Numerische Mathematik (2020) 146:369–400 Numerische

https://doi.org/10.1007/s00211-020-01141-z Mathematik

Efficient algorithms for solving the p-Laplacian in

polynomial time

Sébastien Loisel1

Received: 15 August 2018 / Revised: 19 March 2020 / Published online: 24 August 2020
© The Author(s) 2020

Abstract
The p-Laplacian is a nonlinear partial differential equation, parametrized by p ∈
[1, ∞]. We provide new numerical algorithms,
√ based on the barrier method, for solv-
ing the p-Laplacian numerically in O( n log n) Newton iterations for all p ∈ [1, ∞],
where n is the number of grid points. We confirm our estimates with numerical exper-
iments.

Mathematics Subject Classification 65H20 · 65N22 · 90C25

1 Introduction

Let  ⊂ Rd . For 1 ≤ p < ∞, the p−Laplace equation is

p−2
∇ · (∇v2 ∇v) = f in  and v = g on ∂, (1)
 1/2
d
where w2 = j=1 |w j |
2 is the usual 2−norm on Rd . Prolonging g from ∂
to the interior  and setting u = v − g, the variational form is
 
1, p 1 p
Find u ∈ W0 () such that J (u) = ∇(u + g)2 − f u is minimized.
p  
(2)

A similar definition can be made in the case p = ∞ and will be discussed in Sect. 3.1.
For p = 1, the p-Laplacian is also known as Mean Curvature, and a solution with
f = 0 is known as a minimal surface [31]. The 1-Laplacian is related to a certain
“pusher-chooser” game [19] and compressed sensing [7]. The general p-Laplacian is

B Sébastien Loisel
S.Loisel@hw.ac.uk

1 Department of Mathematics, Heriot-Watt University, Riccarton EH14 4AS, UK

123
370 S. Loisel

used for nonlinear Darcy flow [11], modelling sandpiles [2] and image processing [8].
We also mention the standard text of Heinonen et al. [16]; as well as the lecture notes
of Lindqvist [21].
One may discretize the variational form (2) using finite elements; we briefly outline
this procedure in Sect. 2.1 and refer to Barrett and Liu [3] for details. 
One chooses
piecewise linear basis functions {φ j (x)} on  and we let u h (x) = j u j φ j (x).
The energy J (u h ) can be approximated by quadrature; the quadrature is exact if the
elements are piecewise linear. This leads to a finite-dimensional energy functional

Find u ∈ Rn such that J (u) = c T u

⎛ ⎞p
2
1 m d
+ ωi ⎝ ( j)
(D u + b )i( j) 2 ⎠
is minimized, (3)
p
i=1 j=1

where D ( j) is a numerical partial derivative, b( j) = D ( j) g comes from the boundary

conditions g and c comes from the forcing term f .
Several algorithms have been proposed to minimize the convex functional J (u).
Huang et al. [18] proposed a steepest descent algorithm on a regularized functional
Jh, (u) which works well when p > 2. Tai and Xu [36] proposed a subspace correction
algorithm which works best when p is close to 2 but whose convergence deteriorates
when p → 1 or p → ∞. Algorithms based on a multigrid approach (e.g. Huang
et al. [9]) suffer from the same problems when p approaches 1 or ∞. The algorithm
of Oberman [30] also works for p ≥ 2, although the convergence factor deteriorates
after several iterations so it is difficult to reach high accuracy with this method.
The problem of minimizing J (u) has much in common with the problem of min-
imizing a p-norm, which is by now well-understood. The motivation for optimizing
a p-norm is often given as a facility location problem [1,6]. Efficient algorithms for
solving such problems can be obtained within the framework of convex optimization
and barrier methods; see Hertog et al. [17] and Xue and Ye [38] specifically for p-norm
optimization; and for general convex optimization, see Nesterov and Nemirovskii [28],
Boyd and Vandenberghe [5] and Nesterov [27] and references therein.
Given a ν-self-concordant√barrier for a convex problem, it is well-known that the
solution can be found in O( ν log ν) Newton iterations. However, the “hidden con-
stant” in the big-O notation depends on problem parameters, including the number n
of grid points in a finite element discretization. Our main result is to estimate these
hidden
√ constants and show that the overall performance of our algorithm is indeed
O( n log n).

123
Efficient algorithms for solving the p-Laplacian 371

Theorem 1 Assume that  ⊂ Rd is a polytope and that Th is a quasi-uniform tri-

angulation of , parametrized by 0 < h < 1 and with quasi-uniformity parameter
1 ≤ ρ < ∞. Let 1 ≤ p < ∞. Assume g ∈ W 1, p () is piecewise linear on Th and let
1, p
Vh ⊂ W0 () be the piecewise linear finite element space on Th whose trace vanishes.
Let R ≥ R ∗ := 2(1 + g X p ), where || is the volume of  and
p


p p
g X p = ∇g2 d x.


Let  > 0 be a tolerance. In exact arithmetic, the barrier method of Sect. 2.3, with barrier
(25), to minimize J (u) over u ∈ Vh , starting from (u, s) = (0, ŝ) given by (31), converges
in at most N ∗ iterations, where
 
N ∗ ≤ 14.4 ||h −d d! log h −1−17d R 5 (1 + g X p ) −1 + K ∗ . (4)

Here, the constant K ∗ = K ∗ (, ρ) depends on the domain  and on the quasi-uniformity
parameter ρ of Th . At convergence, u satisfies

J (u) ≤ min J (v) + .

v∈Vh
1 p
p v+g X p ≤R

The global minimizer in Vh can be found by choosing a sufficiently large value of R. We

give two cases where R is sufficiently large (1 < p < ∞ and p = 1.)
Case 1 < p < ∞: For any 1 < p < ∞, assume that f ∈ L q () where 1p + q1 = 1.
p
Assume that  fits inside a strip of width L. The value of R = R1< p<∞ = 2 + 8g X p +
  1 q
4L q 2p 1− p ( p − 1) f  L q always produces the minimizer u of the energy J (u) in the
finite element space, and the number of iterations is bounded by

N1< p<∞ (5)

 
≤ 14.4 ||h −d d! log h −1−17d (1 + g X p + pL q  f  L q )5  −1 + K ∗ .
p q

Case p = 1: Assume that L f  L ∞ < 1. The choice R = R p=1 = 2 + 2g X 1 /(1 −

L f  L ∞ ) always produces a global minimizer u of J (u) and the number of iterations is
bounded by
   5  
2g X 1
N p=1 ≤ 14.4 ||h −d d! log h −1−17d 2+  −1 ∗
+ K . (6)
1 − L f  L ∞

Computational complexity: When we vary 0 < √ h < 1 while freezing all other parame-
ters, the three estimates (4), (5) and (6) are O( n log n), where n is the number of grid
points in Th .
√
The iteration count O( n log n) also holds if  is not frozen, provided that  −1 grows
at most polynomially in n.

