ vanishes at the boundary, we transform (24.2)
into Z Z
∇u·∇ϕ dx = f ϕ dx, ∀ϕ ∈ C0∞ (D). (24.5)
D D
We now recast (24.5) into a functional framework. Let us take f ∈ L2 (D). We observe that a
natural solution space is
H 1 (D) := {v ∈ L2 (D) | ∇v ∈ L2 (D)}. (24.6)
1
Recall from Proposition
R 2.9 that HR (D) is a Hilbert space when equipped with the inner
R prod-
uct (u, v)H (D)
1 := D
uv dx + ℓD D ∇u·∇v dx with associated norm kvkH 1 (D) := ( D v 2 dx +
2
R 1
ℓ2D D k∇vk2ℓ2 dx) 2 , where k·kℓ2 denotes the Euclidean norm in Rd and ℓD is a length scale associ-
ated with the domain D, e.g., ℓD := diam(D) (one can take ℓD := 1 when working in nondimen-
sional form). In order to account for the boundary condition in (24.1), we consider the subspace
spanned by those functions in H 1 (D) that vanish at the boundary. It turns out that this space
is H01 (D); see Theorem 3.10. Finally, we can extend the space of the test functions in (24.5) to
the closure of C0∞ (D) in H 1 (D), which is by definition H01 (D) (see Definition 3.9). To see this,
Part V. Weak formulations and well-posedness 3
we consider any test function w ∈ H01 (D), observe that there is a sequence (ϕn )n∈N in C0∞ (D)
converging to w in H01 (D), and pass to the limit in (24.5) with ϕn used as the test function. To
sum up, a weak formulation of the Poisson equation with homogeneous Dirichlet condition is as
follows:
Find u ∈ V := H01 (D) such that
R R (24.7)
D
∇u·∇w dx = D f w dx, ∀w ∈ V.
A function u solving (24.7) is called weak solution to (24.1).
We now investigate whether a solution to (24.7) (i.e., a weak solution to (24.1)) satisfies the
PDE and the boundary condition in (24.1). Similarly to Definition 2.3, we say that a vector-valued
field σ ∈ L1loc (D) := L1loc (D; Rd ) has a weak divergence ψ ∈ L1loc (D) if
Z Z
σ·∇ϕ dx = − ψϕ dx, ∀ϕ ∈ C0∞ (D), (24.8)
D D
and we write ∇·σ := ψ. The argument of Lemma 2.4 shows that the weak divergence of a vector-
valued field, if it exists, is uniquely defined.
Proposition 24.1 (Weak solution). Assume that u solves (24.7) with f ∈ L2 (D). Then −∇u
has a weak divergence equal to f , the PDE in (24.1) is satisfied a.e. in D, and the boundary
condition a.e. in ∂D.
Proof. Let u be a weak solution. Then ∇u ∈ L2 (D) ⊂ L1loc (D). Taking as a test function in (24.7)
an arbitrary function ϕ ∈ C0∞ (D) ⊂ H01 (D) and observing that f ∈ L2 (D) ⊂ L1loc (D), we infer
from the definition (24.8) of the weak divergence that the vector-valued field σ := −∇u has a weak
divergence equal to f . Hence, the PDE is satisfied in the sense that −∇·(∇u) = f in L2 (D), i.e.,
both functions are equal a.e. in D. Since u ∈ H01 (D), u vanishes a.e. in ∂D owing to the trace
theorem (Theorem 3.10).
The crucial advantage of the weak formulation (24.7) with respect to the original formula-
tion (24.1) is that, as we will see in the next chapter, there exist powerful tools that allow us to
assert the existence and uniqueness of weak solutions. It is noteworthy that uniqueness is not a
trivial property in spaces larger than H 1 (D), and existence is nontrivial in spaces smaller than
H 1 (D). For instance, one can construct domains in which uniqueness does not hold in L2 (D), and
existence does not hold in H 2 (D); see Exercise 24.2.
24.1.2 Second weak formulation

To derive our second formulation, we introduce the vector-valued function σ := −∇u. To avoid
notational collisions, we use the letter p instead of u to denote the scalar-valued unknown function,
and we use the symbol u to denote the pair (σ, p). In many applications, p plays the role of a
potential and σ plays the role of a (diffusive) flux. More generally, p is called primal variable and
σ dual variable.
Since σ = −∇p and −∆p = f , we obtain ∇·σ = f . Therefore, the model problem is now
written as follows:
σ + ∇p = 0 in D, ∇·σ = f in D, p = 0 on ∂D. (24.9)
This is the mixed formulation of the original problem (24.1). The PDEs in (24.9) are often called
Darcy’s equations (in the context of porous media flows, p is the hydraulic head and σ the filtration
velocity).
We multiply the first PDE in (24.9) by a vector-valued test function τ and integrate over D to
obtain Z Z
σ·τ dx + ∇p·τ dx = 0. (24.10)
D D
We multiply the second PDE in (24.9) by a scalar-valued test function q and integrate over D to
obtain Z Z
(∇·σ)q dx = f q dx. (24.11)
D D
No integration by parts is performed in this approach.
We now specify a functional framework. We consider H 1 (D) as the solution space for p (so
that ∇p ∈ L2 (D) and p ∈ L2 (D)), and H(div; D) as the solution space for σ with kσkH(div;D) :=
1
(kσk2L2 (D) + ℓ2D k∇·σk2L2 (D) ) 2 (recall that ℓD is a characteristic length associated with D, e.g.,
ℓD := diam(D)). Moreover, we enforce the boundary condition explicitly by restricting p to be
in the space H01 (D). With this setting, the test function τ can be taken in L2 (D) and the test
function q in L2 (D). To sum up, a second weak formulation is as follows:

Find u := (σ, p) ∈ V such that
R R (24.12)
D (σ·τ + ∇p·τ + (∇·σ)q) dx = D f q dx, ∀w := (τ , q) ∈ W,
with the functional spaces V := H(div; D)×H01 (D) and W := L2 (D)×L2 (D). Note that the space
where the solution is expected to be (trial space) differs from the space where the test functions
are taken (test space).
Proposition 24.2 (Weak solution). Assume that u solves (24.12) with f ∈ L2 (D). Then the
PDEs in (24.9) are satisfied a.e. in D, and the boundary condition a.e. in ∂D.
Proof. Left as an exercise.
24.1.3 Third weak formulation

We start with the mixed formulation (24.9), and we now perform an integration by parts on the
term involving ∇·σ. Proceeding informally, we obtain
Z Z Z
− σ·∇q dx + (n·σ)q ds = f q dx. (24.13)
D ∂D D
We take the test function q in H 1 (D) for the first integral to make sense. Moreover, to eliminate
the boundary integral, we restrict q to be in the space H01 (D). Now the dual variable σ can be
taken in L2 (D). To sum up, a third weak formulation is as follows:

Find u := (σ, p) ∈ V such that
R R (24.14)
D (σ·τ + ∇p·τ + σ·∇q) dx = − D f q dx, ∀w := (τ , q) ∈ V,
with the same functional space V := L2 (D)×H01 (D) for the trial and test spaces. The change of
sign on the right-hand side has been introduced to make the left-hand side symmetric with respect
to (σ, p) and (τ , q).
Proposition 24.3. Let u solve (24.14) with f ∈ L2 (D). Then the PDEs in (24.9) are satisfied
a.e. in D, and the boundary condition a.e. in ∂D.
Proof. Left as an exercise.
24.2 A first-order PDE

For simplicity, we consider a one-dimensional model problem (a more general setting is covered in
Chapter 56). Let D := (0, 1) and let f : D → R be a smooth function. The problem we want to
solve consists of seeking a function u : D → R such that
u′ = f in D, u(0) = 0. (24.15)
Proceeding informally, the solution to this problem is the function defined as follows:
Z x
u(x) := f (t) dt, ∀x ∈ D. (24.16)
0
To give a precise mathematical meaning to this statement, we assume that f ∈ L1 (D), and we
introduce the Sobolev space (see Definition 2.8)
W 1,1 (D) := {v ∈ L1 (D) | v ′ ∈ L1 (D)}, (24.17)
where as usual we interpret the derivatives in the weak sense.

Lemma 24.4 (Solution in W 1,1 (D)). If f ∈ L1 (D), the problem (24.15) has a unique solution
in W 1,1 (D) which is given by (24.16).
Proof. Let u be defined in (24.16).
(1) Let us first show that u ∈ C 0 (D) (recall that D = [0, 1]). Let x ∈ D and let (xn )n∈N be a
sequence converging to x in D. This gives
Z x Z xn Z x Z
u(x) − u(xn ) = f (t) dt − f (t) dt = f (t) dt = 1[xn ,x] (t)f (t) dt,
0 0 xn D
where 1[xn ,x] is the indicator function of the interval [xn , x]. Since 1[xn ,x] f → 0 and |1[xn ,x] f | ≤ |f |
a.e. in D, Lebesgue’s dominated convergence theorem (Theorem 1.23) implies that u(xn ) → u(x).
This shows that u ∈ C 0 (D). Hence, the boundary condition u(0) = 0 is meaningful.
(2) Let us now prove that u′ = f a.e. in D. One can verify (see Exercise 24.7) that
Z 1 Z x Z 1
f (t) dt ϕ′ (x) dx = − f (x)ϕ(x) dx, ∀ϕ ∈ C0∞ (D). (24.18)
0 0 0
R1
Since the left-hand side is equal to 0 u(x)ϕ′ (x) dx and f ∈ L1 (D) ⊂ L1loc (D), we infer that u has
a weak derivative in L1loc (D) equal to f . This implies that the PDE in (24.15) is satisfied a.e. in
D.
(3) Uniqueness of the solution is a consequence of Lemma 2.11 since the difference of two weak
solutions is constant on D (since it has zero weak derivative) and vanishes at x = 0.
We now present two possible mathematical settings for the weak formulation of the prob-
lem (24.15).
24.2.1 Formulation in L1 (D)

Since f ∈ L1 (D) and u ∈ W 1,1 (D) with u(0) = 0, a first weak formulation is obtained by just
multiplying the PDE in (24.15) by a test function w and integrating over D:
Z Z
u′ w dt = f w dt. (24.19)
D D
This equality is meaningful for all w ∈ W (∞) := L∞ (D). Moreover, the boundary condition u(0) =
0 can be explicitly enforced by considering the solution space V (1) := {v ∈ W 1,1 (D) | v(0) = 0}.
Thus, a first weak formulation of (24.15) is as follows:
(
Find u ∈ V (1) such that
R ′ R (24.20)
D
u w dt = D f w dt, ∀w ∈ W (∞) .
Remark 24.5 (Literature). Solving first-order PDEs using L1 -based formulations has been
introduced by Lavery [276, 277]; see also Guermond [227], Guermond and Popov [228], and the
references therein.
24.2.2 Formulation in L2 (D)

Although the weak formulation (24.20) gives a well-posed problem (as we shall see in §25.4.2),
the dominant viewpoint in the literature consists of using L2 -based formulations. This leads us to
consider a second weak formulation where the source term f has slightly more smoothness, i.e.,
f ∈ L2 (D) instead of just f ∈ L1 (D), thereby allowing us to work in a Hilbertian setting. Since
L2 (D) ⊂ L1 (D), we have f ∈ L1 (D), and we can still consider the function u defined in (24.16).
This function turns out to be in H 1 (D) if f ∈ L2 (D). Indeed, the Cauchy–Schwarz inequality and
Fubini’s theorem imply that
Z 1 Z 1 Z x 2 Z 1 Z x

|u(x)|2 dx = f (t) dt dx ≤ |f (t)| 2
dt x dx

0 0 0 0 0
Z 1 Z 1 Z 1 Z 1
2 2
= dx |f (t)| dt = (1 − t)|f (t)| dt ≤ |f (t)|2 dt,
0 t 0 0
′
which shows that kukL2(D) ≤ kf kL2(D) . Moreover, ku kL2 (D) = kf kL2(D) . Hence, u ∈ H 1 (D). We
can then restrict the test functions to the Hilbert space W (2) := L2 (D) and use the Hilbert space
V (2) := {v ∈ H 1 (D) | v(0) = 0} as the solution space. Thus, a second weak formulation of (24.20),
provided f ∈ L2 (D), is as follows:
(
Find u ∈ V (2) such that
R ′ R (24.21)
D
u w dt = D f w dt, ∀w ∈ W (2) .
The main change with respect to (24.20) is in the trial and test spaces.
24.3 A complex-valued model problem

Some model problems are formulated using complex-valued functions. A salient example is Maxwell’s
equations in the time-harmonic regime; see §43.1. For simplicity, let us consider here the PDE
iu − ν∆u = f in D, (24.22)
2
with u : D → C, f : D → C, i = −1, and a real number ν > 0. To fix the ideas, we enforce a
homogeneous Dirichlet condition on u at the boundary.
When working with complex-valued functions, one uses the complex conjugate of the test
function in the weak problem, i.e., the starting point of the weak formulation is the identity
Z Z Z
iuw dx + ν ∇u·∇w dx = f w dx. (24.23)
D D D
One can then proceed as in §24.1.1 (for instance). The functional setting uses the functional space
V := H01 (D; C), and the weak formulation is as follows:

Find u ∈ V such that
R R R (24.24)
D
iuw dx + ν D ∇u·∇w dx = D f w dx, ∀w ∈ V.
Proposition 24.1 is readily adapted to this setting.

The reason for using the complex conjugate of test functions is that itR allows us to infer
R positivity
properties on the real and imaginary parts of the quantity a(u, w) := D iuw dx + ν D ∇u·∇w dx
by taking w := u as the test function. Indeed, we obtain
Z Z
a(u, u) = i |u|2 dx + ν k∇uk2ℓ2 (Cd ) dx = ikuk2L2 (D;C) + νk∇uk2L2 (D;Cd ) .
D D
This means that ℜ(a(u, u)) = νk∇uk2L2 (D;Cd ) and ℑ(a(u, u)) = kuk2L2 (D;C) . These results imply
that
π 1
ℜ(e−i 4 a(u, u)) ≥ √ min(1, νℓ−2 2
D )kukH 1 (D;C) , (24.25)
2
Rwhere we recall that the Hilbert space L2 (D; C) is equipped with the inner product (v, w)L2 (D) :=
1
R Hilbert space H (D; C) is equipped with the inner product (v, w)H 1 (D) :=
RD vw dx and2 the
D vw dx + ℓD D ∇v·∇w dx, where ℓD is a characteristic length associated with D, e.g., ℓD
:=
diam(D).
24.4 Toward an abstract model problem

We conclude this chapter by casting all of the above weak formulations into a unified setting. We
consider complex-valued functions since it is in general simpler to go from complex to real numbers
than the other way around. Whenever relevant, we indicate the (minor) changes to apply in this
situation (apart from replacing C by R).
The above weak formulations fit into the following abstract model problem:

(24.26)
a(u, w) = ℓ(w), ∀w ∈ W,
with maps a : V ×W → C and ℓ : W → C, where V, W are complex vector spaces whose elements
are functions defined on D. V is called trial space or solution space, and W is called test space.
Members of V are called trial functions and members of W are called test functions. The maps a
and ℓ are called forms since their codomain is C (or R in the real case).
Recall that a map A : V → C is said to be linear if A(v1 + v2 ) = A(v1 ) + A(v2 ) for all v1 , v2 ∈ V
and A(λv) = λA(v) for all λ ∈ C and all v ∈ V, whereas a map B : W → C is said to be antilinear
if B(w1 + w2 ) = B(w1 ) + B(w2 ) for all w1 , w2 ∈ W and B(λw) = λB(w) for all λ ∈ C and all
w ∈ W. Then ℓ in (24.26) is an antilinear form, whereas a is a sesquilinear form (that is, the map
a(·, w) is linear for all w ∈ W, and the map a(v, ·) is antilinear for all v ∈ V ). In the real case, ℓ is
a linear form and a is a bilinear form (that is, it is linear in each of its arguments).
Remark 24.6 (Linearity). The linearity of a w.r.t. to its first argument is a consequence of the
linearity of the problem, whereas the (anti)linearity of a w.r.t. its second argument results from
the weak formulation.
Remark 24.7 (Bilinearity). Bilinear forms and linear forms on V ×W are different objects. For
instance, the action of a linear form on (v, 0) ∈ V ×W is not necessarily zero, whereas a(v, 0) = 0
if a is a bilinear form.
Remark 24.8 (Test functions). The role of the test functions in the weak formulations (24.20)
and (24.26) are somewhat different. Since L∞ (D) is the dual space of L1 (D) (the reverse is not
true), the test functions w ∈ L∞ (D) in (24.20) act on the function f ∈ L1 (D). Hence, in principle
it should be more appropriate to write w(ℓ) instead of ℓ(w) in (24.26). Although this alternative
viewpoint is not often considered in the literature, it actually allows for a more general setting
regarding well-posedness. We return to this point in §25.3.2. This distinction is not relevant for
model problems set in a Hilbertian framework.
Exercises
Exercise 24.1 (Forms). RLet D := (0, 1). Which of these R maps are linear or bilinear
R forms on
L2 (D)×L2 (D):
R a 1 (f, g) := D (f +g +1) dx, a 2 (f, g) := D x(f −g) dx, a 3 (f, g) := D (1+x2
)f g dx,
2
a4 (f, g) := D (f + g) dx?
Exercise 24.2 ((Non)-uniqueness). Consider the domain D in R2 whose definition in polar

coordinates is D := {(r, θ) | r ∈ (0, 1), θ ∈ ( πα , 0)} with α ∈ (−1, − 21 ). Let ∂D1 := {(r, θ) | r =
π
1, θ ∈ ( α , 0)} and ∂D2 := ∂D\∂D1 . Consider the PDE −∆u = 0 in D with the Dirichlet conditions
u = sin(αθ) on ∂D1 and u = 0 on ∂D2 . (i) Let ϕ1 := rα sin(αθ) and ϕ2 := r−α sin(αθ). Prove
that ϕ1 and ϕ2 solve the above problem. (Hint : in polar coordinates ∆ϕ = r1 ∂r (r∂r ϕ) + r12 ∂θθ ϕ.)
(ii) Prove that ϕ1 and ϕ2 are in L2 (D) if α ∈ (−1, − 21 ). R (iii) Consider the problem of seeking
u ∈ H 1 (D) s.t. u = sin(αθ) on ∂D1 , u = 0 on ∂D2 , and D ∇u·∇v = 0 for all v ∈ H01 (D). Prove
that ϕ2 solves this problem, but ϕ1 does not. Comment.
1
Exercise 24.3 (Poisson in 1D). Let D := (0, 1) and f (x) := x(1−x) . Consider the PDE
2
−∂x ((1 + sin(x) )∂x u) = f in D with the Dirichlet conditions u(0) = u(1) = 0. Write a weak
formulation of this problem with both trial and test spaces equal to H01 (D) and show that the
linear form on the right-hand side is bounded on H01 (D). (Hint : notice that f (x) = x1 + 1−x
1
.)
Exercise 24.4 (Weak formulations). Prove Propositions 24.2 and 24.3.
Exercise 24.5 (Darcy). (i) Derive another variation on (24.12) and (24.14) with the functional
spaces V = W := H(div; D)×L2 (D). (Hint : use Theorem 4.15.) (ii) Derive yet another variation
with the functional spaces V := L2 (D)×L2 (D) and W := H(div; D)×H01 (D).
Exercise 24.6 (Variational formulation). Prove that u solves (24.7) if and only if u minimizes
over H01 (D) the energy functional
Z Z
1 2
E(v) := |∇v| dx − f v dx.
2 D D
R R R
(Hint : show first that E(v + tw) = E(v) + t D ∇v·∇w dx − D f w dx + 12 t2 D |∇w|2 dx for all
v, w ∈ H01 (D) and all t ∈ R.)
Exercise 24.7 (Derivative of primitive). Prove (24.18). (Hint : use Theorem 1.38 and Lebesgue’s
dominated convergence theorem.)
Exercise 24.8 (Biharmonic problem). Let D be an open, bounded, set in Rd with smooth
boundary. Derive a weak formulation for the biharmonic problem
∆(∆u) = f in D, u = ∂n u = 0 on ∂D,
with f ∈ L2 (D). (Hint : use Theorem 3.16.)

Chapter 25
Main results on well-posedness
The starting point of this chapter is the model problem derived in §24.4. Our goal is to specify
conditions under which this problem is well-posed. Two important results are presented: the Lax–
Milgram lemma and the more fundamental Banach–Nečas–Babuška theorem. The former provides
a sufficient condition for well-posedness, whereas the latter, relying on slightly more sophisticated
assumptions, provides necessary and sufficient conditions. The reader is invited to review the
material of Appendix C on bijective operators in Banach spaces before reading this chapter.
25.1 Mathematical setting

To stay general, we consider complex vector spaces. The case of real vector spaces is recovered
by replacing the field C by R, by removing the real part symbol ℜ(·) and the complex conjugate
symbol ·, and by interpreting the symbol |·| as the absolute value instead of the modulus.
We consider the following model problem:

(25.1)
a(u, w) = ℓ(w), ∀w ∈ W.
The spaces V and W are complex Banach spaces equipped with norms denoted by k·kV and k·kW ,
respectively. In many applications, V and W are Hilbert spaces. The map a : V ×W → C is a
sesquilinear form (bilinear in the real case). We assume that a is bounded, which means that
|a(v, w)|
kakV ×W := sup sup < ∞. (25.2)
v∈V w∈W kvkV kwkW
It is henceforth implicitly understood that this type of supremum is taken over nonzero arguments
(notice that the order in which the suprema are taken in (25.2) does not matter). Furthermore,
the map ℓ : W → C is an antilinear form (linear in the real case). We assume that ℓ is bounded,
and we write ℓ ∈ W ′ . The boundedness of ℓ means that
|ℓ(w)|
kℓkW ′ := sup < ∞. (25.3)
w∈W kwkW
Notice that it is possible to replace the modulus by the real part in (25.2) and (25.3) (replace w
by ξw with a unitary complex number ξ), and in the real case, the absolute value is not needed
(replace w by ±w).
12 Chapter 25. Main results on well-posedness
Definition 25.1 (Well-posedness, Hadamard [236]). We say that the problem (25.1) is well-
posed if it admits one and only one solution for all ℓ ∈ W ′ , and there is c, uniform with respect to
ℓ, s.t. the a priori estimate kukV ≤ c kℓkW ′ holds true.
The goal of this chapter is to study the well-posedness of (25.1). The key idea is to introduce
the bounded linear operator A ∈ L(V ; W ′ ) that is naturally associated with the bilinear form a on
V ×W by setting
hA(v), wiW ′ ,W := a(v, w), ∀(v, w) ∈ V ×W. (25.4)
This definition implies that A is linear and bounded with norm kAkL(V ;W ′ ) = kakV ×W . The
problem (25.1) can be reformulated as follows: Find u ∈ V such that A(u) = ℓ in W ′ . Hence,
proving the existence and uniqueness of the solution to (25.1) amounts to proving that the operator
A is bijective. Letting A∗ : W ′′ → V ′ be the adjoint of A, the way to do this is to prove the following
three conditions:
⇐⇒ A is surjective
z }| {
(i) A is injective, (ii) im(A) is closed, (iii) A∗ is injective . (25.5)
| {z }
⇐⇒ ∃α>0, kA(v)kW ′ ≥αkvkV , ∀v∈V
The conditions (ii)-(iii) in (25.5) are equivalent to A being surjective since the closure of im(A)
is (ker(A∗ ))⊥ ⊂ W ′ owing to Lemma C.34 (see also (C.14b)). That the conditions (i)-(ii) are
equivalent to the existence of some α > 0 s.t. kA(v)kW ′ ≥ αkvkV , for all v ∈ V, is established in
Lemma C.39 (these two conditions are also equivalent to the surjectivity of A∗ ).
25.2 Lax–Milgram lemma

The Lax–Milgram lemma is applicable only if the solution and the test spaces are identical. As-
suming W = V, the model problem (25.1) becomes

(25.6)
a(u, w) = ℓ(w), ∀w ∈ V.
Lemma 25.2 (Lax–Milgram). Let V be a Hilbert space, let a be a bounded sesquilinear form on
V ×V, and let ℓ ∈ V ′ . Assume the following coercivity property: There is a real number α > 0 and
a complex number ξ with |ξ| = 1 such that
ℜ (ξa(v, v)) ≥ αkvk2V , ∀v ∈ V. (25.7)

1
Then (25.6) is well-posed with the a priori estimate kukV ≤ α kℓkV .
′
Proof. Although this lemma is a consequence of the more abstract BNB theorem (Theorem 25.9),
we present a direct proof for completeness. Let A : V → V ′ be the bounded linear operator defined
in (25.4) and let us prove the three conditions (i)-(ii)-(iii) in (25.5). Since ξa(v, v) = a(v, ξv), the
coercivity property (25.7) implies that
ℜ(a(v, ξv)) ℜ(a(v, ξw)) |a(v, w)|

αkvkV ≤ ≤ sup = sup = kA(v)kV ′ ,
kvkV w∈V kwkV w∈V kwkV
so that the conditions (i)-(ii) hold true. Since V is reflexive, we identify V and V ′′ , so that the
adjoint operator A∗ : V → V ′ is such that hA∗ (v), wiV ′ ,V = hA(w), viV ′ ,V for all v, w ∈ V. Let
v ∈ V and assume that A∗ (v) = 0. Then 0 = 0 = hA∗ (v), ξviV ′ ,V = ξa(v, v). We then infer
from (25.7) that αkvk2V ≤ ℜ (ξa(v, v)) = 0, i.e., v = 0. This proves that A∗ is injective. Hence,
ℜ(a(u,ξu))
the condition (iii) also holds true. Finally, the a priori estimate follows from αkukV ≤ kukV =
ℜ(ℓ(ξu))
kukV ≤ kℓkV ′ .
Remark 25.3 (Hilbertian setting). An important observation is that the Lax–Milgram lemma
relies on the notion of coercivity which is applicable only in Hilbertian settings; see Proposi-
tion C.59.
Example 25.4 (Laplacian). Consider the weak formulation (24.7) of the Poisson equation with
homogeneous Dirichlet condition. The functional Rsetting is V = W := H01 (D) equipped with
Rthe norm k·kH 1 (D) , the bilinear form is a(v, w) := D ∇v·∇w dx, and the linear form is ℓ(w) :=
D
f w dx. Owing to the Cauchy–Schwarz inequality, the forms a and ℓ are bounded on V ×V and
V, respectively. Moreover, the Poincaré–Steklov inequality (3.11) (with p := 2) implies that (see
Remark 3.29)
2
Cps
a(v, v) = k∇vk2L2 (D) = |v|2H 1 (D) ≥ ℓ−2
D 2
kvk2H 1 (D) ,
1 + Cps
2
Cps
for all v ∈ V. Hence, (25.7) holds true with α := ℓ−2
D 1+Cps 2 and ξ := 1, and by the Lax–Milgram
lemma, the problem (24.7) is well-posed. Alternatively one can equip V with the norm kvkV :=
−1
ℓD k∇vkL2 (D) which is equivalent to the norm k·kH 1 (D) owing to the Poincaré–Steklov inequality.
The coercivity constant of a is then α := ℓ−2
D .
Example 25.5 (Complex case). Consider the PDE iu − ν∆u = f in D with i2 = −1, a real
number ν > 0, a source term f ∈ L2 (D; C), and a homogeneous Dirichlet condition. The functional
1
R = W := H
setting is V R 0 (D; C) equipped with the norm k·kH 1 (D;C) , the
R sesquilinear form is
a(v, w) := D ivw dx + ν D ∇v·∇w dx, and the antilinear form is ℓ(w) := D f w dx. Then (24.25)
π
shows that the coercivity property (25.7) holds true with ξ := e−i 4 and α := √12 min(1, νℓ−2
D ).
Remark 25.6 (Definition of coercivity). The coercivity property can also be defined in the
following way: There is a real number α > 0 such that |a(v, v)| ≥ αkvk2V for all v ∈ V. It is shown
in Lemma C.58 that this definition and (25.7) are equivalent.
Definition 25.7 (Hermitian/symmetric form). Let V be a Hilbert space. In the complex case,
we say that a sesquilinear form a : V ×V → C is Hermitian whenever a(v, w) = a(w, v) for all
v, w ∈ V. In the real case, we say that a bilinear form a is symmetric whenever a(v, w) = a(w, v)
for all v, w ∈ V.
Whenever the sesquilinear form a is Hermitian and coercive (with ξ := 1 for simplicity), setting
((·, ·))V := a(·, ·) one defines an inner product in V, and the induced norm is equivalent to k·kV
owing to the coercivity and the boundedness of a. Then solving the problem (25.6) amounts to
finding the representative u ∈ V of the linear form ℓ ∈ V ′ , i.e., ((u, w))V = ℓ(w) for all w ∈ V. This
problem is well-posed by the Riesz–Fréchet theorem (Theorem C.24). Thus, the Lax–Milgram
lemma can be viewed as an extension of the Riesz–Fréchet theorem to non-Hermitian forms.
Whenever V is a real Hilbert space and the bilinear form a is symmetric and coercive with
ξ := 1, the problem (25.6) can be interpreted as a minimization problem (or a maximization
problem if ξ := −1). In this context, (25.6) is called variational formulation.
Proposition 25.8 (Variational formulation). Let V be a real Hilbert space, let a be a bounded
bilinear form on V ×V, and let ℓ ∈ V ′ . Assume that a is coercive with ξ := 1. Assume that a is
symmetric, i.e.,
a(v, w) = a(w, v), ∀v, w ∈ V. (25.8)
Then introducing the energy functional E : V → R such that

1
E(v) := a(v, v) − ℓ(v), (25.9)
2
u solves (25.6) iff u minimizes E over V.
Proof. The proof relies on the fact that for all u, w ∈ V and all t ∈ R,
1
E(u + tw) = E(u) + t(a(u, w) − ℓ(w)) + t2 a(w, w), (25.10)
2
which results from the symmetry of a. (i) Assume that u solves (25.6). Since a(w, w) ≥ 0 owing to
the coercivity of a with ξ := 1, (25.10) implies that u minimizes E over V. (ii) Conversely, assume
that u minimizes E over V. The right-hand side of (25.10) is a quadratic polynomial in t reaching
its minimum value at t = 0. Hence, the derivative of this polynomial vanishes at t = 0, which
amounts to a(u, w) − ℓ(w) = 0. Since w is arbitrary in V, we conclude that u solves (25.6).
25.3 Banach–Nečas–Babuška (BNB) theorem

The BNB theorem plays a fundamental role in this book. We use this terminology since, to our
knowledge, the BNB theorem was stated by Nečas in 1962 [310] and Babuška in 1970 in the
context of finite element methods [33]. From a functional analysis point of view, the BNB theorem
is a rephrasing of two fundamental results by Banach: the closed range theorem and the open
mapping theorem. We present two settings for the BNB theorem depending on whether the test
functions in the model problem belong to a reflexive Banach space or to the dual of a Banach
space. Recall from Definition C.18 that a Banach space W is said to be reflexive if the canonical
isometry JW : W → W ′′ is an isomorphism. This is always the case if W is a Hilbert space.
25.3.1 Test functions in reflexive Banach space

Theorem 25.9 (Banach–Nečas–Babuška (BNB)). Let V be a Banach space and let W be a
reflexive Banach space. Let a be a bounded sesquilinear form on V ×W and let ℓ ∈ W ′ . Then the
problem (25.1) is well-posed iff:
|a(v, w)|
(bnb1) inf sup =: α > 0, (25.11a)
v∈V w∈W kvkV kwkW
(bnb2) ∀w ∈ W, [ ∀v ∈ V, a(v, w) = 0 ] =⇒ [ w = 0 ]. (25.11b)
(It is implicitly understood that the argument is nonzero in the above infimum and supremum.)
Moreover, we have the a priori estimate kukV ≤ α1 kℓkW ′ .
Proof. Let A ∈ L(V ; W ′ ) be defined by (25.4) and let us prove that the three conditions (i)-(ii)-(iii)
in (25.5) are equivalent to (bnb1)-(bnb2). The conditions (i)-(ii) are equivalent to (bnb1) since
for all v ∈ V,
|hA(v), wiW ′ ,W | |a(v, w)|
kAvkW ′ = sup = sup .
w∈W kwk W w∈W kwkW
Since hA∗ (JW (w)), viV ′ ,V = hJW (w), A(v)iW ′′ ,W ′ = hA(v), wiW ′ ,W = a(v, w) for all (v, w) ∈
V ×W, stating that a(v, w) = 0 for all v ∈ V is equivalent to stating that (A∗ ◦JW )(w) = 0. Hence,
(bnb2) is equivalent to stating that A∗ ◦JW is injective. Furthermore, since W is reflexive, the
canonical isometry JW : W → W ′′ from Proposition C.17 is an isomorphism. Hence, (bnb2) is
equivalent to stating that A∗ : W ′′ → V ′ is injective, which is the condition (iii) in (25.5). Finally,
the a priori estimate follows from the inequalities αkukV ≤ supw∈W |a(u,w)|kwkW = supw∈W kwkW =
|ℓ(w)|
kℓkW ′ .
Remark 25.10 ((bnb1)). Condition (bnb1) is called inf-sup condition and it is equivalent to the
following statement:
|a(v, w)|
∃α > 0, αkvkV ≤ sup , ∀v ∈ V. (25.12)
w∈W kwkW
Establishing (25.12) is usually done by finding two positive real numbers c1 , c2 s.t. for all v ∈ V,
one can find a “partner” wv ∈ W s.t. kwv kW ≤ c1 kvkV and |a(v, wv )| ≥ c2 kvk2V . If this is indeed
the case, then (25.12) holds true with α := cc12 . Establishing coercivity amounts to asserting that
wv = ζv is a suitable partner for some ζ ∈ C with |ζ| = 1.
Remark 25.11 ((bnb2)). The statement in (bnb2) is equivalent to asserting that for all w in
W, either there exists v in V such that a(v, w) 6= 0 or w = 0. In view of the proof Theorem 25.9,
(bnb2) says that for all w in W, either A∗ ◦JW (w) 6= 0 or w = 0.
Remark 25.12 (Two-sided bound). Since kℓkW ′ = kA(u)kW ′ ≤ ωkukV where ω := kakV ×W =
kAkL(V ;W ′ ) is the boundedness constant of the sesquilinear form a on V ×W, we infer the two-sided
bound
1 1
kℓkW ′ ≤ kukV ≤ kℓkW ′ .
kakV ×W α
Since α−1 = kA−1 kL(W ′ ;V ) owing to Lemma C.51, the quantity
kakV ×W
κ(a) = = kAkL(V ;W ′ ) kA−1 kL(W ′ ;V ) ≥ 1
α
can be viewed as the condition number of the sesquilinear form a (or of the associated operator
A). A similar notion of conditioning is developed for matrices in §28.2.1.
Remark 25.13 (Link with Lax–Milgram). Let V be a Hilbert space and let a be a bounded
and coercive bilinear form on V ×V. The proof of the Lax–Milgram lemma shows that a satisfies
the conditions (bnb1) and (bnb2) (with W = V ). The converse is false: the conditions (bnb1) and
(bnb2) do not imply coercivity. Hence, (25.7) is not necessary for well-posedness, whereas (bnb1)-
(bnb2) are necessary and sufficient. However, coercivity is both necessary and sufficient for well-
posedness when the bilinear form a is Hermitian and positive semidefinite; see Exercise 25.7.
Remark 25.14 (T -coercivity). Let V, W be Hilbert spaces. Then (bnb1)-(bnb2) are equivalent
to the existence of a bijective operator T ∈ L(V ; W ) and a positive real number η such that
ℜ(a(v, T (v))) ≥ ηkvk2V , ∀v ∈ V.
This property is called T -coercivity in Bonnet-Ben Dhia et al. [72, 73]; see Exercise 25.10. The
advantage of this notion over coercivity is the possibility of treating different trial and test spaces
and using a test function different from v ∈ V to estimate kvk2V . Note that the bilinear form
(u, v) 7→ a(u, T (v)) is bounded and coercive on V ×V. Proposition C.59 then implies that V is
necessarily a Hilbert space. This argument proves that T -coercivity is a notion relevant in Hilbert
spaces only. The BNB theorem is more general than T -coercivity since it also applies to Banach
spaces.
25.3.2 Test functions in dual Banach space

The requirement on the reflexivity of the space W in the BNB theorem can be removed if the model
problem is reformulated in such a way that the test functions act on the problem data instead of
the data acting on the test functions. Assume that we are given a bounded operator A ∈ L(V ; W )
and some data f ∈ W, and we want to assert that there is a unique u ∈ V s.t. A(u) = f . To
recast this problem in the general setting of (25.1) using test functions, we define the bounded
sesquilinear form on V ×W ′ such that
a(v, w′ ) := hw′ , A(v)iW ′ ,W , ∀(v, w′ ) ∈ V ×W ′ , (25.13)
and we consider the following model problem:
(
(25.14)
a(u, w′ ) = hw′ , f iW ′ ,W , ∀w′ ∈ W ′ .
Then u ∈ V solves (25.14) iff hw′ , A(u) − f iW ′ ,W = 0 for all w′ ∈ W ′ , that is, iff A(u) = f .
In (25.14), the data is f is in W and the test functions belong to W ′ , whereas in the original model
problem (25.1) the data is ℓ ∈ W ′ and the test functions belong to W. The functional setting
of (25.14) is useful, e.g., when considering first-order PDEs; see §24.2.1.
Theorem 25.15 (Banach–Nečas–Babuška (BNB)). Let V, W be Banach spaces. Let A ∈
L(V ; W ) and let f ∈ W. Let a be the bounded sesquilinear form on V ×W ′ defined in (25.13). The
problem (25.14) is well-posed iff:
|a(v, w′ )|
(bnb1’) inf sup := α > 0, (25.15)
v∈V w ′ ∈W ′ kvkV kw′ kW ′
′ ′
(bnb2’) ∀w ∈ W , [ ∀v ∈ V, a(v, w′ ) = 0 ] =⇒ [ w′ = 0 ]. (25.16)
1
Moreover, we have the a priori estimate kukV ≤ α kf kW .
Proof. The well-posedness of (25.14) is equivalent to the bijectivity of A : V → W, and this

property is equivalent to the three conditions (i)-(ii)-(iii) in (25.5) with W in lieu of W ′ and
|hw ′ ,A(v)i ′ |
A∗ : W ′ → V ′ . Since kA(v)kW = supw′ ∈W ′ W ,W
kw ′ kW ′ owing to Corollary C.14, the condition
(bnb1’) means that kA(v)kW ≥ αkvkV for all v ∈ V. This condition is therefore equivalent to the
conditions (i)-(ii). Moreover, since a(v, w′ ) = hw′ , A(v)iW ′ ,W = hA∗ (w′ ), viV ′ ,V , (bnb2’) amounts
to the condition (iii) (i.e., the injectivity of A∗ ).
Remark 25.16 (A vs. a). In the first version of the BNB theorem (Theorem 25.9), it is equivalent
to assume that we are given an operator A ∈ L(V ; W ′ ) or a bounded sesquilinear form a on V ×W.
But, in the second version of the BNB theorem (Theorem 25.15), we are given an operator A ∈
L(V ; W ), and the bounded sesquilinear form a on V ×W ′ is defined from A. If we were given instead
a bounded sesquilinear form a on V ×W ′ , proceeding as in (25.4) would be awkward since it would
lead to an operator Ã ∈ L(V ; W ′′ ) s.t. hÃ(v), w′ iW ′′ ,W ′ := a(v, w′ ) for all (v, w′ ) ∈ V ×W ′ .
Remark 25.17 (Literature). Inf-sup conditions in nonreflexive Banach spaces are discussed in
Amrouche and Ratsimahalo [9].
25.4 Two examples

In this section, we present two examples illustrating the above abstract results.
25.4.1 Darcy’s equations

The weak formulation (24.12) fits the setting of the model problem (25.1) with
V := H(div; D)×H01 (D), W := L2 (D)×L2 (D),

1
where kσkH(div;D) := (kσk2L2 (D) + ℓ2D k∇·σk2L2 (D) ) 2 (recall that ℓD is a characteristic length scale
associated with D, e.g., ℓD := diam(D)), and with the bilinear and linear forms
Z Z

a(v, w) := σ·τ + ∇p·τ + (∇·σ)q dx, ℓ(w) := f q dx, (25.17)
D D
with v := (σ, p) ∈ V and w := (τ , q) ∈ W.
Proposition 25.18. Problem (24.12) is well-posed.

1
Proof. We equip the Hilbert spaces V and W with the norms kvkV := kσk2H(div;D) + |p|2H 1 (D) 2
21
and kwkW := kτ k2L2 (D) + ℓ−2 2
D kqkL2 (D) with v := (σ, p) and w := (τ , q), respectively. That k·kV
is indeed a norm follows from the Poincaré–Steklov inequality (3.11) (see Remark 3.29). Since the
bilinear form a and the linear form ℓ are obviously bounded, it remains to check the conditions
(bnb1) and (bnb2).
(1) Proof of (bnb1). Let (σ, p) ∈ V and define S := sup(τ ,q)∈W |a((σ,p),(τ
k(τ ,q)kW
,q))|
. Since V ⊂ W, we
can take (σ, p) as the test function. Since p vanishes at the boundary, a((σ, p), (σ, p)) = kσk2L2 (D) ,
whence we infer that
a((σ, p), (σ, p))
kσk2L2 (D) = k(σ, p)kW ≤ S k(σ, p)kW .
k(σ, p)kW
−1
Since k·kW ≤ γk·kV on V with γ := max(1, Cps ), we infer that kσk2L2 (D) ≤ γ S k(σ, p)kV . More-
over, we have
12 R
2 2 2 | D {∇p·τ + (∇·σ)q} dx|
k∇pkL2 (D) + ℓD k∇·σkL2 (D) = sup
(τ ,q)∈W k(τ , q)kW
R
|a((σ, p), (τ , q))| | D σ·τ dx|
≤ sup + sup .
(τ ,q)∈W k(τ , q)kW (τ ,q)∈W k(τ , q)kW
1
Hence, k∇pk2L2(D) + ℓ2D k∇·σk2L2 (D) 2 ≤ S + kσkL2(D) . Squaring this inequality and combining it
with the above bound on kσkL2 (D) , we infer that
k(σ, p)k2V = k∇pk2L2 (D) + kσk2H(div;D) ≤ 2S2 + 3kσk2L2(D) ≤ 2S2 + 3γ S k(σ, p)kV .
1
Hence, the inf-sup condition (bnb1) holds true with α ≥ (4 + 9γ 2 )− 2 .
(2) Proof of (bnb2). Let (τ ,Rq) ∈ W be such that a((σ, p), (τ , q)) = 0 for all (σ, p) ∈ V. This
1
means on the one hand R that D ∇p·τ dx = 0 for all p ∈ H0 (D), so that ∇·τ = 0. ∞On the other
hand we obtain that D {σ·τ + (∇·σ)q} dx = 0 for all σ ∈ H(div; D). Taking σ ∈ C0 (D) we infer
that Rq ∈ H 1 (D) and ∇q = τ . Observing that τ ∈ H(div; D) and taking σ := τ , we infer that
0 = D {τ ·τ + (∇·τ )q} dx = kτ k2L2 (D) since ∇·τ = 0. Hence, τ = 0. Finally, ∇q = τ = 0, which
R
implies that q is constant on D. Since D (∇·σ)q dx = 0 for all σ ∈ H(div; D), q is identically zero
in D (take for instance σ(x) := x).
25.4.2 First-order PDE

Consider the weak formulation (24.20) on D := (0, 1). This formulation fits the setting of the
model problem (25.14) with the spaces
V := {v ∈ W 1,1 (D) | v(0) = 0}, W := L1 (D). (25.18)
The data is f ∈ W and we consider the bounded operator A : V → W s.t. A(v) := dv dt for all v ∈ V.
d
(Here, we denote derivatives by dt and reserve the primes to duality.) Recalling that W ′ = L∞ (D),
the bilinear form a associated with the operator A is s.t.
Z 1
′ dv ′
a(v, w ) := w dt, ∀(v, w′ ) ∈ V ×W ′ , (25.19)
0 dt
R1
and the right-hand side is hw′ , f iW ′ ,W := 0 w′ f dt with f ∈ W.
Proposition 25.19. Problem (24.20) is well-posed.
Proof. We equip the Banach spaces V and W ′ with the norms kvkV := kvkL1 (D) + k dv dt kL (D) and
1
kw′ kW ′ := kw′ kL∞ (D) , and we verify the conditions (bnb1’) and (bnb2’) from Theorem 25.15.
(1) Proof of (bnb1’). Let v ∈ V and set D± := {t ∈ D | ± dv dt (t) > 0}. Taking wv
′ :=
1D+ − 1D− ,
where 1S denotes the indicator function of a measurable set S, we infer that
R1 Z 1
|a(v, w′ )| |a(v, wv′ )| | 0 dv ′
dt wv dt|
dv
dt = dv 1
sup ′
≥ ′
= ′
= dt dt L (D) .
w ′ ∈W ′ kw kW ′ kwv kW ′ kwv kL∞ (D) 0
Invoking the extended Poincaré–Steklov inequality on V (with p := 1 and the bounded linear form
v 7→ v(0) in (3.13)) yields (bnb1’).
R1
(2) Proof of (bnb2’). Let w′ ∈ W ′ be such that 0 dv ′ ∞
dt w dt = 0 for all v ∈ V. Taking v in C0 (D), we
′ ′
infer that the weak derivative of w vanishes. Lemma 2.11 implies that w is a constant. Choosing
R1
v(t) := t as a test function leads to 0 w′ dt = 0. Hence, we have w′ = 0.
Exercises
Exercise 25.1 (Riesz–Fréchet). The objective is to prove the Riesz–Fréchet theorem (Theo-
rem C.24) by using the BNB theorem. Let V be a Hilbert space with inner product (·, ·)V . (i)
Show that for every v ∈ V, there is a unique JVrf (v) ∈ V ′ s.t. hJVrf (v), wiV ′ ,V := (v, w)V for all
w ∈ V. (ii) Show that JVrf : V ′ → V is a linear isometry.
Exercise 25.2 (Reflexivity). Let V, W be two Banach spaces such that there is an isomorphism
A ∈ L(V ; W ). Assume that V is reflexive. Prove that W is reflexive. (Hint : consider the map
A∗∗ ◦ JV ◦ A−1 .)
Exercise 25.3 (Space VR ). Let V be a set and assume that V has a vector space structure over
the field C. By restricting the scaling λv to λ ∈ R and v ∈ V, V has also a vector space structure
over the field R, which we denote by VR (V and VR are the same sets, but they are equipped with
different vector space structures); see Remark C.11. Let V ′ be the set of the bounded anti-linear
forms on V and VR′ be the set of the bounded linear forms on VR . Prove that the map I : V ′ → VR′
such that for all ℓ ∈ V ′ , I(ℓ)(v) := ℜ(ℓ(v)) for all v ∈ V, is a bijective isometry. (Hint : for ψ ∈ VR′ ,
set ℓ(v) := ψ(v) + iψ(iv) with i2 = −1.)
Exercise 25.4 (Orthogonal projection). Let V be a Hilbert space with inner product (·, ·)V
and induced norm k·kV . Let U be a nonempty, closed, and convex subset of V. Let f ∈ V. (i)
Show that there is a unique u in U such that kf − ukV = minv∈U kf − vkV . (Hint : recall that
1 2 1 2 1 2 1 2
4 (a − b) = 2 (c − a) + 2 (c − b) − (c − 2 (a + b)) and show that a minimizing sequence is a
Cauchy sequence.) (ii) Show that u ∈ U is the minimizer if and only if ℜ((f − u, v − u)V ) ≤ 0
for all v ∈ U . (Hint : proceed as in the proof of Proposition 25.8.) (iii) Assuming that U is a
(nontrivial) subspace of V, prove that the unique minimizer is characterized by (f − u, v)V = 0 for
all v ∈ U , and prove that the map ΠU : V ∋ f 7→ u ∈ U is linear and kΠU kL(V ;U) = 1. (iv) Let a
be a bounded, Hermitian, and coercive sesquilinear form (with ξ := 1 for simplicity). Let ℓ ∈ V ′ .
Set E(v) := 21 a(v, v) − ℓ(v). Show that there is a unique u ∈ V such that E(u) = minv∈U E(v) and
that u is the minimizer if and only if ℜ(a(u, v − u) − ℓ(v − u)) ≥ 0 for all v ∈ U .
Exercise 25.5 (Inf-sup constant). Let V be a Hilbert space, U a subset of V, and W a closed
|(u,w)V |
subspace of V. Let β := inf u∈U supw∈W kukV kwkW
. (i) Prove that β ∈ [0, 1]. (ii) Prove that
kΠW (u)kV
β = inf u∈U kukV , where ΠW is the orthogonal projection onto W. (Hint : use Exercise 25.4.)
1
(iii) Prove that ku − ΠW (u)kV ≤ (1 − β 2 ) 2 kukV . (Hint : use the Pythagorean identity.)
Exercise 25.6 (Fixed-point argument). The goal of this exercise is to derive another proof of
the Lax–Milgram lemma. Let A ∈ L(V ; V ) be defined by (A(v), w)V := a(v, w) for all v, w ∈ V
(note that we use an inner product to define A). Let L be the representative in V of the linear form
ℓ ∈ V ′ . Let λ be a positive real number. Consider the map Tλ : V → V s.t. Tλ (v) := v−λξ(A(v)−L)
for all v ∈ V. Prove that if λ is small enough, kTλ (v) − Tλ (w)kV ≤ ρλ kv − wkV for all v, w ∈ V
with ρλ ∈ (0, 1), and show that (25.6) is well-posed. (Hint : use Banach’s fixed-point theorem.)
Exercise 25.7 (Coercivity as necessary condition). Let V be a reflexive Banach space
and let A ∈ L(V ; V ′ ) be a monotone self-adjoint operator; see Definition C.31. Prove that A
is bijective if and only if A is coercive (with ξ := 1). (Hint : prove that ℜ(hA(v), wiV ′ ,V ) ≤
1 1
hA(v), viV2 ′ ,V hA(w), wiV2 ′ ,V for all v, w ∈ V.)
Exercise 25.8 (Darcy). Prove that the problem (24.14) is well-posed. (Hint : adapt the proof
of Proposition 25.18.)
Exercise 25.9 (First-order PDE). Prove that the problem (24.21) is well-posed. (Hint : adapt
the proof of Proposition 25.19.)
Exercise 25.10 (T -coercivity). Let V, W be Hilbert spaces. Prove that (bnb1)-(bnb2) are
equivalent to the existence of a bijective operator T ∈ L(V ; W ) and a real number η > 0 such
−1
that ℜ(a(v, T (v))) ≥ ηkvk2V for all v ∈ V. (Hint : use JW , (A−1 )∗ , and the map JVrf from the
Riesz–Fréchet theorem to construct T .)
Exercise 25.11 (Sign-changing diffusion). Let D be a Lipschitz domain D in Rd partitioned
into two disjoint Lipschitz subdomains D1 and D2 . Set Σ := ∂D1 ∩∂D2 , each having an intersection
with ∂D of positive measure. Let κ1 , κ2 be two real numbers s.t. κ1 > 0 and κ2 < 0. Set
κ(x) := κ1 1D1 (x)+κ2 1D2 (x) for all x ∈ D. Let V := RH01 (D) be equipped with the norm k∇vkL2 (D) .
The goal is to show that the bilinear form a(v, w) := D κ∇v·∇w satisfies conditions (bnb1)-(bnb2)
on V ×V ; see Chesnel and Ciarlet [118]. Set Vm := {v|Dm | v ∈ V } for all m ∈ {1, 2}, equipped
with the norm k∇vm kL2 (Dm ) for all vm ∈ Vm , and let γ0,m be the traces of functions in Vm on
Σ. (i) Assume that there is S1 ∈ L(V1 ; V2 ) s.t. γ0,2 (S1 (v1 )) = γ0,1 (v1 ). Define T : V → V s.t.
for all v ∈ V, T (v)(x) := v(x) if x ∈ D1 and T (v)(x) := −v(x) + 2S1 (v|D1 )(x) if x ∈ D2 . Prove
that T ∈ L(V ) and that T is an isomorphism. (Hint : verify that T ◦ T = IV , the identity in
V.) (ii) Assume that |κκ21 | > kS1 k2L(V1 ;V2 ) . Prove that the conditions (bnb1)-(bnb2) are satisfied.
(Hint : use T -coercivity from Remark 25.14.) (iii) Let D1 := (−a, 0)×(0, 1) and D2 := (0, b)×(0, 1)
with a > b > 0. Show that if |κκ12 | 6∈ [1, ab ], (bnb1)-(bnb2) are satisfied. (Hint : consider the map
S1 ∈ L(V1 ; V2 ) s.t. S1 (v1 )(x, y) := v1 (− ab x, y) for all v1 ∈ V1 , and the map S2 ∈ L(V2 ; V1 ) s.t.
S2 (v2 )(x, y) := v2 (−x, y) if x ∈ (−b, 0) and S2 (v2 )(x, y) := 0 otherwise, for all v2 ∈ V2 .)
Chapter 26
Basic error analysis
In Part VI, composed of Chapters 26 to 30, we introduce the Galerkin approximation technique
and derive fundamental stability results and error estimates. We also investigate implementation
aspects of the method (quadratures, linear algebra, assembling, storage). In this chapter, we
consider the following problem, introduced in Chapter 25, and study its approximation by the
Galerkin method:

(26.1)
a(u, w) = ℓ(w), ∀w ∈ W.
Here, V and W are Banach spaces, a is a bounded sesquilinear form on V ×W, and ℓ is a bounded
antilinear form on W. We focus on the well-posedness of the approximate problem, and we derive
a bound on the approximation error in a simple setting. This bound is known in the literature as
Céa’s lemma. We also characterize the well-posedness of the discrete problem by using the notion
of Fortin operator.
To stay general, we consider complex vector spaces. The case of real vector spaces is recovered
symbol ·, and by interpreting the symbol |·| as the absolute value instead of the modulus. Moreover,
sesquilinear forms become bilinear forms, and antilinear forms are just linear forms. We denote
by α and kakV ×W the inf-sup and the boundedness constants of the sesquilinear form a on V ×W,
i.e.,
|a(v, w)| |a(v, w)|
α := inf sup ≤ sup sup =: kakV ×W . (26.2)
v∈V w∈W kvkV kwkW v∈V w∈W kvkV kwkW
We assume that (26.1) is well-posed, i.e., 0 < α and kakV ×W < ∞. Whenever the context is
unambiguous, we write kak instead of kakV ×W .
26.1 The Galerkin method

The central idea in the Galerkin method is to replace in (26.1) the infinite-dimensional spaces V
and W by finite-dimensional spaces Vh and Wh (we always assume that Vh 6= {0} and Wh 6= {0}).
The subscript h ∈ H refers to the fact that these spaces are constructed as explained in Volume I
using finite elements and a mesh Th belonging to some sequence of meshes (Th )h∈H . The discrete
22 Chapter 26. Basic error analysis
problem takes the following form:

Find uh ∈ Vh such that
(26.3)
ah (uh , wh ) = ℓh (wh ), ∀wh ∈ Wh ,
where ah is a bounded sesquilinear form on Vh ×Wh and ℓh is a bounded antilinear form on Wh .

Notice that ah and ℓh possibly differ from a and ℓ, respectively. Since the spaces Vh and Wh are
finite-dimensional, (26.3) is called discrete problem. The space Vh is called discrete trial space (or
discrete solution space), and Wh discrete test space.
Definition 26.1 (Standard Galerkin, Petrov–Galerkin). The discrete problem (26.3) is called
standard Galerkin approximation when Wh = Vh and Petrov–Galerkin approximation otherwise.
Definition 26.2 (Conforming setting). The approximation is said to be conforming if Vh ⊂ V

and Wh ⊂ W.
There are circumstances when considering nonconforming approximations is useful. Two im-
portant examples are discontinuous Galerkin methods where discrete functions are discontinuous
across the mesh interfaces (see Chapters 38 and 60) and boundary penalty methods where bound-
ary conditions are enforced weakly (see Chapters 37 for elliptic PDEs and Chapters 57–59 for
Friedrichs’ systems). Very often, nonconforming approximations make it necessary to work with
discrete
R forms that differ from their continuous counterparts. For instance, the bilinear form
D ∇v·∇w dx does not make sense if the functions v and w are discontinuous. Another important
example leading to a modification of the forms at the discrete level is the use of quadratures (see
Chapter 30).
26.2 Discrete well-posedness

Our goal in this section is to study the well-posedness of the discrete problem (26.3). We equip Vh
and Wh with norms denoted by k·kVh and k·kWh , respectively. These norms can differ from those
of V and W. One reason can be that the approximation is nonconforming and the norm k·kV is
meaningless on Vh . This is the case for instance if the norm k·kV includes the H 1 -norm and the
discrete functions are allowed to jump across the mesh interfaces.
26.2.1 Discrete Lax–Milgram

Lemma 26.3 (Discrete Lax–Milgram). Let Vh be a finite-dimensional space. Assume that
Wh = Vh in (26.3). Let ah be a bounded sesquilinear form on Vh ×Vh and let ℓh ∈ Vh′ . Assume
that ah is coercive on Vh , i.e., there is a real number αh > 0 and a complex number ξ with |ξ| = 1
such that
ℜ (ξah (vh , vh )) ≥ αh kvh k2Vh , ∀vh ∈ Vh . (26.4)
1
Then (26.3) is well-posed with the a priori estimate kuh kVh ≤ αh kℓh kVh .
′
Proof. A simple proof just consists of invoking the Lax–Milgram lemma (see Lemma 25.2). We
now propose an elementary proof that relies on Vh being finite-dimensional. Let Ah : Vh → Vh′ be
the linear operator such that hAh (vh ), wh iVh′ ,Vh := ah (vh , wh ) for all vh , wh ∈ Vh . Problem (26.3)
amounts to seeking uh ∈ Vh such that Ah (uh ) = ℓh in Vh′ . Hence, (26.3) is well-posed iff Ah is an
isomorphism. Since dim(Vh ) = dim(Vh′ ) < ∞ this is equivalent to require that Ah be injective, i.e.,
Part VI. Galerkin approximation 23
ker(Ah ) = {0}. Let vh ∈ ker(Ah ) so that 0 = ξhAh (vh ), vh iVh′ ,Vh = ξah (vh , vh ). From coercivity,
we deduce that 0 ≥ αh kvh k2Vh , which proves that vh = 0. Hence, ker(Ah ) = {0}, thereby proving
that Ah is bijective.
Example 26.4 (Sufficient condition). (26.4) holds true if Vh ⊂ V (conformity), ah := a|Vh ×Vh ,
and a is coercive on V ×V.
Remark 26.5 (Variational formulation). As in the continuous setting (see Proposition 25.8),
if Vh is a real Hilbert space and if ah is symmetric and coercive (with ξ := 1 and Wh = Vh ), then
uh solves (26.3) iff uh minimizes the functional Eh (vh ) := 21 ah (vh , vh ) − ℓh (vh ) over Vh . If Vh ⊂ V,
ah := a|Vh ×Vh , and ℓh := ℓ|Vh , then Eh = E|Vh (E is the exact energy functional), and E(uh ) ≥ E(u)
since u minimizes E over the larger space V.
26.2.2 Discrete BNB

Theorem 26.6 (Discrete BNB). Let Vh , Wh be finite-dimensional spaces. Let ah be a bounded
sesquilinear form on Vh ×Wh and let ℓh ∈ Wh′ . Then the problem (26.3) is well-posed iff
|ah (vh , wh )|
inf sup =: αh > 0, (26.5a)
vh ∈Vh wh ∈Wh kvh kVh kwh kWh
dim(Vh ) = dim(Wh ). (26.5b)
(Recall that arguments in the above infimum and supremum are understood to be nonzero.) More-
over, we have the a priori estimate kuh kVh ≤ α1h kℓh kWh′ .
Proof. Let Ah : Vh → Wh′ be the linear operator such that
hAh (vh ), wh iWh′ ,Wh := ah (vh , wh ), ∀(vh , wh ) ∈ Vh ×Wh . (26.6)
The well-posedness of (26.3) is equivalent to Ah being an isomorphism, which owing to the finite-
dimensional setting and the rank nullity theorem, is equivalent to (i) ker(Ah ) = {0} (i.e., Ah is
injective) and (ii) dim(Vh ) = dim(Wh′ ). Since dim(Wh ) = dim(Wh′ ), (26.5b) is equivalent to (ii).
Let us prove that (i) is equivalent to the inf-sup condition (26.5a). By definition, we have
|ah (vh , wh )| |hAh (vh ), wh iWh′ ,Wh |
sup = sup =: kAh (vh )kWh′ .
wh ∈Wh kwh kWh wh ∈Wh kwh kWh
Assume first that (26.5a) holds true and let vh ∈ Vh be s.t. Ah (vh ) = 0. Then we have αh kvh kVh ≤
kAh (vh )kWh′ = 0, which shows that vh = 0. Hence, (26.5a) implies the injectivity of Ah . Conversely,
assume ker(Ah ) = {0} and let us prove (26.5a). An equivalent statement of (26.5a) is that there
is n0 ∈ N∗ such that for all vh ∈ Vh with kvh kVh = 1, one has kAh (vh )kWh′ > n10 . Reasoning
by contradiction, consider a sequence (vhn )n∈N∗ in Vh with kvhn kVh = 1 and kAh (vhn )kWh′ ≤
1
n . Since Vh is finite-dimensional, its unit sphere is compact. Hence, there is vh ∈ Vh such
that, up to a subsequence, vhn → vh . The limit vh satisfies kvh kVh = 1 and Ah (vh ) = 0, i.e.,
vh ∈ ker(Ah ) = {0}, which contradicts kvh kVh = 1. Hence, the injectivity of Ah implies (26.5a).
In conclusion, ker(Ah ) = {0} iff (26.5a) holds true. Finally, the a priori estimate follows from
αh kuh kVh ≤ kAh (uh )kWh′ = kℓh kWh′ .
Remark 26.7 (Link with BNB theorem). Condition (26.5a) is identical to (bnb1) from
Theorem 25.9 applied to (26.3), and it is equivalent to the following inf-sup condition:
|ah (vh , wh )|
∃αh > 0, αh kvh kVh ≤ sup , ∀vh ∈ Vh . (26.7)
wh ∈Wh kwh kWh
Condition (26.5b) seemingly differs from (bnb2) applied to (26.3), which reads
∀wh ∈ Wh , [ ah (vh , wh ) = 0, ∀vh ∈ Vh ] =⇒ [ wh = 0 ]. (26.8)
To see that (26.5b) is equivalent to (26.8) provided (26.5a) holds true, let us introduce the adjoint
operator A∗h : Wh → Vh′ (note that the space Wh is reflexive since it is finite-dimensional) such
that
hA∗h (wh ), vh iVh′ ,Vh = ah (vh , wh ), ∀(vh , wh ) ∈ Vh ×Wh . (26.9)
Then (26.8) says that A∗h is injective, and this statement is equivalent to (26.5b) if ker(Ah ) =
{0}; see Exercise 26.1. In summary, when the setting is finite-dimensional, the key property
guaranteeing well-posedness is (26.5a), whereas the other condition (26.5b) is very simple to verify.
Remark 26.8 (A∗h ). Ah is an isomorphism iff A∗h is an isomorphism; see Exercise 26.2. Moreover,
owing to Lemma C.53 (note that the space Vh is reflexive since it is finite-dimensional), Ah and
A∗h satisfy the inf-sup condition (26.5a) with the same constant αh , i.e.,
|hAh (vh ), wh iWh′ ,Wh | |hAh (vh ), wh iWh′ ,Wh |
inf sup = inf sup . (26.10)
vh ∈Vh wh ∈Wh kvh kVh kwh kWh wh ∈Wh vh ∈Vh kvh kVh kwh kWh
Note that hAh (vh ), wh iWh′ ,Wh = hA∗h (wh ), vh iVh′ ,Vh . As shown in Remark C.54, the identity (26.10)
may fail if Ah is not an isomorphism.
26.2.3 Fortin’s lemma

We focus on a conforming approximation, i.e., Vh ⊂ V and Wh ⊂ W, we equip the spaces Vh
and Wh with the norms of V and W, respectively, and we assume that ah := a|Vh ×Wh . Our
goal is to devise a criterion to ascertain that ah satisfies the inf-sup condition (26.5a). To this
purpose, we would like to use the inf-sup condition (26.2) satisfied by a on V ×W. Unfortunately,
this condition does not imply its discrete counterpart on Vh ×Wh . Since Vh ⊂ V, (26.2) implies
that αkvh kV ≤ supw∈W |a(v h ,w)|
kwkW for all vh ∈ Vh , but it is not clear that the bound still holds
when restricting the supremum to the subspace Wh . The Fortin operator provides the missing
ingredient.
Lemma 26.9 (Fortin). Let V, W be Hilbert spaces and let a be a bounded sesquilinear form on
V ×W. Let α and kak be the inf-sup and boundedness constants of a defined in (26.2). Let Vh ⊂ V
and let Wh ⊂ W be equipped with the norms of V and W, respectively. Consider the following two
statements:
(i) There exists a map Πh : W → Wh , called Fortin operator such that: (i.a) a(vh , Πh (w)−w) = 0
for all (vh , w) ∈ Vh ×W ; (i.b) There is γΠh > 0 such that γΠh kΠh (w)kW ≤ kwkW for all
w ∈ W.
(ii) The discrete inf-sup condition (26.5a) holds true.
αh
Then (i) =⇒ (ii) with αh ≥ γΠh α. Conversely, (ii) =⇒ (i) with γΠh ≥ kak and Πh can be
constructed to be linear and idempotent (Πh ◦ Πh = Πh ).
Proof. (1) Let us assume (i). Let ǫ > 0. We have for all vh ∈ Vh ,
|a(vh , wh )| |a(vh , Πh (w))| |a(vh , w)|
sup ≥ sup = sup
wh ∈Wh kwh kW w∈W kΠh (w)kW + ǫkwkW w∈W kΠh (w)kW + ǫkwkW
|a(vh , w)| γ Πh
≥ γΠh sup ≥ α kvh kV ,
w∈W kwkW (1 + ǫγΠh ) 1 + ǫγΠh
since a satisfies (bnb1) and Vh ⊂ V. This proves (26.5a) with αh ≥ γΠh α since ǫ can be taken
arbitrarily small. (Since Πh cannot be injective, we introduced ǫ > 0 to avoid dividing by zero
whenever w ∈ ker(Πh ).)
(2) Conversely, let us assume that a satisfies (26.5a). Let Ah : Vh → Wh′ be defined in (26.6).
Condition (26.5a) means that kAh (vh )kWh′ ≥ αh kvh kV for all vh ∈ Vh (k·kWh′ should not be
confused with k·kW ′ ). Hence, the operator B := Ah satisfies the assumptions of Lemma C.44
with Y := Vh , Z := Wh′ , and β := αh . We infer that A∗h : Wh → Vh′ has a (linear) right inverse
A∗† ′ ∗† −1
h : Vh → Wh such that kAh kL(Vh ,Wh ) ≤ αh . Let us now consider the operator B : W → Vh
′
′
∗†
s.t. hB(w), vh iVh′ ,Vh := a(vh , w) for all (vh , w) ∈ Vh ×W, and let us set Πh := Ah ◦ B : W → Wh .
We have
a(vh , Πh (w)) = hAh (vh ), A∗†
h (B(w))iWh ,Wh = hB(w), vh iVh ,Vh = a(vh , w),
′ ′
kak
so that a(vh , Πh (w) − w) = 0. Moreover, we have kΠh (w)kW ≤ αh kwkW since kA∗†
h kL(Vh ;Wh ) =
′
∗† ∗†
α−1 ∗
h and kBkL(W ;Vh ) ≤ kak. Finally, since B|Wh = Ah , we have Πh ◦ Πh = (Ah ◦ B) ◦ (Ah ◦ B) =
′
A∗† ∗ ∗†
h ◦ (Ah ◦ Ah ) ◦ B = Πh , which proves that Πh is idempotent.
Remark 26.10 (Dimension, equivalence). We did not assume that Vh and Wh have the
same dimension. This level of generality is useful to apply Lemma 26.9 to mixed finite element
approximations; see Chapter 50. The implication (i) =⇒ (ii) in Lemma 26.9 is known in the
literature as Fortin’s lemma [201], and is useful to analyze mixed finite element approximations
(see, e.g., Chapter 54 on the Stokes equations). The converse implication can be found in Girault
and Raviart [217, p. 117]. This statement is useful in the analysis of Petrov–Galerkin methods; see
Carstensen et al. [111], Muga and van der Zee [308], and also Exercise 50.7. Note that the gap in
the stability constant γΠh between the direct and the converse statements is equal to the condition
number κ(a) := kak α of the sesquilinear form a (see Remark 25.12). Finally, we observe that the
Fortin operator is not uniquely defined.
Remark 26.11 (Banach spaces). Lemma 26.9 can be extended to Banach spaces. Such a
construction is done in [187], where Lemma C.42 is invoked to build a (bounded) right inverse of
A∗h , and where the proposed map Πh is nonlinear. Whether one can always construct a Fortin
operator Πh that is linear in Banach spaces seems to be an open question.
26.3 Basic error estimates

In this section, we assume that the exact problem (26.1) and the discrete problem (26.3) are well-
posed. Our goal is to bound the approximation error (u − uh ) in the simple setting where the
approximation is conforming (Vh ⊂ V, Wh ⊂ W, ah := a|Vh ×Wh , and ℓh := ℓ|Wh ).
26.3.1 Strong consistency: Galerkin orthogonality

The starting point of the error analysis is to make sure that the discrete problem (26.3) is consistent
with the original problem (26.1). Loosely speaking one way of checking consistency is to insert the
exact solution into the discrete problem and to verify that the discrepancy is small. We say that
there is strong consistency whenever this operation is possible and the discrepancy is actually zero.
A more general definition of consistency is given in the next chapter. The following result, known
as the Galerkin orthogonality property, expresses the fact that strong consistency holds true in the
present setting.
Lemma 26.12 (Galerkin orthogonality). Assume that Vh ⊂ V, Wh ⊂ W, ah := a|Vh ×Wh , and

ℓh := ℓ|Wh . The following holds true:
a(u, wh ) = ℓ(wh ) = a(uh , wh ), ∀wh ∈ Wh . (26.11)
In particular, we have a(u − uh , wh ) = 0 for all wh ∈ Wh .
Proof. The first equality follows from Wh ⊂ W and the second one from ah := a|Vh ×Wh and
ℓh := ℓ|Wh .
26.3.2 Céa’s and Babuška’s lemmas

Lemma 26.13 (Céa). Assume that Wh = Vh ⊂ V = W, ah := a|Vh ×Vh , and ℓh := ℓ|Vh . Assume
that the sesquilinear form a is V -coercive with constant α > 0 and let kak be its boundedness
constant defined in (26.2) (with W = V ). Then the following error estimate holds true:
kak
ku − uh kV ≤ inf ku − vh kV . (26.12)
α vh ∈Vh
Moreover, if the sesquilinear form a is Hermitian, the error estimate (26.12) can be sharpened as
follows:
1
kak 2
α vh ∈Vh
Proof. Invoking the coercivity of a (stability), followed by the Galerkin orthogonality property
(strong consistency) and the boundedness of a, gives
α ku − uh k2V ≤ ℜ(ξa(u − uh , u − uh ))
= ℜ(ξa(u − uh , u − vh ))
≤ kak ku − uh kV ku − vh kV ,
for all vh in Vh . This proves the error estimate (26.12). Assume now that the sesquilinear form a
is Hermitian. Let vh be arbitrary in Vh . Let us set e := u − uh and ηh := uh − vh . The Galerkin
orthogonality property and the Hermitian symmetry of a imply that a(e, ηh ) = a(ηh , e) = 0. Hence,
we have
a(u − vh , u − vh ) = a(e + ηh , e + ηh ) = a(e, e) + a(ηh , ηh ),
and the coercivity of a implies that ℜ(ξa(e, e)) ≤ ℜ(ξa(u − vh , u − vh )). Combining this bound
with the stability and boundedness properties of a yields
α ku − uh k2V ≤ ℜ(ξa(u − uh , u − uh )) = ℜ(ξa(e, e))
≤ ℜ(ξa(u − vh , u − vh )) ≤ kak ku − vh k2V .
Taking the infimum over vh ∈ Vh proves the error estimate (26.13).
We now extend Céa’s lemma to the more general case where stability relies on a discrete inf-sup
condition rather than a coercivity argument. Thus, the discrete spaces Vh and Wh can differ.
Lemma 26.14 (Babuška). Assume that Vh ⊂ V, Wh ⊂ W, ah := a|Vh ×Wh , ℓh := ℓ|Wh , and
dim(Vh ) = dim(Wh ). Equip Vh and Wh with the norms of V and W, respectively. Assume the
following discrete inf-sup condition:
|a(vh , wh )|
inf sup =: αh > 0. (26.14)
vh ∈Vh wh ∈Wh kvh kV kwh kW
Let kak be the boundedness constant of a defined in (26.2). The following error estimate holds
true:
kak
ku − uh kV ≤ 1 + inf ku − vh kV . (26.15)
αh vh ∈Vh
Proof. Let vh ∈ Vh . Using stability (i.e., (26.14)), strong consistency (i.e., the Galerkin orthogo-
nality property), and the boundedness of a, we infer that
|a(uh − vh , wh )|
αh kuh − vh kV ≤ sup
wh ∈Wh kwh kW
|a(u − vh , wh )|
= sup ≤ kak ku − vh kV ,
wh ∈Wh kwh kW
and (26.15) follows from the triangle inequality.

The error estimates from Lemma 26.13 and from Lemma 26.14 are said to be quasi-optimal
since inf vh ∈Vh ku − vhkV is the best-approximation error of u by an element in Vh , and by definition
ku − uh kV cannot be smaller than the best-approximation error, i.e., the following two-sided error
bound holds:
inf ku − vh kV ≤ ku − uh kV ≤ c inf ku − vh kV , (26.16)
vh ∈Vh vh ∈Vh
with c := kak
α for Céa’s lemma and c
:= 1 + kak
αh for Babuška’s lemma. One noteworthy consequence
of (26.16) is that uh = u whenever the exact solution turns out to be in Vh .
Corollary 26.15 (Convergence). We have limh→0 ku − uhkV = 0 if the assumptions of Lemma
26.14 hold true together with the following properties:
(i) Uniform stability: αh ≥ α0 > 0 for all h ∈ H.
(ii) Approximability: limh→0 (inf vh ∈Vh kv − vh kV ) = 0 for all v ∈ V.
Proof. Direct consequence of the assumptions.
Remark 26.16 (Céa). In the context of Céa’s lemma, uniform stability follows from coercivity.
Thus, approximability implies convergence.
Remark 26.17 (Literature). Lemma 26.13 is derived in [114, Prop. 3.1] and is usually called
Céa’s lemma in the literature; see, e.g., Ciarlet [124, Thm. 2.4.1], Brenner and Scott [87, Thm. 2.8.1].
Lemma 26.14 is derived in Babuška [33, Thm. 2.2].
26.3.3 Approximability by finite elements

Let us present an important example where the approximability property identified in Corol-
lary 26.15 holds true. Let V := H 1 (D) where D is a Lipschitz polyhedron in Rd . Let Vh :=
Pkg (Th ) ⊂ H 1 (D) be the H 1 -conforming finite element space of degree k ≥ 1 (see (20.1)), where
(Th )h∈H is a shape-regular sequence of affine meshes so that each mesh covers D exactly. One
way to prove approximability is to consider the Lagrange interpolation operator or the canonical
interpolation operator (see §19.3), i.e., let us set either Ih := IhL or Ih := Ihg (we omit the subscript
k for simplicity), so that Ih : V g (D) → Pkg (Th ) with domain V g (D) := H s (D), s > d2 (see (19.19)
with p := 2). Let l be the smallest integer s.t. l > d2 . Setting r := min(l − 1, k), Corollary 19.8
with m := 1 (note that r ≥ 1) implies that
inf kv − vh kH 1 (D) ≤ kv − Ih (v)kH 1 (D) ≤ c hr ℓD |v|H 1+r (D) ,

vh ∈Vh
for all v ∈ H 1+r (D), where ℓD is a characteristic length of D, e.g., ℓD := diam(D). Another
possibility consists of using the quasi-interpolation operator Ihg,av : L1 (D) → Vh from Chapter 22
since Theorem 22.6 implies that
inf kv − vh kH 1 (D) ≤ kv − Ihg,av (v)kH 1 (D) ≤ c hr ℓD |v|H 1+r (D) ,

vh ∈Vh
for all v ∈ H 1+r (D) and all r ∈ (0, k]. We now establish approximability by invoking a density
argument. Let v ∈ V and let ǫ > 0. Since H 1+r (D) is dense in V for all r > 0, there is
vǫ ∈ H 1+r (D) s.t. kv − vǫ kH 1 (D) ≤ ǫ. Using the triangle inequality and the above interpolation
estimates, we infer that
inf kv − vh kH 1 (D) ≤ kv − Ihg,av (vǫ )kH 1 (D)

vh ∈Vh
≤ kv − vǫ kH 1 (D) + kvǫ − Ihg,av (vǫ )kH 1 (D)

≤ ǫ + c hr ℓD |vǫ |H 1+r (D) .
Letting h → 0 shows that lim suph→0 (inf vh ∈Vh kv − vh kH 1 (D) ) ≤ ǫ, and since ǫ > 0 is arbitrary, we
conclude that approximability holds true, i.e., the best-approximation error in Vh of any function
v ∈ V tends to zero as h → 0. The above arguments can be readily adapted when homogeneous
Dirichlet conditions are strongly enforced.
26.3.4 Sharper error estimates

We now sharpen the constant appearing in the error estimate (26.15) from Lemma 26.14. Let
Vh ⊂ V and Wh ⊂ W with dim(Vh ) = dim(Wh ), and let a be a bounded sesquilinear form on
V ×W satisfying the discrete inf-sup condition (26.14) on Vh ×Wh . We define the discrete solution
map Gh : V → Vh s.t. for all v ∈ V, Gh (v) is the unique element in Vh satisfying
a(Gh (v) − v, wh ) = 0, ∀wh ∈ Wh . (26.17)
Note that Gh (v) is well defined owing to the discrete inf-sup condition (26.14) and since a(v, ·) :
Wh → C is a bounded antilinear form on Wh . Moreover, Gh is linear and Vh is pointwise invariant
under Gh .
Lemma 26.18 (Xu–Zikatanov). Let {0} ( Vh ( V and Wh ⊂ W with dim(Vh ) = dim(Wh )
where V, W are Hilbert spaces, and let a be a bounded sesquilinear form on V ×W with constant
kak defined in (26.2) satisfying the discrete inf-sup condition (26.14) on Vh ×Wh with constant αh .
Then,
kak
αh vh ∈Vh
Proof. Since Gh is linear and Vh is pointwise invariant under Gh , we have
u − uh = u − Gh (u) = (u − vh ) − Gh (u − vh ),
for all vh ∈ Vh . We infer that
ku − uh kV ≤ kI − Gh kL(V ) ku − vh kV = kGh kL(V ) ku − vh kV ,
where the last equality follows from the fact that in any Hilbert space H, any operator T ∈ L(H)
such that 0 6= T ◦ T = T 6= I verifies kT kL(H) = kI − T kL(H) (see the proof of Theorem 5.14). We
can apply this result to the discrete solution map since Gh 6= 0 (since Vh 6= {0}), Gh ◦ Gh = Gh
(since Vh is pointwise invariant under Gh ), and Gh 6= I (since Vh 6= V ). To conclude the proof, we

bound kGh kL(V ) as follows: For all v ∈ V,
|a(Gh (v), wh )| |a(v, wh )|

αh kGh (v)kV ≤ sup = sup ≤ kak kvkV ,
wh ∈Wh kwh kW wh ∈Wh kwh kW
kak
which shows that kGh kL(V ) ≤ αh .
ku−uh kV
Let Λ be the smallest c so that the inequality inf vh ∈Vh ku−vh kV ≤ c holds for every u ∈ V. Then
ku−Gh (u)kV
Λ = supu∈V supvh ∈Vh ku−vh kV since uh = Gh (u). But the proof of Lemma 26.18 shows that
Λ = kI − Gh kL(V ) = kGh kL(V ) . Hence, kGh kL(V ) is the smallest constant such that the following
quasi-optimal error estimate holds:
ku − uh kV ≤ kGh kL(V ) inf ku − vh kV .

vh ∈Vh
Thus, sharp estimates on kGh kL(V ) are important to determine whether the approximation error
is close or not to the best-approximation error. The following result shows in particular that
kGh kL(V ) is, up to a factor in the interval [α, kak], proportional to the inverse of the discrete
inf-sup constant αh .
Lemma 26.19 (Tantardini–Veeser). Under the assumptions of Lemma 26.18, the following
holds true:

|a(v, wh )|
sup
v∈V kvkV
kGh kL(V ) = sup ≥ 1, (26.19a)
wh ∈Wh |a(vh , wh )|
sup
vh ∈Vh kvh kV
α kak
≤ kGh kL(V ) ≤ . (26.19b)
αh αh
Proof. (1) Let A ∈ L(V ; W ′ ) be the operator associated with the sesquilinear form a, i.e.,
hA(v), wiW ′ ,W := a(v, w), ∀(v, w) ∈ V ×W,
and let A∗ ∈ L(W ; V ′ ) be its adjoint (where we used the reflexivity of W ). We have
|a(v, w)|
α kwkW ≤ kA∗ (w)kV ′ = sup ≤ kak kwkW , (26.20)
v∈V kvkV
for all w ∈ W. Indeed, the first bound follows from Lemma C.53 and the inf-sup stability of a, and
the second one follows from the boundedness of a. This shows that the norms k·kW and kA∗ (·)kV ′
are equivalent on W.
(2) Since Wh ⊂ W, we have A∗ (wh ) ∈ V ′ for all wh ∈ Wh . Upon setting
|a(vh , wh )|
γh := inf sup ,
wh ∈Wh vh ∈Vh kvh kV kA∗ (wh )kV ′
αh
we have γh ≥ kak > 0 owing to the inf-sup condition satisfied by a on Vh ×Wh , the norm equiv-
|a(v,wh )|
alence (26.20), and Lemma C.53. Recalling that kA∗ (wh )kV ′ = supv∈V kvkV , the assertion
(26.19a) amounts to kGh kL(V ) = γh−1 ≥ 1.
(3) Let wh ∈ Wh . Using the definition (26.17) of the discrete solution map and the definition of
the dual norm kA∗ (wh )kV ′ , we have
|a(Gh (v), wh )|
kA∗ (wh )kV ′ = sup
v∈V kvkV
|a(Gh (v), wh )| kGh (v)kV
≤ sup sup
v∈V kGh (v)kV v∈V kvkV
|a(vh , wh )|
≤ sup kGh kL(V ) .
vh ∈Vh kvh kV
Rearranging the terms and taking the infimum over wh ∈ Wh shows that γh ≥ kGh k−1
L(V ) , i.e.,
−1
kGh kL(V ) ≥ γh .
(4) Since γh > 0, Remark 26.8 implies that
|a(vh , wh )|
γh = inf sup . (26.21)
vh ∈Vh wh ∈Wh kvh kV kA∗ (wh )kV ′
Let v ∈ V. Applying the above identity to Gh (v) ∈ Vh , we infer that

|a(Gh (v), wh )| |a(v, wh )|
γh kGh (v)kV ≤ sup = sup ≤ kvkV ,
wh ∈Wh kA∗ (wh )kV ′ ∗
wh ∈Wh kA (wh )kV ′
since |a(v, wh )| = |hA∗ (wh ), viV ′ ,V | ≤ kA∗ (wh )kV ′ kvkV . Taking the supremum over v ∈ V shows
that kGh kL(V ) ≤ γh−1 . Thus, we have proved that kGh kL(V ) = γh−1 , and the lower bound in
(26.19a) is a direct consequence of Vh ⊂ V.
(5) It remains to prove (26.19b). Using the norm equivalence (26.20) in (26.21) to bound from
below and from above kA∗ (wh )kV ′ , we infer that
1 |a(vh , wh )| 1 |a(vh , wh )|
inf sup ≤ γh ≤ inf sup ,
kak vh ∈Vh wh ∈Wh kvh kV kwh kW α vh ∈Vh wh ∈Wh kvh kV kwh kW
so that αh
kak ≤ γh ≤ αh
α , and (26.19b) follows from kGh kL(V ) = γh−1 .
Remark 26.20 (Literature). Lemma 26.18 is proved in Xu and Zikatanov [397, Thm. 2], and
Lemma 26.19 in Tantardini and Veeser [361, Thm. 2.1]. See also Arnold et al. [18] for the lower
bound ααh ≤ kGh kL(V ) .
Remark 26.21 (Discrete dual norm). For all wh ∈ Wh , A∗ (wh ) ∈ V ′ can be viewed as a
member of Vh′ by restricting its action to the subspace Vh ⊂ V. We use the same notation and
simply write A∗ (wh ) ∈ Vh′ . The statement (26.19a) in Lemma 26.19 can be rewritten as follows:
kA∗ (wh )kV ′

kGh kL(V ) = sup , (26.22)
wh ∈Wh kA∗ (wh )kVh′
|hA∗ (wh ),vh iV ′ ,V | |a(vh ,wh )|
where kA∗ (wh )kVh′ := supvh ∈Vh kvh kV = supvh ∈Vh kvh kV .
Example 26.22 (Orthogonal projection). Let V ֒→ L be two Hilbert spaces with continuous
and dense embedding. Using the Riesz–Fréchet theorem (Theorem C.24), we identify L with
its dual L′ by means of the inner product (·, ·)L in L. This allows us to define the continuous
embedding EV ′ : V → V ′ s.t. hEV ′ (v), wiV ′ ,V := (v, w)L for all v, w ∈ V. Note that EV ′ is self-
adjoint. Consider a subspace {0} ( Vh ( V. Let Ph be the discrete solution map associated with
the sesquilinear form a(v, w) := hEV ′ (v), wiV ′ ,V for all v, w ∈ V. Note that Ph is the L-orthogonal
projection onto Vh since
(Ph (v), wh )L = hEV ′ (Ph (v)), wh iV ′ ,V := hEV ′ (v), wh iV ′ ,V = (v, wh )L ,
for all v ∈ V and all wh ∈ Vh . Then Lemma 26.19 provides a precise estimate on the V -stability
of Ph in the form
kEV ′ (wh )kVh′ |(wh , vh )L |
kPh k−1
L(V ) = inf = inf sup . (26.23)
wh ∈Vh kEV ′ (wh )kV ′ wh ∈Vh vh ∈Vh kEV ′ (wh )kV ′ kvh kV
See also Tantardini and Veeser [361, Prop. 2.5], Andreev [11, Lem. 6.2]. An important example
is V := H01 (D) and L := L2 (D). The reader is referred to §22.5 for further discussion on the
L2 -orthogonal projection onto conforming finite element spaces (see in particular Remark 22.23
for sufficient conditions on the underlying mesh to ensure H 1 -stability).
Exercises
Exercise 26.1 ((bnb2)). Prove that (26.8) is equivalent to (26.5b) provided (26.5a) holds true.
(Hint : use that dim(Wh ) = rank(Ah ) + dim(ker(A∗h )) (A∗h is defined in (26.9)) together with the
rank nullity theorem.)
Exercise 26.2 (Bijectivity of A∗h ). Prove that Ah is an isomorphism if and only if A∗h is
an isomorphism. (Hint : use dim(Vh ) = rank(A∗h ) + dim(ker(Ah )) and dim(Wh ) = rank(Ah ) +
dim(ker(A∗h )).)
Exercise 26.3 (Petrov–Galerkin). Let V, W be real Hilbert spaces, let A ∈ L(V ; W ′ ) be an
isomorphism, and let ℓ ∈ W ′ . Consider a conforming Petrov–Galerkin approximation with a
rf −1
finite-dimensional subspace Vh ⊂ V and Wh := (JW ) AVh ⊂ W , where JW rf
: W → W ′ is the
Riesz–Fréchet isomorphism. The discrete bilinear form is ah (vh , wh ) := hA(vh ), wh iW ′ ,W , and the
discrete linear form is ℓh (wh ) := ℓ(wh ) for all vh ∈ Vh and all wh ∈ Wh . (i) Prove that the discrete
problem (26.3) is well-posed. (ii) Show that its unique solution minimizes the residual functional
R(v) := kA(v) − ℓkW ′ over Vh .
Exercise 26.4 (Fortin’s lemma). (i) Prove that Πh in the converse statement of Lemma 26.9
is idempotent. (Hint : prove that B ◦ A∗† h = IVh .) (ii) Assume that there are two maps Π1,h , Π2,h :
′
W → Wh and two uniform constants c1 , c2 > 0 such that kΠ1,h (w)kW ≤ c1 kwkW , kΠ2,h ((I −
Π1,h )(w))kW ≤ c2 kwkW and a(vh , Π2,h (w) − w) = 0 for all vh ∈ Vh , w ∈ W. Prove that Πh :=
Π1,h +Π2,h (I−Π1,h ) is a Fortin operator. (iii) Write a variant of the direct statement in Lemma 26.9
assuming V, W reflexive, A ∈ L(V ; W ′ ) bijective, and using this time an operator Πh : V → Vh
such that a(Πh (v) − v, wh ) = 0 for all (v, wh ) ∈ V ×Wh and γΠh kΠh (v)kV ≤ kvkV for all v ∈ V for
some γΠh > 0. (Hint : use (26.10) and Lemma C.53.)
Exercise 26.5 (Compact perturbation). Let V, W be Banach spaces with W reflexive. Let
A0 ∈ L(V ; W ′ ) be bijective, let T ∈ L(V ; W ′ ) be compact, and assume that A := A0 + T is
injective. Let a0 (v, w) := hA0 (v), wiW ′ ,W and a(v, w) := hA(v), wiW ′ ,W for all (v, w) ∈ V ×W. Let
Vh ⊂ V and Wh ⊂ W be s.t. dim(Vh ) = dim(Wh ) for all h ∈ H. Assume that approximability
holds, and that the sesquilinear form a0 satisfies the inf-sup condition
|a0 (vh , wh )|
inf sup =: α0 > 0, ∀h ∈ H.
Following Wendland [392], the goal is to show that there is h0 > 0 s.t.
|a(vh , wh )|
inf sup =: α > 0, ∀h ∈ H ∩ (0, h0 ].
(i) Prove that A ∈ L(V ; W ′ ) is bijective. (Hint : recall that a compact operator is bijective iff it is
injective; this follows from the Fredholm alternative, Theorem 46.13.) (ii) Consider Rh ∈ L(V ; Vh )
s.t. for all v ∈ V, Rh (v) ∈ Vh satisfies a0 (Rh (v) − v, wh ) = 0 for all wh ∈ Wh . Prove that
Rh ∈ L(V ; Vh ) and that Rh (v) converges to v as h ↓ 0 for all v ∈ V. (Hint : proceed as in the proof
of Céa’s lemma.) (iii) Set L := IV + A−1 −1
0 T and Lh := IV + Rh A0 T where IV is the identity
operator in V (observe that both L and Lh are in L(V )). Prove that Lh converges to L in L(V ).
(Hint : use Remark C.5.) (iv) Show that if h ∈ H is small enough, Lh is bijective and there is C,
independent of h ∈ H, such that kL−1 h kL(V ) ≤ C. (Hint : observe that L
−1
Lh = IV − L−1 (L − Lh )
and consider the Neumann series.) (v) Conclude.
Chapter 27
Error analysis with variational

crimes
We have shown in the previous chapter how the Galerkin method can be used to approximate
the solution to the model problem (26.1), and we have derived an error estimate in the simple
setting where Vh ⊂ V, Wh ⊂ W, ah := a|Vh ×Wh , and ℓh := ℓ|Wh . Departures from this setting
are often called variational crimes in the literature. In this chapter, we perform the error analysis
when variational crimes are committed. The main results, Lemma 27.5 and Lemma 27.8, will be
invoked frequently in this book. They give an upper bound on the approximation error in terms of
the best-approximation error of the exact solution by members of the discrete trial space. These
error estimates are based on the notions of stability and consistency/boundedness. Combined
with an approximability property, they allow us to conclude that the approximation method is
convergent. Two simple examples illustrate the theory: a first-order PDE approximated by the
Galerkin/least-squares technique and a second-order PDE approximated by a boundary penalty
method.
27.1 Setting
In the entire chapter, we suppose that the assumptions of the BNB theorem (Theorem 25.9 or its
variant Theorem 25.15) are satisfied, so that the exact problem (26.1) is well-posed. The inf-sup
and boundedness constants on V ×W of the exact sesquilinear form a are denoted by α and kak;
see (26.2). The exact solution is denoted by u ∈ V.
Recall that the Galerkin approximation (26.3) relies on the discrete trial space Vh and the dis-
crete test space Wh . These spaces are equipped with the norms k·kVh and k·kWh , respectively. The
discrete problem uses a discrete sesquilinear form ah defined on Vh ×Wh and a discrete antilinear
form ℓh defined on Wh . The sesquilinear form ah and the antilinear form ℓh must be viewed,
respectively, as some approximations to a and ℓ. The solution to the discrete problem (26.3) is
denoted by uh ∈ Vh . We always assume that dim(Vh ) = dim(Wh ), so that the well-posedness of
the discrete problem is equivalent to the following inf-sup condition:
|ah (vh , wh )|
inf sup =: αh > 0. (27.1)
vh ∈Vh wh ∈Wh kvh kVh kwh kWh
We say that the approximation (26.3) is stable whenever (27.1) holds true, i.e., αh > 0.
34 Chapter 27. Error analysis with variational crimes
The goal of this chapter is to bound the error, i.e., we want to estimate how far the discrete
solution uh ∈ Vh lies from the exact solution u ∈ V. We say that the method converges if the
error tends to zero as the approximation capacity of the discrete trial space Vh increases. The
approximation capacity of Vh increases by refining an underlying mesh. We will see that there are
three key properties to establish convergence: (i) stability, (ii) consistency/boundedness, and (iii)
approximability. Stability and approximability have already emerged as important notions in the
error analysis presented in §26.3. The notion of consistency was present in the simple form of the
Galerkin orthogonality property, and the boundedness of the sesquilinear form a on V ×W was
also invoked.
Remark 27.1 (Lax principle). A loose principle in numerical analysis, known as Lax Principle, is
that stability and consistency imply convergence. The fact that boundedness and approximability
are not mentioned does not mean that these properties should be taken for granted. We refer the
reader to the upcoming chapters for numerous examples.
Remark 27.2 (Norms). Since all the norms are equivalent in finite-dimensional vector spaces,
if (27.1) holds true for one choice of norms in Vh and Wh , it holds true also for every other choice.
The goal is to select norms s.t. (i) ah is uniformly stable, i.e., αh ≥ α0 > 0 for all h ∈ H, and (ii)
ah is uniformly bounded on Vh ×Wh with respect to h ∈ H.
27.2 Main results

This section contains our two main abstract error estimates.
27.2.1 The spaces Vs and V♯

In a nonconforming approximation setting where Vh 6⊂ V, the exact solution u and the discrete
solution uh may be objects of different nature. This poses the question of how to measure the
approximation error. For instance, does the expression (u − uh ) make sense? We are going to
assume that it is possible to define a common ground between u and uh to evaluate the error. A
simple way to do this is to assume that it is meaningful to define the linear space (V + Vh ). If it
is indeed the case, then the error belongs to this space.
However, we will see in numerous examples that the error analysis often requires to assume
that the exact solution has slightly more smoothness than just being a member of V. We formalize
this assumption by introducing a functional space Vs such that u ∈ Vs ⊆ V. Our setting for the
error analysis is therefore as follows:
u ∈ Vs ⊆ V, u − uh ∈ V♯ := Vs + Vh . (27.2)
Note that this setting allows for Vs := V, and in the conforming setting, where Vh ⊂ V, this then
implies that V♯ := V.
27.2.2 Consistency/boundedness
A crucial notion in the error analysis is that of consistency/boundedness. Loosely speaking the
idea behind consistency is to insert the exact solution into the discrete equations and to verify
that the discrepancy is small. This may not be possible in a nonconforming approximation setting
because it may turn out that the discrete sesquilinear form ah is not meaningful when its first
argument is the exact solution. To stay general, we are going to define a consistency error for
every discrete trial function vh ∈ Vh with the expectation that this error is small if the difference
(u − vh ) ∈ V♯ is small. Let us now formalize this idea. Recall that the norm of any antilinear form
φh ∈ Wh′ := L(Wh ; C) is defined by kφh kWh′ := supwh ∈Wh |φ h (wh )|
kwh kW . h
Definition 27.3 (Consistency/boundedness). Let δh : Vh → Wh′ be defined by setting

hδh (vh ), wh iWh′ ,Wh := ℓh (wh ) − ah (vh , wh ) = ah (uh − vh , wh ). (27.3)
The quantity kδh (vh )kWh′ is called consistency error for the discrete trial function vh ∈ Vh . We say
that consistency/boundedness holds true if the space V♯ can be equipped with a norm k·kV♯ such
that there is a real number ω♯h , uniform w.r.t. u ∈ Vs , such that for all vh ∈ Vh and all h ∈ H,
kδh (vh )kWh′ ≤ ω♯h ku − vh kV♯ . (27.4)
Example 27.4 (Simple setting). Assume conformity (i.e., Vh ⊂ V and Wh ⊂ W ), ah :=
a|Vh ×Wh , and ℓh := ℓ|Wh . Take Vs := V, so that V♯ := V, and take k·kV♯ := k·kV . The consistency
error (27.3) is such that
hδh (vh ), wh iWh′ ,Wh = ℓ(wh ) − a(vh , wh ) = a(u − vh , wh ),
where we used that ℓ(wh ) = a(u, wh ) (i.e., the Galerkin orthogonality property). Since a is bounded
on V ×W, (27.4) holds true with ω♯h := kak.
27.2.3 Error estimate using one norm

We can now establish our first abstract error estimate. This estimate will be applied to various
nonconforming approximation settings of elliptic PDEs. It hinges on the assumption that there is
a real number c♯ , uniform w.r.t. h ∈ H, s.t.
kvh kV♯ ≤ c♯ kvh kVh , ∀vh ∈ Vh . (27.5)
Recall that k·kVh is the stability norm on Vh used in (27.1) and k·kV♯ is the consistency/boundedness
norm on V♯ used in (27.4).
Lemma 27.5 (Quasi-optimal error estimate). Assume the following: (i) Stability, i.e., (27.1)
holds true; (ii) Consistency/boundedness, i.e., u ∈ Vs and (27.4) holds true. Assume that (27.5)
holds true. Then we have

ω♯h
ku − uh kV♯ ≤ 1 + c♯ inf ku − vh kV♯ . (27.6)
αh vh ∈Vh
Proof. Owing to the assumptions, we infer that for all vh ∈ Vh ,
ku − uh kV♯ ≤ ku − vh kV♯ + kvh − uh kV♯
≤ ku − vh kV♯ + c♯ kvh − uh kVh
c♯ |ah (uh − vh , wh )|
≤ ku − vh kV♯ + sup
αh wh ∈Wh kwh kWh
c♯
= ku − vh kV♯ + kδh (vh )kWh′
αh
c♯ ω♯h
≤ ku − vh kV♯ + ku − vh kV♯ .
αh
Taking the infimum over vh ∈ Vh yields (27.6).
Example 27.6 (Simple setting). In the setting of Example 27.4, we can equip Vh and V♯ with
the norm k·kV , so that c♯ = 1. Since ω♯h = kak, the error estimate (27.6) coincides with the error
estimate in Lemma 26.14.
Remark 27.7 (Literature). A general framework for the error analysis of nonconforming meth-
ods for elliptic PDEs can be found in Veeser and Zanotti [373]. This framework introduces a
different notion of consistency and leads to quasi-optimal error estimates in the k·kV -norm with-
out any smoothness assumption on the exact solution u ∈ V (or equivalently for all data ℓ ∈ V ′ ),
i.e., the space Vs and the norm k·kV♯ are not invoked. This remarkable result is achieved at the
expense of a specific design of the discrete form ℓh . We also refer the reader to the gradient dis-
cretization method discussed in Droniou et al. [172] which can be used to analyze nonconforming
methods.
27.2.4 Error estimate using two norms

It turns out that the assumption (27.5) on the k·kV♯ -norm cannot be satisfied when one considers the
approximation of first-order PDEs using stabilization techniques. A more general setting consists
of introducing a second norm on V♯ , say k·kV♭ , and assuming that there exists a real number c♭ s.t.
kvh kV♭ ≤ c♭ kvh kVh , ∀vh ∈ Vh , kvkV♭ ≤ c♭ kvkV♯ , ∀v ∈ V♯ , (27.7)
where k·kVh is the stability norm on Vh used in (27.1) and k·kV♯ is the consistency/boundedness
norm on V♯ used in (27.4).
Lemma 27.8 (Error estimate). Assume the following: (i) Stability, i.e., (27.1) holds true; (ii)
Consistency/boundedness, i.e., u ∈ Vs and (27.4) holds true. Assume that (27.7) holds true. Then
we have
ω♯h
ku − uh kV♭ ≤ c♭ 1 + inf ku − vh kV♯ . (27.8)
αh vh ∈Vh
Proof. The proof is similar to that of Lemma 27.5. Owing to the assumptions, we infer that for
all vh ∈ Vh ,
ku − uh kV♭ ≤ ku − vh kV♭ + kvh − uh kV♭

≤ c♭ ku − vh kV♯ + c♭ kvh − uh kVh
c♭ |ah (uh − vh , wh )|
≤ c♭ ku − vh kV♯ + sup
αh wh ∈Wh kwh kWh
c♭
= c♭ ku − vh kV♯ + kδh (vh )kWh′
αh
c ω♯h
≤ c♭ ku − vh kV♯ + ♭ ku − vh kV♯ .
αh
Taking the infimum over vh ∈ Vh yields (27.8).
Remark 27.9 (Lemma 27.5 vs. Lemma 27.8). Lemma 27.5 estimates the approximation error
by the best-approximation error using the same norm k·kV♯ . We say that this estimate is quasi-
optimal over the whole computational range. In contrast, Lemma 27.8 estimates the approximation
error in the k·kV♭ -norm by the best-approximation error in the stronger k·kV♯ -norm. We will see
numerous examples where the best-approximation errors in both norms actually exhibit the same
decay rate in terms of the meshsize h ∈ H for smooth solutions. In this situation, we say that the
error estimate from Lemma 27.8 is quasi-optimal in the asymptotic range.
27.2.5 Convergence
We are now ready to state a convergence result. The last missing ingredient that we introduce
now is approximability.
Corollary 27.10 (Convergence). We have limh→0 ku − uhkV♯ = 0 in the setting of Lemma 27.5
and limh→0 ku − uh kV♭ = 0 in the setting of Lemma 27.8, provided the following properties hold
true:
(i) Uniform stability: αh ≥ α0 > 0 for all h ∈ H;
(ii) Uniform consistency/boundedness: ω♯h ≤ ω♯0 < ∞ for all h ∈ H;
(iii) Approximability: limh→0 (inf vh ∈Vh kv − vh kV♯ ) = 0 for all v ∈ Vs .
Proof. Direct consequence of the assumptions.
27.3 Two simple examples

This section presents two one-dimensional examples illustrating how to use the above error esti-
mates: (i) a boundary penalty method applied to an elliptic PDE where Lemma 27.5 is applied;
(ii) a stabilized approximation applied to a first-order PDE where Lemma 27.8 is applied.
27.3.1 Boundary penalty method for an elliptic PDE

Consider the PDE −u′′ = f in D := (0, 1) with u(0) = u(1) = 0, f ∈ L2 (D). The trial and test
R1
spaces are V = W := H01 (D). The corresponding bilinear and linear forms are a(v, w) := 0 v ′ w′ dt
R1
and ℓ(w) := 0 f w dt. Consider the standard Galerkin approximation using as discrete trial and
test spaces the spaces Vh = Wh built using continuous P1 Lagrange finite elements on a uniform
mesh Th of step h ∈ H. We do not enforce any boundary condition on Vh . As a result, the
approximation setting is nonconforming. Let us define the discrete forms
Z 1
ah (vh , wh ) := vh′ wh′ dt − (vh′ (1)wh (1) − vh′ (0)wh (0)) + h−1 (vh (1)wh (1) + vh (0)wh (0)),
0
Z 1
ℓh (wh ) := f wh dt.
0
One can show that coercivity holds true with the stability norm
kvh k2Vh := kvh′ k2L2 (D) + h−1 |vh (0)|2 + h−1 |vh (1)|2 ,
i.e., ah (vh , vh ) ≥ α0 kvh k2Vh with α0 := 83 for all vh ∈ Vh ; see Exercise 27.2 and Chapter 37.
Let us perform the error analysis using Lemma 27.5. The assumption u ∈ Vs := H 2 (D)∩H01 (D)
is natural here since f ∈ L2 (D) and −u′′ = f . We equip the space V♯ := Vs + Vh with the norm
kvk2V♯ := kv ′ k2L2 (D) + h−1 |v(0)|2 + h−1 |v(1)|2 + h|v ′ (0)|2 + h|v ′ (1)|2 .
(Recall that H 2 (D) ֒→ C 1 (D) in one dimension.) Using a discrete trace inequality shows that the
norms k·kVh and k·kV♯ are equivalent on Vh uniformly w.r.t. h ∈ H. Hence, (27.5) holds true. It
remains to establish consistency/boundedness. Since u ∈ H 2 (D), integrating by parts leads to
Z 1 Z 1
ℓh (wh ) = − u′′ wh dt = u′ wh′ dt − (u′ (1)wh (1) − u′ (0)wh (0)),
0 0
so that letting η := u − vh and since u(0) = u(1) = 0, we obtain
hδh (vh ), wh iVh′ ,Vh = ℓh (wh ) − ah (vh , wh )

Z 1
= η ′ wh′ dt − (η ′ (1)wh (1) − η ′ (0)wh (0)) + h−1 (η(1)wh (1) + η(0)wh (0)).
0
Using the Cauchy–Schwarz inequality, we conclude that (27.4) holds true with ω♯h = 1. In conclu-
sion, Lemma 27.5 implies that
ku − uh kV♯ ≤ c inf ku − vh kV♯ . (27.9)

vh ∈Vh
Since u ∈ H 2 (D), we use the approximation properties of finite elements to obtain inf vh ∈Vh ku −
vh kV♯ ≤ ch|u|H 2 (D) , so that
ku − uh kV♯ ≤ c h|u|H 2 (D) . (27.10)
This shows that the error in the k·kV♯ -norm tends to zero at rate h.
27.3.2 Stabilized approximation of a first-order PDE

Consider the PDE u′ = f in D := (0, 1) with u(0) = 0 and f ∈ L2 (D). Following §24.2.2, we
consider the L2 -based weak formulation with the trial space V := {v ∈ H 1 (D) | v(0) = 0} and
R1 R1
the test space W := L2 (D). The exact forms are a(v, w) := 0 v ′ w dt and ℓ(w) := 0 f w dt. The
model problem consists of seeking u ∈ V such that a(u, w) = ℓ(w) for all w ∈ W. This problem is
well-posed; see Exercise 25.9.
Consider the standard Galerkin approximation using as discrete trial and test spaces the space
Vh built by using continuous P1 Lagrange finite elements on a uniform mesh Th of step h ∈ H and
by enforcing the boundary condition vh (0) = 0. The discrete problem consists of seeking uh ∈ Vh
such that a(uh , wh ) = ℓ(wh ) for all wh ∈ Vh . (The reader is invited to verify that the resulting
linear system is identical to that obtained with centered finite differences.) The approximation
setting is conforming since Vh ⊂ V and Wh = Vh ⊂ W. Unfortunately, it turns out that the
bilinear form a is not uniformly stable on Vh ×Vh . Indeed, one can show (see Exercise 27.3) that
there are 0 < c1 ≤ c2 s.t. for all h ∈ H,
|a(vh , wh )|
c1 h ≤ inf sup =: αh ≤ c2 h. (27.11)
vh ∈Vh wh ∈Vh kvh kH 1 (D) kwh kL2 (D)
This result shows that the above naive Galerkin approximation of first-order PDEs cannot produce
optimal error estimates, even though it yields an invertible linear system (c1 6= 0). In practice,
this problem manifests itself through the presence of spurious wiggles in the approximate solution.
To circumvent this difficulty, let us define the discrete bilinear and linear forms
Z 1 Z 1
ah (vh , wh ) := (vh′ wh + hvh′ wh′ ) dt, ℓh (wh ) := f (wh + hwh′ ) dt,
0 0
for all vh , wh ∈ Vh . Referring to Exercise 27.4 (see also §57.3 and §61.4), one can establish the
uniform inf-sup condition
|ah (vh , wh )|
inf sup ≥ α0 > 0, (27.12)
vh ∈Vh wh ∈Vh kvh kVh kwh kVh
with the stability norm
kvh k2Vh := ℓ−1 2 2 ′ 2

D kvh kL2 (D) + |vh (1)| + hkvh kL2 (D) ,
where we introduced the length scale ℓD := 1 to be dimensionally consistent.

Let us perform the error analysis using Lemma 27.8. We set Vs := V so that V♯ = V + Vh = V,
and we equip V♯ with the following norms (recall that H 1 (D) ֒→ C 0 (D) in one dimension):
−1
kvkV♭ := ℓD kvk2L2 (D) + |v(1)|2 + hkv ′ k2L2 (D) , (27.13)
−1
kvk2V♯ := h kvk2L2 (D) + |v(1)| + 2
hkv ′ k2L2 (D) , (27.14)
so that (27.7) holds true with c♭ := 1 since h ≤ ℓD . Notice that there is no uniform constant c♯ s.t.
(27.5) holds true, i.e., we cannot apply Lemma 27.5. To apply Lemma 27.8, it remains to establish
consistency/boundedness. Since u′ = f in D, letting η := u − vh , we infer that
hδh (vh ), wh iVh′ ,Vh = ℓh (wh ) − ah (vh , wh )

Z 1 Z 1
′
= f (wh + hwh ) dt − (vh′ wh + hvh′ wh′ ) dt
0 0
Z 1
= (η ′ wh + hη ′ wh′ ) dt =: T1 + T2 .
0
Integrating by parts, we obtain

Z 1 Z 1
T1 = η ′ wh dt = − ηwh′ dt + η(1)wh (1),
0 0
since η(0) = 0. Using the Cauchy–Schwarz inequality, we infer that

1 1
|T1 | ≤ h− 2 kηkL2 (D) h 2 kwh′ kL2 (D) + |η(1)| |wh (1)| ≤ kηkV♯ kwh kVh ,
1 1
|T2 | ≤ h 2 kη ′ kL2 (D) h 2 kwh′ kL2 (D) ≤ kηkV♯ kwh kVh ,
which shows that (27.4) holds true with ω♯h := 2. In conclusion, Lemma 27.8 implies that
ku − uh kV♭ ≤ c inf ku − vh kV♯ . (27.15)

vh ∈Vh
Assuming that u ∈ H 1+r (D), r ∈ [0, 1], we use the approximation properties of finite elements to
1
obtain inf vh ∈Vh ku − vh kV♯ ≤ ch 2 +r |u|H 1+r (D) , so that
1
ku − uh kV♭ ≤ c h 2 +r |u|H 1+r (D) . (27.16)
The error estimate (27.15) is quasi-optimal in the asymptotic range since the best-approximation
errors in the k·kV♭ - and k·kV♯ -norms converge to zero at the same rate (see Remark 27.9 for the
terminology).
27.4 Strang’s lemmas

We review in this section results due to Strang [358] and often called Strang’s lemmas in the
literature. These lemmas are historically important for the development of the analysis of finite
element methods. In this book, we are going to use systematically Lemma 27.5 and Lemma 27.8
and only use Strang’s lemmas at a few instances.
There are two Strang’s lemmas: the first one is tailored to conforming approximations but allows
for ah 6= a and ℓh 6= ℓ, and the second one can be applied to nonconforming approximations.
Both lemmas can be seen as variants of Lemma 27.5 and Lemma 27.8, where the consistency
error kδh (vh )kWh′ is further decomposed by adding/subtracting some terms so as to separate the
approximation of a by ah and the approximation of ℓ by ℓh (these contributions are sometimes
called consistency error in the literature) from the best-approximation error of u by a function in
Vh .
Remark 27.11 (Consistency). One should bear in mind that the notion of consistency in
Strang’s lemmas is somewhat arbitrary. This is illustrated in §27.4.3, where each lemma leads to a
different notion of consistency for the same approximation method. We think that it is preferable
to use the quantity kδh (vh )kWh′ defined in (27.3) as the only notion of consistency. This is the
convention we are going to follow in the rest of the book.
27.4.1 Strang’s first lemma

Strang’s first lemma is tailored to conforming approximations. It has been devised to estimate
the error due to quadratures when approximating elliptic PDEs by H 1 -conforming finite elements
(see §33.3).
Lemma 27.12 (Strang 1). Assume: (i) Conformity: Vh ⊂ V and Wh ⊂ W, and set Vs := V so
that V♯ := V + Vh = V ; (ii) Stability: (27.1) holds true; (iii) Boundedness: the sesquilinear form
a is bounded on V ×Wh , and set
|a(v, wh )|
kak♯h := sup sup , (27.17)
v∈V wh ∈Wh kvkV♯ kwh kWh
where the norm k·kV♯ satisfies (27.5). Let δhst1 : Vh → Wh′ be defined by
hδhst1 (vh ), wh iWh′ ,Wh := ℓh (wh ) − ℓ(wh ) + a(vh , wh ) − ah (vh , wh ). (27.18)
Then the following holds true:

kak♯h c♯ st1
ku − uh kV♯ ≤ inf 1 + c♯ ku − vh kV♯ + kδh (vh )kWh′ . (27.19)
vh ∈Vh αh αh
Proof. Proceeding as in the proof of Lemma 27.5 leads to
c♯
ku − uh kV♯ ≤ ku − vh kV♯ + kδh (vh )kWh′ .
αh
We write the consistency error as follows:
hδh (vh ), wh iWh′ ,Wh := ℓh (wh ) − ah (vh , wh )

= ℓh (wh ) − ℓ(wh ) + a(u, wh ) − ah (vh , wh )
= ℓh (wh ) − ℓ(wh ) + a(u, wh ) − ah (vh , wh ) + [a(vh , wh ) − a(vh , wh )]
= hδhst1 (vh ), wh iWh′ ,Wh + a(u − vh , wh ),
where we used that a(u, wh ) = ℓ(wh ) since Wh ⊂ W. Using the triangle inequality and the bound-
edness property (27.17), we infer that
kδh (vh )kWh′ ≤ kδhst1 (vh )kWh′ + kak♯h ku − vh kV♯ .
Rearranging the terms leads to the expected estimate.

Remark 27.13 (Comparison). In the original statement of Strang’s first lemma, one takes
k·kV♯ := k·kV , and one equips Vh with the k·kV -norm, so that the error estimate (27.19) holds true
with kak♯h := kak. Moreover, the terms ℓh (wh ) − ℓ(wh ) and a(vh , wh ) − ah (vh , wh ) composing
hδhst1 (vh ), wh iWh′ ,Wh are separated, and the term kℓh − ℓkWh′ is taken out of the infimum over
vh ∈ Vh in (27.19). The original statement is sufficient to analyze quadrature errors in the H 1 -
conforming approximation of elliptic PDEs, but as illustrated in §27.4.3, Strang’s first lemma is
not well adapted to analyze stabilized finite element approximations of first-order PDEs, since in
this case one needs to invoke the two norms k·kV♭ and k·kV♯ defined in (27.13)-(27.14).
Remark 27.14 (Nonconforming setting). It is possible to derive an error estimate in the

spirit of Strang’s first lemma in some nonconforming settings. Following Gudi [226], the idea is to
introduce an operator T : Wh → W acting on the discrete test functions. This operator can be
built using the averaging operators analyzed in §22.2. We refer the reader to [226] and Exercise 27.5
for error estimates obtained with this technique.
27.4.2 Strang’s second lemma

Contrary to Strang’s first lemma, the second lemma is applicable to nonconforming approximation
settings.
Lemma 27.15 (Strang 2). Let Vs := V so that V♯ := V + Vh . Assume: (i) Stability: (27.1)
holds true; (ii) Bounded extendibility: There exists a bounded sesquilinear form a♯ on V♯ ×Wh that
extends ah originally defined on Vh ×Wh , i.e., a♯ (vh , wh ) = ah (vh , wh ) for all (vh , wh ) ∈ Vh ×Wh
and
|a♯ (v, wh )|
ka♯ k♯h := sup sup < ∞, (27.20)
v∈V♯ wh ∈Wh kvk V♯ kwh kWh
with a norm k·kV♯ satisfying (27.5). The following holds true:

ka♯ k♯h c♯ st2
ku − uh kV♯ ≤ 1 + c♯ inf ku − vh kV♯ + kδ (u)kWh′ , (27.21)
αh vh ∈Vh αh h
with hδhst2 (u), wh iWh′ ,Wh := ℓh (wh ) − a♯ (u, wh ).
Proof. The starting point is again the bound

c♯
ku − uh kV♯ ≤ ku − vh kV♯ + kδh (vh )kWh′ .
αh
Now we write the consistency error as follows:
hδh (vh ), wh iWh′ ,Wh := ℓh (wh ) − ah (vh , wh ) = ℓh (wh ) − a♯ (vh , wh )

= ℓh (wh ) − a♯ (vh , wh ) + [a♯ (u, wh ) − a♯ (u, wh )]
= a♯ (u − vh , wh ) + hδhst2 (u), wh iWh′ ,Wh .
Using the triangle inequality and the boundedness property (27.20), we infer that
kδh (vh )kWh′ ≤ kδhst2 (u)kWh′ + ka♯ k♯h ku − vh kV♯ .
Rearranging the terms leads to the expected estimate.

Remark 27.16 (Strong consistency, quasi-optimality). Recalling the Galerkin orthogonality

terminology introduced in the context of conforming approximations (see §26.3.1), we say that
strong consistency holds true if δhst2 (u) vanishes identically on Wh , i.e., if the exact solution satisfies
the discrete equations rewritten using the extended sesquilinear form a♯ . In this case, (27.21) leads
to a quasi-optimal error estimate.
Remark 27.17 (Bounded extendibility). Lemma 27.15 has been originally devised to ana-
lyze the Crouzeix–Raviart approximation of elliptic PDEs (see Chapter 36). In this context, the
bounded extendibility assumption is indeed reasonable. However, it is no longer satisfied if a bound-
ary penalty method or a discontinuous Galerkin method is used (see Chapters 37 and 38). For
such methods, it is possible to recover the bounded extendibility assumption (and to prove strong
consistency) provided the exact solution satisfies an additional smoothness assumption which is
typically of the form u ∈ H 1+r (D) with regularity pickup r > 12 . We will see that the error analysis
based on Lemma 27.5 is more general since it only requires a regularity pickup r > 0 in the Sobolev
scale. There are also other situations where the bounded extendibility assumption is simply not
reasonable, e.g., when considering quadratures using point values or for stabilization techniques
based on a two-scale hierarchical decomposition of the discrete spaces that is not meaningful for
nondiscrete functions (see Chapter 59).
27.4.3 Example: first-order PDE

Let us consider the first-order PDE and the discrete setting introduced in §27.3.2, and let us briefly
illustrate how to estimate the error using Strang’s lemmas in this context. Using Strang’s first
lemma, one finds that
hδhst1 (vh ), wh iVh′ ,Vh := ℓh (wh ) − ℓ(wh ) + a(vh , wh ) − ah (vh , wh )

Z 1 Z 1
= h(f − vh′ )wh′ dt = hη ′ wh′ dt,
0 0
1
since f = u′ and η := u − vh , so that kδhst1 (vh )kVh′ ≤ h kη ′ kL2 (D) ≤ kηkV♯ , where k·kV♯ is
2
1 1 1 1
defined in (27.14). One also has kak♯h ≤ ℓD
2
h− 2 . In conclusion, ku − uh kV♯ ≤ (1 + α−1
0 (ℓD h
2 −2
+
1
1)) inf vh ∈Vh ku − vh kV♯ , which yields the suboptimal error estimate ku − uh kV♯ ≤ chr ℓD 2
|u|H 1+r (D)
for all r ∈ [0, 1] (compare with (27.16)). Using instead Strang’s second lemma, one finds that
hδhst2 (u), wh iVh′ ,Vh := ℓh (wh ) − a♯ (u, wh ) = 0 for all wh ∈ Vh , i.e., strong consistency holds true,
1
and one obtains again the suboptimal error estimate ku − uhkV♯ ≤ chr ℓD
2
|u|H 1+r (D) . This example
shows that the two Strang lemmas may lead to different notions of consistency, and, if applied
blindly, they may yield suboptimal error estimates.
Exercises
Exercise 27.1 (Error identity). Assume stability, i.e., (27.1) holds true. Let V♯ be defined
in (27.2) and equip this space with a norm k·kV♭ s.t. there is c♭ s.t. kvh kV♭ ≤ c♭ kvh kVh for all
vh ∈ Vh . Prove that

c♭
ku − uh kV♭ = inf ku − vh kV♭ + kδh (vh )kWh′ .
vh ∈Vh αh
2
0 −β
Exercise 27.2 (Boundary penalty). (i) Prove that x2 −2βxy +η0 y 2 ≥ η1+η 0
(x2 +y 2 ) for all real
numbers x, y, η0 ≥ 0 and β ≥ 0. (ii) Using the notation of §27.3.1, prove that ah (vh , vh ) ≥ 83 kvh k2Vh
1
for all vh ∈ Vh . (Hint : prove that |vh′ (0)vh (0)| ≤ kvh′ kL2 (0,h) h− 2 |vh (0)|.)
Exercise 27.3 (First-order PDE). The goal is to prove (27.11). (i) Prove that
1 |a(vh , wh )| √ − 1
h− 2 kG(vh )kℓ2 (RI ) ≤ sup ≤ 6h 2 kG(vh )kℓ2 (RI ) ,
wh ∈Vh kwh kL2 (D)
where Gi (vh ) := a(vh , ϕi ) for all i ∈ {1:I} with I := dim(Vh ). (Hint : use Simpson’s rule to
compare Euclidean norms of component vectors and L2 -norms of functions.) (ii) Assume that I
is even (the odd case is treated similarly). Prove that αh ≤ c2 h. (Hint : consider the oscillating
function vh s.t. vh (x2i ) := 2ih for all i ∈ {1: I2 } and vh (x2i+1 I
P ) := 1 for all i ∈ {0: 2 −1}.) (iii) Prove
that αh ≥ c1 h. (Hint : prove that maxi∈{1: I} |vh (xi )| ≤ 2 k∈{1: I} |Gk (vh )|.) (iv) Prove that
|a(vh , wh )|
inf sup ≥ α0 > 0
vh ∈Vh wh ∈Wh kvh kW 1,1 (D) kwh kL∞ (D)
with Wh := {wh ∈ L∞ (D) | ∀i ∈ {0:I−1}, wh|[xi ,xi+1 ] ∈ P0 }. (Hint : see Proposition 25.19.)
Exercise 27.4 (GaLS 1D). The goal is to prove (27.12). Let vh ∈ Vh . (i) Compute ah (vh , vh ).
(ii) Let ζ(x) := −2x/ℓD , set ζh := Ihb (ζ), and show that ah (vh , Jhav (ζh vh )) ≥ 12 ℓ−1 2
D kvh kL2 (D) −
c1 a(vh , vh ) uniformly w.r.t. h ∈ H, Jh is the averaging operator defined in (22.9), and Ihb is the
av
L2 -projection on the functions that are piecewise constant over the mesh. (iii) Prove (27.12).
(Hint : use the test function zh := 2Jhav (ζh vh ) + 2(c1 + 1)vh .)
Exercise 27.5 (Nonconforming Strang 1). Let T : Wh → W ∩ Wh . Let Vs := V so that
V♯ := V + Vh , and assume that V♯ is equipped with a norm k·kV♯ satisfying (27.5). (i) Assume that
ah can be extended to Vh ×(W + Wh ). Assume that there is kak♯h s.t. consistency/boundedness
holds true in the form |a(u, T (wh )) − ah (vh , T (wh ))| ≤ kak♯h ku − vh kV♯ kwh kWh . Prove that

kak♯h c♯ st1
ku − uh kV♯ ≤ inf 1 + c♯ ku − vh kV♯ + kδ̂h (vh )kWh′ ,
vh ∈Vh αh αh
with kδ̂hst1 (vh )kWh′ := kℓh − ℓ ◦ T + ah (vh , T (·)) − ah (vh , ·)kWh′ . (Hint : add/subtract ah (vh , T (wh )).)
(ii) We now derive another error estimate that avoids extending ah but restricts the discrete trial
functions to Vh ∩ V (this is reasonable provided the subspace Vh ∩ V has approximation properties
that are similar to those of Vh ). Assuming that there is kakV ×Wh s.t. boundedness holds true in
the form |a(u − vh , T (wh ))| ≤ kakV ×Wh ku − vh kV♯ kwh kWh , prove that

kakV ×Wh c♯ st1
ku − uh kV♯ ≤ inf 1 + c♯ ku − vh kV♯ + kδ̌ (vh )kWh′ ,
vh ∈Vh ∩V αh αh h
with kδ̌hst1 (vh )kWh′ := kℓh − ℓ ◦ T + a(vh , T (·)) − ah (vh , ·)kWh′ . (Hint : add/subtract a(vh , T (wh )).)
Exercise 27.6 (Orthogonal projection). Consider the setting of Exercise 25.4 with real vector
spaces and coercivity with ξ := 1 for simplicity. Let u be the unique element in V such that
a(u, v − u) ≥ ℓ(v − u) for all v ∈ U . Let Vh be a finite-dimensional subspace of V, and let Uh be
a nonempty, closed, and convex subset of Vh . We know from Exercise 25.4 that there is a unique
uh in Vh such that a(uh , vh − uh ) ≥ ℓ(vh − uh ) for all vh ∈ Uh . (i) Show that there is c1 (u) such
that for all (v, vh ) ∈ U ×Vh ,

ku − uh k2V ≤ c1 (u) ku − vh kV + kuh − vkV + ku − uh kV ku − vh kV .
(Hint : prove αku − uh k2V ≤ a(u, v − uh ) + ℓ(uh − v) + a(uh , vh − u) + ℓ(u − vh ).) (ii) Show that
there is c2 (u) such that
21
ku − uh kV ≤ c2 (u) inf ku − vh kV + ku − vh k2V + inf kuh − vkV .
vh ∈Uh v∈U
Chapter 28
Linear algebra
In this chapter, we first show that the discrete problem generated by the Galerkin approximation
can be reformulated as a linear system once bases for the discrete trial space and the discrete test
space are chosen. Then, we investigate important properties of the system matrix, which is called
stiffness matrix, and we also introduce the mass matrix, which is relevant when computing L2 -
orthogonal projections. We derive various estimates on the norm, the spectrum, and the condition
number of both matrices. Finally, we give a brief overview of direct and iterative solution methods
for linear systems.
28.1 Stiffness and mass matrices

Recall that the discrete problem (26.3) consists of seeking uh ∈ Vh s.t. ah (uh , wh ) = ℓh (wh ) for
all wh ∈ Wh , where ah is sesquilinear (bilinear in the real case) and ℓh is antilinear (linear in the
real case). We assume that the discrete problem is well-posed, i.e., the inf-sup condition (26.5a)
holds true and dim(Vh ) = dim(Wh ) =: I. We show in this section that the discrete problem can
be reformulated as a linear system once bases for Vh and Wh are chosen.
28.1.1 Main definitions

Let {ϕi }i∈{1: I} be a basis of Vh and {ψi }i∈{1: I} be a basis of Wh . Let Rϕ : CI → V
Ph be the iso-
morphism that reconstructs functions in Vh from coordinate vectors, i.e., Rϕ (V) := i∈{1: I} Vi ϕi
for all V := (Vi )i∈{1: I} ∈ CI . A similar isomorphism Rψ is considered for the discrete space Wh
equipped with the basis {ψi }i∈{1: I} . These isomorphisms are instrumental to go back and forth
from the functional viewpoint to the algebraic viewpoint.
Let A ∈ CI×I be the stiffness matrix with entries
Aij := ah (ϕj , ψi ), ∀i, j ∈ {1:I}, (28.1)
(note the position of the indices i and j in (28.1)) and let B ∈ CI be the column vector with
components
Bi := ℓh (ψi ), ∀i ∈ {1:I}. (28.2)
The link between the discrete sesquilinear form ah and the stiffness matrix A can be formalized
as follows:
WH AV = ah (Rϕ (V), Rψ (W)), ∀V, W ∈ CI . (28.3)
46 Chapter 28. Linear algebra
Similarly, we have WH B = ℓh (Rψ (W)). The above definitions imply that
[ uh solves (26.3)) ] ⇐⇒ [ AU = B ] with uh := Rϕ (U). (28.4)
We observe that the number of equations and the number of unknowns in the linear system AU = B
is equal to dim(Wh ) and dim(Vh ), respectively. Thus, the linear system is square if and only if
dim(Vh ) = dim(Wh ).
Remark 28.1 (Well-posedness). One easily verifies the following: (i) The inf-sup condition (26.5a)
holds true if and only if ker(A) = {0}. (ii) The condition (26.8) is equivalent to rank(A) = I.
(iii) The sesquilinear form ah is coercive on Vh if and only if the matrix A is Hermitian positive
definite, i.e., ℜ(ξVH AV) ≥ 0 for all V ∈ CI and VH AV = 0 implies that V = 0.
In many applications, Vh and Wh are discrete subspaces of L2 (D). It is then meaningful to
define the following mass matrices:
Mϕ,ij := (ϕj , ϕi )L2 (D) , Mψ,ij := (ψj , ψi )L2 (D) , ∀i, j ∈ {1:I}. (28.5)
Notice that both matrices are Hermitian positive definite. When the basis functions are real-
valued, these matrices are symmetric positive definite. The mass matrices are useful to evaluate
L2 -orthogonal projections. For instance, consider the L2 -orthogonal projection ΠVh : L2 (D) → Vh
such that for all v ∈ L2 (D), ΠVh (v) is the unique function in Vh satisfying (ΠVh (v)−v, yh )L2 (D) = 0
for all yh ∈ Vh . One easily verifies that ΠVh (v) = Rϕ (X) where X ∈ CI solves the linear system
Mϕ X = Y with right-hand side vector Y := ((v, ϕi )L2 (D) )i∈{1: I} ∈ CI .
R
Example 28.2 (P1 Lagrange, 1D). Consider the bilinear form a(v, w) := D v ′ w′ dx for all
v, w ∈ H01 (D) with D := (0, 1). Consider a uniform mesh Th of D specified by its vertices
1
xi := ih for all i ∈ {0:(I + 1)} with h := I+1 . Let Vh be spanned by piecewise affine functions
on Th vanishing at the two endpoints of D. The global shape functions in Vh are the hat basis
functions s.t. ϕi (x) := 1 − h−1 |x − xi | for x ∈ [xi−1 , xi+1 ] and ϕi (x) := 0 otherwise, for all
i ∈ {1:I}. With Wh := Vh and ah := a|Vh ×Vh , the stiffness matrix A ∈ RI×I is tridiagonal.
The diagonal entries are equal to 2h−1 and the upper- and lower-diagonal entries are equal to
−h−1 . We write A = h−1 tridiag(−1, 2, −1). The mass matrix M ∈ RI×I is also tridiagonal with
M = h6 tridiag(1, 4, 1).
28.1.2 Static condensation

The idea behind static condensation is that one can eliminate from the linear system in (28.4) all
the unknowns corresponding to the global basis functions that are supported in one mesh cell only.
This elimination is a cost-effective approach to reduce the size of the linear system since it can be
realized by performing only local computations in each mesh cell.
For simplicity, we present the technique in the case where Vh = Wh . Recall that for every
cell K ∈ Th , the degrees of freedom (dofs) of the finite element (K, PK S , ΣK ) are enumerated by
using the set N . We consider the partition N = N ◦ ∪ N ∂ where N ∂ := F ∈FK NK,F (recall that
n ∈ NK,F iff the local shape function θK,n has a nonzero γ-trace on the face F ∈ FK ; see §20.1).
Thus, n ∈ N ◦ iff the γ-trace of the local shape function θK,n vanishes on the boundary of K. Note
that both sets N ◦ and N ∂ only depend on the reference finite element. Let us now partition the
global set of dofs Ih := {1:I} in the form Ih = ITh ∪ IFh where
ITh := {i ∈ Ih | i = j dof(K, n), K ∈ Th , n ∈ N ◦ }, (28.6a)

∂
IFh := {i ∈ Ih | i = j dof(K, n), K ∈ Th , n ∈ N }, (28.6b)
where j dof is the connectivity array introduced in §19.1 (the sets ITh and IFh are disjoint owing
to the injectivity of the map j dof(K, ·); see (19.3)). We first enumerate the global dofs in ITh
(the associated global shape functions are often called bubble functions), and then we enumerate
those in IFh . This leads to the following block-decomposition of the linear system (with obvious
notation): " #
ATh Th ATh Fh UTh BTh
= . (28.7)
AFh Th AFh Fh UFh BFh
The key observation is that ATh Th is block-diagonal and each block has size card(N ◦ ), If the
method is conforming, the entries of each block are a(θK,n′ , θK,n ) for all n, n′ ∈ N ◦ , and it can
be shown that each of these small matrices is invertible. Hence, ATh Th is easy to invert, and this
leads to
UTh = −(ATh Th )−1 ATh Fh UFh + (ATh Th )−1 BTh . (28.8)
Substituting this expression into the second equation of (28.7), we infer that

AFh Fh − AFh Th (ATh Th )−1 ATh Fh UFh = B′Fh , (28.9)
where B′Fh := BFh − AFh Th (ATh Th )−1 BTh . The matrix on the left-hand side of (28.9) is called
Schur complement of ATh Th . One proceeds as follows to solve the linear system (28.7): one first
computes UFh by solving (28.9), then one computes UTh by solving (28.8). This technique is called
static condensation; see Guyan [232], Irons [252]. Static condensation makes sense only if N ◦ is
nonempty, and it is computationally effective if card(N ◦ ) ∼ card(N ). Referring to §29.1 for further
insight, we note that static condensation reduces the size of the stiffness matrix without altering
its sparsity pattern.
Example 28.3 (Lagrange elements). For Lagrange finite elements of degree k ≥ 1, the internal
◦
◦ k−1
at the nodes located inside K. Then if k ≥ d + 1, the set N is nonempty and
dofs are evaluations
card(N ) = d . Static condensation is effective when k is large.
28.2 Bounds on the stiffness and mass matrices

In this section, we introduce the notion of condition number for a nonsingular matrix, and we
derive various bounds on the spectrum and the norm of the stiffness and mass matrices.
28.2.1 Condition number

We denote by k·kℓ2 (CI ) the Euclidean norm in CI . The induced matrix norm is denoted similarly.
1 1
Recall that for every square matrix Z ∈ CI×I , we have kZkℓ2 (CI ) := ρ(Z H Z) 2 = ρ(ZZ H ) 2 , where
ρ(·) denotes the spectral radius and Z H the Hermitian transpose of Z, i.e., (Z H )ij := Zji . The
Euclidean condition number of any nonsingular matrix Z ∈ CI×I is defined by
κℓ2 (Z) := kZkℓ2 (CI ) kZ −1 kℓ2 (CI ) . (28.10)
(Condition numbers can be defined for every matrix norm induced by a vector norm.) Observe
that κℓ2 (Z) ≥ kZZ −1 kℓ2 (CI ) = 1. The Euclidean condition number of Z is the ratio of the
maximal and the minimal singular values of Z. In particular, if Z is Hermitian (or symmetric in
the real case), i.e., Z H = Z (or Z T = Z), its eigenvalues are real and κℓ2 (Z) is the ratio of the
maximal and the minimal eigenvalues of Z (in absolute value). In other words, letting σ(Z) ⊂ R
denote the spectrum of Z and setting λmin := minλ∈σ(Z) |λ| and λmax := maxλ∈σ(Z) |λ|, we have
κℓ2 (Z) := λλmax
min
.
Definition 28.4 (Ill-conditioning). A matrix Z ∈ CI×I is said to be ill-conditioned whenever
κℓ2 (Z) ≫ 1.
A large condition number often indicates that numerical difficulties are to be expected when
solving a linear system. We refer the reader to Exercise 28.6 for further insight into the influence
of the condition number on the sensitivity to perturbations and to Proposition 28.21 for further
insight into the convergence rate of iterative methods.
28.2.2 Spectrum of the mass matrix

In this section, we investigate the spectrum of the mass matrix Mϕ . The results are the same
for Mψ . Since the mass matrix is Hermitian positive definite, its spectrum lies on the positive
real half-line, i.e., σ(Mϕ ) ⊂ [0, ∞). Let µϕ ϕ
min := minµ∈σ(Mϕ ) µ and µmax := maxµ∈σ(Mϕ ) µ be the
smallest and the largest eigenvalue of Mϕ , respectively. Since V Mϕ V = kRϕ (V)k2L2 (D) , we infer
H
that
kRϕ (V)k2L2 (D) kRϕ (V)k2L2 (D)
µϕ
min = minI ≤ max = µϕ
max . (28.11)
V∈C kVk2ℓ2 (CI ) V∈CI kVk2ℓ2 (CI )
Moreover, we have
µϕ
µϕ
max = kMϕ kℓ2 (CI ) , µϕ −1 −1
min = kMϕ kℓ2 (CI ) , κℓ2 (Mϕ ) = max
. (28.12)
µϕ
min
We assume that the basis functions {ϕi }i∈{1: I} spanning Vh are finite element global shape
functions built using a sequence of affine meshes (Th )h∈H and a reference finite element with
shape functions {θbn }n∈I . In every mesh cell K ∈ Th , the local shape functions are defined by
−1 b
θK,n := ψK (θn ) for all n ∈ N , and we assume that the functional transformation is such that
ψK (v) := AK (v ◦ TK ), where TK is the geometric mapping and AK ∈ Rq×q for some integer q ≥ 1.
The global basis functions ϕi are such that
ϕj dof(K,n)|K = θK,n , ∀(K, n) ∈ Th ×N , (28.13)
where j dof is the connectivity array introduced in Chapter 19.
Proposition 28.5 (Local spectrum). Assume that the mesh Th is affine (the regularity of the
mesh sequence is not needed). Let K ∈ Th and let MK ∈ Rnsh ×nsh be the local mass matrix with
′
entries MK K K
nn′ = (θK,n′ , θK,n )L2 (K) for all n, n ∈ N . Let µmin and µmax be the smallest and the
K
largest eigenvalue of M . Then there are 0 < c♭ ≤ c♯ s.t. for all K ∈ Th and all h ∈ H,
c♭ kAK k−2 K K −1 2
ℓ2 |K| ≤ µmin ≤ µmax ≤ c♯ kAK kℓ2 |K|. (28.14)
Proof. Norm equivalence in Cnsh implies that there are 0 < b
c♭ ≤ b
c♯ s.t.
c♭ kVkℓ2 (Cnsh ) ≤ kRθb(V)kL2 (K)
b b ≤b c♯ kVkℓ2 (Cnsh ) , ∀V ∈ Cnsh ,
P b b
with the reconstructed function Rθb(V) := P Vn θn in K. Let K ∈ Th be a mesh cell and
n∈N
consider the reconstructed function RθK (V) := n∈N Vn θK,n in K. Owing to the linearity of the
map ψK , we have ψK (RθK (V)) = Rθb(V). Lemma 11.7 and the above norm equivalence imply that
1 kRθK (V)kL2 (K) 1
c♭ kAK k−1
ℓ2 |K| ≤
2 ≤ c♯ kA−1
K kℓ2 |K| ,
2
kVkℓ (C sh )
2 n
with 0 < c♭ ≤ c♯ uniform w.r.t. K ∈ Th and h ∈ H. We conclude the proof by invoking (28.11).
Proposition 28.6 (Global spectrum). Assume that the mesh sequence is shape-regular and
that kAK kℓ2 is uniformly equivalent to hrK for some r ≥ 0. Then there are 0 < c1 ≤ c2 s.t. for all
K ∈ Th and all h ∈ H,
d−2r
c1 min hK ≤ µϕ ϕ d−2r
min ≤ µmax ≤ c2 max hK . (28.15)
K∈Th K∈Th
Moreover, if the mesh sequence is quasi-uniform (see Definition 22.20), there are 0 < c′1 ≤ c′2 ,
uniform w.r.t. h ∈ H, such that
c′1 hd−2r ≤ µϕ ϕ ′ d−2r

min ≤ µmax ≤ c2 h , (28.16)
c′2
implying the bound κℓ2 (Mϕ ) ≤ c′1 .
Proof. For all V ∈ CI and all K ∈ Th , let VK ∈ Cnsh be the components of V associated with the
local dofs in K, i.e., VnK := Vj dof(K,n) for all n ∈ N . The regularity of the mesh sequence implies
that there is c s.t. card({(K,Pn) ∈ Th ×N | i = j dof(K, n)}) ≤ c for all i ∈ {1:I} and all h ∈ H.
We infer that kVk2ℓ2 (CI ) ≤ K∈Th kVK k2ℓ2 (Cnsh ) ≤ ckVk2ℓ2 (CI ) . Since Rϕ (V)|K = RθK (VK ) owing
to (28.13), we infer that
X X
kRϕ (V)k2L2 (D) = kRθK (VK )k2L2 (K) ≥ c♭ kAK k−2 K 2
ℓ2 |K|kV kℓ2 (Cnsh )
K∈Th K∈Th
X
≥ c♭ min kAK k−2
ℓ2 |K| kVK k2ℓ2 (Cnsh ) ≥ c♭ min hd−2r
K kVk2ℓ2 (CI ) ,
K∈Th K∈Th
K∈Th
by our assumption on kAK kℓ2 and since |K| is uniformly equivalent to hdK for shape-regular mesh
sequences. This yields the lower bound in (28.15). The upper bound is proved similarly using that
−1
kAK kℓ2 kAK kℓ2 is uniformly bounded. The upper and lower bounds in (28.16) follow from the
quasi-uniformity assumption on the mesh sequence.
Example 28.7 (P1 Lagrange, 1D). Recall from Example 28.2 the mass matrix
h
M= 6 tridiag(1, 4, 1).
1
Letting η := I+1 , one can show that the eigenvalues of a I×I tridiagonal matrix tridiag(b, a, b),
a, b ∈ R, are λl := a + 2b cos(πlη) with associated eigenvectors Vl := (sin(πlmη))m∈{1: I} , for
all l ∈ {1:I}. Hence, the eigenvalues of the mass matrix are µl = 13 h (2 + cos(πlη)), for all
l ∈ {1:I}. This implies that µmin = µI = 13 h (2 − cos(πη)) ≈ 31 h if I is large, and µmax = µ1 =
1
3 h (2 + cos(πη)) ≈ h if I is large.
Remark 28.8 (Exponent r). For Lagrange and canonical hybrid elements, one has r = 0
in (28.15) since ψK is just the pullback by TK . For Nédélec and Raviart–Thomas elements, one
has r = 1 and r = 2, respectively.
Remark 28.9 (Broken spaces). If Vh is a broken finite element space, the support of the basis
functions ϕi is localized to a single mesh cell. This implies that the mass matrix Mϕ is block-
diagonal, each block being of size nsh . Thus, Mϕ is easy to invert. Although this special structure
is lost if Vh is a conforming finite element space, the mass matrix Mϕ remains in general easy to
invert. In particular, Proposition 28.6 shows that Mϕ is well-conditioned (at least on quasi-uniform
mesh sequences).
Remark 28.10 (Literature). Other bounds on the eigenvalues of the mass matrix are derived
in Wathen [389].
28.2.3 Bounds on the stiffness matrix

Let us introduce the following real numbers:
kAVkℓ2 (CI ) |WH AV|
αℓ2 := inf = inf sup , (28.17a)
V∈CI kVkℓ2 (CI ) V∈CI W∈CI kVkℓ2 (CI ) kWkℓ2 (CI )
kAVkℓ2 (CI ) |WH AV|

ωℓ2 := sup = sup sup . (28.17b)
V∈CI kVkℓ2 (CI ) V∈CI W∈CI kVkℓ2 (CI ) kWkℓ2 (CI )
We have ωℓ2 = kAkℓ2 (CI ) , and one can verify that αℓ2 = kA−1 k−1
ℓ2 (CI )
(see Exercise 28.2). The real
numbers αℓ2 and ωℓ2 are called smallest and largest singular values of A, respectively. Our goal is
derive upper and lower bounds on ωℓ2 and αℓ2 . To this purpose, we introduce the following real
numbers:
|ah (vh , wh )|
αL2 := inf sup , (28.18a)
vh ∈Vh wh ∈Wh kvh kL2 (D) kwh kL2 (D)
|ah (vh , wh )|
ωL2 := sup sup . (28.18b)
vh ∈Vh wh ∈Wh kvh kL2 (D) kwh kL2 (D)
Note that we are not using the natural norms in Vh and Wh but the L2 -norm.
Proposition 28.11 (Bounds on A and A−1 ). The following holds true:
1 1
(µϕ ψ 2 ϕ ψ
min µmin ) ωL2 ≤ ωℓ2 = kAkℓ2 (CI ) ≤ (µmax µmax ) ωL2 ,
2
1 1
(µϕ ψ
max µmax )
− 2 −1
αL2 ≤ α−1
ℓ2 = kA
−1
kℓ2 (CI ) ≤ (µϕ ψ
min µmin )
− 2 −1
αL2 ,
where the µ’s denote (with obvious notation) the minimal/maximal eigenvalues of the mass matrices
Mϕ and Mψ .
Proof. Since Rϕ and Rψ are isomorphisms, we infer that
|WH AV|
αL2 = inf sup ,
V∈CI W∈CI kRϕ (V)kL2 (D) kRψ (W)kL2 (D)
|WH AV|
ωL2 = sup sup .
V∈CI W∈CI kRϕ (V)kL2 (D) kRψ (W)kL2 (D)
|WH AV| |WH AV|
Let ξℓ2 (V, W) := kVkℓ2 (CI ) kWkℓ2 (CI ) , ξL2 (V, W) := kRϕ (V)kL2 (D) kRψ (W)kL2 (D) . We have ξℓ2 (V, W) =
kRϕ (V)k 2 kRψ (W)k 2
ξL2 (V, W) kVk 2 LI (D) kWk 2 LI (D) . Owing to (28.11), we infer that
ℓ (C ) ℓ (C )
1 1
ξL2 (V, W)(µϕ ψ 2 ϕ ψ
min µmin ) ≤ ξℓ2 (V, W) ≤ ξL2 (V, W)(µmax µmax ) .
2
The expected bounds follow by taking the supremum over W and then the infimum or the supre-
mum over V.
Recalling the definition (28.10) of the Euclidean condition number, Proposition 28.10 implies
that
ωL 2 ωL 2
c−1
M ≤ κℓ2 (A) ≤ cM , (28.19)
αL2 αL2
1
with cM := (κℓ2 (Mϕ )κℓ2 (Mψ )) 2 . Since the mass matrices Mϕ and Mψ are expected to be
relatively well-conditioned (see in particular Proposition 28.6), it is reasonable to expect that
sharp bounds for κℓ2 (A) can be obtained once sharp estimates of the real numbers αL2 and ωL2
are available.
R
Example 28.12 (Elliptic PDEs). Consider the bilinear form ah (vh , wh ) := D ∇vh ·∇wh dx on
Vh ×Vh with quasi-uniform meshes. The global Poincaré–Steklov inequality together with the exis-
tence of large-scale discrete functions in Vh (i.e., some interpolant of the distance to the boundary
of D) imply that αL2 is uniformly equivalent to ℓ−2 D , where ℓD is a characteristic length of D, e.g.,
ℓD := diam(D). Moreover, a discrete inverse inequality together with the existence of small-scale
functions in Vh (e.g., the global shape functions) implies that ωL2 is uniformly equivalent to h−2 .
Hence, the Euclidean condition number of the stiffness matrix is uniformly equivalent to ℓ2D h−2 . In
dimension one, this can be verified explicitly for P1 Lagrange elements on a uniform mesh since the
eigenvalues of A = h1 tridiag(−1, 2, −1) are { h2 (1 − cos(πlh))}l∈{1: I} (compare with Example 28.7
for the eigenvalues of the 1D mass matrix). See Strang and Fix [359] for early results on elliptic
PDEs and [186] for further insight and more examples.
Remark 28.13 (Ill-conditioning). In general, the Euclidean condition number of the stiffness
matrix grows as the mesh is refined. This growth may have an important impact on the efficiency
of iterative solvers when it comes to solving the linear system AU = B; see §28.3.2. Note though
that the sensitivity to perturbations induced by the growth of the condition number is usually not
a major concern since a discrete stability property can be formulated by using suitable norms; see
Exercise 28.7.
Remark 28.14 (Choice of basis). The condition number of the stiffness matrix and that of
the mass matrix depend on the choice made for the global shape functions. Using well-chosen
hierarchical bases leads to a stiffness matrix having a condition number κℓ2 (A) uniformly bounded
in h ∈ H; see, e.g., Hackbusch [235], Bramble et al. [81]. However, if κℓ2 (A) is bounded, then
κℓ2 (M) must explode as h → 0, i.e., it is not possible to find bases for which both A and M are
well-conditioned.
Remark 28.15 (Dependence on polynomial degree). Estimates on the condition number of
the stiffness matrix on a single mesh cell with high-order polynomials can be found in Olsen and
Douglas [320], Hu et al. [249].
28.2.4 Max-norm estimates

The notion of M -matrix is important in the real case when discretizing a PDE that enjoys a
maximum principle; see §33.2 for an example.
Definition 28.16 (M -matrix and Z-matrix). A matrix A ∈ RI×I is said to be a Z-matrix if
Aij ≤ 0 for all i, j ∈ {1:I} with i 6= j. A matrix A is said to be a nonsingular M -matrix if it is a
Z-matrix, invertible, and (A−1 )ij ≥ 0 for all i, j ∈ {1:I}.
A nonsingular M -matrix A enjoys several interesting properties: AV ≥ 0 implies V ≥ 0 for all
V ∈ RI (where V ≥ 0 means Vi ≥ 0 for all i ∈ {1:I}); all the eigenvalues of A have positive real
part; and all the diagonal entries of A are positive; see e.g., in Plemmons [326].
Lemma 28.17 (Majorizing vector). Let A ∈ RI×I be a Z-matrix. Then A is a nonsingular
I
P is a vector Y ∈ R called majorizing vector s.t. Y > 0 and AY > 0, i.e., Yi > 0
M -matrix iff there
and (AY)i := j∈{1: I} Aij Yj > 0 for all i ∈ {1:I}.
Proof. See Grossmann and Roos [225, p. 70].

Let us recall that kVkℓ∞ (RI ) := maxj∈{1: I} |Vj | for all V ∈ RI , and that thePinduced matrix
norm, which we also denote by k·kℓ∞ (RI ) , is such that kZkℓ∞ (RI ) = maxi∈{1: I} j∈{1: I} |Zij | for
all Z ∈ RI×I . It is possible to estimate kA−1 kℓ∞ (RI ) as follows if A is a nonsingular M -matrix.
Proposition 28.18 (Bound on k·kℓ∞ (RI ) -norm). Let A be a nonsingular M -matrix. Let Y be
a majorizing vector for A. The following holds true:
kYkℓ∞ (RI )
kA−1 kℓ∞ (RI ) ≤ . (28.20)
mini∈{1: I} (AY)i
Proof. See Exercise 28.8.
28.3 Solution methods

This section briefly reviews some methods to solve the linear system AU = B resulting from the
Galerkin approximation. We will see in Chapter 29 that finite element-based matrices are generally
sparse. This means that the number of nonzero entries in A is significantly smaller than the total
number of entries. It is important to keep this property in mind when considering solution methods.
28.3.1 Direct methods

The best-known example of direct method for solving the linear system AU = B consists of
constructing the LU factorization of A. Recall that a matrix Z ∈ CI×I is said to be lower
(resp., upper) triangular if Zij = 0 for all 1 ≤ i < j ≤ I (resp., Zij = 0 for all 1 ≤ j < i ≤ I).
Let σ be any permutation of the set {1:I}, the matrix P ∈ CI×I with entries Pij = δσ(i)j (δ is
the Kronecker symbol) is called permutation matrix. The LU factorization of A with complete
pivoting takes the form
PAQ = TL TU , (28.21)
where P and Q are permutation matrices, TL is lower triangular, and TU is upper triangular. These
matrices can be constructed by using Gaussian elimination; see, e.g., Golub and van Loan [218,
pp. 96-119]. The permutation matrices P and Q are needed to avoid divisions by zero or divisions
by quantities that are very small compared to the other entries of the matrix. When the matrix
A is Hermitian (symmetric in the real case), the right-hand side of (28.21) can be obtained in the
form TL D TLH , where TL is lower triangular with unit diagonal and D is diagonal; see [218, p. 137].
When A is Hermitian positive definite (symmetric positive definite in the real case), the matrix D
can be incorporated in the matrix TL by means of Choleski’s factorization so that TL has real and
positive diagonal entries; see [218, p. 141].
Once the LU factorization (28.21) of A has been constructed, the linear system AU = B is
solved by performing the following three steps: (i) Solve the lower triangular system TL U′ = PB.
(ii) Solve the upper triangular system TU U′′ = U′ . (iii) Set U := QU′′ . In practice, the cost of
computing the LU factorization dominates that of solving the triangular systems (for a dense I×I
matrix, the former scales as 13 I 3 and the latter as 21 I 2 ).
An important issue in the context of sparse matrices is the fill-in induced by the LU decompo-
sition. The (l, u)-bandwidth of a sparse matrix A is obtained from the two integers l, u such that
l := max{p | Aij = 0, i > j + p} and u := max{q | Aij = 0, j > i + q}. For instance, l = u = 0 for
a diagonal matrix, l = u = 1 for a tridiagonal matrix, and so on. It can be shown that if A has
(l, u)-bandwidth and if A has a LU factorization without pivoting, then TL has (l, 0)-bandwidth
and TU has (0, u)-bandwidth; see [218, Thm. 4.3.1]. Hence, the LU factorization does not increase
the bandwidth, but the matrices TL and TU have more nonzero entries than A. This fill-in can
be partly tamed by using reordering techniques; see §29.3. Broadly speaking sparse direct solvers
can be competitive alternatives to iterative methods (see below) for linear systems obtained by
approximating PDEs posed in dimension two, but this is no longer the case in dimension three on
sufficiently fine meshes.
Remark 28.19 (Literature). We refer the reader to George and Liu [214], Duff et al. [177],
Demmel et al. [160], Davis [156], Björck [58] for an overview of sparse direct solvers.
28.3.2 Iterative methods

Using an iterative method for solving a large sparse linear system presents the twofold advantage
of avoiding additional matrix storage and taking full advantage of sparsity by only performing
matrix-vector products. An iterative method is initialized by some vector U0 ∈ CI and then
produces a sequence (Um )m≥1 of vectors in CI that is expected to converge to the solution of the
linear system.
When the stiffness matrix is Hermitian positive definite, the conjugate gradient (CG) method
designed by Hestenes and Stiefel in 1952 [242] is particularly effective. The CG method is presented
in Algorithm 28.1. One matrix-vector product needs to be performed at each iteration (as well as
Algorithm 28.1 Conjugate gradient.

choose U0 ∈ CI , set R0 := B − AU0 and P0 := R0
choose a tolerance tol and set m := 0
while kRm kℓ2 (CI ) > tol do
αm := RH H
m Rm /Pm APm
Um+1 := Um + αm Pm
Rm+1 := Rm − αm APm
βm := RH H
m+1 Rm+1 /Rm Rm
Pm+1 := Rm+1 + βm Pm
m←m+1
end while
two inner products and three vector updates, but these operations induce a marginal computational
cost compared to the matrix-vector product). One can show by induction that Rm = B−AUm , that
{R0 , . . . , Rm−1 } is an ℓ2 -orthogonal set, and that {P0 , . . . , Pm−1 } is an A-orthogonal set; see Saad
[339, Prop. 6.13]. The crucial property of CG is the following [339, Prop. 5.2].
Proposition 28.20 (Optimality of CG). Let CI be equipped with the energy norm k·kA :=
1
(A·, ·)ℓ22 (CI ) . Then, at step m ≥ 1 of CG (provided no termination has occurred), Um satisfies the
following optimality property:
kU − Um kA = min kU − YkA , (28.22)

Y∈U0 +Km
with the Krylov subspace Km := span{R0 , AR0 , . . . , Am−1 R0 }.

The optimality property guarantees that CG terminates in at most I steps (in the absence
of roundoff errors). In practice, termination often occurs much earlier. It is a remarkable fact
that CG provides an optimality property over the whole affine subspace U0 + Km without needing
to store an entire basis of Km . This nice property is unfortunately lost if A is not Hermitian;
see Faber and Manteuffel [198], Voevodin [379]. When A is symmetric but indefinite, it is still
possible to achieve some optimality property over U0 + Km ; see §50.3.2 for mixed finite element
approximations. Solving the normal equations AH AU = AH B by CG is generally not a good idea
since it leads to very poor convergence rates.
There are two broad classes of Krylov subspace methods for non-Hermitian matrices. On the
one hand there are those that guarantee an optimality property over a subspace but require to
store a complete basis, thereby making storage and computational costs grow linearly with the
number of iterations. An important example is the generalized minimal residual (Gmres) method
(Marchuk and Kuznetsov [292, 293], Saad and Schultz [340]) where Um minimizes the Euclidean
norm of the residual over U0 + Km . The computational costs of Gmres can often be tamed by
using a restarted version. On the other hand there are those methods that give up optimality,
but employ short-term recurrences to compute the iterates. Examples are the conjugate gradient
squared (Cgs) (Sonneveld [350]) and the bi-conjugate gradient stabilized (Bi-CGStab) (van der
Vorst [370]) methods. These methods often work well in practice, although convergence is not
guaranteed.
The convergence rate of Krylov subspace methods depends on the spectrum of A. Sharp bounds
can be derived in the normal case (i.e., A commutes with AH ), whereas bounds in the nonnormal
case are more delicate and involve also the eigenvectors. We state the following result for CG when
A is Hermitian positive definite (see Saad [339, p. 193] or Elman et al. [185, p. 75]).
Proposition 28.21 (Convergence rate of CG). Let A ∈ CI×I be a Hermitian positive definite
matrix and let κ(A) be its (Euclidean) condition number. The following holds true for the CG
iterates: !
1 m
κ(A) 2 − 1
kU − Um kA ≤ 2 1 kU − U0 kA . (28.23)
κ(A) 2 + 1
Remark 28.22 (Clustering of eigenvalues). A sharper bound is kU − Um kA ≤ cm kU − U0 kA

with cm := minp∈Pm ,p(0)=1 maxλ∈σ(A) |p(λ)|, showing that clustering of the eigenvalues around a
few points (even spread out) is favorable to fast convergence. Note that (28.23) is derived from
this bound by writing cm ≤ minp∈Pm ,p(0)=1 kpkC 0 ([s♭ ,s♯ ]) with σ(A) ⊂ [s♭ , s♯ ] and constructing a
suitable minimizing polynomial (recall that σ(A) ⊂ [0, ∞) denotes the spectrum of A).
Preconditioning can be very effective to speed-up the convergence of Krylov subspace methods,
the ideal goal being to achieve computational costs that grow linearly with the size of the linear
system. Let PL , PR ∈ CI×I be two nonsingular matrices and assume that linear systems of the
form PL X = Y and PR X′ = Y′ are relatively inexpensive to solve. Then U solves AU = B if and
e := PR U solves AeU
only if U e = B,
e where Ae := P −1 AP −1 and B e := P −1 B. When A is Hermitian,
L R L
one can take PR = PL , and CG can be implemented by just considering the matrix P := PL PLH ;
H
see Exercise 28.10.