123
372 S. Loisel

We emphasize that the p = 1, ∞ cases have up to now been considered to be

especially hard and there are no other algorithm that offers any performance guarantees
in these situations. Estimate (6) is the first algorithm that is known to converge even
when p = 1. We also have an algorithm for the p = ∞ case and we provide a
corresponding iteration count estimate in Theorem 3.
The algorithm mentioned in Theorem 1 is the barrier method of convex optimiza-
tion. Consider the problem of minimizing c T x subject to x ∈ Q where Q is some
convex set. Assume we have a ν-self-concordant barrier F for Q. The barrier method
works by minimizing tc T x + F(x) for larger and larger values of the barrier parameter
t, which is increased iteratively according to some schedule. In the short step vari-
ant, t increases slowly and the method is very robust; the estimates of Theorem 1 are
for the short step variant of the barrier method. It is well-known that the short step
barrier method has the theoretically best convergence estimates, but that long step
variants (where t increases more rapidly) work better in practice. However, long step
algorithms have theoretically worse convergence estimates and, as we will see in the
numerical experiments, can sometimes require a large number of Newton iterations
to converge.
In order to get the best convergence, we have devised a new, very simple adaptive
stepping algorithm for the barrier method. There are already many adaptive stepping
algorithms (see e.g. Nocedal et al. [29] and references therein). It is often difficult
to prove “global convergence” of these algorithms, and we are not aware of global
estimates of Newton iteration counts. With our new, highly innovative algorithm,
we are able to prove “quasi-optimal” convergence of our adaptive √ scheme. Here,
quasi-optimal means that our adaptive algorithm requires Õ( n) Newton iterations,
neglecting logarithmic terms, which is the same as the theoretically optimal short-step
algorithm of Theorem 1.
The p-Laplacian is subject to roundoff problems when p is large but finite, as
p
we now briefly describe. Consider the problem of minimizing v p in the space
{v = (1, y)}. Assume that we are given a machine  (for example,  ≈ 2.22 × 10−16
in double precision) and consider an arbitrary vector v = [1, δ]T . In this situation
p
v p = 1 p + δ p = 1 in machine precision, provided δ <  1/ p . This means that
a region of size  1/ p near the minimum is numerically indistinguishable from the
true minimum when computing the energy, causing a very large relative error in the
solution. This phenomenon becomes worse in higher dimensions and when composing
with matrices with large condition numbers as in (3). This means that all algorithms,
including our own, will struggle to produce highly accurate solutions when p 2.
In particular, for p = 5, we see that δ < 7.4 × 10−4 is best possible, and this is
made worse by the condition number of the differential matrices. However, although
the problem is numerically challenging for finite p 2, the problem becomes easy
again when p = ∞. Our second main result is an estimate for the p = ∞ case in
Sect. 3.1, and we confirm by numerical experiments that there are no numerical issues
for p = ∞.
Our algorithm is an iterative scheme for a high-dimensional problem arising from a
partial differential equation. Each iteration involves the solution of a linear problem that
can be interpreted as a numerical elliptic boundary value problem. One can estimate
pessimistically that solving each linear problem requires O(n 3 ) FLOPS, for a total

123
Efficient algorithms for solving the p-Laplacian 373

cost of O(n 3.5 log n) FLOPS for our entire algorithms. This estimate can be improved
by using an O(n 2.373 ) FLOPS fast matrix inverse [20], making our overall algorithms
O(n 2.873 log n) FLOPS; we mention that this matrix inversion algorithm mostly of
theoretical interest since it is not practical for any reasonable value of n. We have
taken special care to preserve the sparsity
√ of this problem so that, if one assumes a
bandwidth of b (e.g. typically b = O( n) for d = 2 and b = O(n 2/3 ) for d = 3), one
obtains an O(b2 n) sparse matrix solve algorithm, resulting in O(n 2.5 log n) (d = 2)
or O(n 2.84 log n) (d = 3) FLOPS for our overall algorithms. In addition, we mention
many preconditioning opportunities [4,10,12,14,15,22–26,35]. Although solution by
preconditioning is possible, it is difficult to estimate the number of iterations a priori
since the diffusion coefficient of the stiffness matrix is difficult to estimate a priori;
in the best case (“optimal preconditioning”) where the elliptic solve at each Newton
iteration can be done in O(n) FLOPS, our algorithms are then O(n 1.5 log n) FLOPS.
Our paper is organized as follows. In Sect. 2, we give some preparatory material
on the p-Laplacian and the barrier method. In Sect. 3, we prove our main theorem for
1 ≤ p < ∞ and a separate theorem for the case p = ∞. In Sect. 4, we validate our
algorithms with numerical experiments. We end with some conclusions.

2 Preparatory material

We now discuss some preparatory material regarding the p−Laplacian, for 1 ≤ p ≤

∞. The ∞-Laplacian can be interpreted as the problem of minimizing

J (u) = u + g X ∞ () − f u where v X ∞ () = sup ∇v(x)2 . (7)
 x∈

Note that (7) is not a limit as p → ∞ of (2), e.g. because (2) uses the pth power of
 ·  X p in its definition.

Lemma 1 For 1 ≤ p ≤ ∞, J (u) is convex on W 1, p (). For 1 < p < ∞, J (u) is

1, p
strictly convex on W0 ().

Proof We consider the case 1 ≤ p < ∞ in detail, the case p = ∞ is done in

a similar fashion. Convexity (and strict convexity) is unaffected by linear shifts so
without loss of generality we assume that f = 0. Let 0 ≤ t ≤ 1. We must show that
J (tu + (1 − t)v) ≤ t J (u) + (1 − t)J (v). To simplify the notation, let q = ∇u, r = ∇v
and s = ∇g.
 
p (∗)
J (tu + (1 − t)v) = tq + (1 − t)r + s2 ≤ (tq + s2 + (1 − t)r + s2 ) p
 
 
(∗∗) p p
≤ tq + s2 + (1 − t)r + s2 = t J (u) + (1 − t)J (v),


where we have used the triangle inequality for  · 2 at (∗) and the convexity of
φ(z) = z p at (∗∗).

123
374 S. Loisel

We now prove strict convexity for the 1 < p < ∞ case. If we have equality at
(∗) then q(x) + s(x) and r (x) + s(x) are non-negative multiples of one another, i.e.
q + s = aw and  r + s = bw where a(x), b(x) ≥ 0 and w(x) is vector-valued. Then
p
(∗∗) becomes  ((ta + (1 − t)b)w2 ) p ≤  (ta p + (1 − t)b p )w2 . Note that
(ta + (1 − t)b) p < ta p + (1 − t)b p unless a = b so the inequality (**) is strict unless
1, p
∇u = ∇v almost everywhere. Since u, v ∈ W0 can be identified by their gradients,
we have proven strict convexity. 

 
max 0, 1p − q1
From the norm equivalence u p ≤ d uq for x ∈ Rd , one obtains
   
− max 0, 1p − 21 max 0, 21 − 1p
d |u|W 1, p ≤ u X p () ≤ d |u|W 1, p . (8)

We can give a modified Friedrichs inequality for  ·  X p .