Choosing a preconditioner is a compromise between computational cost per iteration and im-
proving the spectral properties of the preconditioned matrix by clustering its eigenvalues. Rel-
atively simple preconditioners can be derived by using the splitting A = A+ − A− (this type
of splitting arises naturally in the context of stationary fixed-point iterations) and by using A+
as a preconditioner. Incomplete LU (ILU) preconditioning is generally a robust choice, the idea
being to discard entries in the LU factorization that do not match the sparsity pattern of A; see
[339, §10.3]. Many other preconditioning techniques are available in the literature. A particularly
important class is that of the multilevel (or multigrid) preconditioners where the solution is ex-
panded over a more or less hierarchical basis; see Bramble et al. [80, 81], Briggs [94], Elman et al.
[185], Hackbusch [235], Trottenberg et al. [365], Wesseling [393] for further insight into this topic.
Remark 28.23 (From complex to real linear systems). Solving the complex linear system
AU = B can be avoided by rewriting it as a linear system of twice the size of the real (or imaginary)
part of U. Using the obvious notation (R+ iS)(U1 + iU2 ) = B1 + iB2 with i2 = −1 and R, S ∈ RI×I ,
two possible rewritings of AU = B are

R −S U1 B1 R S U1 B1
= , = .
S R U2 B2 S −R −U2 B2
Let us denote by A∗ and A∗∗ the above two real matrices of size 2I×2I. Note that in general both
matrices are nonsymmetric and/or indefinite. Unfortunately, the distribution of the eigenvalues
in the spectra σ(A∗ ) and σ(A∗∗ ) are typically unfavorable for the convergence of Krylov subspace
methods as the eigenvalues embrace the origin of the complex plane with a large number of eigen-
values straddling the origin. Specifically, one can show that σ(A∗ ) = σ(A) ∪ σ(A) is symmetric
with respect to the real line, whereas σ(A∗∗ ) is symmetric with respect to both the real and imag-
inary lines and we have σ(A∗∗ ) = {λ ∈ C | λ2 ∈ σ(AA)}; see Freund [208, Prop. 5.1]. Hence, it is
in general preferable to deploy Krylov subspace methods on the complex linear system than on the
equivalent real ones. Effective algorithms can be devised by exploiting some particular structure
of A, e.g., if A is complex symmetric (i.e., A = AT instead of A = AH ); see Freund [208], Axelsson
and Kucherov [31].
Exercises
Exercise 28.1 (Matrix representation of operators). Let H be a (complex) Hilbert space
with inner product (·, ·)H . Let Vh be a finite-dimensional subspace of H with basis {ϕi }i∈{1: I} .
Let Z : Vh → Vh be a linear operator. Let M ∈ CI×I be the mass matrix s.t. Mij := (ϕj , ϕi )H ,
and let B, D ∈ CI×I be s.t. Bij := (Z(ϕj ), ϕi )H , Dij := (Z(ϕ
P j ), Z(ϕi ))H for all i, j ∈ {1:I}. Prove
that D = B H M−1 B. (Hint : use Z ∈ CI×I s.t. Z(ϕj ) := k∈{1: I} Zkj ϕk .)
Exercise 28.2 (Smallest singular value). Prove that the real number αℓ2 defined (28.17a)
is equal to kA−1 k−1 ℓ2 (CI )
. (Hint : to bound αℓ2 , consider a vector V∗ ∈ CI s.t. kA−1 V∗ kℓ2 (CI ) =
kA−1 kℓ2 (CI ) kV∗ kℓ2 (CI ) .)
Exercise 28.3 (ℓ2 -condition number). Let Z ∈ RI×I be the upper triangular matrix such that
Zii := 1 for all i ∈ {1:I}, and Zij := −1 for all i, j ∈ {1:I}, i 6= j. Let X ∈ RI have coordinates
Xi := 21−i for all i ∈ {1:I}. Compute ZX, kZXkℓ2 (RI ) , and kXkℓ2 (RI ) . Show that kZkℓ2 (RI ) ≥ 1
and derive a lower bound for κℓ2 (Z). What happens if I is large?
Exercise 28.4 (Local mass matrix, 1D). Evaluate the local mass matrix for one-dimensional
P1 and P2 Lagrange finite elements on a cell of length h.
Exercise 28.5 (Stiffness matrix). (i) Let {λ b1 , λ
b2 , λ
b3 } be the shape functions of the P1 Lagrange
element with the cell K b shown on the leftmost part of Figure 28.1. Here, λ b1 is associated with
the vertex (1, 0), b2 with the vertex (0, 1), and λ
λ b3 with the vertex (0, 0). Evaluate the stiffness
R
matrix for Kb ∇v·∇w dx. Same question for the Q1 Lagrange element with the shape functions
{θb1 , θb2 , θb3 , θb4 } associated with the vertices (1, 0), (1, 1), (0, 1), (0, 0), respectively (see the central
part of Figure 28.1). (ii) Consider the meshes R of D := (0, 3)×(0, 2) shown in the right part of
Figure 28.1. Evaluate the stiffness matrix for D ∇v·∇w dx.
Exercise 28.6 (Sensitivity to perturbations). Let Z ∈ CI×I be invertible and let X ∈ CI

solve ZX = B with B 6= 0. Set κ := κℓ2 (Z). (i) Let X̃ ∈ CI solve Z X̃ = B̃. Prove that
kX̃−Xkℓ2 (CI ) kB̃−Bkℓ2 (CI ) kX̌−Xkℓ2 (CI ) kŽ−Zkℓ2 (CI )
kXkℓ2 (CI ) ≤κ kBkℓ2 (CI ) . (ii) Let X̌ ∈ CI solve Ž X̌ = B. Prove that kX̌kℓ2 (CI )
≤κ kZkℓ2 (CI ) .
(iii) Explain why the above bounds are sharp.
(0,1) (0,1) (1,1)
(0,0) (1,0) (0,0) (0,1)
Figure 28.1: Illustration for Exercise 28.5. Left and central panels: reference triangle and square
considered in Step (i). Right panel: three meshes for Step (ii).
Exercise 28.7 (Stability). Let AU = B be the linear system resulting from the Galerkin ap-
H
proximation. Equip the vector space CI with the norm kVk∗ := supY∈CI kRψ|V(Y)k
Y|
W
. Show that
h
kuh −vh kVh
kuh kVh ≤ kaαhhk kB−AVk
kBk∗
∗
for all V ∈ CI , where uh := Rϕ (U) and vh := Rϕ (V). (Hint : show that
αh kuh − vh kVh ≤ kA(U − V)k∗ and that kBk∗ ≤ kah kkuhkVh , where αh and kah k are the stability
and boundedness constants of ah on Vh ×Wh .)
Exercise 28.8 (ℓ∞ -norm). (i) Prove Proposition 28.18. (Hint : use that AY ≥ minj∈{1: I} (AY)j U,
where U ∈ RI has all entries equal to 1.) (ii) Derive a bound on kA−1 kℓ∞ (RI ) with A :=
h−1 tridiag(−1, 2, −1). (Hint : consider the function x 7→ x(1 − x) on (0, 1) to build a majorizing
vector.) (iii) Let (E1 , . . . , EI ) be the canonical basis of RI . Let α ∈ R and consider the matrix
Z := I + αE1 ⊗ EI with entries Zij := δij + αδi1 δjI . Verify that Z −1 = I − αE1 ⊗ EI and evaluate
the condition number κℓ∞ (Z). What happens if α is large?
Exercise 28.9 (Lumped mass matrix). Let D be a two-dimensional polygonal set and consider
an affine mesh Th of D composed of triangles and P1 LagrangeR elements. (i) Let K be a cell in
Th . Compute the local mass matrix MK with entries MK ij := K θK,i (x)θK,j (x) dx, i, j ∈ {1:3}.
K K P
(ii) Compute the lumped local mass matrix M with Mij := δij l∈{1: 3} MK il . (iii) Compute
K K
the eigenvalues of (M )−1 (M − MK ). (iv) Let M be the global mass matrix and M be the
lumped mass matrix. Show that the largest eigenvalue of (M)−1 (M − M) is 34 .
Exercise 28.10 (CG). Let A ∈ RI×I be a real symmetric positive definite matrix and let
J : RI → R be such that J(V) := 12 VT AV − BT V. Let Um be the iterate at step m ≥ 1 of
the CG method. (i) Prove that Um minimizes J over U0 + Km . (Hint : use Proposition 28.20.)
(ii) Let ηm := arg minη∈C J(Um + ηPm ). Show that ηm = αm in the CG method. (iii) Write the
preconditioned CG method by just invoking the matrix P := PL PLT .
Exercise 28.11 (Complex symmetric system). Let A := T + iσI where T is symmetric real,
σ > 0, and I is the identity matrix of size I×I. Let A∗ and A∗∗ be the two rewritings of A as a real
matrix of size 2I×2I (see Remark 28.23). Determine the spectra σ(A), σ(A∗ ), and σ(A∗∗ ), and
comment on their position with respect to the origin. What happens if one considers the rotated
linear system −iAU = −iB instead?
Chapter 29
Sparse matrices
A matrix is said to be sparse if the number of its nonzero entries is significantly smaller than the
total number of its entries. The stiffness matrix is generally sparse as a consequence of the global
shape functions having local support. This chapter deals with important computational aspects
related to sparsity: storage, assembling, and reordering.
29.1 Origin of sparsity

Let us assume for simplicity that the discrete trial and the test spaces coincide. Recalling (28.1),
the entries of the stiffness matrix A ∈ CI×N glob are given by Aij := ah (ϕj , ϕi ) for all i, j ∈ {1:I},
for some sesquilinear form ah evaluated by computing an integral over D. Let Th be a mesh of D.
Decomposing the integral as a sum over the mesh cells we write
X Z
Aij = AK (x, ϕj|K , ϕi|K ) dx, ∀i, j ∈ {1:I}, (29.1)
K∈Th K
for some local functional AK acting on the restriction of the global shape functions to K. A crucial
consequence of (29.1) is that
[ Aij 6= 0 ] =⇒ [ | supp(ϕi ) ∩ supp(ϕj )| > 0 ], (29.2)
where supp(f ) denotes the support in D of the function f : D → R (i.e., the closure in D of the
subset {x ∈ D | f (x) 6= 0}).
The support of a global shape function ϕi depends on the associated global degree of freedom,
which is typically an evaluation at a node or an integral over an edge, a face, or a cell of the mesh.
As a result, supp(ϕi ) coincides with the set of the mesh cells containing the corresponding vertex,
edge, face, or cell, respectively. Let nmesh := maxi∈{1: I} card{K ⊂ supp(ϕi )}. This number is
bounded uniformly w.r.t. h ∈ H owing to the regularity of the mesh sequence. Let nsh be the
number of local shape functions. Let Nrow be the maximum number of nonzero entries per row of
A. A consequence of (29.2) is that Nrow is bounded from above as follows:
Nrow ≤ nmesh ×(nsh − 1) + 1. (29.3)
The right-hand side of (29.3) being independent of h ∈ H, we infer that the stiffness matrix becomes
sparser as the mesh is refined. The bound (29.3) can be made sharper by considering the type of
58 Chapter 29. Sparse matrices
support for the various global shape functions. For instance, the support of a global shape function
associated with a vertex is larger than the support of a global shape function associated with an
edge or a face.
Example 29.1 (Lagrange P1 and P2 , 2D). Consider a two-dimensional matching simplicial

mesh. Let nvtx be the maximum number of edges arriving at a vertex. For Lagrange P1 finite
elements, the global shape functions are attached to the mesh vertices, and each shape function
interacts with at most nvtx other shape functions, so that Nrow = nvtx +1 (left panel of Figure 29.1).
For Lagrange P2 finite elements, the global shape functions are attached either to the mesh vertices
or to the edge midpoints. The vertex shape functions interact with at most 3nvtx other shape
functions, whereas the edge shape functions interact with at most 8 other shape functions. Hence,
Nrow = 3nvtx + 1 since nvtx ≥ 3, (central and right panels of Figure 29.1).
Figure 29.1: Left (Lagrange P1 element): the global shape function attached to the vertex in black
interacts with the shape functions attached to the vertices in gray. Center and right (Lagrange
P2 element): the global shape function attached to the vertex (center) or edge midpoint (right) in
black interacts with the shape functions attached to the nodes in gray.
Example 29.2 (Sparsity pattern, structured mesh). The sparsity pattern of a 16×16 stiffness
matrix using P1 Lagrange finite elements on a two-dimensional structured mesh is shown in the
left panel of Figure 29.2. The mesh is shown in the right panel. The black squares in the sparsity
pattern are the nonzero entries. There are at most seven nonzero entries per row, i.e., Nrow = 7.
More generally, on a structured mesh that is built by using M nodes in each direction (M = 4
above), the stiffness matrix is of size M 2 ×M 2 , its entries are organized into a tridiagonal block-
structure with M blocks of size M ×M , and each block is tridiagonal or bidiagonal. As a result,
we also have Nrow = 7 in this case, independently of M .
13 14 15 16
9 10 11 12
5 6 7 8
1 2 3 4
Figure 29.2: Sparsity pattern (left) and underlying mesh (right).

29.2 Storage and assembling

Compressed storage formats are crucial in large-scale applications to avoid wasting memory space
by just storing zero entries. Assembling refers to the procedure in a finite element code where the
entries of the stiffness matrix (and those of the right-hand side) are computed.
29.2.1 CSR and CSC formats

One of the most frequently used storage techniques is probably the Compressed sparse rows or
Compressed row storage format (resp., columns), usually referred to as CSR or CRS format (resp.,
CSC or CCS). The CSR and CSC formats are very similar, the role played by rows and columns
being simply interchanged. We only present the CSR format for brevity.
Let A(1:I, 1:I ′ ) be a sparse matrix not necessarily square and containing nnz nonzero en-
tries. We define three arrays to store this matrix in the CSR format: ia(1:I+1), ja(1:nnz), and
aa(1:nnz).
Array ia. The integer array ia stores the number of nonzero entries in each row. More precisely,
conventionally setting ia(1) := 1, the value of ia(i + 1) is defined such that ia(i + 1) − ia(i) is
equal to the number of nonzero entries in the i-th row of the matrix A, i ∈ {1:I}. Note that
nnz = ia(I + 1) − ia(1).
Array ja. The integer array ja gives the column indices of the nonzero entries. More precisely, for
all i ∈ {1:I}, the list (ja(p))p∈{ia(i): ia(i+1)−1} contains the column indices of the nonzero entries
in row i. A usual convention is to store the column indices in ja in increasing order for every row.
Array aa. The array aa contains the nonzero entries of the matrix. For every row i, the list
(aa(p))p∈{ia(i): ia(i+1)−1} contains all the nonzero entries of the row i. The same ordering is used
for ja and aa, so that aa(p) := Ai,ja(p) .
Example 29.3. The CSR arrays for the following 5×5 matrix:
 
1. 0. 0. 2. 0.
3. 4. 0. 5. 0. 
 
A := 6. 0. 7. 8. 9. 
 (29.4)
0. 0. 10. 11. 0. 
0. 0. 0. 0. 12.
are aa = [1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12.], ja = [1 4 1 2 4 1 3 4 5 3 4 5], and ia = [1 3 6 10 12 13].

Note that nnz = ia(6) − ia(1) = 12.
Matrix-vector multiplication. Matrix-vector multiplication is an operation that is invoked re-

peatedly in iterative solution methods (see §28.3.2). Algorithm 29.1 shows how to perform the
matrix-vector multiplication with the CSR format. The technique is optimal in the sense that it
involves only the number of operations that are necessary.
29.2.2 Ellpack format

The CSR format has some drawbacks since the rows of the compressed matrix are not of constant
length. Moreover, the fact that the index of the first nonzero entry has to be computed for each
row may hamper vectorization. The purpose of the Ellpack format is to solve these difficulties.
This format, which is based on the hypothesis that each row of the matrix contains almost the
same number of nonzero entries, is well-adapted for meshes that are almost structured. Let A be
Algorithm 29.1 Matrix-vector multiplication in CSR format.

for i ∈ {1:I} do; yi := 0
for p ∈ {ia(i): ia(i + 1) − 1} do
yi := yi + aa(p) ∗ x(ja(p))
end for
y(i) := yi
end for
a matrix of size I×I ′ and let nrow (i) be the number of nonzero entries in the i-th row of A for
all i ∈ {1:I}. Let Nrow := maxi∈{1: I} nrow (i) be the maximum number of nonzero entries per row
in A. For instance, for continuous Q1 finite elements, Nrow = 9 in dimension two and Nrow = 27
in dimension three. The storage is done with two arrays aa(1:I, 1:Nrow ) and ja(1:I, 1:Nrow ) as
follows:
Array aa. The array aa contains the nonzero entries of A. For every row i, aa(i, 1:nrow (i))
contains all the nonzero entries in row i, and if nrow (i) < Nrow , the entries aa(i, (nrow (i) + 1):Nrow )
are set to zero by convention.
Array ja. The integer array ja contains the column indices of the nonzero entries in the matrix A.
For every row i, ja(i, 1:nrow (i)) contains all the column indices of the nonzero entries in row i. The
ordering of the indices in ja is the same as that in aa. The simplest convention consists of ordering
the column indices in increasing order. If nrow (i) < Nrow , the entries ja(i, (nrow (i) + 1):Nrow ) are
given an arbitrary value, say ja(i, nrow (i)).
29.2.3 Assembling
Let us see how the formula (29.1) can be implemented to evaluate the entries of the stiffness matrix
when it is stored in some compressed format, e.g., the CSR format. Let j dof(1:Nc , 1:nsh ) be the
double-entry connectivity array introduced in Chapter 19. Recall that this array is defined such
that
ϕj dof(m,n)|Km = θKm ,n , (29.5)
for every integers n ∈ {1:nsh } and m ∈ {1:Nc }. The assembling of the matrix A stored in the
CSR format is described in Algorithm 29.2. The temporary array tmp in each mesh cell stores the
local stiffness matrix. We will see in §30.3 how to compute the entries of this array by means of
quadratures.
29.3 Reordering
Reordering a square matrix A means replacing A by the matrix B := PAP T , where the permutation
P has entries Pij := δσ(i)j and σ is a permutation of the set {1:I}, so that Bij = Aσ(i)σ(j) . Since
P T = P −1 , the Euclidean condition number of a matrix is invariant by reordering. The goal of
reordering techniques modify the sparsity pattern of the matrix by clustering nonzero entries as
close as possible to the diagonal to reduce the bandwidth. We present in this section various
reordering techniques based on the concept of adjacency graph.
Example 29.4 (8×8 matrix). To illustrate how reordering can affect the fill-in resulting from the
LU factorization (see §28.3.1), let us consider the 8×8 matrix A whose sparsity pattern is shown
Algorithm 29.2 Matrix assembling in CSR format.

aa := 0
for m ∈ {1:Nc } do
for ni ∈ {1:nsh } do
for nj ∈ {1:nsh } Rdo
tmp(ni, nj) := Km AKm (x, θKm ,nj , θKm ,ni ) dx
end for
end for
for ni ∈ {1:nsh } do; i := j dof(m, ni)
for nj ∈ {1:nsh } do; j := j dof(m, nj)
for p ∈ {ia(i): ia(i+1)−1} do
if ja(p) := j then
aa(p) := aa(p) + tmp(ni, nj); Exit loop on p
end if
end for
end for
end for
end for
in the left panel of Figure 29.3. One can verify that the LU factorization of A (without pivoting)
results in complete fill-in, i.e., the lower and upper triangular matrices TL and TU in (28.21)
(such that A = TL TU ) are filled. Let us now consider the permutation σ : (1, 2, 3, 4, 5, 6, 7, 8) 7→
(8, 7, 6, 5, 4, 3, 2, 1). Let B be the 8×8 matrix such that Bij := Aσ(i)σ(j) for all i, j ∈ {1:8}. The
sparsity pattern of B is shown in the right panel of Figure 29.3. It is straightforward to check that
no fill-in occurs when computing the LU factorization of B. This simple example illustrates that
significant savings in memory and computational time can be achieved by enumerating properly
the degrees of freedom (dofs) in a finite element code.
Initial matrix A Reordered matrix B
Figure 29.3: Two different orderings for a sparse matrix.
Remark 29.5 (Literature). We refer the reader to Saad [339] for reordering techniques applied
to iterative solvers and to George and Liu [214], George et al. [215], Davis [156, Chap. 7] for direct
solution methods.
29.3.1 Adjacency graph

As illustrated in Example 29.4, it is often important to reorder the unknowns and the equations
before solving a linear system. Of course, the reordering technique to be used depends on the
strategy chosen to solve the linear system (direct, iterative, parallel, etc.). Choosing optimal
reordering strategies is a difficult branch of graph theory.
To better understand the enumeration issue, it is convenient to introduce the notion of ad-
jacency graph. Let V be a set, let ℜ be a binary relation on V, and denote E := {(x, y) ∈
V ×V | xℜy}. The pair G := (V, E) is called graph. The elements of V are called graph vertices or
nodes and the members of E are called graph edges. We say that G is an undirected graph if ℜ is
symmetric. A vertex y is said to be adjacent to x if (x, y) ∈ E. For a subset X ⊂ V, the adjacent
set of X is defined as Adj(X) := {y ∈ V \X | ∃x ∈ X, (x, y) ∈ E}. The set Adj(x), defined as
Adj({x}), is called neighborhood of x. The cardinality of Adj(x) is called degree of x. A common
way of representing graphs is to associate with each vertex in V a point in the plane and to draw
a directed line between two points (possibly identical) whenever their associated vertices are in E.
Let A be a I×I matrix. The adjacency graph of A is the pair (V, E) where V := {1:I} and
E := {(i, j) ∈ V ×V | Aij 6= 0}. Thus, we have
Adj(i) = {j ∈ {1:I}\{i} | Aij 6= 0}, ∀i ∈ {1:I}. (29.6)
We say that (E, V ) is the undirected adjacency graph of A when (i, j) ∈ E iff Aij 6= 0 or Aji 6= 0.
Figure 29.4 shows the adjacency graph of a 8×8 sparse matrix. A circle around a number means
that the corresponding diagonal entry in the matrix is not zero.
2 5
1 7
3 4
6
Figure 29.4: Sparsity pattern (left) and adjacency graph (right) of a 8×8 sparse matrix.
29.3.2 Level-set ordering

Assume that V is finite, let G := (V, E) be a graph, and let x ∈ V be a vertex. The elements of an
indexed collection of disjoint subsets of V,Ssay L1 , L2 , L3 , . . ., are said to be level sets associated
with x if L1 := {x} and Lk+1 := Adj(Lk )\( l∈{1: k} Ll ) for all k ≥ 1. Lk is said to be the k-th level
set. The list L1 , L2 , L3 , . . . is finite since V is finite. Moreover, L1 , L2 , L3 , . . . forms a partition of
V if G is a strongly connected graph, i.e., if there exists a path from each vertex to every other
vertex.
For every vertex y in V, we define the distance from x to y, dx (y), as follows: If there is k such
that y ∈ Lk , then dx (y) := k − 1. Otherwise, dx (y) := ∞ (note that dx (y) < ∞ if the graph is
strongly connected). In general, dx (y) 6= dy (x) unless the graph is undirected (think of V := {x, y}
and E := {(x, y)} so that dx (y) = 1 and dy (x) = ∞).
Algorithm 29.3 shows a possible way to evaluate the level sets associated with a vertex i1 in
the adjacency graph of a I×I sparse matrix A. The integer max levelset is the number of level
sets associated with the vertex i1 . The level sets are deduced from the arrays perm and stride as
Algorithm 29.3 Evaluation of the level sets of i1 .

Input: i1
Output: perm, max levelset, stride
k := 2; count := 1; virgin(1:I) :=.true.
perm(1) := i1 ; stride(1) := 1; stride(2) := 2
loop
nb vert in Lk := 0
for l ∈ {stride(k−1):stride(k)−1} do
for all j ∈ Adj(perm(l)) do
if (virgin(j)) then
virgin(j) := .false.; nb vert in Lk = nb vert in Lk + 1
count := count + 1; perm(count) := j
end if
end for
end for
if (nb vert in Lk = 0) then
max levelset := k − 1; exit loop
end if
stride(k + 1) := stride(k) + nb vert in Lk; k := k + 1
end loop
if count 6= I then G is not strongly connected
follows:
{perm(1)}, . . . , {perm(stride(k)), . . . , perm(stride(k + 1) − 1)}, . . . .
| {z } | {z }
=:L1 =:Lk
If there is a vertex i2 which is not in any of the level sets associated with i1 , i.e., i1 is not connected
to i2 , then the level sets associated with i2 are constructed by using Algorithm 29.3 again. The
process is repeated until the union of all the level sets forms a partition of V. At the end all the
permutation arrays are collected in a single array still denoted by perm.
Example 29.6. Let us consider the undirected graph shown in Figure 29.5 to illustrate the level
set concept. The level sets associated with vertex 2 are
L1 = {2}, L2 = {5, 7}, L3 = {9, 11, 14}, L4 = {1, 3, 12, 15},

L5 = {8, 10, 13}, L6 = {4, 6}.
Hence, max levelset = 6, stride = (1, 2, 4, 7, 11, 14, 16), and a possible choice for perm is perm =
(2, 5, 7, 9, 11, 14, 1, 3, 12, 15, 8, 10, 13, 4, 6).
The simplest reordering for A consists of setting Bij := Aperm(i)perm(j) . This technique is known
as the breadth-first-search (BFS) reordering. One interest of this reordering is the following result.
Proposition 29.7. Assume G is undirected and max levelset ≥ 3. The array stride defines a
tridiagonal block structure of B, i.e.,


i ∈ {stride(k): stride(k+1)−1} i.e., perm(i) ∈ Lk ,
Bij = 0 if j ∈ {stride(k ′ ): stride(k ′ +1)−1} i.e., perm(j) ∈ Lk′ ,


|k − k ′ | ≥ 2.
8 12 14 5 2
4 10 1 15 11 7
6 13 3 9
Figure 29.5: Sparsity pattern and adjacency graph.
Proof. The proof proceeds by contradiction. Assume that Bij 6= 0 and |k − k ′ | ≥ 2 with perm(i) ∈
Lk and perm(j) ∈ Lk′ . Then Aperm(i)perm(j) 6= 0. This means perm(i) ∈ Adj(perm(j)) and perm(j) ∈
Adj(perm(i)),
S since G is undirected and i 6= j. Assume further that k ′ ≥ k. Then perm(j) 6∈
l∈{1: k} Ll since the level sets are disjoint. Moreover, perm(j) ∈ Adj(perm(i)) and perm(i) ∈
Lk means S perm(j) ∈ Adj(Lk ). Combining the above two statements yields perm(j) ∈ Lk+1 =
Adj(Lk )\( l∈{1: k} Ll ). This means k ′ = k + 1, which contradicts |k − k ′ | ≥ 2. The argument
applies also if k ′ ≤ k since the graph is undirected.
Proposition 29.7 shows that choosing level sets with max levelset as large as possible minimizes
the bandwidth of B. This can be achieved by picking the initial vertex i1 such that maxy∈V di1 (y)
is maximal.
The ordering depends on the way the vertices are traversed in each level set. In the BFS
reordering, the vertices are traversed in the natural order. Another strategy consists of ordering
the vertices in each level set by increasing degree. This ordering technique is known as the Cuthill–
McKee (CMK) ordering. Another popular strategy consists of reversing the CMK ordering. It has
been observed that the reversing strategy yields a better scheme for sparse Gaussian elimination.
We refer the reader to George and Liu [214], George et al. [215] for further insight into these
techniques and their many generalizations.
Example 29.8 (CMK reordering). Figure 29.6 shows the adjacency graph and the sparsity
pattern of the CMK-reordered matrix corresponding to the matrix shown in Figure 29.5. The
reordering has been done by using the level sets associated with vertex 2. In each level set, the
nodes are ordered by increasing degree. The permutation array is
perm = (2, 5, 7, 9, 14, 11, 12, 3, 15, 1, 8, 10, 13, 4, 6).
The reordered matrix has a tridiagonal block structure, and the size of the k-th block is stride(k +
1) − stride(k) with the array stride evaluated in Example 29.6.
29.3.3 Independent set ordering (ISO)

The aim of ISO techniques is to find a permutation of the vertices such that the reordered matrix
has the following 2×2 block structure:

D E
B= ,
F H
11 9 6 2 1
14 12 7 10 5 3
15 13 8 4
Figure 29.6: Sparsity pattern and adjacency graph after using the CMK reordering for the matrix
shown in Figure 29.5.
Algorithm 29.4 Independent set ordering.

S := ∅; virgin :=.true.
for i ∈ {1:I} do
j := traverse(i)
if (virgin(j)) then
S := S ∪ {j}; virgin(j) :=.false.
for all k ∈ Adj(j) do
virgin(k) :=.false.
end for
end if
end for
where D is diagonal and as large as possible. To this purpose, we introduce the notion of indepen-
dent set. Let G := (V, E) be a graph. S ⊂ V is said to be an independent set if for all x 6= y ∈ S,
the edge (x, y) is not in E. An independent set is said to be maximal if it is maximal with respect
to the inclusion order.
Assume that G is the adjacency graph of a square matrix A. Let S be an independent set.
Let perm be any permutation array of {1:I} such that S = {perm(1), . . . , perm(card(S))}. Define
the reordered matrix B such that Bij = Aperm(i)perm(j) for all i, j ∈ {1:I}. We readily infer the
following result.
Proposition 29.9. The triple (1, card(S), I) defines a 2×2 block structure of B where the top left
block is diagonal.
Let traverse be a permutation array of {1:I}. Algorithm 29.4 presents a simple strategy to
construct an independent set. A possible choice for traverse consists of setting traverse(i) := i,
but in general traverse is set to maximize the cardinality of S. Since card(S) is equal to the
number of times the statement (virgin(j)) is true in Algorithm 29.4, a possible technique to
maximize this number is to choose j s.t. card(Adj(j)) is small, i.e., among all the nodes left, j
must be one of those having the lowest degree. A simple strategy consists of sorting the nodes in
increasing degree in traverse.
29.3.4 Multicolor ordering

A third standard reordering method uses graph coloring. Assume that G is an undirected and
strongly connected graph. Then the map C : V → N is said to be a graph coloring if C(x) 6= C(y)
for all (x, y) ∈ E s.t. x 6= y. For x ∈ V, C(x) is referred to as the color of x. The goal of graph
coloring is to find a map C s.t. the cardinality of the range of C is as small as possible, i.e., the
number of colors to color V is as small as possible.
In the context of linear algebra, optimality is not a major issue and one is usually satisfied by
using simple heuristics. For instance, given a permutation array traverse of {1:I}, Algorithm 29.5
describes a basic coloring strategy. The simplest choice consists of setting traverse(i) := i, but
more sophisticated choices are possible. For instance, it can be shown that if the graph can be
colored with two colors only and if BFS is used to initialize traverse, then Algorithm 29.5 finds a
two-color partitioning; see Exercise 29.5. Independently of traverse, the number of colors found
by Algorithm 29.5 is at most equal to 1 plus the largest degree in the graph; see Exercise 29.5.
Let G be the undirected adjacency graph of a matrix A. Assume that we have colored G. Denote
by k max the number of colors that are used, and let C : V → {1:k max} be the corresponding color
mapping. Let col part(1 : k max) be the array such that col part(1) := 1 and col part(k + 1) :=
col part(k) + card(C −1 (k)) for all k ∈ {1:k max}. Let perm be any permutation array s.t. the
color of the vertices in the set {perm(col part(k)), . . . , perm(col part(k + 1) − 1)} is k. Define
the reordered matrix B such that Bij = Aperm(i)perm(j) . Multicolor ordering partially finds its
justification in the following result.
Proposition 29.10. The array col part defines a k max×k max block structure of B where the
diagonal blocks are diagonal.
Proof. Left as an exercise; see also Adams and Jordan [5].
Exercises
Exercise 29.1 (Retrieving a nonzero entry in CSR format). Write an algorithm to retrieve
the value Aij from the array aa stored in CSR format.
Exercise 29.2 (Ellpack (ELL)). Write the arrays needed to store the matrix from Example 29.3
in the Ellpack format. Write an algorithm that performs a matrix-vector multiplication in this
format.
Exercise 29.3 (Coordinate format (COO)). Let A be a I×I sparse matrix. Consider the
storage format where one stores the nonzero entries Aij in the array aa(1:nnz) and stores in the
same order the row and columns indices in the integer arrays ia(1:nnz) and ja(1:nnz), respectively.
(i) Use this format to store the matrix defined in (29.4). (ii) Write an algorithm to perform a
matrix-vector product in this format. Compare with the CSR format.
Algorithm 29.5 Greedy coloring.

color := 0
for i ∈ {1:I} do
j := traverse(i) {Since G is strongly connected Adj(j) cannot be empty.}
color(j) := min{k > 0 | k 6∈ color(Adj(j))}
end for
Exercise 29.4 (Storage). Consider the storage format for sparse I×I matrices where one stores
the nonzero entries Aij in the array aa(1:nnz) and stores in the same order the integer (i − 1)I + j
in the integer array ja(1:nnz). (i) Use this format for the matrix defined in (29.4). (ii) Write an
algorithm to do matrix-vector products in this format. Compare with the CSR format.
Exercise 29.5 (Greedy coloring). (i) Prove that the total number of colors found by Algo-
rithm 29.5 is at most equal to 1 plus the largest degree in the graph. (ii) Assume that a graph G
can be colored with two colors only. Prove that if the BFS reordering is used to initialize traverse,
then Algorithm 29.5 finds a two-color partitioning. (Hint : by induction on the number of level
sets.)
Exercise 29.6 (Multicolor ordering). Prove Proposition 29.10.
Exercise 29.7 (CMK reordering). Give the sparsity pattern and the CMK reordering for the
matrix shown in Figure 29.4.
Chapter 30
Quadratures
Implementing the finite element method requires evaluating the entries of the stiffness matrix and
the right-hand side vector, which in turn requires computing integrals over the cells and (possibly)
the faces of the mesh. In practice, these integrals must often be evaluated approximately by means
of quadratures. In this chapter, we review multidimensional quadratures that are frequently used
in finite element codes, and we derive bounds on the quadrature error. We also describe the
implementation of quadratures in conjunction with the assembling of the stiffness matrix. Recall
that one-dimensional quadratures are presented in Chapter 6.
30.1 Definition and examples

Let D be a Lipschitz polyhedron in Rd , d ∈ {2, 3}, let Th be a mesh of D that covers D exactly,
R and
let φ : D → R be a smooth function. Suppose that we want to evaluate the integral D φ(x) dx.
Since Z X Z
φ(x) dx = φ(x) dx,
D K∈Th K
R
this problem reduces to evaluating the integral K φ(x) dx over each mesh cell K ∈ Th . Since
computing
R integrals exactly is often impossible, one needs to use quadratures to approximate
K φ(x) dx.
Definition 30.1 (Quadrature nodes and weights). Let K be a compact, connected, Lipschitz
subset of Rd with nonempty interior. Let lQ ≥ 1 be an integer. A quadrature in K with lQ nodes
is specified through a set of lQ points {ξl }l∈{1: lQ } in K, called quadrature nodes or Gauss nodes
and a set of lQ real numbers {ωl }l∈{1: lQ } , called quadrature weights. The quadrature consists of
the approximation Z X
φ(x) dx ≈ ωl φ(ξl ). (30.1)
K l∈{1: lQ }
The largest integer k such that (30.1) is an equality for every polynomial in Pk,d is called quadrature
order and is denoted by kQ .
b and a mesh Th , a quadrature on every cell
Given a quadrature on the reference element K
K ∈ Th can be generated by using the geometric mapping TK : Kb → K. Let JK denote the
Jacobian matrix of TK .
70 Chapter 30. Quadratures
b with nodes {ξbl }l∈{1: l }

Proposition 30.2 (Quadrature generation). Consider a quadrature in K Q
and weights {b
ωl }l∈{1: lQ } . Setting
ξlK := TK (ξbl ) and bl |det(JK (ξbl ))|,

ωlK := ω (30.2)
for all l ∈ {1:lQ }, generates a quadrature on K. If the quadrature on K b is of order kQ and the
geometric mapping TK is affine, then the quadrature on K is also of order kQ .
R
Proof. Since TK is a C 1 -diffeomorphism, the change of variables x = TK (bx) yields K φ(x) dx =
R
b φ TK (bx) |det(JK (b
x))| db b to the right-hand side. The
x, and we can apply the quadrature over K
K
statement on the quadrature order is immediate to verify. Indeed, if TK is affine, JK is constant
and φ ◦ TK is in Pk,d iff φ ∈ Pk,d .
Remark 30.3 (Surface quadrature). When generating a surface quadrature from a quadrature
on a reference surface, Lemma 9.12 must be used to account for the transformation of the surface
measure; see Exercise 30.5.
Example 30.4 (Literature). The literature on quadratures is abundant; see Abramowitz and
Stegun [3, Chap. 25], Hammer and Stroud [237], Stroud [360], Davis and Rabinowitz [155], Brass
and Petras [85]. We refer the reader to §6.2 for a review of one-dimensional quadratures using the
Gauss–Legendre, Gauss–Lobatto, and Gauss–Radau nodes.
Example 30.5 (Cuboids). Quadratures on cuboids can be deduced from one-dimensional quadra-
tures by taking the Gauss nodes in tensor-product form. Note though that tensor-product formulas
are not optimal in the sense of using the fewest function evaluations for a given order. Although
no general formula for non-tensor-product quadratures for the cube is known, many quasi-optimal
quadratures are available in the literature; see, e.g., Cools [139, §2.3.1] and Cools and Rabinowitz
[140, §4.1].
Example 30.6 (Quadratures on the triangle). Table 30.1 lists some quadratures on the
triangle (see, e.g., Dunavant [178]). In this table, we call multiplicity the number of permutations
to be performed on the barycentric coordinates to obtain the list of all the Gauss nodes of the
quadrature. For instance, the first-order formula in the second line has three Gauss nodes with
barycentric coordinates and weights {1, 0, 0; 13 S}, {0, 1, 0; 31 S}, {0, 0, 1; 13 S}, where S denotes the
surface of the triangle.
Example 30.7 (Quadratures on the tetrahedron). Table 30.2 lists some quadratures on the
tetrahedron (see, e.g., Keast [265]). As above, the multiplicity is the number of permutations
to perform on the barycentric coordinates to obtain all the Gauss nodes of the quadrature. For
instance, the third-order formula has five Gauss nodes which are the node ( 41 , 14 , 14 , 14 ) with the
weight − 45 V and the four nodes ( 61 , 16 , 16 , 12 ), ( 61 , 16 , 12 , 16 ), ( 61 , 12 , 16 , 16 ), ( 12 , 16 , 61 , 16 ) with the weight
9
20 V, where V denotes the volume of the tetrahedron.
Example 30.8 (Integral of barycentric coordinates). Let {λi }i∈{0: d} be the barycentric
coordinates in a simplex K in Rd . We have
Z
α0 ! . . . αd !d!
λα0 αd
0 . . . λd dx = |K| , (30.3)
K (α0 + . . . + αd + d)!
for every natural numbers α0 , . . . , αd . This formula is useful to verify numerically the order of a
quadrature.
kQ lQ Barycentric coord. Multiplicity Weights ωl

1 1 1

1 1 3, 3, 3 1 S
1
1 3 (1, 0, 0) 3 3S
1 1 2
1
2 3 6, 6, 3 3 3S
1 1
1
2 3 2, 2, 0 3 3S
1 1 1
9
3 4 3, 3, 3 1 − 16 S
1 1 3 25
5, 5, 5 3 48 S
1 1 1
9
3 7 3 , 3 , 3 1 20 S
1 1 2
2, 2, 0 3 15 S
1
(1, 0, 0) 3 20 S
4 6 (ai , ai , 1 − 2ai ) for i = 1, 2 3 ωi for i = 1, 2

a1 = 0.445948490915965 ω1 = S×0.223381589678010
a2 = 0.091576213509771 ω2 = S×0.109951743655322
1 1 1
9
5 7 3, 3, 3 1 40 S
(ai , ai , 1 − 2ai ) for i = 1, 2 3
√ √
a1 = 6−21 15 155− 15
1200 S
√ √
6+ 15 155+ 15
a2 = 21 1200 S
6 12 (ai , ai , 1 − 2ai ) for i = 1, 2 3

a1 = 0.063089014491502 S×0.050844906370206
a2 = 0.249286745170910 S×0.116786275726378
(a, b, 1 − a − b) 6
a = 0.310352451033785 S×0.082851075618374
b = 0.053145049844816
Table 30.1: Nodes and weights for quadratures on a triangle of area S.
30.2 Quadrature error

Let (Th )h∈H be a shape-regular sequence of affine meshes. Proposition 30.2 allows us to generate a
quadrature in each mesh cell K ∈ Th from a reference quadrature by using the geometric mapping
TK : Kb → K. Let {ξlK }l∈{1: l } and {ωlK }l∈{1: l } be the nodes and weights of the quadrature on
Q Q
K thus obtained. Let kQ ≥ 0 be the order of the quadrature. For every function φ that is smooth
enough to have point values, say φ ∈ C 0 (K), we define the quadrature error in the mesh cell K as
follows:
Z X
EK (φ) := φ(x) dx − ωlK φ(ξlK ). (30.4)
K l∈{1: lQ }
Lemma 30.9 (Quadrature error). Let p ∈ [1, ∞] and let m ∈ N be such that m > dp . Assume
kQ lQ Barycentric coord. Multiplicity Weights ωl

1 1 1 1

1 1 4, 4, 4, 4 1 V
1
1 4 (1, 0, 0, 0) 4 4V
1
2 4 (a, a, a, 1 − 3a) 4 4V
√
a = 5−20 5
1 1
1
2 10 2 , 2 , 0, 0 6 5V
1
(1, 0, 0, 0) 4 − 20 V
1 1 1 1

3 5 4, 4, 4, 4 1 − 45 V
1 1 1 1 9
6, 6, 6, 2 4 20 V
1 1 1 1
16
5 15 4, 4, 4, 4 1 135 V
(ai , ai , ai , 1 − 3ai ) for i = 1, 2 4
√ √
7− 15 2665+14 15
a1 = 34 37800√ V
√
7+ 15 2665−14 15
a2 = 34 37800 V
(a, a, 21 − a, 21 − a) 6
√
a = 10−2 40
15 10
189 V
Table 30.2: Nodes and weights for quadratures on a tetrahedron of volume V.
that kQ + 1 ≥ m. There is c s.t. for all φ ∈ W m,p (K), all K ∈ Th , and all h ∈ H,
1
m+d(1− p )
|EK (φ)| ≤ c hK |φ|W m,p (K) . (30.5)
Proof. Let φ ∈ W m,p (K). Since m > dp , the embedding W m,p (K) ֒→ C 0 (K) from Theorem 2.31
implies that the function φ is continuous. Moreover, since the meshes are affine, we have EK (φ) =
|det(JK )| E( b where φb := φ ◦ TK and E(
b φ), b is the quadrature error on K.
b φ) b By definition, E b :
0 b b b b
C (K) → R is a bounded linear form, i.e., |E(φ)| ≤ ckφkC 0 (K)
b . Using the embedding W
m,p b
(K) ֒→
0 b b b b b
C (K), we infer that |E(φ)| ≤ ckφk m,p b . Since E(b p) = 0 for all pb in Pm−1,d ⊂ PkQ ,d (since kQ +
W (K)
1 ≥ m), we deduce from the Bramble–Hilbert/Deny–Lions lemma (more precisely Corollary 11.11
b ≤ c|φ|
b φ)|
with k := m − 1) that |E( b m,p b . Since the geometric mapping is affine and the mesh
W (K)
sequence is shape-regular, we infer from (11.7a) in Lemma 11.7 that
1
b m,p b ≤ c kJK km2 d |det(JK )|− p |φ|W m,p (K) .
|φ| W (K) ℓ (R )
hK |K|
We conclude by using (11.3), i.e., kJK kℓ2 (Rd ) ≤ ρK
c
and |det(JK )| ≤ b .
|K|
In the analysis of finite element methods with quadrature, it is useful to estimate the quadrature
−1
error EK (φp), where p ∈ Pn,d ◦ TK for some integer n ≥ 0; see §33.3 for an application.
Lemma 30.10 (Quadrature error with polynomial factor). Let m ∈ N, let n ∈ N. (i)
Assume that kQ ≥ m + n − 1. There is c s.t.
|EK (φp)| ≤ c hm
K |φ|W m,∞ (K) kpkL1 (K) , (30.6)
−1
for all φ ∈ W m,∞ (K), all p ∈ Pn,d ◦ TK , all K ∈ Th , and all h ∈ H. (ii) Assume that n ≥ 1,
m ≥ 1, and kQ ≥ n + m − 2. There is c s.t.

|EK (φp)| ≤ c hm
K |φ|W m,∞ (K) kpkL1 (K) + |φ|W m−1,∞ (K) k∇pkL1 (K) , (30.7)
−1
for all φ ∈ W m,∞ (K), all p ∈ Pn,d ◦ TK , all K ∈ Th , and all h ∈ H.
Proof. We only prove (30.6), and the reader is referred to Exercise 30.4 for the proof of (30.7).
Let pb := p ◦ TK ∈ Pn,d and φb := φ ◦ TK . After making a change of variable, we obtain EK (φp) =
bp) since Th is affine (with obvious notation). Assuming first m ≥ 1, we infer that
b φb
|det(JK )|E(
b
E(b g ∈ Pm−1,d since pb ∈ Pn,d and kQ ≥ n + m − 1. Hence, E(
gpb) = 0 for all b bp) = E((
b φb b φb − b
g)b
p) for
g ∈ Pm−1,d . Therefore, we obtain
all b
b φb
|EK (φp)| = |det(JK )||E( bp)| = |det(JK )| inf |E(( b φb − gb)bp)|
g ∈Pm−1,d
b

≤ c |det(JK )| inf kφb − gbkC 0 (K)
b kb
pkC 0 (K)
b
g ∈Pm−1,d
b

≤ c |det(JK )| inf kφb − gbkW m,∞ (K)b kb
pkL1 (K)
b ,
g ∈Pm−1,d
b
where we used norm equivalence in Pn,d for pb. Since
inf kφb − gbkW m,∞ (K) b

b ≤ c |φ|W m,∞ (K)
b ,
g∈Pm−1,d
b
owing to the Bramble–Hilbert/Deny–Lions lemma (see Lemma 11.9), we infer that
b m,∞ b |det(JK )| kb
|EK (φp)| ≤ c |φ| pkL1 (K)
b .
W (K)
We conclude by using |det(JK )| kb

pkL1 (K) b
b = kpkL1 (K) and |φ|W m,∞ (K)
m
b ≤ ckJK kℓ2 |φ|W m,∞ (K) with
kJK kℓ2 ≤ chK (see Lemma 11.7). Finally, if m = 0, we have
b ∞ b |det(JK )| kb
|EK (φp)| ≤ ckφk pkL1 (K)
b ,
L (K)
b ∞ b = kφkL∞ (K) .
and we conclude by using that kφk L (K)
30.3 Implementation
This section addresses practical implementation aspects of quadratures in the assembling modules
of a finite element code.
30.3.1 Nodes and weights

Let {ξbl }l∈{1: lQ } and {b
ωl }l∈{1: lQ } be the quadrature nodes and weights on the reference element.
Let m ∈ {1:Nc } and let Km be the corresponding mesh cell. The nodes and weights of the
quadrature on Km are defined in Proposition 30.2. Recall from Definition 8.1 that the geometric
mapping TKm is built from the reference shape functions of the geometric element, {ψbn }n∈{1: ngeo } .
This leads us to define the two-dimensional array psi(1:ngeo , 1:lQ ) that contains the values of the
geometric shape functions at the quadrature nodes:
psi(n, l) := ψbn (ξbl ).
The k-th Cartesian component of the Gauss node ξlKm := TKm (ξbl ) is then given by
X
(ξlKm )k = coord(k, j geo(n, m)) psi(n, l),
n∈{1: ngeo }
where the arrays coord and j geo are defined in §8.3. We also need the three-dimensional array
dpsi dhatK(1:d, 1:ngeo , 1:lQ ) providing the derivatives of the geometric shape functions at the
Gauss nodes:
∂ ψbn b
dpsi dhatK(k, n, l) := (ξl ).
∂bxk
Then the entries of the Jacobian matrix JKm at ξbl can be computed as follows:
X
JKm (ξbl ) = coord(k1 , j geo(n, m)) dpsi dhatK(k2 , n, l),
k1 ,k2
n∈{1: ngeo }
for all k1 , k2 ∈ {1:d}. Since det(JKm (ξbl )) is always multiplied by the weight ω
bl in the quadratures,
it can be useful to store this product once and for all in the two-dimensional array of weights
weight K(1:lQ , 1:Nc ):
weight K(l, m) := ω bl |det(JKm (ξbl ))|.
Notice that when the mesh is affine, the partial derivatives of the shape functions ψbn are constant
b so that the size of the array dpsi dhatK can be reduced to d×ngeo . Further memory
on K,
space can be saved by storing separately the reference quadrature weights ω bl and the determinants
|det(JKm )|. The choice between storing and recomputing on the fly depends on the hardware at
hand. For instance, it is preferable to recompute a quantity if accessing the memory is slower than
the actual computing.
30.3.2 Shape functions

Let {θbn }n∈N be the reference shape functions. Since these functions and their derivatives need to
be evaluated many times at the Gauss nodes in K, b it can be useful to compute these values once and
for all and store them in the two-dimensional array theta(1:nsh , 1:lQ ) and in the three-dimensional
array dtheta dhatK(1:d, 1:nsh , 1:lQ ) such that
∂ θbn b
theta(n, l) := θbn (ξbl ), dtheta dhatK(k, n, l) := (ξl ).
∂bxk
Let us assume for simplicity that the linear bijective map used to generate the local shape functions
is the pullback by the geometric mapping. Then the values of the local shape functions at the
Gauss nodes in the mesh cell Km are given by
θn (ξlKm ) := θbn (ξbl ) = theta(n, l),
for all n ∈ {1:nsh }, all l ∈ {1:lQ }, and all m ∈ {1:Nc } (notice that the value of θn (ξlKm ) is
independent of Km ). Let us now consider the first-order derivatives of the local shape functions
at the Gauss nodes. Using the chain rule, we infer that for all k1 ∈ {1:d},
∂θn X ∂ θbn
(ξlKm ) = (ξbl ) J−1
Km (ξbl ) ,
∂xk1 ∂b
xk2 k2 ,k1
k2 ∈{1: d}
−1
∂(TK )
where we used that m k2
(ξlKm ) = (J−1 b
∂xk1 Km (ξl ))k2 ,k1 . One can evaluate the partial deriva-
tives of the local shape functions once and for all and store them in the four-dimensional array
dtheta dK(1:d, 1:nsh , 1:lQ , 1:Nc ) such that
∂θn
dtheta dK(k, n, l, m) := (ξlKm ).
∂xk
Notice that the size of this array, d×nsh ×lQ ×Nc , can be very large. If the mesh is affine, one can
adopt another strategy since the Jacobian matrix JKm and its inverse do not depend on the Gauss
nodes. In this case, one can store the inverse of the Jacobian matrix in the three-dimensional array
inv jac K(1:d, 1:d, 1:Nc ) such that
−1
inv jac K(k1 , k2 , m) := JKm .
k1 ,k2
∂θn
Then the following operations must be performed each time the quantity ∂xk1 (ξlKm ) is needed:
∂θn X
(ξlKm ) = dtheta dhatK(k2 , n, l) inv jac K(k2 , k1 , m).
∂xk1
k2 ∈{1: d}
The array inv jac K has d×d×Nc entries, which is smaller than the number of entries of dtheta dK
if d ≪ nsh ×lQ . In this situation, storing inv jac K will save memory space at the prize of some
additional computations. But again, depending of the hardware at hand, one must be aware that
a balance must be struck between storage and recomputing on the fly.
30.3.3 Assembling
For simplicity, we assume that a standard Galerkin formulation is considered with the bilinear form
a and the linear form ℓ. The discrete trial and test spaces are identical. Let {ϕi }i∈{1: I} be the
global shape functions. In the absence of quadratures, the entries of the stiffness matrix A ∈ RI×I
are Aij := a(ϕj , ϕi ) for all i, j ∈ {1:I}, and those of the right-hand side vector are Bi := ℓ(ϕi ) for
all i ∈ {1:I}; see §28.1.1. The goal of this section is to revisit the assembling of A and B when
quadratures are employed.
Let us first consider the assembling of the stiffness matrix. To fix the ideas, we consider the
Rbilinear form associated with a diffusion-advection-reaction model problem for which a(vh , wh ) :=
D A(x, vh , wh ) dx, with
A(x, ϕ, ψ) := d(x)∇ϕ(x)·∇ψ(x) + ψ(x)β(x)·∇ϕ(x) + ϕ(x)µ(x)ψ(x),
with smooth fields d, β, and µ taking values in Rd×d , Rd , and R, respectively. This model problem
is considered, e.g., in Chapter 31. In Chapter 24, we considered the simpler setting where d := Id ,
β := 0, and µ := 0; see (24.7). In Cartesian notation, the quantity A(x, ϕ, ψ) is expressed as
follows:
X ∂ϕ ∂ψ
A(x, ϕ, ψ) := (x)dk1 k2 (x) (x)
∂xk1 ∂xk2
k1 ,k2 ∈{1: d}
X ∂ϕ
+ ψ(x) βk1 (x) (x) + ϕ(x)µ(x)ψ(x).
∂xk1
k1 ∈{1: d}
Using quadratures to integrate A(x, ϕ, ψ) over all the mesh cells, we obtain the approximate
bilinear form aQ s.t.
X X
aQ (vh , wh ) := ωlKm A(ξlKm , vh|Km , wh|Km ).
m∈{1: Nc } l∈{1: lQ }
This leads to the approximate stiffness matrix AQ with entries AQ,ij := aQ (ϕj , ϕi ) for all i, j ∈
{1:I}.
Algorithm 30.1 Assembling of AQ for analytic data.

AQ = 0
for l ∈ {1:lQ } do; tmp := 0
for k ∈ {1:d} doX
xi l(k) := coord(k, j geo(n, m)) ∗ psi(n, l)
n∈{1: ngeo }
end for
for nj ∈ {1:n
X sh } do
x1 := dtheta dK(k1 , nj, l, m)∗dk1 k2 (xi l)∗ dtheta dK(k2 , ni, l, m)
k1 ,k2 ∈{1: d}
X
x2 := theta(ni, l) βk1 (xi l) ∗ dtheta dK(k1 , nj, l, m)
k1 ∈{1: d}
x3 := theta(ni, l) ∗ µ(xi l) ∗ theta(nj, l)
tmp(ni, nj) := tmp(ni, nj) + [x1 + x2 + x3 ] ∗ weight K(l, m)
end for
end for
end for
Accumulate tmp in AQ as in Algorithm 29.2
end for
A general assembling procedure for the stiffness matrix A (stored in the CSR format) has been
outlined in Algorithm 29.2. Our goal is now to detail the evaluation of the array tmp used in this al-
gorithm by means of quadratures. We assume for simplicity that the coefficients (dk1 k2 )k1 ,k2 ∈{1: d} ,
(βk1 )k1 ∈{1: d} , and µ are known analytically; see Exercise 30.7 for discrete data. The assembling
procedure of the approximate stiffness matrix AQ is shown in Algorithm 30.1. Notice that we first
evaluate and store the coordinates of the Gauss nodes ξlKm since we need to evaluate the values
of the coefficients at these nodes.
The assembling of the right-hand side vector R can be performed similarly. To fix the ideas,
we consider a linear form such that ℓ(wh ) := D F (x, wh ) dx, with F (x, ψ) := f (x)ψ(x) and
f : D → R is a smooth function. Using quadratures to integrate F (x, ψ) over all the mesh cells,
we obtain the approximate linear form ℓQ s.t.
X X
ℓQ (wh ) := ωlKm F (ξlKm , wh|Km ).
m∈{1: Nc } l∈{1: lQ }
This leads to the approximate right-hand side vector BQ with entries BQ,i := ℓQ (ϕi ) for all
i ∈ {1:I}. The assembling procedure of the vector BQ is presented in Algorithm 30.2.
Algorithm 30.2 Assembling of BQ for analytic data.

BQ = 0
for l ∈ {1:lQ } do; tmp := 0
for k1 ∈ {1:d} doX
xi l(k1 ) := coord(k1 , j geo(n, m)) psi(n, l)
n∈{1: ngeo }
end for
tmp(ni) := tmp(ni) + f (xi l) ∗ theta(ni, l) ∗ weight K(l, m)
end for
end for
for ni ∈ {1:nsh } do; i := j dof(m, ni)
BQ,i := BQ,i + tmp(ni)
end for
end for
Exercises
Exercise 30.1 (Quadratures on simplices). Let K be a simplex in Rd . Let zK be the barycen-
ter of K, let {zi }i∈{0: d} be the vertices of K, and let {mi }i∈{0: d} be the midpoints of the edges of K.
1 1
Consider the following quadratures: {zK }, {|K|}; {zi }i∈{0: d} , { d+1 |K|}; {mi }i∈{0: d} , { d+1 |K|}.
(i) Prove that the first and the second quadratures are of order one. (ii) Prove that the third one
is of order two for d = 2.
Exercise 30.2 (Quadrature for Q2,d ). Let K b := [0, 1]d be the unit hypercube. Let a b i1 ...id :=
i1 id
R P
( 2 , . . . , 2 ), i1 , . . . , id ∈ {0:2}. Show that the quadrature K b f (b x ≈ i1 ,...id wi1 ...id f (b
x) db ai1 ...id )
Qd
where wi1 ...id := 61d k=1 (3ik (2 − ik ) + 1) is exact for all f ∈ Q2,d . (Hint : write the Q2,d Lagrange
shape functions in tensor-product form and use Simpson’s rule in each direction.)
Exercise 30.3 (Global quadrature error). Prove that
Z X X
1
φ(x) dx − ωlK φ(ξlK ) ≤ chm |D|1− p |φ|W m,p (D) ,

D K∈Th l∈{1: lQ }
for all φ ∈ W m,p (D) and all h ∈ H. (Hint : use Lemma 30.9.)
Exercise 30.4 (Quadrature error with polynomial). The goal is to prove (30.7). We are
going to make use of (30.6) formulated as follows: |EK (ψq)| ≤ c hµK |ψ|W µ,∞ (K) kqkL1 (K) for all
q ∈ Pν,d ◦TK where µ+ν −1 ≤ kQ , µ, ν ∈ N. (i) Prove that |EK (φpK )| ≤ chmK |φ|W m,∞ (K) kpkL1 (K) ,
where pK is the mean value of p over K. (ii) Prove (30.7). (Hint : use Step (i) with µ := m − 1.)
Exercise 30.5 (Surface quadrature). Assume d = 3. Let F be a face of a mesh cell. Let Fb ⊂ R2
be a reference face and let TF : Fb → F be the geometric mapping for F . Let t1 (b s), t2 (b
s) be the two
column vectors of the Jacobian matrix of TF (b s) := [t1 (b
s), say JF (b s)] ∈ R3×2 . (i) Compute
s), t2 (b
the metric tensor gF := JT F JF ∈ R
2×2
in terms of the dot products ti ·tj , i, j ∈ {1, 2}. (ii) Show that
ds = kt1 (b
s)×t2 (b s. (Hint : use Lagrange’s identity, that is, kak2ℓ2 (R3 ) kbk2ℓ2 (R3 ) − (a·b)2 =
s)kℓ2 (R3 ) db
p
ka×bk2ℓ2 (R3 ) for any pair of vectors a, b ∈ R3 , and recall that ds = det(gF ) db s). (iii) Given a
quadrature {b b
bl }l∈{1: l∂ } on F , generate the quadrature on F .
sl , w Q
Exercise 30.6 (Asssembling). Let D := (0, 1)2 . Consider the problem −∆u + u = 1 in D and
u|∂D = 0. (i) Approximate its solution with P1 H 1 -conforming finite elements on the two meshes
shown in Figure 30.1. (ii) Evaluate the discrete solution in both cases. (Hint : there is only one
degree of freedom in both cases, see Exercise 28.5 for computing the gradient part of the stiffness
coefficient and use a quadrature from Table 30.1 for the zero-order term.) (iii) For a fine mesh
composed of 800 elements, we have uh ( 21 , 12 ) ≈ 0.0702. Comment.
1 1
0 1 0 1
Figure 30.1: Illustration for Exercise 30.6.
Exercise 30.7 (Discrete data). Adapt Algorithm 30.1 to the case where (dk1 k2 )k1 ,k2 ∈{1: d} ,
(βk1 )k1 ∈{1: d} , and µ are known in the discrete space Vh . (Hint
P : let dif, beta, and mu be the corre-
sponding coordinate vectors, and observe that µ(ξlKm ) = n∈{1: nsh } mu(j dof(m, i))× theta(n, l),
etc.)
Exercise 30.8 (Assembling of RHS). Write P the assembling algorithm for the right-hand side
∂wh
vector in the case where F (ξ, wh ) := f (ξ)wh (ξ)+ k1 ∈{1: d} βk1 (ξ) ∂x k1
(ξ) with analytically known
data.
Chapter 31
Scalar second-order elliptic PDEs
In Part VII, composed of Chapters 31 to 35, we study the approximation of scalar second-order
elliptic PDEs by H 1 -conforming finite elements. Among the topics we address in this part are
weak formulations and well-posedness, a priori error analysis, the discrete maximum principle,
the impact of quadratures, and a posteriori error analysis. In Chapters 31 to 34, we focus on
weak formulations endowed with a coercivity property, so that well-posedness hinges on the Lax–
Milgram lemma and the error analysis on Céa’s lemma (and its variants). In Chapter 35, we study
the Helmholtz problem as an example of elliptic PDE without coercivity.
The present chapter addresses fundamental properties of scalar-valued second-order elliptic
PDEs endowed with a coercivity property. The prototypical example is the Laplacian with homo-
geneous Dirichlet conditions. More generally, we consider PDEs including lower-order terms, such
as the diffusion-advection-reaction equation, where the lower-order terms are small enough so as
not to pollute the coercivity provided by the diffusion operator. We also study in some detail how
various boundary conditions (Dirichlet, Neumann, Robin) can be enforced in the weak formulation.
Moreover, important smoothness properties of the solutions to scalar second-order elliptic PDEs
are listed at the end of the chapter. These results will be useful later to establish error estimates
for the finite element approximation.
31.1 Model problem

Let D be a domain in Rd , i.e., D is a nonempty, open, bounded, connected subset of Rd (see
Definition 3.1). Let d, β, and µ be functions defined on D that take values in Rd×d , Rd , and
R, respectively. Given a function f : D → R, we look for a function u : D → R that solves the
following linear PDE:
−∇·(d∇u) + β·∇u + µu = f in D. (31.1)
Boundary conditions
are discussed later. Using Cartesian coordinates, the PDE (31.1) amounts
P P
to i,j∈{1: d} ∂xi dij ∂xj + i∈{1: d} βi ∂x
∂ ∂u ∂u
i
+ µu = f. The PDE reduces to the Poisson equation
−∆u = f studied in §24.1 if d is the identity tensor in Rd and β and µ vanish identically.
More generally, (31.1) is a diffusion-advection-reaction equation modeling for instance heat or
mass transfer or flows in porous media. The first term on the left-hand side of (31.1) accounts for
diffusion processes, the second one for advection processes, and the third one for reaction processes
(depletion occurs when µ is positive).
80 Chapter 31. Scalar second-order elliptic PDEs
31.1.1 Ellipticity and assumptions on the data

We assume that d ∈ L∞ (D) := L∞ (D; Rd×d ) and that d takes symmetric values. We also assume
that β ∈ W 1,∞ (D) := W 1,∞ (D; Rd ), µ ∈ L∞ (D), and f ∈ L2 (D). For dimensional consistency,
1
we equip the space H 1 (D) with the norm kvkH 1 (D) := (kvk2L2 (D) + ℓ2D k∇vk2L2 (D) ) 2 , where ℓD is a
length scale associated with D, e.g., ℓD := diam(D). A key notion for the second-order PDE (31.1)
is that of ellipticity.
Definition 31.1 (Ellipticity). For a.e. x ∈ D, let [λmin (x), λmax (x)] be the smallest interval
containing the eigenvalues of d(x). We say that the PDE (31.1) is elliptic if
0 < λ♭ := ess inf λmin (x) ≤ ess sup λmax (x) =: λ♯ < ∞. (31.2)
x∈D x∈D
Example 31.2 (Anisotropic diffusion). We say that the diffusion process is anisotropic if the
2 2 2
diffusion matrix is not proportional to the identity, as in the PDE − ∂∂xu2 + 2κ ∂x∂1 ∂x
u
2
− ∂∂xu2 = f
1 2
which is is elliptic if κ ∈ (−1, 1).
Remark 31.3 (Divergence form, Cordes condition). The PDE (31.1) is said to be in diver-
gence form because of the way the second-order term is written. One can also consider the PDE
P
in nondivergence form −d:D2 u + β·∇u + µu = f , where d:D2 u = i,j∈{1: d} dij ∂x∂i ∂x
2
u
j
. In this
case, one usually adds the Cordes condition [141] to the ellipticity assumption: There is ǫ ∈ (0, 1]
kdk2
uniformly in D, where kdkF = (d:d) 2 is the Frobenius norm of d and tr(d)
1 1
s.t. (tr(dF))2 ≤ d−1+ǫ
its trace (note that tr(d) > 0 owing to the ellipticity condition). We refer the reader to Smears
and Süli [348] for further insight in the context of Hamilton–Jacobi–Bellman equations.
The following important result, which is similar to the unique continuation principle for real
analytic functions, hinges on the ellipticity property.
Theorem 31.4 (Unique continuation principle). Let D be a connected subset of Rd with
0 ∈ D. Assume that d satisfies the ellipticity condition (31.2), dij ∈ C 0 (D; R), dij is Lipschitz
continuous in D\{0}, and there are c > 0 and δ > 0 such that k∇dij (x)kℓ2 ≤ ckxkℓδ−1 2 for all
1
x ∈ D. Let u ∈ Hloc (D) and assume that
X
|d:D2 u| ≤ c
δ+|α|−2 α
kxkℓ2 |∂ u|, (31.3a)
|α|≤1
Z
1
lim d u2 (x) dx = 0. (31.3b)
ǫ→0 ǫ kxkℓ2 <ǫ
Then u = 0 in D.
Proof. See Hörmander [247, Thm. 17.2.6]. We also refer the reader to Reed and Simon [332,
Thm. XIII.57&63] for variations on the unique continuation principle that are somewhat easier to
grasp.
Example 31.5 (Application to (31.1)). The above result, known in the literature as the
Aronszajn–Cordes uniqueness theorem, can be used to establish the uniqueness of the solution
to the PDE (31.1). Assume that u1 , u2 are two solutions of (31.1), and assume that one can show
1 1
that u1 ∈ Hloc (D), u2 ∈ Hloc (D), and there is an open set S ⊂ D s.t. (u1 − u2 )|S = 0. One can
always assume that 0 ∈ S. Setting u := u1 − u2 , one has −d:D2 u = (∇·d − β)·∇u − µu. Let
us assume that d satisfies the assumptions of Theorem 31.4. Then one immediately deduces that
(31.3a) holds true with some appropriate constant c. Using that u|S = 0, the second condition
(31.3b) is trivially satisfied, and uniqueness follows.
Part VII. Elliptic PDEs: conforming approximation 81
31.1.2 Toward a weak formulation

Proceeding informally as in §24.1, e.g., assuming u ∈ H 2 (D), we multiply (31.1) by a test function
w ∈ H 1 (D) and integrate over D to obtain
Z Z

− ∇·(d∇u)w + (β·∇u)w + µuw dx = f w dx. (31.4)
D D
Integrating by parts the first term on the left-hand side leads to

Z Z Z
−∇·(d∇u)w dx = (d∇u)·∇w dx − (n·(d∇u))w ds, (31.5)
D D ∂D
where n denotes the outward unit normal to D. We then arrive at

Z Z
a(u, w) − (n·(d∇u))w ds = f w dx, ∀w ∈ H 1 (D), (31.6)
∂D D
where a is defined for all (v, w) ∈ H 1 (D)×H 1 (D) as follows:

Z

a(v, w) := (d∇v)·∇w + (β·∇v)w + µvw dx. (31.7)
D
Notice
P in passing that using
P Cartesian∂vcoordinates, the symmetry of d implies (d∇v)·∇w =
i,j∈{1: d} dij ∂xj ∂xi = i,j∈{1: d} dij ∂xi ∂xj = ∇v·(d∇w), that is, (d∇v)·∇w = ∇v·(d∇w).
∂v ∂w ∂w
Moreover, using the Cauchy–Schwarz inequality for the three integrals leads to
−2
|a(v, w)| ≤ (λ♯ ℓD + β♯ ℓ−1
D + µ♯ )kvkH 1 (D) kwkH 1 (D) , (31.8)
for all v, w ∈ H 1 (D), with β♯ := kβkL∞ (D) and µ♯ := kµkL∞ (D) , which proves that the bilinear form
a is bounded on H 1 (D)×H 1 (D). Equation (31.6) is the starting point to derive weak formulations
for the PDE (31.1) with various types of boundary conditions.
31.2 Dirichlet boundary condition

Our goal is now to prove the well-posedness of the weak formulation when a Dirichlet boundary
′
condition is enforced. In what follows, we identify L2 (D) with its dual space L2 (D) so that we
are in the situation where
′ ′
H01 (D) ֒→ L2 (D) ≡ L2 (D) ֒→ H −1 (D) = H01 (D) , (31.9)
with bounded and densely defined embeddings (recall that the notation V ֒→ W means that the
embedding of V into W is bounded).
31.2.1 Homogeneous Dirichlet condition

We consider the homogeneous Dirichlet condition
u = 0 on ∂D, (31.10)
which we are going to enforce strongly by using the space H01 (D) for both the trial and the test
spaces. Recall from the trace theorem (Theorem 3.10) that u ∈ H01 (D) implies that γ g (u) = 0,
1
where γ g : H 1 (D) → H 2 (∂D) is the trace map such that γ g (v) = v|∂D if the function v is smooth.
Since the test functions vanish at the boundary, we can drop the boundary term on the left-hand
side of (31.6), leading to the following weak formulation:

R (31.11)
a(u, w) = D f w dx, ∀w ∈ V.
Proposition 31.6 (Weak solution). Let f ∈ L2 (D). If the function u ∈ H01 (D) solves (31.11),
then it satisfies the PDE (31.1) a.e. in D and the boundary condition (31.10) a.e. on ∂D.
Proof. Let u be a weak solution. Testing the weak formulation (31.11) against an arbitrary function
R
ϕ ∈ C0∞ (D) ⊂ H01 (D) and using the notion of weak derivatives leads to h−∇·(d∇u), ϕi = D (f −
β·∇u − µu)ϕ dx since f ∈ L2 (D) and β·∇u + µu ∈ L2 (D) owing to the assumptions on the data.
Hence, −∇·(d∇u) defines a bounded linear form on L2 (D) with Riesz–Fréchet representative equal
to f − β·∇u − µu. This means that u solves the PDE (31.1) a.e. in D. Moreover, u ∈ H01 (D)
1
implies that γ g (u) = 0 in H 2 (∂D) ֒→ L2 (∂D), i.e., the boundary condition (31.10) holds a.e. on
∂D.
R
Remark 31.7 (f ∈ H −1 (D)). When f ∈ H −1 (D), the term D f w dx in (31.11) must be under-
stood as hf, wiH −1 (D),H01 (D) . More specifically, recalling from Theorem 4.12 that the assumption
f ∈ H −1 (D) is equivalent
R to assuming that there are g0 ∈ L2 (D) and g1 ∈ L2 (D) such that
hf, wiH −1 (D),H01 (D) = D (g0 w + g1 ·∇w) dx, each time we write a(u, w) = hf, wiH −1 (D),H01 (D) , we
R
actually mean a(u, w) = D (g0 w + g1 ·∇w) dx, and the PDE we actually solve is −∇·(d∇u) +
β·∇u + µu = g0 − ∇·g1 in H −1 (D).
We now make assumptions on the PDE coefficients that are sufficient to prove the well-
posedness of (31.11) by invoking a coercivity property. Recall the Poincaré–Steklov inequal-
ity (3.11) (with p := 2), i.e., there is Cps > 0 such that
Cps kvkL2 (D) ≤ ℓD k∇vkL2 (D) , ∀v ∈ H01 (D). (31.12)
Owing to (31.12), we can equip the space V := H01 (D) with the norm
kvkV := k∇vkL2 (D) = |v|H 1 (D) . (31.13)
The space V equipped with this norm is a Hilbert space since kvkV ≤ ℓ−1
D kvkH 1 (D) ≤ (1 +
−2 21
Cps ) kvkV for all v ∈ V.
Proposition 31.8 (Well-posedness). Assume the ellipticity condition (31.2). Assume that there
exists θ > 0 such that
1
2 −2
µ♭ := ess inf µ − ∇·β (x) ≥ −(1 − θ)Cps ℓD λ♭ . (31.14)
x∈D 2
(i) The bilinear form a is V -coercive:
a(v, v) ≥ λ♭ min(1, θ)kvk2V , ∀v ∈ V. (31.15)
(ii) The problem (31.11) is well-posed.

Proof. The boundedness property (31.8) of a can be rewritten as
−1 −2 2

|a(v, w)| ≤ λ♯ + Cps ℓD β♯ + Cps ℓD µ♯ kvkV kwkV ,
R
for all v, w ∈ V. Moreover, the linear form ℓ(w) := D f w dx is bounded on V since |ℓ(w)| ≤
−1
kf kL2(D) kwkL2 (D) ≤ kf kL2 (D) Cps ℓD kwkV for all w ∈ V. Let us now prove the coercivity prop-
erty (31.15). Using the divergence formula for the field ( 12 v 2 )β, we infer that
Z Z Z
1 1
v(β·∇v) dx = − (∇·β)v 2 dx + (β·n)v 2 ds, (31.16)
D 2 D 2 ∂D
for all smooth functions v ∈ C ∞ (D). A density argument then shows that the formula (31.16)
remains valid for all v ∈ H 1 (D). Using the definition of µ♭ , the identity
(31.16), and that v
R 2 2
vanishes at the boundary, we obtain a(v, v) ≥ D λ♭ k∇vkℓ2 (Rd ) + µ♭ |v| dx for all v ∈ V. The
assumptions on λ♭ and µ♭ imply that

2 −2
a(v, v) ≥ λ♭ k∇vk2L2 (D) − (1 − θ)Cps ℓD kvk2L2 (D) .
If θ > 1, the last term is positive, whereas if θ ∈ (0, 1], we have

2 −2
k∇vk2L2 (D) − (1 − θ)Cps ℓD kvk2L2 (D)
2 −2
= θk∇vk2L2 (D) + (1 − θ)(k∇vk2L2 (D) − Cps ℓD kvk2L2 (D) ) ≥ θk∇vk2L2 (D) ,
where the last bound follows from the Poincaré–Steklov inequality. The coercivity property (31.15)
then results from kvkV := k∇vkL2 (D) . Finally, the well-posedness of (31.11) follows from the Lax–
Milgram lemma.
Example 31.9 (Pure diffusion). Coercivity for a purely diffusive problem (so that µ♭ = 0)
holds true with θ := 1.
Remark 31.10 (Variational formulation). Assume that Rβ = 0 in D. Then u solves (31.11)
iff u minimizes in H01 (D) the energy functional ED (v) := 12 D ∇v·(d∇v) + µv 2 − 2f v dx; see
Proposition 25.8.
Remark 31.11 (Helmholtz). The condition (31.14) is only sufficient to ensure the well-posedness
of (31.11) by means of a coercivity argument. We will see in Chapter 35, which deals with the
Helmholtz problem, that well-posedness can also hold without invoking (31.14). In this case, we
will establish well-posedness by means of an inf-sup argument.
31.2.2 Non-homogeneous Dirichlet condition

1
Let g ∈ H (∂D). We consider the non-homogeneous Dirichlet condition
2
u=g on ∂D. (31.17)

1
Since the map γ g : H 1 (D) → H 2 (∂D) is surjective, there is a uniform constant Cγ g and ug ∈
H 1 (D) such γ g (ug ) = g and kug kH 1 (D) ≤ Cγ g kgk 21 ; see Theorem 3.10(iii). Setting u0 :=
H (∂D)
u − ug , we obtain γ g (u − ug ) = g − g = 0, i.e., u0 ∈ H01 (D). This leads to the following weak
formulation:

Find u ∈ H 1 (D) such that u0 := u − ug ∈ V := H01 (D) satisfies
R (31.18)
a(u0 , w) = D f w dx − a(ug , w), ∀w ∈ V.
The right-hand side in (31.18) defines a bounded linear form on V owing to the boundedness of a
on H 1 (D)×H 1 (D) and the above bound on ug . Proceeding as in the homogeneous case, one can
prove the following result.
1
Proposition 31.12 (Well-posedness). Let f ∈ L2 (D) and g ∈ H 2 (∂D). (i) If the function u ∈
H 1 (D) solves (31.18), then it satisfies the PDE (31.1) a.e. in D and the boundary condition (31.17)
a.e. on ∂D. (ii) Under the assumptions of Proposition 31.8, (31.18) is well-posed.
31.3 Robin/Neumann conditions

The Dirichlet conditions are called essential boundary conditions since they are imposed explicitly
in the solution space. The Robin and the Neumann conditions belong to the class of natural
boundary conditions. These conditions are not explicitly enforced in the solution space, but they
are enforced in the weak formulations by using test functions that are not zero at the boundary.
31.3.1 Robin condition

∞
Let ρ ∈ L (∂D) and g ∈ L2 (∂D). We consider the Robin boundary condition
ρu + n·(d∇u) = g on ∂D. (31.19)
Starting from (31.6) and still proceeding informally, we consider test functions in H 1 (D) (i.e., they
are no longer in H01 (D) as for the Dirichlet conditions), and we use the Robin condition in the
boundary integral
R R side of (31.6), thereby replacing n·(d∇u) by g − ρu. This leads
on the left-hand
to a(u, w) + ∂D (g − ρu)w ds = D f w dx. Introducing the trace map γ g in the boundary term and
rearranging the expression, we obtain the following weak formulation:

Find u ∈ V := H 1 (D) such that
R R (31.20)
aρ (u, w) = D f w dx + ∂D gγ g (w) ds, ∀w ∈ V,
with the bilinear form aρ on H 1 (D)×H 1 (D) s.t.

Z
aρ (v, w) := a(v, w) + ργ g (v)γ g (w) ds. (31.21)
∂D
The boundedness of the trace map (see Theorem 3.10) implies that there is Mγ g s.t. kγ g (v)kL2 (∂D) ≤
−1
Mγ g ℓD 2 kvkH 1 (D) . Using the Cauchy–Schwarz inequality yields
Z
ργ g (v)γ g (w) ds ≤ ρ♯ Mγ2g ℓ−1
D kvkH 1 (D) kwkH 1 (D)
∂D
M with ρ♯ := kρkL∞ (∂D) . Since a is bounded on H 1 (D)×H 1 (D), so is aρ . Similarly, the right-hand
side in (31.20) defines a bounded linear form in H 1 (D).
We identify L2 (∂D) with its dual space L2 (∂D)′ in order to interpret the boundary condi-
1
tion satisfied by weak solutions to (31.20). Hence, we have H 2 (∂D) ֒→ L2 (∂D) ≡ L2 (∂D)′ ֒→
1 1 1
H − 2 (∂D) with dense embeddings, where H − 2 (∂D) is the dual space of H 2 (∂D). Recall from
1
Theorem 4.15 the normal trace map γ d : H(div; D) → H − 2 (∂D) defined such that the following
identity holds true for all φ ∈ H(div; D) and all w ∈ H 1 (D):
Z

hγ d (φ), γ g (w)i − 12 12 = φ·∇w + (∇·φ)w dx. (31.22)
H ,H
D
We have γ d (φ) = n·φ whenever φ is smooth, e.g., if φ ∈ H s (D), s > 21 .

Proposition 31.13 (Weak solution). Let f ∈ L2 (D), ρ ∈ L∞ (∂D), and g ∈ L2 (∂D). If the
function u ∈ H 1 (D) solves (31.20), then it satisfies the PDE (31.1) a.e. in D, and the boundary
condition (31.19) a.e. in ∂D in the sense that ργ g (u) + γ d (d∇u) = g in L2 (∂D).
Proof. As in the proof of Proposition 31.6, one can show that the PDE (31.1) is satisfied a.e. in
D. In particular, introducing the diffusive flux σ := −d∇u, we obtain σ ∈ L2 (D) and ∇·σ =
f − β·∇u − µu ∈ L2 (D), i.e., σ ∈ H(div; D). Using the weak formulation, we infer that
Z
−hγ d (σ), γ g (w)i − 12 21 + (ργ g (u) − g)γ g (w) ds = 0,
H ,H
∂D
1
for all w ∈ H (D). Since the trace operator γ : H 1 (D) → H 2 (∂D) is surjective and the above
1 g
equality is valid for all w ∈ H 1 (D), we infer that γ d (σ) defines a bounded linear form on L2 (∂D)
with Riesz–Fréchet representative equal to ργ g (u) − g. Hence, the boundary condition is satisfied
a.e. on ∂D.
Remark 31.14 (Data smoothness). Notice that f ∈ L2 (D) is needed to establish that ∇·σ ∈
1
L2 (D). It is possible to assume that g is only in H − 2 (∂D). Then the boundary term in (31.20)
1
becomes hg, γ g (w)i − 12 21 , and the Robin boundary condition is satisfied only in H − 2 (∂D).
H ,H
We now address the well-posedness of (31.20). One can show (see Exercise 31.2 and (3.15))
that there is Čps > 0 such that for all v ∈ H 1 (D),
Čps kvkL2 (D) ≤ ℓD kvkV ,

n o 12 (31.23)
with kvkV := k∇vk2L2 (D) + ℓ−1
D kγ g
(v)k 2
L 2 (∂D) .
1 1
−2 − 2
Thus, (1 + Čps ) kvkH 1 (D) ≤ ℓD kvkV ≤ (1 + Mγ2g ) 2 kvkH 1 (D) , so that the space V := H 1 (D)
equipped with the norm kvkV is a Hilbert space. Let µ♭ := ess inf x∈D (µ − 12 ∇·β)(x) and ν♭ :=
ess inf x∈∂D (ρ + 21 β·n)(x).
Proposition 31.15 (Coercivity, well-posedness). Assume that the ellipticity assumption (31.2)
holds. Assume that either µ♭ > 0, ν♭ ≥ 0 or µ♭ ≥ 0, ν♭ > 0. (i) The bilinear form aρ is V -coercive.
(ii) The problem (31.20) is well-posed.
Proof. The boundedness of aρ follows from
−1 −2 2
|aρ (v, w)| ≤ (λ♯ + β♯ Čps ℓD + µ♯ Čps ℓD + ρ♯ ℓD )kvkV kwkV ,
R R
for all v, w ∈ V. Moreover, the linear form ℓ(w) := D f w dx + ∂D gγ g (w) ds is bounded on V since
1
−1
|ℓ(w)| ≤ (Čps ℓD kf kL2(D) + ℓD
2
kgkL2 (∂D) )kwkV for all w ∈ V. Let us now prove the V -coercivity
of aρ . Let v ∈ V. Using (31.16), we infer that
aρ (v, v) ≥ λ♭ k∇vk2L2 (D) + µ♭ kvk2L2 (D) + ν♭ kγ g (v)k2L2 (∂D) . (31.24)
If µ♭ > 0 and ν♭ ≥ 0, we can drop the term multiplied by ν♭ in (31.24), and coercivity follows from
aρ (v, v) ≥ min(λ♭ , µ♭ ℓ2D )ℓ−2 2 2 2 −1

D kvkH 1 (D) ≥ min(λ♭ , µ♭ ℓD )(1 + Mγ g ) kvk2V .
If ν♭ > 0 and µ♭ ≥ 0, we can drop the term multiplied by µ♭ in (31.24), and coercivity follows from
aρ (v, v) ≥ min(λ♭ , ν♭ ℓD )kvk2V .
That (31.20) is well-posed follows from the Lax–Milgram lemma.

Example 31.16 (Pure diffusion). For a purely diffusive problem, coercivity holds if ρ is uni-
formly bounded from below away from zero.
Remark 31.17 (Variational formulation). Assume that β is identically zero in D. Owing
toR Proposition 25.8, u solves (31.20) Riff u minimizes in H 1 (D) the energy functional ER (v) :=
1
2 D ∇v·( d ∇v) + µv 2
− 2f v dx + 1
2 ∂D ργ g
(v)2
− 2gγ g
(v) ds.
31.3.2 Neumann condition

The Neumann condition is a particular case of the Robin condition in which ρ vanishes identically
on ∂D, i.e., we want to enforce
n·(d∇u) = g on ∂D. (31.25)
The following weak formulation is obtained by setting ρ to zero in (31.20):

R R (31.26)
a(u, w) = D f w dx + ∂D gγ g (w) ds, ∀w ∈ V.
Proposition 31.18 (Weak solution, well-posedness). Let f ∈ L2 (D) and g ∈ L2 (∂D). (i)
If the function u ∈ H 1 (D) solves (31.26), then it satisfies the PDE (31.1) a.e. in D and the
boundary condition (31.25) a.e. in ∂D in the sense that γ d (d∇u) = g in L2 (∂D). (ii) If the
ellipticity assumption (31.2) holds true and if µ♭ > 0 and ess inf x∈∂D (β·n)(x) ≥ 0, the bilinear
form a is V -coercive. (iii) The problem (31.26) is well-posed.
Proof. Set ρ := 0 in Propositions 31.13 and 31.15.
The coercivity assumption invoked in Proposition 31.18 fails when µ and β vanish identically
in D, i.e., for the purely diffusive problem
−∇·(d∇u) = f in D, n·(d∇u) = g on ∂D. (31.27)
Indeed, we observe that if u is a solution, then u + c is also a solution for all c ∈ R. A simple way
to deal with this arbitrariness is to restrict the
solution space to functions
whose mean value
R over
D is zero, i.e., we consider the space H∗1 (D) := v ∈ H 1 (D) | v D = 0 where v D := |D|−1 D v dx.
Note that a necessary condition for a solution to exist is the following compatibility condition on
f and g: Z Z
f dx + g ds = 0. (31.28)
D ∂D
R R R
Indeed, (31.22) implies that if (31.27) has a solution u, then D f dx+ ∂D g ds = − D ∇·(d∇u) dx+
hγ d (d∇u), 1i − 21 21 = 0.
H ,H
We now consider the following weak formulation:

Find u ∈ V := H∗1 (D) such that
R R (31.29)
ad (u, w) = D f w dx + ∂D gγ g (w) ds, ∀w ∈ V,
R
with the bilinear form ad (v, w) := D (d∇v)·∇w dx. Note that the test functions in (31.29) have
also zero mean value over D.
Proposition 31.19 (Well-posedness). Let f ∈ L2 (D) and g ∈ L2 (∂D) satisfy (31.28). (i)
If the function u ∈ H∗1 (D) solves (31.29), then it satisfies the PDE (31.27) a.e. in D and the
boundary condition a.e. in ∂D in the sense that γ d (d∇u) = g in L2 (∂D). (ii) Under the ellipticity
condition (31.2), ad is V -coercive. (iii) The problem (31.29) is well-posed.

1
d −2
Recall from (4.12) that the R normal trace operator γ : H(div; D) → H (∂D) is defined by
setting hγ d (v), γ g (w)i∂D := D (v·∇w + w∇·v) dx for all w ∈ H 1 (D), where h·, ·i∂D denotes the
1 1
duality pairing between H − 2 (∂D) and H − 2 (∂D). This definition makes sense since the full trace
1 R
operator γ g : H 1 (D) → H 2 (∂D) is surjective (see Theorem 3.10(iii)) and D (v·∇w+w∇·v) dx = 0
for all w ∈ H01 (D) and all v ∈ H(div; D), i.e., we have hγ d (v), γ g (w1 )i∂D = hγ d (v), γ g (w2 )i∂D if
γ g (w1 ) = γ g (w2 ).
Corollary 31.20 (Surjectivity of normal trace operator). Let D be a Lipschitz domain. The
1
normal trace operator γ d : H(div; D) → H − 2 (∂D) is surjective.
1 R −2
Proof. Let an ∈ H − 2 (∂D) and φ ∈ H 1 (D) solve D (∇φ·∇w + ℓD φw) dx = han , γ g (w)i
R ∂D for all
w ∈ H 1 (D). We have seen above that this problem has a unique solution. Since D (∇φ·∇w +
ℓ−2 ∞
D φw) dx = 0 for all w ∈ C0 (D),
−2
R we infer that ∆φ = ℓDR φ a.e. in D. −2 Hence, ∇φ ∈ H(div; D).
Moreover, hγ (∇φ), γ (w)i∂D := D (∇φ·∇w+w∆φ) dx = D (∇φ·∇w+ℓD wφ) dx = han , γ g (w)i∂D
d g
1
for all w ∈ H 1 (D). This proves that hγ d (∇φ) − an , li∂D = 0 for all l ∈ H 2 (∂D) since γ g is surjec-
d d
tive. In conclusion, we have established γ (∇φ) = an , i.e., γ is surjective.
31.3.3 Mixed Dirichlet–Neumann conditions

It is possible to combine the Dirichlet and the Neumann conditions. Let ∂Dd be a closed subset of
∂D and set ∂Dn := ∂D\∂Dd . We assume that both subsets ∂Dd and ∂Dn have positive (surface)
measures, and we enforce a Dirichlet and a Neumann condition on ∂Dd and ∂Dn , respectively:
u = gd on ∂Dd , n·(d∇u) = gn on ∂Dn , (31.30)
with gd and gn defined on ∂Dd and ∂Dn , respectively. We assume that there exists a bounded
1 1 1
extension operator H 2 (∂Dd ) → H 2 (∂D), i.e., there exists C∂Dd > 0 s.t. for all α ∈ H 2 (∂Dd ),
1
there is α̌ ∈ H 2 (∂D) s.t. α̌|∂Dd := α and C∂D d kα̌k 12 ≤ kαk 21 . Owing to Theorem 2.30,
H (∂D) H (∂Dd )
this assumption holds true if the interface between ∂Dd and ∂Dn is Lipschitz. Then let ǔd ∈ H 1 (D)
be s.t. γ g (ǔd ) = ǧd and let V := {v ∈ H 1 (D) | γ g (v) = 0 a.e. on ∂Dd }. Consider the weak
formulation:
(
Find u0 ∈ V such that
R R (31.31)
a(u0 , w) = D f w dx + ∂Dn gn γ g (w) ds − a(ǔd , w), ∀w ∈ V.
e 12 (∂Dn ) := {v ∈ H 12 (∂Dn ) | ve ∈ H 21 (∂D)}, where ve is the zero-extension of v to ∂D.

Let H
1
Proposition 31.21 (Well-posedness). Let f ∈ L2 (D), gd ∈ H 2 (∂Dd ), and gn ∈ L2 (∂Dn ).
(i) Under the above assumptions, if the function u0 ∈ H01 (D) solves (31.31), then the function
u := u0 + ǔd ∈ H 1 (D) satisfies the PDE (31.1) a.e. in D. It also satisfies the Dirichlet condition
a.e. on ∂Dd and the Neumann condition a.e. on ∂Dn in the sense that hγ d (d∇u), vei − 21 12 =
H ,H
R
gv ds for all v ∈ e 21 (∂Dn ). (ii) The problem (31.31) is well-posed under the assumptions of
H
∂Dn
Proposition 31.8.
Proof. We only sketch the proof.
(i) That u := u0 + ǔd satisfies the PDE in D is shown as above. The Dirichlet condition results
from γ g (u)|∂Dd = γ g (u0 )|∂Dd + γ g (ǔd )|∂Dd = γ g (ǔd )|∂Dd = ǧd|∂Dd = gd . To obtain the Neumann
condition, we observe that for all v ∈ H e 12 (∂Dn ), there is w ∈ V s.t. γ g (w) = ṽ. Using w as a test
R
function in (31.31), we infer that hγ d (d∇u), ṽi 1 1 = gv ds.
H − 2 ,H 2 ∂Dn
(ii) To prove the well-posedness of (31.31), we first notice that V is a closed subspace of H 1 (D).
Indeed, if (vn )n∈N is a Cauchy sequence in V, then vn → v in H 1 (D) as n → ∞. This implies that
1
γ g (v) = 0 a.e. on ∂Dd since γ g (vn ) → γ g (v) in H 2 (∂D). To conclude the proof, we use the following
Poincaré–Steklov inequality in V : There is C̃ps > 0 such that C̃ps kvkL2 (D) ≤ ℓD k∇vkL2 (D) for all
Rv ∈ V.g This inequality is a consequence of Lemma R 3.30 applied with the linear form f (v) :=
∂Dd γ (v) ds and p
:= 2 (notice that V ∋ v 7−→ ∂Dd γ g (v) ds restricted to constant functions is
nonzero since ∂Dd has positive measure).
Remark 31.22 (Data in H e 12 (∂Dn )′ ). The weak formulation (31.31) still makes sense if the
R
e 12 (∂Dn )′ , since the
boundary integral ∂Dn gn γ g (w) ds is replaced by gn (γ g (w)|∂D n ) where gn ∈ H
map V ∋ w 7→ γ g (w)|∂D ∈ H e 2 (∂Dn ) is bounded.
1
n
1 1
Remark 31.23 (H e 21 (∂Dn ) vs. H 2 (∂Dn )). In the literature, the interpolation space H 2 (∂Dn )
00 00
introduced in Lions and Magenes [286, Thm. 11.7] is sometimes invoked instead of H e 21 (∂Dn ).
′ 1
More precisely, if U is a Lipschitz domain in Rd (think of d′ := d − 1), we define H00 2
(U ) :=
1
2 1
[L (U ), H (U )] 1 (see Definition A.22). Then H (U ) ֒→ H
2 e 2 (U ) follows from Theorem A.27,
1
0 2 ,2 00
since the zero-extension operator maps boundedly L2 (U ) to L2 (Rd ) and H01 (D) to H 1 (Rd ) (since
1
1 − 12 6∈ N). Moreover, as observed in [286, Thm. 11.7] and Tartar [362, p. 160], “H00 2
(U ) is
1 p 2
characterized as the space of functions u in H 2 (U ) such that u/ d(x) ∈ L (U ), where d(x) is
the distance to the boundary ∂U ”, which according to Theorem 3.18 is also the characterization
1
e 12 (U ). Hence, H
of H e 12 (U ) = H 2 (U ). We also refer the reader to Chandler-Wilde et al. [115,
00
Cor. 4.10], where it is shown that {H e s (U ) | s ∈ R} is an interpolation scale (i.e., for all s1 < s2
and all s ∈ (s1 , s2 ), we have H e s (U ) = [H e s2 (U )]θ,2 with θ := (s − s1 )/(s2 − s1 )). The above
e s1 (U ), H
1
argument leads us to conjecture that the spaces H e 21 (∂Dn ) and H 2 (∂Dn ) are identical provided
00
the interface between ∂Dn and ∂Dd is smooth enough. Since we do not know any precise result
from the literature establishing this equality, we prefer to work with the space H e 12 (∂Dn ).
31.4 Elliptic regularity

The solution space V for scalar second-order elliptic PDEs is such that H01 (D) ⊆ V ⊆ H 1 (D)
depending on the type of boundary condition that is enforced. Since functions in V may not have
weak second-order derivatives, a natural question is whether it is possible to prove that the weak
solution enjoys higher regularity. The elliptic regularity theory provides theoretical results allowing
one to assert that under suitable assumptions on the smoothness of the domain and the data, the
weak solution sits indeed in a Sobolev space with higher regularity, e.g., in H 1+r (D) with r > 0.
We say that r is the index of elliptic regularity pickup. These results are important for the finite
element error analysis since convergence rates depend on the smoothness of the weak solution. In
this section, we consider elliptic regularity results in the interior of the domain and then up to the
boundary, with a particular attention paid to the case of Lipschitz domains. Most of the results
are just stated and we provide pointers to the literature for the proofs.
Besides the hypotheses on the PDE coefficients from §31.1.1, we implicitly assume that the
lower-order terms β and µ are s.t. the advection-reaction term β·∇v + µv has the same smoothness
as that requested for the source f for all v ∈ H 1 (D). For instance, when we assume f ∈ L2 (D),
we also implicitly assume that β ∈ L∞ (D) and µ ∈ Lr (D), with r > 2 and r ≥ d, so that
β·∇v + µv ∈ L2 (D) for all v ∈ H 1 (D).
31.4.1 Interior regularity

We first present a general result concerning interior regularity, i.e., regularity in any subset S ⊂⊂ D
(meaning that S ( D). Notice that we do not make any assumption on the boundary condition
satisfied by the weak solution or on the smoothness of D.
Theorem 31.24 (Interior regularity). Let D be a bounded open set. Assume that d ∈ C1 (D)
and f ∈ L2 (D). Let u ∈ H 1 (D) be any of the above weak solutions. Then for every open subset
S ⊂⊂ D, there are C1 , C2 (depending on S, D, and the PDE coefficients) such that
kukH 2 (S) ≤ C1 kf kL2 (D) + C2 kukL2 (D) . (31.32)
Proof. See Evans [196, §6.3.1]. The main tool for the proof is the technique of difference quotients
by Nirenberg [312], Agmon et al. [6].
Remark 31.25 (Sharper bound). If the weak formulation is well-posed, the bound (31.32)
takes the form kukH 2 (S) ≤ Ckf kL2 (D) owing to the a priori estimate kukH 1 (D) ≤ C ′ kf kL2(D) .
Remark 31.26 (Higher-order interior regularity). Let m be a nonnegative integer. Assume
that d ∈ Cm+1 (D), that the coefficients {βi }i∈{1: d} and µ are in C m (D), and that f ∈ H m (D).
Let u ∈ H 1 (D) be any of the above weak solutions. Then for every open subset S ⊂⊂ D, there
are C1 , C2 (depending on S, D, m, and the PDE coefficients) s.t. kukH m+2 (S) ≤ C1 kf kH m (D) +
C2 kukL2 (D) ; see [196, §6.3.1].
31.4.2 Regularity up to the boundary

We are now concerned with the smoothness of the weak solution up to the boundary. In this
context, the smoothness of ∂D and the nature of the boundary condition enforced on ∂D play a
role. The following theorems gather results established over the years by many authors. We refer
the reader to the textbooks by Grisvard [223, 224], Dauge [152] for more detailed presentations.
We consider three situations: domains having a smooth boundary, convex domains, and Lipschitz
domains. In what follows, we assume that the weak formulations are well-posed.
Theorem 31.27 (Smooth domain). Let D be a domain in Rd with a C 1,1 -boundary. Assume
that d is Lipschitz in D, i.e., there is L s.t.
kd(x) − d(y)kℓ2 (Rd×d ) ≤ Lkx − ykℓ2 (Rd ) , ∀x, y ∈ D. (31.33)
Let p ∈ (1, ∞) and assume that f ∈ Lp (D). (i) The weak solution to the Dirichlet problem with
1
boundary data g ∈ W 2− p ,p (∂D) is in W 2,p (D). (ii) The weak solution to the Neumann problem
1
with boundary data g ∈ W 1− p ,p (∂D) is in W 2,p (D). The same conclusion holds true for the Robin
problem if ρ is Lipschitz on ∂D.
Proof. See [223, Thm. 2.4.2.5-2.4.2.7] (see also [196, §6.3.2] for the Dirichlet problem and p :=
2).
Remark 31.28 (Neumann problem). Elliptic regularity for the Neumann problem is often
established under the assumptions of Proposition 31.18, i.e., the coefficient µ is uniformly bounded
from below away from zero. The Neumann problem (31.27) with the compatibility condition (31.28)
can be treated by observing that if u is the weak solution to this problem, then u is also the weak
solution to the Neumann problem set in H 1 (D) with the coefficient µ := µ0 > 0 and the source
term f replaced by f + µ0 u, where µ0 is any nonzero constant with the appropriate units.
Theorem 31.29 (Higher-order regularity). Let m be a positive integer. Assume that ∂D
is of class C m+1,1 , d ∈ Cm,1 (D), and f ∈ W m,p (D). Assume that the coefficients {βi }i∈{1: d}
1
and µ are in C m (D). Then the weak solution to the Dirichlet problem with g ∈ W m+2− p ,p (∂D)
is in W m+2,p (D). The same conclusion holds true for the Robin and Neumann problems if g ∈
1
W m+1− p ,p (∂D) and ρ ∈ C m,1 (∂D).
Proof. See [223, Thm. 2.5.1.1].
The smoothness assumption on ∂D can be relaxed if the domain D is convex. Notice that a
convex domain is Lipschitz; see [223, Cor. 1.2.2.3].
Theorem 31.30 (Convex domain). Let D be a convex domain. Assume that d is Lipschitz in
D. Let f ∈ L2 (D). (i) The weak solution to the Dirichlet problem with g := 0 is in H 2 (D). (ii)
The weak solution to the Robin or Neumann problem with g := 0 is in H 2 (D).
Proof. See [223, Thm. 3.2.1.2, 3.2.1.3, 3.2.3.1].
Elliptic regularity in Lipschitz domains is widely studied in the literature; see, e.g., Kondrat′ ev
[270], Maz′ ja and Plamenevskiı̆ [296], Jerison and Kenig [255, 256]. We first consider polygons in
R2 and quote results from [223, Chap. 4].
Theorem 31.31 (Polygon, d = I). Let D ⊂ R2 be a polygon with boundary vertices {Sj }1≤j≤J
where the segment joining Sj to Sj+1 corresponds to the boundary face denoted by Fj (setting
conventionally J + 1 := 1). Let θj ∈ (0, 2π) be the interior angle formed by the faces Fj and Fj+1 .
Assume that d is the identity matrix and that θj 6= π for all j ∈ {1:J}. Let f ∈ L2 (D). (i)
There is s0 ∈ ( 12 , 1] such that the weak solution to the Dirichlet problem enforcing u|Fj = gj , with
3
gj ∈ H 2 (Fj ) and gj (Sj ) = gj+1 (Sj ) for all j ∈ {1:J}, is in H 1+s (D) for all s ∈ [0, s0 ] and s0 = 1
∂u
if D is convex. (ii) The same conclusion holds true for Neumann problem enforcing ∂n |Fj = gj
1
with gj ∈ H 2 (Fj ) for all j ∈ {1:J}.
Proof. See [223, Cor. 4.4.4.14] (which treats mixed Dirichlet–Neumann conditions and Lp -Sobolev
spaces). The weak solution is in H 2 (D) up to singular perturbations that behave in radial coordi-
π
nates as r 2θj sin(θ 2θπj + ϕj ) in the vicinity of Sj with ϕj ∈ R; see Exercise 31.5.
Remark 31.32 (Variable coefficients). This case can be treated by freezing the diffusion tensor
at each polygon vertex and applying locally a coordinate transformation to recover the Laplace
operator; see [223, §5.2].
The analysis of elliptic regularity in a polyhedron is more intricate since vertex, edge, and
edge-vertex singularities can occur; see Grisvard [223, §8.2], Dauge [152, §5], Lubuma and Nicaise
[288, 289], Nicaise [311], Guo and Babuška [229, 230], Costabel et al. [147, 148]. For s ∈ (0, 1), let
us define the space H −1+s (D) either by interpolation between L2 (D) and H −1 (D) or as the dual
of H01−s (D) (the subspace of H 1−s (D) spanned by functions with zero trace on ∂D for s ∈ (0, 12 )).
These two definitions give the same space with equivalent norms.
Theorem 31.33 (Polyhedron, d = I, Dirichlet). Let D ⊂ R3 be a Lipschitz polyhedron.
There exists s0 > 12 , depending on D, such that the Laplace operator is an isomorphism from
H 1+s (D) ∩ H01 (D) to H −1+s (D) for all s ∈ [0, s0 ].
Proof. This is a consequence of Theorem 18.13 in Dauge [152, p. 158].

Theorem 31.34 (Lipschitz domain, Lipschitz diffusion). Let D be a Lipschitz domain in
Rd . Assume that d is Lipschitz in D. There is s0 ∈ (0, 12 ) such that the following holds true for all
1
s ∈ [0, s0 ]: (i) The weak solution to the Dirichlet problem with f ∈ L2 (D) and g ∈ H 2 +s (∂D) is in
1
H 1+s (D). (ii) The weak solution to the Neumann problem with f ∈ L2 (D) and g ∈ H − 2 +s (∂D)
is in H 1+s (D).
Proof. See Theorems 3 and 4 in Savaré [341]. Notice also that the lowest-order terms in the
PDE are in L2 (D) and that L2 (D) ⊂ H −1+s (D) for all s ≤ 1, so that f can be replaced by
f − β·∇u − µu.
Remark 31.35 (Very weak solution). It is possible to extend the notion of elliptic regularity
to the very weak solutions. Such solutions do not necessarily belong to the space H 1 (D). For
instance, using the transposition technique from Lions and Magenes [286, Chap. 2], it is shown in
Savaré [341] that the statement of Theorem 31.34 also holds true for all s ∈ (− 21 , 0).
The Lipschitz property of d is rather restrictive since it excludes domains composed of different
materials. Following Jochmann [258], it is possible to replace this hypothesis by a (usually called)
multiplier assumption, which consists of assuming that there is s0 ∈ (0, 21 ) such that
the map H s0 (D) ∋ ξ 7−→ dξ ∈ H s0 (D) is bounded. (31.34)
It is shown in Jochmann [258, Lem. 2] (see also Bonito et al. [70, Prop. 2.1]) that this property
holds true if D is partitioned into M disjoint Lipschitz subdomains {Dm }m∈{1: M} and if there is
a realPnumber α ∈ (s0 , 1] and there are diffusion tensors dm ∈ C0,α (Dm ) for all m ∈ {1:M }, s.t.
d := m∈{1: M} 1Dm dm, where 1Dm is the indicator function of Dm .
Theorem 31.36 (Piecewise smooth diffusion). Assume that there is s0 ∈ (0, 12 ) such that the
multiplier assumption (31.34) holds true. Then there is s ∈ (0, s0 ), depending on D and d, s.t. the
weak solution to the homogeneous Dirichlet problem or to the Neumann problem with f ∈ L2 (D)
(and g := 0) is in H 1+s (D).
Proof. See Theorem 3 in [258] or Lemma 3.2 in [70]. The statement also holds true for f in the
dual space of H01−s (D) for the Dirichlet problem and for f in the dual space of H 1−s (D) for the
Neumann problem. See also Bernardi and Verfürth [55] for Dirichlet conditions and piecewise
constant (or pcw. twice continuously differentiable) isotropic diffusion.
Theorem 31.36 also holds true for the mixed Dirichlet–Neumann problem. We refer the reader
to Jochmann [258] for more details on this question.
Exercises
Exercise 31.1 (Cordes). Prove that ellipticity implies the Cordes condition if d = 2. (Hint : use
that kdk2F = (tr(d))2 − 2 det(d).)
Exercise 31.2 (Poincaré–Steklov). Prove (31.23). (Hint : use (3.12).)
Exercise 31.3 (Potential flow). Consider the PDE ∇·(−κ∇u+βu) = f in D with homogeneous
Dirichlet conditions and assume that κ is a positive real number. Assume that β := ∇ψ for some
smooth function ψ (we say that β is a potential flow). Find a functional E : H01 (D) → R of which
the weak solution u is a minimizer on H01 (D). (Hint : consider the function e−ψ/κ u.)
Exercise 31.4 (Purely diffusive Neumann). Prove Proposition 31.19. (Hint : for all w ∈
H 1 (D), the function w̃ := w − w D is in H∗1 (D), use also the Poincaré–Steklov inequality from
Lemma 3.24.)
Exercise 31.5 (Mixed Dirichlet–Neumann). The goal is to show by a counterexample that
one cannot assert that the weak solution is in H 2 (D) for the mixed Dirichlet–Neumann problem
even if the domain and the boundary data are smooth. Using polar coordinates, set D := {(r, θ) ∈
(0, 1) × (0, π)}, ∂Dn := {r ∈ (0, 1), θ = π}, and ∂Dd := ∂D\∂Dn . Verify that the function
1 1
∂u
u(r, θ) := r 2 sin( 21 θ) satisfies −∆u = 0 in D, ∂n 1
|Dn = 0, and u|Dd = r 2 sin( 2 θ). (Hint : in polar
1 ∂ ∂u
1 ∂2u 2
coordinates, ∆u = r ∂r r ∂r + r2 ∂θ2 .) Verify that u 6∈ H (D).
Exercise 31.6 (H 2 (Rd )-seminorm). Prove that |φ|H 2 (Rd ) = k∆φkL2 (Rd ) for all φ ∈ C0∞ (Rd ).
(Hint : use Theorem B.3.)
Exercise 31.7 (Counterexample to elliptic regularity in W 2,∞ (D)). Let D be the unit disk
in R2 . Consider the function u(x1 , x2 ) := x1 x2 ln(r) with r2 := x21 + x22 (note that u|∂D = 0).
Verify that ∆u ∈ L∞ (D), but that u 6∈ W 2,∞ (D). (Hint : consider the cross-derivative.)
Exercise 31.8 (Domain with slit). Let D := {r ∈ (0, 1), θ ∈ (0, 2π)}, where (r, θ) are the polar
coordinates, i.e., D is the closed ball of radius 1 centered at 0. Let u(r, θ) := r cos( 12 θ) for all r > 0
and θ ∈ [0, 2π). (i) Let p ∈ [1, ∞). Is u|D in W 1,p (D)? Is u|int(D) in W 1,p (int(D))? (Hint : recall
Example 4.3.) (ii) Is the restriction to D of the functions in C 1 (D) dense in W 1,p (D)? (Hint :
argue by contradiction and use that kv|D kW 1,p (D) = kv|int(D) kW 1,p (int(D)) for all v ∈ C 1 (D).)
Exercise 31.9 (A priori estimate). Consider the PDE −κ0 ∆u + β·∇u + µ0 u = f with ho-
mogeneous Dirichlet conditions. Assume that κ0 , µ0 ∈ R, κ0 > 0, ∇·β = 0, β|∂D = 0, and
f ∈ H01 (D). Let ∇s β := 21 (∇β + (∇β)T ) denote the symmetric part of the gradient of β, and
assume that there is µ′0 > 0 s.t. ∇s β + µ0 Id ≥ µ′0 Id in the sense of quadratic forms. Prove
1
that |u|H 1 (D) ≤ (µ′0 )−1 |f |H 1 (D) and k∆ukL2(D) ≤ (4µ′0 κ0 )− 2 |f |H 1 (D) . (Hint : use −∆u as a test
function.) Note: these results are established in Beirão da Veiga [49], Burman [97].
Exercise 31.10 (Complex-valued diffusion). Assume that the domain D is partitioned into
two disjoint subdomains D1 and D2 . Let κ1 , κ2 be two complex numbers, both with positive
modulus and such that κκ12 6∈ R− . Set κ(x) := κ1 1D1 (x) + κ2 1D2 (x) for all x ∈ D. Let f ∈ L2 (D).
1
R
RShow that the problem of seeking u ∈ V := H0 (D; C) such that a(u, w) := D κ∇u·∇w dx =
D f w dx for all w ∈ V is well-posed. (Hint : use (25.7).)
Exercise 31.11 (Dependence on diffusion coefficient). Consider two numbers 0 < λ♭ ≤ λ♯ <
∞ and define the set K := {κ ∈ L∞ (D; R) | κ(x) ∈ [λ♭ , λ♯ ], a.e. x ∈ D}. Let V := H01 (D) equipped
with the norm kvkV := k∇vkL2 (D) and V ′ = H −1 (D). Consider the operator Tκ : V → V ′
s.t. Tκ (v) := −∇·(κ∇v) for all v ∈ V and all κ ∈ K. (i) Prove that λ♭ ≤ kTκ kL(V ;V ′ ) ≤ λ♯
and that TκR is an isomorphism. (Hint : use Proposition 31.8 with θ := 1 and the bilinear form
a(v, w) := D κ∇v·∇w dx on V × V.) (ii) Prove that kTκ − Tκ′ kL(V ;V ′ ) = kκ − κ′ kL∞ (D) for all
κ, κ′ ∈ K ∩C 0 (D; R). (Hint : if kκ−κ′ kL∞ (D) > 0, for all ǫ > 0 there is an open subset Dǫ ⊂ D such
that the sign of (κ−κ′ )|Dǫ is constant and |κ−κ′ | ≥ kκ−κ′ kL∞ (D) −ǫ in Dǫ ; then consider functions
in H01 (Dǫ ).) (iii) Let Sκ := Tκ−1 ∈ L(V ′ ; V ). Prove that λ2♭ kSκ − Sκ′ kL(V ′ ;V ) ≤ kκ − κ′ kL∞ (D) ≤
λ2♯ kSκ − Sκ′ kL(V ′ ;V ) for all κ, κ′ ∈ K ∩ C 0 (D; R). (Hint : Sκ − Sκ′ = Sκ (Tκ′ − Tκ )Sκ′ .)
Chapter 32
H 1-conforming approximation (I)
The goal of this chapter is to analyze the approximation of second-order elliptic PDEs using
H 1 -conforming finite elements. We focus the presentation on homogeneous Dirichlet boundary
conditions for simplicity. The well-posedness of the discrete problem follows from the Lax–Milgram
lemma and the error estimate in the H 1 -norm from Céa’s lemma. We also introduce a duality
argument due to Aubin and Nitsche to derive an improved error estimate in the (weaker) L2 -norm.
Some further topics on the H 1 -conforming approximation of second-order elliptic PDEs are covered
in the next chapter.
32.1 Continuous and discrete problems

Let D be a Lipschitz domain in Rd and let f ∈ L2 (D). We assume for simplicity that D is a
polyhedron. The model problem we want to approximate is the homogeneous Dirichlet problem:
−∇·(d∇u) + β·∇u + µu = f in D, u=0 on ∂D, (32.1)
with d ∈ L∞ (D), β ∈ W 1,∞ (D), µ ∈ L∞ (D). We assume that d is a symmetric second-order

tensor field and that its smallest eigenvalue is uniformly bounded from below by λ♭ > 0. We also
assume that (µ − 12 ∇·β) takes nonnegative values a.e. in D. The model problem is formulated as
follows:
(32.2)
a(u, w) = ℓ(w), ∀w ∈ V,
with the following bilinear and linear forms on V ×V and V, respectively:
Z Z

a(v, w) := (d∇v)·∇w + (β·∇v)w + µvw dx, ℓ(w) := f w dx.
D D
This problem is well-posed owing to the Lax–Milgram lemma. We equip the space V with the
norm kvkV := k∇vkL2 (D) = |v|H 1 (D) . This is legitimate owing to the Poincaré–Steklov inequality
Cps kvkL2 (D) ≤ ℓD k∇vkL2 (D) for all v ∈ H01 (D) (see (3.11) with p := 2), where ℓD is a characteristic
length of D, e.g., ℓD := diam(D). With this choice of norm, the coercivity and the boundedness
constants of the bilinear form a on V ×V are
−1 −2 2
α := λ♭ , kak := λ♯ + β♯ Cps ℓD + µ♯ Cps ℓD , (32.3)
94 Chapter 32. H 1 -conforming approximation (I)
with λ♯ := kdkL∞ (D) , β♯ := kβkL∞ (D) , and µ♯ := kµkL∞ (D) .

Let (Th )h∈H be a shape-regular sequence of affine meshes so that each mesh covers D exactly.
We approximate (32.2) with H 1 -conforming finite elements of some degree k ≥ 1. Let (K, b Pb , Σ)
b
b is a cuboid (see Chapter 6) or
be the reference finite element, e.g., a Qk,d Lagrange element if K
b
a Pk,d Lagrange element or the canonical hybrid element if K is a simplex (see Chapter 7). For
all K ∈ Th , let TK : Kb → K be the geometric mapping and let ψ g (v) := v ◦ TK be the pullback
K
g −1 b
by the geometric mapping. Let us define the local (polynomial) space PK := (ψK ) (P ). Let
b g 1 g
Pk (Th ) be the broken finite element space, Pk (Th ) the H -conforming subspace, and Pk,0 (Th ) its
zero-trace subspace. Recalling the construction from Chapter 19, we have
Pkb (Th ) := {vh ∈ L∞ (D) | vh|K ∈ PK , ∀K ∈ Th }, (32.4a)

Pkg (Th ) := {vh ∈ Pkb (Th ) | [[vh ]]F = 0, ∀F ∈ Fh◦ }, (32.4b)
g
Pk,0 (Th ) := {vh ∈ Pkg (Th ) | vh|∂D = 0}, (32.4c)
where Fh◦ (resp., Fh∂ ) is the collection of the mesh interfaces (resp., boundary faces) and [[vh ]]F
g
denotes the jump of vh across F . In other words, Pk,0 (Th ) is composed of functions that are
piecewise in PK , that are continuous across the mesh interfaces, and that vanish at the boundary.
Recalling Theorem 18.8, we have Pkg (Th ) ⊂ H 1 (D) and Pk,0
g
(Th ) ⊂ H01 (D).
The discrete problem is as follows:
g
Find uh ∈ Vh := Pk,0 (Th ) such that
(32.5)
a(uh , wh ) = ℓ(wh ), ∀wh ∈ Vh .
g
Since Pk,0 (Th ) ⊂ H01 (D), this problem is well-posed owing to the Lax–Milgram lemma. Note that
we enforce the homogeneous Dirichlet condition in an essential manner in (32.5). An alternative
technique weakly enforcing the Dirichlet condition by means of a boundary penalty method is
studied in Chapter 37. Using the notation from §26.3.4, we introduce the discrete solution map
Gh : V → Vh so that for all v ∈ V,
a(Gh (v) − v, wh ) := 0, ∀wh ∈ Vh . (32.6)
It follows from the Lax–Milgram lemma that Gh (v) is uniquely defined, and since u solves (32.2),
one readily sees that uh = Gh (u) iff uh solves (32.5). The main properties of the discrete solution
map are investigated in §26.3.4 in the abstract context of Galerkin methods. It is observed therein
that Gh is a projection and that kGh kL(V ) ≤ kakα .
Remark 32.1 (Variants). One must use the entire space Pkg (Th ) to enforce Robin/Neumann
conditions; see §31.3. When working with Dirichlet–Neumann conditions (see §31.3.3), one must
construct meshes that are compatible with the boundary partition ∂D = ∂Dd ∪∂Dn , i.e., boundary
faces cannot be split, they must belong either to ∂Dd or to ∂Dn .
32.2 Error analysis and best approximation in H 1

Let u solve (32.2) and let uh solve (32.5). Our goal is to bound the approximation error (u − uh ).
Recall that kvkV := k∇vkL2 (D) = |v|H 1 (D) .
Theorem 32.2 (H 1 -error estimate). The following holds true:
kak
k∇(u − uh )kL2 (D) ≤ inf k∇(u − vh )kL2 (D) , (32.7)
α vh ∈Vh
with the coercivity and boundedness constants α and kak defined in (32.3). Moreover, limh→0 k∇(u−
uh )kL2 (D) = 0, and assuming u ∈ H 1+r (D) with r ∈ (0, k], there is c s.t. for all h ∈ H,
X 21
k∇(u − uh )kL2 (D) ≤ c h2r 2
K |u|H 1+r (K) ≤ c hr |u|H 1+r (D) . (32.8)
K∈Th
Proof. The bound (32.7) follows from Céa’s lemma. We use a density argument and proceed as
in §26.3.3 to prove that limh→0 k∇(u − uh )kL2 (D) = 0. To prove (32.8), we start from (32.7)
g,av g,av
and estimate the infimum from above by taking vh := Ih0 (u), where Ih0 : L1 (D) → H01 (D) is
the quasi-interpolation operator with zero boundary trace introduced in §22.4.2. Using the esti-
g,av
mate (22.29) from Theorem 22.14 (with m := 1 and p := 2), we infer that k∇(u−Ih0 (u))kL2 (K) ≤
r 2
P 2
chK |u|H 1+r (ŤK ) , where |·|H 1+r (Ť ) := K ′ ∈ŤK |·|H 1+r (K ′ ) and ŤK is the collection of all the mesh
K
cells sharing at least one vertex with K. We obtain (32.8) by invoking the regularity of the mesh
sequence which implies that all the cells in ŤK have a diameter uniformly equivalent to hK and
that card(ŤK ) is uniformly bounded.
Remark 32.3 (Canonical or Lagrange interpolant). If 1 + r > d2 , one can also prove (32.8)
by replacing vh in (32.7) by either the canonical interpolant of u or the Lagrange interpolant of u,
both with zero boundary trace (see §19.4). This leads to k∇(u − vh )kL2 (K) ≤ chrK |u|H 1+r (K) for
all K ∈ Th , i.e., this argument circumvents the use of the subset ŤK .
Remark 32.4 (Condition number). The ratio kak α , which represents the condition number of
the bilinear form a (see Remark 25.12), can become very large when the lower-order terms in the
PDE (32.1) dominate the diffusive term. One then says that the PDE is singularly perturbed.
In this situation, one needs to use stabilized finite elements to obtain an accurate approximate
solution on a reasonably fine mesh. Examples can be found in Chapter 61.
Remark 32.5 (W 1,p -estimate). The reader is referred to the seminal work by Rannacher and
Scott [329] for W 1,p -error estimates on convex polygonal domains (d = 2) and quasi-uniform mesh
families with p ∈ [2, ∞]. Extensions to dimension three can be found in Guzmán et al. [234], and
extensions to graded meshes can be found in Demlow et al. [159].
Theorem 32.2 shows that the approximation error in the H 1 -seminorm is controlled by the best-
approximation error of u in Vh in the same norm, that is, by the quantity inf vh ∈Vh k∇(u−vh )kL2 (D) .
It is therefore interesting to investigate the behavior of this quantity. A question one may ask is
whether the broken finite element space Pkb (Th ) and its H01 (D)-conforming counterpart Vh :=
g
Pk,0 (Th ) have the same capacity to approximate a given function v ∈ H01 (D). In other words, did
we sacrifice anything in terms of best-approximation error by working with Vh rather than with
Pkb (Th )? We are going to show that, remarkably, this is not the case.
To better understand the above question, let us look at how the best-approximation errors in
Vh and in Pkb (Th ) are evaluated for a given function v ∈ H01 (D). When working in Vh , we need to
find a function vhg ∈ Vh s.t.
k∇(v − vhg )k2L2 (D) = min k∇(v − vh )k2L2 (D) . (32.9)

vh ∈Vh
Since k∇(v − vh )k2L2 (D) = k∇vh k2L2 (D) − 2(∇v, ∇vh )L2 (D) + k∇vk2L2 (D) and the function v is
kept fixed in our reasoning, we want to minimize over Vh the functional E : Vh → R defined by
E(vh ) := k∇vh k2L2 (D) − 2(∇v, ∇vh )L2 (D) . Owing to Proposition 25.8, this problem has a unique
minimizer in Vh characterized by the equations (∇(vhg − v), ∇wh )L2 (D) = 0 for all wh ∈ Vh . (Note
that uniqueness follows from the Poincaré–Steklov inequality since Vh ⊂ H01 (D).) In practice, one
can find vhg by inverting the global stiffness matrix associated with the global shape functions in Vh
(see §28.1.1). On the other hand, when working in Pkb (Th ), we need to find a function vhb ∈ Pkb (Th )
such that X X
k∇(v − vhb )k2L2 (K) = min k∇(v − vh )k2L2 (K) . (32.10)
vh ∈Pkb (Th )
K∈Th K∈Th
(We sum over the mesh cells since functions in Pkb (Th )
do not necessarily have a weak gradient in
′
L2 (D).) Since for all vh ∈ Pkb (Th ) and all K 6= K ∈ Th , the restrictions vh|K and vh|K ′ can be
chosen independently in the local polynomial spaces PK and PK ′ , we have
X X
k∇(v − vhb )k2L2 (K) = min k∇(v − q)k2L2 (K) , (32.11)
q∈PK
K∈Th K∈Th
b := b
and thus we need to find a function vK vh|K ∈ PK for all K ∈ Th s.t.
b
k∇(v − vK )k2L2 (K) = min k∇(v − q)k2L2 (K) . (32.12)
q∈PK
b b
Invoking Proposition 25.8, the above argument shows that the function vK is such that (∇(vK −
v), ∇q)L2 (K) = 0 for all q ∈ PK , and it is therefore uniquely defined up to an additive constant.
b b
It is convenient to require that (vK − v, 1)L2 (K) := 0. In practice, one finds each function vK by
inverting the local stiffness matrix associated with the local shape functions in PK .
Theorem 32.6 (Best-approximation error). There is a constant c such that the following
two-sided bounds hold true for all v ∈ H01 (D) and all h ∈ H:
X
min k∇(v − vh )k2L2 (K) ≤ min k∇(v − vh )k2L2 (D)
vh ∈Pkb (Th ) vh ∈Vh
K∈Th
X
≤c min k∇(v − vh )k2L2 (K) .
vh ∈Pkb (Th )
K∈Th
Proof. Let vP ∈ H01 (D). The first inequality follows from Vh being a subspace of Pkb (Th ) and
the identity K∈Th k∇(v − vh )k2L2 (K) = k∇(v − vh )k2L2 (D) if vh ∈ Vh . Let us prove the second
inequality. Recalling that the minimizers are denoted by vhg and vhb respectively, we need to prove
that X
k∇(v − vhg )k2L2 (D) ≤ c k∇(v − vhb )k2L2 (K) .
K∈Th
g,av
Let Jh,0 : Pkb (Th ) → Vh be the averaging operator defined in §22.4.1. Owing to Lemma 22.12
(with p := 2, r := 2, and m := 1) and since [[v]]F = 0 for all F ∈ Fh because v ∈ H01 (D), we have
g,av b −1
X
k∇(vhb − Jh,0 (vh ))kL2 (K) ≤ c hK 2 k[[v − vhb ]]F kL2 (F ) ,
F ∈F̌K
where FˇK is the collection of the mesh faces (interfaces and boundary faces) sharing at least one
vertex with K. Since the jump is the difference of the values from both sides of the interface (the
jump is the actual value for the boundary faces), we bound the jump by the triangle inequality.
Then we apply the multiplicative trace inequality (12.16) (with p := 2) and invoke the local
Poincaré–Steklov inequality (12.13) in all the cells having a face in F̌K (recall that vhb and v share
the same mean value in every mesh cell). This leads to
k∇(vhb − Jhg,av (vhb ))kL2 (K) ≤ c k∇(v − vhb )kL2 (DK ) ,
where DK is the set of the points composing the cells sharing at least one vertex with K. Since
Jhg,av (vhb ) ∈ Vh , the minimization property of vhg over Vh implies that k∇(v − vhg )k2L2 (D) ≤ k∇(v −
Jhg,av (vhb ))k2L2 (D) . Hence, we have
X
k∇(v − vhg )k2L2 (D) ≤ 2 k∇(v − vhb )k2L2 (K) + k∇(vhb − Jhg,av (vhb ))k2L2 (K)
K∈Th
X
≤2 k∇(v − vhb )k2L2 (K) + c k∇(v − vhb )k2L2 (DK ) .
K∈Th
We conclude by invoking the regularity of the mesh sequence.
Remark 32.7 (Literature). Theorem 32.6 is due to Veeser [371]. The present proof makes a
direct use of the averaging operator from §22.4.1.
32.3 L2-error analysis: the duality argument

The goal of this section is to derive an improved error estimate in a norm that is weaker than
that of V := H01 (D). This type of estimate is important, in particular, in the approximation
of eigenvalue problems (see Chapter 48). More precisely, the question we want to investigate
is whether it is possible to find some exponent γ > 0, uniform w.r.t. h ∈ H and u, such that
ku − uh kL2 (D) ≤ chγ ℓ1−γ
D k∇(u − uh )kL2 (D) ? (The length scale ℓD := diam(D) is introduced to
make the constant c dimensionless.)
32.3.1 Abstract duality argument

The above question can be formulated in a context more general than that of the boundary value
problem (32.1). Let us for a moment adopt an abstract point of view. Let V and L be two Banach
spaces such that V embeds continuously in L, i.e., V ֒→ L. Let a : V ×V → C be a bounded
sesquilinear form satisfying the assumptions of the BNB theorem (Theorem 25.9). Let Vh ⊂ V be
a finite-dimensional subspace equipped with the norm of V, and assume that the restriction of a
to Vh ×Vh satisfies a uniform inf-sup condition with constant αh , i.e., αh ≥ α0 > 0 for all h ∈ H.
A first important step toward answering the above question for the error measured in the L- and
V -norms is given by the following result due to Sayas [342].
Theorem 32.8 (Improved estimate ⇔ compactness). Let Gh : V → Vh ⊂ L be the discrete

solution map defined in (32.6), i.e., a(Gh (v) − v, wh ) := 0 for all wh ∈ Vh . Then the following
holds true if and only if the embedding V ֒→ L is compact:
!
kGh (v) − vkL
lim sup = 0. (32.13)
h→0 v∈V \Vh kGh (v) − vkV
Proof. See [342, Thm. 1.1] and Exercise 32.3.
An immediate consequence of Theorem 32.8 is that it is necessary that the embedding V ֒→ L

be compact to get a better convergence rate on kGh (u)−ukL than on kGh (u)−ukV . We now present
a result due to Aubin [28, 29] and Nitsche [313] that gives an estimate of the gain in convergence
rate that one should expect. Let ιL,V denote the operator norm of the above embedding, i.e.,
kvkL ≤ ιL,V kvkV for all v ∈ V. Recall that α and kak denote the coercivity and boundedness
constants of a on V ×V.
Definition 32.9 (Adjoint problem). Assume that L is a Hilbert space with inner product (·, ·)L .
For any g ∈ L, we denote by ζg ∈ V the unique solution to the following adjoint problem:
a(v, ζg ) = (v, g)L , ∀v ∈ V. (32.14)
The adjoint problem is well-posed since a : V ×V → C is a bounded sesquilinear form satisfying

the assumptions of the BNB theorem.
Lemma 32.10 (Aubin–Nitsche, abstract setting). Let ζu−Gh (u) solve the adjoint problem (32.14)
with data g := u − Gh (u), i.e., a(v, ζu−Gh (u) ) = (v, u − Gh (u))L for all v ∈ V. The following holds
true:
kζu−Gh (u) − wh kV
ku − Gh (u)kL ≤ kak inf ku − Gh (u)kV . (32.15)
wh ∈Vh ku − Gh (u)kL
Proof. Using g := u − Gh (u) and the test function v := u − Gh (u) in (32.14), and using the
definition of Gh (u) (that is, the Galerkin orthogonality property), we obtain
ku − Gh (u)k2L = a(u − Gh (u), ζu−Gh (u) ) = a(u − Gh (u), ζu−Gh (u) − wh ),
for all wh ∈ Vh . The assertion follows readily.
The factor that leads to an improved rate of convergence on the L-error is the infimum on the
right-hand side of (32.15). Assume that there is a subspace Y ֒→ V composed of functions that
can be approximated at a rate hγ in the V -norm by a function in Vh , that is, inf wh ∈Vh kζ − wh kV ≤
capp hγ ℓ−γ
D ιV,Y kζkY for all ζ ∈ Y, where ιV,Y is the operator norm of the above embedding, i.e.,
kykV ≤ ιV,Y kykY for all y ∈ Y. Assume that the adjoint solution to (32.14) enjoys a smoothness
kgkL . Setting c := capp csmo kak
ι
property of the form αkζg kY ≤ csmo ιL,VV,Y α , we conclude from (32.15)
that
ku − Gh (u)kL ≤ c hγ ℓ−γD ιL,V ku − Gh (u)kV . (32.16)
32.3.2 L2 -error estimate

Let us now return to the H 1 -conforming approximation of the elliptic PDE (32.1). Let u solve (32.2)
and let uh solve (32.5). Since the embedding H 1 (D) ֒→ L2 (D) is compact (this is the Rellich–
Kondrachov theorem), Theorem 32.8 says that it is possible to obtain a convergence rate on
ku − uh kL2 (D) that is better than that on k∇(u − uh )kL2 (D) . It is important to realize that the
compactness property is essential here (see Theorem 2.35).
Let us apply Lemma 32.10 with L := L2 (D) and V := H01 (D) equipped with the norms
−1
kvk := kvkL2 (D) and kvkV := k∇vkL2 (D) , respectively, so that ιL,V := Cps ℓD . For all g ∈ L2 (D),
the adjoint problem consists of seeking the function ζg ∈ H01 (D) s.t.
a(v, ζg ) = (v, g)L2 (D) , ∀v ∈ H01 (D). (32.17)

We assume that there is s ∈ (0, 1] and a constant csmo s.t. the following smoothing property holds
true:
kζg kH 1+s (D) ≤ csmo α−1 ℓ2D kgkL2 (D) , ∀g ∈ L2 (D). (32.18)
In the setting of §32.3.1, we have Y := H 1+s (D) with kζkY := kζkH 1+s (D) so that ιV,Y := ℓ−1 D .
Since d is symmetric, a distribution argument shows that −∇·(d∇ζg ) − β·∇ζg + µζg = g in D and
ζg = 0 on ∂D. Sufficient conditions for the smoothness property (32.18) to hold true then follow
from the elliptic regularity theory of §31.4. For instance, this property holds true with s ∈ ( 12 , 1]
if D is a Lipschitz polyhedron and the fields d, β, and µ are smooth. The maximal value s = 1 is
obtained for convex domains.
Lemma 32.11 (Aubin–Nitsche). Let u solve (32.2) and let uh solve (32.5). Assume that the
−1 kak
smoothing property (32.18) holds true for some s ∈ (0, 1]. There is c, depending linearly on Cps α ,
s.t. for all h ∈ H,
ku − uh kL2 (D) ≤ c hs ℓ1−s
D k∇(u − uh )kL2 (D) . (32.19)
−1
Proof. Direct consequence of (32.16) since ιL,V := Cps ℓD .
Remark 32.12 (Adjoint operator). Let A ∈ L(V ; V ′ ) be the operator associated with the
bilinear form a, i.e., hA(v), wiV ′ ,V := a(v, w) for all (v, w) ∈ V ×V. The adjoint operator A∗ ∈
L(V ; V ′ ) is s.t. hA∗ (v), wiV ′ ,V = hA(w), viV ′ ,V = a(w, v). Hence, the adjoint solution solves
A∗ (ζg ) = g.
Remark 32.13 (Best approximation in L2 and L∞ ). It is in general not true that there is c
s.t. ku − uh kL2 (D) ≤ c inf wh ∈Vh ku − wh kL2 (D) for all h ∈ H; see Babuška and Osborn [37, p. 58]
for a one-dimensional counterexample. However, if the mesh sequence is quasi-uniform, it is shown
in Schatz and Wahlbin [344, Thm. 5.1] that ku − uh kL∞ (D) ≤ c inf wh ∈Vh ku − wh kL∞ (D) if k ≥ 2,
and ku − uh kL∞ (D) ≤ c ln(ℓD /h) inf wh ∈Vh ku − wh kL∞ (D) if k = 1.
32.4 Elliptic projection

The operator defined below is a useful tool we are going to invoke often; see, e.g., §66.3.1 for
parabolic problems.
Definition 32.14 (Elliptic projection). Let V ⊂ H 1 (D) be a Hilbert space and assume that
v 7→ k∇vkL2 (D) is a norm on V. The discrete solution map Gh : V → Vh defined in (32.6) with
a(v, w) := (∇v, ∇w)L2 (D) is called elliptic projection and is denoted by ΠEh : V → Vh . Thus, for
all v ∈ V, we have
(∇(v − ΠEh (v)), ∇wh )L2 (D) = 0, ∀wh ∈ Vh . (32.20)
The two main properties of ΠEh are the following: (i) ΠEh is a projection, i.e., ΠEh (ΠEh (v)) = ΠEh (v)
for all v ∈ V ; (ii) Since by definition a(ΠEh (v) − v, wh ) = 0 for all wh ∈ Vh , one always has
|ΠEh (v)|H 1 (D) ≤ |v|H 1 (D) .
Theorem 32.15 (Approximation). Let s ∈ (0, 1] be the elliptic regularity index (i.e., there is
csmo s.t. for all ξ ∈ L2 (D), the solution to the adjoint problem a(v, z(ξ)) = (v, ξ)L2 (D) for all v ∈ V,
satisfies kzkH 1+s (D) ≤ csmo ℓ2D kξkL2 (D) ). There is c such that for all h ∈ H,
k∇(ΠEh (v) − v)kL2 (D) ≤ inf k∇(v − vh )kL2 (D) , (32.21a)

vh ∈Vh
kΠEh (v) − vkL2 (D) ≤ c hs ℓ1−s E

D k∇(Πh (v) − v)kL2 (D) . (32.21b)
Proof. The estimate (32.21a) is a consequence of
|ΠEh (v) − v|2H 1 (D) + |vh − ΠEh (v)|2H 1 (D) = |v − vh |2H 1 (D) , ∀vh ∈ Vh .
The estimate (32.21b) follows from (32.19) since ΠEh is the solution operator associated with the
bilinear form a(v, w) := (∇v, ∇w)L2 (D) on Vh .
Remark 32.16 (Other BCs). The elliptic projection is unambiguously defined when Dirichlet
conditions are applied on some part of ∂D with positive measure. In the case of Neumann con-
ditions, ΠEh acts only on H∗1 (D) := {v ∈ H 1 (D) | v = 0}, where v denotes the average of v over
D. We can extend ΠEh to H 1 (D) by setting ΠE∗h (v) := ΠEh (v − v) + v for all v ∈ H 1 (D). The
approximation properties of ΠE∗h in H 1 (D) are exactly the same as those of ΠEh in H∗1 (D).
Exercises
Exercise 32.1 (Discrete solution map). Let Gh be defined in (32.6). (i) Prove that k∇(v −
Gh (v))kL2 (D) ≤ chr |v|H 1+r (D) for all r ∈ (0, k], all v ∈ H 1+r (D), and all h ∈ H. (Hint : observe
g,av g,av
that Gh (Ih0 (v)) = Ih0 (v).) (ii) Assume that the adjoint operator A∗ has a smoothing property
in H 1+s
(D) for some real number s ∈ (0, 1]. Prove that kv − Gh (v)kL2 (D) ≤ chr+s ℓ1−s D |v|H 1+r (D) .
(Hint : consider the adjoint problem A∗ (ζ) = v − Gh (v).)
Exercise 32.2 (H −1 -estimate). Assume that for all g ∈ H 1 (D), the adjoint solution ζ ∈ H01 (D)
s.t. A∗ (ζ) = g satisfies kζkH 2+s (D) ≤ csmo α−1 ℓ2D kgkH 1 (D) with s ∈ ( 12 , 1]. Assume that k ≥
(v,z)L2 (D)
1 + s. Let kvkH −1 (D) := supz∈H01 (D) |z|H 1 (D) for all v ∈ L2 (D). Prove that ku − uh kH −1 (D) ≤
ch1+s ℓ1−s ∗
D k∇(u − uh )kL2 (D) . (Hint : consider the adjoint problem A (ζ) = z.)
Exercise 32.3 (Compactness). The goal is to prove Theorem 32.8. Let I : V → L be the natural
embedding and define ǫ(h) := supv∈V \Vh kG h (v)−vkL kak
kGh v−vkV . (i) Prove that kGh − IkL(V ;L) ≤ α ǫ(h),
where α and kak are the coercivity and the boundedness constants of a on V × V. (ii) Assume
that limh→0 ǫ(h) = 0. Prove that I is compact. (Hint : use (i).) (iii) Let R : L → V be s.t.
a(y, R(f )) := (y, f )L for all y ∈ V and all f ∈ L. Assuming that I is compact, prove that R is
compact. (Hint : prove that R = (A∗ )−1 I ∗ and use Schauder’s theorem; see Theorem C.48.) (iv)
Let PhV : V → Vh be the V -orthogonal projection onto Vh . Let Rh : L → Vh be the operator
defined by a(vh , Rh (f )) := (vh , f )L , for all vh ∈ Vh and all f ∈ L. Prove that kR − Rh kL(L;V ) ≤
kak V
α kR − Ph ◦RkL(L;V ) . (v) Assuming that I is compact, prove that limh→0 kR − Rh kL(L;V ) = 0.
(Hint : use (iii)-(iv) and proceed as in Remark C.5.) (vi) Assuming that I is compact, prove that
limh→0 ǫ(h) = 0.
Exercise 32.4 (Source approximation). Let f ∈ L2 (D), let Ihb (f ) be theR L2 -projection of f
onto Pkb′ (Th ). Consider the discrete problem (32.5) with Rthe right-hand side D Ihb (f )wh dx, that
g
is: Find uh ∈ Vh := Pk,0 (Th ) s.t. a(uh , wh ) = ℓh (wh ) := D Ihb (f )wh dx for all wh ∈ Vh . (i) How
should (32.7) be rewritten? Show that k ′ := k − 1 leads to an optimal H 1 -norm error estimate.
(ii) How should (32.19) be rewritten? Assuming full elliptic regularity, show that k ′ := k leads to
an optimal L2 -norm error estimate.
Exercise 32.5 (Advection-diffusion, 1D). Let D := (0, 1). Let ν, b be positive real numbers.
Let f : D → R be a smooth function. Consider the model problem −νu′′ + bu′ = f in D, u(0) = 0,
u(1) = 0. Consider H 1 -conforming P1 Lagrange finite elements on the uniform grid Th with nodes
1
xi := ih, ∀i ∈ {0:I}, and meshsize h := I+1 . (i) Evaluate the stiffness matrix. (Hint : factor out
the ratio h and introduce the local Péclet number γ := bh
ν
ν .) (ii) Solve the linear system when
f := 1 and plot the solutions for h := 10−2 and γ ∈ {0.1, 1, 10}. (Hint : wrtite U = U0 + Ũ ∈ RI
with U0i := b−1 ih and Ũi := ̺ + θδ i for some constants ̺, θ, δ.) (iii) Consider now the boundary
conditions u(0) = 0 and u′ (1) = 0. Write the weak formulation and show its well-posedness.
Evaluate the stiffness matrix. (Hint : the matrix is of order (I + 1).) Derive the equation satisfied
by h−1 (UI+1 − UI ), and find the limit values as h → 0 with fixed ν > 0 and as ν → 0 with fixed
h ∈ H.
Chapter 33
H 1-conforming approximation (II)
In this chapter, we study the following questions regarding the approximation of second-order
elliptic PDEs by H 1 -conforming finite elements: (i) How can non-homogeneous Dirichlet conditions
be taken into account in the error analysis, and how can they be implemented in practice; (ii) Can
the discrete problem reproduce the maximum principle, which is an important property enjoyed
by the exact problem; (iii) How quadratures impact the well-posedness and error analysis of the
discrete problem. Two other important topics treated in the forthcoming chapters are: (iv) The
derivation of a posteriori error estimates and their use for mesh adaptation (Chapter 34); (v) A
local post-processing technique to recover an H(div; D)-conforming flux approximating the exact
flux σ := −∇u (Chapter 52).
33.1 Non-homogeneous Dirichlet conditions

In this section, we consider the PDE (32.1), i.e.,
−∇·(d∇u) + β·∇u + µu = f in D, (33.1)
with the same assumptions on d, β, µ, and f as in §32.1, but with the non-homogeneous Dirichlet
1 1
condition γ g (u) = g on ∂D with g ∈ H 2 (∂D), where γ g : H 1 (D) → H 2 (∂D) is the trace map
such that γ g (v) = v|∂D for every smooth function v. Following §31.2.2, we invoke the surjectivity
1
of the trace map γ g to infer that there is Cγ g such that for all g ∈ H 2 (∂D), there is ug ∈ H 1 (D)
satisfying γ g (ug ) = g and kug kH 1 (D) ≤ Cγ g kgk 12 ; see Theorem 3.10(iii). The function ug is
H (∂D)
called lifting of the Dirichlet condition. By making the change of variable u0 := u − ug , we now
look for a function u0 satisfying the homogeneous Dirichlet condition γ g (u0 ) = 0. Let Ve := H 1 (D)
and V := H01 (D). The above considerations lead to the following weak formulation:
(
Find u ∈ Ve such that u0 := u − ug ∈ V satisfies
(33.2)
a(u0 , w) = ℓ(w) − e a(ug , w), ∀w ∈ V,
where Z

a(v, w) :=
e (d∇v)·∇w + (β·∇v)w + µvw dx, ∀(v, w) ∈ Ve ×V, (33.3)
D
R
a|V ×V , and ℓ(w) := D f w dx for all w ∈ V. We equip the space V with the norm kvkV :=
a := e
k∇vkL2 (D) . Let ℓD be a characteristic length of D, e.g., ℓD := diam(D), and let Cps be the
104 Chapter 33. H 1 -conforming approximation (II)
Poincaré–Steklov constant s.t. Cps kwkL2 (D) ≤ ℓD k∇wkL2 (D) = ℓD |w|H 1 (D) for all w ∈ H01 (D)
(see (3.11) with p := 2). Then the assumptions on d, β, and µ from §32.1 imply that a is
V -coercive, i.e., there is α > 0 s.t. a(v, v) ≥ αk∇vk2L2 (D) for all v ∈ V. Moreover, we have
|e
a(v, w)| ≤ ke
akℓD −1
kvkH 1 (D) k∇wkL2 (D) for all (v, w) ∈ Ve ×V, with ke −1
ak := λ♯ +β♯ Cps −1 2
ℓD +µ♯ Cps ℓD ,
λ♯ := kdkL∞ (D) , β♯ := kβkL∞ (D) , µ♯ := kµkL∞ (D) .
33.1.1 Discrete problem and well-posedness

We want to approximate the model problem (33.2) using the H 1 -conforming finite element space
Pkg (Th ) defined in (32.4b) and its zero-trace subspace Pk,0g
(Th ) defined in (32.4c). Since the function
g g
g may not be in γ (Pk (Th )), we need to approximate the non-homogeneous Dirichlet condition in
the discrete problem. To this purpose, we assume for simplicity that g ∈ C 0 (∂D), and we define
an approximation gh of g by using the boundary degrees of freedom (dofs) of the finite element.
Recall that the dofs are point-values for Lagrange elements, whereas they are point-values or
integrals over edges, faces, or cells for the canonical hybrid element. Let {ϕa }a∈Ah and {σa }a∈Ah
be, respectively, the global shape functions and dofs in Pkg (Th ) (see §19.2.1). Let s > d2 and
Ih : V g (D) := H s (D) → Pkg (Th ) denote either the canonical interpolation
P operator Ihg from §19.3,
or the Lagrange interpolation operator IhL . We have Ih (v) := a∈Ah σa (v)ϕa for all v ∈ V g (D).
Recall from Definition 19.11 that the set A∂h is the collection of the boundary dofs, i.e., a ∈ A∂h
iff γ g (ϕa ) = ϕa|∂D 6= 0. Then σa (v) only depends on γ g (v) for all a ∈ A∂h , i.e., we can write
σa (v) = (σa∂ ◦ γ g )(v) for all v ∈ V g (D), where σa∂ can be a value at a boundary point or an integral
over a boundary edge or a boundary face. Let us set
X
gh := σa∂ (g)ϕa|∂D . (33.4)
a∈A∂
h
We consider the following discrete problem:

(
Find uh ∈ Veh := Pkg (Th ) such that uh|∂D = gh and
g (33.5)
a(uh , wh ) = ℓ(wh ), ∀wh ∈ Vh := Pk,0
e (Th ).
At this stage, the discrete trial space includes boundary dofs (these dofs are fixed for uh by setting
uh|∂D = gh ), whereas the boundary dofs vanish for the discrete test functions. Upon introducing
P
the discrete lifting uhg := a∈A∂ σa∂ (g)ϕa ∈ Veh and making the change of variable uh0 := uh −uhg ,
h
we notice that uh0 ∈ Vh since uh|∂D = gh = uhg|∂D , i.e., uh0|∂D = 0. The discrete problem (33.5)
can then be recast in a form that is similar to (33.2), that is,
(
Find uh ∈ Veh such that uh0 := uh − uhg ∈ Vh satisfies
(33.6)
a(uh0 , wh ) = ℓ(wh ) − e
a(uhg , wh ), ∀wh ∈ Vh .
Lemma 33.1 (Well-posedness). The discrete problems (33.5) and (33.6) are well-posed.
Proof. Since Vh ⊂ V, the bilinear form a is bounded and coercive on Vh , and the linear form
ℓhg (·) := ℓ(·) − e
a(uhg , ·) is bounded on Vh . The Lax–Milgram lemma implies that the discrete
solution uh0 ∈ Vh is uniquely defined. Since the problems (33.5) and (33.6) are equivalent, the
discrete solution uh ∈ Veh is also uniquely defined.
33.1.2 Error analysis

The approximation setting leading to (33.5) is conforming since Veh ⊂ Ve and Vh ⊂ V. However,
there is a consistency error resulting from the fact that the non-homogeneous Dirichlet condition
is interpolated in (33.5).
d
Theorem 33.2 (H 1 -estimate). Let u solve (33.2) and let uh solve (33.5). Assume that k +1 > 2
and u ∈ H 1+r (D) with r ∈ ( d2 − 1, k]. There is c s.t. for all h ∈ H,
X 12
k∇(u − uh )kL2 (D) ≤ c h2r 2
K |u|H 1+r (K) ≤ c hr |u|H 1+r (D) . (33.7)
K∈Th
Proof. The proof is similar to that of Céa’s Lemma, except that we need to account for the
interpolation of the Dirichlet condition. Since 1 + r > d2 by assumption, we have u ∈ V g (D) and
Ih (u) is well defined. Owing to (33.4) and since σa (u) = σa∂ (γ g (u)) = σa∂ (g) for all a ∈ A∂h , we
infer that
X X
Ih (u)|∂D = σa (u)ϕa|∂D = σa∂ (g)ϕa|∂D = gh = uh|∂D .
a∈A∂
h a∈A∂
h
Hence, Ih (u) − uh ∈ Vh . Moreover, (33.2) and (33.5) imply that e

a(u − uh , wh ) = 0 for all
wh ∈ Vh ⊂ V := H01 (D). The coercivity of a on V and the boundedness of e a on Ve ×V imply
that
|a(Ih (u) − uh , wh )| |e
a(Ih (u) − uh , wh )|
αk∇(Ih (u) − uh )kL2 (D) ≤ sup = sup
wh ∈Vh k∇wh kL2 (D) wh ∈Vh k∇wh kL2 (D)
|e
a(Ih (u) − u, wh )|
= sup akℓ−1
≤ ke D ku − Ih (u)kH 1 (D) .
wh ∈Vh k∇wh kL2 (D)
ke
ak
The triangle inequality leads to k∇(u − uh )kL2 (D) ≤ cℓ−1 D ku − Ih (u)kH 1 (D) with c := 1 + α .
P −2 2(r+1) 1
Finally, Corollary 19.8 yields ℓ−1
D ku − Ih (u)kH 1 (D) ≤ ( K∈Th (ℓD hK + h2r 2
K )|u|H r+1 (K) ) , and
2
(33.7) follows since hK ≤ ℓD for all K ∈ Th .

We now use duality techniques to derive an improved L2 -norm error estimate. Recall from §32.3
that for all g ∈ L2 (D), the adjoint solution ζg ∈ H01 (D) is s.t. a(v, ζg ) = (v, g)L2 (D) for all
v ∈ H01 (D). Notice that ζ satisfies a homogeneous Dirichlet condition.
Theorem 33.3 (L2 -estimate). Assume that there is s ∈ ( 12 , 1] and csmo > 0 s.t. the adjoint
solution satisfies kζg kH 1+s (D) ≤ csmo α−1 ℓ2D kgkL2(D) . Assume that d is Lipschitz. Assume that
k + 1 > d2 and u ∈ H 1+r (D) with r ∈ ( d2 − 1, k]. There is c s.t. for all h ∈ H,
1
ku − uh kL2 (D) ≤ c hr+s ℓ1−s
D |u|H 1+r (D) + ℓD kg − gh kL2 (∂D) ,
2
(33.8)
where c depends linearly on ke

ak
α and the Lipschitz constant of ℓD λ♯−1 d.
Proof. The proof is similar to that of the Aubin–Nitsche lemma, except that we need to account for
the interpolation of the Dirichlet condition. Let ζ ∈ H01 (D) be the adjoint solution s.t. e a(v, ζ) =
(v, Ih (u) − uh )L2 (D) for all v ∈ H01 (D). The smoothness property implies that |ζ|H 1+s (D) ≤
ℓ−1−s
D kζkH 1+s (D) ≤ csmo α−1 ℓ1−s 1
D kIh (u) − uh kL2 (D) . Since Ih (u) − uh ∈ Vh ⊂ H0 (D), we obtain
kIh (u) − uh k2L2 (D) = a(Ih (u) − uh , ζ) = e

a(Ih (u) − u, ζ) + e
a(u − uh , ζ) =: T1 + T2 .
The term T1 is bounded using Lemma 33.4, leading to

1
akℓ−2
|T1 | ≤ c ke D ku − Ih (u)kH 1−s (D) + ℓD kg − gh kL2 (∂D) kζkH 1+s (D) ,
2
where we used that γ g (Ih (u) − u) = gh − g. For the term T2 , letting e := u − uh , we have
a on Ve ×V and the approximation properties of
a(e, vh ) = 0 for all vh ∈ Vh . The boundedness of e
e
g,av
Ih0 then give
g,av
akℓ−1
|T2 | ≤ ke D kekH 1 (D) k∇(ζ − Ih0 (ζ))kL2 (D)
akℓ−1
≤ c hs ke s
D kekH 1 (D) |ζ|H 1+s (D) ≤ c h kakℓD
−2−s
kekH 1 (D) kζkH 1+s (D)
1
akℓ−2−s
≤ c′ hs ke D ℓD k∇ekL2 (D) + ℓD 2
kg − gh kL2 (∂D) kζkH 1+s (D)
1
akℓ−2
≤ c′ ke s 1−s
D h ℓD k∇ekL2 (D) + ℓD kg − gh kL2 (∂D) kζkH 1+s (D) ,
2
1
where we used that kekH 1 (D) ≤ c(ℓD k∇ekL2 (D) + ℓD
2
kg − ghkL2 (∂D) ) owing to the Poincaré–Steklov
inequality (31.23), and h ≤ ℓD for the boundary term in the last line. Using the above bounds on
T1 and T2 , and the smoothness property of ζ, we infer that
1
kIh (u) − uh kL2 (D) ≤ c hs ℓ1−s
D k∇ekL2 (D) + ku − Ih (u)kH 1−s (D) + ℓD kg − gh kL2 (∂D) ,
2
where c depends linearly on keαak . We now use Theorem 33.2 to bound k∇ekL2 (D) , and the es-
timate ku − Ih (u)kH 1−s (D) ≤ cℓ1−s
D h
r+s
|u|H 1+r (D) which results from the Riesz–Thorin theorem
(Theorem A.27). We conclude by using the triangle inequality.
Lemma 33.4 (H 1+s -boundedness). Assume that d is Lipschitz. There is c, depending linearly
♯ d, s.t.
on the Lipschitz constant of ℓD λ−1
1
e akℓ−2
a(v, ζ) ≤ c ke g
D kvkH 1−s (D) + ℓD kγ (v)kL2 (∂D) kζkH 1+s (D) ,
2
(33.9)
for all v ∈ Ve := H 1 (D) and all ζ ∈ H01 (D) ∩ H 1+s (D) with s ∈ ( 21 , 1].
Proof. Since d is Lipschitz and ∇ζ ∈ H s (D), we infer that d∇ζ ∈ H s (D), i.e., there is c (depend-
♯ d) s.t. kd∇ζkH s (D) ≤ cλ♯ k∇ζkH s (D) . This implies
ing linearly on the Lipschitz constant of ℓD λ−1
that d∇ζ has a trace in H s− 12
(∂D), and hence that n·(d∇ζ) ∈ L2 (∂D) (since s > 21 ). Moreover,
∇·(d∇ζ) ∈ H −1+s
(D) and
k∇·(d∇ζ)kH −1+s (D) ≤ c ℓ−1
D kd∇ζkH s (D) ≤ c ℓD λ♯ k∇ζkH s (D)
′ −1
≤ c′′ λ♯ ℓ−2 ′′′

D kζkH 1+s (D) ≤ c ke akℓ−2
D kζkH 1+s (D) .
Here, we used that ∇ : H 1+s (D) → H s (D) and ∇· : H s (D) → H −1+s (D) are bounded owing to
the Riesz–Thorin theorem. Observing that H 1−s (D) = H01−s (D) (since 1 − s < 12 ), the linear
form ∇·(d∇ζ) can act on any v ∈ H 1−s (D) even if v does not have a zero trace at the boundary.
Denoting by h·, ·i the corresponding duality product between H −1+s (D) and H 1−s (D), we infer
that Z Z
h∇·(d∇ζ), vi + ∇v·(d∇ζ) dx = (n·(d∇ζ))γ g (v) ds.
D ∂D
As a result, the bilinear form e
a can be rewritten as
Z
a(v, ζ) = − h∇·(d∇ζ), vi +
e (n·(d∇ζ))γ g (v) ds
∂D
Z
+ (−β·∇ζ + (µ − ∇·β)ζ)v dx =: T1 + T2 + T3 .
D
The three terms on the right-hand side can be bounded as follows:
|T1 | ≤ k∇·(d∇ζ)kH −1+s (D) kvkH 1−s (D) ≤ c ke

akℓ−2
D kζkH 1+s (D) kvkH 1−s (D) ,
|T2 | ≤ kn·(d∇ζ)kL2 (∂D) kγ g (v)kL2 (∂D) ≤ c ke

−3
a(kℓD 2 kζkH 1+s (D) kγ g (v)kL2 (∂D) ,
−2
|T3 | ≤ c ke
akℓD akℓ−2
kζkH 1 (D) kvkL2 (D) ≤ c ke D kζkH 1+s (D) kvkH 1−s (D) .
33.1.3 Algebraic viewpoint

Let us enumerate the dofs using the set Ih := {1:I}. We identify a block structure by enumerating
first the internal dofs by using the set Ih◦ := {1:I ◦ }, then we enumerate the boundary dofs by
using the set Ih∂P:= {1:I ∂ }. Notice that I = I ◦ + I ∂ . Introducing the decomposition of the discrete
solution uh := i∈Ih Ui ϕi , the algebraic realization of (33.5) is the linear system AU = B, where
the stiffness matrix A and the load vector B have entries given by A◦◦ ◦
ij := a(ϕj , ϕi ) and Bi := ℓ(ϕi )
◦ ◦ ◦∂ := ◦ ∂
for all (i, j) ∈ Ih ×Ih , and Aij a(ϕj , ϕi ) for all (i, j) ∈ Ih ×Ih . Note that the row index
associated with the test function takes values in Ih◦ only. Moreover, the boundary prescription
uh|∂D = gh in (33.5) leads to Ui = B∂i := σi∂ (g) for all i ∈ Ih∂ . Thus, we obtain the following
block-decomposition (with obvious notation)
" # ◦
A◦◦ A◦∂ U◦ B
= ∂ , (33.10)
O II ∂ U∂ B
where O is a zero rectangular matrix of order I ∂ × I ◦ and II ∂ is the identity matrix of order I ∂ .
The matrix A◦◦ is of size I ◦ and is invertible owing to the H01 -coercivity of a.
A first option to solve (33.10) is to eliminate U∂ , i.e., to solve the linear system A◦◦ U◦ =
B − A◦∂ B∂ . The advantage is that the final size of the linear system is optimal since only the
◦
internal dofs are unknown. However, this technique requires assembling two matrices, A◦◦ and
A◦∂ , instead of one, and the two matrices have a different sparsity profile. An alternative technique
consists of assembling first the stiffness matrix for all the dofs in Ah (this is the stiffness matrix
for the Neumann problem) and then correcting the rows for a ∈ A∂h by setting the entries to zero
except the diagonal ones which are set to 1. The right-hand side of (33.10) is assembled similarly.
Despite the slight increase in the number of unknowns, this technique is computationally effective.
It has the apparent drawback of breaking the symmetry of the model problem (recall that A◦◦ is
symmetric if the advective velocity is zero) since the matrix in (33.10) is not symmetric. Actually,
when using an iterative solution method based on a Krylov space, if the initial residual is zero
for the boundary dofs, it is always zero during the iterations, and the iterative algorithm behaves
exactly as if the boundary dofs are eliminated; see Exercise 33.2. Of course, in practice the way
the boundary and interior dofs are enumerated does not matter.
Remark 33.5 (Penalty method). Another way of enforcing Dirichlet conditions without elim-
ination is to use a penalty method. First, one assembles the matrix and the right-hand side
of the homogeneous Neumann problem. Then, for each row associated with a ∈ A∂h , one adds
ǫ−1 to the diagonal entry of the stiffness matrix and ǫ−1 σa∂ (g) to the right-hand side; see Lions
[285], Babuška [35]. If ǫ−1 is not large enough, the method suffers from a lack of consistency.
As shown in Chapter 37, this problem can be avoided by adding extra boundary terms ensuring
consistency; see Nitsche [314].
33.2 Discrete maximum principle

The maximum principle is an important property of scalar second-order elliptic PDEs that sets
them apart from higher-order PDEs and systems of PDEs. We focus here on the PDE (33.1)
equipped with Dirichlet boundary conditions. Thus, the weak formulation is again (33.2).
Theorem 33.6 (Maximum principle). Let D be a Lipschitz domain in Rd . Let d, β, and µ

satisfy the assumptions in §31.1.1. Let f ∈ L2 (D) and let u ∈ H 1 (D) satisfy (33.1). (i) If µ = 0
in D, then f ≤ 0 a.e. in D implies that u ≤ ess sup∂D u a.e. in D, and f = 0 in D implies that
ess inf ∂D u ≤ u ≤ ess sup∂D u a.e. in D. (ii) Assume min(ess inf D (µ − ∇·β), ess inf D µ) > 0. Then
the following holds true a.e. in D:
min(ess inf u, ess inf (µ−1 f )) ≤ u ≤ max(ess sup u, ess sup(µ−1 f )). (33.11)
∂D D ∂D D
Proof. For the proof of (i), see Gilbarg and Trudinger [216, Thm. 8.1], Brezis [89, Prop. 9.29], or
Evans [196, §6.4]. Let us prove (ii) by following Brezis [89, Prop. 9.29], i.e., we use Stampacchia’s
truncation technique. Let G ∈ C 1 (R) be such that G(t) = 0 for all t ≤ 0, and 0 < G(t),
0 < G′ (t) < M for all t > 0. Let K := max(ess sup∂D u, ess supD µ−1 f ) and assume that K < ∞,
otherwise there is nothing to prove. Note that ζK (u) := G(u − K) ∈ H 1 (D) and ζK (u)|∂D = 0 a.e.
(since (u − K)|∂D ≤ 0 a.e. in ∂D), so that ζK (u) ∈ H01 (D). Testing the weak formulation (33.2)
with ζK (u), we infer that
Z Z
kd ∇ukℓ2 (Rd ) G (u − K) + (β·∇u + µu)ζK (u) dx =
1
2 ′
2 f ζK (u) dx.
D D
This proves that

Z Z

(β·∇u + µ(u − K))ζK (u) dx ≤ (f − µK)ζK (u) dx ≤ 0,
D D
where the last bound follows from f − µK ≤ 0 and 0 ≤ ζK (u) a.e. in D by definition of G. Let
Rt R R
F (t) := 0 G(z) dz. We have D (β·∇u)ζK (u) dx = D (β·∇(ηK (u)) dx with ηK (u) := F (u − K).
Integrating by parts the advective derivative and since ηK (u)|∂D = 0 (because F (t) = 0 for all
t ≤ 0), we infer that Z

−ηK (u)∇·β + µ(u − K)ζK (u) dx ≤ 0. (33.12)
D
Rt
The definition of F implies that F (t) + 0 zG′ (z) dz = tG(t), which applied to t := u − K yields
R u−K
ηK (u) + 0 zG′ (z) dz = (u − K)ζK (u). Using this identity in (33.12) implies that
Z Z u−K
′
ηK (u)(µ − ∇·β) + µ zG (z) dz dx ≤ 0.
D 0
R u−K
Using the assumption min(ess inf D (µ−∇·β), ess inf D µ) > 0 together with ηK (u) ≥ 0 and 0 zG′ (z) dz ≥
R
0 a.e. in D, we conclude that D ηK (u) dx = 0. This means that ηK (u) = 0 a.e. in D, i.e., u−K ≤ 0
a.e. in D.
Remark 33.7 (Sign change). Owing to the linearity of the PDE, Theorem 33.6(i) can be adapted
to a sign change, e.g., if µ = 0 in D, f ≥ 0 a.e. in D implies that u ≥ ess inf ∂D u a.e. in D.
The discrete analogue of Item (i) in Theorem 33.6 is called discrete maximum principle (DMP)
(see Ciarlet and Raviart [127] for one of the pioneering works on this topic). As in [127], we
only consider linear finite elements on simplicial meshes with homogeneous Dirichlet conditions
(see Vejchodský and Šolı́n [374] for a 1D example where the DMP is shown to hold with higher-
g
order elements). Let P1,0 (Th ) be the H01 (D)-conforming finite element space using linear finite
g
elements. Let {ϕi }i∈{1: I} denote the global shape functions in P1,0 (Th ), where I now denotes the
I×I
number of interior mesh vertices, and let A ∈ R be the stiffness matrix.
P For a vector V ∈ RI ,
the notation V ≤ 0 means that Vi ≤ 0 for all i ∈ {1:I}. Then uh := i∈{1: I} Ui ϕi ≤ 0 on D iff
U ≤ 0 in RI since the linear shape functions are nonnegative (this equivalence is no longer valid
for higher-order finite elements).
Definition 33.8 (DMP). We say R that the DMP holds true for the discrete problem (32.5) with
g
Vh := P1,0 (Th ) and Bi := ℓ(ϕi ) := D f ϕi dx for all i ∈ {1:I}, if
[ B ≤ 0 in RI ] =⇒ [ U ≤ 0 in RI ]. (33.13)
We now formulate conditions on A that are equivalent to, or imply, the DMP. Let us recall
from Definition 28.16 the notions of Z-matrix and M -matrix. A matrix A ∈ RI×I is said to be a
Z-matrix if Aij ≤ 0 for all i, j ∈ {1:I} with i 6= j. A matrix A is said to be a nonsingular M -matrix
if it is a Z-matrix, invertible, and (A−1 )ij ≥ 0 for all i, j ∈ {1:I}.
Lemma 33.9 (Stiffness matrix). (i) The DMP holds iff (A−1 )ij ≥ 0 for all i, j ∈ {1:I}. (ii)
The DMP holds if A is a Z-matrix.
Proof. The statement (i) follows from the fact that A−1 B ≤ 0 for all B ≤ 0 iff (A−1 )ij ≥ 0 for all
i, j ∈ {1:I}. Let us now prove the statement (ii). We follow Jiang and Nochetto [257]. Assume
that A is a Z-matrix. Letting z + := max(0, z), z − := z − z + = min(0, z) for all z ∈ R, and
Π+ : RI → RI be such that (Π+ (V))i = Vi+ for all i ∈ {1:I}, we have for all V ∈ RI ,
X X
Π+ (V)T A(V − Π+ (V)) = Vi+ Aij Vj− = Vi+ Aij Vj− ≥ 0,
i,j∈{1: I} i,j∈{1: I},i6=j
since A is a Z-matrix, i.e., Aij ≤ 0 for all i 6= j. Let now B ≤ 0 and assume that U ∈ RI solves
AU = B. We want to prove that U ≤ 0. We have
0 ≥ Π+ (U)T B = Π+ (U)T AU ≥ Π+ (U)T AΠ+ (U),
which implies that Π+ (U) = 0 since A is positive definite (owing to the coercivity of the bilinear
form a). In other words, U ≤ 0.
Remark 33.10 (Nonsingular M -matrix). A consequence of Lemma 33.9 is that if A is a Z-
matrix, then it is a nonsingular M -matrix. Indeed, if A is a Z-matrix, Item (ii) implies that the
DMP is satisfied, and Item (i) then implies that (A−1 )ij ≥ 0 for all i, j ∈ {1:I}. This shows that
A is a nonsingular M -matrix (see Definition 28.16); see also Exercises 33.3 and 33.5.
Lemma 33.9 shows that a sufficient condition for the DMP to hold is that the stiffness matrix
A is a Z-matrix. Since ensuring this property on general diffusion-advection-reaction PDEs is
delicate, we continue the discussion by focusing on the Poisson equation, i.e., d := Id , β := 0, and
µ := 0. For all i ∈ {1:I}, let I(i) := {j ∈ {1:I} | ϕi ϕj 6≡ 0} and I ∗ (i) := I(i)\{i}.
Definition 33.11 (Weakly acute meshes). A simplicial mesh is said to be weakly acute if the
stiffness matrix of the Laplacian is a Z-matrix, i.e.,
Z
∇ϕi ·∇ϕj dx ≤ 0, ∀i ∈ {1:I}, ∀j ∈ I ∗ (i). (33.14)
D
The condition (33.14) is always satisfied in dimension d = 1. In higher dimension, (33.14) boils
down to a geometric restriction on the mesh. Notice that for all i ∈ {1:I} and all j ∈ I ∗ (i), the
collection of the mesh cells having both zi and zj as vertices, say Tij , is nonempty. Let K ∈ Tij .
Let FK,i (resp. FK,j ) be the face of K opposite to zi (resp. zj ). Let nK,i , nK,j be the two unit
∗ ∗
normal vectors to FK,i and FK,j , respectively, pointing outward. Let zK,i (resp. zK,j ) be the
2 −1 ∗
ℓ -orthogonal projection of zi onto FK,i (resp. FK,j ). We have ϕi|K (x) = hK,i (zK,i − x)·nK,i
and ∇ϕi|K = −h−1 ∗ 1
K,i nK,i with hK,i := kzi − zK,i kℓ2 . Recalling that |K| = d |FK,i |hK,i and setting
cos(αK,ij ) := −nK,i ·nK,j (i.e., αK,ij ∈ (0, π) is the dihedral angle between FK,i and FK,j ), we
infer that Z
|FK,i ||FK,j |
∇ϕi ·∇ϕj dx = − cos(αK,ij ). (33.15)
K d2 |K|
Thus, a sufficient condition for (33.14) to hold true, that is, for the mesh to be weakly acute, is that
for all K ∈ Th and for every pair of distinct faces of K, say F, F ′ , we have nK|F ·nK|F ′ ≤ 0, i.e., the
dihedral angle between F and F ′ is in (0, π2 ]. We say in this case that the mesh is nonobtuse; see
also Brandts et al.P [84]. However,
R a weakly acute mesh is not necessarily nonobtuse since (33.14)
only requires that K∈Tij K ∇ϕi ·∇ϕj dx ≤ 0, whereas (33.15) requires that each term in the sum
is nonpositive. This leads us to look for a necessary and sufficient condition so that (33.14) holds
true.
Lemma 33.12 (Geometric identity). The following holds true (with the convention that |FK,i ∩
FK,j | = 1 for d = 2): Z
|FK,i ∩ FK,j |
∇ϕi ·∇ϕj dx = − cot(αK,ij ). (33.16)
K d(d − 1)
Proof. Since K is fixed, we drop the index K in the proof. Since Fj is a simplex in Rd−1 , we have
1 ∗ ∗
|Fj | = d−1 |Fi ∩ Fj |hi,j where hi,j := kzi,j − zi kℓ2 and zi,j is the projection of zi onto Fi ∩ Fj .
Thus, we have
cos(αij ) cos(αij )|Fj | |Fi ∩ Fj | cos(αij ) hi,j
∇ϕi ·∇ϕj |K = − =− =− ,
hi hj hi d|K| d(d − 1)|K| hi
h 1
and it remains to show that hi,j i
= sin(α ij )
. Letting m′j = nj −(nj ·ni )ni so that km′j kℓ2 = sin(αij ),
1
we set n′j := sin(α ij )
(nj − (nj ·ni )ni ). The set {ni , n′j } is an orthonormal basis of the plane
orthogonal to the (d − 2)-dimensional manifold Fi ∩ Fj . Let zk be one of the (d − 1) vertices in
Fi ∩ Fj . By definition, zi∗ − zk = zi − zk − ((zi − zk )·ni )ni and zi,j ∗
− zk = zi − zk − ((zi −
′ ′
zk )·ni )ni − ((zi − zk )·nj )nj . Hence, we have
h2i,j = kzi∗ − zi − ((zi − zk )·n′j )n′j k2ℓ2 = h2i + |(zi − zk )·n′j |2 .

cos2 (α )
But |(zi − zk )·n′j |2 = sin2 (αij
ij 2 2 2
) hi since (zi − zk )·nj = 0 and |(zi − zk )·ni | = hi . Thus, we have
2
cos (αij ) 1
h2i,j = 1 + sin2 (αij 2 2
) hi = sin(αij )2 hi .
Corollary 33.13 (Necessary and sufficient condition). The condition (33.14) is fulfilled iff
X
|FK,i ∩ FK,j | cot(αK,ij ) ≥ 0, ∀i ∈ {1:I}, ∀j ∈ I ∗ (i). (33.17)
K∈Tij
If d = 2, Tij consists of only two cells, and the identity (cot(α)+cot(β)) sin(α) sin(β) = sin(α+β)
shows that (33.14) holds true iff the sum of the two angles opposite to any interior face is less than
or equal to π; see Xu and Zikatanov [396, Eq. (2.5)].
Remark 33.14 (Obstructions). It is noticed in Brandts et al. [83, §5.2] that (33.14) cannot
hold true in dimension d ≥ 4, and that the strict inequality cannot hold true in dimension three
on Cartesian meshes.
Remark 33.15 (Nonlinear stabilization). An alternative approach to enforce the discrete

maximum principle that avoids geometric requirements on the mesh is to add a nonlinear viscosity
term to the discrete problem; see Burman and Ern [98, 99], Barrenechea et al. [45].
33.3 Discrete problem with quadratures

In this section, we study the influence of quadratures when approximating a scalar elliptic PDE
by means of finite elements.
33.3.1 Continuous and discrete settings

For simplicity, we drop the lower-order terms in the PDE which becomes −∇·(d∇u) = f in D,
and we consider homogeneous Dirichlet boundary conditions, i.e., u = 0 on ∂D. Thus, the weak
formulation is posed in V := H01 (D), which we equip with the norm kvkV := k∇vkL2 (D) = |v|H 1 (D) ,
and the bilinear and linear forms are
Z Z
a(v, w) = (d∇v)·∇w dx, ℓ(w) = f w dx. (33.18)
D D
We assume that the model problem is well-posed, i.e., the conditions of the BNB theorem (or the
|a(v,w)|
Lax–Milgram lemma) are fulfilled. Hence, there is α > 0 s.t. αk∇vkL2 (D) ≤ supw∈V k∇wk for
L2 (D)
all v ∈ V.
g
We consider the H01 -conforming finite element space Vh := Pk,0 (Th ) ⊂ H01 (D) defined in (32.4c).
If the integrals defining the forms a and ℓ are evaluated exactly in the discrete problem, well-
posedness follows automatically in the setting of the Lax–Milgram lemma if a is V -coercive, and
well-posedness holds in the setting of the BNB theorem if the bilinear form a satisfies the following
uniform inf-sup condition on Vh ×Vh for all h ∈ H (see Chapter 32):
|a(vh , wh )|
∃α0 > 0, ∀vh ∈ Vh , α0 k∇vh kL2 (D) ≤ sup . (33.19)
Moreover, in both situations one obtains H 1 -seminorm error estimates of order hr if u is in H 1+r (D)
with r ∈ (0, k]. But in practice the integrals defining the forms a and ℓ have to be evaluated
approximately by means of quadratures as described in Chapter 30. Therefore, a natural question
is whether the quadratures impact the well-posedness of the discrete problem and its error analysis.
Recalling Definition 30.1, we consider a quadrature in K b with nodes {ξbl }l∈{1: l } and weights
Q
{bωl }l∈{1: lQ } . The largest integer k such that the quadrature is exact for any polynomial in Pk,d
is called quadrature order and is denoted by kQ . Recalling Proposition 30.2, we construct a
quadrature in every mesh cell K ∈ Th by setting {ξlK := TK (ξbl )}l∈{1: lQ } for the nodes and
{ωlK := ω bl |det(JK (ξbl ))|}l∈{1: lQ } for the weights, where TK : K
b → K is the geometric mapping.
RThis quadrature P allows us to approximate the integral of any continuous function φ in K as
φ(x) dx ≈ l∈{1: lQ } ωlK φ(ξlK ). The quadrature error EK : C 0 (K) → R is defined by setting
K R P
EK (φ) := K φ(x) dx − l∈{1: lQ } ωlK φ(ξlK ).
33.3.2 Well-posedness with quadratures

For simplicity, we assume that the mesh Th is affine and that the diffusion coefficients dij and the
source term f are continuous in every mesh cell K ∈ Th . The use of a quadrature in every mesh
cell to evaluate the exact forms a and ℓ defined in (33.18) leads to the following approximate forms:
X X
aQ (vh , wh ) := ωlK d(ξlK )∇vh (ξlK ) ·∇wh (ξlK ), (33.20a)
K∈Th l∈{1: lQ }
X X
ℓQ (wh ) := ωlK f (ξlK )wh (ξlK ), (33.20b)
K∈Th l∈{1: lQ }
for all (vh , wh ) ∈ Vh ×Vh . It is not possible in general to extend aQ and ℓQ to H01 (D), since
functions in H01 (D) are not necessarily defined pointwise. The discrete problem with quadratures
is formulated as follows:
(33.21)
aQ (uh , wh ) = ℓQ (wh ), ∀wh ∈ Vh .
Lemma 33.16 (Well-posedness). Assume that Pb ⊂ Pl,d for some integer l ≥ 1. Assume that
kQ ≥ 2l − 2 and that d ∈ W1,∞ (Th ) := W 1,∞ (Th ; Rd×d ). Assume that the inf-sup condition
α0
(33.19) is satisfied. Define the length scale ℓ0 := |d| 1,∞ . (i) There is ̺ > 0 such that for all
W (Th )
h ∈ H ∩ (0, ̺ℓ0 ], the approximate bilinear form aQ satisfies an inf-sup condition on Vh ×Vh with
constant αQ := 12 α0 . (ii) The discrete problem (33.21) is well-posed.
Proof. We only need to establish the item (i) since the item (ii) follows from (i). Let vh ∈ Vh .
Owing to (33.19), we have
|aQ (vh , wh )| |a(vh , wh )| |(a − aQ )(vh , wh )|
sup ≥ sup − sup
wh ∈Vh k∇wh kL2 (D) wh ∈Vh k∇wh kL2 (D) wh ∈Vh k∇wh kL2 (D)
|(a − aQ )(vh , wh )|
≥ α0 k∇vh kL2 (D) − sup .
Recalling that EK (·) denotes the quadrature error, we have
X
(a − aQ )(vh , wh ) = EK (d∇vh )·∇wh .
K∈Th
For all i, j ∈ {1:d}, let us set p := ∂i vh ∂j wh . Since ∂i′ vbh ◦TK ∈ Pl−1,d and JK is constant
P
over K b (recall that the mesh is affine) we have ∂i vh ◦TK = −T
bh ) ∈ Pl−1,d . A
i′ ∈{1: d} JK,ii′ (∂i′ v
similar argument shows that ∂j wh ◦TK ∈ Pl−1,d . This proves that p ◦ TK ∈ P2l−2,d . We now
use Lemma 30.10 with φ := dij and p := ∂i vh ∂j wh . The assumptions of the lemma are met
with m := 1 and n := 2l − 2 (so that n + m − 1 = 2l − 2 ≤ kQ ). Since k∂i vh ∂j wh kL1 (K) ≤
k∂i vh kL2 (K) k∂j wh kL2 (K) , we infer that there is cQ such that for all vh , wh ∈ Vh , all K ∈ Th , and
all h ∈ H,
|EK ((d∇vh )·∇wh | ≤ cQ hK |d|W1,∞ (K) k∇vh kL2 (K) k∇wh kL2 (K) .
(Notice that it is natural that the above estimate depends on |d|W1,∞ (K) because EK ((d∇vh )·∇wh )
is zero if d is constant over K.) Owing to the Cauchy–Schwarz inequality, we infer that
|(a − aQ )(vh , wh )| ≤ cQ h|d|W1,∞ (Th ) k∇vh kL2 (D) k∇wh kL2 (D) .
Taking ̺ := 1
2cQ
α0
and assuming that h ∈ (0, ̺ℓ0 ] with ℓ0 := |d| 1,∞ implies that cQ h|d|W1,∞ (Th ) ≤
W (Th )
α0 |aQ (vh ,wh )| α0
2 . Combining the above estimates then yields supwh ∈Vh k∇wh kL2 (D) ≥ 2 k∇vh kL (D) ,
2 which is
the expected bound.
33.3.3 Error analysis with quadratures

Theorem 33.17 (Error estimate). Assume Pk,d ⊂ Pb ⊂ Pl,d with the integers l ≥ k ≥ 1.
Assume that kQ ≥ l + k − 2, d ∈ Wk,∞ (Th ), and f ∈ W k,∞ (Th ). Assume that (33.21) is well-
posed and let αQ denote the inf-sup constant of aQ on Vh ×Vh (see Lemma 33.16). Assume that
u ∈ H k+1 (D). There is c s.t. for all h ∈ H ∩ (0, ̺ℓ0 ],

Q (CQ (d, u) + CQ (f )) ,
|u − uh |H 1 (D) ≤ c hk |u|H k+1 (D) + α−1 (33.22)
where we have set

X
CQ (d, u) := |d|Wk−m,∞ (Th ) |u|H m+1 (D) ,
m∈{0: k}
1
CQ (f ) := |D| 2 max(ℓD |f |W k,∞ (Th ) , |f |W k−1,∞ (Th ) ),
and ℓD is a length scale associated with D, e.g., ℓD := diam(D).