Lemma 2 (Friedrichs inequality for  ·  X p ) Assume that  ⊂ Rd fits inside of a
1, p
strip of width and assume that φ ∈ W0 (), where 1 ≤ p ≤ ∞. Then, φ L p ≤
− 1p 1
p φ X p , where we define ∞− ∞ = 1.
Proof Without loss of generality, assume that  is inside the strip 0 ≤ x1 ≤ . From
the fundamental theorem of calculus, the following argument proves the p = ∞
x 
case: |φ(x1 , . . . , xd )| ≤ 0 1 |φx1 (t, x2 , . . . , xd )| d x ≤ 0 supx∈ ∇φ(x)2 d x1 ≤
φ X ∞ . Now assume 1 ≤ p < ∞.
   x 1 p
 
|φ| d x1 =
p  φx1 (t, x2 , . . . , xd ) dt  d x1 (9)

0 0 0
  x1
p−1
≤ |φx1 (t, x2 , . . . , xd )| p x1 dt d x1 (Jensen’s ineq.) (10)
0 0
 
p p−1
≤ ∇φ(t, x2 , . . . , xd )2 x1 d x1 dt (11)
0 t

p 1
= ∇φ(t, x2 , . . . , xd )2 ( p
− t p ) dt (12)
0 p
p 
p
≤ ∇φ(t, x2 , . . . , xd )2 dt. (13)
p 0

The result follows by integrating over x2 , . . . , xd . 


We now give an a priori estimate on the magnitude of the minimizer of J (u). This
estimate will be important in the design of our algorithm in order to limit the search
volume to some ball of reasonable size.
Lemma 3 Let 1 < p < ∞ and 1p + q1 = 1 and assume that  ⊂ Rd is a domain of
p  p 1, p
width L. Let v X p =  ∇v2 . Assume {u k } ⊂ W0 () is a minimizing sequence
for J (u). Then, for large enough k,
 p 1
p p 1− p q
u k  X p ≤ 4g X p + 2L q ( p − 1) f  L q . (14)
2

123
Efficient algorithms for solving the p-Laplacian 375

If { p, q} = {1, ∞} and L f  L q < 1 then a minimizing sequence must eventually lie

in u k  X p ≤ g X p /(1 − L f  L q ).
p 
Proof Case 1 < p < ∞: For convenience, we write J (u) = p u
1
+ g X p −  f u.
Assume u X p ≥ g X p ; then:

1
J (u) ≥ (u X p − g X p ) p −  f  L q u L p
p
1 1
≥ u X p − g X p −  f  L q L p −1/ p u X p .
p p
p p

Next, we use Young’s inequality ab ≤ 1 q

qa + 1 p
pb with a = 21/ p p −1/ p L f  L q ,
b = 2−1/ p u X p to obtain

1
1 p 2 p L q ( p/2) 1− p q
J (u) − J (0) ≥ u X p − g X p −  f L q . (15)
2p p q

p p 1 q
Hence, if u X p > 4g X p + 2L q ( p/2) 1− p ( p − 1) f  L q , then J (u) − J (0) > 0
and hence a minimizing sequence must satisfy (14).
The p = 1 case is as follows:

J (u) − J (0) ≥ u X 1 − g X 1 −  f  L ∞ Lu X 1 > 0,

if u X 1 > g X 1 /(1 −  f  L ∞ L). The p = ∞ case is done in a similar fashion. 


The a priori estimate above can also be used to show the existence of a minimizer
of J (u).
1, p
Lemma 4 Let 1 < p < ∞. There is a unique u ∈ V ⊂ W0 () that minimizes J (u).

Proof Let α = inf v J (v). We now show how to produce a minimizing sequence {u k } ⊂
1, p 1, p
W0 (). For k = 1, 2, . . ., let Bk = {u ∈ W0 () | J (u) < α + 1/k and u X p <
p q
4g X p + 4L ( p − 1) f  L q + 1}, see (14). Note that each Bk is open and nonempty
2

so pick u k ∈ Bk . Furthermore, the Bk are nested: B1 ⊃ B2 ⊃ . . .; the convexity of J

implies that the Bk are also convex.
According to (8), we see that each Bk is contained  in a closed ball F = {u ∈
1, p max 0, 1p − 21  p q 
W0 () | |u|W 1, p () ≤ r } where r = d 4g X p +4L 2 ( p−1) f  L q +1 .
Recall that F is weakly compact. Passing to a subsequence if necessary, we may
assume that {u k } converges weakly to some u. By Mazur’s lemma, we can now find
 J (k)
some convex linear combinations vk = j=k α j u j ∈ Bk such that {vk } converges
to u strongly. This shows that u belongs to every Bk and hence J (u) is minimal.
Uniqueness follows by strict convexity. 


123
376 S. Loisel

2.1 Finite elements

Assume that  is a polygon. We introduce a triangulation Th of , parametrized by

0 < h < 1, and piecewise linear finite element basis functions {φ1 (x), . . . , φn (x)} ⊂
W 1, p (). As  usual, we define a “reference element” K̂ = {x ∈ Rd : xi ≥ 0 for i =
d
1, . . . , d and i=1 xi ≤ 1}. Each simplex K k in Th can be written as K k = k ( K̂ ) =
P K̂ + q , where P (k) ∈ Rd×d and q (k) ∈ Rd . If Th is a uniform lattice of
(k) (k)

squares or d-cubes, then each P (k) is of the form diag(±h, . . . , ±h), and ~P (k) ~2 =
~[P (k) ]−1 ~−1 2 = h. In general, if Th is not necessarily a uniform lattice, we say that
the family of triangulations  Th , parametrized by 0 < h < 1, is quasi-uniform with
parameter ρ < ∞ if h ≤ ~P (k) ~2 , ~[P (k) ]−1 ~−1 2 ≤ ρh. Note that on the reference

simplex, the basis functions are φ̂i (x̂) = x̂i for i = 1, . . . , d and φ̂0 (x̂) = 1 − i x̂i .
√ √
As a result, ∇ φ̂2 ≤ d and, from the chain rule, ∇φi (x)2 ≤ h −1 d.
Let span{φk (x) | k = 1, . . . , n} ⊂ W1, p () be the finite m element(i)space of
piecewise linear elements over Th and let  w(x) d x ≈ i=1 ωi w(x ) be the
midpoint quadrature rule, which is exact for piecewise linear or piecewise constant
functions. We can construct a “discrete derivative” matrix D ( j) whose (i, k) entry is
( j)
Di,k = ∂φ (i)
∂ x j (x ). Then,
k

⎛ ⎞p
 
m d  2 2
1 ωi ⎝ (D ( j) (u + g))i ⎠ ;
p
∇(u + g)2 =
p  p
i=1 j=1

note that the quadrature is exact provided that g is also piecewise linear. For the
midpoint rule, ωi is the volume of the simplex K i ; if the triangulation Th is quasi-
uniform then we find that

hd ρd hd
≤ ωi ≤ ; (16)
d! d!

we write ωi = (h d ), which means both that ωi = O(h d ) and h d = O(ωi ). We abuse

the notation and use the same symbol u to represent both the finite element coefficient
vector [u 1 , . . . , u n ]T and the finite element function u(x) = nk=1 u k φk (x).
( j)
We further denote by D the columns of D ( j) corresponding to the vertices of
( j)
Th in ∂, and D I corresponding to the interior vertices in , such that D ( j) =
( j) ( j)  T  T
D I D . Denoting u = u TI u T = u TI 0T , note that D ( j) (u + g) =
( j)
D I u I + D ( j) g. Putting b( j) = D ( j) g and dropping the subscript I leads to
d (k) T (k)
the discretized system (3). The matrix A = k=1 [D I ] diag(ω1 , . . . , ωm )D I
is the usual discretization
 of the Laplacian or Poisson problem, and we have that
u T Au = |u|2H 1 =  ∇u22 d x. For a domain of width , the Friedrichs inequality
u L 2 ≤ |u| H 1 (see [33, (18.1) and (18.19)]) proves that the smallest eigenvalue of
the Laplacian differential operator is at least −2 ; however, the smallest eigenvalue

123
Efficient algorithms for solving the p-Laplacian 377

of the finite-dimensional matrix A is actually (h 2 ) because the relevant Rayleigh

quotient in Rn is u ∗ Au/u ∗ u = |u|2H 1 /u2L 2 .
We now prove that the finite element method converges for the p-Laplacian.
Lemma 5 Assume that  is a polytope and 1 < p < ∞. Let u h be the finite element
minimizer of J (u) in a finite element space that contains the piecewise linear functions.
Then, J (u h ) converges to inf v J (v) as h → 0.
1, p
Proof Let u be a minimizer of J (u) and denote by Vh ⊂ W0 () the finite element
space with grid parameter 0 < h < 1. Recall that finite element functions are dense in
1, p
W0 () [13, Proposition 2.8, page 316]. Hence, we can find finite element functions
1, p
{vh ∈ Vh } that converge to u in the W0 () norm as h → 0. Since J is continuous
and since u h minimizes J (u) in the finite element space Vh , we find that J (u h ) ≤
J (vh ) → J (u) = inf v J (v), as required. 

Lemma 5 is very general (no regularity assumptions are made on the solution u) but
also very weak since it does not give a rate of convergence. If one assumes regularity
of the solution then one can use quasi-interpolation [34] to estimate the convergence
more precisely. However, we will see in Sect. 2.2 (Example 1) that it is difficult to
prove regularity. Since the present paper focuses on the numerical solver, and not
in the discretization, we do not investigate this aspect any further. The theorem also
does not specify whether u h converges as h tends to 0. In the case 1 < p < ∞, the
strict convexity of J ensures that u h will indeed converge to a u in W 1, p () but for
p = 1 there may be multiple minimizers and then u h could oscillate between the many
minimizers or converge to a “minimizer” in the double-dual of W 1,1 ().

2.2 Pathological situations for extreme p values

The p-Laplace problem varies in character as p ranges from 1 ≤ p ≤ +∞. When

p = 2, minimizing J (u) is equivalent to solving a single linear problem, which is
clearly faster than solving hundreds of linear problems as required by a barrier method.
As p gets further away from p = 2, naive solvers work less well and proper convex
optimization algorithms are required, such as our barrier methods. The extreme cases
p = 1 and p = ∞ have traditionally been considered hardest. For example, J (u) may
not be differentiable at u when p ∈ {1, ∞}, typically when ∇(u + g) vanishes at some
point x ∈ . In Sect. 2, we have introduced several lemmas, some of which work for
all cases 1 ≤ p ≤ ∞, others are restricted to 1 < p < ∞. Briefly speaking, we have
shown that for all 1 ≤ p ≤ ∞, J (u) is convex and possesses minimizing sequences
(with some restrictions on the forcing f when p ∈ {1, ∞}.) These facts are sufficient
to deploy barrier methods, because barrier methods do not require the objective to be
differentiable, be strictly convex, or have a unique minimizer. As a “bonus”, we have
also shown that J (u) is strictly convex and has a unique minimum when 1 < p < ∞,
but this is not required for the successful application of our barrier methods.
We now illustrate the pathological behavior for extreme values of p with several
simple examples. For 1 < p < ∞, strict convexity ensures the uniqueness of the
minimum of J (u). However, in the case p = 1, the minimizer may be “outside” of
W 1,1 () or nonunique.

123
378 S. Loisel

Example 1 Consider  = (0, 1) in dimension d = 1 and f = 0, with boundary

conditions u(0) = 0 and u(1) = 1 and with p = 1. Then,
 1
J (u) = |u  (x)| d x = T V (u) ≥ 1,
0

where T V (u) denotes the usual total variation of u. Any monotonically nondecreasing
function u(x) with u(0) = 0 and u(1) = 1 will minimize J (u) and satisfy the boundary
conditions.
A minimizing sequence for J (u) is the piecewise linear functions u n (x) =
min(1, max(0, 0.5 + n(x − 0.5))). This sequence converges to the indicating function
of [0.5, 1), which is not in W 1,1 (0, 1). This is because W 1,1 (0, 1) is not reflexive
and hence its unit ball is not weakly compact. Instead, the limit of u n is in BV , the
double-dual of W 1,1 (0, 1).
We now briefly show why the minimization of J (u) for u ∈ Vh is numerically
challenging.
Example 2 Consider J (x) = |x| p where x ∈ R and 1 ≤ p < ∞; this corresponds to
a 1-dimensional discrete p-Laplacian with a single grid point. The Newton iteration
xk+1 = xk − J  (xk )/J  (xk ) is

sgn(xk ) p|xk | p−1 p−2

xk+1 = xk − = xk .
p( p − 1)|xk | p−2 p−1

Hence, the Newton iteration converges linearly for p ∈ (1.5, 2) ∪ (2, ∞) and diverges
for 1 < p ≤ 1.5. The Newton iteration is undefined for p = 1 since J  = 0.
The p-Laplacian for p = 1 is particularly hard; we now show two types of difficulties.
First, the Hessian may be singular, and regularizing the Hessian leads to gradient
descent.

Example 3 Consider J (x) = x12 + x22 = x2 ; this correspond to a 2-dimensional

1-Laplacian discretized with a single grid point. The gradient is J  (x) = x
x
2
and the
Hessian is
 2 
 1 x2 −x1 x2
J (x) = .
x32 −x1 x2 x12

The Hessian matrix J  (x) is singular which makes the Newton iteration undefined. To
make matters worse, the kernel of J  is spanned by J  and hence any “regularization”
J  +  I leads to a simple gradient descent.
Yet another difficulty is that the 1-Laplacian may have nonunique solutions or no
solutions when the forcing is nonzero.
Example 4 Let c ∈ R and J (x) = |x| + cx; this corresponds to a 1-dimensional
1-Laplacian with a nonzero forcing term, discretized with a single grid point. Then,

123
Efficient algorithms for solving the p-Laplacian 379

J (x) is convex for all c ∈ R. However, J (x) has a unique minimum x = 0 if and only
if |c| < 1. When |c| = 1, J (x) has infinitely many minima. When |c| > 1, J (x) is
unbounded below and there is no minimum.
As a result, the energy J (u) of the 1-Laplacian may not be bounded below when
the forcing f = 0; see also Lemma 3.