Proof. Since the discrete problem with quadratures is stable, we bound the error using Strang’s
g,av g,av
first lemma (Lemma 27.12). We consider vh := Ih0 (u), where Ih0 is the quasi-interpolation
operator introduced in §22.4.2. Recalling that k·kV := k∇(·)kL2 (D) , this yields
g,av
k∇(u − uh )kL2 (D) ≤ c k∇(u − Ih0 (u))kL2 (D) + α−1 st1
Q kδh (vh )kVh ,
′
where the consistency error δhst1 (vh ) ∈ Vh′ is such that
hδhst1 (vh ), wh iVh′ ,Vh := (a − aQ )(vh , wh ) − (ℓ − ℓQ )(wh ).

P
(1) Bound on (a − aQ ). Since (aQ − a)(vh , wh ) = K∈Th EK (d∇vh )·∇wh , we can use the
bound (30.6) from Lemma 30.10 with φ := dij ∂i vh , p := ∂j wh , m := k, and n := l − 1 (so that
n + m − 1 = l + k − 2 ≤ kQ ) for all i, j ∈ {1:d} to infer that
X
|EK (d∇vh )·∇wh | ≤ c hkK |dij ∂i vh |W k,∞ (K) k∂j wh kL1 (K) .
i,j∈{1: d}
Combining the Leibniz product rule with the inverse inequality (12.3) (with p := ∞ and r := 2)
leads to X
|dij ∂i vh |W k,∞ (K) ≤ c |dij |W k−m,∞ (K) |K|− 2 |∂i vh |H m (K) .
1
m∈{0: k}
Applying again an inverse inequality, we infer that

X
|EK (d∇vh )·∇wh | ≤ c hkK |d|Wk−m,∞ (K) k∇vh kH m (K) k∇wh kL2 (K) .
m∈{0: k}
As a result, we have
X
k(a − aQ )(vh , wh )kVh′ ≤ c hk |d|Wk−m,∞ (Th ) |u|H m+1 (D) ,
m∈{0: k}
g,av
where we used the estimate |vh |H m (Th ) = |Ih0 (u)|H m (Th ) ≤ c|u|H m (D) which follows from Theo-
rem 22.14. P
(2) Bound on (ℓ − ℓQ ). We have (ℓQ − ℓ)(wh ) = K∈Th EK (f wh ). We cannot apply the
bound (30.6) from Lemma 30.10 with φ := f , p := wh , m := k, and n := l since n+m−1 = l+k −1
may be larger than kQ . Instead, we use the bound (30.7) with m := k and n := l (since
n + m − 2 = l + k − 2 ≤ kQ ) yielding
X
|(ℓQ − ℓ)(wh )| ≤ c hkK (|f |W k,∞ (K) kwh kL1 (K) + |f |W k−1,∞ (K) k∇wh kL1 (K) )
K∈Th
1
−1
≤ c hk max(Cps ℓD |f |W k,∞ (Th ) , |f |W k−1,∞ (Th ) )|D| 2 k∇wh kL2 (D) ,
where Cps is the global Poincaré–Steklov constant from (3.11). The rest of the proof follows
readily.
Remark 33.18 (Literature). The above analysis is inspired from Ciarlet [124, §4.1], Ciarlet and
Raviart [126], Dautray and Lions [154, §XII.5]. It is possible to refine the analysis by assuming
f ∈ W k,q (Th ) with q > kd and q ≥ 2. The analysis with approximate Neumann conditions can be
done by assuming that surface quadratures of order at least k + l − 1 are used to approximate the
boundary integrals; see [154, §XII.5].
Exercises
Exercise 33.1 (Regularity assumption). Let uh solve (33.5). Assume that u ∈ H 1+r (D)
P 1
with r ∈ (0, k]. Prove that ku − uh kH 1 (D) ≤ c(hr |u|H 1+r (D) + ( F ∈F ∂ h−1 2
F kg − gh kL2 (F ) ) ).
2
g,av P ∂
h
(Hint : consider vh := Ih0 (u) + a∈A∂ σa (g)ϕa , and follow the proof of Theorem 22.14 to bound
h
ku − vh kH 1 (D) .)
Exercise 33.2 (Non-homogeneous Dirichlet). Let A denote the system matrix in (33.10).
Let R ∈ RI and let k ≥ 1. Consider the Krylov space Sk := span{R, AR, . . . , Ak−1 R}. For all
V ∈ RI , write V := (V◦ , V∂ )T . Assume that R∂ = 0. (i) Prove that Y∂ = 0 for all Y ∈ Sk . (ii)
Prove that if A◦◦ is symmetric, the restriction of A to Sk is symmetric.
Exercise 33.3P (DMP). Assume that the stiffness matrix is a Z-matrix. Assume P the following:
(i) Aii ≥ − j6=i Aij for all i ∈ {1:I}; (ii) ∃i∗ ∈ {1:I} such that Ai∗ i∗ > − j6=i∗ Ai∗ j ; (iii)
For all i ∈ {1:I}, i 6= i∗ , there exists a path [i =: i1 , . . . , iJ := i∗ ] such that Aij ij+1 < 0 for all
j ∈ {1:J−1}. Prove that A is a nonsingular M -matrix. (Hint : let B ≤ 0, let U := A−1 B, and
proceeding by contradiction, assume that there is i ∈ {1:I} s.t. Ui = maxj∈{1: I} Uj > 0.)
Exercise 33.4 (Obtuse mesh). The mesh shown in Figure 33.1 contains three interior nodes
with coordinates z1 := (1, 1), z2 := (3, 1), and z3 := (2, 32 ). The sum of the two angles opposite the
edge linking z1 and z2 is larger than π. (i) Assemble the 3×3 stiffness matrix A generated by the
three shape functions associated with the three interior nodes z1 , z2 , z3 . Is A a Z-matrix? (Hint :
the local stiffness matrix is translation- and scale-invariant, there are four shapes of triangles in
the mesh, and one can work on triangles with vertices ((0, 0), (1, 0), (0, 1)), ((0, 0), (1, 0), (0, 12 )),
((−1, 0), (1, 0), (0, 21 )), and ((−1, 1), (1, 1), (0, 1)).) (ii) Compute A−1 . Is A an M -matrix?
Exercise 33.5 (1D DMP). Consider the equation µu + βu′ − νu′′ = f in D := (0, 1). Let Th be
1
the uniform mesh composed of the cells [ih, (i+1)h], ∀i ∈ {0:I}, with uniform meshsize h := I+1 .
1
P g
Assume µ ∈ R+ , β ∈ R, ν ∈ R+ and f ∈ L (D). Let uh := i∈{0: I+1} Ui ϕRi ∈ P1 (Th )R be such
R R
that D ((µuh + βu′h )ϕi + νu′h ϕ′i ) dx = D f ϕi dx for all i ∈ {1:I}. Let Fi := D f ϕi dx/ D ϕi dx.
Assume that hν ≥ |β| µh Fi Fi
2 + 6 . (i) Show that min(Ui−1 , Ui+1 , µ ) ≤ Ui ≤ max(Ui−1 , Ui+1 , µ ) for all
i ∈ {1:I}. (Hint : write the linear system as µUi + αi−1 (µ, β, ν)(Ui − Ui−1 ) + αi+1 (µ, β, ν)(Ui −
Figure 33.1: Illustration for Exercise 33.4.
minj∈{1: I} Fj maxj∈{1: I} Fj
Ui+1 ) = Fi .) (ii) Show that min(U0 , UI+1 , µ ) ≤ Ui ≤ max(U0 , UI+1 , µ ) for
all i ∈ {1:I}.
Exercise 33.6 (1D DMP, pure diffusion). Let D := (0, 1), f ∈ L∞ (D), and a nonuniform R ′ ′ mesh
g
T
R h of D with nodes {xi } i∈{0: I+1} . Let u h ∈ P 1 (T h ) be s.t. u h (0) = a, u h (1) = b, and D uh vh dx =
g 1
D f vh dx for all vh ∈ P (T
1,0 h ). (i) Show that max x∈D h u (x) ≤ max(a, b) + 4 ess sup x∈D f (x).
g x 1−x
(Hint : test with φh ∈ P1,0 (Th ) s.t. φh|[0,xi ] := xi and φh|[xi ,1]
:= 1−xi for all i ∈ {1:I}.) (ii) Let
φh be the function defined in the hint. Compute −∂xx φh . Comment on the result.
Exercise 33.7 (Maximum principle). Let D be a bounded Lipschitz domain in Rd . Let x0 ∈ D
1
and R ∈ R be s.t. maxx∈D kx − x0 kℓ2 ≤ R. (i) Let φ(x) := − 2d kx − x0 k2ℓ2 . Compute −∆φ. Give
an upper bound on maxx∈D φ(x) and a lower bound on minx∈∂D φ(x). (ii) Let f ∈ L∞ (D) and
let u ∈ H 1 (D) solve −∆u = f . Let M := ess supx∈D f (x). Give an upper bound on −∆(u − M φ).
2
(iii) Prove that maxx∈D u(x) ≤ maxx∈∂D u(x) + M+ R 2d with M+
:= max(M, 0). (Hint : use (i)
from Theorem 33.6.)
Chapter 34
A posteriori error analysis
An a posteriori error estimate is an upper bound on the approximation error that can be computed
by using only the discrete solution and the problem data. Such an estimate can serve the twofold
purpose of judging the quality of the discrete solution and of guiding an adaptive procedure that
modifies the discretization iteratively in order to diminish the approximation error. A posteriori
error estimates should involve constants that are all computable, or sharp estimates from above
of these constants. For the purpose of mesh adaptation, the error estimate should be a sum
of local contributions (usually called indicators) that can be used to mark those cells requiring
further refinement at the next iteration of the adaptive procedure. It is then important that the
indicators represent a local lower bound on the error. A posteriori and a priori error estimates
are conceptually different. A priori error estimates rely on the stability of the discrete problem
to provide decay estimates of the error that depend on high-order Sobolev norms of the exact
solution which are inaccessible to computation. A posteriori error estimates rely on the stability
of the continuous problem and provide computable upper bounds on the error.
34.1 The residual and its dual norm

A key notion in a posteriori error analysis is the residual and its dual norm.
34.1.1 Model problem and residual

For simplicity, we focus on the purely diffusive version of the model problem (32.1) with homoge-
1
neous Dirichlet boundary conditions. We denote by Ru the unique function in V := R H0 (D) such
that a(u, w) = ℓ(w) for all w ∈ V, where a(v, w) := D (d∇v)·∇w dx and ℓ(w) := D f w dx with
f ∈ L2 (D). As in §32.1, we assume that d is defined on D with values in Rd×d and that d(x) is
symmetric with all its eigenvalues in the interval [λ♭ , λ♯ ] for a.e. x ∈ D, where 0 < λ♭ ≤ λ♯ < ∞.
g
Let (Th )h∈H be a shape-regular family of matching simplicial meshes of D, let Vh := Pk,0 (Th )
be the H01 (D)-conforming finite element space of some degree k ≥ 1, and let uh ∈ Vh be the
corresponding approximate solution such that a(uh , wh ) = ℓ(wh ) for all wh ∈ Vh (see §32.1). Let
V ′ := L(V ; R) be the real space of bounded linear forms acting on V. In the present setting, we
have V ′ := L(H01 (D); R) = H −1 (D); see Definition 4.10. We denote the action of an element
of V ′ on a function in V by using brackets. We equip the space V with the H 1 -seminorm, i.e.,
118 Chapter 34. A posteriori error analysis
kvkV := k∇vkL2 (D) = |v|H 1 (D) . The Poincaré–Steklov inequality implies that this seminorm is
indeed a norm on V ; see Lemma 3.27 (with p := 2).
Definition 34.1 (Residual). The residual of the discrete solution uh ∈ Vh is the element ρ(uh ) ∈
V ′ := H −1 (D) acting as follows:
hρ(uh ), ϕi := ℓ(ϕ) − a(uh , ϕ), ∀ϕ ∈ V := H01 (D). (34.1)
The boundedness of the bilinear form a together with the assumption f ∈ L2 (D) implies
that ρ(uh ) is bounded on V. Using the embedding L2 (D) ֒→ H −1 (D), (34.1) is equivalent to
ρ(uh ) := f + ∇·(d∇uh ) ∈ H −1 (D). Moreover, since u satisfies a(u, ϕ) = ℓ(ϕ) for all ϕ ∈ H01 (D),
we infer that
hρ(uh ), ϕi := a(u − uh , ϕ), ∀ϕ ∈ H01 (D), (34.2)
which is equivalent to saying that ρ(uh ) := ∇·(d∇(uh − u)) ∈ H −1 (D).
Remark 34.2 (Extensions). We refer the reader to Verfürth [378, §4.3-4.4] for other boundary
conditions and lower-order terms in the PDE, to Verfürth [377] for the analysis of singularly
perturbed regimes and a precise tracking of the model parameters in the error constants, to Ciarlet
and Vohralı́k [122] for sign-changing diffusion coefficients, and to Cohen et al. [138] for a source
term in H −1 (D).
34.1.2 The residual dual norm and the error

The dual space V ′ := H −1 (D) is equipped with the norm
|hη, ϕi|
kηkH −1 (D) := sup , (34.3)
ϕ∈H01 (D) k∇ϕkL2 (D)
for all η ∈ H −1 (D). Our first important observation is that the dual norm of the residual is closely
related to the H 1 -seminorm of the error.
Lemma 34.3 (Error and residual). Let α and M denote, respectively, the stability and bound-
edness constants of the bilinear form a with respect to the H 1 -seminorm. The following holds
true:
1 1
kρ(uh )kH −1 (D) ≤ k∇(u − uh )kL2 (D) ≤ kρ(uh )kH −1 (D) . (34.4)
M α
|a(u−uh ,ϕ)|
Proof. Owing to (34.2), we have kρ(uh )kH −1 (D) = supϕ∈H01 (D) k∇ϕkL2 (D) , and the stability and
boundedness of a imply that
|a(u − uh , ϕ)|
αk∇(u − uh )kL2 (D) ≤ sup ≤ M k∇u − uh kL2 (D) .
ϕ∈H01 (D) k∇ϕkL2 (D)
Remark 34.4 (Stability). The coercivity of a is not needed to prove (34.4), it is just the inf-sup
stability that is invoked (see the BNB theorem).
Lemma 34.3 is fundamental since it provides two-sided bounds on the approximation error in
terms of the residual. These two bounds are computable since they only depend on uh and on the
problem data (f and d). Considering the H −1 (D)-norm is a key ingredient in the argumentation
since it allows us to get rid of u by replacing a(u, ϕ) by ℓ(ϕ) for all ϕ ∈ H01 (D). The price to pay
for this replacement is that the H −1 (D)-norm is not computable since it invokes the supremum
over an infinite-dimensional space. We will circumvent this difficulty in §34.2.
34.1.3 Localization of dual norms

An objection that is often raised with dual norms is that they are not localizable. While this is
generally true for arbitrary elements of H −1 (D), localization is actually possible when the linear
form vanishes on a set of functions with local support forming a partition of unity. Let us first
observe that we can define the restriction of η ∈ H −1 (D) to an open Lipschitz subset U ⊂ D to
be the bounded linear form η|U ∈ H −1 (U ) such that hη|U , ψiU := hη, EU (ψ)i for all ψ ∈ H01 (U ),
where EU (ψ) ∈ H01 (D) denotes the zero-extension of ψ to D and h·, ·iU the duality pairing between
H −1 (U ) and H01 (U ). Note that kη|U kH −1 (U) ≤ kηkH −1 (D) . We abuse the notation by just writing
η ∈ H −1 (U ) when the context is unambiguous.
Consider the vertices z ∈ Vh of the mesh Th and the global shape functions {ψz }z∈Vh associated
with the P1 Lagrange finite elements (called hat or Courant basis functions); S see §19.2.1. For all
z ∈ Vh , let Tz be the collection of the mesh cells sharing z and let Dz := int( K∈Tz K). The set
Dz is called finite element star and its diameter is denoted by hDz . Recall that ψz is supported
D
Pz (see the left panel of Figure 21.1). The hat basis functions P form a partition of unity since
z∈Vh ψ z = 1 in D. We even have a local partition of unity since z∈VK ψz|K = 1 for all K ∈ Th ,
where VK is the collection of the (d+1) vertices of K. To handle homogeneous Dirichlet conditions,
we define the subset Vh◦ ⊂ Vh composed of the interior vertices (i.e., not lying on ∂D) and the
subset Vh∂ := Vh \Vh◦ composed of the boundary vertices.
Definition
R 34.5 (Poincaré–Steklov constant CPS,z ). For all z ∈ Vh◦ , let H∗1 (Dz ) := {v ∈
H (Dz ) | Dz v dx = 0}, and for all z ∈ Vh , let H∗1 (Dz ) := {v ∈ H 1 (Dz ) | v|∂Dz ∩∂D = 0}. We
1 ∂
define
kvkL2 (Dz )
CPS,z := h−1
Dz sup . (34.5)
1
v∈H∗ (Dz ) k∇vk L2 (Dz )
Remark 34.6 (CPS,z ). The constant CPS,z in Definition 34.5 is uniformly bounded on shape-
regular mesh sequences. For z ∈ Vh◦ , one has CPS,z ≤ π −1 if Dz is convex; see (12.13). Sharp
estimates in the nonconvex case can be found in Eymard et al. [197], Repin [334], Veeser and
Verfürth [372], Šebestová and Vejchodský [346]; see also Exercise 22.3. For z ∈ Vh∂ , one has
CPS,z ≤ 1 if there is a vector t ∈ Rd such that for a.e. x ∈ Dz , the straight line drawn from x
in the direction of t first hits ∂Dz at a point in ∂D; see Vohralı́k [381]. We refer the reader to
[381, 334] for the general case.
Proposition 34.7 (Localization). Let η ∈ H −1 (D). (i) We have
X
kηk2H −1 (Dz ) ≤ (d + 1)kηk2H −1 (D) . (34.6)
z∈Vh
(ii) If η does not have low-frequency components, i.e., if the following holds:
hη, ψz i = 0, ∀z ∈ Vh◦ , (34.7)
then letting ČPS := maxz∈Vh (1 + hDz k∇ψz kL∞ (Dz ) CPS,z ), we have
X
kηk2H −1 (D) ≤ (d + 1)ČPS
2
kηk2H −1 (Dz ) . (34.8)
z∈Vh
Proof. For all z ∈ Vh , let vz ∈ H01 (Dz ) be the Riesz–Fréchet representative P of η|Dz . Then
kηk2H −1 (Dz ) = hη, vz iDz = hη, EDz (vz )i = k∇vz k2L2 (Dz ) . Let us set v := z∈Vh EDz (vz ). Since
v ∈ H01 (D), we infer that
X X
kηk2H −1 (Dz ) = hη, EDz (vz )i = hη, vi ≤ kηkH −1 (D) k∇vkL2 (D) .
z∈Vh z∈Vh
Using the Cauchy–Schwarz inequality and rearranging the sums leads to

2
X X
X X
k∇vk2L2 (D) = ∇vz ≤ (d + 1) k∇vz k2L2 (K)

K∈Th z∈VK L2 (K) K∈Th z∈VK
X X
2
= (d + 1) k∇vz kL2 (Dz ) = (d + 1) kηk2H −1 (Dz ) .
z∈Vh z∈Vh
Combining the above bounds yields (34.6). Let us prove (34.8), i.e., we assume now that hη, ψz i = 0
for all z ∈ Vh◦ , i.e., that (34.7) holds true. Let ϕ ∈ H01 (D). Since ψz ϕ ∈ H01 (Dz ), the partition of
unity implies that
X X X
hη, ϕi = hη, ψz ϕiDz = hη, ψz (ϕ − ϕz )iDz + hη, ψz ϕiDz , (34.9)
z∈Vh ◦
z∈Vh ∂
z∈Vh
1
R
with ϕz := |Dz | Dz
ϕ dx, since hη, ψz ϕz i = ϕz hη, ψz i = 0 for all z ∈ Vh◦ owing to (34.7). We have
k∇(ψz (ϕ − ϕz ))kL2 (Dz ) ≤ (1 + hDz k∇ψz kL∞ (Dz ) CPS,z )k∇ϕkL2 (Dz )
≤ ČPS k∇ϕkL2 (Dz ) ,
where we used that ∇(ψz (ϕ − ϕz )) = ψz ∇ϕ + (ϕ − ϕ z )∇ψz , the triangle inequality, kψz kL∞ (Dz ) =
1, and the definitions of CPS,z and ČPS . Proceeding similarly, we infer the same bound on
k∇(ψz ϕ)kL2 (Dz ) for all z ∈ Vh∂ . Using the Cauchy–Schwarz inequality, we infer that
! 21 ! 12
X X
|hη, ϕi| ≤ kηk2H −1 (Dz ) ČPS k∇ϕk2L2 (Dz ) .
z∈Vh z∈Vh
P
This gives (34.8) since z∈Vh k∇ϕk2L2 (Dz ) = (d + 1)k∇ϕk2L2 (D) .
The bound (34.8) means that we can consider local test functions in {H01 (Dz )}z∈Vh to explore
the action on the whole space H01 (D) of a linear form η ∈ H −1 (D) satisfying (34.7). Note that
the residual ρ(uh ) satisfies (34.7): this is the Galerkin orthogonality property for the hat basis
functions.
Remark 34.8 (Value of ČPS ). Using an inverse inequality to bound k∇ψz kL∞ (Dz ) , we can see
that the constant ČPS from Lemma 34.7 is uniformly bounded on shape-regular mesh sequences.
See also Remark 34.6.
Remark 34.9 (Literature). The proof of Proposition 34.7 is inspired by Carstensen and Funken
[108]; see also Babuška and Miller [36] (for the idea of working on finite element stars), Cohen
et al. [138], Ciarlet and Vohralı́k [122], Blechta et al. [59].
34.2 Global upper bound

We derive a computable upper bound on the error by using Lemma 34.3. Inspired by Nochetto and
Veeser [315], Veeser and Verfürth [372], we achieve this by relying on two key ideas: (i) the residual
is the sum of an L2 -function and a measure supported in the mesh interfaces; (ii) localization is
achieved by exploiting that the residual vanishes on the hat basis functions in the same spirit as
Proposition 34.7. The error upper bound derived herein belongs to the class of residual-based a
posteriori estimates pioneered by Babuška and Rheinbolt [39]. Another class of a posteriori error
estimates based on local flux equilibration in finite element stars is discussed in Chapter 52.
Let us first observe that the residual ρ(uh ) ∈ H −1 (D) admits the following representation (see
Exercise 34.1): For all ϕ ∈ H01 (D),
X Z X Z
v
hρ(uh ), ϕi = r (uh )ϕ dx + rs (uh )ϕ ds, (34.10)
K∈Th K F ∈Fh◦ F
with densities rv (uh ) ∈ L2 (D) and rs (uh ) ∈ L∞ (Fh◦ ) defined by

rv (uh )|K := f|K + (∇·(d∇uh ))|K , ∀K ∈ Th , (34.11a)
rs (uh )|F := [[d∇uh ]]F ·nF , ∀F ∈ Fh◦ , (34.11b)
and [[·]]F denotes the jump across F using the orientation of the unit normal nF (see Definitions 8.10
and 18.2).
Definition 34.10 (Trace inequality constant). Let z ∈ Vh , let Fz◦ be the collection of the
interfaces sharing z, and let H∗1 (Dz ) be defined as in Definition 34.5. Then we set
−1 kvkL2 (Fz◦ )
Ctr,z := hDz2 sup , (34.12)
v∈H∗1 (Dz ) k∇vkL2 (Dz )
P 1
with the notation kvkL2 (Fz◦ ) := ( F ∈Fz◦ kvk2L2 (F ) ) 2 .
Remark 34.11 (Ctr,z ). The constant Ctr,z in Definition 34.10 is uniformly bounded on shape-
regular mesh sequences; see Exercise 34.2.
Theorem 34.12 (Upper bound). Define the vertex-based error indicators
1 1 1
ηzv (uh ) := hDz kψz2 rv (uh )kL2 (Dz ) , ηzs (uh ) := hD 2
z
kψz2 rs (uh )kL2 (Fz◦ ) ,
1
ηz (uh ) := (d + 1) 2 CPS,z ηzv (uh ) + Ctr,z ηzs (uh ) , (34.13)
with CPS,z defined in (34.5) and Ctr,z defined in (34.12). The following global a posteriori estimate
holds true: ! 21
X
2
αk∇(u − uh )kL2 (D) ≤ ηz (uh ) . (34.14)
z∈Vh
Proof. Our starting point is the error upper bound from Lemma 34.3, i.e., αk∇(u − uh )kL2 (D) ≤
|hρ(uh ),ϕi| P
supϕ∈H01 (D) k∇ϕk . Using (34.9), we infer that hρ(uh ), ϕi = z∈V ◦ hρ(uh ), ψz (ϕ − ϕz )iDz +
P L 2 (D) h
z∈Vh∂ hρ(u h ), ψz ϕi D z , where ϕz

is the mean value of ϕ over D z . Consider z ∈ Vh◦ . Exploiting
the representation (34.10) and since ψz is supported in Dz , we infer that
X Z XZ
hρ(uh ), ψz (ϕ − ϕz )iDz = rv (uh )ψz (ϕ − ϕz ) dx + rs (uh )ψz (ϕ − ϕz ) ds.
K∈Tz K F ∈Fz◦ F
Let T1 , T2 denote the two terms on the right-hand side. Using the Cauchy–Schwarz inequality,
kψz kL∞ (Dz ) = 1, Definition 34.5, and ∇ϕz = 0 yields
1
|T1 | ≤ kψz rv (uh )kL2 (Dz ) kϕ − ϕz kL2 (Dz ) ≤ CPS,z hDz kψz2 rv (uh )kL2 (Dz ) k∇ϕkL2 (Dz ) .
Proceeding similarly and invoking Definition 34.10 leads to

1 1
|T2 | ≤ Ctr,z hD
2
z
kψz2 rs (uh )kL2 (Fz◦ ) k∇ϕkL2 (Dz ) .
Similar boundsPare verified for z ∈ Vh∂ . We conclude the proof by using the Cauchy–Schwarz
inequality and z∈Vh k∇ϕk2L2 (Dz ) = (d + 1)k∇ϕk2L2 (D) .
Remark 34.13 (Variants). The factor (d + 1) in ηz (uh ) can be avoided by invoking Poincaré–
Steklov inequalities based on norms weighted by the hat basis functions (see Veeser and Verfürth
1
[372]). The weights ψz2 in ηzv (uh ) and ηzs (uh ) are not essential, but they will help to deduce Corol-
lary 34.14. Another route to derive an upper bound similar to (34.14) consists of combining the
upper bound from Lemma 34.3 with the localization property (34.8) and the residual represen-
tation (34.10). The above proof of Theorem 34.12 is slightly more direct and therefore leads to
somewhat sharper values for the constants weighting the error indicators.
Since adaptive procedures usually mark cells rather than vertices (see Morin et al. [306] for
an example of vertex-based marking), we reformulate the global upper bound (34.14) in terms of
cell-based error indicators.
Corollary 34.14 (Cell-based upper bound). The following holds true:
X 21
v
αk∇(u − uh )kL2 (D) ≤ CGUB ηK (uh )2 + s
ηK (uh )2 , (34.15)
K∈Th
1 1 1
with the constant CGUB := (d + 1) 2 maxz∈Vh (2 2 CPS,z ϑz , Ctr,z ̺z2 ), the geometric factors ϑz :=
h h
maxK∈Tz hDKz and ̺z := maxF ∈Fz◦ hDFz , and the cell-based error indicators
1
v
ηK (uh ) := hK krv (uh )kL2 (K) , s
ηK (uh ) := hK
2
krs (uh )kL2 (FK◦ ) , (34.16)
P 1
◦
where kvkL2 (FK◦ ) := ( F ∈F ◦ kvk2L2 (F ) ) 2 and FK is the collection of the faces of K that are inter-
K
faces.
Proof. Since (a + b)2 ≤ 2(a2 + b2 ) for all a, b ∈ R, recalling (34.13) we have
X X 1
ηz (uh )2 ≤ 2
2(d + 1)CPS,z h2Dz kψz2 rv (uh )k2L2 (Dz )
z∈Vh z∈Vh
X 1
2
+ 2(d + 1)Ctr,z hDz kψz2 rs (uh )k2L2 (Fz◦ ) .
z∈Vh
Let T1 , T2 denote the two terms on the right-hand side. We have

X X 1
2
T1 ≤ 2(d + 1) max (CPS,z ϑ2z ) h2K kψz2 rv (uh )k2L2 (K)
z∈Vh
z∈Vh K∈Tz
X
2
= 2(d + 1) max (CPS,z ϑ2z ) h2K krv (uh )k2L2 (K) ,
z∈Vh
K∈Th
h S
where we used ϑz := maxK∈Tz hDKz and Dz := int( K∈Tz K) in the first line and where the identity
in the second line follows by exchanging the two summations and using that the restrictions to
any mesh cell of the hat basis functions form a partition of unity. The reasoning for T2 is similar
since the restrictions to any interface of the hat basis functions also form a partition of unity. This
leads to X
2
T2 ≤ (d + 1) max (Ctr,z ̺z ) 2hF krs (uh )k2L2 (F ) .
z∈Vh
F ∈Fh◦
Finally, we remove the factor 2 by introducing a cell-based summation and observing that every
interface F := ∂Kl ∩ ∂Kr ∈ Fh◦ is shared by two mesh cells and that hF ≤ hK for all K ∈
{Kl , Kr }.
Remark 34.15 (CGUB ). The constant CGUB in Corollary 34.14 is uniformly bounded on shape-
regular mesh sequences.
Remark 34.16 (Dual-weighted residual estimate). Let ψ ∈ H −1 (D) be some linear output
functional. Let zψ ∈ H01 (D) solve the dual problem a(ϕ, zψ ) := hψ, ϕi for all ϕ ∈ H01 (D). Then
hψ, u − uh i = a(u − uh , zψ ) = hρ(uh ), zψ i, i.e., the output error hψ, u − uh i is equal to the residual
tested against the dual solution zψ . For instance, we have ku − uh k2L2 (D) = hρ(uh ), zψ i with
zψ ∈ H01 (D) s.t. a(ϕ, zψ ) := (u − uh , ϕ)L2 (D) for all ϕ ∈ H01 (D). See Becker and Rannacher [48]
for further insight on the whole approach. Notice though that one must approximate zψ to obtain
a computable estimate.
34.3 Local lower bound

v s
Our goal in this section is to bound the cell-based error indicators ηK (uh ) and ηK (uh ) defined
in (34.16) by the approximation error in the mesh cell K (and some of its neighbors) for all
K ∈ Th . This will give lower bounds on the approximation error. These lower bounds differ from
the upper bounds on two aspects: they are local (recall that the upper bounds are global), and
they involve generic constants whose value may depend on the regularity of the mesh sequence and
the polynomial degree k. The symbol c denotes these generic constants (the value can change at
each occurrence).
To put the upcoming results into perspective, we observe that for every subset U of D, defining
MU := kdkL∞ (U;Rd×d ) , the bound
kρ(uh )kH −1 (U) ≤ MU k∇(u − uh )kL2 (U) (34.17)

R
follows from hρ(uh ), ϕiU = U (d∇(u − uh ))·∇ϕ dx for all ϕ ∈ H01 (U ) and the Cauchy–Schwarz
inequality. After observing that rv (uh )|K = ρ(uh )|K (where we used the (slightly abusive) notation
on the restriction to H −1 (K) of the residual), what we now need is a bound of the form
krv (uh )kL2 (K) ≤ c h−1 v

K kr (uh )kH −1 (K) , ∀K ∈ Th . (34.18)
If (34.18) were indeed true, we would immediately deduce that

v
ηK (uh ) := hK krv (uh )kL2 (K) ≤ c krv (uh )kH −1 (K) ≤ c MK k∇(u − uh )kL2 (K) .
Unfortunately, (34.18) is an inverse-like inequality (where one norm is that of a dual space). Hence,
it cannot be valid for every function in L2 (K) (recall that rv (uh ) depends on the source term f
and the tensor d). We refer the reader to [315, Pbm. 32] for a concrete example.
To address this problem, we introduce the notion of oscillation. For all K ∈ Th and every integer
lv ≥ 0, we use the notation Plv (K) := Plv ,d ◦ TK−1 b → K is the geometric mapping.
, where TK : K
−1
Similarly, for all F ∈ Fh◦ and every integer ls ≥ 0, we use the notation Pls (F ) := Pls ,d−1 ◦ TK,F ,
where K is a cell having F as face, TK,F := TK|Fb ◦ TFb : Fb d−1
→ F , Fb := TK (F ), Fb
−1 d−1
is the
reference simplex in R , and TFb : Fb
d−1 d−1
→ Fb is an affine bijective mapping (recall that Pls (F )
is independent of the choice of K); see §20.2.
Definition 34.17 (Oscillation). Let lv , ls ∈ N, K ∈ Th , and F ∈ Fh◦ . Let r̄v (uh ) be the L2 -
orthogonal projection of rv (uh ) onto Plv (K). Let r̄s (uh ) be the L2 -orthogonal projection of rs (uh )
onto Pls (F ). The oscillation indicators are defined by
φvK (uh , f, d) := hK krv (uh ) − r̄v (uh )kL2 (K) , (34.19a)
φsF (uh , f, d) := hF krs (uh ) − r̄s (uh )kL2 (F ) .
1
2
(34.19b)
v 2 −1
Lemma 34.18 (Verfürth’s
R inverse inequalities). (i) Let : L (K) → H (K) be defined
TK
v
by hTK (r), ϕiK := K rϕ dx for all ϕ ∈ H01 (K) and all r ∈ L (K). There is c, depending on lv ,
2
such that for all K ∈ Th , all h ∈ H, and all q ∈ Plv (K),

kqkL2 (K) ≤ c h−1 v
K kTK (q)kH −1 (K) . (34.20)
s 2 −1 s
R 1
(ii) Let TF : L (F ) → H (DF ) be defined by hTF (r), ϕiDF := F rϕ ds for all ϕ ∈ H0 (DF ) and
all r ∈ L2 (F ), where DF := int(Kl ∪ Kr ) with F := ∂Kl ∩ ∂Kr . There is c, depending on ls , such
that for all F ∈ Fh◦ , all h ∈ H, and all g ∈ Pls (F ),
−1
kgkL2 (F ) ≤ c hF 2 kTFs (g)kH −1 (DF ) . (34.21)
Proof. The proof hinges on the use of suitable cell- and face-based bubble functions introduced
by Verfürth [378, §3.6]. These functions vanish on the boundary of K and DF , respectively.
(1) Proof of (34.20). Let K ∈ Th . The cell-based bubble function bvK := (d + 1)d+1 λK K
0 . . . λd ,
K 2 v 12 2
where {λi }i∈{0: d} are the barycentric coordinates in K, is such that kqkL2 (K) ≤ c1 k(bK ) qkL2 (K)
and k∇(bvK q)kL2 (K) ≤ c2 h−1
K kqkL2 (K) for all q ∈ Plv (K) (both inequalities are established on the
reference element by invoking norm equivalence in polynomial spaces and then transferred back to
K by the geometric mapping TK ). Noticing that bvK q ∈ H01 (K), we infer that
1
kqk2L2 (K) ≤ c1 k(bvK ) 2 qk2L2 (K) = c1 hTK
v
(q), bvK qiK
v
≤ c1 kTK (q)kH −1 (K) k∇(bvK q)kL2 (K)
≤ c1 c2 h−1 v
K kTK (q)kH −1 (K) kqkL2 (K) .
(2) Proof of (34.21). Let F := ∂Kl ∩ ∂Kr ∈ Fh◦ , and assume without loss of generality that in both
cells, the vertex opposite to F is associated with the barycentric coordinate λK 0 with K ∈ TF :=
s s := d K K
{Kl , Kr }. The face-based bubble function bF such that bF |K d λ1 . . . λd , for all K ∈ TF , is in
1 −1
H01 (DF ) and is such that kgk2L2 (F ) ≤ c3 k(bsF ) 2 gk2L2 (F ) and k∇(bsF e
g)kL2 (DF ) ≤ c4 hF 2 kgkL2 (F ) for
−1
all g ∈ Pls (F ), where eg is the extension of g to DF defined by ge|K := (((g ◦ (TK )|Fb ) ◦ ΠFb ) ◦ TK ) for
all K ∈ TF , where F b := −1 b
TK (F ) and ΠFb is the orthogonal projection onto F . Note that ge|F = g.
The above bounds are again proved on the reference element by using the pullback by TK . Since
bsF ge ∈ H01 (DF ), we conclude that
1
kgk2L2(F ) ≤ c3 k(bsF ) 2 gk2L2 (F ) = c3 hTFs (g), bsF geiDF
≤ c3 kTFs (g)kH −1 (DF ) k∇(bsF e
g )kL2 (DF )
−1
≤ c3 c4 h̃F 2 kTFs (g)kH −1 (DF ) kgkL2 (F ) .
The operator TK v
: L2 (K) → H −1 (K) is nothing but the natural injection of L2 (K) into
H −1 (K). Observe also that TKv
(rv (uh )) = ρ(uh )|K in H −1 (K). We now establish a local lower
v s
bound on the error using the cell-based indicators ηK (uh ) and ηK (uh ) defined in (34.16).
Theorem 34.19 (Local lower bound). For all K ∈ Th , Slet TKf be the set composed of K and
f :=
those cells sharing an interface with K, and let DK int( K ′ ∈T f K ′ ). There is c such that for
K
all K ∈ Th and all h ∈ H,

f k∇(u − uh )kL2 (D f ) + φT f (uh , f, d)
v s
ηK (uh ) + ηK (uh ) ≤ c MDK K K
, (34.22)
P P
with φTKf (uh , f, d) := K ′ ∈T f φvK ′ (uh , f, d) + F ∈F ◦ φsF (uh , f, d), where φvK ′ and φsF are defined
f := kdkL∞ (D f ;Rd×d ) .
K K
in (34.19), and MDK K
Proof. Let K ∈ Th . Owing to (34.20) and the triangle inequality, we infer that
hK kr̄v (uh )kL2 (K) ≤ c kTK

v v
(r̄ (uh ))kH −1 (K)
v v v v

≤ c kTK (r (uh ))kH −1 (K) + kTK (r (uh ) − r̄v (uh ))kH −1 (K)

≤ c′ kρ(uh )kH −1 (K) + φvK (uh , f, d) ,
v
since TK (rv (uh )) := ρ(uh )|K and kTK v
(r)kH −1 (K) ≤ chK krkL2 (K) for all r ∈ L2 (K) (see Ex-
ercise 34.3). Using the triangle inequality and kρ(uh )kH −1 (K) ≤ MK k∇(u − uh )kL2 (K) owing
to (34.17), we infer that

v
ηK (uh ) := hK krv (uh )kL2 (K) ≤ c MK k∇(u − uh )kL2 (K) + φvK (uh , f, d) .
1
◦
Let F ∈ FK . By using (34.21) and kTFs (g)kH −1 (DF ) ≤ ch̃F2 kgkL2 (F ) for all g ∈ L2 (F ) (see
Exercise 34.3) and proceeding similarly, we infer that

hF2 krs (uh )kL2 (F ) ≤ c kTFs (rs (uh ))kH −1 (DF ) + φsF (uh , f, d) .
1
R
Since hTFs (rs (uh )), ϕi := − DF rv (uh )ϕ dx + hρ(uh ), ϕi and kϕkL2 (DF ) ≤ chF k∇ϕkL2 (DF ) for all
ϕ ∈ H01 (DF ) owing to the Poincaré-Steklov inequality in H01 (DF ) (the constant c is independent
on F and h), we infer that

kTFs (rs (uh ))kH −1 (DF ) ≤ c hF krv (uh )kL2 (DF ) + kρ(uh )kH −1 (DF ) .
1
s
Since ηK (uh ) := hK2
krs (uh )kL2 (FK◦ ) , we conclude by using the regularity of the mesh sequence, the
v
above bound on r (uh ), and (34.17).
Remark 34.20 (Oscillation). The oscillation term somehow pollutes the local lower bound in
Theorem 34.19. As emphasized above, this is the price to pay to have computable error indicators.
It is usually recommended in the literature to take lv := 2k − 2 and ls := 2k − 1 for general d,
where k is the polynomial degree of the finite elements. When the diffusion tensor d is piecewise
constant, taking ls := k − 1 makes the face-based oscillation φsF to vanish (i.e., it is not necessary to
invoke φsF ), and taking ls := k − 1 transforms the cell-based oscillation into a data oscillation since
v
in this case φvK = hK kf −f kL2 (K) . With the above choices for lv and ls , the oscillation is expected
to be of higher-order than the approximation error (see Cascón et al. [113] and Exercise 34.4). On
coarse meshes however the oscillation can be the dominant (or even be the only) contribution to
the approximation error.
34.4 Adaptivity
This section outlines important ideas and results on adaptive mesh refinement driven by a posteriori
error estimates. We do not consider mesh coarsening, even though this is also a practically impor-
tant topic. The analysis of adaptive finite element methods (i.e., finite element solvers employing
adaptive mesh refinement) has witnessed extensive progress over the years. The convergence of
the adaptive procedure and its (quasi-)optimality in terms of error decay rates as a function of the
number of the degrees of freedom is now well understood. Seminal contributions include those in
Dörfler [171], Morin et al. [305], Binev et al. [56], Stevenson [356, 357], Cascón et al. [113]. Com-
prehensive surveys can be found in Nochetto et al. [316], Nochetto and Veeser [315], and Verfürth
[378, p. 264]. An axiomatic presentation with numerous references is proposed in Carstensen et al.
[112].
Algorithm 34.1 Adaptive finite element solver.

Build an initial grid T0 and choose a tolerance TOL
for n = 0, 1, . . . until η(un , Tn ) ≤ TOL do
un ← SOLVE(Tn )
{ηK (un )}K∈Tn ← ESTIMATE(un , Tn )
Mn ← MARK({ηK (un )}K∈Tn , Tn )
Tn+1 ← REFINE(Mn , Tn )
n← n+1
end for
The core of an adaptive finite element solver is outlined in Algorithm 34.1, which generates
a sequence of (matching simplicial) meshes T0 , T1 , . . . (we omit the subscript h to simplify the
notation). The module SOLVE consists of building the finite element space Vn from the current
mesh Tn and solving for the discrete solution un ∈ Vn . The module ESTIMATE computes the cell-
based error estimators {ηK (un )}K∈Tn defined in (34.16). The module MARK uses these estimators
to mark some cells in Tn for refinement. The marked cells are collected in the set Mn ⊂ Tn . The
fourth module REFINE uses the marked cells in Mn and the current mesh Tn to build a new mesh
Tn+1 for the next iteration.PThe termination criterion of the adaptive loop compares the global
1
upper bound η(un , Tn ) := ( K∈Tn ηK (un )2 ) 2 to the user-prescribed tolerance TOL.
The modules SOLVE and ESTIMATE have been already discussed. The module MARK selects mesh
cells for refinement using Dörfler’s marking [171] (also called bulk chasing criterion) as follows:
Given a fixed parameter θ ∈ (0, 1), MARK determines a set Mn ⊂ Tn of (almost) minimal cardinality
such that
η(un , Mn ) ≥ θ η(un , Tn ), (34.23)
P 1
where η(un , Mn ) := ( K∈Mn ηK (un )2 ) 2 . This marking means that the set Mn contains a sub-
stantial part of the total (or bulk) error. Taking θ small typically means that few mesh cells
are marked. A mesh of minimal cardinality M∗n is one such that card(Mn ) ≥ card(M∗n ) for all
Mn ⊂ Tn s.t. η(un , Mn ) ≥ θ η(un , Tn ). Building a set M∗n of minimal cardinality entails sorting
all the mesh cells, which is of superlinear complexity (for instance the complexity of the merge-sort
algorithm is card(Tn ) ln(card(Tn ))). By relaxing the minimality requirement, one can use a sort-
ing algorithm of linear complexity based on binning, thereby producing a set Mn of cardinality
card(Mn ) ≤ c card(M∗n ) for some uniform constant c.
The module REFINE refines all the marked cells in Mn at least once. Refining mesh cells is
usually done by using a double labeling technique indicating how the cells are to be subdivided
and giving a rule to label the newly created subcells. An important example in dimension two
Figure 34.1: Newest vertex bisection: mesh cell with flagged vertex and dashed line indicating the
bisecting method (left). If the cell is marked for refinement, two new cells are produced (right)
and the newest vertex is flagged in both cells.
is the Newest vertex bisection (NVB), where one vertex of the cell is labeled to indicate that the
opposite edge is to be bisected if refinement is required. If two new subcells are indeed created, the
midpoint of the bisected edge is in turn labeled; see Figure 34.1. A three-dimensional extension of
the NVB exists; see Stevenson [357]. One can verify that a sequence of meshes produced by NVB
is shape-regular. However, since the new mesh Tn+1 must remain a matching mesh, the module
REFINE cannot be completely local. The propagation of refinement beyond the set of marked cells
is a rather delicate issue. A crucial result on the cumulative effect of refinement propagation shown
in Binev et al. [56] for d = 2 and [357] for d > 2 is that, provided the initial labeling of T0 satisfies
some suitable requirements, there is a uniform constant c such that any sequence of successively
bisected meshes satisfies the following bound:
X
card(Tn+1 ) − card(T0 ) ≤ c card(Mm ), (34.24)
m∈{0: n}
whereas single-step uniform bounds of the form card(Tm+1 ) − card(Tm ) ≤ c card(Mm ) may not
hold true.
The first important result for the adaptive finite element solver is a contraction property (im-
plying convergence with geometric rate). This property can be stated on the quasi-error defined as
a weighted sum of the approximation error plus the estimator. In particular, it is shown in Cascón
et al. [113] that using Dörfler’s marking and bisecting marked elements at least once, there exist
γ > 0 and ρ ∈ (0, 1) such that
En+1 ≤ ρEn , (34.25)
where En := k∇(u − un )kL2 (D) + γη(un , Tn ). The proof uses the global error upper bound, but not
the lower bound. The symmetry and coercivity of the bilinear form a and the nesting of the finite
element spaces are also used. Strict error reduction (γ = 0 in the definition of En ) is not true in
general as shown in Morin et al. [305].
The second important result deals with convergence rates. For simplicity, we first discuss the
case without oscillation, and we consider the Laplacian with piecewise polynomial source term on
the initial mesh T0 . For a real number s > 0 and a function y ∈ H01 (D), we consider the following
quantity:
|y|As := sup N s inf inf k∇(y − v)kL2 (D) , (34.26)
N >0 T ∈TN v∈V (T )
where TN is the set of matching simplicial meshes that are refinements of the initial mesh T0 with
cardinality such that card(T ) − card(T0 ) ≤ N , and where V (T ) is the H01 (D)-conforming finite
element space (of some order k) built using the mesh T . Observe that inf v∈V (T ) k∇(y − v)kL2 (D)
represents the best-approximation error of y in V (T ). Moreover, the discrete solution uT ∈ V (T )
from the Galerkin approximation delivers the quasi-optimal error bound k∇(u − uT )kL2 (D) ≤
M
α inf v∈V (T ) k∇(u − v)kL (D) . Using (34.26), we define the approximation class
2
As := {y ∈ H01 (D) | |y|As < ∞}. (34.27)

Membership in As informs on the optimal decay rate one can expect for the approximation error.
Specifically, if u is in As , then there is c s.t. for every N > 0, there is an optimal mesh TN∗ such that
k∇(u − uTN∗ )kL2 (D) ≤ cN −s |u|As . Finding an optimal mesh TN∗ is computationally untractable.
Fortunately, it turns out that the adaptive finite element procedure from Algorithm 34.1 selects
meshes {Tn }n≥0 delivering optimal decay rates. Indeed, it is shown in Cascón et al. [113] (see also
Binev et al. [56], Stevenson [357]) that using Dörfler’s marking with a parameter θ small enough
together with a sorting algorithm such that card(Mn ) ≤ c card(M∗n ) for some uniform constant c,
and if the complexity estimate (34.24) for REFINE holds true, there is c such that for all n ≥ 1,
1
k∇(u − un )kL2 (D) ≤ c |u|As (card(Tn ) − card(T0 ))− s . (34.28)
Note that the error lower bound is used in this proof. In the general case with oscillations, the
problem data f and d are included in the definition of the approximation class, and the decay rate
is established in Cascón et al. [113] for the total error defined as the sum of the approximation
error and the data oscillation. An alternative viewpoint (see Carstensen et al. [112]) is to introduce
approximation classes and decay rates for the estimator, and then use the error lower bound to
infer decay rates for the approximation error.
Exercises
Exercise 34.1 (Residual). Prove (34.10). (Hint : integrate by parts.)
Exercise 34.2 (Trace inequality in stars). Let Ctr,z be defined in (34.12). Prove that Ctr,z ≤
1
2 1 |F |
̟z2 (dCPS,z + 2CPS,z ) 2 with ̟z := hDz maxF ∈Fz◦ |DF | and DF := int(Kl ∪ Kr ) with F := ∂Kl ∩
∂Kr . (Hint : see the proof of Lemma 12.15.)
v
Exercise 34.3 (Bound on dual norm). (i) Prove that kTK (f )kH −1 (K) ≤ chK kf kL2 (K) for
all f ∈ L (K). (Hint : use a scaled Poincaré–Steklov inequality for functions ϕ ∈ H01 (K).) (ii)
2
1
Prove that kTFs (g)kH −1 (DF ) ≤ chF2 kgkL2 (F ) for all g ∈ L2 (F ). (Hint : use the multiplicative trace
inequality from Lemma 12.15.)
(p)
Exercise 34.4 (Oscillation). (i) Let Pm : Lp (K) → Pm be the best-approximation operator in
Lp (K) for p ∈ [1, ∞] and m ∈ N. Prove that
(2) (∞)
k(I − Pm )(θvh )kL2 (K) ≤ k(I − Pm−n )(θ)kL∞ (K) kvh kL2 (K) ,
for all θ ∈ L∞ (K) and all vh ∈ Pn with n ≤ m. (ii) Consider the oscillation indicators defined
in (34.19) with lv := 2k − 2 and ls := 2k − 1. Prove that φvK (uh , f, d) ≤ hK k(I − P2k−2 )(f )kL2 (K) +
(2)
P
c(k(I − Pk−1 )(∇·d)kL∞ (K) + k(I − Pk )(d)kL∞ (K) )k∇uh kL2 (K) with (∇·d)i := j∈{1: d} ∂x dji
(∞) (∞) ∂
j
for all i ∈ {1:d}. Prove that φsF (uh , f, d) ≤ ck(I − Pk )(d)kL∞ (F ) k∇uh kL2 (DF ) with best-
(∞)
(∞)
approximation operator Pk mapping to L∞ (F ). What are the decay rates of the oscillation
terms for smooth f and d? (iii) What happens if lv := k and ls := k − 1 for piecewise constant d?
Exercise 34.5 (Error reduction). Consider two discrete spaces Vh1 ⊂ Vh2 ⊂ H01 (D) with
1
corresponding discrete solutions uh1 and uh2 , respectively. Consider the norm kvka := a(v, v) 2
for all v ∈ H01 (D). Prove that ku − uh1 k2a = ku − uh2 k2a + kuh2 − uh1 k2a . (Hint : use the Galerkin
orthogonality property.)
Exercise 34.6 (Approximation class for smooth solution). Let D be a Lipschitz polyhedron
in Rd . Prove that H k+1 (D) ⊂ Ak/d . (Hint : consider uniformly refined meshes.)
Exercise 34.7 (Graded mesh). Let D := (0, 1) and let (xi )i∈{0: I} , I ≥ 2, be a mesh of D.
Let u ∈ W 1,1 (D) and consider the piecewise constant function uI such that uI (x) := u(xi−1 )
for all x ∈ (xi−1 , xi ) and all i ∈ {1:I}. (i) Assume u ∈ W 1,∞ (D). Prove that the decay rate
ku − uI kL∞ (D) ≤ I1 ku′ kL∞ (D) is achieved using a uniform mesh. (ii) Assume now u ∈ W 1,1 (D).
Prove that the decay rate ku − uI kL∞ (D) ≤ I1 ku′ kL1 (D) is achieved using a graded mesh such that
Rs ′
xi := Φ(−1) ( Ii ), where Φ(s) := ku′ k 1 1 0
|u (t)| dt for all s ∈ (0, 1) and all i ∈ {0:I}.
L (D)
Chapter 35
The Helmholtz problem
The objective of this chapter is to give a brief overview of the analysis of the Helmholtz problem
and its approximation using H 1 -conforming finite elements. The Helmholtz problem arises when
modeling electromagnetic or acoustic scattering problems in the frequency domain. One specificity
of this elliptic problem is that one cannot apply the Lax-Milgram lemma to establish well-posedness.
The correct way to tackle the Helmholtz problem is to invoke the BNB theorem (Theorem 25.9).
In the entire chapter, D is a Lipschitz domain in Rd with d ≥ 1, i.e., a nonempty open bounded
and connected subset of Rd with a Lipschitz boundary.
35.1 Robin boundary conditions

We investigate in this section the Helmholtz problem with Robin boundary conditions. Given
f ∈ L2 (D), g ∈ L2 (∂D), and κ ∈ R, our goal is to find a function u : D → C such that
−∆u − κ2 u = f in D, ∂n u − iκu = g on ∂D, (35.1)
with i2 = −1. Notice that the Robin boundary condition couples the real and imaginary parts of
u. The sign of the parameter κ is irrelevant in what follows, but to simplify some expressions, we
henceforth assume that κ > 0. All that is said below remains valid when κ < 0 by replacing κ by
|κ| in the definitions of the norms and in the upper bounds. Note that κ−1 is a length scale. The
problem (35.1) can be reformulated as follows in weak form:

(35.2)
a(u, w) = ℓ(w), ∀w ∈ V,
with the sesquilinear form

Z Z
2
a(v, w) := (∇v·∇w − κ vw) dx − iκ γ g (v)γ g (w) ds, (35.3)
D ∂D
R R 1
and the antilinear form ℓ(w) := D f w dx + ∂D gγ g (w) ds, where γ g : H 1 (D) → H 2 (∂D) is the
trace map.
Remark 35.1 (Sommerfeld radiation condition). The Helmholtz problem is in general posed
on unbounded domains, and the proper “boundary condition to set at infinity” is the Sommerfeld
132 Chapter 35. The Helmholtz problem
d−1
radiation condition limr→∞ r 2 (e·∇u(re) − iκu(re)) = 0 for every unit vector e ∈ Rd and the
convergence must be uniform with respect to n. One usually simplifies this problem by truncating
the domain and replacing the Sommerfeld radiation condition by a Robin boundary condition as
in (35.1).
Remark 35.2 (Wave equation). The Helmholtz problem can be derived by considering the
wave equation ∂tt v − c2 ∆v = g(x) cos(ωt) in D×(0, T ) with appropriate initial data and boundary
conditions; see §46.2.1 and §46.2.2. Here, c is the wave speed and g is some forcing. Assuming that
the solution is of the form v(x, t) = ℜ(u(x)eiωt ), the complex amplitude u solves ω 2 u − c2 ∆u = g.
We then recover (35.1) by setting κ := ωc .
35.1.1 Well-posedness
Contrary to what was done in the previous chapters, we cannot apply the Lax–Milgram lemma
to establish that the weak formulation (35.2) is well-posed since the sesquilinear form a is not
coercive. We are going to invoke instead the BNB theorem (Theorem 25.9), and with this goal in
mind, we first establish an abstract result.
Lemma 35.3 (Gårding). Let V ֒→ L be two Banach spaces with compact embedding. Let a :
V ×V → C be a bounded sesquilinear form. Assume that there exist two real numbers β, γ > 0 such
that the following holds true:
|a(v, v)| + βkvk2L ≥ γkvk2V , ∀v ∈ V, (35.4a)
[ a(v, w) = 0, ∀w ∈ V ] =⇒ [ v = 0 ]. (35.4b)
|a(v,w)|
Then there is α > 0 such that inf v∈V supw∈V kvkV kwkV ≥ α.
Proof. Let us argue by contradiction like in the proof of the Peetre–Tartar lemma (Lemma A.20).
Assume that for every integer n ≥ 1, there is vn ∈ V with kvn kV = 1 and supw∈V |a(vn , w)|/kwkV ≤
1
n . Since the embedding V ֒→ L is compact, there is a subsequence (vl )l∈S , S ⊂ N, such that (vl )l∈S
converges strongly to some v in L. The assumption (35.4a) implies that
γkvm − vn k2V ≤ βkvm − vn k2L + |a(vm − vn , vm − vn )|

≤ βkvm − vn k2L + |a(vm , vm )| + |a(vm , vn )| + |a(vn , vm )| + |a(vn , vn )|.
Since |a(vl , vl′ )| = |a(vl , vl′ )|/kvl′ kV ≤ 1l , for all l, l′ ∈ {m, n}, we infer that γkvm − vn k2V ≤
βkvm − vn k2L + 2(m−1 + n−1 ), which in turn implies that (vl )l∈S is a Cauchy sequence in V. As a
result, v ∈ V and supw∈V |a(v, w)|/kwkV = 0, which means that a(v, w) = 0 for all w ∈ V. The
assumption (35.4b) implies that v = 0, which contradicts 1 = limS∋l→∞ kvl kV = kvkV .
Remark 35.4 (Gårding’s inequality). Inequalities like (35.4a) are called Gårding’s inequality
in the literature.
Theorem 35.5 (BNB, Robin BCs). Let V := H 1 (D) be equipped with the norm kvkV :=
1
{k∇vk2L2(D) + κkvk2L2 (∂D) } 2 . The sesquilinear form a defined in (35.3) satisfies the conditions of
the BNB theorem.
Proof. We are going to verify (35.4a) and (35.4b) from Lemma 35.3.
(1) Let v ∈ V. The real and imaginary parts of a(v, v) are
ℜ(a(v, v)) = k∇vk2L2 (D) − κ2 kvk2L2 (D) , (35.5a)
ℑ(a(v, v)) = −κkvk2L2(∂D) . (35.5b)
√ 2 1
Using that 2(x + y 2 ) 2 ≥ x − y for all x, y ∈ R, this implies that
√
2|a(v, v)| ≥ kvk2V − κ2 kvk2L2 (D) .
Hence, (35.4a) holds true with β := √12 κ2 and γ := √12 .

(2) Let us now assume that a(v, w) = 0 for all w ∈ V. We are going to prove that v = 0 by arguing
by contradiction. The inequality |a(v, v)| ≥ −ℑ(a(v, v)) = κkvk2L2 (∂D) implies that γ g (v) = 0.
Hence, v ∈ H01 (D). Let us embed D into a ball of radius R large enough, say R > R0 := diam(D),
and without loss of generality, we assume that this ball is centered at 0. Let BR be the ball in
c :=
question and let us set DR Dc ∩ BR , where Dc denotes the complement of D in Rd . Since
c
v|∂D = 0, we can extend v by zero over DR , and we denote by veR the extension in question.
1 c
We have e vR ∈ H0 (BR ), (∇e vR )|D ∈ H(div; D), and (∇e vR )|DRc ∈ H(div; DR ). Since the Robin
boundary condition implies that ∂n v|∂D = 0, we infer that the normal component of ∇e vR is
continuous across ∂D. Reasoning as in the proof of Theorem 18.10, we conclude that ∇e vR is a
member of H(div; BR ). This means that ∆e vR ∈ L2 (BR ). Since veR ∈ H01 (BR ) and veR vanishes on
an open subset of BR , we can invoke the unique continuation principle (see Theorem 31.4) to infer
that veR = 0 in BR . Hence, v = 0 in D and the property (35.4b) holds true.
Remark 35.6 (Alternative proof ). Instead of invoking the unique continuation principle in
the above proof, one can use the spectral theorem for symmetric compact operators (see Theo-
rem 46.21). The above reasoning shows that veR ∈ H01 (BR ) and −∆e vR = κ2 veR in BR . Hence,
2
if e
vR is not zero, then κ is an eigenvalue of the Laplace operator equipped with homogeneous
Dirichlet boundary conditions on every ball centered at 0 in Rd with radius larger than R0 . How-
ever, Theorem 46.21 says that the eigenvalues of the Laplace operator in H01 (BR ) are countable
with no accumulation point and are of the form (R−2 λn )n∈N for every R > 0, where (λn )n∈N are
the eigenvalues of the Laplace operator in H01 (B1 ). Assuming that the eigenvalues are ordered in
increasing order, let R0′ > R0 be large enough so that there is some n ∈ N such that κ2 (R0′ )2 = λn
with λn < λn+1 . Let δ be defined by κ2 (R0′ + δ)2 := 21 (λn + λn+1 ). Then κ2 (R0′ + δ)2 cannot be
in the set {λn }n∈N , but this is a contradiction since the above reasoning with R := R0′ + δ shows
that κ2 R2 = κ2 (R0′ + δ)2 is a member of the sequence (λn )n∈N if veR is not zero. This proves that
e
vR = 0.
35.1.2 A priori estimates on the solution

In this section, we derive a priori estimates on the weak solution of (35.2). We are particularly
interested in estimating the possible dependence of the upper bound on the (nondimensional)
quantity κℓD with ℓD := diam(D). The following result, established in Melenk [299, Prop. 8.1.4]
and Hetmaniuk [243], delivers a sharp upper bound on the V -norm of the weak solution that relies
on the relatively strong assumption that the domain D is star-shaped with respect to some point
in D which we take to be 0.
Lemma 35.7 (A priori estimate). Assume that D is a bounded Lipschitz domain and star-
shaped w.r.t. 0, i.e., there exists r > 0 s.t. x·n > rℓD for all x ∈ ∂D. Let V := H 1 (D) be equipped
1
with the norm kvkV := {k∇vk2L2 (D) + κkvk2L2 (∂D) } 2 . There is a constant c that depends only on
D (i.e., it is independent of κℓD ) such that the weak solution of (35.2) satisfies
1
κkukL2(D) + kukV ≤ c (ℓD kf kL2 (D) + ℓD
2
kgkL2 (∂D) ). (35.6)
Proof. We only give the proof when κ is bounded away from zero, say κℓD ≥ 1 since the proof in
the other case is similar; see [299, 243]. Since we assume that 0 ∈ D, we have kxkℓ2 ≤ ℓD for all
1
x ∈ D. We write C(f, g) := c(ℓD kf kL2 (D) + ℓD2
kgkL2 (∂D) ), where as usual the value of the constant
c can change at each occurrence as long as it is independent of κ.
(1) In the first step of the proof, we assume that ∇u|∂D ∈ L2 (∂D) (we establish this smoothness
property in the second step). Let us multiply the PDE −∆u − κ2 u = f with x·∇u and integrate
over D. The identity (35.11) from Lemma 35.8 with m := x implies that
Z
d
−ℜ ∆u x·∇u dx = 1 − k∇uk2L2(D)
D 2
Z Z
1 2
+ (x·n)k∇ukℓ2 ds − ℜ (∂n u)(x·∇u) ds ,
2 ∂D ∂D
since ∇x = (∇x)T = Id and ∇·x = d so that e(x) = (1 − d2 )Id (see Lemma 35.8). This identity
is often called Rellich’s identity in the literature. Using the PDE −∆u − κ2 u = f , the Robin
boundary condition ∂n u = iκu + g, and the assumption x·n > rℓD on ∂D, we obtain
d Z
rℓD 2 2 2
k∇ukL2 (∂D) ≤ − 1 k∇ukL2(D) + ℜ κ u(x·∇u) dx
2 2 D
Z Z
+ℜ f (x·∇u) dx + ℜ (iκu + g)(x·∇u) ds .
D ∂D
R R
Since ℜ( D u(x·∇u) dx) = − d2 kuk2L2(D) + 1
2 ∂D
(x·n)|u|2 ds, this leads to
rℓD dκ2 d κ2 ℓ D
k∇uk2L2(∂D) + kuk2L2 (D) ≤ − 1 k∇uk2L2(D) + kuk2L2(∂D)
2 2 2 Z 2 Z
+ℜ f (x·∇u) dx + ℜ (iκu + g)(x·∇u) ds .
D ∂D
We now bound the last two terms on the right-hand side by using Young’s inequality, which yields
Z Z
ℜ f (x·∇u) dx + ℜ (iκu + g)(x·∇u) ds ≤ γ1 k∇uk2L2 (D)
D ∂D
1 2 rℓD 2ℓD 2
+ ℓD kf k2L2 (D) + k∇uk2L2 (∂D) + κ kuk2L2(∂D) + kgk2L2(∂D) ,
4γ1 4 r
where γ1 > 0 can be chosen as small as needed. Rearranging the terms gives
rℓD dκ2 d
k∇uk2L2 (∂D) + kuk2L2(D) ≤ − 1 + γ1 k∇uk2L2 (D)
4 2 2
r+4 2
+ κ ℓD kuk2L2(∂D) + C(f, g)2 . (35.7)
2r
Let us now bound the norms k∇uk2L2 (D) and kuk2L2 (∂D) appearing on the right-hand side. Owing
to (35.5a) and Young’s inequality, we infer that

k∇uk2L2 (D) = κ2 kuk2L2(D) + ℜ (f, u)L2 (D) + (g, γ g (u))L2 (∂D)
1 1 1
≤ (1 + γ2 )κ2 kuk2L2 (D) + kf k2L2 (D) + kgk2L2(∂D) + κkuk2L2 (∂D) ,
4γ2 κ2 2κ 2
where γ2 > 0 can be chosen as small as needed. Since we assumed above that κℓD ≥ 1, we obtain
1
k∇uk2L2(D) ≤ (1 + γ2 )κ2 kuk2L2 (D) + κkuk2L2(∂D) + C(f, g)2 . (35.8)
2
Owing to (35.5b), we infer that

κkuk2L2 (∂D) = −ℑ (f, u)L2 (D) + (g, γ g (u))L2 (∂D) ,
and applying Young’s inequality with a positive real number θ gives

1 1 1
κkuk2L2 (∂D) ≤ θκkuk2L2 (D) + kf k2L2 (D) + kgk2L2(∂D) .
2 4θκ 2κ
Taking θ := γ3 κ with γ3 > 0 as small as needed leads to (recall that κℓD ≥ 1)
1
κkuk2L2(∂D) ≤ γ3 κ2 kuk2L2 (D) + C(f, g)2 . (35.9)
2
1 r r+4
In addition, taking θ := 2ℓD r+4 and multiplying by r κℓD yields
r+4 2 1
κ ℓD kuk2L2(∂D) ≤ κ2 kuk2L2 (D) + C(f, g)2 . (35.10)
2r 2
Inserting (35.9) into (35.8) gives k∇uk2L2 (D) ≤ (1 + γ2 + γ3 )κ2 kuk2L2 (D) + C(f, g)2 , and inserting
this bound into (35.7), we obtain
rℓD dκ2 d
k∇uk2L2 (∂D) + kuk2L2(D) ≤ − 1 + γ1 (1 + γ2 + γ3 )κ2 kuk2L2(D)
4 2 2
r+4 2
+ κ ℓD kuk2L2 (∂D) + C(f, g)2 .
2r
Using now the bound on kuk2L2 (∂D) from (35.10), we infer that

rℓD d d 1 2
k∇uk2L2(∂D) + κ2 kuk2L2 (D) ≤ − 1 + γ1 (1 + γ2 + γ3 ) + κ kuk2L2 (D) + C(f, g)2 .
4 2 2 2
1 1
Letting γ1 := 4d , γ2 = γ3 := 8d , we observe that ( d2 − 1 + γ1 )(1 + γ2 + γ3 ) = d
2 − 7
8 + 1
16d2 ≤ d
2 − 1
4
for all d ≥ 1. We conclude that
rℓD κ2
k∇uk2L2(∂D) + kuk2L2(D) ≤ C(f, g)2 .
4 4
Invoking once again the bounds (35.8) and (35.9), we infer that
κ2 kuk2L2(D) + κkuk2L2(∂D) + k∇uk2L2 (D) + ℓD k∇uk2L2 (∂D) ≤ C(f, g)2 ,
which shows that the a priori estimate (35.6) holds true.

(2) It remains to prove that indeed ∇u|∂D ∈ L2 (∂D). Recall that u is in the functional space
Y := {y ∈ H 1 (D) | ∆y ∈ L2 (D), ∂n y ∈ L2 (∂D)} owing to (35.1) and our assumption that
f ∈ L2 (D) and g ∈ L2 (∂D). We are going to show by means of a density argument that any
function y ∈ Y is such that ∇y|∂D ∈ L2 (∂D). Let (ϕm )m∈N be a sequence in C ∞ (D) converging
to y in Y (such a sequence can be constructed by using mollifying operators, as in §23.1). Let
us set fm := −∆ϕm − ϕm and gm := ∂n ϕm − iκϕm . Then (fm )m∈N and (gm )m∈N are Cauchy
sequences in L2 (D) and L2 (∂D), respectively. Moreover, the bound from Step (1) implies that
1
k∇(ϕm − ϕp )kL2 (∂D) ≤ c(ℓD
2
kfm − fp kL2 (D) + kgm − gp kL2 (∂D) ) for all m, p ∈ N, which shows that
(∇ϕm )m∈N is a Cauchy sequence in L2 (∂D). The uniqueness of the limit in the distribution sense
finally shows that ∇y|∂D ∈ L2 (∂D).
Lemma 35.8 (Special identity). For all q ∈ {v ∈ H 1 (D; C) | ∆v ∈ L2 (D; C), ∇v ∈ L2 (∂D; Cd )}
and all m ∈ W 1,∞ (D; Rd ), letting e(m) := 21 (∇m + (∇m)T − (∇·m)Id ), we have
Z Z
−ℜ ∆q(m·∇q) dx = ℜ ∇q·(e(m)∇q) dx
D D
Z Z
1
+ (m·n)k∇qk2ℓ2 ds − ℜ (n·∇q)(m·∇q) ds . (35.11)
2 ∂D ∂D
Proof. See Exercise 35.4 and Hetmaniuk [243, Lem. 3.2].

A detailed analysis of the Helmholtz problem (35.2) using integral representations is done in
Esterhazy and Melenk [195, §2]. The following result is established therein.
Theorem 35.9 (BNB, Robin BCs). Let D be a Lipschitz domain in Rd , d ∈ {2, 3}. Let
V := H 1 (D) be equipped with the norm kvkV := κkvkL2 (D) + k∇vkL2 (D) . Let k0 > 0 be a fixed
number and set κ0 := k0 ℓ−1 D . Then there is c > 0, depending on D and k0 , such that the following
holds true for all κ ≥ κ0 :
|a(v, w)|
inf sup ≥ c (κℓD )−s , (35.12)
v∈V w∈V kvkV kwkV
7
with s := 2 in general, and s := 1 if D is convex or if D is star-shaped or if ∂D is smooth.
1
This theorem implies, in particular, that for every f ∈ V ′ := (H 1 (D))′ and g ∈ H − 2 (∂D) =
1
(H 2 (∂D))′ , the problem (35.2) is uniquely solvable in V, and its solution satisfies the a priori
7
bound kukV ≤ c(κℓD ) 2 (kf kV ′ + kgk − 12 ). If f ∈ L2 (D) and g ∈ L2 (∂D), this estimate can
H (∂D)
5 1
be improved to kukV ≤ c(κℓD ) 2 (ℓD kf kL2(D) + κ− 2 kgkL2 (∂D) ); see [195, Thm. 2.5].
35.2 Mixed boundary conditions

We consider in this section the Helmholtz problem with mixed Dirichlet and Robin boundary
conditions. The problem is formulated as follows: For f ∈ L2 (D), g ∈ L2 (∂Dr ), and κ ∈ R, find a
complex-valued function u such that
−∆u − κ2 u = f in D, u = 0 on ∂Dd , ∂n u − iκu = g on ∂Dr , (35.13)
where {∂Dd , ∂Dr } is a partition of ∂D. We assume that the subsets ∂Dd and ∂Dr have a Lipschitz
boundary and have positive (surface) measure. As before, we assume that κ > 0 for simplicity.
The above problem is reformulated as follows:

Find u ∈ V := {v ∈ H 1 (D) | γ g (v)|∂D d = 0} such that
(35.14)
a(u, w) = ℓ(w), ∀w ∈ V,
with the sesquilinear form

Z Z
a(v, w) := (∇v·∇w − κ2 vw) dx − iκ vw ds, (35.15)
D ∂Dr
R R
and the antilinear form ℓ(w) := D f w dx + ∂Dr gw ds. Here again, we cannot apply the Lax–
Milgram lemma since a is not coercive on V. We are going to invoke instead the BNB theorem.
Theorem 35.10 (BNB, mixed BCs). Let the space V defined in (35.14) be equipped with the
norm kvkV := k∇vkL2 (D) . The sesquilinear form a defined in (35.15) satisfies the conditions of
the BNB theorem.
Proof. We are going to invoke Lemma 35.3. We can proceed as in the proof of Theorem 35.5 to
prove the Gårding inequality (35.4a), but we proceed slightly differently to prove (35.4b). Let us
assume that a(v, w) = 0 for all w ∈ V. The inequality |a(v, v)| ≥ κkvk2L2 (∂Dr ) implies that v|∂Dr = 0.
Since |∂Dr | > 0, there exists a point x0 ∈ ∂Dr and there is r0 > 0 such that B(x0 , r0 ) ∩ ∂D ⊂ ∂Dr .
Let Drc0 := Dc ∩ B(x0 , r0 ). We extend v by zero over Drc0 , denote the extension in question
by evr0 and set D e r0 := int(D ∪ Dc ). We have ver0 ∈ H 1 (D e r0 ), (∇evr0 )|D ∈ H(div; D), and
r0 0
(∇evr0 )|Drc ∈ H(div; Drc0 ). Since the Robin boundary condition implies that (∂n v)|∂D r = 0, we
0
infer that the normal component of ∇e vr0 is continuous across ∂Dr ∩ B(x0 , r0 ). Reasoning as in the
proof of Theorem 18.10, we conclude that ∇e vr0 is a member of H(div; D e r0 ), i.e., ∆e e r0 ).
vr0 ∈ L2 (D
1 e 2 e r0 , and ver |Dc = 0. The unique
In conclusion, we have ver0 ∈ H0 (Dr0 ), −∆e vr0 = κ e vr0 in D 0 r0
continuation principle (Theorem 31.4) implies that ver0 = 0. Hence, v = 0.
Following Ihlenburg and Babuška [251], we now set D := (0, ℓD ) and investigate the one-
dimensional version of the problem (35.13). A homogeneous Dirichlet boundary condition is en-
forced at {x = 0}, and a homogeneous Robin condition is enforced at {x = ℓD }. The space V
becomes V := {v ∈ H 1 (D) | v(0) = 0}.
Theorem 35.11 (BNB, mixed BCs, 1D). Let D := (0, ℓD ). Let the space V be equipped with
the norm kvkV := k∂x vkL2 (D) . There are two constants 0 < c♭ ≤ c♯ , both uniform with respect to
κ, such that
c♭ |a(v, w)| |a(v, w)| c♯
≤ inf sup ≤ sup sup ≤ .
1 + κℓD v∈V w∈V kvkV kwkV v∈V w∈V kvkV kwkV 1 + κℓD
Proof. (1) Let us start with the lower bound. Let v ∈ V, v 6= 0, and let z ∈ V solve a(w, z) =
(w, κ2 v)L2 (D) for all w ∈ V. It is shown in Exercise 35.1 that this problem has a unique solution in
V, and it is shown in Exercise 35.2 that kzkV ≤ 4κℓD kvkV . Then we have
|a(v, v + z)| ≥ ℜ(a(v, v + z)) = ℜ(a(v, v)) + κ2 kvk2L2 (D)

1
= kv ′ k2L2 (D) = kvk2V = kvkV (kvkV + 4κℓD kvkV )
4κℓD + 1
1 1
≥ kvkV (kvkV + kzkV ) ≥ kvkV kv + zkV .
4κℓD + 1 4κℓD + 1
This shows that the lower bound holds true.
(2) Let us now prove the upper bound. Let v ∈ V.
(2.a) If κℓD ≤ 2, then we can invoke the following Poincaré–Steklov inequality in V : there is a
−1
constant C̃ps > 0 s.t. C̃ps (ℓ−1
D kvkL2 (D) + ℓD |v(ℓD )|) ≤ kvkV (see the proof of Proposition 31.21).
2
Using the Cauchy–Schwarz inequality in (35.3) implies that
|a(v, w)| ≤ kvkV kwkV + κ2 kvkL2 (D) kwkL2 (D) + κ|v(ℓD )||w(ℓD )|
−2
≤ max(1, C̃ps )(1 + κℓD + (κℓD )2 )kvkV kwkV .
Since we assumed κℓD ≤ 2, this leads to the bound |a(v, w)| ≤ c(1 + κℓD )−1 kvkV kwkV with
−2
c := max(1, C̃ps ) maxt∈[0,2] (1 + t + t2 )(1 + t).
(2.b) Let us now assume that κℓD ≥ 2. Let ϕ be a smooth nonnegative function equal to 1 on
[0, 12 ℓD ] and such that ϕ(ℓD ) = ∂x ϕ(ℓD ) = 0. Let us set w(x) := ϕ(x) sin(κx)/κ so that w ∈ V,
Rx
w(0) = 0, w(ℓD ) = 0, and ∂x w(ℓD ) = 0. Let us set η(x) := ∂x w(x) − ∂x w(0) + κ2 0 w(s) ds, and
cϕ := max(2ℓD k∂x ϕkL∞ (D) , ℓ2D k∂xx ϕkL∞ (D) ). Since w is real-valued and vanishes at x = ℓD and
v(0) = 0, we have
Z ℓD Z ℓD
a(v, w) = ∂x v∂x w dx − κ2 vw dx
0 0
Z ℓD Z ℓD Z x
= (∂x v)η dx + v(ℓD )∂x w(0) − κ2 vw + ∂x v w(s) ds dx.
0 0 0
Rℓ
The last term is equal to −κ2 v(ℓD ) 0 D w(s) ds since v(0) = 0. Since η(ℓD ) = −∂x w(0) +
Rℓ 1
κ2 0 D w(s) ds and |v(ℓD )| ≤ ℓD
2
kvkV , we infer that
Z ℓD
|a(v, w)| = (∂x v)η dx − v(ℓD )η(ℓD )
0
1 1
≤ kvkV (kηkL2 (D) + ℓD
2
|η(ℓD )|) ≤ 2ℓD
2
kvkV kηkL∞ (D) .
Since η(0) = 0, we have kηkL∞ (D) ≤ ℓD k∂x ηkL∞ (D) . After observing that
∂x η(x) = ∂xx ϕ(x) sin(κx)/κ + 2∂x ϕ(x) cos(κx)
and recalling the above bounds on the derivatives of ϕ, we deduce that kηkL∞ (D) ≤ cϕ (1+(κℓD )−1 ).
1
Hence, we have |a(v, w)| ≤ 2cϕ (1 + (κℓD )−1 )ℓD
2
kvkV . After observing that
Z 1
2 ℓD ℓD 1 ℓD
kwk2V ≥ cos(κx)2 dx ≥ − ≥ ,
0 4 4κ 8
1
since κℓD ≥ 2, we conclude that kwkV ≥ ( 18 ℓD ) 2 . Hence, |a(v, w)| ≤ c(1 + κℓD )−1 kvkV kwkV , and
the proof is complete.
Remark 35.12 (Literature). Theorem 35.11 has been derived in Ihlenburg and Babuška [251,
Thm. 1], and we refer the reader to this work for an exhaustive analysis of the continuous problem
in one dimension with g := 0. Two- and three-dimensional versions of Lemma 35.7 for mixed
boundary conditions are established in Hetmaniuk [243].
35.3 Dirichlet boundary conditions

We consider in this section the Helmholtz problem with Dirichlet boundary conditions: For f ∈
L2 (D; R) and κ ∈ R, find u such that
−∆u − κ2 u = f in D, u = 0 on ∂D. (35.16)
As before, we assume that κ > 0 for simplicity. Note that the solution is now real-valued. We
reformulate the above problem as follows:

(35.17)
a(u, w) = ℓ(w), ∀w ∈ V,
with the bilinear form Z

a(v, w) := (∇v·∇w − κ2 vw) dx, (35.18)
D
R
and the linear form ℓ(v) := D f v dx. As above, we are going to rely on the BNB theorem to
establish the well-posedness (35.17) since a is not coercive. But contrary to the case with Robin or
mixed boundary conditions, the enforcement of Dirichlet conditions leads to a conditional stability
depending on the value of κ. In other words, resonance phenomena can occur if κ takes values
in some discrete subset of R+ associated with the spectrum of the Laplacian operator in D with
Dirichlet conditions.
Since the embedding H01 (D) ֒→ L2 (D) is compact and the operator (−∆)−1 : L2 (D) → L2 (D)
is self-adjoint, there exists a Hilbertian basis of L2 (D) composed of eigenvectors of the Laplace
operator (see Theorem 46.21). Let (ψl )l∈N be the basis in question and let (λl )l∈N be the cor-
responding eigenvalues with the normalization
P Then every function v ∈PH01 (D)
kψl kL2 (D) = 1. P
:= 2 2 2 2
admits a unique expansion v l∈N vl ψl with k∇vkL2 (D) = l∈N λl vl , kvkL2 (D) = l∈N vl .
P 2
P P 1
Notice that a(v, w) = l∈N (λl − κ )vl wl for all v = l∈N vl ψl , w = l∈N wl ψl in H0 (D). Let
us denote by l(κ) the largest integer such that λl(κ) < κ2 with the convention that l(κ) = −1 if
κ2 ≤ λ0 . The well-posedness of the problem (35.17) follows from the following result.
Theorem 35.13 (BNB, Dirichlet BCs). Let V := H01 (D) be equipped with the norm kvkV :=
k∇vkL2 (D) . Assume that κ2 6∈ {λl }l∈N . Then the bilinear form a satisfies the conditions of the
BNB theorem with the constant α(κ) := minl∈N |λl − κ2 |/λl > 0.
P P P
Proof. Let v ∈ H01 (D) with v := l∈N vl ψl . Let us set w := l≤l(κ) −vl ψl + l(κ)<l vl ψl with the
convention that l ∈ N in the sums. Then we have
X X X
a(v, w) = (κ2 − λl )vl2 + (λl − κ2 )vl2 ≥ α(κ) λl vl2 = α(κ)kvk2V .
l≤l(κ) l(κ)<l l∈N
The assertion follows readily from kwkV = kvkV . The reader is referred to Ciarlet [120, §3.1] for
more details on this problem.
In general, α(κ) behaves like α0 γ(κ)(κℓD )−1 , where γ(κ) ∈ (0, 1] and α0 only depends on D.
For D := (0, ℓD ), the eigenvalues of the Laplace operator are λl := πl2 ℓ−2 D . Let β ∈ (0, 1) and
L ∈ N \ {0} be s.t. κ2 := π(L + β)2 ℓ−2
D . Then α(κ) = min(β(2L+β)/L 2
, (1+β)(2L+1+β)/(L+1) 2
),
and the claim follows readily. Notice that γ(κ) becomes arbitrarily small as κ approaches an
eigenvalue of the Laplace operator, i.e., if β is close to 0.
35.4 H 1 -conforming approximation

We now formulate an H 1 -conforming approximation of the Helmholtz problem with one of the
boundary conditions discussed in the previous sections (Robin, mixed or Dirichlet). At this stage,
we do not specify the norm with which we equip the space V : we just assume that it is an H 1 -like
norm that can contain some lower-order terms depending on κ (see Example 35.18).
Let (Th )h∈H be a shape-regular mesh sequence so that each mesh covers D exactly. In the
case of mixed boundary conditions, we also assume that the meshes are compatible with the
corresponding partition of the boundary ∂D. Let k ≥ 1 be the degree of the underlying finite
element. Let Pkg (Th ) be the H 1 -conforming finite element space considered in §18.2.3 and §32.1.
For the Robin problem, we set Vh := Pkg (Th ), and for the mixed and the Dirichlet problems we set
Vh := {vh ∈ Pkg (Th ) | vh|∂Dd = 0}. (35.19)
We construct an approximation of the Helmholtz problem as follows:

(35.20)
a(uh , wh ) = ℓ(wh ), ∀wh ∈ Vh .
A first way to investigate the stability of the discrete problem (35.20) consists of reasoning
by perturbation using the fact that the continuous problem is well-posed. Such a result can be
obtained by invoking a variation of Fortin’s lemma (a more abstract version of this variation is
discussed in Exercise 35.3). Recall that the elliptic projection ΠEh : V → Vh is defined for all v ∈ V
s.t. (∇(v − ΠEh (v)), ∇wh )L2 (D) = 0 for all wh ∈ Vh (see §32.4).
Lemma 35.14 (Modified Fortin). Assume that there are positive real numbers γstb , capp , s
such that the elliptic projection satisfies for all v ∈ V,
γstb kΠEh (v)kV ≤ kvkV , kv − ΠEh (v)kL2 (D) ≤ capp hs ℓ1−s

D kvkV . (35.21)
Let α be the inf-sup constant of a on V ×V. Let ιL,V > 0 be such that
kvkL2 (D) ≤ ιL,V ℓD kvkV . (35.22)

−1 s−2 −2 s 1
Assume that h ∈ H ∩ (0, ℓ0 (κ)] with ℓ0 (κ) := ( 12 c−1
app ιL,V αℓD κ ) . Then the restriction of a to
Vh ×Vh satisfies the following inf-sup condition:
|a(vh , wh )| 1
inf sup ≥ α0 := γstb α > 0. (35.23)
vh ∈Vh wh ∈Wh kvh kV kwh kV 2
Proof. Using that ΠEh (V ) ⊂ Vh and the assumptions on ΠEh , we have
−1 |a(vh , wh )| −1 |a(vh , ΠEh (w))| |a(vh , ΠEh (w))|

γstb sup ≥ γstb sup ≥ sup
wh ∈Vh kwh kV w∈V kΠEh (w)kV w∈V kwkV
|a(vh , w) + κ2 (vh , w − ΠEh (w))L2 (D) |
≥ sup
w∈V kwkV
|a(vh , w)|
≥ sup − capp ιL,V hs ℓ2−s 2 s 2−s 2
D κ kvh kV ≥ (α − capp ιL,V h ℓD κ )kvh kV .
w∈V kwkV
−1 |a(vh ,wh )|
Since h ≤ ℓ0 (κ), using the definition of ℓ0 (κ) yields γstb supwh ∈Vh kwh kV ≥ 21 αkvh kV , i.e., (35.23)
1
holds true with α0 := 2 γstb α.
The above result can be applied with s := 1 when full elliptic regularity is available. One
always has s > 21 in polyhedra (see Theorem 31.31).
Remark 35.15 (Duality argument). A duality argument is implicitly present in the assump-
tions of Lemma 35.14 since duality has to be invoked to establish the approximation property
kv − ΠEh (v)kL2 (D) ≤ capp hs ℓ1−s
D kvkV (see Theorem 32.15).
A second way to investigate the stability of the discrete problem (35.20) is a technique intro-
duced by Schatz [343] based on the Aubin–Nitsche duality argument.
Lemma 35.16 (Schatz). Let V, W be two Banach spaces, W being reflexive. Let a be a bounded
sesquilinear form on V ×W satisfying the conditions of the BNB theorem with inf-sup and bound-
edness constants 0 < α ≤ kak. Let L be a Hilbert space such that kvkL ≤ ιL,V kvkV for all v ∈ V
(i.e., V ֒→ L). Let (Vh )h∈H , (Wh )h∈H be sequences of finite-dimensional subspaces equipped,
respectively, with the norm of V and the norm of W. Assume the following:
|a(vh ,wh )|
(i) (Gårding’s inequality) There are cV > 0, cL ≥ 0 s.t. cV kvh kV −cL kvh kL ≤ supwh ∈Wh kwh kW
for all vh ∈ Vh .
(ii) (Duality argument) There is a subspace Ws ֒→ W and real numbers csmo , capp , and s ∈ (0, 1]
s.t. inf wh ∈Wh kz − wh kW ≤ capp hs kzkWs for all z ∈ Ws and all h ∈ H. Moreover, for all
g ∈ L, the unique solution z ∈ W to the adjoint problem a(v, z) = (v, g)L for all v ∈ V,
satisfies kzkWs ≤ csmo kgkL .
1
Assume that h ∈ H ∩ (0, ℓ0 (κ)] with ℓ0 (κ) := ( 12 cV c−1
L kak
−1 −1 −1 s
capp csmo ) . Then the restriction of a
to Vh ×Wh satisfies the discrete inf-sup condition (35.23) with α0 ≥ 2(kak+cLcιVL,V + 1 cV ) α.
2
′
Proof. Let vh 6= 0 be a member of Vh . Consider the antilinear form ℓh ∈ (Wh ) defined by ℓh (wh ) :=
a(vh , wh ) for all wh ∈ Wh . (Note that ℓh := Ah (vh ) with Ah ∈ L(Vh ; Wh′ ) s.t. hAh (yh ), wh iWh′ ,Wh :=
a(yh , wh ) for all (yh , wh ) ∈ Vh ×Wh .) Owing to the Hahn–Banach theorem (Theorem C.13), we
can extend ℓh to W. Let ℓeh be the extension in question with kℓeh kW ′ = kℓh kWh′ . Since a satisfies
the conditions of the BNB theorem, there exists u ∈ V such that a(u, w) := ℓeh (w) for all w ∈ W.
(Notice that u := A−1 (ℓeh ) with A ∈ L(V ; W ′ ) s.t. hA(y), wiW ′ ,W := a(y, w) for all (y, w) ∈ V ×W.)
Using the inf-sup condition satisfied by a on V ×W , we infer that
|a(vh , wh )| |ℓh (wh )| |a(u, w)|
sup = sup = kℓh kWh′ = kℓeh kW ′ = sup ≥ αkukV .
wh ∈Wh kwh kW wh ∈Wh kw h k W w∈W kwkW
The rest of the proof consists of showing that there is c s.t. kukV ≥ ckvh kV for all h ∈ H. Invoking
Gårding’s inequality on Vh gives
|a(vh , wh )| |a(u, wh )|
cV kvh kV − cL kvh kL ≤ sup = sup ≤ kakkukV ,
wh ∈Wh kwh kW wh ∈Wh kwh kW
where we used that a(u − vh , wh ) = 0 for all wh ∈ Wh (Galerkin orthogonality property) and the
boundedness of the sesquilinear form a on V ×W. Since kvkL ≤ ιL,V kvkV for all v ∈ V, we infer
that
cV kvh kV ≤ cL kvh − ukL + (cL ιL,V + kak)kukV .
We now establish an upper bound on kvh − ukL . Let z ∈ W solve a(v, z) = (v, u − vh )L for all v in
V. The Galerkin orthogonality property implies that ku − vh k2L = a(u − vh , z) = a(u − vh , z − zh )
for all zh ∈ Wh . Hence, we have
ku − vh k2L ≤ kakku − vh kV ca hs kzkWs ≤ kakku − vh kV capp csmo hs ku − vh kL ,
so that ku − vh kL ≤ kakcappcsmo hs ku − vh kV . This in turn implies that
cV kvh kV ≤ cL kvh − ukL + (cL ιL,V + kak)kukV

≤ cL kakcapp csmo hs ku − vh kV + (cL ιL,V + kak)kukV .
Using the triangle inequality gives

(cV − cL kakcappcsmo hs )kvh kV ≤ kak + cL ιL,V + cL kakcappcsmo hs kukV .
Provided h ≤ ℓ0 (κ) we obtain cL kakcappcsmo hs ≤ 21 cV , so that

cV
kvh kV ≤ kukV .
2(kak + cL ιL,V + 21 cV )
This concludes the proof.
Both Lemma 35.14 and Lemma 35.16 imply that there is ℓ0 (κ) such that, if h ∈ H ∩ (0, ℓ0 (κ)],
the discrete inf-sup condition (35.23) holds true with a constant that is uniform with respect to
the meshsize but may depend on κ. To emphasize this dependency, let us write this constant as
α0 (κ). We can now invoke Babuška’s lemma (Lemma 26.14) to infer a quasi-optimal bound on the
approximation error.
Corollary 35.17 (Error estimate). There is ℓ0 (κ) s.t. the following quasi-optimal error estimate
holds true for all h ∈ H ∩ (0, ℓ0 (κ)]:

kak
ku − uh kV ≤ 1 + inf ku − vh kV . (35.24)
α0 (κ) vh ∈Vh
Example 35.18 (Dependence on κ). In order to illustrate the above results, let us assume that
we impose Robin boundary conditions with the norm kvkV := k∇vkL2 (D) + κkvkL2 (D) . Let us also
assume that full elliptic regularity holds true, i.e., the conclusion of Theorem 35.9 is fulfilled with
s := 1. Then α(κ) ∼ (κℓD )−1 for all κ ≥ κ0 . Moreover, we have capp ∼ 1, s := 1, ιL,V ∼ κℓD
in Lemma 35.14, so that ℓ0 (κ) ∼ ℓ−1 D κ
−2
κℓD (ℓD κ)−1 = κ−2 ℓ−1
D , and α0 (κ) ∼ (κℓD )
−1
. The error
estimate (35.24) gives ku − uh kV ≤ (1 + κℓD ) inf vh ∈Vh ku − vh kV . Let us now use Lemma 35.16
with kak ∼ 1, cV := 1, cL := κ, ιL,V := κ−1 , capp ∼ 1, s := 1. In this case, it can be shown that
csmo ∼ κℓD . Then we have again ℓ0 (κ) ∼ cV c−1 L kak ca csmo ∼ κ−2 ℓ−1
−1 −1 −1
D and α0 (κ) ∼ (κℓD )
−1
leading to the same error estimate.

Remark 35.19 (Literature). The reader is referred to Ihlenburg and Babuška [251] for an
exhaustive analysis of the one-dimensional Helmholtz problem with mixed boundary conditions and
its Galerkin approximation in one dimension with g := 0. In particular, the following statements
are proved therein: (i) For piecewise linear continuous finite elements on a uniform mesh, αh scales
exactly like (κℓD )−1 uniformly in h ∈ H, i.e., the discrete problem is well-posed for all h ∈ H
(see [251, Thm. 4]); (ii) The P1 Galerkin method delivers a quasi-optimal error estimate in the
H 1 -seminorm with a constant proportional to κℓD if κh < 1 < κℓD (see [251, Cor. 2]).
Remark 35.20 (Dispersion error). It is shown in [251, Thm. 5] that k∇(u − uh )kL2 (D) ≤
ℓD (hκ/π)(1 + chκ2 ℓD )kf kL2 , where c is independent of h ∈ H and κ ≥ 0. The term proportional
to hκ2 ℓD is usually called pollution error or dispersion error. This term grows unboundedly when
κ grows even if hκ < 1. The question whether the pollution error could be reduced or eliminated
by using stabilization techniques (i.e., discontinuous approximation techniques or methods similar
to those presented in Chapters 57–60) has been extensively addressed in the literature. We refer
the reader to Burman et al. [102], Feng and Wu [200], Melenk and Sauter [300], Peterseim [325],
and the literature therein for more details. For instance, it is shown in [300, Thm. 5.8] that under
some appropriate assumptions the pollution effect can be suppressed if one assumes that κh/k is
sufficiently small and that the polynomial degree k is at least O(ln(κ)). It is shown in [102, Thm. 6]
that the pollution error disappears in one dimension for some specific κ-dependent choices of the
penalty parameter of the CIP method (see §58.3 for details on CIP). The pollution error is also
shown to disappear in [325, Thm. 6.2] for a localized Petrov-Galerkin method where the global
shape functions each have a support of size rh with the oversampling condition r & ln(κℓD ).
Exercises
Exercise 35.1 (1D Helmholtz, well-posedness). Let D := (0, ℓD ), κ > 0, and consider the
Helmholtz problem with mixed boundary conditions: −∂xx u − κ2 u = f in D, u(0) = 0, and
∂x u(ℓD ) − iκu(ℓD ) = 0. (i) Give a weak formulation in V := {v ∈ H 1 (D) | v(0) = 0}. (ii) Show
by invoking an ODE argument that if the weak formulation has a solution, then it is unique. (iii)
Show that the weak problem is well-posed. (Hint : use Lemma 35.3.)
Exercise 35.2 (Green’s function, 1D). Let G : D×D → C be the function defined by
(
iκs
−1 sin(κx)e if x ∈ [0, s],
G(x, s) := κ iκx
sin(κs)e if x ∈ [s, 1].
(i) Prove that for all x ∈ D, the function D ∋ s 7→ G(x, s) ∈ C solves the PDE −∂ss u − κ2 u = δs=x
in D with the boundary conditions u(0) = 0 and ∂s u(ℓD ) − iκu(ℓD ) = 0 (i.e., G is the Green’s
function of the Helmholtz problem from Exercise 35.1). (ii) Find H(x, s) s.t. ∂s H(x, s) = ∂x G(x, s).
Rℓ
(iii) Let u(x) := 0 D G(x, s)f (s) ds. Prove that kukL2(D) ≤ κ−1 kf kL2(D) , |u|H 1 (D) ≤ kf kL2(D) ,
Rℓ
and |u|H 2 (D) ≤ (κ + 1)kf kL2(D) . (iv) Let v ∈ L2 (D) and let ze(x) := κ2 0 D G(x, s)v(s) ds. What
Rℓ
is the PDE solved by ze? Same question for z(x) := κ2 0 D G(x, s)v(s) ds. Note: The function z is
invoked in Step (1) of the proof of Theorem 35.11. (v) Assume now that v ∈ H 1 (D) with v(0) = 0,
and let z and ze be defined as above. Prove that max(|z|H 1 (D) , |ez |H 1 (D) ) ≤ 4κℓD |v|H 1 (D) . (Hint :
see Ihlenburg and Babuška [251, p. 14] (up to the factor 4).)
Exercise 35.3 (Variation on Fortin’s lemma). Let V, W be two Banach spaces and let a
be a bounded sesquilinear form on V ×W like in Fortin’s Lemma 26.9. Let (Vh )h∈H , (Wh )h∈H be
sequences of subspaces of V and W equipped with the norm of V and W, respectively. Assume that
there exists a map Πh : W → Wh and constants γΠh > 0, c(h) > 0 such that |a(vh , w − Πh (w))| ≤
c(h)kvh kV kwkW , γΠh kΠh (w)kW ≤ kwkW for all vh ∈ Vh , all w ∈ W, and all h ∈ H. Assume
that limh→0 c(h) = 0. Prove that the discrete inf-sup condition (26.5a) holds true for h ∈ H small
enough.
Exercise 35.4 (Lemma 35.8). (i) Prove that ℜ((m·∇v)v) = 21 m·∇|v|2 for all v ∈ H 1 (D; C) and
m ∈ Rd . (ii) Prove that ℜ(m·((∇v)T v)) = 21 m·∇kvk2ℓ2 (Cd ) for all v ∈ H 1 (D; Cd ) and m ∈ Rd .
(iii) Let q ∈ H 2 (D; C) and let D2 q denote the Hessian matrix of q, i.e., (D2 q)ij = ∂x2i xj q for all
i, j ∈ {1:d}. Show that ℜ(m·((D2 q)∇q)) = 21 m·∇k∇qk2ℓ2 (Cd ) . (iv) Prove that (35.11) holds true
for all q ∈ {v ∈ H 1 (D; C) | ∆v ∈ L2 (D; C), ∇v ∈ L2 (∂D; Cd )} and all m ∈ W 1,∞ (D; Rd ). (Hint :
assume first that q ∈ H 2 (D; C).)
Chapter 36
Crouzeix–Raviart approximation
In Part VIII, composed of Chapters 36 to 41, we study various nonconforming approximations

of an elliptic model problem. We first study the Poisson equation with a homogeneous Dirichlet
condition and then address a diffusion PDE with contrasted coefficients. Nonconformity means that
the discrete trial and test spaces are not subspaces of H 1 (D). Nonconformity has many sources.
It may be that the discrete shape functions have nonzero jumps across the mesh interfaces. It
may be that the Dirichlet conditions are enforced weakly. Another possible reason is that the
approximation involves discrete unknowns associated with the mesh faces as in hybrid methods.
All of these situations are studied in the following chapters. The objective of the present chapter
is to study the nonconforming approximation of the Poisson equation by Crouzeix–Raviart finite
elements. Another objective is to illustrate the abstract error analysis of Chapter 27.
36.1 Model problem

Let D be a Lipschitz domain in Rd . We assume for simplicity that D is a polyhedron. We focus
on the Poisson equation with homogeneous Dirichlet boundary conditions:
−∆u = f in D, u = 0 on ∂D, (36.1)
with source term f ∈ L2 (D). The weak formulation is as follows:

(36.2)
a(u, w) = ℓ(w), ∀w ∈ V,
with Z Z
a(v, w) := ∇v·∇w dx, ℓ(w) := f w dx. (36.3)
D D
Owing to the Poincaré–Steklov inequality (see (3.11) with p := 2), there is Cps > 0 such that
Cps kvkL2 (D) ≤ ℓD k∇vkL2 (D) for all v ∈ V, where ℓD is a length scale associated with D, e.g.,
ℓD := diam(D). Hence, V equipped with the norm kvkV := k∇vkL2 (D) = |v|H 1 (D) is a Hilbert
space, and the bilinear form a coincides with the inner product in V. Owing to the Lax–Milgram
lemma, (36.2) is well-posed. We refer the reader to §41.2 for the more general PDE −∇·(λ∇u) = f
with contrasted diffusivity λ.
146 Chapter 36. Crouzeix–Raviart approximation
36.2 Crouzeix–Raviart discretization

In this section, we recall Crouzeix–Raviart finite element, we define the corresponding approxima-
tion space, we formulate the discrete problem, and we establish its well-posedness. We also derive
some important stability estimates for Crouzeix–Raviart finite elements.
36.2.1 Crouzeix–Raviart finite elements

The Crouzeix–Raviart finite element is introduced in §7.5; see [151] for the original work to approx-
imate the Stokes equations. Let K b be the unit simplex in Rd with vertices {b
zi }i∈{0: d} . Let Fbi be
b opposite to zbi . The Crouzeix–Raviart finite element is defined by setting Pb := P1,d
the face of K
and by using the following degrees of freedom (dofs) on Pb:
Z
1
bicr (b
σ p) := pb ds, ∀i ∈ {0:d}. (36.4)
|Fbi | Fbi
Let (Th )h∈H be a shape-regular matching mesh sequence composed of affine simplices so that
each mesh covers D exactly. Let Th be a mesh and let K be a cell in Th . Using the Crouzeix–
Raviart element as reference finite element and letting the transformation ψK be the pullback by
the geometric mapping, i.e., ψK (v) := v ◦ TK , Proposition 9.2 allows us to generate a Crouzeix–
−1 b −1
Raviart finite element in K. We have PK := ψK (P ) = P1,d ◦ TK = P1,d since TK is affine, and
the local dofs in K are for all p ∈ PK ,
Z Z
cr 1 1
σK,i bicr (ψK (p)) =
(p) := σ p ◦ TK db
s= p ds, (36.5)
|Fbi | Fbi |FK,i | FK,i
for all i ∈ {1:d}, where {FK,i := TK (Fbi )}i∈{0: d} are thePfaces of K. The local interpolation operator
IKcr
: V (K) := W 1,1 (K) → PK is such that IK cr cr
(v) := i∈{0: d} σK,i cr
(v)θK,i for all v ∈ V (K), where
cr cr
{θK,i }i∈{0: d} are the local shape functions in K s.t. σK,i (θK,j ) = δij for all i, j ∈ {0:d}. Recall
that θicr := 1 − dλi , where {λi }i∈{0: d} are the barycentric coordinates in K.
Lemma 36.1 (Local interpolation). There is c s.t. for all r ∈ [0, 1], all p ∈ [1, ∞], all v ∈
W 1+r,p (K), all K ∈ Th , and all h ∈ H,
cr
kv − IK cr
(v)kLp (K) + hK |v − IK (v)|W 1,p (K) ≤ c h1+r
K |v|W 1+r,p (K) . (36.6)
Proof. Let v ∈ W 1+r,p (K). The error estimates for r ∈ {0, 1} follow from Theorem 11.13 with
k := 1 and l := 1 since V (K) := W 1,1 (K). For r ∈ (0, 1), we use Corollary 12.13, the W 1,p -stability
cr cr
of IK , and the fact that PK := P1,d is pointwise invariant under IK to infer that
cr cr
|v − IK (v))|W 1,p (K) ≤ inf |v − p − IK (v − p))|W 1,p (K)
p∈P1,d
≤ c inf |v − p|W 1,p (K) ≤ c′ hrK |v|W 1+r,p (K) .

p∈P1,d
cr cr
The bound on kv − IK (v)kLp (K) follows by proceeding similarly and using that kIK (v)kLp (K) ≤
kvkLp(K) + chK |v|W 1,p (K) .
36.2.2 Crouzeix–Raviart finite element space

Consider the broken finite element space defined in (18.4) with k := 1,
P1b (Th ) := {vh ∈ L∞ (D) | vh|K ∈ P1,d , ∀K ∈ Th }.

Part VIII. Elliptic PDEs: nonconforming approximation 147
Recall that the set Fh◦ is the collection of the interior faces (interfaces) in the mesh, and the
faces are oriented by the unit normal vector nF (see Chapter 10 on mesh orientation). For all
F ∈ Fh◦ , there are two cells Kl , Kr s.t. F := ∂Kl ∩ ∂Kr and nF points from Kl to Kr , i.e.,
nF := nKl = −nKr . The notion of jump across F is defined by setting [[v]]F := v|Kl − v|Kr . It is
convenient to use a common notation for interfaces and boundary faces by writing [[v]]F := v|Kl for
every boundary face F := ∂Kl ∩ ∂D ∈ Fh∂ . The Crouzeix–Raviart finite element space is defined
as Z
P1cr (Th ) := {vh ∈ P1b (Th ) | [[vh ]]F ds = 0, ∀F ∈ Fh◦ }. (36.7)
F
R
The condition F [[vh ]]F ds = 0 is equivalent to the continuity of vh at the barycenter xF of F .
Note that P1cr (Th ) is not H 1 -conforming since membership in H 1 (D) requires having zero-jumps
pointwise (see Theorem 18.8).
Let F ∈ Fh be a mesh face. Let us denote by TF := {K ∈ Th | F ∈ FK } the collection of the
mesh cells having F as face (TF contains two cells for F ∈ Fh◦ and one cell for F ∈ Fh∂ ). Let ϕcr F
be the function such that ϕcr F |K is the local shape function in K associated with FSif K ∈ TF and
ϕcr
F |K
:= 0 otherwise; see Figure 36.1 for d = 2. Note that supp(ϕcr F ) = DF := int( K∈TF K), i.e.,
DF is the collection of all the points in the (one orR two) mesh cells containing F . Let γFcr be the
linear form on P1cr (Th ) such that γFcr (vh ) := |F |−1 F vh ds for all cr
R vh ∈ P1 (Th ). Although vh may
cr
be multivalued at F , the quantity γF (vh ) is well defined since F [[vh ]]F ds = 0.
Figure 36.1: Global shape function for the Crouzeix–Raviart finite element. The support is mate-
rialized by thick lines and the graph by thin lines. Bullets indicate the barycenter of the edges.
Proposition 36.2 (Global dofs). {ϕcr cr cr

F }F ∈Fh is a basis of P1 (Th ), and {γF }F ∈Fh is a basis of
L(P1cr (Th ); R).
Proof. ϕcr cr cr
F is a member of P1 (Th ) since ϕF is piecewise affine by construction and P its mean value
on a mesh face is 0 or 1. Consider now real numbers {αF }F ∈Fh s.t. the function w := F ∈Fh αF ϕcr F
′
vanishes identically. Observing that γFcr′ (ϕcrF ) = δF F ′ for all F, F ∈ Fh , where δF F ′ denotes
the Kronecker symbol, we infer that αF ′ = γFcr′ (w) = 0 for all F ′ ∈ Fh . Hence, P the functions
{ϕcr
F } F ∈F h
are linearly independent. Finally, let vh ∈ P1
cr
(Th ) and set wh := cr cr
F ∈Fh γF (vh )ϕF .
Then, vh|K and wh|K are in PK for all K ∈ Th , and σK,i (wh|K ) = σK,i (vh|K ) for all i ∈ {0:d}.
Unisolvence implies that vh|K = wh|K , so that vh = wh since K ∈ Th is arbitrary. This shows that
{ϕcr cr cr
F }F ∈Fh is a basis of P1 (Th ). By using similar arguments, it follows that {γF }F ∈Fh is a basis
cr
of L(P1 (Th ); R).
Proposition 36.2 implies that the dimension of P1cr (Th ) is equal to the number of faces (edges
in dimension two) in the mesh. Moreover, the global Crouzeix–Raviart interpolation operator acts
on every function v in W 1,1 (D) as follows: For all x ∈ D,

X X 1 Z
Ihcr (v)(x) := γFcr (v)ϕcr
F (x) = v ds ϕcr
F (x).
|F | F
F ∈Fh F ∈Fh
Since Ihcr (v)|K = IK

cr
(v|K ) for all K ∈ Th , the approximation results of Lemma 36.1 can be
rephrased in terms of Ihcr .

We account for the homogeneous Dirichlet boundary condition by considering the following sub-
space of P1cr (Th ): Z
n o
cr cr
P1,0 (Th ) := vh ∈ P1 (Th ) | vh ds = 0, ∀F ∈ Fh∂ , (36.8)
F
where Fh∂ is the collection of the mesh faces located at the boundary. By proceeding as in Propo-
sition 36.2, one can verify that {ϕcr cr cr
F }F ∈Fh is a basis of P1,0 (Th ), and {γF }F ∈Fh is a basis of
◦ ◦
cr cr
L(P1,0 (Th ); R). The dimension of P1,0 (Th ) is the number of internal faces (edges if d = 2) in the
mesh.
cr
The bilinear form a introduced in (36.3) is not well defined on P1,0 (Th ) since this space is not
1 cr
H -conforming. Since functions in P1,0 (Th ) are piecewise smooth, we can localize their gradient to
the mesh cells. To this purpose, we introduce the notion of broken gradient on the broken Sobolev
space W 1,p (Th ) with p ∈ [1, ∞]. Recall from Definition 18.1 that a function v ∈ W 1,p (Th ) is s.t.
∇(v|K ) ∈ Lp (K) for all K ∈ Th .
Definition 36.3 (Broken gradient). Let p ∈ [1, ∞]. The broken gradient operator ∇h :
W 1,p (Th ) → Lp (D) is defined by setting (∇h v)|K := ∇(v|K ) for all K ∈ Th .
A crucial consequence of Lemma 18.9 is that ∇h v = ∇v whenever v ∈ W 1,p (D). This property
will be often used for the solution to the model problem (36.2) since u ∈ H01 (D). We define the
following discrete bilinear and linear forms on Vh ×Vh and on Vh , respectively:
Z Z
ah (vh , wh ) := ∇h vh ·∇h wh dx, ℓh (wh ) := f wh dx, (36.9)
D D
and we consider the following discrete problem:

cr
Find uh ∈ Vh := P1,0 (Th ) such that
(36.10)
ah (uh , wh ) = ℓh (wh ), ∀wh ∈ Vh .
Lemma 36.4 (Coercivity, well-posedness). (i) The map

1
vh 7→ kvh kVh := ah (vh , vh ) 2 = k∇h vh kL2 (D) (36.11)
cr
is a norm on P1,0 (Th ). (ii) Equipping Vh with this norm, the bilinear form ah is coercive on Vh
with αh := 1. (iii) The discrete problem (36.10) is well-posed.
Proof. (i) The only nontrivial property is to prove that kvh kVh = 0 implies thatR vh = 0 for all
vh ∈ Vh . If kvh kVh = 0, then vh is piecewise constant. The additional property F R[[vh ]]F ds = 0 for
all F ∈ Fh◦ implies that vh is globally constant on D. That vh = 0 follows from F vh ds = 0 for
all F ∈ Fh∂ .
(ii)-(iii) Since k·kVh is a norm on Vh , coercivity follows from the definition of k·kVh , and well-
posedness follows from the Lax–Milgram lemma.
Remark 36.5 (Nonsmooth right-hand side). We observe that it is not clear how one should
account for a source term f in H −1 (D) in (36.10), since it is not clear how f would act on
(discrete) functions that are not in H01 (D). One possibility is to consider the discrete linear form
av av g
ℓh (wh ) := hf, Jh,0 (wh )iH −1 (D),H01 (D) where Jh,0 : P1b (Th ) → P1,0 (Th ) is the averaging operator
with boundary conditions introduced in §22.4.1. A general theory addressing this type of difficulty
is developed in Veeser and Zanotti [373].
36.2.4 Discrete Poincaré–Steklov inequality

g
On the H01 -conforming subspace P1,0 cr
(Th ) := P1,0 (Th ) ∩ H01 (D), the norm k·kVh defined in (36.11)
1
coincides with the H -seminorm. Owing to the Poincaré–Steklov inequality, we know that there is
g
Cps > 0 s.t. Cps kvh kL2 (D) ≤ ℓD k∇vh kL2 (D) = ℓD kvh kVh for all vh ∈ P1,0 (Th ). We now prove that
cr
a similar inequality is available on the larger space P1,0 (Th ).
cr
Lemma 36.6 (Discrete Poincaré–Steklov inequality). There is Cps > 0 s.t. for all vh ∈
cr
P1,0 (Th ) and all h ∈ H,
cr
Cps kvh kL2 (D) ≤ ℓD k∇h vh kL2 (D) . (36.12)
Proof. Let vh ∈ P1,0 cr

(Th ). Let φ ∈ H01 (D) solve ∆φ = vh and let σ := ∇φ. Then ∇·σ = vh .
Elliptic regularity implies that there is s > 12 such that φ ∈ H 1+s (D) (see Theorem 31.33) so that
σ ∈ H s (D). Moreover, there is γD > 0 such that γD (kσkL2 (D) + ℓsD |σ|H s (D) ) ≤ ℓD kvh kL2 (D) .
Integrating by parts cellwise, we infer that
Z X Z
2
kvh kL2 (D) = vh ∇·σ dx = vh|K ∇·σ dx
D K∈Th K
X Z X X Z
=− σ·∇(vh|K ) dx + σ·nK vh|K ds
K∈Th K K∈Th F ∈FK F
Z X X Z
=− σ·∇h vh dx + σ·nK vh|K ds =: T1 + T2 ,
D K∈Th F ∈FK F
where FK is the collection of the faces of K and nK the outward unit normal to K (observe that
σ is single-valued on F since σ ∈ H s (D) with s > 21 ). The Cauchy–Schwarz inequality implies
that
|T1 | ≤ kσkL2 (D) k∇h vh kL2 (D) .
R now T2 . If RF := ∂Kl ∩ ∂Kr is an interface, the integral over F appears twice in the sum.
Consider
Since F vh|Kl ds = F vh|Kr ds by definition of P1cr (Th ) and sinceR nKl = −nKr , we can subtract
from σ a constant function on F that we take equal to σ F := |F1 | F σ ds. The same conclusion is
R cr
valid for the boundary faces since F vh ds = 0 on such faces by definition of P1,0 (Th ). This leads
to
X X Z
T2 = (σ − σ F )·nK vh|K ds
K∈Th F ∈FK F
X X Z
= (σ − σ F )·nK (vh|K − v F ) ds,
K∈Th F ∈FK F
R
where the subtraction of the single-valued quantity v F := |F1 | F vh ds is justified as above. Ap-
plying Lemma 36.8 below to σ|K and to vh|K , using hK ≤ ℓD for all K ∈ Th , and invoking the
Cauchy–Schwarz inequality yields

X s− 1 1
|T2 | ≤ c hK 2 |σ|H s (K) hK
2
k∇(vh|K )kL2 (K)
K∈Th
X
≤ c ℓsD |σ|H s (K) k∇(vh|K )kL2 (K) ≤ c ℓsD |σ|H s (D) k∇h vh kL2 (D) ,
K∈Th
P
since K∈Th |σ|2H s (K) ≤ |σ|2H s (D) . Combining the above bounds on T1 and T2 , we infer that

kvh k2L2 (D) ≤ kσkL2 (D) + c ℓsD |σ|H s (D) k∇h vh kL2 (D) ,
and (36.12) follows from γD (kσkL2 (D) + ℓsD |σ|H s (D) ) ≤ ℓD kvh kL2 (D) .
Remark 36.7 (Literature). The above proof is adapted from Temam [363, Prop. 4.13]; see also
Croisille and Greff [150].
Lemma 36.8 (Poincaré–Steklov on faces). Let s ∈ ( 12 , 1]. There is c s.t.
s− 1
kψ − ψ F kL2 (F ) ≤ c hK 2 |ψ|H s (K) , (36.13)
R
for all ψ ∈ H s (K) with ψ F := |F1 | F ψ ds; all K ∈ Th , all F ∈ FK , and all h ∈ H (the constant c
grows unboundedly as s ↓ 12 ).
1
R
Proof. Let ψ̃ := ψ − |K| K ψ dx. With obvious notation, we have ψ − ψ F = ψ̃ − ψ̃ F . The triangle
inequality and the Cauchy–Schwarz inequality imply that kψ − ψ F kL2 (F ) ≤ 2kψ̃kL2 (F ) . Using the
trace inequality (12.17) yields
−1 s− 1
kψ − ψ F kL2 (F ) ≤ c(hK 2 kψ̃kL2 (K) + hK 2 |ψ̃|H s (K) ).
The expected bound follows from |ψ̃|H s (K) = |ψ|H s (K) and the Poincaré–Steklov inequality ((12.13)
if s = 1 or (12.14) if s ∈ ( 12 , 1)) on K, which gives kψ̃kL2 (K) ≤ chsK |ψ|H s (K) .
36.2.5 Bound on the jumps

cr
Bounding the jumps of functions in P1,0 (Th ) is useful in many situations. The following result will
be invoked in the next section.
cr
Lemma 36.9 (Bound on the jumps). There is c s.t. for all vh ∈ P1,0 (Th ) and all h ∈ H,
X
c−1 h−1 2
F k[[vh ]]kL2 (F ) ≤ inf k∇h (v − vh )k2L2 (D)
v∈H01 (D)
F ∈Fh
X
≤c h−1 2
F k[[vh ]]kL2 (F ) . (36.14)
F ∈Fh
R
cr
Proof. Let vh ∈ P1,0 (Th ). For all K ∈ Th , let us set H∗1 (K) := {φ ∈ H 1 (K) | K φ dx = 0} and
let FK be the collection of the faces of K. For all F ∈ FK , let ψK,F ∈ H∗1 (K) solve the local
Neumann problem:
Z Z
∇ψK,F ·∇φ dx = ǫK,F [[vh ]]F φ ds, ∀φ ∈ H∗1 (K), (36.15)
K F
R
where ǫK,F := nK ·nF = ±1. This problem is well-posed since F [[vh ]]F ds = 0 for all F ∈ Fh . Since
ψK,F ∈ H∗1 (K), the multiplicative trace inequality (12.17) (with s := 1 and p := 2) together with
1
the Poincaré–Steklov inequality (12.13) implies that kψK,F kL2 (F ) ≤ chK 2
k∇ψK,F kL2 (K) . Taking
φ := ψK,F as a test function in (36.15), we infer that
Z
k∇ψK,F k2L2 (K) = ǫK,F [[vh ]]F ψK,F ds ≤ k[[vh ]]kL2 (F ) kψK,F kL2 (F )
F
1
≤ c hK
2
k[[vh ]]F kL2 (F ) k∇ψK,F kL2 (K) .
Owing to the regularity of the mesh sequence, we infer that

1
k∇ψK,F kL2 (K) ≤ c hF2 k[[vh ]]F kL2 (F ) .
(1) Let us prove the first bound in (36.14). Let v ∈ H01 (D). Let cK be the mean value of the
function (vh − v) over K. The restriction of (vh − v − cK ) to K is in H∗1 (K). Let F ∈ Fh . Taking
φK := (vh − v)|K − cK as a test function in (36.15) and summing over K ∈ TF , we infer that
X Z X Z
∇ψK,F ·∇(vh − v)|K dx = ∇ψK,F ·∇φK dx
K∈TF K K∈TF K
X Z X Z
= ǫK,F [[vh ]]F φK ds = ǫK,F [[vh ]]F (vh|K − v − cK ) ds
K∈TF F K∈TF F
Z Z Z
= [[vh ]]F [[vh − v − cK ]]F ds = [[vh ]]F [[vh − v]]F ds = [[vh ]]2F ds,
F F F
R
where we used that F [[vh ]]F ds = 0 to eliminate cK and the fact that v ∈ H01 (D) to eliminate
[[v]]F . Using the Cauchy–Schwarz inequality and the above bound on k∇ψK,F kL2 (K) , we obtain
X
h−1 2
F k[[vh ]]F kL2 (F ) ≤ c k∇(v − vh|K )k2L2 (K) . (36.16)
K∈TF
Summing over F ∈ Fh leads to the first bound in (36.14).

(2) To prove the second bound in (36.14), we estimate the infimum over v ∈ H01 (D) by taking
g,av g,av g
v := Jh,0 (vh ) where Jh,0 : P1b (Th ) → P1,0 (Th ) ⊂ H01 (D) is the averaging operator with zero
trace introduced in §22.4.1. Then the second bound in (36.14) follows from Lemma 22.12 and the
regularity of the mesh sequence.
The bound (36.14) can be adapted to the case where vh ∈ P1cr (Th ), i.e., without any boundary
prescription. The summations over the mesh faces are then restricted to the mesh interfaces, and
the infimum is taken over the functions v in H 1 (D). The idea of introducing the local Neumann
problem (36.15) has been considered in Achdou et al. [4].
36.3 Error analysis

In this section, we first establish an error estimate by using the coercivity norm and the abstract
error estimate from Lemma 27.5. Then we derive an improved L2 -error estimate by adapting the
duality argument from §32.3.
36.3.1 Energy error estimate

We perform the error analysis under the assumption that the solution to the model problem (36.2)
is in H 1+r (D) with r > 21 , i.e., we set
1
Vs := H 1+r (D) ∩ H01 (D), r> . (36.17)
2
The assumption u ∈ Vs is reasonable in the setting of the Poisson equation with Dirichlet con-
ditions in a Lipschitz polyhedron since it is consistent with the elliptic regularity theory (see
Theorem 31.33). The important property of a function v ∈ Vs that we use here is that its normal
derivative nK ·∇v is meaningful in L2 (∂K) for all K ∈ Th . Actually, the full trace of ∇v on ∂K is
meaningful on L2 (∂K), and this trace is single-valued on any interface F ∈ Fh◦ (see Remark 18.4).
Therefore, we have [[∇v]]F = 0 for all v ∈ Vs and all F ∈ Fh◦ .
cr
The discrete space Vh := P1,0 (Th ) is equipped with the norm k·kVh defined in (36.11), and we
introduce the space V♯ := Vs + Vh equipped with the norm k·kV♯ defined by
X
kvk2V♯ := k∇vk2L2 (K) + hK knK ·∇v|K k2L2 (∂K) . (36.18)
K∈Th
A discrete trace inequality shows that there is c♯ s.t. kvh kV♯ ≤ c♯ kvh kVh for all vh ∈ Vh and all
h ∈ H, i.e., (27.5) holds true. Using the forms ah and ℓh defined in (36.9), the consistency error is
s.t.
hδh (vh ), wh iVh′ ,Vh := ℓh (wh ) − ah (vh , wh ), ∀vh , wh ∈ Vh . (36.19)
Lemma 36.10 (Consistency/boundedness). Assume (36.17). There is ω♯ , uniform w.r.t.
u ∈ Vs , s.t. for all vh ∈ Vh and all h ∈ H,
kδh (vh )kVh′ ≤ ω♯ ku − vh kV♯ . (36.20)
Proof. Let vh , wh ∈ Vh . Since the normal derivative nK ·∇u is meaningful in L2 (∂K) for all
K ∈ Th , we have
X Z X Z
ℓh (wh ) = f wh|K dx = −(∆u)wh|K dx
K∈Th K K∈Th K
X Z X X Z
= ∇u·∇wh|K dx − (nK ·∇u)wh|K ds
K∈Th K K∈Th F ∈FK F
Z X X Z
= ∇u·∇h wh dx − (nK ·∇u)wh|K ds.
D K∈Th F ∈FK F
Note that we write nK ·∇u instead of nK ·∇u|K since ∇u is single-valued on F because u ∈ Vs .

We want to exchange the order of the summations on the right-hand side. Recalling that for every
interface F := ∂Kl ∩ ∂Kr ∈ Fh◦ with nF pointing from Kl to Kr , i.e., nF := nKl = −nKr , we
have
(nKl ·∇u)wh|Kl + (nKr ·∇u)wh|Kr = (nKl ·∇u)[[wh ]]F .
For every boundary face F := ∂Kl ∩ ∂D ∈ Fh∂ , recall that we have conventionally set [[wh ]]F :=
wh|Kl . Thus, we infer that
X X Z X Z
(nK ·∇u)wh|K ds = (nKl ·∇u)[[wh ]]F ds.
K∈Th F ∈FK F F ∈Fh F
Setting η := u − vh , we can write the consistency error as follows:

Z X Z
hδh (vh ), wh iVh′ ,Vh = ∇h η·∇h wh dx − (nKl ·∇u)[[wh ]]F ds
D F ∈Fh F
Z X Z
= ∇h η·∇h wh dx − (nKl ·∇η|Kl )[[wh ]]F ds,
D F ∈Fh F
R
where we used that F (nKl ·∇vh|Kl )[[wh ]]F ds = 0 for all F ∈ Fh by definition of the Crouzeix–
cr
Raviart space Vh = P1,0 (Th ). We conclude by invoking the Cauchy–Schwarz inequality, the first
P
bound on the jumps in (36.14) which implies that F ∈Fh h−1 2 2
F k[[wh ]]F kL2 (F ) ≤ ckwh kVh (bound
the infimum by taking v := 0), and the regularity of the mesh sequence.
Theorem 36.11 (Convergence). Let u solve (36.2) and let uh solve (36.10). Assume (36.17).
(i) There is c s.t. the following quasi-optimal error estimate holds true for all h ∈ H,

vh ∈Vh
(ii) Letting t := min(1, r), we have

X 21
ku − uh kV♯ ≤ c h2t 2
K |u|H 1+t (K) . (36.22)
K∈Th
Proof. (i) The estimate (36.21) follows from Lemma 27.5 combined with stability (Lemma 36.4)
and consistency/boundedness (Lemma 36.10).
(ii) The bound (36.22) follows from (36.21) by taking vh := Ihcr (u). Letting η := u − Ihcr (u),
we indeed have k∇η|K kL2 (K) ≤ chtK |u|H 1+t (K) for all K ∈ Th owing to Lemma 36.1. Moreover,
invoking the multiplicative trace inequality (12.17), we obtain
1
hK
2
knK ·∇η|K kL2 (∂K) ≤ k∇η|K kL2 (K) + htK |η|K |H 1+t (K) ,
and we have |η|K |H 1+t (K) = |u|H 1+r (K) since Ihcr (u) is affine in K.
Remark 36.12 (Strang 2). The analysis can also be done by invoking Strang’s second lemma
(Lemma 27.15). Let us set V♯ := H01 (D) + P1,0
cr
(Th ) and let us equip this space with the norm k·kV♯
defined in (36.18). The discrete bilinear form ah can be extended to a bilinear form a♯ having
boundedness constant equal to 1 on V♯ ×Vh . Lemma 27.15 leads to the error bound

ku − vh kV♯ ≤ c inf ku − vh kV♯ + kδhst2 (u)kVh′ ,
vh ∈Vh
cr
with the consistency error s.t. for all wh ∈ P1,0 (Th ),
X Z
hδhst2 (u), wh iVh′ ,Vh := ℓh (wh ) − ah (u, wh ) = (f wh − ∇u·∇wh|K ) dx
K∈Th K
X Z
=− (nK ·∇u)wh|K ds.
K∈Th ∂K
Thus, the consistency error does not vanish identically, i.e., the Crouzeix–Raviart finite element
method is not strongly consistent in the sense defined in Remark 27.16. Since we have
X Z X Z
(nK ·∇u)wh|K ds = (nK ·∇(u − vh ))wh|K ds,
K∈Th ∂K K∈Th ∂K
cr
for all vh ∈ P1,0 (Th ), by proceeding as in the proof of Theorem 36.11, we infer again that the
quasi-optimal error estimate (36.21) holds true.
36.3.2 L2 -error estimate

The goal of this section is to derive an improved L2 -error estimate of the form ku − uh kL2 (D) ≤
chγ ℓ1−γ
D ku − uh kV♯ for some real number γ > 0, where ℓD is a length scale associated with D, e.g.,
ℓD := diam(D).
Proceeding as in §32.3, we invoke a duality argument. We consider for all g ∈ L2 (D) the adjoint
solution ζg ∈ V := H01 (D) such that
a(v, ζg ) = (v, g)L2 (D) , ∀v ∈ V. (36.23)
Notice that −∆ζg = g in D and γ g (ζg ) = 0. Owing to the elliptic regularity theory (see §31.4),
there is s ∈ (0, 1] and a constant csmo such that kζg kH 1+s (D) ≤ csmo ℓ2D kgkL2 (D) for all g ∈ L2 (D).
In the present setting of the Poisson equation with Dirichlet conditions in a Lipschitz polyhedron,
it is reasonable to assume that s ∈ ( 12 , 1].
Theorem 36.13 (L2 -estimate). Let u solve (36.2) and let uh solve (36.10). Assume that the
elliptic regularity index satisfies s ∈ ( 12 , 1]. There is c s.t. for all h ∈ H,

D ku − uh kV♯ . (36.24)
g
Proof. Let e := u − uh and set Yh := P1,0 cr
:= P1,0
(Th ) (Th ) ∩ H01 (D). Then (∇h e, ∇yh )L2 (D) =
(∇u, ∇yh )L2 (D) − (∇h uh , ∇yh )L2 (D) = 0 for all yh ∈ Yh . Since kek2L2(D) = −(e, ∆ζe )L2 (D) , we have

kek2L2 (D) = (∇h e, ∇ζe )L2 (D) − (e, ∆ζe )L2 (D) + (∇h e, ∇ζe )L2 (D)
= (∇h e, ∇(ζe − yh ))L2 (D) − hδ adj (ζe ), eiV♯′ ,V♯ ,
where we introduced δ adj (ζe ) ∈ V♯′ s.t. hδ adj (ζe ), viV♯′ ,V♯ := (v, ∆ζe )L2 (D) + (∇h v, ∇ζe )L2 (D) and
|hδ adj (ζe ),viV ′ ,V |
♯
used that (∇h e, ∇yh )L2 (D) = 0 for all yh ∈ Vh . Let us set kδ adj (ζe )kV♯′ := supv∈V♯ kvkV♯
♯
.
The Cauchy–Schwarz inequality and the definition of the k·kV♯ - and k·kV♯′ -norms imply that

kek2L2(D) ≤ inf k∇(ζe − yh )kL2 (D) + kδ adj
(ζe )kV♯′ kekV♯ .
yh ∈Yh
It remains to bound the two terms between parentheses on the right-hand side. Using the quasi-
g,av
interpolation operator Ih0 from §22.4, we infer that
g,av
inf k∇(ζe − yh )kL2 (D) ≤ k∇(ζe − Ih0 (ζe ))kL2 (D)
yh ∈Yh
≤ c hs |ζe |H 1+s (D) ≤ c hs ℓ−1−s

D kζe kH 1+s (D) ≤ c csmo hs ℓ1−s
D kekL2 (D) ,
g,av
where we used the approximation properties of Ih0 from Theorem 22.14 and the elliptic regularity
theory to bound kζe kH 1+s (D) by kekL2(D) . Let us now estimate kδ adj (ζe )kV♯′ . By proceeding as in
the proof of Lemma 36.10 (observe that [[∇ζe ]]F = 0 for all F ∈ Fh◦ ), we infer that we have, for all
v := vs + vh ∈ V♯ := Vs + Vh with vs ∈ Vs and vh ∈ Vh , and all zh ∈ Vh ,
X Z
hδ adj (ζe ), viV♯′ ,V♯ = nKl ·∇ζe [[vh ]]F ds
F ∈Fh F
X Z
= nKl ·∇(ζe − zh )|Kl [[vh ]]F ds
F ∈Fh F
X 21
≤ c kζe − zh kV♯ h−1 2
F k[[vh ]]F kL2 (F ) ,
F ∈Fh
where we used that nKl ·∇zh|Kl is constant on F . Using the leftmost inequality in (36.14) with
P
inf w∈H01 (D) k∇h (w − vh )k2L2 (D) ≤ k∇h (vs + vh )k2L2 (D) , we infer that F ∈Fh h−1 2
F k[[vh ]]F kL2 (F ) ≤
2 2 ′
ckvs + vh kV♯ = ckvkV♯ . Thus, kδ (ζe )kV♯′ ≤ c inf zh ∈Vh kζe − zh kV♯ . Using the approximation
adj
properties of Vh , we conclude that kδ adj (ζe )kV♯′ ≤ chs |ζe |H 1+s (D) , and reasoning as above yields
kδ adj (ζe )kV♯′ ≤ chs ℓ1−s
D kekL2 (D) .
36.3.3 Abstract nonconforming duality argument

Let us finish with an abstract formulation of the above duality argument that can be applied in the
context of nonconforming approximation techniques. Let V a Banach space, L be a Hilbert space,
and assume that V embeds continuously into L (i.e., V ֒→ L) and V is dense in L. Identifying L
with L′ , we are in the situation where
V ֒→ L ≡ L′ ֒→ V ′ , (36.25)
with continuous and dense embeddings. Let a : V ×V → C be a bounded sesquilinear form

satisfying the assumptions of the BNB theorem (Theorem 25.9). For all f ∈ L we denote by ξf
the unique solution to the problem
a(ξf , v) = (f, v)L , ∀v ∈ V. (36.26)
Similarly, for all g ∈ L we denote by ζg ∈ V the unique solution to the adjoint problem
a(v, ζg ) = (v, g)L , ∀v ∈ V. (36.27)
These two problems are well-posed since a satisfies the assumptions of the BNB theorem. Let
Aadj ∈ L(V ; V ′ ) be s.t. hAadj (w), viV ′ ,V = a(v, w) for all (v, w) ∈ V ×V. Owing to (36.25) and
(36.27), we have Aadj (ζg ) = g in L.
We assume that we have at hand two subspaces Vs ⊂ V and Zs ⊂ V s.t. the maps V ′ ∋ f 7→
ξf ∈ Vs and V ′ ∋ g 7→ ζg ∈ Zs are bounded. Let Vh ⊂ L be a finite-dimensional subspace of L
(but not necessarily of V ). Let Yh ⊆ Vh . We set V♯ := Vs + Vh and Z♯ := Zs + Yh , and we equip
these spaces with norms denoted by k·kV♯ and k·kZ♯ .
Lemma 36.14 (L-norm estimate). Let a♯ be a bounded sesquilinear form on V♯ ×Z♯ . Let ka♯ k
be the norm of a♯ on V♯ ×Z♯ . Let u ∈ Vs and uh ∈ Vh . Assume that the following Galerkin
orthogonality property holds true:
a♯ (u − uh , yh ) = 0, ∀yh ∈ Yh . (36.28)
Let e := u − uh and let δ adj (ζe ) ∈ V♯′ be the adjoint consistency error:
hδ adj (ζe ), viV♯′ ,V♯ := (v, Aadj (ζe ))L − a♯ (v, ζe ), ∀v ∈ V♯ . (36.29)
Then the following estimate holds true:

kδ adj (ζe )kV♯′ kζe − yh kZ♯
kekL ≤ + ka♯ k inf kekV♯ , (36.30)
kekL yh ∈Yh kekL
Proof. Using the identity Aadj (ζe ) = e and the Galerkin orthogonality property (36.28), we infer
that
kek2L = (e, Aadj (ζe ))L = (e, Aadj (ζe ))L − a♯ (e, ζe ) + a♯ (e, ζe )
= hδ adj (ζe ), eiV♯′ ,V♯ + a♯ (e, ζe − yh ).
The boundedness of a♯ on V♯ ×Z♯ and the definition of the dual norm kδ adj (ζe )kV♯′ imply that (36.30)
holds true.
Example 36.15 (Crouzeix–Raviart). Lemma 36.14 can be applied to the Crouzeix–Raviart ap-
proximation with Vs := H 1+r (D) ∩ H01 (D), Zs := H 1+s (D) ∩ H01 (D), a♯ (v, w) := (∇h v, ∇h w)L2 (D) ,
and equipping the spaces V♯ := Vs + Vh , Z♯ := Zs + Yh , Yh := Vh ∩ H01 (D), with the broken
energy norm. Note that the adjoint consistency error is nonzero, and that the proof of Theo-
rem 36.13 shows that both terms on the right-hand side of (36.30) converge with the same rate
w.r.t. h ∈ H.
Exercises
Exercise 36.1 (Commuting properties). Let K be a simplex in Rd and let Π0K denote the
L2 -orthogonal projection onto constants. Prove that ∇(IK cr
(p)) = Π0K (∇p) and ∇·(II cr
K (σ)) =
Π0K (∇·σ) for all p ∈ H 1 (K) and all σ ∈ L2 (K) with ∇·σ ∈ L1 (K) and I cr
K defined componentwise
using Ihcr .
Exercise 36.2 (Best approximation). Let v ∈ H 1 (D). A global best-approximation of v in

P1cr (Th ) in the broken H 1 -seminorm is a function vhcr ∈ P1cr (Th ) s.t.
X X
k∇(v − vhcr )k2L2 (K) = min
cr
k∇(v − vh )k2L2 (K) .
vh ∈P1 (Th )
K∈Th K∈Th
(i) Write a characterization of vhcr in weak form and show that vhcr is unique up to an addi-
tive constant. (Hint : adapt Proposition 25.8.) (ii) Let vhb be a global
P best-approximation of v
in the broken finite element space P1b (Th ); see §32.2. Prove that K∈Th k∇(v − vhcr )k2L2 (K) =
P b 2 cr cr
K∈Th k∇(v − vh )kL2 (K) . (Hint : using Exercise 36.1, show that vh = Ih (v) up to an additive
constant.)
Exercise 36.3 (H(div)-flux recovery). Let uh solve (36.10). Assume that f is piecewise con-
stant on Th . Set σh|K := −∇uh|K + d1 f|K (x − xK ), where xK is the barycenter of K for all
K ∈ Th . Prove that σh is in the lowest-order Raviart–Thomas
R finite element space P0d (Th ) and
that ∇·σ = f ; see Marini [295] (Hint : evaluate F [[σh ]]·nF ϕF ds for all F ∈ Fh◦ .)
cr
Exercise 36.4 (Discrete Helmholtz). Let D ⊂ R2 be a simply connected polygon. Prove that
P0b (Th ) = ∇P1g (Th ) ⊕ ∇h⊥ P1,0
cr
(Th ), where
∇h⊥ P1,0
cr
(Th ) := {vh ∈ P0b (Th ) | ∃qh ∈ P1,0
cr
(Th ) | vh|K = ∇⊥ (qh|K ), ∀K ∈ Th },
and ∇⊥ is the two-dimensional curl operator defined in Remark 16.17. (Hint : prove that the
decomposition is L2 -orthogonal and use a dimension argument based on Euler’s relations.)
Exercise 36.5 (Rannacher–Turek). Let K := [−1, 1]d . For all i ∈ {1:d} and α ∈ {l, r}, let
Fi,α be the face of K corresponding to {xi = −1} when α = l and to {xi = 1} when α = r.
Observe that there are 2d such faces, each of measure 2d−1 . Let P be spanned by the R 2d functions
{1, x1 , . . . , xd , x21 − x22 , . . . , x2d−1 − x2d }. Consider the linear forms σi,α (p) := 21−d Fi,α p ds for all
i ∈ {1:d} and α ∈ {l, r}. Setting Σ := {σi,α }i∈{1: d},α∈{l,r} , prove that (K, P, Σ) is a finite element.
Note: this element has been introduced by [330] for the mixed discretization of the Stokes equations
on Cartesian grids.
Exercise 36.6 (Quadratic space). Let Th be a triangulation of a simply connected domain
D ⊂ R2 and let
Z
cr b
P2 (Th ) := {vh ∈ P2 (Th ) | [[vh ]]F (q ◦ TF−1 ) ds = 0, ∀F ∈ Fh◦ , ∀q ∈ P1,1 },
F
where TF is an affine bijective mapping from the unit segment Sb1 = [−1, 1] to F . Orient all √ the
±
faces F ∈ Fh and define the two Gauss points gF on F that are the image by TF of b g ± :=
± 33 ,
− +
in such a way that the orientation of F goes from gF to gF . For all K ∈ Th , let {λ0,K , λ1,K , λ2,K }
be the barycentric coordinates in K and set bK := 2 − 3(λ20,K + λ21,K + λ22,K ) (this function is
usually called Fortin–Soulié bubble [204]). One can verify that a polynomial p ∈ P2,2 vanishes
at the six points {gF± }F ∈FK if and only if p = αbK for some α ∈ R. Note: this shows that
these six points, which lie on an ellipse, cannot be taken as nodes of a P2,2 Lagrange element.
(i) Extending bK Rby zero outside K, verify that bK ∈ P2cr (Th ). (ii) Set B := spanK∈Th {bK } and
B∗ := {vh ∈ B | D vh dx = 0}. Prove that P2g (Th ) + B∗ ⊂ P2cr (Th ) and that P2g (Th ) ∩ B∗ = {0}.
(iii) Define J : P2cr (Th ) → R2Nf s.t. J(vh ) := (vh (gF− ), vh (gF+ ))F ∈Fh for all vh ∈ P2cr (Th ). Prove
that dim(ker(J)) = Nc and dim(im(J)) ≤ 2Nf − Nc . (Hint : any polynomial p ∈ P2,2 satisfies
P + − cr g
F ∈FK (p(gF ) − p(gF )) = 0 for all K ∈ Th .) (iv) Prove that P2 (Th ) = P2 (Th ) ⊕ B∗ ; see Greff
[222]. (Hint : use a dimensional argument and Euler’s relation from Remark 8.13.)
Chapter 37
Nitsche’s boundary penalty

method
The main objective of this chapter is to present a technique to treat Dirichlet boundary conditions
in a natural way using a penalty method. This technique is powerful and has many extensions.
In particular, the idea is reused in the next chapter for discontinuous Galerkin methods. Another
objective of this chapter is to illustrate the abstract error analysis of Chapter 27.
37.1 Main ideas and discrete problem

Let D be a Lipschitz domain in Rd . We assume for simplicity that D is a polyhedron. Let
1
f ∈ L2 (D) be the source term, and let g ∈ H 2 (∂D) be the Dirichlet boundary data. We consider
the Poisson equation with Dirichlet conditions
−∆u = f in D, γ g (u) = g on ∂D, (37.1)
1
where γ g : H 1 (D) → H (∂D) is the trace map. Let ug ∈ H 1 (D) be a lifting of g, i.e., γ g (ug ) = g
2
1
(recall that γ g : H 1 (D) → H 2 (∂D) is the trace map). We seek u0 ∈ H01 (D) s.t. a(u0 , w) =
1
ℓ(w) − a(ug , w) for all w ∈ H0 (D), with
Z Z
a(v, w) := ∇v·∇w dx, ℓ(w) := f w dx. (37.2)
D D
This problem is well-posed in H01 (D)owing to the Lax–Milgram lemma and the Poincaré–Steklov
inequality in H01 (D). Then the unique weak solution to (37.1) is u := u0 + ug (see §31.2.2).
In this chapter, we take a route that is different from the above approach to construct an
approximation of the solution. Instead of enforcing the Dirichlet boundary condition strongly,
we are going to construct an H 1 -conforming discretization of (37.1) that enforces this condition
naturally. This means that we no longer require that the discrete test functions vanish at the
boundary. The discrete counterpart of the bilinear form a must then be modified accordingly. To
motivate the modification in question, let us proceed informally by assuming that all the functions
we manipulate are sufficiently smooth. Testing (37.1) with a function w which we do not require
to vanish at the boundary, the integration by parts formula (4.8b) gives
Z
a(u, w) − (n·∇u)w ds = ℓ(w). (37.3)
∂D
160 Chapter 37. Nitsche’s boundary penalty method
R
The idea of Nitsche is to modify (37.3) by adding a term proportional to ∂D uw ds on both sides
of the above identity. This leads to
Z Z Z
a(u, w) − (n·∇u)w ds + ̟ uw ds = ℓ(w) + ̟ gw ds, (37.4)
∂D ∂D ∂D
where the boundary value of u has been replaced by g in the boundary integral on the right-hand
side. The yet unspecified parameter ̟ is assumed to be positive. Heuristically, if u satisfies (37.4)
and if ̟ is large, one expects u to be close to g at the boundary. For this reason, ̟ is called
penalty parameter.
The above ideas lead to an approximation method employing discrete trial and test spaces
composed of functions that are not required to vanish at the boundary. Let Th be a mesh from a
shape-regular sequence of meshes so that each mesh covers D exactly. Let Fh∂ be the collection of
the boundary faces. Let Pkg (Th ) be the H 1 -conforming finite element space of degree k ≥ 1 based
on Th ; see (19.10). We consider the following discrete problem:

Find uh ∈ Vh := Pkg (Th ) such that
(37.5)
ah (uh , wh ) = ℓh (wh ), ∀wh ∈ Vh ,
where the discrete forms ah and ℓh are inspired from (37.4):

Z X Z
ah (vh , wh ) := a(vh , wh ) − (n·∇vh )wh ds + ̟(hF ) vh wh ds,
∂D F
F ∈Fh∂
X Z
ℓh (wh ) := ℓ(wh ) + ̟(hF ) gwh ds.
F
F ∈Fh∂
The second term in the definition of ah is called consistency term (this term plays a key role
when estimating the consistency error) and the third one is called penalty term. The penalty
parameter ̟(hF ) > 0, yet to be defined, depends on the diameter of the face F (or a uniformly
equivalent local length scale). The stability analysis will reveal that ̟(hF ) should scale like h−1
F .
The approximation setting associated with Nitsche’s boundary penalty method is nonconforming,
i.e., Vh 6⊂ V := H01 (D), since functions in Vh may not vanish at the boundary, whereas functions
in V do.
Remark 37.1 (Literature, extensions). The boundary penalty method has been introduced
by Nitsche [314] to treat Dirichlet boundary conditions. It was extended in Juntunen and Stenberg
[262] to Robin boundary conditions. We refer the reader to §41.3 where the more general PDE
−∇·(λ∇u) = f with contrasted diffusivity λ is treated.
37.2 Stability and well-posedness

The main objective of this section is to prove that the discrete bilinear form ah is coercive on
Vh if the penalty parameter is large enough. This is done by showing that the consistency term
can be appropriately bounded. For all F ∈ Fh∂ , let us denote by Kl the unique mesh cell having
F as a face, i.e., F := ∂Kl ∩ ∂D.S Let Th∂D be the collection of the mesh cells having at least
one boundary face, i.e., Th∂D := F ∈F ∂ {Kl }. (The set Th∂D should not be confused with the
h
larger set Th∂ defined in (22.28), which is the collection of the mesh cells touching the boundary.)
Let n∂ denote the maximum number of boundary faces that a mesh cell in Th∂D can have, i.e.,
n∂ := maxK∈Th∂D card(FK ∩ Fh∂ ) (n∂ ≤ d for simplicial meshes). Owing to the regularity of the
mesh sequence, the discrete trace inequality from Lemma 12.8 (with p = q := 2) implies that there
is cdt such that for all vh ∈ Vh , all F ∈ Fh∂ , and all h ∈ H,
−1
kn·∇vh kL2 (F ) ≤ cdt hF 2 k∇vh kL2 (Kl ) . (37.6)
Lemma 37.2 (Bound on consistency term). The following holds true for all vh ∈ Vh :
Z X 12 X 12
1 1

(n·∇vh )vh ds ≤ n∂ cdt
2 2
k∇vh kL2 (K) 2
kvh kL2 (F ) .
hF
∂D ∂D ∂
K∈Th F ∈Fh
Proof. Let vh ∈ Vh . Let F ∈ Fh∂ . Using the Cauchy–Schwarz inequality, bounding the normal
component of the gradient by its Euclidean norm, and using the discrete trace inequality (37.6)
componentwise, we infer that
Z X 21 X 12
1
(n·∇vh )vh ds ≤ 2
hF kn·∇vh kL2 (F ) 2
kvh kL2 (F )
hF
∂D ∂ ∂
F ∈Fh F ∈Fh
X 21 X 12
1
≤ cdt k∇vh k2L2 (Kl ) kvh k2L2 (F ) .
hF
F ∈Fh∂ F ∈Fh∂
P P P
Finally, we have F ∈Fh∂ k·k2L2 (Kl ) = K∈Th∂D card(FK ∩Fh∂ )k·k2L2 (K) ≤ n∂ K∈Th∂D k·k2L2 (K) .
We equip the space Vh with the following norm:

X 1
kvh k2Vh := k∇vh k2L2 (D) + |vh |2∂ , |vh |2∂ := kvh k2L2 (F ) . (37.7)
hF
F ∈Fh∂
Note that kvh kVh = 0 implies that vh is constant on D and vanishes on ∂D, so that vh = 0. Hence,
k·kVh is a norm on Vh . Note also that the two terms composing the norm k·kVh are dimensionally
consistent.
Lemma 37.3 (Coercivity, well-posedness). Assume that the penalty parameter ̟(hF ) is de-
fined s.t.
1
̟(hF ) := ̟0 , ∀F ∈ Fh∂ , (37.8)
hF
with ̟0 > 41 n∂ c2dt . (i) The following coercivity property holds true:
ah (vh , vh ) ≥ αkvh k2Vh , ∀vh ∈ Vh , (37.9)
̟0 − 14 n∂ c2dt
with α := 1+̟0 > 0, (ii) The discrete problem (37.5) is well-posed.
Proof. Let vh ∈ Vh . We have

Z
ah (vh , vh ) = k∇vh k2L2 (D) − (n·∇vh )vh ds + ̟0 |vh |2∂ .
∂D
P 1 P 1
Setting z := ( K∈Th \T ∂D k∇vh k2L2 (K) ) 2 , x := ( K∈T ∂D k∇vh k2L2 (K) ) 2 , and y := |vh |∂ , and using
h h
Lemma 37.2, we infer that
1
ah (vh , vh ) ≥ z 2 + (x2 − n∂2 cdt xy + ̟0 y 2 ).
1
̟0 −β 2
Coercivity follows from the inequality x2 −2βxy+̟0 y 2 ≥ 2 2
1+̟0 (x +y ) applied with β := 21 n∂2 cdt
̟0 −β 2 ̟0
(see Exercise 37.2) and since 1+̟0 ≤ 1+̟0 ≤ 1. Finally, well-posedness follows from the Lax–
Milgram lemma.
Remark 37.4 (Choice of penalty parameter). Ensuring the stability condition ̟0 > 41 n∂ c2dt
requires in practice to know a reasonable upper bound on the constant cdt . The results of §12.2
show that cdt scales like the polynomial degree k. More precisely, Lemma 12.10 shows that for
1
simplices one can take cdt := ((k + 1)(k + d)/d) 2 with hF := |Kl |/|F |.
37.3 Error analysis

In this section, we derive an energy error estimate, that is, we bound the error by using the
coercivity norm and the abstract error estimate from Lemma 27.5. We also derive an improved
L2 -error estimate by means of a duality argument.
37.3.1 Energy error estimate

We perform the error analysis under the assumption that the solution to (37.1) is in H 1+r (D) with
r > 21 , i.e., we set
1
Vs := H 1+r (D), r > . (37.10)
2
Theorem 31.33). The important property that we use is that for any function v ∈ Vs , the normal
derivative n·∇v at the boundary is meaningful in L2 (∂D). We consider the space V♯ := Vs + Vh
equipped with the norm
X
kvk2V♯ := k∇vk2L2 (D) + |v|2∂ + hF kn·∇vk2L2 (F ) , (37.11)
F ∈Fh∂
P
with |v|2∂ := F ∈F ∂ h1F kvk2L2 (F ) . A discrete trace inequality shows that there is c♯ s.t. kvh kV♯ ≤
h
c♯ kvh kVh for all vh ∈ Vh and all h ∈ H, i.e., (27.5) holds true. Recall from Definition 27.3 that the
consistency error is defined by setting hδh (vh ), wh iVh′ ,Vh := ℓh (wh ) − ah (vh , wh ) for all vh , wh ∈ Vh .
Lemma 37.5 (Consistency/boundedness). Assume (37.10). There is ω♯ , uniform w.r.t. u ∈

Vs , s.t. for all vh ∈ Vh and all h ∈ H,

Proof. Let vh , wh ∈ Vh . Since the normal derivative n·∇u is meaningful at the boundary, using
the PDE and the boundary condition in (37.1), we infer that
Z X Z
1
ℓh (wh ) = −(∆u)wh dx + ̟0 gwh ds
D hF F
F ∈Fh∂
Z Z X Z
1
= ∇u·∇wh dx − (n·∇u)wh ds + ̟0 uwh ds.
D ∂D ∂
hF F
F ∈Fh
Letting η := u − vh , this implies that

Z Z X Z
1
hδh (vh ), wh iVh′ ,Vh = ∇η·∇wh dx − (n·∇η)wh ds + ̟0 ηwh ds.
D ∂D ∂
h F F
F ∈Fh
Using the Cauchy–Schwarz inequality, we obtain the estimate (37.12) with ω♯ := max(1, ̟0 ).
Theorem 37.6 (Convergence). Let u solve (37.1) and let uh solve (37.5) with the penalty
parameter ̟0 > 14 n∂ c2dt . Assume (37.10). (i) There is c s.t. the following quasi-optimal error
estimate holds true for all h ∈ H:

vh ∈Vh
(ii) Letting t := min(k, r), we have

X 21
K |u|H 1+t (K) . (37.14)
K∈Th
(ii) The proof of (37.14) is left as an exercise.
37.3.2 L2 -norm estimate

We derive an improved error estimate of the form ku − uh kL2 (D) ≤ chγ ℓ1−γD ku − uh kV♯ for some
real number γ > 0, where ℓD is a length scale associated with D, e.g., ℓD := diam(D). Proceeding
as in §36.3.2, we invoke a duality argument. We consider the adjoint solution ζr ∈ V := H01 (D) for
all r ∈ L2 (D) such that
a(v, ζr ) = (v, r)L2 (D) , ∀v ∈ V, (37.15)
i.e., ζr solves −∆ζr = r in D and γ g (ζr ) = 0. (Note that we enforce a homogeneous Dirichlet
condition on the adjoint solution.) Owing to the elliptic regularity theory (see §31.4), there is
s ∈ (0, 1] and a constant csmo such that
kζr kH 1+s (D) ≤ csmo ℓ2D krkL2 (D) , ∀r ∈ L2 (D). (37.16)
Theorem 37.7 (L2 -estimate). Let u solve (37.1) and let uh solve (37.5). Assume that the elliptic
regularity index satisfies s ∈ ( 21 , 1]. There is c s.t. for all h ∈ H,
1 1
ku − uh kL2 (D) ≤ c h 2 ℓD
2
ku − uh kV♯ . (37.17)
Proof. Set e := u − uh . We apply the abstract error estimate of Lemma 36.14 with V♯ := Vs + Vh
as above, Zs := H 1+s (D) ∩ H01 (D), Yh := Vh ∩ H01 (D), and Z♯ := Zs + Yh equipped with the H 1 -
seminorm. We consider the bilinear form a♯ (v, w) := (∇v, ∇w)L2 (D) . Notice that a♯ is bounded on
V♯ ×Z♯ . Moreover, a♯ (e, yh ) = 0 for all yh ∈ Yh since Yh ⊂ H01 (D), i.e., the Galerkin orthogonality
property (36.28) holds true. Lemma 36.14 implies that
adj
kδ (ζe )kV♯′ k∇(ζe − yh )kL2 (D)
kekL2(D) ≤ + inf kekV♯ ,
kekL2 (D) yh ∈Yh kekL2 (D)
where the first and the second term between parentheses are the adjoint consistency error and the
interpolation error on the adjoint solution, respectively. Let us first bound the adjoint consistency
error. Recall that δ adj (ζe ) is defined in such a way that the following identity holds true: For all
v ∈ V♯ ,
hδ adj (ζe ), viV♯′ ,V♯ = −(v, ∆ζe )L2 (D) − a♯ (v, ζe ) = −(v, n·∇ζe )L2 (∂D) .
The Cauchy–Schwarz inequality implies that
1 1
|hδ adj (ζe ), viV♯′ ,V♯ | ≤ h 2 k∇ζe kL2 (∂D) |v|∂ ≤ h 2 k∇ζe kL2 (∂D) kvkV♯
1 −3
≤ c h 2 ℓD 2 kζe kH 1+s (D) kvkV♯ ,
1 1
since s > 21 . Using (37.16), we infer that kδ adj (ζe )kV♯′ ≤ ch 2 ℓD
2
kekL2(D) . To bound the interpolation
g,av
error on the adjoint solution, we consider the quasi-interpolation operator Ih0 from §22.4. Since
g,av
Ih0 (ζe ) ∈ Yh , we deduce that
g,av
inf k∇(ζe − yh )kL2 (D) ≤ k∇(ζe − Ih0 (ζe ))kL2 (D)
yh ∈Yh
≤ c hs |ζe |H 1+s (D) ≤ c hs ℓ−1−s

D kζe kH 1+s (D) ≤ c csmo hs ℓ1−s
D kekL2 (D) ,
g,av
where we used the approximation properties of Ih0 from Theorem 22.14 and the estimate (37.16).
1 1
Since s > 1
2 and h ≤ ℓD , we have hs ℓ1−s
D ≤ h 2 ℓD
2
, and this concludes the proof.
37.3.3 Symmetrization
1
The estimate (37.17) is suboptimal by a factor hs− 2 , and this loss of optimality is caused by
1
the adjoint consistency error which is only of order h 2 . This shortcoming can be avoided by
symmetrizing ah and modifying ℓh consistently. More precisely, we define
Z Z
sym
ah (vh , wh ) := a(vh , wh ) − (n·∇vh )wh ds − vh (n·∇wh ) ds
∂D ∂D
X Z
1
+ ̟0 vh wh ds,
h F F
F ∈Fh∂
Z X Z
1
ℓsym
h (wh ) := ℓ(w h ) − g(n·∇w h ) ds + ̟ 0 gwh ds.
∂D ∂
hF F
F ∈Fh
Consider the following discrete problem:

(37.18)
asym sym
h (uh , wh ) = ℓh (wh ), ∀wh ∈ Vh .
Adapting the proof of Lemma 37.3, one can show that the problem (37.18) is well-posed if one
chooses the stabilization parameter s.t. ̟0 > n∂ c2dt .
Theorem 37.8 (L2 -estimate). Let u solve (37.1) and let uh solve (37.18). Assume ̟0 > n∂ c2dt
and that there is s ∈ ( 21 , 1] s.t. the adjoint solution satisfies the a priori estimate (37.16). There is
c s.t. for all h ∈ H,
D ku − uh kV♯ . (37.19)
Proof. We proceed as in the proof of Theorem 37.7 with the same spaces V♯ , Z♯ , and Yh , but now
we set a♯ (v, w) := (∇v, ∇w)L2 (D) − (v, n·∇w)L2 (∂D) . We equip Z♯ with the same norm as V♯ , so
that a♯ is bounded on V♯ ×Z♯ . The Galerkin orthogonality property still holds true for a♯ . Indeed,
we have
a♯ (u, yh ) = (f, yh )L2 (D) − (g, n·∇yh )L2 (∂D)

= ℓsym sym
h (yh ) = ah (uh , yh ) = a♯ (uh , yh ), ∀yh ∈ Yh ,
since γ g (u) = g and yh vanishes on ∂D. Now the adjoint consistency error vanishes, and we still
g,av
have kζe − Ih0 (ζe )kZ♯ ≤ c hs |ζe |H 1+s (D) .
Exercises
Exercise 37.1 (Poincaré–Steklov). Let Čps be defined in (31.23). Prove that Čps ℓ−1
D kvkL2 (D) ≤
2 2 12 1
(k∇vkL2 (D) + |v|∂ ) for all v ∈ H (D). (Hint : use h ≤ ℓD and (31.23).)
̟0 −β 2
Exercise 37.2 (Quadratic inequality). Prove that x2 − 2βxy + ̟0 y 2 ≥ 1+̟0 (x
2
+ y 2 ) for all
real numbers x, y, ̟0 ≥ 0 and β ≥ 0.
Exercise 37.3 (Error estimate). Prove (37.14). (Hint : consider the quasi-interpolation opera-
tor from §22.3.)
Exercise 37.4 (Gradient). Let U be an open bounded set in Rd , let s ∈ (0, 1), and set H00 s
(U ) :=
2 1 1−s s ′
[L (U ), H0 (U )]s,2 . (i) Show that ∇ : H (U ) → (H00 (U )) is bounded for all s ∈ (0, 1). (Hint :
use Theorems A.27 and A.30.) (ii) Assume that U is Lipschitz. Show that ∇ : H 1−s (U ) → H −s (U )
is bounded for all s ∈ (0, 1), s 6= 12 . (Hint : see (3.7), Theorem 3.19; see also Grisvard [223,
Lem. 1.4.4.6].)
Exercise 37.5 (L2 -estimate). (i) Modify the proof of Theorem 37.7 by measuring the inter-
polation error on the adjoint solution with the operator Ihg,av instead of Ih0 g,av
, i.e., use Yh :=
1
Vh instead of Yh := Vh ∩ H0 (D). (Hint : set a♯ (v, w) := (∇v, ∇w)L2 (D) − (n·∇v, w)L2 (∂D) +
P 1
F ∈F ∂ ̟0 hF (v, w)L2 (F ) .) (ii) Do the same for the proof of Theorem 37.8.
h
Chapter 38
Discontinuous Galerkin
The goal of this chapter is to study the approximation of an elliptic model problem by the discon-
tinuous Galerkin (dG) method. The distinctive feature of dG methods is that the trial and the test
spaces are broken finite element spaces (see §18.1.2). Inspired by the boundary penalty method
from Chapter 37, dG formulations are obtained by adding a consistency term at all the mesh
interfaces and boundary faces, boundary conditions are weakly enforced à la Nitsche, and conti-
nuity across the mesh interfaces is weakly enforced by penalizing the jumps. The dG method we
study here is called symmetric interior penalty (SIP) because the consistency term is symmetrized
to maintain the symmetry of the discrete bilinear form. Incidentally, the symmetry property is
important to derive optimal L2 -error estimates assuming full elliptic regularity pickup. We also
discuss a useful reformulation of the dG method by lifting the jumps, leading to the important
notion of discrete gradient reconstruction.
38.1 Model problem

For simplicity, we focus on the Poisson equation with homogeneous Dirichlet boundary conditions:

(38.1)
a(u, w) = ℓ(w), ∀w ∈ V,
R R
with a(v, w) := D ∇v·∇w dx, ℓ(w) := D f w dx, f ∈ L2 (D), and D is a Lipschitz polyhedron
in Rd . This problem is well-posed owing to the Lax–Milgram lemma and the Poincaré–Steklov
inequality in H01 (D). We refer the reader to §41.4 for the more general PDE −∇·(λ∇u) = f with
contrasted diffusivity λ.
38.2 Symmetric interior penalty

In this section, we derive the dG approximation of the model problem (38.1) using the SIP method
and show that the discrete problem is well-posed.
168 Chapter 38. Discontinuous Galerkin
38.2.1 Discrete problem

Although dG methods can be used on general meshes composed of polyhedral cells, we consider for
simplicity a shape-regular sequence (Th )h∈H of affine matching meshes so that each mesh covers
D exactly. Let W 1,1 (Th ; Rq ), q ≥ 1, be the broken Sobolev space introduced in Definition 18.1.
Recall that every interface F := ∂Kl ∩ ∂Kr ∈ Fh◦ is oriented by the fixed unit normal vector
nF pointing from Kl to Kr , i.e., nF := nKl = −nKr , and that the jump across F of a function
v ∈ W 1,1 (Th ; Rq ) is defined by setting [[v]]F := v|Kl − v|Kr a.e. on F . We also need the following
notion of face average.
Definition 38.1 (Average). For all F := ∂Kl ∩ ∂Kr ∈ Fh◦ , the average of a function v ∈
W 1,1 (Th ; Rq ) on F is defined as
1
{v}F := v|Kl + v|Kr a.e. on F . (38.2)
2
As for jumps, the subscript F is dropped when the context is unambiguous.
To be more concise, it is customary in the dG literature dedicated to elliptic PDEs to define

the jump and the average of a function at the boundary faces by setting [[v]]F := {v}F := v|Kl a.e.
on F := ∂Kl ∩ ∂D ∈ Fh∂ (i.e., Kl is the unique mesh cell having the boundary face F among its
faces).
b Pb, Σ)
Let (K, b be the reference finite element which we assume to be of degree k ≥ 1. Let us
consider the broken finite element space (see (18.4)) s.t.
Vh := Pkb (Th ) := {vh ∈ L∞ (D) | ψK (vh|K ) ∈ Pb , ∀K ∈ Th }, (38.3)
where ψK (v) := v ◦ TK is the pullback by the geometric mapping TK . The approximation setting
in dG methods is nonconforming since functions in Vh can jump across the mesh interfaces and can
have nonzero boundary values, whereas membership in V := H01 (D) requires continuity across the
interfaces (see Theorem 18.8) and zero boundary values. Nonconformity implies that we cannot
work with the bilinear form a. The construction of the discrete bilinear form ah on Vh ×Vh is a
bit more involved than for the Crouzeix–Raviart finite element method from Chapter 36, where it
was sufficient to replace the weak gradient ∇ by the broken gradient ∇h (see Definition 36.3) to
build ah from a. Instead, the SIP method hinges on the following discrete bilinear form:
Z X Z
ah (vh , wh ) := ∇h vh ·∇h wh dx − {∇h vh }·nF [[wh ]] ds
D F ∈Fh F
X Z X Z
− [[vh ]]{∇h wh }·nF ds + ̟(hF ) [[vh ]][[wh ]] ds, (38.4)
F ∈Fh F F ∈Fh F
where the second and the fourth terms on the right-hand side are reminiscent of Nitsche’s boundary
penalty method. The second term is called consistency term since it is important to establish
consistency/boundedness (see Lemma 38.9). The third term, which is called adjoint consistency
term, makes the discrete bilinear form ah symmetric and it is important to establish an improved
L2 -error estimate (see Theorem 38.12). The fourth term is important to establish coercivity (see
Lemma 38.6). It penalizes jumps across interfaces and values at boundary faces and is, therefore,
called penalty term. Coercivity requires that the penalty parameter be s.t. ̟(hF ) := ̟0 h−1
F , where
̟0 > 0 has to be chosen large enough, and on shape-regular mesh sequences, the local length scale
hF can be taken to be the diameter of F .
We consider the following discrete problem:

(38.5)
ah (uh , wh ) = ℓh (wh ), ∀wh ∈ Vh ,
where the discrete linear form is given by
Z
ℓh (wh ) := f wh dx, ∀wh ∈ Vh . (38.6)
D
This choice for ℓh is possible since the source term in the model problem (38.1) is assumed to be
in L2 (D). A more general setting, e.g., f ∈ H −1 (D), is discussed in Remark 36.5. Furthermore,
it is legitimate to extend ah to (H 1+r (D) + Vh )×Vh , r > 12 , since ∇u ∈ H r (D) implies that
(∇u)|F is well defined as an integrable function for all F ∈ Fh . To motivate the appearance of the
consistency term in the definition of ah , let us prove the following important result.
Lemma 38.2 (Consistency term). Assume that u ∈ H 1+r (D), r > 21 . Then we have ah (u, wh ) =
ℓh (wh ) for all wh ∈ Vh .
Proof. We have [[u]]F = 0 a.e. on all F ∈ Fh (use Theorem 18.8 for F ∈ Fh◦ and γ g (u) = 0 for
F ∈ Fh∂ ) and ∇h u = ∇u (see Lemma 18.9). Since ∇u ∈ H r (D), r > 12 , we also have [[∇u]]·nF = 0
a.e. on all F ∈ Fh◦ (see Remark 18.4). We infer that
Z X Z
ah (u, wh ) = ∇u·∇h wh dx − (∇u·nF )[[wh ]] ds.
D F ∈Fh F
We conclude by performing elementwise integration by parts as follows:

Z Z X X Z
∇u·∇h wh dx = −(∆u)wh dx + (∇u·nK )wh|K ds
D D K∈Th F ∈FK F
X Z
= ℓh (wh ) + (∇u·nF )[[wh ]] ds.
F ∈Fh F
Remark 38.3 (Literature). The SIP approximation has been analyzed in Arnold [15] (see also
Baker [44], Wheeler [394]).
Remark 38.4 (Nonmatching meshes). It is possible to consider nonmatching meshes if the
diameter of each interface F ∈ Fh◦ is uniformly equivalent to the diameter of the two cells sharing
F.
38.2.2 Coercivity and well-posedness

We equip the space Vh with the following norm:
X 1
kvh k2Vh := k∇h vh k2L2 (D) + |vh |2J , |vh |2J := k[[vh ]]k2L2 (F ) . (38.7)
hF
F ∈Fh
That k·kVh is a norm on Vh (and not just a seminorm) can be verified directly: If kvh kVh = 0,
then vh is piecewise constant and [[vh ]]F = 0 for all F ∈ Fh . This means that vh is constant on
D and vanishes at ∂D, so that vh = 0. Our first step in the analysis is to bound from above the
consistency term. Recall that TF := {K ∈ Th | F ∈ FK } is the collection of the mesh cells having
F as face. Let |TF | denote the cardinality of the set TF (|TF | = 2 for all F ∈ Fh◦ and |TF | = 1 for
all F ∈ Fh∂ ).
Lemma 38.5 (Consistency term). Let us set for all (vh , wh ) ∈ Vh ×Vh ,
X Z
nh (vh , wh ) := − {∇h vh }·nF [[wh ]] ds. (38.8)
F ∈Fh F
Then the following holds true for all vh ∈ Vh :

! 21
|nh (vh , wh )| X 1 X
sup ≤ hF knF ·∇(vh|K )k2L2 (F ) . (38.9)
wh ∈Vh |wh |J |TF |
F ∈Fh K∈TF
Proof. The Cauchy–Schwarz inequality leads to

X 1
−1
|nh (vh , wh )| ≤ hF2 knF ·{∇h vh }kL2 (F ) × hF 2 k[[wh ]]kL2 (F )
F ∈Fh
! 12
X
≤ hF knF ·{∇h vh }k2L2 (F ) |wh |J ,
F ∈Fh
1 P
Letting gh := ∇h vh , (38.9) follows from {gh }F = |TF | K∈TF gh|K and
2
1 X 1 X

knF ·{gh }k2L2 (F ) = nF ·gh|K ≤ knF ·gh|K k2L2 (F ) .
|TF |2 |TF |
K∈TF L2 (F ) K∈TF
We shall use the same discrete trace inequality as in Chapter 37 to prove a coercivity property.
Let cdt be the smallest constant such that
−1
knF ·∇h wh|K kL2 (F ) ≤ cdt hF 2 k∇h wh kL2 (K) , (38.10)
for all wh ∈ Vh , all K ∈ Th , all F ∈ FK , and all h ∈ H. Let n∂ := maxK∈Th |FK | be the largest
number of faces per mesh cell, i.e., n∂ ≤ d + 1 for simplicial meshes (the definition of n∂ differs
from that of Chapter 37).
Lemma 38.6 (Coercivity, well-posedness). Let the penalty parameter be s.t. ̟(hF ) := ̟0 h−1
F
with ̟0 > n∂ c2dt . (i) We have
ah (vh , vh ) ≥ αkvh k2Vh , ∀vh ∈ Vh , (38.11)
̟0 −n∂ c2dt
with α := 1+̟0 > 0. (ii) The discrete problem (38.5) is well-posed.
Proof. Let vh ∈ Vh . Our starting observation is that
ah (vh , vh ) = k∇h vh k2L2 (D) + 2nh (vh , vh ) + ̟0 |vh |2J .
Using (38.9) and (38.10), we infer that

|nh (vh , wh )| 1
|nh (vh , vh )| ≤ sup |vh |J ≤ n∂2 cdt k∇h vh kL2 (D) |vh |J ,
wh ∈Vh |wh |J
P P P P
since |TF | ≥ 1, F ∈Fh K∈TF (·) = K∈Th F ∈FK (·), and |FK | ≤ n∂ , so that
X X X
kgh|K k2L2 (K) ≤ n∂ kgh|K k2L2 (K) = n∂ kgh k2L2 (D)
F ∈Fh K∈TF K∈Th
with gh := ∇h vh . This leads to the lower bound

1
ah (vh , vh ) ≥ k∇h vh k2L2 (D) − 2n∂2 cdt k∇h vh kL2 (D) |vh |J + ̟0 |vh |2J ,
whence we infer the coercivity property (38.11) by using the quadratic inequality from Exer-
cise 37.2. Finally, the well-posedness of (38.5) follows from the Lax–Milgram lemma.
Remark 38.7 (Penalty parameter). As in the boundary penalty method from Chapter 37,
one needs a (reasonable) upper bound on the constant cdt to choose a value of ̟0 that guarantees
coercivity. The results of §12.2 show that cdt scales essentially as k 2 . An alternative penalty strat-
egy allowing for an easy-to-compute value of ̟0 is discussed in Remark 38.17, but this technique
requires local inversions of small mass matrices.
Remark 38.8 (Discrete Sobolev inequality). Let ℓD be a length scale associated with D,
e.g., ℓD := diam(D). One can show that there is Csob > 0 such that Csob kvh kLq (D) ≤ ℓD kvh kVh
2d
for all vh ∈ Vh , all h ∈ H, and all q ∈ [1, ∞) if d = 2 and q ∈ [1, d−2 ] if d ≥ 3; see Buffa and
Ortner [95], Di Pietro and Ern [164]. The reader is referred to Arnold [15], Brenner [86] for similar
estimates in broken Hilbert Sobolev spaces (q = 2).
38.2.3 Variations on boundary conditions

1
The non-homogeneous Dirichlet boundary condition u = g on ∂D with g ∈ H 2 (∂D) is discretized
by modifying the right-hand side in (38.5) as follows:
X Z
ℓnD
h (wh ) := ℓ(wh ) − g(nF ·∇h wh − ̟(hF )wh ) ds. (38.12)
F
F ∈Fh∂
For the Robin boundary condition γu + n·∇u = g on ∂D with g ∈ L2 (∂D) and γ ∈ L∞ (∂D)
taking nonnegative values on ∂D (γ := 0 corresponds to the Neumann problem), the discrete
bilinear form and the right-hand side become
Z X Z
aRb
h (vh , wh ) := ∇h vh ·∇h wh dx − {∇h vh }·nF [[wh ]] ds (38.13a)
D F ∈Fh◦ F
X Z X Z X Z
− [[vh ]]{∇h wh }·nF ds + ̟(hF ) [[vh ]][[wh ]] ds + γvh wh ds,
F ∈Fh◦ F F ∈Fh◦ F F
F ∈Fh∂
X Z
ℓRb
h (wh ) := ℓ(wh ) + gwh ds. (38.13b)
F
F ∈Fh∂
One can verify that Lemma 38.2 still holds true in both cases.
38.3 Error analysis

In this section, we derive an energy error estimate, that is, we bound the error by using the
coercivity norm and the abstract error estimate from Lemma 27.5. We also derive an improved
L2 -error estimate by means of a duality argument. We assume that u ∈ Vs with
1
Vs := H 1+r (D) ∩ H01 (D), r> . (38.14)
2
Theorem 31.33). The important property that we use is that for any function v ∈ Vs the nor-
mal derivative nK ·∇v is meaningful in L2 (∂K) for all K ∈ Th . Recall that the discrete space is
Vh := Pkb (Th ) equipped with the k·kVh -norm defined in (38.7). We set V♯ := Vs + Vh and we equip
this space with the norm
X
kvk2V♯ := k∇h vk2L2 (D) + |v|2J + hK knK ·∇v|K k2L2 (∂K) , (38.15)
K∈Th
P 1
with |v|2J := F ∈Fh hF k[[v]]k2L2 (F ) . A discrete trace inequality shows that there is c♯ s.t. kvh kV♯ ≤
c♯ kvh kVh for all vh ∈ Vh and all h ∈ H, i.e., (27.5) holds true. Using the discrete bilinear forms ah
and ℓh defined in (38.4) and (38.6), respectively, the consistency error is s.t. hδh (vh ), wh iVh′ ,Vh :=
ℓh (wh ) − ah (vh , wh ) for all vh , wh ∈ Vh .
Lemma 38.9 (Consistency/boundedness). Assume (38.14). There is ω♯ , uniform w.r.t. u ∈
Vs , s.t. for all vh ∈ Vh and all h ∈ H,
Proof. Let vh ∈ Vh and let us set η := u − vh . Owing to Lemma 38.2 and since [[u]]F = 0 for all
F ∈ Fh , we infer that for all wh ∈ Wh ,
Z
hδh (vh ), wh iVh′ ,Vh = ∇h η·∇h wh dx + n♯ (η, wh )
D
X Z X ̟0 Z
− [[η]]{∇h wh }·nF ds + [[η]][[wh ]] ds,
F hF F
F ∈Fh F ∈Fh
P R
where n♯ (v, wh ) := − F ∈Fh F {∇h v}·nF [[wh ]] ds is understood as an extension to V♯ ×Vh of the
discrete bilinear form nh originally defined on Vh ×Vh by (38.8). (Note that the assumption r > 12
in the definition of Vs is crucial for this extension to make sense.) The Cauchy–Schwarz inequality
implies that
Z X ̟0 Z

∇h η·∇h wh dx + [[η]][[wh ]] ds
hF F
D F ∈Fh
≤ k∇h ηkL2 (D) k∇h wh kL2 (D) + ̟0 |η|J |wh |J ≤ max(1, ̟0 )kηkV♯ kwh kVh .
Since the bound (38.9) is still valid for n♯ (η, wh ), we also have
! 12
X 1 X
|n♯ (η, wh )| ≤ hF knF ·∇(η|K )k2L2 (F ) |wh |J
|TF |
F ∈Fh K∈TF
≤ c kηkV♯ |wh |J ≤ c kηkV♯ kwh kVh .