2.3 Convex optimization by the barrier method

In this section, we briefly review the theory and algorithms of convex optimization and
refer to Nesterov [27, Section 4.2] for details, including the notion of self-concordant
barriers.
Let Q ⊂ Rn be a bounded closed convex set that is the closure of its interior, c ∈ Rn
be a vector and consider the convex optimization problem

c∗ = min{c T x : x ∈ Q}. (17)

The barrier method (or interior point method) for solving (17) is to minimize tc T x +
F(x) for increasing values of t → ∞, where the barrier function F(x) tends to ∞
when x → ∂ Q. The minimizer x ∗ (t), parametrized by t ≥ 0, of tc T x + F(x), is
called the central path, and x ∗ (t) forms a minimizing sequence1 for (17) as t → ∞.
Assume
 we have a ν-self-concordant barrier F(x) for Q. Define the norm v∗x =
v [F (x)]−1 v. The main path-following scheme is
T 

1. Set t0 = 0, β = 1/9 and γ = 5/36. Choose an accuracy  > 0 and x (0) ∈ Q such
that F  (x (0) )∗x (0) ≤ β.
2. The kth iteration (k ≥ 0) is
γ
tk+1 = tk + and x (k+1) = x (k) − [F  (x (k) )]−1 (tk+1 c + F  (x (k) )). (18)
c∗x (k)
 √ 
(β+ ν)β
3. Stop if tk ≥ ν + 1−β  −1 =: tol−1 .
The invariant of this algorithm is that, if tk c + F  (x (k) )∗x (k) ≤ β then also
tk+1 c + F  (x (k+1) )∗x (k+1) ≤ β. The stopping criterion guarantees that, at conver-
gence, c T x (k) − c∗ ≤ . Starting this iteration can be difficult, since it is not always
obvious how to find an initial point x (0) ∈ Q such that F  (x (0) )∗x (0) ≤ β. Define the
analytic center x F∗ by F  (x F∗ ) = 0. We use an auxiliary path-following scheme2 to
approximate the analytic center x F∗ of Q:
1. Choose x (0) ∈ Q and set t0 = 1 and G = −F  (x (0) ).
2. For the kth iteration (k ≥ 0):
γ
tk+1 = tk − and (19)
G∗x (k)

1 Perhaps one should say “minimizing filter”.

2 Our presentation corrects some misprints in the auxiliary path algorithm of Nesterov [27, Section 4.2].

123
380 S. Loisel

x (k+1) = x (k) − [F  (x (k) )]−1 (tk+1 G + F  (x (k) )). (20)

√
β
3. Stop if F  (x (k) )∗x (k) ≤ √
1+ β
. Set x̄ = x (k) − [F  (x (k) )]−1 F  (x (k) ).

The invariant of the auxiliary scheme is that tk G + F  (x (k) )∗x (k) ≤ β for every k. At
convergence, one can show that F  (x̄)∗x̄ ≤ β. Let x̂ ∈ Q be some starting point for
the auxiliary path-following scheme. Combining the auxiliary path-following scheme
to find the approximate analytic center x̄ of Q, followed by the main path-following
scheme to solve the optimization problem (17) starting from x (0) = x̄, completes in
at most N iterations, where
√
N = 7.2 ν 2 log ν + log F  (x̂)∗x ∗ + log x̂∗x ∗ + log(1/) . (21)
F F

2.3.1 Long-step algorithms

The path-following schemes of Sect. 2.3 are so-called “short step”, meaning that the
barrier parameter t increases fairly slowly when ν is large. It is well-known that long-
step algorithms, where t increases more rapidly, often converge faster overall than
short-step algorithms, even√though the worst case estimate O(ν log ν) is worse than
the short-step estimate O( ν log ν), see Nesterov and Nemirovskii [28] for details.
The main path-following scheme can be made “long-step” as follows:
1. Assume x (0) ∈ Q such that F  (x (0) )∗x (0) ≤ β and let t0 = 0.
2. Set
⎧ " #
⎨max κt , t + γ if tk c + F  (x (k) )∗x (k) ≤ β,
tk+1 =
k k c∗ (k) (22)
x
⎩
tk otherwise;
x (k+1) = x (k) − rk [F  (x (k) )]−1 (tk+1 c + F  (x (k) )), (23)

where 0 < rk ≤ 1 is found by line search, see e.g. Boyd and Vandenberghe [5,
 with α √
Algorithm 9.2 = 0.01 and β = 0.25].
3. Stop if tk ≥ ν + (β+1−β
ν)β
 −1 = tol−1 .

The parameter κ ≥ 1 determines the step size of the scheme. In convex optimization,
step sizes κ = 10 or even κ = 100 are often used, but we will see in Sect. 4 that
shorter step sizes are better suited for the p-Laplacian.
The long-step variant of the auxiliary path-following scheme is implemented in a
similar fashion; the criterion for decreasing tk+1 is then tk G + F  (x (k) ∗x (k) ≤ β.

2.3.2 Adaptive stepping

We finally introduce an algorithm whose step parameter κk is indexed by the iteration

counter k. We first introduce some terminology. If tk c + F  (x (k) )∗x (k) ≤ β (main
phase) or tk G + F  (x (k) )∗x (k) ≤ β (auxiliary phase), we say that x (k) was accepted,

123
Efficient algorithms for solving the p-Laplacian 381

else we say that x (k) was a slow step. Let κ0 be an initial step size (we will take
κ0 = 10.)
1. If x (k) is accepted after 2 or fewer slow steps, put κk+1 = min{κ0 , κk2 }.
√
2. If x (k) is accepted after 8 or more slow steps, put κk+1 = κk .
3. If x (k) is still not accepted after 15 slow steps, replace x (k+1) and tk+1 by the most
1/4
recently accepted step and put κk+1 = κk . We call this procedure a rejection.
4. Otherwise, put κk+1 = κk .
The quantity tk+1 is computed as in the long step algorithm (22), with κ = κk+1 . Note
that whenever tk+1 coincides with the short step (20), then the step is automatically
accepted. The rejection is “wasteful” in that it discards possibly useful information, but
we will see in the numerical experiments that this adaptive scheme is quite efficient in
practice. Furthermore, the rejection step is the key that unlocks a very simple analysis
of our algorithm.

Theorem 2 For given c, F, , let N S and N A be the number of Newton steps of the
short step and adaptive step algorithms, respectively. Then,
√
N A ≤ 160.76 + 0.73 log(1 + 9 ν)N S . (24)

Proof By construction, on each accepted step of the main path-following algorithm,

we find that tk+1 ≥ tk + cγ∗ , the short step size, see (22). Thus, we only need to
x (k)
estimate the maximum number of slow steps before a step is accepted. According to
[27, p.202], the short step size satisfies

κmin
$
 %& '
5
tk+1 ≥ 1+ √ tk .
4 + 36 ν

Starting from κ = 10, after r rejections, the step size is κ = 10(1/4) . When κ ≤ κmin ,
r

the short step is automatically accepted and hence the maximum number of rejections
is r = r− , where

log(log(κmin )/ log(10))
10(1/4)
r−
= κmin ⇒ r− = − .
log 4

Hence,
√
r ≤ 0.76 + 0.73 log(1 + 9 ν).

Since all the adaptive steps are at least as large as the short steps and the stopping
criterion is purely based on the barrier parameter tk , and noting that each rejection
corresponds to 15 slow steps (plus the initial accepted step), we obtain the estimate for
the main phase. The estimate for the auxiliary phase is obtained in a similar fashion.



123
382 S. Loisel

Theorem 2 states that the adaptive algorithm cannot be much worse than the short
√
step algorithm, which means that the adaptive algorithm scales at worse like Õ( ν),
where we have neglected some logarithms. The reader may be surprised that the
estimate for the adaptive scheme is slightly worse than the estimate for the short step
scheme, but this is a well-known phenomenon in convex optimization. The long step
estimates are quite pessimistic and in practice, long step and adaptive schemes work
much better than the theoretical estimates. Our result is especially interesting because
it is well-known that estimates
√for long step algorithms scale like Õ(ν), whereas our
new algorithm scales like Õ( ν).