(This is where we use the contribution of the normal derivative to the k·kV♯ -norm.) Proceeding as
in the proof of Lemma 38.5, we finally infer that
X 21
X Z 1 X

[[η]]{∇h wh }·nF ds ≤ |η|J hF k∇(wh|K )k2L2 (F )
F |TF |
F ∈Fh F ∈Fh K∈TF
1 1
≤ n∂ cdt |η|J k∇h wh kL2 (D) ≤ n∂2 cdt kηkV♯ kwh kVh ,
2
where we used the discrete trace inequality (38.10) as in the proof of Lemma 38.6. Collecting the
above bounds shows that |hδh (vh ), wh iVh′ ,Vh | ≤ ckηkV♯ kwh kVh , i.e., (38.16) holds true.
Theorem 38.10 (Convergence). Let u solve (38.1) and let uh solve (38.5) with the penalty
parameter ̟0 > c2dt n∂ . Assume (38.14). (i) There is c s.t. the following holds true for all h ∈ H:

vh ∈Vh
(ii) Letting t := min(k, r), we have

! 21
X
K |u|H 1+t (K) . (38.18)
K∈Th
(ii) We bound the infimum in (38.17) by taking vh := Ih♯ (u), where Ih♯ : L1 (D) → Pkb (Th ) is the L1 -
1
stable interpolation operator from §18.3. We need to bound k∇(η|K )kL2 (K) +hK
2
k∇(η|K )kL2 (∂K) for
−1
all K ∈ Th and hF 2 k[[η]]F kL2 (F ) for all F ∈ Fh , with η := u − Ih♯ (u). Theorem 18.14 implies that
1
k∇(η|K )kL2 (K) ≤ chtK |u|H 1+t (K) . Moreover, Corollary 18.15 implies that hK
2
k∇(η|K )kL2 (∂K) ≤
t+ 1
chtK |u|H 1+t (K) and that kη|K kL2 (F ) ≤ chK 2 |u|H 1+t (K) for any face F ∈ FK . Since [[η]]F := η|Kl
for all F := ∂Kl ∩ ∂D ∈ Fh∂ and [[η]]F := η|Kl − η|Kr for all F := ∂Kl ∩ ∂Kr ∈ Fh◦ , we can use
the shape-regularity of the mesh sequence and the triangle inequality for the jump to infer that
−1 P
hF 2 k[[η]]F kL2 (F ) ≤ c K∈TF htK |u|H 1+t (K) for all F ∈ Fh . This leads to (38.18).
Remark 38.11 (L2 -orthogonal projection). Note that, as shown in Remark 18.18, Ih♯ is the
L2 -orthogonal projection onto Pkb (Th ) since ψK is the pullback by the geometric mapping TK .
We now derive an L2 -error estimate by invoking a duality argument as in §36.3.3. For all
g ∈ L2 (D), we consider the adjoint solution ζg ∈ V := H01 (D) s.t. a(v, ζg ) = (v, g)L2 (D) for all
v ∈ V, i.e., −∆ζg = g in D and γ g (ζg ) = 0. Owing to the elliptic regularity theory (see §31.4),
there is s ∈ (0, 1] and a constant csmo such that kζg kH 1+s (D) ≤ csmo ℓ2D kgkL2 (D) for all g ∈ L2 (D).
Theorem 38.12 (L2 -estimate). Under the assumptions of Theorem 38.10 and assuming that the
elliptic regularity index satisfies s ∈ ( 12 , 1], there is c such that for all h ∈ H,

D ku − uh kV♯ . (38.19)
Proof. Apply Lemma 36.14 and use exact adjoint consistency; see Exercise 38.3.
Remark 38.13 (Variations on symmetry). Let us set

Z X Z
ah (vh , wh ) := ∇h vh ·∇h wh dx − {∇h vh }·nF [[wh ]] ds
D F ∈Fh F
X Z X Z
−θ {∇h wh }·nF [[vh ]] ds + ̟(hF ) [[vh ]][[wh ]] ds, (38.20)
F ∈Fh F F ∈Fh F
where θ is a real number (θ := 1 corresponds to the SIP formulation). The choice θ := −1

gives the method usually called nonsymmetric interior penalty (NIP). This choice is interesting
since it simplifies the analysis of the coercivity in that the consistency term cancels with the
added nonsymmetric term. The original idea can be traced back to the method in Oden et al.
[318], where the nonsymmetric method is introduced without the penalty term. The convergence
analysis when the penalty term is included can be found in Rivière et al. [335, 336], where it is
shown that coercivity only requires ̟0 > 0; see also Larson and Niklasson [274] for the inf-sup
stability analysis. The incomplete interior penalty (IIP) method corresponds to the choice θ := 0.
Similarly to SIP, a minimal threshold on the penalty parameter ̟0 is required for the coercivity; see
Dawson et al. [157]. Whenever θ 6= 1, the analysis of the L2 -error estimate proceeds as in §37.3.2
1
(accounting for an adjoint consistency error), and one only obtains ku − uhkL2 (D) ≤ ch 2 ku − uhkV♯
even if full elliptic regularity holds true (s = 1).
Remark 38.14 (L∞ -estimates). Pointwise dG error estimates are found in Kanschat and Ran-
nacher [264], Chen and Chen [117], Guzmán [233].
38.4 Discrete gradient and fluxes

In this section, we introduce the notion of discrete gradient and use it to derive an alternative
viewpoint on the SIP bilinear form. One interesting outcome is a reformulation of the discrete
problem (38.5) in terms of local problems with numerical fluxes.
38.4.1 Liftings
Loosely speaking the discrete gradient consists of the broken gradient plus a correction associated
with the jumps. This correction is formulated in terms of local liftings introduced in Bassi and
Rebay [46] and analyzed in Brezzi et al. [93] (see also Perugia and Schötzau [323] for the hp-
analysis). Let F ∈ Fh and an integer l ≥ 0. Consider the local lifting operator L lF : L2 (F ) →
Plb (Th ) := Plb (Th ; Rd ) s.t. for all ϕ ∈ L2 (F ), the discrete function L lF (ϕ) is defined as
Z Z
L lF (ϕ)·τh dx := {τh }·nF ϕ ds, ∀τh ∈ Plb (Th ). (38.21)
D F
By localizing the support of τh to a single S mesh cell, we infer that LlF (ϕ) is collinear to nF
and is supported in the set DF := int( K∈TF K). In practice, the Cartesian components of the
polynomial function L lFR (ϕ) can be computed in each K ∈ TF by inverting the local mass matrix
with entries MK,ij := K θK,i θK,j dx, where the functions θK,i are the local shape functions in K.
Consider now a function v ∈ H 1 (Th ). We define the global lifting of the jumps of v as follows:
X l
L lh ([[v]]) := L F ([[v]]).
F ∈Fh
This makes sense since 2

P [[v]]F ∈l L (F ) for all F ∈ Fh . A consequence of supp(L LlF ([[v]])) = DF
l
is that L h ([[v]])|K := F ∈FK L F ([[v]]) for all K ∈ Th , i.e., only the jumps across the faces of K
contribute to the restriction to K of the global lifting L lh ([[v]]).
Lemma 38.15 (Stability). The following holds true for all l ≥ 0:
−1
LlF (ϕ)kL2 (DF ) ≤ cdt hF 2 kϕkL2 (F ) ,
kL ∀ϕ ∈ L2 (F ), ∀F ∈ Fh , (38.22a)
1
Llh ([[v]])kL2 (D) ≤ n∂ cdt |v|J ,
kL 2
∀v ∈ H 1 (Th ), (38.22b)
where cdt is the constant from the discrete trace inequality (38.10).
Proof. The proof of (38.22a) is proposed in Exercise 38.4. To prove (38.22b), we use the Cauchy–
Schwarz inequality and the definition of n∂ to infer that
Z X 2
X

Llh ([[v]])k2L2 (K)
kL =
l
L F ([[v]]) dx ≤ n∂ LlF ([[v]])k2L2 (K) ,
kL
K F ∈FK

F ∈FK
for all K ∈ Th . Summing over the mesh cells, recalling that the support of L lF ([[v]]) is DF , and
using (38.22a) yields (38.22b).
Definition 38.16 (Discrete gradient). Let l ≥ 0. The discrete gradient operator Glh : H 1 (Th ) →
L2 (D) is defined as follows:
Glh (v) := ∇h v − L lh ([[v]]), ∀v ∈ H 1 (Th ). (38.23)
We can now use Definition 38.16 to derive alternative expressions for the SIP bilinear form ah
defined in (38.4). Recalling that Vh := Pkb (Th ), k ≥ 1, we choose the polynomial degree of the
liftings such that l ∈ {k − 1, k}. Since ∇h vh , ∇h wh ∈ Pbk−1 (Th ) ⊂ Pbl (Th ) for all vh , wh ∈ Vh , we
infer that
Z X Z
∇h vh ·∇h wh dx − {∇h vh }·nF [[wh ]] + [[vh ]]{∇h wh }·nF ds
D F ∈Fh F
Z Z
= Glh (vh )·Glh (wh ) dx − L lh ([[vh ]])·L
Llh ([[wh ]]) dx. (38.24)
D D
Recalling the expression (38.4) of ah , we obtain

Z
ah (vh , wh ) := Glh (vh )·Glh (wh ) dx + s̃h (vh , wh ), (38.25)
D
P R R
with s̃h (vh , wh ) := F ∈Fh ̟(hF ) F [[vh ]][[wh ]] ds− D L lh ([[vh ]])·L
Llh ([[wh ]]) dx. The estimate (38.22b)
from Lemma 38.15 implies that
ah (vh , vh ) ≥ kGlh (vh )k2L2 (D) + (̟0 − n∂ c2dt )|vh |2J , (38.26)
for all vh ∈ Vh , showing again the relevance of the condition ̟0 > n∂ c2dt for coercivity (see
Lemma 38.6).
Remark 38.17 (Alternative penalty strategy). It is possible to penalize the liftings of the
jumps instead of the jumps, leading to the following modification of the SIP bilinear form:
Z X Z
ǎh (vh , wh ) := ∇h vh ·∇h wh dx − {∇h vh }·nF [[wh ]] ds
D F ∈Fh F
X Z X Z
− [[vh ]]{∇h wh }·nF ds + ̟0 L lF ([[vh ]])·L
LlF ([[wh ]]) dx.
F ∈Fh F F ∈Fh D
The main advantage of this formulation is that coercivity holds true as soon as ̟0 > n∂ , thereby
avoiding the constant cdt from (38.10). However, the discretization stencil is larger since the dofs
in two cells K, K ′ ∈ Th are coupled if there is K ′′ ∈ Th s.t. ∂K ∩ ∂K ′′ ∈ Fh◦ and ∂K ′ ∩ ∂K ′′ ∈ Fh◦
(for the usual penalty strategy the coupling condition is ∂K ∩ ∂K ′ ∈ Fh◦ ).
Remark 38.18 (Choosing l). The computation of the discrete gradient can be done with any
l ≥ k − 1. The minimal choice is l = k − 1, but choosing l = k may be more interesting from
the implementation point of view since it does not require the user to construct the finite element
b
space Pk−1 (Th ).
Remark 38.19 (Literature). The discrete gradient is an important notion in the design and
analysis of dG methods for nonlinear problems. We refer the reader to Ten Eyck and Lew [364] for
nonlinear mechanics, to Burman and Ern [100], Buffa and Ortner [95] for Leray–Lions operators,
and to Di Pietro and Ern [164] for the incompressible Navier–Stokes equations. Moreover, an
important stability result established in John et al. [260] is that if l = k + 1, then there is c s.t.
kvh kVh ≤ ckGlh (vh )kL2 (D) for all vh ∈ Vh and all h ∈ H. Since the proof of this result invokes
Raviart–Thomas functions, simplicial meshes are required, but hanging nodes are still allowed
under some assumptions. An interesting consequence R of this stability result is that for l = k + 1,
replacing ah defined in (38.4) by ãh (vh , wh ) := D Glh (vh )·Glh (wh ) dx gives a stable and optimally
convergent dG discretization without any penalty parameters. Notice that ãh does not deliver
exact consistency because liftings are discrete objects; see Exercise 38.6. For the same reason, the
bilinear form ah defined in (38.4) coincides with the right-hand side of (38.25) on Vh ×Vh , but the
two sides of the equality produce different results on V♯ ×Vh .
38.4.2 Local formulation with fluxes

Let K ∈ Th and consider a smooth function ξ ∈ C 1 (K). Integration by parts shows that the
solution to (38.1), if it is smooth enough, satisfies
Z Z Z Z
f ξ dx = −(∆u)ξ dx = ∇u·∇ξ dx − (∇u·nK )ξ ds.
K K K ∂K
Splitting the boundary integral over the faces F ∈ FK yields

Z X Z Z
∇u·∇ξ dx + ǫK,F ΦF (u)ξ ds = f ξ dx, (38.27)
K F ∈FK F K
where ΦF (u) := −∇u·nF , ǫK,F = nK ·nF , and nK is the outward normal to K (nK ·nF = ±1, for
all F ∈ FK , depending on the orientation of F ). The function ΦF is called exact flux since (38.27)
expresses a balance between the source term in K, the diffusion processes in K, and the fluxes
across all the faces in FK . An interesting feature of dG methods is that one obtains a discrete
counterpart of (38.27) when the test function is supported only in the mesh cell K.
Lemma 38.20 (Local formulation). Let uh solve (38.5). Let the numerical flux on a mesh face
F ∈ Fh be defined by
b F (uh ) := −{∇h uh }·nF + ̟(hF )[[uh ]].
Φ (38.28)
Then the following holds true for all q ∈ PK and all K ∈ Th :
Z X Z Z
l
Gh (uh )·∇q dx + ǫK,F b
ΦF (uh )q ds = f q dx. (38.29)
K F ∈FK F K
Proof. Let 1K be the indicator function of K and

R let q be arbitrary in PK . Using the test function
wh := q1K in (38.5), we obtain ah (uh , q1K ) = K f q dx. Then, (38.29) follows by invoking (38.24)
and by making use of the identity [[q1K ]]F = ǫK,F q if F ∈ FK and [[q1K ]]F = 0 otherwise.
The numerical flux Φ b F (uh ) consists of a centered flux, −{∇h uh }·nF , originating from the
consistency term, plus a stabilization term, ̟(hF )[[uh ]], originating from the penalty term. A
unified presentation of dG methods for the Poisson equation based on fluxes can be found in Arnold
et al. [21].
38.4.3 Equilibrated H(div) flux recovery

The vector-valued function σ := −∇u is called diffusive flux. This function is important in many
applications where the underlying PDE expresses a conservation principle in the form ∇·σ = f in
D. Since σ ∈ H(div; D), Theorem 18.10 implies that [[σ]]·nF = 0 for all F ∈ Fh◦ (possibly in a weak
sense if σ is not smooth enough). From a physical viewpoint, this zero-jump condition expresses
the fact that what flows out of a mesh cell through one of its faces flows into the neighboring mesh
cell.
The local formulation (38.29) provides a natural way of reconstructing a discrete diffusive flux
σh in H(div; D) that closely approximates σ. Assuming that the mesh is matching and simplicial,
we now describe a way to reconstruct σh in the Raviart–Thomas finite element space Pld (Th )
defined in (19.16) with l ∈ {k − 1, k}. The reconstruction is explicit and amounts to prescribing
the global degrees of freedom of σh in Phd (Th ); see Ern et al. [193], Kim [268].
Lemma 38.21 (Flux recovery). Let σh ∈ Phd (Th ) be such that
Z Z
−1
(σh ·nF )(q ◦ TF ) ds = b F (uh )(q ◦ T −1 ) ds,
Φ ∀F ∈ Fh , ∀q ∈ Pl,d−1 ,
F
F F
Z Z
and if l ≥ 1, σh ·r dx = − Glh (uh )·r dx, ∀K ∈ Th , ∀r ∈ P l−1,d ,
K K
where TF is an affine bijective mapping from the unit simplex of Rd−1 to F for all F ∈ Fh . Let
Ihb denote the L2 -orthogonal projection onto Plb (Th ). Then we have
∇·σh = Ihb (f ). (38.30)
Proof. Integrating by parts on a cell K ∈ Th and using (38.29), we infer that
Z Z X Z Z
−1
(∇·σh )q dx = − σh ·∇q dx + (σh ·nK )(q|F ◦TF )◦TF ds = f q dx,
K K F ∈FK F K
for all q ∈ Pl,d since ∇q ∈ P l−1,d and q|F ◦ TF ∈ Pl,d−1 (see Lemma 7.10). Then (38.30) is a
consequence of ∇·σh ∈ Plb (Th ).
Equation (38.30) shows that ∇·σh optimally approximates the source term. By proceeding as
in Di Pietro and Ern [165, §5.5.3], it is possible to show that kσ − σh kL2 (D) ≤ c(ku − uh kV♯ +
hkf − Ihb (f )kL2 (D) ).
Exercises
Exercise 38.1 (Elementary dG identities). (i) Let F := ∂Kl ∩ ∂Kr ∈ Fh◦ . Prove that
2{σ·nK q} = ({σ}[[q]] + [[σ]]{q})·nF . (ii) Let θl , θr ∈ [0, 1] such that θl + θr = 1. Let [[a]]θ :=
2(θr al − θl ar ) and {a}θ := θl al + θr ar . Show that {ab} = {a}{b}θ + 41 [[a]]θ [[b]].
Exercise 38.2 (Boundary conditions). (i) Assume that u solves the Poisson problem (38.1)
with the non-homogeneous Dirichlet condition u = g on ∂D. Let aθh be defined in (38.20). Devise
ℓθ,nD
h so that exact consistency holds for the following formulation: Find uh ∈ Vh such that
ah (uh , wh ) = ℓθ,nD
θ
h (wh ) for all wh ∈ Vh . (ii) Assume that u solves the Poisson problem with the
Robin condition γu + n·∇u = g on ∂D. Let ℓRb Rb
h be defined in (38.13b). Devise ah so that exact
θ,Rb
consistency holds for the following formulation: Find uh ∈ Vh such that ah (uh , wh ) = ℓRb h (wh )
for all wh ∈ Vh .
Exercise 38.3 (L2 -estimate). Prove Theorem 38.12. (Hint : see the proof of Theorem 37.8.)
Exercise 38.4 (Local lifting). Prove (38.22a). (Hint : use (38.10).)
Exercise 38.5 (Local formulation). Write the local formulation of the OBB, NIP, and IIP dG
methods discussed in Remark 38.13.
Exercise 38.6 (Extending (38.25)). Let ãh (resp., P ah ) be defined
R by extending (38.25) (resp.,
(38.4)) to V♯ ×Vh . Show that ãh (v, wh ) = ah (v, wh ) + F ∈Fh F {∇h v − Ihb (∇h v)}·nF [[wh ]] ds for
all (v, wh ) ∈ V♯ ×Vh .
Exercise 38.7 (Discrete gradient). Let (vh )h∈H be a sequence in (Vh )h∈H (meaning that vh ∈
Vh for all h ∈ H). Assume that there is C s.t. kvh kVh ≤ C for all h ∈ H. One can show that there is
v ∈ L2 (D) such that, up to a subsequence, vh → v in L2 (D) as h → 0; see [165, Thm. 5.6]. (i) Show
that, up to a subsequence, Glh (vh ) weakly converges to some G in L2 (D) as h → 0. (Hint : bound
kGlh (vh )kL2 (D) .) (ii) Show that 1
R G l = ∇v and thatRv ∈ H0 (D). (Hint : extend
R functions by zero
P
outside D and prove first that Rd Gh (vh )·Φ dx = − Rd vh ∇·Φ dx+ F ∈Fh F {Φ−Ihb Φ}·nF [[vh ]] ds
for all Φ ∈ C0∞ (Rd ).)
Chapter 39
Hybrid high-order method
As in Chapter 38, we want to approximate the Poisson equation with homogeneous Dirichlet
conditions, but this time we use the hybrid high-order (HHO) method. In this method, the
discrete solution is composed of a pair: a face component that approximates the trace of the
solution on the mesh faces and a cell component that approximates the solution in the mesh
cells. The cell unknowns can be eliminated locally by static condensation. The two key ideas
behind the HHO method are a local reconstruction operator and a local stabilization operator.
Altogether the approximation setting is nonconforming since the solution is approximated by
piecewise polynomials that can jump across the mesh interfaces. The error analysis leads to
O(hk+1 ) convergence rates in H 1 for smooth solutions if polynomials of degree k ≥ 0 are used for
the face and the cell unknowns. Moreover, we show that the HHO method is closely related to the
hybridizable discontinuous Galerkin (HDG) method.
39.1 Local operators

Local reconstruction and stabilization operators associated with each mesh cell lie at the heart
of the HHO method. Although these operators can be defined on general meshes composed of
cells having a polyhedral shape, for simplicity, we are going to restrict our attention to simplicial
meshes.
39.1.1 Discrete setting

Let K ∈ Th be a mesh cell, where Th is a member of a shape-regular sequence of affine simplicial
meshes. Let k ≥ 0 be the polynomial degree. We consider a pair v̂K := (vK , v∂K ), where vK is
defined on K and v∂K is defined on the faces F ∈ FK composing the boundary ∂K of K. We
write v̂K := (vK , v∂K ) ∈ V̂Kk := VKk ×V∂K
k
with
Y
−1
VKk := Pk,d ◦ TK , k
V∂K := Pk,d−1 ◦ TF−1 , (39.1)
F ∈FK
where TK : Sbd → K, TF : Sbd−1 → F are affine geometric mappings defined on the reference
simplices of Rd and Rd−1 , respectively (see Figure 39.1). We have dim(V̂Kk ) = k+d k+d−1
d +(d+1) d−1 .
k
Functions in V∂K are defined independently on each face composing the boundary ∂K of K. More
180 Chapter 39. Hybrid high-order method
k
precisely, if v∂K ∈ V∂K and F1 , F2 ∈ FK are two distinct faces of K, then v∂K|F1 and v∂K|F2 may
not have the same restriction on F1 ∩ F2 .
k=0 k=1 k=2
Figure 39.1: Local unknowns for the HHO method (d = 2). Each bullet on the faces and in the
cell conventionally represents one basis function, which can be of modal or nodal type. The face
basis functions are not necessarily continuous at the cell vertices.
39.1.2 Local reconstruction and stabilization

−1
Let VKk+1 := Pk+1,d ◦ TK . We define a reconstruction operator R : V̂Kk → VKk+1 s.t., for every pair
v̂K = (vK , v∂K ) ∈ V̂K , the function R(v̂K ) ∈ VKk+1 is s.t. for all q ∈ VKk+1 ,
k
(∇R(v̂K ), ∇q)L2 (K) := −(vK , ∆q)L2 (K) + (v∂K , nK ·∇q)L2 (∂K) (39.2)
= (∇vK , ∇q)L2 (K) − (vK − v∂K , nK ·∇q)L2 (∂K) ,
and (R(v̂K ) − vK , 1)L2 (K) := 0. This local Neumann problem (where the unknown is R(v̂K )) makes
sense since the right-hand side of (39.2) vanishes when the function q is constant. The second
equality in (39.2) is obtained by integration by parts. The reconstruction operator depends on K
and k, but for simplicity we just write R. Notice that R(v̂K ) = vK if v∂K = vK|∂K . In practice,

the computation of R(v̂K ) requires inverting the local stiffness matrix in K of order k+d+1
d − 1.
k 1 k
Let ÎK : H (K) → V̂K be the local interpolation operator s.t.
k
ÎK (v) := (ΠkK (v), Πk∂K (v|∂K )) ∈ V̂Kk , ∀v ∈ H 1 (K), (39.3)
where ΠkK : L2 (K) → VKk is the L2 -orthogonal projection onto VKk and Πk∂K : L2 (∂K) → V∂K
k
is
2 k
the L -orthogonal projection onto V∂K .
Lemma 39.1 (Elliptic projection). EK := R ◦ ÎK k

: H 1 (K) → VKk+1 is the elliptic projection
onto VK , i.e., (∇(EK (v) − v), ∇q)L2 (K) = 0 and (EK (v) − v, 1)L2 (K) = 0 for all q ∈ VKk+1 and
k+1
all v ∈ H 1 (K).
Proof. Let v ∈ H 1 (K) and φ := R(ÎKk

(v)) = R(ΠkK (v), Πk∂K (v|∂K )). Using the definition (39.2) of
the reconstruction operator, we infer that
(∇φ, ∇q)L2 (K) = −(ΠkK (v), ∆q)L2 (K) + (Πk∂K (v|∂K ), nK ·∇q)L2 (∂K)
= −(v, ∆q)L2 (K) + (v, nK ·∇q)L2 (∂K) = (∇v, ∇q)L2 (K) ,
for all q ∈ VKk+1 , since ∆q ∈ VKk and nK ·∇q ∈ V∂Kk

(recall that all the faces are planar so that nK
k
is piecewise constant). Moreover, (R(ÎK (v)), 1)L2 (K) = (ΠkK (v), 1)L2 (K) = (v, 1)L2 (K) owing to the
k
definition of R and ÎK .
The main issue with the reconstruction operator is that ∇R(v̂K ) = 0 does not imply that vK and
v∂K are constant functions taking the same value. This can be seen from a dimension argument: We
have ker(R) ⊂ {v̂K ∈V̂Kk | ∇R(v̂K ) = 0} and dim(ker(R)) = dim(V̂Kk ) − dim(im(R)) ≥ dim(V̂Kk ) −
dim(VKk+1 ) = k+d−1
d−1
kd+1
k+1 > 1 (unless k = 0 and d = 1 where the difference is equal to 1). To fix
this issue, a local stabilization operator is introduced. Among various possibilities, we focus on an
operator that maps V̂Kk to face-based functions S : V̂Kk → V∂K
k
s.t. for all v̂K ∈ V̂Kk ,

S(v̂K ) := Πk∂K vK|∂K − v∂K + ((I − ΠkK )R(v̂K ))|∂K , (39.4)
where I is the identity. The stabilization operator depends on K and k, but for simplicity we just
write S. Letting δ∂K := vK|∂K − v∂K be the difference between the trace of the cell component
and the face component on ∂K, the operator S in (39.4) can be rewritten as

S(v̂K ) = Πk∂K δ∂K − ((I − ΠkK )R(0, δ∂K ))|∂K , (39.5)
where we used R(v̂K ) = R(vK , vK|∂K )− R(0, δ∂K ), by linearity, and that (I − ΠkK )R(vK , vK|∂K ) = 0
since R(vK , vK|∂K ) = vK and vK ∈ VKk . The identity (39.5) shows that S(v̂K ) only depends
(linearly) on the difference (vK|∂K − v∂K ). The role of S is to help enforce the matching between
the trace of the cell component and the face component. In the discrete problem, this matching is
enforced in a least-squares manner (see §39.2.1). In practice, computing S(v̂K ) requires to evaluate
L2 -orthogonal projections in the cell and on its faces, which entails inverting the mass matrix in
K, which is of size k+d d , and inverting the mass matrix in each face F ∈ FK , which is of size
k+d−1

d−1 .
We now show that the operator S leads to an important stability result. We equip the space
V̂Kk with the following H 1 -like seminorm: For all v̂K ∈ V̂Kk ,
|v̂K |2V̂ k := k∇vK k2L2 (K) + h−1 2

K kvK − v∂K kL2 (∂K) . (39.6)
K
Lemma 39.2 (Stability). There are 0 < α ≤ ω s.t. for all v̂K ∈ V̂Kk , all K ∈ Th , and all h ∈ H,
α |v̂K |2V̂ k ≤ k∇R(v̂K )k2L2 (K) + h−1 2 2

K kS(v̂K )kL2 (∂K) ≤ ω |v̂K |V̂ k . (39.7)
K K
Proof. Let v̂K = (vK , v∂K ) ∈ V̂Kk and set rK := R(v̂K ).

(1) Lower bound. Let us first bound k∇vK kL2 (K) . Taking q := vK in (39.2) and using the
Cauchy–Schwarz inequality yields
k∇vK k2L2 (K) = (∇rK , ∇vK )L2 (K) + (vK − v∂K , nK ·∇vK )L2 (∂K)
−1 1
≤ k∇rK kL2 (K) k∇vK kL2 (K) + hK 2 kvK − v∂K kL2 (∂K) hK
2
knK ·∇vK kL2 (∂K) .
1
A discrete trace inequality yields hK
2
knK ·∇vK kL2 (∂K) ≤ ck∇vK kL2 (K) . These bounds imply that
−1
k∇vK kL2 (K) ≤ c (k∇rK kL2 (K) + hK 2 kvK − v∂K kL2 (∂K) ). (39.8)
−1
Let us now bound hK 2 kvK − v∂K kL2 (∂K) . We have

kΠk∂K ((I − ΠkK )rK )|∂K kL2 (∂K) ≤ k(I − ΠkK )rK kL2 (∂K)
−1 1
≤ c hK 2 k(I − ΠkK )rK kL2 (K) ≤ c′ hK
2
k∇rK kL2 (K) ,
owing to a discrete trace inequality and the local Poincaré–Steklov inequality (12.13) since (I −
ΠkK )rK has zero mean value in K. Using the definition of S and the fact that vK|∂K − v∂K is in
k
V∂K , we infer that vK|∂K − v∂K = S(v̂K ) − Πk∂K (((I − ΠkK )rK )|∂K ). The triangle inequality and
the above bound imply that
−1 −1
hK 2 kvK − v∂K kL2 (∂K) ≤ hK 2 kS(v̂K )kL2 (∂K) + c k∇rK kL2 (K) .
Combining this estimate with (39.8) proves the lower bound in (39.7).
(2) Upper bound. Using the definition (39.2) of R, we have
(∇rK , ∇q)L2 (K)

k∇rK kL2 (K) = sup
k+1
q∈VK k∇qkL2 (K)
(∇vK , ∇q)L2 (K) − (vK − v∂K , nK ·∇q)L2 (∂K)
= sup
k+1
q∈VK k∇qkL2 (K)
−1
≤ k∇vK kL2 (K) + c hK 2 kvK − v∂K kL2 (∂K) ,
where the last bound follows from the Cauchy–Schwarz inequality and a discrete trace inequality.
Moreover, the triangle inequality and the L2 -stability of Πk∂K imply that
kS(v̂K )kL2 (∂K) ≤ kvK − v∂K kL2 (∂K) + kΠk∂K ((I − ΠkK )rK )|∂K kL2 (∂K) .
Invoking the above bound on kΠk∂K (((I − ΠkK )rK )|∂K )kL2 (∂K) yields
−1 −1
hK 2 kS(v̂K )kL2 (∂K) ≤ hK 2 kvK − v∂K kL2 (∂K) + c′ k∇rK kL2 (K) .
Combining the above bounds proves the upper bound in (39.7).
Another important property of the stabilization operator is that it leads to optimal approxi-
k
mation properties when combined with the interpolation operator ÎK .
k
Lemma 39.3 (Approximation property of S ◦ ÎK ). There is c s.t. for all v ∈ H 1 (K), all
K ∈ Th , and all h ∈ H,
−1 k
hK 2 kS(ÎK (v))kL2 (∂K) ≤ c k∇(v − EK (v))kL2 (K) . (39.9)
Proof. Let v ∈ H 1 (K) and set η := v − EK (v). Owing to the definitions of S and ÎK
k
and the fact
k
that R ◦ ÎK = EK , we infer that
k

S(ÎK (v)) = Πk∂K ΠkK (v)|∂K − Πk∂K (v|∂K ) + ((I − ΠkK )EK (v))|∂K
= ΠkK (η)|∂K − Πk∂K (η|∂K ),
where we used that Πk∂K (ΠkK (η)|∂K ) = ΠkK (η)|∂K and Πk∂K ◦ Πk∂K = Πk∂K . Invoking the triangle
inequality, the L2 -stability of the projections ΠkK and Πk∂K , and a discrete trace inequality leads
to
−1 1
kS(ÎKk
(v))kL2 (∂K) ≤ c (hK 2 kηkL2 (K) + kηkL2 (∂K) ) ≤ c′ hK
2
k∇ηkL2 (K) ,
where the last bound follows from the multiplicative trace inequality (12.16) and the Poincaré–
Steklov inequality (12.13) (since (η, 1)L2 (K) = 0).
Remark 39.4 (Literature). The HHO method was introduced in Di Pietro and Ern [166],
Di Pietro et al. [168]. Its algebraic realization and implementation are discussed in Cicuttin et al.
[128], and an open-source library is available at https://github.com/wareHHOuse/diskpp.
Remark 39.5 (Variants). Let k ≥ 0 be the face polynomial degree. As observed in Cockburn
et al. [137], the cell polynomial degree can be taken equal to (k − 1), if k ≥ 1, or to k (as considered
above), or to (k + 1). In the first case, the HHO method is related, up to an equivalent form of the
stabilization, to the nonconforming Virtual Element method of Ayuso de Dios et al. [32] (which
adopts a viewpoint close in spirit to §39.1.3). In the third case, the stabilization operator can be
simplified to S(v̂K ) := Πk∂K (vK|∂K − v∂K ), as considered by Lehrenfeld and Schöberl for the HDG
method (see [280, 281] and Oikawa [319]). All these methods have also close connections with the
weak Galerkin method of Wang and Ye [386].
39.1.3 Finite element viewpoint

In this section, we briefly outline how the above setting can be understood within the finite element
viewpoint by identifying a triple (K, P, Σ) (see Definition 5.2). Recall that k ≥ 0 is the polynomial
degree. Consider the space
n o
k −1
VK := v ∈ H 1 (K) | ∆v ∈ Pk,d ◦ TK k
, nK ·(∇v)|∂K ∈ V∂K , (39.10)
−1
k
with V∂K defined in (39.1). We observe that VKk+1 := Pk+1,d ◦ TK ⊂ VKk
, but there are functions
k k+1
in VK that are not in VK and these functions are not accessible to direct computation (they can
be approximated by solving a subgrid problem in K). A key observation is the following.
k
Lemma 39.6 (VK ↔ V̂Kk ). The functional space VK k
is finite-dimensional and the restriction of
k k k k k
ÎK to VK , i.e., ÎK|V k : V K → V̂K , is an isomorphism.
K
k k k k k
Proof. (1) Let us first prove that ÎK|V k : VK → V̂K is injective. Let v ∈ VK be s.t. ÎK (v) =
K
(0, 0) ∈ V̂Kk . Integrating by parts, we infer that

k∇vk2L2 (K) = −(v, ∆v)L2 (K) + (v, nK ·∇v)L2 (∂K)
k k
= −(ÎK (v)K , ∆v)L2 (K) + (ÎK (v)∂K , nK ·∇v)L2 (∂K) = 0,
k k
where we used the definitions of ÎK and of VK . Hence, v is constant on K and since ΠkK (v) = 0,
the mean value of v in K vanishes. Thus, v = 0.
(2) Consider the map ΦK : V̂Kk → VK k
s.t. for all v̂K = (vK , v∂K ) ∈ V̂Kk , the function ΦK (v̂K ) is the
unique solution in H (K) of the well-posed Neumann problem −∆(ΦK (v̂K )) = vK − v̄K + |∂K|
1
|K| v̄∂K
in K, −nK ·∇(ΦK (v̂K )) = v∂K on ∂K, and (ΦK (v̂K ) − vK , 1)L2 (K) = 0, where v̄K and v̄∂K denote
k
the mean value of vK and v∂K on K and ∂K, respectively. By definition, ΦK (v̂K ) ∈ VK . Moreover,
ΦK (v̂K ) is clearly injective.
k
(3) Combining Steps (1) and (2), the rank nullity theorem implies dim(VK ) = dim(V̂Kk ) and
k k k
ÎK|V k : VK → V̂K is an isomorphism.
K
k
Let Σ be the collection of the following linear forms acting on VK :
Z
1
f
σF,m (v) := v (ζm ◦ TF−1 ) ds, ∀F ∈ FK , (39.11a)
|F | F
Z
c 1 −1
σm (v) := v (ψm ◦ TK ) dx, (39.11b)
|K| K
k+d−1

where {ζm }m∈{1: nfsh } is a basis of Pk,d−1 with nfsh := dim(Pk,d−1 ) = d−1 and {ψm }m∈{1: ncsh }

is a basis of Pk,d with ncsh := dim(Pk,d ) = d+k
d .
k
Lemma 39.7 (Finite element). The triple (K, VK , Σ) is a finite element.
Proof. Direct consequence of Lemma 39.6.
39.2 Discrete problem

We now show how to assemble the discrete problem, how to reduce its size by static condensation,
and how the HHO and HDG methods are connected. Let D be a Lipschitz polyhedron in Rd and
f ∈ L2 (D). The model problem is

(39.12)
a(u, w) = ℓ(w), ∀w ∈ V,
with a(v, w) := (∇v, ∇w)L2 (D) and ℓ(w) := (f, w)L2 (D) for all v, w ∈ V.
39.2.1 Assembling and well-posedness

Let k ≥ 0. Let Th be a member of a shape-regular family of affine simplicial meshes. Let Fh =
Fh◦ ∪ Fh∂ be the collection of the faces of Th , where Fh◦ is the collection of the mesh interfaces and
Fh∂ the collection of the mesh boundary faces. Let us set V̂hk := VTkh ×VFkh , where
VTkh := {vTh ∈ L2 (D) | vTh |K ∈ VKk , ∀K ∈ Th }, (39.13a)

VFkh 2
:= {vFh ∈ L (Fh ) | vFh |∂K ∈ k
V∂K , ∀K ∈ Th }. (39.13b)
Note that functions in VFkh are single-valued on the mesh interfaces. For every pair v̂h :=
(vTh , vFh ) ∈ V̂hk and all K ∈ Th , it is convenient to define vK := vTh |K and v∂K := vFh |∂K ,
so that v̂K := (vK , v∂K ) ∈ V̂Kk . For all K ∈ Th , we define the local forms âK and ℓK s.t.
âK (v̂K , ŵK ) := (∇R(v̂K ), ∇R(ŵK ))L2 (K) + h−1

K (S(v̂K ), S(ŵK ))L2 (∂K) ,
ℓK (wK ) := (f, wK )L2 (K) ,
for all v̂K , ŵK ∈ V̂Kk . We define the global forms âh and ℓh s.t.
X X
âh (v̂h , ŵh ) := âK (v̂K , ŵK ), ℓh (wTh ) := ℓK (wK ), (39.14)
K∈Th K∈Th
for all v̂h , ŵh ∈ V̂hk . Notice that only the cell component of the test function is used to evaluate
ℓh . We enforce strongly the homogeneous Dirichlet boundary condition by zeroing out the discrete
unknowns associated with the boundary faces, i.e., we consider
k
V̂h,0 := VTkh ×VFkh ,0 , VFkh ,0 := {vFh ∈ VFkh | vFh |F := 0, ∀F ∈ Fh∂ }. (39.15)
The discrete problem is as follows:

(
k
Find ûh ∈ V̂h,0 such that
k
(39.16)
âh (ûh , ŵh ) = ℓh (wTh ), ∀ŵh ∈ V̂h,0 .
In other words, the HHO method produces a discrete solution having two components: a piecewise
polynomial function in the mesh cells and a piecewise polynomial function on the mesh faces. The
second component is single-valued at the mesh interfaces, vanishes at the boundary faces, and its
value can jump from one interface to a neighboring one.
To establish the well-posedness of (39.16), we prove that the bilinear form âh is coercive on
k
V̂h,0 . We equip this space with the norm
X
kv̂h k2V̂ k := |v̂K |2V̂ k , k
∀v̂h ∈ V̂h,0 . (39.17)
h,0 K
K∈Th
The only nontrivial property to verify that we have indeed defined a norm is that kv̂h kV̂ k = 0
h,0
k
implies v̂h = (0, 0). Let vh ∈ V̂h,0 be s.t. kv̂h kV̂ k = 0, i.e., |v̂K |V̂ k = 0 for all K ∈ Th . Then
h,0 K
recalling (39.6) we infer that vK and v∂K are constant functions taking the same value in each
mesh cell. On cells having a boundary face, this value must be zero since vFh vanishes on the
boundary faces. We can repeat the argument for the cells sharing an interface with those cells,
and we can move inward and reach all the cells in Th by repeating this process a finite number of
times. Thus, v̂h = (0, 0).
k
Lemma 39.8 (Coercivity, well-posedness). (i) The bilinear form âh is coercive on V̂h,0 . (ii)
The discrete problem (39.16) is well-posed.
Proof. The coercivity of âh follows by summing the lower bound from Lemma 39.2 over the mesh
cells, which yields
âh (v̂h , v̂h ) ≥ α kv̂h k2V̂ k , k
∀v̂h ∈ V̂h,0 . (39.18)
h,0
Well-posedness is a consequence of the Lax–Milgram lemma.
Remark 39.9 (Finite element viewpoint). The role of the stabilization in the HHO method can
also be understood by taking inspiration from the ideas at the heart of the virtual element method
k
(Beirão da Veiga et al. [50]). Since manipulating functions v ∈ VK (see (39.10)) is unpractical
because these functions are not known explicitly, one would like to manipulate only the projection
EK (v) ∈ VKk+1 which is computable from the dofs {σ(v)}σ∈Σ of v, that is, from the polynomial
k
pair ÎK (v) ∈ V̂Kk . The local bilinear form on VKk
×VK k
is aK (v, w) := (∇EK (v), ∇EK (w))L2 (K) +
−1
hK (S(ÎK (v), S(ÎK (v))L2 (∂K) . To prove that aK (v, v) controls k∇vk2L2 (K) , one needs to control
k k
k∇(v − EK (v))k2L2 (K) , and this is where the stabilization comes into play (see Exercise 39.2).
39.2.2 Static condensation and global transmission problem

The problem (39.16) can be solved by using a Schur complement technique consisting of locally
eliminating all the cell unknowns (this technique is also known as static condensation; see §28.1.2).
In other words, (39.16) can be reformulated in the form of local problems patched together by
a global transmission problem. To see this, we define Uµ ∈ VKk for all µ ∈ V∂K k
, and we define
k 2
Ur ∈ VK for all r ∈ L (K), s.t.
âK ((Uµ , 0), (q, 0)) := −âK ((0, µ), (q, 0)), ∀q ∈ VKk , (39.19a)
âK ((Ur , 0), (q, 0)) := (r, q)L2 (K) , ∀q ∈ VKk . (39.19b)
These problems are well-posed since âK is coercive on VKk ×{0}.

k
Proposition 39.10 (Transmission problem). The pair ûh ∈ V̂h,0 solves (39.16) iff uK =
k
Uu∂K + Uf|K for all K ∈ Th , and uFh ∈ VFh ,0 solves the following global transmission problem: For
all wFh ∈ VFkh ,0 ,
X X
âK ((Uu∂K , u∂K ), (Uw∂K , w∂K )) = ℓK (Uw∂K ). (39.20)
K∈Th K∈Th
Proof. Assume that ûh solves (39.16). Let K ∈ Th and wK ∈ VKk . Since âK ((uK , u∂K ), (wK , 0)) =
ℓK (wK ) = âK ((Uf|K , 0), (wK , 0)), we infer that
âK ((uK − Uf|K , u∂K ), (wK , 0)) = âK ((uK , u∂K ) − (Uf|K , 0), (wK , 0))
= 0 = âK ((Uu∂K , u∂K ), (wK , 0)),
k
showing that uK − Uf|K = Uu∂K . This implies that for all w∂K ∈ V∂K ,
âK ((Uu∂K , u∂K ), (Uw∂K , w∂K )) = âK ((uK , u∂K ), (Uw∂K , w∂K )) − âK ((Uf|K , 0), (Uw∂K , w∂K ))
= âK ((uK , u∂K ), (Uw∂K , w∂K )),
where we used the symmetry of âK and that âK (Uw∂K , w∂K ), (q, 0)) = 0 for all q ∈ VKk . Summing
over K ∈ Th shows that uFh solves the transmission problem (39.20). The converse statement is
proved in Exercise 39.6.
Remark 39.11 (Transmission problem). Following Cockburn [130], one can show that the
problem (39.12) can be reformulated as a transmission problem. Let aK (φ, ψ) := (∇φ, ∇ψ)L2 (K)
and ℓK (ψ) := (f, ψ)L2 (K) be the restrictions to K of the exact forms a and ℓ. There is a unique
1
lifting Uµ ∈ H 1 (K) for all µ ∈ H 2 (∂K) s.t. Uµ|∂K := µ and aK (Uµ , ψ) := 0 for all ψ ∈ H01 (K).
Similarly, there is a unique lifting Ur ∈ H01 (K) for all r ∈ L2 (K) s.t. Ur|∂K := 0 and aK (Ur , ψ) :=
(r, ψ)L2 (K) for all ψ ∈ H01 (K). For all v ∈ H 1 (D), we slightly abuse the notation by writing v|Fh
for the restriction of v to the mesh faces, and for every function λ defined on the mesh faces, we
write λ∂K for its restriction to the boundary of any mesh cell K ∈ Th . Since the weak solution u
is in H01 (D), u|Fh is in the trace space Λ defined as
1
Λ := {λ ∈ L2 (Fh ) | λ∂K ∈ H 2 (∂K), ∀K ∈ Th ; λ|Fh∂ = 0}. (39.21)
By definition, functions in Λ are single-valued on the mesh interfaces. Then the function u is
the weak solution iff there exists λ ∈ Λ s.t. u|K = Uλ∂K + Uf|K for all K ∈ Th and λ solves the
transmission problem
X X
aK (Uλ∂K , Uµ∂K ) = ℓK (Uµ∂K ), ∀µ ∈ Λ. (39.22)
K∈Th K∈Th
Moreover, assuming that nF ·(∇u)|F is in L2 (F ) for all F ∈ Fh (this is the case if u ∈ H 1+r (D),
r > 12 ), we have for all µ ∈ Λ,
X X
([[∇u]]F ·nF , µ)L2 (F ) = aK (Uλ∂K , Uµ∂K ) − ℓK (Uµ∂K ) = 0. (39.23)
F ∈Fh◦ K∈Th
The identity (39.23) shows that the global transmission problem (39.22) expresses the continuity
of the normal component of ∇u across the mesh interfaces. In conclusion, the problem (39.22)
consists of seeking λ ∈ Λ such that [[Uf + Uλ ]]F = 0 and nF ·[[∇(Uf + Uλ )]]F = 0 for all F ∈ Fh◦ ,
(Uf + Uλ )|∂D = 0, and −∆(Uf + Uλ )|K = f|K for all K ∈ Th . We refer the reader to Exercise 39.6
for more details.
Remark 39.12 (Finite element viewpoint). Recalling §39.1.3, we define the high-order Crou-
zeix–Raviart-type finite element space
Vhk := {v ∈ L2 (D) | v|K ∈ VK

k
, ∀K ∈ Th , ([[v]], q ◦ TF−1 )L2 (F ) = 0, ∀q ∈ Pk,d−1 , ∀F ∈ Fh◦ },
and
k
Vh,0 := {v ∈ Vhk | (v, q ◦ TF−1 )L2 (F ) = 0, ∀q ∈ Pk,d−1 , ∀F ∈ Fh∂ }.
P
For all vh , wh ∈ Vhk , let ah (vh , wh ) := K∈Th aK (vh , wh ), with aK defined in Remark 39.9, and
P k
ℓh (wh ) := K∈Th (f, wh )L2 (K) . Then the problem consisting of seeking uh ∈ Vh,0 s.t. ah (uh , wh ) =
k k
ℓh (wh ) for all wh ∈ Vh,0 is well-posed, and Îk (uh|K ) is the HHO solution in K.
39.2.3 Comparison with HDG and flux recovery

In this section, we compare the HHO method with the hybridizable discontinuous Galerkin (HDG)
method. In the HDG method, one approximates a triple, whereas one approximates a pair in the
HHO method. Let us consider the dual variable σ := −∇u (sometimes called flux), the primal
variable u, and its trace λ := u|Fh on the mesh faces. HDG methods approximate the triple
(σ, u, λ) by introducing local spaces SK , VK , and VF for all K ∈ Th and all F ∈ Fh◦ , and by
defining a numerical flux trace that includes a stabilization operator. Let us define the broken
spaces
STh := {τTh ∈ L2 (D) | τK ∈ SK , ∀K ∈ Th }, (39.24a)

2
VTh := {vTh ∈ L (D) | vK ∈ VK , ∀K ∈ Th }, (39.24b)
2
VFh := {µFh ∈ L (Fh ) | µF ∈ VF , ∀F ∈ Fh }, (39.24c)
with τK := τTh |K , vK := vTh |K , and µF := µFh |F , and let us set VFh ,0 := {µFh ∈ VFh | µFh |F =
0, ∀F ∈ Fh∂ }. The HDG method consists of seeking (σTh , uTh , λFh ) ∈ STh ×VTh ×VFh ,0 s.t. the
following holds true for all (τK , wK , µF ) ∈ SK ×VK ×VF , all K ∈ Th , and all F ∈ Fh◦ :
(σK , τK )L2 (K) − (uK , ∇·τK )L2 (K) + (λ∂K , τK ·nK )L2 (∂K) = 0, (39.25a)
− (σK , ∇wK )L2 (K) + (φ∂K ·nK , wK )L2 (∂K) = (f, wK )L2 (K) , (39.25b)
([[φh ]]·nF , µF )L2 (F ) = 0, (39.25c)
where the numerical flux trace φ∂K is defined by
φ∂K := σK|∂K + τ∂K (uK|∂K − λ∂K )nK on ∂K, (39.26)
for all K ∈ Th , with λ∂K := (λF )F ∈FK and τ∂K is a linear stabilization operator. The equa-
tion (39.25a) is the discrete counterpart of σ = −∇u, the equation (39.25b) that of ∇·σ = f , and
the equation (39.25c) weakly enforces the continuity of the normal component of the numerical
flux trace. Various HDG methods are realized by choosing the local spaces SK , VK , VF , and the
stabilization operator τ∂K . In general, the stabilization operator τ∂K in the HDG method acts
pointwise on ∂K. We will see in Proposition 39.13 that in the HHO method τ∂K acts collectively
on ∂K.
k k
Let S̃ : V∂K → V∂K be s.t. S̃(µ) := Πk∂K (µ − ((I − ΠkK )R(0, µ))|∂K ). The HHO stabilization
operator satisfies S(v̂K ) = S̃(vK|∂K − v∂K ); see (39.5). By definition, the adjoint of S̃, say S̃∗ :
k
V∂K → V∂K k
, is s.t. (S̃∗ (λ), µ)L2 (∂K) := (λ, S̃(µ))L2 (∂K) for all λ, µ ∈ V∂K
k
.
−1
Proposition 39.13 (HHO vs. HDG). Let SK := ∇VKk+1 , VK := Pk,d ◦TK , VF := Pk,d−1 ◦TF−1 ,
and τ∂K := h−1 ∗ ◦
K S̃ ◦ S̃ for all K ∈ Th and all F ∈ Fh . (i) If ûh := (uTh , uFh ) solves the HHO problem
(39.16), then (σTh , uTh , uFh ) solves the HDG problem (39.25) with σK := −∇R(ûK ) for all K ∈ Th .
(ii) Conversely, if (σTh , uTh , λFh ) solves the HDG problem (39.25), then σK = −∇R(uK , λ∂K ) for
all K ∈ Th , and (uTh , λFh ) solves the HHO problem (39.16).
Proof. We only prove the forward statement since the proof of the converse statement follows by
k
similar arguments. Let ûh ∈ V̂h,0 solve (39.16). Let σTh be s.t. σK := −∇R(ûK ) for all K ∈ Th .
Note that σK ∈ SK . For all τK := ∇q ∈ SK = ∇VKk+1 with q ∈ VKk+1 , using the definition (39.2)
of R shows that σK solves (39.25a). Let now wK be an arbitrary function in VKk . Since (39.2)
implies that (σK , ∇R(wK , 0))L2 (K) = (σK , ∇wK )L2 (K) − (σK ·nK , wK )L2 (∂K) , we have
(f, wK )L2 (K) = âK (ûK , (wK , 0))
= −(σK , ∇R(wK , 0))L2 (K) + h−1
K (S̃(uK|∂K − u∂K ), S̃(wK|∂K ))L2 (∂K)
= −(σK , ∇wK )L2 (K) + (σK ·nK + τ∂K (uK|∂K − u∂K ), wK )L2 (∂K) .
This shows that (39.25b) holds true with φ∂K defined in (39.26). Finally, let µF be an arbitrary
function in VF and let us denote by µ eF the extension by zero of µF to all the faces in Fh except F .
By definition, we have (∇R(0, µ eF ), ∇wK )L2 (K) = (µF , nK ·∇wK )L2 (F ) for all wK ∈ VK . Hence,
letting TF := {K ∈ Th | F ∈ FK }, the proof of (39.25c) follows from
X
0=− eF ))
âK (ûK , (0, µ
K∈TF
X
= eF ))L2 (K) + h−1
(σK , ∇R(0, µ K (S̃(uK|∂K − u∂K ), S̃(e
µF ))L2 (∂K)
K∈TF
X
= (φh|∂K ·nK , µ
eF )L2 (∂K) = ([[φh ]]·nF , µF )L2 (F ) .
K∈TF
Remark 39.14 (Flux recovery). Proposition 39.13 shows that one can post-process the HHO
method by computing the numerical flux traces
φ∂K (ûK ) := −∇R(ûK )|∂K + h−1 ∗
K (S̃ S̃(uK|∂K − u∂K ))nK . (39.27)
Defining the global flux trace φh (ûh )|∂K := φ∂K (ûK ) for all K ∈ Th gives ([[φh (ûh )]]·nF , µF )L2 (F ) =
0 for all µF ∈ VF . Since both factors are polynomials of degree at most k, we infer that
[[φh (ûh )]]·nF = 0. Finally, the above flux traces can be lifted as Raviart–Thomas vector-valued
functions defined in the mesh cells as was done for HDG methods in Cockburn et al. [135].
Remark 39.15 (Literature). HDG methods were introduced in Cockburn et al. [134]; see also
[130] for a review and [135] for the convergence analysis. The link between the HHO and HDG
methods is explored in Cockburn et al. [137]. With the simple polynomial spaces defined in (39.1),
the HHO stabilization operator yields optimal error estimates for all k ≥ 0 even on polyhedral
meshes. Achieving this result with the HDG method with the simpler operator Sk := Πk∂K (vK|∂K −
v∂K ) on polyhedral meshes requires a subtle design of the local spaces; see Cockburn et al. [136].
39.3 Error analysis

This section is devoted to the error analysis of the HHO method. We adopt a point of view
similar to that of Lemma 27.5, where the notions of stability and consistency/boundedness were
essential. Since stability has been established in Lemma 39.8, we now turn our attention to
consistency/boundedness. We slightly adapt the notion of the consistency error since the solution
to (39.12) is a function defined on D, whereas the discrete solution is a pair composed of a function
defined on D and a function defined on the mesh faces. Let Îhk : V := H01 (D) → V̂h,0 k
be the global
k k
interpolation operator s.t. for all v ∈ V, Îh (v) ∈ V̂h,0 is specified as follows: For all K ∈ Th ,
((Îhk (v))K , (Îhk (v))∂K ) := ÎK

k
(v|K ) = (ΠkK (v|K ), Πk∂K (v|∂K )) ∈ V̂Kk . (39.28)
Notice that Îhk (v) is well defined in V̂h,0

k
since v has zero jumps across the mesh interfaces (see
Theorem 18.8) and zero traces at the boundary faces. We define the consistency error δh (v̂h ) ∈
k ′ k
(V̂h,0 ) s.t. for all v̂h , ŵh ∈ V̂h,0 ,
hδh (v̂h ), ŵh i(V̂ k ′ ,V̂ k := ℓh (wTh ) − âh (v̂h , ŵh ). (39.29)
h,0 ) h,0
To avoid distracting technicalities, we assume in the error analysis that u ∈ H 1+r (D), r > 12 .
This assumption can be removed as discussed in §41.5. Recall from Lemma 39.1 the local elliptic
projection EK : H 1 (K) → VKk+1 .
Lemma 39.16 (Consistency/boundedness). Assume that the solution to (39.12) satisfies u ∈

H 1+r (D) ∩ H01 (D), r > 12 . There is c s.t. for all h ∈ H,
X
kδh (Îhk (u))k2(V̂ k ′ ≤c ku − EK (u)k2♯,K , (39.30)
h,0 )
K∈Th
where we defined for all K ∈ Th and all v ∈ H 1+r (K), r > 12 ,

1
kvk♯,K := k∇vkL2 (K) + hK
2
k∇vkL2 (∂K) . (39.31)
k
Proof. Let ŵh ∈ V̂h,0 . Integrating by parts cellwise, we observe that
X X
ℓh (wTh ) = ℓK (wK ) = −(∆u, wK )L2 (K)
K∈Th K∈Th
X
= (∇u, ∇wK )L2 (K) − (nK ·∇u, wK )L2 (∂K)
K∈Th
X
= (∇u, ∇wK )L2 (K) − (nK ·∇u, wK − w∂K )L2 (∂K) ,
K∈Th
P
where we used that K∈Th (nK ·∇u, w∂K )L2 (∂K) = 0 since ∇u and wFh are single-valued on the
k
mesh interfaces and wFh vanishes at the boundary faces. Moreover, since EK = R ◦ ÎK , using the
definition of R(ŵK ) leads to
k
(∇R(ÎK (u)), ∇R(ŵK ))L2 (K) = (∇EK (u), ∇R(ŵK ))L2 (K)
= (∇EK (u), ∇wK )L2 (K) − (nK ·∇EK (u), wK − w∂K )L2 (∂K) .
Using the definition of âK and since (∇(u − EK (u)), ∇wK )L2 (K) = 0, we have
X
hδh (Îhk (u)), ŵh i(V̂ k ′ k =− (T1,K + T2,K )
h,0 ) ,V̂h,0
K∈Th
with
T1,K := (nK ·∇ηK , wK − w∂K )L2 (∂K) ,

T2,K := h−1 k
K (S(ÎK (u)), S(ŵK ))L2 (∂K) ,
and ηK := u|K − EK (u) for all K ∈ Th . The Cauchy–Schwarz inequality and the definition of
|ŵK |V̂ k imply that
K
1
|T1,K | ≤ k∇ηK kL2 (∂K) kwK − w∂K kL2 (∂K) ≤ hK
2
k∇ηK kL2 (∂K) |ŵK |V̂ k .
K
−1 k −1
Moreover, |T2,K | ≤ hK 2 kS(ÎK (u))kL2 (∂K) hK 2 kS(ŵK )kL2 (∂K) owing to the Cauchy–Schwarz in-
equality. The first factor is bounded in Lemma 39.3, and the second one in Lemma 39.2. Hence,
|T2,K | ≤ ck∇ηK kL2 (K) |ŵK |V̂ k . Collecting these bounds and summing over the mesh cells leads
K
to (39.30).
Theorem 39.17 (Error estimate). Let u be the solution to (39.12) and let ûh := (uTh , uFh ) ∈
k
V̂h,0 solve (39.16). Assume that u ∈ H 1+r (D) ∩ H01 (D), r > 21 . Recall the notation uK := uTh |K ,
u∂K := uFh |∂K , and ûK := (uK , u∂K ) for all K ∈ Th . (i) There is c s.t. for all h ∈ H,
X X
k∇(u − R(ûK ))k2L2 (K) ≤ c ku − EK (u)k2♯,K . (39.32)
K∈Th K∈Th
(ii) Letting t := min(k + 1, r), we have

X X
k∇(u − R(ûK ))k2L2 (K) ≤ c h2t 2
K |u|H 1+t (K) . (39.33)
K∈Th K∈Th
Proof. (i) We adapt the proof of Lemma 27.5 to account for the use of the reconstruction operator.
Set ζ̂hk := Îhk (u) − ûh ∈ V̂h,0 k k
so that ζ̂K k
= ÎK (u|K ) − ûK for all K ∈ Th . Notice that ah (ζ̂hk , ζ̂hk ) =
−δh (Îhk (u)), ζ̂hk i(V̂ k )′ ,V̂ k . Then the coercivity property (39.18) implies that
h,0 h,0
X âh (ζ̂hk , ζ̂hk ) X

k
α k∇R(ζ̂K )k2L2 (K) ≤ k
k∇R(ζ̂K )k2L2 (K)
K∈Th kζ̂hk k2V̂ k K∈Th
h,0
hδh (Îhk (u)), ζ̂hk i2(V̂ k

âh (ζ̂hk , ζ̂hk )2 h,0 )
′ ,V̂ k
≤ = h,0
≤ kδh (Îhk (u))k2(V̂ k ′ .
kζ̂hk k2V̂ k kζ̂hk k2V̂ k h,0 )
h,0 h,0
P k
P
Lemma 39.16 yields K∈Th k∇R(ζ̂K )k2L2 (K) ≤ c K∈Th ku − EK (u)k2♯,K . Since R(ÎK
k
(u)) = EK (u)
for all K ∈ Th , we infer that
k
u|K − R(ûK ) = u|K − EK (u) + R(ζ̂K ).
Then the estimate (39.32) follows from the triangle inequality.

(ii) The estimate (39.33) results from the approximation properties of the local elliptic projection;
see Exercise 39.3.
Remark 39.18 (L2 -estimate). Improved L2 -error estimates can be established if elliptic regu-
larity pickup can be invoked; see [168].
Exercises
Exercise 39.1 (Stabilization). Prove that âK (v̂K , v̂K ) is equivalent to k∇rK k2L2 (K) + θ̂K (v̂K , v̂K )
for all v̂K ∈ V̂Kk , with rK := R(v̂K ) and
θ̂K (v̂K , v̂K ) := h−2 k 2 −1 k 2

K kvK − ΠK (rK )kL2 (K) + hK kv∂K − Π∂K (rK )kL2 (∂K) .
(Hint : note that S(v̂K ) = Πk∂K (vK − ΠkK (rK ))|∂K − (v∂K − Πk∂K (rK )), and to bound âK (v̂K , v̂K )
1 −1
from below, prove that θ̂K (v̂K , v̂K ) 2 ≤ c h−1
K kvK − rK kL2 (K) + hK kS(v̂K )kL2 (∂K) , then invoke the
2
Poincaré–Steklov inequality, the triangle inequality, and the lower bound from Lemma 39.2.)
k
Exercise 39.2 (Finite element viewpoint). Let VK be defined in (39.10). Let EK : H 1 (K) →
k+1 k
VK be the elliptic projection and set δ := v − EK (v) for all v ∈ VK . (i) Prove that
− 21
h−1 k k
K kΠK (δ)kL2 (K) ≤ c k∇EK (v)kL2 (K) + hK kS(ÎK (v))kL2 (∂K) .
(Hint : use the Poincaré–Steklov inequality in K and the lower bound from Lemma 39.2.) (ii)
Prove that
−1 k

k∇δkL2 (K) ≤ c k∇EK (v)kL2 (K) + hK 2 kS(ÎK (v))kL2 (∂K) .
(Hint : integrate by parts k∇δk2L2 (K) and accept as a fact that a discrete trace inequality and an
k k
inverse inequality are valid on VK , then use that S(ÎK (v)) = Πk∂K (ΠkK (δ)|∂K ) − Πk∂K (δ|∂K ).) (iii)
Let aK (v, w) := (∇EK (v), ∇EK (w))L2 (K) +h−1 k k k k
K (S(ÎK (v)), S(ÎK (v)))L2 (∂K) on VK ×VK . Prove that
2
aK (v, v) ≥ ck∇vkL2 (K) with c > 0.
Exercise 39.3 (Elliptic projection). Prove the second bound in Theorem 39.17. (Hint : intro-
duce the L2 -orthogonal projection Πk+1
K .)
−1
Exercise 39.4 (Reconstruction). (i) Let G : V̂Kk → VKk := P k,d ◦ TK be s.t. (G(v̂K ), q)L2 (K) =
k
−(vK , ∇·q)L2 (K) + (v∂K , nK ·q)L2 (∂K) for all q ∈ VK . Prove that Π∇V k+1 G = ∇R, where Π∇V k+1
K K
is the L2 -orthogonal projection onto ∇VKk+1 . (ii) Let Grt : V̂Kk → VKk := (ψK d −1 RT
) (RT k,d ) be
k d
s.t. (Grt (v̂K ), q)L2 (K) = −(vK , ∇·q)L2 (K) + (v∂K , nK ·q)L2 (∂K) for all q ∈ VK , where ψK is the
contravariant Piola transformation defined in (9.9c), and RT k,d is the Raviart–Thomas polynomial
space. Prove that kGrt (v̂K )kL2 (K) ≥ c|v̂K |V̂ k with c > 0. (Hint : use the dofs of the Raviart–
K
Thomas element; see John et al. [260] for the seminal idea in the context of dG methods.)
Exercise 39.5 (k = 0). (i) Derive the HHO method in 1D for k = 0, as well as the global trans-
P |F |
mission problem. (ii) Prove that, in dimension d ≥ 2 for k = 0, R(v̂K )(x) = vK + F ∈FK |K| (vF −
vK )nK|F ·(x − xK ) for all x ∈ K, with vF := v∂K|F for all F ∈ FK , and xK is the barycenter of
K, and S(v̂K )|F = vK − vF − ∇R(v̂K )·(xK − xF ), where xF is the barycenter of F for all F ∈ FK
−1
(Hint : any function q ∈ P1,d ◦ TK is of the form q(x) = qK + Gq ·(x − xK ), where qK := q(xK )
is the mean value of q over K and Gq := ∇q, and use also (7.1).)
Exercise 39.6 (Transmission problem). (i) Prove the converse statement in Proposition 39.10.
(Hint : write ŵK = (wK − Uw∂K , 0) + (Uw∂K , w∂K ).) (ii) Justify Remark 39.11. (Hint : for the
converse statement show that aK (u, w) − ℓK (w) = aK (Uλ∂K , Uµ ) − ℓK (Uµ ) with µ := w∂K .) (iii)
Adapt the statement if aK is nonsymmetric. (Hint : consider Uλ∗ ∈ H 1 (K) s.t. Uλ|∂K∗
= λ and
∗ 1
aK (ψ, Uλ ) = 0 for all ψ ∈ H0 (K).) (iv) Prove (39.23).
Exercise 39.7 (HDG).QConsider the HDG method. Assume the following: if (vK , µ∂K ) ∈
VK ×V∂K with V∂K := F ∈FK VF is s.t. (τ∂K (vK|∂K − µ∂K ), vK|∂K − µ∂K )L2 (∂K) = 0 and
(vK , ∇·τK )L2 (K) − (µ∂K , τK ·nK )L2 (∂K) = 0 for all τK ∈ SK , then vK and µ∂K are constant
functions taking the same value. Prove that the discrete problem (39.25) is well-posed. (Hint :
derive an energy identity.)
g
Exercise 39.8 (Space Λ). Let Λ be defined in (39.21). Recall that the trace map γ∂K : H 1 (K) →
1
H 2 (∂K) is surjective. (i) Prove that there are constants 0 < c1 ≤ c2 s.t. c1 k∇Uµ kL2 (K) ≤
1
|µ| 21 ≤ c2 k∇Uµ kL2 (K) for all µ ∈ H 2 (∂K), all K ∈ Th , and all h ∈ H. (Hint : prove first the
H (∂K)
b (ii) Set kλk2 := P 2
bounds on the reference cell K.) Λ K∈Th |λ∂K | 1 . Verify that k·kΛ indeed
H 2 (∂K)
defines a norm on Λ, and that Λ is a Hilbert space. (Hint : for all λ ∈ Λ, consider the function
Uλ : D → R s.t. Uλ|K := Uλ∂K for all K ∈ Th , and prove that Uλ ∈ H01 (D).)
Exercise 39.9 (Liftings, 1D). Consider a uniform mesh of D := (0, 1) with nodes xi := ih,
1
i := I+1 for all i ∈ {0:(I+1)}. Consider the PDE −u′′ = f in D with u(0) = u(1) = 0.
−1
R xi+1that (39.22) amounts to AX = B with A = h tridiag(−1, 2, −1), Xi = λi , and
(i) Prove
Bi = xi−1 ϕi f ds for all i ∈ {1:I}. (Hint : prove that Uλ is affine on every cell Ki = [xi−1 , xi ].)
Prove that λi = u(xi ). (Hint : write f = −u′′ and integrate by parts. This remarkable fact only
1
happens in 1D.) (ii) Let k ≥ 2. For all m ≥ 1, set φm := (2(2m + 1))− 2 (Lm+1 − Lm−1 ), where
Lm is the Legendre polynomial of degree m (see §6.1). Verify that {φm }m∈{1: k−1} is a basis of
R
P◦k := {p ∈ Pk | p(±1) = 0}. Prove that Uf|K b := [−1, 1] with the
(x) = Kb G(x, s)f (s) ds on K
P c
discrete Green’s function G(x, s) := m∈{1: k−1} φm (x)φm (s). (Hint : observe that φ′m = Lm .)
Infer the expression of Uf|Ki for every cell Ki .
Chapter 40
Contrasted diffusivity (I)
The goal of Chapters 40 and 41 is to investigate the approximation of a diffusion model problem
with contrasted diffusivity and revisit the error analysis of the various nonconforming approxi-
mation methods presented in the previous chapters. The essential difficulty is that the elliptic
regularity theory (see §31.4) tells us that the Sobolev smoothness index of the solution u may be
just barely larger than one. For this reason, we are going to perform the error analysis under the
assumption that u ∈ H 1+r (D), r > 0 (recall that we assumed r > 12 in the previous chapters).
This will be done by using again the abstract error analysis from §27.2, but the lack of smoothness
will require that we invoke the tools devised in Chapter 17 to give a proper meaning to the normal
derivative of the solution at the mesh faces. We assume that d ≥ 2 since the analysis is much
simpler if d = 1.
40.1 Model problem

Let D be a Lipschitz domain in Rd , which we assume for simplicity to be a polyhedron. We
consider the following scalar model problem:
−∇·(λ∇u) = f in D, γ g (u) = g on ∂D, (40.1)

1 1
with the trace map γ g : H 1 (D) → H 2 (∂D), the Dirichlet boundary data g ∈ H 2 (∂D), and the
scalar-valued diffusion coefficient λ ∈ L∞ (D) which we assume to be uniformly bounded from
below away from zero. For simplicity, we also assume that λ is piecewise constant in D, i.e., there
is a partition of D into M disjoint Lipschitz polyhedra {Di }i∈{1: M} s.t. λ|Di is a positive real
number for all i ∈ {1:M }. A central notion in this chapter is the diffusive flux which is defined as
follows:
σ(v) := −λ∇v ∈ L2 (D), ∀v ∈ H 1 (D). (40.2)
In the previous chapters, we considered elliptic PDEs with a source term f ∈ L2 (D). We are now
2d
going to relax a bit this hypothesis by only assuming that f ∈ Lq (D) with q > 2+d . Note that
2d
q > 1 since d ≥ 2. Since 2+d < 2, we are going to assume without loss of generality that q ≤ 2 in
the entire chapter. Readers who wish to simplify some arguments can think that q = 2 in what
follows.
194 Chapter 40. Contrasted diffusivity (I)
In the case of the homogeneous Dirichlet condition (g := 0), the weak formulation of the model
problem (40.1) is as follows:

(40.3)
a(u, w) = ℓ(w), ∀w ∈ V,
with the bilinear and linear forms

Z Z
a(v, w) := λ∇v·∇w dx, ℓ(w) := f w dx. (40.4)
D D
The bilinear form a is coercive on V owing to the Poincaré–Steklov inequality, and it is also
bounded on V ×V owing to the Cauchy–Schwarz inequality. The linear form ℓ is bounded on V
since the Sobolev embedding theorem (Theorem 2.31) and Hölder’s inequality imply that |ℓ(w)| ≤
kf kLq (D) kwkLq′ (D) ≤ ckf kLq (D) kwkH 1 (D) with q1 + q1′ = 1. Note that q ≥ 2+d
2d
is the minimal
integrability requirement on f for this boundedness property to hold true. The above coercivity
and boundedness properties combined with the Lax–Milgram lemma imply that (40.3) is well-
posed. For the non-homogeneous Dirichlet boundary condition, one invokes the surjectivity of the
trace map γ g to infer the existence of a lifting of g, say ug ∈ H 1 (D), and one decomposes the
solution to (40.1) as u := ug + u0 , where u0 ∈ H01 (D)R solves the weak problem (40.3) with ℓ(w)
replaced by ℓg (w) := ℓ(w) − e a(ug , w) with ea(u, v) := D λ∇v·∇w dx. The weak formulation thus
modified is well-posed since ℓg is bounded on H01 (D).
Lemma 40.1 (A priori regularity). If the solution to (40.3) is s.t. u ∈ H 1+r (D), r > 0, and if
2d
the source term f is in Lq (D), 2+d < q ≤ 2, then
u ∈ Vs := {v ∈ H01 (D) | σ(v) ∈ Lp (D), ∇·σ(v) ∈ Lq (D)}, (40.5)
where the real numbers p, q are such that

2d
2 < p, < q ≤ 2. (40.6)
2+d
Proof. The Sobolev embedding theorem implies that there is p > 2 s.t. H r (D) ֒→ Lp (D). Indeed,
2d 2d
if 2r < d, we have H r (D) ֒→ Ls (D) for all s ∈ [2, d−2r ] and we can take p := d−2r , whereas if
d
2r ≥ d, we have H r (D) ֒→ H 2 (D) ֒→ Ls (D) for all s ∈ [2, ∞), and we can take any p > 2. The
above argument implies that ∇u ∈ Lp (D), and since λ is piecewise constant and σ(u) = −λ∇u,
2d
we have σ(u) ∈ Lp (D). Since ∇·σ(u) = f and f ∈ Lq (D) with q > 2+d by assumption, we have
q
∇·σ(u) ∈ L (D).
The smoothness assumption u ∈ H 1+r (D), r > 0, is reasonable owing to the elliptic regularity
theory (see Theorem 31.36). In general, one expects that r ≤ 12 whenever u is supported in at least
two contiguous subdomains where λ takes different values since otherwise the normal derivative of
u would be continuous across the interface separating the two subdomains in question, and owing
to the discontinuity of λ, the normal component of the diffusive flux σ(u) would be discontinuous
across the interface, which would contradict the fact that σ(u) has a weak divergence. It is possible
that r > 12 when u is supported in one subdomain only. If r ≥ 1, we notice that we must have
f ∈ L2 (D) (since f|Di = −λ|Di (∆u)|Di for all i ∈ {1:M }), i.e., we can assume that q = 2 if r ≥ 1.
Remark 40.2 (Extensions). One can also consider lower-order terms in (40.1), e.g., −∇·(λ∇u)+
β·∇u + µu = f with β ∈ W 1,∞ (D) and µ ∈ L∞ (D) s.t. µ − 12 ∇·β ≥ 0 a.e. in D (for simplic-
ity). The present error analysis still applies provided the lower-order terms are not too large, e.g.,
λ ≥ max(hkβkL∞ (D) , h2 kµkL∞ (D) ), where h ∈ H denotes the meshsize. Stabilization techniques
like those discussed in Chapter 61 have to be invoked when the lower-order terms are large. Fur-
thermore, the error analysis can be extended to account for a piecewise constant tensor-valued
diffusivity d. Then the various constants in the error estimate depend on the square-root of the
anisotropy ratios measuring the contrast between the largest and the smallest eigenvalue of d in
each subdomain Di . Finally, one can consider that the diffusion tensor d is piecewise smooth
instead of being piecewise constant, and a reasonable requirement is that d|Di is Lipschitz for all
i ∈ {1:M }. Notice though that this last extension entails some subtleties in the analysis because
the discrete diffusive flux is no longer a piecewise polynomial function.
40.2 Discrete setting

We introduce in this section the discrete setting that we are going to use to approximate the
solution to (40.3). Let Th be a mesh from a shape-regular sequence. We assume that Th is oriented
in a generation-compatible way and that Th covers each of the subdomains {Di }i∈{1: M} exactly,
b Pb, Σ)
so that λK := λ|K is constant for all K ∈ Th . Let (K, b be the reference finite element. We
b
assume that Pk,d ⊂ P ⊂ W k+1,∞ b
(K) for some k ≥ 1. For all K ∈ Th , let TK : K b → K be the
g
geometric mapping and let ψK (v) := v ◦ TK be the pullback by the geometric mapping. Recall the
broken finite element space defined as
Pkb (Th ) := {vh ∈ L∞ (D) | vh|K ∈ PK , ∀K ∈ Th }, (40.7)

g −1 b
with the local space PK := (ψK ) (P ) ⊂ W k+1,∞ (K). For all vh ∈ Pkb (Th ), we define the broken
2
diffusive flux σ(vh ) ∈ L (D) by setting σ(vh )|K := −λK ∇(vh|K ) for all K ∈ Th . Recalling the
notion of broken gradient (see Definition 36.3), we have σ(vh ) := −λ∇h vh .
Recall that the set Fh◦ is the collection of the mesh interfaces and the set Fh∂ is the collection of
the mesh faces at the boundary. For all F ∈ Fh◦ , there are two cells Kl , Kr ∈ Th s.t. F := ∂Kl ∩∂Kr ,
and F is oriented by the unit normal vector nF pointing from Kl to Kr , i.e., nF := nKl = −nKr .
For all F ∈ Fh∂ , we write F := ∂Kl ∩ ∂D, and F is oriented by the unit normal vector pointing
toward the outside of D, i.e., nF := nKl = n. For all F ∈ Fh , let TF be the collection of the one
or two mesh cells sharing F , i.e., TF := {Kl , Kr } for all F ∈ Fh◦ and TF := {Kl } for all F ∈ Fh∂ .
For all K ∈ Th , let FK be the collection of the faces of K and let ǫK,F := nK ·nF = ±1. The
jump of a function v ∈ W 1,1 (Th ) across the mesh face F ∈ Fh is defined a.e. on F by setting
[[v]]F := v|Kl − v|Kr if F := ∂Kl ∩ ∂Kr ∈ Fh◦ (see §18.1.1) and [[v]]F := v|Kl if F := ∂Kl ∩ ∂D ∈ Fh∂ .
It is also useful to consider weighted averages at the mesh faces. For all F ∈ Fh◦ , we assume that
we have at hand two real numbers such that
θKl ,F , θKr ,F ∈ [0, 1] and θKl ,F + θKr ,F = 1. (40.8)
We then set
{v}F,θ := θKl ,F v|Kl + θKr ,F v|Kr , (40.9a)

{v}F,θ̄ := θKr ,F v|Kl + θKl ,F v|Kr . (40.9b)
Whenever θKl ,F = θKr ,F := 12 , these two definitions coincide with the usual arithmetic average (see
Definition 38.1). In order to use a common notation for interfaces and boundary faces, we write
for all F := ∂Kl ∩ ∂D ∈ Fh∂ , θKl ,F := 1 and {v}θ = v|Kl . We omit the subscript F in the jump
and the average whenever the context is unambiguous. The following identity (see Exercise 40.3)
will be useful:
[[vw]] = {v}θ [[w]] + [[v]]{w}θ̄ . (40.10)
40.3 The bilinear form n♯

In this section, we give a proper meaning to the normal trace of the diffusive flux of the solution to
(40.3) over each mesh face. To this purpose, we are going to rely on the face-to-cell lifting operator
introduced in §17.1.
40.3.1 Face localization of the normal diffusive flux

zd
Let p, q be two real numbers satisfying the requirement (40.6). Since z 7→ z+d is an increasing
p
ed
function, there is pe ∈ (2, p] such that q ≥ pe+d . With the three numbers p, q, pe in hand, we now
invoke the existence of a face-to-cell lifting operator that has been established in Lemma 17.1. Let
us recall this result for completeness.
Lemma 40.3 (Face-to-cell lifting). For every mesh cell K ∈ Th and every face F ∈ FK , there
1
p′
e ,e p′
exists a lifting operator LK
F :W
p (F ) → W 1,e (K) s.t.
g g
γ∂K (LK
F (φ))|∂K\F = 0, γ∂K (LK
F (φ))|F = φ, (40.11)
1
p′
e ,e
for all φ ∈ W p (F ). Moreover, there is c s.t.
d
−1+ d −1+d
hK |LK
p
F (φ)|W 1,p′ (K) + hK
q
kLK
F (φ)kLq′ (K) ≤ c hK
p
e p
e
kφk 1 ,p
p
e
e′ , (40.12)
W (F )
1 ′ 1
for all φ ∈ W pe ,e
p
(F ) with kφk 1 ,p
p
e
e′ := kφkLpe′ (F ) + hFpe |φ| 1 ,p
p
e
e′ , all K ∈ Th , all F ∈ FK , and
W (F ) W (F )
all h ∈ H.
Let K ∈ Th be a mesh cell and consider the functional space
S d (K) := {τ ∈ Lp (K) | ∇·τ ∈ Lq (K)}, (40.13)
where the superscript d refers to the divergence operator. We equip S d (K) with the following
dimensionally consistent norm:
1
1+d( p − q1 )
kτ kS d (K) := kτ kLp (K) + hK k∇·τ kLq (K) . (40.14)
With the lifting operator LK d
F in hand, the normal trace on any face F of K of any field τ ∈ S (K),
1 ′
,e
p ′
denoted by (τ ·nK )|F , is defined to be the linear form in (W pe (F )) whose action on any function
1
p′
φ ∈ W pe ,e (F ) is Z
h(τ ·nK )|F , φiF := τ ·∇LK K
F (φ) + (∇·τ )LF (φ) dx. (40.15)
K
1 ′ 1 ′
Here, h·, ·iF denotes the duality pairing between (W pe ,e
p
(F ))′ and W pe ,e
p
(F ). Notice that the right-
hand side of (40.15) is well defined owing to Hölder’s inequality and (40.12). Owing to (40.11),
we readily verify that
R we have indeed defined an extension of the normal trace since we have
h(τ ·nK )|F , φiF = F (τ ·nK )φ ds whenever the field τ is smooth. Let us now derive an important
bound on the linear form (τ ·nK )|F when it acts on a function from the space PF which is composed
1
p′
of the restrictions to F of the functions in PK . Notice that PF ⊂ W pe ,e (F ).
Lemma 40.4 (Bound on normal component). There is c s.t.

d( 1 − p
1
) −1
|h(τ ·nK )|F , φh iF | ≤ c hK 2 kτ kS d (K) hF 2 kφh kL2 (F ) , (40.16)
for all τ ∈ S d (K), all φh ∈ PF , all K ∈ Th , all F ∈ FK , and all h ∈ H.
Proof. A direct consequence of (40.15), Hölder’s inequality, and Lemma 40.3 is that the following
1
p′
holds true for all φ ∈ W pe ,e (F ):
− 1 +d( p
1 1
e−p)
|h(τ ·nK )|F , φiF | ≤ c hK pe kτ kS d (K) kφk 1 ,p
p
e
e′ .
W (F )
1
Since kφk 1 ,p
p
e′
e (F )
:= kφkLpe′ (F ) + hFpe |φ| 1 ,p
p
e
e′ , the assertion (40.16) follows from the inverse
W W (F )
(d−1)( 21 − p
1
e)
inequality kφh k 1 ,p
p
e
e′ ≤ chF kφh kL2 (F ) , which is valid for all φh ∈ PF , and the regularity
W (F )
of the mesh sequence.
40.3.2 Definition of n♯ and key identities

Let us consider the functional space Vs defined in (40.5). For all v ∈ Vs , Lemma 40.1 shows that
σ(v)|K ∈ S d (K) for all K ∈ Th , and Lemma 40.4 implies that it is possible to give a meaning by
duality to the normal component of σ(v)|K on all the faces of K separately. Since we have set
σ(vh )|K = −λK ∇(vh|K ) for all vh ∈ Pkb (Th ), and since we have PK ⊂ W k+1,∞ (K) with k ≥ 1, we
infer that σ(vh )|K ∈ S d (K) as well. Thus, σ(v)|K ∈ S d (K) for all v ∈ V♯b := Vs + Pkb (Th ). Let us
now introduce the bilinear form n♯ : V♯b ×Pkb (Th ) → R defined as follows:
X X
n♯ (v, wh ) := ǫK,F θK,F h(σ(v)|K ·nK )|F , [[wh ]]iF , (40.17)
F ∈Fh K∈TF
where the (yet to be defined) weights {θK,F }F ∈Fh ,K∈TF are assumed to satisfy (40.8). The def-
inition (40.17) is meaningful since [[wh ]]F ∈ PF for all wh ∈ Pkb (Th ). The factor ǫK,F in (40.17)
handles the relative orientation of nK and nF , whereas the weights θK,F will help achieve robust-
ness w.r.t. the diffusivity contrast. We will see in the next section how these weights must depend
on the diffusion coefficient.
The following lemma is fundamental to understand the role that the bilinear form n♯ will play
in the next section in the analysis of various nonconforming approximation methods. Recall the
definition (40.9) of the weighted average {·}θ .
Lemma 40.5 (Identities for n♯ ). For every choice of weights {θK,F }F ∈Fh ,K∈TF , we have
X Z
n♯ (vh , wh ) = {σ(vh )}θ ·nF [[wh ]] ds, (40.18a)
F ∈Fh F
X Z
n♯ (v, wh ) = σ(v)·∇wh|K + (∇·σ(v))wh|K dx, (40.18b)
K∈Th K
for all vh , wh ∈ Pkb (Th ) and all v ∈ Vs .
Proof. (1) Proof of (40.18a). Let vh , wh ∈ Pkb (Th ). Since the restriction of σ(vh ) to each mesh cell
is smooth, and since the restriction of LKF ([[wh ]]) to ∂K is nonzero only on the face F ∈ FK where
it coincides with [[wh ]], we have

Z
h(σ(vh )|K ·nK )|F , [[wh ]]iF = σ(vh )|K ·∇LK K
F ([[wh ]]) + (∇·σ(vh )|K )LF ([[wh ]]) dx
ZK Z
= σ(vh )|K ·nK LK
F ([[wh ]]) ds = σ(vh )|K ·nK [[wh ]] ds,
∂K F
where we used the divergence formula in K. After using the definitions of ǫK,F and of θK,F , we
obtain
X X Z
n♯ (vh , wh ) = ǫK,F θK,F σ(vh )|K ·nK [[wh ]] ds
F ∈Fh K∈TF F
X Z
= {σ(vh )}θ ·nF [[wh ]] ds.
F ∈Fh F
(2) Proof of (40.18b). Let v ∈ Vs and wh ∈ Pkb (Th ). Let Kδd : L1 (D) → C ∞ (D) and Kδb : L1 (D) →
C ∞ (D) be the mollification operators introduced in §23.1. Recall the following key commuting
property:
∇·(Kδd (τ )) = Kδb (∇·τ ), (40.19)
for all τ ∈ L1 (D) s.t. ∇·τ ∈ L1 (D). It is important to realize that this property can be applied to
σ(v) for all v ∈ Vs since ∇·σ(v) ∈ L1 (D) by definition of Vs . (Note that this property cannot be
applied to σ(vh ) with vh ∈ Pkb (Th ), since the normal component of σ(vh ) is in general discontinuous
across the mesh interfaces, i.e., σ(vh ) does not have a weak divergence; see Theorem 18.10.) Let
us consider the mollified bilinear form
X X
n♯δ (v, wh ) := ǫK,F θK,F h(Kδd (σ(v))|K ·nK )|F , [[wh ]]iF .
F ∈Fh K∈TF
Owing to the commuting property (40.19), we infer that

Z
h(Kδd (σ(v))|K ·nK )|F , [[wh ]]iF = Kδd (σ(v))·LK b K
F ([[wh ]]) + Kδ (∇·σ(v))LF ([[wh ]]) dx.
K
Theorem 23.4 implies that

Z
lim Kδd (σ(v))·LK b K
F ([[wh ]]) + Kδ ((∇·σ(v)))LF ([[wh ]]) dx =
δ→0 K
Z
σ(v)·LK F ([[wh ]]) + (∇·σ(v))L K
F ([[wh ]]) dx = h(σ(v)|K ·nK )|F , [[wh ]]iF .
K
Summing over the mesh faces and the associated mesh cells, we infer that
lim n♯δ (v, wh ) = n♯ (v, wh ).

δ→0
Since the mollified function Kδd (σ(v)) is smooth, by repeating the calculation done in Step (1), we
also have
X Z
n♯δ (v, wh ) = {Kδd (σ(v))}θ ·nF [[wh ]] ds.
F ∈Fh F
Using the identity (40.10), [[Kδd (σ(v))]]·nF = 0 for all F ∈ Fh◦ , the divergence formula in K, and
the commuting property (40.19), we obtain
X Z X Z
n♯δ (v, wh ) = {Kδd (σ(v))}θ ·nF [[wh ]] ds + [[Kδd (σ(v))]]·nF {wh }θ̄ ds
F ∈Fh F F ∈Fh◦ F
X Z X Z
= [[wh Kδd (σ(v))]]·nF ds = Kδd (σ(v))·nK wh|K ds
F ∈Fh F K∈T ∂K
h
X Z
d b
= Kδ (σ(v))·∇wh|K + Kδ (∇·σ(v))wh|K dx.
K∈Th K
Invoking again Theorem 23.4 leads to the assertion since

X Z
lim n♯δ (v, wh ) = σ(v)·∇wh|K + (∇·σ(v))wh|K dx.
δ→0 K
K∈Th
Remark 40.6 (Identity (40.18b)). We are going to use the identity (40.18b) to assert that σ(v)·n
is continuous across the mesh interfaces without assuming that v is smooth, say v ∈ H 1+r (D) with
r > 12 .
We now establish an important boundedness estimate on the bilinear form n♯ . Since σ(v)|K ∈
S d (K) for all K ∈ Th and all v ∈ V♯b , we can equip the space V♯b with the seminorm
X 2d( 1 − 1 ) 2d( 2+d − 1 )

|v|2n♯ := λ−1
K hK
2 p
kσ(v)|K k2Lp (K) + hK 2d q k∇·σ(v)|K k2Lq (K) . (40.20)
K∈Th
We notice that this seminorm is dimensionally consistent with the classical energy-norm defined as
P 2
1 1
− 12 d( 2 − p )
K∈Th λK k∇v|K kL2 (K) . The reader is invited to verify that |v|♯ ≤ cλ♭ (ℓD kσ(v)kLp (D) +
d( 2+d 1
2d − q )
ℓD k∇·σ(v)kLq (D) ), for all v ∈ Vs ; see Exercise 40.2.
In order to get robust error estimates with respect to λ, we need to avoid any dependency
on the ratio of the values taken by λ in two adjacent subdomains since otherwise the error es-
timates become meaningless when the diffusion coefficient λ is highly contrasted. To avoid such
dependencies, we introduce the following diffusion-dependent weights for all F := ∂Kl ∩∂Kr ∈ Fh◦ :
λKr λKl
θKl ,F := , θKr ,F := . (40.21)
λKl + λKr λKl + λKr
We also define
2λKl λKr
λF := if F ∈ Fh◦ and λF := λKl if F ∈ Fh∂ . (40.22)
λKl + λKr
The two key properties we are going to use are that, for all F ∈ Fh and all K ∈ TF , |TF |λK θK,F =
λF and λF ≤ |TF | minK∈TF λK (recall that |TF | is the cardinality of the set TF ).
Lemma 40.7 (Boundedness of n♯ ). With the weights defined in (40.21) and λF defined in
(40.22) for all F ∈ Fh , there is c, uniform w.r.t. λ, s.t. for all v ∈ V♯b , all wh ∈ Pkb (Th ), and all
h ∈ H,
X 12
−1 2
|n♯ (v, wh )| ≤ c |v|n♯ λF hF k[[wh ]]kL2 (F ) . (40.23)
F ∈Fh
Proof. Let v ∈ Vs + Pkb (Th ) and wh ∈ Pkb (Th ). Owing to the definition (40.17) of n♯ and the
estimate (40.16) from Lemma 40.4, we infer that
X X 1
d( 1 − p ) −1
|n♯ (v, wh )| ≤ c θK,F hK 2 kσ(v)|K kS d (K) hF 2 k[[wh ]]kL2 (F )
F ∈Fh K∈TF
X X
−1
1
d( 1 − p ) 1 1
−1
≤c λK 2 hK 2 kσ(v)|K kLp (K) |TF |− 2 λF2 hF 2 k[[wh ]]kL2 (F )
F ∈Fh K∈TF
X X
−1 d( 2+d 1
2d − q ) 1 1
−1
+ λK 2 hK k∇·σ(v)|K kLq (K) |TF |− 2 λF2 hF 2 k[[wh ]]kL2 (F ) ,
F ∈Fh K∈TF
1
where we used that θK,F ≤ θK,F2
(since θK,F ≤ 1), |TF |λK θK,F = λF , the definition of k·kS d (K) ,
1 1 2+d 1
and 1 + d( 2 − q ) = d( 2d − q ). Owing to the Cauchy–Schwarz inequality, we infer that
X X 1
X 1
X 1
aK |TF |− 2 bF ≤ ( |FK |a2K ) 2 ( b2F ) 2 ,
F ∈Fh K∈TF K∈Th F ∈Fh
for all real numbers {aK }K∈Th , {bF }F ∈Fh , where we used that
X X X X
(·) = (·)
F ∈Fh K∈TF K∈Th F ∈FK
for the term involving the aK ’s. Since |FK | is uniformly bounded (|FK | = d + 1 for simplicial
meshes), applying this bound to the two terms composing the estimate on |n♯ (v, wh )| leads to the
expected boundedness property.
Remark 40.8 (Literature). Diffusion-dependent averages have been introduced in Dryja [173]
for discontinuous Galerkin methods and have been analyzed in various contexts in Burman and
Zunino [101], Dryja et al. [174], Di Pietro et al. [167], Ern et al. [194].
Exercises
Exercise 40.1 (Normal flux). Let σ ∈ {τ ∈ Lp (K) | ∇·τ ∈ L2 (K)}, p > 2. Let γ∂K d
(σ) ∈
− 12 d
R R 1
H (∂K) be s.t. hγ∂K (σ), φi∂K := K σ·∇v(φ) dx + K (∇·σ)v(φ) dx for all φ ∈ H 2 (∂K), where
1 g d
P ∈ H (K) is a lifting of φ, i.e., γ∂K (v(φ)) = φ (see (4.12)). Prove that hγ∂K (σ), φi∂K =
v(φ)
F ∈FK h(σ·nK )|F , φ|F iF . (Hint : reason as in the proof of (40.18b).)
−1 d( 1 − p1 )
Exercise 40.2 (Bound on |v|♯ ). Prove that for all v ∈ Vs , |v|n♯ ≤ cλ♭ 2 (ℓD 2 kσ(v)kLp (D) +
d( 2+d − 1 ) p
ℓD 2d q k∇·σ(v)kLq (D) ). (Hint : for the sum with L -norms, use Hölder’s inequality after observ-
P 1 P 1
ing that hdK ≤ c|K|, and for the sum with Lq norms, use that ( K∈Th atK ) t ≤ ( K∈Th asK ) s for
real numbers t ≥ s.)
Exercise 40.3 (Jump identity). Let F := ∂Kl ∩ ∂Kr ∈ Fh◦ . Let θl , θr ∈ [0, 1] be s.t. θl + θr = 1.
Set {a}θ := θl al + θr ar and {a}θ̄ := θr al + θl ar . (i) Show that [[ab]] = {a}θ̄ [[b]] + [[a]]{b}θ . (ii) Show
that [[ab]] = {a}θ [[b]] + [[a]]{b}θ̄ .
Chapter 41
Contrasted diffusivity (II)
In this chapter, we continue the study of the model elliptic problem (40.3) with contrasted diffu-
sivity. Now that we have in hand our key tool, that is, the bilinear form n♯ introduced in §40.3,
we perform the error analysis when the model problem (40.3) is approximated by one of the
nonconforming methods introduced previously, i.e., Crouzeix–Raviart finite elements, Nitsche’s
boundary penalty method, the discontinuous Galerkin (dG) method, and the hybrid high-order
(HHO) method.
41.1 Continuous and discrete settings

1
Recall that the model problem R (40.3) consists of seeking Ru ∈ V := H0 (D) s.t. a(u, w) = ℓ(w) for
all w ∈ V, with a(v, w) := D λ∇v·∇w dx and ℓ(w) := D f w dx. We assume that the solution
to (40.3) is in the functional space Vs defined in (40.5) with the real numbers p, q satisfying (40.6),
i.e.,
u ∈ Vs := {v ∈ H01 (D) | σ(v) ∈ Lp (D), ∇·σ(v) ∈ Lq (D)}, (41.1)
where σ(v) := −λ∇v and the real numbers p, q are such that
2d
2 < p, < q ≤ 2. (41.2)
2+d
Let (Th )h∈H be a shape-regular sequence of simplicial affine meshes so that each mesh covers
D exactly. Let k ≥ 1 and consider the broken polynomial space Pkb (Th ) defined in (40.7). The
discrete problem takes the generic form

(41.3)
ah (uh , wh ) = ℓh (wh ), ∀wh ∈ Vh ,
where the subspace Vh ⊂ Pkb (Th ) and the forms ah and ℓh depend on the approximation method.
For all the approximation methods, the error analysis relies on Lemma 27.5, and the main
issue is to prove consistency/boundedness. Recall from Definition 27.3 that the consistency error
is defined by setting
hδh (vh ), wh iVh′ ,Vh := ℓh (wh ) − ah (vh , wh ), ∀vh , wh ∈ Vh . (41.4)

202 Chapter 41. Contrasted diffusivity (II)
The key tool to prove consistency/boundedness is the bilinear form n♯ : V♯b ×Pkb (Th ) → R with
V♯b := Vs + Pkb (Th ) s.t.
X X
n♯ (v, wh ) := ǫK,F θK,F h(σ(v)|K ·nK )|F , [[wh ]]iF , (41.5)
F ∈Fh K∈TF
with the orientation factor ǫK,F := nK ·nF = ±1 and the diffusion-dependent weights s.t. for all
F := ∂Kl ∩ ∂Kr ∈ Fh◦ ,
λKr λKl
θKl ,F := , θKr ,F := , (41.6)
λKl + λKr λKl + λKr
with the convention θKl ,F := 1 for all F := ∂Kl ∩ ∂D ∈ Fh∂ . Notice that θKl ,F , θKr ,F ∈ [0, 1] and
θKl ,F + θKr ,F = 1. For all v ∈ W 1,1 (Th ), we define weighted averages a.e. on every face F ∈ Fh as
follows: If F ∈ Fh◦ ,
{v}F,θ := θKl ,F v|Kl + θKr ,F v|Kr , (41.7a)
{v}F,θ̄ := θKr ,F v|Kl + θKl ,F v|Kr , (41.7b)
and {v}θ := v|Kl if F := ∂Kl ∩ ∂D ∈ Fh∂ .
We omit the subscript F in the jump and the average
whenever the context is unambiguous.
The key properties of the bilinear form n♯ we are going to invoke are the following: For all
vh , wh ∈ Pkb (Th ) and all v ∈ Vs , we have (see Lemma 40.5)
X Z
n♯ (vh , wh ) = {σ(vh )}θ ·nF [[wh ]] ds, (41.8a)
F ∈Fh F
X Z
n♯ (v, wh ) = σ(v)·∇wh|K + (∇·σ(v))wh|K dx, (41.8b)
K∈Th K
and (see Lemma 40.7) there is c, uniform w.r.t. λ, s.t. for all v ∈ V♯b , all wh ∈ Pkb (Th ), and all
h ∈ H,
X 21
1 2
|n♯ (v, wh )| ≤ c |v|n♯ λF k[[wh ]]kL2 (F ) , (41.9)
hF
F ∈Fh
where |·|n♯ is defined in (40.20) and
2λKl λKr
λF := if F ∈ Fh◦ and λF := λKl if F ∈ Fh∂ . (41.10)
λKl + λKr
1
We consider the dimensionally consistent seminorm |v|2λ,p,q := kλ 2 ∇h vk2L2 (D) + |v|2n♯ for all v ∈ V♯b .
Since λ is piecewise constant, we have
X 2d( 1 − 1 ) 2d( d+2 − 1 )

|v|2λ,p,q := λK k∇(v|K )k2L2 (K) + hK 2 p k∇(v|K )k2Lp (K) + hK 2d q k∆(v|K )k2Lq (K) .
K∈Th
(41.11)
41.2 Crouzeix–Raviart approximation

We consider in this section the Crouzeix–Raviart finite element space introduced in Chapter 36,
that is, Z
n o
cr
P1,0 (Th ) := vh ∈ P1b (Th ) | [[vh ]]F ds = 0, ∀F ∈ Fh . (41.12)
F
cr
The discrete problem takes the form (41.3) with Vh := P1,0 (Th ) and
Z Z
ah (vh , wh ) := λ∇h vh ·∇h wh dx, ℓh (wh ) := f wh dx. (41.13)
D D
1
We equip Vh with the norm kvh kVh := kλ 2 ∇h vh kL2 (D) . Adapting Lemma 36.4 to the present
setting leads to the following result.
Lemma 41.1 (Coercivity, well-posedness). (i) The bilinear form ah is coercive on Vh with
constant α := 1. (ii) The discrete problem (41.3) is well-posed.
Let V♯ := Vs + Vh be equipped with the norm kvkV♯ := |v|λ,p,q with |v|λ,p,q defined in (41.11)
(this is indeed a norm on V♯ since |v|λ,p,q = 0 implies that v is piecewise constant and hence
vanishes identically owing to the definition of Vh ). Invoking inverse inequalities shows that there
is c♯ s.t. kvh kV♯ ≤ c♯ kvh kVh for all vh ∈ Vh and all h ∈ H, i.e., (27.5) holds true.
Lemma 41.2 (Consistency/boundedness). There is ω♯ , uniform w.r.t. u ∈ Vs and λ, s.t.

kδh (vh )kVh′ ≤ ω♯ ku − vh kV♯ for all vh ∈ Vh and all h ∈ H.
Proof. Let vh , wh ∈ Vh . Since Vh ⊂ Pkb (Th ), the identity (41.8a) implies that
X Z
n♯ (vh , wh ) = {σ(vh )}θ ·nF [[wh ]] ds = 0,
F ∈Fh F
because {σ(vh )}θ ·nF is constant on F . Invoking the identity (41.8b) with v := u and since
f = ∇·σ(u), we also have
Z
ℓh (wh ) = − σ(u)·∇h wh dx + n♯ (u, wh ).
D
Combining the above two identities and letting η := u − vh , we obtain

Z
hδh (vh ), wh iVh′ ,Vh = λ∇h η·∇h wh dx + n♯ (u, wh )
D
Z
= λ∇h η·∇h wh dx + n♯ (η, wh ).
D
The assertion follows

P by invoking the Cauchy–Schwarz inequality, the boundedness of n♯ (see (41.9)),
−1 2 2
and the bound λ h
F ∈Fh F F k[[w h ]]kL2 (F ) ≤ ckwh kVh which follows by adapting the proof of
Lemma 36.9 with v := 0 and using that λF ≤ |TF | minK∈TF λK .
Theorem 41.3 (Error estimate). Let u solve (40.3) and uh solve (41.3). Assume that u ∈
H 1+r (D), r > 0. (i) There is c, uniform w.r.t. λ, s.t. for all h ∈ H,

vh ∈Vh
(ii) Letting t := min(1, r), we have

X d+2 1
12
−1 2d( 2d − q )
ku − uh kV♯ ≤ c λK h2t 2
K |u|H 1+t (K) + λK hK kf k2Lq (K) . (41.15)
K∈Th
(ii) We bound the infimum in (41.14) by considering η := u − Ihcr (u). For all K ∈ Th , Lemma 36.1
implies that k∇(η|K )kL2 (K) ≤ chtK |u|H 1+t (K) . Moreover, invoking the embedding H t (K) b ֒→
b we obtain the bound (see (17.19))
Lp (K)
d( 1 − p
1
)
hK 2 k∇(η|K )kLp (K) ≤ c k∇(η|K )kL2 (K) + htK |∇(η|K )|H t (K) . (41.16)
Observing that |∇(η|K )|H t (K) = |u|H 1+t (K) since Ihcr (u) is affine on K and using again Lemma 36.1
d( 1 − p1 )
gives hK 2 k∇(η|K )kLp (K) ≤ chtK |u|H 1+t (K) . Finally, we have ∆(η|K ) = −λ−1
K f in K.
Remark 41.4 (Convergence). Note that the rightmost term in the estimate (41.15) converges
2d
as O(h) when q = 2. Note also that convergence is lost when q ≤ d+2 , which is somewhat natural
R
since in this case the linear form w 7→ D f w dx is no longer bounded on H 1 (D).
Remark 41.5 (Weights). Although the weights introduced in (40.21) are not used in the
Crouzeix–Raviart discretization, they play a role in the error analysis. More precisely, we used
the boundedness of the bilinear form n♯ together with λF ≤ |TF | minK∈TF λK in the proof of
Lemma 41.2.
41.3 Nitsche’s boundary penalty method

We consider in this section the boundary penalty method introduced in Chapter 37. Recall that
Vh := Pkg (Th ), k ≥ 1, i.e., Vh is H 1 -conforming with
Pkg (Th ) := {vh ∈ Pkb (Th ) | [[vh ]]F = 0, ∀F ∈ Fh◦ }. (41.17)
The discrete problem is (41.3) with Vh := Pkg (Th ),
X Z λKl

ah (vh , wh ) := a(vh , wh ) + σ(vh )·n + ̟0 vh wh ds, (41.18)
F hF
F ∈Fh∂
P λ R
and ℓh (wh ) := ℓ(wh )+ F ∈F ∂ ̟0 hKFl F gwh ds, where the exact forms a and ℓ are defined in (40.4),
h
F := ∂Kl ∩ ∂D, and the user-dependent penalty parameter ̟0 is yet to be chosen large enough.
1 P λ
We equip Vh with the norm kvh k2Vh := kλ 2 ∇vh k2L2 (D) +|vh |2∂ with |vh |2∂ := F ∈F ∂ hKFl kvh k2L2 (F ) .
h
−1
Recall the discrete trace inequality stating that there is cdt s.t. kn·∇vh kL2 (F ) ≤ cdt hF 2 k∇vh kL2 (Kl )
for all vh ∈ Vh , all F ∈ Fh∂ , and all h ∈ H. Let n∂ denote the maximum number of boundary faces
that a mesh cell can have (n∂ ≤ d for simplicial meshes), and let Th∂D be the collection of the mesh
cells having at least one boundary face.
Lemma 41.6 (Coercivity, well-posedness). Assume that the penalty parameter satisfies ̟0 >
1 2 ̟ − 14 n∂ c2dt
4 n∂ cdt . (i) ah is coercive on Vh with constant α
:= 0 1+̟ 0
> 0. (ii) The discrete problem (41.3)
is well-posed.
Proof. Adapting the proof of Lemma 37.2, we infer that for all vh ∈ Vh ,
! 21
X Z X
1
λKl (n·∇vh )vh ds ≤ n∂ cdt
2 2
λK k∇vh kL2 (K) |vh |∂ .

F ∈F ∂ F K∈T ∂D
h h
The rest of the proof is similar to that of Lemma 37.3.

Let V♯ := Vs + Vh be equipped with the norm kvk2V♯ := |v|2λ,p,q + |v|2∂ with |v|λ,p,q defined
in (41.11) (the summations involving the terms k∇(v|K )kLp (K) and k∆(v|K )kLq (K) in the definition
P λ
of |·|λ,p,q in (41.11) can be restricted here to K ∈ Th∂D ) and |v|2∂ := F ∈F ∂ hKFl kvk2L2 (F ) . Invoking
h
inverse inequalities shows that there is c♯ s.t. kvh kV♯ ≤ c♯ kvh kVh for all vh ∈ Vh and all h ∈ H, i.e.,
(27.5) holds true.
Lemma 41.7 (Consistency/boundedness). There is ω♯ , uniform w.r.t. u ∈ Vs and λ, but
depending on p and q, s.t. kδh (vh )kVh′ ≤ ω♯ ku − vh kV♯ for all vh ∈ Vh and all h ∈ H.
Proof. Let vh , wh ∈ Vh . Using the identity (41.8a) for n♯ , [[wh ]]F = 0 for all F ∈ Fh◦ (since Vh is
H 1 -conforming), and the definition of the weights at the boundary faces, we infer that n♯ (vh , wh ) =
P R P λKl R
F ∈Fh∂ F σ(vh )·nwh ds. Hence, ah (vh , wh ) = a(vh , wh ) + n♯ (vh , wh ) + F ∈Fh∂ ̟0 hF F vh wh ds.
RInvoking the identity (41.8b) for the exact solution ug and observing that f = ∇·σ(u), we infer that
D
f wh dx = a(u, wh ) + n♯ (u, wh ). Recalling that γ (u) = g, and letting η := u − vh , we obtain
X Z
λKl
hδh (vh ), wh iVh′ ,Vh = a(η, wh ) + n♯ (η, wh ) + ̟0 ηwh ds.
∂
hF F
F ∈Fh
We conclude using the Cauchy–Schwarz inequality and the boundedness of n♯ from (41.9), where
the summation in |v|n♯ can be restricted to the mesh cells in Th∂D since [[wh ]]F = 0 across all the
mesh interfaces.
Theorem 41.8 (Error estimate). Let u solve (40.3) and uh solve (41.3) with the penalty pa-
rameter ̟0 > 14 n∂ c2dt . Assume that u ∈ H 1+r (D), r > 0. (i) There is c, uniform w.r.t. λ, s.t. for
all h ∈ H,
vh ∈Vh
(ii) Letting t := min(r, k) and χt := 1 if t ≤ 1 and χt := 0 if t > 1, we have

X 12
χt 2d( d+2 −1)
K |u|H 1+t (ŤK ) + hK 2d q kf k2Lq (K) , (41.20)
λK
K∈Th
where ŤK is the collection of the mesh cells having at least a common vertex with K, and |u|H 1+t (ŤK )
can be replaced by |u|H 1+t (K) if 1 + t > d2 .
(ii) We bound the infimum in (41.19) by considering η := u − Ihg,av (u), where Ihg,av is the quasi-
interpolation operator introduced in §22.3. We take the polynomial degree of Ihg,av to be ℓ := ⌈t⌉
(recall that ⌈t⌉ is the smallest integer n ∈ N s.t. n ≥ t). Notice that ℓ ≥ 1 because r > 0 and k ≥ 1,
and ℓ ≤ k because t ≤ k. Hence, Ihg,av (u) ∈ Vh . We need to bound all the terms composing the
norm kηkV♯ . Owing to Theorem 22.6 (with m := 1), we have k∇(η|K )kL2 (K) ≤ chtK |u|H 1+t (ŤK ) for
−1
all K ∈ Th . Moreover, using Exercise 22.5, we have hF 2 kηkL2 (F ) ≤ chtKl |u|H 1+t (ŤK ) for all F ∈ Fh∂ .
l
d( 1 − 1 ) d( d+2 − 1 )
It remains to estimate hK 2 p k∇(η|K )kLp (K)
and for all K ∈ Th∂D .
hK 2d q k∆(η|K )kLq (K)
Using (41.16), the above bound on k∇(η|K )kL2 (K) , and |∇(η|K )|H t (K) = |∇u|H t (K) = |u|H 1+t (K)
d( 1 − p
1
)
since ℓ < 1 + t, we infer that hK 2 k∇(η|K )kLp (K) ≤ chtK |u|H 1+t (ŤK ) . Moreover, if t ≤ 1, we
have ℓ = 1 so that k∆(η|K )kLq (K) = k∆ukLq (K) = λ−1

K kf kLq (K) . But, if t > 1, we infer that r > 1
so that we can set q := 2 (recall that f|Di = −λ|Di (∆u)|Di for all i ∈ {1:M }, and u ∈ H 2 (D) if
r ≥ 1), and we estimate k∆(η|K )kL2 (K) by using Theorem 22.6 (with m := 2). Finally, if 1 + t > d2 ,
we can use the canonical interpolation operator Ihg instead of Ihg,av , and this allows us to replace
|u|H 1+t (ŤK ) by |u|H 1+t (K) in (41.20).
Remark 41.9 (Localization). One obtains the same upper bound as in (41.20) when using
conforming
P finite elements for the approximation,
P i.e., Vh ⊂ H01 (D); see Exercise 41.1. Notice also
2t 2
that K∈Th λK hK |u|H 1+t (Ť ) ≤ ckλkL∞ (D) K∈Th h2t 2
K |u|H 1+t (K) .
K
Remark 41.10 (Literature). An alternative analysis based on the approach of Gudi [226] is
developed in Lüthen et al. [291].
41.4 Discontinuous Galerkin

We consider in this section the symmetric interior penalty (SIP) discontinuous Galerkin method
introduced in Chapter 38 (i.e., θ := 1 in (38.20)). The discrete problem is (41.3) with Vh := Pkb (Th ),
k ≥ 1, the bilinear form
Z X Z
ah (vh , wh ) := λ∇h vh ·∇h wh dx + {σ(vh )}θ ·nF [[wh ]] ds
D F ∈Fh F
X Z X λF
Z
+ [[vh ]]{σ(wh )}θ ·nF ds + ̟0 [[vh ]][[wh ]] ds,
F hF F
F ∈Fh F ∈Fh
P λ R
and the linear form ℓh (wh ) := ℓ(wh ) + F ∈F ∂ ̟0 hKFl F gwh ds, where ℓ is defined in (40.4), and
h
the user-specified penalty parameter ̟0 is yet to be chosen P large enough. We equip Vh with the
1
norm kvh k2Vh := kλ 2 ∇h vh k2L2 (D) + |vh |2J with |vh |2J := F ∈Fh λhFF k[[vh ]]k2L2 (F ) . Recall the discrete
−1
trace inequality stating that there is cdt s.t. knF ·∇vh kL2 (F ) ≤ cdt hF 2 k∇vh kL2 (K) for all vh ∈ Vh ,
all K ∈ Th , all F ∈ FK , and all h ∈ H. Let n∂ denote the maximum number of faces that a mesh
cell can have (n∂ ≤ d + 1 for simplicial meshes).
Lemma 41.11 (Coercivity, well-posedness). Assume that the penalty parameter satisfies ̟0 >
̟ −n c2
n∂ c2dt . (i) ah is coercive on Vh with constant α := 01+̟∂0 dt > 0. (ii) The discrete problem (41.3)
is well-posed.
Proof. Let vh ∈ Vh . Our starting observation is that
1
ah (vh , vh ) = kλ 2 ∇h vh k2L2 (D) + 2n♯ (vh , vh ) + ̟0 |vh |2J .
Proceeding as in the proof of Lemma 38.5 and Lemma 38.6, and using that |TF |θK,F λK λ−1
F = 1
for all F ∈ Fh and all K ∈ TF , the reader is invited to verify that
1 1
|n♯ (vh , wh )| ≤ n∂2 cdt kλ 2 ∇h vh kL2 (D) |wh |J . (41.21)
We can then conclude as in the proof of Lemma 38.6.
Let V♯ := Vs + Vh be equipped with the norm kvk2V♯ := |v|2λ,p,q + |v|2J with |v|λ,p,q defined
P
in (41.11) and |v|2J := F ∈Fh λhFF k[[v]]k2L2 (F ) . Invoking inverse inequalities shows that there is c♯
s.t. kvh kV♯ ≤ c♯ kvh kVh for all vh ∈ Vh and all h ∈ H, i.e., (27.5) holds true.
Lemma 41.12 (Consistency/boundedness). There is ω♯ , uniform w.r.t. u ∈ Vs and λ, s.t.

kδh (vh )kVh′ ≤ ω♯ ku − vh kV♯ for all vh ∈ Vh and all h ∈ H.
, wh ∈ Vh . Owing to the identity (41.8b) and since f = ∇·σ(u), we infer that
RProof. Let vhP
D f wh dx = K∈Th −(σ(u), ∇h wh )L2 (K) + n♯ (u, wh ). Hence, we have
Z X Z
λF
ℓh (wh ) = n♯ (u, wh ) − σ(u)·∇h wh dx + ̟0 gwh ds.
D ∂ F hF
F ∈Fh
Using the identity (41.8a), we obtain

Z
ah (vh , wh ) = −σ(vh )·∇h wh dx + n♯ (vh , wh )
D
X Z X Z
λF
− [[vh ]]{σ(wh )}θ ·nF ds + ̟0 [[vh ]][[wh ]] ds.
F h
F F
F ∈Fh F ∈Fh
Setting η := u − vh and using that [[u]]F = 0 for all F ∈ Fh◦ and [[u]]F = g for all F ∈ Fh∂ , we obtain
Z
hδh (vh ), wh iVh′ ,Vh = λ∇η·∇h wh dx + n♯ (η, wh )
D
X Z X Z
λF
− [[η]]{σ(wh )}θ ·nF ds + ̟0 [[η]][[wh ]] ds.
F
F ∈Fh F hF
F ∈Fh
We conclude by using the Cauchy–Schwarz inequality for the first and the fourth terms on the
right-hand side, using the boundedness estimate on n♯ from (41.9) for the second term, and by
proceeding as in the proof of (41.21) to bound the third term.
Theorem 41.13 (Error estimate). Let u solve (40.3) and uh solve (41.3) with the penalty
parameter ̟0 > n∂ c2dt . Assume that u ∈ H 1+r (D), r > 0. (i) There is c, uniform w.r.t. λ, s.t. for
all h ∈ H,
vh ∈Vh
(ii) Letting t := min(r, k) and χt := 1 if t ≤ 1 and χt := 0 if t > 1, we have

X 12
χt 2d( d+2 −1)
K |u|H 1+t (K) + hK 2d q kf k2Lq (K) . (41.23)
λK
K∈Th
Proof. Proceed as in the proof of Theorem 41.8, where we now use the L1 -stable interpolation
operator Ih♯ : L1 (D) → Pkb (Th ) from §18.3 to estimate the best-approximation error.
41.5 The hybrid high-order method

We consider in this section the hybrid high-order (HHO) method from Chapter 39. The discrete
k :=
space V̂h,0 VTkh ×VFkh ,0 is defined in (39.15) with k ≥ 0. Recall that
VTkh := {vTh ∈ L2 (D) | vh|K ∈ VKk , ∀K ∈ Th }, (41.24a)

VFkh ,0 2
:= {vFh ∈ L (Fh ) | vFh |∂K ∈ k
V∂K , ∀K ∈ Th ; vFh |Fh∂ = 0}, (41.24b)
−1 Q −1
with VKk := Pk,d ◦ TK and V∂Kk :=
F ∈FK Pk,d−1 ◦ TF , where TK and TF are affine geometric
k
mappings. For every pair v̂h := (vTh , vFh ) ∈ V̂h,0 , vTh is a collection of cell polynomials of degree
at most k, and vFh is a collection of face polynomials of degree at most k which are single-valued at
the mesh interfaces and vanish at the boundary faces (so as to enforce strongly the homogeneous
Dirichlet condition). Recall the notation v̂K := (vK , v∂K ) ∈ V̂Kk := VKk ×V∂Kk
with vK := vTh |K and
v∂K := vFh |∂K for all K ∈ Th . The (λ-independent) local bilinear form âK on V̂Kk ×V̂Kk is defined
as follows:
âK (v̂K , ŵK ) := (∇R(v̂K ), ∇R(ŵK ))L2 (K) + h−1
K (S(v̂K ), S(ŵK ))L2 (∂K) ,
with the local reconstruction and stabilization operators R and S defined in (39.2) and in (39.4),
respectively.
k
The discrete problem is as follows: Find ûh ∈ V̂h,0 s.t.
k
âh (ûh , ŵh ) = ℓh (wTh ), ∀ŵh ∈ V̂h,0 , (41.25)
with the forms
X X
âh (v̂h , ŵh ) := λK âK (v̂K , ŵK ), ℓh (wTh ) := (f, wK )L2 (K) .
K∈Th K∈Th
k
P
Recalling (39.6), we equip the discrete space V̂h,0 with the norm kv̂h k2V̂ k := K∈Th λK |v̂K |2V̂ k ,
h,0 K
i.e., we set
X
−1
kv̂h k2V̂ k := 2 2
λK k∇vK kL2 (K) + λK hK kvK − v∂K kL2 (∂K) . (41.26)
h,0
K∈Th
k
A straightforward consequence of Lemma 39.2 is that the bilinear form âh is coercive on V̂h,0 .
Owing to the Lax–Milgram lemma, the discrete problem (41.25) is, therefore, well-posed.
k
As in Chapter 39, the local interpolation operator ÎK : H 1 (K) → V̂Kk for all K ∈ Th is s.t.
k
ÎK (v) := (ΠkK (v), Πk∂K (v|∂K )) ∈ V̂Kk for all v ∈ H 1 (K), where ΠkK and Πk∂K are the L2 -orthogonal
projections onto VKk and V∂K k
, respectively. The local elliptic projection EK : H 1 (K) → VKk+1 :=
−1
Pk+1,d ◦ TK is s.t. (∇(EK (v) − v), ∇w)L2 (K) = 0 for all w ∈ VKk+1 , and (EK (v) − v, 1)L2 (K) = 0.
We define global counterparts of these operators, Îhk : H 1 (D) → VTkh ×VFkh and Eh : H 1 (D) →
b
Pk+1 (Th ), that are simply defined locally by setting Îhk (v)|K := ÎK k
(v|K ) and Eh (v)|K := EK (v|K ).
Recalling the duality pairing h·, ·iF defined in (40.15), the generalization to the HHO method
b k
of the bilinear form n♯ defined in (41.5) is the bilinear form defined on (Vs + Pk+1 (Th ))×V̂h,0 that
acts as follows: X X
n♯ (v, ŵh ) := h(σ(v)·nK )|F , (wK − w∂K )|F iF . (41.27)
K∈Th F ∈FK
We now establish the counterparts of the identities (41.8a)-(41.8b) and the boundedness estimate
(41.9).
Lemma 41.14 (Identities and boundedness for n♯ ). The following holds true for all ŵh ∈
k b
V̂h,0 , all vh ∈ Pk+1 (Th ), and all v ∈ Vs :
XZ
n♯ (vh , ŵh ) = λK ∇vh|K ·∇(R(ŵK ) − wK ) dx, (41.28a)
K∈Th K
X Z
n♯ (v, ŵh ) = σ(v)·∇wK + (∇·σ(v))wK dx. (41.28b)
K∈Th K
b k
Moreover, there is c, uniform w.r.t. λ, s.t. for all v ∈ Vs + Pk+1 (Th ), all ŵh ∈ V̂h,0 , and all h ∈ H,
X 21
−1
|n♯ (v, ŵh )| ≤ c |v|n♯ λK hK kwK − w∂K k2L2 (∂K) , (41.29)
K∈Th
with the seminorm |·|n♯ defined in (40.20).

Proof. See Exercise 41.3 for the proof of (41.28). The proof of (41.29) uses the same arguments
as the proof of Lemma 40.7.
Remark 41.15 ((41.28b)). The right-hand side of (41.28b) does not depend on the face-based
function wFh . This identity replaces the argument in the proof of Lemma 39.16 invoking the
continuity of the normal component of σ(u) across the mesh interfaces, which makes sense only
when the solution to (40.3) is smooth enough, say σ(u) ∈ H r (D) with r > 12 .
b
Let V♯ := Vs + Pk+1 (Th ) be equipped with the seminorm kvkV♯ := |v|λ,p,q with |v|λ,p,q defined
in (41.11). Note that kvkV♯ = 0 implies that v = 0 if v has zero mean value in each mesh cell
K ∈ Th . This is the case for instance if one takes v := u − Eh(u). Recalling (39.29), the consistency
error is defined s.t. hδh (Îh (u)), ŵh i(V̂ k )′ ,V̂ k := ℓh (wTh ) − âh (Îhk (u), ŵh ) for all ŵh ∈ Vh,0
k
.
h,0 h,0
Lemma 41.16 (Consistency/boundedness). There is ω♯ , uniform w.r.t. u ∈ Vs and λ, s.t. for

all h ∈ H,
kδh (Îh (u))k(V̂ k )′ ≤ ω♯ ku − Eh (u)kV♯ . (41.30)
h,0
R
Proof.
P Since σ(u) = −λ∇u, ∇·σ(u) = f , and u ∈ Vs , (41.28b) implies that D f wh dx =
K∈Th aK (u, wK ) + n♯ (u, ŵh ), where aK (u, wK ) := (−σ(u), ∇wK )L2 (K) . This implies that
X
ℓ̂h (ŵh ) = aK (u, wK ) + n♯ (u, ŵh ).
K∈Th
k
Using first the definition of âh , then the identity R◦ ÎK = EK , and finally (41.28a) with vh := Eh (u),
we obtain
X X
âh (Îhk (u), ŵh ) = aK (EK (u), wK ) + n♯ (Eh (u), ŵh ) + λK h−1 k
K (S(ÎK (u)), S(ŵK ))L2 (∂K) .
K∈Th K∈Th
Subtracting these two identities and using that aK (u − EK (u), wK ) = 0 for all K ∈ Th leads to
hδh (Îhk (u)), ŵh i(V̂ k )′ ,V̂ k = T1 + T2 with
h,0 h,0
T1 := n♯ (u − EK (u), ŵh ),
X
T2 := − λK h−1 k
K (S(ÎK (u)), S(ŵK )L2 (∂K) .
K∈Th
P
To bound T1 , we invoke (41.29) and use K∈Th λK h−1 2 2
K kwK − w∂K kL2 (∂K) ≤ kŵh kV̂ k owing
h,0
to (41.26). We bound T2 as in the proof of Lemma 39.16.
k
Theorem 41.17 (Error estimate). Let u solve (40.3) and ûh ∈ V̂h,0 solve (41.25). Assume that
1+r
u∈H (D), r > 0. (i) There is c, uniform w.r.t. λ, s.t. for all h ∈ H,
X
λK k∇(u − R(ûK )k2L2 (K) ≤ c ku − Eh (u)k2V♯ . (41.31)
K∈Th
(ii) Let t := min(r, k + 1) and χt := 1 if t ≤ 1 and χt := 0 if t > 1. We have

X X χt 2d( d+2 −1)
λK k∇(u − R(ûK ))k2L2 (K) ≤ c λK h2t 2
K |u|H 1+t (K) + hK 2d q kf k2Lq (K) . (41.32)
λK
K∈Th K∈Th

Remark 41.18 (HHO vs. dG). The HHO method is somewhat simpler than the dG method
when it comes to solving problems with contrasted coefficients. For the HHO method, one assem-
bles cellwise the local bilinear forms âK weighted by the local diffusion coefficient λK , whereas for
the dG method one has to invoke interface-based values of the diffusion coefficient to construct the
penalty term.
Exercises
Exercise 41.1 (Conforming finite elements). Consider the approximation of (40.3) by con-
g
forming finite elements. Let V := H01 (D), Vh := Pk,0 (Th ) ⊂ V, k ≥ 1, and consider the norm
1
1+r
kvkV := kλ 2 ∇vkL2 (D) . Assume u ∈ H (D), r > 0, and set t := min(r, k). Prove that there is
P 1
c, uniform w.r.t. λ, s.t. ku − uh kV ≤ c( K∈Th λK h2t 2
K |u|H 1+t (ŤK ) ) for all h ∈ H, where ŤK is the
2
collection of the mesh cells sharing at least a vertex with K, and that |u|H 1+t (ŤK ) can be replaced
by |u|H 1+t (K) if 1 + t > d2 .
Exercise 41.2 (dG). Prove the estimate (41.21).
Exercise 41.3 (HHO). (i) Prove (41.28a) (Hint : adapt the proof of (40.18a), i.e., use the
definition of the pairing h·, ·iF together with the definition (39.2) for R). (ii) Prove (41.28b).
(Hint : adapt the proof of (40.18b). (iii) Prove the error bound (41.31). (Hint : see the proof
of (39.32) in Theorem 39.17.) (iv) Prove (41.32). (Hint : set ℓ = ⌈t⌉ and consider the elliptic
ℓ
projection of degree ℓ, say EK , for all K ∈ Th .)
Chapter 42
Linear elasticity
The four chapters composing Part IX deal with the approximation of vector-valued elliptic PDEs
endowed with a multicomponent coercivity property either in H 1 (linear elasticity) or in H(curl)
(Maxwell’s equations in some specific regimes). The present chapter is concerned with the linear
elasticity equations where the main tool to establish coercivity is Korn’s inequality. We consider
H 1 -conforming and nonconforming approximations, and we address the robustness of the approx-
imation in the incompressible limit.
42.1 Continuum mechanics

Let D be a Lipschitz domain in Rd , d = 3. We assume that D represents a deformable medium,
initially at equilibrium, that is subjected to an external load f : D → R3 . Our goal is to determine
the displacement field u : D → R3 induced by f once the system has reached equilibrium again.
Let s : D → R3×3 be the stress tensor in the medium. We write s(u) since this tensor depends on
the displacement field. The equilibrium conditions under the external load f can be expressed as
∇·s(u) + f = 0 in D, (42.1)
and the balance of the angular momentum requires that s(u) be symmetric, i.e., s(u) = s(u)T .
We assume that the deformations are small enough so that the linear elasticity theory applies. Let
e(u) : D → R3×3 be the (linearized) strain rate tensor defined as
1
e(u) := (∇u + ∇uT ). (42.2)
2
In the framework of linear isotropic elasticity, the stress tensor is related to the strain rate tensor
by the relation
s(u) = 2µe(u) + λ tr(e(u))Id , (42.3)
where λ and µ are the phenomenological parameters called Lamé coefficients, and Id is the identity
tensor in Rd×d . Using (42.2), we also have
s(u) = µ(∇u + ∇uT ) + λ(∇·u)Id .
Owing to thermodynamic stability, the Lamé coefficients are such that µ > 0 and λ + 32 µ > 0. We
henceforth assume that there are µmin , κmin > 0 s.t.
µ, λ ∈ L∞ (D), µ(x) ≥ µmin , λ(x) + 23 µ(x) ≥ κmin , a.e. x in D. (42.4)
212 Chapter 42. Linear elasticity
42.1 (Cauchy–Navier). If µ and λ are constant in D, the identity ∇·e(u) = 2 ∆u +

1
Remark
∇(∇·u) implies that (42.1) can be rewritten −µ∆u − (µ + λ)∇(∇·u) = f in D. This PDE is
called Cauchy–Navier formulation in linear elasticity.
Remark 42.2 (Incompressibility). The coefficient κ := λ + 23 µ, called bulk modulus, describes

the compressibility of the material. Very large values w.r.t. µ, i.e., λ ≫ µ, correspond to almost
incompressible materials.
Remark 42.3 (Material parameters). Instead of using λ and µ, it is sometimes more convenient
to consider the Young modulus, E, and the Poisson coefficient, ν, defined as follows:
3λ + 2µ λ
E := µ ν := . (42.5)
λ+µ 2(λ + µ)
The Poisson coefficient is such that −1 < ν < 12 . An almost incompressible material corresponds
to a Poisson coefficient very close to 21 .
Remark 42.4 (Linearity). The linear isotropic elasticity model is in general valid for problems
involving infinitesimal strains. In this case, the medium responds linearly to externally applied
loads so that one can normalize the problem and consider arbitrary loads.
Remark 42.5 (A bit of history). The finite element method was originally developed in the
1950s by aeronautical engineers to solve problems of continuum mechanics that could not be
easily handled by classical finite difference techniques since they involved complex geometries; see,
e.g., Levy [282], Argyris and Kelsey [14], and the references cited in Oden [317]. At the same
time, theoretical researches on the approximation of the linear elasticity equations were carried
out by Turner et al. [367], and eventually in 1960, Clough [129] coined the terminology “finite
elements”.
Definition 42.6 (Rigid displacement). A rigid displacement r : D → R3 is a global motion

of the medium D consisting of a translation and a rotation, i.e., r is a member of the following
six-dimensional vector space:
R := P 0,3 + x×P
P0,3 = N 0,3 , (42.6)
where N 0,3 is the lowest-order Nédélec polynomial space defined in §15.1.
Lemma 42.7 (Kernel of strain rate). For all r ∈ L1loc (D), r ∈ R iff e(r) = 0.
Proof. Let r ∈ R. Then ∇r is skew-symmetric so that e(r) = 0. Conversely, let r ∈ L1loc (D) be
such that e := e(r) = 0. Since ∂ij rk = ∂ji rk in the distribution sense for all i, j, k ∈ {1:3}, we
have
∂k (∂j ri ) = ∂k eij + 21 ∂k ∂j ri − 12 ∂k ∂i rj
= ∂k eij + 21 ∂j (∂k ri + ∂i rk ) − 21 ∂i (∂k rj + ∂j rk )
= ∂k eij + ∂j eik − ∂i ejk = 0.
This implies that all the Cartesian components of r are first-order polynomials, i.e., r(x) = α+Bx,
with α ∈ R3 and B ∈ R3×3 . Moreover, e(r) = 0 implies that B + BT = 0, i.e., the matrix B is
skew-symmetric. Therefore, there exists a vector β ∈ R3 such that Bx = β×x. Thus, r ∈ R.
Lemma 42.7 implies that if the displacement field u satisfies the equilibrium condition (42.1),
then u + r, where r is a rigid displacement, also satisfies this equation. We will see below that the
rigid displacements can be controlled by the boundary conditions.
Part IX. Vector-valued elliptic PDEs 213
42.2 Weak formulation and well-posedness

In this section, we present a weak formulation for the linear elasticity problem and establish its
well-posedness in the framework of the Lax–Milgram lemma. The key tool to prove coercivity are
Korn’s inequalities.
42.2.1 Weak formulation

The model problem (42.1)-(42.3) must be supplemented with boundary conditions. We consider a
boundary partition ∂D = ∂Dd ∪ ∂D n such that |∂Dd | > 0. The displacement is imposed to vanish
on ∂Dd (we say that the medium is clamped at ∂Dd ), and a normal load g : ∂Dn → R3 is imposed
on ∂Dn . This leads to the following problem:
∂Dn g
11
00
00
11
00
11 f u ∇·s(u) + f = 0 in D, (42.7a)
∂Dd 00
11
00
11 D s(u) = 2µe(u) + λ tr(e(u))Id in D, (42.7b)
00
11 u=0 on ∂Dd , (42.7c)
s(u)n = g on ∂Dn . (42.7d)
Clamping the medium at ∂Dd allows one to control the rigid displacements since the only field
r ∈ R such that r|∂Dd = 0 is r = 0 provided |∂Dd | > 0.
To derive a weak formulation for (42.7), we take the Rscalar product of the equilibrium
R equa-
tion
R with a smooth test function v : D → R 3
. Since D ∇· s(u) ·v dx = − D s(u):∇v dx +
∂D
v· s (u)n ds and s(u):∇v
P
= s (u):e (v) owing to the symmetry of s(u) (here the double dot
product is defined as e:s := j,k∈{1: d} ejk sjk = tr(e sT )), we have
Z Z Z
s(u):e(v) dx − v·s(u)n ds = f ·v dx.
D ∂D D
The displacement u and the test function v are taken in the functional space
Vd := {v ∈ H 1 (D) | γ g (v)|∂D d = 0}, (42.8)
1
where γ g : H 1 (D) → H 2 (∂D) acts componentwise as the trace map from Theorem 3.10, i.e.,
γ g (v) = v|∂D if v is a smooth function. Since the measure of ∂Dd is positive, the following
Poincaré–Steklov inequality holds true on Vd : There is C̃ps > 0 such that
C̃ps kvkL2 (D) ≤ ℓD k∇vkL2 (D) , ∀v ∈ Vd , (42.9)
where ℓD is a length scale associated with D, e.g., ℓD := diam(D). Therefore, Vd ∋ v 7→ kvkVd :=
k∇vkL2 (D) = |v|H 1 (D) is a norm on Vd . This norm is equivalent to the H 1 -norm in Vd since
−2 12 1
kvkVd ≤ ℓ−1 2 2 2
D kvkH 1 (D) ≤ (1 + C̃ps ) kvkVd , where kvkH 1 (D) := (kvkL2 (D) + ℓD k∇vkL2 (D) ) for all
2
v ∈ Vd . A possible weak formulation of (42.7) is as follows:

Find u ∈ Vd such that
(42.10)
a(u, w) = ℓ(w), ∀w ∈ Vd ,
with the following bilinear and linear forms:
Z Z

a(v, w) := s(v):e(w) dx = 2µe(v):e(w) + λ(∇·v)(∇·w) dx,
ZD Z D
ℓ(w) := f ·w dx + g·w ds.

D ∂Dn
In the language of continuum mechanics, the test function w plays the role of a virtual displace-
ment, and the weak formulation
P (42.10) expresses the principle of virtual work. Moreover, recalling
that ke(v)k2ℓ2 = e(v):e(v) = i,j∈{1: d} |e(v)ij |2 , the quantity E(v) := 12 a(v, v) − ℓ(v), i.e.,
Z Z Z
1
E(v) := 2µke (v)k2ℓ2 + λ|∇·v| 2
dx − f ·v dx − g·v ds, (42.11)
2 D D ∂Dn
represents the total energy of the deformed medium at equilibrium. The quadratic terms corre-
spond to the energy of the elastic deformation, and the linear terms represent the potential energy
associated with the volume and boundary loads. Note that E(v) is not bounded from below over
the whole space H 1 (D) since a(r, r) = 0 for all r ∈ R and ℓ(r) may be arbitrarily large for some
rigid displacements. Proceeding as in §31.3.3 for scalar elliptic PDEs leads to the following result.
Proposition 42.8 (Weak solution). Let D be a Lipschitz domain in R3 with ∂D = ∂Dd ∪
∂Dn . Let f ∈ L2 (D) and g ∈ L2 (∂Dn ). If the function u ∈ Vd solves (42.10), then it satis-
fies (42.7a)-(42.7b) a.e. in D, (42.7c) a.e. on ∂Dd , and (42.7d) a.e. on ∂Dn in the sense that
R
hs(u)n, v f 12 (∂Dn ) := {v ∈ H 21 (∂Dn ) | v 1
ei − 21 12 = ∂Dn g·v ds for all v ∈ H e ∈ H 2 (∂D)}, where
H ,H
e is the zero extension of v to ∂D.
v
42.2.2 Korn’s inequalities and well-posedness

There are two Korn’s inequalities. These inequalities will be invoked to establish the coercivity of
the bilinear form a. The first one deals with the simpler situation where the displacement field
vanishes on the whole boundary. The second one does not say anything on the boundary values.
To unify the notation, we use the same symbol Ck to denote the constant associated with the first
and the second Korn inequality.
Theorem 42.9 (Korn’s first inequality). Let D be a Lipschitz domain in Rd . Setting Ck := √1 ,
2
the following holds true:
Ck k∇vkL2 (D) ≤ ke(v)kL2 (D) , ∀v ∈ H01 (D). (42.12)
Proof. Let v ∈ C0∞ (D). Since v vanishes at the boundary, we have

Z X Z X Z
2
∇v:∇v T dx = (∂i vj )(∂j vi ) dx = − (∂ij vj )vi dx
D i,j∈{1: d} D i,j∈{1: d} D
X Z Z
= (∂i vi )(∂j vj ) dx = (∇·v)2 dx ≥ 0.
i,j∈{1: d} D D
A density argument then shows that the above inequality holds true for all v ∈ H01 (D). As a
result, we infer that for all v ∈ H01 (D),
Z Z
1
e(v):e(v) dx = (∇v + ∇v T ):(∇v + ∇v T ) dx
D 4
ZD Z
1 1
= ∇v:∇v dx + ∇v:∇v T dx
2 D 2 D
Z
1 1
≥ ∇v:∇v dx = k∇vk2L2 (D) .
2 D 2
Theorem 42.10 (Korn’s second inequality). Let D be a Lipschitz domain in Rd . There is

Ck > 0 s.t.
Ck k∇vkL2 (D) ≤ ke(v)kL2 (D) + ℓ−1
D kvkL2 (D) , ∀v ∈ H 1 (D). (42.13)
Moreover, for every closed subspace V of H 1 (D) s.t. V ∩ R = {0}, there is Ck > 0 s.t.
Ck k∇vkL2 (D) ≤ ke(v)kL2 (D) , ∀v ∈ V . (42.14)
Proof. For (42.13), see Ciarlet [123, p. 11], Duvaut and Lions [183, p. 110], McLean [298, Thm. 10.2].
The inequality (42.14) results from (42.13) and the Peetre–Tartar lemma (Lemma A.20). Let us
define X := V , Y := L2 (D), A : X → Y with A(v) := e(v), Z := L2 (D), and let T be the
compact injection from X into Z. Lemma 42.7 implies that ker(A) ⊂ R. But V ∩ R = {0},
so that ker(A) = {0}, i.e., A is injective. Moreover, (42.13) implies that there is c > 0 s.t.
kvkH 1 (D) ≤ c(ℓD kA(v)kL2 (D) + kT (v)kL2 (D) ). Then the Peetre–Tartar lemma asserts that there
is c > 0 s.t. kvkH 1 (D) ≤ cℓD kA(v)kL2 (D) for all v ∈ X. Since ℓD k∇vkL2 (D) ≤ kvkH 1 (D) , this
proves (42.14).
Theorem 42.11 (Well-posedness). Let D be a Lipschitz domain in R3 with ∂D = ∂Dd ∪ ∂Dn .
Assume that |∂Dd | > 0. Let f ∈ L2 (D) and, if |∂Dn | > 0, let g ∈ L2 (∂Dn ). Let λ, µ satisfy (42.4).
(i) The problem (42.10) is well-posed. (ii) (42.10) is equivalent to the variational formulation
u = arg minv∈Vd E(v) with the energy functional E defined in (42.11).
Proof. (i) We apply the Lax–Milgram lemma. The linear form ℓ is bounded on H 1 (D), and
the boundedness of the bilinear form a on H 1 (D)×H 1 (D) is a consequence of the assumption
λ, µ ∈ L∞ (D). Let us verify the coercivity of a on Vd . If λ ≥ 0, we have 2µke(v)k2ℓ2 + λ|∇·v|2 ≥
P e(v)kℓ2 . 2If λ < inequality ke(v)k2ℓ2 ≥ 13 (∇·v)2 (which follows from ke(v)k2ℓ2 =
2
2µk P0, we use the P
i |∂i vi | ≥ 3 (∇·v) ) to infer that 2µke(v)kℓ2 + λ|∇·v| ≥
2 2 1 2 2 2
i,j |e(v)ij | ≥ i |e(v)ii | =
3κke(v)kℓ2 . Recalling the assumption (42.4), this shows that in all the cases we obtain
2
2µke(v)k2ℓ2 + λ|∇·v|2 ≥ ρmin ke(v)k2ℓ2 ,
with ρmin := 2µmin if λ ≥ 0 a.e. in D and ρmin := min(2µmin , 3κmin) otherwise. Notice that we
have ρmin > 0. The above bounds imply that
Z
a(v, v) = (2µke(v)k2ℓ2 + λ|∇·v|2 ) dx
D
Z
≥ ρmin ke(v)k2ℓ2 dx = ρmin ke(v)k2L2 (D) , ∀v ∈ Vd .
D
If ∂Dd = ∂D, then we have Vd := H01 (D), and we invoke Korn’s first inequality (see (42.12)).
Otherwise, we invoke Korn’s second inequality (see (42.14)) since |∂Dd | > 0 implies that Vd ∩ R =
{0}. In both cases, there is Ck > 0 s.t.
a(v, v) ≥ ρmin Ck2 k∇vk2L2 (D) = ρmin Ck2 kvk2Vd , ∀v ∈ Vd , (42.15)
i.e., a is Vd -coercive. This shows that the problem (42.10) is well-posed.