3 Proof of Theorem 1

The proof of Theorem 1 is rather technical, so we begin by outlining the plan of our
proof. The idea is to estimate all the quantities in the bound (21) for the number N of
Newton iterations. The barrier parameter ν is estimated in Lemma 6. Some “uniform”
or “box” bounds are given for the central path in Lemma 7; these are an intermediate
step in converting as many estimates as possible into functions of h. Because (21)
depends on the Hessian F  of the barrier, the lowest eigenvalue λmin of F  is estimated
in Lemma 8. This bound itself depends on extremal singular values of the discrete
derivative matrices, which are estimated in Lemma 9, and these bounds are rephrased
in terms of h in Lemma 10. In Lemma 11, we establish the connection between the
number m of simplices and the grid parameter h, which is used in Lemma 12 to
estimate the quantities x̂2 and F  (x̂)2 , which can be converted to estimates for
x̂∗x ∗ and F  (x̂)∗x ∗ in (21) by dividing by λmin ; here x̂ is a starting point for the
F F
barrier method. Finally, the quantities R appearing in Theorem 1 are obtained by
starting from the estimates of Lemma 3, adding 1, and doubling them. This ensures
that the central path will be well inside the ball of radius R.
In the present section, we treat in detail the case 1 ≤ p < ∞. The case p = 1,
which is considered especially difficult, poses no special difficulty in the present
section, provided that the hypotheses of Lemma 3 are satisfied. The case p = ∞ is
deferred to Sect. 3.1.
Let 1 ≤ p < ∞ and define the barrier

  
F(u, s) = F p (u, s) = − log z i − σ log si − log τi where (25)
i i i
y ( j)

d $ %& '
[( D ( j) u + D ( j) g)i ]2 ,
2/ p
z i = si − τi = R − ωi si and (26)
j=1
(
2 if 1 ≤ p < 2 and
σ = σ ( p) = (27)
1 if p ≥ 2.

123
Efficient algorithms for solving the p-Laplacian 383

Lemma 6 The function F(u, t) is an m(σ + 2)-self-concordant barrier for the set
p
Q = {(u, s) : si ≥ ∇(u + g)| K i 2 , si ≥ 0 and max ωi si ≤ R}, (28)
i

The problem of minimizing J (u) over u ∈ Vh subject to the additional constraint that
maxi ωi ∇(u + g)| K i 2 ≤ R is equivalent to:
 
−f
min c x subject to x ∈ Q where c =
T
. (29)
ω

Here, we have abused the notation and used the symbol f for the vector whose ith
component is  f (x)φi (x) d x.
Proof The functions B p (x, s) = − log(s 2/ p − x T x) are σ + 1-self-concordant so
 m  ( j)
− i log τi + i=1 B p ([ k Di,k (u + g)k ]dj=1 , si ) is m(σ + 2) self-concordant, see
Nesterov and Nemirovskii [28]. The rest is proved by inspection. 


From Lemma 3, it is tempting to use a bound such as u X p < R, i.e. i ωu si ≤ R,
but this leads to a dense Hessian Fss . Instead, we have used the “uniform” bound:
 
ωi si ≤ ωi si = s ≤ R.
i 

With this “looser” bound, the Hessian Fss is sparse.3 Furthermore, by using the R
value from the a-priori estimate Lemma 3, one can ensure that Q is non-empty and
contains minimizing sequences for J (u). Thus, put:
⎛ ⎞p
2
d
R ≥ R = 2(1 + g X p ) = 2 + 2 ⎝ [(D g)i ] ⎠ .
∗ (
p j) 2
(30)
j=1

Set
⎛ ⎞p
2
d
R
ŝi = 1 + ⎝ ( j)
[(D g)i ] 2⎠
; hence ŝi ≤ . (31)
2
j=1

In this way, (0, ŝ) ∈ Q.

Lemma 7 For all (u, s) ∈ Q,
 2
R R p
τi ≤ R, si ≤ , zi ≤ . (32)
ωi ωi
3 The alternative barrier term − log(R −  ω s ) corresponding to

i i i  s < R leads to a rank-1 dense
Hessian matrix so one could also use this more “natural” barrier by combining it with a Woodbury identity
to invert the Hessian. This trades a slightly more complex implementation in exchange for a slightly “smaller”
set Q.

123
384 S. Loisel

2/ p
Proof From w T s ≥ 0 and (26), we find τ ≤ R. From (26), we find z i ≤ si and
from 0 ≤ τi = R − ωi s, we find ωi si ≤ R.

We further find that:

τ̂i ≥ R/2, ŝi ≥ 1, ẑ i ≥ 1. (33)

The gradient of F is:

   ( j) ]T y
( j)

F 2[D
F = u
= j z , (34)
Fs − 2p 1z s 2/ p−1 − σs + ωτ

where vector algebra is defined entrywise.

The Hessian of F is
   
 Fuu Fus Fuu Fus
F = = T where (35)
Fsu Fss Fus Fss

d 
d
Fuu = 2 [D ( j) ]T Z −1 D ( j) + 4 (Y ( j) D ( j) )T Z −2 (Y (r ) D (r ) ), (36)
j=1 j,r =1

4  ( j) ( j) T −2 2/ p−1
d
Fus = − (Y D ) Z S , (37)
p
j=1
 
2 2 4
Fss = − − 1 Z −1 S 2/ p−2 + 2 Z −2 S 4/ p−2 + σ S −2 + W 2 Z −2 , (38)
p p p
S = diag(s), W = diag(ω), Y = diag(y), Z = diag(z). (39)
d (k) T (k) and assume 0 <
k=1 [D ] D
Lemma 8 Let dmin 2 be the smallest eigenvalue of

h < 1. Let ωmin = mini ωi , and similarly for (z ∗F )max , etc... The smallest eigenvalue
λmin of F  (u ∗F , s F∗ ) is bounded below by
 
λmin ≥ min 2(z ∗F )−1 2 2 ∗ −2
max dmin , ωmin (z F )max . (40)

Proof We consider the “Rayleigh quotient” x T F x/xT x, the extremal values of which
v
are the extremal eigenvalues of F  . We put x = so that
w

x T F  x = v T Fuu v + 2v T Fus w + w T Fss w

We use the Cauchy-Schwarz inequality together with Young’s inequality to find that
⎛ ⎞ 
  
4  d T ( j) ( j) T −1  −1 2/ p−1 
2|v T Fus w| = ⎝ v (Y D ) Z ⎠ Z S w 

p  j=1 

123
Efficient algorithms for solving the p-Laplacian 385

)
* d 
8* 
≤ + v T (Y ( j) D ( j) )T Z −2 (Y (r ) D (r ) )v w T S 2/ p−1 Z −2 S 2/ p−1 w
p
j,r =1


d
4 T 2/ p−1 −2 2/ p−1
≤4 v T (Y ( j) D ( j) )T Z −2 (Y (r ) D (r ) )v + w S Z S w.
p2
j=1,r

Hence we find:


d    
 ( j) T −1 ( j) 2 2 −1 2/ p−2
x F x ≥2
T
v [DT
] Z D v+ −1 w − Z S
T
w
p p
j=1

+ w T σ S −2 w + w T W 2 Z −2 w.