(ii) The equivalence of the problem (42.10) and the variational formulation minimizing the energy
functional E follows from Proposition 25.8, since the bilinear form a is symmetric.
Remark 42.12 (Regularity pickup). The elliptic regularity theory applies to the linear elas-
ticity equations with smooth Lamé parameters. In particular, if ∂D is smooth and either the
homogeneous Dirichlet or Neumann boundary condition is applied, there exist s ∈ ( 21 , 1] and csmo
s.t. kukH 1+s (D) ≤ csmo µ−1 2 2

min ℓD kf kL2 (D) for all f ∈ L (D); see e.g., Mazzucato and Nistor [297,
§10]. This property is not valid for mixed Dirichlet–Neumann boundary conditions, as already
illustrated in Exercise 31.5 for scalar elliptic PDEs. For nonsmooth domains and mixed Dirichlet–
Neumann boundary conditions, it is nevertheless possible to show that there exists s (usually in
−1 2
(0, 12 )) such that kukH 1+s (D) ≤ csmo µmin ℓD kf kL2 (D) .
Remark 42.13 (Pure traction). The pure traction (or Neumann) problem is obtained when
R n = ∂D Rand ∂Dd = ∅ in (42.7). In this case, the data must satisfy the compatibility condition
∂D
D
f ·r dx+ ∂D g·r ds R= 0 for all r ∈ R. Uniqueness
R of the weak solution is obtained by additionally
prescribing, e.g., that D v dx := 0 and D ∇×v dx := 0; see Exercise 42.3.
Remark 42.14 (Elasticity functionals). The weak formulation (42.10) is posed in terms of one
dependent variable: the displacement field. The strain and the stress fields are evaluated from the
displacement field by means of (42.2) and (42.3). As shown in Theorem 42.11, the weak solution
is the minimizer of the energy functional E defined in (42.11). It is possible to characterize the
solution of the linear elasticity equations by an optimality condition using other functionals. One
important example consists of using both the stress and the displacement fields as independent
variables and to look for a critical point of the Hellinger–Reissner functional (see [241, 333] and
Exercise 42.1). This approach constitutes the basis of the mixed stress-displacement finite element
methods in elasticity (see §42.4.2). Among other possible approaches are the three-field formulation
that consists of treating the displacement, the strain tensor, and the stress tensor as independent
variables (see Fraejis de Veubeke [205], Hu [248], Washizu [388]) and the intrinsic formulation
where the only dependent variable is the strain tensor (see Ciarlet and Ciarlet [125]).

Let D be a polyhedron in R3 . Let (Th )h∈H be a shape-regular sequence of affine matching meshes
so that each mesh covers D exactly. For simplicity, we assume that the material is clamped at ∂D,
i.e., ∂Dd = ∂D so that Vd := H01 (D). Let Pkg (Th ) := {vh ∈ C 0 (D) | vh|K ◦ TK ∈ Pb, ∀K ∈ Th }
be the scalar-valued H 1 -conforming finite element space constructed in §19.2.1, where k ≥ 1 is
the degree of the reference finite element (K,b Pb , Σ).
b g
Let Pk,0 (Th ) := Pkg (Th ) ∩ H01 (D) be the
corresponding H01 (D)-conforming subspace. We define the H01 -conforming approximation space
g g g
Vh0 := Pk,0 (Th ) × Pk,0 (Th ) × Pk,0 (Th ), (42.16)
and consider the following discrete problem:

Find uh ∈ Vh0 such that
(42.17)
a(uh , wh ) = ℓ(wh ), ∀wh ∈ Vh0 .
Since a is coercive and the approximation setting is conforming, i.e., Vh0 ⊂ Vd , the discrete
problem (42.17) is well-posed. Recalling the energy functional E defined in (42.11), we have
uh = arg minv∈Vh0 E(vh ) and E(uh ) ≥ E(u) owing to the conformity of the approximation setting.
g
Remark 42.15 (Collocation). Let {ϕa }a∈Ah be the global shape functions of Pk,0 (Th ), and
let {e1 , e2 , e3 } be the canonical basis of R3 . Then one can use {ei ϕa }a∈Ah ,i∈{1: 3} as the global
shape functions of Vh0 . With this choice, (42.17) leads to a collocalized scheme since the three
components of the discrete displacement field uh are associated with the same scalar-valued global
shape function.
Theorem 42.16 (Error estimate). Let the assumptions of Theorem 42.11 hold true. Let u
solve (42.10) and let uh solve (42.17). (i) There is c s.t.
|u − uh |H 1 (D) ≤ c inf |u − vh |H 1 (D) , (42.18)

vh ∈Vh0
for all h ∈ H and limh→0 |u − uh |H 1 (D) = 0. (ii) If u ∈ H 1+r (D) for some r ∈ (0, k], the following
holds true: X 12
|u − uh |H 1 (D) ≤ c h2r
K |u| 2
H 1+r (K) ≤ c hr |u|H 1+r (D) . (42.19)
K∈Th
(iii) Letting s be the index of the elliptic regularity pickup, we have

D |u − uh |H 1 (D) . (42.20)
Proof. The proof goes along the same lines as those presented in Chapter 32 for scalar elliptic
PDEs, i.e., (42.18) follows from Céa’s lemma (Lemma 26.13), (42.19) from the approximation
properties of the quasi-interpolation operator with zero trace from §22.4 (and the regularity of the
mesh sequence), and (42.20) from the Aubin–Nitsche lemma (Lemma 32.11).
A shortcoming of low-order H 1 -conforming finite elements is their poor performance when
approximating nearly-incompressible materials. This phenomenon is known in the literature as
volume or dilatation locking. Other types of locking can occur in linear elasticity problems, such
as shear locking in plate models when the plate thickness is very small. For simplicity, we focus on
volume locking and on how to avoid it. Nearly-incompressible materials are characterized by the
fact that the ratio µλ of the Lamé parameters is very large (or equivalently the Poisson coefficient
ν is very close to 21 , see (42.5)). In this situation, the displacement field is nearly divergence-free.
It has long been known that the H 1 -conforming approximation of nearly-incompressible mate-
rials on triangular meshes may not behave properly on meshes that are not fine enough if k = 1.
Moreover, the method converges sub-optimally for k ∈ {2, 3} and delivers optimal-order conver-
gence for k ≥ 4. On quadrilateral meshes, volume locking cannot be avoided for all k ≥ 1. We refer
the reader, e.g., to Vogelius [380], Scott and Vogelius [345], Babuška and Suri [40]. To understand
why H 1 -conforming finite elements may fail, let us inspect how the error estimate (42.18) depends
on the Lamé parameters µ and λ. To simplify the discussion, we assume that these parameters are
constant, and since we are concerned with the case µλ ≫ 1, we assume that λ is nonnegative. We
first observe that the bilinear form a is H01 -coercive with coercivity constant being proportional
to µ since Korn’s first inequality (see (42.12)) and λ ≥ 0 imply that
a(v, v) ≥ 2µke(v)k2L2 (D) ≥ µ|v|2H 1 (D) .
Moreover, since ke(v)kL2 (D) ≤ |v|H 1 (D) , the Cauchy–Schwarz inequality implies the following
boundedness property:
a(v, w) ≤ µ|v|H 1 (D) |w|H 1 (D) + λk∇·vkL2 (D) k∇·wkL2 (D) .
Following the proof of Céa’s lemma, we infer that

λ
|u − uh |H 1 (D) ≤ c inf |u − vh |H 1 (D) + k∇·(u − vh )kL2 (D) , (42.21)
vh ∈Vh0 µ
where the punchline is that, in contrast to (42.18), this error estimate features a constant c that is
uniform w.r.t. µ and λ. The first term on the right-hand side decays as the best-approximation error
√
of u in Vh0 . This is also the case for the second term (since k∇·(u−vh )kL2 (D) ≤ 3|u−vh |H 1 (D) ),
but the scaling by the multiplicative factor µλ ≫ 1 causes this term to be very large on practically
feasible meshes. In other words, the second term tends to zero as h → 0, but this asymptotic range
is only visible on meshes that are refined enough to beat the large constant µλ .
There are several possibilities to circumvent this bottleneck and to devise approximation meth-
ods that are robust w.r.t. volume locking. The first route consists of introducing the auxiliary
variable p := λ∇·u which plays the role of the pressure in the incompressible limit. The idea is
then to devise a mixed finite element approximation for the pair (u, p). This approach requires
some care in choosing the finite element spaces to approximate the displacement and the pressure,
and is analyzed in Chapters 53 and 54 in the context of the Stokes equations. One can also con-
sider mixed finite element methods that approximate both the stress and the displacement fields;
see §42.4.2. Another route consists of using a nonconforming approximation for the displacement
in such a way that the error on the approximation of ∇·u only depends on the smoothness of ∇·u.
Examples include the nonconforming finite element methods in Fortin and Soulié [204], Fortin
[202], Falk [199], Brenner and Sung [88], the discontinuous Galerkin methods in Hansbo and Lar-
son [239, 240], Wihler [395], Cockburn et al. [133], the hybridizable discontinuous Galerkin methods
in Soon et al. [351], Fu et al. [209], the discontinuous Petrov–Galerkin method in Carstensen and
Hellwig [110], and the hybrid high-order method in Di Pietro and Ern [166] that we briefly present
in §42.4.3.
42.4 Further topics

This section briefly reviews some other discretization techniques to approximate the model prob-
lem (42.10): Crouzeix–Raviart elements, mixed finite elements, and hybrid high-order (HHO)
methods.
42.4.1 Crouzeix–Raviart approximation

Let P1cr (Th ) := P1cr (Th ; R3 ) be the vector-valued Crouzeix–Raviart finite element space (see Chap-
ter 36 for the scalar-valued space P1cr (Th ) := P1cr (Th ; R)). Using P1cr (Th ) to approximate the
components of the displacement field leads to the desirable property on the approximation of
I cr 0 0 2
the divergence since ∇·(I K (u)) = ΠK (∇·u) for all K ∈ Th , where ΠK is the L -orthogonal
cr
I K (u)) is equal to the mean value of ∇·u in K (see
projection onto constants in K, i.e., ∇·(I
Exercise 36.1). Unfortunately, the Crouzeix–Raviart finite element fails to satisfy the broken ver-
sion of Korn’s inequality, since it is possible to find nonzero discrete fields vh ∈ P1cr (Th ) such
that locally in each mesh cell K ∈ Th , e(vh|K ) vanishes identically on K. This is a striking
difference with the scalar-valued case where a broken version of the Poincaré–Steklov inequality
holds true (see Lemma 36.6). For pure traction (Neumann) boundary conditions, the failure to
satisfy a discrete Korn inequalityRcan be shownPby theRfollowing dimension argument; see [199].
cr
Let P1,∗ (Th ) := {vh ∈ P1cr (Th ) | D vh dx = 0, K∈Th K ∇×vh dx = 0}, where the two integral
conditions (altogether, six scalar conditions) are meant to remove global rigid-body motions from
the space (see Remark 42.13). Let Nc , Nf , and Nf∂ denote the number of cells, faces, and bound-
ary faces in the mesh. Observe that 4Nc = 2Nf − Nf∂ (indeed, separating all the mesh cells, we
obtain 4Nc faces, and this number is equal to 2Nf − Nf∂ since there are two faces contributing to
each interface but one face contributing to each boundary face). Let eh : P1,∗ cr
(Th ) → L2 (D) be
s.t. eh (vh )|K = e(vh|K ). Since eh (vh ) is piecewise constant on Th and takes symmetric values in
R3×3 , we have dim(im(eh )) ≤ 6Nc . We infer that
dim(ker(eh )) = dim(P1,∗
cr
(Th )) − dim(im(eh ))
3 ∂
= 3Nf − 6 − dim(im(eh )) ≥ 3Nf − 6 − 6Nc = (N − 4),
2 f
which is positive as soon as the mesh is composed of more than one cell. The discrete Korn
inequality can also fail on some meshes when enforcing pure displacement (Dirichlet) boundary
conditions as shown in Figure 42.1. Here, D := (−1, 1)2 . Let z0 := 0, z5 = z1 := (1, 1),
z2 := (−1, 1), z3 := (−1, −1), z4 := (1, −1), S and let Ki be the triangle with vertices z0 , zi , zi+1
for all i ∈ {1:4}. Consider the mesh Th := i∈{1: 4} Ki . The vector field shown in Figure 42.1
is piecewise linear and defined by vh|K1 := −2(y, −x) + (0, −2), vh|K2 := 2(y, −x) + (−2, 0),
vh|K3 := −2(y, −x) + (0, R2), and vh|K4 := 2(y, −x) + (2, 0). One readily verifies that vh is in
cr
P1,∗ (Th ) and is such that F vh ds = 0 for all F ∈ Fh , but e(vh )|Ki = O for all i ∈ {1:4}.
Figure 42.1: Failure of the discrete Korn inequality using Crouzeix–Raviart displacements on a
mesh composed of four cells. The bullets symbolize zero displacement at the midpoint of the four
boundary edges, and the arrows show the displacement at the midpoint of the four internal edges.
42.4.2 Mixed finite elements

The idea in mixed finite element methods for linear elasticity is to approximate both the stress and
the displacement fields. Besides robustness w.r.t. volume locking, mixed finite element methods
ensure a direct approximation of the equilibrium condition (42.1), and the discrete strain can be
recovered locally from the discrete stress by inverting the constitutive relation (42.3). However, the
relation between displacement and stress, i.e., (42.2), is less direct (i.e., it is only obtained in a weak
form). In contrast, using the displacement-based formulation (42.17) ensures that (42.2) is satisfied
locally (i.e., one can define the discrete strain as e(uh )), but the equilibrium condition (42.1) and
the constitutive relation (42.3) are only satisfied in a weak sense. One difficulty with mixed finite
element methods for elasticity is the devising of discrete spaces with symmetric stresses. The idea
of relaxing this symmetry constraint by means of an auxiliary variable (that can be interpreted
as a rotation) was originally proposed in Fraejis de Veubeke [206] and was further developed and
analyzed in Amara and Thomas [8], Arnold et al. [19], Stenberg [353, 355], Morley [307]; see also
the more recent and comprehensive presentation in Arnold et al. [24], Boffi et al. [64]. Mixed finite
elements with symmetric stresses have been proposed in Arnold and Winther [17] in dimension two
and extended to dimension three in Arnold et al. [25], but the number of local degrees of freedom
is fairly substantial.
42.4.3 Hybrid high-order (HHO) approximation

We refer the reader to Chapter 39 for a detailed presentation and analysis of the HHO method
when approximating a scalar-valued elliptic PDE. For simplicity, we consider here homogeneous
Dirichlet conditions on the displacement, and we suppose that the Lamé coefficients take constant
values.
Let D be a polyhedron in Rd and (Th )h∈H be a shape-regular sequence of affine meshes so that
each mesh covers D exactly. To fix the ideas, we assume that d = 3. Let k ≥ 1 be the polynomial
degree. The local unknowns are Rd -valued polynomials of degree at most k on the mesh cells and
the mesh faces. For all K ∈ Th , we let V̂Kk := VKk × V∂K
k
with
Y
−1
VKk := P k,d ◦ TK , V∂Kk
:= P k,d−1 ◦ TF−1 , (42.22)
F ∈FK
where FK is the collection of the faces of K and TK : Sbd → K and TF : S d−1 → F are affine
geometric mappings defined on the reference simplices of Rd and Rd−1 , respectively. Pairs in V̂Kk
are denoted by v̂K := (vK , v∂K ).
There are three key ingredients to devise the HHO method for linear elasticity (the first and the
third ones are similar to those introduced in §39.1): (i) a displacement reconstruction operator,
(ii) a divergence reconstruction operator, and (iii) a stabilization operator. The displacement
−1
reconstruction operator R : V̂Kk → VKk+1 := P k+1,d ◦ TK is defined by solving the following local
Neumann problem: For all v̂K ∈ V̂K , the R -valued polynomial function d := R(v̂K ) ∈ VKk+1 is
k d
s.t.
(e(d), e(w))L2 (K) := −(vK , ∇·e(w))L2 (K) + (v∂K , e(w)nK )L2 (∂K) , (42.23)
for all w ∈ VKk+1 . To obtain a well-posed problem, we recall the space of rigid displacements R :=
c −1 c
N0,d and Lemma 42.7. Let RK := (ψK ) (R), where ψK is the covariant Piola transformation
(see (9.9b)), and observe that RRK = R sinceR the geometric
R mapping is affine.
R Then d ∈ VKk+1 is
uniquely definedRby prescribing RK d dx := K vK dx and K ∇×d dx := ∂K nK ×v∂K ds (indeed,
if r ∈ RK is s.t. K r dx = 0 and K ∇×r dx = 0, then r = 0; see Exercise 42.3). Furthermore, the
−1
divergence reconstruction operator D : V̂Kk → VKk := Pk,d ◦ TK is defined by solving the following
well-posed problem:
(D(v̂K ), q)L2 (K) := −(vK , ∇q)L2 (K) + (v∂K , qnK )L2 (K) , (42.24)
for all v̂K ∈ V̂Kk and all q ∈ VKk . Recalling the definition (39.3), we observe that this operator
satisfies the following important commuting property:
k
D(ÎK (v)) = ΠkK (∇·v), ∀v ∈ H 1 (K). (42.25)
This property is the key argument to ensure robustness w.r.t. volume locking. Finally, the stabiliza-
tion operator S : V̂Kk → V∂K
k
is defined as follows: For all v̂K ∈ V̂Kk , letting δ∂K := vK|∂K − v∂K ,

S(v̂K ) := Πk∂K vK|∂K − v∂K + ((I − ΠkK )R(v̂K ))|∂K

= Πk∂K δ∂K − ((I − ΠkK )R(0, δ∂K ))|∂K , (42.26)
where I is the identity, Πk∂K : L2 (∂K) → V∂K k

is the L2 -orthogonal projection onto V∂Kk
, and
k 2 k 2 k k 1 k
ΠK : L (K) → VK is the L -orthogonal projection onto VK . Let ÎK : H (K) → V̂K be the local
interpolation operator s.t. ÎK k
(v) := (ΠkK (v), Πk∂K (v|∂K )). Let E K : H 1 (K) → VKk+1 be the local
R
elliptic projection s.t. (e(E K (v)− v), e(w))L2 (K) = 0 for all w ∈ VKk+1 , and K (E K (v)− v) dx = 0,
R k
K
∇×(E K (v) − v) dx = 0. As in Lemma 39.1 and Lemma 39.3, we have R ◦ ÎK = E K and there
is c s.t.
−1 k
hK 2 kS(ÎK (v))kL2 (∂K) ≤ c |v − E K (v)|H 1 (K) , (42.27)
for all v ∈ H 1 (K), all K ∈ Th , and all h ∈ H.
For all K ∈ Th , we define the bilinear form on V̂Kk × V̂Kk such that
âK (v̂K , wK ) := 2µ(e(R(v̂K )), e(R(ŵK )))L2 (K)
+ λ(D(v̂K ), D(ŵK ))L2 (K) + 2µh−1
K (S(v̂K ), S(ŵK ))L2 (∂K) .
We introduce the global discrete spaces

VTkh := {vTh ∈ L2 (D) | vK := vTh |K ∈ VKk , ∀K ∈ Th },
k
VFkh ,0 := {vFh ∈ L2 (Fh ) | v∂K := vFh |∂K ∈ V∂K , ∀K ∈ Th ; vFh |Fh∂ = 0},
k :=
and the product space V̂h,0 VTkh × VFkh ,0 . For every pair v̂h := (vTh , vFh ) ∈ V̂h,0
k
, we denote by
v̂K := (vK , v∂K ) ∈ V̂Kk its local components in the mesh cell K ∈ Th . The discrete problem is as
follows: (
k
Find ûh ∈ V̂h,0 such that
k
(42.28)
âh (ûh , ŵh ) = ℓh (wTh ), ∀ŵh ∈ V̂h,0 ,
P
where the formsP âh and ℓh are assembled cellwise by setting âh (v̂h , ŵh ) := K∈Th âK (v̂K , ŵK )
and ℓh (wTh ) := K∈Th (f , wK )L2 (K) .
Remark 42.17 (Elimination of cell unknowns). As in §39.1, the cell unknowns can be elim-
inated locally in the discrete problem (42.28) by using a Schur complement technique, i.e., static
condensation. The global transmission problem coupling the face unknowns is of size 3 k+2 2 Nf
(for d = 3), where Nf is the number of mesh interfaces (that is, 9Nf in the lowest-order case k = 1).
Moreover, one can define as in §39.2.3 face-based tractions in each cell that are in equilibrium with
the applied load and the internal efforts and that comply at the interfaces with the law of action
and reaction.
Remark 42.18 (Polynomial degree). The minimal value k ≥ 1 is needed to control the rigid
displacements since P 0,d ( R ( P 1,d (recall that the minimal value of the polynomial degree for
scalar elliptic PDEs is k ≥ 0).
Remark 42.19 (Literature). HHO methods for linear elasticity were introduced in Di Pietro
and Ern [166]. Applications to nonlinear mechanics are developed in Botti et al. [75], Abbas et al.
[1, 2].
We equip the local HHO space V̂Kk with the strain-seminorm
|v̂K |e2 ,K := 2ke(vK )k2L2 (K) + h−1 2
K kvK − v∂K kL2 (∂K) , (42.29)
k
and the global HHO space V̂h,0 with the norm
X
kv̂h k2V̂ k := µ|v̂K |e2 ,K + λkD(v̂K )k2L2 (K) . (42.30)
h,0
K∈Th
k
This indeed defines a norm on V̂h,0 since kv̂h kV̂ k = 0 implies that for all K ∈ Th , vK is a rigid
h,0
displacement whose trace on ∂K is v∂K . Since two rigid displacements that coincide on a face are
identical, we infer that wTh is a global rigid displacement, and the Dirichlet condition enforced on
vFh ∈ VFkh ,0 at the boundary faces implies that vTh and vFh are zero.
Lemma 42.20 (Stability, well-posedness). (i) There are 0 < η ≤ ω s.t.
η kv̂K ke2 ,K ≤ 2ke(R(v̂K )k2L2 (K) + h−1 2 2

K kS(v̂K )kL2 (∂K) ≤ ω kv̂K ke,K ,
for all v̂K ∈ V̂Kk , all K ∈ Th , and all h ∈ H, and we have
âh (v̂h , v̂h ) ≥ min(1, η) kv̂h k2V̂ k , (42.31)

h,0
k
for all v̂h ∈ Vh,0 . (ii) The discrete problem (42.28) is well-posed.
k ′
To derive an error estimate, we introduce the consistency error δI (u) ∈ (V̂h,0 ) s.t.
hδI (u), ŵh i(V̂ k ′ ,V̂ k := ℓ̂h (ŵh ) − âh (Îhk (u), ŵh ), k
∀ŵh ∈ V̂h,0 ,
h,0 ) h,0
with Îhk : H01 (D) → V̂h,0k k

s.t. (Îhk (v))K := ÎK (v|K ) for all v ∈ H01 (D) and all K ∈ Th . Note that
Îhk (H01 (D)) ⊂ V̂h,0
k
since functions in H01 (D) have zero jumps across the mesh interfaces and zero
traces at the boundary faces.
Lemma 42.21 (Consistency). Assume that u ∈ H 1+r (D), r > 21 . There is c, uniform w.r.t. µ
and λ, such that for all h ∈ H,
X
2 k 2
kδI (u)k(Vhk )′ ≤c µku − E K (u)k♯,K + λk∇·u − ΠK (∇·u)kL2 (K) ,
K∈Th
where kvk♯,K := ke(v)kL2 (K) + hK ke(v)kL2 (∂K) for all v ∈ H 1+r (K).
1
2
Proof. See Exercise 42.5
k
Theorem 42.22 (Error estimate). Let u solve (42.10) and let ûh ∈ V̂h,0 solve (42.28). Assume
1
that u ∈ H 1+r (D), r > 21 . (i) Letting kφk†,K := kφkL2 (K) + hK
2
kφkL2 (∂K) , there is c s.t. for all
h ∈ H,
X X
µke(u − rK )k2L2 (K) ≤ c µku − E K (u)k2♯,K + λk∇·u − ΠkK (∇·u)k2†,K ,
K∈Th K∈Th
with rK := R(ûK ). (ii) If u ∈ H k+2 (D) and ∇·u ∈ H k+1 (D), then
X X
µke(u −
2(k+1)
rK )k2L2 (K) ≤c hK 2 2
µ|u|H k+2 (K) + λ|∇·u|H k+1 (K) .
K∈Th K∈Th
Proof. Use the approximation properties of the local elliptic projection E K and the L2 -orthogonal
projection ΠkK (see the proof of Theorem 39.17).
Exercises
Exercise 42.1 (Compliance). (i) Let s(e) be defined in (42.3) (i.e., s(e) := 2µe + λtr(e)Id ) and
let A be the fourth order tensor s.t. s(e) = Ae. Verify that A is symmetric positive definite. (Hint :
compute the quadratic form Ae:f .) Compute A 2 e. (Hint : find α, β ∈ R s.t. A 2 e = αe + β tr(e)I.)
1 1
(ii) Invert (42.3), i.e., express e as a function of s (the fourth-order tensor C s.t. e = Cs is
called compliance tensor ). (Hint : compute first tr(s).) Compute e:s in terms of s′ and tr(s)
where t′ := t − 13 tr(t)I is the deviatoric (i.e.,
R trace-free) part of the tensor t. (iii) Consider the
Hellinger–Reissner functional LHR (t, v) := D ( 4µ t :t + 18κ
1 ′ ′ 1
tr(t)2 + (∇·t)·v − f ·v) dx on H × V
where H := {t ∈ L2 (D) | t = tT , ∇·t ∈ L2 (D)} and V := L2 (D). Find the equations (in weak
form) satisfied by a critical point (s, u) of LHR . Verify that (s, u) satisfies (42.1) and (42.3) a.e.
in D. (Hint : use a density argument.)
Exercise 42.2 P (Second-order system). (i) Find matrices P Ajk ∈ Rd×d for all j, k ∈ {1:d}
s.t. ∇·s(u) = jk
j,k ∂j (A ∂k u). (Hint : verify that j,k ∂j (λ(ej ⊗ ek )∂k u) = ∇(λ∇·u) and
P
j,k ∂j (µ(e k ⊗ e j )∂ k u) = ∇·(µ∇u T
) where (e j )j∈{1: d} is the canonical basis of Rd .) (ii) Verify
R
that (Ajk )T = Akj . What is the consequence on the bilinear form a(v, w) := D ∂j wT Ajk ∂k v dx?
Exercise 42.3 (Pure traction). The pure traction problem is ∇·s(u)+f = 0 in D and R s(u)·n =
1
g
R on ∂D. (i) Write a weak formulation in H (D). (ii) Show that it is necessary that
R D f ·r dx +
R∂D g·r ds = 0 for a weak solution to exist. (iii) Assume 1that r R∈ R satisfies R D r dx = 0 and
D
∇×r dx = 0. Show that r = 0. (iv) Let V := {v ∈ H (D) | D v dx = 0, D ∇×v dx = 0}.
Show that the weak formulation is well-posed in V .
Exercise 42.4 (Timoshenko beam). Consider a horizontal beam D := (0, L) clamped at x = 0
and subjected to a (vertical) force distribution f and to a bending moment distribution m. A
(vertical) shear force F and a bending moment M are applied at x = L. The unknowns are the
γ
vertical displacement u and the rotation angle of the transverse section θ s.t. −(u′′ − θ′ ) = EI f
γ
and −γθ′′ − (u′ − θ) = EI m in D, where E is the Young modulus, I is the area moment of inertia,
and γ := 2(1+ν)I
Sκ (S is the cross section area and κ is an empirical correction factor usually set
γ
to 56 ). The boundary conditions are u(0) = 0, θ(0) = 0, (u′ − θ)(L) = EI F , and θ′ (L) = EI1
M.
2
(i) Assuming f, m ∈ L (D), write a weak formulation for the pair (u, θ) in Y := X × X with
1
X := {v R ∈ H′ (D) | v(0)2 = 0}. (ii) Prove the well-posedness of the weak formulation. (Hint : use
that 2 D θu dx ≤ µkθkL2 (D) + µ1 |u|2H 1 (D) with µ sufficiently close to 1 and the Poincaré–Steklov
inequality.) (iii) Write an H 1 -conforming finite element approximation and derive H 1 - and L2 -error
estimates for u and θ.
Exercise 42.5 (HHO). (i) Prove (42.25). (ii) Prove Lemma 42.20. (Hint : see Lemma 39.2 and
use the local Korn inequality kvkL2 (K) ≤ chK ke(v)kL2 (K) for all v ∈ H 1 (K) s.t. (v, r)L2 (K) = 0
for all r ∈ RK ; see Horgan [246], Kim [269].) (iii) Prove Lemma 42.21. (Hint : adapt the proof of
Lemma 39.16.)
Chapter 43
Maxwell’s equations:
H(curl)-approximation
The objective of this chapter is to introduce some model problems derived from Maxwell’s equations
that all fit the Lax-Milgram formalism in H(curl). The approximation is performed using H(curl)-
conforming edge (Nédélec) finite elements. The analysis relies on a coercivity argument in H(curl)
that exploits the presence of a uniformly positive zero-order term in the formulation. A more
robust technique controlling the divergence of the approximated field is presented in Chapter 44.
The space dimension is 3 in the entire chapter (d = 3), and D is a Lipschitz domain in R3 .
43.1 Maxwell’s equations

We start by recalling some basic facts about Maxwell’s equations. The reader is referred to Bossavit
[74, Chap. 1], Monk [303, Chap. 1], Assous et al. [27, Chap. 1] for a detailed discussion on this
model. Maxwell’s equations are partial differential equations providing a macroscopic description of
electromagnetic phenomena. These equations describe how the electric field E, the magnetic field
H, the electric displacement field D, and the magnetic induction B (sometimes called magnetic
flux density) interact through the action of currents j and charges ρ:
∂t D − ∇×H = −j (Ampère’s law), (43.1a)

∂t B + ∇×E = 0 (Faraday’s law of induction), (43.1b)
∇·D = ρ (Gauss’s law for electricity), (43.1c)
∇·B = 0 (Gauss’s law for magnetism). (43.1d)
Notice that if (∇·B)|t=0 = 0, taking the divergence of (43.1b) implies that (43.1d) is satisfied
at all times. Similarly, assuming (∇·D)|t=0 = ρ|t=0 and that the charge conservation equation
∂t ρ + ∇·j = 0 is satisfied at all times implies that (43.1c) is satisfied at all times. This shows that
if the data ρ, j, B|t=0 , and D|t=0 satisfy the proper constraints, Gauss’s laws are just consequences
of Ampère’s law and Faraday’s law.
The system (43.1) is closed by relating the fields through constitutive laws describing micro-
scopic mechanisms of polarization and magnetization:
D − ε0 E = P , B = µ0 (H + M ), (43.2)
226 Chapter 43. Maxwell’s equations: H(curl)-approximation
where ε0 and µ0 are the electric permittivity and the magnetic permeability of vacuum, and P
and M are the polarization and the magnetization fields, respectively. These quantities are the
average representatives at the macroscopic scale of complex microscopic interactions that must
be modeled. The models in question always involve parameters that need to be identified by
measurements or other techniques like homogenization or multiscale models. We have P := 0 and
M := 0 in vacuum, and it is common to use P := ε0 εr E and M := µr H to model isotropic
homogeneous dielectric and magnetic materials, where εr is the electric susceptibility and µr is the
magnetic susceptibility. In the rest of the chapter, we assume that
D := ǫE and B := µH, (43.3)
where ǫ and µ are given coefficients that may be space-dependent. The current j and charge
density ρ are a priori given, but it is also possible to make these quantities depend on the other
fields through phenomenological mechanisms. For instance, it is possible to further decompose the
current into one component that depends on the material and another one that is a source. The
simplest model doing that is Ohm’s law, j = js + σE, where σ is the electrical conductivity and
js an imposed current.
We now formulate Maxwell’s equations in two different regimes: the time-harmonic regime and
the eddy current limit.
43.1.1 The time-harmonic regime

We first consider Maxwell’s equations in the time-harmonic regime where the time-dependence is
assumed to be of the form eiωt with i2 = −1 and ω is a given angular frequency. The time-harmonic
version of (43.1a)-(43.1b) is
iωǫE + σE − ∇×H = −js , in D, (43.4a)
iωµH + ∇×E = 0, in D, (43.4b)
H|∂Dd ×n = ad , E|∂Dn ×n = an , on ∂D, (43.4c)
where {∂Dd , ∂Dn } forms a partition of the boundary ∂D of D. The dependent variables are the
electric field E and the magnetic field H. The data are the conductivity σ, the permittivity ǫ, the
permeability µ, the current js , and the boundary data ad and an . The material coefficients ǫ and
µ can be complex-valued. The system (43.4) models for instance a microwave oven; see e.g., [74,
Chap. 9]. The conditions H|∂Dd ×n = 0 and E|∂Dn ×n = 0 are usually called perfect magnetic
conductor and perfect electric conductor boundary conditions, respectively.
Let us assume that the modulus of the magnetic permeability µ is bounded away from zero
uniformly in D. It is then possible to eliminate H by using H = i(ωµ)−1 ∇×E. The system then
takes the following form:
(−ω 2 ǫ + iωσ)E + ∇×(µ−1 ∇×E) = −iωjs , in D, (43.5a)
(∇×E)|∂Dd ×n = −iωµad , E|∂Dn ×n = an , on ∂D. (43.5b)
Notice that Gauss’s law for electricity is contained in (43.5a) since taking the divergence of the
equation yields ∇·((−ω 2 ǫ+iωσ)E) = ∇·(−iωjs ), which is the time-harmonic counterpart of (43.1c)
combined with (43.1a). The system (43.5) is often used to model the propagation of electromagnetic
waves through various media.
43.1.2 The eddy current problem

When the time scale of interest, say τ , is such that the ratio ǫ/(τ σ) ≪ 1, it is legitimate to neglect
the displacement current in Ampère’s law (i.e., Maxwell’s correction ∂t D). This situation occurs in
particular in systems with moving parts (either solid or fluids) whose characteristic speed is much
slower than the speed of light. The resulting system, called eddy current problem, is as follows:
σE − ∇×H = −js , in D, (43.6a)
∂t (µH) + ∇×E = 0, in D, (43.6b)
H|∂Dd ×n = ad , E|∂Dn ×n = an , on ∂D, (43.6c)
where {∂Dd , ∂Dn } forms a partition of the boundary ∂D of D. The system (43.6) arises in
magneto-hydrodynamics (MHD). In this case, js is further decomposed into js = js′ + σu×B,
where u is the velocity of the fluid occupying the domain D, i.e., the actual current is decomposed
into j = js′ + σ(E + u×B).
Let us assume that σ is bounded from below away from zero uniformly in D. It is then possible
to eliminate the electric field from (43.6) by using E = σ −1 (∇×H − js ). The new system to be
solved is rewritten as follows:
∂t (µH) + ∇×(σ −1 ∇×H − u×(µH)) = ∇×(σ −1 js′ ), in D, (43.7a)
−1
H|∂Dd ×n = ad , (σ ∇×H − u×(µH))|∂Dn ×n = cn , on ∂D, (43.7b)
where cn := an + (σ −1 js′ )|∂Dn ×n. At this point, it is possible to further simplify the problem
by assuming that either the time evolution is harmonic, i.e., H(x, t) := Hsp (x)eiωt , or the time
derivative is approximated as ∂t H(x, t) ≈ τ −1 (H(x, t) − H(x, t − τ )), where τ is the time step
of the time discretization. After appropriately renaming the dependent variable and the data, say
e := iωµ and f := ∇×(σ −1 js′ ), or µ
either µ e := µτ −1 and f := ∇×(σ −1 js′ ) + µ
eH(x, t − τ ), the above
system reduces to solving the following problem:
eH + ∇×(σ −1 ∇×H − u×(µH)) = f ,
µ in D, (43.8a)
−1
H|∂Dd ×n = ad , (σ ∇×H − u×(µH))|∂Dn ×n = cn , on ∂D. (43.8b)
Notice that ∇·f = 0 in both cases. Hence, Gauss’s law for magnetism is contained in (43.8a) since
taking the divergence of the equation yields ∇·(µH) = 0 whether µ e := µτ −1 .
e := iωµ or µ
43.2 Weak formulation

The time-harmonic problem and the eddy current problem have a very similar structure. After
lifting the boundary condition (either on ∂Dn for the time-harmonic problem or on ∂Dd for the
eddy current problem) and making appropriate changes of notation, the above two problems (43.5)
and (43.8) can be reformulated as follows: Find A : D → C3 such that
νA + ∇×(κ∇×A) = f , A|∂Dd ×n = 0, (κ∇×A)|∂Dn ×n = 0, (43.9)
where ν, κ, and f are complex-valued. We have taken u := 0 in the MHD problem for simplicity.
We have also assumed that the Neumann data is zero to avoid unnecessary technicalities. We have
ν := −ω 2 ǫ + iωσ and κ := µ−1 for the time-harmonic problem, and ν := iωµ or ν := µτ −1 and
κ := σ −1 for the eddy current problem.
43.2.1 Functional setting

Let us assume that f ∈ L2 (D) := L2 (D; C3 ) and ν, κ ∈ L∞ (D; C). A weak formulation of (43.9)
is obtained by multiplying the PDE by the complex conjugate of a smooth test function b with
zero tangential component over ∂Dd and integrating by parts. Recalling (4.11), we obtain
Z Z
(νA·b + κ∇×A·∇×b) dx = f ·b dx.
D D
The integral on the left-hand side makes sense if A, b ∈ H(curl; D). To be dimensionally coherent,
1
we equip H(curl; D) with the norm kbkH(curl;D) := (kbk2L2 (D) + ℓ2D k∇×bk2L2 (D) ) 2 , where ℓD is
some characteristic length of D, e.g., ℓD := diam(D).
1
Let γ c : H(curl; D) → H − 2 (∂D) denote the tangential trace operator introduced in (4.11)
1 1
and let h·, ·i∂D denote the duality pairing between H − 2 (∂D) and H 2 (∂D). Since the Dirichlet
c
condition γ (A) = 0 is enforced on ∂Dd only, we must consider the restriction of the linear
1
forms in H − 2 (∂D) to functions that are only defined on ∂Dd . Let H f 12 (∂Dd ) be composed of
1
the functions θ defined on ∂Dd whose zero-extension to ∂D, say θ̃, is in H 2 (∂D). Then for all
b ∈ H(curl; D), the restriction γ c (b)|∂D d is defined in H f 12 (∂Dd )′ by using the duality product
f 2 (∂Dd ). A weak formulation of (43.9) is the
hγ c (b)|∂D d , θi∂Dd := hγ c (b), θ̃i∂D for all θ ∈ H
1
following:
Find A ∈ Vd := {b ∈ H(curl; D) | γ c (b)|∂D d = 0} such that
(43.10)
aν,κ (A, b) = ℓ(b), ∀b ∈ Vd ,
with the following sesquilinear and antilinear forms:
Z Z
aν,κ (a, b) := (νa·b + κ∇×a·∇×b) dx, ℓ(b) := f ·b dx. (43.11)
D D
We assume that there are real numbers θ, ν♭ > 0, and κ♭ > 0 s.t.

ess inf ℜ eiθ ν(x) ≥ ν♭ and ess inf ℜ eiθ κ(x) ≥ κ♭ . (43.12)
x∈D x∈D
Let us set ν♯ := kνkL∞ (D;C) and κ♯ := kκkL∞ (D;C) .

Theorem 43.1 (Coercivity, well-posedness). (i) Assume f ∈ L2 (D), ν, κ ∈ L∞ (D; C),
and (43.12). Then the sesquilinear form aν,κ is coercive and bounded:

ℜ eiθ aν,κ (b, b) ≥ min(ν♭ , ℓ−2 2
D κ♭ )kbkH(curl;D) , (43.13a)
|aν,κ (a, b)| ≤ max(ν♯ , ℓ−2
D κ♯ )kakH(curl;D) kbkH(curl;D) , (43.13b)
for all a, b ∈ H(curl; D). (ii) The problem (43.10) is well-posed.

Proof. Let us first verify that Vd is a closed subspace of H(curl; D). Let (bn )n∈N be a Cauchy
f 12 (∂Dd ), we have
sequence in Vd . Then bn → b in H(curl; D), and for all θ ∈ H
0 = hγ c (bn )|∂Dd , θi∂Dd := hγ c (bn ), θ̃i∂D → hγ c (b), θ̃i∂D =: hγ c (b), θi∂Dd ,

1
so that b ∈ Vd . (Recall that (4.11) implies that γ c : H(curl; D) → H − 2 (∂D) is continuous.)
Moreover, coercivity follows from (43.12) since we have
Z
ℜ(eiθ aν,κ (b, b)) = ℜ(eiθ ν)|b|2 + ℜ(eiθ κ)|∇×b|2 dx
ZD
≥ ν♭ |b|2 + κ♭ |∇×b|2 dx ≥ min(ν♭ , ℓ−2 2
D κ♭ )kbkH(curl;D) .
D
Similarly, the boundedness of aν,κ follows from ν, κ ∈ L∞ (D; C), and the boundedness of ℓ follows
from f ∈ L2 (D). Finally, well-posedness follows from the complex version of the Lax–Milgram
lemma.
Example 43.2 (Property (43.12)). Assume to fix the ideas that κ is real and uniformly positive.
If ν is also real and uniformly positive, (43.12) is satisfied with θ := 0, κ♭ := ess inf x∈D κ(x), and
ν♭ := ess inf x∈D ν(x). If instead ν is purely√imaginary with a uniformly positive
√
imaginary part,
π 2 2
(43.12) is satisfied with θ := − 4 , κ♭ := 2 ess inf x∈D κ(x), and ν♭ := 2 ess inf x∈D ℑ(ν(x)).
More generally, if ν := ρν eiθν with ess inf x∈D ρν (x) =: ρ♭ > 0 and θν (x) ∈ [θmin , θmax ] ⊂ (−π, π)
a.e. in D, then setting δ := θmax − θmin and assuming that δ < π, (43.12) is satisfied with
π
θ := − 21 (θmin +θmax ) 2π−δ , ν♭ := min(cos(θmin +θ), cos(θmax +θ))ρ♭ and κ♭ := cos(θ) ess inf x∈D κ(x)
(see Exercise 43.3). An important example where the condition (43.12) fails is when the two
complex numbers ν and κ are collinear and point in opposite directions. In this case, resonances
may occur and (43.10) has to be replaced by an eigenvalue problem.
43.2.3 Regularity
In the case of constant or smooth coefficients, a smoothness property on the solution to (43.10)
can be inferred from the following important result.
Lemma 43.3 (Regularity). Let D be a Lipschitz domain in R3 . (i) There is c > 0 s.t. the
following holds true:
c ℓsD |v|H s (D) ≤ kvkL2 (D) + ℓD k∇×vkL2 (D) + ℓD k∇·vkL2 (D) , (43.14)
with s := 21 , for all vector fields v ∈ H(curl; D) ∩ H(div; D) with either zero normal trace or
zero tangential trace over ∂D. (ii) The estimate remains valid with s ∈ ( 12 , 1] if D is a Lipschitz
polyhedron, and with s := 1 if D is convex.
Proof. (i) For the proof of (43.14), see Birman and Solomyak [57, Thm. 3.1] and Costabel [142,
Thm. 2]. (ii) See Amrouche et al. [10, Prop. 3.7] when D is a Lipschitz polyhedron and [10,
Thm. 2.17] when D is convex.
Let us consider the problem (43.10) and assume that f ∈ H(div; D) and ν is constant (or
smooth) over D. Then the unique solution A is such that ∇·A = ν −1 ∇·f ∈ L2 (D). Hence, A ∈
H(curl; D) ∩ H(div; D). Moreover, (43.9) implies that ∇×(κ∇×A) ∈ L2 (D) so that, assuming
that κ is constant (or smooth) over D, we infer that ∇×A ∈ H(curl; D) ∩ H(div; D). In addition
to the above assumptions on ν and κ, let us also assume that ∂Dn = ∅ (i.e., A has a zero tangential
trace, which implies that ∇×A has a zero normal trace a.e. on ∂D). Lemma 43.3 implies that
there exists r > 0 so that
A ∈ H r (D), ∇×A ∈ H r (D), (43.15)
with r := 21 in general, r ∈ ( 12 , 1] if D is a Lipschitz polyhedron, and r := 1 if D is convex. In the

more general case of heterogeneous coefficients, we will see in the next chapter (see Lemma 44.2)
that the smoothness assumption (43.15) is still valid with a smoothness index r > 0 under appro-
priate assumptions on ν. In the rest of this chapter, we are going to assume that (43.15) holds
true with r > 0.
43.3 Approximation using edge elements

We assume that the hypotheses of Theorem 43.1 are satisfied so that the boundary-value problem
(43.10) is well-posed.
43.3.1 Discrete setting

We consider a shape-regular sequence of affine meshes (Th )h∈H of D. We assume that D is a
Lipschitz polyhedron so that each mesh covers D exactly. We also assume that the meshes are
compatible with the partition of the boundary into {∂Dd , ∂Dn }. We consider the Nédélec (or edge)
finite elements of some order k ≥ 0 from Chapter 15 and the corresponding H(curl)-conforming
finite element space Pkc (Th ) built in Chapter 19. Let Vhd be the subspace of Pkc (Th ) defined by
Vhd := {bh ∈ Pkc (Th ) | bh|∂Dd ×n = 0}. (43.16)
Since the Dirichlet boundary condition is strongly enforced in Vhd , the approximation setting is
conforming, i.e., Vhd ⊂ Vd . The discrete formulation of (43.10) is

Find Ah ∈ Vhd such that
(43.17)
aν,κ (Ah , bh ) = ℓ(bh ), ∀bh ∈ Vhd .
The Lax–Milgram lemma together with the conformity of the approximation setting implies that
(43.17) has a unique solution.
43.3.2 H(curl)-error estimate

Theorem 43.4 (H(curl)-error estimate). (i) Under the assumptions of Theorem 43.1, there is
c s.t. for all h ∈ H,
kA − Ah kH(curl;D) ≤ c inf kA − bh kH(curl;D) . (43.18)

bh ∈Vhd
(ii) Assuming that either ∂Dd = ∂D or ∂Dn = ∂D and that there is r ∈ (0, k + 1] s.t. A ∈ H r (D)
and ∇×A ∈ H r (D), where k ≥ 0 is the degree of the finite element used to build Vhd , we have
kA − Ah kH(curl;D) ≤ c hr (|A|H r (D) + ℓD |∇×A|H r (D) ). (43.19)
Proof. (i) The estimate (43.18) is a direct consequence of Céa’s lemma.

c
(ii) We prove the estimate (43.19) when ∂Dd = ∂D, that is, when Vhd := Pk,0 (Th ) := {bh ∈
c c
Pk (Th ) | bh|∂D ×n = 0}. We estimate the infimum in (43.18) by taking bh := Jh0 (A), where
c 1 c
Jh0 : L (D) → Pk,0 (Th ) is the commuting quasi-interpolation operator with zero tangential trace
introduced in §23.3.3. Owing to the items (ii) and (iii) in Theorem 23.12, we infer that
c c c
kA − Jh0 (A)kH(curl;D) ≤ kA − Jh0 (A)kL2 (D) + ℓD k∇×(A − Jh0 (A))kL2 (D)
c d
= kA − Jh0 (A)kL2 (D) + ℓD k∇×A − Jh0 (∇×A)kL2 (D)
≤c inf
c (T )
kA − bh kL2 (D) + c′ ℓD inf k∇×A − dh kL2 (D)
bh ∈Pk,0 h
d (T )
dh ∈Pk,0 h
≤ c′′ hr (|A|H r (D) + ℓD |∇×A|H r (D) ),
where the last step follows from Corollary 22.16. The proof for ∂Dn = ∂D is similar if one uses
Jhc , Jhd instead of Jh0
c d
, Jh0 .
Remark 43.5 (ν♭ -dependency). The coercivity and boundedness properties in (43.13) show
max(ν♯ ,ℓ−2
D κ♯ )
that the constant in the error estimate (43.18) is c = min(ν♭ ,ℓ−2
, which becomes unbounded
D κ♭ )
when ν♭ is very small. This difficulty is addressed in Chapter 44.
Remark 43.6 (Variants). It is possible to localize (43.19) by using Theorem 22.14 instead
of Corollary 22.16 when ∂Dd = ∂D, and using Theorem 22.6 instead of Corollary 22.9 when
∂Dn = ∂D. Using that A ∈ H0 (curl; D), ∇×A ∈ H0 (div; D), and the regularity of the mesh
sequence, Theorem 22.14 and Theorem 22.6 imply that
X 12
kA − Ah kH(curl;D) ≤ c h2r
K (|A|H r (K) + ℓD |∇×A|H r (K) ) 2
,
K∈Th
when r > 12 . The seminorm |·|H r (K) has to be replaced by |·|H r (DK ) whenever r ≤ 21 , where
DK is the set of the points composing the mesh cells sharing a degree of freedom with K. One
can also extend the estimate (43.19) to the case of mixed boundary conditions by adapting the
construction of the quasi-interpolation operator and of the commuting projection from Chapters 22
and 23. Finally, we refer the reader to Ciarlet [121, Prop. 4] for an alternative proof of (43.19).
43.3.3 The duality argument

Recalling the material from §32.3, we would like to apply the Aubin–Nitsche duality argument
to deduce an improved error estimate on kA − Ah kL2 (D) . It is at this point that we realize that
the approach we have taken so far is too simplistic. To better understand the problem, let us
consider the case ∂Dd = ∂D. In this context, we have Vd := H0 (curl; D) and L := L2 (D), and
Theorem 32.8 tells us that the Aubin–Nitsche argument provides a better rate of convergence in the
L2 -norm if and only if the embedding H0 (curl; D) ֒→ L2 (D) is compact, which is not the case as
shown in Exercise 43.1. The conclusion of this argumentation is that the estimates we have derived
so far cannot yield an improved error estimate on kA − Ah kL2 (D) . A way around this obstacle
is to find a space smaller than H0 (curl; D), where the weak solution A lives and that embeds
compactly into L2 (D), and to show that Ah is a convergent nonconforming approximation of A in
that space. We are going to see in Chapter 44 that a good candidate is H0 (curl; D) ∩ H(div; D),
as pointed out in Weber [391, Thm. 2.1-2.3]. Recall that the unknown field A stands for E or H,
and that the Gauss laws (43.1c)-(43.1d) combined with (43.3) imply that ∇·(ǫE) = ∇·D = ρ and
that ∇·(µH) = ∇·B = 0. Thus, it is reasonable to expect some control on the divergence of A
and, therefore, to hope for an improved estimate on kA − Ah kL2 (D) provided ∇·Ah is controlled
in some sense. This question is addressed in Chapter 44.
Exercises
Exercise 43.1 (Compactness). Let D := (0, 1)3 be the unit cube in R3 . Show that the em-
bedding H0 (curl; D) ֒→ L2 (D) is not compact. (Hint : consider vn := ∇φn with φn (x1 , x2 , x3 ) :=
1
nπ sin(nπx1 ) sin(nπx2 ) sin(nπx3 ), n ≥ 1, and prove first that (vn )n≥1 weakly converges to zero in
L2 (D) (see Definition C.28), then compute kvn kL2 (D) and argue by contradiction.)
Exercise 43.2 (Curl). (i) Let v be a smooth field. Show that k∇×vk2ℓ2 ≤ 2∇v:∇v. (Hint :
relate ∇×v to the components of (∇v − ∇v T ).) (ii) Show that k∇×vkL2 (D) ≤ |v|H 1 (D) for all
v ∈ H01 (D). (Hint : use an integration by parts.)
Exercise 43.3 (Property (43.12)). Prove the claim in Example 43.2, i.e., for [θmin , θmax ] ⊂
(−π, π) with δ := θmax − θmin < π, letting θ := − 21 (θmin + θmax ) 2π−δ
π
, prove that θ ∈ (− π2 , π2 ) and
[θmin + θ, θmax + θ] ⊂ (− π2 , π2 ).
Exercise 43.4 (Dirichlet/Neumann). Let v be Ra smooth vector field in D such that v|∂Dd ×n =
0. Prove that (∇×v)|∂D d ·n = 0. (Hint : compute D (∇×v)·∇q dx with q well chosen.)
Chapter 44
Maxwell’s equations: control on

the divergence
The analysis of Chapter 43 requires a coercivity property in H(curl). There is, however, a loss of
coercivity when the lower bound on the model parameter ν becomes very small. This situation
occurs in the following two situations: (i) in the low frequency limit (ω → 0) when ν := iωµ as in
the eddy current problem; (ii) if κ ∈ R and σ ≪ ωǫ when ν := −ω 2 ǫ + iωσ as in the time-harmonic
problem. We have also seen in Chapter 43 that a compactness property needs to be established
to deduce an improved L2 -error estimate by the duality argument. We show in this chapter that
robust coercivity and compactness can be achieved by a weak control on the divergence of the
discrete solution. The material of this chapter is based on [188].
44.1 Functional setting

In this section, we present the assumptions on the model problem and introduce a functional
setting leading to a key smoothness result on the curl operator.
44.1.1 Model problem

We consider the model problem (43.9) on a Lipschitz domain D in R3 . For simplicity, we restrict
the scope to the homogeneous Dirichlet boundary condition A|∂D ×n = 0 (so that ∂Dd = ∂D).
The weak formulation is

Find A ∈ V0 := H0 (curl; D) such that
(44.1)
aν,κ (A, b) = ℓ(b), ∀b ∈ V0 ,
R R
with aν,κ (a, b) := D (νa·b + κ∇×a·∇×b) dx and ℓ(b) := D f ·b dx. We assume that f ∈ L2 (D)
and that ∇·f = 0. The divergence-free condition on f implies the following important property
on the solution A:
∇·(νA) = 0. (44.2)
Concerning the material properties ν and κ, we make the following assumptions: (i) Bounded-
ness: ν, κ ∈ L∞ (D; C) and we set ν♯ := kνkL∞ (D;C) and κ♯ := kκkL∞ (D;C) . (ii) Rotated positivity:
234 Chapter 44. Maxwell’s equations: control on the divergence
there are real numbers θ, ν♭ > 0, and κ♭ > 0 s.t. (43.12) is satisfied, i.e.,

ess inf ℜ eiθ ν(x) ≥ ν♭ , ess inf ℜ eiθ κ(x) ≥ κ♭ . (44.3)
x∈D x∈D
ν κ
We define the contrast factors ν♯/♭ := ν♯♭ and κ♯/♭ := κ♯♭ . We also define the magnetic Reynolds
number γν,κ := ν♯ ℓ2D κ♯−1 . Several magnetic Reynolds numbers can be defined if the material is
highly contrasted, but we will not explore this situation further. (iii) Piecewise smoothness: there
is a partition of D into M disjoint Lipschitz polyhedra {Dm }m∈{1: M} s.t. ν|Dm , κ|Dm ∈ W 1,∞ (Dm )
for all m ∈ {1:M }. The reader who is not comfortable with this assumption may think of ν, κ
being constant without missing anything essential in the analysis.
44.1.2 A key smoothness result on the curl operator

Let us define the (complex-valued) functional spaces
M0 := H01 (D), M∗ := {q ∈ H 1 (D) | (q, 1)L2 (D) = 0}, (44.4)
as well as the following subspaces of H(curl; D):
X0ν := {b ∈ H0 (curl; D) | (νb, ∇m)L2 (D) = 0, ∀m ∈ M0 }, (44.5a)

−1
X∗κ−1 := {b ∈ H(curl; D) | (κ b, ∇m)L2 (D) = 0, ∀m ∈ M∗ }, (44.5b)
where (·, ·)L2 (D) denotes the inner product in L2 (D). The main motivation for introducing the
above subspaces is that A ∈ X0ν owing to (44.2). Moreover, we will see below that κ∇×A ∈
X∗κ−1 . Taking m ∈ C0∞ (D) in (44.5a) shows that for all b ∈ X0ν , the field νb has a weak divergence
in L2 (D) and ∇·(νb) = 0. Similarly, the definition (44.5b) implies that for all b ∈ X∗κ−1 , the
field κ−1 b has a weak divergence in L2 (D) and ∇·(κ−1 b) = 0. Invoking the integration by parts
1
formula (4.12) and the surjectivity of the trace map γ g : H 1 (D) → H 2 (∂D) then shows that
γ d (κ−1 b) = 0 for all b ∈ X∗κ−1 , where γ d is the normal trace operator (recall that γ d (v) = v|∂D ·n
if the field v is smooth).
Let us first state a simple result related to the Helmholtz decomposition of vector fields in
V0 := H0 (curl; D) using the subspace X0ν (a similar result is available on H(curl; D) using the
subspace X∗κ−1 ).
Lemma 44.1 (Helmholtz decomposition). The following holds true:
V0 = X0ν ⊕ ∇M0 . (44.6)
Proof. Let b ∈ V0 and let p ∈ M0 solve (ν∇p, ∇q)L2 (D) = (νb, ∇q)L2 (D) for all q ∈ M0 . Our
R v := b−∇p and
assumptions on ν imply that there is a unique solution to this problem. Then we set
observe that v ∈ X0ν . The sum is direct because if 0 = v + ∇p, then the identity D ν∇p·v dx = 0,
which holds true for all p ∈ M0 and all v ∈ X0ν , implies that ∇p = 0 = v.
We can now state the main result of this section. This result extends Lemma 43.3 to heteroge-
1
neous domains. Given a smoothness index s > 0, we set kbkH s (D) := (kbk2L2 (D) + ℓ2s 2
D |b|H s (D) ) ,
2
where ℓD is some characteristic length of D, e.g., ℓD := diam(D).

Lemma 44.2 (Regularity pickup). Let D be a Lipschitz domain in R3 . (i) Assume that the
boundary ∂D is connected and that ν is piecewise smooth. There exist s > 0 and Č > 0 (depending
on D and the contrast factor ν♯/♭ but not on ν♭ alone) such that
Čℓ−1
D kbkH s (D) ≤ k∇×bkL2 (D) , ∀b ∈ X0ν . (44.7)
(ii) Assume that D is simply connected and that κ is piecewise smooth. There exist s′ > 0 and
Č ′ > 0 (depending on D and the contrast factor κ♯/♭ but not on κ♭ alone) such that
Č ′ ℓ−1
D kbkH s (D) ≤ k∇×bkL2 (D) ,
′ ∀b ∈ X∗κ−1 . (44.8)
Proof. See Jochmann [259], Bonito et al. [70].

Remark 44.3 (Smoothness index). There are some situations where the smoothness indices
s, s′ can be larger than 21 . One example is that of isolated inclusions in an otherwise homogeneous
material. We refer the reader to Ciarlet [121, §5.2] for further insight and examples.
Lemma 44.2 has two important consequences. First, by restricting the smoothness index s to
zero in (44.7), we obtain the following important stability result on the curl operator.
Lemma 44.4 (Poincaré–Steklov). Assume that the boundary ∂D is connected and that ν is
piecewise smooth. There is Ĉps > 0 (depending on D and the contrast factor ν♯/♭ ) such that the
following Poincaré–Steklov inequality holds true:
Ĉps ℓ−1
D kbkL2 (D) ≤ k∇×bkL2 (D) , ∀b ∈ X0ν . (44.9)
The bound (44.9) is what we need to establish a coercivity property on X0ν that is robust
w.r.t. ν♭ . Indeed, we have

ℜ eiθ aν,κ (b, b) ≥ ν♭ kbk2L2 (D) + κ♭ k∇×bk2L2 (D) ≥ κ♭ k∇×bk2L2 (D)
1 2 −2
≥ κ♭ (k∇×bk2L2 (D) + Ĉps ℓD kbk2L2 (D) )
2
1
≥ κ♭ ℓ−2 2
min(1, Ĉps )kbk2H(curl;D) , (44.10)
2 D
for all b ∈ X0ν , where we recall that H(curl; D) is equipped with the norm kbkH(curl;D) :=
1
(kbk2L2 (D) + ℓ2D k∇×bk2L2 (D) ) 2 . This shows that the sesquilinear form aν,κ is coercive on X0ν
with a coercivity constant depending on the contrast factor ν♯/♭ but not on ν♭ alone (whereas the
coercivity constant on the larger space V0 is min(ν♭ , ℓ−2
D κ♭ ) (see (43.13a))).
Let us now examine the consequences of Lemma 44.2 on the Sobolev smoothness index of A and
∇×A. Owing to (44.7), there is s > 0 s.t. A ∈ H s (D). We will see in §44.3 that the embedding
H s (D) ֒→ L2 (D) is the compactness property that we need to apply the duality argument and
derive an improved L2 -error estimate. Furthermore, the field R := κ∇×A is in X∗κ−1 (notice in
particular that ∇×R = f − νA ∈ L2 (D)), so that we deduce from (44.8) that there is s′ > 0
′
s.t. R ∈ H s (D). In addition, the material property κ being piecewise smooth, we infer that the
following multiplier property holds true (see [259, Lem. 2] and [70, Prop. 2.1]): There exists τ > 0
and Cκ−1 s.t.
|κ−1 ξ|H τ ′ (D) ≤ Cκ−1 |ξ|H τ ′ (D) , ∀ξ ∈ H τ (D), ∀τ ′ ∈ [0, τ ]. (44.11)

′′
Letting s′′ := min(s′ , τ ) > 0, we conclude that ∇×A ∈ H s (D).
44.2 Coercivity revisited for edge elements

In this section, we revisit the H(curl)-error analysis for the approximation of the weak prob-
lem (44.1) using Nédélec (or edge) elements (see Chapters 15 and 19). The key tool we are going
to use is a discrete counterpart of the Poincaré–Steklov inequality (44.9). We consider a shape-

regular sequence of affine meshes (Th )h∈H of D. We assume that D is a Lipschitz polyhedron and
that each mesh covers D exactly.
44.2.1 Discrete Poincaré–Steklov inequality

Let Vh0 be the H0 (curl)-conforming space using Nédélec elements of order k ≥ 0 defined by
c
Vh0 := Pk,0 (Th ) := {bh ∈ Pkc (Th ) | bh|∂D ×n = 0}. (44.12)
Observe that the Dirichlet condition is enforced strongly in Vh0 . The discrete problem is formulated
as follows:
Find Ah ∈ Vh0 such that
(44.13)
aν,κ (Ah , bh ) = ℓ(bh ), ∀bh ∈ Vh0 .
Since it is not reasonable to consider the space {bh ∈ Vh0 | ∇·(νbh ) = 0}, because the normal
component of νbh may jump across the mesh interfaces, we are going to consider instead the
subspace
Xh0ν := {bh ∈ Vh0 | (νbh , ∇mh )L2 (D) = 0, ∀mh ∈ Mh0 }, (44.14)
g
where Mh0 := Pk+1,0 (Th ; C) is conforming in H01 (D; C). Note that the polynomial degrees of the
finite element spaces Mh0 and Vh0 are compatible in the sense that ∇Mh0 ⊂ Vh0 . Using this
property and proceeding as in Lemma 44.1 proves the following discrete Helmholtz decomposition:
Vh0 = Xh0ν ⊕ ∇Mh0 . (44.15)
Lemma 44.5 (Discrete solution). Let Ah ∈ Vh0 be the unique solution to (44.13). Then
Ah ∈ Xh0ν .
Proof. We must show that (νAh , ∇mh )L2 (D) = 0 for all mh ∈ Mh0 . Since ∇mh ∈ ∇Mh0 ⊂ Vh0 ,
∇mh is an admissible test function in (44.13). Recalling that ∇·f = 0, we infer that
0 = ℓ(∇mh ) = aν,κ (Ah , ∇mh ) = (νAh , ∇mh )L2 (D) ,
since ∇×(∇mh ) = 0. This completes the proof.
We now establish a discrete counterpart to the Poincaré–Steklov inequality (44.9). This result
is not straightforward since Xh0ν is not a subspace of X0ν . The key tool that we are going to
invoke is the stable commuting quasi-interpolation projections from §23.3.3.
Theorem 44.6 (Discrete Poincaré–Steklov). Under the assumptions of Lemma 44.4, there is
′
a constant Ĉps > 0 (depending on Ĉps , the polynomial degree k, the regularity of the mesh sequence,
and the contrast factor ν♯/♭ , but not on ν♭ alone) s.t. for all xh ∈ Xh0ν and all h ∈ H,
′ −1
Ĉps ℓD kxh kL2 (D) ≤ k∇×xh kL2 (D) . (44.16)
Proof. Let xh ∈ Xh0ν be a nonzero discrete field. Let φ(xh ) ∈ M0 := H01 (D) be the solution to
the following well-posed Poisson problem:
(ν∇φ(xh ), ∇m)L2 (D) = (νxh , ∇m)L2 (D) , ∀m ∈ M0 .
Let us define the curl-preserving lifting of xh s.t. ξ(xh ) := xh − ∇φ(xh ), and let us notice that
c d
ξ(xh ) ∈ X0ν . Upon invoking the quasi-interpolation operators Jh0 and Jh0 introduced in §23.3.3,
we observe that
c c c g
xh − Jh0 (ξ(xh )) = Jh0 (xh − ξ(xh )) = Jh0 (∇(φ(xh ))) = ∇(Jh0 (φ(xh ))),
c g c
where we used that Jh0 (xh ) = xh and the commuting properties of Jh0 and Jh0 . Since xh ∈ Xh0ν ,
g
we infer that (νxh , ∇(Jh0 (φ(xh ))))L2 (D) = 0, so that
c c
(νxh , xh )L2 (D) = (νxh , xh − Jh0 (ξ(xh )))L2 (D) + (νxh , Jh0 (ξ(xh )))L2 (D)
c
= (νxh , Jh0 (ξ(xh )))L2 (D) .
Multiplying by eiθ , taking the real part, and using the Cauchy–Schwarz inequality, we infer that
c
ν♭ kxh k2L2 (D) ≤ ν♯ kxh kL2 (D) kJh0 (ξ(xh ))kL2 (D) .
c
The uniform boundedness of Jh0 on L2 (D), together with the Poincaré–Steklov inequality (44.9)
and the identity ∇×ξ(xh ) = ∇×xh , implies that
c c
kJh0 (ξ(xh ))kL2 (D) ≤ kJh0 kL(L2 ;L2 ) kξ(xh )kL2 (D)
c −1
≤ kJh0 kL(L2 ;L2 ) Ĉps ℓD k∇×xh kL2 (D) ,
′ := −1 c −1
so that (44.16) holds true with Ĉps ν♯/♭ kJh0 kL(L2 ;L2 ) Ĉps .
Remark 44.7 (Literature). There are many ways to prove the discrete Poincaré–Steklov in-
equality (44.16). One route described in Hiptmair [244, §4.2] consists of invoking subtle regularity
estimates from Amrouche et al. [10, Lem. 4.7]. Another one, which avoids invoking regularity
estimates, is based on an argument by Kikuchi [267] which is often called discrete compactness;
see also Monk and Demkowicz [304], Caorsi et al. [106]. The proof is not constructive and is based
on an argument by contradiction. The technique used in the proof of Theorem 44.6, inspired
from Arnold et al. [23, Thm. 5.11] and Arnold et al. [26, Thm. 3.6], is more recent, and uses the
stable commuting quasi-interpolation projections Jhc and Jh0 c
. It was already observed in Boffi
[61] that the existence of stable commuting quasi-interpolation operators would imply the discrete
compactness property.
44.2.2 H(curl)-error analysis

We are now in a position to revisit the error analysis of §43.3. Let us first show that Xh0ν has the
same approximation properties as Vh0 in X0ν .
Lemma 44.8 (Approximation in Xh0ν ). There is c, uniform w.r.t. the model parameters, s.t.
for all A ∈ X0ν and all h ∈ H,
inf kA − xh kH(curl;D) ≤ c ν♯/♭ inf kA − bh kH(curl;D) . (44.17)
xh ∈Xh0ν bh ∈Vh0
c
Proof. Let A ∈ X0ν . We start by computing the Helmholtz decomposition of Jh0 (A) in Vh0
as stated in (44.15). Let ph ∈ Mh0 be the unique solution to the discrete Poisson problem
c c
(ν∇ph , ∇qh )L2 (D) = (νJh0 (A), ∇qh )L2 (D) for all qh ∈ Mh0 . Let us define yh := Jh0 (A) − ∇ph .
c
By construction, yh ∈ Xh0ν and ∇×yh = ∇×Jh0 (A). Hence, k∇×(A − yh )kL2 (D) = k∇×(A −
c
Jh0 (A))kL2 (D) . Since ∇·(νA) = 0, we also infer that
c c
(ν∇ph , ∇ph )L2 (D) = (νJh0 (A), ∇ph )L2 (D) = (ν(Jh0 (A) − A), ∇ph )L2 (D) ,
c
which in turn implies that k∇ph kL2 (D) ≤ ν♯/♭ kJh0 (A) − AkL2 (D) . The above argument shows that
c c
kA − yh kL2 (D) ≤ kA − Jh0 (A)kL2 (D) + kJh0 (A) − yh kL2 (D)
c
≤ kA − Jh0 (A)kL2 (D) + k∇ph kL2 (D)
c
≤ (1 + ν♯/♭ )kA − Jh0 (A)kL2 (D) .
In conclusion, we have proved that

c
inf kA − xh kH(curl;D) ≤ kA − yh kH(curl;D) ≤ (1 + ν♯/♭ )kA − Jh0 (A)kH(curl;D) .
xh ∈Xh0ν
Invoking the commutation and approximation properties of the quasi-interpolation operators, we

infer that
c
kA − Jh0 (A)k2H(curl;D) = kA − Jh0
c
(A)k2L2 (D) + ℓ2D k∇×(A − Jh0
c
(A))k2L2 (D)
c
= kA − Jh0 (A)k2L2 (D) + ℓ2D k∇×A − Jh0
d
(∇×A)k2L2 (D)
≤c inf kA − bh k2L2 (D) + c′ ℓ2D inf k∇×A − dh k2L2 (D)
bh ∈P0c (Th ) dh ∈P0d (Th )
≤c inf kA − bh k2L2 (D) + c′ ℓ2D inf k∇×(A − bh )k2L2 (D) ,

bh ∈P0c (Th ) bh ∈P0c (Th )
where the last bound follows by restricting the minimization set to ∇×P0c (Th ) since ∇×P0c (Th ) ⊂
P0d (Th ). The conclusion follows readily.
Theorem 44.9 (H(curl)-error estimate). Let A solve (44.1) and let Ah solve (44.13). Assume
that ∂D is connected and that ν is piecewise smooth. There is c, which depends on the discrete
′
Poincaré–Steklov constant Ĉps and the contrast factors ν♯/♭ and κ♯/♭ , s.t. for all h ∈ H,
kA − Ah kH(curl;D) ≤ c γ̂ν,κ inf kA − bh kH(curl;D) , (44.18)
bh ∈Vh0
with γ̂ν,κ := max(1, γν,κ ) and the magnetic Reynolds number γν,κ := ν♯ ℓ2D κ−1
♯ .
Proof. Owing to Lemma 44.5, Ah also solves the following problem: Find Ah ∈ Xh0ν s.t.
aν,κ (Ah , xh ) = ℓ(xh ), ∀xh ∈ Xh0ν .
Using the discrete Poincaré–Steklov inequality (44.16) and proceeding as in (44.10), we infer that
1
ℜ eiθ aν,κ (xh , xh ) ≥ κ♭ ℓ−2 min(1, (Ĉps′ 2
) )kxh k2H(curl;D) ,
2 D
for all xh ∈ Xh0ν . Hence, the above problem is well-posed. Recalling the boundedness prop-
erty (43.13b) of the sesquilinear form aν,κ and invoking the abstract error estimate (26.18) leads
to
2 max(ν♯ , ℓ−2
D κ♯ )
kA − Ah kH(curl;D) ≤ −2
inf kA − xh kH(curl;D) .
κ♭ ℓD min(1, (Ĉps′ )2 ) xh ∈Xh0ν
We conclude the proof by invoking Lemma 44.8.

Remark 44.10 (Neumann boundary condition). The above analysis can be adapted to han-
dle the Neumann condition (κ∇×A)|∂D ×n = 0; see Exercise 44.3. This condition implies that
(∇×(κ∇×A))|∂D ·n = 0. Moreover, assuming f|∂D ·n = 0 and taking the normal component of
the equation νA + ∇×(κ∇×A) = f at the boundary gives A|∂D ·n = 0. Since ∇·f = 0, we also
have ∇·(νA) = 0. In other words, we have
A ∈ X∗ν := {b ∈ H(curl; D) | (νb, ∇m)L2 (D) = 0, ∀m ∈ M∗ }.
g
The discrete spaces are now Vh := Pkc (Th ) and Mh∗ := Pk+1 (Th ; C)∩M∗ . Using Vh for the discrete
trial and test spaces, we infer that
Ah ∈ Xh∗ν := {bh ∈ Vh | (νbh , ∇mh )L2 (D) = 0, ∀mh ∈ Mh∗ }.
The Poincaré–Steklov inequality (44.16) still holds true provided the assumption that ∂D is con-
nected is replaced by the assumption that D is simply connected. The error analysis from Theo-
rem 44.9 can be readily adapted.
44.3 The duality argument for edge elements

Our goal is to derive an improved error estimate in the L2 -norm using a duality argument that
invokes a weak control on the divergence. The subtlety is that, as already mentioned, the setting is
nonconforming since Xh0ν is not a subspace of X0ν . We assume in the section that the boundary
∂D is connected and that the domain D is simply connected. Recalling the smoothness indices
s, s′ > 0 from Lemma 44.2 together with the index τ > 0 from the multiplier property (44.11) and
′′
letting s′′ := min(s′ , τ ), we have A ∈ H s (D) and ∇×A ∈ H s (D) with s, s′′ > 0. In what follows,
we set
σ := min(s, s′′ ). (44.19)
Let us first start with an approximation result on the curl-preserving lifting operator ξ : Xh0ν →
X0ν defined in the proof of Theorem 44.6. Recall that for all xh ∈ Xh0ν , the field ξ(xh ) ∈ X0ν is
s.t. ξ(xh ) := xh − ∇φ(xh ) with φ(xh ) ∈ H01 (D), implying that ∇×ξ(xh ) = ∇×xh .
Lemma 44.11 (Curl-preserving lifting). Let s > 0 be the smoothness index introduced in
(44.7). There is c, depending on the constant ČD from (44.7) and the contrast factor ν♯/♭ , s.t. for
all xh ∈ Xh0ν and all h ∈ H,
kξ(xh ) − xh kL2 (D) ≤ c hs ℓ1−s

D k∇×xh kL2 (D) . (44.20)
c
Proof. Let us set eh := ξ(xh ) − xh . We have seen in the proof of Theorem 44.6 that Jh0 (ξ(xh )) −
c
xh ∈ ∇Mh0 , so that (νeh , Jh0 (ξ(xh )) − xh )L2 (D) = 0 since ξ(xh ) ∈ X0ν , Mh0 ⊂ M0 , and
c c
xh ∈ Xh0ν . Since eh = (I − Jh0 )(ξ(xh )) + (Jh0 (ξ(xh )) − xh ), we infer that
c
(νeh , eh )L2 (D) = (νeh , (I − Jh0 )(ξ(xh )))L2 (D) ,
c
thereby implying that keh kL2 (D) ≤ ν♯/♭ k(I − Jh0 )(ξ(xh ))kL2 (D) . Using the approximation prop-
c
erties of Jh0 yields
keh kL2 (D) ≤ c ν♯/♭ hs |ξ(xh )|H s (D) ,
and we conclude using the bound |ξ(xh )|H s (D) ≤ ČD ℓ1−s
D k∇×xh kL2 (D) which follows from (44.7)
since ξ(xh ) ∈ X0,ν and ∇×ξ(xh ) = ∇×xh .
Lemma 44.12 (Adjoint solution). Let y ∈ X0ν and let ζ ∈ X0ν solve the (adjoint) problem
νζ + ∇×(κ∇×ζ) := ν♭−1 νy. There is c, depending on the constants Ĉps from (44.9), Č, Č ′
from (44.7)-(44.8), and the contrast factors ν♯/♭ , κ♯/♭ , and κ♯ Cκ−1 , s.t. for all h ∈ H,
|ζ|H σ (D) ≤ c ν♯−1 γν,κ ℓD

−σ
kykL2 (D) , (44.21a)
|∇×ζ|H σ (D) ≤ c ν♯−1 γν,κ γ̂ν,κ ℓ−1−σ
D kykL2 (D) . (44.21b)
Proof. Proof of (44.21a). Testing the adjoint problem with e−iθ ζ leads to κ♭ k∇×ζk2L2 (D) ≤
ν♯/♭ kykL2 (D) kζkL2 (D) . Using the Poincaré–Steklov inequality (44.9), we can bound kζkL2 (D) by
k∇×ζkL2 (D) , and altogether this gives
k∇×ζkL2 (D) ≤ κ−1 −1

♭ ν♯/♭ Ĉps ℓD kykL (D) .
2 (44.22)
Invoking (44.7) with σ ≤ s yields

−1 1−σ
|ζ|H σ (D) ≤ ČD ℓD k∇×ζkL2 (D) ≤ κ−1 −1 −1 2−σ
♭ ν♯/♭ ČD Ĉps ℓD kykL (D) ,
2
which proves (44.21a) since κ♭−1 ℓ2D = κ♯/♭ ν♯−1 γν,κ .

Proof of (44.21b). Invoking (44.8) with σ ≤ s′ for b := κ∇×ζ, which is a member of X∗κ−1 , we
infer that
′ −1+σ
ČD ℓD |b|H σ (D) ≤ k∇×bkL2 (D) = k∇×(κ∇×ζ)kL2 (D)
≤ ν♯/♭ kykL2 (D) + ν♯ kζkL2 (D) ,
by definition of the adjoint solution ζ and the triangle inequality. Invoking again the Poincaré–
Steklov inequality (44.9) to bound kζkL2 (D) by k∇×ζkL2 (D) and using (44.22) yields kζkL2 (D) ≤
κ−1 −2 2
♭ ν♯/♭ Ĉps ℓD kykL (D) . As a result, we obtain
2
′ −1+σ
ČD ℓD |b|H σ (D) ≤ ν♯/♭ (1 + ν♯ κ−1 −2 2
♭ Ĉps ℓD )kykL (D) ,
2
and this concludes the proof of (44.21b) since |∇×ζ|H σ (D) ≤ Cκ−1 |b|H σ (D) owing to the multiplier
property (44.11) and σ ≤ τ .
We can now state the main result of this section.
Theorem 44.13 (Improved L2 -error estimate). Let A solve (44.1) and let Ah solve (44.13).
There is c, depending on the constants Ĉps from (44.9), Č, Č ′ from (44.7)-(44.8), and the contrast
factors ν♯/♭ , κ♯/♭ , and κ♯ Cκ−1 , s.t. for all h ∈ H,
kA − Ah kL2 (D) ≤ c 3
inf (kA − vh kL2 (D) + γ̂ν,κ hσ ℓ−σ
D kA − vh kH(curl) ).
vh ∈Vh0
Proof. In this proof, we use the symbol c to denote a generic positive constant that can have
the same parametric dependencies as in the above statement. Let vh ∈ Xh0ν and let us set
xh := Ah − vh . We observe that xh ∈ Xh0ν . Let ξ(xh ) be the image of xh by the curl-preserving
lifting operator and let ζ ∈ X0ν be the solution to the following adjoint problem:
νζ + ∇×(κ∇×ζ) := ν♭−1 νξ(xh ).
(1) Let us first bound kξ(xh )kL2 (D) from above. Recalling that ξ(xh ) − xh = −∇φ(xh ) and that
(νξ(xh ), ξ(xh ) − xh )L2 (D) = −(νξ(xh ), ∇φ(xh ))L2 (D) = 0, we infer that
(ξ(xh ), νξ(xh ))L2 (D) = (xh , νξ(xh ))L2 (D)

= (A − vh , νξ(xh ))L2 (D) + (Ah − A, νξ(xh ))L2 (D)
= (A − vh , νξ(xh ))L2 (D) + ν♭ aν,κ (Ah − A, ζ)
c
= (A − vh , νξ(xh ))L2 (D) + ν♭ aν,κ (Ah − A, ζ − Jh0 (ζ)),
where we used the Galerkin orthogonality property on the fourth line. Since we have |aν,κ (a, b)| ≤
κ♯ ℓ−2
D γ̂ν,κ kakH(curl;D) kbkH(curl;D) by (43.13b), we infer from the commutation and approximation
properties of the quasi-interpolation operators that
kξ(xh )k2L2 (D) ≤ ν♯/♭ kA − vh kL2 (D) kξ(xh )kL2 (D)

+ c κ♯ ℓ−2 σ
D γ̂ν,κ h kA − Ah kH(curl;D) (|ζ|H σ (D) + ℓD |∇×ζ|H σ (D) ).
Owing to the bounds from Lemma 44.12 on the adjoint solution with y := ξ(xh ), we conclude that
2
kξ(xh )kL2 (D) ≤ ν♯/♭ (kA − vh kL2 (D) + c γ̂ν,κ hσ ℓ−σ
D kA − Ah kH(curl;D) ).
(2) The triangle inequality and the identity A − Ah = A − vh − xh imply that
kA − Ah kL2 (D) ≤ kA − vh kL2 (D) + kξ(xh ) − xh kL2 (D) + kξ(xh )kL2 (D) .
We use Lemma 44.11 to bound the second term on the right-hand side as
kξ(xh ) − xh kL2 (D) ≤ c hσ ℓ1−σ

D k∇×xh kL2 (D)
≤ c hσ ℓ1−σ
D (k∇×(A − vh )kL2 (D) + k∇×(A − Ah )kL2 (D) ),
and we use (44.18) to infer that kA − Ah kH(curl;D) ≤ cγ̂ν,κ kA − vh kH(curl;D) . For the third term
on the right-hand side, we use the bound on kξ(xh )kL2 (D) from Step (1). We conclude by taking
the infimum over vh ∈ Xh0ν , and we use Lemma 44.8 to extend the infimum over Vh0 .
Remark 44.14 (Literature). The construction of the curl-preserving lifting operator invoked in
the proof of Theorem 44.6 and Theorem 44.13 is done in Monk [302, pp. 249-250]. The statement
in Lemma 44.11 is similar to that in Monk [303, Lem. 7.6], but the present proof is simplified by
the use of the commuting quasi-interpolation operators. The curl-preserving lifting of A − Ah
is invoked in Arnold et al. [23, Eq. (9.9)] and denoted therein by ψ. The estimate of kψkL2 (D)
given one line above [23, Eq. (9.11)] is similar to (44.3) and is obtained by invoking the commuting
quasi-interpolation operators constructed in [23, §5.4] for natural boundary conditions. Note that
contrary to the above reference, we invoke the curl-preserving lifting of Ah − vh instead of A − Ah
and make use of Lemma 44.11, which simplifies the argument. Furthermore, the statement of
Theorem 44.13 is similar to that of Zhong et al. [405, Thm. 4.1], but the present proof is simpler
and does not require the smoothness index σ to be larger than 12 .
Exercises
Exercise 44.1 (Gradient). Let φ ∈ H01 (D). Prove that ∇φ ∈ H0 (curl; D)
Exercise 44.2 (Vector potential). Let v ∈ L2 (D) with (νv, ∇mh )L2 (D) = 0 for all mh ∈ Mh0 .
Prove that (νv, wh )L2 (D) = (∇×zh , ∇×wh )L2 (D) for all wh ∈ Vh0 , where zh solves a curl-curl
problem on Xh0ν .
Exercise 44.3 (Neumann condition). Recall Remark 44.10. Assume that D is simply con-
nected so that there is Ĉps > 0 such that Ĉps ℓ−1D kbkL2 (D) ≤ k∇×bkL2 (D) for all b ∈ X∗ν . Prove
′ ′ −1
that there is Ĉps > 0 such that Ĉps ℓD |bh kL2 (D) ≤ k∇×bh kL2 (D) for all bh ∈ Xh∗ν . (Hint : adapt
the proof of Theorem 44.6 using Jhc .)
Exercise 44.4 (Discrete Poincaré–Steklov for ∇·). Let ν be as in §44.1.1. Let Y0ν := {v ∈
H0 (div; D) | (νv, ∇×φ)L2 (D) = 0, ∀φ ∈ H0 (curl; D)} and accept as a fact that there is Ĉps > 0
such that Ĉps ℓ−1
D kvkL2 (D) ≤ k∇·vkL2 (D) for all v ∈ Y0ν . Let k ≥ 0 and consider the discrete
d c
space Yh0ν := {vh ∈ Pk,0 (Th ) | (νvh , ∇×φh )L2 (D) = 0, ∀φh ∈ Pk,0 (Th ; C)}. Prove that there is
′ ′
Ĉps > 0 such that Ĉps kvh kL2 (D) ≤ ℓD k∇·vh kL2 (D) for all vh ∈ Yh0ν . (Hint : adapt the proof of
d
Theorem 44.6 using Jh0 .)
Chapter 45
Maxwell’s equations: further

topics
In this chapter, we investigate two additional topics on the approximation of Maxwell’s equa-
tions. First, we study the use of a boundary penalty method inspired by Nitsche’s method for
elliptic PDEs (see Chapter 37 and §41.3) to enforce the boundary condition on the tangential
component. We combine this method with H(curl)-conforming elements and with the all-purpose
H 1 -conforming elements. The use of a boundary penalty method is motivated for H 1 -conforming
elements whenever some faces of the domain D are not parallel to the canonical Cartesian planes
in R3 since in this case the boundary condition couples the Cartesian components of the discrete
solution. For simplicity, we study the boundary penalty method under the assumption that there
is a uniformly positive zero-order term in the model problem. The second topic we explore in this
chapter is the use of a least-squares penalty technique to control the divergence in the context
of H 1 -conforming elements. We will see that this technique works well for smooth solutions, but
there is an approximability obstruction for nonsmooth solutions.
45.1 Model problem

We consider the weak formulation (44.1) in a Lipschitz domain D in R3 :

Find A ∈ V0 := H0 (curl; D) such that
(45.1)
aν,κ (A, b) = ℓ(b), ∀b ∈ V0 ,
R R
with aν,κ (a, b) := D (νa·b + κ∇×a·∇×b) dx, ℓ(b) := D f ·b dx, and f ∈ L2 (D). As in §44.1.1,
we assume that (i) ν, κ ∈ L∞ (D; C) and we set ν♯ := kνkL∞ (D;C) , κ♯
:= kκkL∞ (D;C) ; (ii) There are

real numbers θ, ν♭ > 0, and κ♭ > 0 s.t. i.e., ess inf x∈D ℜ eiθ ν(x) ≥ ν♭ , ess inf x∈D ℜ eiθ κ(x) ≥ κ♭ ;
(iii) There is a partition of D into M disjoint Lipschitz polyhedra {Dm }m∈{1: M} s.t. ν|Dm , κ|Dm
are constant for all m ∈ {1:M }. Recall that ℓD := diam(D).
244 Chapter 45. Maxwell’s equations: further topics
45.2 Boundary penalty method in H(curl)

In this section, we apply Nitsche’s boundary penalty method (see Chapter 37 and §41.3) to the
approximation of the model problem (43.10) using edge (Nédélec) finite elements.
45.2.1 Discrete problem

Let (Th )h∈H be a shape-regular sequence of affine meshes so that each mesh covers D exactly.
We assume that each mesh is compatible with the partition {Dm }m∈{1: M} so that ν and κ are
piecewise constant on Th . We set κK := κ|K , νK := ν|K , κr,K := ℜ(eiθ κK ), and νr,K := ℜ(eiθ νK )
for all K ∈ Th (notice that κr,K ≥ κ♭ > 0 and νr,K ≥ ν♭ > 0). For every boundary face F ∈ Fh∂ ,
we denote by Kl the unique mesh cell having F as a face, i.e., F := ∂Kl ∩ ∂D. To simplify the
|κKl |2
notation, we set λF := κr,K .
l
In Nitsche’s boundary penalty method, the degrees of freedom associated with the tangential
component at the boundary of the trial functions and of the test functions are kept in the trial and
in the test spaces. Hence, we set Vh := Pkc (Th ) ⊂ H(curl; D), k ≥ 0. Since Vh is not a subspace
of H0 (curl; D), the approximation setting is nonconforming. The discrete formulation is

Find Ah ∈ Vh such that
(45.2)
aν,κ,h (Ah , bh ) = ℓ(bh ), ∀bh ∈ Vh .
The sesquilinear form aν,κ,h : Vh ×Vh → C is such that
X Z
λF
aν,κ,h (ah , bh ) := aν,κ (ah , bh ) − nh (ah , bh ) + η0 e−iθ (ah ×n)·(bh ×n) ds, (45.3)
hF F
F ∈Fh∂
where η0 is a user-dependent parameter to be chosen large enough (see Lemma 45.1), and using
the notation σ(a) := κ∇×a for all a ∈ H(curl; D),
Z
nh (ah , bh ) := (σ(ah )×n)·bh ds. (45.4)
∂D
45.2.2 Stability and well-posedness

We equip the space Vh with the following stability norm: For all bh ∈ Vh ,
X
kbh k2Vh := νr,K kbh k2L2 (K) + κr,K k∇×bh k2L2 (K) + |bh |2∂ , (45.5a)
K∈Th
X λF
|bh |2∂ := kbh ×nk2L2 (F ) . (45.5b)
∂
h F
F ∈Fh
Let Th∂D be the collection of all the mesh cells having a boundary face. Let n∂ := maxK∈Th∂D |FK ∩
Fh∂ | denote the maximum number of boundary faces that a mesh cell in Th∂D can have (n∂ ≤ d for
simplicial meshes). As in (37.6), let cdt be the smallest constant such that kn×(∇×vh )kL2 (F ) ≤
−1
cdt hF 2 k∇×vh kL2 (F ) for all F ∈ Fh∂ and all vh ∈ Vh .
Lemma 45.1 (Coercivity, well-posedness). Assume that η0 > 41 n∂ c2dt . (i) The following
coercivity property holds true:

ℜ eiθ aν,κ,h (bh , bh ) ≥ αkbh k2Vh , ∀bh ∈ Vh , (45.6)
η0 − 41 n∂ c2dt
with α := 1+η0 > 0. (ii) The discrete problem (45.2) is well-posed.
Proof. We only sketch the proof since it is similar to that of Lemma 37.3. For all bh ∈ Vh , we have

ℜ eiθ aν,κ,h (bh , bh ) ≥ kbh k2Vh − ℜ eiθ nh (bh , bh ) .
The last term on the right-hand side is bounded by proceeding as in the proof of Lemma 37.2.
Using that (σ(bh )×n)·bh = −σ(bh )·(bh ×n) and |κr,Kl | ≤ |κKl |, we infer that
X 21
1
|nh (bh , bh )| ≤ n∂2 cdt κr,Kl k∇×bh k2L2 (K) |bh |∂ .
K∈Th∂D
Then we use the same quadratic inequality as in the proof of Lemma 37.3 to conclude that (45.6)
holds true. Finally, the well-posedness of (45.2) follows from the Lax–Milgram lemma.
Remark 45.2 (Sesquilinear form aν,κ,h ). If in the penalty term in (45.3) one takes the κ-
|κ |
dependent factor equal to hKFl instead of hλFF , the minimal value for the parameter η0 in Lemma 45.1
|κKl |
depends on maxK∈Th κr,Kl , which is not convenient in general. Furthermore, one can add the term
−nh (bh , ah ) to the right-hand side of (45.3) to make aν,κ,h Hermitian. Then the coercivity property
η −n c2
(45.6) is valid if η0 > n∂ c2dt with α := 0 1+η∂0 dt > 0.

We perform the error analysis by making only a minimal regularity assumption on A, i.e., we
assume that (43.15) holds true for some r > 0. Our first step consists of extending the tangential
trace of σ(a) := κ∇×a. Just like in §40.3.1, we introduce two real numbers p, q such that
2d
2 < p, < q ≤ 2, (45.7)
2+d
and consider pe ∈ (2, p] such that q ≥ pepe+d
d xd
. Notice that pe always exists since x 7→ x+d is an
increasing function. Let K ∈ Th be a mesh cell with outward unit normal nK , and let F ∈ FK be
a face of K. We consider the functional space V c (K) := {v ∈ Lp (K) | ∇×v ∈ Lq (K)} equipped
with the norm
1+d( 1 − 1 )
kvkV c (K) := kvkLp (K) + hK p q k∇×vkLq (K) . (45.8)
Recalling that LK F is the face-to-cell lifting operator defined in Lemma 17.1, the tangential trace
on F of any field v ∈ V c (K) is denoted by (v×nK )|F , and is defined as the antilinear form in
1
p′ 1
p′
(W pe ,e (F ))′ s.t. for all φ ∈ W pe ,e (F ),
Z
h(v×nK )|F , φiF := v·∇×LK
F (φ) − (∇×v)·L K
F (φ) dx, (45.9)
K
1 ′ 1 ′
where h·, ·iF denotes the duality pairing between (W pe ,e
p
(F ))′ and W pe ,e
p
(F ). The right-hand
p′
side of (45.9) is well defined since v ∈ L (K), ∇×v ∈ L (K), and LK
p q
F (φ) ∈ W
1,e
(K) ֒→
′ ′
Lq (K) ∩ W 1,p (K).
We now properly extend the sesquilinear form nh defined in (45.4). For this purpose, we set
Vs := {a ∈ H0 (curl; D) | σ(a) ∈ Lp (D), ∇×σ(a) ∈ Lq (D)}, (45.10)
and observe that σ(a)|K ∈ V c (K) for all a ∈ Vs and all K ∈ Th . It turns out that the assumption
f ∈ L2 (D) implies that one can always take q := 2.
Lemma 45.3 (Smoothness). Let A solve (45.1). Assume that there is r > 0 s.t.
A ∈ H r (D), ∇×A ∈ H r (D). (45.11)

6 3
Then A ∈ Vs with q := 2 and with either any p ∈ (2, 3−2r ] if r < 2 or all p ∈ (2, ∞) otherwise.
Proof. The property σ(A) ∈ Lp (D) with p as in the assertion follows from ∇×A ∈ H r (D) ֒→
Lp (D) owing to the Sobolev embedding theorem and the assumption κ ∈ L∞ (D; C). Since
∇×σ(A) = f − νA, f ∈ L2 (D), and ν ∈ L∞ (D; C), we conclude that ∇×σ(A) ∈ L2 (D).
We define V♯ := Vs + Vh and the sesquilinear form on V♯ × Vh such that

X
n♯ (a, bh ) := h(σ(a)|Kl ×n)|F , ΠF (bh )iF , (45.12)
F ∈Fh∂
where ΠF is the ℓ2 -orthogonal projection onto the hyperplane tangent to F , i.e., ΠF (bh ) :=
n×(bh ×n) (note that n = nF for boundary faces). We observe that (45.12) is meaningful since
1
p′
ΠF (bh ) is in W pe ,e (F ).
Lemma 45.4 (Properties of n♯ ). The following holds true for all ah , bh ∈ Vh and all a ∈ Vs :
n♯ (ah , bh ) = nh (ah , bh ), (45.13a)

Z
n♯ (a, bh ) = σ(a)·∇×bh − (∇×σ(a))·bh dx. (45.13b)
D
Moreover, there is c, uniform w.r.t. κ, such that the following boundedness property holds true for
all a ∈ V♯ , all bh ∈ Vh , and all h ∈ H:
|n♯ (a, bh )| ≤ c |a|n♯ |bh |∂ , (45.14)
with |bh |∂ defined in (45.5b) and

X 2d( 1 − 1 ) 2d( d+2 − 1 )

−1
|a|2n♯ := λF hKl 2 p kσ(a)k2Lp (Kl ) + hKl 2d q k∇×σ(a)k2Lq (Kl ) .
F ∈Fh∂
Proof. (1) Proof of (45.13a). Let ah , bh ∈ Vh . We have for all F ∈ Fh∂ ,

Z
h(σ(ah )|Kl ×n)|F , ΠF (bh )iF := σ(ah )·∇×LK
F
l
(ΠF (b h )) − (∇×σ(a h ))·L Kl
F (ΠF (b h )) dx
Kl
Z
= (σ(ah )|Kl ×n)·LK
F (ΠF (bh )) ds
l
∂Kl
Z Z
= (σ(ah )|Kl ×n)·ΠF (bh ) ds = (σ(ah )|Kl ×n)·bh ds,
F F
since LK Kl Kl
F (ΠF (bh ))|∂Kl \F = 0, LF (ΠF (bh ))|F = ΠF (bh ) by definition of LF , and owing to the
l
identity (v×n)·ΠF (w) = (v×n)·w for all v, w ∈ R . Summing the above identity over F ∈ Fh∂
d
proves (45.13a).
(2) See Exercise 45.1 for the proof of (45.13b).
(3) We prove (45.14) by proceeding as in the proof of Lemma 40.4. We have

d( 1 − 1 ) −1
|h(σ(a)|Kl ×n)|F , ΠF (bh )iF | ≤ c hKl2 p kσ(a)kV c (Kl ) hF 2 kΠF (bh )kL2 (F )
1 1
− 1 d( − ) 1
−1
≤ c λF 2 hKl2 p kσ(a)kLp(Kl ) λF2 hF 2 kΠF (bh )kL2 (F )
d+2

− 1 d( −1) 1
−1
+ λF 2 hKl 2d q k∇×σ(a)kLq (Kl ) λF2 hF 2 kΠF (bh )kL2 (F ) ,
for all F ∈ Fh∂ , where we used that kΠF (bh )kℓ2 = kbh ×nkℓ2 and the definition (45.8) of the
norm kσ(a)kV c (Kl ) . The rest of the proof is identical to that of Lemma 40.7 by invoking the
Cauchy–Schwarz inequality.
Recalling that q := 2, we equip the space V♯ with the norm
X |νK |2 |κK |2

2 2
kbkV♯ := kbkL2 (K) + k∇×bkL2 (K) + |b|2∂
2
(45.15)
νr,K κr,K
K∈Th
X 2d( 1 − 1 )
+ κr,Kl hKl 2 p k∇×bk2Lp (Kl ) + h2Kl k∇×(∇×b)k2Lq (Kl ) ,
F ∈Fh∂
for all b ∈ V♯ , where |b|2∂ := |bh |2∂ with b := bs + bh , bs ∈ Vs , bh ∈ Vh (this definition is

meaningful since functions in Vs have a zero tangential trace at the boundary, so that as + ah =
bs + bh implies ah|∂D = bh|∂D ). Invoking inverse inequalities shows that there is c♯ s.t. kbh kV♯ ≤
c♯ kbh kVh for all bh ∈ Vh and all h ∈ H, i.e., (27.5) holds true. Note that c♯ depends on the factor
maxK∈Th ( κ|κr,K
K|
, ν|νr,K
K|
).
Lemma 45.5 (Consistency/boundedness). Let the consistency error be defined by
hδh (ah ), bh iVh′ ,Vh := ℓ(bh ) − aν,κ,h (ah , bh ), ∀ah , bh ∈ Vh .
There is ω♯ , uniform w.r.t. A and κ, s.t. kδh (ah )kVh′ ≤ ω♯ kA − ah kV♯ for all ah ∈ Vh and all
h ∈ H.
Proof. See Exercise 45.2
Theorem 45.6 (Error estimate). Let A solve (45.1). Let Ah solve (45.2) with the penalty
parameter η0 as in Lemma 45.1. Assume that the smoothness property (43.15) holds true with
r > 0. (i) There is c, which can depend on maxK∈Th ( κ|κr,K
K|
, ν|νr,K
K|
), such that for all h ∈ H,
kA − Ah kV♯ ≤ c inf kA − ah kV♯ . (45.16)

ah ∈Vh
(ii) Letting t = min(r, k + 1), χt := 1 if t ≤ 1 and χt := 0 if t > 1, we have

1 1
kA − Ah kV♯ ≤ c (h/ℓD )t (ν♯2 kAkH t (D) + κ♯2 k∇×AkH t (D)
1

+ χt κ♯2 hkκ−1 (f − νA)kL2 (D) . (45.17)
Proof. (i) Lemma 45.3 implies that A ∈ Vs with q := 2. Then the error estimate (45.16)
follows from Lemma 27.5 combined with stability (Lemma 45.1) and consistency/boundedness
(Lemma 45.5).
c c
(ii) To prove (45.17), we bound the infimum in (45.16) by taking ah := Jh0 (A), where Jh0 is
the commuting quasi-interpolation operator with boundary prescription from Chapter 23, which
we take of degree ℓ := ⌈t⌉ − 1. Note that contrary to the proof of Theorem 41.8, here the best-
approximation error is estimated by using an interpolant with boundary prescription so as to
facilitate the estimate on the boundary penalty seminorm. Since ℓ < t ≤ k + 1, we have ℓ ≤ k, i.e.,
c
ah ∈ Pℓ,0 (Th ) ⊂ Vh . Let us set η := A − ah , so that we need to bound the five terms composing
kηkV♯ (see (45.15)). For the first term, we have
X 21
|νK |2 1 1
2
kηkL2 (K) ≤ c ν♯2 kηkL2 (D) ≤ c′ ν♯2 inf kA − bh kL2 (D)
νr,K c (T )
bh ∈Pℓ,0 h
K∈Th
1
≤ c′′ ν♯2 (h/ℓD )t kAkH t (D) ,
where we used the approximation properties of Jhc from Theorem 23.12 in the first line and Corol-
lary 22.16 together with ℓ < t in the second line. Considering the second term and using the
c d
commuting property ∇×(Jh0 (ah )) = Jh0 (∇×ah ), we infer that
X 12
|κK |2 2
1
k∇×ηkL2 (K) ≤ c κ♯2 (h/ℓD )t k∇×AkH t (D) .
κr,K
K∈Th
The estimate on the third term is straightforward since η|∂D ×n = 0. For the fourth term, we
invoke the embedding inequality (41.16), and we infer that for all K ∈ Th ,
1
d( 1 − p )
hK 2 k∇×ηkLp (K) ≤ c (k∇×ηkL2 (K) + htK |∇×η|H t (K) )
= c (k∇×ηkL2 (K) + htK |∇×A|H t (K) ),
since ℓ < t implies that |∇×η|H t (K) = |∇×A|H t (K) . Using the above bound on k∇×ηkL2 (K) ,
together with κr,K ≤ |κK | ≤ κ♯ and |∇×A|H t (K) ≤ ℓ−t
D k∇×AkH t (K) for all K ∈ Th , we infer that
X 12
2d( 1 − 1 ) 1
κr,Kl hKl 2 p k∇×ηk2Lp (Kl ) ≤ c κ♯2 (h/ℓD )t k∇×AkH t (D) .
F ∈Fh∂
P 1
Consider the fifth term, ( F ∈F ∂ κr,Kl h2Kl k∇×(∇×η)k2L2 (Kl ) ) 2 . If t ≤ 1, we have ℓ = 0, so that
h
∇×(∇×η) = ∇×(∇×A) = κ−1 (f −νA) in each cell Kl since κ is piecewise constant. If t > 1, using
k∇×(∇×η)kL2 (Kl ) ≤ 2|∇×η|H 1 (Kl ) (see Exercise 43.2), and owing to the commuting property
c d d
∇×(Jh0 (ah )) = Jh0 (∇×ah ), we obtain k∇×(∇×η)kL2 (Kl ) ≤ 2|∇×A − Jh0 (∇×A)|H 1 (Kl ) . Since
t > 1, we infer that
X 12
1
κr,Kl h2Kl k∇×(∇×η)k2L2 (Kl ) ≤ cκ♯2 (h/ℓD )t k∇×AkH t (D) .
F ∈Fh∂
Collecting the above estimates leads to the assertion.
Remark 45.7 (Estimate (45.17)). If r > 12 , then t > 21 and the terms ℓ−t D k·kH t (D) can be
replaced by |·|H t (D) in (45.17). These terms can also be localized as a sum over the mesh cells.
45.3 Boundary penalty method in H 1

One can also combine Nitsche’s boundary penalty method with the use of H 1 -conforming finite
elements. The discrete problem is still (45.2), but the discrete trial and test space is now Vh :=
Pkg (Th ) := [Pkg (Th )]3 , k ≥ 1, where Pkg (Th ) is the scalar-valued H 1 -conforming finite element space
built in Chapter 19. Letting (e1 , e2 , e3 ) be the canonical basis of R3 and {ϕi }i∈Ah be the global
shape functions in Pkg (Th ), the basis that we use for Pkg (Th ) is {ϕi ek }i∈Ah ,k∈{1: 3} . Notice that
working with the H 1 -conforming space Vh leads to a collocalized scheme, i.e., the three components
of the discrete field Ah are associated with the same global shape function.
Invoking stability and consistency/boundedness arguments as in §45.2 leads to a quasi-optimal
error estimate that is identical to (45.16), except that the best-approximation error is measured
with respect to a smaller discrete space since H 1 -conformity is required. To bound this error,
we follow the arguments from Bonito and Guermond [69], Bonito et al. [71], where a mollification
operator is considered. For simplicity, we focus on the case with mild smoothness where r ∈ (0, 21 )
in (45.11), so that the optimal choice for the polynomial degree is k := 1.
Corollary 45.8 (Convergence). Let A solve (45.1) and assume that (45.11) holds true with
r ∈ (0, 21 ). Assume that the mesh sequence is quasi-uniform and that the polynomial degree is
k := 1. There are h0 > 0 and c s.t. for all h ∈ H ∩ (0, h0 ],

r 1 1 1
kA − Ah kV♯ ≤ c (h/ℓD ) 2 κ♯2 k∇×AkH r (D) + (κ♯2 ℓ−1
D + ν♯ )kAkH r (D)
2

1
+ κ♯2 hkκ−1 (f − νA)kL2 (D) . (45.18)
g c
Proof. To prove (45.18), we bound the infimum in (45.16) by taking ah := Ih,0 (Kδ,0 (A)), where
g g g c
Ih,0 is the canonical interpolation operator associated with P1,0 (Th ) ⊂ P1 (Th ) and Kδ,0 : L1 (D) →
∞
C0 (D) is the mollification operator from §23.4. To simplify the notation, we use here the parame-
ter δ as a length scale (that is, it corresponds to the length scale 2ζδ from Lemma 23.15). We take
1
c
δ := (ℓD h) 2 , and we assume that h is small enough so that Kδ,0 is well defined and Lemma 23.15
can be applied. Using the inverse inequality from the hint of Exercise 23.8 and the approximation
properties of Kδ,0c
from Corollary 23.5, we have δ 2 |Kδ,0c
(A)|H 2 (D) ≤ cδ r ℓ−r
D kAkH r (D) . Letting
θh := A − ah , we need to bound the five terms composing kθh kV♯ (see (45.15)). Considering the
second term, we have
c c
k∇×θh kL2 (D) ≤ k∇×(A − Kδ,0 (A))kL2 (D) + k∇×(Kδ,0 (A) − ah )kL2 (D)
d c g c
≤ k∇×A − Kδ,0 (∇×A)kL2 (D) + 2|Kδ,0 (A) − Ih,0 (Kδ,0 (A))|H 1 (D)
r −r c

≤ c δ ℓD k∇×AkH r (D) + h|Kδ,0 (A)|H 2 (D)

≤ c δ r ℓ−r
D k∇×AkH r (D) + hδ
r−2 −r
ℓD kAkH r (D)
r
= c (h/ℓD ) 2 k∇×AkH r (D) + ℓ−1
D kAkH r (D) ,
1
where we used Exercise 43.2 in the second line. The bound on the first term concerning ν♯2 kθh kL2 (D)
is similar. The estimate on the third term related to the boundary penalty term is zero. To bound
the fourth and fifth terms, we proceed as in the proof of Theorem 45.6 using k = 1.
Remark 45.9 (Estimate (45.18)). We refer the reader to [69, 71] for further developments. The
r
error estimate (45.18) is only of order h 2 . This is the price to pay to invoke only the smoothness
of ∇×A when using H 1 -conforming elements.
45.4 H 1 -approximation with divergence control

We now consider the model problem (45.1) when ν is significantly smaller than κℓ−2 D and ∇·f =
0. Recall that this situation was successfully treated in the previous chapter in the context of
H(curl)-conforming edge elements. In the present section, we consider instead an approximation
using H 1 -conforming elements. For simplicity, we assume that ν is constant, so that the solution
to (45.1) satisfies ∇·A = 0. Moreover, we are going to enforce the boundary condition strongly by
considering the discrete trial and test space Vh0 := Pkg (Th ) ∩ H0 (curl; D), k ≥ 1, with Pkg (Th ) :=
[Pkg (Th )]3 . Working with the discrete space Vh0 is viable provided the faces of D are parallel to the
canonical Cartesian planes in R3 , so that the boundary condition does not couple the Cartesian
components. Recall also that working with Pkg (Th ) leads to a collocalized scheme, i.e., the three
components of the discrete field Ah are associated with the same global shape function.
45.4.1 A least-squares technique

c
In the context of edge elements, a weak discrete control on the divergence of Ah ∈ Pk,0 (Th ) was
g 1
achieved by using that (νAh , ∇qh )L2 (D) = 0 for all qh ∈ Pk+1,0 (Th ). When using H -conforming
elements, it is no longer legitimate to use ∇qh as a test function. This difficulty can be circumvented
by employing a least-squares technique to control the divergence of Ah since ∇·Ah is an integrable
function when Ah is discretized with H 1 -conforming elements.
The functional setting we have in mind hinges on the space
Z := H(curl; D) ∩ H(div; D), (45.19)

1
with the norm kbkZ := (kbk2L2 (D) + ℓ2D k∇×bk2L2 (D) + ℓ2D k∇·bk2L2 (D) ) 2 . Consider the closed sub-
space Z0 := H0 (curl; D) ∩ H(div; D) and the following weak formulation:

Find A ∈ Z0 such that
(45.20)
aν,κ,η (A, b) = ℓ(b), ∀b ∈ Z0 ,
with aν,κ,η := aν,κ + aη and aη (a, b) := ηe−iθ κ♭ (∇·a, ∇·b)L2 (D) , where η > 0 is a user-defined
penalty parameter.
Proposition 45.10 (Equivalence). A solves (45.1) iff A solves (45.20).
Proof. See Exercise 45.3(i).
Owing to the Cauchy–Schwarz inequality, the sesquilinear form aν,κ,η satisfies the following
boundedness property:
|aν,κ,η (a, b)| ≤ max(ν♯ , κ♯ ℓ−2 −2

D , ηκ♭ ℓD )kakZ kbkZ , (45.21)
for all a, b ∈ Z. Moreover, owing to the following Poincaré–Steklov inequality (see Exercise 45.3(ii)),
we have 1
′′ −1
Ĉps ℓD kbkL2 (D) ≤ k∇×bk2L2 (D) + k∇·bk2L2 (D) 2 , ∀b ∈ Z0 , (45.22)
′′
where Ĉps > 0 only depends on D, we infer that
ℜ(eiθ aν,κ,η (b, b)) ≥ ℜ(eiθ aν,κ (b, b)) + ηκ♭ k∇·bk2L2 (D)
≥ min(1, η)κ♭ (k∇×bk2L2 (D) + k∇·bk2L2 (D) )
1
≥ min(1, η)κ♭ ℓ−2 ′′ 2 2
D min(1, (Ĉps ) )kbkZ , (45.23)
2
for all b ∈ Z0 , thus proving the ν-robust coercivity of aν,κ,η on Z0 .

A conforming approximation of the model problem (45.20) is realized using H 1 -conforming
elements as follows:

Find Ah ∈ Vh0 := Pkg (Th ) ∩ H0 (curl; D) such that
(45.24)
aν,κ,η (Ah , bh ) = ℓ(bh ), ∀bh ∈ Vh0 .
Since the approximation setting is conforming and the Galerkin orthogonality property holds true,
the basic error estimate (26.18) combined with the above boundedness and coercivity properties
of aν,κ,η yields
kA − Ah kZ ≤ c inf kA − bh kZ , (45.25)
bh ∈Vh0
where c is uniform w.r.t. ν♭−1 . Assuming that A ∈ H 1+r (D) ∩ H0 (curl; D), 0 ≤ r ≤ k, and using
the approximation results from §19.3 and §19.4, we infer the optimal error estimate kA − Ah kZ ≤
cℓD hr |A|H 1+r (D) .
45.4.2 The approximability obstruction

The H(curl)-conforming approximation method (based on edge elements) converges optimally
when A ∈ H r (D), ∇×A ∈ H r (D) and r ≥ 0. In contrast, the H 1 -conforming least-squares
technique converges optimally when A ∈ Z0 ∩ H r (D) with r ≥ 1, but it may fail to converge if A
is just in Z0 ∩ H r (D) with r < 1, which can be the case when D is a nonconvex polyhedron. The
bottleneck comes from the approximability property, i.e., inf bh ∈Vh0 kA − bh kZ may not go to zero
as h → 0.
Lemma 45.11 (Costabel). Let D be a nonconvex polyhedron. Then the space H0 (curl; D) ∩
H 1 (D) is a closed proper subspace of Z0 := H0 (curl; D) ∩ H(div; D), and the space H0 (div; D) ∩
H 1 (D) is a closed proper subspace of H0 (div; D) ∩ H(curl; D) (also equipped with the k·kZ -norm).
Proof. See [143, p. 541] and [144, Cor. 2.5].
Corollary 45.12 (Approximability obstruction). Let D be a nonconvex polyhedron. Then

(Vh0 )h∈H cannot have the approximability property in Z0 .
Proof. Since Vh0 ⊂ H0 (curl; D) ∩ H 1 (D) =: Z0,1 and Z0,1 is closed in Z0 , the limit of all the
Cauchy sequences in Vh0 are in Z0,1 . Moreover, Lemma 45.11 implies that there are functions of
Z0 that lie at a positive distance from Z0,1 . Thus, Cauchy sequences in Vh0 ⊂ Z0,1 cannot reach
these functions.
The striking consequence of the above arguments is that using H 1 -conforming finite elements
together with the formulation (45.24) produces a method that is convergent if the solution to (45.1)
is at least in H 1 (D), as it happens if D is a convex polyhedron (see Lemma 43.3(ii)), but the method
may fail to converge if it is used to approximate fields that are not in H 1 (D). This example shows
that there are situations where the approximability property should not be treated too lightly.
Remark 45.13 (Beyond the approximability obstruction). The source of the problem
is that the L2 -based least-squares penalty on ∇·Ah is too strong. Convergence can be recov-
R weakening this control. For instance, one can consider the sesquilinear form (a, b) 7→
ered by
ηe−iθ D dγ (x)κ♭ ∇·a∇·b dx, where d is the distance to the set of the reentrant edges of D (assumed
to be a three-dimensional polyhedron) and γ > 0 depends on the strength of the singularities in-
duced by the reentrant edges; see Costabel and Dauge [145], Buffa et al. [96]. An alternative
method developed in Bramble and Pasciak [79], Bramble et al. [82] involves a least-squares ap-
proximation of a discrete problem with different test and trial spaces. The numerical method
uses piecewise constant functions for the trial space and piecewise linear functions enriched with
face bubbles for the test space. Furthermore, a technique based on a local L2 -stabilization of the
divergence is introduced in Duan et al. [175, 176]. A method based on the stabilization of the
divergence in H −α (D) with α ∈ ( 12 , 1) has been proposed in Bonito and Guermond [69], Bonito
et al. [71]. All these methods have been proved to be quasi-optimal for solving the boundary value
problem (45.1) and for solving the corresponding eigenvalue problem (see Chapter 46). A similar
method, where the stabilization is done in H −1 (D), has been proposed in Badia and Codina [43].
However, as reported in [71, §6.4], it seems that controlling the divergence in H −1 (D) may not be
sufficient to guarantee that the spectrum of the Maxwell operator is well approximated.
Exercises
Exercise 45.1 (Identity for n♯ ). Prove (45.13b). (Hint : use the mollification operators Kδc :
L1 (D) → C ∞ (D) and Kδd : L1 (D) → C ∞ (D) from §23.1, and adapt the proof of Lemma 40.5.)
Exercise 45.2 (Consistency/boundedness). Prove Lemma 45.5. (Hint : adapt the proof of
Lemma 41.7 and use Lemma 45.4.)
Exercise 45.3 (Least-squares penalty on divergence). (i) Prove Proposition 45.10. (Hint :
use Lemma 44.1 to write A := A0 + ∇p, where A0 is divergence-free and p ∈ H01 (D), and prove
that p = 0.) (ii) Prove (45.22). (Hint : use Lemma 44.4 for A − ∇p.)
Chapter 46
Symmetric elliptic eigenvalue

problems
The three chapters composing Part X deal with the finite element approximation of the spectrum
of elliptic differential operators. Ellipticity is crucial here to provide a compactness property that
guarantees that the spectrum of the operators in question is well structured. We start by recalling
fundamental results on compact operators and symmetric operators in Hilbert spaces. Then, we
study the finite element approximation of the spectrum of compact operators. We first focus on the
H 1 -conforming approximation of symmetric operators, then we treat the (possibly nonconforming)
approximation of nonsymmetric operators.
The present chapter contains a brief introduction to the spectral theory of compact operators
together with illustrative examples. Eigenvalue problems occur when analyzing the response of
devices, buildings, or vehicles to vibrations, or when performing the linear stability analysis of
dynamical systems.
46.1 Spectral theory

We briefly recall in this section some essential facts regarding the spectral theory of linear operators.
Most of the proofs are omitted since the material is classical and can be found in Brezis [89,
Chap. 6], Chatelin [116, pp. 95-120], Dunford and Schwartz [179, Part I, pp. 577-580], Lax [278,
Chap. 21&32], Kreyszig [271, pp. 365-521]. In the entire section, L is a complex Banach space, we
use the shorthand notation L(L) := L(L; L), and IL denotes the identity operator in L.
46.1.1 Basic notions and examples

Definition 46.1 (Resolvent, spectrum, eigenvalues, eigenvectors). Let T ∈ L(L). The
resolvent set of T , ρ(T ), and the spectrum of T , σ(T ), are subsets of C defined as follows:
ρ(T ) := {µ ∈ C | µIL − T is bijective}, (46.1a)

σ(T ) := C\ρ(T ) = {µ ∈ C | µIL − T is not bijective}. (46.1b)
254 Chapter 46. Symmetric elliptic eigenvalue problems
(Since L is a Banach space, µ ∈ ρ(T ) iff (µIL − T )−1 ∈ L(L).) The spectrum of T is decomposed
into the following disjoint union:
σ(T ) = σp (T ) ∪ σc (T ) ∪ σr (T ), (46.2)
where the point spectrum, σp (T ), the continuous spectrum, σc (T ), and the residual spectrum,
σr (T ), are defined as follows:
σp (T ) := {µ ∈ C | µIL − T is not injective},
σc (T ) := {µ ∈ C | µIL − T is injective, not surjective, im(µIL − T ) = L},
σr (T ) := {µ ∈ C | µIL − T is injective, not surjective, im(µIL − T ) 6= L}.
Whenever σp (T ) is nonempty, members of σp (T ) are called eigenvalues, and the nonzero vectors in
ker(µIL − T ) are called eigenvectors associated with µ, i.e., 0 6= z ∈ L is an eigenvector associated
with µ iff T (z) = µz.
Example 46.2 (Finite dimension). If L is finite-dimensional, ker(µIL − T ) 6= {0} iff (µIL − T )
is not invertible. In this case, the spectrum of T only consists of eigenvalues, i.e., σ(T ) = σp (T )
and σc (T ) = σr (T ) = ∅.
Example 46.3 (Volterra operator). Let L := L2 ((0, 1); C) and let us identify L and L′ by
R1 Rx
setting hl′ , liL′ ,L := 0 l′ (x)l(x) dx. Let T : L → L be s.t. T (f )(x) := 0 f (t) dt for a.e. x ∈ (0, 1).
We have ρ(T ) = C\{0}, σp (T ) = ∅, σc (T ) = {0}, and σr (T ) = ∅; see Exercise 46.4.
Theorem 46.4 (Spectral radius). Let T ∈ L(L). (i) The subsets ρ(T ) and σ(T ) are both
nonempty. (ii) σ(T ) is a compact subset of C. (iii) Let
r(T ) := max |µ| (46.3)
µ∈σ(T )
be the spectral radius of T . Then

1
r(T ) = lim kT n kL(L)
n
≤ kT kL(L). (46.4)
n→∞
Proof. See Kreyszig [271], Thm. 7.5.4 for (i), Thm. 7.3.4 for (ii), and Thm. 7.5.5 for (iii).
1
Remark 46.5 ((46.4)). The identity r(T ) = limn→∞ kT n kL(L)
n
is often called Gelfand’s formula
1
(see [213, p. 11]). The inequality limn→∞ kT n kL(L)
n
≤ kT kL(L) may sometimes be strict. For
instance, r(T ) = 0 if σ(T ) = {0}, but it can happen in that case that kT kL(L) > 0. A simple
example is the operator T : R2 → R2 s.t. T (X) := AX with A := ( 00 10 ).
Let us consider more specifically the eigenvalues of T . Assume that σp (T ) 6= ∅ and let µ ∈
σp (T ). Let us set Ki := ker(µIL − T )i for all i ∈ N\ {0}. One readily verifies that the spaces Ki are
invariant under T . Moreover, K1 ⊂ K2 . . ., and if there is an integer j ≥ 1 such that Kj = Kj+1 ,
then Kj = Kj ′ for all j ′ > j.
Definition 46.6 (Ascent, algebraic and geometric multiplicity). Assume that σp (T ) 6= ∅
and let µ ∈ σp (T ). We say that µ has finite ascent if there is j ∈ N \ {0} such that Kj = Kj+1 ,
and the smallest integer satisfying this property is called ascent of µ and is denoted by α(µ) (or
simply α). Moreover, if Kα is finite-dimensional, then the algebraic multiplicity of µ, say m, and
the geometric multiplicity of µ, say g, are defined as follows:
m := dim(Kα ) ≥ dim(K1 ) =: g. (46.5)
Whenever α ≥ 2, nonzero vectors in Kα are called generalized eigenvectors.
Part X. Eigenvalue problems 255
If the eigenvalue µ has finite ascent α and if Kα is finite-dimensional, then elementary arguments
from linear algebra show that α + g − 1 ≤ m ≤ αg (note that α = 1 iff m = g). This inequalities
are proved by showing that g1 + i − 1 ≤ gi ≤ gi−1 + g1 for all i ∈ {1:α} with gi := dim(Ki );
see Exercise 46.2. All the eigenvalues have a finite ascent and a finite multiplicity if L is finite-
dimensional, or if the operator T is compact (see Theorem 46.14(iv)), but this may not be the case
in general.
Example 46.7 (Ascent, algebraic and geometric multiplicity). To illustrate Definition 46.6
in a finite-dimensional setting, we consider the operator T : R4 → R4 defined by T (X) := AX for
all X ∈ L := R4 , where  
1 1 0 0
0 1 1 0 
A := 0 0 1 0  .

0 0 0 1
Then µ = 1 is the only eigenvalue of T . Since
   
0 1 0 0 0 0 1 0
0 0 1 0  0
I4 − A =   2 0 0 0 , (I4 − A)3 = O4 ,
0 0 0 0 , (I4 − A) = 0 0 0 0
0 0 0 0 0 0 0 0
we have ker(IL − T ) = span{e1 , e4 }, ker(IL − T )2 = span{e1 , e2 , e4 }, and ker(IL − T )3 = ker(IL −

T )4 = span{e1 , e2 , e3 , e4 }, where {e1 , e2 , e3 , e4 } is the canonical Cartesian basis of R4 . Thus, the
ascent of µ = 1 is α = 3, its algebraic multiplicity is m = dim(ker(IL − T )3 ) = 4, and its geometric
multiplicity is g = dim(ker(IL − T )) = 2. Notice that α + g − 1 = 4 = m ≤ 6 = αg.
Let us finally explore the relation between the spectrum of T and that of its adjoint T ∗ : L′ → L′
s.t. hT ∗ (l′ ), liL′ ,L := hl′ , T (l)iL′ ,L for all l ∈ L and all l′ ∈ L′ (see Definition C.29). Recall that we
have adopted the convention that dual spaces are composed of antilinear forms (see Definition A.11
and §C.4), so that (λT )∗ = λT ∗ for all λ ∈ C. (The reader should be aware that a usual convention
in the mathematical physics literature is that dual spaces are composed of linear forms, in which
case (λT )∗ = λT ∗ .) Moreover, for any subset A ⊂ C, we denote conj(A) := {µ ∈ C | µ ∈ A}.
Lemma 46.8 (Spectrum of T ∗ ). Let T ∈ (L). The following holds true:
σ(T ∗ ) = conj(σ(T )), σr (T ) ⊂ conj(σp (T ∗ )) ⊂ σr (T ) ∪ σp (T ). (46.6)
Proof. Corollary C.52 implies that µIL −T is not bijective iff (µIL −T )∗ = µIL′ −T ∗ is not bijective.
This proves the first equality. See Exercise 46.1 for the proof of the other two inclusions.
Example 46.9 (Left and right shifts). Let p ∈ (1,P ∞) and let ℓp be the Banach space composed
of the complex-valued sequences x := (xn )n∈N s.t. n∈N |xn |p < ∞. We can identify the dual
′ P
space of ℓp with ℓp , where p1 + p1′ = 1, by setting hx, yiℓp′ ,ℓp := n∈N xn yn with x := (xn )n∈N
′ ′
and y := (yn )n∈N . Consider the left shift operator L : ℓp → ℓp defined by L(x) := (x1 , x2 , . . .)
p p
and
P the right shift Poperator R : ℓ → ℓ defined by R(x) := (0, x0 , x1 , . . .). ∗
Then hx, R(y)iℓp′ ,ℓp :=
∗
n≥1 xn y n−1 = n≥0 xn+1 y n = hL(x), yi p ′ p . This shows that L = R . Similarly, R = L
ℓ ,ℓ once
′
p p
the dual of ℓ is identified with ℓ . Observe that kRkL(ℓp;ℓp ) = kLkL(ℓp′ ;ℓp′ ) = 1, so that both σ(R)
and σ(L) are contained in the unit disk {λ ∈ C | |λ| ≤ 1} owing to Theorem 46.4(iii). Notice that
0 6∈ σp (R) since R is injective. Assume that there exists µ ∈ σp (R), i.e., there is a nonzero x ∈ ℓp s.t.
(µx0 , µx1 −x0 , µx2 −x1 , . . .) = 0. Then xn = 0 for all n ∈ N, i.e., x = 0, which is absurd (recall that
µ 6= 0). Hence, σp (R) = ∅. Lemma 46.8 in turn implies that σr (L) = ∅ because L∗ = R. Similarly,
Lemma 46.8 implies that σr (R) ⊂ conj(σp (L)) ⊂ σr (R), i.e., σr (R) = conj(σp (L)). Assuming that
′
µ ∈ σp (L), there is a nonzero vector x ∈ ℓp s.t. L(x) = µx, which means that x = x0 (1, µ, µ2 , . . .).
′
This vector is in ℓp iff |µ| < 1. Hence, σp (L) = {µ ∈ C | |µ| < 1}. Since σp (L) is invariant
under complex conjugation, we conclude that σr (R) = σp (L). Finally, since σ(L) is closed (see
Theorem 46.4(ii)) and kLkL(ℓp′ ;ℓp′ ) = 1, we have σ(L) ⊂ {µ ∈ C | |µ| ≤ 1}. But σ(L) must also
contain the closure in C of σp (L) = {µ ∈ C | |µ| < 1}. Hence, σ(L) = {µ ∈ C | |µ| ≤ 1}. This, in
turn, implies that σc (L) = {µ ∈ C | |µ| = 1}. In conclusion, we have established that
σp (L) = {µ ∈ C | |µ| < 1} = σr (R),

σc (L) = {µ ∈ C | |µ| = 1} = σc (R),
σr (L) = ∅ = σp (R).
46.1.2 Compact operators in Banach spaces

Since we are going to focus later our attention on the approximation of the eigenvalues and
eigenspaces of compact operators, we now recall important facts about such operators. Given
two Banach spaces V, W, we say that T ∈ L(V ; W ) is compact if T maps the unit ball of V into a
relatively compact set in W (see Definition A.17). Let us also recall (see Theorem A.21) that if there
exists a sequence (Tn )n∈N of operators in L(V ; W ) of finite rank s.t. limn→∞ kT − Tn kL(V ;W ) = 0,
then T is compact. Conversely, if W is a Hilbert space and T ∈ L(V ; W ) is a compact op-
erator, then there exists a sequence of operators in L(V ; W ) of finite rank, (Tn )n∈N , such that
limn→∞ kT − Tn kL(V ;W ) = 0.
Example 46.10 (Rellich–Kondrachov). For every bounded Lipschitz domain D, the Rellich–
pd
Kondrachov theorem states that the injection W s,p (D) ֒→ Lq (D) is compact for all q ∈ [1, d−sp )
if sp ≤ d (see Theorem 2.35).
Example 46.11 (Hilbert–Schmidt operators). Let K ∈ L2 (D×D; C), where D is a bounded
d 2 2
Rset in R . Then the Hilbert–Schmidt operator T : L (D; C) → L (D; C) defined s.t. T∗ (f )(x) :=
f (y)K(x, y) dy a.e. in D is compact (see Brezis [89, Thm. 6.12]). Note that T (f )(x) :=
RD
D (y)K(y, x) dy.
f
Example 46.12 (Identity). The identity Iℓp : ℓp → ℓp , p ∈ [1, ∞], is not compact. Indeed,
consider the sequence en := (δmn )m∈N . For all N ≥ 0 and n, m ≥ N, n 6= m, we have ken −em kℓp =
1
2 p for all p ∈ [1, ∞), and ken − em kℓ∞ = 1. Hence, one cannot extract any Cauchy subsequence
in ℓp from (en )n∈N .
Let us now state some important results on compact operators.
Theorem 46.13 (Fredholm alternative). Let T ∈ L(L) be a compact operator. The following
properties hold true for all µ ∈ C\{0}:
(i) µIL − T is injective iff µIL − T is surjective.
(ii) ker(µIL − T ) is finite-dimensional.
(iii) im(µIL − T ) is closed, i.e., im(µIL − T ) = ker(µIL′ − T ∗ )⊥ .
(iv) dim ker(µIL − T ) = dim ker(µIL′ − T ∗ ).
Proof. See Brezis [89, Thm. 6.6].
The Fredholm alternative usually refers to Item (i), which implies that every nonzero member
of the spectrum of T is an eigenvalue when T is compact. The key result for compact operators is
the following theorem.
Theorem 46.14 (Spectrum of compact operators). Let T ∈ L(L) be a compact operator with
dim(L) = ∞. The following holds true:
(i) 0 ∈ σ(T ).
(ii) σ(T )\{0} = σp (T )\{0}.
(iii) One of the following three cases holds: (1) σ(T ) = {0}; (2) σ(T )\{0} is a finite set; (3)
σ(T )\{0} is a sequence converging to 0.
(iv) Any µ ∈ σ(T )\{0} has a finite ascent α, and the space ker(µIL − T )α is finite-dimensional,
i.e., µ has finite algebraic and geometric multiplicity.
(v) µ ∈ σ(T ) iff µ ∈ σ(T ∗ ), i.e., σ(T ∗ ) = conj(σ(T )). The ascent, algebraic and geometric
multiplicities of µ ∈ σ(T )\{0} and of µ are equal.
Proof. See Brezis [89, Thm. 6.8], Lax [278, p. 238], or Kreyszig [271, Thm. 8.3.1 & 8.4.4] for (i)-(iii)
and [271, Thm. 8.4.3] for (iv)-(v).
The first two items in Theorem 46.14 imply that either T is not injective (i.e., 0 ∈ σp (T ))
and then σ(T ) = σp (T ) (and σc (T ) = σr (T ) = ∅), or T is injective (i.e., 0 6∈ σp (T )) and then
σ(T ) = σp (T ) ∪ {0} (and σc (T ) = {0}, σr (T ) = ∅ or σr (T ) = {0}, σc (T ) = ∅).
46.1.3 Symmetric operators in Hilbert spaces

In this section, L denotes a complex Hilbert space. The reader is invited to review §C.3 for basic
facts about Hilbert spaces. Let T ∈ L(L). The (Hermitian) transpose of T , say T H ∈ L(L), is
defined by setting
(T H (w), v)L := (w, T (v))L , ∀v, w ∈ L. (46.7)
Let (JLrf )−1 : L′ → L be the Riesz–Fréchet representation operator (see Theorem C.24), that is,
((JLrf )−1 (l′ ), l)L := hl′ , liL′ ,L for all l′ ∈ L′ and l ∈ L. We recall that JLrf and (JLrf )−1 are linear
operators because we have chosen dual spaces to be composed of antilinear forms (see Exercise 46.5
and Remark C.26).
Lemma 46.15 (Transpose and adjoint). Let T ∈ L(L). We have T H = (JLrf )−1 ◦ T ∗ ◦ JLrf , and
σp (T ∗ ) = σp (T H ), σc (T ∗ ) = σc (T H ), σr (T ∗ ) = σr (T H ). (46.8)
Finally, if the duality paring is identified with the inner product of L, i.e., if L and L′ are identified,
we have T H = T ∗ .
Proof. The identities ((JLrf )−1 T ∗ (l′ ), l)L = hT ∗ (l′ ), liL′ ,L = hl′ , T (l)iL′ ,L = ((JLrf )−1 (l′ ), T (l))L
show that T H = (JLrf )−1 ◦ T ∗ ◦ JLrf . This proves the first assertion. To prove (46.8), we observe
that for all µ ∈ C, we have µIL′ − T ∗ = µIL′ − JLrf ◦ T H ◦ (JLrf )−1 = JLrf ◦ (µIL − T H ) ◦ (JLrf )−1 . The
assertion (46.8) on the spectrum follows readily. Finally, if L and L′ are identified, JLrf becomes
the identity operator so that T H = T ∗ .
Definition 46.16 (Symmetric operator). Let T ∈ L(L). We say that T is (Hermitian) sym-
metric if T = T H .
Theorem 46.17 (Spectrum, spectral radius, ascent). Let T ∈ L(L) be a symmetric operator.
The following holds true: (i) σ(T ) ⊂ R, σr (T ) = ∅, and
{a, b} ⊂ σ(T ) ⊂ [a, b], (46.9)
with a := inf v∈L,kvkL =1 (T (v), v)L and b := supv∈L,kvkL =1 (T (v), v)L . (ii) kT kL(L) = r(T ) =
max(|a|, |b|). (iii) The ascent of any µ ∈ σp (T ) is equal to 1, i.e., every generalized eigenvec-
tor is an eigenvector, and if T is compact, the algebraic multiplicity and the geometric multiplicity
of µ are equal.
Proof. See Lax [278, p. 356], Kreyszig [271, §9.2], and Exercise 46.6 for a proof of (i). See Exer-
cise 46.6(iii) for a proof of (iii).
Corollary 46.18 (Characterization of σ(T )). Let T ∈ L(L) be a symmetric operator. Then µ ∈
σ(T ) iff there is a sequence (vn )n∈N in L such that kvn kL = 1 for all n ∈ N and kT (vn )−µvn kL → 0
as n → ∞.
Proof. Identifying L and L′ , we apply Corollary C.50 which says that (µIL − T ) is not bijective iff
1
there exists a sequence (vn )n∈N in L such that kvn kL = 1 and kµvn − T (vn )kL ≤ n+1 .
For the reader’s convenience, we now recall the notion of Hilbert basis in a separable Hilbert
space (separability is defined in Definition C.8).
Definition 46.19 (Hilbert basis). Let L be a separable Hilbert space. A sequence (en )n∈N in L
is said to be a Hilbert basis of L if it satisfies the following two properties:
(i) (em , en )L = δmn for all m, n ∈ N.
(ii) The linear space composed of all the finite linear combinations of the vectors in (en )n∈N is
dense in L.
Not every Hilbert space is separable, but all the Hilbert spaces encountered in this book are
separable (or by default are always assumed to be separable).
Lemma 46.20 (Pareseval). P Let L be a separable Hilbert space and let (en )n∈N be a Hilbert basis
of L. For all u ∈ L, set un := k∈{0: n} (u, ek )L ek . The following holds true:
X
lim ku − un kL = 0 and kuk2L = |(u, ek )L |2 . (46.10)
n→∞
k∈N
2
P
Conversely, let (αn )n∈N be a sequence in ℓ (C) and set uα,n := k∈{0: n} αk ek . Then the sequence
P to some uα in L such that (uα , en )V = αn for all n ∈ N, and we have
(uα,n )n∈N converges
kuα k2L = limn→∞ k∈{0: n} α2k .
Proof. See Brezis [89, Cor. 5.10].
Theorem 46.21 (Symmetric compact operator). Let L be a separable Hilbert space and let
T ∈ L(L) be a symmetric compact operator. Then there exists a Hilbert basis of L composed of
eigenvectors of T .
Proof. See [89, Thm. 6.11].
The above results mean that the eigenvectors of a symmetric compact operator T form a
P (vm , vn )L = δmn for all m, n ∈ N. Moreover, for all u ∈ L, letting αn2 :=
sequence (vn )n∈N s.t.
(u, v ) and u := k∈{0: n} αk vk , the sequence (un )n∈N converges to u in L and we have kukL =
P n L 2 n
k∈N |αk | .
46.2 Introductory examples

We review in this section some typical examples that give rise to an eigenvalue problem, and we
illustrate some of the concepts introduced in §46.1.
46.2.1 Example 1: Vibrating string

Consider a vibrating string of linear density ρ, length ℓ, attached at x = 0 and x = ℓ, and
maintained under tension with the force τ . Let us set D := (0, ℓ), J := (0, Tmax ), Tmax > 0, and
denote by u : D×J → R the displacement of the string in the direction orthogonal to the x-axis.
Denoting by u0 (x) and u1 (x) the initial position and the initial velocity (i.e., the time derivative
of the displacement), the displacement of the string can be modeled by the linear wave equation
∂tt u(x, t) − c2 ∂xx u(x, t) = 0 in D×J, (46.11a)
u(0, t) = 0, u(ℓ, t) = 0 in J, (46.11b)
u(x, 0) = u0 (x), ∂t u(x, 0) = u1 (x) in D, (46.11c)
1
where the wave speed is c := ( τρ ) 2 . The method of the separation of variables gives the following
representation of the solution:
X
u(x, t) = (αn cos(ωn t) + βn sin(ωn t)) ψn (x), (46.12)
n≥1
1
n2 π 2
with ωn := cλn , λn :=
2
ℓ2 , ψn (x) := sin(nπ xℓ ),
Z ℓ Z ℓ
2 2
αn := u0 (x)ψn (x) dx, , βn := u1 (x)ψn (x) dx.
ℓ 0 cnπ 0
A remarkable fact is that for all n ≥ 1, (λn , ψn ) is an eigenpair for the Laplace eigenvalue problem
−∂xx ψn (x) = λn ψn (x), ψn (0) = 0, ψn (ℓ) = 0. (46.13)
The eigenfunctions ψn are called normal modes. In musical language, they are called harmonics of
Rℓ Rℓ Rℓ Rℓ
the string. Note that αn = 0 u0 (x)ψn (x) dx/ 0 ψn2 (x) dx, ωn βn = 0 u1 (x)ψn (x) dx/ 0 ψn2 (x) dx.
We say that (46.13) is the spectral problem associated with the vibrating string. This problem
can be reformulated in the following weak form:

Find ψ ∈ H01 (D)\{0} and λ ∈ R such that
R R (46.14)
D ∂x ψ∂x w dx = λ D ψw dx, ∀w ∈ H01 (D).
2
R L := L (D) and let
Let R T : L → L be defined so that for all f ∈ L, T (f ) ∈ H01 (D) solves
1
:= D f w dx for all w ∈ H0 (D). The operator T is compact since the in-
D ∂x (T (f ))∂x w dx
jection H01 (D) ֒→ L2 (D) is compact owing to the Rellich–Kondrachov theorem. This R compact-
ness
R property will be important
R for approximation purposes. Upon observing that D f T (g) dx =
D ∂x (T (f ))∂ x (T (g)) dx = D T (f )g dx, we infer that T is symmetric according to Definition 46.16.
Owing to Theorem 46.17, all the eigenvalues of T are real and σr (T ) = 0. According to Theo-
rem 46.14, the eigenvalues of T are well separated and form a sequence that goes to 0. Note that T is
injective, that is, 0 is not an eigenvalue. According
R to Theorem
R 46.14, this means Rthat σc (T ) = {0}.
Let (µ, ψ) be an eigenpair of T . Then µ D ∂x ψ∂x w dx = D ∂x (T (ψ))∂x w dx = D ψw dx. Hence,
(µ−1 , ψ) solves (46.14). Conversely, one readily sees that if (λ, ψ) solves (46.14), then (λ−1 , ψ) is
an eigenpair of T . Thus, we have established that (λ, ψ) solves (46.14) iff (λ−1 , ψ) is an eigenpair
of T . Finally, Theorem 46.21 asserts that there exists a Hilbert basis of L consisting of eigenvectors
1
of T , and the basis in question is (( 2ℓ ) 2 ψn )n≥1 .
46.2.2 Example 2: Vibrating drum

Consider a two-dimensional elastic homogeneous membrane occupying at rest the domain D ⊂ R2
and attached to a rigid frame on ∂D, as shown in Figure 46.1. We assume that D is embedded
f D
1111111111111111111111111
0000000000000000000000000
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000∂D
0000000000000000000000000
1111111111111111111111111
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
0000000000000000000000000
1111111111111111111111111
Figure 46.1: Vibrating membrane attached to a rigid frame. Left: reference configuration D,
externally applied load f , and equilibrium displacement u. Right: one normal mode.
in R3 and denote by ez the third direction. Assume that the membrane is of uniform thickness,
has area density ρ, and that the tension tensor in the membrane, t, is uniform and isotropic, i.e.,
it is of the form t = τ I2 for some positive real number τ (force per unit surface). Consider a
time-dependent load f (x, t) := ρg(x) cos(ωt) with angular frequency ω for all (x, t) ∈ D×J with
J := (0, Tmax ), Tmax > 0. Under the small strain assumption, the time-dependent displacement of
the membrane in the ez direction, u : D×J → R, is modeled by the two-dimensional wave equation
∂tt u − c2 ∆u = g(x) cos(ωt) in D×J, (46.15a)

u(·, t)|∂D = 0 in J, (46.15b)
u(x, 0) = u0 (x), ∂t u(x, 0) = u1 (x) in D, (46.15c)
1
where the wave speed is c := ( τρ ) 2 . As in §46.2.1, the solution to this problem can be expressed in
terms of the normal modes (eigenmodes) of the membrane, (λn , ψn )n≥1 , which satisfy
−∆ψn = λn ψn in D, ψn|∂D = 0. (46.16)

1
Setting ωn := cλn2 , a straightforward calculation shows that if ω 6∈ {ωn }n≥1 ,
( )
X γn sin ω−ω ω+ωn
2 t sin 2 t
n
u(x, t) = αn cos(ωn t) + βn sin(ωn t) + ω−ωn ω+ωn ψn (x),

2 2 2
n≥1
where
(u0 , ψn )L2 (D) (u1 , ψn )L2 (D) (g, ψn )L2 (D)
αn := , ωn βn := , γn := .
kψn k2L2 (D) kψn k2L2 (D) kψn k2L2 (D)
As the forcing angular frequency ω gets close to one of the ωn ’s, a resonance phenomenon occurs.
When ω = ωn , |u(x, t)| grows linearly in time like t| sin(ωn t)|.
The spectral problem associated with the vibrating drum can be rewritten in weak form as
follows:
R R (46.17)
D
∇ψ·∇w dx = λ D ψw dx, ∀w ∈ H01 (D).
If the tension tensor t in the membrane is not uniform and/or not isotropic (think of a membrane
made of composite materials), and if the area density ρ is not uniform, the above spectral problem
takes the following form:

R R (46.18)
D
(t∇ψ)·∇w dx = λ D ρψw dx, ∀w ∈ H01 (D).
By proceeding as in §46.2.1 and under reasonable assumptions on t and ρ, one can show that the
solution operator associated with (46.18) is symmetric and compact from L2 (D) to L2 (D). Hence,
the eigenvalues associated with the eigenvalue problem (46.18) are countable, isolated, and grow
to infinity.
46.2.3 Example 3: Stability analysis of PDEs

It is common that one has to study the stability of physical systems modeled by PDEs. For instance,
the following nonlinear reaction-diffusion equation (sometimes referred to as the Kolmogorov–
Petrovsky–Piskounov equation):
∂t u − ∆g(u) − f (u) = 0 in D×J, (46.19)
models the spreading of biological populations when f (u) := u(1 − u), the Rayleigh–Benard con-
vection when f (u) := u(1 − u2 ), and combustion processes when f (u) := u(1 − u)(u − α) with
α ∈ (0, 1). We assume here that D := (0, 1)d , periodic boundary conditions are enforced, f and g
are smooth, and g ′ is bounded from below by some positive constant. Assuming that this problem
admits a particular time-independent solution (a standing wave), usw , the natural question that
follows is to determine whether this solution is stable under infinitesimal perturbations. Writing
u(x, t) := usw (x) + ψ(x)e−λt , λ ∈ C, where ψ is assumed to be small compared to usw , one obtains
the following linearized form of the PDE:
−λψ − ∆(g ′ (usw )ψ) − f ′ (usw )ψ = 0 in D×J. (46.20)
Since ∇(g ′ (usw )ψ) = g ′ (usw )∇ψ + ψg ′′ (usw )∇usw , this problem leads to the following eigenvalue
problem:
1 1
Find ψ ∈ Hper (D)\{0} and λ ∈ C such that ∀w ∈ Hper (D),
R ′ ′′ ′
R (46.21)
D
(g (u sw )∇ψ + ψ g (u sw )∇u sw )·∇w − f (u sw )ψw dx = λ D ψw dx,
1
where Hper (D) is composed of the functions in H 1 (D) that are periodic over D. The particular
solution usw is said to be linearly stable if all the eigenvalues have a positive real part. Here
again, it is the solution operatorR T : L2 (D) → L2 (D) that is of interest, where for all s ∈ L2 (D),
1 2
R (s) ∈ Hper (D) ⊂ L (D)
T solves D (g ′ (usw )∇T (s)+T (s) g ′′ (usw )∇usw )·∇w −T (s)f ′ (usw )w dx =
1
D sw dx for all w ∈ Hper (D). Under reasonable assumptions on f, g, usw , the operator T can be
shown to be compact.
46.2.4 Example 4: Schrödinger equation and hydrogen atom

The vibrating string and the drum are typical examples where compactness directly results from
the boundedness of the domain D. We now give an example where compactness results from an
additional potential in the PDE.
An important example of eigenvalue problem in physics is the Schrödinger equation. For

instance, the normalized Schrödinger equation takes the following form for the one-dimensional
quantum harmonic oscillator over R:
1 1
− ψ ′′ + x2 ψ = Eψ in R. (46.22)
2 2
The function ψ is the wave function of the oscillator, and the quantity ψ 2 is its probability distri-
bution function. The eigenvalue E is called energy. This problem has a countable (quantified) set
of eigenpairs
1 2
− x2 1
ψn (x) := 1 1 e Hn (x), En := n + , (46.23)
(2n n!) 2 π 4 2
2
dn −x2
where Hn (x) := (−1)n ex dxn e is the Hermite polynomial of order n. A natural functional
space for this problem is
B 1 (R) := {v ∈ H 1 (R) | xv ∈ L2 (R)}. (46.24)
1 1
R 2 2
In addition to being in H (R), functions in B (R) satisfy R x v (x) dx < ∞. It is shown in
Exercise 46.8 that the embedding B 1 (R) ֒→ L2 (R) is compact, whereas it is shown in Exercise 46.7
Rthat ′ the embedding H 1 (R) ֒→ L2 (R) is not compact. Hence, the sesquilinear form a(v, w) =
′
R
(v w + x2 vw) 1 1
R dx is bounded and 1coercive on B (R), and the operator T : B (R) → B (R) s.t.
1
a(T (u), w) = R uw dx for all w ∈ B (R), is symmetric and compact.

The hydrogen atom is a model for which the Schrödinger equation has the following simple
form:
~2 q2
− ∆ψ − ψ = Eψ in R3 . (46.25)
2me 4πǫ0 r
Here, ~ is the Planck constant, me the mass of the electron, ǫ0 the permittivity of free space, q the
electron charge, and r := kxkℓ2 the Euclidean distance of the electron to the nucleus. This problem
is far more difficult than the one-dimensional quantum harmonic oscillator because the Coulomb
q2
potential − 4πǫ 0r
is negative and vanishes at infinity. The sign problem can be handled as for the
Helmholtz problem (see Chapter 35) by invoking Gårding’s inequality after making use of Hardy’s
2 R
u2
inequality |u|2H 1 (Rd ) ≥ (d−2)
4 Rd r 2
dx for all u ∈ H 1 (Rd ). The spectrum of the solution operator
is composed of the point spectrum and the continuous spectrum. The residual spectrum is empty
because the solution operator is symmetric. There is a countable (quantified) set of eigenpairs.
Using spherical coordinates, they are given for all n ≥ 1 by
−3 ρ
ψn,l,m (r, θ, φ) := Cn,l a0 2 e− 2 ρl L2l+1 m
n−l−1 (ρ)Yl (θ, φ),
~2 1
En := − ,
2me a20 n2
3
2 (n−l−1)!
21
4πǫ0 ~2
where l ∈ {0:n−1}, m ∈ {−l:l}, Cn,l := n
2
2n((n+l)!)3 , a0 := me q 2 is the Bohr radius,
−γ r β
2r
ρ := na 0
, Lγβ (r) := r β!e dr
d
β (e
−r γ+β
r ) is the generalized Laguerre polynomial of degree β, and
m
Yl is the spherical harmonic function of degree l and order m.
Exercises
Exercise 46.1 (Spectrum). Let L be a complex Banach space. Let T ∈ L(L). (i) Show that
(λT )∗ = λT ∗ for all λ ∈ C. (ii) Show that σr (T ) ⊂ conj(σp (T ∗ )) ⊂ σr (T ) ∪ σp (T ). (Hint : use
1
Corollary C.15.) (iii) Show that the spectral radius of T verifies r(T ) ≤ lim supn→∞ kT nkL(L)
n
.
P −1 n
P
(Hint : consider n∈N (µ T ) and use the root test: the complex-valued series n∈N an converges
1
absolutely if lim supn→∞ |an | n < 1.)
Exercise 46.2 (Ascent, algebraic and geometric multiplicities). (i) Let T ∈ L(L). Let µ
be an eigenvalue of T and let Ki := ker(µIL − T )i for all i ∈ N\{0}. Prove that K1 ⊂ K2 . . ., and
assuming that there is j ≥ 1 s.t. Kj = Kj+1 , show that Kj = Kj ′ for all j ′ > j. (ii) Assume that
µ has a finite ascent α, and finite algebraic multiplicity m and geometric multiplicity g. Show that
α + g − 1 ≤ m ≤ αg. (Hint : letting gi := dim(Ki ) for all i ∈ {1:α}, prove that g1 + i − 1 ≤ gi and
gi ≤ gi−1 + g1 .) (iii) Compute the ascent, algebraic multiplicity, and geometric multiplicity of the
eigenvalues of following matrices and verify the two inequalities from Step (i):
     
1 1 0 0 1 1 0 0 1 1 0 0
0 1 2 0 0 1 0 0 0 1 0 0
 ,  ,  .
0 0 1 0 0 0 1 1 0 0 1 0
0 0 0 1 0 0 0 1 0 0 0 1
Exercise 46.3 (Eigenspaces). The following three questions are independent. (i) Suppose V =
V1 ⊕ V2 and consider T ∈ L(V ) defined by T (v1 + v2 ) := v1 for all v1 ∈ V1 and all v2 ∈ V2 . Find
all the eigenvalues and eigenspaces of T . (ii) Let T ∈ L(V ). Assume that S is invertible. Prove
that S −1 T S and T have the same eigenvalues. What is the relationship between the eigenvectors
of T and those of S −1 T S? (iii) Let V be a finite-dimensional vector space. Let {vn }n∈{1: m} ⊂ V,
m ≥ 1. Show that the vectors {vn }n∈{1: m} are linearly independent iff there exists T ∈ L(V ) such
that {vn }n∈{1: m} are eigenvectors of T corresponding to distinct eigenvalues.
RExercise 46.4 (Volterra operator). Let L := L2 ((0, 1); C) and let T : L → L be s.t. T (f )(x) :=
x
0 f (t) dt for a.e. x ∈ (0, 1). Notice that T is a Hilbert–Schmidt R1
operator, but this exercise is
meant to be done without using this fact. (i) Show that T H (g) = x g(t) dt for all g ∈ L2 ((0, 1); C).
(ii) Show that T is injective. (Hint : use Theorem 1.32.) (iii) Show that 0 ∈ σc (T ). (iv) Show that
σp (T ) = ∅. (v) Prove that µIL − T is bijective if µ 6= 0. (vi) Determine ρ(T ), σp (T ), σc (T ), σr (T ).
Do the same for T H .
Exercise 46.5 (Riesz–Fréchet). Let H be a finite-dimensional complex Hilbert space with

orthonormal basis {ei }i∈{1: n} and inner P product (·, ·)H . (i) Let g be an antilinear form on H,
i.e., g ∈ H ′ . Show that (JH rf −1
) (g) = i∈{1: n} g(ei )ei with g(ei ) := hg, ei iH ′ ,H , ∀i ∈ {1:n}. Is
rf −1
(JH ) : H ′ → H linear or antilinear? (ii) Let g be a linear form on H. Show that xg :=
P ′
i∈{1: n} g(ei )ei is s.t. hg, yiH ′ ,H = (xg , y)H . Is the map H ∋ g 7→ xg ∈ H linear or antilinear?
Exercise 46.6 (Symmetric operator). Let L be a complex Hilbert space and T ∈ L(L) be
a symmetric operator. (i) Show that σ(T ) ⊂ R. (Hint : compute ℑ((T (v) − µv, v)L and show
that |ℑ(µ)|kvk2L ≤ |(T (v) − µv, v)L | for all v ∈ L.) (ii) Prove that σr (T ) = ∅. (Hint : apply
Corollary C.15.) (iii) Show that the ascent of each µ ∈ σp (T ) is equal to 1. (Hint : compute
k(µIL − T )(x)k2L with x ∈ ker(µIL − T )2 .)
Exercise 46.7 (H 1 (R) ֒→ L2 (R) is not compact). (i) Let χ(x) := 1 + x if −1 ≤ x ≤ 0,

χ(x) := 1−x if 0 ≤ x ≤ 1 and χ(x) := 0 if |x| ≥ 1. Show that χ ∈ H 1 (R). (ii) Let vn (x) := χ(x−n)
for all n ∈ N. Show that (vn )n∈N converges weakly to 0 in L2 (R) (see Definition C.28). (iii)
Show that the embedding H 1 (R) ֒→ L2 (R) is not compact. (Hint : argue by contradiction using
Theorem C.23.)
Exercise 46.8 (B 1 (R) ֒→ L2 (R) is compact). (i) Show that the embedding B 1 (R) ֒→ L2 (R) is
compact, where B 1 (R) := {v ∈ H 1 (R) | xv ∈ L2 (R)}. (Hint : let (un )n∈N be a bounded sequence
in B 1 (R), build nested subsets Jk ⊂ N, ∀k ∈ N\{0}, s.t. the sequence (un|(−k,k) )n∈Jk converges in
L2 (−k, k).) (ii) Give a sufficient condition on α ∈ R so that Bα1 (R) ֒→ L2 (R) is compact, where
Bα1 (R) := {v ∈ H 1 (R) | |x|α v ∈ L2 (R)}.
Exercise 46.9 (Hausdorff–Toeplitz theorem). The goal of this exercise is to prove that the
numerical range of a bounded linear operator in a Hilbert space is convex; see also Gustafson [231].
Let L be a complex Hilbert space and let SL (1) := {x ∈ L | kxkL = 1} be the unit sphere in L.
Let T ∈ L(L) and let W (T ) := {α ∈ C | ∃x ∈ SL (1), α = (T (x), x)L } be the numerical range
of T . Let γ, µ ∈ W (T ), γ 6= µ, and x1 , x2 ∈ SL (1) be s.t. (T (x1 ), x1 )L = γ, (T (x2 ), x2 )L = µ.
1
Let T ′ := µ−γ (T − γIL ). (i) Compute (T ′ (x1 ), x1 )L and (T ′ (x2 ), x2 )L . (ii) Prove that there
exists θ ∈ [0, 2π) s.t. ℑ(eiθ (T ′ (x1 ), x2 )L + e−iθ (T ′ (x2 ), x1 )L ) = 0. (iii) Let x′1 := eiθ x1 . Compute
(T ′ (x′1 ), x′1 )L . (iv) Let λ ∈ [0, 1]. Show that the following problem has at least one solution: Find
α, β ∈ R s.t. kαx′1 + βx2 kL = 1 and (T ′ (αx′1 + βx2 ), αx′1 + βx2 )L = λ. (Hint : view the two
equations as those of an ellipse and an hyperbola, respectively, and determine how these curves
cross the axes.) (v) Prove that W (T ) is convex. (Hint : compute (T (αx′1 + βx2 ), αx′1 + βx2 )L .)
Chapter 47
Symmetric operators, conforming

approximation
The objective of this chapter is to study the approximation of eigenvalue problems associated
with symmetric coercive differential operators using H 1 -conforming finite elements. The goal is
to derive error estimates on the eigenvalues and the eigenfunctions. The analysis is adapted from
Raviart and Thomas [331] and uses relatively simple geometric arguments. The approximation of
nonsymmetric eigenvalue problems using nonconforming techniques is studied in Chapter 48 using
slightly more involved arguments.
47.1 Symmetric and coercive eigenvalue problems

In this section, we reformulate the eigenvalue problems introduced in §46.2 in a unified setting.
This abstract setting will be used in §47.2 to analyze the approximation of these problems using
H 1 -conforming finite elements. We restrict ourselves to the real-valued setting since we are going
to focus on symmetric operators.
47.1.1 Setting
Let D be a Lipschitz domain
R in Rd . Let L2 (D) be the real Hilbert space equipped with the inner
product (v, w)L2 (D) := D vw dx. Let V be a closed subspace of H 1 (D) which, depending on
the boundary conditions that are enforced, satisfies H01 (D) ⊆ V ⊆ H 1 (D). We assume that V
is equipped with a norm that is equivalent to that of H 1 (D). We also assume that the V -norm
is rescaled so that the operator norm of the embedding V ֒→ L2 (D) is at most one, e.g., one
−1
could set kvkV := Cps ℓD k∇vkL2 (D) if V := H01 (D), where Cps is the constant from the Poincaré–
Steklov inequality (31.12) in H01 (D) and ℓD is a characteristic length associated with D, e.g.,
ℓD := diam(D).
Let a : V ×V → R be a symmetric bilinear form, i.e., a(v, w) = a(w, v), satisfying the following
coercivity and boundedness properties:
α kvk2V ≤ a(v, v), |a(v, w)| ≤ kak kvkV kwkV , (47.1)

R
for all v, w ∈ V, with 0 < α ≤ kak < ∞. For instance, we have a(v, w) := D(t∇v)·∇w dx and
266 Chapter 47. Symmetric operators, conforming approximation
V := H01 (D) in (46.18), so that we can take α := τ♭ ℓ−2 −2

D and kak := τ♯ ℓD , where τ♭ and τ♯ are the
smallest and the largest eigenvalues of t in D.
Our goal is to investigate the H 1 -conforming approximation of the following eigenvalue problem:

Find ψ ∈ V \{0} and λ ∈ R such that
(47.2)
a(ψ, w) = λ(ψ, w)L2 (D) , ∀w ∈ V.
Let T : L2 (D) → L2 (D) be the solution operator such that for all u ∈ L2 (D),
a(T (u), w) := (u, w)L2 (D) , ∀w ∈ V. (47.3)
By proceeding as in §46.2.1, we conclude that T is symmetric and compact. We are then in the
setting of Theorem 46.14 and Theorem 46.21.
Theorem 47.1 (Hilbert basis). Under the above assumptions on the bilinear form a, the fol-
lowing properties hold true:
(i) (λ, ψ) ∈ (0, ∞)×V is an eigenpair for the eigenvalue problem (47.2) iff (λ−1 , ψ) ∈ (0, ∞)×V
is an eigenpair for T .
(ii) σp (T ) ⊂ (0, α1 ].
(iii) The eigenvalue problem (47.2) has a countable sequence of isolated real positive eigenvalues
that grows to infinity.
(iv) It is possible to construct a Hilbert basis (ψn )n≥1 of L2 (D), where (λn , ψn )n≥1 are the eigen-
pairs solving (47.2) (see Definition 46.19). (It is customary to enumerate the eigenpairs
starting with n ≥ 1.)
−1
(v) (λn 2 ψn )n≥1 is a Hilbert basis of V equipped with the inner product a(·, ·).
Proof. (i) Let (µ, ψ) be an eigenpair of T . Then kψk2L2 (D) = a(T (ψ), ψ) = µa(ψ, ψ), which implies
that µ > 0. This proves that σp (T ) = σ(T )\{0} and σp (T ) ⊂ (0, ∞) (see Theorem 46.14(ii) and
recall that dim(L2 (D)) = ∞). Let (µ, ψ) be an eigenpair for T . Then a(T (ψ), w) = µa(ψ, w) =
(ψ, w)L2 (D) for all w ∈ V. Since µ 6= 0, we conclude that a(ψ, w) = µ−1 (ψ, w)L2 for all w ∈ V, that
is, (µ−1 , ψ) solves (47.2). The converse is also true: if (λ, ψ) is an eigenpair for (47.2), then the
coercivity of a implies that λ 6= 0, and reasoning as above shows that (λ−1 , ψ) is an eigenpair of
T.
(ii) Let (µ, ψ) be an eigenpair of T . The coercivity of a implies that kψk2L2 (D) = a(T (ψ), ψ) =
µa(ψ, ψ) ≥ µαkψk2V ≥ µαkψk2L2 (D) , where the last bound follows from our assuming that the norm
of the embedding V ֒→ L2 (D) is at most one. Hence, µ ∈ (0, α1 ].
(iii) The number of eigenvalues of T cannot be finite since the eigenspaces are finite-dimensional
(see Theorem 46.13(ii)) and there exists a Hilbert basis of L2 (D) composed of eigenvectors of T
(see Theorem 46.21). We are then in the third case described in Theorem 46.14(iii): the eigenvalues
of T form a (countable) sequence that converges to zero. Hence, the eigenvalues of (47.2) grow to
infinity.
(iv) This is a consequence of Theorem 46.21 and Item (iii) proved above.
(v) Let ψm , ψn be two members of the Hilbert basis (ψk )k≥1 of L2 (D). Recalling that (λm , ψm )
and (λn , ψn ) are eigenpairs of (47.2), we infer that
−1 −1 1
−1
a(λm 2 ψm , λn 2 ψn ) = λm
2
λn 2 (ψm , ψn )L2 (D) = δmn .
Let W be the vector space composed of all the finite linear combinations of vectors in {ψn }n≥1 .
We have to prove that W is dense in V. Let f ∈ V ′ and assume that f annihilates W. Denoting by
(JVrf )−1 (f ) the Riesz–Fréchet representative of f in V equipped with the inner product a(·, ·), we
have
−1 −1 −1
0 = hf, λn 2 ψn iV ′ ,V = a((JVrf )−1 (f ), λn 2 ψn ) = a(λn 2 ψn , (JVrf )−1 (f ))
1
= λn2 (ψn , (JVrf )−1 (f ))L2 (D) ,
for all n ≥ 1, where we used the symmetry of a. The above identity implies that (JVrf )−1 (f ) = 0
since W is dense in L2 (D). Hence, f = 0. Corollary C.15 then implies that W is dense in V as
claimed.
The eigenvalues are henceforth counted with their multiplicity and ordered as follows: λ1 ≤
λ2 ≤ . . .. Moreover, the associated eigenfunctions ψ1 , ψ2 , . . . are chosen and normalized as in The-
orem 47.1(iv) so that kψn kL2 (D) = 1. The coercivity property of a implies that the eigenvalues are
all positive and larger than or equal to α. Notice that since T is symmetric, the notions of algebraic
and geometric multiplicity coincide, and for every eigenvalue λ−1 ∈ σp (T ), the multiplicity of λ is
equal to dim(λ−1 IL2 (D) − T ).
47.1.2 Rayleigh quotient

We introduce in this section the notion of Rayleigh quotient which will be instrumental in the
analysis of the H 1 -conforming approximation technique presented in §47.2.
Definition 47.2 (Rayleigh quotient). The Rayleigh quotient of a function v ∈ V \{0}, relative
to the bilinear form a, is defined as
a(v, v)
R(v) := . (47.4)
kvk2L2 (D)
In this chapter, all the expressions involving R(v) are understood with v 6= 0. For any functional
J : V → R, we write minv∈V J (v) instead of inf v∈V J (v) to indicate that the infimum is attained,
i.e., if there exists a minimizer v∗ ∈ V such that J (v∗ ) = inf v∈V J (v).
Proposition 47.3 (First eigenvalue). Let λ1 be the smallest eigenvalue of the problem (47.2)
and let ψ1 be a corresponding eigenfunction. Then we have
α ≤ λ1 = R(ψ1 ) = min R(v). (47.5)
v∈V
Proof. We have λ1 = R(ψ1 ) ≥ inf v∈V R(v) ≥ α, where the first equality results from a(ψ1 , ψ1 ) =
λ1 kψ1 k2L2 (D) and the second from Theorem 47.1(ii). It remains to prove that inf v∈V R(v) ≥ λ1 (this
also proves that the infimum of R over V is attained at ψ1 since λ1 = R(ψP 1 )). Let v ∈ V \{0}. Since
(ψn )n≥1 is a Hilbert basis of L2 (D) (see Theorem 47.1(iv)), the series ( k∈{1: n} Wk ψk )n≥1 , with
P
Wk := (v, ψk )L2 (D) , converges to v in L2 (D) and we have kvk2L2 (D) = n≥1 Wn2 . Furthermore, since
−1
(λn 2 ψn )n≥1 is a Hilbert basis of V equipped with the inner product a(·, ·) (see Theorem 47.1(v)),
P −1 −1
the series ( k∈{1: n} Vk λk 2 ψk )n≥1 , with Vk := a(v, λk 2 ψk ), converges to v in V , and we have
P 2 − 12 1 1
a(v, v) = n≥1 Vn . But we also have Vn = a(v, λn ψn ) = λn (v, ψn )L2 (D) = λn Wn . Since
2 2
λ1 ≤ λn for all n ≥ 1, we conclude that

P 2
P 2
n≥1 Vn n≥1 λn Wn
R(v) = P 2
= P 2
≥ λ1 .
n≥1 Wn n≥1 Wn
Proposition 47.4 (Min-max principle). Let Vm denote the set of the subspaces of V having
dimension m. For all m ≥ 1, we have
λm = min max R(v) = max min R(v), (47.6)

Em ∈Vm v∈Em Em−1 ∈Vm−1 v∈Em−1
⊥
⊥
where for all m > 1, Em−1 denotes the orthogonal of Em−1 in L2 (D) w.r.t. the L2 -inner product
and E0 := {0} by convention.
Proof. Let Wm := span{ψ1 , . . . , ψm }. Using the notation Wk := (v, ψk )L2 (D) , a direct computation
shows that P 2
n∈{1: m} λn Wn
min max R(v) ≤ max R(v) = max P 2
= λm .
Em ∈Vm v∈Em v∈Wm v∈Wm n∈{1: m} Wn
⊥
Consider now any Em ∈ Vm . A dimensional argument shows that there exists w 6= 0 in Em ∩Wm−1
2
(apply the rank nullity theorem to the L -orthogonal projection from Em onto Wm−1 ). Since w
P P 1
−1
can be written in the form w = n≥m Wn ψn = n≥m λn2 Wn λn 2 ψn , one shows by proceeding
as in the proof of Proposition 47.3 that R(w) ≥ λm . As a result, maxv∈Em R(v) ≥ λm . Hence,
minEm ∈Vm maxv∈Em R(v) ≥ λm . This concludes the proof of the first equality in (47.6). See
Exercise 47.4 for the proof of the second equality.
Remark 47.5 (Poincaré–Steklov constant). The best Poincaré–Steklov constant in H01 (D)
ℓD k∇vkL2 (D)
is Cps := inf v∈H01 (D)\{0} kvkL2 (D) . Letting λ1 be the smallest eigenvalue of the Laplacian with
1
Dirichlet boundary conditions, Proposition 47.3 shows that Cps = ℓD λ12 , and the Poincaré–Steklov
inequality becomes an equality when applied to the first eigenfunction ψ1 .

In this section, we investigate the H 1 -conforming finite element approximation of the spectral
problem (47.2).
47.2.1 Discrete setting and algebraic viewpoint

We assume that D is a Lipschitz polyhedron in Rd , and we consider a shape-regular sequence
(Th )h∈H of affine meshes so that each mesh covers D exactly. Depending on the boundary condi-
tions that are imposed in V, we denote by Vh the H 1 -conforming finite element space based on Th
g
such that Vh ⊂ V and Pk,0 (Th ) ⊆ Vh ⊆ Pkg (Th ) with k ≥ 1 (see §19.2.1 or §19.4). The approximate
eigenvalue problem we consider is the following:

Find ψh ∈ Vh \{0} and λh ∈ R such that
(47.7)
a(ψh , wh ) = λh (ψh , wh )L2 (D) , ∀wh ∈ Vh .
Let I := dim Vh , let {ϕi }i∈{1: I} be the global shape functions in Vh , and let Uh ∈ RI be the
coordinate vector of ψh relative to this basis. The discrete eigenvalue problem (47.7) can be recast
as follows:
Find Uh ∈ RI \{0} and λh ∈ R such that
(47.8)
AUh = λh MUh ,
where the stiffness matrix A and the mass matrix M have entries
Aij := a(ϕj , ϕi ) and Mij := (ϕj , ϕi )L2 (D) . (47.9)
Both matrices are symmetric positive definite since they are Gram matrices (see also §28.1). Be-
cause M is not the identity matrix, the problem (47.8) is called generalized eigenvalue problem.
Proposition 47.6 (Spectral problems). (i) (47.7) and (47.8) admit I (positive) eigenvalues
(counted with their multiplicity) {λhi }i∈{1: I} . (ii) The eigenfunctions {ψhi }i∈{1: I} ⊂ Vh in (47.7)
can be chosen so that a(ψhj , ψhi ) = λhi δij and (ψhj , ψhi )L2 (D) = δij . Equivalently, the eigenvectors
{Uhi }i∈{1: I} ⊂ RI in (47.8) can be chosen so that UT T
hj AUhi = λhi δij and Uhj MUhi = δij .
Proof. (i) Since A is symmetric and M is symmetric positive definite, these two matrices can be
simultaneously diagonalized. Let us recall the process for completeness. Let QQT be the Cholesky
factorization of M−1 , i.e., M = Q−T Q−1 . Since QT AQ is real and symmetric, there exists an
orthogonal matrix P (with PP T = II ), and a diagonal matrix Λ with diagonal entries (λhi )i∈{1: I} ,
such that QT AQ = PΛP −1 . Then AQP = Q−T PΛ = MQPΛ. Let us set U := QP and let
(Uhi )i∈{1: I} be the columns of the matrix U. The identity AU = MUΛ is equivalent to
AUhi = λhi MUhi , ∀i ∈ {1:I},
showing that the λhi ’s are the eigenvalues of the generalized eigenvalue problem (47.8) and the
Uhi ’s are the corresponding eigenvectors.
(ii) One readily sees that U T AU = P T QT Q−T PΛ = Λ and U T MU = P T QT Q−T Q−1 QP = II .
This proves the identities on the eigenvectors, and those on the eigenfunctions follow from the
definitions of A and M.
It is henceforth assumed that the eigenvalues are enumerated in increasing order λh1 ≤ . . . ≤
λhI , where each eigenvalue appears in this list as many times as its multiplicity. Moreover, the
eigenfunctions are chosen and normalized as in Proposition 47.6(ii) so that kψhi kL2 (D) = 1.
47.2.2 Eigenvalue error analysis

Let m ≥ 1 be a fixed natural number. We assume that h is small enough so that m ≤ I (recall
that I := dim(Vh ) grows roughly like (ℓD /h)d as h → 0). Our objective is to estimate |λhm − λm |.
Let us introduce the discrete solution map Gh : V → Vh defined s.t. a(Gh (v) − v, vh ) = 0 for all
v ∈ V and all vh in Vh (see §26.3.4 and §32.1). Let Wm := span{ψi }i∈{1: m} and let Sm be the
unit sphere of Wm in L2 (D). We define
kGh (v)kL2 (D)

σhm := min = min kGh (v)kL2 (D) . (47.10)
v∈Wm \{0} kvkL2 (D) v∈Sm
(Note that kGh (v)kL2 (D) attains its infimum over Sm since Sm is compact.)
Lemma 47.7 (Comparing λm and λhm ). Let m ∈ {1:I}. Assume that σhm 6= 0. The following
holds true:
−2
λm ≤ λhm ≤ σhm λm . (47.11)
P
Proof. Let wh = i∈{1: m} Wi ψhi ∈ Whm := span{ψhi }i∈{1: m} , where the eigenfunctions are
chosen and normalized
P P as in Proposition 47.6(ii), so that kψhi kL2 (D) = 1. Then R(wh ) =
2 2
i∈{1: m} λhi Wi / i∈{1: m} Wi . We infer that λhm = maxwh ∈Whm R(wh ), and the first inequality
in (47.11) is a consequence of Proposition 47.4. Let us now prove the second inequality. We observe
that ker(Gh ) ∩ Wm = {0} since σhm 6= 0 by assumption. Hence, the rank nullity theorem implies
that dim(Gh (Wm )) = m. Let Wh,m−1 = span{ψhi }i∈{1: m−1} and consider the L2 -projection
from Gh (Wm ) onto Wh,m−1 . The rank nullity theorem implies that P there is a nonzero vector
vh ∈ Gh (Wm ) such that vh is L2 -orthogonal to Wh,m−1 , so that vh = i∈{m: I} Vi ψhi . It follows
that R(vh ) ≥ λhm . As a result, we have
a(wh , wh ) a(Gh (v), Gh (v))

λhm ≤ R(vh ) ≤ max = max .
wh ∈Gh (Wm ) kwh k2L2 (D) v∈Wm kGh (v)k2 2
L (D)
1 1
Using that a(Gh (v), Gh (v)) = a(v, Gh (v)) ≤ a(v, v) 2 a(Gh (v), Gh (v)) 2 since a is symmetric and
coercive, we infer that a(Gh (v), Gh (v)) ≤ a(v, v). Recalling that maxv∈Wm R(v) = λm , we conclude
that
a(v, v) kvk2L2 (D)

λhm ≤ max ≤ max max R(v)
v∈Wm kGh (v)k2L2 (D) v∈Wm kGh (v)k2 2
L (D)
v∈Wm
−2 −2
= σhm max R(v) = σhm λm .
v∈Wm
Remark 47.8 (Guaranteed upper bound). It is remarkable that independently of the approx-
imation space, but provided conformity holds true, i.e., Vh ⊂ V, each eigenvalue of the discrete
problem (47.8) is larger than the corresponding eigenvalue of the exact problem (46.17). In other
words, the discrete eigenvalue λhm is a guaranteed upper bound on the exact eigenvalue λm for all
m ∈ {1:I}. Estimating computable lower bounds on the eigenvalues using conforming elements
is more challenging. We refer the reader to Cancès et al. [104] for a literature overview and to
Remark 48.13 when the approximation setting is nonconforming.
Lemma 47.9 (Lower bound on σhm ). Let m ∈ {1:I}. Recall that Sm is the unit sphere of
Wm := span{ψi }i∈{1: m} in L2 (D) and recall that Gh : Vh → V is the discrete solution operator.
The following holds true:
2
√ kak
σhm ≥1−2 m max kv − Gh (v)k2V . (47.12)
λ1 v∈Sm
Proof. Let v ∈ Sm . Let (Vi )i∈{1: m}Pbe the coordinate vector of v relative to the basis {ψi }i∈{1: m} .
Since (ψi , ψj )L2 (D) = δij , we have i∈{1: m} Vi2 = kvk2L2 (D) = 1. In addition, kGh (v)k2L2 (D) can be
bounded from below as
kGh (v)k2L2 (D) = kvk2L2 (D) − 2(v, v − Gh (v))L2 (D) + kv − Gh (v)k2L2 (D)
≥ kvk2L2 (D) − 2(v, v − Gh (v))L2 (D)
= 1 − 2(v, v − Gh (v))L2 (D) . (47.13)
Using that (λi , ψi ) is an eigenpair, the symmetry of a, and the Galerkin orthogonality property
satisfied by the discrete solution map, we have
X
(v, v − Gh (v))L2 (D) = Vi (ψi , v − Gh (v))L2 (D)
i∈{1: m}
X Vi X Vi
= a(ψi , v − Gh (v)) = a(ψi − Gh (ψi ), v − Gh (v)).
λi λi
i∈{1: m} i∈{1: m}
This implies that
kak X
(v, v − Gh (v))L2 (D) ≤ kv − Gh (v)kV |Vi |kψi − Gh (ψi )kV
λ1
i∈{1: m}
kak X
≤ max kw − Gh (w)k2V |Vi |
λ1 w∈Sm
i∈{1: m}
√ kak
≤ m max kw − Gh (w)k2V ,
λ1 w∈Sm
where we used the boundedness of a and λ1 ≤ λi for all i ∈ {1:m} in the first P bound, that
v∪{ψi }i∈{1: m} ⊂ Sm in the second bound, and the Cauchy–Schwarz inequality and i∈{1: m} Vi2 =
1 in the third bound. The expected estimate is obtained by inserting this bound into (47.13) and
taking the infimum over v ∈ Sm (recall that σhm := minv∈Sm kGh (v)kL2 (D) ).
√
Theorem 47.10 (Error on eigenvalues). Let m ∈ N\{0} and c1 (m) := 4 m kak kak
λ1 α . There is
h0 (m) > 0 s.t. for all h ∈ H ∩ (0, h0 (m)], we have σhm ≥ 21 and
0 ≤ λhm − λm ≤ λm c1 (m) max min kv − vh k2V . (47.14)

v∈Sm vh ∈Vh
Proof. (1) Since I grows unboundedly as h ↓ 0, there is h′0 (m) > 0 s.t. m ∈ {1:I} for all h ∈ H ∩
(0, h′0 (m)], i.e., the pair (λhm , ψhm ) exists for all h ∈ H∩(0, h′0 (m)]. Moreover, since the unit sphere
Sm is compact, there is v∗ (m) ∈ Sm such that maxv∈Sm kv − Gh (v)k2V = kv∗ (m) − Gh (v∗ (m))k2V .
The approximation property of the sequence (Vh )h∈H implies that there is h′′0 (m) > 0 such that
√
c0 (m)kv∗ (m) − Gh (v∗ (m))k2V ≤ 21 for all h ∈ H ∩ (0, h′′0 (m)], with c0 (m) := 2 m kak λ1 . We now
′ ′′ 1 1
set h0 (m) := min(h0 (m), h0 (m)). Observing that 1−x ≤ 1 + 2x for all x ∈ [0, 2 ], and applying
−2
this inequality to (47.12) with x := c0 (m) maxv∈Sm kv − Gh (v)k2V ≤ 12 , we infer that σhm ≤
2 1 1
1 + 2c0 (m) maxv∈Sm kv − Gh (v)kV . This implies in particular that σhm ≥ √2 ≥ 2 for all h ∈ H ∩
(0, h0 (m)].
(2) Inserting the above bound into (47.11) yields
−2
λhm − λm ≤ (σhm − 1)λm ≤ 2λm c0 (m) max kv − Gh (v)k2V .
v∈Sm
Since a is symmetric and coercive, Céa’s lemma (Lemma 26.13) implies that
21
kak
kv − Gh (v)kV ≤ min kv − vh kV . (47.15)
α vh ∈Vh
The assertion follows readily.
Remark 47.11 (Units). One readily sees that kak −2 −2d 2

λ1 scales as k·kL2 (D) , i.e., as ℓD . Since k·kV
2d 2
also scales like ℓD owing to our assumption on the boundedness of the embedding V ֒→ L (D),
we infer that the factor c1 (m) maxv∈Sm minvh ∈Vh kv − vh k2V is nondimensional.
Remark 47.12 (Double rate). The elliptic regularity theory implies that for all m ≥ 1, there
are s(m) > 0 and cm s.t. kψm kH 1+s(m) (D) ≤ cm . Here, the value of s(m) is not restricted to
the interval (0, 1] since there is a bootstrapping phenomenon that allows s(m) to be large. To
illustrate this property, assume that D is of class C r+1,1 , r ∈ N, and the bilinear form a is
associated with an operator A satisfying the assumptions of Theorem 31.29. Let s := r mod 2 ∈
{0, 1} and let l♯ ∈ N\{0} be s.t. 2(l♯ − 1) + s = r. Theorem 31.29 implies that there is c0 (r)
such that kA−1 (v)kH s (D) ≤ c0 (r)ℓ2D kvkL2 (D) for all v ∈ L2 (D), and there are cl (r), such that
kA−1 (v)kH 2l+s (D) ≤ cl (r)ℓ2D kvkL2(l−1)+s (D) for all v ∈ H 2(l−1)+s (D) and all l ∈ {1:l♯ }. Since
A(ψm ) = λm ψm , we obtain
♯
kψm kH r+2 (D) = kψm kH 2l♯ +s (D) ≤ cl♯ (r) . . . c1 (r)c0 (r)(λm ℓ2D )l +1
kψm kL2 (D) .
Recalling the normalization kψm kL2 (D) = 1, this argument shows that if D is of class C r+1,1 , we
♯
have kψm kH 1+s(m) (D) ≤ cm with s(m) := r + 1 and cm := cl♯ (r) . . . c1 (r)c0 (r)(λm ℓ2D )l +1 . Recalling
that k is the approximation degree of Vh , let s♭ (m) := min(s(1), . . . , s(m), k) for all m ≥ 1, and
1+s (m)
χ(m) := maxv∈Sm ℓD ♭ |v|H 1+s♭ (m) (D) (recall that Sm is the unit sphere of Wm in L2 (D)). The
best-approximation estimates established in §22.3 and §22.4 imply that there exists capp such that
the following holds true for all h ∈ H ∩ (0, h0 (m)]:
max min kv − vh kV ≤ capp χ(m)(h/ℓD )s♭ (m) .

v∈Sm vh ∈Vh
Owing to Theorem 47.10, this implies that
0 ≤ λhm − λm ≤ λm c1 (m)c2app χ(m)2 (h/ℓD )2s♭ (m) . (47.16)
In the best-case scenario where s(n) ≥ k for all n ∈ {1:m}, we have s♭ (m) = k so that the
convergence rate for the error on λm is O(h2k ), i.e., this error converges at a rate that is double
that of the best-approximation error on the eigenvectors in the H 1 -norm; see Remark 47.16 below.
Note that the convergence rate on λm in (47.16) depends on the smallest smoothness index of all
the eigenfunctions {ψn }n∈{1: m} . This shortcoming is circumvented with the more general theory
presented in Chapter 48, where the convergence rate on λm only depends on the smoothness index
of the eigenfunctions associated with λm . Note also that since c1 (m) grows unboundedly with m,
(47.16) shows that when h is fixed the accuracy of the approximation decreases as m increases.
Example 47.13 (1D Laplacian). Let us consider the eigenvalue problem for the one-dimensional
Laplacian discretized using P1 Lagrange elements on a uniform mesh on D := (0, 1). It is shown in
Exercise 47.5 that λm = m2 π 2 and λhm = h62 1−cos(mπh)
2+cos(mπh) for all m ≥ 1. The left panel of Figure 47.1
shows the first 50 exact eigenvalues and the 50 discrete eigenvalues on a mesh having I := 50
internal vertices. The exact eigenvalues are approximated from above as predicted in Lemma 47.7.
Observe that only the first eigenvalues are approximated accurately. The reason for this is that the
eigenfunctions corresponding to large eigenvalues oscillate too much to be represented accurately
on the mesh as √
illustrated in the right panel of Figure 47.1. A rule of thumb is that a meshsize
ǫ
smaller than m must be used to approximate the m-th eigenvalue with relative accuracy ǫ, i.e.,
|λhm −λm | < ǫλm . For instance, only the first 10 eigenvalues are approximated within 1% accuracy
when I := 100. We refer the reader to Exercise 47.5 for further details.
47.2.3 Eigenfunction error analysis

The goal of this section is to estimate the approximation error on the eigenfunctions. We first
estimate this error in the L2 -norm and then in the H 1 -norm. Let m ≥ 1 be a fixed natural
number, and let us assume as in the previous section that the meshsize h ∈ H is small enough
so that m ≤ I and σhm > 0 (see Theorem 47.10). For the sake of simplicity, we also assume
that the eigenvalue λm is simple, and we set γm := 2 maxi∈N\{0,m} |λmλ−λ
m
i|
. Observe that γm =
λm λm
2 max( λm −λ ,
m−1 λm+1 −λm
). Since λhi → λi as h → 0 for all i ∈ {1:m+1} (see Theorem 47.10),
4
x 10
3.5 Eigenvalues 1
0.5
2.5
λ
hm
2 λ
m
0
1.5
1 -0.5
0.5
-1
0
0 10 20 30 40 50 0.4 0.45 0.5 0.55 0.6
Figure 47.1: P1 approximation of the eigenvalues of the Laplacian in one dimension. Left: discrete
and exact eigenvalues, I := 50. Right: Graph of the 80th exact (dashed line) and discrete (solid
line) eigenfunctions in the interval (0.4, 0.6), I := 100.
there exists h0 (m) > 0 so that |λmλ−λ

m
hi |
≤ γm for all i ∈ {1:m + 1}\{m} and all h ∈ H ∩ (0, h0 (m)].
Moreover, using that |λm − λhi | ≤ |λm − λm+1 | for all i ≥ m + 1, we infer that the following holds
true for all h ∈ H ∩ (0, h0 (m)]:
λm
max ≤ γm . (47.17)
i∈{1: I} |λm − λhi |
i6=m
Theorem 47.14 (L2 -error on eigenfunctions). Let m ∈ N\{0}. Assume that λm is simple
and let h0 (m) > 0 be s.t. (47.17) holds true. Let c2 (m) := 2(1 + γm ). There is an eigenfunction
ψm such that the following holds true for all h ∈ H ∩ (0, h0 (m)]:
kψm − ψhm kL2 (D) ≤ c2 (m)kψm − Gh (ψm )kL2 (D) . (47.18)

P
Proof. Recall that Gh (ψm ) = i∈{1: I} Vi ψhi with Vi := (Gh (ψm ), ψhi )L2 (D) . Let us set vhm :=
P
Vm ψhm so that Gh (ψm ) − vhm = i∈{1: I}\{m} Vi ψhi . Since the bilinear form a is symmetric and
(λhi , ψhi ) is a discrete eigenpair, we have
1 1
Vi = a(ψhi , Gh (ψm )) =
λhi λhi a(Gh (ψm ), ψhi )
1 λm
= a(ψm , ψhi ) = (ψm , ψhi )L2 (D) ,
λhi λhi
where we used the definition of Gh and that (λm , ψm ) is an eigenpair. This implies that
(λhi − λm )Vi = λhi Vi − λm Vi = λm (ψm , ψhi )L2 (D) − λm Vi

= λm (ψm , ψhi )L2 (D) − λm (Gh (ψm ), ψhi )L2 (D)
= λm (ψm − Gh (ψm ), ψhi )L2 (D) .
Hence, we have Vi = λhiλ−λ m

m
(ψm − Gh (ψm ), ψhi )L2 (D) for all i ∈ {1:I}\{m}. Since the discrete
eigenfunctions {ψhi }i∈{1: I} are L2 -orthonormal, we obtain
X X
kGh (ψm ) − vhm k2L2 (D) = Vi2 ≤ γm2
(ψm − Gh (ψm ), ψhi )2L2 (D)
i∈{1: I} i∈{1: I}
i6=m i6=m
2
≤ γm kψm − Gh (ψm )k2L2 (D) , (47.19)
P
where the first bound follows from (47.17) and the last one from Bessel’s inequality i∈{1: I} (ψm −
Gh (ψm ), ψhi )2L2 (D) ≤ kψm − Gh (ψm )k2L2 (D) . Let us now estimate kψhm − vhm kL2 (D) . Since
kψhm kL2 (D) = 1, we have
kψhm − vhm kL2 (D) = k(1 − Vm )ψhm kL2 (D) = |Vm − 1|

= |(Gh (ψm ), ψhm )L2 (D) − 1|.
Assume that ψhm is chosen so that Vm = (Gh (ψm ), ψhm )L2 (D) ≥ 0. Then we have kvhm kL2 (D) =
|Vm | = (Gh (ψm ), ψhm )L2 (D) , and kψhm − vhm kL2 (D) = |kvhm kL2 (D) − 1|. Since the triangle in-
equality implies that
kψm kL2 (D) − kψm − vhm kL2 (D) ≤ kvhm kL2 (D) ≤ kψm kL2 (D) + kψm − vhm kL2 (D) ,
and since kψm kL2 (D) = 1, we infer that |kvhm kL2 (D) − 1| ≤ kψm − vhm kL2 (D) . This implies that
kψhm − vhm kL2 (D) = |kvhm kL2 (D) − 1| ≤ kψm − vhm kL2 (D) .
Invoking the triangle inequality, the above bound, and the triangle inequality one more time gives
kψm − ψhm kL2 (D) ≤ kψm − Gh (ψm )kL2 (D) + kGh (ψm ) − vhm kL2 (D) + kψhm − vhm kL2 (D)
≤ kψm − Gh (ψm )kL2 (D) + kGh (ψm ) − vhm kL2 (D) + kψm − vhm kL2 (D)
≤ 2(kψm − Gh (ψm )kL2 (D) + kGh (ψm ) − vhm kL2 (D) )
≤ 2(1 + γm )kψm − Gh (ψm )kL2 (D) ,
where the last bound follows from (47.19). Using the definition of c2 (m) leads to the expected
estimate.
Theorem 47.15 (H 1 -error on eigenfunctions). Let m ∈ N\{0}. Assume that λm is simple and
let h0 (m) > 0 be s.t. (47.14) and (47.17) hold for all h ∈ H ∩ (0, h0 (m)]. There is an eigenfunction
ψm such that the following holds true for all h ∈ H ∩ (0, h0 (m)]:
kψm − ψhm kV ≤ c3 (m) max min kv − vh kV , (47.20)

v∈Sm vh ∈Vh
where c3 (m) := ( λαm ) 2 (c1 (m) + c2 (m)2 kak

1 1
α ) is independent of h ∈ H.
2
Proof. Owing to the coercivity of a, we infer that
αkψm − ψhm k2V ≤ a(ψm − ψhm , ψm − ψhm )

= λhm + λm − 2λm (ψm , ψhm )L2 (D)
= λhm − λm + λm kψm − ψhm k2L2 (D) ,
since kψm kL2 (D) = kψhm kL2 (D) = 1 implies that kψm − ψhm k2L2 (D) = 2 − 2(ψm , ψhm )L2 (D) . The
inequality (47.20) is obtained by estimating (λhm − λm ) and kψm − ψhm k2L2 (D) . The estimate on
(λhm − λm ) is given by (47.14) in Theorem 47.10, and Theorem 47.14 gives kψm − ψhm kL2 (D) ≤
c2 (m)kψm − Gh (ψm )kL2 (D) . We observe that
kψm − Gh (ψm )kL2 (D) ≤ kψm − Gh (ψm )kV ≤ max kv − Gh (v)kv

v∈Sm
21
kak
≤ min kv − vh kV ,
α vh ∈Vh
where the last bound follows from (47.15) (Céa’s lemma). Putting everything together leads to
the expected estimate.
Remark 47.16 (Convergence rates). Let us use the notation of Remark 47.12. Assume that the
eigenvalue λm is simple. We can then invoke the estimates from Theorem 47.14 and Theorem 47.15.
The best-approximation estimates in the H 1 -norm established in §22.3 and §22.4 and the Aubin–
Nitsche lemma (Lemma 32.11) imply that the following holds true for all h ∈ H ∩ (0, h0 (m)]:
kψm − ψhm kL2 (D) ≤ č2 (m)χ(m)(h/ℓD )s♭ (m)+s , (47.21a)

s♭ (m)
kψm − ψhm kH 1 (D) ≤ č3 (m)χ(m)(h/ℓD ) , (47.21b)
where the constants č2 (m), č3 (m) have the same dependencies w.r.t. m as the constants c2 (m), c3 (m),
and χ(m) is defined in Remark 47.12. The best possible convergence rates are obtained when
sn (m) ≥ k for all n ∈ {1:m} so that s♭ (m) = k, yielding the rates O(hk+1 ) in the L2 -norm
and O(hk ) in the H 1 -norm. Moreover, it can be shown that if λm has multiplicity p, i.e.,
†
λm = λm+1 = . . . = λm+p−1 , then there exists an eigenfunction ψm ∈ span{ψm , . . . , ψm+p−1 }
† †
with kψm kL2 (D) = 1 such that (47.21) holds true with ψm replaced by ψm . Note that (47.21)
shows that when h is fixed, the accuracy of the approximation decreases as m increases, since
c2 (m), c3 (m) grow unboundedly with m.
Exercises
Exercise 47.1R(Real eigenvalues). Consider
R the eigenvalue problem: Find ψ ∈ H01 (D; C)\{0}
and λ ∈ C s.t. D (∇ψ·∇w + ψw) dx = λ D ψw dx for all w ∈ H01 (D; C). Prove directly that λ is
real. (Hint : test with w := ψ.)
Exercise 47.2 (Smallest eigenvalue). Let D1 ⊂ D2 be two Lipschitz domains in Rd . Let
ai : H01 (Di )×H01 (Di ) → R, i ∈ {1, 2}, be two symmetric, coercive, bounded bilinear forms. Assume
that a1 (v, w) = a2 (e e for all v, w ∈ H01 (D1 ), where ve, w
v , w) e denote the extension by zero of v, w, re-
spectively. Let λ1 (Di ) be the smallest eigenvalue of the eigenvalue problem: Find ψ ∈ H01 (Di )\{0}
and λ ∈ R s.t. ai (ψ, w) = λ(ψ, w)L2 (Di ) for all w ∈ H01 (Di ). Prove that λ1 (D2 ) ≤ λ1 (D1 ). (Hint :
use Proposition 47.3.)
Exercise 47.3 (Continuity of eigenvalues). Consider the setting defined in §47.1. Let a1 , a2 :
V ×V → R be two symmetric, coercive, bounded bilinear forms. Let A1 , A2 : V → V ′ be the linear
operators defined by hAi (v), wiV ′ ,V := ai (v, w), i ∈ {1, 2}, for all v, w ∈ A. Let λk (a1 ) and λk (a2 )
be the k-th eigenvalues, respectively. Prove that |λk (a1 )− λk (a2 )| ≤ supv∈S |h(A1 − A2 )(v), viV ′ ,V |,
where S is the unit sphere in L2 (D). (Hint : use the min-max principle.)
Exercise 47.4 (Max-min principle). Prove the second equality in (47.6). (Hint : let Em−1 ∈
⊥
Vm−1 and observe that Em−1 ∩ Wm 6= {0}.)
Exercise 47.5 (Laplacian, 1D). Consider the spectral problem for the 1D Laplacian on D :=
(0, 1). (i) Show that the eigenpairs (λm , ψm ) are λm = m2 π 2 , ψm (x) = sin(mπx), for all x ∈ D
1
and all m ≥ 1. (ii) Consider a uniform mesh of D of size h := I+1 and H 1 -conforming P1 finite
elements. Compute the stiffness matrix A and the mass matrix M. (iii) Show that the eigenvalues
1−cos(mπh)
of the discrete problem (47.8) are λhm = h62 ( 2+cos(mπh) ) for all m ∈ {1:I}. (Hint : consider the
vectors (sin(πhml))l∈{1: I} for all m ∈ {1:I}.)
Exercise 47.6 (Stiffness matrix). Assume that the mesh sequence (Th )h∈H is quasi-uniform.
Estimate from below the smallest eigenvalue of the stiffness matrix A defined in (47.9) and estimate
from above its largest eigenvalue. (Hint : see §28.2.3.)
Chapter 48
Nonsymmetric problems
In this chapter, we continue our investigation of the finite element approximation of eigenvalue
problems, but this time we do not assume symmetry and we explore techniques that can handle
nonconforming approximation settings. The main abstract results used in the present chapter are
based on a theory popularized in the landmark review article by Babuška and Osborn [38]. Some
results are simplified to avoid invoking spectral projections. Our objective is to show how to apply
this abstract theory to the conforming and nonconforming approximation of eigenvalue problems
arising from variational formulations.
48.1 Abstract theory

In this section, we present an abstract theory for the approximation of the spectrum of compact
operators in complex Banach spaces, and we show how to apply it to spectral problems arising
from variational formulations.
48.1.1 Approximation of compact operators

Let L be a complex Banach space and T ∈ L(L) be a compact operator. We assume that we have
at hand a sequence of compact operators Tn : L → L, n ∈ N, that converges in norm to T i.e., we
assume that
lim kT − Tn kL(L) = 0. (48.1)
n→∞
We want to estimate how the eigenpairs of each member in the sequence (Tn )n∈N approximate
some of the eigenpairs of T .
Recall that σ(T )\{0} = σp (T )\{0} and that the nonzero eigenvalues of T are isolated since
T is compact; see Items (ii)-(iii) in Theorem 46.14. Let µ ∈ σp (T )\{0} be a nonzero eigenvalue
of T . Let α be the ascent of µ. Recall that α is the smallest integer with the property that
ker(µIL − T )α = ker(µIL − T )α+1 . Denoting by T ∗ : L′ → L′ the adjoint of T , we set
Gµ := ker(µIL − T )α , G∗µ := ker(µIL′ − T ∗ )α , (48.2a)

m := dim(Gµ ) = dim(G∗µ ). (48.2b)
Members of Gµ and G∗µ are called generalized eigenvectors. The generalized eigenvectors are all
eigenvectors only if α = 1. Recall that m is the algebraic multiplicity of µ and that m ≥ α; see
278 Chapter 48. Nonsymmetric problems
(46.5). Owing to the above assumption on norm convergence, it can be shown that there are m
eigenvalues of Tn , say {µn,j }j∈{1: m} (counted with their algebraic multiplicities), that converge to
µ as n → ∞. Let αn,j be the ascent of µn,j and let us set
X
Gn,µ := ker(µn,j IL − Tn )αn,j . (48.3)
j∈{1: m}
We want to evaluate how close the subspaces Gµ and Gn,µ are, and for this purpose we define the no-
tion of gap. Given two closed subspaces of L, Y, and Z, we define δ(Y, Z) := supy∈Y ;kykL =1 dist(y, Z),
where dist(y, Z) := inf z∈Z ky − zkL . The gap between Y and Z is defined by
b Z) := max(δ(Y, Z), δ(Z, Y )).
δ(Y,
Theorem 48.1 (Bound on eigenspace gap). Assume (48.1). Let µ ∈ σp (T )\{0}. Let Gµ be
defined in (48.2a) and let Gn,µ be defined in (48.3). There is c, depending on µ, such that for all
n ∈ N,
b µ , Gn,µ ) ≤ c k(T − Tn )|G kL(G ;L) .
δ(G (48.4)
µ µ
Proof. See Osborn [321, Thm. 1] or Babuška and Osborn [38, Thm. 7.1].
Let us now examine the convergence of the eigenvalues. When α, the ascent of µ, is larger than
one, it is interesting to consider the convergence of the arithmetic mean of the eigenvalues µn,j .
We will see that this quantity converges faster than any of the µn,j (for instance, compare (48.5)
and (48.6), and see (48.21) in Theorem 48.8).
Theorem 48.2 (Convergence of eigenvalues). Assume (48.1). Let µ ∈ σp (T )\{0} with al-
gebraic multiplicity m. Let {µn,j }j∈{1: m} be the eigenvalues of Tn that converge to µ and set
1
P
hµn i := m j∈{1: m} µn,j . There is c, depending on µ, such that for all n ∈ N,
1 |hw, (T − Tn )(v)iL′ ,L |
|µ − hµn i| ≤ max ∗
m (v,w)∈Gµ ×Gµ kwkL′ kvkL
+ c k(T − Tn )|Gµ kL(Gµ ;L) k(T − Tn )∗|G∗µ kL(G∗µ ;L′ ) , (48.5)
and for all j ∈ {1:m},

|hw, (T − Tn )(v)iL′ ,L |
|µ − µn,j | ≤ c max
(v,w)∈Gµ ×G∗
µ kwkL′ kvkL
α1
+ k(T − Tn )|Gµ kL(Gµ ;L) k(T − Tn )∗|G∗µ kL(G∗µ ;L′ ) . (48.6)
Proof. See [321, Thm. 3&4], [38, Thm. 7.2&7.3], and Exercise 48.3.
Finally, we evaluate how the vectors in Gn,µ approximate those in Gµ .
Theorem 48.3 (Convergence of eigenvectors). Assume (48.1). Let µ ∈ σp (T )\{0} with
algebraic multiplicity m. Let {µn,j }j∈{1: m} be the eigenvalues of Tn that converge to µ. For
all integers j ∈ {1:m} and ℓ ∈ {1:α}, let wn,j be a unit vector in ker(µn,j IL − Tn )ℓ . There is c,
′
depending on µ, such that for every integer ℓ′ ∈ {ℓ:α}, there is a unit vector uℓ′ ∈ ker(µIL − T )ℓ ⊂
Gµ such that for all n ∈ N,
ℓ′ −ℓ+1
kuℓ′ − wn,j kL ≤ c k(T − Tn )|Gµ kL(Gαµ ;L) . (48.7)
Proof. See [321, Thm. 5] or [38, Thm. 7.4].

Remark 48.4 (Literature). The above theory has been developed by Bramble and Osborn [78],
Osborn [321], Descloux et al. [161, 162]; see Vainikko [368, 369], Strang and Fix [359] for earlier
references. Overviews can also be found in Boffi [62], Chatelin [116, Chap. 6].
Remark 48.5 (Sharper bounds). The bounds in Theorem 48.2 are simplified versions of the
|hw,(T −Tn )(v)iL′ ,L |
estimates given in [321, Thm. 3&4]. Therein, instead of max(v,w)∈Gµ ×G∗µ kwkL′ kvkL , one
P ∗ ∗
has j∈{1: m} |hφj , (T − Tn )(φj )iL′ ,L |, where {φj }j∈{1: m} is a basis of Gµ and {φj }j∈{1: m} is a
dual basis of G∗µ , i.e., hφ∗j , φk iL′ ,L = δjk and the action of the forms φ∗j outside Gµ is defined by
selecting an appropriate complement of Gµ . The expressions given in Theorem 48.2 will suffice for
our purpose.
48.1.2 Application to variational formulations

Let V ֒→ L be a complex Banach space with compact embedding and let a : V ×V → C be a
bounded sesquilinear form. We assume that the sesquilinear form a satisfies the two conditions of
the BNB theorem (Theorem 25.9), but we do not assume that a is Hermitian. Let b : L×L → C
be another bounded sesquilinear form. We now consider the following eigenvalue problem:

Find ψ ∈ V \{0} and λ ∈ C such that
(48.8)
a(ψ, w) = λb(ψ, w), ∀w ∈ V.
If (λ, ψ) solves (48.8), we say that (λ, ψ) is an eigenpair of the form a relative to the form b, or
simply (λ, ψ) is an eigenpair of (48.8) when the context is unambiguous.
To reformulate (48.8) so as to fit the approximation theory of the spectrum of compact operators
from §48.1.1, we define the solution operator T : L → V ֒→ L such that
a(T (v), w) := b(v, w), ∀v ∈ L, ∀w ∈ V. (48.9)
Note that T (v) is well defined for all v ∈ L since a satisfies the two BNB conditions. Notice also
that im(T ) ⊂ V and that T is injective.
Proposition 48.6 (Spectrum of T ). (i) 0 6∈ σp (T ). (ii) (µ, ψ) ∈ C×V is an eigenpair of T iff
(µ−1 , ψ) ∈ C×V is an eigenpair of (48.8).
Proof. (i) If (0, ψ) is an eigenpair of T (i.e., ψ 6= 0), then a(ψ, v) = 0 for all v ∈ V, and the inf-sup
condition on a implies that ψ = 0, which is a contradiction.
(ii) Let (µ, ψ) be an eigenpair of T , i.e., µ−1 T (ψ) = ψ (notice that µ 6= 0 since T is injective). We
infer that
µ−1 b(ψ, w) = b(µ−1 ψ, w) = a(T (µ−1 ψ), w) = a(µ−1 T (ψ), w) = a(ψ, w),
for all w ∈ V. Hence, (µ−1 , ψ) is an eigenpair of (48.8). The proof of the converse statement is
identical.
We refer the reader to §46.2 for various examples of spectral problems that can be put into the
variational form (48.8). For instance, the model
R problem (46.21) leads to a sesquilinear form a that
is not Hermitian since we haveRa(v, w) := D (g ′ (usw )∇v·∇w + vg ′′ (usw )∇usw ·∇w − f ′ (usw )vw) dx,
1
V := Hper (D), and b(v, w) := D vw dx. An example with a sesquilinear form b that is not the L2 -
inner product is obtained from the vibrating string model from §46.2.1 by assuming that the string
has a nonuniform bounded linear density ρ. In this case, one recovers the model Rproblem (48.8)
with V := H01 (D; R), D := (0, ℓ), where ℓ is the length of the string,
R a(v, w) := D τ ∂x v∂x w dx,
where τ > 0 is the uniform tension of the string, and b(v, w) := D ρvw dx.
48.2 Conforming approximation

The goal of this section is to illustrate the approximation theory from §48.1 when applied to the
conforming approximation of the model problem (48.8). Let V be a closed subspace of H 1 (D)
which, depending on the boundary conditions that are enforced, satisfies H01 (D) ⊆ V ⊆ H 1 (D).
We assume that V is equipped with a norm that is equivalent to that of H 1 (D). We assume also
that the V -norm is rescaled so the operator norm of the embedding V ֒→ L2 (D) is at most one,
−1
e.g., one could set kvkV := Cps ℓD k∇vkL2 (D) if V := H01 (D), where Cps is the constant from the
Poincaré–Steklov inequality (31.12) in H01 (D) and ℓD is a characteristic length associated with D,
e.g., ℓD := diam(D).
Let T : L2 (D) → L2 (D) be the compact operator defined in (48.9). We identify L and L′ ,
so that T ∗ = T H (see Lemma 46.15). We want to approximate the spectrum of T assuming that
we have at hand an H 1 -conforming approximation setting. More precisely, assume that D is a
Lipschitz polyhedron and let (Th )h∈H be a shape-regular sequence of affine meshes so that each
mesh covers D exactly. Let k ≥ 1 be the polynomial degree of the approximation. We denote by
g
Vh the H 1 -conforming finite element space based on Th such that Pk,0 (Th ) ⊆ Vh ⊆ Pkg (Th ) and
Vh ⊂ V (see §19.2.1 or §19.4). To avoid being specific on the type of finite element we use, we
assume the following best-approximation result:
min kv − vh kV ≤ c hr ℓD |v|H 1+r (D) , (48.10)

vh ∈Vh
for all v ∈ H 1+r (D) ∩ V and all r ∈ [0, k]. We assume that there is α0 > 0 such that for all h ∈ H,
|a(vh , wh )|
inf sup ≥ α0 . (48.11)
vh ∈Vh wh ∈Vh kvh kH 1 (D) kwh kH 1 (D)
Since the sesquilinear form b may differ from the L2 -inner product, we additionally introduce
the linear operator S∗ : L2 (D) → V ֒→ L2 (D) s.t.
a(v, S∗ (w)) = (v, w)L2 (D) , ∀v ∈ V, ∀w ∈ L2 (D). (48.12)
Notice that we use the L2 -inner product on the right-hand side of (48.12) instead of the sesquilinear
form b as we did for the definition of T in (48.9). We also assume that the following elliptic regularity
pickup holds true for T and S∗ (see §31.4.2): There are real numbers τ, τ ∗ ∈ (0, 1] such that
∗
T ∈ L(L2 (D); H 1+τ (D)), S∗ ∈ L(L2 (D); H 1+τ (D)). (48.13)
We have τ = τ ∗ := 1 when maximal elliptic regularity occurs.

The discrete counterpart of the eigenvalue problem (48.8) is formulated as follows:

Find ψh ∈ Vh \{0} and λh ∈ C such that
(48.14)
a(ψh , wh ) = λh b(ψh , wh ), ∀wh ∈ Vh .
We define the discrete solution operator Th : L2 (D) → Vh ⊂ L2 (D) s.t. for all v ∈ L2 (D),
Th (v) ∈ Vh is the unique solution to the following problem:
a(Th (v), wh ) = b(v, wh ), ∀wh ∈ Vh .
Notice that 0 cannot be an eigenvalue of (48.14) owing to the inf-sup condition (48.11) satisfied
by a on Vh ×Vh . Moreover, (λh , ψh ) is an eigenpair of (48.14) iff (λ−1
h , ψh ) is an eigenpair of Th .
Lemma 48.7 (Bound on (T − Th )). There is c such that for all t, t∗ ∈ [0, k], all v ∈ L2 (D) s.t.
∗
T (v) ∈ H 1+t (D), all w ∈ L2 (D) s.t. S∗ (w) ∈ H 1+t (D), and all h ∈ H,

((T −Th )(v), w)L2 (D) ≤ c ht+t∗kakℓ2D |T (v)|H 1+t (D) |S∗ (w)|H 1+t∗ (D) . (48.15)
Proof. Lemma 26.14 and the best-approximation property (48.10) imply that k(T − Th )(v)kV ≤
cht ℓD |T (v)|H 1+t (D) . Since (T − Th )(v) ∈ V, the Galerkin orthogonality property and the bound-
edness of a imply that

((T − Th )(v), w)L2 (D) = |a((T − Th )(v), S∗ (w))|
≤ inf |a((T − Th )(v), S∗ (w) − wh )|
wh ∈Vh
≤ kak k(T − Th )(v)kV inf kS∗ (w) − wh kV .

wh ∈Vh
Using the above bound on (T − Th )(v) and the best-approximation property (48.10) to bound
kS∗ (w) − wh kV leads to the expected estimate.
The estimate (48.15) with t := τ and t∗ := τ ∗ combined with the regularity property (48.13)
implies that
∗
kT − Th kL(L2 ;L2 ) ≤ c hτ +τ kakℓ2D kT kL(L2;H 1+τ ) kS∗ kL(L2 ;H 1+τ ∗ ) . (48.16)
Since τ + τ ∗ > 0, this means that Th → T in operator norm as h → 0, that is, the key assump-
tion (48.1) holds true. It is then legitimate to use the approximation results for compact operators
stated in Theorems 48.1 to 48.3.
Let µ be a nonzero eigenvalue of T of ascent α and algebraic multiplicity m, and let
Gµ := ker(µIL2 − T )α , G∗µ := ker(µIL2 − T H )α , (48.17)
so that m := dim(Gµ ) = dim(G∗µ ) (see (48.2)). Recall that Proposition 48.6 implies that λ := µ−1
is an eigenvalue for (48.8). Since the smoothness of the generalized eigenvectors may differ from
one eigenvalue to the other, we now define τµ and τµ∗ to be the two largest real numbers in (0, k]
such that
∗
T|Gµ ∈ L(Gµ ; H 1+τµ (D)), S∗|G∗µ ∈ L(G∗µ ; H 1+τµ (D)), (48.18)
where Gµ and G∗µ are equipped with the L2 -norm. The two real numbers τµ and τµ∗ measure the
smoothness of the generalized eigenvectors in Gµ and G∗µ , respectively. Notice that τµ ∈ [τ, k] and
τµ∗ ∈ [τ ∗ , k], where τ and τ ∗ are defined in (48.13) and are both in (0, 1]. We can set τµ = τµ∗ := k
when maximal smoothness is available. It may Rhappen that τµ < τµ∗ even if a is Hermitian.
For instance, this may be the case if b(v, w) := D ρvw dx, where the function ρ is a bounded
discontinuous function.
Owing to the norm convergence of Th to T as h → 0, there are m eigenvalues of Th , say
{µh,j }j∈{1: m} (counted with their algebraic multiplicities), that converge to µ as h → 0. Let
X
Gh,µ := ker(µh,j IL2 − Th )αh,j , (48.19)
j∈{1: m}
where αh,j is the ascent of µh,j . We are now in the position to state the main result of this section.
Theorem 48.8 (Convergence of eigenspace gap, eigenvalues, and eigenvectors). Let

µ ∈ σp (T ) \ {0} with algebraic multiplicity m and let {µh,j }j∈{1: m} be the eigenvalues of Th that
converge to µ. Let Gµ be defined in (48.17) and let Gh,µ be defined in (48.19). There is c, depending
on µ, such that for all h ∈ H,
b µ , Gh,µ ) ≤ c hτµ +t∗ ,
δ(G (48.20)
1
P
and letting hµh i := m j∈{1: m} µh,j , we have
∗ 1 ∗
|µ − hµh i| ≤ c hτµ +τµ , |µ − µh,j | ≤ c h α (τµ +τµ ) , ∀j ∈ {1:m}. (48.21)
Moreover, for all integers j ∈ {1:m} and ℓ ∈ {1:α}, let wh,j be a unit vector in ker(µh,j IL2 − Th )ℓ .
There is c, depending on µ, such that for every integer ℓ′ ∈ {ℓ:α}, there is a unit vector uℓ′ ∈
′
ker(µIL2 − T )ℓ ⊂ Gµ such that for all h ∈ H,
ℓ′ −ℓ+1
(τµ +τ ∗ )
kuℓ′ − wh,j kL2 (D) ≤ c h α . (48.22)
In the above estimates, the constant c depends on kakℓ2D and on the operator norms resulting
from (48.13) and (48.18).
Proof. Using t := τµ and t∗ := τ ∗ in (48.15), we infer that
((T − Th )(v), w)L2 ∗

k(T − Th )|Gµ kL(Gµ ;L2 ) = sup sup ≤ c hτµ +τ .
v∈Gµ w∈L2 kvkL2 kwkL2
Similarly, using t := τ and t∗ := τµ∗ in (48.15), and recalling that T ∗ = T H in the present case, we
infer that
(v, (T H − ThH )(w))L2
k(T − Th )∗ |G∗µ kL(G∗µ ;L2 ) = sup sup
v∈L2 w∈G∗
µ
kvkL2 kwkL2
((T − Th )(v), w)L2 ∗
= sup sup ≤ c hτ +τµ .
v∈L2 w∈G∗
µ
kvkL2 kwkL2
Finally, using t := τµ and t∗ := τµ∗ in (48.15), we infer that
((T − Th )(v), w)L2 ∗

sup sup ≤ c hτµ +τµ .
v∈Gµ w∈G∗
µ
kvkL2 kwkL2
The conclusion follows by applying Theorems 48.1-48.3.
Remark 48.9 (Convergence rates). Notice that among the two terms that compose the right-
hand side in (48.5), it is ∗the first one that dominates when the meshsize
∗
goes∗ to zero. The first
term scales like O(hτµ +τµ ), whereas the second one scales like O(hτµ +τµ +τ +τ ) with τ + τ ∗ > 0.
The same observation is valid for (48.6).
Remark 48.10 (Symmetric case). The estimate (48.21) coincides with the estimate (47.16),
and the estimate (48.22) (with α = ℓ = ℓ′ := 1) coincides with the estimate (47.21) when T is
symmetric. Notice though that the estimates from Chapter 47 for the i-th eigenpair depend on
the smoothness of all the unit eigenfunctions {ψn }n∈{1: i} (counting the multiplicities), whereas
the estimates (48.21)-(48.22) depend only on the smoothness of the unit eigenvectors in Gµi ; see
Remark 47.12.
48.3 Nonconforming approximation

We revisit the theory presented above in a nonconforming context. Typical examples we have
in mind are the Crouzeix–Raviart approximation from Chapter 36, Nitsche’s boundary penalty
technique from Chapter 37, and the discontinuous Galerkin method from Chapter 38. The theory
is also applicable to the hybrid high-order method from Chapter 39.
48.3.1 Discrete formulation

We consider again the model problem (48.8) and we want to approximate the spectrum of the op-
erator T : L2 (D) → L2 (D) defined in (48.9) using an approximation setting that is not conforming
in V.
To stay general, we assume that we have at hand a sequence of discrete spaces (Vh )h∈H with
Vh 6⊂ V. For all h ∈ H, the sesquilinear form a is approximated by a discrete sesquilinear form
ah : Vh ×Vh → C, and for simplicity we assume that the sesquilinear form b is meaningful on
Vh ×Vh , i.e., we assume that Vh ⊂ L2 (D). The discrete eigenvalue problem is formulated as
follows:

Find ψh ∈ Vh \{0} and λh ∈ C such that
(48.23)
ah (ψh , wh ) = λh b(ψh , wh ), ∀wh ∈ Vh .
The discrete solution operator Th : L2 (D) → Vh ⊂ L2 (D) and the adjoint discrete solution operator
S∗h : L2 (D) → Vh ⊂ L2 (D) are defined as follows:
ah (Th (v), wh ) := b(v, wh ), ∀(v, wh ) ∈ L2 (D)×Vh , (48.24a)

2
ah (vh , S∗h (w)) := (vh , w)L2 (D) , ∀(vh , w) ∈ Vh ×L (D). (48.24b)
We assume that Th and S∗h are both well defined, i.e., we assume that ah satisfies an inf-sup
condition on Vh ×Vh uniformly w.r.t. h ∈ H. As above, (λh , ψh ) is an eigenpair of (48.23) iff
(λ−1
h , ψh ) is an eigenpair of Th .
To avoid unnecessary technicalities and to stay general, we make the following assumptions:
(i) There exists a dense subspace Vs ֒→ V such that the solution operators T and S∗ satisfy
T (v) ∈ Vs , S∗ (w) ∈ Vs , ∀v, w ∈ L2 (D). (48.25)
(ii) There is a sesquilinear form a♯ extending ah to V♯ ×V♯ , with V♯ := Vs + Vh , i.e., a♯ (vh , wh ) =

ah (vh , wh ) for all vh , wh ∈ Vh . The space V♯ is equipped with a norm k·kV♯ s.t. there is ka♯ k such
that
|a♯ (v, w)| ≤ ka♯ k kvkV♯ kwkV♯ , ∀v, w ∈ V♯ , ∀h ∈ H. (48.26)
(iii) The sesquilinear forms a♯ and a coincide on Vs ×Vs so that
a♯ (T (v), S∗ (w)) = a(T (v), S∗ (w)), ∀v, w ∈ L2 (D). (48.27)
(iv) Restricted Galerkin orthogonality and restricted adjoint Galerkin orthogonality, i.e., we have
the following identities:
a♯ (T (v), wh ) = ah (Th (v), wh ), ∀(v, wh ) ∈ L2 (D)×(Vh ∩ V ), (48.28a)

2
a♯ (vh , S∗ (w)) = ah (vh , S∗h (w)), ∀(vh , w) ∈ (Vh ∩ V )×L (D). (48.28b)
(Notice that discrete test functions are restricted to Vh ∩ V.)

(v) There is c such that for all h ∈ H,
kT (v) − Th (v)kV♯ ≤ c inf kT (v) − vh kV♯ , (48.29a)
vh ∈Vh ∩V
kS∗ (w) − S∗h (w)kV♯ ≤ c inf kS∗ (w) − wh kV♯ . (48.29b)

wh ∈Vh ∩V
Moreover, there is an integer k ≥ 1, and there is c such that the following best-approximation
property holds true for all t ∈ [0, k], all v ∈ H 1+t (D) ∩ V, and all h ∈ H:
inf kv − vh kV♯ ≤ c ℓD ht |v|H 1+t (D) . (48.30)
vh ∈Vh ∩V
The reader is invited to verify whether all the above conditions are satisfied, with Vs := V ∩
H 1+r (D) and r > 21 , by the Crouzeix–Raviart approximation from Chapter 36, Nitsche’s boundary
penalty technique from Chapter 37, and the Discontinuous Galerkin method from Chapter 38.

We are going to use the general approximation results for compact operators stated in Theo-
rems 48.1-48.3. Let t0 ≥ 0 be the smallest real number such that H 1+t0 (D) ∩ V ⊂ Vs . We assume
that t0 ≤ k, i.e., the interval [t0 , k] is nonempty. In the applications we have in mind, t0 is a
number close to 21 and k ≥ 1.
Lemma 48.11 (Bound on (T − Th )). There is c s.t. for all t, t∗ ∈ [t0 , k], all v ∈ L2 (D) s.t.
∗
T (v) ∈ H 1+t (D), all w ∈ L2 (D) s.t. S∗ (w) ∈ H 1+t (D), and all h ∈ H,

((T − Th )(v), w)L2 (D) ≤ c ht+t∗ ka♯ kℓ2 |T (v)|H 1+t (D) |S∗ (w)|H 1+t∗ (D) . (48.31)
D
∗
Proof. Let v ∈ L2 (D) be s.t. T (v) ∈ H 1+t (D), and let w ∈ L2 (D) be s.t. S∗ (w) ∈ H 1+t (D). We
∗
have T (v) ∈ H 1+t (D) ∩ V ⊂ Vs since t ≥ t0 , and S∗ (w) ∈ H 1+t (D) ∩ V ⊂ Vs since t∗ ≥ t0 . Using
the definitions of S∗ and S∗h , the assumption (48.27), i.e., that a♯ and a coincide on Vs ×Vs (and
that a♯ and ah coincide over Vh ×Vh ), and elementary manipulations, we infer that
((T − Th )(v), w)L2 (D) = a(T (v), S∗ (w)) − ah (Th (v), S∗h (w))
= a♯ (T (v), S∗ (w)) − a♯ (Th (v), S∗h (w))
= a♯ (T (v) − Th (v), S∗ (w)) + a♯ (Th (v), S∗ (w) − S∗h (w))
= a♯ (T (v) − Th (v), S∗ (w) − S∗h (w)) + a♯ (T (v) − Th (v), S∗h (w))
+ a♯ (Th (v), S∗ (w) − S∗h (w)) =: T1 + T2 + T3 .
Owing to the boundedness of a♯ on V♯ ×V♯ and the approximation properties (48.29)-(48.30), we
have ∗
|T1 | ≤ c ht+t ka♯ kℓ2D |T (v)|H 1+t (D) |S∗ (w)|H 1+t∗ (D) .
The other two terms have a similar structure that can be dealt with by invoking the restricted
Galerkin orthogonality (48.28). For instance, we have
|T2 | = inf |a♯ ((T − Th )(v), S∗h (w) − wh )|
wh ∈Vh ∩V
≤ ka♯ kk(T − Th )(v)kV♯ inf kS∗h (w) − wh kV♯

wh ∈Vh ∩V
≤ c ka♯ kk(T − Th )(v)kV♯ (kS∗h (w) − S∗ (w)kV♯ + inf kS∗ (w) − wh kV♯ )
vh ∈Vh ∩V
∗
≤ c′ ht+t ka♯ kℓ2D |T (v)|H 1+t (D) |S∗ (w)|H 1+t∗ (D) .
The term T3 is estimated similarly.
We assume now that the following elliptic regularity pickup holds true for T and S∗ (see §31.4.2):
There are real numbers τ ∈ (0, 1] and τ ∗ ∈ (0, 1] such that (48.13) holds true. The estimate (48.31)
with t := τ and t∗ := τ ∗ implies that
∗
kT − Th kL(L2 ,L2 ) ≤ c ka♯ kℓ2D kT kL(L2;H 1+τ ) kS∗ kL(L2 ;H 1+τ ∗ ) hτ +τ .
Since τ + τ ∗ > 0, this means that Th → T in operator norm as h → 0, that is, the key assump-
tion (48.1) holds true. It is then legitimate to use the approximation results for compact operators
stated in Theorems 48.1-48.3.
Let µ be a nonzero eigenvalue of T of ascent α and algebraic multiplicity m, and let Gµ , Gµ be
defined in (48.17). Proposition 48.6 implies that λ := µ−1 is an eigenvalue for (48.8). Let τµ and
τµ∗ be the two largest real numbers less than or equal to k satisfying (48.18). Recall that τµ ∈ [τ, k]
and τµ∗ ∈ [τ ∗ , k]. Moreover, we can set τµ = τµ∗ := k when maximal smoothness is available.
Owing to the norm convergence Th to T as h → 0, there are m eigenvalues of Th , say
{µh,j }j∈{1: m} (counted with their algebraic multiplicities), that converge to µ as h → 0. Let
Gh,µ be defined in (48.19). We are now in the position to state the main result of this section.
Theorem 48.12 (Convergence of eigenspace gap, eigenvalues, and eigenvectors). Let
µ ∈ σp (T )\{0} with algebraic multiplicity m and let {µh,j }j∈{1: m} be the eigenvalues of Th that
converge to µ. There is c, depending on µ, s.t. for all h ∈ H,
b µ , Gh,µ ) ≤ c hτµ +τµ∗ ,
δ(G (48.32)
1
P
and letting hµh i := m j∈{1: m} µh,j , we have
∗ 1 ∗
|µ − hµh i| ≤ c hτµ +τµ , |µ − µh,j | ≤ c h α (τµ +τµ ) , ∀j ∈ {1:m}. (48.33)
Moreover, for all integers j ∈ {1:m} and ℓ ∈ {1:α}, let wh,j be a unit vector in ker(µh,j IL2 − Th )ℓ .
There is c, depending on µ, such that for every integer ℓ′ ∈ {ℓ:α}, there is a unit vector uℓ′ ∈
′
ker(µIL2 − T )ℓ ⊂ Gµ such that for all h ∈ H,
ℓ′ −ℓ+1
(τµ +τ ∗ )
kuℓ′ − wh,j kL2 (D) ≤ c h α . (48.34)
In the above estimates, the constant c depends on ka♯ kℓ2D and on the operator norms defined
in (48.13) and (48.18).
Remark 48.13 (Literature). The nonconforming approximation of the elliptic eigenvalue prob-
lem has been studied in Antonietti et al. [12] for discontinuous Galerkin (dG) methods, Gopalakr-
ishnan et al. [220] for hybridizable discontinuous Galerkin (HDG) methods, and Calo et al. [103] for
hybrid high-order (HHO) methods. We refer the reader to Canuto [105], Mercier et al. [301], Durán
et al. [182], Boffi et al. [63] for mixed and hybrid mixed methods and to Carstensen and Gedicke
[109], Liu [287] for guaranteed eigenvalue lower bounds using Crouzeix–Raviart elements.
Exercises
Exercise 48.1 (Linearity). Consider the setting of §48.1.2. Let V ֒→ L be two complex Banach
spaces and a : V ×V → C be a bounded sesquilinear form satisfying the two conditions of the
BNB theorem. Let b : L×L → C be bounded sesquilinear form. (i) Let T : L → L be such that
a(T (v), w) := b(v, w) for all v ∈ L and all w ∈ V. Show that T is well defined and linear. (ii) Let
T∗ : L → L be such that a(v, T∗ (w)) := b(v, w) for all v ∈ V and all w ∈ L. Show that T∗ is well
defined and linear.
Exercise 48.2 (Invariant sets). (i) Let S, T ∈ L(V ) be such that ST = T S. Prove that ker(S)
and im(S) are invariant under T . (ii) Let T ∈ L(V ) and T let W1 , . . . , Wm be subspaces of V that
are invariant under T . Prove that W1 + . . . + Wm and i∈{1: m} Wi are invariant under T . (iii) Let
T ∈ L(V ) and let {v1 , . . . , vn } be a basis of V. Show that the following statements are equivalent:
(a) The matrix of T with respect to {v1 , . . . , vn } is upper triangular; (b) T (vj ) ∈ span{v1 , . . . , vj }
for all j ∈ {1:n}; (c) span{v1 , . . . , vj } is invariant under T for all j ∈ {1:n}. (iv) Let T ∈ L(V ). Let
µ be an eigenvalue of T . Prove that im(µIV − T ) is invariant under T . Prove that ker(µIV − T )α
is invariant under T for every integer α ≥ 1.
Exercise 48.3 (Trace). (i) Let V be a complex Banach space. Let G ⊂ V be a subspace of V
of dimension m. Let {φj }j∈{1: m} and {ψj }j∈{1: m} be two bases of G, and let {φ′j }j∈{1: m} and
{ψj′ }j∈{1: m} be corresponding dual bases, i.e., hφ′i , φj iV ′ ,V = δij , etc. (the way the antilinear forms
{φ′j }j∈{1: m} and {ψj′ }j∈{1: m} are extendedPto V does not matter). Let P T ∈ L(V )′ and assume
that G is invariant under T . Show that j∈{1: m} hψj′ , T (ψj )iV ′ ,V = j∈{1: m} hφj , T (φj )iV ′ ,V .
P
(ii) Let B ∈ Cm×m be s.t. T (φi ) =: j∈{1: m} Bji φj (recall that G is invariant under T ). Let
P
V := (hφ′j , viV ′ ,V )T α
j∈{1: m} for all v ∈ G. Prove that T (v) =
α
j∈{1: m} (B V)j φj for all α ∈ N.
(Hint : use an induction argument.) (iii) Let µ ∈ C, α ≥ 1, and S ∈ L(V ). Assume that
α
P := ker(µI′V − S) is finite-dimensional and nontrivial (i.e., dim(G) := m ≥ 1). Prove
G
′
that
j∈{1: m} jhφ , S(φ j )iV,V ′ = mµ. (Hint : consider the m×m matrix A with entries hφ i , (µIV −
S)(φj )iV ′ ,V and show that Aα = 0.)
Exercise 48.4 (Theorem 48.12). Prove the estimates in Theorem 48.12. (Hint : see the proof
of Theorem 48.8.)
Exercise 48.5 (Nonconforming approximation). Consider the Laplace operator with R homo-
geneous Dirichlet boundary conditions in a Lipschitz polyhedron D with b(v, w) := D ρvw dx,
where ρ ∈ C ∞ (D; R). Verify that the assumptions (48.25) to (48.30) hold true for the Crouzeix–
Raviart approximation.
Chapter 49
Well-posedness for PDEs in mixed

form
In Part XI, composed of Chapters 49 to 55, we study the well-posedness and the finite element
approximation of PDEs formulated in mixed form. Mixed formulations are often obtained from
elliptic PDEs by introducing one or more auxiliary variables. One reason for introducing these
variables can be that they have some physical relevance. For instance, one can think of the flux
in Darcy’s equations (see Chapter 51). Another reason can be to relax a constraint imposed on
a variational formulation. This is the case for the Stokes equations where the pressure results
from the incompressibility constraint enforced on the velocity field (see Chapter 53). The PDEs
considered in this part enjoy a coercivity property on the primal variable, but not on the auxiliary
variable, so that the analysis relies on inf-sup conditions. The goal of the present chapter is to
identify necessary and sufficient conditions for the well-posedness of a model problem that serves
as a prototype for PDEs in mixed form. The finite element approximation of this model problem
is investigated in Chapter 50.
49.1 Model problems

We introduce in this section some model problems illustrating the concept of PDEs in mixed form.
Let D be a domain in Rd , i.e., D is a nonempty, open, bounded, connected subset of Rd (see
Definition 3.1).
49.1.1 Darcy
Consider the elliptic PDE −∇·(d∇p) = f in D; see §31.1. Introducing the flux (or filtration veloc-
ity) σ := −d∇p leads to the following mixed formulation known as Darcy’s equations (see §24.1.2):
d−1 σ + ∇p = 0 in D, (49.1a)
∇·σ = f in D. (49.1b)
Here, (49.1a) is a phenomenological law relating the flux to the gradient of the primal unknown
p (a nonzero right-hand side can be considered as well). The equation (49.1b) expresses mass
288 Chapter 49. Well-posedness for PDEs in mixed form
conservation. For simplicity, we assume that (49.1) is equipped with the boundary condition
p|∂D = 0.
Let us give a more abstract form to the above problem by setting
V := L2 (D), Q := H01 (D). (49.2)
Consider the linear operators A : V → V ′ = L2 (D) (owing to the Riesz–Fréchet representation

theorem) and B : V → Q′ = H −1 (D) defined by setting A(τ ) := d−1 τ and B(τ ) := −∇·τ . Un-
der appropriate boundedness assumptions on d−1 , the linear operators A and B are bounded.
Using the identification (H01 (D))′′ = H01 (D), we have B ∗ : Q → V ′ and hB ∗ (q), τ iV ′ ,V =
hq, B(τ )iH01 (D),H −1 (D) = (∇q, τ )L2 (D) for all q ∈ H01 (D) and all τ ∈ L2 (D). Hence, B ∗ (q) = ∇q
for all q ∈ H01 (D).
An alternative point of view consists of setting
V := H(div; D), Q := L2 (D). (49.3)
Then A : V → V ′ is defined by setting A(τ ) := d−1 τ (where we use that V ֒→ L2 (D) ≡

′
L2 (D) ֒→ V ′ ) and B : V → Q′ = L2 (D) (owing to the Riesz–Fréchet representation theorem)
is defined by setting B(τ ) := −∇·τ . The adjoint operator B ∗ : Q → V ′ is s.t. hB ∗ (q), τ iV ′ ,V =
(q, B(τ ))L2 (D) = (q, −∇·τ )L2 (D) for all q ∈ L2 (D) and all τ ∈ H(div; D). Let us have a closer look
at B ∗ . Let q ∈ L2 (D) and assume that there exists g ∈ L2 (D) so that hB ∗ (q), τ iV ′ ,V = (g, τ )L2 (D) .
This implies that (q, −∇·τ )L2 (D) = (g, τ )L2 (D) for all τ ∈ H(div; D). Taking τ ∈ C0∞ (D) shows
that q has a weak derivative and ∇q = g. Hence, q ∈ H 1 (D) and (q, ∇·τ )L2 (D) + (∇q, τ )L2 (D) = 0
1
for all τ ∈ H(div; D). Moreover, considering the trace of q on ∂D, γ g (q) ∈ H 2 (∂D), and using the
1
surjectivity of the normal trace operator γ d : H(div; D) → H − 2 (∂D) (see Theorem 4.15), we infer
1
that for all φ ∈ H − 2 (∂D), there is τφ ∈ H(div; D) s.t. γ d (τφ ) = φ. Then hφ, γ g (q)i − 12 12 =
H ,H
1
hγ d (τφ ), γ g (q)i 1 1 = (q, ∇·τφ )L2 (D) + (∇q, τφ )L2 (D) = 0 for all φ ∈ H − 2 (∂D). Hence, γ g (q) =
H− 2 ,H 2
0, i.e., q|∂D = 0. This shows that B ∗ (q) ∈ L2 (D) encodes the boundary condition q|∂D = 0 in a
weak sense.
In conclusion, regardless of the chosen setting, the Darcy problem (49.1) can be reformulated
as follows: 

 Find σ ∈ V and p ∈ Q such that
A(σ) + B ∗ (p) = 0, (49.4)

 B(σ) = −f.
The mixed finite element approximation of (49.4) is studied in Chapter 51.
49.1.2 Stokes
The Stokes equations model steady incompressible flows in which inertia forces are negligible. The
problem is written in the following mixed form:
∇·(−µe(u)) + ∇p = f in D, (49.5a)
∇·u = 0 in D, (49.5b)
where µ > 0 is the viscosity, u : D → Rd the velocity field with the (linearized) strain rate tensor
e(u) := 21 (∇u + (∇u)T ) : D → Rd×d , p : D → R the pressure, and f : D → Rd the body force.
The equation (49.5a) expresses the momentum balance in the flow, and (49.5b) the mass balance.
For simplicity, we assume that (49.5) is equipped with the boundary condition u|∂D = 0.
Part XI. PDEs in mixed form 289
Let us set
V := H01 (D), Q := L2 (D), (49.6)
and let us define A : V → V ′ = H −1 (D), B : V → Q′ = L2 (D) (owing to the Riesz–Fréchet
representation theorem) by setting A(v) := −∇·(µe(v)) and B(v) := −∇·v. The adjoint operator
B ∗ : Q → V ′ is s.t. hB ∗ (q), viV ′ ,V = (q, −∇·v)L2 (D) for all v ∈ V and all q ∈ Q. This means that
B ∗ (q) = ∇q for all q ∈ L2 (D). In conclusion, the Stokes problem (49.5) can be reformulated as
follows: 

 Find v ∈ V and p ∈ Q such that
A(v) + B ∗ (p) = f , (49.7)

 B(v) = 0.
The finite element approximation of (49.7) is studied in Chapters 53 to 55.
49.1.3 Maxwell
Consider the model problem (43.4) for Maxwell’s equations in the time-harmonic regime (see §43.1.1)
in the limit ω → 0 and with the boundary condition H|∂D ×n = 0. Ampère’s equation (43.4a)
gives −E + σ1 ∇×H = σ1 js , and Faraday’s equation (43.4b) gives ∇·(µH) = 0 and ∇×E = 0
(since ω → 0). Setting κ := σ −1 the strong form of this problem consists of looking for a field
H : D → R3 such that ∇×(κ∇×H) = f , with f := ∇×(κjs ) : D → R3 , H|∂D ×n = 0, and
under the constraint ∇·(µH) = 0. The dual variable of this constraint is a scalar-valued function
φ : D → R with the boundary condition φ|∂D = 0, leading to the following mixed formulation (see,
e.g., Kanayama et al. [263]):
∇×(κ∇×H) + ν∇φ = f in D, (49.8a)

∇·(µH) = 0 in D. (49.8b)
Let us set
V := H0 (curl; D), Q := H01 (D), (49.9)
and let us define A : V → V ′ , B : V → Q′ = H −1 (D) by setting A(v) := ∇×(κ∇×v) and
B(v) := −∇·(νv). Using the identification (H01 (D))′′ = H01 (D), the adjoint operator B ∗ : Q → V ′
is s.t. hB ∗ (ψ), viV ′ ,V = hψ, −∇·(νv)iH01 (D),H −1 (D) = (ν∇ψ, v)L2 (D) for all v ∈ V and all ψ ∈ Q.
This means that B ∗ (ψ) = ν∇ψ for all ψ ∈ H01 (D). In conclusion, the Maxwell problem (49.8) can
be reformulated as follows:


 Find H ∈ V and φ ∈ Q such that
A(H) + B ∗ (φ) = f , (49.10)

 B(H) = 0.
Some further aspects of this problem are considered in Exercise 49.6.
49.2 Well-posedness in Hilbert spaces

Consider two real Hilbert spaces V and Q, and two operators A ∈ L(V ; V ′ ), B ∈ L(V ; Q′ ). We
identify Q′′ = Q. Our goal in this section is to investigate the well-posedness of the following
mixed model problem: 


 Find u ∈ V and p ∈ Q such that
A(u) + B ∗ (p) = f, (49.11)

 B(u) = g,
for all (f, g) ∈ V ′ ×Q′ . We assume in the entire section that A is self-adjoint and coercive. This
assumption simplifies many arguments. In particular, we establish the well-posedness of (49.11) by
means of a coercivity argument on the Schur complement. The complete theory for well-posedness
in Banach spaces is done in §49.4. Let α be the coercivity constant of A,
hA(v), viV ′ ,V
inf =: α > 0. (49.12)
v∈V kvk2V
We also assume that B is surjective, i.e., recalling Lemma C.40,
kB ∗ (q)kV ′
inf =: β > 0. (49.13)
q∈Q kqkQ
We denote kak := kAkL(V ;V ′ ) and kbk := kBkL(V ′ ;Q) .
49.2.1 Schur complement

Let JQ : Q → Q′ be the Riesz–Fréchet isometric isomorphism (see Theorem C.24), i.e.,
hJQ (q), riQ′ ,Q := (q, r)Q , ∀q, r ∈ Q.
We call Schur complement of A on Q the linear operator S : Q → Q defined by

−1
S := JQ BA−1 B ∗ . (49.14)
(S is sometimes defined with the opposite sign in the literature.)

Lemma 49.1 (Coercivity and boundedness of S). Let S be defined in (49.14). Then S is
symmetric and bijective with
β2 kbk2
kqk2Q ≤ (S(q), q)Q ≤ kqk2Q , ∀q ∈ Q. (49.15)
kak α
Proof. (1) Symmetry. Since A−1 is self-adjoint, we infer that for all q, r ∈ Q,
(S(q), r)Q = hBA−1 B ∗ (q), riQ′ ,Q = hA−1 B ∗ (q), B ∗ (r)iV,V ′

= hA−1 B ∗ (r), B ∗ (q)iV,V ′ = (S(r), q)Q .
(2) Bounds (49.15). The self-adjointness and coercivity of A imply that kA−1 kL(V ′ ;V ) = α−1 (see
1
Lemma C.51) and hA−1 (φ), φiV,V ′ ≥ kak kφk2V ′ for all φ ∈ V ′ (see Lemma C.63). Moreover, the
definitions of kbk and β mean that kB kL(Q;V ′ ) = kbk and kB ∗ (q)kV ′ ≥ βkqkQ for all q ∈ Q. Since
∗
(S(q), q)Q = hA−1 (B ∗ (q)), B ∗ (q)iV,V ′ for all q ∈ Q, we conclude that (49.15) holds true. Finally,
S is bijective since S is Q-coercive and bounded.
Lemma 49.2 (Equivalence with (49.11)). Let (u, p) ∈ V ×Q. Then the pair (u, p) solves (49.11)
iff (u, p) solves
−1
S(p) = JQ (BA−1 (f ) − g), A(u) = f − B ∗ (p). (49.16)
Proof. Let (u, p) ∈ V ×Q solve (49.11). Since A is bijective, we have u = A−1 (f − B ∗ (p)), so that
−1 −1
B(u) = BA−1 (f − B ∗ (p)) = g. This in turn implies that JQ BA−1 (f − B ∗ (p)) = JQ (g), finally
−1 −1 ∗
giving S(p) = JQ (BA (f ) − g) and A(u) = f − B (p). This means that (u, p) solves (49.16).
Conversely, assume that (u, p) ∈ V ×Q solves (49.16). Then BA−1 B ∗ (p) = BA−1 (f ) − g, that is,
BA−1 (f − B ∗ (p)) = g. But A−1 (f − B ∗ (p)) = u. Hence, B(u) = g and A(u) = f − B ∗ (p), which
means that (u, p) solves (49.11).
Theorem 49.3 (Well-posedness). The problem (49.11) is well-posed.
Proof. Owing to Lemma 49.2, it suffices to show that (49.16) is well-posed, but this is a consequence
of Lemma 49.1 and the Lax–Milgram lemma.
49.2.2 Formulation with bilinear forms

We now reformulate the mixed problem (49.11) using bilinear forms. This formalism will be used
in Chapters 50 to 55 where we consider various Galerkin approximations to (49.11). Let us set
a(v, w) := hA(v), wiV ′ ,V , b(w, q) := hq, B(w)iQ,Q′ , (49.17)
for all v, w ∈ V and all q ∈ Q (recall that we have identified Q′′ and Q). The assumed boundedness
of A and B implies that a and b are bounded on V ×V and on V ×Q, respectively. The abstract
problem (49.11) can then be reformulated in the following equivalent form:


a(u, w) + b(w, p) = f (w), ∀w ∈ V, (49.18)

 b(u, q) = g(q), ∀q ∈ Q,
with the shorthand notation f (w) := hf, wiV ′ ,V and g(q) := hg, qiQ′ ,Q . The definitions (49.12) and
(49.13) of α and β are then equivalent to
|a(v, v)| |b(v, q)|
inf =: α > 0, inf sup =: β > 0. (49.19)
v∈V kvk2V q∈Q v∈V kvkV kqkQ
Let X := V ×Q and consider the bilinear form t : X×X → R defined by
t((v, q), (w, r)) := a(v, w) + b(w, q) + b(v, r), (49.20)
for all (v, q), (w, r) ∈ X. Then (u, p) ∈ X solves (49.11) iff
t((u, p), (w, r)) = f (w) + g(r), ∀(w, r) ∈ X. (49.21)
49.2.3 Sharper a priori estimates

We collect in this section some additional results regarding the operator S, and we give a priori
estimates on the solution to the mixed problem (49.11). Recall from Definition 46.1 the notions of
spectrum and eigenvalues.
β 2 kbk2
Corollary 49.4 (Spectrum of S). The spectrum of S is such that σ(S) ⊂ kak , α , and
kbk2 kak
kSkL(Q;Q) ≤ α , kS −1 kL(Q;Q) ≤ β2 .
Proof. These statements are consequences of (49.15) and Theorem 46.17. Recall that Theo-
rem 46.17 asserts in particular that σ(S) ⊂ R, kSkL(Q;Q) = supλ∈σ(S) |λ| and kS −1 kL(Q;Q) =
supλ∈σ(S) |λ|−1 . See also Exercise 49.3.
Remark 49.5 (Spectrum of S). Corollary 49.4 can be refined by equipping V with the energy
1
norm k·k2a := a(·, ·) 2 (recall that a is symmetric and coercive) which is equivalent to k·kV . Setting
|b(v,q)k |b(v,q)|
βa := inf q∈Q supv∈V kvk a kqkQ
and kbka := supq∈Q supv∈V kvk a kqkQ
, we have kbk2a = kSkL(Q;Q) ,
βa−2 = kS −1 kL(Q;Q) , and {βa2 , kbk2a } ⊂ σ(S) ⊂ [βa2 , kbk2a ].
We define the linear operator T : X → X such that
−1
T (v, q) := (v + A−1 B ∗ (q), S −1 JQ B(v)), (49.22)
for all (v, q) ∈ V ×Q. With a slight abuse of notation regarding the column vector convention, we
IV A−1 B ∗
can also write T := S −1 J −1 B 0 , where IV is the identity in V. We have T ∈ L(X; X), and
Q
upon introducing the weighted inner product (x, y)Xe := a(v, w) + (S(q), r)Q for all x := (v, q) and
y := (w, r) in X, we also have (T (x), y)Xe = a(v, w) + b(v, r) + b(w, q), that is, (T (x), y)Xe = t(x, y).
This identity implies that T is symmetric with respect to the weighted inner product (x, y)Xe . The
following result, due to Bacuta [42], provides a complete characterization of the spectrum of T .
We refer the reader to §50.3.2 for the algebraic counterpart of this result.
√
Theorem 49.6 (Spectrum of T ). Let ̺ := 1+2 5 be the golden ratio. Assume that ker(B) is
nontrivial. Then
σ(T ) = σp (T ) = {−̺−1 , 1, ̺}. (49.23)
Proof. Let λ ∈ σ(T ). Owing to Corollary 46.18, there is a sequence (xn )n∈N in X such that
kxn kX = 1 for all n ∈ N and T (xn ) − λxn → 0 in X as n → ∞. Writing xn := (vn , qn ), we infer
−1
that (1 − λ)vn + A−1 B ∗ (qn ) → 0 in V and S −1 JQ B(vn ) − λqn → 0 in Q as n → ∞. Applying
−1
the bounded operator JQ B to the first limit and the bounded operator S to the second one, we
infer that −1
1−λ 1 JQ B(vn )
→ 0.
1 −λ S(qn )
Assume that λ 6∈ {−̺−1 , 1, ̺}. The matrix on the left-hand side is invertible since λ 6∈ {−̺−1 , ̺}.
This implies that S(qn ) → 0 in Q, so that kqn kQ → 0 since S is a bounded bijective operator.
Since λ 6= 1 and recalling that (1 − λ)vn + A−1 B ∗ (qn ) → 0 in V, we conclude that kvn kV → 0,
providing the expected contradiction with kxn kX = 1. Hence, σ(T ) ⊂ {−̺−1 , 1, ̺}. Finally, we
observe that λ = 1 is an eigenvalue associated with the eigenvectors (v, 0)T for all v ∈ ker(B) \ {0},
and ±̺±1 is an eigenvalue associated with the eigenvectors (±̺±1 A−1 B ∗ (q), q)T for all q ∈ Q \ {0}.
This proves (49.23).
Theorem 49.6 allows us to derive sharp stability estimates for the solution of (49.18) in the
1
weighted norm k(v, q)kXe := (a(v, v)+(S(q), q)Q ) 2 induced by the weighted inner product for which
T is symmetric. Equipping X with this weighted norm and since kT kL(X;X) = supλ∈σ(T ) |λ| and
kT −1kL(X;X) = supλ∈σ(T ) |λ|−1 (owing to Theorem 46.17), we infer from Theorem 49.6 that
kT kL(X;X) = kT −1 kL(X;X) = ̺. (49.24)
Recalling that t((v, q), (w, r)) = (T (v, q), (w, r))Xe , we infer that
|t(x, y)|
inf sup = kT −1 k−1
L(X;X) = ̺
−1
, (49.25a)
x∈X y∈Y kxkXe kykXe
|t(x, y)|
sup sup = kT kL(X;X) = ̺. (49.25b)
x∈X y∈Y kxkXe kykXe
Corollary 49.7 (Stability). Let (u, p) ∈ X solve (49.18). The following holds true:
12 21
1 α 1 kak
̺−1 kf k2V ′ + kgk2Q′ ≤ k(u, p)kXe ≤ ̺ kf k2V ′ + 2 kgk2Q′ .
kak kbk2 α β
−1
Proof. Let us set x := (u, p). Then (49.18) amounts to T (x) = y with y := (A−1 (f ), S −1 JQ (g)).
Hence, we have
kT k−1
L(X;X) kykX
e ≤ kxkX
e ≤ kT
−1
kL(X;X) kykXe ,
and we use (49.24) to infer that ̺−1 kykXe ≤ kxkXe ≤ ̺kykXe . Finally, the bounds in the proof of
1 α 1 kak
Lemma 49.1 imply that kyk2Xe ≥ kak kf k2V ′ + kbk 2 2 2 2
2 kgkQ′ and kyk e ≤ α kf kV ′ + β 2 kgkQ′ .
X
Proposition 49.8 (Stability). Let (u, p) ∈ X solve (49.18). The following holds true:

4 1 2 2
kbk2 1
kf k V ′ + kgk Q ′ ≤ a(u, u) + kpk2Q
((4 α + 1) + 1)
2 2 kak

4 1 2 2
≤ 2 1 kf k V ′ + kgk Q .
′ (49.26)
((4 β + 1) 2 − 1)2 α
kak
Proof. The proof is similar to that of Corollary 49.7 but uses the operator Te ∈ L(X; X) s.t.
Te(v, q) := (v + A−1 B ∗ (q), JQ
−1
B(v)) for all (v, q) ∈ V ×Q; see [42] and Exercise 49.4.
49.3 Saddle point problems in Hilbert spaces

In this section, we assume again that V and Q are real Hilbert spaces and the bilinear form a is
symmetric and coercive, i.e., A is self-adjoint and coercive. We show that the mixed problem (49.18)
has a saddle point structure.
49.3.1 Finite-dimensional constrained minimization

We start by introducing some simple ideas in the finite-dimensional setting of linear algebra. Let
N, M be two positive integers, let A be a symmetric positive definite matrix in RN ×N and let
F ∈ RN . Consider the functional E : RN → R such that E(V) := 12 (AV, V)ℓ2 (RN ) − (F, V)ℓ2 (RN ) .
Then E admits a unique global minimizer over RN , say U, which is characterized by the Euler
condition DE(U)(W) = (AU − F, W)ℓ2 (RN ) = 0 for all W ∈ RN , i.e., AU = F (see Proposition 25.8
and Remark 26.5 for similar results expressed in terms of bilinear forms).
Let now B ∈ RM×N , let G ∈ im(B) ⊂ RM , and consider the affine subspace K := {V ∈
N
R | BV = G}. Then E admits a unique global minimizer over K, say U, which is characterized by
the Euler condition DE(U)(W) = (AU−F, W)ℓ2 (RN ) = 0 for all W ∈ ker(B), i.e., AU−F ∈ ker(B)⊥ .
Since ker(B)⊥ = im(B T ), we infer that there is P ∈ RM such that AU + B T P = F. Recalling that
BU = G, the optimality condition is equivalent to solving the system

A BT U F
= , (49.27)
B O P G
where O is the zero matrix in RM×M . Moreover, P is unique if ker(B T ) = {0}, i.e., if B has full
row rank (note that this implies that N ≥ M ). This argument shows that the problem (49.27)
(which is similar to (49.11)) is actually the optimality condition characterizing the minimizer of
the functional V 7→ E(V) under the constraint V ∈ K.
Another way to look at the above problem consists of introducing the Lagrange multiplier
associated with the constraint V ∈ K, say Q, and considering the Lagrangian functional
L(V, Q) := E(V) + (Q, BV − G)ℓ2 (RM ) . (49.28)
Then the optimality conditions for a saddle point of L, say (U, P), are DV L(U, P)(W) = (AU −
F, W)ℓ2 (RN ) + (B T P, W)ℓ2 (RM ) = 0 for all W ∈ RN , and DQ L(U, P)(R) = (R, BU − G)ℓ2 (RM ) = 0 for
all R ∈ RM , which again gives (49.27).
49.3.2 Lagrangian
Let us now reformulate in a more general framework the computations we have done in the previous
section in finite dimension.
Definition 49.9 (Saddle point). Let V and Q be two sets and consider a map F : V ×Q → R.
A pair (u, p) ∈ V ×Q is said to be a saddle point of F if
∀q ∈ Q, F (u, q) ≤ F (u, p) ≤ F (v, p), ∀v ∈ V. (49.29)
Equivalently, we have supq∈Q F (u, q) = F (u, p) = inf v∈V F (v, p).
Lemma 49.10 (Inf-sup). The pair (u, p) ∈ V ×Q is a saddle point of F iff
inf sup F (v, q) = F (u, p) = sup inf F (v, q). (49.30)

v∈V q∈Q q∈Q v∈V
Proof. According to Definition 49.9, (u, p) ∈ V ×Q is a saddle point of F iff
inf sup F (v, q) ≤ sup F (u, q) = F (u, p) = inf F (v, p) ≤ sup inf F (v, q).
v∈V q∈Q q∈Q v∈V q∈Q v∈V
But independently of the existence of a saddle point, one can prove that
sup inf F (v, q) ≤ inf sup F (v, q). (49.31)

q∈Q v∈V v∈V q∈Q
Indeed, inf v∈V F (v, r) ≤ F (w, r) ≤ supq∈Q F (w, q) for all (w, r) ∈ V ×Q. The assertion (49.31) fol-
lows by taking the supremum over r ∈ Q on the left and the infimum over w ∈ V on the right. Thus,
the existence of a saddle point is equivalent to supq∈Q inf v∈V F (v, q) = inf v∈V supq∈Q F (v, q).
Proposition 49.11 (Lagrangian). Let V and Q be two real Hilbert spaces. Let a be a bounded,
symmetric, and coercive bilinear form on V ×V. Let b be a bounded bilinear form on V ×Q satisfy-
ing (49.19). Let f ∈ V ′ and g ∈ Q′ . The following three statements are equivalent: (i) u minimizes
the quadratic functional E(v) := 12 a(v, v) − f (v) on the affine subspace Vg := {v ∈ V | b(v, q) =
g(q), ∀q ∈ Q}. (ii) There is (a unique) p ∈ Q such that the pair (u, p) ∈ V ×Q is a saddle point of
the Lagrangian L s.t.
1
L(v, q) := a(v, v) + b(v, q) − f (v) − g(q). (49.32)
2
(iii) The pair (u, p) is the unique solution of (49.18).

49.4 Babuška–Brezzi theorem

Let V and M be two real Banach spaces. Consider two bounded linear operators A : V → V ′ and
B : V → M , and the model problem

 ′
 Find u ∈ V and p ∈ M such that
∗
A(u) + B (p) = f, (49.33)

 B(u) = g,
where B ∗ : M ′ → V ′ is the adjoint of B, f ∈ V ′ , and g ∈ M . The goal of this section is

to characterize the well-posedness of (49.33), reformulate it in terms of inf-sup conditions and
bilinear forms associated with the operators A and B, and relate this well-posedness result to the
BNB theorem (Theorem 25.9). The theory exposed here is due to Babuška and Brezzi [34, 90].
(In the Hilbert setting considered in §49.2-§49.3, the spaces M and Q are related by Q = M ′ .)
49.4.1 Setting with Banach operators

Let ker(B) be the null space of B and let JB be the canonical injection from ker(B) into V and
∗
JB : V ′ → ker(B)′ be its adjoint. Let Aπ : ker(B) → ker(B)′ be such that hAπ (v), wiV ′ ,V :=
∗
hA(v), wiV ′ ,V for all v, w ∈ ker(B), i.e., Aπ := JB AJB .
Theorem 49.12 (Well-posedness). Problem (49.33) is well-posed if and only if Aπ is an iso-

morphism and B is surjective.
Proof. (1) Assume first that (49.33) is well-posed.

(1.a) Let g ∈ M and let us denote by (u, p) the solution to (49.33) with data (0, g). Since u satisfies
B(u) = g, we infer that B is surjective.
(1.b) Let us show that Aπ is surjective. Let h ∈ ker(B)′ . Owing to the Hahn–Banach theorem, there
is an extension e h ∈ V ′ s.t. he
h, viV ′ ,V = hh, viV ′ ,V for all v in ker(B) and ke hkV ′ = khkker(B)′ . Let
e
(u, p) be the solution to (49.33) with f := h and g := 0. Then u ∈ ker(B). Since hB ∗ (p), viV ′ ,V =
hp, B(v)iM ′ ,M = 0 for all v ∈ ker(B), we infer that hAπ (u), viV ′ ,V = hA(u), viV ′ ,V = he h, viV ′ ,V =
hh, viV ′ ,V for all v ∈ ker(B). As a result, Aπ (u) = h.
(1.c) Let us show that Aπ is injective. Let u ∈ ker(B) be s.t. Aπ (u) = 0. Then hA(u), viV ′ ,V = 0
for all v ∈ ker(B), so that A(u) ∈ ker(B)⊥ . B being surjective, im(B) is closed and owing to
Banach’s theorem (Theorem C.35), im(B ∗ ) = ker(B)⊥ . As a result, A(u) ∈ im(B ∗ ), i.e., there is
p ∈ M ′ such that A(u) = −B ∗ (p). Hence, A(u) + B ∗ (p) = 0 and B(u) = 0, which shows that (u, p)
solves (49.33) with f := 0 and g := 0. Uniqueness of the solution to (49.33) implies that u = 0.
(2) Conversely, assume that Aπ is an isomorphism and B is surjective.
(2.a) For all f ∈ V ′ and all g ∈ M , let us show that there is at least one solution to (49.33).
The operator B being surjective, there is ug ∈ V s.t. B(ug ) = g. Denote by hf,g the bounded
linear form on ker(B) s.t. hhf,g , viV ′ ,V = hf, viV ′ ,V − hA(ug ), viV ′ ,V for all v ∈ ker(B). Since Aπ :
ker(B) → ker(B)′ is an isomorphism, Aπ is surjective, so that there is φ ∈ ker(B) s.t. Aπ (φ) = hf,g .
Set u := φ + ug . The linear form f − A(u) is in ker(B)⊥ . Since B is surjective, ker(B)⊥ = im(B ∗ ),
i.e., there is p ∈ M ′ such that B ∗ (p) = f − A(u). Moreover, B(u) = B(φ + ug ) = B(ug ) = g.
Hence, we have constructed a solution to (49.33).
(2.b) Let us show that the solution is unique. Let (u, p) be such that B(u) = 0 and A(u)+B ∗ (p) = 0,
so that u ∈ ker(B) and Aπ (u) = 0. Since πA is injective, u = 0. As a result, B ∗ (p) = 0. Since B
is surjective, B ∗ is injective, which implies p = 0.
49.4.2 Setting with bilinear forms and reflexive spaces

Let us now assume that V and M are reflexive Banach spaces and let us set Q := M ′ . Notice that
this implies that Q′ = M ′′ = M . Thus, we have B ∈ L(V ; Q′ ) and B ∗ ∈ L(Q; V ′ ). Consider the
two bounded bilinear forms a and b defined, respectively, on V ×V and on V ×Q s.t. a(v, w) :=
hA(v), wiV ′ ,V and b(v, q) := hB(v), qiQ′ ,Q . Let us set
|a(v, w)| |b(v, q)|

kak := sup sup , kbk := sup sup . (49.34)
v∈V w∈W kvkV kwkV v∈V q∈Q kvkV kqkQ
Let f ∈ V ′ and g ∈ Q′ . With the shorthand notation f (v) := hf, viV ′ ,V and g(q) := hg, qiQ′ ,Q , the
abstract problem (49.33) is reformulated as follows:


a(u, w) + b(w, p) = f (w), ∀w ∈ V, (49.35)

 b(u, q) = g(q), ∀q ∈ Q.
Theorem 49.13 (Babuška–Brezzi). (49.35) is well-posed if and only if



 inf |a(v, w)|
sup =: α > 0,
v∈ker(B) w∈ker(B) kvkV kwkV (49.36)


∀w ∈ ker(B), [ ∀v ∈ ker(B), a(v, w) = 0 ] =⇒ [ w = 0 ],
and the following inequality, usually called Babuška–Brezzi condition, holds:
|b(v, q)|
inf sup =: β > 0. (49.37)
q∈Q v∈V kvkV kqkQ
Furthermore, we have the following a priori estimates:
kukV ≤ c1 kf kV ′ + c2 kgkQ′ , (49.38a)

kpkQ ≤ c3 kf kV ′ + c4 kgkQ′ , (49.38b)
1 1 kak 1 kak kak kak

with c1 := α, c2 := β (1 + α ), c3 := β (1 + α ), and c4 := β 2 (1 + α ).
Proof. (1) Since ker(B) ⊂ V is reflexive, we infer form Theorem C.49 that (49.36) is equivalent
to Aπ being an isomorphism. Furthermore, (49.37) is equivalent to B being surjective owing to
(C.17) in Lemma C.40 since Q is reflexive. We invoke Theorem 49.12 to conclude that (49.35) is
well-posed iff (49.36)-(49.37) hold true.
(2) Let us now prove the a priori estimates (49.38). From the condition (49.37) and Lemma C.42
(since Q is reflexive), we deduce that there exists ug ∈ V such that B(ug ) = g and βkug kV ≤ kgkQ′ .
Setting φ := u − ug ∈ ker(B) yields a(φ, v) = f (v) − a(ug , v) for all v ∈ ker(B). Since

kak
|a(φ, v)| ≤ (kf kV ′ + kak kug kV )kvkV ≤ kf kV ′ + kgkQ′ kvkV ,
β
taking the supremum over v in ker(B) yields αkφkV ≤ kf kV ′ + kakβ kgkQ owing to (49.36). The
′
estimate on u results from this inequality and the triangle inequality. To prove the estimate
on p, we deduce from (49.37) and Lemma C.40 that βkpkQ ≤ kB ∗ (p)kV ′ , yielding βkpkQ ≤
kakkukV + kf kV ′ . The estimate on kpkQ then results from that on kukV .
Remark 49.14 (Coercivity). The conditions in (49.36) are automatically fulfilled if the bilinear
form a is coercive on ker(B) or coercive on X.
Let us recall that we have adopted the convention that suprema and infima in expressions like
(49.34)-(49.36)-(49.37) are taken over nonzero arguments. To relate the conditions (49.36) and
(49.37) with the conditions (bnb1) and (bnb2) from the BNB theorem (Theorem 25.9), let us
introduce the space X := V ×Q equipped with the norm k(v, q)kX := kvkV + kqkQ and let us recall
from (49.20) the bounded bilinear form t on X×X defined by
t((v, q), (w, r)) := a(v, w) + b(w, q) + b(v, r). (49.39)
Theorem 49.15 (Link with BNB). The bilinear form t satisfies the conditions (bnb1) and
(bnb2) if and only if (49.36) and (49.37) are satisfied.
Proof. (1) Let us prove that (49.36) and (49.37) imply (bnb1) and (bnb2). (1a) Proof of (bnb1).
Let (v, q) ∈ X and set S := sup(w,r)∈X |t((v,q),(w,r))|
k(w,r)kX . Lemma C.42 implies that there exists vb ∈ V
such that B(bv ) = B(v) and βkbv kV ≤ kB(v)kQ′ . We infer that
|b(v, r)| |t((v, q), (0, r))|

v kV ≤ kB(v)kQ′ = sup
βkb = sup ≤ S.
r∈Q krkQ (0,r)∈X k(0, r)kX
Observing that v − vb ∈ ker(B) we also infer that
|a(v − vb, w)|

αkv − vbkV ≤ sup
w∈ker(B) kwkV
|a(v − vb, w) + b(w, q) + b(v, 0)|
= sup
w∈ker(B) kwkV
|t((v, q), (w, 0))| |a(b
v , w)|
≤ sup + sup
(w,0)∈X k(w, 0)k X w∈ker(B) kwk V

kak
≤ S + kak kb v kV ≤ 1 + S.
β

Using the triangle inequality yields kvkV ≤ β1 + α1 1 + kak β S. Then we proceed as follows to
bound kqkQ :
|b(w, q)| |a(v, w) + b(w, q) + b(v, 0)| |a(v, w)|

βkqkQ ≤ sup ≤ sup + sup .
w∈V kwkV w∈V k(w, 0)kX w∈V kwkV
This estimate implies that βkqkQ ≤ S + kakkvkV , and we conclude that

1 1 1 kak
kqkQ ≤ 1 + kak + 1+ S.
β β α β
This proves (bnb1).

(1b) Let (w, r) ∈ X be s.t. t((v, q), (w, r)) = 0 for all (v, q) ∈ X, i.e.,
a(v, w) + b(v, r) = 0, ∀v ∈ V, (49.40a)

b(w, r) = 0, ∀r ∈ Q. (49.40b)
Then (49.40b) implies that w ∈ ker(B), and taking v ∈ ker(B) in (49.40a), we infer that a(v, w) = 0,
for all v ∈ ker(B). The second statement in (49.36) implies that w = 0. Finally, (49.40a) yields
b(v, r) = 0 for all v ∈ V, and (49.37) implies that r = 0. This proves (bnb2).
(2) Let us now prove that the conditions (bnb1) and (bnb2) on the bilinear form t imply the
conditions (49.36) and (49.37) on the bilinear forms a and b. Let γ denote the inf-sup constant of
the bilinear form t on X×X.
(2a) Let us start with (49.37). For all q ∈ Q, we have
|t((0, q), (w, r))| |b(w, q)|
γkqkQ = γk(0, q)kX ≤ sup = sup
(w,r)∈X k(w, r)kX (w,r)∈X k(w, r)kX
|b(w, q)| |b(w, q)|
= sup sup = sup ,
w∈V r∈Q k(w, r)kX w∈V kwkV
since the supremum over r ∈ Q is reached for r = 0. This proves (49.37) with β ≥ γ > 0.
′ ′
(2b) Let us prove the first statement in (49.36). For all w ∈ V , we define (ww , rw ) ∈ X to
′ ′
be the solution to the adjoint problem t((v, q), (ww , rw )) = a(v, w) for all (v, q) ∈ X. Owing to
′ ′
Lemma C.53, this problem is well-posed. Moreover, we have ww ∈ ker(B) and γkww kV ≤ kakkwkV .
′ ′ ′
Let v ∈ ker(B). We have a(v, w) = t((v, q), (ww , rw )) = a(v, ww ) for all q ∈ Q. We infer that
|t((v, 0), (w, r))| |a(v, w)|
γ kvkV = γ k(v, 0)kX ≤ sup = sup
(w,r)∈X k(w, r)kX (w,r)∈X k(w, r)kX
′ ′
|a(v, w)| |a(v, ww )| kak |a(v, ww )|
= sup = sup ≤ sup ′ k
.
w∈V kwkV w∈V kwk V γ w∈V kww V
′ γ2 |a(v,w)|
Since ww ∈ ker(B), this finally gives kak kvkV ≤ supw∈ker(B) kwkV , which is the first statement
γ2
in (49.36) with α ≥ kak > 0.
(2c) Let us now prove the second statement in (49.36). We first recall that we have already
established that (49.37) holds true. From Lemma C.40, we then infer that im(B ∗ ) is closed in
V ′ . Let w ∈ ker(B) be s.t. a(v, w) = 0 for all v ∈ ker(B). This implies that A∗ (w) ∈ ker(B)⊥ =
im(B ∗ ) = im(B ∗ ). Then there is rw ∈ Q s.t. B ∗ (rw ) = A∗ (w). For all (v, q) ∈ X, we then have
t((v, q), (w, −rw )) = a(v, w) + b(w, q) − b(v, rw ) = a(v, w) − b(v, rw )
= hA∗ (w) − B ∗ (rw ), viV ′ ,V = 0.
The condition (bnb2) on t implies that (w, −rw ) = 0, so that w = 0.
Remark 49.16 (Sharper estimate). Sharper estimates on the inf-sup stability constant of t
have been derived in Corollary 49.7 and Proposition 49.8 under the assumption that the bilinear
form a is symmetric and coercive.
′
Remark 49.17 (Direct sums). Notice that the map V ∋ w 7→ ww ∈ ker(B) introduced in step
(2b) of the proof of Theorem 49.15 implies that any w ∈ V can be uniquely decomposed into w =
′
ww +(A∗ )−1 B ∗ rw
′
. This means that we have the direct decomposition V = ker(B)⊕im((A∗ )−1 B ∗ ).
Note also that the same argument implies that V = ker(B) ⊕ im(A−1 B ∗ ).
Exercises
Exercise 49.1 (Algebraic setting). (i) Derive the counterpart of Theorem 49.12 in the setting
of §49.3.1. (Hint : assume that the matrix B has full row rank and consider a basis of ker(B).) (ii)
What happens if the matrix A is symmetric positive definite?

Exercise 49.2 (Constrained minimization). The goal is to prove Proposition 49.11. (i) Prove
that if u minimizes E over Vg , there is (a unique) p ∈ Q such that (u, p) solves (49.35). (Hint :
proceed as in §49.3.1.) (ii) Prove that (u, p) solves (49.35) if and only if (u, p) is a saddle point
of L. (Hint : consider Ep : V → R s.t. Ep (v) := L(v, p) with fixed p ∈ Q.) (iii) Prove that if
(u, p) is a saddle point of L, then u minimizes E over Vg . (iv) Application: minimize E(v) :=
2v12 + 2v22 − 6v1 + v2 over R2 under the constraint 2v1 + 3v2 = −1.
Exercise 49.3 (Symmetric operator). Let X be a Hilbert space and let T ∈ L(X; X) be a
bijective symmetric operator. (i) Prove that T −1 is symmetric. (ii) Prove that [ λ ∈ σ(T ) ] ⇐⇒
[ λ−1 ∈ σ(T −1 ) ]. (Hint : use Corollary 46.18.) (iii) Prove that σ(T ) ⊂ R. (Hint : consider the
sesquilinear form tλ (x, y) := ((T − λIX )(x), y)X and use the Lax–Milgram lemma.)
Exercise 49.4 (Sharp stability). The goal is to prove Proposition 49.8. (i) Assume that ker(B)
is nontrivial. Verify that 1 ∈ σp (Te). (ii) Let λ 6= 1 be in σ(Te). Prove that λ(λ − 1) ∈ σ(S). (Hint :
consider the sequence xn := (vn , qn ) in X from Corollary 46.18, then observe that (S(qn ), qn )Q =
(1 − λ)2 hA(vn ), vn iV ′ ,V + δn , with δn := hB ∗ (qn ) + (1 − λ)A(vn ), A−1 B ∗ (qn ) − (1 − λ)vn iV ′ ,V ,
and prove that S(qn ) − λ(λ − 1)qn → 0 and lim inf n→∞ kqn kQ > 0.) (iii) Prove that σ(Te) ⊂
β2 1 kbk2 1
[λ− − + + ± 1 ± 1
♯ , λ♭ ] ∪ {1} ∪ [λ♭ , λ♯ ] with λ♭ = 2 (1 ± (4 kak + 1) ), and λ♯ = 2 (1 ± (4 α + 1) ). (Hint : use
2 2
Lemma 49.1.) (iv) Conclude. (Hint : Te is symmetric with respect to the weighted inner product
(x, y)Xe := a(v, w) + (q, r)Q .)
Exercise 49.5 (Abstract Helmholtz decomposition). Consider the setting of §49.2 and
equip V with the bilinear form a as inner product. (i) Prove that im(A−1 B ∗ ) is closed and
that V = ker(B) ⊕ im(A−1 B ∗ ), the sum being a-orthogonal. (Hint : use Lemma C.39.) (ii)
Let f ∈ ker(B)⊥ . Prove that solving b(v, p) = f (v) for all v ∈ V is equivalent to solving
−1
(S(p), q)Q = (JQ BA−1 (f ), q)Q for all q ∈ Q.
Exercise 49.6 (Maxwell’s equations). Consider the following problem: For f ∈ L2 (D), find
A and φ such that

 ∇×(κ∇×A) + ν∇φ = f ,

∇·(νA) = 0,


A|∂Dd ×n = 0, φ|∂Dd = 0, (κ∇×A)|∂Dn ×n = 0, A|∂Dn ·n = 0,
where κ, ν are real and positive constants (for simplicity), and |∂Dd | > 0 (see §49.1.3; here we
write A in lieu of H and we consider mixed Dirichlet–Neumann conditions). (i) Give a mixed weak
formulation of this problem. (Hint : use the spaces Vd := {v ∈ H(curl; D) | γ c (v)|∂D d = 0}, where
the meaning of the boundary condition is specified in §43.2.1, and Qd := {q ∈ H 1 (D) | γ g (q)|∂Dd =
0}.) (ii) Let B : Vd → Q′d be s.t. hB(v), qiQ′d ,Qd := (νv, ∇q)L2 (D) . Let v ∈ ker(B). Show that
∇·v = 0 and, if v ∈ H 1 (D), γ g (v)|∂D n ·n = 0. (Hint : recall that ν is constant.) (iii) Accept as a
fact that D, ∂Dd , ∂Dn have topological and smoothness properties such that there exists c > 0
s.t. ℓD k∇×vkL2 (D) ≥ ckvkL2 (D) , for all v ∈ ker(B), with ℓD := diam(D). Show that the above
weak problem is well-posed. (Hint : use Theorem 49.13.) (iv) Let (Th )h∈H be a shape-regular
g
sequence of affine meshes. Let k ≥ 0, let Vh := Pkc (Th ) ∩ Vd , and let Qh := Pk+1 (Th ) ∩ Qd . Show
that ∇Qh ⊂ Vh . (v) Show that the discrete mixed problem is well-posed in Vh ×Qh assuming that
∂Dd = ∂D. (Hint : invoke Theorem 44.6.)
Chapter 50
Mixed finite element

approximation
This chapter is concerned with the approximation of the model problem analyzed in Chapter 49.
We focus on the Galerkin approximation in the conforming setting. We establish necessary and suf-
ficient conditions for well-posedness, and we derive error bounds in terms of the best-approximation
error. Then we consider the algebraic viewpoint, and we discuss augmented Lagrangian methods
in the context of saddle point problems. Finally, we examine iterative solvers, including Uzawa
iterations and Krylov subspace methods.
50.1 Conforming Galerkin approximation

Let V and Q be two reflexive (real) Banach spaces. Let a and b be two bounded bilinear forms
on V ×V and on V ×Q respectively. Let f ∈ V ′ and let g ∈ Q′ . We consider the following model
problem:


a(u, w) + b(w, p) = f (w), ∀w ∈ V, (50.1)

 b(u, q) = g(q), ∀q ∈ Q.
We introduce the associated operators A ∈ L(V ; V ′ ) and B ∈ L(V ; Q′ ) such that a(v, w) :=
hA(v), wiV ′ ,V and b(v, q) := hB(v), qiQ′ ,Q . We assume that (50.1) is well-posed. Owing to Theo-
rem 49.13, this means that the bilinear form a satisfies the conditions (49.36) (implying the inf-sup
|a(v,w)|
condition inf v∈ker(B) supw∈ker(B) kvk V kwkV
=: α > 0) and that the bilinear form b satisfies the
|b(v,q)|
inf-sup condition (49.37), i.e., inf q∈Q supv∈V kvkV kqkQ =: β > 0.
A conforming Galerkin approximation of (50.1) is obtained by considering two finite-dimensional
subspaces Vh ⊂ V, Qh ⊂ Q. The discrete problem is


 Find uh ∈ Vh and ph ∈ Qh such that
a(uh , wh ) + b(wh , ph ) = f (wh ), ∀wh ∈ Vh , (50.2)

 b(u , q ) = g(q ),
h h h ∀qh ∈ Qh .
302 Chapter 50. Mixed finite element approximation
Let Bh : Vh → Q′h be the discrete counterpart of the operator B : V → Q′ , that is,
hBh (vh ), qh iQ′h ,Qh := hB(vh ), qh iQ′ ,Q = b(vh , qh ), ∀(vh , qh ) ∈ Vh ×Qh .
The null space of Bh is such that
ker(Bh ) = {vh ∈ Vh | ∀qh ∈ Qh , b(vh , qh ) = 0}. (50.3)
One important aspect of the discretization is that the surjectivity of B does not imply that of
Bh . One rare occasion where this is nevertheless the case is when B ∗ (Qh ) ⊂ Vh , i.e., Bh∗ = B|Q
∗
h
.
This exceptional situation is illustrated in Exercise 49.6. Note also that in general, ker(Bh ) is not
necessarily a subspace of ker(B).
Proposition 50.1 (Well-posedness). (50.2) is well-posed if and only if
|a(vh , wh )|
inf sup := αh > 0, (50.4a)
vh ∈ker(Bh ) wh ∈ker(Bh ) kvh kV kwh kV
|b(vh , qh )|
inf sup := βh > 0. (50.4b)
qh ∈Qh vh ∈Vh kvh kV kqh kQ
Proof. Apply Theorem 49.13 and use the fact that (50.4a) implies both conditions in (49.36) since
Vh is finite-dimensional; see Remark 26.7.
The condition (50.4a) holds true for all conforming subspaces Vh and Qh if a is V -coercive
(the coercivity of a on ker(B) may not be sufficient since it may happen that ker(Bh ) 6⊂ ker(B)).
Note that verifying the inf-sup condition for a on Vh ×Vh is not sufficient to prove (50.4a) (think of
an invertible matrix having a square diagonal sub-block that is not invertible; see Exercise 50.1).
Furthermore, the condition (50.4b) is equivalent to Bh being surjective, which is also equivalent
to Bh∗ being injective since the setting is finite-dimensional. In practice, it is important that both
(50.4a) and (50.4b) hold true uniformly w.r.t. h ∈ H, i.e., inf h∈H αh =: α0 > 0 and inf h∈H βh =:
β0 > 0.

Our goal is to estimate the approximation errors (u − uh ) and (p − ph ) in terms of the best-
approximation error on u by a discrete field in Vh and the best-approximation error on p by
a discrete function in Qh . Céa’s lemma (Lemma 26.13) could be applied to the bilinear form
t((v, q), (w, r)) := a(v, w) + b(w, q) + b(v, r) introduced in §49.4.2 (see Exercise 50.2). But here, we
present a more specific analysis distinguishing the errors on u and on p. We say that Πh ∈ L(V ; Vh )
is a Fortin operator for the bilinear form b if b(Πh (v) − v, qh ) = 0 for all qh ∈ Qh and all v ∈ V. (We
do not assume Vh to be pointwise invariant under Πh .) This class of operators is investigated in
§26.2.3. In particular, Lemma 26.9 shows that the inf-sup condition (50.4b) implies the existence
of a Fortin operator with kΠh kL(V ;Vh ) ≤ kbk βh .
Lemma 50.2 (Error estimate). Let (u, p) solve (50.1). Assume (50.4) and let (uh , ph ) solve the
discrete problem (50.2). Let Πh ∈ L(V ; Vh ) be any Fortin operator. The following error estimates
hold true:
ku − uh kV ≤ c1h inf ku − vh kV + c2h inf kp − qh kQ , (50.5a)
vh ∈Πh (u)+ker(Bh ) qh ∈Qh
kp − ph kQ ≤ c3h inf ku − vh kV + c4h inf kp − qh kQ , (50.5b)

vh ∈Πh (u)+ker(Bh ) qh ∈Qh
kak kbk
with c1h := (1 + αh ), c2h
:= αh if ker(Bh )6⊂ ker(B) and c2h := 0 otherwise, c3h := c1h kak
βh , and
kbk
c4h := 1 + βh + c2h kak
βh .
Proof. (1) Estimate on (u−uh). Let vh ∈ Πh (u)+ker(Bh ), i.e., vh := Πh (u)+γh with γh ∈ ker(Bh ).
Then uh − vh ∈ ker(Bh ) since we have
b(uh − vh , qh ) = b(uh − Πh (u), qh ) + b(γh , qh ) = b(uh − u, qh ) = 0,
for all qh ∈ Qh , where we used the Galerkin orthogonality property for the second equation
in (50.2). Owing the inf-sup condition (50.4a), we infer that
|a(uh − vh , yh )|
αh kuh − vh kV ≤ sup
yh ∈ker(Bh ) kyh kV
|b(yh , p − ph ) + a(u − vh , yh )|
= sup ,
yh ∈ker(Bh ) kyh kV
where the equality follows from the Galerkin orthogonality property for the first equation in (50.2),
i.e., we have a(u−uh , yh )+b(yh , p−ph ) = 0 for all yh ∈ Vh . If ker(Bh ) ⊂ ker(B), then b(yh , p−ph ) =
0 for all yh ∈ ker(Bh ), yielding
αh kuh − vh kV ≤ kak ku − vh kV .
In the general case, we have b(yh , ph ) = 0 = b(yh , qh ) for all qh ∈ Qh , since yh is in ker(Bh ). This
implies that
αh kuh − vh kV ≤ kak ku − vh kV + kbk kp − qh kQ .
Hence, both cases are summarized by the following estimate:
kak
kuh − vh kV ≤ ku − vh kV + c2h kp − qh kQ ,
αh
with c2h as in the assertion. Using the triangle inequality and taking the infimum over vh ∈
Πh (u) + ker(Bh ) and over qh ∈ Qh leads to (50.5a).
(2) Estimate on (p − ph ). Using again the Galerkin orthogonality property for the first equation
in (50.2), we have
b(vh , qh − ph ) = a(uh − u, vh ) + b(vh , qh − p), ∀(vh , qh ) ∈ Vh ×Qh .
Combined with the inf-sup condition (50.4b), this implies that

|b(vh , qh − ph )|
βh kqh − ph kQ ≤ sup
vh ∈Vh kvh kV
≤ kak ku − uh kV + kbk kp − qh kQ .
The bound (50.5b) follows from the triangle inequality, the bound on (u − uh ), and by taking the
infimum over qh ∈ Qh .
The estimate on (u − uh ) involves the best-approximation error on u by a member of the
affine subspace Πh (u) + ker(Bh ). This error may not be easy to estimate in practice, and it is
sometimes preferable to bound it by the best-approximation error on u by a member of Vh since
Πh (u) + ker(Bh ) ⊂ Vh . Of course, the best-approximation error in Πh (u) + ker(Bh ) is larger
than the best-approximation error in Vh . The following lemma quantifies the discrepancy. (Recall
that (50.4b) is equivalent to the existence of Fortin operators.)
Lemma 50.3 (Best-approximation in Vh ). Assume (50.4b). The following holds true for all
u ∈ V and any Fortin operator Πh ∈ L(V ; Vh ):
inf ku − vh kV ≤ (1 + kΠh kL(V ;Vh ) ) inf ku − yh kV . (50.6)

vh ∈Πh (u)+ker(Bh ) yh ∈Vh
Proof. Let u ∈ V. Let yh ∈ Vh and set zh := Πh (u − yh ). Then yh + zh = Πh (u) + yh − Πh (yh ) ∈

Πh (u) + ker(Bh ) since b(yh − Πh (yh ), qh ) = 0 for all qh ∈ Qh . This implies that
inf ku − vh kV ≤ ku − (yh + zh )kV ≤ ku − yh kV + kzh kV

vh ∈Πh (u)+ker(Bh )
≤ (1 + kΠh kL(V ;Vh ) )ku − yh kV ,
and we conclude by taking the infimum over yh ∈ Vh .

Remark 50.4 (g = 0). If g = 0, then Πh (u) ∈ ker(Bh ), and the infimum in (50.5) and (50.6)
reduces to vh ∈ ker(Bh ).
Corollary 50.5 (Error estimate). Let (u, p) solve (50.1). Assume (50.4) and let (uh , ph ) solve
the discrete problem (50.2). The following error estimates hold true:
ku − uh kV ≤ c′1h inf ku − vh kV + c2h inf kp − qh kQ , (50.7a)

vh ∈Vh qh ∈Qh
kp − ph kQ ≤ c′3h inf ku − vh kV + c4h inf kp − qh kQ , (50.7b)

vh ∈Vh qh ∈Qh
with c′1h := (1 + kak ′

αh )(1 + kΠh kL(V ;Vh ) ) for every Fortin operator Πh ∈ L(V ; Vh ), c3h
:= c′1h kak
βh ,
and c2h , c4h are as in Lemma 50.2.
Proof. Combine Lemma 50.2 with Lemma 50.3.
Remark 50.6 (c′1h ). Lemma 26.9 asserts the existence of a Fortin operator with kΠh kL(V ;Vh ) ≤
kbk ′ kak kbk
βh . Hence, the upper bound c1h ≤ (1 + αh )(1 + βh ) always holds true. However, the estimate
kΠh kL(V ;Vh ) ≤ kbk
βh can be pessimistic. For instance, for the Stokes equations in elongated domains,
the boundedness constant of the bilinear form b(v, p) = (∇·v, p)L2 (D) on H01 (D)×L2 (D) is kbk =
1, and the inf-sup constant βh can be shown to be very small (see Chizhonkov and Olshanskii
[119], Dobrowolski [170]), whereas for some of these domains it is possible to construct a Fortin
operator with norm of order unity (see Mardal et al. [294], Linke et al. [284]).
Remark 50.7 (ker(Bh )). We refer the reader to Theorem 51.16 for an example of error estimate
exploiting the approximation properties in ker(Bh ) in the context of Darcy’s equations.
Remark 50.8 (c2h ). The constant c2h vanishes whenever ker(Bh ) ⊂ ker(B). Using a discrete pair
(Vh , Qh ) that guarantees that ker(Bh ) ⊂ ker(B) may be interesting when the best approximation
error on p is (much) larger than that on u. A simple example where this occurs is when f = B ∗ (p)
for some p ∈ Q and g = 0, so that the solution to (50.1) is (0, p). If ker(Bh ) ⊂ ker(B), the
estimate (50.7a) implies that uh = u = 0. But if ker(Bh ) 6⊂ ker(B), then uh is generally nonzero
and grows linearly with the size of p, which is not a desirable property. More generally, the well-
posedness of (50.1) with g := 0 implies the abstract Helmholtz decomposition V ′ = Y0 ⊕ Y1 with
Y0 := A(ker(B)) and Y1 = im(B ∗ ). Whenever the component of f in Y1 is much larger than
that in Y0 , the best-approximation error on p dominates the approximation error on u unless the
discretization satisfies ker(Bh ) ⊂ ker(B). See also Remark 53.22 for further insight in the context
of the Stokes equations.
Remark 50.9 (Stabilization). It is possible to approximate (50.1) using discrete spaces Vh and
Qh that violate the inf-sup condition (50.4b) by replacing the bilinear forms a and b by some
stabilized versions ah and bh ; see Chapters 62 and 63.
We now establish an error estimate on u in a norm that is weaker than that in V. We do so by
using a duality argument in the spirit of the Aubin–Nitsche lemma (Lemma 32.11).
Definition 50.10 (Smoothing property). The problem (50.1) is said to have a smoothing
property if there is a Hilbert space H ֒→ V with inner product (·, ·)H , two Banach spaces Y ֒→ V
and N ֒→ Q, and a constant csmo such that the following adjoint problem:


 Find ϕ(g) ∈ V and ϑ(g) ∈ Q such that
a(v, ϕ(g)) + b(v, ϑ(g)) = (g, v)H , ∀v ∈ V,

 b(ϕ(g), q) = 0, ∀q ∈ Q,
has a unique solution for all g ∈ H and satisfies the a priori estimate kϕ(g)kY + kϑ(g)kN ≤
csmo kgkH .
In addition to the smoothing property, we assume that the spaces H, Y, and N satisfy an
additional approximation property, i.e., there are s > 0 and c such that the following holds true
for all (v, q) ∈ Y ×N and all h ∈ H:
inf (kv − vh kV + kq − qh kQ ) ≤ c hs (kvkY + kqkN ). (50.8)

(vh ,qh )∈Vh ×Qh
Lemma 50.11 (Improved error estimate in weaker norm). Let (u, p) solve (50.1). As-
sume (50.4) and let (uh , ph ) solve the discrete problem (50.2). Assume that (50.1) has a smoothing
property and that (50.8) holds true. Then we have
ku − uh kH ≤ c hs (ku − uh kV + kp − ph kQ ),
where c is independent of (u, p), (uh , ph ) and h ∈ H.

Proof. Set V := V ×Q, Z := Y ×N , and L := H×Q, each equipped with the product norm. Define
the symmetric positive bilinear form l((v, q), (w, r)) := (v, w)H and the seminorm |(v, q)|L := kvkH .
Apply Lemma 32.11 in the conforming setting with the bilinear form t((u, p), (v, q)) := a(u, v) +
b(v, p) + b(u, q) to conclude.
50.2 Algebraic viewpoint

In this section, we study the linear system associated with the discrete problem (50.2) assuming
that the well-posedness conditions (50.4a)-(50.4b) are satisfied. We also assume that the bilinear
form a satisfies an inf-sup condition on Vh ×Vh . For simplicity, we consider real vector spaces.
50.2.1 The coupled linear system

Let N := dim(Vh ) and M := dim(Qh ). Let {ϕi }i∈{1: N } be a basis for Vh and let {ψk }k∈{1: M}
be a basis for Qh . Recall that these bases consist of global shape functions when Vh and Qh
are finite element spaces. Proceeding as in §28.1.1, for every column vectors U = (U1 , . . . , UN )T
in RN and P = (P1 , . . . , PM )T in RM , we define the functions Rϕ (U) ∈ Vh and Rψ (P) ∈ Qh by
P P
Rϕ (U) := i∈{1: N } Ui ϕi and Rψ (P) := k∈{1: M} Pk ψk . The correspondences between Rϕ (U) and
U and between Rψ (P) and P are one-to-one since {ϕi }i∈{1: N } and {ψk }k∈{1: M} are bases.
Inserting the expansions of Rϕ (U) and Rψ (P) into (50.2) and choosing the basis functions of Vh
and Qh to test (50.2), we obtain the linear system

U F A BT
C = , C := , (50.9)
P G B O
where the matrices A ∈ RN ×N and B ∈ RM×N are such that Aij := a(ϕj , ϕi ) and Bki := b(ϕi , ψk )
for all i ∈ {1:N } and all k ∈ {1:M }, O is the zero matrix in RM×M , and the vectors F ∈ RN and
G ∈ RM are such that Fi = f (ϕi ) and Gk = g(ψk ) for all i, j ∈ {1:N } and all k ∈ {1:M }.
The matrix C is invertible since (50.2) is well-posed owing to (50.4a)-(50.4b). Notice also that
(50.4b) implies that B T has full column rank and B has full row rank (these ranks are equal to
M ). Moreover, A is invertible since we additionally assumed that a satisfies an inf-sup condition
on Vh ×Vh . Algebraic counterparts of the boundedness and inf-sup conditions on the bilinear forms
a and b can be established by using the dual norm
UT Y
kUkℓ2ϕ := sup , ∀U ∈ RN . (50.10)
Y∈RN kRϕ (Y)kV
Proposition 50.12 (Norm equivalence). The following holds true:
αh kRϕ (U)kV ≤ kAUkℓ2ϕ ≤ kak kRϕ(U)kV , ∀U ∈ RN , (50.11a)

βh kRψ (P)kQ ≤ kB T Pkℓ2ϕ ≤ kbk kRψ (P)kQ , ∀P ∈ RM . (50.11b)
50.2.2 Schur complement

Since the matrix A is invertible, the vector U can be eliminated from the linear system (50.9)
yielding
SP = BA−1 F − G, S := BA−1 B T . (50.12)
Once P is known, U is obtained by solving AU = F − B T P. The matrix S ∈ RM×M (up to a sign
convention) is called Schur complement of A; see §49.2.1 for the infinite-dimensional counterpart.
Notice that the matrix S is invertible (if SP = 0, setting U := −A−1 B T P, we infer that C(U, P)T =
(0, 0)T , and C being invertible, this implies that U = 0 and P = 0).
Additional properties of the Schur complement matrix S are available when the bilinear form
a is symmetric and coercive, since in this case the matrix A is symmetric positive definite.
Proposition 50.13 (Symmetry and positivity of S). If A is symmetric positive definite, so
is S.
Proof. The definition of S implies that S T = B(A−1 )T B T , but (A−1 )T = (AT )−1 . Hence, S is
symmetric if A is symmetric. Let now P ∈ RM . Then PT SP = PT BA−1 B T P = (B T P)T A−1 B T P ≥
0. This proves that S is positive semidefinite. Moreover, SP = 0 implies that B T P = 0, so that
P = 0 since B T has full column rank. Hence, S is positive definite.
Note that even if A is symmetric positive definite, the matrix C is symmetric but indefinite.
Observing that

IN ON,M A ON,M IN A−1 B T
C= ,
BA−1 IM OM,N −S ON,M IM
we infer from the Sylvester Law of Inertia (stating that two symmetric matrices C and C ′ satisfying
C = PC ′ P T with P invertible have the same number of positive, zero, and negative eigenvalues;
see Golub and van Loan [218, p. 403]) that C has N positive eigenvalues and M negative ones.
Upper and lower bounds on the clusters of positive and negative eigenvalues of C are derived
in Rusten and Winther [338], Wathen and Silvester [390]. In practice, the matrix C can be very
poorly conditioned. We return to this issue in §50.3.2. Note that changing the lower-left block of
C into −B produces a positive semidefinite, but nonsymmetric, matrix.
Let us now examine more closely the eigenvalues of S (see Verfürth [375]). To this purpose,
let MQ ∈ RM×M be the matrix with entries MQ,kl := (ψk , ψl )Q for all k, l ∈ {1:M }. The matrix
MQ is symmetric by construction, and the identity PT MQ P = (Rψ (P), Rψ (P))Q = kRψ (P)k2Q for
all P ∈ RM shows that MQ is positive definite. Since Q is the L2 -space in many applications,
the matrix MQ is called mass matrix (see §28.1.1). Let µmin and µmax be the lowest and largest
eigenvalues of MQ . Recall from §28.2.1 that the (Euclidean) condition number κ(Z) of a symmetric
invertible matrix Z is the ratio of the largest to the smallest eigenvalues of Z in absolute value.
Proposition 50.14 (Spectrum of S). Assume that the bilinear form a is symmetric and coercive
on Vh with constant αh and that the inf-sup condition (50.4b) for b is satisfied with constant βh .
Then the matrices S and MQ are spectrally equivalent, i.e., the following holds true for all P ∈ RM :
βh2 PT SP kbk2
≤ T ≤ . (50.13)
kak P MQ P αh
β2 kbk2 βh2 2
Moreover, σ(MQ−1
S) ⊂ kak , αh , and σ(S)
h
⊂ µmin kak , µmax kbk −1
αh , which implies that κ(MQ S) ≤
2 2
kak kbk kak kbk
αh βh and κ(S) ≤ αh βh κ(MQ ).
Proof. (1) Proof of (50.13). For all P ∈ RM , we observe that

1 1
(B T P)T Y (B T P)T A− 2 Y (A− 2 B T P)T Y
sup 1 = sup = sup
Y∈RN (YT AY) 2 Y∈RN kYkℓ2 (RN ) Y∈RN kYkℓ2 (RN )
1 1
= kA− 2 B T Pkℓ2 (RN ) = (PT SP) 2 ,
1 kR (Y)k2 1
since A is symmetric positive definite. Observing that kak ≤ YϕT AY V ≤ αh for all Y ∈ RN , we
infer that
1 1
kB T Pk2ℓ2ϕ ≤ PT SP ≤ kB T Pk2ℓ2ϕ .
kak αh
Finally, (50.13) follows from (50.11b) using PT MQ P = kRψ (P)k2Q .
(2) The spectrum and condition number for M−1 Q S readily follow from (50.13), and the results for
S follow from the fact that µmin kPk2ℓ2 (RM ) ≤ PT MQ P ≤ µmax kPk2ℓ2 (RM ) for all P ∈ RM .
50.2.3 Augmented Lagrangian for saddle point problems

Assume that the matrix A is symmetric positive definite and that B T has full column rank. Re-
ferring to §49.3.2 for the infinite-dimensional setting we infer that the pair (U, P) solves the linear
system (50.9) iff it is a saddle point of the Lagrangian
1 T
L(Y, R) := Y AY − FT Y + RT (BY − G). (50.14)
2
Recall that
inf sup L(Y, R) = L(U, P) = sup inf L(Y, R). (50.15)
Y∈RN R∈RM R∈RM Y∈R
N
The optimization problem on the left-hand side of (50.15) amounts to minimizing the convex
energy functional E(Y) := 21 YT AY − FT Y over the affine subspace {Y ∈ RN | BY = G} since
supR∈RM L(Y, R) = ∞ if BY 6= G. Consider now the optimization problem on the right-hand
side of (50.15). The minimization of L(Y, R) over Y ∈ RN leads to the optimal solution Y∗ :=
A−1 (F − B T R), and we are left with maximizing the following concave functional over RM :
1 1
R 7→ L(Y∗ , R) = − RT SR + (BA−1 F − G)T R − FT A−1 F,
2 2
where S := BA−1 B T . The optimal solution to this maximization problem solves SR = BA−1 F − G,
i.e., we recover the Schur complement system (50.12).
The main idea of augmented Lagrangian methods (see Fortin and Glowinski [203]) is to add
to the Lagrangian a least-squares penalty on the constraint. Specifically, letting ρ > 0 be a real
parameter and recalling the mass matrix MQ ∈ RM×M , the augmented Lagrangian is defined as
ρ
Lρ (Y, R) := L(Y, R) + (BY − G)T M−1
Q (BY − G).
2
Since we also have BU = G, the solution to (50.9) is also the unique saddle point of the augmented
Lagrangian Lρ , i.e., (U, P) can be found by solving the following linear system:

Aρ B T U Fρ Aρ := A + ρB T M−1 Q B,
= , (50.16)
B O P G Fρ := F + ρB T M−1 Q G.
−1
The augmented Schur complement is defined as Sρ := BA−1 T
ρ B . Recall that σ(MQ S) ⊂ [s♭ , s♯ ],
2
βh kbk2
with s♭ := kak and s♯ := αh .
Proposition 50.15 (Spectrum of Sρ ). The following holds true:
Sρ−1 = ρM−1
Q +S
−1
, (50.17)
ρ+s−1
and σ(M−1 −1 −1
Q Sρ ) ⊂ [(ρ + s♭ ) , (ρ + s−1
♯ )
−1
] and κ(M−1
Q Sρ ) ≤
♭
ρ+s−1
.
♯
Proof. See Exercise 50.5 for the proof of (50.17). The properties on the spectrum and the condition
number of Sρ follow readily.
Remark 50.16 (Value of ρ). Proposition 50.15 shows that taking ρ ≫ 1 improves the condition
number of the Schur complement Sρ . A large value of ρ however deteriorates the conditioning of
the matrix Aρ which makes it more difficult to invert iteratively. In practice, it is necessary to
strike a balance between these two criteria.
Remark 50.17 (Hilbert setting). The notion of augmented Lagrangian can be extended to
the infinite-dimensional setting. The mass matrix MQ is then replaced by the Riesz–Fréchet
isomorphism JQ : Q → Q′ .
The augmented Lagrangian technique is in general preferable to the following unconstrained
penalty method:
A BT Uǫ F
= , (50.18)
B −ǫMQ Pǫ G
where ǫ > 0 is a small parameter. This technique is often referred to as artificial compressibility
in the fluid mechanics literature. Eliminating Pǫ from the first equation yields
A 1ǫ Uǫ = F + ǫ−1 B T M−1
Q G. (50.19)
The advantage of (50.19) with respect to (50.9) (or to (50.16)) is that the system matrix (i.e., A 1ǫ )
is symmetric positive definite. The solution (Uǫ , Pǫ ) however differs from (U, P). In particular, U
fails to satisfy the constraint BU = G, although the difference (U − Uǫ , P − Pǫ ) tends to zero as
ǫ → 0. Unfortunately, taking ǫ ≪ 1 makes the linear system (50.19) ill-conditioned.
Proposition 50.18 (Penalty). Let ǫ > 0. Let (U, P) solve (50.9) and (Uǫ , Pǫ ) solve (50.18). The
2
αh βh αh βh
kak kRϕ (U − Uǫ )kV + kak2 kRψ (P − Pǫ )kQ ≤ ǫkRψ (P)kQ . (50.20)
50.3 Iterative solvers

In this section, we discuss iterative solvers for the linear system (50.9) (or its augmented Lagrangian
version (50.16)). First, we discuss the Uzawa algorithm as an example of a technique based on
stationary iterations. Then, we present more efficient techniques based on preconditioned Krylov
subspaces. We assume that the matrix A is invertible and that the matrix B T has full column
rank.
50.3.1 Uzawa algorithm

The Uzawa algorithm is an iterative method where U and P are updated one after the other. Given
P0 ∈ RM and a parameter η > 0, the algorithm consists of constructing the iterates (Um , Pm ) for
m = 1, 2, . . . as follows:
AUm = F − B T Pm−1 , (50.21a)

MQ Pm = MQ Pm−1 + η(BUm − G). (50.21b)
This makes sense since A and MQ are invertible. Eliminating Um gives
MQ Pm = MQ Pm−1 − η(SPm−1 − BA−1 F + G). (50.22)
In other words, the Uzawa algorithm is equivalent to the Richardson iteration applied to the
linear system (50.12) left-preconditioned by the mass matrix MQ . (Recall that for a generic linear
system ZX = Y, the Richardson iteration reads Xm = Xm−1 + η(ZXm−1 − Y).) If A is symmetric
positive definite, we can use the bounds on the spectrum of M−1
Q S from Proposition 50.14, that
2 2
βh kbk
is, s♭ := kak ≤ M−1 Q S ≤ αh := s♯ in the sense of quadratic forms. We then infer that the
Richardson iteration (50.22) converges geometrically provided we take 0 < η < s2♯ , and the error
2
reduction factor is maximized by taking the optimal value ηopt := s♭ +s ♯
; see Saad [339, p. 106]. It
is often easier to estimate s♯ than s♭ since βh is more difficult to estimate than αh .
Remark 50.19 (Implementation). The matrices A and B are sparse (see §29.1), but S is a
dense matrix owing to the presence of A−1 in the definition of S. Since precomputing A−1 is
generally too expensive, an inner iteration has to be employed to compute the action of A−1 on
vectors in RN . The matrix B can be assembled and stored once and for all, or its action on a
given vector in RM can be computed on the fly whenever needed. In practice, one must often find
a compromise between many (often conflicting) criteria: the memory space available; the number
of times the linear system has to be solved; the ratio between the speed to access memory and the
speed to perform floating point operations; parallelization; etc.
Remark 50.20 (Variants). The mass matrix MQ can be replaced by the identity matrix IM
in (50.21b). The advantage is that this avoids computing the inverse of the mass matrix (although
this matrix is generally easy to invert since it is well-conditioned). The drawback is that the
choice of the relaxation parameter η now depends on the spectrum of the unpreconditioned Schur
complement matrix, which requires some information on the (mesh-dependent) spectrum of MQ .
Another variant is to consider an approximate inverse of A that is easy to compute, say H, and to
replace (50.21a) by Um = Um−1 +H(F−AUm−1 −B TPm−1 ) leading to an inexact Uzawa algorithm;
see Bacuta [41] for a convergence analysis.
50.3.2 Krylov subspace methods

Krylov subspace methods for solving (preconditioned) linear systems of the form (50.9) or variations
thereof constitute an active area of research. In this section, we sketch a few important ideas and
refer to Benzi et al. [52, §9] for a broader treatment and to Elman et al. [185, Chap. 6&8], Turek
[366] for applications to fluid mechanics.
In the context of saddle point problems, the matrix C in (50.9) is symmetric, but indefinite
(recall that the matrix A is symmetric positive definite by assumption). In this case, Minres is
a method of choice to solve (50.9); see [185, p. 289]. The attractive feature of Minres is that
it achieves an optimality property on the residual while employing only short-term recurrences.
Specifically, at step m ≥ 1, the iterate Xm ∈ RN +M with residual Rm := (F, G)T − CXm satisfies the
following optimality property (compare with Proposition 28.20 for the conjugate gradient method
applied to symmetric positive definite linear systems):
kRm kℓ2 (RN +M ) = min k(F, G)T − CYkℓ2 (RN +M ) , (50.23)

Y∈U0 +Km
with the Krylov subspace Km := span{R0 , CR0 , . . . , C m−1 R0 }. The convergence rate of Minres
depends on the spectrum of C. More precisely, defining c̃m := minp∈Pm ,p(0)=1 maxλ∈σ(C) |p(λ)|, one
can prove that kRm kℓ2 (RN +M ) ≤ c̃m kR0 kℓ2 (RN +M ) (this bound is sharp). The constant c̃m can be
estimated under the assumption that σ(C) ⊂ [−a, −b] ∪ [c, d] with positive real numbers a, b, c, d
such that the two intervals have the same length (i.e., d − c = a − b). One can show that (see
Greenbaum [221, Chap. 3])
√ √ !m
ad − bc
kR2m kℓ2 (RN +M ) ≤ 2 √ √ kR0 kℓ2 (RN +M ) . (50.24)
ad + bc
The minimization property of Minres implies that kR2m+1 kℓ2 (RN +M ) ≤ kR2m kℓ2 (RN +M ) , but it is
possible that no reduction of the norm of the residual occurs in every other step, leading to a stair-
casing behavior of the iterates. A comparison of (28.23) with (50.24) shows that Minres requires
twice as many iterations as the Conjugate Gradient to reach a given threshold for a symmetric
positive definite matrix with condition number κ2 . Hence, solving linear systems like (50.9) is a
significant computational challenge, and preconditioning is essential. Before addressing this ques-
tion let us observe that Minres is bound to fail if A is not symmetric, since symmetry is essential
for Minres to work properly. This happens, for instance, in fluid mechanics when solving the
Oseen or (linearized) Navier–Stokes equations. One alternative is to use the Gmres method which
retains an optimality property over the Krylov subspace at the price of storing a complete basis
thereof; see Saad [339, §6.5] for a thorough description.
Preconditioning is a very important ingredient of Krylov subspace methods, especially for linear
systems of the form (50.9). Here, we only discuss block preconditioners and refer the reader to [52,
§10] and references therein for further insight into preconditioned Krylov methods. Block diagonal
and triangular preconditioners are, respectively, of the form

Â O Â B T
Pd := , Pt := , (50.25)
O Ŝ O Ŝ
where Â and Ŝ are easy-to-invert approximations of A and S. In the ideal case where Â := A
and Ŝ := S, a direct calculation shows that the left-preconditioned
√
matrices Pd−1 C and Pt−1 C
1± 5
are zeroes of the polynomials pd (λ) := (λ − 1)(λ − 2 ) and pt (λ) := λ2 − 1, respectively (see
Kuznetsov [272], Murphy et al. [309]; see also (49.22) and Theorem 49.6), implying convergence
in at most three (resp., two) steps for every preconditioned Krylov subspace method. The block
triangular preconditioner Pt breaks the symmetry of the system even if A is symmetric, but this
preconditioner is still quite effective in many cases, particularly for Oseen and Navier–Stokes flows
(where the convective term breaks the symmetry of A anyway). Note also that the costs of the two
preconditioners in (50.25) are essentially identical since the cost of the additional multiplication
by B T is often marginal.
Effective choices for Â and Ŝ are often driven by the application at hand. For Darcy’s and
Maxwell’s equations (see Examples in §49.1.1 and §49.1.3), A represents a zeroth-order differential
operator (multiplication by a material property), and choosing a diagonal lumping for Â together
with some multilevel technique for Ŝ often works well if the material coefficients are smooth (see,
e.g., Perugia and Simoncini [324] for magnetostatics problems). For the Stokes equations (see
§49.1.2), A represents a second-order differential operator, and the preconditioner Â is typically
based on some multilevel technique. The mass matrix associated with p can be used for Ŝ and
a detailed eigenvalue analysis of the resulting block-diagonal preconditioned system can be found
in Silvester and Wathen [347]. The approximation of the Schur complement becomes more delicate
in the unsteady case and in the presence of convection. Preconditioners devised from the structure
of the steady Navier–Stokes equations can be found in Elman et al. [185, Chap. 8] and the refer-
ences therein. Furthermore, an attractive idea for transient and high-Reynolds number flows is to
consider a block triangular preconditioner based on the augmented Lagrangian formulation (50.16)
for the (1, 1)-block, together with the (scaled) mass matrix for the (2, 2)-block (thereby avoiding
to consider the Schur complement); see Benzi and Olshanskii [51], Benzi et al. [53].
Exercises
√
Exercise 50.1 (Algebraic setting). Let A := √1 2
and B := (1, 0)T . Show that
2 0
WT AV WT AV
inf sup < inf 2 sup .
V∈ker(B) W∈ker(B) kWkℓ2 (R2 ) kVkℓ2 (R2 ) V∈R W∈R2 kWkℓ2 (R2 ) kVkℓ2 (R2 )
(Hint : one number is equal to 0 and the other is equal to 1.)

Exercise 50.2 (Saddle point problem). Let V, Q be Hilbert spaces and let a be a symmetric,
coercive, bilinear form. Consider the discrete problem (50.2) and the bilinear form t(y, z) :=
a(v, w) + b(w, q) + b(v, r) for all y := (v, q), z := (w, r) ∈ X := V ×Q. Let Xh := Vh ×Qh and
consider the linear map Ph ∈ L(X; Xh ) such that for all x ∈ X, Ph (x) ∈ Xh is the unique
solution of t(Ph (x), yh ) = t(x, yh ) for all yh ∈ Xh . Equip X and Xh with the norm k(v, q)kXe :=
2 1
1 (4 kbk
α +1) +1
2
(kvk2a + kqk2Q ) 2 with kvk2a := a(v, v). (i) Prove that kPh kL(X;X) ≤ c̃h := β2 1
. (Hint : use
h +1) 2
(4 kak −1
Proposition 49.8.) (ii) Prove that ku − uh k2a + kp − ph k2Q ≤ c̃2h (inf vh ∈Vh ku − uh k2a + inf qh ∈Qh kp −
qh k2Q ). (Hint : see the proof of Theorem 5.14.)
Exercise 50.3 (Error estimate). (i) Prove directly the estimate (50.7a) with c′1h replaced by
c′′1h := (1 + kak kbk
αh )(1 + βh ). (Hint : consider zh ∈ Vh s.t. Bh (zh )
:= Bh (uh − vh ) with vh ∈ Vh
arbitrary.) (ii) Assume that V is a Hilbert space, ker(Bh ) ⊂ ker(B), and g := 0. Prove that
ku − uh kV ≤ kak
αh inf vh ∈ker(Bh ) ku − vh kV .
Exercise 50.4 (Bound on A and B). (i) Prove Proposition 50.12. (Hint : observe that (AU)T Y =
a(Rϕ (U), Rϕ (Y)).) (ii) Let JV ∈ RN ×N be the symmetric positive definite matrix with entries
JV,ij := (ϕi , ϕj )X for all i, j ∈ {1:N }. Let k·kℓ2 (RN ) denote the Euclidean norm in RN . Verify
1
−1
that kRϕ (U)kV = kJV2 Ukℓ2 (RN ) and kUkℓ2 (RN ) = kJV 2 Ukℓ2 (RN ) for all U ∈ RN .
Exercise 50.5 (Sρ ). The goal is to prove the identity (50.17). (i) Verify that A−1 ρ = A−1 −
−1 T −1 −1
ρA B (MQ + ρS) BA . (Hint : multiply the right-hand side by Aρ and develop the product.)
(ii) Infer that Sρ = S − ρS(MQ + ρS)−1 S. (iii) Conclude. (Hint : multiply the right-hand side by
ρM−1 Q +S
−1
.)
Exercise 50.6 (Penalty). (i) Prove Proposition 50.18. (Hint : verify that C(U − Uǫ , P − Pǫ )T =
(0, −ǫMQ Pǫ )T and use Proposition 50.12.) (ii) Replace the mass matrix MQ by the identity
matrix IM times a positive coefficient λ in (50.18). Does the method still converge? Is there any
interest of doing so? Can you think of another choice?
Exercise 50.7 (Inexact Minres and DPG). Let V, Y be Hilbert spaces and B ∈ L(V ; Y ′ ) be s.t.
βkvkV ≤ kB(v)kY ′ ≤ kbkkvkV for all v ∈ V with 0 < β ≤ kbk < ∞. Set b(v, y) := hB(v), yiY ′ ,Y .
Let f ∈ Y ′ . Let JY : Y → Y ′ denote the isometric Riesz–Fréchet isomorphism. (i) Show that
the Minres problem minv∈V kf − B(v)kY ′ has a unique solution u ∈ V. (Hint : introduce the
sesquilinear form a(v, w) := hB(v), JY−1 (B(w))iY ′ ,Y and invoke the Lax–Milgram Lemma.) (ii) Let
{Vh ⊂ V }h∈H and {Yh ⊂ Y }h∈H be sequences of subspaces approximating V and Y, respectively.
Assume that there is β0 > 0 s.t. for all h ∈ H,
|b(vh , yh )|
inf sup ≥ β0 . (50.26)
vh ∈Vh yh ∈Yh kvh kV kyh kY
Let Ih : Yh → Y be the canonical injection and Ih∗ : Y ′ → Yh′ . Show that the inexact Minres
problem minvh ∈Vh kIh∗ (f − B(vh ))kYh′ has a unique solution uh ∈ Vh . (Hint : introduce the residual
representative rh := JY−1 I ∗ (f − B(uh )) ∈ Vh and show that the pair (uh , rh ) ∈ Vh × Yh solves a
h h
saddle point problem.) (iii) Show that the residual representative rh ∈ Yh is the unique solution of
the following constrained minimization problem: minzh ∈Yh ∩(Ih∗ (B(Vh )))⊥ 21 kzh k2Y − hIh∗ (f ), zh iYh′ ,Yh .
(Hint : see Proposition 49.11.) (iv) Assume now that f ∈ im(B) so that B(u) = f . Prove that
there is c s.t. ku − uhkV ≤ c inf wh ∈Vh ku − wh kV for all h ∈ H. (Hint : use a Fortin operator.) Note:
since βkvh kV ≤ kB(vh )kY ′ for all vh ∈ Vh , it is natural to expect that the inf-sup condition (50.26)
is satisfied if the subspace Yh ⊂ Y is chosen rich enough. The inexact residual minimization in
a discrete dual norm is at the heart of the discontinuous Petrov–Galerkin (dPG) method; see
Demkowicz and Gopalakrishnan [158], Gopalakrishnan and Qiu [219], Carstensen et al. [111]. The
extension to reflexive Banach spaces is studied in Muga and van der Zee [308].
Chapter 51
Darcy’s equations
Darcy’s equations consist of the following PDEs in the domain D ⊂ Rd :
d−1 σ + ∇p = f in D, (51.1a)
∇·σ = g in D. (51.1b)
The unknowns are the primal variable p and the dual variable σ. In the literature, p is also called
potential and σ flux. The PDEs (51.1) are used to model porous media flows, e.g., fluid flows in
aquifers and petroleum reservoirs. In this context, σ is the seepage velocity, p the pressure, and d
the material permeability, the equation (51.1a) is called Darcy’s law, and (51.1b) expresses mass
conservation. Eliminating the dual variable σ leads to −∇·(d∇p) = g − ∇·(df ) in D, which is
a PDE where the only unknown is the primal variable p. This PDE can be approximated using,
e.g., H 1 -conforming finite elements as in Chapter 32. The approach we follow here is conceptually
different since our aim is to approximate simultaneously the primal and the dual variables. In
this chapter, we derive well-posed weak formulations for (51.1) with various boundary conditions.
Then we study mixed finite element approximations using H(div)-conforming spaces for the dual
variable.
51.1 Weak mixed formulation

The data in (51.1) are d, f , and g, where d is a second-order tensor if the material is anisotropic and
it may depend on x if the material is non-homogeneous. We assume that f ∈ L2 (D), g ∈ L2 (D) and
that d is symmetric and the eigenvalues of d are bounded from below and from above, respectively,
by λ♭ and λ♯ uniformly in D. We assume that λ♭ > 0. We will consider Dirichlet, Neumann,
and mixed Dirichlet–Neumann conditions for (51.1), and we will see that contrary to the primal
formulation studied in Chapter 31, Dirichlet conditions on p are enforced weakly, whereas Neumann
conditions on σ are enforced strongly.
51.1.1 Dirichlet boundary condition

In this section, we consider the Dirichlet condition γ g (p) = ad on ∂D. Let us first proceed
informally by assuming that all the functions are smooth enough. Multiplying (51.1a) by a smooth
vector-valued test function τ , integrating over D, integrating by parts the term with ∇p, and using
316 Chapter 51. Darcy’s equations
the Dirichlet boundary condition, we infer that

Z Z Z
((d−1 σ)·τ − p∇·τ ) dx = f ·τ dx − ad τ ·n ds. (51.2)
D D ∂D
Furthermore, multiplying (51.1b) by a smooth scalar-valued test function q and integrating over
D gives Z Z
(∇·σ)q dx = gq dx. (51.3)
D D
Since f ∈ L2 (D) and g ∈ L2 (D), the volume integrals make sense if we assume that σ, τ ∈
H(div; D) := {ς ∈ L2 (D) | ∇·ς ∈ L2 (D)} and we assume that p, q ∈ L2 (D). To be dimensionally
1
coherent, we equip H(div; D) with the norm kςkH(div;D) := (kςk2L2 (D) + ℓ2D k∇·ςk2L2 (D) ) 2 , where
ℓD is a length scale associated with D, e.g., ℓD := diam(D).
A rigorous meaning can be given to the boundary integral in (51.2) once we recall that any
vector field in H(div; D) has a normal component over ∂D that can be defined by the normal trace
1
map γ d : H(div; D) → H − 2 (∂D) such that the following integration by parts formula holds true
(see (4.12)): Z
hγ d (τ ), γ g (q)i∂D := (q∇·τ + τ ·∇q) dx, (51.4)
D
for all q ∈ H 1 (D) and all τ ∈ H(div; D), where h·, ·i∂D denotes the duality pairing between
1 1 1
H − 2 (∂D) and H 2 (∂D). More precisely, assuming that ad ∈ H 2 (∂D), the boundary R integral
in (51.2) is understood as hγ d (τR), ad i∂D . We can now define the linear forms Fd (τ ) := D f ·τ dx −
hγ d (τ ), ad i∂D and Gd (q) := − D gq dx and the bilinear forms
Z Z
a(ς, τ ) := (d−1 ς)·τ dx, b(ς, q) := − (∇·ς)q dx, (51.5)
D D
and we consider the following problem:


 2
 Find σ ∈ V := H(div; D) and p ∈ Q := L (D) such that
a(σ, τ ) + b(τ , p) = Fd (τ ), ∀τ ∈ V , (51.6)

 b(σ, q) = G (q),
d ∀q ∈ Q.
The above assumptions imply that Fd ∈ V ′ , Gd ∈ Q′ ≡ Q, a is bounded on V ×V , and b is

bounded on V ×Q. The negative sign in the definition of b and Gd is not essential. This choice
leads to a symmetric weak problem.
1
Proposition 51.1 (Well-posedness). Assume f ∈ L2 (D), g ∈ L2 (D), and ad ∈ H 2 (∂D). (i)
The problem (51.6) is well-posed. (ii) The pair (σ, p) satisfies the PDEs (51.1) a.e. in D, p is
in H 1 (D), and p satisfies the boundary condition γ g (p) = ad a.e. on ∂D, where γ g : H 1 (D) →
1
H 2 (∂D) is the trace map defined in Theorem 3.10.
Proof. (i) We apply Theorem 49.13. Set B := ∇· : V → Q so that ker(B) = {v ∈ V | ∇·v = 0}.
We have already seen that the bilinear forms a and b are continuous. Moreover, the inequality
a(ς, ς) ≥ λ−1 2
♯ kςkL2 (D) implies that the bilinear form a is coercive on ker(B) since kςkL2 (D) =
kςkH(div;D) if ς ∈ ker(B). Hence, the two conditions (49.36) on a hold true. Moreover, the inf-sup
condition (49.37) on b follows from Lemma 51.2 below. Hence, all the required assumptions for
well-posedness are met.
(ii) Testing (51.6) with an arbitrary function τ ∈ C0∞ (D) := C0∞ (D; Rd ) and with q := 0, we
R R
infer that D p∇·τ dx = D (d−1 σ − f )·τ dx. This proves that p has a weak derivative in L2 (D)
and ∇p = f − d−1 σ. Hence, p ∈ H 1 (D). Since p ∈ H 1 (D), we invoke the integration by parts
formula (51.4) and infer that
Z Z

hγ d (τ ), γ g (p)i∂D = (p∇·τ + τ ·∇p) dx = p∇·τ − τ ·(d−1 σ − f ) dx
D D
= −a(σ, τ ) − b(p, τ ) + Fd (τ ) + hγ d (τ ), ad i∂D = hγ d (τ ), ad i∂D ,
for all τ ∈ H(div; D). The surjectivity of γ d from Theorem 4.15 implies that hφ, γ g (p) − ad i∂D = 0
1 1
for all φ ∈ H − 2 (∂D). Hence, γ g (p) = adR in H 2 (∂D). Testing (51.6) with τ := 0 and an
arbitrary function q ∈ C0∞ (D) finally yields D (∇·σ − g)q dx = 0. Invoking Theorem 1.32 proves
∇·σ = g.
Lemma 51.2 (Surjectivity of divergence). Let D be a Lipschitz domain in Rd . The operator
∇· : H(div; D) → L2 (D) is surjective, and we have
R
| D q∇·τ dx|
inf sup ≥ ℓ−1
D βD , (51.7)
q∈L2 (D) τ ∈H(div;D) kqkL2 (D) kτ kH(div;D)
1
−2
with βD := (Cps + 1)− 2 where Cps is the constant from the Poincaré–Steklov inequality (3.11)
with p := 2, i.e., Cps kvkL2 (D) ≤ ℓD k∇vkL2 (D) for all v ∈ H01 (D).
Proof. Let q ∈ L2 (D). Let φ ∈ H01 (D) be such that (∇φ, ∇ψ)L2 (D) = (q, ψ)L2 (D) for all ψ ∈
−1
H01 (D), so that k∇φkL2 (D) ≤ Cps ℓD kqkL2 (D) . Setting ςq := −∇φ, we have ςq ∈ H(div; D),
∇·ςq = q, and
−2
kςq k2H(div;D) = k∇φk2L2 (D) + ℓ2D kqk2L2 (D) ≤ (Cps + 1)ℓ2D kqk2L2 (D) ,
−1
so that kςq kH(div;D) ≤ ℓD βD kqkL2 (D) . As a result, we have
R R
q∇·τ dx q∇·ςq dx kqk2L2 (D)
sup D
≥ D = ≥ ℓ−1
D βD kqkL2 (D) ,
τ ∈H(div;D) kτ kH(div;D) kςq kH(div;D) kςq kH(div;D)
and this proves (51.7).
51.1.2 Neumann boundary condition

We now consider the Neumann boundary condition σ·n = an on ∂D, and we assume that an ∈
1
H − 2 (∂D). We still look for d
R σ ∈ H(div; D), and we interpret the boundary condition as γ (σ) =
d
an . Since hγ (σ), 1i∂D = D ∇·σ dx, the data an and g must satisfy the compatibility condition
Z
han , 1i∂D = g dx. (51.8)
D
Since only the gradient of the primal

R variable p now appears in the problem, 1we additionally require
that p ∈ L2∗ (D) := {q ∈ L2 (D) | D q dx = 0}. Since γ d : H(div; D) → H − 2 (∂D) is surjective (see
Theorem 4.15(ii) or Corollary 31.20), there exists a field σn ∈ H(div; D) s.t. γ d (σn ) = an . We
now make the change of variable σ0 := σ − σn . Note that σ0 satisfies the homogeneous Neumann
boundary condition γ d (σ0 ) = 0, i.e.,
σ0 ∈ H0 (div; D) := {ς ∈ H(div; D) | γ d (ς) = 0} = ker(γ d ). (51.9)

H0 (div; D) is a Hilbert space when equipped with the norm k·kH(div;D) . The weak formulation is
as follows: 
 2
 Find σ0 ∈ V := H0 (div; D) and p ∈ Q := L∗ (D) such that
a(σ0 , τ ) + b(τ , p) = Fn (τ ), ∀τ ∈ V , (51.10)

 b(σ , q) = G (q),
0 n ∀q ∈ Q,
R R
with the linear forms Fn (τ ) := D (f − d−1 σn )·τ dx and Gn (q) := − D (g −∇·σn )q dx. The bilinear
forms a and b have been defined in (51.5).
1
Proposition 51.3 (Well-posedness). Assume that f ∈ L2 (D), g ∈ L2 (D), an ∈ H − 2 (∂D) and
that the compatibility condition (51.8) holds true. (i) The problem (51.10) is well-posed. (ii) The
pair (σ := σ0 + σn , p) satisfies the PDEs (51.1) a.e. in D, and the boundary condition γ d (σ) = an
1
is satisfied in H − 2 (∂D).
Remark 51.4 (Choice of σn ).R One possibility to define σn ∈ H(div; D) is to set σ R n := ∇φ, where
φ R∈ H∗1 (D) := {q ∈ H 1 (D) | D q dx = 0} solves the pure Neumann problem D ∇φ·∇r dx :=
− D gr dx + han , γ g (r)i∂D for all r ∈ H∗1 (D). The compatibility condition (51.8) implies that it is
legitimate to take any test function r in H 1 (D) in the above equation. Taking first r ∈ C0∞ (D)
yields ∇·σn = ∆φ = g, whence hγ d (σn ) − an , γ g (r)i∂D = 0 for all r ∈ H 1 (D). The trace map
1
γ g : H 1 (D) → H 2 (∂D) being surjective, we conclude that γ d (σn ) = an . This construction of σn
gives Gn = 0.
51.1.3 Mixed Dirichlet–Neumann boundary conditions

Let ∂Dd ∪ ∂Dn be a partition of the boundary ∂D with |∂Dd | 6= 0 and |∂Dn | 6= 0. We want
to enforce the mixed Dirichlet–Neumann conditions p = ad on ∂Dd and σ·n = an on ∂Dn . A
rigorous mathematical setting for these conditions entails some subtleties.
Concerning the Dirichlet condition, we assume that there exists a bounded extension operator
1 1 1
H 2 (∂Dd ) → H 2 (∂D) (see §31.3.3). We assume that ad ∈ H 2 (∂Dd ), and we denote by ǎd ∈
1
H 2 (∂D) the extension of ad . Concerning the Neumann condition, we have seen in §51.1.2 that
Neumann conditions on ∂D are enforced using the normal trace operator γ d : H(div; D) →
1
H − 2 (∂D). When the Neumann condition is enforced only on ∂Dn , we need to consider the
1
restriction to ∂Dn of linear forms in H − 2 (∂D). Let H e 21 (∂Dn ) be composed of the functions θ
1
defined on ∂Dn whose extension by zero to ∂D, say θ̃, is in H 2 (∂D). Let us denote by h·, ·i∂Dn
the duality pairing between H e 2 (∂Dn )′ and H
1
e 2 (∂Dn ) i.e., the action of a ∈ H
1
e 12 (∂Dn )′ on r ∈
He 2 (∂Dn ) is denoted by ha, ri∂D . Then for all ς ∈ H(div; D), the restriction γ d (ς)|∂D is defined
1
n n
in H e 12 (∂Dn )′ by setting
hγ d (ς)|∂Dn , θi∂D n := hγ d (ς), θ̃i∂D . (51.11)
Lemma 51.5 (Surjectivity of restricted normal trace). The restricted normal trace operator
e 21 (∂Dn )′ is surjective.
γ d (·)|∂Dn : H(div; D) → H
Proof. We proceed as in Remark 51.4. Let an ∈ H e 2 (∂Dn )′ and let us set H 1 (D) := {r ∈
1
d
H 1 (D) | γ g (r)|∂D d = 0}. Notice that for all r ∈ Hd1 (D), the zero extension of γ g (r)|∂D n to ∂D
1
coincides with γ g (r), which is in H 2 (∂D). g e 21 1
R Hence, γ (r)|∂D n g∈ H (∂Dn ). Let φ ∈ Hd1(D) solve
the mixed Dirichlet–Neumann problem D ∇φ·∇r dx = han , γ (r)|∂D n i∂Dn for all r ∈ Hd (D). We
now set σn := ∇φ and observe that ∇·σn = 0. Since for all θ ∈ H e 21 (∂Dn ) there is rθ ∈ H 1 (D)
d
such that γ g (rθ ) = θ̃ (the zero-extension of θ to ∂D), we infer that
Z Z
hγ d (σn )|∂Dn − an , θi∂D n = hγ d (σn ), θ̃i∂D − ∇φ·∇rθ = rθ ∇·σn dx = 0.
D D
e 2 (∂Dn ), we conclude that γ d (σn )|∂D = an , i.e., γ d (·)|∂D
Since this identity holds true for all θ ∈ H
1
n n
is surjective.
Owing to Lemma 51.5, it is natural to assume that an ∈ H e 2 (∂Dn )′ . Referring again to this
1
lemma and its proof, we then infer the existence of σn ∈ H(div; D) with ∇·σn = 0 such that
γ d (σn )|∂Dn = an . Making the change of variable σ0 := σ − σn gives
σ0 ∈ Hn (div; D) := {ς ∈ H(div; D) | γ d (ς)|∂Dn = 0}. (51.12)
Notice that Hn (div; D) is a Hilbert space when equipped with the natural norm. The weak
formulation we now consider is as follows:

 2
 Find σ0 ∈ V := Hn (div; D) and p ∈ Q := L (D) such that
a(σ0 , τ ) + b(τ , p) = Fdn (τ ), ∀τ ∈ V , (51.13)

 b(σ , q) = G (q),
0 n ∀q ∈ Q,
R R
with the linear forms Fdn (τ ) := D (f − d−1 σn )·τ dx − hγ d (τ ), ǎd i∂D and Gn (q) := − D (g −
1
∇·σn )q dx. The bilinear forms a, b are defined in (51.5). Recall that ǎd ∈ H 2 (∂D) is an extension
1
of ad over ∂D, and notice that hγ d (τ ), ǎd i∂D is independent on the way ǎd is extended to H 2 (∂D).
Indeed, considering two extensions ǎd and âd , we have ǎd − âd ∈ H e 2 (∂Dn ), so that hγ d (τ ), ǎd −
1
âd i∂D = 0 for all τ ∈ V .