We use that Fs = 0, which implies that − 2p Z −1 S 2/ p−1 = T −1 W − σ S −1 , where

T = diag(τ ) and hence


d  
2
x T F  x ≥ 2 v T [D ( j) ]T Z −1 D ( j) v +
− 1 w T T −1 W S −1 w + w T W 2 Z −2 w
p
j=1
 
−1 2 −2
≥ x22 min 2z max dmin , ωmin
2
z max .



A domain  is said to be of width L when  ⊂ S, where S is a strip of width L. The
1, p
Friedrichs inequality states that, for domains of width L > 0 and for u ∈ W0 (),
u L 2 ≤ L|u| H 1 () .
Lemma 9 Let  be a polytope of width L < ∞, and assume that the triangulation
Th , which depends on the grid parameter 0 < h < 1, is quasi-uniform. Then, there is
a constant c > 0, which depends on  and
 the(k)quasi-uniformity parameter ρ of Th ,
2 of
such that the smallest eigenvalue dmin k [D ]T D (k) satisfies

2
dmin ≥ c > 0. (41)
d (k) ]T W D (k)
Proof Consider the matrix A = k=1 [D and note that
   
 (ρh)d ∗  (k) T (k)
∗ ∗ (k) T (k)
u Au ≤ ωmax u [D ] D u≤ u [D ] D u.
d!
k k

Furthermore, u T Au = |u|2H 1 , and the Friedrichs inequality states that u L 2 ≤

L|u| H 1 . Furthermore, according to [32, Proposition 6.3.1], there is a constant K 
such that u T u ≤ K  h −d u2L 2 . Finally, we use that the quadrature weights {ωi }
are (h d ) to find that u T u ≤ C h −d |u|2H 1 = C h −d u T Au ≤ C h −d (ρh)
d
∗
  d! u
(k) T (k) u, as required.
k [D ] D 


123
386 S. Loisel

Lemma 10 Assume Th is quasi-uniform and that R ≥ 1 and 0 < h ≤ 1. There is a

 > 0, which depends on , such that
constant c


λmin ≥ c R −4 h 6d . (42)

Proof Using (32) and (40), we arrive at:

(   2   4,
R −p 2 R −p
λmin ≥ min 2 dmin , ωmin
2
.
ωmin ωmin

Note that R ≥ R ∗ ≥ 1 = ||−1 || ≥ ||−1 ωmin and hence R/ωmin ≥ ||−1 and

− 4p 2+ 4
λmin ≥ min{2c , 1} max{||, 1}2 R ωminp .

Since R ≥ 1 and ωmin ≤ 1 (because h ≤ 1), we can find a lower bound by putting
p = 1 in the exponents. Under the quasi-uniform hypothesis, all the quadrature weights
are bounded below by ωi ≥ h d /(d!), which yields (42). 

Lemma 11 For 0 < h < 1, assume Th is a quasi-uniform triangulation of . The
number n of vertices of Th inside  and the number m of simplices in Th satisfy
n
≤ m ≤ ||h −d d!, (43)
d +1

where || is the volume of .

Proof The inequality n ≤ (d + 1)m follows from the fact that each of the m simplices
has precisely d + 1 vertices; we may indeed have n < (d + 1)m since some vertices
may beshared between multiple simplices. The upper bound for m follows from
m
|| = i=1 ωi ≥ mωmin ≥ mh d /(d!). 

Lemma 12 Consider the point x̂ = (0, ŝ).
∗ −1.5d
x̂2 ≤ C h R, and F  (x̂)2 ≤ C
∗ −1−1.5d
h R(1 + g X p ), (44)

where C ∗ < ∞ is a constant that depends on  and the quasi-unifomity parameter

ρ of Th .
Proof From (31) we have


m  2
R
x̂22 = ŝi2 ≤ m . (45)
2ωmin
i=1

Using (43), we obtain (44).

We now estimate F  (x̂). Using (34), we find

F  (x̂)2 ≤ Fu 2 + Fs 2 (46)

123
Efficient algorithms for solving the p-Laplacian 387

 2 −1 2/ p−1 1
≤ [D ( j) ]T Ẑ −1 D ( j) g2 + ẑ min ŝ 2 + σ ŝ −1 2 + ω2 .
p τ̂min
j
(47)

We bound the first term as follows:

⎛ ⎞1/2 ⎛ ⎞1/2
  
[D ( j) ]T Ẑ −1 D ( j) g2 ≤ ⎝ ~[D ( j) ]T Ẑ −1 ~22 ⎠ ⎝ D ( j) g22 ⎠ (48)
j j j
≤1 ⎛ ⎞1/2
$%&' 
−1 ⎝
−1
≤ δ2 ẑ min ωmin D ( j) g22 ⎠ (49)
j

−1
Here, δ 2j = ~D ( j) ~22 ≤ ωmin ρ([D ( j) ]T W D ( j) ), where ρ(·) is the spectral radius. We
estimate the spectral radius as follows:

w T [D ( j) ]T W D ( j) w = wx2 j d x ≤ |w|2H 1 ≤ C I2S h −2 w2L 2 ≤ C I2S [K 
 2 2d−2
] h w22 ,


where we have used the inverse Sobolev inequality |w| H 1 ≤ C I S h −1 w L 2 for w ∈ Vh
(see e.g. Toselli and Widlund [37, Lemma B.27]) and the norm equivalence w L 2 ≤
K h d w2 . Thus,
2

−1/2 √ 
δ2 ≤ ωmin dC I S [K  ]h d−1 .

Furthermore, using equivalence of p−norms in m dimensions,

⎛ ⎞1/2 ⎛ ⎡⎛ ⎞1/2 ⎤2 ⎞1/2

 −1/2 ⎜
m
⎢ d
⎥ 
⎝ D ( j) g22 ⎠ ≤ ωmin ⎝ ωk ⎣⎝ (D ( j) g)2k ⎠ ⎦ ⎠ (50)
j k=1 j=1
⎛ ⎡⎛ ⎞1/2 ⎤ p ⎞1/ p
⎜ ⎢⎝ ( j) 2 ⎠ ⎥ 
m d
−1/2
≤ ωmin m 1/2 ⎝ ωk ⎣ (D g)k ⎦ ⎠ (51)
k=1 j=1
−1/2
= ωmin m 1/2 g X p . (52)

As a result,
   
−1/2 √ −1/2
[D ( j) ]T Ẑ −1 D ( j) g2 ≤ ωmin dC I S [K 

]h d−1 ωmin m 1/2 g X p (53)
j
 −1−0.5d
≤ C I S K h g X p . (54)

123
388 S. Loisel

From (31) and (33), we further estimate

⎧  2/ p−1
⎨√ 2/ p−1 √
m ŝmax ≤ m 2ωRmin if 1 ≤ p < 2
ŝ 2/ p−1
2 ≤ √ (55)
⎩ m if p ≥ 2
 
√ R
≤ m and (56)
2ωmin
√ √
ŝ −1 2 ≤ m and ω2 ≤ m(ρh)d /(d!). (57)

Hence,

F  (x̂)2 ≤ C I S K 
 −1−0.5d
h g X p
 
√ R √ 2√
+ m +2 m+ m(ρh)d /(d!).
2ωmin R




Proof of Theorem 1 Using (42), we find

 −1  
x̂∗x ∗ ≤ λ−1
min  x̂2 ≤ c

R −4 6d
h C ∗ −1.5d
 h R  −1 ∗ 5 −7.5d
= [c ] C R h . (58)
F

Also
 −1  
F  (x̂)∗x ∗ ≤ λ−1
min F 
( x̂)2 ≤ c

R −4 6d
h C ∗ −1−1.5d
 h R(1 + g X p ) (59)
F
 −1 ∗ −1−7.5d 5
= [c ] C h R (1 + g X p ). (60)

We substitute these estimates into (21) to get

√ 
N ∗ ≤ 7.2 4m 2 log(4m) (61)
 
 −1 ∗ −1−7.5d 5
+ log [c ] C h R (1 + g X p ) (62)
 
 −1 ∗ 5 −7.5d
+ log [c ] C R h + log(1/) (63)

Using m ≤ ||d!h −d , we get

 
≤ 14.4 ||d!h −d log h −1−17d R 5 (1 + g X p ) −1 + K ∗ . (64)

The estimates R p=1 , R1< p<∞ were obtained by starting from the estimates of
Lemma 3, adding 1, and doubling them. Substituting these into N ∗ produces N p=1
and N1< p<∞ . 