1
Proposition 51.6 (Well-posedness). Assume f ∈ L2 (D), g ∈ L2 (D), ad ∈ H 2 (∂Dd ), and
an ∈ He 12 (∂Dn )′ . (i) The problem (51.13) is well-posed. (ii) The pair (σ := σ0 + σn , p) satisfies
the PDEs (51.1) a.e. in D, the Dirichlet condition γ g (p)|∂Dd = ad is satisfied a.e. on ∂Dd , and
the Neumann condition γ d (σ)|∂D n = an is satisfied in (H e 12 (∂Dn ))′ .
Proof. We just sketch the differences with the proof of Proposition 51.1. The surjectivity of
∇· : Hn (div; D) → L2 (D) follows by defining φr ∈ H 1 (D) such that ∆φr = r, φr|∂Dd = 0,
∂n φr|∂Dn = 0 for all r ∈ L2 (D) and observing that ∇·(∇φr ) = r and ∇φr ∈ Hn (div; D). To
recover the Dirichlet boundary condition, we observe, as in the proof of Proposition 51.1, that
∇p = f − d−1 (σn − σ0 ), which in turn implies that p ∈ H 1 (D) and
hγ d (τ ), γ g (p)i∂D = hγ d (τ ), ǎd i∂D , ∀τ ∈ Hn (div; D). (51.14)
Let ψ ∈ C0∞ (∂Dd ) d
and let ψ̃ be the extension by zero of ψ to ∂D. Recalling that γ : H(div; D) →
− 12
H (∂D) is surjective (see Theorem 4.15(ii) or Corollary 31.20), there is τψ ∈ H(div; D) s.t.
γ d (τψ ) = ψ̃.R Notice that γ d (τψ )|∂Dn = ψ̃|∂Dn = 0, i.e., τψ ∈ Hn (div; D). Using τψ in (51.14)
shows that ∂Dd (γ g (p) − ad )ψ ds = 0, which in turn gives γ g (p)|∂D d = ad since ψ is arbitrary in
C0∞ (∂Dd ).
51.2 Primal, dual, and dual mixed formulations

In this section, we consider alternative formulations where either the primal variable p or the dual
variable σ is eliminated. We focus on homogeneous Dirichlet boundary conditions for simplicity.
The material readily extends to other types of (non-homogeneous) boundary conditions.
The primal formulation of the PDEs (51.1) with the boundary condition p = 0 on ∂D is obtained
by eliminating σ using the identity σ = d(f − ∇p). This leads to the following formulation:

Find p ∈ H01 (D) such that
R (51.15)
a♯ (p, r) := D d∇p·∇r dx = F♯ (r), ∀r ∈ H01 (D),
R
with the linear form F♯ (r) := D (df ·∇r + gr) dx. This problem has been analyzed in Chapter 31
(see Remark 31.7 since F♯ ∈ H −1 (D)). In particular, p is the unique minimizer of the energy
functional
1
E♯ (q) := a♯ (q, q) − F♯ (q) (51.16)
2
over H01 (D) (see Remark 31.10).
The dual formulation
R is obtained by eliminating p using divergence-free test functions in Darcy’s
law (observe that D ∇p·τ dx = 0 if τ is divergence-free since p|∂D = 0), and enforcing the mass
conservation equation explicitly. This leads to the following formulation:

Find σ ∈ H(div; D) with ∇·σ = g such that
R (51.17)
a♭ (σ, τ ) := D (d−1 σ)·τ dx = F♭ (τ ), ∀τ ∈ H(div = 0; D),
with
R the space H(div = 0; D) := {ς ∈ H(div; D) | ∇·ς = 0} and the linear form F♭ (τ ) :=
D
f ·τ dx. The well-posedness of (51.17) can be established by lifting the divergence constraint
using σg ∈ H(div; D) such that ∇·σg = g, making the change of variable σ0 := σ − σg ∈
H(div = 0; D), and observing that the bilinear form a♭ is coercive on H(div = 0; D) equipped
with the natural norm. Moreover, defining the complementary energy functional
Z
1
E♭ (ς) := − (ς − df )·d−1 (ς − df ) dx, (51.18)
2 D
and since (51.17) amounts to DE♭ (σ)(τ ) = 0 for all τ ∈ H(div = 0; D), we infer that the field σ
solving (51.17) is the unique maximizer of E♭ over the affine subspace of H(div; D) with divergence
equal to g. The dual formulation is seldom used for approximation purposes since it requires to
manipulate divergence-free vector fields. An interesting application related to flux recovery for
H 1 -conforming finite elements is presented in §52.2.
We can now relate the primal formulation (51.15) and the dual formulation (51.17) to the mixed
formulation (51.6), which in the present context is called dual mixed formulation.
Proposition 51.7 (Equivalence, energy identity). The primal formulation (51.15), the dual
formulation (51.17), and the dual mixed formulation (51.6) are equivalent in the sense that the
solutions p from (51.15) and (51.6) coincide, the solutions σ from (51.17) and (51.6) coincide,
and we have d−1 σ + ∇p = f . Moreover, the following energy identity holds true:
min E♯ (q) = E♯ (p) = E♭ (σ) = max E♭ (ς). (51.19)
q∈H01 (D) ς∈H(div;D),∇·ς=g

Remark 51.8 (Lagrangian). Proposition 49.11 implies that the pair (σ, p) solving the dual
mixed formulation (51.6) is the unique saddle point of the Lagrangian
Z
1
L(ς, q) := ( ς·d−1 ς − q∇·ς) dx − Fd (ς) − Gd (q).
D 2
R R
Since
R L(ς, q) = −E♭ (ς)− 12 D f ·df dx+ D q(g−∇·ς) dx, we infer that E♯ (p) = E♭ (σ) = −L(σ, p)−
2 D f ·df dx.
1
Remark 51.9 (Linear elasticity). The above formalism can be applied to the linear elasticity
equations studied in Chapter 42. Denoting by u the displacement and s the stress tensor, Darcy’s
law is replaced by the constitutive equation C:s(u) = e(u) = 12 (∇u + (∇u)T ), where C is the
(fourth-order) compliance tensor (see Exercise 42.1), and the mass conservation equation is replaced
by the equilibrium equation ∇·s(u) = g. We refer the reader to, e.g., Gatica [212, §2.4.3] for some
weak mixed formulations and to §42.4.2 for a brief literature review of their mixed finite element
approximation.
51.3 Approximation of the mixed formulation

In this section, we analyze an H(div)-conforming approximation of the weak formulation (51.6)
focusing on Dirichlet boundary conditions for simplicity.

Let (Th )h∈H be a shape-regular sequence of affine simplicial meshes so that each mesh covers D
exactly, let Pkd (Th ) be the H(div)-conforming Raviart–Thomas finite element space of order k ≥ 0
from §19.2.3, and let Pkb (Th ) be the broken finite element space built using piecewise polynomials
in Pk,d . We recall that
Pkd (Th ) := {ςh ∈ H(div; D) | ψK

d
(ςh|K ) ∈ RT k,d , ∀K ∈ Th }, (51.20a)
g
Pkb (Th ) 2
:= {qh ∈ L (D) | ψK (qh|K ) ∈ Pk,d , ∀K ∈ Th }, (51.20b)
d g
where ψK is the contravariant Piola transformation and ψK is the pullback by the geometric
mapping TK (see Definition 9.8). Notice that ςh|K ∈ RT k,d and qh|K ∈ Pk,d since Th is affine. The
discrete counterpart of (51.6) is

 d b
 Find σh ∈ Vh := Pk (Th ) and ph ∈ Qh := Pk (Th ) such that
a(σh , τh ) + b(τh , ph ) = Fd (τh ), ∀τh ∈ Vh , (51.21)

 b(σ , q ) = G (q ),
h h d h ∀qh ∈ Q h .
Let ker(Bh ) = {ςh ∈ Vh | b(ςh , qh ) = 0, ∀qh ∈ Qh } and recall that ker(B) = {ς ∈ H(div; D) | ∇·ς =
0}.
Lemma 51.10 (Discrete inf-sup). We have
ker(Bh ) ⊂ ker(B), (51.22)
and the following holds true:

R
| D qh ∇·ςh dx|
inf sup ≥ ℓ−1 ♭
D βD , (51.23)
qh ∈Qh ςh ∈Vh kqh kL2 (D) kςh kH(div;D)
1
−2
♭ :=
with βD (Cps kJhd k2L(L2 ;L2 ) + 1)− 2 > 0, where Jhd is the L2 -stable commuting projection
from §23.3.3.
Proof. Since ∇· : Vh := Pkd (Th ) → Pkb (Th ) =: Qh , we have ker(Bh ) ⊂ ker(B). Let us prove (51.23).
Let qh ∈ Qh . Using Lemma 51.2, we know that there is ςqh ∈ H(div; D) such that ∇·ςqh =
−1
qh and kςqh kL2 (D) ≤ Cps ℓD kqh kL2 (D) . Let us set ςh∗ := Jhd (ςqh ). The commuting property
in Theorem 23.12 implies that ∇·ςh∗ = Jhb (∇·ςqh ) = Jhb (qh ) = qh (since Pkb (Th ) is pointwise
invariant under Jhb ). Since kςh∗ kL2 (D) ≤ kJhd kL(L2 ;L2 ) kςqh kL2 (D) , we infer that kςh∗ kH(div;D) ≤
♭ −1
ℓD (βD ) kqh kL2 (D) . As a result, we have
R R
D qh ∇·ςh dx ∗
D qh ∇·ςh dx
kqh k2L2 (D)
sup ≥ ∗ = ≥ ℓ−1 ♭
D βD kqh kL2 (D) ,
ςh ∈Vh kςh kH(div;D) kςh kH(div;D) kςh∗ kH(div;D)
and this proves (51.23).
Corollary 51.11 (Well-posedness). (51.21) is well-posed.
Proof. We apply Proposition 50.1. The condition (50.4a) on the bilinear form a follows from the
coercivity of a on ker(B) and ker(Bh ) ⊂ ker(B), whereas the condition (50.4b) on the bilinear
form b is just (51.23).
Remark 51.12 (Discrete inf-sup). Using the L2 -norm of ςh instead of the H(div)-norm in the
proof of Lemma 51.10, one can show that
R
| D qh ∇·ςh dx| ♯
inf sup ≥ ℓ−1
D βD ,
qh ∈Qh ςh ∈Vh kqh kL2 (D) kςh kL2 (D)
♯
where βD := Cps kJhd k−1 ♭
L(L2 ;L2 ) > βD . We will use this somewhat sharper bound in the proof of the
error estimate in Theorem 51.16.
Remark 51.13 (Ihd vs. Jhd ). One can use the canonical interpolation operator Ihd instead of Jhd
to prove (51.23). Owing to the theory of elliptic regularity in Lipschitz domains, the function ςqh
constructed in Lemma 51.2 is indeed in Lp (D) for some p > 2. Proposition 17.3 then implies that
ςqh is in the domain of Ihd , and the commuting property results from Lemma 19.6.
Remark 51.14 (Fortin operator). Since ∇· : H(div; D) → L2 (D) is surjective, ∇· admits a

bounded right inverse which we denote by (∇·)† . Then Πh = Jhd ◦ (∇·)† ◦ Ihb ◦ (∇·) is a Fortin
operator, where Ihb is the L2 -orthogonal projection onto the broken polynomial space Pkb (Th ); see
Exercise 51.6.
Remark 51.15 (Variants). Other boundary conditions can be treated. For mixed Dirichlet–
Neumann conditions, for instance, one assumes that the boundary faces are located on either ∂Dd
or ∂Dn . Then an Hn (div; D)-conforming subspace is built by taking fields in Pkd (Th ) with zero
normal component on ∂Dn . It is also possible to work with rectangular meshes using Cartesian
Raviart–Thomas elements of degree k ≥ 0 for the dual variable and Qk,d polynomials for the primal
variable.

The error analysis presented in this section follows the general ideas of §50.1.2. We also exploit
the particular structure of Darcy’s equations to derive more specific estimates that give bounds on
the dual variable in H(div) that are independent of the discrete inf-sup constant, which is not the
case of the estimate derived in §50.1.2.
Theorem 51.16 (Error estimate). Let (σ, p) and (σh , ph ) solve (51.6) and (51.21), respectively.
Let Vh,g := {ςh ∈ Vh | Bh (ςh ) = Ihb (g)}. (i) We have
kσ − σh kL2 (D) ≤ c1 inf kσ − ςh kL2 (D) ,

ςh ∈Vh,g
k∇·(σ − σh )kL2 (D) = inf k∇·σ − φh kL2 (D) ,

φh ∈Pkb (Th )
kp − ph kL2 (D) ≤ c3 inf ℓD kσ − ςh kL2 (D) + 2 inf kp − qh kL2 (D) ,

ςh ∈Vh,g qh ∈Qh
λ ♯ −1 ♯
with c1 := λ♭♯ and c3 := λc1♭ (βD ) with βD from Remark 51.12. (ii) If σ ∈ H r (D), ∇·σ ∈ H r (D),
r
and p ∈ H (D) with r ∈ (0, k + 1], then
X 12
kσ − σh kL2 (D) ≤ c h2r 2
K |σ|H r (DK ) ,
K∈Th
X 12
k∇·(σ − σh )kL2 (D) ≤ c h2r 2
K |∇·σ|H r (K) ,
K∈Th
X 1
2
kp − ph k L2 (D) ≤c h2r
K ℓ2D |σ|2H r (DK ) + |p|2H r (K) ,
K∈Th
where DK is the set of the points composing the mesh cells that share at least one face with K ∈ Th
(one can replace DK by K if r > 21 ). In particular, we have kσ − σh kL2 (D) ≤ chr |σ|H r (D) ,
k∇·(σ − σh )kL2 (D) ≤ chr |∇·σ|H r (D) , and kp − ph kL2 (D) ≤ chr (ℓD |σ|H r (D) + |p|H r (D) ).
Proof. (1) We first observe that we have the following Galerkin orthogonality:
a(σ − σh , τh ) + b(τh , p − ph ) = 0, ∀τh ∈ Vh , (51.24a)

b(σ − σh , qh ) = 0, ∀qh ∈ Qh . (51.24b)
Since (Bh (σh ) − Ihb (g), qh )L2 (D) = b(σh − σ, qh ) = 0 for all qh ∈ Qh , owing to (51.24b), and
Bh (σh ) − Ihb (g) ∈ Qh , we infer that σh ∈ Vh,g .
(2) Let ςh ∈ Vh,g so that σh −ςh ∈ ker(Bh ) ⊂ ker(B). Since (51.24a) implies that a(σ −σh , τh ) = 0
for all τh ∈ ker(Bh ), we infer that
λ−1 2
♯ kσ − σh kL2 (D) ≤ a(σ − σh , σ − σh )
= a(σ − σh , σ − ςh ) + a(σ − σh , ςh − σh )
= a(σ − σh , σ − ςh ) ≤ λ−1
♭ kσ − σh kL (D) kσ − ςh kL (D) .
2 2
Thus, kσ − σh kL2 (D) ≤ c1 kσ − ςh kL2 (D) , and the expected bound on kσ − σh kL2 (D) follows by
taking the infimum over ςh ∈ Vh,g .
(3) Since ∇·σh = Bh (σh ) = Ihb (g) and Ihb is the L2 -orthogonal projection onto Pkb (Th ), we have
∇·(σh − σ) = Ihb (g) − g. The optimal bound on k∇·(σ − σh )kL2 (D) follows readily.
(4) Let qh ∈ Qh . Owing to Remark 51.12, we infer that there is τh ∈ Vh such that ∇·τh = −(ph −qh )
♯
and ℓ−1
D βD kτh kL2 (D) ≤ kph − qh kL2 (D) . We infer that
kph − qh k2L2 (D) = b(τh , ph − qh ) = b(τh , ph − p) + b(τh , p − qh )

= a(σ − σh , τh ) + b(τh , p − qh )
≤ λ−1
♭ kσ − σh kL (D) kτh kL (D) + k∇·τh kL (D) kp − qh kL (D) ,
2 2 2 2
owing to (51.24a). The above properties of τh combined with the above bound on kσ − σh kL2 (D)
and the triangle inequality lead to the expected bound on kp − ph kL2 (D) .
(5) The convergence rate on kσ − σh kL2 (D) is obtained by taking ςh := Jhd (σ) and by using
the approximation properties of Jhd , which result from Item (iii) in Theorem 23.12 and from
Theorem 22.6. Note that ςh ∈ Vh,g since ∇·Jhd (σ) = Ihb (∇·σ) = Ihb (g). The other two estimates
follow from the estimate on kσ − σh kL2 (D) and the approximation properties of Ihb .
We now use duality techniques to derive an estimate on the primal variable with a better rate
of convergence. One difference with the primal formulation analyzed in §32.2 is that we now bound
the (discrete) error Ihb (p) − ph , instead of the full error p − ph . As in §32.2, we assume that the
following smoothing property is satisfied: There are real numbers s ∈ (0, 1] and csmo such that
the (adjoint) solution zφ ∈ H01 (D) to the PDE −∇·(d∇zφ ) = φ for all φ ∈ L2 (D), satisfies the
a priori bound kzφ kH 1+s (D) ≤ csmo ℓ−2D kφkL2 (D) . Sufficient conditions for this smoothing property
are given by the elliptic regularity theory; see §31.4. We also assume that d is s.t. the map
H s (D) ∋ ξ 7→ d·ξ ∈ H s (D) is bounded (see (31.34)).
Theorem 51.17 (Potential supercloseness). Under the above smoothing property and multi-
plier assumption, the following holds true:
kIhb (p) − ph kL2 (D) ≤ c hs ℓ1−s

D (kσ − σh kL2 (D) + hk∇·(σ − σh )kL2 (D) ). (51.25)
Proof. Let zφ ∈ H 1+s (D) ∩ H01 (D) be the adjoint solution with data φ := Ihb (p) − ph , i.e.,
−∇·(d∇zφ ) = Ihb (p) − ph , and let us set ξ := −d∇zφ so that ∇·ξ = Ihb (p) − ph . We observe
that
kIhb (p) − ph k2L2 (D) = (p − ph , Ihb (p) − ph )L2 (D) = (p − ph , ∇·ξ)L2 (D)
= (p − ph , ∇·Jhd (ξ))L2 (D) = (d−1 (σ − σh ), Jhd (ξ))L2 (D)
= (d−1 (σ − σh ), Jhd (ξ) − ξ)L2 (D) + (d−1 (σ − σh ), ξ)L2 (D) ,
where in the first line we used that (p−Ihb (p), qh )L2 (D) = 0 for all qh ∈ Pkb (Th ) and ∇·ξ = Ihb (p)−ph ,
and in the second line we used that ∇·Jhd (ξ) = Jhb (∇·ξ) = ∇·ξ, since ∇·ξ ∈ Pkb (Th ), and the
identity (51.24a) with τh := Jhd (ξ). The first term on the right-hand side, say T1 , is bounded as
follows:
|T1 | ≤ λ♭−1 kσ − σh kL2 (D) kJhd (ξ) − ξkL2 (D) ≤ c kσ − σh kL2 (D) hs |ξ|H s (D) ,
and |ξ|H s (D) ≤ c|zφ |H 1+s (D) owing to the multiplier assumption. For the second term, say T2 ,
(51.24b) implies that
(σh − σ, ∇zφ )L2 (D) = (∇·(σ − σh ), zφ )L2 (D)

= (∇·(σ − σh ), zφ − Ihb (zφ ))L2 (D) .
Hence, |T2 | ≤ ck∇·(σ − σh )kL2 (D) h1+s |zφ |H 1+s (D) . Finally, we have
|zφ |H 1+s (D) ≤ ℓ−1−s

D kzφ kH 1+s (D) ≤ csmo ℓ1−s b
D kIh (p) − ph kL2 (D) ,
owing to the smoothing property.
Remark 51.18 (L∞ -bounds). L∞ -bounds on the dual and primal errors can be derived as in,
e.g., Gastaldi and Nochetto [211].
Exercises
Exercise 51.1 (Compactness). Let D := (0, 1)3 be the unit cube in R3 . Show that the embed-
ding H0 (div; D) ֒→ L2 (D) is not compact. (Hint : let
1
φ1,n (x1 , x2 , x3 ) := sin(nπx2 ) sin(nπx3 ),
nπ
1
nπ
1
nπ
for all n ≥ 1, set vn := ∇×φn , and prove first that (vn )n≥1 weakly converges to zero in L2 (D)
(see Definition C.28), then compute kvn kL2 (D) and argue by contradiction.)
Exercise 51.2 (Neumann condition). Prove Proposition 51.3. (Hint : for the surjectivity of
the divergence, solve a pure Neumann problem.)
Exercise
R 51.3 (Integration by parts). Let Hd1 (D) and Hn (div; D) be defined in §51.1.3. Prove
that D (∇q·ς + q∇·ς) dx = 0 for all q ∈ Hd1 (D) and all ς ∈ Hn (div; D). (Hint : observe that
e 21 (∂Dn ).)
γ g (q)|∂D n ∈ H
Exercise 51.4 (Primal, dual formulations). Prove Proposition 51.7.
Exercise 51.5 (Primal mixed formulation). Consider the problem: Find p ∈ H 1 (D) such
1
that −∆p = f and γ g (p) = g with f ∈ L2 (D) and g ∈ H 2 (∂D). Derive a mixed formulation of
1
this problem with unknowns (p, λ) ∈ H 1 (D)×H − 2 (∂D) and show that it is well-posed. (Hint : set
1
b(v, µ) := hµ, γ g (v)i∂D and observe that B = γ g : H 1 (D) → H 2 (∂D).) Recover the PDE and the
boundary condition. Note: this method is introduced in Babuška [34].
Exercise 51.6 (Fortin operator). Justify Remark 51.14. (Hint : use arguments similar to those
of the proof of Lemma 51.10.)
Exercise 51.7 (Inf-sup condition). The goal is to prove the inf-sup condition (51.23) using the
canonical Raviart–Thomas interpolation operator. (i) Do this by using elliptic regularity. (Hint :
solve a Dirichlet problem.) (ii) Do this again by using the surjectivity of ∇· : H 1 (D) → L2 (D).
Exercise 51.8 (Error estimate). (i) Prove that
kσ − σh kH(div;D) ≤ c′1 inf kσ − ςh kH(div;D) ,

ςh ∈Vh
kp − ph kL2 (D) ≤ c′3 inf kσ − ςh kH(div;D) + 2 inf kp − qh kL2 (D) ,

ςh ∈Vh qh ∈Qh
λ♯ c′1
with c′1 := (1 + λ♭ )(1 + β1 ) and c′3 := λ♭ β ′ . (ii) Assuming that σ ∈ H r (D), ∇·σ ∈ H r (D), and
L2
p ∈ H r (D) with r ∈ (0, k + 1], prove that
kσ − σh kH(div;D) ≤ c hr (|σ|H r (D) + |∇·σ|H r (D) ),

kp − ph kL2 (D) ≤ c hr (|σ|H r (D) + |∇·σ|H r (D) + |p|H r (D) ).
(Hint : use the commuting projection Jhd .)

Exercise 51.9 (Box scheme). Let d := λ0 Id , λ0 > 0, and enforce the boundary condition
γ g (p) = 0. Let Vh := P0d (Th )×P1,0 cr cr
(Th ), where P1,0 (Th ) is the Crouzeix–Raviart space defined
b b
in (36.8). Let Wh := P0 (Th )×P0 (Th ). Consider the bilinear form ah : Vh ×Wh → R defined
by ah (vh , wh ) := λ−1
0 (σh , τh )L2 (D) + (∇·σh , qh )L2 (D) + (∇h ph , τh )L2 (D) with vh := (σh , ph ) and
wh := (τh , qh ) (see Definition 36.3 for the broken gradient ∇h ). (i) Prove that dim(Vh ) = dim(Wh )
and that there is α > 0 s.t. for all vh ∈ Vh and all h ∈ H, αkvh kVh ≤ supwh ∈Wh |akw h (vh ,wh )|
h kW
with
h
kvh k2Vh := λ−1 2 2 2 −1 2 −2 2
0 kσh kH(div;D) + λ0 k∇h ph kL2 (D) and kwh kWh := λ0 kτh kL2 (D) + λ0 ℓD kqh kL2 (D) .
(Hint : test with (σ h + λ0 ∇h ph , 2ph + ℓ2D λ−1 2
0 ∇·σh ), where (σ h , ph ) is the L -orthogonal projection
of (σh , ph ) onto Wh .) (ii) Consider the discrete problem: Find uh ∈ Vh such that ah (uh , wh ) =
(f , τh )L2 (D) + (g, qh )L2 (D) for all wh ∈ Wh . Show that this problem is well-posed, prove a quasi-
optimal error estimate, and show that the error converges to zero with rate h if the exact solution
is smooth enough. (Hint : use Lemma 27.5.) Note: the scheme has been introduced in Croisille
[149] to approximate (51.1). It is a Petrov–Galerkin scheme with only local test functions.
Chapter 52
Potential and flux recovery
This chapter addresses topics related to the approximation of Darcy’s equations using either mixed
or H 1 -conforming finite elements. Mixed finite elements approximate the flux (i.e., the dual variable
σ) in H(div; D), but the connection to the gradient of the potential (i.e., the primal variable p)
sitting in H01 (D) is enforced weakly. We show here how this connection can be made explicit using
hybridization techniques. Alternatively, H 1 -conforming finite elements approximate the primal
variable in H01 (D), but the connection to the dual variable σ sitting in H(div; D) is enforced
weakly. We show here how this connection can be made explicit by using a local post-processing
technique. In the whole chapter, we consider homogeneous Dirichlet boundary conditions on the
potential for simplicity, and we assume that (Th )h∈H is a shape-regular sequence of affine simplicial
meshes so that each mesh covers the domain D ⊂ Rd exactly.
52.1 Hybridization of mixed finite elements

Hybridization was introduced by Fraejis de Veubeke in 1965 (see [207] for a reprint) as a compu-
tationally effective technique to transform the symmetric indefinite linear system on the potential
and the flux into a symmetric positive definite system on an auxiliary variable playing the role of
a Lagrange multiplier associated with the continuity constraint on the normal component of the
flux. As shown in the seminal work of Arnold and Brezzi [16], viewing the auxiliary variable as
a potential trace on the mesh faces allows one to devise a post-processed potential with better
approximation properties.
52.1.1 From hybridization to static condensation

Let us focus on the same discrete setting as in §51.3, where we employed simplicial RT k,d Raviart–
Thomas elements for the flux and broken Pk,d finite elements for the potential with some polynomial
degree k ≥ 0, i.e.,
Vh := Pkd (Th ) := {ςh ∈ H(div; D) | ψK

d
(ςh|K ) ∈ RT k,d , ∀K ∈ Th }, (52.1a)
g
Qh := Pkb (Th ) 2
:= {qh ∈ L (D) | ψK (qh|K ) ∈ Pk,d , ∀K ∈ Th }, (52.1b)
d g
where ψK is the contravariant Piola transformation and ψK is the pullback by R the geometric
mapping TK (see Definition 9.8). Recall that the bilinear forms are a(ςh , τh ) := D (d−1 ςh )·τh dx
328 Chapter 52. Potential and flux recovery
R R
R b(ςh , qh ) := − D (∇·ςh )qh dx, and that the linear forms are F (τh ) := D τh ·f dx and G(qh ) :=
and
− D qh g dx. Our starting point is the discrete problem (51.21) which consists of seeking the pair
(σh , ph ) ∈ Vh ×Qh such that
a(σh , τh ) + b(τh , ph ) = F (τh ), ∀τh ∈ Vh , (52.2a)

b(σh , qh ) = G(qh ), ∀qh ∈ Qh . (52.2b)
Recall that this problem is well-posed (see Corollary 51.11) and gives optimal error estimates (see
Theorem 51.16).
Let Λh be the space composed of the functions that are piecewise polynomials of degree at
most k on the mesh interfaces and are extended by zero on all the boundary faces, i.e.,
Λh := {λh ∈ L2 (Fh ) | λh ◦ TF ∈ Pk,d−1 , ∀F ∈ Fh◦ , λh|∂D = 0}, (52.3)
where TF is an affine bijective mapping from the unit simplex of Rd−1 to F . Let Vhhy :=
Pkd,b (Th ) := {ςh ∈ L2 (D) | ψK
d
(ςh|K ) ∈ RT k,d , ∀K ∈ Th } be the broken Raviart–Thomas space
(this space is composed of piecewise RT k,d polynomials in each mesh cell since the mesh is affine).
Recall that Vh = H(div; D) ∩ Vhhy (see §18.2.3). We define the bilinear form bh (ςh , qh ) :=
P R
− K∈Th K (∇·ςh|K )qh dx on Vhhy ×Qh and observe that bh|Vh ×Qh = b. The hybridized version of
the discrete problem (51.21) consists of seeking the triple (σh′ , p′h , λh ) ∈ Vhhy ×Qh ×Λh such that
a(σh′ , τh ) + bh (τh , p′h ) + ch (τh , λh ) = F (τh ), ∀τh ∈ Vhhy , (52.4a)

bh (σh′ , qh ) = G(qh ), ∀qh ∈ Qh , (52.4b)
ch (σh′ , µh ) = 0, ∀µh ∈ Λh , (52.4c)
with the bilinear form

X Z
ch (τh , µh ) := [[τh ]]·nF µh ds. (52.5)
F ∈Fh◦ F
Proposition 52.1 (Equivalence). The discrete problem (52.4) admits a unique solution
(σh′ , p′h , λh ) ∈ Vhhy ×Qh ×Λh .
Moreover, σh′ ∈ Vh , and the pair (σh′ , p′h ) is the unique solution to (52.2).
Proof. The well-posedness of (52.4) is treated in Exercise 52.1. Assume that (σh′ , p′h , λh ) ∈
Vhhy ×Qh ×Λh is the unique solution to (52.4). Equation (52.4c) implies that the normal com-
ponent of σh′ is continuous, since Lemma 14.7 shows that (σh′ ·nF ) ◦ TF is in Pk,d−1 for all
F ∈ Fh . Owing to Theorem 18.10, we infer that σh′ ∈ H(div; D), i.e., σh′ ∈ Vh . Restricting
the test functions in (52.4a) to be in Vh so that the bilinear form bh can be replaced by b, and
using (52.4b), shows that the pair (σh′ , p′h ) solves (52.2). Uniqueness of the solution to (52.2) shows
that (σh′ , p′h ) = (σh , ph ).
Owing to the equivalence result stated in Proposition 52.1, we drop the primes from now on
in the discrete problem (52.4). The advantage of (52.4) over (52.2) is that the pair (σh , ph ) can
be eliminated locally. This operation is called static condensation (see §28.1.2 and §39.2.2). Let
d −1 RT g −1
VkK := (ψK ) (RT k,d )×(ψK ) (Pk,d ) (since the mesh is affine, we have VkK := RT k,d ×Pk,d ). We
define the following local bilinear form on VkK ×VkK :
âK ((σ, p), (τ , q)) := (d−1 σ, τ )L2 (K) − (∇·τ , p)L2 (K) − (∇·σ, q)L2 (K) .
For all µ ∈ L2 (∂K), we define the polynomial pair (Sµ , Pµ ) ∈ VkK by solving the following local
problem: For all (τ , q) ∈ VkK ,
âK ((Sµ , Pµ ), (τ , q)) = −(µ, τ ·nK )L2 (∂K) . (52.6)
Notice that this definition implies, in particular, that ∇·Sµ = 0. For all (φ, ψ) ∈ L2 (K)×L2 (K),
we define the polynomial pair (Sφ,ψ , Pφ,ψ ) ∈ VkK by solving the following local problem: For all
(τ , q) ∈ VkK ,
âK ((Sφ,ψ , Pφ,ψ ), (τ , q)) = (φ, τ )L2 (K) − (ψ, q)L2 (K) . (52.7)
Notice that both local problems are well-posed since âK satisfies an inf-sup condition on VkK ×VkK .
For all µh ∈ Λh and all K ∈ Th , we denote by µ∂K := (µh|F )F ∈FK the restriction of µh to the
mesh faces in ∂K.
Proposition 52.2 (Static condensation). (i) (σh , ph , λh ) solves (52.4) if and only if (σh , ph )|K =
(Sλ∂K , Pλ∂K ) + (Sf|K ,g|K , Pf|K ,g|K ) for all K ∈ Th , where λh ∈ Λh is the unique solution of
X
(d−1 Sλ∂K , Sµ∂K )L2 (K) = ℓ(µh ), ∀µh ∈ Λh , (52.8)
K∈Th
P
with ℓ(µh ) := K∈Th (g, Pµ∂K )L2 (K) −(f , Sµ∂K )L2 (K) . (ii) The algebraic realization of (52.8) leads
to a symmetric positive definite matrix.
Proof. (i) Assume that (σh , ph , λh ) solves (52.4) and let us show that (σh , ph )|K = (Sλ∂K , Pλ∂K ) +
(Sf|K ,g|K , Pf|K ,g|K ). Let K ∈ Th and (τ , q) ∈ VkK . Extending these functions by zero to D leads
to a pair (τ̃h , q̃h ) ∈ Vhhy ×Qh that we can use as a test function in (52.4a)-(52.4b). This leads to
âK ((σh|K , ph|K ), (τ , q)) = a(σh , τ̃h ) + b(τ̃h , ph ) + b(σh , q̃h )
= F (τ̃h )+G(q̃h )−ch (τ̃h , λh )
= (f , τ )L2 (K) −(g, q)L2 (K) −(λh , τ ·nK )L2 (∂K)
= âK ((Sf|K ,g|K , Pf|K ,g|K ), (τ , q)) + âK ((Sλ∂K , Pλ∂K ), (τ , q)).
Since (τ , q) ∈ VkK is arbitrary, this proves that (σh , ph )|K = (Sλ∂K , Pλ∂K ) + (Sf|K ,g|K , Pf|K ,g|K ).
Let us now establish (52.8). We first observe that (52.4c) together with the definition (52.5) of the
bilinear form ch implies that the following identity holds true for all µh ∈ Λh :
X
0 = ch (σh , µh ) = (σh|K ·nK , µ∂K )L2 (∂K) (52.9)
K∈Th
X
= (Sλ∂K ·nK , µ∂K )L2 (∂K) + (Sf|K ,g|K ·nK , µ∂K )L2 (∂K) .
K∈Th
Using the definition of (Sµ∂K , Pµ∂K ), observing that ∇·Sλ∂K = ∇·Sµ∂K = 0, and since d is
symmetric, we also infer that
(Sλ∂K ·nK , µ∂K )L2 (∂K) = − âK ((Sµ∂K , Pµ∂K ), (Sλ∂K , Pλ∂K ))
= −(d−1 Sλ∂K , Sµ∂K )L2 (K) . (52.10)
Using the definition of (Sµ∂K , Pµ∂K ), the symmetry of âK , and the definition of (Sf|K ,g|K , Pf|K ,g|K ),
we finally infer that
(Sf|K ,g|K ·nK , µ∂K )L2 (∂K) = −âK ((Sµ∂K , Pµ∂K ), (Sf|K ,g|K , Pf|K ,g|K ))
= −âK ((Sf|K ,g|K , Pf|K ,g|K ), (Sµ∂K , Pµ∂K ))
= −(f , Sµ∂K )L2 (K) + (g, Pµ∂K )L2 (K) . (52.11)
Adding the identities (52.10)-(52.11), summing the result over K ∈ Th , and adding the new
identity thus obtained to (52.9) leads to (52.8). The proof of the converse statement employs
similar arguments and is left to the reader.
(ii) The matrix associated with (52.8) is symmetric positivePsemidefinite by construction. Let
K∈Th (d
−1
us show the definiteness. Assume that λh ∈ Λh satisfies Sλ∂K , Sλ∂K )L2 (K) = 0.
Then Sλ∂K = R0 for all K ∈ Th . Owing
R to (52.6) and the definition of the spaces Vh and VkK ,
we infer that K Pλ∂K ∇·τ|K dx = ∂KP λ∂K τ|KR·nK ds for all τ ∈ Vh . Summing over the mesh
cells and since τ ∈ Vh , we infer that K∈Th K Pλ∂K ∇·τ dx = 0. Since ∇·Vh = Qh , choosing
τ so that ∇·τ|K = Pλ∂K for all K ∈ Th implies that Pλ∂K = 0. This in turn implies that
R d −1 RT d d −1 RT
∂K λ∂K τ ·nK ds = 0 for all τ ∈ (ψK ) (RT k,d ). Since λ∂K ∈ γ∂K ((ψK ) (RT k,d )), this argument
shows that λ∂K = 0.
Remark 52.3 (Literature). The above proof is inspired from Cockburn [130], Boffi et al. [65,
§7.2-7.3], and Cockburn and Gopalakrishnan [131].
Remark 52.4 (Lowest-order (k = 0)). There is a close link between the lowest-order Raviart–
Thomas elements and the Crouzeix–Raviart elements from Chapter 36 when d and g are piecewise
constant and f := 0 in (52.2); see Marini [295] and Exercise 52.2. The implementation of the
lowest-order Raviart–Thomas method with one unknown per cell and connections to finite volume
and mimetic finite difference methods are discussed in Younes et al. [399], Vohralı́k and Wohlmuth
[385].
52.1.2 From hybridization to post-processing

Let λh be the solution to the global “skeleton” problem (52.8). Let (σh , ph ) := (Sλh +Sf|K ,g|K , Pλh +
Pf|K ,g|K ). Recall that we have shown that (σh , ph ) solves (52.2) and (σh , ph , λh ) solves (52.4). We
are now going to post-process ph and λh to construct a potential mnc h in a piecewise polynomial
space of higher order. The superscript refers to the fact that mnc h is a nonconforming function, i.e.,
mnch ∈
6 H 1
0 (D). However, we will see below that the jumps of m nc
h across the mesh interfaces and
its trace on the boundary faces have vanishing moments against polynomials of degree at most k.
Let QK and Λ∂K be composed of the restriction to K and ∂K of the functions in Qh and Λh ,
−1
respectively. For Raviart–Thomas elements, QK ◦TK := Pk,d and Λ∂K is composed of piecewise
polynomials of degree k on the faces of K, i.e., λ∂K|F ◦TF ∈ Pk,d−1 for all F ∈ FK (recall that
the mesh is affine by assumption). Let ΠQK and ΠΛ∂K denote the corresponding L2 -orthogonal
projections. The post-processed potential mnc h is built locally in each mesh cell. One first picks
a polynomial space Mk,k′ satisfying Pk,d ⊂ Mk,k′ ⊂ Pk′ ,d and such that the problem of seeking
−1
mK ◦TK ∈ Mk,k′ s.t. ΠQK (mK ) := qK and ΠΛ∂K (mK ) := λ∂K is solvable for all qK ∈ QK and all
λ∂K ∈ Λ∂K . Then mnc h is defined such that for all K ∈ Th ,
mnc
h|K ∈ Mk,k′ , ΠQK (mnc
h|K ) := ph|K , ΠΛ∂K (mnc
h|K ) := λh|∂K . (52.12)
There are in general various admissible choices for the polynomial space Mk,k′ . For instance, if
one works with simplicial Raviart–Thomas elements of degree k, one can set Mk,k′ := Pk′ ,d with
′
k ′ := k + 2 since we have dim(Pk′ ,d ) = k d+d > k+d d + (d + 1) k+d−1
d−1 = dim(Pk,d ) + (d +
1) dim(Pk,d−1 ) = dim(QK ) + dim(Λ∂K ). Alternatively, one can take a smaller space Mk,k′ so that
dim(Mk,k′ ) = dim(QK ) + dim(Λ∂K ). For the lowest-order Raviart–Thomas elements in R2 , one
can set M0,2 := span{1, b b1 , λ
λ0 λ b2 , λ
b1 λ b0 }, where {λ
b2 λ b0 , λ
b1 , λ
b2 } are the barycentric coordinates on
′
the reference element, i.e., k = 0 and k = 2. A similar choice can be made in dimension 3 with
k ′ = 3. Further examples can be found in Arnold and Brezzi [16], Vohralı́k [382] for k := 0 and
more generally in Arbogast and Chen [13] for all k ≥ 0.
Proposition 52.5 (Post-processed potential). Let mnc

h satisfy (52.12). The following holds
true for all τ ∈ RT k,d and all K ∈ Th :
(d−1 σh − f + ∇mnc
h , τ )L2 (K) = 0. (52.13)
Moreover, recalling the convention that [[v]]F = v|F if F ∈ Fh∂ , the following holds true for all
ζ ∈ Pk,d−1 and all F ∈ Fh : Z
−1
[[mnc
h ]]F (ζ ◦ TF ) ds = 0. (52.14)
F
Proof. Let us take a test function τK supported in a single mesh cell K ∈ Th in (52.4a). Integrating
by parts in K and using (52.12) together with ∇·τK ∈ QK and τK|∂K ·nK ∈ Λ∂K , we infer that
(d−1 σh − f + ∇mnc
h , τK )L2 (K)
= (d−1 σh − f , τK )L2 (K) − (mnc nc
h , ∇·τK )L2 (K) + (mh , τK ·nK )L2 (∂K)
= (d−1 σh − f , τK )L2 (K) − (ph , ∇·τK )L2 (K) + (λh , τK ·nK )L2 (∂K) = 0.
This proves (52.13). Observing that λh is single-valued on the Rinterfaces and vanishes on the
−1
boundary faces, the rightmost equation in (52.12) implies that 0 = F [[mnc
h − λh|F ]]F (ζ ◦ TF ) ds =
R −1
F
[[mnc
h ]]F (ζ ◦ TF ) ds. This proves (52.14).
The post-processed potential mnc

h can be used in the a priori and a posteriori analysis of mixed
finite element methods; see Vohralı́k [383].
52.2 Flux recovery for H 1-conforming elements

In this section, we return to the primal formulation of the model elliptic problem considered in
Chapter 32, i.e., −∇·(d∇p) = g in D with p|∂D = 0 (we write p instead of u and g instead of f for
consistency with the present notation). As before, we assume that the eigenvalues of d are in the
interval [λ♭ , λ♯ ] a.e. in D with λ♭ > 0. The exact flux is σ := −d∇p ∈ L2 (D) (this corresponds to
setting f := 0 in Darcy’s law). A crucial observation is that
σ ∈ H(div; D), ∇·σ = g. (52.15)
g
Let ph ∈ Pk,0 (Th ) be the discrete solution obtained from the H01 (D)-conforming finite element
approximation of order k ≥ 1. Recall that ph satisfies (d∇ph , ∇wh )L2 (D) = (g, wh )L2 (D) for all
wh ∈ Pk,0g
(Th ); see (32.5). The approximate flux σh := −d∇ph ∈ L2 (D) delivers an accurate
approximation of the exact flux σ. We indeed have kσ − σh kL2 (D) ≤ λ♯ |p − ph |H 1 (D) , and we
have seen in §32.2 that the error |p − ph |H 1 (D) converges to zero as h → 0 with the rate O(hr )
provided p ∈ H 1+r (D) and r ∈ (0, k]. But for this approximation we do not have σh ∈ H(div; D).
Since it is desirable for some applications to have a discrete flux in H(div; D), we now present a
post-processing technique to build a post-processed flux σh∗ ∈ H(div; D) s.t.
(∇·σh∗ , q)L2 (K) = (g, q)L2 (K) , ∀q ∈ Pl,d , (52.16)
for all K ∈ Th , l ∈ {k − 1, k}, and such that kσh − σh∗ kL2 (K) is bounded by the local H 1 -seminorm
of the error p − ph (up to some data oscillation). Notice that the post-processing is local and does
not require any additional global solve. The flux σh∗ can be used as the transport velocity field in
underground flow applications (see Bastian and Rivière [47]). The post-processed flux can also be
used to evaluate sharp a posteriori error estimates (see §52.2.3).
52.2.1 Local flux equilibration

We locally construct a flux σh∗ that satisfies (52.16) in this section, and we show that kσh −σh∗ kL2 (K)
behaves as expected in §52.2.2. Let z ∈ Vh be a mesh vertex, let ψz be the corresponding P1
Lagrange basis function (also called hat or Courant basis function). The support of ψz is denoted
by Dz and consists of all the mesh cells in the set Tz having z as vertex. This set is often called
finite element star.
d
Let l ≥ 0 and Pl,∗ (Tz ) be the (local) Raviart–Thomas finite element space of order l in the star
d
Dz with the additional requirement that every function τz ∈ Pl,∗ (Tz ) is such that (τz ·nDz )|∂Dz = 0
◦ ∂
if z ∈ Vh or (τz ·nDz )|∂Dz \∂D = 0 if z ∈ Vh , where nDz is the outward unit normal to Dz . Let
b
Pl,∗ (Tz ) be the (local) broken space of scalar-valued finite elements of order l in the star Dz with
the constraint that every function qz ∈ Pl,∗ b
(Tz ) satisfies (qz , 1)L2 (Dz ) = 0 if z ∈ Vh◦ . Let I d,b
l
d,b
be the interpolation operator in the broken Raviart–Thomas space Pl,∗ (Tz ) (without boundary
conditions) and let Ilb be the L2 -orthogonal projection onto the broken space Plb (Tz ). Let us set
fz := −ψz d∇ph , gz := ψz g − (d∇ph )·∇ψz , and consider the constrained minimization problem
σz∗ := arg min kτz − I d,b
l (fz )kL2 (Dz ) . (52.17)
d (T ),∇·τ =I b (g )
τz ∈Pl,∗ z z l z
Following the discussion in §51.2, the problem (52.17) can be efficiently solved by considering the
following dual mixed formulation:
 ∗ d ∗ b
 Find σz ∈ Pl,∗ (Tz ) and rz ∈ Pl,∗ (Tz ) such that

I d,b
(σz∗ , τz )L2 (Dz ) − (∇·τz , rz∗ )L2 (Dz ) = (I l (fz ), τz )L2 (Dz ) ,
d
∀τz ∈ Pl,∗ (Tz ),

 (∇·σ ∗ , q ) 2 b
z z L (Dz ) = (gz , qz )L2 (Dz ) , ∀qz ∈ Pl,∗ (Tz ).
We obtain a pure Neumann problem if z ∈ Vh◦ and a mixed Dirichlet–Neumann problem if z ∈ Vh∂ .
The pure Neumann problem is well-posed owing to the compatibility condition
(Ilb (gz ), 1)L2 (Dz ) = (gz , 1)L2 (Dz ) = (g, ψz )L2 (Dz ) − (d∇ph , ∇ψz )L2 (Dz ) = 0,
which is the Galerkin orthogonality property on the hat basis functions.
Theorem 52.6 (Equilibrated flux). Let l ≥ 0. Let σz∗ be Pdefined∗by (52.17)∗for all z ∈ Vh , and
e z∗ be the zero extension of σz∗ outside Dz . Set σh∗ := z∈Vh σ
let σ e z . Then σh ∈ H(div; D), and
the divergence of σh∗ satisfies (52.16).
Proof. For every z ∈ Vh , the normal component of σz∗ is continuous across all the interfaces in the
mesh Tz since σz∗ ∈ Pl,∗ d
(Tz ). Recall also that by definition the normal component of σz∗ is zero
on all the boundary faces F of Tz that are not in Fh∂ (i.e., those in Fh◦ ). Invoking Theorem 18.10
shows that σe z∗ ∈ H(div;PD). This argument implies that σh∗ ∈ H(div; D). Furthermore, after
∗ ∗
observing that σh|K = z∈VK σz|K , where VK is the collection of all the vertices of K, we have
X X
(∇·σh∗ , q)L2 (K) = (∇·σz∗ , q)L2 (K) = (gz , q)L2 (K) = (g, q)L2 (K) ,
z∈VK z∈VK
P P
for all q ∈ Pl,d , since the local partition of unity z∈VK ψz|K = 1 implies that z∈VK gz = g.
Remark 52.7 (Local vs. global). The post-processed flux σh∗ is a member of the Raviart–
Thomas finite element space Pld (Th ) of order l (see (51.20a)). Yet, the construction of σh∗ is local,
so the technique discussed above is an inexpensive alternative to the global equilibration procedure
defined by σh∗glob := arg minτh ∈Pld (Th ),∇·τh =Ihb (g) kτh − σh kL2 (D) which requires to solve a global
Darcy problem using mixed elements.
Remark 52.8 (Literature). That H 1 -conforming finite elements can be post-processed to give
quantities that enjoy local conservation properties has been highlighted in Hughes et al. [250]. The
present construction of σh∗ is inspired by the seminal ideas in Braess and Schöberl [76], Braess
et al. [77]; see also Ern and Vohralı́k [190, 191, 192] for further results. A somewhat simpler
construction in the lowest-order case can be found in Destuynder and Métivet [163]; see also Larson
and Niklasson [275]. An alternative approach consists of performing the flux equilibration on a
dual (barycentric) mesh; see Luce and Wohlmuth [290], Ern and Vohralı́k [189], Vohralı́k [384] for
the lowest-order case and see Hannukainen et al. [238] for a higher-order extension.
52.2.2 L2 -norm estimate

In this section, we assume for simplicity that d is piecewise constant. We also assume that there
exists a polynomial space Mk,k′ as considered in §52.1.2.
Lemma 52.9 (L2 -estimate). Let σh∗ be defined as in Theorem 52.6 with l ≥ k − 1. Recall that
σh := −d∇ph . There is a constant c such that for all K ∈ Th and all h ∈ H,
X
kσh∗ − σh kL2 (K) ≤ c hK ′ kg + ∇·(d∇ph )kL2 (K ′ )
K ′ ∈ŤK
X
hF ′ k[[d∇ph ]]·nF k
1
+ 2
L2 (F ′ ) , (52.18)
◦
F ′ ∈F̌K
◦
where ŤK and F̌K are the collections of those cells and interfaces that share at least one vertex with
v :=
K, respectively. Moreover, defining the oscillation term ωK ′ hK ′ kg − Ilb (g)kL2 (K ′ ) , we have
X
kσh∗ − σh kL2 (K) ≤ c v
(|p − ph |H 1 (K ′ ) + ωK ′ ). (52.19)
K ′ ∈ŤK
Proof. Let K ∈ Th and let us estimate kσh∗ + d∇ph kL2 (K) . Since d is piecewise constant and
l ≥ k − 1, we infer that (d∇ph )|K is in RT l,d . Hence, I d,b
l (d∇ph ) = d∇ph . Recalling that fz :=
P
−ψz d∇ph , using the local partition of unity and the linearity of I d,b ∗
l , and since σh|K =
∗
z∈VK σz ,
we infer that
!
X X X
(σh + d∇ph )|K =
∗ ∗
σz + I ld,b
ψz d∇ph = (σz∗ − I d,b
l (fz )).
z∈VK z∈VK z∈VK
Invoking the triangle inequality leads to

X
kσh∗ + d∇ph kL2 (K) ≤ kσz∗ − I d,b
l (fz )kL2 (Dz ) ,
z∈VK
which shows that we are left with estimating kσz∗ − I d,b l (fz )kL2 (Dz ) for all z ∈ VK . Owing to
Exercise 52.5, we infer that there is c > 0 such that for all z ∈ Vh and all h ∈ H,
X X 1
c kσz∗ − I d,b
l (fz )kL2 (Dz ) ≤ hK ′ kδzv kL2 (K ′ ) + hF2 ′ kδzs kL2 (F ′ ) , (52.20)
K ′ ∈Tz F ′ ∈Fz◦
with δzv := ∇·fz − gz and δzs := [[fz ]]·nF ′ . This leads to the bound (52.18) since δzv = −ψz (g +
∇·(d∇ph )), δzs = ψz [[d∇ph ]]·nF , and kψz kL∞ (Dz ) = 1. We refer the reader to Exercise 52.4 for
the proof of (52.19).
Remark 52.10 (Choice of l). Lemma 52.9 shows that σh∗ ∈ Pld (Th ) approximates the exact
flux in L2 with the rate O(hk ). Hence, the choice l := k − 1 is optimal from this viewpoint.
Choosing l := k leads to a (slightly) more precise recovered flux since (52.16) is satisfied up to
l = k. Moreover, if g is smooth, i.e., g ∈ H l (Th ), the data oscillation term in (52.19) converges like
O(hl+2 ), which for l = k − 1 and for l = k, respectively, is one order or two orders faster than the
approximation error p − ph .
Remark 52.11 (Literature). The proof of (52.18) essentially follows Ern and Vohralı́k [190] (up
to minor variations). A different proof of (52.19) is devised in Braess et al. [77] (see also Ern and
Vohralı́k [191]) in dimension two with uniform diffusion, allowing one to prove that the constant c
is independent of the polynomial degree k. The proof in dimension three can be found in Ern and
Vohralı́k [192].
52.2.3 Application to a posteriori error analysis

An important application of local flux recovery is the a posteriori error analysis of H 1 -conforming
finite elements. Recall from Chapter 34 that a posteriori error estimates provide two-sided, fully
computable bounds on the approximation error (p−ph ). We continue to denote the primal variable
by p and the source term by g instead of using u and f as in Chapter 34.
Lemma 52.12 (Two-sided bound). Let p ∈ H01 (D) solve −∇·(d∇p) = g, and let ph ∈ H01 (D)
be its H 1 -conforming approximation. Set σh := −d∇ph . We have the following upper and lower
bounds on k∇(p − ph )kL2 (D) :
λ♭ k∇(p − ph )kL2 (D) ≤ inf kσ ∗ − σh kL2 (D)

σ ∗ ∈H(div;D),∇·σ ∗ =g
≤ λ♯ k∇(p − ph )kL2 (D) . (52.21)
Proof. Recall that Lemma 34.3 gives λ♭ k∇(p−ph )kL2 (D) ≤ kρ(ph )kH −1 (D) with the residual defined
s.t. hρ(ph ), ϕi := (g, ϕ)L2 (D) − (d∇ph , ∇ϕ)L2 (D) and kϕkH01 (D) := |ϕ|H 1 (D) for all ϕ ∈ H01 (D). For
all σ ∗ ∈ H(div; D) such that ∇·σ ∗ = g, we then have
hρ(ph ), ϕi = (g, ϕ)L2 (D) + (σh , ∇ϕ)L2 (D)

= (∇·σ ∗ , ϕ)L2 (D) + (σh , ∇ϕ)L2 (D) = (σh − σ ∗ , ∇ϕ)L2 (D) .
Hence, kρ(ph )kH −1 (D) ≤ kσ ∗ −σh kL2 (D) , and the first bound in (52.21) follows since σ ∗ is arbitrary.
For the second bound, it suffices to pick σ ∗ := −d∇p, which is in H(div; D) and its weak divergence
is equal to g.
Solving the infinite-dimensional constrained minimization problem in (52.21) is unfeasible. Let
us consider instead the milder constraint
(∇·σ ∗ , 1)L2 (K) = (g, 1)L2 (K) , ∀K ∈ Th . (52.22)
Lemma 52.13 (Flux-equilibrated upper bound). The following holds true for all σ ∗ ∈
H(div; D) satisfying (52.22):
! 12
X 2
λ♭ k∇(p − ph )kL2 (D) ≤ kσ ∗ − σh kL2 (K) + ηosc,K , (52.23)
K∈Th
1
with the data oscillation term ηosc,K := π hK kg − ∇·σ ∗ kL2 (K) .
P
Proof. Let ϕ ∈ H01 (D). Owing to (52.22), we infer that (g −∇·σ ∗ , ϕ)L2 (D) = K∈Th (g −∇·σ ∗ , ϕ−
cK )L2 (D) , and we choose the constant cK equal to the mean value of ϕ in K. Recalling that
the mesh cells are convex sets, the Poincaré–Steklov inequality (12.13) gives kϕ − cK kL2 (K) ≤
1
π hK k∇ϕkL (K) . We now adapt the proof of Lemma 52.12 to infer that
2
X
hρ(ph ), ϕi ≤ (kσ ∗ − σh kL2 (K) + ηosc,K )k∇ϕkL2 (K) .
K∈Th
We conclude by using the Cauchy–Schwarz inequality.
An ideal candidate for σ ∗ is the locally post-processed flux σh∗ introduced in Theorem 52.6 and
satisfying (52.16) for some l ≥ 0, which is a (possibly) higher-order version of (52.22). Moreover,
Lemma 52.9 shows that kσh∗ −σh kL2 (K) is also a local lower bound on the error, up to the oscillation
term ηosc,K . Lemma 52.13 is actually valid for all ph ∈ H01 (D). That ph solves a discrete problem
is exploited in the actual construction of σh∗ by means of the Galerkin orthogonality property on
the hat basis functions.
Remark 52.14 (Comparison). The constants in Lemma 52.13 are simpler to estimate than those
in Corollary 34.14. Indeed, the leading term in (52.23) has constant 1, and the data oscillation
term only depends on the constant from the Poincaré–Steklov inequality in mesh cells (as opposed
to the vertex-based stars which have a more complex geometry). However, equilibrated-flux a
posteriori error estimates depend on data oscillations both for the lower and the upper bounds,
i.e., not just for the lower bound as in §34.3.
Remark 52.15 (Literature). Equilibrated-flux a posteriori error estimation for H 1 -conforming

finite elements has a long history. Invoking H(div)-fluxes leads to guaranteed upper bounds on the
error, as shown by Prager and Synge [327], Hlaváček et al. [245] (see also Exercise 52.6). Building
the flux by means of a local equilibration procedure on finite element stars leads in turn to local
efficiency, i.e., to local lower bounds on the error; see Ladevèze and Leguillon [273], Ainsworth
and Oden [7], Parés et al. [322]. Unfortunately, the estimators proposed in these references are
not computable since they require solving an infinite-dimensional problem locally. Inexpensive
local flux equilibration where local finite-dimensional problems are solved on finite element stars
are devised in Destuynder and Métivet [163], Braess and Schöberl [76]. The idea of working on
stars for a posteriori error analysis can be traced back to Babuška and Miller [36], where infinite-
dimensional Dirichlet problems are posed on stars. Finite-dimensional Dirichlet problems inspired
by Carstensen and Funken [107] are considered in Morin et al. [306].
Exercises
Exercise 52.1 (Hybridization). Consider the discrete problem (52.4). (i) Let Q e h := Qh ×Λh
e hy e ′ e
and Bh : Vh → Qh s.t. hBh (τh ), (qh , µh )iQe′ ,Qeh := bh (τh , qh ) + ch (τh , µh ) for all τh ∈ Vhhy and
h
e h . Prove that B
(qh , µh ) ∈ Q e ∗ is injective. (Hint : integrate by parts and use the degrees of freedom
h
of the RT k,d element.) (ii) Prove that (52.4) admits a unique solution.
Exercise 52.2 (Crouzeix–Raviart). Assume that d|K and g|K are constant over each mesh
cr
cell K ∈ Th . Let ∇h denote the broken gradient (see Definition 36.3). Let P1,0 (Th ) be the
nonconforming Crouzeix–Raviart finite element space with homogeneous Dirichlet conditions (see
R R
h ∈ P1,0 (Th ) solve D (d∇h ph )·∇h qh dx = D gqh dx for all qh ∈ P1,0 (Th ). Let
(36.8)) and let pcr cr cr cr cr cr cr
xK be the barycenter of K for all K ∈ Th . Define
σh|K := −(d∇pcr
h )|K + d
−1
g|K (x − xK )|K ,
ph|K := pcrh (xK ) + d
−2
|K|−1 g|K (d−1 (x − xK ), x − xK )L2 (K) .
R
(i) Prove that σh ∈ P0d (Th ). (Hint : compute R Fcr [[σh ]]·nF ϕcr cr
F ds with ϕF the Crouzeix–Raviart
basis function attached to F .) (ii) Prove that D (qh ∇·τh + ∇h qh ·τh ) dx = 0 for all qhcr ∈ P1,0
cr cr
(Th )
d
and all τh ∈ P0 (Th ). (iii) Prove that the pair (σh , ph ) solves (51.21) for k := 0 and f := 0. (Hint :
any function τh ∈ P0d (Th ) is such that τh|K = τK + d−1 (∇·τh )|K (x − xK ), where τK is the mean
value of τh on K.)
Exercise 52.3 (Post-processed potential). Let k ≥ 0. Consider the simplicial Raviart–Thomas
element RT k,d . Assume that it is possible to find a polynomial space Mk,k′ so that for all m ∈ Mk,k′ ,
ΠQK (m) = ΠΛ∂K (m) = 0 implies that m = 0 for all K ∈ Th . Prove that (∇m, τ )L2 (K) = 0 for all
τ ∈ RT k,d implies that m = 0. (Hint : integrate by parts and use the degrees of freedom in RT k,d .)
Let now mnc h be the post-processed potential from the dual mixed formulation (52.2). Show that
2 (K) ≤ ckd
−1
k∇mnch k L σh − f kL2 (K) for all K ∈ Th . (Hint : use norm equivalence on the reference
element, then (52.13); see also Vohralı́k [383, Lem. 5.4].)
Exercise 52.4 (Bound (52.19)). Prove (52.19). (Hint : use Theorem 34.19.)
Exercise 52.5 (Inverse inequality). Prove (52.20). (Hint : consider the dual mixed formula-
tion of (52.17) and introduce the post-processed variable mnc z , use (52.13), accept as a fact that
kmnc
z k L 2 (D ) ≤ chD k∇h m
z z
nc
z k L 2 (D ) , and bound traces of m
z
nc
z using Lemma 12.15.)
Exercise 52.6 (Prager–Synge equality). Let u ∈ H01 (D) be such that −∆u = f in L2 (D).
Let uh ∈ H01 (D), and let σ ∗ ∈ H(div; D) be such that ∇·σ ∗ = f . Prove that k∇(u − uh )k2L2 (D) +
k∇u + σ ∗ k2L2 (D) = k∇uh + σ ∗ k2L2 (D) . (Hint : compute (∇(u − uh ), ∇u + σ ∗ )L2 (D) .)
Chapter 53
Stokes equations: Basic ideas
The Stokes equations constitute the basic linear model for incompressible fluid mechanics. We
first derive a weak formulation of the Stokes equations and establish its well-posedness. The
approximation is then realized by means of mixed finite elements, that is, we consider a pair of
finite elements, where the first component of the pair is used to approximate the velocity and the
second component is used to approximate the pressure. Following the ideas of Chapter 50, the
finite element pair is said to be stable whenever the discrete velocity and the discrete pressure
spaces satisfy an inf-sup condition. In this chapter, we list some classical unstable pairs. Examples
of stable pairs are reviewed in the following two chapters.
53.1 Incompressible fluid mechanics

Let D be a Lipschitz domain in Rd . We are interested in modeling the behavior of incompressible
fluid flows in D in the time-independent Stokes regime, i.e., the inertial forces are assumed to be
negligible. Given a vector-valued field f : D → Rd (the body force acting on the fluid) and a
scalar-valued field g : D → R (the mass production rate), the Stokes problem consists of seeking
the velocity field u : D → Rd and the pressure field p : D → R such that the following balance
equations hold true:
− ∇·s(u) + ∇p = f in D, (53.1a)
∇·u = g in D, (53.1b)
u|∂Dd = ad , s(u)|∂D n
n − p|∂Dn n = an on ∂D. (53.1c)
The equations (53.1a)-(53.1b) express, respectively, the balance of momentum and mass. The
second-order tensor s(u) in (53.1a) is the viscous stress tensor. Notice that we abuse the notation
in (53.1a) since we should write ∇·(s(u)) instead of ∇·s(u). As for linear elasticity (see §42.1),
the principle of conservation of angular momentum implies that s(u) is symmetric and, assuming
the fluid to be Newtonian, Galilean invariance implies that
1
s(u) = 2µe(u) + λ(∇·u)I, e(u) := (∇u + (∇u)T ), (53.2)
2
where I is the d×d identity tensor. The quantity e(u) is called (linearized) strain rate tensor,
and the constants µ > 0, λ ≥ 0 are the dynamic and bulk viscosities, respectively. In (53.1c),
338 Chapter 53. Stokes equations: Basic ideas
the subsets ∂Dd , ∂Dn form a partition of the boundary ∂D, and we assume for simplicity that
|∂Dd | > 0. The boundary data are the prescribed velocity ad on ∂Dd (Dirichlet condition) and
the prescribed normal force an on ∂Dn (Neumann condition).
Remark 53.1 (Total stress tensor). After introducing the total stress tensor r(u, p) := s(u) −
pI, one can rewrite the momentum balance equation (53.1a) in the form −∇·r(u, p) = f , and the
Neumann condition on ∂Dn as r(u, p)|∂Dn n = an .
Remark 53.2 (Incompressibility). The field u is said to be incompressible, or divergence-free,

if ∇·u = g = 0. In the incompressible regime, (53.2) simplifies to s(u) = 2µe(u).
Remark 53.3 (Laplacian/Cauchy–Navier form). When g = 0 and the dynamic viscosity is

constant, the momentum equation can be simplified by observing that ∇·((∇u)T ) = ∇(∇·u) =
0. The momentum equation can then be rewritten in the Laplacian (or Cauchy–Navier ) form
−µ∆u + ∇p = f , and the Neumann boundary condition becomes µ∂n u|∂Dn − p|∂Dn n = an .
Remark 53.4 (Pressure R constant).

R When ∂D = ∂Dd , the data fields g and ad must satisfy the
compatibility condition D g dx = ∂D ad ·n ds, and the pressure is determined
R up to an additive
constant. This indetermination is usually removed by assuming that D p dx = 0.
Remark 53.5 (λ = 0). Since ∇·(λ(∇·u)I) = ∇(λ∇·u), we can redefine the pressure and the
viscous stress tensor by setting p′ := p − λ∇·u and s′ (u) := 2µe(u). Then the momentum balance
equation (53.1a) becomes −∇·s′ (u) + ∇p′ = f . We adopt this change of variable in what follows,
i.e., we assume that s(u) := 2µe(u) from now on.
Remark 53.6 (Homogeneous Dirichlet condition). Let us assume that there is a function ud
(smooth enough) s.t. (ud )|∂Dd = ad . Then we can make the change of variable u′ := u−ud so that
u′ satisfies the homogeneous boundary condition u′|∂Dd = 0. Upon denoting f ′ := f + ∇·(s(ud )),
g ′ := g − ∇·ud , and inserting the definition u = u′ + ud into (53.1), one observes that the pair
(u′ , p) solves a Stokes problem with homogeneous Dirichlet data and with source terms f ′ and g ′ .
From now on, we abuse the notation and use the symbols u, f , g instead of u′ , f ′ , g ′ . This is
equivalent to assuming that ad = 0.
53.2 Weak formulation and well-posedness

In this section, we present a weak formulation of the Stokes equations and we establish its well-
posedness using the Babuška–Brezzi theorem (Theorem 49.13).
53.2.1 Weak formulation

Let w be a sufficiently smooth Rd -valued test function. Since the velocity u vanishes on ∂Dd , we
only consider test functions w that vanish on ∂Dd . Multiplying (53.1a) by w and integrating over
D gives Z Z Z
− (∇·s(u))·w dx + ∇p·w dx = f ·w dx.
D D D
Integrating by parts the term involving the viscous stress tensor, we obtain
Z Z Z
− (∇·s(u))·w dx = s(u):∇w dx − (s(u)n)·w ds,
D D ∂Dn
where n := (n1 , . . . , nd )T is the outward unit normal to D. The boundary integral over ∂Dd is zero
since w vanishes
R on ∂Dd . The symmetry R of s(u) Rimplies that s(u):∇w = s(u):e(w). Similarly,
the term D ∇p·w dx is equal to − D p∇·w dx + ∂Dn pn·w ds. Combining the above equations
and using the Neumann boundary condition s(u)|∂Dn n − p|∂Dn n = an , the weak form of the
momentum equation is
Z Z Z
(s(u):e(w) − p∇·w) dx = f ·w dx + an ·w ds.
D D ∂Dn
The three integrals are well defined if p ∈ L2 (D), f ∈ L2 (D), an ∈ L2 (∂Dn ), and if u, w are in
the space
Vd (D) := {v ∈ H 1 (D) | γ g (v)|∂Dd = 0}, (53.3)
1
with the Rd -valued trace operator γ g : H 1 (D) → H 2 (∂D) acting componentwise as the scalar-
1
valued trace operator γ g : H 1 (D) → H 2 (∂D). We equip the space Vd with the norm kvkVd :=
|v|H 1 (D) = k∇vkL2 (D) . Since |∂Dd | > 0, we infer from the Poincaré–Steklov inequality (42.9) that
there is a constant C̃ps > 0 s.t. C̃ps kvkL2 (D) ≤ ℓD k∇vkL2 (D) for all v ∈ Vd (recall that ℓD is a
length scale associated with D, e.g., ℓD := diam(D)). This argument shows that k·kVd is a norm
on Vd , equivalent to the k·kH 1 (D) -norm.
A weak formulation of the mass conservation (53.1b) is obtained as above by testing the equa-
tion against a sufficiently smooth scalar-valued function q. No integration by parts needs to be
performed, and we simply write Z Z
q∇·u dx = gq dx.
D D
The left-hand side is wellR defined provided q ∈ L2 (D) and Ru is in Vd . Note

R that if ∂D = ∂Dd , the
compatibility condition D g dx = 0 implies the equality D ∇·u dx = D g dx, meaning that the
mass conservation equation need not be tested against constant functions. In this particular case,
the test functions q must be restricted to be of zero mean over D. This motivates the following
definition: (
L2 (D) if ∂D 6= ∂Dd ,
Q := 2 2
R (53.4)
L∗ (D) := {q ∈ L (D) | D q dx = 0} if ∂D = ∂Dd .
We equip the space Q with the L2 -norm. Let us define the bilinear forms
Z Z
a(v, w) := s(v):e(w) dx, b(w, q) := − q∇·w dx, (53.5)
D D
R
R Vd ×Vd and Vd ×Q, respectively.
on R We also define the linear forms F (w) := D f ·w dx +
∂Dn n
a ·w ds, G(q) := − D gq dx on Vd and Q, respectively. Assuming enough smoothness on
f , an , and g, it is reasonable to expect that F ∈ L(Vd ; R) and G ∈ L(Q; R). We obtain the
following weak formulation:

 Find u ∈ Vd and p ∈ Q such that

a(u, w) + b(w, p) = F (w), ∀w ∈ Vd , (53.6)


b(u, q) = G(q), ∀q ∈ Q.
Proposition 53.7 (Weak solution). Assume f ∈ L2 (D), g ∈ Q, and an ∈ L2 (∂Dn ). Any weak
solution (u, p) to (53.6) satisfies (53.1a)-(53.1b) a.e. in D and satisfies the boundary condition
(53.1c) a.e. on ∂D.
Proof. Let us set r(u, p) := s(u) − pI ∈ L2 (D). Testing the momentum equation in (53.6) against
an arbitrary function w ∈ C0∞ (D), we infer that r(u, p) has a weak divergence in L2 (D) equal to
−f . Since ∇·r(u, p) = ∇·s(u) − ∇p, we infer that (53.1a) is satisfied a.e. in D. Testing the mass
equation in (53.6) against an arbitrary function q R∈ C0∞ (D), we infer that (53.1b) is satisfied a.e.
in D (if ∂D = ∂Dd , the compatibility condition D g dx = 0 implies that b(u, q) = G(q) for all
q ∈ L2 (D)). The Dirichlet boundary condition u|∂Dd = 0 is a natural consequence of the trace
theorem (Theorem 3.10) and u being in Vd . To derive the Neumann condition, we proceed as in
§31.3.3. Since ∇·r(u, p) = −f ∈ L2 (D), we have r(u, p) ∈ H(div; D) (i.e., each row of r(u, p) is in
H(div; D)). Owing to Theorem 4.15, we infer that r(u, p)n ∈ H − 2 (∂D). As a result, we have
1
Z
hr(u, p)n, γ g (w)i∂D = (r(u, p):∇w + (∇·r(u, p))·w) dx
ZD Z
= (s(u):e(w) − p∇·w − f ·w) dx = an ·γ g (w) ds, ∀w ∈ Vd ,
D ∂Dn
which implies that the Neumann condition r(u, p)n = an is satisfied in H f 12 (∂Dn ) :=
f 21 (∂Dn )′ , with H
1 1
e ∈ H 2 (∂D)} (recall that v
{v ∈ H 2 (∂Dn ) | v e is the zero extension of v to ∂D). Actually, the
Neumann condition is satisfied a.e. on ∂Dn since we assumed an ∈ L2 (∂Dn ).
Remark 53.8 (Neumann data). The above proof shows that it is possible to take more generally
f 21 (∂Dn )′ .
an ∈ H
One readily sees that the bilinear form a(v, w) := (s(v), e(w))L2 (D) defined in (53.5) is coercive and
bounded on Vd ×Vd . The coercivity of a has been established in Theorem 42.11 as a consequence
of Korn’s inequalities. In particular, there is Ck > 0 s.t. ke(v)kL2 (D) ≥ Ck |v|H 1 (D) for all v ∈ Vd ,
and this implies that (see (42.15) with ρmin := 2µ in the present setting)
a(v, v) ≥ 2µCk2 |v|2H 1 (D) , ∀v ∈ Vd . (53.7)
Moreover, the Cauchy–Schwarz inequality and the bound ke(v)kL2 (D) ≤ |v|H 1 (D) show that the
boundedness constant of the bilinear form a satisfies kak ≤ 2µ. Hence, the key argument for the
well-posedness of the Stokes problem is the surjectivity of the divergence operator ∇· : Vd → Q.
This result is a bit more subtle than Lemma 51.2 since Vd is a smaller space than H(div; D).
Lemma 53.9 (∇· is surjective). Let D be a Lipschitz domain in Rd . (i) Case ∂D = ∂Dd .
∇· : H01 (D) → L2∗ (D) is surjective. (ii) Case ∂D 6= ∂Dd . Consider the partition ∂D = ∂Dd ∪ ∂Dn
with |∂Dd | > 0. Assume that |∂Dn | > 0 and that there exists a subset O of ∂Dn with |O| > 0 and
1
n|O ∈ H 2 (O). Then the operator ∇· : X := {v ∈ Vd | γ g (v)|∂Dn ×n = 0} → L2 (D) is surjective.
(iii) In all the cases, identifying Q′ with Q we have
R
| D q∇·v dx|
inf sup := βD > 0. (53.8)
q∈Q v∈Vd kqkL2 (D) |v|H 1 (D)
Proof. (i) We refer the reader to Girault and Raviart [217, pp. 18-26] for a proof of the surjectivity
of ∇· : H01 (D) → L2∗ (D), (see also Exercise 53.1 if D is a smooth domain). (ii) Let us now consider
the second case.RLet q be in L2 (D). Let ρ be a smooth nonnegative function compactly supported
in O such that O ρ ds > 0 (this is Rpossible since R |O| > 0). Let g := cρn be a vector field in O,
where the constant c is chosen s.t. O g·n ds = D q dx. Let ge be the zero extension of g to ∂D.
1
Since n|O ∈ H 2 (O), we have ρn ∈ H f 2 (O). Hence, ge is in H 2 (∂D) so that it is possible to find
1 1
a function w in H (D) s.t. γ (w) = ge on ∂D. We have γ g (w)|∂Dd = 0 and γ g (w)|∂D n ×n = 0,

1 g
i.e., w ∈ X. NowRlet q0 := ∇·w − q. The above definitions and the divergence formula imply that
q0 ∈ L2 (D) and D q0 dx = 0. Hence, q0 is in L2∗ (D). Since ∇· : H01 (D) → L2∗ (D) is surjective,
there is w0 ∈ H01 (D) such that ∇·w0 = −q0 . Thus, for all q in L2 (D) the function w + w0 is in
X with ∇·(w + w0 ) = q, that is, ∇·X → L2 (D) is surjective. This also implies that ∇· : Vd → Q
is surjective. (iii) The inf-sup condition (53.8) follows from the surjectivity of ∇· : Vd → Q and
Lemma C.40.
′
′
Remark 53.10 (Inf-sup condition in W 1,p -Lp ). Let p ∈ (1, ∞) andR let p ∈ (1, ∞) be s.t.
1 1 1,p p p
p + p′ = 1. Then the operator ∇· : W0 (D) → L∗ (D)
:= {q ∈ L (D) | D q dx = 0} is surjective
′
(see Auscher et al. [30, Lem. 10]), that is, identifying (Lp∗ (D))′ with Lp∗ (D), we have
R
| D q∇·v dx|
inf′ sup := βD,p > 0. (53.9)
q∈Lp (D) kqkLp (D) |v|W
1,p ′ 1,p (D)
∗ (D) v∈W 0
The assumption that D is Lipschitz can be weakened. For instance, the inf-sup condition (53.9)
holds true also if D is a bounded open set in Rd and if D is star-shaped with respect to an open
ball B ⊂ D, i.e., for all x ∈ D and z ∈ B, the segment joining x and z is contained in D; see
Bogovskiı̆ [66], Galdi [210, Lem. 3.1, Chap. III], Durán and Muschietti [181], Durán et al. [180],
Solonnikov [349, Prop. 2.1], Costabel and McIntosh [146].
R
Let B : Vd → Q′ be s.t. hB(v), qiQ′ ,Q := b(v, q) = − D q(∇·v) dx. Identifying Q and Q′ , we
have B(v) = −∇·v, and ker(B) := {v ∈ Vd | ∇·v = 0}.
Theorem 53.11 (Well-posedness). (i) The weak formulation (53.6) of the Stokes problem is
well-posed. (ii) There is c such that for all f ∈ L2 (D), all g ∈ Q, and all an ∈ L2 (∂Dn ),
1

2µ|u|H 1 (D) + kpkL2 (D) ≤ c ℓD kf kL2 (D) + µkgkL2(D) + ℓD2
kan kL2 (∂Dn ) .
Proof. We apply the Babuška–Brezzi theorem (Theorem 49.13). The inf-sup condition (49.37) on
the bilinear form b follows from Lemma 53.9. The two conditions in (49.36) are satisfied owing
to the coercivity of the bilinear form a on Vd (see (53.7)). Finally, the stability estimate follows
from (49.38).
One can formulate a more precise stability result on the product space Y := Vd ×Q equipped
with the norm k(v, q)k2Y := µ|v|2H 1 (D) + µ−1 kpk2L2(D) , and the bilinear form t((v, q), (w, r)) :=
a(v, w) + b(w, q) − b(v, r) on Y ×Y.
Lemma 53.12 (Inf-sup condition). The following holds true:
|t((v, q), (w, r))|
inf sup =: γ > 0, (53.10)
(v,q)∈Y (w,r)∈Y k(v, q)kY k(w, r)kY
where γ is uniform w.r.t. µ > 0.

|t((v,q),(w,r))|
Proof. Let (v, q) ∈ Y and let us set S := sup(w,r)∈Y k(w,r)kY . Owing to (53.7), we have
2µCk2 |v|2H 1 (D) ≤ a(v, v) = t((v, q), (v, q)) ≤ Sk(v, q)kY . (53.11)
Moreover, owing to Lemma 53.9, there is wq ∈ Vd s.t.
∇·wq = −µ−1 q, |wq |H 1 (D) ≤ (βD µ)−1 kqkL2 (D) .

We obtain
µ−1 kqk2L2 (D) = −(q, ∇·wq ) = −t((v, q), (wq , 0)) + a(v, wq )
1 1 1
≤ Sµ 2 |wq |H 1 (D) + 2µ 2 |v|H 1 (D) µ 2 |wq |H 1 (D)
1 1 1
≤ c′ (S + S 2 k(v, q)kY2 )µ 2 |wq |H 1 (D) ,
where we used that |a(v, w)| ≤ 2µ|v|H 1 (D) |w|H 1 (D) and then (53.11). Using the bound on
|wq |H 1 (D) and Young’s inequality leads to
µ−1 kqk2L2 (D) ≤ c (S2 + Sk(v, q)kY ).
We can now combine this bound with (53.11) to infer that
k(v, q)k2Y ≤ c (S2 + Sk(v, q)kY ).
Applying one more time Young’s inequality yields k(v, q)kY ≤ cS.
Remark 53.13 (Helmholtz decomposition). Letting H∗1 (D) := H 1 (D) ∩ L2∗ (D) and H :=
{v ∈ L2 (D) | ∇·v = 0, v|∂D ·n = 0}, the following L2 -orthogonal Helmholtz decomposition holds
true: L2 (D) = H⊕∇(H∗1 (D)) (see Lemma 74.1). The L2 -orthogonal projection PH : L2 (D) → H
resulting from this decomposition is often called Leray projection. Let (u, p) solve (53.6). Assume
for simplicity that the homogeneous Dirichlet condition u|∂D = 0 is enforced over the whole
boundary and assume that g = 0. Since u is divergence-free and vanishes at the boundary, we
have (f , u)L2 (D) = (PH (f ), u)L2 (D) . Then taking w := u in (53.6) and invoking the coercivity of a
shows that 2µCk2 |u|2H 1 (D) ≤ a(u, u) = (PH (f ), u)L2 (D) . Owing to the Cauchy–Schwarz inequality
and the Poincaré–Steklov inequality, we get
−1 −2
2µ|u|H 1 (D) ≤ Cps Ck ℓD kPH (f )kL2 (D) .
This a priori estimate on the velocity is sharper than the one from Theorem 53.11 since kPH (f )kL2 (D)
appears on the right-hand side instead of kf kL2 (D) . One should bear in mind that, even if
p ∈ H∗1 (D), the fields −∇·s(u) and PH (f ) are generally different since the normal component
of ∇·s(u) at ∂D is generally nonzero.
53.2.3 Regularity pickup

Regularity properties for the Stokes problems can be established when µ and λ are both constant
(or smooth) and |∂Dn | = 0. For instance, if ∂D is of class C ∞ , for all s > 0 there is c, depending
on D and s, such that
µℓ−1
D kukH 1+s (D) + kpkH s (D) ≤ c (ℓD kf kH s−1 (D) + µkgkH s (D) ). (53.12)
There is an upper limit on s when D is not smooth. For instance, let D be a two-dimensional
convex polygon. Let ρ : D → R be the distance to the closest vertex of D. It is shown in Kellogg
and Osborn [266, Thm. 2] that there is a constant c that depends only on D such that

µ|u|H 2 (D) + |p|H 1 (D) ≤ c kf kL2 (D) + µℓ−1
D (|g|H 1 (D) + kρ
−1
gkL2 (D) ) . (53.13)
The situation is a bit more complicated in dimension three. We refer to Dauge [153] for an overview
of the problem. Assuming that g = 0, it is shown in [153, p. 75] that (53.12) holds true in the
following situations: (i) For all s ≤ 1 if D is a convex polyhedron; (ii) For all s < 23 if D is any
convex domain with wedge angles ≤ 32 π; (iii) For all s < 21 if D has a piecewise smooth boundary,
and its faces meet two by two or three by three with independent normal vectors at the meeting
points.
53.3 Conforming approximation

In the rest of this chapter, we assume that D is a polyhedron in Rd and (Th )h∈H is a shape-regular
sequence of matching meshes so that each mesh covers D exactly. We also assume that ∂Dd is
a union of mesh faces. Let (Vhd ⊂ Vd )h∈H and (Qh ⊂ Q)h∈H be sequences of finite-dimensional
spaces built using (Th )h∈H . Notice that the inclusion Vhd ⊂ Vd means that the homogeneous
Dirichlet condition on the velocity is strongly enforced on ∂Dd . The discrete counterpart of the
problem (53.6) is as follows:


 Find uh ∈ Vhd and ph ∈ Qh such that
a(uh , vh ) + b(vh , ph ) = F (vh ), ∀vh ∈ Vhd , (53.14)

 b(u , q ) = G(q ),
h h h ∀qh ∈ Q h .
Since Vhd is Vd -conforming, the discrete formulation inherits the coercivity of a. Unfortunately,
there is no reason a priori for the discrete formulation to inherit the surjectivity of the divergence
operator established in Lemma 53.9. Verifying this condition is the crucial step in devising stable
mixed finite elements for the Stokes problem.
Proposition 53.14 (Well-posedness). The discrete problem (53.14) is well-posed if and only if
the following inf-sup condition holds true:
R
| D qh ∇·vh dx|
inf sup =: βh > 0. (53.15)
qh ∈Qh vh ∈Vhd kqh kL2 (D) |vh |H 1 (D)
Proof. Apply Proposition 50.1.

We henceforth say that the inf-sup condition (53.15) holds uniformly w.r.t. h ∈ H if inf h∈H βh =:
β0 > 0.
Definition 53.15 (Stable/unstable pair). We say that a pair of finite elements used to approx-
imate the velocity and the pressure is stable if the inf-sup condition (53.15) holds true uniformly
w.r.t. h ∈ H, and we say that it is unstable otherwise.
′
Remark 53.16 (Inf-sup condition in W 1,p -Lp ). Let p ∈ (1, ∞) and let p′ ∈ (1, ∞) be s.t.
1 1
p + p′ = 1. As in Remark 53.10, a more general variant of the inf-sup condition (53.15) is
R
| D qh ∇·vh dx|
inf sup =: βh > 0. (53.16)
qh ∈Qh vh ∈Vhd kqh kLp′ (D) |vh |W 1,p (D)
We will see in the next chapters that many stable finite element pairs for the Stokes equations
satisfy this more general inf-sup condition.
Let us define the discrete operator Bh : Vhd → Q′h s.t. hBh (vh ), qh iQ′h ,Qh := b(vh , qh ) =
R
− D qh ∇·vh dx for all (vh , qh ) ∈ Vhd ×Qh . We have
(53.15) ⇐⇒ Bh is surjective, (53.17a)

ker(Bh ) = {vh ∈ Vhd | (qh , ∇·vh )L2 (D) = 0, ∀qh ∈ Qh }. (53.17b)
The operator Bh : Vhd → Q′h is the discrete counterpart of the divergence operator B : Vd → Q′
introduced just above Theorem 53.11. We observe that the inf-sup condition (53.15) is equivalent
to asserting the surjectivity of Bh . Moreover, assuming for simplicity that g = 0 in the mass
conservation equation, the discrete Stokes problem (53.14) produces a velocity field uh ∈ ker(Bh ).
One then says that the discrete velocity field is weakly divergence-free. However, ker(Bh ) may not
be a subspace of ker(B), i.e., the discrete velocity field uh is not necessarily strongly (or pointwise)
divergence-free.
Several techniques are available to prove the inf-sup condition (53.15), and we refer the reader
to the next two chapters for various examples. Recall in particular that (53.15) is equivalent to
the existence of a Fortin operator Πh ∈ L(Vd ; Vhd ) s.t. b(Πh (v) − v, qh ) = 0 for all qh ∈ Qh (see
Lemma 26.9).
Theorem 53.17 (Error estimate). Let (u, p) solve (53.6). Assume (53.15) and let (uh , ph )
solve (53.14). Then we have
|u − uh |H 1 (D) ≤ c1h inf |u − vh |H 1 (D) + c2h inf kp − qh kL2 (D) ,

vh ∈Vhd qh ∈Qh
kp − ph kL2 (D) ≤ c3h inf |u − vh |H 1 (D) + c4h inf kp − qh kL2 (D) ,

vh ∈Vhd qh ∈Qh
where c1h := (1 + kak

α )(1 + kΠh kL(Vd ;Vhd ) ) for any Fortin operator Πh ∈ L(Vd ; Vhd ), c2h
:= 0 if
ker(Bh ) ⊂ ker(B) and c2h := kbkα otherwise, c 3h := c kak
1h βh , and c 4h := 1 + kbk
βh + c kak
2h βh . Here,
2
α ≥ 2µCk is the coercivity constant of the bilinear form a on Vd ×Vd , kak ≤ 2µ its norm, and
kbk ≤ 1 the norm of the bilinear form b on Vd ×Q.
Proof. This is a direct application of Corollary 50.5.
Remark 53.18 (βh vs. β0 ). The estimates from Theorem 53.17 show that it is important that
the inf-sup condition (53.15) be satisfied uniformly w.r.t. h ∈ H. Indeed, the factor β1h appears
in the coefficients c3h and c4h in the pressure error bound, and a factor β1h may appear in the
kbk
constant c1h affecting both error bounds if kΠh kL(Vd ;Vhd ) ∼ βh for every Fortin operator.
We say that the pair (ξ(r), φ(r)) ∈ Vd ×Q is theRsolution to the adjoint problem of (53.6) with
source term r ∈ L2 (D) if a(v, ξ(r)) + b(v, φ(r)) = D r·v dx for all v ∈ Vd and b(ξ(r), q) = 0 for
all q ∈ Q.
Theorem 53.19 (L2 -velocity error estimate). Let (u, p) solve (53.6). Assume (53.15) and let
(uh , ph ) solve (53.14). Assume that there exist real numbers csmo and s ∈ (0, 1] s.t.
µℓ−1
D kξ(r)kH 1+s (D) + kφ(r)kH s (D) ≤ csmo ℓD krkL2 (D) , ∀r ∈ L2 (D),
and that there is c such that for all h ∈ H, inf vh ∈Vhd |v − vh |H 1 (D) ≤ chs |v|H 1+s (D) for all v ∈
Vd ∩ H 1+s (D) and inf qh ∈Qh kq − qh kL2 (D) ≤ chs |q|H s (D) for all q ∈ Q ∩ H s (D). Then there is c
s.t. for all h ∈ H,

kbk
D inf |u − v | 1
h H (D) + inf kp − q k
h L (D) .
2
vh ∈Vhd kak qh ∈Qh
Proof. Apply Lemma 50.11 or see Exercise 53.3.
Let us give some further insight into the velocity error estimate from Theorem 53.17. For
simplicity, we assume that g = 0. Let us define the projection operator PhS : Vd → ker(Bh ) such
that
a(PhS (v), wh ) = a(v, wh ), ∀(v, wh ) ∈ Vd × ker(Bh ). (53.18)
Lemma 53.20 (Quasi-optimality). Assume (53.15). The following holds true for all v ∈ Vd
and any Fortin operator Πh ∈ L(Vd ; Vhd ):
|v − PhS (v)|H 1 (D) ≤ c̃1h inf |v − vh |H 1 (D) , (53.19)

vh ∈Vhd
kak
with c̃1h := α (1 + kΠh kL(Vd ;Vhd ) ).
Proof. Since the bilinear form a is bounded and coercive, we have
kak
|v − PhS (v)|H 1 (D) ≤ inf |v − vh |H 1 (D) .
α vh ∈ker(Bh )
The assertion then follows from Lemma 50.3 (notice that Πh (u) ∈ ker(Bh ) since ∇·u = g = 0 by
assumption, see Remark 50.4).
Lemma 53.21 (Discrete velocity estimate). Let (u, p) solve (53.6). Assume (53.15) and let
(uh , ph ) solve (53.14). As in Theorem 53.17, set c2h := 0 if ker(Bh ) ⊂ ker(B) and c2h := kbk
α
otherwise. The following holds true:
|uh − PhS (u)|H 1 (D) ≤ c2h inf kp − qh kL2 (D) . (53.20)

qh ∈Qh
Proof. The proof follows a similar, yet simpler, path to that of Lemma 50.2. Since a(uh , wh ) +
b(wh , ph ) = F (wh ) = a(u, wh ) + b(wh , p) = a(PhS (u), wh ) + b(wh , p) for all wh ∈ ker(Bh ) ⊂
Vhd ⊂ Vd , setting eh := uh − PhS (u) ∈ ker(Bh ), we infer that a(eh , wh ) = b(wh , p − ph ) for all
wh ∈ ker(Bh ). Since eh ∈ ker(Bh ), invoking the coercivity of a then yields
α |eh |2H 1 (D) ≤ b(eh , p − ph ).
If ker(Bh ) ⊂ ker(B), then |eh |H 1 (D) = 0 which proves (53.20). Otherwise, we use that eh ∈
ker(Bh ) to write α|eh |2H 1 (D) ≤ b(eh , p − qh ) for all qh ∈ Qh , and invoke the boundedness of b to
prove (53.20).
The bound (53.20) implies that uh = PhS (u) whenever ker(Bh ) ⊂ ker(B). Moreover, in the
general case, combining (53.20) with (53.19) and using the triangle inequality we obtain again the
velocity error estimate from Theorem 53.17 with the slightly sharper constant c̃1h instead of c1h .
Remark 53.22 (Well-balanced scheme). In the particular case where f = ∇φ for some φ ∈
H 1 (D) ∩ L2∗ (D), the solution to the Stokes problem (53.6) is (u, p) = (0, φ). This situation is
encountered with hydrostatic (or curl-free) body forces. One says that the discrete problem (53.14)
is well-balanced w.r.t. hydrostatic body forces if uh = 0 as well. (One also sometimes says that
the discretization is pressure robust.) A well-balanced discretization of the Stokes equations can be
desirable even if f is not curl-free, but has a relatively large curl-free component. In this case, a
discretization that is not well-balanced can lead to a rather poor velocity approximation, even on
meshes that seem rather fine. Lemma 53.21 shows that (53.14) is well-balanced whenever ker(Bh ) ⊂
ker(B). The scheme can be made well-balanced when ker(Bh ) 6⊂ ker(B) by slightly modifying
the discrete momentum equation. Considering Dirichlet conditions over the whole boundary for
simplicity, one introduces a lifting operator L : Vhd → Vd such that L(ker(Bh )) ⊂ ker(B) and then
replaces the first equation in (53.14) by a(uh , wh ) + b(wh , ph ) = (f , L(wh ))L2 (D) for all wh ∈ Vhd .
The lifting operator L must satisfy some consistency conditions to preserve the optimal decay rates
of the error estimate. This idea has been introduced by Linke [283] and explored more thoroughly
by Lederer et al. [279] in the context of mixed finite elements with continuous pressures; see also
John et al. [261] for an overview. Examples of curl-free body forces in fluid mechanics are the
Coriolis force if d = 2, the gravity, and the centrifugal force. Obviously, if f ≈ ∇φ and φ is
explicitly known, one can always make the change of variable p → p − φ to alleviate the above
difficulty if the scheme is not well-balanced.
53.4 Classical examples of unstable pairs

We study in this section three pairs of finite elements that look appealing at first sight, but
that unfortunately do not satisfy the inf-sup condition (53.15). For simplicity, we consider a
homogeneous Dirichlet condition on the velocity over the whole boundary, so that Vd := H01 (D)
and we write Vh0 instead of Vhd for the discrete velocity space. Since the approximation setting
is conforming, we have Vh0 ⊂ H01 (D) in all cases.
Recall that the inf-sup condition (53.15) is not satisfied if and only if Bh∗ : Qh → Vh0 ′
is not
injective (or, once global shape functions have been chosen, the associated matrix does not have
full column rank). In this case, a nonzero pressure field in ker(Bh∗ ) is called spurious pressure mode.
Equivalently, the inf-sup condition is not satisfied if and only if Bh : Vh0 → Q′h is not surjective.
53.4.1 The (Q
Q1 , P0 ) pair: Checkerboard instability
A well-known pair of incompatible finite elements is the (QQ 1 , P0 ) pair obtained when approximat-
ing the velocity with continuous piecewise bilinear polynomials and the pressure with piecewise
constants. This pair produces an instability often called checkerboard instability.
Let us restrict ourselves to the two-dimensional setting and assume that D := (0, 1)2 . We
define a uniform Cartesian mesh on D as follows: Let N be an integer larger than 2. Set h := 1I ,
and for all i, j ∈ {0:I−1}, denote by aij the point with Cartesian coordinates (ih, jh). Let Kij be
the square
S cell whose bottom left node is aij ; see Figure 53.1. The resulting mesh is denoted by
Th := i,j Kij . Consider the following finite element spaces:
Vh0 := {vh ∈ C 0 (D) | ∀Kij ∈ Th , vh ◦ TKij ∈ Q 1,d , vh|∂D = 0}, (53.21a)

Qh := {qh ∈ L2∗ (D) | ∀Kij ∈ Th , qh ◦ TKij ∈ P0,d }. (53.21b)
Recall that for all K ∈ Th , TK : K b → K denotes the geometric mapping; see §8.1. For all ph ∈ Qh ,
set pi+ 12 ,j+ 21 := ph|Kij , and for all vh ∈ Vh0 , denote by (uij , vij ) the values of the two Cartesian
components of vh at the node aij .
To prove that the inf-sup constant R is zero, it is sufficient to prove the existence of a nonzero
pressure field ph ∈ ker(Bh∗ ), i.e., D ph ∇·vh dx = 0 for all vh ∈ Vh0 . Since ph is constant on each
cell, we have
Z Z
ph ∇·vh dx = pi+ 12 ,j+ 12 vh ·n ds
Kij ∂Kij
1
= 2 hpi+ 2 ,j+ 2 (ui+1,j
1 1 + ui+1,j+1 + vi+1,j+1 + vi,j+1
−ui,j − ui,j+1 − vi,j − vi+1,j ) .
Summing over all the cells and rearranging the sum yields
Z X
ph ∇·vh dx = −h2 (ui,j G1,ij (ph ) + vi,j G2,ij (ph )),
D i,j∈{0: N −1}
−1 +1 −1 +1 −1
+1 −1 +1 −1 +1
−1 +1 −1 +1 −1
Kij +1 −1 +1 −1 +1
X
aij
−1 +1 −1 +1 −1
Figure 53.1: (Q
Q1 , P0 ) pair: mesh (left) and spurious pressure mode (right).
where
1
G1,ij (ph ) := 2h (pi+ 2 ,j+ 2
1 1 + pi+ 12 ,j− 12 − pi− 12 ,j+ 12 − pi− 12 ,j− 12 ),
1
G2,ij (ph ) := 2h (pi+ 2 ,j+ 2
1 1 + pi− 12 ,j+ 21 − pi+ 12 ,j− 12 − pi− 12 ,j− 12 ).
R
We infer that D
ph ∇·vh dx = 0 for all vh ∈ Vh0 if and only if
pi+ 21 ,j+ 12 = pi− 21 ,j− 12 and pi− 12 ,j+ 21 = pi+ 12 ,j− 21 .
The solution set of this linear system is a two-dimensional vector space. One dimension is spanned
by the constant field ph = 1, but span{1} must be excluded from the solution set since the
elements in Qh must have a zero mean. The other dimension is spanned by the field whose value is
alternatively +1 and −1 on adjacent cells in a checkerboard pattern, as shown on the right panel
of Figure 53.1. This is a spurious pressure mode, and if N is even, this spurious mode is in Qh
(i.e., it satisfies the zero-mean condition). In this case, the inf-sup condition is not satisfied, i.e.,
the (QQ1 , P0 ) pair is incompatible for the Stokes problem.
Remark 53.23 (Filtering). Since the (Q Q 1 , P0 ) pair is very simple to program, one may be
tempted to cure its deficiencies by restricting the size of Qh . For instance, one could enforce the
pressure to be orthogonal (in the L2 -sense) to the space spanned by the spurious pressure mode.
Unfortunately, this remedy is not strong enough to produce a healthy finite element pair, since it
can be shown that in this case there are positive constants c, c′ s.t. ch ≤ βh ≤ c′ h uniformly w.r.t.
h ∈ H; see Boland and Nicolaides [68] or Girault and Raviart [217, p. 164]. This shows that the
method may not converge since the factor β1h appears in the error bound on the velocity and the
factor β12 appears in the error bound on the pressure (see Theorem 53.17).
h
53.4.2 The (P
P1 , P1 ) pair: Checkerboard-like instability
Because it is very simple to program, the continuous P1 finite element for both the velocity and
the pressure is a natural choice for approximating the Stokes problem. Unfortunately, the (PP 1 , P1 )
pair does not satisfy the inf-sup condition (53.15). To understand the origin of the problem, let us
construct a two-dimensional counterexample in D := (0, 1)2 . Consider a uniform Cartesian mesh
composed of squares of side h and split each square along one diagonal as shown in the left panel
of Figure 53.2. Let Th be the resulting triangulation and let the velocity and the pressure finite
element spaces be
g
Vh0 := P1,0 (Th ), Qh := P1g (Th ) ∩ L2∗ (D), (53.22)
g g
where P1,0 (Th ) is the vector-valued version of the space P1,0 (Th ) defined in §19.4. Let {zn,K }n∈{0: 2}
−1 0 +1 −1 0 +1
+1 −1 0 +1 −1 0
0 +1 −1 0 +1 −1
−1 0 +1 −1 0 +1
+1 −1 0 +1 −1 0
0 +1 −1 0 +1 −1
Figure 53.2: (P
P1 , P1 ) pair: the mesh (left) and one spurious pressure mode (right).
be the three vertices of the mesh cell K ∈ Th . Now consider a pressure field ph such that
P
n∈{0: 2} ph (zn,K ) is zero on each triangle K. An example of such a spurious pressure mode
is shown in the right panel of Figure 53.2. Then we have for all vh ∈ Vh0 ,
Z X Z
ph ∇·vh dx = (∇·vh )|K ph dx,
D K∈Th K
X |K| X
= (∇·vh )|K ph (zn,K ) = 0.
3
K∈Th n∈{0: 2}
R
Hence, ph satisfies D ph ∇·vh dx = 0 for all vh ∈ Vh0 . In other words, the field ph is a spurious
pressure mode, and the inf-sup constant is zero.
53.4.3 The (P
P1 , P0 ) pair: Locking effect
A simple alternative to the (Q P1 , P0 ) pair, i.e., assuming that Th is
Q 1 , P0 ) pair consists of using the (P
composed of simplices, the velocity is approximated with continuous, piecewise linear polynomials
and the pressure with piecewise constants. We observe that ker(Bh ) ⊂ ker(B) in this case since the
divergence of the velocity is piecewise constant. Unfortunately, the (P P1 , P0 ) pair does not satisfy
the inf-sup condition (53.15). Let us produce a two-dimensional counterexample. Assume that
D is a simply connected polygon. Let Nc , Nvi , and Ne∂ denote the number of elements, internal
vertices, and boundary edges in Th , respectively. The Euler relations give Nc = 2Nvi + Ne∂ − 2 (see
Remark 8.13 and Exercise 8.2). Since dim(Qh ) = Nc − 1 and dim(Vh0 ) = 2Nvi , the rank nullity
theorem implies that
dim(ker(Bh∗ )) = dim(Qh ) − dim(im(Bh∗ )) ≥ dim(Qh ) − dim(Vh0 )
= Nc − 1 − 2Nvi = Ne∂ − 3.
Hence, there are at least Ne∂ − 3 spurious pressure modes. This means that the space Qh is far too
rich for Bh to be surjective. Actually, in some cases, it can be shown that Bh is injective, i.e., the
only member of ker(Bh ) is zero. This situation is referred to as locking in the literature.
Remark 53.24 (Comparison with (P P1 , P1 )). Note that the dimension of the pressure finite
P1 , P1 ) pair (where dim(Qh ) = Nv − 1) than for the (P
element space is smaller for the (P P1 , P0 ) pair
(where dim(Qh ) = Nc − 1). Indeed, we have Nc ∼ 2Nv on fine meshes (see Exercise 8.2).
Exercises
Exercise 53.1 (∇· is surjective). Let D ⊂ R2 be a domain of class C 2 . Prove that ∇· : H01 (D) →
L2∗ (D) is continuous and surjective. (Hint : construct v ∈ H01 (D) such that v = ∇q + ∇×ψ, where
q solves a Poisson problem, ψ solves a biharmonic problem, and ∇×ψ := (∂2 ψ, −∂1 ψ)T .)
Exercise 53.2 (de Rham). Let D be a bounded open set in Rd and assume that D is star-shaped
with respect to an open ball B ⊂ D. Prove that the continuous linear forms on W01,p (D) that
′
are zero on ker(∇·) are gradients of functions in Lp∗ (D). (Hint : use Remark 53.10 and the closed
range theorem.)
Exercise 53.3 (L2 -estimate). Prove Theorem 53.19 directly, i.e., without invoking Lemma 50.11.
Exercise 53.4 (Projection). Let (Vh0 , Qh )h∈H be a sequence of pairs of finite element spaces.
Let p ∈ [1, ∞] and let p′ ∈ [1, ∞] be s.t. p1 + p1′ = 1. Let ΠZh : Qh → Zh be an operator, where Zh is
p
a finite-dimensional
R subspace of L (D). Assume that there are β1 , β2 > 0 such thatR for all h ∈ H,
| D qh ∇·vh dx| Z | q ∇·v dx|
supvh ∈Vh0 |vh | 1,p ≥ β1 kqh − Πh (qh )kLp′ (D) for all qh ∈ Qh and supvh ∈Vh0 |vDh |h 1,ph ≥
W (D) W (D)
β2 kqh kLp′ (D) for all qh ∈ Zh . (i) Show that ΠZ

h is bounded uniformly w.r.t. h ∈ H. (ii) Show that
the (Vh0 , Qh ) pair satisfies an inf-sup condition uniformly w.r.t. h ∈ H.
Exercise 53.5 (Spurious mode for the (Q b := [0, 1]2 be the unit square.
Q1 , Q1 ) pair). (i) Let K
i j
R P
Let ab ij := ( 2 , 2 ), for all i, j ∈ {0:2}. Show that the quadrature K b f (b x ≈ i,j wij f (b
x) db aij ),
:= 1
:= 1 b := 1
where wij 36 (3i(2 − i) + 1)(3j(j − 2) + 1) (wij 36 for the four vertices of K, wij 9 for
the four edge midpoints, and wij := 49 at the barycenter of K) b is exact for all f ∈ Q2 . (Hint :
write the Q2 Lagrange shape functions in tensor-product form and use Simpson’s rule in each
direction.) (ii) Consider D := (0, 1)2 and a mesh composed of I×I squares, I ≥ 2. Consider the
points alm := ( 2Il , 2Im
) for all l, m ∈ {0:2I}. Let ph be the continuous, piecewise bilinear function
such that ph (a2k,2n ) := (−1)k+n for all k, n ∈ {0:I}. Show that ph is a spurious pressure mode
for the (QQ1 , Q1 ) pair (continuous velocity and pressure).
Chapter 54
Stokes equations: Stable pairs (I)
This chapter reviews various stable finite element pairs that are suitable to approximate the Stokes
equations, i.e., the discrete velocity space and the discrete pressure space satisfy the inf-sup con-
′
dition (53.15) (or its W 1,p -Lp version (53.16)) uniformly with respect to h ∈ H. We first review
two standard techniques to prove the inf-sup condition, one based on the Fortin operator and one
hinging on a weak control of the pressure gradient. Then we show how these techniques can be
applied to finite element pairs where the discrete pressure space is H 1 -conforming. The two main
examples are the mini element based on the (P P1 -bubble, P1 ) pair and the Taylor–Hood element
based on the (PP2 , P1 ) pair. In the next chapter, we introduce another technique based on macroele-
ments to prove the inf-sup condition and we review stable finite element pairs where the discrete
pressures are discontinuous. We assume in the entire chapter that Dirichlet conditions are enforced
on the velocity over the whole boundary, that D is a polyhedron in Rd , and that (Th )h∈H is a
shape-regular sequence of affine meshes so that each mesh covers D exactly.
54.1 Proving the inf-sup condition

We briefly review two standard techniques to prove the inf-sup condition (53.15): one uses a Fortin
operator and the other uses a weak control on the pressure gradient. Since this section is only
meant to be a short introduction, the reader is referred to Boffi et al. [65, Chap. 8], Girault and
Raviart [217, §II.1.4] for thorough reviews of the topic.
54.1.1 Fortin operator

One way to prove the inf-sup condition (53.15) consists of using the notion of Fortin operator.
The theory behind the Fortin operator theory is investigated in detail §26.2.3. We now briefly
summarize the main features of this theory and adapt the notation to the setting of the Stokes
equations.
Let V , Q be two complex Banach spaces and let b be a bounded sesquilinear form on V ×Q. Let
β and kbk be the inf-sup and the boundedness constants of b. Let Vh0 ⊂ V and let Qh ⊂ Q be finite-
dimensional subspaces equipped, respectively, with the norms of V and Q. A map Πh : V → Vh0 ,
is called a Fortin operator if b(Πh (v) − v, qh ) = 0 for all (v, qh ) ∈ V ×Qh , and there is real
number γh > 0 such that γh kΠh (v)kV ≤ kvkV for all v ∈ V . The key result we are going to
352 Chapter 54. Stokes equations: Stable pairs (I)
use is the following statement (see Lemma 26.9, Boffi et al. [65, Prop. 8.4.1], and the work by the
authors [187, Thm. 1]).
Lemma 54.1 (Fortin operator). If there exists a Fortin operator, then the inf-sup condition
|b(vh , qh )|
inf sup =: βh > 0, (54.1)
qh ∈Qh vh ∈Vh0 kvh kV kqh kQ
holds true with βh ≥ γh β. Conversely, if the inf-sup condition (54.1) holds true, then there exists
βh
a Fortin operator with γh ≥ kbk .
Hence, proving the inf-sup condition (54.1) can be done by constructing a Fortin operator. A
practical way to do this is given by the following result.
Lemma 54.2 (Decomposition). Let Π1h , Π2h : V → Vh0 be two operators. Assume the follow-
ing: (i) Π2h is linear; (ii) b(v − Π2h (v), qh ) = 0 for all (v, qh ) ∈ V ×Qh ; (iii) The real numbers
kΠ1h (v)kV kΠ2h (v − Π1h (v))kV

c1h := sup and c2h := sup (54.2)
v∈V kvkV v∈V kvkV
are finite. Then (recalling that IV : V → V is the identity)
Πh := Π1h + Π2h (IV − Π1h ) (54.3)
is a Fortin operator with γh ≥ (c1h + c2h )−1 .
Proof. Since the operator Π2h is linear owing to the assumption (i), we have
b(v − Πh (v), qh ) = b(v − Π2h (v), qh ) − b(Π1h (v) − Π2h (Π1h (v)), qh ),
for all (v, qh ) ∈ V ×Qh , and both terms on the right-hand side are zero owing to the assumption (ii).
Furthermore, we have supv∈V kΠkvk h (v)kV
V
≤ c1h + c2h , i.e., γh kΠh (v)kV ≤ kvkV for all v ∈ V with
γh ≥ (c1h + c2h )−1 > 0 owing to the assumption (iii).
54.1.2 Weak control on the pressure gradient

A second possibility to prove the inf-sup condition (54.1) consists of establishing a weak control
on the gradient of the pressure. This technique can be used when the discrete pressure space is
H 1 -conforming. Let us focus more specifically on the bilinear form b(v, q) := −(∇·v, q)L2 (D) . Let
′
p ∈ (1, ∞), V := W01,p (D) equipped with the norm kvkV := |v|W 1,p (D) , and Q := Lp∗ (D) := {q ∈
′ R
Lp (D) | D q dx = 0} equipped with the norm kqkQ := kqkLp′ (D) with p′ ∈ (1, ∞) s.t. p1 + p1′ = 1.
The discrete velocity space is Vh0 ⊂ V , and the discrete pressure space is Qh ⊂ Q.
Lemma 54.3 (Pressure gradient control). Assume that the discrete pressure space Qh is
H 1 -conforming, and that there is c such that the following holds true for all p ∈ (1, ∞) and all
h ∈ H:
! 1′
|b(vh , qh )| X p′ p′
p
sup ≥c hK k∇qh kLp′ (K) . (54.4)

vh ∈Vh0 |vh |W 1,p (D)
K∈Th
Then the inf-sup condition (54.1) holds true uniformly w.r.t. h ∈ H.

Proof. Let qh ∈ Qh . Since Qh ⊂ Q, the continuous inf-sup condition (53.9) implies that
|b(v, qh )| I av
|b(I h (v), qh )| |b(v − I av
h (v), qh )|
βD kqh kQ ≤ sup ≤ sup + sup ,
v∈V |v|W 1,p (D) v∈V |v|W 1,p (D) v∈V |v|W 1,p (D)
1,p
where I av d
h0 is the R -valued version of the W0P-conforming quasi-interpolation P operator intro-
duced in §22.4.2. This means that I av h0 (v) :=
av
i∈{1: d} Ih0 (vi )ei , where v := i∈{1: d} vi ei and
{ei }i∈{1: d} is the canonical Cartesian basis of Rd . Let T1 , T2 denote the two terms on the right-
hand side. Owing to the W01,p -stability of I av h0 , we have |I I av
h0 (v)|W 1,p (D) ≤ cI |v|W 1,p (D) . Since
av
I h0 (v) ∈ Vh0 , we infer that
I av
|b(I h0 (v), qh )| |b(vh , qh )|
|T1 | ≤ cI sup av ≤ cI sup .
v∈V I h0 (v)kW 1,p (D)
kI vh ∈Vh0 |v h |W 1,p (D)
Moreover, using that Qh is H 1 -conforming to integrate by parts, and then invoking Hölder’s
inequality and the approximation properties of I av h0 , we infer that

|b(v − I av av
h0 (v), qh )| = (∇qh , v − I h0 (v))L2 (D)

X
≤c k∇qh kLp′ (K) hK k∇vkLp (DK ) ,
K∈Th
P
where DK is the set of the points composing the mesh cells touching K. Since K∈Th k∇vkpLp (DK ) ≤
ck∇vkpLp(D) = c|v|pW 1,p (D) owing to the regularity of the mesh sequence, Hölder’s inequality com-
bined with the assumption (54.4) implies that
! 1′
X ′ ′
p
|b(vh , qh )|
|T2 | ≤ c′ hpK k∇qh kpLp′ (K) ≤ c′′ sup .
K∈Th
This completes the proof of the inf-sup condition.

Remark 54.4 (Literature). The technique presented above is based on Bercovier and Pironneau
[54, Prop. 1], Verfürth [376] (for p := 2).
54.2 Mini element: the (P

P1 -bubble, P1) pair
Let (Th )h∈H be a shape-regular sequence of affine simplicial meshes. Recall from §53.4.2 that
the reason for which the (P P1 , P1 ) pair does not satisfy the inf-sup condition (53.15) is that the
velocity space is not rich enough (or equivalently the pressure space is too rich). To circumvent
this difficulty, we are going to enlarge the velocity space by adding one more degree of freedom per
simplex for each Cartesian component of the velocity.
Let K b be the reference simplex and x bK b
b be its barycenter, and let b be a function such that
bb ∈ W 1,∞ (K),
b 0 ≤ bb ≤ 1, bb(b
xKb ) = 1. (54.5)
0
Q
One can use bb(b
x) := (d + 1)d+1 i∈{0: d} λbi (b
x), where {λbi }i∈{0: d} are the barycentric coordinates
b This function is usually called bubble function in reference to the shape of its graph as shown
on K.
in Figure 54.1. Another possibility consists of dividing the simplex K b into (d + 1) subsimplices
Velocity Pressure
P1 -bubble 3P1 or 4P1 P1
Figure 54.1: Conventional representation of the (P1 -bubble, P1 ) pair in dimensions two (top) and
three (bottom). The degrees of freedom for the velocity are shown in the first column (P1 -bubble)
and in the second column (3P1 in dimension two and 4P1 in dimension three). Some isolines of
the two-dimensional bubble function are drawn. The pressure degrees of freedom are shown in the
third column.
b to x
by connecting the (d + 1) vertices of K bKb . Then bb is defined to be the continuous piecewise
affine function on Kb that is equal to one at x bKb and zero at the vertices of K.b We introduce
b b d b
the finite-dimensional space P := P 1,d ⊕ (span{b}) and define Σ to be the Lagrange degrees of
b plus x
freedom associated with the vertices of K bKb for each Cartesian component of the velocity.
Let (Th )h∈H be a shape-regular sequence of affine simplicial meshes so that each mesh covers
D exactly. Recalling that we are enforcing homogeneous Dirichlet conditions on the velocity, the
approximation spaces are defined by
Vh0 := P g1,0 (Th ) ⊕ Bh , Qh := P1g (Th ) ∩ L2∗ (D), (54.6)

L
where Bh := K∈Th (span{bK })d and bK := bb ◦ TK being the bubble function associated with the
mesh cell K ∈ Th . Notice that
Vh0 = {vh ∈ C 0 (D) | ∀K ∈ Th , vh ◦TK ∈ Pb , vh|∂D = 0}, (54.7)
and that Vh0 ⊂ W01,p (D) for all p ∈ (1, ∞). We now show that the (P
P1 -bubble, P1 ) pair is stable.
We do so by constructing a Fortin operator as in Lemma 54.2.
1 1
Lemma 54.5 (Stability). Let p ∈ (1, ∞) and let p′ ∈ (1, ∞) be s.t. p + p′ = 1. Let Vh0 and Qh
be defined in (54.6). There is β0 such that for all h ∈ H,

R
| D qh ∇·vh dx|
inf sup ≥ β0 > 0. (54.8)
qh ∈Qh vh ∈Vh0 kqh kLp′ (D) |vh |W 1,p (D)
Proof. Let us build a Fortin operator by means of the construction devised in Lemma 54.2 with
V := W01,p (D) equipped with the norm kvkV := |v|W 1,p (D) . We define the operator Π2h : V →
Vh0 by setting R
X
R K v dx
Π2h (v) := bK ∈ Bh ⊂ Vh0 .
K∈T K bK dx h
This operator is linear in agreement

R with theR assumption (i) of Lemma 54.2. Moreover, the
definition of Π2h implies that K Π2h (v) dx = K v dx for all v ∈ V . Then for all (v, qh ) ∈ V ×Qh
we infer that
Z Z X Z
b(v, qh ) = − qh ∇·v dx = v·∇qh dx = ∇qh|K · v dx
D D K∈Th K
X Z Z
= ∇qh|K · Π2h (v) dx = Π2h (v)·∇qh dx = b(Π2h (v), qh ),
K∈Th K D
which proves the assumption (ii) of Lemma 54.2. We now set Π1h := I av av
h0 , where I h0 : V → Vh0 is
d 1,p
the R -valued version of the W0 -conforming quasi-interpolation operator introduced in §22.4.2.
W |Π1h (v)|
(D) 1,p
We observe that the real number c1h := supv∈V |v|W 1,p (D) is uniformly bounded w.r.t. h ∈ H.
Moreover, the regularity of the mesh sequence and Lemma 11.7 imply that for all K ∈ Th ,
1 1
|bK |W 1,p (K) ≤ c kJ−1 p b ′ −1
K kℓ2 |det(JK )| |b|W 1,p (K)
b ≤ c hK |K| .
p
R R
Similar arguments show that K
bK dx ≥ c|K| and Hölder’s inequality implies that | K
v dx| ≤
1
|K| kvk
p′
Lp (K) . Putting these estimates together shows that
−1
|Π2h (v)|W 1,p (K) ≤ c hK kvkLp (K) .
Then the approximation properties of I av

h0 (see Theorem 22.14) yield
|Π2h (v − Π1h (v))|W 1,p (K) = |Π2h (v − I av

h0 (v))|W 1,p (K)
≤ c h−1 av ′
K kv − I h0 (v)kLp (K) ≤ c |v|W 1,p (DK ) ,
where DK is the set of the points composing the mesh cells touching K. Summing the above
bound over K ∈ Th and using the regularity of the mesh sequence, we infer that
|Π2h (v − Π1h (v))|W 1,p (D) ≤ c |v|W 1,p (D) .
|Π2h (v−Π1h (v))|W 1,p (D)

This shows that the real number c2h := supv∈V is uniformly bounded w.r.t.
|v|W 1,p (D)
h ∈ H. In conclusion, all the assumptions of Lemma 54.2 are met, showing that Πh := Π1h +
Π2h (IV − Π1h ) is a Fortin operator with γh ≥ (c1h + c2h )−1 . Notice that γh is bounded from
below away from zero uniformly w.r.t. h ∈ H. Invoking Lemma 54.1, we conclude that the inf-sup
condition (54.8) holds true uniformly w.r.t. h ∈ H.
Remark 54.6 (Convergence rate). Assume that the solution to (53.6) is such that u ∈ H 2 (D)∩
H01 (D) and p ∈ H 1 (D) ∩ L2∗ (D). Owing to Theorem 53.17, the discrete solution (uh , ph ) to
(53.14) with (Vh0 , Qh ) defined in (54.6) satisfies µ|u − uh|H 1 (D) + kp − phkL2 (D) ≤ ch(µ|u|H 2 (D) +
|p|H 1 (D) ). If the assumptions of Theorem 53.19 additionally hold true with s := 1, then µku −
uh kL2 (D) ≤ ch2 (µ|u|H 2 (D) +|p|H 1 (D) ). Notice that the convergence rate of the error on the velocity
is that associated with the finite element space P g1,0 (Th ), i.e., the bubble functions introduced to
approximate the velocity do not contribute to the approximation error, they contribute only to
the stability of the discretization (see also Exercise 54.2).
Remark 54.7 (Literature). The idea of using bubble functions has been introduced by Crouzeix
and Raviart [151]. The analysis of the mini element is due to Arnold et al. [20].
54.3 Taylor–Hood element: the (P

P2 , P1) pair
This section is dedicated to the analysis of the Taylor–Hood element based on the (P P2 , P1 ) pair.
Compared to the mini element which is based on the (P1 -bubble, P1 ) pair, the idea is to further
enrich the discrete velocity space so as to improve by one order the convergence rate of the error.
Let (Th )h∈H be a shape-regular family of affine simplicial meshes. Recalling that we are enforcing
homogeneous Dirichlet conditions on the velocity, the approximation spaces are defined by
g
Vh0 := P2,0 (Th ), Qh := P1g (Th ) ∩ L2∗ (D), (54.9)
i.e., the velocity is approximated using continuous P2 elements and the pressure is approximated us-
ing continuous P1 elements. The conventional representation of this element is shown in Figure 54.2.
We are going to prove the inf-sup condition (54.1) by using the technique described in §54.1.2, i.e.,
we first establish a weak control on the pressure gradient, then we invoke Lemma 54.3. As above,
′
we set V := W01,p (D) and Q := Lp∗ (D) with p, p′ ∈ (1, ∞) and p1 + p1′ = 1. Notice that Vh0 ⊂ V
and Qh ⊂ Q.
Lemma 54.8 (Bound on pressure gradient). Let Vh0 , Qh be defined in (54.9). Assume that
d ∈ {2, 3} and that every mesh cell has at least d internal edges (i.e., at most one face in ∂D).
There is c such that the following holds true for all p ∈ (1, ∞) and all h ∈ H:
R ! 1′
| X p
D qh ∇·vh dx| ′ ′
sup ≥c hpK k∇qh kpLp′ (K) . (54.10)
K∈Th
Proof. We only give the proof for d = 3 since the proof for d = 2 is similar. Let us number all the
internal mesh edges from 1 to Nei . Consider an oriented edge Ei with i ∈ {1:Nei }, and denote its
two endpoints by zi± and its midpoint by mi . Set li := kzi+ − zi− kℓ2 and τi := li−1 (zi+ − zi− ), so
that li is the length of Ei and τi is the unit tangent vector orienting Ei . Let qh be a function in
Qh and let sgn be the sign function. Let vh ∈ Vh0 be (uniquely) defined by prescribing its global
degrees of freedom in Vh0 as follows:

 v (a ) := 0 if aj is a mesh vertex,
 h j
′ ′
 vh (mi ) := −lip sgn(∂τi qh )|∂τi qh |p −1 τi if Ei 6⊂ ∂D,


vh (mi ) := 0 if Ei ⊂ ∂D,
Velocity Pressure Velocity Pressure

P2 P1 Q2 Q1
Figure 54.2: Conventional representation of the (P

P2 , P1 ) pair (left) and of the (Q
Q2 , Q1 ) pair (right)
in dimensions two (top) and three (bottom, only visible degrees of freedom are shown).
where ∂τi qh := τi ·∇qh denotes the tangential derivative of qh along the oriented edge Ei . Note
that vh (mi ) depends only on the values of qh on Ei . Let K ∈ Th . Using the quadrature formula
Z !
X φ(m) X φ(a)
φ dx = |K| − , ∀φ ∈ P2 ,
K 5 20
m∈MK a∈VK
where MK is the set of the midpoints of the edges of K and VK is the set of the vertices of K and
since Qh is H 1 -conforming, we infer that
Z Z X Z
qh ∇·vh dx = − vh ·∇qh dx = − vh ·∇qh dx
D D K∈Th K
X X 1
=− |K| vh (mi )·∇qh (mi )
5
K∈Th mi ∈K
X X 1 ′ ′ X p′ ′
= |K| |∂τi qh (mi )|p lip ≥ c hK k∇qh kpLp′ (K) .
5
K∈Th mi ∈K K∈Th
The last inequality results from the fact that li ≥ chK owing to the regularity of the mesh sequence,
and that every tetrahedron K ∈ Th has at least three edges in D, i.e., the quantities |∂τi qh (mi )|,
where mi spans the midpoints of the edges of K that are not in ∂D, control k∇qh kℓ2 . Finally, the
inverse inequality from Lemma 12.1 (with r := p, l := 1, m := 0) together with Proposition 12.5
implies that for all K ∈ Th ,
X
|vh |pW 1,p (K) ≤ c h−p
K |K| kvh (m)kpℓ2 ,
m∈MK
′ ′
and since li ≤ chK , we have kvh (m)kℓ2 ≤ chpK k∇qh kℓp2−1 . Since p(p′ − 1) = p′ , combining these
′ ′
bounds shows that |vh |pW 1,p (K) ≤ chpK k∇qh kpLp′ (K) for all K ∈ Th . This proves (54.10).
Lemma 54.9 (Stability). For all p ∈ (1, ∞) and under the hypotheses of Lemma 54.8, the
P2 , P1 ) pair satisfies the inf-sup condition (54.8) uniformly w.r.t. h ∈ H.
(P
Proof. Apply Lemma 54.3.
Remark 54.10 (Convergence rate). Owing to Theorem 53.17 and assuming that the solu-
tion to (53.6) is smooth enough, the solution to (53.14) with (Vh0 , Qh ) defined in (54.9) satisfies
µ|u−uh |H 1 (D) +kp−ph kL2 (D) ≤ ch2 (µ|u|H 3 (D) +|p|H 2 (D) ). Moreover, if the assumptions of Theo-
rem 53.19 are met for some s ∈ (0, 1], then µku − uhkL2 (D) ≤ ch2+s ℓ1−s
D (µ|u|H 3 (D) + |p|H 2 (D) ).
Remark 54.11 (Literature). Further insight and alternative proofs can be found in Bercovier
and Pironneau [54, Prop. 1], Girault and Raviart [217, p. 176], Stenberg [354]. We refer the reader
to Mardal et al. [294] for the construction of a Fortin operator associated with the Taylor–Hood
element in dimension two. Well-balanced schemes (see Remark 53.22) using Taylor–Hood mixed
finite elements are analyzed in Lederer et al. [279].
54.4 Generalizations of the Taylor–Hood element

In this section, we briefly review some generalizations of the Taylor–Hood element: extension to
quadrangles, higher-order extensions, and the use of a submesh to build the discrete velocity space.
54.4.1 The (P
Pk , Pk−1) and (Q
Qk , Qk−1 ) pairs
It is possible to generalize the Taylor–Hood element to quadrangles and hexahedra. For instance,
the (QQ2 , Q1 ) pair has the same properties as the Taylor–Hood element; see Figure 54.2.
It is also possible to use higher-degree polynomials. For k ≥ 2, the (P Pk , Pk−1 ) pair and
the (QQ k , Qk−1 ) pair are stable in dimensions two and three. Provided the solution to (53.6)
is smooth enough, these elements yield the error estimates µ|u − uh |H 1 (D) + kp − ph kL2 (D) ≤
c hk (µ|u|H k+1 (D) + |p|H k (D) ) and µku − uh kL2 (D) ≤ chk+s ℓ1−s
D (µ|u|H k+1 (D) + |p|H k (D) ) if the as-
sumptions of Theorem 53.19 are met for some s ∈ (0, 1]. Proofs and further insight can be found
in Stenberg [352, p. 18], Brezzi and Falk [92], Boffi et al. [65, p. 494], Boffi [60].
54.4.2 The (P
P1 -iso-P
P2 , P1 ) and (Q
Q1 -iso-Q
Q2 , Q1 ) pairs
An alternative to the Taylor–Hood element consists of replacing the P 2 approximation of the ve-
locity by a P 1 approximation on a finer simplicial mesh. This finer mesh, say T h , is constructed
2
as follows. In two dimensions, each triangle in Th is divided into four new triangles by connecting
the midpoints of the three edges. In three dimensions, each tetrahedron in Th is divided into eight
new tetrahedra (all having the same volume) by dividing each face into four new triangles and by
connecting the midpoints of one pair of nonintersecting edges (there are three pairs of noninter-
secting edges). This construction is illustrated in the top and bottom left panels of Figure 54.3.
The discrete spaces are
g
Vh0 := P1,0 (T h ), Qh := P1g (Th ) ∩ L2∗ (D). (54.11)
2
Velocity Pressure Velocity Pressure

P 1 -iso-P
P2 P1 Q 1 -iso-Q
Q2 Q1
Figure 54.3: (P
P1 -iso-P
P2 , P1 ) (left) and (Q
Q1 -iso-Q
Q2 , Q1 ) (right) pairs in dimensions two (top) and
three (bottom, only visible degrees of freedom are shown for the (Q Q 1 -iso-Q
Q 2 , Q1 ) pair).
These finite element pairs are often called (P P1 -iso-P

P2 , P1 ), or (4 P 1 , P1 ) in two dimensions and
(8 P 1 , P1 ) in three dimensions.
The (P P1 -iso-P
P2 , P1 ) pair can be generalized to quadrangles in two dimensions and hexahedra
in three dimensions. Assume that (Th )h∈H is a shape-regular sequence of meshes composed of
quadrangles or hexahedra. A new mesh T h is defined in two dimensions by dividing each quadrangle
2
in Th into four new quadrangles and by connecting the midpoints of all the pairs of nonintersecting
edges. In three dimensions, we divide each hexahedron in Th into eight new hexahedra by dividing
each face into four quadrangles and by connecting the barycenters of all the pairs of nonintersecting
faces. This construction is illustrated in the top and bottom right panels of Figure 54.3. The
discrete spaces are
Vh0 := {vh ∈ C 0 (D) | ∀K ∈ T h , vh ◦ TK ∈ Q 1 , vh|∂D = 0}, (54.12a)

2
0
Qh := {qh ∈ C (D) ∩ L2∗ (D) | ∀K ∈ Th , qh ◦ TK ∈ Q1 }. (54.12b)
These finite elements are often called (Q

Q 1 -iso-Q
Q2 , Q1 ), or (4 Q 1 , Q1 ) in dimension two and (8 Q 1 , Q1 )
in dimension three.
Lemma 54.12 (Stability). For all p ∈ (1, ∞), and under the hypotheses of Lemma 54.8 if Th
is composed of simplices, the (PP1 -iso-P
P2 , P1 ) and (Q
Q1 -iso-Q
Q2 , Q1 ) pairs satisfy the inf-sup condi-
tion (54.8) uniformly w.r.t. h ∈ H.
Proof. Adapt the proof of Lemma 54.8; see Bercovier and Pironneau [54] (for d = 2 and p = 2)
and Exercise 54.4.
Remark 54.13 (Convergence rate). Owing to Theorem 53.17 and assuming that the solution
to (53.6) is smooth enough, the discrete solution to (53.14) with (Vh0 , Qh ) defined in either (54.11)
or (54.12) satisfies µ|u−uh |H 1 (D) +kp−ph kL2 (D) ≤ ch(µ|u|H 2 (D) +|p|H 1 (D) ), and if the assumptions
of Theorem 53.19 are met for some s ∈ (0, 1], we have µku − uh kL2 (D) ≤ ch1+s ℓ1−s D (µ|u|H 2 (D) +
|p|H 1 (D) ).
Exercises
Exercise 54.1 (Mini element). Show that the P FortinPoperator iΠh constructed in the proof of
Lemma 54.5 is of the form Πh (v) := I avh0 (v) + K∈Th i∈{1: d} γK (v)bK ei , for some coefficients
i
γK (v) to be determined. Here, {ei }i∈{1: d} is the canonical Cartesian basis of Rd .
Exercise 54.2 (Bubble⇔Stabilization). Consider the mini element defined in §54.2 and assume
that the viscosity µ is constant over D. Recall that Vh0 := Vh0 1
⊕ Bh and Qh := P1g (Th ) ∩ L2∗ (D)
1 := g
with Vh0 P 1,0 (Th ). Let (uh , ph ) be the solution to the discrete Stokes problem (53.14). (i) Show
1
that a(vh , bh ) = 0 for all vh ∈ Vh0 and all bh ∈ Bh . (ii) Set uh := u1h + ubh ∈ Vh0 . Show that
a(u1h , vh ) + b(vh , ph ) = F (vh ), 1

∀vh ∈ Vh0 . (54.13)
(iii) Let bK := bb◦TK be the bubble function on K ∈ Th . Let {ei }i∈{1: d} be the canonical Cartesian
K := R 1
basis of Rd . Let S K ∈ Rd×d be defined by Sij b dx
a(bK ej , bK ei ) for all i, j ∈ {1:d}. Let
b
P i K −1
K K
i := R 1
uh|K := i∈{1: d} cK ei bK . Show that cK = (S ) (FK − ∇ph|K ), where FK b dx
F (bK ei ),
P K −1
R K K
for all i ∈ {1:d}. (iv) Set ch (ph , qh ) := K∈Th ∇qh|K (S ) ∇ph|K K bK dx and Rh (qh ) :=
P K −1
R
K∈Th ∇q h|K (S ) FK K
b K dx. Show that the mass conservation equation becomes
b(u1h , qh ) − ch (ph , qh ) = G(qh ) − Rh (qh ), ∀qh ∈ Qh . (54.14)

P h2K R
Note: since (SK )−1 scales like µ−1 h2K , ch (ph , qh ) behaves like K∈Th µ K ∇qh ·∇ph dx, and
P h2K R
Rh (qh ) scales like K∈Th µ K ∇qh|K ·FK dx. This shows that, once the bubbles are eliminated,
the system (54.13)-(54.14) is equivalent to a stabilized form of the Stokes system for the (P P 1 , P1 )
pair; see Chapters 62 and 63.
Exercise 54.3 (Singular vertex). Let K ⊂ R2 be a quadrangle and let z be the intersection of
the two diagonals of K. Let K1 , . . . , K4 be the four triangles formed by dividing K along its two
diagonals (assume that K1 ∩ K3 = {z} and K2 ∩ K4 = {z}). (i) LetPφ be a scalar field continuous
over K and of class C 1 over the triangles K1 , . . . , K4 . Prove that i∈{1: 4} (−1)i n·∇φ|Ki (z) = 0
for every unit vector n. (ii) Let vPbe a vector field continuous over K and of class C 1 over the
triangles K1 , . . . , K4 . Prove that i∈{1: 4} (−1)i ∇·v|Ki (z) = 0. (iii) Assume that v is linear over
R
each triangle. Show that the four equations Ki ∇·v dx = 0 for all i ∈ {1:4} are linearly dependent.
Exercise 54.4 (P P1 -iso-P

P2 , P1 ). Consider the setting of Lemma 54.12 with the (PP 1 -iso-PP2 , P1 ) pair
in dimension three. (i) Let K ∈ Th . Let VK be the set of the vertices of K. Let MK be the mid-
points of the six edges of K. Let M1K be the set of the two midpoints that are connected to create
the 8 new tetrahedra. Let M2K be the set of the remaining midpoints. Let Vh0 be the P 1 velocity
space basedRof Th/2 . Find the P coefficients α, β, γ so
P that the following quadrature
P is exact for all
wh ∈ Vh0 : K wh dx = |K|(α z∈VK wh (z) + β m∈M1 wh (m) + γ m∈M2 wh (m)). (Hint :
K K
R P
on a tetrahedron K ′ with vertices {z ′ }z′ ∈VK ′ , the quadrature K ′ wh dx = |K ′ | z′ ∈VK ′ 41 wh (z ′ )
is exact on P 1 .) (ii) Prove Lemma 54.12 for the (P P1 -iso-P
P2 , P1 ) pair in dimension three for all
p ∈ (1, ∞). (Hint : adapt the proof of Lemma 54.8.)
Chapter 55
Stokes equations: Stable pairs (II)
In this chapter, we continue the study of stable finite element pairs that are suitable to approximate
the Stokes equations. In doing so, we introduce another technique to prove the inf-sup condition
that is based on a notion of macroelement. Recall that we assume that Dirichlet conditions are
enforced on the velocity over the whole boundary, that D is a polyhedron in Rd , and that (Th )h∈H
is a shape-regular sequence of affine meshes so that each mesh covers D exactly. In this chapter,
we focus more specifically on the case where the discrete pressure space is a broken finite element
space.
55.1 Macroelement techniques

In addition to the Fortin operator technique described in Lemma 54.1 and the method consisting of
weakly controlling the pressure gradient described in Lemma 54.3, we now present a third method
to establish the inf-sup condition between the discrete velocity space and the discrete pressure
space. This method is based on a notion of macroelement.
We return to the abstract setting and consider two complex Banach spaces V and Q and a
bounded sesquilinear form b on V ×Q. Let Vh0 ⊂ V and Qh ⊂ Q. Recall that kbk denotes the
boundedness constant of b on V ×Q and that the inf-sup condition (54.1) takes the form
|b(vh , qh )|
inf sup =: βh > 0. (55.1)
1 2
Lemma 55.1 (Partition lemma). Let Vh0 , Vh0 be two subspaces of Vh0 and Q1h , Q2h be two
1 2
subspaces of Q such that Qh = Qh + Qh . Let
|b(vh , qh )| |b(vh , qh )|
β1 := inf 1 sup , β2 := inf 2 sup ,
qh ∈Qh vh ∈V 1
h0
kvh kV kqh kQ qh ∈Qh vh ∈V 2
h0
kvh kV kqh kQ
|b(vh , qh )| |b(vh , qh )|
b12 := sup sup , b21 := sup sup .
qh ∈Q1h vh ∈Vh0
2 kvh kV kqh kQ qh ∈Q2h vh ∈Vh0
1 kvh kV kqh kQ
Assume that 0 < β1 β2 and λ1 λ2 < 1 with λ1 := bβ122 , λ2 := bβ211 . Then the inf-sup condition (55.1)
holds true with βh ≥ 14 min(β1 , β2 ) if λ1 + λ2 ≤ 1 and with βh ≥ 64 1
(1 − λ1 λ2 )kbk−2 min(β1 , β2 )3
otherwise.
364 Chapter 55. Stokes equations: Stable pairs (II)
Proof. Let qh := qh1 + qh2 ∈ Qh \{0}. The definition of β1 , β2 together with the assumption 0 < β1 β2
implies that there exists vhl ∈ Vhl so that b(vhl , qhl ) = kqhl k2Q and βl kvhl kV ≤ kqhl kQ for all l ∈ {1, 2}.
We now investigate two cases: either λ1 + λ2 ≤ 1 or λ1 + λ2 > 1.
(1) Let us assume that λ1 + λ2 ≤ 1. Then, setting vh := vh1 + vh2 we have
b(vh , qh ) = b(vh1 , qh1 ) + b(vh2 , qh1 ) + b(vh1 , qh2 ) + b(vh2 , qh2 )

≥ kqh1 k2Q + kqh2 k2Q − b12 kvh2 kV kqh1 kQ − b21 kvh1 kV kqh2 kQ
≥ kqh1 k2Q + kqh2 k2Q − (β2−1 b12 + β1−1 b21 )kqh1 kQ kqh2 kQ .
Using that β2−1 b12 + β1−1 b21 = λ1 + λ2 ≤ 1, we infer that

1 1 2 1 1 2
b(vh , qh ) ≥ kqh kQ + kqh2 k2Q ≥ kqh1 kQ + kqh2 kQ
2 2 4
1 1
≥ kqh kQ β1 kvh kV + β2 kvh2 kV ≥ min(β1 , β2 )kqh kQ kvh kV ,
1
4 4
where we used the triangle inequality and the above bounds on kvhl kV for all l ∈ {1, 2}. The
assertion then follows with βh ≥ 41 min(β1 , β2 ).
(2) Let us now assume that λ1 + λ2 > 1. Without loss of generality, we assume that λ2 ≥ λ1 . Let
σ ∈ R, let vh := vh1 + σvh2 , let ǫ > 0, and let us minorize b(vh , qh ) as follows:
b(vh , qh ) = b(vh1 , qh1 ) + σb(vh2 , qh1 ) + b(vh1 , qh2 ) + σb(vh2 , qh2 )

≥ kqh1 k2Q + σkqh2 k2Q − b12 kvh2 kV kqh1 kQ − b21 σkvh1 kV kqh2 kQ
≥ kqh1 k2Q + σkqh2 k2Q − (β2−1 b12 + σβ1−1 b21 )kqh1 kQ kqh2 kQ
ǫ 1
≥ 1 − (λ1 + σλ2 ) kqh1 k2Q + σ − (λ1 + σλ2 ) kqh2 k2Q .
2 2ǫ
Let us show that we can choose σ and ǫ so that 2ǫ (λ1 + σλ2 ) < 1 and 2ǫ 1
(λ1 + σλ2 ) < σ. We
2
consider the quadratic equation Ψ(t) := (λ1 + tλ2 ) − 4t = 0. Since the discriminant, 16(1 − λ1λ2 ),
is positive and λ2 6= 0, Ψ(t) has two distinct roots, t− , t+ , and Ψ is minimal at 21 (t− +t+ ) = 2−λ1 λ2
λ2
.
2
Therefore, if we choose σ := 2−λ 1 λ2
λ22
, we have Ψ(σ) < 0, i.e., 12 (λ1 + σλ2 ) < λ1 +σλ
2σ
2
. We then define
1 1 2σ 2σ
ǫ by setting ǫσ := 2 ( 2 (λ1 + σλ2 ) + λ1 +σλ2 ). This choice in turn implies that ǫσ < λ1 +σλ 2
, i.e.,
ǫ 1 1
2 (λ1 + σλ2 ) < 1 and that 2 (λ1 + σλ2 ) < ǫσ, i.e., 2ǫ (λ1 + σλ2 ) < σ. We have thus proved that
c1 := 1 − 2ǫ (λ1 + σλ2 ) > 0 and c2 := σ − 2ǫ 1
(λ1 + σλ2 ) > 0. Then we conclude as above
1
b(vh , qh ) ≥ min(c1 , c2 )kqh kQ (β1 kvh1 kV + β2 kvh2 kV )
2
1
≥ min(c1 , c2 ) min(β1 , σ −1 β2 )kqh kQ kvh kV ,
2
and the assertion follows with βh ≥ 1

2 min(c1 , c2 ) min(β1 , σ −1 β2 ). Notice that λ2 ∈ [ 12 , kbk
β1 ] because
2−λ1 λ2 λ2 (3−λ1 λ2 )
2λ2 ≥ λ1 + λ2 ≥ 1 and b21 ≤ kbk. Moreover, since σ = λ22
and ǫ = 2(2−λ1 λ2 ) , we obtain
1 − λ1 λ2 (1 − λ1 λ2 )(2 − λ1 λ2 )
c1 = , c2 = ,
2(2 − λ1 λ2 ) λ22 (3 − λ1 λ2 )
β12 β2
so that c1 ≥ 41 (1 − λ1 λ2 ), c2 ≥ 2kbk2 (1 − λ1 λ2 ), σ −1 ≥ 81 . Hence, we have βh ≥ 1
32 min( 21 , kbk1 2 )(1 −
3
λ1 λ2 ) min(β1 , β2 ) ≥ 1
64 (1 − λ1 λ2 ) min(β 1 ,β2 )
kbk2 .
Remark 55.2 (Inequality λ1 λ2 < 1). This inequality, which amounts to b12 b21 < β1 β2 , is
trivially satisfied if b12 b21 = 0, which is the case in many applications; see, e.g., Corollary 55.3
below.
Let us illustrate the above result with the Stokes problem. We set V := H01 (D), Q := L2∗ (D),
kvkV := |v|H 1 (D) , kqkQ := kqkL2 (D) , and b(v, q) := −(∇·v, q)L2 (D) . Let Th be a mesh in the
sequence (Th )h∈H . Let Uh be a partition of the set Th . We call Uh macroelement partition and the
members of Uh macroelements. For every macroelement U ∈ Uh , we abuse the notation by writing
U also for the set of the points composing the cells in the macroelement U . For all U ∈ Uh , we
define the following spaces:
Vh0 (U ) := {vh ∈ Vh0 | vh|U ∈ H01 (U ), vh|D\U = 0} ⊂ Vh0 , (55.2a)

Qh (U ) := {1U qh | qh ∈ Qh }, (55.2b)
R
Qh (U ) := span(1U ), Qe h (U ) := {qh ∈ Qh (U ) | qh dx = 0}, (55.2c)
U
where 1U is the indicator function of U . We additionally define

X X
e h :=
Q Qe h (U ), Qh := Qh (U ). (55.3)
U∈Uh U∈Uh
Corollary 55.3 (Macroelement partition). Assume that for all h ∈ H, there exists a partition
of Th , say Uh , such that
|b(vh , qh )|
∀U ∈ Uh , inf sup =: β1h (U ) > 0, (55.4a)
e h (U) vh ∈Vh0 (U)
qh ∈Q kvh kV kqh kQ
|b(vh , qh )|
inf sup =: β2h > 0. (55.4b)
(i) The inf-sup condition (55.1) is satisfied. (ii) If inf h∈H β2h > 0 and inf h∈H minU∈Uh β1h (U ) > 0,
the inf-sup condition (55.1) holds uniformly w.r.t. h ∈ H.
Proof. The idea is to show that the assumptions of Lemma 55.1 are R met.
1
(1) For all qh ∈ Qh and all U ∈ Uh , let us denote q hU := |U| U qh dx. The identities qh =
P 1 2 1 e
U∈Uh 1U qh and 1U qh = 1U (qh − q hU ) + q hU 1U show that Qh = Qh + Qh , with Qh := Qh and
Q2h := Qh . Notice that this decomposition holds true whether Qh is composed of discontinuous
functions or not.
(2) Let us prove the first inf-sup condition from Lemma 55.1. Let qh ∈ Q1h = Q e h . Then (55.4a)
implies that for all U ∈ Uh there isP vh (U ) ∈ Vh0 (U ) s.t. ∇·(vPh (U )) = 1 q
U h and β 1h (U )kvh (U )kV ≤
1 := 1
k1U qh kQ = kqh kL2 (U) . Set vh := U∈Uh vh (U ) ∈ Vh0 V
U∈Uh h0 (U ). Notice that Vh0 ⊂ Vh0
P 2 2
1
by construction. Using that U∈Uh kvh (U )kV = kvh kV , we infer that
Z X Z X
qh ∇·vh dx = qh ∇·vh (U ) dx = kqh k2L2 (U)
D U∈Uh U U∈Uh
X 12 X 12
= kqh k L2 (D) kqh k2L2 (U) ≥ kqh kQ 2
(β1h (U )) kvh (U )k2V
U∈Uh U∈Uh
X 12
≥ β1h kqh kQ kvh (U )k2V = β1h kqh kQ kvh kV ,
U∈Uh
|b(vh ,qh )|
β1h := minU∈Uh β1h (U ) > 0. Hence, inf qh ∈Q1h supvh ∈Vh0
1
kvh kV kqh kQ ≥ β1h .
2 :=
(3) The second inf-sup condition from Lemma 55.1 holds by assumption with Vh0 Vh0 , Q2h :=
Qh , and the constant β2h > 0.
(4) Finally, let us verify the last assumption by showing that λ1 λ2 := βb1h 12 b21
β2h = 0 < 1. Let
P 1
P 2
vh := U∈Uh vh (U ) ∈ Vh0 and qh := U∈Uh qU 1U ∈ Qh . We obtain
X Z
b(vh , qh ) = qU ∇·vh (U ) dx = 0,
U∈Uh U
R
since vh (U ) ∈ H01 (U ) implies that U ∇·vh (U ) dx = 0 for all U ∈ Uh . Hence, b21 = 0. This
completes the proof.
Remark 55.4 (Assumption (55.4a)). For all qh ∈ Qh , let q h ∈ Qh be defined s.t. q h|U := q hU :=
1
R R R
|U| U qh dx for all U ∈ Uh . Since U qh ∇·vh dx = U (qh − q hU )∇·vh dx for all vh ∈ Vh0 (U ) and
all U ∈ Uh , the assumption (55.4a) means that for all qh ∈ Qh , we have supvh ∈Vh0 (U) |b(v h ,qh )|
kvh kV ≥
β1h (U )kqh|U − q hU kQ . Then the argument in Step (2) of the proof of Corollary 55.3 shows that
|b(vh ,qh )|
supvh ∈Vh0
1
kvh kV ≥ β1h kqh −q h kQ for all qh ∈ Qh , where we have set β1h := minU∈Uh β1h (U ).
e h ⊂ Qh and Qh ⊂ Qh when Qh is composed of discontinuous functions, but the
Notice that Q
above theory does not require that Qh be composed of discontinuous finite elements. It turns out
that the assumption (55.4b) can be relaxed if Qh is H 1 -conforming.
Proposition 55.5 (Macroelement, continuous pressures). Let (Th )h∈H be a shape-regular
mesh sequence. Assume that there exists a macroelement partition Uh for every mesh Th . Assume
b and that the sequence
that every U ∈ Uh can be mapped by an affine mapping to a reference set U
{Uh }h∈H is shape-regular. Assume that inf h∈H maxU∈Uh card{K ⊂ U } < ∞. Assume that Qh ⊂
H 1 (D) ∩ L2∗ (D) and that the following holds true that for all h ∈ H:
|b(vh , qh )|
∀U ∈ Uh , inf sup =: β1h (U ) > 0. (55.5)
e h (U) vh ∈Vh0 (U)
qh ∈Q kvh kV kqh kQ
(i) The inf-sup condition (55.1) is satisfied. (ii) If inf h∈H minU∈Uh β1h (U ) > 0, the inf-sup condi-
tion (55.1) holds uniformly w.r.t. h ∈ H.
Proof. See Brezzi and Bathe [91, Prop. 4.1] and Exercise 55.7.
Remark 55.6 (Literature). Macroelement techniques have been introduced in a series of works
by Boland and Nicolaides [67], Girault and Raviart [217, §II.1.4], Stenberg [352, 354, 353]. This
theory is further refined in Qin [328, Chap. 3]. In particular, Lemma 55.1 is established in [328,
Thm. 3.4.1]. It is possible to generalize the macroelement technique to situations where the
macroelements are not disjoint provided one assumes that each cell K belongs to a finite set
of macroelements with cardinality bounded from above uniformly w.r.t. h ∈ H. This type of
technique can be used in particular to prove the stability of the generalized Taylor–Hood elements
(P Qk , Qk−1 ), k ≥ 2. We refer the reader to Boffi et al. [65, §8.8] for a thorough discussion
Pk , Pk−1 ), (Q
on this topic.
55.2 Discontinuous pressures and bubbles

We investigate in this section finite element pairs based on simplicial meshes. The pressure ap-
proximation is discontinuous and stability is achieved by enriching the velocity space.
55.2.1 Discontinuous pressures

Since the functional space for the pressure is Q := L2∗ (D), the approximation setting remains
conforming for the pressure. The discrete pressure space is typically the broken polynomial space
(see §18.1.2)
b
Pl,∗ (Th ) := {qh ∈ L2∗ (D) | ∀K ∈ Th , qh ◦ TK ∈ Pl,d }, (55.6)
for some l ∈ N and where TK : K b → K is the geometric mapping. The (P Pk , Pbl ) pair refers to the
g b
choice of finite element space Vh0 := Pk,0 (Th ) for the velocity and Qh := Pl,∗ (Th ) for the pressure.
b P2 -bubble, Pb1 ) pairs.
The stable finite element pairs investigated herein are the (P P2 , P0 ) and the (P
Remark 55.7 (Local mass balance). Working with discontinuous pressures is interesting since
it becomes possible to test the discrete mass conservation equation against R a gfunction supported
−1
in
R a single mesh cell K ∈ Th . This leads to the local mass balance K (ψK ) (q)∇·u h dx =
g −1 g
K
(ψK ) (q)g dx for all q ∈ P k,d with ψK (q) := q ◦ T K , see Exercise 55.1.
55.2.2 P2 , Pb0 ) pair

The (P
Let (Th )h∈H be a shape-regular family of affine simplicial meshes. Recalling that we are enforcing
homogeneous Dirichlet conditions on the velocity, the (P P2 , Pb0 ) pair gives to the following approxi-
mation spaces:
g b
Vh0 := P2,0 (Th ), Qh := P0,∗ (Th ). (55.7)
This simple finite element pair satisfies the inf-sup condition (55.1) uniformly w.r.t. h ∈ H in
dimension two, but it has little practical interest since it is does not provide optimal convergence
results. Nevertheless it is an important building block for other more useful finite element pairs.
′
Let V := W01,p (D) be equipped with the norm kvkV := |v|W 1,p (D) and let Q := Lp∗ (D) be
equipped with the norm kqkQ := kqkLp′ (D) , where p, p′ ∈ (1, ∞) are s.t. p1 + p1′ = 1.
P2 , Pb0 ) pair satisfies the inf-sup condi-

Lemma 55.8 (Stability). Assume that d = 2. The (P
tion (55.1) uniformly w.r.t. h ∈ H.
Proof. We construct a Fortin operator by using the decomposition defined in Lemma 54.2 and by
invoking Lemma 54.1 to conclude. The operator Π2h : V → Vh0 is defined as follows. Let v ∈ Vh0 .
We set Π2h (v)(z) := 0 forR all z ∈ Vh◦ (Vh◦ is the collection of the internal vertices of the mesh),
3 ◦ ◦
and Π2h (v)(mF ) := 2|F | F v ds for all F ∈ Fh (Fh is the collection of the mesh interfaces),
where mF is the barycenter of F . This entirely defines Π2h (v) in Vh0 since d = 2. Notice that
v|F ∈ L1 (F ) for all v ∈ W01,p (D) and all F ∈ Fh◦ so that the above construction is meaningful.
1,p
Then we set Π1h := I av av d
h0 , where I h0 is the R -valued version of thePW0 -conforming quasi-
av av
interpolation operator introduced in §22.4.2. This means that I h0 (v) := i∈{1: d} Ih0 (vi )ei , where
P d
v := i∈{1: d} vi ei and {ei }i∈{1: d} is the canonical Cartesian basis of R . The rest of the proof
consists of verifying that the assumptions (i)–(iii) from Lemma 54.2 are met; see Exercise 55.2.
Remark 55.9 (Literature). The reader is referred to Boffi et al. [65, §8.4.3] for other details
P2 , Pb0 ) pair. In general, this pair is not stable in dimension 3, but it is shown in Zhang
on the (P
and Zhang [404] that one can construct special families of tetrahedral meshes for which stability
holds.
55.2.3 P2 -bubble, Pb1 ) pair

The (P
Let bb be the bubble function defined in (54.5) and Pb := P2,d ⊕ (span{bb})d . Let (Th )h∈H be a shape-
regular family of affine simplicial meshes. Recalling that we are enforcing homogeneous Dirichlet
conditions on the velocity, the (PP2 -bubble, Pb1 ) pair gives the following approximation spaces:
g b
Vh0 := P2,0 (Th ) ⊕ Bh , Qh := P1,∗ (Th ), (55.8)
L
with Bh := K∈Th (span{bK })d and bK := bb ◦ TK is the bubble function associated with the mesh
cell K ∈ Th . Notice that
Vh0 := {vh ∈ C 0 (D) | ∀K ∈ Th , vh ◦ TK ∈ Pb , vh|∂D = 0}. (55.9)
Since the pressure is locally P1 on each simplex and globally discontinuous, its local degrees of
freedom can be taken to be its mean value and its gradient in each mesh cell. A conventional
representation is shown in Figure 55.1. We have the following result (see Boffi et al. [65, p. 488]).
Proposition 55.10 (P P2 -bubble, Pb1 ). The (PP2 -bubble, Pb1 ) pair satisfies the inf-sup condition (55.1)
uniformly w.r.t. h ∈ H. Moreover, this pair leads to the same error estimates as the Taylor–Hood
element, that is, µ|u − uh |H 1 (D) + kp − ph kL2 (D) ≤ ch2 (µ|u|H 3 (D) + |p|H 2 (D) ), and if the assump-
tions of Theorem 53.19 are met for some s ∈ (0, 1], then µku − uhkL2 (D) ≤ ch2+s ℓ1−s D (µ|u|H 3 (D) +
|p|H 2 (D) ).
Dimension 2 Dimension 3
velocity pressure velocity pressure
Figure 55.1: Conventional representation of the (P P2 -bubble, Pb1 ) pair in dimensions two (left) and
three (right, only visible degrees of freedom of the velocity are shown). Among various possibilities,
the degrees of freedom for the pressure here are the mean value (indicated by a dot) and the d
components of the gradient (indicated by arrows).
Remark 55.11 (Literature). The (P P2 -bubble, Pb1 ) pair is also called conforming Crouzeix–
Raviart mixed finite element [151].
55.3 Scott–Vogelius elements and generalizations

g
Let k ≥ 1. The (PPk , Pbk−1 ) pair is interesting since ∇·Pk,0 b
(Th ) ⊂ Pk−1,∗ (Th ), which implies that any
g 2 b
vector field in Pk,0 (Th ) whose divergence is L -orthogonal to Pk−1,∗ (Th ) is exactly divergence-free.
55.3.1 Special meshes
Pk , Pbk−1 ) pair does not satisfy the inf-sup condition (55.1) (e.g., we have seen in
In general, the (P
§53.4.3 that for k = 1, this pair suffers from locking). However, it is possible to construct special
meshes so that this element satisfies the inf-sup condition (55.1) uniformly w.r.t. h ∈ H for some
k. Let us now introduce some special meshes to substantiate this claim. Various two-dimensional
examples of such meshes are shown in Figure 55.2.
Irregular crisscross: A two-dimensional triangulation Th is said to be an irregular crisscross
mesh if it is obtained from a matching mesh of D ⊂ R2 composed of quadrangles, where each
quadrilateral cell is divided along its two diagonals; see the leftmost panel in Figure 55.2.
Simplicial barycentric (d + 1)-sected: We say that Th is a simplicial barycentric (d + 1)-
sected mesh in Rd if Th is obtained after refinement of a simplicial matching mesh by subdividing
each initial simplex into (d+1) sub-simplices by connecting the barycenter with the (d+1) vertices.
Simplicial barycentric (d + 1)-sected meshes are also called Hsieh–Clough–Tocher (HCT) meshes
in the literature; see the second panel from the left in Figure 55.2.
Twice quadrisected crisscrossed: We say that a two-dimensional triangulation Th is twice
quadrisected crisscrossed if it is formed as follows. First, the polygon D is partitioned into a
matching mesh of quadrangles, say Q4h . Then, each quadrangle in Q4h is divided into four new
quadrangles by connecting the point at the intersection of its two diagonals with the midpoint on
each of its edges. The mesh Q2h thus formed is subdivided once more by repeating this process.
Finally, Th is obtained by dividing each quadrangle in Qh along its two diagonals, thereby giving
4 triangles per quadrangular cell in Qh , or 64 triangles for each quadrangle in Q4h ; see the third
panel from the left in Figure 55.2.
Powell–Sabin: A simplicial mesh of a polygon or polyhedron D is said to be a Powell–Sabin
mesh if it is constructed as follows. For instance, assuming that the space dimension is two, let
Th be an affine simplicial matching mesh of D. For each triangle K ∈ Th , let cK be the center of
the inscribed circle of K and assume that cK ∈ K for all K ∈ Th . We then divide K into three
triangles by connecting cK to the three vertices of K (this is similar to an HCT triangulation).
Each of the newly created triangles is divided again by connecting cK to cK1 , cK2 , and cK3 , where
K1 , K2 , and K3 are the three neighbors of K (or cK is connected to the midpoint of the edge if
the corresponding neighbor does not exist). The same construction can be done in Rd as shown in
Zhang [403, Fig. 1]. This construction is illustrated in the rightmost panel in Figure 55.2.
Figure 55.2: Irregular crisscross mesh (left). Simplicial barycentric trisected mesh also called
Hsieh–Clough–Tocher (HCT) mesh (center left). One quadrangular cell that is twice quadrisected
and crisscrossed (center right). Powell–Sabin mesh (right).
55.3.2 Pk , Pbk−1) pairs on special meshes

Stable (P
Pk , Pbk−1 ) pair has been thoroughly investigated in dimension two by Scott
The stability of the (P
and Vogelius [345].
Lemma 55.12 ((P Pk , Pbk−1 ), k ≥ 4, d = 2). Let d = 2 and k ≥ 4. Assume that the mesh sequence
(Th )h∈H is quasi-uniform. Assume also that any pair of edges meeting at an internal vertex does
not form a straight line. (An internal vertex violating this property is called singular vertex; see
Exercise 54.3.) The (P Pk , Pbk−1 ) pair satisfies the inf-sup condition (55.1) uniformly w.r.t. h ∈ H.
P3 , Pb2 ) pair, the (P

There are extensions of the above result to the (P P2 , Pb1 ) pair, and the (P
P1 , Pb0 )
pair in dimension two on some of the special meshes described above; see Qin [328].
Lemma 55.13 (Crisscross meshes, k ∈ {2, 3}, d = 2). Let (Th )h∈H be a shape-regular sequence
P 2 , Pb1 ) pair and the (P
of irregular crisscross meshes. Then the (P P3 , Pb2 ) pair have as many spurious
pressure modes as singular vertices, but the velocity approximation is optimal, and the pressure
approximation in the L2 -orthogonal complement to the spurious modes is optimal.
Proof. See [328, Thm. 4.3.1 & 6.2.1].
Lemma 55.14 (HCT meshes, k ∈ {2, 3}, d = 2). Let (Th )h∈H be a shape-regular sequence of
P2 , Pb1 ) pair and the (P
barycentric trisected triangulations. Then the (P P3 , Pb2 ) pair satisfy the inf-sup
condition (55.1) uniformly w.r.t. h ∈ H, and therefore lead to optimal error estimates.
Proof. These statements are proved in Qin [328, Thm. 4.6.1 & 6.4.1]. We detail the proof for the
P2 , Pb1 ) pair since it illustrates the use of the macroelement technique from Corollary 55.3. Here,
(P
Vh0 := P g2,0 (Th ) and Qh := Pb1,∗ (Th ).
(1) Let (Uh )h∈H be the sequence of triangulations that is used to create (Th )h∈H by barycentric
trisection. For every triangle U ∈ Uh , we consider the spaces Vh0 (U ), Qh (U ), Qh (U ), and Q e h (U )
e
defined in (55.2). We also consider the spaces Qh , Qh defined in (55.3). We are going to prove the
inf-sup conditions (55.4a) and (55.4b) in Corollary
P 55.3.
b
(2) Proof of (55.4b). We have Qh := U∈Uh Q h (U ) = P0,∗ (Uh ). Since, as established in
g b
Lemma 55.8, the (P2,0 (Uh ), P0,∗ (Uh )) pair satisfies an inf-sup condition uniformly w.r.t. h ∈ H,
g g
and P2,0 (Uh ) ⊂ P2,0 (Th ) =: Vh0 , we infer that the inf-sup condition (55.4b) is satisfied uniformly
w.r.t. h ∈ H.
(3) Proof of (55.4a). Let U b be the reference simplex in R2 . For every U ∈ Uh , let TU : U b → U be
the corresponding affine geometric mapping. Let us set
b ) := {ψU
V (U d
(vh ) | vh ∈ Vh0 (U )},
Q(Ub ) := {ψ (qh ) | qh ∈ Qh (U )},
g e U
Q( b ) := {ψ g (qh ) | qh ∈ Q
e h (U )},
U U
g d g
where ψU is the pullback by TU and ψU is the contravariant Piola transformation, i.e., ψU (q) :=
d −1
q◦TU and ψU (v) := det(JU )JU (v◦TU ) (see Definition 9.8). One can verify that both spaces V (U b)
e U
and Q( b ) are 8-dimensional, whereas the space Q(U b ) is 9-dimensional. Let B b : V (Ub ) → Q(U b)
R
be defined by (B(bb v ), qb)L2 (U)
b = U b(b
b q x)∇·bv (b
x) db
x for all (b b b
v , qb) ∈ V (U )×Q(U ). A lengthy but
straightforward computation (see Exercise 55.5) shows that im(B) b ⊥ = span(1b ), where ⊥ means
U
b e b
the L -orthogonal complement in Q(U ). Since Q(U ) = (span(1b ))⊥ , this result implies that
2
U
b : V (U
B b ) → Q(
e Ub ) is surjective. (Actually, B b is bijective since dim(V (U b )) = dim(Q(
e Ub )).) Hence,
we have R
| Ub qb(b
x)∇·bv (bx) dbx|
inf sup =: βb1 > 0,
e b
qb∈Q(U) v b)
b∈V (U kb
q k b
Q(U) kb
v k b
V (U)
with kvkV (U)b := |b v |H 1 (Ub ) and kb

qkQ(U) b := kbqkL2 (U)
b . Using the scaling inequality (11.7b) and
the regularity of the mesh sequence (Uh )h∈H , we infer that there is c1 > 0 s.t. c1 kvkV kqkQ ≤
kb
vk b kb q kQ(U)
b for all v ∈ Vh0 (U ), all q ∈ Qh (U ), all U ∈ Uh , and all h ∈ H. Observing that
R V (U) R
b
U
b
q (b
x )∇·b
v (b
x ) x = U q(x)∇·v(x) dx (see Exercise 14.3(i)), we infer that
db
R
| U q(x)∇·v(x) dx|
inf sup =: β1 ≥ c1 βb2 > 0, (55.10)
e
q∈Q(U) v∈Vh0 (U) kqkQh kvkV
i.e., the inf-sup condition (55.4a) is satisfied uniformly w.r.t. h ∈ H.

The analysis of the (P P1 , Pb0 ) pair is a little bit more subtle since filtering the spurious pressure
modes is not enough to approximate the velocity and the pressure properly on general meshes, but
filtering is sufficient on twice quadrisected crisscrossed meshes or Powell–Sabin meshes.
Lemma 55.15 ((P P1 , Pb0 )). Let (Th )h∈H be a shape-regular mesh sequence of either twice quadri-
sected crisscrossed meshes or Powell–Sabin meshes. Then the (P P1 , Pb0 ) pair optimally approximates
1
the velocity of the Stokes problem (i.e., first-order in the H -seminorm) and the approximation of
the pressure is optimal as well after post-processing the spurious pressure modes.
Proof. See Qin [328, Thm. 7.4.2], Zhang [402].
Three-dimensional extensions of the above results are available.
Lemma 55.16 ((P Pk , Pbk−1 ), d = 3). Let (Th )h∈H be a shape-regular sequence of simplicial barycen-
tric quadrisected meshes in R3 . The (P Pk , Pbk−1 ) pair is uniformly stable for all k ≥ 3.
Proof. See Zhang [401, Thm. 5].

Lemma 55.17 ((P P2 , Pb1 ), d = 3). Let (Th )h∈H be a shape-regular sequence of Powell–Sabin simpli-
3 P2 , Pb1 ) pair optimally approximates the velocity and after post-processing
cial meshes in R . The (P
the spurious modes, the approximation of the pressure is optimal as well.
Proof. See Zhang [403, Thm. 4.1].
55.4 Nonconforming and hybrid methods

In this section, we review some nonconforming and some hybrid discretization methods. Let us
start with a nonconforming approximation technique based on the Crouzeix–Raviart finite element
studied in Chapter 36. Let (Th )h∈H be a shape-regular sequence of affine simplicial meshes. Let
cr
P1,0 (Th ) be the Crouzeix–Raviart finite element space with homogeneous Dirichlet conditions (see
cr
(36.8)). Recall that P1,0 (Th ) is composed of piecewise affine functions with continuous mean value
Pcr b
across the mesh interfaces and zero mean value at the boundary faces. The (P 1 , P0 ) pair gives
the following approximation spaces:
cr b
Vh0 := P1,0 (Th ), Qh := P0,∗ (Th ), (55.11)
cr cr
where P1,0 (Th ) is composed of vector-valued functions with each Cartesian component in P1,0 (Th ).
1,p
Observe that Vh0 is nonconforming in W0 (D). The conventional representation of the (P P1 , Pb0 )
cr
pair is shown in Figure 55.3.
Dimension 2 Dimension 3
velocity pressure velocity pressure
Pcr b
Figure 55.3: Conventional representation of the (P 1 , P0 ) pair in dimensions two (left) and three
(right, only visible velocity degrees of freedom are shown). The pressure degree of freedom is the
average over each mesh cell.
To avoid technicalities related to the discrete version of Korn’s inequality in Vh0 (see §42.4.1),
we assume in this section that the momentum equation in the Stokes equations is written in the
Laplacian (or Cauchy–Navier)R form (see Remark 53.3), i.e., we replace the bilinear form a defined
in (53.5) by a(v, w) := D µ∇v:∇w dx. Since Vh0 is nonconforming, we define the following
discrete counterparts of the bilinear forms a and b:
X Z X Z
ah (vh , wh ) := µ∇vh :∇wh dx, bh (vh , qh ) := − qh ∇·vh dx,
K∈Th K K∈Th K
and consider the following discrete problem:



 Find uh ∈ Vh0 and ph ∈ Qh such that
ah (uh , vh ) + bh (vh , ph ) = F (vh ), ∀vh ∈ Vh0 , (55.12)

 b (u , q ) = G(q ),
h h h h ∀qh ∈ Qh ,
R
where the linear R forms on the right-hand side are defined as before as F (vh ) := D f ·vh dx and
G(qh ) := − D P gqh dx. Let p ∈ (1, ∞) and let us equip Vh0 with the mesh-dependent norm
|vh |pW 1,p (Th ) := K∈Th |vh |pW 1,p (K) (the same reasoning as in the proof of Lemma 36.4 shows that
vh 7→ |vh |W 1,p (Th ) is indeed a norm on Vh0 ).
Lemma 55.18 (Stability). Let p, p′ ∈ (1, ∞) be s.t. 1p + p1′ = 1. There is β0 such that for all
h ∈ H,
|bh (vh , qh )|
inf sup ≥ β0 > 0. (55.13)
qh ∈Qh vh ∈Vh0 |vh |W 1,p (Th ) kqh kLp′ (D)
Proof. For all r ∈ Lp∗ (D), there is vr ∈ W01,p (D) s.t. ∇·vr = r and |vr |W 1,p (D) ≤ ckrkLp (D)
1,p
(see Remark 53.10). Let I crh0 : W0 (D) → Vh0 be the vector-valued Crouzeix–Raviart interpo-
lation operator. Owing to the local commuting property established in Exercise 36.1, we have
I cr
bh (I h0 (vr ) − vr , qh ) = 0 for all qh ∈ Qh . Since
Z
I cr
qh r dx = b(vr , qh ) = bh (vr , qh ) = bh (I h0 (vr ), qh ),
D
we infer that
R
| q r dx|
D h I cr
|bh (I h0 (vr ), qh )|
kqh kLp′ (D) ≤ sup = sup
r∈Lp
∗ (D)
krkLp (D) p
r∈L∗ (D) krk Lp (D)
|bh (vh , qh )| I cr
|I h0 (vr )|W 1,p (Th )
≤ sup × sup .
vh ∈Vh0 |vh |W 1,p (Th ) r∈Lp∗ (D) krkLp (D)
Using the W01,p -stability of I cr

h0 (see Lemma 36.1 with r := 0) together with the above bound on
I cr
|I h0 (vr )|W 1,p (Th)
vr , we conclude that supr∈Lp∗ (D) krkLp (D) is uniformly bounded w.r.t. h ∈ H. This proves
the expected inf-sup condition.
Pcr b
Remark 55.19 (Convergence rate). The (P 1 , P0 ) pair is first-order accurate. More precisely,
let (u, p) solve (53.6) and assume that u ∈ H (D) ∩ H01 (D), p ∈ H 1 (D) ∩ L2∗ (D). Then the
2
solution to (53.14) with (Vh0 , Qh ) defined in (55.11) satisfies µk∇h (u − uh)kL2 (D) + kp− phkL2 (D) ≤
ch(µ|u|H 2 (D) + |p|H 1 (D) ). Moreover, if the assumptions of Theorem 53.19 are met for some s ∈
(0, 1], we have µku − uh kL2 (D) ≤ ch1+s ℓ1−s
D (µ|u|H 2 (D) + |p|H 1 (D) ); see Exercise 55.4.
Pcr b
Remark 55.20 (Literature). The (P 1 , P0 ) pair has been introduced by Crouzeix and Raviart
[151]. A quadrilateral nonconforming mixed finite element has been introduced by Rannacher and
Turek [330, 366].
Remark 55.21 (Fortin operator). The proof of Lemma 55.18 shows that the Crouzeix–Raviart
interpolation operator acts as a nonconforming Fortin operator. Indeed, we have ∇·(I I cr
h0 (v)) =
0 1,p
ΠK (∇·v) for all v ∈ W0 (D) and all K ∈ Th (see Exercise 36.1), and since any qh ∈ Qh is
piecewise constant, this implies that bh (I I cr
h0 (v) − v, qh ) = 0. Moreover, there is γ0 > 0 s.t.
I h0 (v)|W 1,p (Th ) ≤ |v|W 1,p (D) for all v ∈ W 1,p (D) and all h ∈ H.
γ0 |I cr
An arbitrary-order discretization of the Stokes equations can be done by using the hybrid high-
order (HHO) method introduced in §39.1. The method uses face-based and cell-based velocities
together with discontinuous cell-based pressures. Let k ∈ N denote the degree of the velocity and
pressure unknowns. As in Di Pietro et al. [169], one can take any k ≥ 0 if one uses the Cauchy–
Navier form of the momentum equation (see Remark 53.3). If one uses instead the formulation
based on the linearized strain tensor (i.e., (53.1a) with (53.2)), then one can adapt the HHO
method for the linear elasticity equations from Di Pietro and Ern [166] (see §42.4.3). In this case,
one takes k ≥ 1 since the analysis invokes a Korn inequality in each mesh cell. In practice, the size
of the linear system can be significantly reduced since one can eliminate locally all the cell-based
velocities and all the (cell-based) pressures up to a constant in each cell. The size of the linear
system is thus reduced to dim(Pk,d−1 )×d×Nf + Nc , where Nf and Nc are the number of mesh
faces and cells, respectively. Other methods using similar discrete unknowns are the hybridizable
discontinuous Galerkin (HDG) methods developed by Egger and Waluga [184], Cockburn and Shi
[132], and the related weak Galerkin methods from Wang and Ye [387]. See also Lehrenfeld and
Schöberl [281] for HDG methods with H(div)-velocities and Jeon et al. [254] for hybridized finite
elements.
Pcr b
Remark 55.22 (Well-balanced scheme). For the (P 1 , P0 ) pair, the discrete velocity fields
are divergence-free locally in each mesh cell, but since Vh0 is nonconforming in H(div; D) (the
normal component of fields in Vh0 can jump across the mesh interfaces), these fields are generally
not divergence-free in D. Recalling Remark 53.22, this means that the discretization is not well-
balanced, and this can lead to a poor velocity approximation in problems with large curl-free body
forces. This issue has been addressed in Linke [283], where a well-balanced scheme is designed by
using a lifting operator mapping the velocity test functions to the lowest-order Raviart–Thomas
space in order to test the body forces in the discrete momentum balance equation. A similar
modification is possible for the HHO discretization by using a lifting operator mapping the veloc-
ity test functions to the Raviart–Thomas space of the same degree as the face-based velocities;
see [169].
55.5 Stable pairs with Q k -based velocities

It is possible to used mixed finite elements based on quadrangular and hexahedral meshes. Since
the literature on the topic is vast and this chapter is just meant to be a brief overview of the field,
we only mention a few results. We assume in the entire section that (Th )h∈H is a shape-regular
sequence of affine meshes composed of cuboids. We start with a negative result.
Lemma 55.23 ((Q Qk , Qbk−1 )). The (Q

Qk , Qbk−1 ) pair composed of continuous Q k elements for the
velocity and discontinuous Qk−1 elements for the pressure does not satisfy the inf-sup condition
for all k ≥ 1.
Proof. This result is established in Brezzi and Falk [92, Thm. 3.2]. A proof is proposed in Exer-
cise 55.3.
It is possible to save the situation by removing some degrees of freedom in the pressure space.
This can be done by considering the polynomial space Pbl instead of Qbl with l ∈ {0, 1}.
Lemma 55.24 ((Q Q 2 , Pb0 )). The (Q

Q2 , Pb0 ) pair satisfies the inf-sup condition (53.15) uniformly
2
w.r.t. h ∈ H in R .
Proof. The proof is the same as that for the (P P2 , Pb0 ) pair. For every face/edge F ∈ Fh and
g
every vh ∈ Q 2,0 (Th ), v|F ·nF is quadratic and one can use Simpson’s quadrature rule to compute
R
F vh ·nF ds; see Exercise 55.2.
Lemma 55.25 ((Q Q 2 , Pb1 )). The (Q

Q2 , Pb1 ) pair satisfies the inf-sup condition (53.15) uniformly
w.r.t. h ∈ H in R2 and yields the same error estimates as the Taylor–Hood mixed finite element.
Proof. The proof is similar to that of the (P P2 , Pb1 ) pair. The reader is referred to Boffi et al. [65,
§8.6.3.1] for other details and a literature review.
Remark 55.26 (Q1 geometric transformation). Let us assume that for all K ∈ Th , the
geometric finite element that is used to construct the cells in Th is the Lagrange Q1 element;
see §8.1. Then the (QQ 2 , Pb1 ) pair satisfies the inf-sup condition (53.15) uniformly w.r.t. h ∈ H
2
in R (the proof is the same as that of Lemma 55.25), but, as shown in Arnold et al. [22], the
approximation properties are suboptimal since in this case the polynomial space P1 is not rich
enough to ensure optimal approximability of the pressure.
Exercises
R
Exercise 55.1 (Local mass balance). Let uh ∈ Vh0 and g ∈ L2∗ (D) satisfy D qh ∇·uh dx =
R b
R g −1 R g −1
D qh g dx for all qh ∈ Pk,∗ (Th ). Show that K (ψK ) (q)∇·uh dx = K (ψK
R R
) (q)g dx for all
g
q ∈ Pk,d and all K ∈ Th with ψK (q) := q ◦ TK . (Hint : use that D ∇·uh dx = D g dx = 0.)
Exercise 55.2 ((P P2 , Pb0 )). Complete the proofR of Lemma 55.8. (Hint : to show that the assump-
tion (ii) from Lemma 54.2 is met, prove that F (v − Π2h (v)) ds = 0 for all F ∈ Fh◦ using Simp-
son’s quadrature rule; to show that the assumption (iii) is met, show first that |Π2h (v)|W 1,p (K) ≤
1
p −1
P
chK F ∈F ◦ kvkLp (F ) and then invoke the multiplicative trace inequality (12.16).)
K
Exercise 55.3 ((Q Q k , Qbk−1 )). (i) Justify Lemma 55.23 for k := 2 by constructing a counterexample.
(Hint : given an interior vertex of a uniform Cartesian mesh, consider the patch composed of
the four square cells sharing this vertex, and find an oscillating pressure field using (ii) from
Exercise 54.3.) (ii) Generalize the argument for all k ≥ 2.
Pcr b
Exercise 55.4 ((P 1 , P0 )). Justify the claim in Remark 55.19. (Hint : see the proof of Theo-
rem 36.11.)
Exercise 55.5 ((P P 2 , Pb1 ), HCT mesh). Using the notation from the proof of Lemma 55.14, the
goal is to prove that im(B) b ⊥ = span(1b ). Let zb1 := (0, 0), zb2 := (1, 0), zb3 := (0, 1), zb4 := ( 1 , 1 ).
U 3 3
Consider the triangles K b 1 := conv(b b 2 := conv(b
z1 , zb2 , zb4 ), K b 3 := conv(b
z2 , zb3 , zb4 ), and K z3 , zb1 , zb4 ).
b ) with the reference macroelement U
Let p ∈ P1b (U b := {K
b 1, K
b2, K b 3 }, and set
p1 := p|K z1 ), p2 := p|K
b 1 (b z2 ), p3 := p|K
b 1 (b b 1 (b
z4 ),
q1 := p|Kb 2 (b
z2 ), q2 := p|K z3 ), q3 := p|K
b 2 (b b 2 (b
z4 ),
s1 := p|K z3 ), s2 := p|K
b 3 (b z1 ), s3 := p|K
b 3 (b b 3 (b
z4 ).
Let m c14 := 21 (b
z1 + zb4 ), m c24 := 21 (bz2 + zb4 ), mc34 := 12 (b
z3 + zb4 ). Let u ∈ P2,0 g b ) and set
(U
T := T := T := T :=
(u7 , v7 ) m14 ), (u8 , v8 )
u(c m24 ), (u9 , v9 )
u(c m34 ), (u10 , v10 )
u(c z4 ). (i) Show (or
u(b
accept as a fact) that
Z
p∇·u db x = (−u7 + u8 + 4v7 + 2v8 )p1
b1
K
+ (−u7 + u8 + v7 + 5v8 )p2 + (−2u7 + 2u8 − v7 + v8 + 3v10 )p3 .
(Hint : compute the P2 shape functions on K b 1 associated with the nodes m

c14 , m
c24 , and zb4 .) (ii)
b b b b
Let TKb 2 : K1 → K2 , TKb 3 : K1 → K3 be the geometric mappings s.t.

−1 −1 0 1
x) := zb2 +
TKb 2 (b x − zb1 ),
(b x) := zb3 +
TKb 3 (b x − zb1 ).
(b
1 0 −1 −1
Verify that TKb i maps the vertices of K b 1 to the vertices of K b i for i ∈ {2, 3}. (iii) Compute the
d d
R
contravariant Piola tranformations ψK b 2 (v) and ψK b 3 (v). (iv) Compute K b i p∇·u dbx for i ∈ {2, 3}.
R R g d
(Hint : use Steps (i) and (iii), and Kb i q∇·v db x= K b 1 ψKi (q)∇·(ψKi (v)) db x (see Exercise 14.3(i)).)
R
b
(v) Write the linear system corresponding to the statement (B(u), p) 2 b := b p∇·u db x = 0 for
L (U) U
g
all u ∈ P2,0 b ), and compute im(B)
(U b ⊥.
Exercise 55.6 (Macroelement partition). Reprove Corollary 55.3 without invoking the par-
tition lemma (Lemma 55.1). (Hint : see Brezzi and Bathe [91, Prop.4.2].)
Exercise 55.7 (Macroelement, continuous pressure). Let the assumptions of Proposition 55.5
hold true. (i) Show that there are c1 , c2 > 0 s.t.
|b(vh , qh )| X 12
sup ≥ c1 βD kqh kQ − c2 h2U |qh |2H 1 (U) ,
vh ∈Vh0 kvh kV
U∈Uh
for all qh ∈ Qh and all h ∈ H. (Hint: use the quasi-interpolation

R operator I av
h0 and proceed as in
1
the proof of Lemma 54.3.) (ii) Setting q hU := |U| q
U h dx, show that there is c s.t. |qh|U |H 1 (U) ≤
ckqh − q hU kL2 (U)
b for all U ∈ Uh and all h ∈ H. (Hint : use Lemma 11.7 and the affine geometric
b
mapping TU : U → U .) (iii) Prove Corollary 55.5. (Hint : use Remark 55.4. See also Brezzi and
Bathe [91, Prop 4.1].)
Appendix C
Bijective operators in Banach

spaces
The goal of this appendix is to recall fundamental results on linear operators (that is, bounded
linear maps) in Banach and Hilbert spaces, and in particular to state conditions allowing us to assert
the bijectivity of these operators. The results collected herein provide a theoretical framework for
the mathematical analysis of the finite element method. We refer the reader to Aubin [29], Brezis
[89], Lax [278], Rudin [337], Yosida [398], Zeidler [400] for further reading.
C.1 Injection, surjection, bijection

Since we are interested in asserting the bijectivity of bounded linear maps in Banach and Hilbert
spaces, let us first recall some basic notions concerning injectivity, surjectivity, and bijectivity, as
well as left and right inverses.
Definition C.1 (Injection, surjection, bijection). Let E and G be two nonempty sets. A
function (or map) f : E → G is said to be injective if every element of the codomain (i.e., G) is
mapped to by at most one element of the domain (i.e., E). The function is said to be surjective
if every element of the codomain is mapped to by at least one element of the domain. Finally, f
is said to be bijective if every element of the codomain is mapped to by exactly one element of the
domain (i.e., f is both injective and surjective).
Definition C.2 (Left and right inverse). Let E and G be two nonempty sets and let f : E → G
be a function. We say that f ‡ : G → E is a left inverse of f if (f ‡ ◦ f )(e) = e for all e ∈ E, and
that f † : G → E is a right inverse of f if (f ◦ f † )(g) = g for all g ∈ G.
A map with a left inverse is necessarily injective. Conversely, if the map f : E → G is injective,
the following holds true: (i) The map f˜ : E → f (E) such that f˜(e) = f (e) for all e ∈ E has a
unique left inverse; (ii) One can construct a left inverse f ‡ : G → E of f by setting f ‡ (g) := e
(with e ∈ E arbitrary) if g 6∈ f (E) and f ‡ (g) := (f˜)‡ (g) otherwise; (iii) If E, G are vector spaces
and the map f is linear, the left inverse of f˜ is also linear. A map with a right inverse is necessarily
surjective. Conversely, one can construct right inverse maps for every surjective map by invoking
the axiom of choice.
378 Appendix C. Bijective operators in Banach spaces
C.2 Banach spaces

Basic properties of Banach and Hilbert spaces are collected in Appendix A. In this section, we
recall these properties and give more details. To stay general, we consider complex vector spaces,
i.e., vector spaces over the field C of complex numbers. The case of real vector spaces is recovered
symbol ·, and by interpreting |·| as the absolute value instead of the complex modulus. Recall that
a complex vector space V equipped with a norm k·kV is said to be a Banach space if every Cauchy
sequence in V has a limit in V.
Let V, W be complex vector spaces. The complex vector space composed of the bounded linear
maps from V to W is denoted by L(V ; W ). Members of L(V ; W ) are often called operators. This
space is equipped with the norm
kA(v)kW
kAkL(V ;W ) := sup < ∞, ∀A ∈ L(V ; W ). (C.1)
v∈V kvkV
In this book, we systematically abuse the notation by implicitly assuming that the argument in
this type of supremum or infimum is nonzero. If W is a Banach space, then L(V ; W ) equipped
with the above norm is also a Banach space (see Rudin [337, p. 87], Yosida [398, p. 111]).
Theorem C.3 (Banach–Steinhaus). Let V, W be Banach spaces and let {Ai }i∈I be a collection
of operators in L(V ; W ) (the set I is not necessarily countable). Assume that supi∈I kAi (v)kW is
a finite number for all v ∈ V. Then there is a real number C such that
sup kAi (v)kW ≤ CkvkV , ∀v ∈ V. (C.2)
i∈I
Proof. See Brezis [89, p. 32], Lax [278, Chap. 10].

Corollary C.4 (Pointwise convergence). Let V, W be Banach spaces. Let (An )n∈N be a se-
quence in L(V ; W ) such that for all v ∈ V, the sequence (An (v))n∈N converges as n → ∞ to a
limit in W denoted by A(v) (one says that the sequence (An )n∈N converges pointwise to A). The
(i) supn∈N kAn kL(V ;W ) < ∞.
(ii) A ∈ L(V ; W ).
(iii) kAkL(V ;W ) ≤ lim inf n→∞ kAn kL(V ;W ) .
Proof. The statement (i) follows from the Banach–Steinhaus theorem. Owing to (C.2), we infer
that kAn (v)kW ≤ CkvkV for all v ∈ V and all n ∈ N. Letting n → ∞ yields kA(v)kW ≤ CkvkV ,
and since A is obviously linear, we infer that the statement (ii) holds true. The statement (iii)
results from the bound kAn (v)kW ≤ kAn kL(V ;W ) kvkV for all v ∈ V and all n ∈ N.
Remark C.5 (Uniform convergence on compact sets). Corollary C.4 does not claim that
(An )n∈N converges to A in L(V ; W ), i.e., uniformly on bounded sets. A standard argument shows
however that (An )n∈N converges uniformly to A on compact sets. Let indeed K ⊂ V be a compact
set. Let ǫ > 0. Set C := supn∈N kAn kL(V ;W ) . The real number C is finite owing to Corollary C.4(i).
The set K being compact, there is a finite set of points {xi }i∈I in K such that for all v ∈ K, there
is i ∈ I such that kv − xi kV ≤ (3C)−1 ǫ. Owing to the pointwise convergence of (An )n∈N to A,
there is Ni such that kAn (xi ) − A(xi )kW ≤ 13 ǫ for all n ≥ Ni . Using the triangle inequality and
the statement (iii) above, we infer that
kAn (v) − A(v)kW ≤ kAn (v − xi )kW + kAn (xi ) − A(xi )kW + kA(v − xi )kW ≤ ǫ,
for all v ∈ K and all n ≥ maxi∈I Ni .
Appendix 379
C.3 Hilbert spaces

Let V be a complex vector space equipped with an inner product (·, ·)V : V × V → C. Recall that
the inner product is linear w.r.t. its first argument and antilinear w.r.t. its second argument, i.e.,
(λv, w)V = λ(v, w)V and (v, λw)V = λ(v, w)V for all λ ∈ C and all v, w ∈ V, and that Hermitian
symmetry means that (v, w)V = (w, v)V . The space V is said to be a Hilbert space if it is a
1
Banach space when equipped with the induced norm kvkV := (v, v)V2 for all v ∈ V. Recall the
Cauchy–Schwarz inequality
|(v, w)V | ≤ kvkV kwkV , ∀v, w ∈ V. (C.3)
Notice that we obtain an equality in (C.3) iff v and w are collinear. This follows from kvkV kwkV −
v w 2

ℜ(ξ(v, w)V ) = kvkV 2kwkV kvk V
− ξ kwk V V
for all nonzero v, w ∈ V and all ξ ∈ C with |ξ| = 1.
Remark C.6 (Arithmetic-geometric and Young’s inequalities). Let x1 , . . . , xn be non-

negative real numbers. Using the convexity of the function x 7→ ex , one can show the following
arithmetic-geometric inequality:
1 1
(x1 x2 . . . xn ) n ≤ (x1 + . . . + xn ). (C.4)
n
Moreover, Young’s inequality states that for every positive real number γ > 0,
γ 1
|(v, w)V | ≤ kvk2V + kwk2V , ∀v, w ∈ V. (C.5)
2 2γ
This follows from the Cauchy–Schwarz inequality and (C.4) with n := 2, x1 := γkvk2V , and
x2 := γ −1 kwk2V .
Definition C.7 (Hilbert basis). A sequence (en )n∈N in V is said to be a Hilbert basis of V if
it satisfies the following two properties:
(i) (em , en )V = δmn for all m, n ∈ N.
(ii) The linear space composed of all the finite linear combinations of the vectors in (en )n∈N is
dense in V.
The existence of Hilbert bases is not a natural consequence of the Hilbert space structure, but
the question of the existence of Hilbert bases can be given a positive answer by introducing the
notion of separability.
Definition C.8 (Separability). A Hilbert space V is said to be separable if it admits a countable

dense subset (vn )n∈N .
Not every Hilbert space is separable, but all the Hilbert spaces encountered in this book are
separable (or by default are always assumed to be separable). The main motivation for the notion
of separability is the following result.
Theorem C.9 (Separability and Hilbert basis). Every separable Hilbert space has a Hilbert
basis.

Lemma C.10 (Pareseval). Let (en )n∈N be a Hilbert basis of V. For all u ∈ V, set un :=
P
k∈{0: n} (u, ek )V ek . The following holds true:
X
lim ku − un kV = 0 and kuk2V = |(u, ek )V |2 . (C.6)
n→∞
k∈N
P
Conversely, let (αn )n∈N be a sequence in ℓ2 (C) and set uα,n := k∈{0: n} αk ek for all n ∈ N. Then
converges to some uα in V such that (uα , en )V = αn for all n ∈ N, and we
the sequence (uα,n )n∈N P
have kuα k2V = limn→∞ k∈{0: n} α2k .

A striking consequence of Lemma C.10 is that all separable Hilbert spaces are isomorphic and
isometric with ℓ2 (C).
Remark C.11 (Space VR ). Let V be a complex vector space. By restricting the scaling operation
(λ, v) 7→ λv to (λ, v) ∈ R×V, V can also be equipped with a vector space structure over R,
which we denote by VR (V and VR are the same sets, but they are equipped with different vector
space structures). For instance, if V = Cm , then dim(V ) = m but dim(VR ) = 2m. Moreover,
the canonical set {ek }k∈{1: m} , where the Cartesian components of ek in Cm are ek,l = δkl (the
Kronecker symbol) for all l ∈ {1:m}, is a basis of V, whereas the set {ek , iek }k∈{1: m} with i2 = −1
is a basis of VR . Finally, if V is a complex Hilbert space with inner product (·, ·)V , then VR is a
real Hilbert space with inner product ℜ(·, ·)V .
C.4 Duality, reflexivity, and adjoint operators

Let V be a complex Banach space. Its dual space V ′ is composed of all the antilinear forms
A : V → C that are bounded. The reason we consider antilinear forms is that we employ the
complex conjugate of test functions in the weak formulation of complex-valued PDEs. The action
of A ∈ V ′ on v ∈ V is denoted by hA, viV ′ ,V ∈ C (and sometimes also A(v)). Equipped with the
norm
|hA, viV ′ ,V |
kAkV ′ := sup , ∀A ∈ V ′ , (C.7)
v∈V kvkV
V ′ is a Banach space. In the real case, the absolute value can be omitted at the numerator since
±v can be considered in the supremizing set. In the complex case, the modulus can be replaced
by the real part since v can be multiplied by any ξ ∈ C with |ξ| = 1.
Remark C.12 (Linear vs. antilinear form). If A : V → C is an antilinear form, then A
(defined by A(v) := A(v) ∈ C for all v ∈ V ) is a linear form.
C.4.1 Fundamental results in Banach spaces

Theorem C.13 (Hahn–Banach). Let V be a normed vector space over C and let W be a subspace
of V. Let B ∈ W ′ . There exists A ∈ V ′ that extends B, i.e., A(w) = B(w) for all w ∈ W, and
such that kAkV ′ = kBkW ′ .
Proof. For the real case, see Brezis [89, p. 3], Lax [278, Chap. 3], Rudin [337, p. 56], Yosida [398,
p. 102]. The above statement is a simplified version of the actual Hahn–Banach theorem. For the
complex case, see Lax [278, p. 27], Brezis [89, Prop. 11.23].
Appendix 381
Corollary C.14 (Norm by duality). The following holds true:
|A(v)| |hA, viV ′ V |

kvkV = sup = sup , (C.8)
A∈V ′ kAkV ′ A∈V ′ kAkV ′
for all v ∈ V, and the supremum is attained.

|A(v)|
Proof. Assume v 6= 0 (the assertion is obvious for v = 0). We first observe that supA∈V ′ kAk ≤
V′
:= ′ :=
kvkV since |A(v)| ≤ kAkV kvkV . Let W
′ span{v} and let B ∈ W be defined as B(λv) λkvkV
for all λ ∈ C. By construction, B ∈ W ′ and kBkW ′ = 1. Owing to the Hahn–Banach theorem,
there exists A ∈ V ′ such that kAkV ′ = 1 and A(v) = B(v) = kvkV .
Corollary C.15 (Characterization of density). Let V be a normed vector space over C and W
be a subspace of V. Then W 6= V (i.e., W is not dense in V ) if and only if there exists f ∈ V ′ \{0}
such that f (w) = 0 for all w ∈ W.
Proof. See Brezis [89, p. 8], Rudin [337, Thm. 5.19].
Definition C.16 (Double dual). The double dual of a Banach space V is denoted by V ′′ and
is defined to be the dual space of its dual space V ′ .
Proposition C.17 (Isometry into double dual). The bounded linear map JV : V → V ′′ such
that
hJV (v), φ′ iV ′′ ,V ′ = hφ′ , viV ′ ,V , ∀(v, φ′ ) ∈ V × V ′ , (C.9)
is an isometry.
Proof. The claim follows from Corollary C.14 since
|hJV (v), φ′ iV ′′ ,V ′ | |hφ′ , viV ′ ,V |

kJV (v)kV ′′ = sup = sup = kvkV .
φ′ ∈V ′ kφ′ kV ′ φ′ ∈V ′ kφ′ kV ′
Definition C.18 (Reflexivity). A Banach space V is said to be reflexive if JV is an isomorphism.
Remark C.19 (Map JV ). Since JV is an isometry, it is injective. Thus, V can be identified with
the subspace JV (V ) ⊂ V ′′ . It may happen that the map JV is not surjective. In this case, V is a
proper subspace of V ′′ .
Example C.20 (Lebesgue spaces). One important consequence of Theorem 1.41 is that the
Lebesgue space Lp (D) is reflexive for all p ∈ (1, ∞). However, L1 (D) and L∞ (D) are not reflexive.
Indeed, L∞ (D) = L1 (D)′ , but L1 (D) ( L∞ (D)′ with strict inclusion; see §1.4 and Brezis [89,
p. 102].
Remark C.21 (Space VR ). Let V be a complex vector space and let VR be defined in Re-
mark C.11. Let VR′ be the dual space of VR , i.e., the normed real vector space composed of
the bounded R-linear maps from V to R. Then the map I : V ′ → VR′ s.t. for all ℓ ∈ V ′ ,
I(ℓ)(v) := ℜ(ℓ(v)), for all v ∈ V, is a bijective isometry; see [89, Prop. 11.22].
Definition C.22 (Weak convergence). Let V be a Banach space. The sequence (vn )n∈N in V
is said to converge weakly to v ∈ V if
hA, vn iV ′ ,V → hA, viV ′ ,V , ∀A ∈ V ′ . (C.10)

It is shown in Brezis [89, Prop. 3.5] that if the sequence (vn )n∈N converges strongly to v (that
is, in the norm topology, i.e., kvn − vkV → 0 as n → ∞), then it also converges weakly to v.
The converse is true if V is finite-dimensional (see [89, Prop. 3.6]). Furthermore, if the sequence
(vn )n∈N converges weakly to v, then it is bounded and kvkV ≤ lim inf n→∞ kvn kV . One important
result on weak convergence is the following (see [89, Thm. 3.18]).
Theorem C.23 (Reflexivity and weak compactness). Let V be a reflexive Banach space.
Then from every bounded sequence (vn )n∈N of V, there exists a subsequence (vnk )k∈N that is weakly
convergent.
C.4.2 Further results in Hilbert spaces

Theorem C.24 (Riesz–Fréchet). The operator JVrf : V → V ′ such that
hJVrf (v), wiV ′ ,V := (v, w)V , ∀v, w ∈ V, (C.11)
is a linear isometric isomorphism.
Proof. See Brezis [89, Thm. 5.5], Lax [278, p. 56], Yosida [398, p. 90], or Exercise 25.1.
Remark C.25 (Riesz–Fréchet representation). Theorem C.24 is often called Riesz–Fréchet

representation theorem. It states that for every antilinear form v ′ ∈ V ′ , there exists a unique vector
v ∈ V such that v ′ = JVrf (v). The vector (JVrf )−1 (v ′ ) ∈ V is called Riesz–Fréchet representative of
the antilinear form v ′ ∈ V ′ . The
action of v ′ on V is represented by (JVrf )−1 (v ′ ) with the identity
′ rf −1 ′
hv , wiV ′ ,V = (JV ) (v ), w V for all w ∈ V.
Remark C.26 (Linear vs. antilinear). Notice that JVrf is a linear operator. If we had adopted
the convention that dual spaces were composed of linear forms, we would have had to define JVrf by
setting hJVrf (v), wiV ′ ,V := (v, w)V for all v, w ∈ V, or, equivalently, hv ′ , wiV ′ ,V := ((JVrf )−1 (v ′ ), w)V
for all w ∈ V and v ′ ∈ V ′ . In this case, JVrf would have been antilinear.
Corollary C.27 (Reflexivity). Hilbert spaces are reflexive.
Owing to the Riesz–Fréchet theorem, the notion of weak convergence (see Definition C.22) can
be reformulated as follows in Hilbert spaces.
Definition C.28 (Weak convergence). Let V be a Hilbert space. The sequence (vn )n∈N in V
is said to converge weakly to v ∈ V if (w, vn )V → (w, v)V as n → ∞, for all w ∈ V.
A useful connection between weak and strong convergence in Hilbert spaces is that if the
sequence (vn )n∈N converges weakly to v ∈ V and if additionally, kvn kV → kvkV as n → ∞, then
the sequence (vn )n∈N converges strongly to v, i.e., kvn − vkV → 0 as n → ∞ (see, e.g., Brezis [89,
Prop. 3.32]).
C.4.3 Adjoint
Definition C.29 (Adjoint operator). Let V, W be complex Banach spaces. Let A ∈ L(V ; W ).
The adjoint operator of A is the bounded linear operator A∗ ∈ L(W ′ ; V ′ ) such that
hA∗ (w′ ), viV ′ ,V := hw′ , A(v)iW ′ ,W , ∀(v, w′ ) ∈ V × W ′ . (C.12)
Note that (λA)∗ = λA∗ for all λ ∈ C.

Appendix 383
Lemma C.30 (Norm of adjoint). Let A ∈ L(V ; W ) and let A∗ ∈ L(W ′ ; V ′ ) be its adjoint.
Then kA∗ kL(W ′ ;V ′ ) = kAkL(V ;W ) .
Proof. We have
kA∗ (w′ )kV ′ |hA∗ (w′ ), viV ′ ,V |
kA∗ kL(W ′ ;V ′ ) = sup = sup sup
w ′ ∈W ′ kw′ kW ′ w ′ ∈W ′ v∈V kvkV kw′ kW ′
′
|hw , A(v)iW ′ ,W | kA(v)kW
= sup sup ′
= sup = kAkL(V ;W ) ,
′
v∈V w ∈W ′ kvkV kw kW ′ v∈V kvkV
where we used that supw′ ∈W ′ supv∈V = supv∈V supw′ ∈W ′ , the definition of A∗ , and Corollary C.14.
Definition C.31 (Self-adjoint operator). Let V be a reflexive Banach space. Let A ∈ L(V ; V ′ ),
so that A∗ ∈ L(V ′′ ; V ′ ). The operator A is said to be self-adjoint if A = A∗ ◦JV , i.e., if the following
holds true:
hA(v), wiV ′ ,V = hA(w), viV ′ ,V , ∀v, w ∈ V. (C.13)
In particular, hA(v), viV ′ ,V takes real values if A is self-adjoint. (Notice that if A is self-adjoint,
λA is not self-adjoint if the imaginary part of λ ∈ C is nonzero.) If the spaces V and V ′′ are
actually identified, we write A∗ ∈ L(V ; V ′ ) and say that A is self-adjoint if A = A∗ .
Remark C.32 (Hermitian transpose). If V and W are finite-dimensional and after choosing
one basis for V and one for W, A can be represented by a matrix with complex-valued entries.
Then A∗ is represented in the same bases by the Hermitian transpose of this matrix. Self-adjoint
operators are represented by Hermitian matrices.
C.5 Open mapping and closed range theorems

Let V, W be complex Banach spaces. For A ∈ L(V ; W ), we denote by ker(A) its kernel and by
im(A) its range. The operator A being bounded, ker(A) is closed in V. Hence, the quotient of V
by ker(A), V /ker(A), can be defined. This space is composed of equivalence classes v̌ such that v
and w are in the same class v̌ if and only if v − w ∈ ker(A), i.e., A(v) = A(w).
Theorem C.33 (Quotient space). The space V /ker(A) is a Banach space when equipped with
the norm kv̌k := inf v∈v̌ kvkV . Moreover, the operator Ǎ : V /ker(A) → im(A) s.t. Ǎ(v̌) := A(v) for
all v in v̌, is an isomorphism.
Proof. See Brezis [89, §11.2], Yosida [398, p. 60].
For subspaces M ⊂ V and N ⊂ V ′ , we define the annihilators of M and N as follows:
M ⊥ := {v ′ ∈ V ′ | ∀m ∈ M, hv ′ , miV ′ ,V = 0}, (C.14a)

⊥ ′ ′
N := {v ∈ V | ∀n ∈ N, hn , viV ′ ,V = 0}. (C.14b)
Let M denote the closure of the subspace M in V. A characterization of ker(A) and im(A) is given
by the following result.
Lemma C.34 (Kernel and range). Let A ∈ L(V ; W ). The following holds true:
(i) ker(A) = (im(A∗ ))⊥ .
(ii) ker(A∗ ) = (im(A))⊥ .

(iii) im(A) = (ker(A∗ ))⊥ .
(iv) im(A∗ ) ⊂ (ker(A))⊥ .
Proof. See Brezis [89, Cor. 2.18], Yosida [398, pp. 202-209].
Showing that the range of an operator is closed is a crucial step towards proving that this
operator is surjective. This is the purpose of the following fundamental theorem.
Theorem C.35 (Banach or closed range). Let A ∈ L(V ; W ). The following statements are
equivalent:
(i) im(A) is closed.
(ii) im(A∗ ) is closed.
(iii) im(A) = (ker(A∗ ))⊥ .
(iv) im(A∗ ) = (ker(A))⊥ .
Proof. See Brezis [89, Thm. 2.19], Yosida [398, p. 205].
We now put in place the second keystone of the edifice.
Theorem C.36 (Open mapping). If A ∈ L(V ; W ) is surjective and U is an open set in V, then
A(U ) is an open set in W.
Proof. See Brezis [89, Thm. 2.6], Lax [278, p. 168], Rudin [337, p. 47], Yosida [398, p. 75].
Theorem C.36, also due to Banach, has far-reaching consequences. In particular, it leads to
the following characterization of the closedness of im(A).
Lemma C.37 (Characterization of closed range). Let A ∈ L(V ; W ). The following state-
ments are equivalent:
(i) im(A) is closed in W.
(ii) A has a bounded right inverse map A† : im(A) → V, i.e., (A ◦ A† )(w) = w for all w ∈ im(A),
and there exists α > 0 such that αkA† (w)kV ≤ kwkW for all w ∈ im(A) (A† is not necessarily
linear).
Proof. (i) ⇒ (ii). Since im(A) is closed in W, im(A) is a Banach space. Applying the open mapping
theorem to A : V → im(A) and U := BV (0, 1) (the open unit ball in V ) proves that A(BV (0, 1)) is
open in im(A). Since 0 ∈ A(BV (0, 1)), there is γ > 0 s.t. BW (0, γ) ⊂ A(BV (0, 1)). Let w ∈ im(A).
Since γ2 kwk
w
W
∈ BW (0, γ), there is z ∈ BV (0, 1) s.t. A(z) = γ2 kwk
w
W
. Setting A† (w) := 2kwk
γ
W
z
† γ †
leads to A(A (w)) = w and 2 kA (w)kV ≤ kwkW .
(ii) ⇒ (i). Let (wn )n∈N be a sequence in im(A) that converges to some w ∈ W. The sequence
(vn := A† (wn ))n∈N in V is such that A(vn ) = wn and αkvn kV ≤ kwn kW . Thus, (vn )n∈N is a
Cauchy sequence in V. Since V is a Banach space, (vn )n∈N converges to a certain v ∈ V. Owing to
the boundedness of A, (A(vn ))n∈N converges to A(v). Hence, w = A(v) ∈ im(A).
Corollary C.38 (Bounded inverse). If A ∈ L(V ; W ) is bijective, then A−1 ∈ L(W ; V ).
Proof. Since A is bijective, im(A) = W is closed. Moreover, the right inverse A† is necessarily
equal to A−1 (apply A−1 to A ◦ A† = IW ). Lemma C.37(ii) shows that A−1 ∈ L(W ; V ) with
kA−1 kL(W ;V ) ≤ α−1 .
Appendix 385
C.6 Characterization of surjectivity

As a consequence of the closed range theorem and of the open mapping theorem, we deduce two
characterizations of surjective operators.
Lemma C.39 (Surjectivity of A∗ ). Let A ∈ L(V ; W ). The following statements are equivalent:
(i) A∗ : W ′ → V ′ is surjective.
(ii) A : V → W is injective and im(A) is closed in W.
(iii) There exists α > 0 such that
kA(v)kW ≥ αkvkV , ∀v ∈ V. (C.15)
Equivalently, there exists α > 0 such that
|hw′ , A(v)iW ′ ,W |
inf sup ≥ α. (C.16)
v∈V w ′ ∈W ′ kw′ kW ′ kvkV
Proof. (i) ⇒ (iii). Since the map A∗ is surjective, Lemma C.37 implies that A∗ has a bounded
right inverse map A∗† : V ′ → W ′ . In particular, A∗ (A∗† (v ′ )) = v ′ for all v ′ ∈ V ′ , and there is
α > 0 such that αkA∗† (v ′ )kW ′ ≤ kv ′ kV ′ . Let now v ∈ V. We infer that
|hv ′ , viV ′ ,V | |hA∗ (A∗† (v ′ )), viV ′ ,V | |hA∗† (v ′ ), A(v)iW ′ ,W |

′
= ′
≤ α−1
kv kV ′ kv kV ′ kA∗† (v ′ )kW ′
|hw′ , A(v)iW ′ ,W |
≤ α−1 sup .
w ′ ∈W ′ kw′ kW ′
|hv ′ ,vi ′ |
Since kvkV = supv′ ∈V ′ kv′ kV ′,V , taking the supremum w.r.t. v ′ ∈ V ′ followed by the infimum
V
w.r.t. v ∈ V proves (C.16). Moreover, (C.15) and (C.16) are equivalent owing to Corollary C.14.
(iii) ⇒ (ii). The bound (C.15) implies that A is injective. Consider a sequence (vn )n∈N such that
(A(vn ))n∈N is a Cauchy sequence in W. Then (C.15) implies that (vn )n∈N is a Cauchy sequence in
V. Let v be its limit. A being bounded implies that A(vn ) → A(v). Hence, im(A) is closed.
(ii) ⇒ (i). Since im(A) is closed, we use Theorem C.35(iv) together with the injectivity of A to
infer that im(A∗ ) = (ker(A))⊥ = {0}⊥ = V ′ . This shows that A∗ is surjective.
Lemma C.40 (Surjectivity of A). Let A ∈ L(V ; W ). The following statements are equivalent:
(i) A : V → W is surjective.
(ii) A∗ : W ′ → V ′ is injective and im(A∗ ) is closed in V ′ .
(iii) There exists α > 0 such that
kA∗ (w′ )kV ′ ≥ αkw′ kW ′ , ∀w′ ∈ W ′ . (C.17)
Equivalently, there exists α > 0 such that
|hA∗ (w′ ), viV ′ ,V |

inf ′ sup ≥ α. (C.18)
′
w ∈W v∈V kw′ kW ′ kvkV
Proof. We only prove the implication (i) ⇒ (iii) since the rest of the proof proceeds as above (the
equivalence between (C.17) and (C.18) now follows from the definition of k·kV ′ ). Since the map
A is surjective, Lemma C.37 implies that A has a bounded right inverse map A† : W → V. In
particular, A(A† (w)) = w for all w ∈ W, and there is α > 0 such that αkA† (w)kV ≤ kwkW . Then
for all w′ ∈ W ′ , we have
|hA∗ (w′ ), viV ′ ,V | |hA∗ (w′ ), A† (w)iV ′ ,V |

kA∗ (w′ )kV ′ = sup ≥ sup
v∈V kvkV w∈W kA† (w)kV
′ ′
|hw , wiW ′ ,W | |hw , wiW ′ ,W |
= sup †
≥ α sup = αkw′ kW ′ .
w∈W kA (w)kV w∈W kwkW
Remark C.41 (Lions’ theorem). The assertion (i) ⇔ (iii) in Lemma C.40 is sometimes called
Lions’ theorem. It means that establishing the a priori estimate (C.17) is a necessary and sufficient
condition to prove that the problem A(u) = f has at least one solution u ∈ V for all f ∈ W.
Lemma C.42 (Right inverse). Let V, W be Banach spaces and let A ∈ L(V ; W ) be a surjective
operator. Assume that V is reflexive. Then A has a bounded right inverse A† : W → V satis-
fying αkA† (w)kV ≤ kwkW , where α is the same constant as in the equivalent statements (C.17)
and (C.18).
Proof. The proof is inspired from ideas by P. Azerad (private communication). Let A ∈ L(V ; W )
be a surjective operator. Lemma C.34(ii) shows that the adjoint operator A∗ : W ′ → V ′ is injective.
Let us equip the subspace R := im(A∗ ) ⊂ V ′ with the norm k·kV ′ . The injectivity of A∗ implies
the existence of a linear left inverse A∗‡ : R → W ′ . Consider its adjoint A∗‡∗ : W ′′ → R′ . Let
HB ′ ′′
ER ′ V ′′ be one Hahn–Banach extension operator from R to V (see Theorem C.13). Using the
reflexivity of V to invoke the inverse of the canonical isometry JV : V → V ′′ , we set
A† := JV−1 ◦ ER
HB
′ V ′′ ◦ A
∗‡∗
◦ JW : W → V.
Let us verify that A† satisfies the expected properties. We have for all (w′ , w) ∈ W ′ × W,
hw′ , A(A† (w))iW ′ ,W = hA∗ (w′ ), A† (w)iV ′ ,V

HB (A∗‡∗ (J (w))), A∗ (w′ )i ′′ ′
= hER ′ V ′′ W V ,V
= hA∗‡∗ (JW (w)), A∗ (w′ )iR′ ,R = hJW (w), A∗‡ (A∗ (w′ ))iW ′′ ,W ′
= hJW (w), w′ iW ′′ ,W ′ = hw′ , wiW ′ ,W ,
where to pass from the second to the third line we used that A∗ (w′ ) ∈ R. Since w′ is arbitrary in
W ′ , this proves that A ◦ A† = IW . Moreover, since R := im(A∗ ), we infer that for all w ∈ W,
|hA∗‡∗ (JW (w)), A∗ (w′ )iR′ ,R |

kA† (w)kV = kA∗‡∗ (JW (w))kR′ = sup
w ′ ∈W ′ kA∗ (w′ )kV ′
|hJW (w), w′ iW ′′ ,W ′ | kw′ kW ′
= sup ∗ ′
≤ sup ∗ ′
kwkW .
w ′ ∈W ′ kA (w )kV ′ w ′ ∈W ′ kA (w )kV ′
kw ′ kW ′
Since A ∈ L(V ; W ) is surjective, we have supw′ ∈W ′ kA∗ (w ′ )kV ′ ≤ α−1 owing to (C.17), and this
shows that kA† (w)kV ≤ α−1 kwkW .
Appendix 387
Remark C.43 (Counterexample). The assumption that V be reflexive in Lemma C.42 cannot
be removed if one insists on having the bound α supkwkW =1 kA† (w)kV ≤ 1. Let us consider the
real sequence spaces ℓp , p ∈ [1, ∞]. Since V := ℓ1 is not reflexive, there exists a linear form
A : ℓ1 → W := R that does not attain its norm on the unit ball of V (this is James’s theorem
[253, Thm. 1]). Notice that A 6= 0, hence A is necessarily surjective. Using ℓ2 as pivot space, it
is well known that ℓ∞ can be identified with the dual of V (seeP e.g., Brezis [89, Thm. 4.14]). Let
t be the nonzero sequence in ℓ∞ such that A(v) = (v, t)ℓ2 := i∈N vi ti for all v ∈ V. A simple
computation shows that the adjoint A∗ : R → V ′ ≡ ℓ∞ is such that A∗ (s) = st for all s ∈ R. Let
|(A∗ (w ′ ),v) | |(t,v) |
us define α := inf w′ ∈R supv∈ℓ1 |w′ |kvk 1ℓ2 . We have α = supv∈ℓ1 kvk ℓ12 = kAkV ′ . Let A† be a
ℓ ℓ
right inverse of A. Then for all s ∈ R, we have s = (A ◦ A† )(s) = (t, A† (s))ℓ2 . For all s ∈ R\{0},
A† (s)
kA† (s)kV
is in the unit ball of V. Since A does not attain its norm on this ball by assumption, we
† †
infer that |A( kAA† (s)k
(s)
V
)| < α. Since A( kAA† (s)k
(s)
V
1
) = kA† (s)k V
s, we can rewrite the above bound as
1 † †
α |s| < kA (s)kV for all s ∈ R\{0}, that is, 1 < α supkwkW =1 kA (w)kV .
We observe that nothing is said in Lemma C.42 on the linearity of the right inverse A† . A
slightly different construction of A† that guarantees linearity is possible in the Hilbertian setting.
Lemma C.44 (Right inverse in Hilbert spaces). Let Y, Z be two nontrivial Hilbert spaces.
Let B : Y → Z ′ be a bounded linear operator such that there exists β > 0 s.t.
kB(y)kZ ′ ≥ βkykY , ∀y ∈ Y. (C.19)
Then B ∗ : Z → Y ′ has a linear right inverse B ∗† : Y ′ → Z such that kB ∗† kL(Y ′ ;Z) ≤ β −1 .
Proof. Owing to Lemma C.39, the assumption (C.19) is equivalent to B ∗ : Z → Y ′ being surjective.
Let us set M := ker(B ∗ )⊥ ⊂ Z, where the orthogonality is defined using the inner product of Z
(note that M 6= {0} since otherwise ker(B ∗ ) = Z, i.e., B ∗ = B = 0 implying by (C.19) that
Y = {0} would be trivial). Let J : M → Z be the canonical injection, and note that J ∗ : Z ′ → M ′
is s.t. for all z ′ ∈ Z ′ and all m ∈ M ,
hJ ∗ (z ′ ), miM ′ ,M = hz ′ , J(m)iZ ′ ,Z = hz ′ , miZ ′ ,Z .
Let us set S := J ∗ ◦ B : Y → M ′ . Let y ′ ∈ Y ′ . The surjectivity of B ∗ together with the definition
of M implies that there is z := m + m⊥ ∈ M ⊕ M ⊥ = Z s.t. y ′ = B ∗ (z) = B ∗ (m) = B ∗ (J(m)) =
S ∗ (m), which proves that S ∗ is surjective. Let m ∈ M and assume that 0 = S ∗ (m) = B ∗ (J(m)) =
B ∗ (m). Then m ∈ ker(B ∗ ) ∩ ker(B ∗ )⊥ , i.e., m = 0, which proves that S ∗ : M → Y ′ is injective.
∗ −1
) (y ′ )kZ
Hence, S ∗ and S are isomorphisms. Moreover, since k(S ∗ )−1 kL(Y ′ ;M) = supy′ ∈Y ′ k(S ky ′k ′ =
Y
kmkZ
supm∈M kS ∗ (m)kY ′ , we have
kS ∗ (m)kY ′ |hS ∗ (m), yiY ′ ,Y |
k(S ∗ )−1 k−1
L(Y ′ ;M) = inf = inf sup
m∈M kmkZ m∈M y∈Y kmkZ kykY
|hS(y), miM ′ ,M | |hB(y), miZ ′ ,Z |
= inf sup = inf sup ,
y∈Y m∈M kykY kmkZ y∈Y m∈M kykY kmkZ
where the first equality on the second line follows from (C.25) below and the bijectivity of S. Using
that Z = M ⊕ M ⊥ and M ⊥ = ker(B ∗ ), we obtain
−1 |hB(y), miZ ′ ,Z |
k(S ∗ )−1 kL(Y ′ ;M) = inf sup
y∈Y m∈M kykY kmkZ
|hB(y), m + m⊥ iZ ′ ,Z |
≥ inf sup = β,
y∈Y m+m⊥ ∈M⊕M ⊥ kykY (kmk2Z + km⊥ k2Z )1/2
which proves that k(S ∗ )−1 kL(Y ′ ;M) ≤ β −1 . (Note that we actually have k(S ∗ )−1 kL(Y ′ ;M) = β −1
|hB(y),miZ ′ ,Z | |hB(y),ziZ ′ ,Z |
since supm∈M kmkZ ≤ supz∈Z kzkZ .) Let us now set
B ∗† := J ◦ (B ∗ ◦ J)−1 = J ◦ (S ∗ )−1 : Y ′ → Z.
Then B ∗ ◦ B ∗† = B ∗ ◦ J ◦ (S ∗ )−1 = S ∗ ◦ (S ∗ )−1 = IY ′ , which proves that B ∗† is indeed a right

inverse of B ∗ . Moreover, kB ∗† kL(Y ′ ,Z) = kJ ◦ (S ∗ )−1 kL(Y ′ ,Z) ≤ k(S ∗ )−1 kL(Y ′ ,Z) = β −1 .
Remark C.45 (Lemma C.44 vs. Lemma C.42). Without the statement on the linearity of
B ∗† , Lemma C.44 would be a direct consequence of Lemma C.42 applied with A := B ∗ , V := Z,
and W := Y ′ . Indeed, the condition (C.19) implies that A is a surjective operator satisfying the
inf-sup condition (C.18) with constant β.
Remark C.46 (Left inverse). The operator B ∗‡ := (J ∗ ◦ B)−1 ◦ J ∗ = S −1 ◦ J ∗ : Z ′ → Y is a

left inverse of B s.t. kB ∗‡ kL(Z ′ ;Y ) ≤ β −1 .
Finally, let us recall two important results on compactness.
Lemma C.47 (Peetre–Tartar). Let X, Y, Z be Banach spaces. Let A ∈ L(X; Y ) be injective

and let T ∈ L(X; Z) be compact. Assume that there is c > 0 such that ckxkX ≤ kA(x)kY +kT (x)kZ
for all x ∈ X. Then im(A) is closed. Equivalently, there is α > 0 such that
αkxkX ≤ kA(x)kY , ∀x ∈ X. (C.20)
Proof. Owing to Lemma C.39 and since A is injective, im(A) is closed iff (C.20) holds true. This
inequality has already been proved in Lemma A.20 (see (A.6)).
Theorem C.48 (Schauder). A bounded linear operator between Banach spaces is compact if and
only if its adjoint is compact.
C.7 Characterization of bijectivity

The following theorem provides the theoretical foundation of the BNB theorem stated in §25.3 and
which is often invoked in this book.
Theorem C.49 (Bijectivity of A). Let A ∈ L(V ; W ). The following statements are equivalent:
(i) A : V → W is bijective.
(ii) A is injective, im(A) is closed, and A∗ : W ′ → V ′ is injective.
(iii) A∗ is injective and there exists α > 0 such that
kA(v)kW ≥ αkvkV , ∀v ∈ V. (C.21)
Equivalently, A∗ is injective and
|hw′ , A(v)iW ′ ,W |
inf sup =: α > 0. (C.22)
v∈V w ′ ∈W ′ kw′ kW ′ kvkV
Appendix 389
Proof. (1) The statements (ii) and (iii) are equivalent since (C.21) is equivalent to A injective and
im(A) closed owing to Lemma C.39.
(2) Let us first prove that (i) implies (ii). Since A is surjective, ker(A∗ ) = im(A)⊥ = {0}, i.e., A∗
is injective. Since im(A) = W is closed and A is injective, this yields (ii). Finally, to prove that
(ii) implies (i), we only need to prove that (ii) implies the surjectivity of A. The injectivity of A∗
implies that im(A) = (ker(A∗ ))⊥ = W. Since im(A) is closed, im(A) = W, i.e., A is surjective.
Corollary C.50 (Self-adjoint bijective operator). Assume that V is reflexive. Let A ∈
L(V ; V ′ ) be a self-adjoint operator. Then A is bijective iff there is a real number α > 0 such that
kA(v)kV ′ ≥ αkvkV , ∀v ∈ V. (C.23)
Proof. Owing to Theorem C.49, the bijectivity of A implies that A satisfies the inequality (C.23).
Conversely, (C.23) means that A is injective. It follows that A∗ is injective since A∗ = A ◦ JV−1
owing to the reflexivity hypothesis. The bijectivity of A then follows from Theorem C.49(iii).
Let A ∈ L(V ; W ) be a bijective operator. We have seen in Corollary C.38 that A−1 ∈ L(W ; V ).
We can now characterize more precisely the constants associated with the boundedness of A−1 and
the closedness of its range.
Lemma C.51 (Bounds on A−1 ). Let A ∈ L(V ; W ) be a bijective operator. Then kA−1 kL(W ;V ) =
α−1 with α defined in (C.22), and
kA−1 (w)kV |hv ′ , A−1 (w)iV ′ ,V |

inf = inf sup = kAk−1
L(V ;W ) . (C.24)
w∈W kwkW w∈W v ′ ∈V ′ kv ′ kV ′ kwkW
Proof. (1) Using the bijectivity of A, we have

−1 −1
kA−1 (w)kV kvkV
sup = sup
w∈W kwkW v∈V kA(v)kW
kA(v)kW |hw′ , A(v)iW ′ ,W |
= inf = inf sup ,
v∈V kvkV v∈V w ′ ∈W ′ kw′ kW ′ kvkV
which shows using (C.22) that kA−1 kL(W ;V ) = α−1 .

(2) Similarly, we have
−1
kA−1 (w)kV kA(v)kW
inf = sup = kAkL(V ;W ) .
w∈W kwkW v∈V kvkV
|hv ′ ,A−1 (w)iV ′ ,V |
Since kA−1 (w)kV = supv′ ∈V ′ kv ′ kV ′ owing to Corollary C.14, this proves the inf-sup
condition in (C.24).
Let us finish this section with some useful results concerning the bijectivity of the adjoint
operator and some bounds on its inverse.
Corollary C.52 (Bijectivity of A∗ ). Let A ∈ L(V ; W ) and consider its adjoint A∗ ∈ L(W ′ ; V ′ ).
Then A is bijective if and only if A∗ is bijective.
Proof. Assume first that A is bijective. Since A is injective and im(A) = W, the equivalence of
Items (i) and (ii) in Lemma C.39 implies that A∗ is surjective. Since A is surjective, the equivalence
of Items (i) and (ii) in Lemma C.40 implies that A∗ is injective. Hence, A∗ is bijective. The converse
statement is proved by invoking the same arguments.
Lemma C.53 (Inf-sup condition). Let A ∈ L(V ; W ) be a bijective operator. Assume that V is
reflexive. The following holds true:
|hw′ , A(v)iW ′ ,W | |hw′ , A(v)iW ′ ,W |
inf sup = inf sup . (C.25)
v∈V w ′ ∈W ′ kw′ kW ′ kvkV w ′ ∈W ′ v∈V kw′ kW ′ kvkV
In other words, the inf-sup constant of A ∈ L(V ; W ) on V × W ′ is equal to the inf-sup constant of
A∗ ∈ L(W ′ ; V ′ ) on W ′ × V.
Proof. The left-hand side, l, and the right-hand side, r, in (C.25) are two positive finite numbers
since A is a bijective bounded operator. The left-hand side being equal to l means that l is the
largest number such that kA(v)kW ≥ l kvkV for all v in V. Let w′ ∈ W ′ and w ∈ W. Since A is
surjective, we can consider its right inverse A† , and the previous statement regarding l implies that
l kA† (w)kV ≤ kwkW . Since A(A† (w)) = w, this implies that
|hw′ , wiW ′ ,W | |hA∗ (w′ ), A† (w)iV ′ ,V |
kw′ kW ′ = sup = sup
w∈W kwkW w∈W kwkW
kA† (w)kV 1 1 |hw′ , A(v)iW ′ ,W |
≤ kA∗ (w′ )kV ′ sup ≤ kA∗ (w′ )kV ′ = sup .
w∈W kwkW l l v∈V kvkV
Taking the infimum w.r.t. w ∈ W ′ proves that l ≤ r. The converse inequality r ≤ l is proved
′
similarly by working with W ′ in lieu of V, V ′ in lieu of W and A∗ in lieu of A (notice that A∗ is

bijective owing to Corollary C.52), leading to
|hv ′′ , A∗ (w′ )iV ′′ ,V ′ | |hv ′′ , A∗ (w′ )iV ′′ ,V ′ |
inf sup ≤ inf sup .
w ∈W ′ v ′′ ∈V ′′
′ kv ′′ kV ′′ kw′ kW ′ v ′′ ∈V ′′ w ′ ∈W ′ kv ′′ kV ′′ kw′ kW ′
Owing to the reflexivity of V, this inequality becomes r ≤ l.
Remark C.54 (Counterexample). The identity (C.25) can fail if A 6= 0 is not bijective. For
instance, if A : (x0 , x1 , x2 . . .) 7→ (0, x0 , x1 , x2 , . . .) is the right shift operator in ℓ2 , then A∗ :
(x0 , x1 , x2 . . .) 7→ (x1 , x2 , x3 , . . .) is the left shift operator. It can be verified that A is injective
but not surjective, whereas A∗ is surjective but not injective. Using the notation of the proof of
Lemma C.53, it can also be shown that l = 1 and r = 0.
Lemma C.55 (Bounds on A−∗ ). Let A ∈ L(V ; W ) be a bijective operator. Assume that V is
reflexive. Let A∗ ∈ L(W ′ ; V ′ ) be the adjoint of A and let A−∗ ∈ L(V ′ ; W ′ ) denote its inverse.
Then kA−∗ kL(V ′ ;W ′ ) = α−1 with α defined in (C.22), and
kA−∗ (v ′ )kW ′ |hA−∗ (v ′ ), wiW ′ ,W | 1
inf ′ ′
= inf sup ′
= . (C.26)
′
v ∈V kv kV ′ ′ ′
v ∈V w∈W kv kV ′ kwkW kAkL(V ;W )
Proof. Notice that the notation A−∗ is meant to reflect that (A−1 )∗ = (A∗ )−1 . Combining the
results from Lemma C.30 and Lemma C.51, we infer that kA−∗ kL(V ′ ;W ′ ) = kA−1 kL(W ;V ) = α−1 .
Moreover, the first equality in (C.26) follows from the definition of k·kW ′ and the second one from
hA−∗ (v ′ ), wiW ′ ,W = hv ′ , A−1 (w)iV ′ ,V , the identity (C.25) (since A−1 is bijective), and the identity
(C.24).
C.8 Coercive operators

We now focus on the more specific class of coercive operators. The notion of coercivity plays a
central role in the analysis of PDEs involving the Laplace operator, and more generally elliptic
operators (see Chapter 31).
Appendix 391
Definition C.56 (Coercive operator). Let V be a complex Banach space. The operator A ∈
L(V ; V ′ ) is said to be a coercive if there exist a real number α > 0 and a complex number ξ ∈ C
with |ξ| = 1 such that
ℜ ξhA(v), viV ′ ,V ≥ α kvk2V , ∀v ∈ V. (C.27)
In the real case, we have either ξ = 1 or ξ = −1.
Remark C.57 (Self-adjoint case). Let A be a coercive self-adjoint operator (see Definition C.31).
Since hA(v), viV ′ ,V is real for all v ∈ V, coercivity means that ℜ(ξ)hA(v), viV ′ ,V ≥ αkvk2V . Thus,
up to rescaling α, one can always take either ξ = 1 or ξ = −1 when A is self-adjoint.
The coercivity condition is sometimes defined as follows: There exists a real number α > 0
such that |hA(v), viV ′ ,V | ≥ αkvk2V for all v ∈ V. Although this variant looks slightly more general
since ℜ ξhA(v), viV ′ ,V ≤ |hA(v), viV ′ ,V |, it is equivalent to (C.27). More precisely, we have the
following result.
Lemma C.58 (Real part vs. module). Let α > 0 and let V be a Hilbert space. The following
two statements are equivalent: (i) |hA(v), viV ′ ,V | ≥ αkvk2V for all v ∈ V. (ii) There is ξ ∈ C with
|ξ| = 1 s.t. (C.27) holds true.
Proof. Let us prove the claim in the real case. It suffices to show that the statement (i) implies
that hA(v), viV ′ ,V has always the same sign for all nonzero v ∈ V. Reasoning by contradiction, if
there are nonzero v, w ∈ V such that hA(v), viV ′ ,V < 0 and hA(w), wiV ′ ,V > 0, then the second-
order polynomial R ∋ λ 7→ hA(v + λw), v + λwiV ′ ,V ∈ R has at least one root λ∗ ∈ R. The
statement (i) yields v + λ∗ w = 0, so that hA(v), viV ′ ,V = λ2∗ hA(w), wiV ′ ,V > 0, which contradicts
hA(v), viV ′ ,V < 0. We refer the reader to Brezis [89, p. 366] for the proof in the complex case (see
also Exercise 46.9 for a proof of the Hausdorff–Toeplitz theorem).
It turns out that the notion of coercivity is relevant only in Hilbert spaces.
Proposition C.59 (Hilbert structure). Let V be a Banach space. V can be equipped with a
Hilbert structure with the same topology if and only if there is a coercive operator in L(V ; V ′ ).
Proof. Setting ((v, w))V := 12 (ξhA(v), wiV ′ ,V + ξhA(w), viV ′ ,V ), we define a sesquilinear form on
V ×V that is Hermitian. The coercivity and boundedness of A imply that
α kvk2V ≤ ((v, v))V ≤ kAkL(V ;V ′ ) kvk2V ,
for all v ∈ V. This shows positive definiteness (so that ((·, ·))V is an inner product in V ) and that
the induced norm is equivalent to k·kV .
Corollary C.60 (Coercivity as a sufficient condition). If the operator A ∈ L(V ; V ′ ) is
coercive, then it is bijective.
Proof. This is the Lax–Milgram lemma which is proved in §25.2.
Definition C.61 (Monotone operator). The operator A ∈ L(V ; V ′ ) is said to be monotone if

ℜ hA(v), viV ′ ,V ≥ 0, ∀v ∈ V. (C.28)
Corollary C.62 (Coercivity as a necessary and sufficient condition). Assume that V is

reflexive. Let A ∈ L(V ; V ′ ) be a monotone self-adjoint operator. Then A is bijective iff it is
coercive (with ξ := 1).
From now on, we assume that V is a Hilbert space. If the operator A ∈ L(V ; V ′ ) is coercive
(and therefore bijective), its inverse A−1 ∈ L(V ′ ; V ) turns out to be coercive as well. Indeed, using
the coercivity of A and the lower bound on A−1 resulting from (C.24), we infer that for all φ ∈ V ′ ,

ℜ ξhφ, A−1 (φ)iV ′ ,V = ℜ ξhA(A−1 (φ)), A−1 (φ)iV ′ ,V
α
≥ α kA−1 (φ)k2V ≥ kφk2V ′ , (C.29)
kAk2
with the shorthand notation kAk := kAkL(V ;V ′ ) . The following results provide more precise char-
acterizations of the coercivity constant of A−1 .
Lemma C.63 (Coercivity of A−1 , self-adjoint case). Let A ∈ L(V ; V ′ ) be a self-adjoint
coercive operator (i.e., (C.27) holds true with either ξ = 1 or ξ = −1 according to Remark C.57).
Then A−1 is coercive with coercivity constant kAk−1 , and we have more precisely
ξhφ, A−1 (φ)iV ′ ,V 1

inf ′ = . (C.30)
φ∈V kφk2V ′ kAk
Proof. Assume that ξ = 1 (the case ξ = −1 is identical). The coercivity of A together with A = A∗
implies that ((v, w))A := hA(v), wiV ′ ,V is an inner product on V. Let v ∈ V and φ ∈ V ′ . Since
hφ, viV ′ ,V = ((A−1 (φ), v))A , the Cauchy–Schwarz inequality implies that
1 1
ℜ hφ, viV ′ ,V ≤ ((v, v))A2 ((A−1 (φ), A−1 (φ)))A2
1 1 1 1
= hA(v), viV2 ′ ,V hφ, A−1 (φ)iV2 ′ ,V ≤ kAk 2 kvkV hφ, A−1 (φ)iV2 ′ ,V ,
where we used the boundedness of A. This implies that

|hφ, viV ′ ,V | 1 1
kφkV ′ = sup ≤ kAk 2 hφ, A−1 (φ)iV2 ′ ,V .
v∈V kvkV
Taking the infimum over φ ∈ V ′ , we infer that

1 hφ, A−1 (φ)iV ′ ,V |hψ, A−1 (φ)iV ′ ,V | 1
≤ inf ′ 2 ≤ inf ′ sup = ,
kAk φ∈V kφkV ′ φ∈V ψ∈V ′ kψkV ′ kφkV ′ kAk
where the last equality follows from (C.24). Thus, all the terms are equal, and this concludes the
proof.
Let us now consider the case where the operator A ∈ L(V ; V ′ ) is not necessarily self-adjoint.
Since V is Hilbert space, V is reflexive. Hence, the adjoint of A is A∗ ∈ L(V ; V ′ ), and we have
hA∗ (v), wiV ′ ,V = hA(w), viV ′ ,V .
Lemma C.64 (Coercivity of A−1 , general case). Let A ∈ L(V ; V ′ ) be a coercive operator
with parameters α > 0 and ξ ∈ C with |ξ| = 1. Let the self-adjoint part of ξA be defined as
(ξA)s := 21 (ξA + (ξA)∗ ) = 21 (ξA + ξA∗ ). The following holds true:

α ℜ ξhφ, A−1 (φ)iV ′ ,V 1
2
≤ inf ′ 2 ≤ . (C.31)
kAk φ∈V kφkV ′ k(ξA)s k
Proof. The lower bound in (C.31) is a restatement of (C.29). To establish the upper bound, let us
set B := ξA. Then B are Bs are coercive (and therefore invertible) operators since

hBs (v), viV ′ ,V = ℜ hB(v), viV ′ ,V = ℜ ξhA(v), viV ′ ,V ≥ α kvk2V ,
Appendix 393
for all v ∈ V. A direct calculation shows that
B −1 (B − Bs )Bs−1 (B ∗ − Bs )B −∗ = (Bs−1 − B −1 )(I − Bs B −∗ )

= Bs−1 − B −1 − B −∗ + B −1 Bs B −∗
1
= Bs−1 − B −1 − B −∗ + B −1 (B + B ∗ )B −∗
2
−1 1 −1 −∗
= Bs − (B + B ).
2
This implies that for all φ ∈ V ′ ,
1
hφ, Bs−1 (φ)iV ′ ,V = hφ, (B −1 + B −∗ )(φ)iV ′ ,V + hφ, B −1 (B − Bs )Bs−1 (B ∗ − Bs )B −∗ (φ)iV ′ ,V
2
= ℜ hφ, B −1 (φ)iV ′ ,V + hψ, Bs−1 (ψ)iV ′ ,V ≥ ℜ hφ, B −1 (φ)iV ′ ,V ,
with ψ := (B ∗ − Bs )B −∗ (φ) and where we used that hψ, Bs−1 (ψ)iV ′ ,V ≥ 0. Applying Lemma C.63
to the operator Bs , which is coercive and self-adjoint, we conclude that

1 hφ, Bs−1 (φ)iV ′ ,V ℜ hφ, B −1 (φ)iV ′ ,V
= inf ≥ inf ′ .
kBs k φ∈V ′ kφk2V ′ φ∈V kφk2V ′
Since hφ, B −1 (φ)iV ′ ,V = (ξ)−1 hφ, A−1 (φ)iV ′ ,V and (ξ)−1 = ξ, this proves the upper bound in (C.31).
Bibliography
[1] M. Abbas, A. Ern, and N. Pignet. Hybrid high-order methods for finite deformations of
hyperelastic materials. Comput. Mech., 62(4):909–928, 2018. pages 221
[2] M. Abbas, A. Ern, and N. Pignet. A hybrid high-order method for incremental associative
plasticity with small deformations. Comput. Methods Appl. Mech. Engrg., 346:891–912, 2019.
pages 221
[3] M. Abramowitz and I. Stegun. Handbook of mathematical functions. Dover Publications

Inc., New York, NY, 9th edition, 1972. pages 70
[4] Y. Achdou, C. Bernardi, and F. Coquel. A priori and a posteriori analysis of finite volume
discretizations of Darcy’s equations. Numer. Math., 96(1):17–42, 2003. pages 151
[5] L. M. Adams and H. F. Jordan. Is SOR color-blind? SIAM J. Sci. Statist. Comput., 7(2):
490–506, 1986. pages 66
[6] S. Agmon, A. Douglis, and L. Nirenberg. Estimates near the boundary for solutions of elliptic
partial differential equations satisfying general boundary conditions. I. Comm. Pure Appl.
Math., 12:623–727, 1959. pages 89
[7] M. Ainsworth and J. T. Oden. A posteriori error estimation in finite element analysis. Wiley,
New York, NY, 2000. pages 335
[8] M. Amara and J. M. Thomas. Equilibrium finite elements for the linear elastic problem.
Numer. Math., 33(4):367–383, 1979. pages 219
[9] C. Amrouche and R. Ratsimahalo. Conditions “inf sup” dans les espaces de Banach non
réflexifs. C. R. Acad. Sci. Paris Sér. I Math., 330(12):1069–1072, 2000. pages 16
[10] C. Amrouche, C. Bernardi, M. Dauge, and V. Girault. Vector potentials in three-dimensional

non-smooth domains. Math. Methods Appl. Sci., 21(9):823–864, 1998. pages 229, 237
[11] R. Andreev. Stability of sparse space-time finite element discretizations of linear parabolic
evolution equations. IMA J. Numer. Anal., 33(1):242–260, 2013. pages 31
[12] P. F. Antonietti, A. Buffa, and I. Perugia. Discontinuous Galerkin approximation of the

Laplace eigenproblem. Comput. Methods Appl. Mech. Engrg., 195(25):3483–3503, 2006. pages
285
[13] T. Arbogast and Z. Chen. On the implementation of mixed methods as nonconforming

methods for second-order elliptic problems. Math. Comp., 64(211):943–972, 1995. pages 330
396 Bibliography
[14] J. H. Argyris and S. Kelsey. Energy theorems and structural analysis. Buttelworths, London,
UK, third edition, 1967. Originally published in Aircraft Engrg., 26, pp. 347–356, 383–387,
394, 410–422, (1954) and 27, pp. 42–58, 80–94, 125–134, 145–158 (1955). pages 212
[15] D. N. Arnold. An interior penalty finite element method with discontinuous elements. SIAM
J. Numer. Anal., 19:742–760, 1982. pages 169, 171
[16] D. N. Arnold and F. Brezzi. Mixed and nonconforming finite element methods: implemen-
tation, postprocessing and error estimates. RAIRO Modél. Math. Anal. Numér., 19(1):7–32,
1985. pages 327, 330
[17] D. N. Arnold and R. Winther. Mixed finite elements for elasticity. Numer. Math., 92(3):
401–419, 2002. pages 219
[18] D. N. Arnold, I. Babuška, and J. Osborn. Finite element methods: principles for their
selection. Comput. Methods Appl. Mech. Engrg., 45(1-3):57–96, 1984. pages 30
[19] D. N. Arnold, F. Brezzi, and J. Douglas, Jr. PEERS: a new mixed finite element for plane
elasticity. Japan J. Appl. Math., 1(2):347–367, 1984. pages 219
[20] D. N. Arnold, F. Brezzi, and M. Fortin. A stable finite element for the Stokes equations.
Calcolo, 21:337–344, 1984. pages 356
[21] D. N. Arnold, F. Brezzi, B. Cockburn, and L. D. Marini. Unified analysis of discontinuous

Galerkin methods for elliptic problems. SIAM J. Numer. Anal., 39(5):1749–1779, 2001/02.
pages 177
[22] D. N. Arnold, D. Boffi, and R. S. Falk. Approximation by quadrilateral finite elements.

Math. Comp., 71(239):909–922, 2002. pages 374
[23] D. N. Arnold, R. S. Falk, and R. Winther. Finite element exterior calculus, homological
techniques, and applications. Acta Numer., 15:1–155, 2006. pages 237, 241
[24] D. N. Arnold, R. S. Falk, and R. Winther. Mixed finite element methods for linear elasticity
with weakly imposed symmetry. Math. Comp., 76(260):1699–1723, 2007. pages 219
[25] D. N. Arnold, G. Awanou, and R. Winther. Finite elements for symmetric tensors in three
dimensions. Math. Comp., 77(263):1229–1251, 2008. pages 219
[26] D. N. Arnold, R. S. Falk, and R. Winther. Finite element exterior calculus: from Hodge
theory to numerical stability. Bull. Amer. Math. Soc. (N.S.), 47(2):281–354, 2010. pages 237
[27] F. Assous, P. Ciarlet, Jr., and S. Labrunie. Mathematical foundations of computational elec-
tromagnetism, volume 198 of Applied Mathematical Sciences. Springer, Cham, Switzerland,
2018. pages 225
[28] J.-P. Aubin. Behavior of the error of the approximate solutions of boundary value problems
for linear elliptic operators by Galerkin’s and finite difference methods. Ann. Scuola Norm.
Sup. Pisa (3), 21:599–637, 1967. pages 98
[29] J.-P. Aubin. Applied functional analysis. Pure and Applied Mathematics. Wiley-Interscience,
New York, NY, second edition, 2000. With exercises by B. Cornet and J.-M. Lasry, Translated
from the French by C. Labrousse. pages 98, 377
References 397
[30] P. Auscher, E. Russ, and P. Tchamitchian. Hardy Sobolev spaces on strongly Lipschitz
domains of Rn . J. Funct. Anal., 218(1):54–109, 2005. pages 341
[31] O. Axelsson and A. Kucherov. Real valued iterative methods for solving complex symmetric
linear systems. Numer. Linear Algebra Appl., 7(4):197–218, 2000. pages 55
[32] B. Ayuso de Dios, K. Lipnikov, and G. Manzini. The nonconforming virtual element method.
ESAIM Math. Model. Numer. Anal., 50(3):879–904, 2016. pages 183
[33] I. Babuška. Error-bounds for finite element method. Numer. Math., 16:322–333, 1970/1971.
pages 14, 27
[34] I. Babuška. The finite element method with Lagrangian multipliers. Numer. Math., 20:
179–192, 1973. pages 295, 325
[35] I. Babuška. The finite element method with penalty. Math. Comp., 27:221–228, 1973. pages
107
[36] I. Babuška and A. Miller. A feedback finite element method with a posteriori error estimation.
I. The finite element method and some basic properties of the a posteriori error estimator.
Comput. Methods Appl. Mech. Engrg., 61(1):1–40, 1987. pages 120, 335
[37] I. Babuška and J. Osborn. Analysis of finite element methods for second order boundary
value problems using mesh dependent norms. Numer. Math., 34(1):41–62, 1980. pages 99
[38] I. Babuška and J. Osborn. Eigenvalue problems. In Handbook of Numerical Analysis, Vol.
II, pages 641–787. North-Holland, Amsterdam, The Netherlands, 1991. pages 277, 278, 279
[39] I. Babuška and W. C. Rheinbolt. Error estimates for adaptive finite element method com-
putations. SIAM J. Numer. Anal., 15:736–754, 1978. pages 121
[40] I. Babuška and M. Suri. Locking effects in the finite element approximation of elasticity
problems. Numer. Math., 62(4):439–463, 1992. pages 217
[41] C. Bacuta. A unified approach for Uzawa algorithms. SIAM J. Numer. Anal., 44(6):2633–
2649, 2006. pages 310
[42] C. Bacuta. Sharp stability and approximation estimates for symmetric saddle point systems.
Appl. Anal., 95(1):226–237, 2016. pages 292, 293
[43] S. Badia and R. Codina. A nodal-based finite element approximation of the Maxwell problem
suitable for singular solutions. SIAM J. Numer. Anal., 50(2):398–417, 2012. pages 252
[44] G. A. Baker. Finite element methods for elliptic equations using nonconforming elements.
Math. Comp., 31(137):45–59, 1977. pages 169
[45] G. R. Barrenechea, E. Burman, and F. Karakatsani. Edge-based nonlinear diffusion for

finite element approximations of convection-diffusion equations and its relation to algebraic
flux-correction schemes. Numer. Math., 135(2):521–545, 2017. pages 111
[46] F. Bassi and S. Rebay. A high-order accurate discontinuous finite element method for the
numerical solution of the compressible Navier-Stokes equations. J. Comput. Phys., 131(2):
267–279, 1997. pages 174
398 Bibliography
[47] P. Bastian and B. Rivière. Superconvergence and H(div) projection for discontinuous
Galerkin methods. Internat. J. Numer. Methods Fluids, 42(10):1043–1057, 2003. pages
331
[48] R. Becker and R. Rannacher. An optimal control approach to a posteriori error estimation
in finite element methods. Acta Numerica, 10:1–102, 2001. pages 123
[49] H. Beirão da Veiga. On a stationary transport equation. Ann. Univ. Ferrara Sez. VII, 32:
79–91, 1986. pages 92
[50] L. Beirão da Veiga, F. Brezzi, A. Cangiani, G. Manzini, L. D. Marini, and A. Russo. Basic
principles of virtual element methods. M3AS Math. Models Methods Appl. Sci., 199(23):
199–214, 2013. pages 185
[51] M. Benzi and M. A. Olshanskii. An augmented Lagrangian-based approach to the Oseen

problem. SIAM J. Sci. Comput., 28(6):2095–2113, 2006. pages 311
[52] M. Benzi, G. H. Golub, and J. Liesen. Numerical solution of saddle point problems. Acta
Numer., 14:1–137, 2005. pages 310, 311
[53] M. Benzi, M. A. Olshanskii, and Z. Wang. Modified augmented Lagrangian preconditioners

for the incompressible Navier-Stokes equations. Internat. J. Numer. Methods Fluids, 66(4):
486–508, 2011. pages 311
[54] M. Bercovier and O. Pironneau. Error estimates for finite element solution of the Stokes
problem in the primitive variables. Numer. Math., 33:211–224, 1979. pages 353, 358, 359
[55] C. Bernardi and R. Verfürth. Adaptive finite element methods for elliptic equations with
non-smooth coefficients. Numer. Math., 85(4):579–608, 2000. pages 91
[56] P. Binev, W. Dahmen, and R. DeVore. Adaptive finite element methods with convergence
rates. Numer. Math., 97(2):219–268, 2004. pages 126, 127, 128
[57] M. S. Birman and M. Z. Solomyak. L2 -theory of the Maxwell operator in arbitrary domains.
Russian Mathematical Surveys, 42(6):75, 1987. pages 229
[58] Å. Björck. Numerical methods in matrix computations, volume 59 of Texts in Applied Math-
ematics. Springer, Cham, Switzerland, 2015. pages 53
[59] J. Blechta, J. Málek, and M. Vohralı́k. Localization of the W −1,q norm for local a posteriori
efficiency. IMA J. Numer. Anal., 40(2):914–950, 2020. pages 120
[60] D. Boffi. Three-dimensional finite element methods for the Stokes problem. SIAM J. Numer.
Anal., 34(2):664–670, 1997. pages 358
[61] D. Boffi. A note on the de Rham complex and a discrete compactness property. Appl. Math.
Lett., 14(1):33–38, 2001. pages 237
[62] D. Boffi. Finite element approximation of eigenvalue problems. Acta Numer., 19:1–120, 2010.
pages 279
[63] D. Boffi, F. Brezzi, and L. Gastaldi. On the problem of spurious eigenvalues in the approxi-
mation of linear elliptic problems in mixed form. Math. Comp., 69(229):121–140, 2000. pages
285
References 399
[64] D. Boffi, F. Brezzi, and M. Fortin. Reduced symmetry elements in linear elasticity. Commun.
Pure Appl. Anal., 8(1):95–121, 2009. pages 219
[65] D. Boffi, F. Brezzi, and M. Fortin. Mixed finite element methods and applications, volume 44
of Springer Series in Computational Mathematics. Springer, Heidelberg, Germany, 2013.
pages 330, 351, 352, 358, 366, 367, 368, 374
[66] M. E. Bogovskiı̆. Solutions of some problems of vector analysis, associated with the operators
div and grad. In Theory of cubature formulas and the application of functional analysis to
problems of mathematical physics, volume 1980 of Trudy Sem. S. L. Soboleva, No. 1, pages
5–40, 149. Akad. Nauk SSSR Sibirsk. Otdel. Inst. Mat., Novosibirsk, Russia, 1980. pages 341
[67] J. M. Boland and R. A. Nicolaides. Stability of finite elements under divergence constraints.
SIAM J. Numer. Anal., 20(4):722–731, 1983. pages 366
[68] J. M. Boland and R. A. Nicolaides. Stability and semistable low order finite elements for
viscous flows. SIAM J. Numer. Anal., 22:474–492, 1985. pages 347
[69] A. Bonito and J.-L. Guermond. Approximation of the eigenvalue problem for the time
harmonic Maxwell system by continuous Lagrange finite elements. Math. Comp., 80(276):
1887–1910, 2011. pages 249, 252
[70] A. Bonito, J.-L. Guermond, and F. Luddens. Regularity of the Maxwell equations in hetero-
geneous media and Lipschitz domains. J. Math. Anal. Appl., 408:498–512, 2013. pages 91,
235
[71] A. Bonito, J.-L. Guermond, and F. Luddens. An interior penalty method with C 0 finite ele-
ments for the approximation of the Maxwell equations in heterogeneous media: convergence
analysis with minimal regularity. ESAIM Math. Model. Numer. Anal., 50(5):1457–1489,
2016. pages 249, 252
[72] A.-S. Bonnet-Ben Dhia, P. Ciarlet, Jr., and C. M. Zwölf. Time harmonic wave diffraction
problems in materials with sign-shifting coefficients. J. Comput. Appl. Math., 234(6):1912–
1919, 2010. pages 15
[73] A.-S. Bonnet-Ben Dhia, L. Chesnel, and P. Ciarlet, Jr. T -coercivity for scalar interface
problems between dielectrics and metamaterials. ESAIM Math. Model. Numer. Anal., 46(6):
1363–1387, 2012. pages 15
[74] A. Bossavit. Computational electromagnetism, variational formulations, complementary,
edge elements, volume 2 of Electromagnetism. Academic Press, New York, NY, 1998. pages
225, 226
[75] M. Botti, D. A. Di Pietro, and P. Sochala. A hybrid high-order method for nonlinear elasticity.
[76] D. Braess and J. Schöberl. Equilibrated residual error estimator for edge elements. Math.
Comp., 77(262):651–672, 2008. pages 333, 335
[77] D. Braess, V. Pillwein, and J. Schöberl. Equilibrated residual error estimates are p-robust.
Comput. Methods Appl. Mech. Engrg., 198(13-14):1189–1197, 2009. pages 333, 334
[78] J. H. Bramble and J. E. Osborn. Rate of convergence estimates for nonselfadjoint eigenvalue
approximations. Math. Comp., 27:525–549, 1973. pages 279
400 Bibliography
[79] J. H. Bramble and J. E. Pasciak. A new approximation technique for div-curl systems. Math.
Comp., 73(248):1739–1762, 2004. pages 252
[80] J. H. Bramble, J. E. Pasciak, and J. Xu. Parallel multilevel preconditioners. Math. Comp.,
55(191):1–22, 1990. pages 54
[81] J. H. Bramble, J. E. Pasciak, and P. S. Vassilevski. Computational scales of Sobolev norms

with application to preconditioning. Math. Comp., 69(230):463–480, 2000. pages 51, 54
[82] J. H. Bramble, T. V. Kolev, and J. E. Pasciak. The approximation of the Maxwell eigenvalue
problem using a least-squares method. Math. Comp., 74(252):1575–1598, 2005. pages 252
[83] J. Brandts, S. Korotov, and M. Křı́žek. The discrete maximum principle for linear simplicial
finite element approximations of a reaction-diffusion problem. Linear Algebra Appl., 429(10):
2344–2357, 2008. pages 111
[84] J. Brandts, S. Korotov, M. Křı́žek, and J. Šolc. On nonobtuse simplicial partitions. SIAM
Rev., 51(2):317–335, 2009. pages 110
[85] H. Brass and K. Petras. Quadrature theory, volume 178 of Mathematical Surveys and Mono-
graphs. American Mathematical Society, Providence, RI, 2011. pages 70
[86] S. C. Brenner. Poincaré-Friedrichs inequalities for piecewise H 1 functions. SIAM J. Numer.

Anal., 41(1):306–324, 2003. pages 171
[87] S. C. Brenner and L. R. Scott. The mathematical theory of finite element methods, volume 15
of Texts in Applied Mathematics. Springer, New York, NY, third edition, 2008. pages 27
[88] S. C. Brenner and L.-Y. Sung. Linear finite element methods for planar linear elasticity.
Math. Comp., 59(200):321–338, 1992. pages 218
[89] H. Brezis. Functional analysis, Sobolev spaces and partial differential equations. Universitext.
Springer, New York, NY, 2011. pages 108, 253, 256, 257, 258, 377, 378, 379, 380, 381, 382,
383, 384, 387, 388, 391
[90] F. Brezzi. On the existence, uniqueness and approximation of saddle-point problems arising
from Lagrange multipliers. RAIRO, Anal. Num., pages 129–151, 1974. pages 295
[91] F. Brezzi and K.-J. Bathe. A discourse on the stability conditions for mixed finite element
formulations. Comput. Methods Appl. Mech. Engrg., 82(1-3):27–57, 1990. pages 366, 376
[92] F. Brezzi and R. S. Falk. Stability of a higher-order Hood–Taylor method. SIAM J. Numer.
Anal., 28:581–590, 1991. pages 358, 374
[93] F. Brezzi, G. Manzini, L. D. Marini, P. Pietra, and A. Russo. Discontinuous Galerkin

approximations for elliptic problems. Numer. Methods Partial Differential Equations, 16(4):
365–378, 2000. pages 174
[94] W. L. Briggs. A multigrid tutorial. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia, PA, 1987. pages 54
[95] A. Buffa and C. Ortner. Compact embeddings of broken Sobolev spaces and applications.
IMA J. Numer. Anal., 4(29):827–855, 2009. pages 171, 176
References 401
[96] A. Buffa, P. Ciarlet, Jr., and E. Jamelot. Solving electromagnetic eigenvalue problems in
polyhedral domains with nodal finite elements. Numer. Math., 113:497–518, 2009. pages 251
[97] E. Burman. Robust error estimates in weak norms for advection dominated transport prob-
lems with rough data. Math. Models Methods Appl. Sci., 24(13):2663–2684, 2014. pages
92
[98] E. Burman and A. Ern. Discrete maximum principle for Galerkin approximations of the
Laplace operator on arbitrary meshes. C. R. Math. Acad. Sci. Paris, 338(8):641–646, 2004.
pages 111
[99] E. Burman and A. Ern. Stabilized Galerkin approximation of convection-diffusion-reaction

equations: discrete maximum principle and convergence. Math. Comp., 74(252):1637–1652,
2005. pages 111
[100] E. Burman and A. Ern. Discontinuous Galerkin approximation with discrete variational
principle for the nonlinear Laplacian. C. R. Math. Acad. Sci. Paris, 346(17–18):1013–1016,
2008. pages 176
[101] E. Burman and P. Zunino. A domain decomposition method for partial differential equations
with non-negative form based on interior penalties. SIAM J. Numer. Anal., 44:1612–1638,
2006. pages 200
[102] E. Burman, H. Wu, and L. Zhu. Linear continuous interior penalty finite element method
for Helmholtz equation with high wave number: one-dimensional analysis. Numer. Methods
Partial Differential Equations, 32(5):1378–1410, 2016. pages 142
[103] V. Calo, M. Cicuttin, Q. Deng, and A. Ern. Spectral approximation of elliptic operators by
the hybrid high-order method. Math. Comp., 88(318):1559–1586, 2019. pages 285
[104] E. Cancès, G. Dusson, Y. Maday, B. Stamm, and M. Vohralı́k. Guaranteed and robust
a posteriori bounds for Laplace eigenvalues and eigenvectors: conforming approximations.
[105] C. Canuto. Eigenvalue approximations by mixed methods. RAIRO Anal. Numér., 12(1):
27–50, 1978. pages 285
[106] S. Caorsi, P. Fernandes, and M. Raffetto. On the convergence of Galerkin finite element
approximations of electromagnetic eigenproblems. SIAM J. Numer. Anal., 38(2):580–607,
2000. pages 237
[107] C. Carstensen and S. A. Funken. Fully reliable localized error control in the FEM. SIAM J.
Sci. Comput., 21(4):1465–1484, 1999/00. pages 335
[108] C. Carstensen and S. A. Funken. Constants in Clément-interpolation error and residual

based a posteriori error estimates in finite element methods. East-West J. Numer. Math., 8
(3):153–175, 2000. pages 120
[109] C. Carstensen and J. Gedicke. Guaranteed lower bounds for eigenvalues. Math. Comp., 83
(290):2605–2629, 2014. pages 285
[110] C. Carstensen and F. Hellwig. Low-order discontinuous Petrov-Galerkin finite element meth-
ods for linear elasticity. SIAM J. Numer. Anal., 54(6):3388–3410, 2016. pages 218
402 Bibliography
[111] C. Carstensen, L. Demkowicz, and J. Gopalakrishnan. A posteriori error control for DPG
methods. SIAM J. Numer. Anal., 52(3):1335–1353, 2014. pages 25, 313
[112] C. Carstensen, M. Feischl, M. Page, and D. Praetorius. Axioms of adaptivity. Comput. Math.
Appl., 67(6):1195–1253, 2014. pages 126, 128
[113] J. M. Cascón, C. Kreuzer, R. H. Nochetto, and K. G. Siebert. Quasi-optimal convergence
rate for an adaptive finite element method. SIAM J. Numer. Anal., 46(5):2524–2550, 2008.
pages 125, 126, 127, 128
[114] J. Céa. Approximation variationnelle des problèmes aux limites. Ann. Inst. Fourier (Greno-
ble), 14(2):345–444, 1964. pages 27
[115] S. N. Chandler-Wilde, D. P. Hewett, and A. Moiola. Interpolation of Hilbert and Sobolev
spaces: quantitative estimates and counterexamples. Mathematika, 61(2):414–443, 2015.
pages 88
[116] F. Chatelin. Spectral approximation of linear operators. Computer Science and Applied
Mathematics. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], New York, NY,
1983. With a foreword by P. Henrici, with solutions to exercises by M. Ahués. pages 253,
279
[117] Z. Chen and H. Chen. Pointwise error estimates of discontinuous Galerkin methods with
penalty for second-order elliptic problems. SIAM J. Numer. Anal., 42(3):1146–1166, 2004.
pages 174
[118] L. Chesnel and P. Ciarlet, Jr. T -coercivity and continuous Galerkin methods: application
to transmission problems with sign changing coefficients. Numer. Math., 124(1):1–29, 2013.
pages 19
[119] E. V. Chizhonkov and M. A. Olshanskii. On the domain geometry dependence of the LBB
condition. M2AN Math. Model. Numer. Anal., 34(5):935–951, 2000. pages 304
[120] P. Ciarlet, Jr. T -coercivity: application to the discretization of Helmholtz-like problems.
Comput. Math. Appl., 64(1):22–34, 2012. pages 139
[121] P. Ciarlet, Jr. On the approximation of electromagnetic fields by edge finite elements. Part
1: Sharp interpolation results for low-regularity fields. Comput. Math. Appl., 71(1):85–104,
2016. pages 231, 235
[122] P. Ciarlet, Jr. and M. Vohralı́k. Localization of global norms and robust a posteriori error
control for transmission problems with sign-changing coefficients. ESAIM Math. Model.
Numer. Anal., 52(5):2037–2064, 2018. pages 118, 120
[123] P. G. Ciarlet. Mathematical elasticity II: Theory of plates, volume 27 of Studies in Mathe-
matics and its Applications. Elsevier, Amsterdam, 1997. pages 215
[124] P. G. Ciarlet. The finite element method for elliptic problems, volume 40 of Classics in Applied
Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA,
2002. Reprint of the 1978 original [North-Holland, Amsterdam, The Netherlands]. pages 27,
114
[125] P. G. Ciarlet and P. Ciarlet, Jr. Another approach to linearized elasticity and a new proof
of Korn’s inequality. Math. Models Methods Appl. Sci., 15(2):259–271, 2005. pages 216
References 403
[126] P. G. Ciarlet and P.-A. Raviart. The combined effect of curved boundaries and numerical
integration in isoparametric finite element methods. In The mathematical foundations of
the finite element method with applications to partial differential equations (Proc. Sympos.,
Univ. Maryland, Baltimore, MD, 1972), pages 409–474. Academic Press, New York, NY,
1972. pages 114
[127] P. G. Ciarlet and P.-A. Raviart. Maximum principle and uniform convergence for the finite
element method. Comput. Methods Appl. Mech. Engrg., 2:17–31, 1973. pages 109
[128] M. Cicuttin, D. A. Di Pietro, and A. Ern. Implementation of discontinuous skeletal methods
on arbitrary-dimensional, polytopal meshes using generic programming. J. Comput. Appl.
Math., 344:852–874, 2018. pages 183
[129] R. W. Clough. The finite element method in plane stress analysis. In Proc. 2nd ASCE
Conference on Electronic Computation, Pittsburgh, PA, 1960. pages 212
[130] B. Cockburn. Static condensation, hybridization, and the devising of the HDG methods. In
G. R. Barrenechea, F. Brezzi, A. Cangiani, and E. H. Georgoulis, editors, Building bridges:
Connections and challenges in modern approaches to numerical partial differential equations,
volume 114 of Lecture Notes in Computational Science and Engineering, pages 129–178,
Cham, Switzerland, 2016. Springer. pages 186, 188, 330
[131] B. Cockburn and J. Gopalakrishnan. A characterization of hybridized mixed methods for
second order elliptic problems. SIAM J. Numer. Anal., 42(1):283–301, 2004. pages 330
[132] B. Cockburn and K. Shi. Devising HDG methods for Stokes flow: An overview. Comput. &
Fluids, 98:221–229, 2014. pages 373
[133] B. Cockburn, D. Schötzau, and J. Wang. Discontinuous Galerkin methods for incompressible
elastic materials. Comput. Methods Appl. Mech. Engrg., 195(25-28):3184–3204, 2006. pages
218
[134] B. Cockburn, J. Gopalakrishnan, and R. Lazarov. Unified hybridization of discontinuous
Galerkin, mixed, and continuous Galerkin methods for second order elliptic problems. SIAM
J. Numer. Anal., 47(2):1319–1365, 2009. pages 188
[135] B. Cockburn, J. Gopalakrishnan, and F.-J. Sayas. A projection-based error analysis of HDG
methods. Math. Comp., 79(271):1351–1367, 2010. pages 188
[136] B. Cockburn, W. Qiu, and K. Shi. Conditions for superconvergence of HDG methods for
second-order elliptic problems. Math. Comp., 81(279):1327–1353, 2012. pages 188
[137] B. Cockburn, D. A. Di Pietro, and A. Ern. Bridging the hybrid high-order and hybridizable
discontinuous Galerkin methods. ESAIM Math. Model Numer. Anal., 50(3):635–650, 2016.
pages 183, 188
[138] A. Cohen, R. DeVore, and R. H. Nochetto. Convergence rates of AFEM with H −1 data.
Found. Comput. Math., 12(5):671–718, 2012. pages 118, 120
[139] R. Cools. Monomial cubature rules since “Stroud”: a compilation. II. J. Comput. Appl.
Math., 112(1-2):21–27, 1999. pages 70
[140] R. Cools and P. Rabinowitz. Monomial cubature rules since “Stroud”: a compilation. J.
Comput. Appl. Math., 48(3):309–326, 1993. pages 70
404 Bibliography
[141] H. O. Cordes. Vereinfachter Beweis der Existenz einer Apriori-Hölderkonstanten. Math.

Ann., 138:155–178, 1959. pages 80
[142] M. Costabel. A remark on the regularity of solutions of Maxwell’s equations on Lipschitz

domains. Math. Methods Appl. Sci., 12(4):365–368, 1990. pages 229
[143] M. Costabel. A coercive bilinear form for Maxwell’s equations. J. Math. Anal. Appl., 157
(2):527–541, 1991. pages 251
[144] M. Costabel and M. Dauge. Maxwell and Lamé eigenvalues on polyhedra. Math. Methods
Appl. Sci., 22(3):243–258, 1999. pages 251
[145] M. Costabel and M. Dauge. Weighted regularization of Maxwell equations in polyhedral

domains. A rehabilitation of nodal finite elements. Numer. Math., 93(2):239–277, 2002.
pages 251
[146] M. Costabel and A. McIntosh. On Bogovskiı̆ and regularized Poincaré integral operators for
de Rham complexes on Lipschitz domains. Math. Z., 265(2):297–320, 2010. pages 341
[147] M. Costabel, M. Dauge, and S. Nicaise. Analytic regularity for linear elliptic systems in
polygons and polyhedra. Math. Models Methods Appl. Sci., 22(8):1250015, 63 pp., 2012.
pages 90
[148] M. Costabel, M. Dauge, and S. Nicaise. Weighted analytic regularity in polyhedra. Comput.
Math. Appl., 67(4):807–817, 2014. pages 90
[149] J.-P. Croisille. Finite volume box schemes and mixed methods. M2AN Math. Model. Numer.
Anal., 34(2):1087–1106, 2000. pages 326
[150] J.-P. Croisille and I. Greff. Some nonconforming mixed box schemes for elliptic problems.
Numer. Methods Partial Differential Equations, 18(3):355–373, 2002. pages 150
[151] M. Crouzeix and P.-A. Raviart. Conforming and nonconforming finite element methods for
solving the stationary Stokes equations. I. Rev. Française Automat. Informat. Recherche
Opérationnelle Sér. Rouge, 7(R-3):33–75, 1973. pages 146, 356, 368, 373
[152] M. Dauge. Elliptic boundary value problems on corner domains, volume 1341 of Lecture
Notes in Mathematics. Springer-Verlag, Berlin, 1988. pages 89, 90, 91
[153] M. Dauge. Stationary Stokes and Navier-Stokes systems on two- or three-dimensional do-
mains with corners. I. Linearized equations. SIAM J. Math. Anal., 20(1):74–97, 1989. pages
342
[154] R. Dautray and J.-L. Lions. Mathematical analysis and numerical methods for science and
technology. Vol. 4. Integral equations and numerical methods. Springer-Verlag, Berlin, Ger-
many, 1990. pages 114
[155] P. J. Davis and P. Rabinowitz. Methods of numerical integration. Computer Science and
Applied Mathematics. Academic Press, New York, NY, 1975. pages 70
[156] T. A. Davis. Direct methods for sparse linear systems, volume 2 of Fundamentals of Algo-
rithms. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA, 2006.
pages 53, 61
References 405
[157] C. Dawson, S. Sun, and M. F. Wheeler. Compatible algorithms for coupled flow and trans-
port. Comp. Meth. Appl. Mech. Eng., 193:2565–2580, 2004. pages 174
[158] L. F. Demkowicz and J. Gopalakrishnan. An overview of the discontinuous Petrov Galerkin
method. In Recent developments in discontinuous Galerkin finite element methods for partial
differential equations, volume 157 of The IMA Volumes in Mathematics and its Applications,
pages 149–180. Springer, Cham, Switzerland, 2014. pages 313
[159] A. Demlow, D. Leykekhman, A. H. Schatz, and L. B. Wahlbin. Best approximation property
1
in the W∞ norm for finite element methods on graded meshes. Math. Comp., 81(278):
743–764, 2012. pages 95
[160] J. W. Demmel, S. C. Eisenstat, J. R. Gilbert, X. S. Li, and J. W. H. Liu. A supernodal

approach to sparse partial pivoting. SIAM J. Matrix Analysis and Applications, 20(3):720–
755, 1999. pages 53
[161] J. Descloux, N. Nassif, and J. Rappaz. On spectral approximation. I. The problem of con-
vergence. RAIRO Anal. Numér., 12(2):97–112, 1978. pages 279
[162] J. Descloux, N. Nassif, and J. Rappaz. On spectral approximation. II. Error estimates for
the Galerkin method. RAIRO Anal. Numér., 12(2):113–119, 1978. pages 279
[163] P. Destuynder and B. Métivet. Explicit error bounds in a conforming finite element method.
Math. Comp., 68(228):1379–1396, 1999. pages 333, 335
[164] D. A. Di Pietro and A. Ern. Discrete functional analysis tools for discontinuous Galerkin
methods with application to the incompressible Navier-Stokes equations. Math. Comp., 79
(271):1303–1330, 2010. pages 171, 176
[165] D. A. Di Pietro and A. Ern. Mathematical aspects of discontinuous Galerkin methods, vol-
ume 69 of Mathématiques & Applications [Mathematics & Applications]. Springer-Verlag,
Berlin, 2012. pages 177, 178
[166] D. A. Di Pietro and A. Ern. A hybrid high-order locking-free method for linear elasticity on
general meshes. Comput. Meth. Appl. Mech. Engrg., 283:1–21, 2015. pages 183, 218, 221,
373
[167] D. A. Di Pietro, A. Ern, and J.-L. Guermond. Discontinuous Galerkin methods for anisotropic
semi-definite diffusion with advection. SIAM J. Numer. Anal., 46(2):805–831, 2008. pages
200
[168] D. A. Di Pietro, A. Ern, and S. Lemaire. An arbitrary-order and compact-stencil discretiza-

tion of diffusion on general meshes based on local reconstruction operators. Comput. Meth.
Appl. Math., 14(4):461–472, 2014. pages 183, 190
[169] D. A. Di Pietro, A. Ern, A. Linke, and F. Schieweck. A discontinuous skeletal method for
the viscosity-dependent Stokes problem. Comput. Methods Appl. Mech. Engrg., 306:175–195,
2016. pages 373, 374
[170] M. Dobrowolski. On the LBB constant on stretched domains. Math. Nachr., 254/255:64–67,
2003. pages 304
[171] W. Dörfler. A convergent adaptive algorithm for Poisson’s equation. SIAM J. Numer. Anal.,
33(3):1106–1124, 1996. pages 126
406 Bibliography
[172] J. Droniou, R. Eymard, T. Gallouët, C. Guichard, and R. Herbin. The gradient discretisation
method, volume 82 of Mathématiques & Applications [Mathematics & Applications]. Springer,
Cham, Switzerland, 2018. pages 36
[173] M. Dryja. On discontinuous Galerkin methods for elliptic problems with discontinuous co-
efficients. Comput. Methods Appl. Math., 3(1):76–85, 2003. pages 200
[174] M. Dryja, J. Galvis, and M. Sarkis. BDDC methods for discontinuous Galerkin discretization
of elliptic problems. J. Complexity, 23(4-6):715–739, 2007. pages 200
[175] H. Duan, P. Lin, and R. C. E. Tan. C 0 elements for generalized indefinite Maxwell equations.
Numer. Math., 122(1):61–99, 2012. pages 252
[176] H. Duan, R. C. E. Tan, S.-Y. Yang, and C.-S. You. A mixed H 1 -conforming finite element
method for solving Maxwell’s equations with non-H 1 solution. SIAM J. Sci. Comput., 40
(1):A224–A250, 2018. pages 252
[177] I. S. Duff, A. M. Erisman, and J. K. Reid. Direct methods for sparse matrices. Monographs
on Numerical Analysis. The Clarendon Press, Oxford University Press, New York, 1986.
pages 53
[178] D. A. Dunavant. High degree efficient symmetrical Gaussian quadrature rules for the triangle.
Internat. J. Numer. Methods Engrg., 21(6):1129–1148, 1985. pages 70
[179] N. Dunford and J. T. Schwartz. Linear operators. I. General theory, volume 7 of Pure and
Applied Mathematics. Interscience Publishers, Inc., New York, NY, 1958. With the assistance
of W. G. Bade and R. G. Bartle. pages 253
[180] R. Durán, M. A. Muschietti, E. Russ, and P. Tchamitchian. Divergence operator and
Poincaré inequalities on arbitrary bounded domains. Complex Var. Elliptic Equ., 55(8-10):
795–816, 2010. pages 341
[181] R. G. Durán and M. A. Muschietti. An explicit right inverse of the divergence operator which
is continuous in weighted norms. Studia Math., 148(3):207–219, 2001. pages 341
[182] R. G. Durán, L. Gastaldi, and C. Padra. A posteriori error estimators for mixed approxima-
tions of eigenvalue problems. Math. Models Methods Appl. Sci., 9(8):1165–1178, 1999. pages
285
[183] G. Duvaut and J.-L. Lions. Les inéquations en mécanique et en physique. Dunod, Paris,
1972. pages 215
[184] H. Egger and C. Waluga. hp analysis of a hybrid DG method for Stokes flow. IMA J. Numer.
Anal., 33(2):687–721, 2013. pages 373
[185] H. C. Elman, D. J. Silvester, and A. J. Wathen. Finite elements and fast iterative solvers:
with applications in incompressible fluid dynamics. Numerical Mathematics and Scientific
Computation. Oxford University Press, New York, 2005. pages 54, 310, 311
[186] A. Ern and J.-L. Guermond. Evaluation of the condition number in linear systems arising in
finite element approximations. M2AN Math. Model. Numer. Anal., 40(1):29–48, 2006. pages
51
[187] A. Ern and J.-L. Guermond. A converse to Fortin’s lemma in Banach spaces. C. R. Math.
Acad. Sci. Paris, 354(11):1092–1095, 2016. pages 25, 352
References 407
[188] A. Ern and J.-L. Guermond. Analysis of the edge finite element approximation of the Maxwell
equations with low regularity solutions. Comput. Math. Appl., 75(3):918–932, 2018. pages
233
[189] A. Ern and M. Vohralı́k. A posteriori error estimation based on potential and flux recon-
struction for the heat equation. SIAM J. Numer. Anal., 48(1):198–223, 2010. pages 333
[190] A. Ern and M. Vohralı́k. Adaptive inexact Newton methods with a posteriori stopping criteria
for nonlinear diffusion PDEs. SIAM J. Sci. Comput., 35(4):A1761–A1791, 2013. pages 333,
334
[191] A. Ern and M. Vohralı́k. Polynomial-degree-robust a posteriori estimates in a unified setting
for conforming, nonconforming, discontinuous Galerkin, and mixed discretizations. SIAM J.
Numer. Anal., 53(2):1058–1081, 2015. pages 333, 334
[192] A. Ern and M. Vohralı́k. Stable broken H 1 and H(div) polynomial extensions for polynomial-
degree-robust potential and flux reconstruction in three space dimensions. Math. Comp., 89
(322):551–594, 2020. pages 333, 334
[193] A. Ern, S. Nicaise, and M. Vohralı́k. An accurate H(div) flux reconstruction for discontinuous
Galerkin approximations of elliptic problems. C. R. Math. Acad. Sci. Paris, 345(12):709–712,
2007. pages 177
[194] A. Ern, A. F. Stephansen, and P. Zunino. A discontinuous Galerkin method with weighted
averages for advection-diffusion equations with locally small and anisotropic diffusivity. IMA
J. Numer. Anal., 29(2):235–256, 2009. pages 200
[195] S. Esterhazy and J. M. Melenk. On stability of discretizations of the Helmholtz equation.
In Numerical analysis of multiscale problems, volume 83 of Lect. Notes Comput. Sci. Eng.,
pages 285–324. Springer, Heidelberg, 2012. pages 136
[196] L. C. Evans. Partial differential equations, volume 19 of Graduate Studies in Mathematics.
American Mathematical Society, Providence, RI, 1998. pages 89, 108
[197] R. Eymard, T. Gallouët, and R. Herbin. Convergence of finite volume schemes for semilinear
convection diffusion equations. Numer. Math., 82(1):91–116, 1999. pages 119
[198] V. Faber and T. Manteuffel. Necessary and sufficient conditions for the existence of a con-
jugate gradient method. SIAM J. Numer. Anal., 21(2):352–362, 1984. pages 53
[199] R. S. Falk. Nonconforming finite element methods for the equations of linear elasticity. Math.
Comp., 57(196):529–550, 1991. pages 218
[200] X. Feng and H. Wu. Discontinuous Galerkin methods for the Helmholtz equation with large
wave number. SIAM J. Numer. Anal., 47(4):2872–2896, 2009. pages 142
[201] M. Fortin. An analysis of the convergence of mixed finite element methods. RAIRO Anal.
Num., 11:341–354, 1977. pages 25
[202] M. Fortin. A three-dimensional quadratic nonconforming element. Numer. Math., 46(2):
269–279, 1985. pages 218
[203] M. Fortin and R. Glowinski. Augmented Lagrangian methods, volume 15 of Studies in Math-
ematics and its Applications. North-Holland Publishing Co., Amsterdam, The Netherlands,
1983. Translated from the French by B. Hunt and D. C. Spicer. pages 308
408 Bibliography
[204] M. Fortin and M. Soulié. A non-conforming piecewise quadratic finite element on triangles.
Internat. J. Numer. Methods Engrg., 19:505–520, 1983. pages 157, 218
[205] B. Fraejis de Veubeke. Diffusion des inconnues hyperstatiques dans les voilures à longerons
couplés. Bull. Serv. Technique de l’Aéronautique, 24:1–56, 1951. pages 216
[206] B. Fraejis de Veubeke. Stress function approach. In World Congress on the Finite Element
Method in Structural Mechanics, pages J.1–J.51, Bournemouth, UK, 1975. Available at
http://orbi.ulg.ac.be/handle/2268/205875. pages 219
[207] B. Fraejis de Veubeke. Displacement and equilibrium models in the finite element method.
Internat. J. Numer. Methods Engrg., 52(3):287–342, 2001. pages 327
[208] R. W. Freund. Conjugate gradient-type methods for linear systems with complex symmetric
coefficient matrices. SIAM J. Sci. Statist. Comput., 13(1):425–448, 1992. pages 55
[209] G. Fu, B. Cockburn, and H. Stolarski. Analysis of an HDG method for linear elasticity.
Internat. J. Numer. Methods Engrg., 102(3-4):551–575, 2015. pages 218
[210] G. P. Galdi. An Introduction to the mathematical theory of the Navier-Stokes equations.

Vol. I, volume 38 of Springer Tracts in Natural Philosophy. Springer-Verlag, New York, NY,
1994. pages 341
[211] L. Gastaldi and R. H. Nochetto. Sharp maximum norm error estimates for general mixed
finite element approximations to second order elliptic equations. RAIRO Modél. Math. Anal.
Numér., 23(1):103–128, 1989. pages 324
[212] G. N. Gatica. A simple introduction to the mixed finite element method. Springer Briefs in
Mathematics. Springer, Cham, Switzerland, 2014. pages 321
[213] I. Gelfand. Zur Theorie der Charaktere der Abelschen topologischen Gruppen. Rec. Math.
[Mat. Sbornik] N. S., 9 (51):49–50, 1941. pages 254
[214] A. George and J. W. H. Liu. Computer solution of large sparse positive definite systems.
Prentice-Hall Series in Computational Mathematics. Prentice-Hall Inc., Englewood Cliffs,
NJ, 1981. pages 53, 61, 64
[215] A. George, J. R. Gilbert, and J. W. H. Liu, editors. Graph theory and sparse matrix computa-
tion, volume 56 of The IMA Volumes in Mathematics and its Applications. Springer-Verlag,
New York, NY, 1993. pages 61, 64
[216] D. Gilbarg and N. S. Trudinger. Elliptic partial differential equations of second order. Classics
in Mathematics. Springer-Verlag, Berlin, 2001. Reprint of the 1998 edition. pages 108
[217] V. Girault and P.-A. Raviart. Finite element methods for Navier–Stokes equations. Theory
and algorithms. Springer Series in Computational Mathematics. Springer-Verlag, Berlin,
Germany, 1986. pages 25, 340, 347, 351, 358, 366
[218] G. H. Golub and C. F. van Loan. Matrix computations. John Hopkins University Press,
Baltimore, MD, second edition, 1989. pages 52, 307
[219] J. Gopalakrishnan and W. Qiu. An analysis of the practical DPG method. Math. Comp., 83
(286):537–552, 2014. pages 313
References 409
[220] J. Gopalakrishnan, F. Li, N.-C. Nguyen, and J. Peraire. Spectral approximations by the
HDG method. Math. Comp., 84(293):1037–1059, 2015. pages 285
[221] A. Greenbaum. Iterative methods for solving linear systems, volume 17 of Frontiers in Applied
Mathematics. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA,
1997. pages 310
[222] I. Greff. Schémas boı̂te : étude théorique et numérique. PhD thesis, University of Metz,
France, 2003. pages 157
[223] P. Grisvard. Elliptic problems in nonsmooth domains, volume 24 of Monographs and Studies
in Mathematics. Pitman (Advanced Publishing Program), Boston, MA, 1985. pages 89, 90,
165
[224] P. Grisvard. Singularities in boundary value problems, volume 22 of Recherches en

Mathématiques Appliquées [Research in Applied Mathematics]. Masson, Paris, France, 1992.
pages 89
[225] C. Grossmann and H.-G. Roos. Numerical treatment of partial differential equations. Uni-
versitext. Springer, Berlin, 2007. Translated and revised from the 3rd (2005) German edition
by Martin Stynes. pages 51
[226] T. Gudi. A new error analysis for discontinuous finite element methods for linear elliptic
problems. Math. Comp., 79(272):2169–2189, 2010. pages 41, 206
[227] J.-L. Guermond. A finite element technique for solving first-order PDEs in Lp . SIAM J.
Numer. Anal., 42(2):714–737, 2004. pages 6
[228] J.-L. Guermond and B. Popov. An optimal L1 -minimization algorithm for stationary
Hamilton-Jacobi equations. Commun. Math. Sci., 7(1):211–238, 2009. pages 6
[229] B. Guo and I. Babuška. Regularity of the solutions for elliptic problems on nonsmooth do-
mains in R3 . I. Countably normed spaces on polyhedral domains. Proc. Roy. Soc. Edinburgh
Sect. A, 127(1):77–126, 1997. pages 90
[230] B. Guo and I. Babuška. Regularity of the solutions for elliptic problems on nonsmooth
domains in R3 . II. Regularity in neighbourhoods of edges. Proc. Roy. Soc. Edinburgh Sect.
A, 127(3):517–545, 1997. pages 90
[231] K. Gustafson. The Toeplitz-Hausdorff theorem for linear operators. Proc. Amer. Math. Soc.,
25:203–204, 1970. pages 264
[232] R. J. Guyan. Reduction of stiffness and mass matrices. J. Am. Inst. Aeron. and Astro., 3:
380, 1965. pages 47
[233] J. Guzmán. Pointwise error estimates for discontinuous Galerkin methods with lifting oper-
ators for elliptic problems. Math. Comp., 75(255):1067–1085, 2006. pages 174
[234] J. Guzmán, D. Leykekhman, J. Rossmann, and A. H. Schatz. Hölder estimates for Green’s
functions on convex polyhedral domains and their applications to finite element methods.
Numer. Math., 112(2):221–243, 2009. pages 95
[235] W. Hackbusch. Multigrid methods and applications, volume 4 of Springer Series in Compu-
tational Mathematics. Springer-Verlag, Berlin, Germany, 1985. pages 51, 54
410 Bibliography
[236] J. Hadamard. Le problème de Cauchy et les équations aux dérivées partielles linéaires hy-
perboliques. Hermann, Paris, France, 1932. pages 12
[237] P. C. Hammer and A. H. Stroud. Numerical integration over simplexes. Math. Tables Aids
Comput., 10:137–139, 1956. pages 70
[238] A. Hannukainen, R. Stenberg, and M. Vohralı́k. A unified framework for a posteriori error
estimation for the Stokes problem. Numer. Math., 122(4):725–769, 2012. pages 333
[239] P. Hansbo and M. G. Larson. Discontinuous Galerkin methods for incompressible and nearly
incompressible elasticity by Nitsche’s method. Comput. Methods Appl. Mech. Engrg., 191
(17-18):1895–1908, 2002. pages 218
[240] P. Hansbo and M. G. Larson. Discontinuous Galerkin and the Crouzeix-Raviart element:
application to elasticity. M2AN Math. Model. Numer. Anal., 37(1):63–72, 2003. pages 218
[241] E. Hellinger. Die allgemeinen Ansätze der Mechanik der Kontinua. In F. Klein and C. Müller,
editors, Enzyklopädie der mathematischen Wissenschaften, volume 4, pages 601–694. Teub-
ner, Leipzig, 1914. pages 216
[242] M. R. Hestenes and E. Stiefel. Method of conjugate gradients for solving linear systems. J.
Res. Nat. Bur. Standards, 49:409–436, 1952. pages 53
[243] U. Hetmaniuk. Stability estimates for a class of Helmholtz problems. Commun. Math. Sci.,
5(3):665–678, 2007. pages 133, 136, 138
[244] R. Hiptmair. Finite elements in computational electromagnetism. Acta Numerica, 11:237–
339, 1 2002. pages 237
[245] I. Hlaváček, J. Haslinger, J. Nečas, and J. Lovı́šek. Solution of variational inequalities in
mechanics, volume 66 of Applied Mathematical Sciences. Springer-Verlag, New York, 1988.
Translated from the Slovak by J. Jarnı́k. pages 335
[246] C. O. Horgan. Korn’s inequalities and their applications in continuum mechanics. SIAM
Rev., 37:491–511, 1995. pages 223
[247] L. Hörmander. The analysis of linear partial differential operators. III. Classics in Math-
ematics. Springer, Berlin, 2007. Pseudo-differential operators, Reprint of the 1994 edition.
pages 80
[248] H.-C. Hu. On some variational methods on the theory of elasticity and the theory of plasticity.
Scientia Sinica, 4:33–54, 1955. pages 216
[249] N. Hu, X. Guo, and I. N. Katz. Bounds for eigenvalues and condition numbers in the p-version
of the finite element method. Math. Comp., 67(224):1423–1450, 1998. pages 51
[250] T. J. R. Hughes, G. Engel, L. Mazzei, and M. G. Larson. The continuous Galerkin method
is locally conservative. J. Comput. Phys., 163(2):467–488, 2000. pages 333
[251] F. Ihlenburg and I. Babuška. Finite element solution of the Helmholtz equation with high
wave number. I. The h-version of the FEM. Comput. Math. Appl., 30(9):9–37, 1995. pages
137, 138, 142, 143
[252] B. M. Irons. Structural eigenvalue problems – elimination of unwanted variables. AIAA J.,
3(5):961–962, 1965. pages 47
References 411
[253] R. C. James. Reflexivity and the supremum of linear functionals. Ann. of Math. (2), 66:
159–169, 1957. pages 387
[254] Y. Jeon, E.-J. Park, and D. Sheen. A hybridized finite element method for the Stokes
problem. Computers and Mathematics with Applications, 68:2222–2232, 2014. pages 373
[255] D. S. Jerison and C. E. Kenig. The Neumann problem in Lipschitz domains. Bull. Amer.
Math. Soc. (N.S.), 4(2):203–207, 1981. pages 90
[256] D. S. Jerison and C. E. Kenig. The inhomogeneous Dirichlet problem in Lipschitz domains.
J. Func. Anal., 130(1):161–219, 1995. pages 90
[257] X. Jiang and R. H. Nochetto. Effect of numerical integration for elliptic obstacle problems.
Numer. Math., 67(4):501–512, 1994. pages 109
[258] F. Jochmann. An H s -regularity result for the gradient of solutions to elliptic equations with
mixed boundary conditions. J. Math. Anal. Appl., 238:429–450, 1999. pages 91
[259] F. Jochmann. Regularity of weak solutions of Maxwell’s equations with mixed boundary-
conditions. Math. Methods Appl. Sci., 22(14):1255–1274, 1999. pages 235
[260] L. John, M. Neilan, and I. Smears. Stable discontinuous Galerkin FEM without penalty
parameters. In Numerical Mathematics and Advanced Applications ENUMATH 2015, volume
112 of Lecture Notes in Computational Science and Engineering, pages 165–173. Springer,
Cham, Switzerland, 2016. pages 176, 191
[261] V. John, A. Linke, C. Merdon, M. Neilan, and L. G. Rebholz. On the divergence constraint
in mixed finite element methods for incompressible flows. SIAM Rev., 59(3):492–544, 2017.
pages 346
[262] M. Juntunen and R. Stenberg. Nitsche’s method for general boundary conditions. Math.
Comp., 78(267):1353–1374, 2009. pages 160
[263] H. Kanayama, R. Motoyama, K. Endo, and F. Kikuchi. Three-dimensional magnetostatic
analysis using Nédélec’s elements. IEEE T. Magn., 26:682–685, 1990. pages 289
[264] G. Kanschat and R. Rannacher. Local error analysis of the interior penalty discontinuous
Galerkin method for second order elliptic problems. J. Numer. Math., 10(4):249–274, 2002.
pages 174
[265] P. Keast. Moderate-degree tetrahedral quadrature formulas. Comput. Methods Appl. Mech.
Engrg., 55(3):339–348, 1986. pages 70
[266] R. B. Kellogg and J. E. Osborn. A regularity result for the Stokes problem in a convex
polygon. J. Functional Analysis, 21(4):397–431, 1976. pages 342
[267] F. Kikuchi. On a discrete compactness property for the Nédélec finite elements. J. Fac. Sci.
Univ. Tokyo Sect. IA Math., 36(3):479–490, 1989. pages 237
[268] K. Y. Kim. A posteriori error estimators for locally conservative methods of nonlinear elliptic
problems. Appl. Numer. Math., 57(9):1065–1080, 2007. pages 177
[269] K. Y. Kim. Guaranteed a posteriori error estimator for mixed finite element methods of
linear elasticity with weak stress symmetry. SIAM J. Numer. Anal., 48(6):2364–2385, 2011.
pages 223
412 Bibliography
[270] V. A. Kondrat′ ev. Boundary value problems for elliptic equations in domains with conical
or angular points. Trudy Moskov. Mat. Obšč., 16:209–292, 1967. pages 90
[271] E. Kreyszig. Introductory functional analysis with applications. John Wiley & Sons, New
York-London-Sydney, 1978. pages 253, 254, 257, 258
[272] Y. A. Kuznetsov. Efficient iterative solvers for elliptic finite element problems on nonmatching
grids. Russian J. Numer. Anal. Math. Modelling, 10(3):187–211, 1995. pages 311
[273] P. Ladevèze and D. Leguillon. Error estimate procedure in the finite element method and
applications. SIAM J. Numer. Anal., 20(3):485–509, 1983. pages 335
[274] M. G. Larson and A. J. Niklasson. Analysis of a nonsymmetric discontinuous Galerkin

method for elliptic problems: stability and energy error estimates. SIAM J. Numer. Anal.,
42(1):252–264, 2004. pages 174
[275] M. G. Larson and A. J. Niklasson. A conservative flux for the continuous Galerkin method
based on discontinuous enrichment. Calcolo, 41(2):65–76, 2004. pages 333
[276] J. E. Lavery. Nonoscillatory solution of the steady-state inviscid Burgers’ equation by math-
ematical programming. J. Comput. Phys., 79(2):436–448, 1988. pages 6
[277] J. E. Lavery. Solution of steady-state one-dimensional conservation laws by mathematical

programming. SIAM J. Numer. Anal., 26(5):1081–1089, 1989. pages 6
[278] P. D. Lax. Functional analysis. Pure and Applied Mathematics. Wiley-Interscience [John
Wiley & Sons], New York, NY, 2002. pages 253, 257, 258, 377, 378, 380, 382, 384
[279] P. Lederer, A. Linke, C. Merdon, and J. Schöberl. Divergence-free reconstruction operators

for pressure-robust Stokes discretizations with continuous pressure finite elements. SIAM J.
Numer. Anal., 55(5):1291–1314, 2017. pages 345, 358
[280] C. Lehrenfeld. Hybrid Discontinuous Galerkin methods for solving incompressible flow prob-
lems. PhD thesis, Rheinisch-Westfälische Technische Hochschule (RWTH) Aachen, Germany,
2010. pages 183
[281] C. Lehrenfeld and J. Schöberl. High order exactly divergence-free hybrid discontinuous
Galerkin methods for unsteady incompressible flows. Comput. Methods Appl. Mech. Engrg.,
307:339–361, 2016. pages 183, 373
[282] S. Levy. Structural analysis and influence coefficients for delta wings. J. Aeronaut. Sci., 20,
1953. pages 212
[283] A. Linke. On the role of the Helmholtz decomposition in mixed methods for incompressible
flows and a new variational crime. Comput. Methods Appl. Mech. Engrg., 268:782–800, 2014.
pages 345, 374
[284] A. Linke, C. Merdon, and W. Wollner. Optimal L2 velocity error estimate for a modified
pressure-robust Crouzeix-Raviart Stokes element. IMA J. Numer. Anal., 37(1):354–374,
2017. pages 304
[285] J.-L. Lions. Problèmes aux limites non homogènes à données irrégulières: Une méthode
d’approximation. In Numerical Analysis of Partial Differential Equations (C.I.M.E. 2 Ciclo,
Ispra, Italy, 1967), pages 283–292. Edizioni Cremonese, Rome, Italy, 1968. pages 107
References 413
[286] J.-L. Lions and E. Magenes. Non-homogeneous Boundary Value Problems and Applications.
Vols. I, II. Springer-Verlag, New York-Heidelberg, 1972. Translated from the French by P.
Kenneth, Die Grundlehren der mathematischen Wissenschaften, Band 181-182. pages 88, 91
[287] X. Liu. A framework of verified eigenvalue bounds for self-adjoint differential operators.
Appl. Math. Comput., 267:341–355, 2015. pages 285
[288] J. M.-S. Lubuma and S. Nicaise. Dirichlet problems in polyhedral domains. I. Regularity of
the solutions. Math. Nachr., 168:243–261, 1994. pages 90
[289] J. M.-S. Lubuma and S. Nicaise. Dirichlet problems in polyhedral domains. II. Approximation
by FEM and BEM. J. Comput. Appl. Math., 61(1):13–27, 1995. pages 90
[290] R. Luce and B. I. Wohlmuth. A local a posteriori error estimator based on equilibrated
fluxes. SIAM J. Numer. Anal., 42(4):1394–1414, 2004. pages 333
[291] N. Lüthen, M. Juntunen, and R. Stenberg. An improved a priori error analysis of Nitsche’s
method for Robin boundary conditions. Numer. Math., 138(4):1011–1026, 2018. pages 206
[292] G. I. Marchuk and Y. A. Kuznetsov. On optimal iteration processes. Soviet. Math. Dokl., 9:
1041–1045, 1968. pages 54
[293] G. I. Marchuk and Y. A. Kuznetsov. Méthodes itératives et fonctionnelles quadratiques. In

J.-L. Lions and G. I. Marchuk, editors, Sur les méthodes numériques en sciences physiques
et économiques, pages 3–131, Paris, France, 1974. Dunod. pages 54
[294] K.-A. Mardal, J. Schöberl, and R. Winther. A uniformly stable Fortin operator for the
Taylor-Hood element. Numer. Math., 123(3):537–551, 2013. pages 304, 358
[295] L. D. Marini. An inexpensive method for the evaluation of the solution of the lowest order
Raviart-Thomas mixed method. SIAM J. Numer. Anal., 22(3):493–496, 1985. pages 156,
330
[296] V. G. Maz′ ja and B. A. Plamenevskiı̆. Weighted spaces with inhomogeneous norms, and
boundary value problems in domains with conical points. In Elliptische Differentialgleichun-
gen (Meeting, Rostock, Germany, 1977), pages 161–190. Wilhelm-Pieck-Univ., Rostock, Ger-
many, 1978. pages 90
[297] A. L. Mazzucato and V. Nistor. Well-posedness and regularity for the elasticity equation
with mixed boundary conditions on polyhedral domains and domains with cracks. Arch.
Ration. Mech. Anal., 195(1):25–73, 2010. pages 216
[298] W. McLean. Strongly elliptic systems and boundary integral equations. Cambridge University
Press, Cambridge, UK, 2000. pages 215
[299] J. M. Melenk. On generalized finite-element methods. ProQuest LLC, Ann Arbor, MI, 1995.
Ph.D. thesis, University of Maryland, MD. pages 133
[300] J. M. Melenk and S. Sauter. Wavenumber explicit convergence analysis for Galerkin dis-
cretizations of the Helmholtz equation. SIAM J. Numer. Anal., 49(3):1210–1243, 2011.
pages 142
[301] B. Mercier, J. E. Osborn, J. Rappaz, and P.-A. Raviart. Eigenvalue approximation by mixed
and hybrid methods. Math. Comp., 36(154):427–453, 1981. pages 285
414 Bibliography
[302] P. Monk. A finite element method for approximating the time-harmonic Maxwell equations.
Numer. Math., 63(2):243–261, 1992. pages 241
[303] P. Monk. Finite element methods for Maxwell’s equations. Numerical Mathematics and
Scientific Computation. Oxford University Press, New York, NY, 2003. pages 225, 241
[304] P. Monk and L. Demkowicz. Discrete compactness and the approximation of Maxwell’s
equations in R3 . Math. Comp., 70(234):507–523, 2001. pages 237
[305] P. Morin, R. H. Nochetto, and K. G. Siebert. Data oscillation and convergence of adaptive
FEM. SIAM J. Numer. Anal., 38(2):466–488, 2000. pages 126, 127
[306] P. Morin, R. H. Nochetto, and K. G. Siebert. Local problems on stars: a posteriori error
estimators, convergence, and performance. Math. Comp., 72(243):1067–1097, 2003. pages
122, 335
[307] M. E. Morley. A family of mixed finite elements for linear elasticity. Numer. Math., 55(6):
633–666, 1989. pages 219
[308] I. Muga and K. G. van der Zee. Discretization of linear problems in Banach
spaces: residual minimization, nonlinear Petrov–Galerkin, and monotone mixed methods.
arXiv:1511.04400v3 [Math.NA], 2018. pages 25, 313
[309] M. F. Murphy, G. H. Golub, and A. J. Wathen. A note on preconditioning for indefinite

linear systems. SIAM J. Sci. Comput., 21(6):1969–1972, 2000. pages 311
[310] J. Nečas. Sur une méthode pour résoudre les équations aux dérivées partielles de type
elliptique, voisine de la variationnelle. Ann. Scuola Norm. Sup. Pisa, 16:305–326, 1962.
pages 14
[311] S. Nicaise. Regularity of the solutions of elliptic systems in polyhedral domains. Bull. Belg.
Math. Soc. Simon Stevin, 4(3):411–429, 1997. pages 90
[312] L. Nirenberg. Remarks on strongly elliptic partial differential equations. Comm. Pure Appl.
Math., 8:649–675, 1955. pages 89
[313] J. Nitsche. Ein Kriterium für die Quasi-Optimalität des Ritzschen Verfahrens. Numer.
Math., 11:346–348, 1968. pages 98
[314] J. Nitsche. Über ein Variationsprinzip zur Lösung von Dirichlet-Problemen bei Verwendung
von Teilräumen, die keinen Randbedingungen unterworfen sind. Abh. Math. Sem. Univ.
Hamburg, 36:9–15, 1971. pages 107, 160
[315] R. H. Nochetto and A. Veeser. Primer of adaptive finite element methods. In Multiscale and
adaptivity: modeling, numerics and applications, volume 2040 of Lecture Notes in Math.,
pages 125–225. Springer, Heidelberg, 2012. pages 120, 123, 126
[316] R. H. Nochetto, K. G. Siebert, and A. Veeser. Theory of adaptive finite element methods: an
introduction. In Multiscale, nonlinear and adaptive approximation, pages 409–542. Springer,
Berlin, 2009. pages 126
[317] J. T. Oden. Finite elements: An introduction, volume II : Finite element methods of Handbook
of Numerical Analysis, chapter 1, pages 3–15. North Holland, Amsterdam, The Netherlands,
1991. P.G. Ciarlet and J.L. Lions, editors. pages 212
References 415
[318] J. T. Oden, I. Babuška, and C. E. Baumann. A discontinuous hp finite element method for
diffusion problems. J. Comput. Phys., 146(2):491–519, 1998. pages 174
[319] I. Oikawa. A hybridized discontinuous Galerkin method with reduced stabilization. J. Sci.
Comput., 65(1):327–340, 2015. pages 183
[320] E. T. Olsen and J. Douglas, Jr. Bounds on spectral condition numbers of matrices arising
in the p-version of the finite element method. Numer. Math., 69(3):333–352, 1995. pages 51
[321] J. E. Osborn. Spectral approximation for compact operators. Math. Comp., 29:712–725,
1975. pages 278, 279
[322] N. Parés, P. Dı́ez, and A. Huerta. Subdomain-based flux-free a posteriori error estimators.
Comput. Methods Appl. Mech. Engrg., 195(4-6):297–323, 2006. pages 335
[323] I. Perugia and D. Schötzau. The hp-local discontinuous Galerkin method for low-frequency
time-harmonic Maxwell equations. Math. Comp., 72(243):1179–1214, 2003. pages 174
[324] I. Perugia and V. Simoncini. Block-diagonal and indefinite symmetric preconditioners for
mixed finite element formulations. Numer. Linear Algebra Appl., 7(7-8):585–616, 2000. pages
311
[325] D. Peterseim. Eliminating the pollution effect in Helmholtz problems by local subscale
correction. Math. Comp., 86(305):1005–1036, 2017. pages 142
[326] R. J. Plemmons. M -matrix characterizations. I. Nonsingular M -matrices. Linear Algebra
and Appl., 18(2):175–188, 1977. pages 51
[327] W. Prager and J. L. Synge. Approximations in elasticity based on the concept of function
space. Quart. Appl. Math., 5:241–269, 1947. pages 335
[328] J. Qin. On the convergence of some low order mixed finite elements for incompressible fluids.
ProQuest LLC, Ann Arbor, MI, 1994. Ph.D. thesis, The Pennsylvania State University, PA.
pages 366, 370, 371
[329] R. Rannacher and R. L. Scott. Some optimal error estimates for piecewise linear finite
element approximations. Math. Comp., 38(158):437–445, 1982. pages 95
[330] R. Rannacher and S. Turek. Simple nonconforming quadrilateral Stokes element. Numer.
Methods Partial Differential Equations, 8(2):97–111, 1992. pages 157, 373
[331] P.-A. Raviart and J.-M. Thomas. Introduction à l’analyse numérique des équations aux
dérivées partielles. Masson, Paris, France, 1983. pages 265
[332] M. Reed and B. Simon. Methods of modern mathematical physics. IV. Analysis of operators.
Academic Press [Harcourt Brace Jovanovich, Publishers], New York-London, 1978. pages 80
[333] E. Reissner. On a variational theorem in elasticity. J. Math. Physics, 29:90–95, 1950. pages
216
[334] S. I. Repin. Computable majorants of constants in the Poincaré and Friedrichs inequalities.
J. Math. Sci., 186(2):307–321, 2012. pages 119
[335] B. Rivière, M. F. Wheeler, and V. Girault. Improved energy estimates for interior penalty,
constrained and discontinuous Galerkin methods for elliptic problems. I. Comput. Geosci.,
8:337–360, 1999. pages 174
416 Bibliography
[336] B. Rivière, M. F. Wheeler, and V. Girault. A priori error estimates for finite element methods
based on discontinuous approximation spaces for elliptic problems. SIAM J. Numer. Anal.,
39(3):902–931, 2001. pages 174
[337] W. Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, NY, third edition,
1987. pages 377, 378, 380, 381, 384
[338] T. Rusten and R. Winther. A preconditioned iterative method for saddlepoint problems.
SIAM J. Matrix Anal. Appl., 13(3):887–904, 1992. pages 307
[339] Y. Saad. Iterative methods for sparse linear systems. PWS Publishing Company, Boston,
MA, 1996. pages 53, 54, 61, 309, 311
[340] Y. Saad and M. H. Schultz. GMRES: a generalized minimal residual algorithm for solving
nonsymmetric linear systems. SIAM J. Sci. Statist. Comput., 7:856–869, 1986. pages 54
[341] G. Savaré. Regularity results for elliptic equations in Lipschitz domains. J. Func. Anal., 152:
176–201, 1998. pages 91
[342] F.-J. Sayas. Aubin-Nitsche estimates are equivalent to compact embeddings. BIT, 44(2):
287–290, 2004. pages 97, 98
[343] A. H. Schatz. An observation concerning Ritz-Galerkin methods with indefinite bilinear
forms. Math. Comp., 28:959–962, 1974. pages 140
[344] A. H. Schatz and L. B. Wahlbin. On the quasi-optimality in L∞ of the Ḣ 1 -projection into
finite element spaces. Math. Comp., 38(157):1–22, 1982. pages 99
[345] L. R. Scott and M. Vogelius. Norm estimates for a maximal right inverse of the divergence
operator in spaces of piecewise polynomials. RAIRO Modél. Math. Anal. Numér., 19(1):
111–143, 1985. pages 217, 370
[346] I. Šebestová and T. Vejchodský. Two-sided bounds for eigenvalues of differential operators
with applications to Friedrichs, Poincaré, trace, and similar constants. SIAM J. Numer.
Anal., 52(1):308–329, 2014. pages 119
[347] D. J. Silvester and A. J. Wathen. Fast iterative solution of stabilised Stokes systems. II.
Using general block preconditioners. SIAM J. Numer. Anal., 31(5):1352–1367, 1994. pages
311
[348] I. Smears and E. Süli. Discontinuous Galerkin finite element approximation of Hamilton-
Jacobi-Bellman equations with Cordes coefficients. SIAM J. Numer. Anal., 52(2):993–1016,
2014. pages 80
[349] V. A. Solonnikov. Lp -estimates for solutions of the heat equation in a dihedral angle. Rend.
Mat. Appl. (7), 21(1-4):1–15, 2001. pages 341
[350] P. Sonneveld. CGS, a fast Lanczos-type solver for nonsymmetric linear systems. SIAM J.
Sci. Statist. Comput., 10(1):36–52, 1989. pages 54
[351] S.-C. Soon, B. Cockburn, and H. K. Stolarski. A hybridizable discontinuous Galerkin method
for linear elasticity. Internat. J. Numer. Methods Engrg., 80(8):1058–1092, 2009. pages 218
[352] R. Stenberg. Analysis of mixed finite elements methods for the Stokes problem: a unified
approach. Math. Comp., 42(165):9–23, 1984. pages 358, 366
References 417
[353] R. Stenberg. On the construction of optimal mixed finite element methods for the linear
elasticity problem. Numer. Math., 48(4):447–462, 1986. pages 219, 366
[354] R. Stenberg. On some three-dimensional finite elements for incompressible media. Comput.
Methods Appl. Mech. Engrg., 63(3):261–269, 1987. pages 358, 366
[355] R. Stenberg. A family of mixed finite elements for the elasticity problem. Numer. Math., 53
(5):513–538, 1988. pages 219
[356] R. Stevenson. Optimality of a standard adaptive finite element method. Found. Comput.
Math., 7(2):245–269, 2007. pages 126
[357] R. Stevenson. The completion of locally refined simplicial partitions created by bisection.
Math. Comp., 77(261):227–241, 2008. pages 126, 127, 128
[358] G. Strang. Variational crimes in the finite element method. In A. Aziz, editor, The math-
ematical foundations of the finite element method with applications to partial differential
equations, New York, NY, 1972. Academic Press. pages 39
[359] G. Strang and G. J. Fix. An analysis of the finite element method. Prentice-Hall Inc.,
Englewood Cliffs, NJ, 1973. Prentice-Hall Series in Automatic Computation. pages 51, 279
[360] A. H. Stroud. Approximate calculation of multiple integrals. Prentice-Hall Inc., Englewood

Cliffs, NJ, 1971. pages 70
[361] F. Tantardini and A. Veeser. The L2 -projection and quasi-optimality of Galerkin methods
for parabolic equations. SIAM J. Numer. Anal., 54(1):317–340, 2016. pages 30, 31
[362] L. Tartar. An introduction to Sobolev spaces and interpolation spaces, volume 3 of Lecture
Notes of the Unione Matematica Italiana. Springer, Berlin, Germany; UMI, Bologna, Italy,
2007. pages 88
[363] R. Temam. Navier–Stokes equations, volume 2 of Studies in Mathematics and its Applica-
tions. North-Holland, Amsterdam, The Netherlands, 1977. pages 150
[364] A. Ten Eyck and A. Lew. Discontinuous Galerkin methods for non-linear elasticity. Internat.
J. Numer. Methods Engrg., 67(9):1204–1243, 2006. pages 176
[365] U. Trottenberg, C. W. Oosterlee, and A. Schüller. Multigrid. Academic Press, Inc., San
Diego, CA, 2001. With contributions by A. Brandt, P. Oswald and K. Stüben. pages 54
[366] S. Turek. Efficient solvers for incompressible flow problems. An algorithmic and compu-
tational approach, volume 6 of Lecture Notes in Computational Science and Engineering.
Springer-Verlag, Berlin, Germany, 1999. pages 310, 373
[367] M. J. Turner, R. W. Clough, H. C. Martin, and L. J. Topp. Stiffness and deflection analysis
of complex structures. J. Aero. Sci., 23:805–823, 1956. pages 212
[368] G. M. Vainikko. Asymptotic error bounds for projection methods in the eigenvalue problem.
Ž. Vyčisl. Mat. i Mat. Fiz., 4:405–425, 1964. pages 279
[369] G. M. Vainikko. Rapidity of convergence of approximation methods in eigenvalue problems.

Ž. Vyčisl. Mat. i Mat. Fiz., 7:977–987, 1967. pages 279
418 Bibliography
[370] H. A. van der Vorst. Bi-CGStab: a more stably converging variant of CG-S for the solution
of nonsymmetric linear systems. SIAM J. Sci. Statist. Comput., 13:631–644, 1992. pages 54
[371] A. Veeser. Approximating gradients with continuous piecewise polynomial functions. Found.
Comput. Math., 16(3):723–750, 2016. pages 97
[372] A. Veeser and R. Verfürth. Poincaré constants for finite element stars. IMA J. Numer. Anal.,
32(1):30–47, 2012. pages 119, 120, 122
[373] A. Veeser and P. Zanotti. Quasi-optimal nonconforming methods for symmetric elliptic
problems. I—Abstract theory. SIAM J. Numer. Anal., 56(3):1621–1642, 2018. pages 36, 149
[374] T. Vejchodský and P. Šolı́n. Discrete maximum principle for Poisson equation with mixed
boundary conditions solved by hp-FEM. Adv. Appl. Math. Mech., 1(2):201–214, 2009. pages
109
[375] R. Verfürth. A combined conjugate gradient-multigrid algorithm for the numerical solution
of the Stokes problem. IMA J. Numer. Anal., 4(4):441–455, 1984. pages 307
[376] R. Verfürth. Error estimates for a mixed finite element approximation of the Stokes equation.
RAIRO, Anal. Num., 18:175–182, 1984. pages 353
[377] R. Verfürth. Robust a posteriori error estimates for nonstationary convection-diffusion equa-
tions. SIAM J. Numer. Anal., 43(4):1783–1802, 2005. pages 118
[378] R. Verfürth. A posteriori error estimation techniques for finite element methods. Numerical
Mathematics and Scientific Computation. Oxford University Press, Oxford, UK, 2013. pages
118, 124, 126
[379] V. V. Voevodin. The question of non-self-adjoint extension of the conjugate gradients method
is closed. USSR Comput. Maths. Math. Phys., 23(2):143–144, 1983. pages 53
[380] M. Vogelius. An analysis of the p-version of the finite element method for nearly incompress-
ible materials. Uniformly valid, optimal error estimates. Numer. Math., 41(1):39–53, 1983.
pages 217
[381] M. Vohralı́k. On the discrete Poincaré-Friedrichs inequalities for nonconforming approxima-

tions of the Sobolev space H 1 . Numer. Funct. Anal. Optim., 26(7-8):925–952, 2005. pages
119
[382] M. Vohralı́k. A posteriori error estimates for lowest-order mixed finite element discretizations
of convection-diffusion-reaction equations. SIAM J. Numer. Anal., 45(4):1570–1599, 2007.
pages 330
[383] M. Vohralı́k. Unified primal formulation-based a priori and a posteriori error analysis of
mixed finite element methods. Math. Comp., 79(272):2001–2032, 2010. pages 331, 336
[384] M. Vohralı́k. Guaranteed and fully robust a posteriori error estimates for conforming dis-
cretizations of diffusion problems with discontinuous coefficients. J. Sci. Comput., 46(3):
397–438, 2011. pages 333
[385] M. Vohralı́k and B. I. Wohlmuth. Mixed finite element methods: implementation with one
unknown per element, local flux expressions, positivity, polygonal meshes, and relations to
other methods. Math. Models Methods Appl. Sci., 23(5):803–838, 2013. pages 330
References 419
[386] J. Wang and X. Ye. A weak Galerkin mixed finite element method for second order elliptic
problems. Math. Comp., 83(289):2101–2126, 2014. pages 183
[387] J. Wang and X. Ye. A weak Galerkin finite element method for the Stokes equations. Adv.
[388] K. Washizu. On the variational principles of elasticity and plasticity. Technical Report 25-18,
MIT, Cambridge, MA, 1955. pages 216
[389] A. J. Wathen. Realistic eigenvalue bounds for the Galerkin mass matrix. IMA J. Numer.
Anal., 7(4):449–457, 1987. pages 49
[390] A. J. Wathen and D. J. Silvester. Fast iterative solution of stabilised Stokes systems. I. Using
simple diagonal preconditioners. SIAM J. Numer. Anal., 30(3):630–649, 1993. pages 307
[391] C. Weber. A local compactness theorem for Maxwell’s equations. Math. Methods Appl. Sci.,
2(1):12–25, 1980. pages 231
[392] W. L. Wendland. Strongly elliptic boundary integral equations. In The state of the art in
numerical analysis (Birmingham, UK, 1986), volume 9 of The Institute of Mathematics and
its Applications Conference Series. New Series, pages 511–562. Oxford Univ. Press, New
York, NY, 1987. pages 32
[393] P. Wesseling. An introduction to multigrid methods. Pure and Applied Mathematics. John
Wiley & Sons, Ltd., Chichester, UK, 1992. pages 54
[394] M. F. Wheeler. An elliptic collocation-finite element method with interior penalties. SIAM
J. Numer. Anal., 15:152–161, 1978. pages 169
[395] T. P. Wihler. Locking-free DGFEM for elasticity problems in polygons. IMA J. Numer.
Anal., 24(1):45–75, 2004. pages 218
[396] J. Xu and L. Zikatanov. A monotone finite element scheme for convection-diffusion equations.
Math. Comp., 68(228):1429–1446, 1999. pages 110
[397] J. Xu and L. Zikatanov. Some observations on Babuška and Brezzi theories. Numer. Math.,
94(1):195–202, 2003. pages 30
[398] K. Yosida. Functional analysis. Classics in Mathematics. Springer-Verlag, Berlin, Germany,

1995. Reprint of the sixth (1980) edition. pages 377, 378, 380, 382, 383, 384
[399] A. Younes, P. Ackerer, and G. Chavent. From mixed finite elements to finite volumes for
elliptic PDEs in two and three dimensions. Internat. J. Numer. Methods Engrg., 59(3):
365–388, 2004. pages 330
[400] E. Zeidler. Applied functional analysis, volume 108 of Applied Mathematical Sciences.
Springer-Verlag, New York, NY, 1995. pages 377
[401] S. Zhang. A new family of stable mixed finite elements for the 3D Stokes equations. Math.
Comp., 74(250):543–554, 2005. pages 371
[402] S. Zhang. On the P1 Powell-Sabin divergence-free finite element for the Stokes equations. J.
420 Bibliography
[403] S. Zhang. Quadratic divergence-free finite elements on Powell-Sabin tetrahedral grids. Cal-
colo, 48(3):211–244, 2011. pages 369, 371
[404] S. Zhang and S. Zhang. C0 P2 -P0 Stokes finite element pair on sub-hexahedron tetrahedral
grids. Calcolo, 54(4):1403–1417, 2017. pages 367
[405] L. Zhong, S. Shu, G. Wittum, and J. Xu. Optimal error estimates for Nedelec edge elements
for time-harmonic Maxwell’s equations. J. Comput. Math., 27(5):563–572, 2009. pages 241
Index
Symbols bijective map, 377

M -matrix, 51, 109 bilinear form, 7
T -coercivity, 15 BNB theorem, 14, 16
Vs , V♯ , 34 boundary condition, 1
Z-matrix, 51, 109 boundary penalty in H(curl), 244
(bnb1)–(bnb2) conditions, 14 broken gradient, 148
bubble function, 353
A bulk chasing, 126
a posteriori (equilibrated flux), 334
a posteriori (residual-based), 118 C
adjacency graph, 62 Céa’s lemma, 26
adjacent set, 62 Cauchy–Navier formulation, 212, 338
CG algorithm, 53, 54
adjoint consistency (dG), 168
checkerboard instability, 346
adjoint consistency (Nitsche), 164
Choleski’s factorization, 52
adjoint operator, 382
closed range theorem, 384
adjoint problem, 98
clustering of eigenvalues, 54
algebraic multiplicity, 254
coercive operator, 391
Ampère’s law, 225
coercivity (form), 12
anisotropic diffusion, 80
coercivity (modulus), 391
annihilator, 383
compliance tensor, 223
antilinear form, 7
condition number (form), 15, 95
antilinear map, 7
condition number (matrix), 47
approximability, 27, 37
conforming approximation, 22, 38
approximability obstruction, 251
conjugate gradient, 53
Aronszajn–Cordes theorem, 80
consistency error, 35
ascent, 254
consistency term (dG), 168
Aubin–Nitsche lemma, 98, 99
consistency term (Nitsche), 160
augmented Lagrangian, 308
consistency/boundedness, 35
average across an interface, 168 continuous spectrum, 254
control on pressure gradient, 352
B COO (coordinate format), 66
Babuška’s lemma, 26 Crouzeix–Raviart finite element, 146
Babuška–Brezzi condition, 296 Crouzeix–Raviart mixed element, 368
Babuška–Brezzi vs. BNB, 297 CSC (compressed sparse columns), 59
Banach closed range theorem, 384 CSR (compressed sparse rows), 59
Banach open mapping theorem, 384 curl-preserving lifting, 236, 239
Banach–Nečas–Babuška theorem, 14 Cuthill–McKee ordering, 64
Banach–Steinhaus theorem, 378
best-approximation in H 1 , 96 D
BFS (breadth-first-search), 63 Darcy’s equations, 3
422 Index
Darcy’s law, 315 Fredholm alternative, 256

Dirichlet boundary condition, 1
Dirichlet condition (algebraic), 107 G
Dirichlet condition (Darcy), 315 Galerkin orthogonality, 26
Dirichlet–Neumann conditions, 87 Gauss law, 225
discontinuous Petrov–Galerkin, 312 Gaussian elimination, 52
discrete BNB theorem, 23 generalized eigenvalue problem, 269
discrete compactness, 237 generalized eigenvectors, 254, 277
discrete gradient, 175 geometric multiplicity, 254
discrete maximum principle, 109 graph (edges, vertices), 62
discrete Poincaré–Steklov (curl), 236 graph coloring, 66
discrete Poincaré–Steklov inequality, 149 Green’s formula, 2
discrete Sobolev inequality, 171 Gårding’s inequality, 132
discrete solution map, 28, 94
discrete solution space, 22 H
discrete test space, 22 Hahn–Banach theorem, 380
discrete trial space, 22 Hellinger–Reissner functional, 216
dispersion error, 142 Helmholtz decomposition, 234, 304, 342
Hermitian sesquilinear form, 13
divergence formula, 2
Hilbert basis, 258, 379
double dual, 381
Hilbert–Schmidt operator, 256
dual mixed formulation, 320
homogeneous Dirichlet condition, 2
dual variable, 3, 315
Hsieh–Clough–Tocher mesh, 369
duality argument, 98, 140, 154, 305
hybrid high-order (HHO), 179, 220, 373
duality argument (dG), 173
hybridizable dG (HDG), 187
duality argument (Maxwell), 231, 239
hybridization (mixed formulation), 327
duality argument (Nitsche), 163
I
E ill-conditioning, 48
eddy current problem, 227 incomplete LU (ILU), 54
eigenvalue, eigenvector, 254 inf-sup condition, 15, 23
elliptic PDE, 80 inf-sup condition (adjoint), 24, 390
elliptic projection, 99 inf-sup condition (projection), 30
elliptic regularity, 88 injective map, 377
Ellpack format, 59 irregular crisscross mesh, 369
energy functional, 14 ISO (independent set ordering), 64
equilibrated flux, 332
essential boundary condition, 84 K
Korn’s inequalities, 214
F Krylov subspace, 53, 310
face localization (diffusive flux), 196
face-to-cell lifting, 196 L
Faraday’s equation, 225 Lagrangian, 294, 320
finite element star, 119, 332 Lamé coefficients, 211
flux (Darcy), 315 Laplace equation, 1
flux recovery (Crouzeix–Raviart), 156 Lax Principle, 34
flux recovery (dG), 177 Lax–Milgram lemma, 12
flux recovery (HHO), 188 left inverse, 377
flux recovery (Lagrange), 331 Leray projection, 342
Fortin operator, 24, 351, 352 level set (graph), 62
Subject Index 423
lifting (Dirichlet condition), 103 pickup (elliptic regularity), 88

lifting (jump), 174 Poincaré–Steklov in H(curl), 235
linear form, 7 Poincaré–Steklov inequality, 119, 268
linear map, 7 point spectrum, 254
linear operator, 378 Poisson coefficient, 212
Lions’ theorem, 386 Poisson equation, 1, 79
local mass balance (Stokes), 367 pollution error, 142
localization of dual norm, 119 post-processing (mixed formulation), 330
locking, 217, 348 potential (Darcy), 315
LU factorization, 52 Powell–Sabin mesh, 369
preconditioning, 54, 311
M pressure robust, 345
macroelement partition, 365 primal variable, 3, 315
macroelement techniques, 363 principle of virtual work, 214
mass matrix, 46, 269
matrix-vector multiplication, 59, 60 Q
maximum principle, 108 quadrature formulas (2D), 71
min-max principle, 268 quadrature formulas (3D), 72
minimal residual, 312 quadrature nodes, 69
Minres, 310 quadrature order, 69
mixed boundary condition (Darcy), 318 quadrature weights, 69
mixed boundary conditions, 87 quasi-optimal estimate, 27
mixed formulation, 3
monotone operator, 391 R
multicolor ordering, 66 Rannacher–Turek mixed element, 373
multiplier assumption, 91 Rayleigh quotient, 267
reflexive Banach space, 381
N regularity pickup (Stokes), 342
natural boundary condition, 84 Rellich identity, 134
Neumann boundary condition, 86 reordering (BFS), 63
Neumann condition (Darcy), 317 reordering (CMK), 64
Nitsche’s boundary penalty, 160 reordering (multicolor), 66
non-homogeneous Dirichlet condition, 83 residual spectrum, 254
nonconforming approximation, 37 resolvent set, 253
nonobtuse mesh, 110 Riesz–Fréchet theorem, 382
nonsingular M -matrix, 51, 109 right inverse, 377, 386, 387
numerical flux (dG), 177 rigid displacement, 212
Robin boundary condition, 84
O
open mapping theorem, 384 S
oscillation indicators, 124 saddle point, 294
Schatz lemma, 140
P Schauder’s theorem, 388
Parseval’s formula, 258, 380 Schur complement, 47, 290, 306
Peetre–Tartar lemma, 388 Scott–Vogelius elements, 368
penalty parameter, 160 self-adjoint operator, 383
penalty term (dG), 168 separable space, 379
penalty term (Nitsche), 160 sesquilinear form, 7
permutation matrix, 52 simplicial barycentric (d+1)-sected, 369
Petrov–Galerkin approximation, 22 singular perturbation, 95
424 Index
singular values, 50 V
SIP, IIP, NIP (dG), 174 variational crimes, 33
smoothing property, 99, 305, 324 variational formulation, 13, 23, 83, 86
solution space, 7 Verfürth’s inverse inequalities, 124
Sommerfeld radiation condition, 131 viscous stress tensor, 337
sparse direct solvers, 52
spectral problem, 259 W
spectral radius of an operator, 254 weak convergence, 381, 382
spectrum (operator), 253 weak solution, 3
spectrum of compact operator, 257 weakly acute mesh, 109
spurious pressure mode, 346, 348 well-balanced mixed scheme, 345, 373
stable pair (Stokes), 343 well-posed problem, 12
Stampacchia’s truncation technique, 108 X
standard Galerkin approximation, 22 Xu–Zikatanov lemma, 28
static condensation, 47, 328
stiffness matrix, 45, 269 Y
strain rate tensor, 211, 337 Young modulus, 212
Strang’s first lemma, 40, 113 Young’s inequality, 379
Strang’s second lemma, 41, 153
stress tensor, 211
strong consistency, 25, 42
strongly connected graph, 62
surjective map, 377
surjectivity of divergence, 317, 340
symmetric (operator), 257
symmetric bilinear form, 13
symmetric interior penalty (dG), 167
T
Tantardini–Veeser lemma, 29
test function, 7
test space, 7
time-harmonic regime (Maxwell), 226
total stress tensor, 338
transmission problem, 186
transpose and adjoint, 257
trial function, 7
trial space, 7
twice quadrisected crisscrossed, 369
U
undirected adjacency graph, 62
undirected graph, 62
uniformly bounded, 34, 37
uniformly stable, 27, 34, 37
unique continuation principle, 80
unstable pair (Stokes), 343
Uzawa algorithm, 309

volII Hal

Uploaded by

Copyright:

Available Formats

volII Hal

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

volII Hal

Uploaded by

Copyright:

Available Formats

Finite Elements II: Galerkin Approximation, Elliptic

and Mixed PDEs

To cite this version:

HAL Id: hal-03226050

HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est

Alexandre Ern Jean-Luc Guermond

May 13, 2021

Part V. Weak formulations and well-posedness

24 Weak formulation of model problems 1

25 Main results on well-posedness 11

Part VI. Galerkin approximation

26 Basic error analysis 21

27 Error analysis with variational crimes 33

Part VII. Elliptic PDEs: conforming approximation

31 Scalar second-order elliptic PDEs 79

32 H 1 -conforming approximation (I) 93

33 H 1 -conforming approximation (II) 103

34 A posteriori error analysis 117

35 The Helmholtz problem 131

Part VIII. Elliptic PDEs: nonconforming approximation

36 Crouzeix–Raviart approximation 145

37 Nitsche’s boundary penalty method 159

38 Discontinuous Galerkin 167

39 Hybrid high-order method 179

40 Contrasted diffusivity (I) 193

41 Contrasted diffusivity (II) 201

Part IX. Vector-valued elliptic PDEs

42 Linear elasticity 211

43 Maxwell’s equations: H(curl)-approximation 225

44 Maxwell’s equations: control on the divergence 233

45 Maxwell’s equations: further topics 243

Part X. Eigenvalue problems

46 Symmetric elliptic eigenvalue problems 253

47 Symmetric operators, conforming approximation 265

48 Nonsymmetric problems 277

Part XI. PDEs in mixed form

49 Well-posedness for PDEs in mixed form 287

50 Mixed finite element approximation 301

51 Darcy’s equations 315

52 Potential and flux recovery 327

53 Stokes equations: Basic ideas 337

54 Stokes equations: Stable pairs (I) 351

55 Stokes equations: Stable pairs (II) 363

C Bijective operators in Banach spaces 377

Weak formulation of model

24.1 A second-order PDE

−∆u = f in D u = 0 on ∂D, (24.1)

where the Laplace operator is defined by ∆u := ∇·(∇u). In Cartesian coordinates, we have

24.1.1 First weak formulation

Equation (24.2) is equivalent to the PDE in (24.1)

24.1.2 Second weak formulation

σ + ∇p = 0 in D, ∇·σ = f in D, p = 0 on ∂D. (24.9)

24.1.3 Third weak formulation

24.2 A first-order PDE

W 1,1 (D) := {v ∈ L1 (D) | v ′ ∈ L1 (D)}, (24.17)

where as usual we interpret the derivatives in the weak sense.

24.2.1 Formulation in L1 (D)

24.2.2 Formulation in L2 (D)

24.3 A complex-valued model problem