123
Efficient algorithms for solving the p-Laplacian 389

3.1 The case p = ∞

Recall the ∞-Laplacian of (7). As in the p = 1 case, J (u) is non-differentiable and

may be unbounded below when
 f is large. As per Lemma 3, assume that L f  L 1 < 1
2g ∞
and set R p=∞ = maxi ωi 2 + 1−LX f ()1 and impose ωi s ≤ R p=∞ . The problem
L
of minimizing J (u) over u ∈ Vh is equivalent to
5 6
min s over Q := (u, s) : s ≥ ∇(u + g)| K i 2 ∀i, and R ≥ ωi s (65)

We notice that this definition of Q coincides with the definition (28) with p = 1
subject to the additional restriction that s1 = . . . = sm and subsequently dropping the
index i from si . As a result, one can obtain a barrier for Q by taking the barrier (25)
with p = 1 on the subspace of constant valued s vectors, hence the barrier F∞ and its
derivatives are


F∞ (u, s) = F1 (u, se), F∞ (u, s) = E F1 (u, se), F∞

(u, s) = E F1 (u, se)E T , (66)
⎡ ⎤
1  
⎢ ⎥
where e = ⎣ ... ⎦ , E =
I O
. (67)
O eT
1

The starting point for the optimization is (û, ŝ) with û = 0 and

⎛ ⎞1
2
d
ŝ = 1 + max ⎝ ( j)
[(D g)i ] 2⎠
. (68)
i
j=1

Theorem 3 With the notation as in Theorem 1, let

 
2g X ∞ ()
R = R p=∞ = max ωi 2 + (69)
i 1 − L f  L 1

and assume p = ∞, L f  L 1 < 1. The barrier method to solve (65) requires at most
N p=∞ Newton iterations, where
   5  
2g ∞ ()
≤ 14.4 ||h −d d! log h −1−6.5d 2 +  −1 + K ∗ .
X
N p=∞
1 − L f  L 1
(70)

The computational complexity

√ as a function of the number n of grid points (and freezing
all other parameters) is O( n log n).

123
390 S. Loisel

Proof The proof of Theorem 3 follows the same logic as that of Theorem 1, so we
merely sketch it here. First, (34) and (35) are replaced by:

   ( j)

 Fu 2[D ( j) ]T y z
F = =  j  ωj , (71)
Fs −2s j z1j − mσ s + j τj
   
F Fus F F
F  = uu = uu T
us
where (72)
Fsu Fss Fus Fss

d 
d
Fuu = 2 [D ( j) ]T Z −1 D ( j) + 4 (Y ( j) D ( j) )T Z −2 (Y (r ) D (r ) ), (73)
j=1 j,r =1


d
Fus = −4 (Y ( j) D ( j) )T z −2 s, (74)
j=1
  
Fss = −2 z −1
j +4 z −2
j s + σ ms
2 −2
+ ω2j z −2
j . (75)
j j j

The proof of (40) holds (changing what must be changed), ending with

x T F  x ≥ x22 min{2z max
−1 2
dmin , ωk2 z k−2 }, (76)
k

which is slightly stronger than (40).

The estimate (44) also holds verbatim. The estimate for x̂2 is by inspection of
(68) and (69). The estimate for Fu (x̂)2 is identical to the proof of (44), and |Fs (x̂)|
is estimated as follows:
 
R −1 −1
|Fs | ≤ 2 z min m + σ m + ||τmin ≤ C Rh −2d , (77)
2ωmin

where we have used (16), 1 ≤ ŝ ≤ R/(2 min ωi ), τ̂i ≥ R/(2 min ωi ), and ẑ j ≥ 1, and
C is some constant that depends only on . Thus,

x̂∗x ∗ ≤ λ−1  −1 ∗ 5 −7.5d

min  x̂2 ≤ [c ] C  R h ,
F

see (58). Futhermore,

F  (x̂)∗x ∗ ≤ λ−1 
min F ( x̂)2 (78)
F
 −1  

≤ c R −4 h 6d  −1−0.5d
C I S K h g X ∞ + C Rh −2d (79)
 
 −1−3d g X ∞
≤ K h , (80)
1 +  f X 1

123
Efficient algorithms for solving the p-Laplacian 391

yielding the final estimate

√ 
N ∗ ≤ 7.2 4m 2 log(4m) (81)
  
 −1−3d g X ∞
+ log K  h (82)
1 +  f X 1
 
 −1 ∗ 5 −7.5d
+ log [c ] C R h (83)

+ log( −1 ) , (84)

as required. 


3.2 Implementation notes


In principle, the vector ( f i ) is defined by f i =  f φi ; we have not analyzed an inexact
scheme for computing these integrals. If f is assumed to be a suitable finite element
space (e.g. piecewise constant), then these integrals can be computed exactly from the
same quadrature we use on the diffusion term. In addition, we can then compute  f  L q
exactly by quadrature since | f |q is also piecewise constant. Assuming g is piecewise
p
linear, the quantities || and g X p can be computed exactly, see (30). Thus, one can
compute R1< p<∞ , R p=1 , etc... exactly.
In the strong form (1), the function g is given on ∂ (i.e. it is a trace), but in
the variational form (2), the function g has domain . Regarding v = u + g as the
solution, the choice of g| doe not affect the value of v, provided that g|∂ is fixed.
The simplest way to choose g| as a piecewise linear function on Th is to set all nodal
values to 0 inside of , but this is a somewhat “rough” prolongation that is furthermore
dependent on h. Using such a prolongation of g causes the estimates (5) and (6) to
become dependent on h where g appears. In order to avoid this dependence on h, one
can proceed in one of two ways. First, if the meshes Th are all included in one coarse
mesh Th 0 , then one can do the prolongation on Th 0 and use the same prolongation on
all Th .
Another method is to solve the linear Laplacian with boundary conditions g|∂
on the mesh Th . This choice of g| does vary slightly with h but it converges to the
continuous solution as h → 0. Furthermore, this choice of g minimizes g X 2 = |g| H 1
so it may result in a smaller value of R than prolongation by truncation. We call this
choice of g the discrete harmonic prolongation of g|∂ to . We use the discrete
harmonic prolongation in all our numerical experiments.

4 Numerical experiments

We consider the p-Laplacian for p = 1, 1.1, 1.2, 1.5, 2, 3, 4, 5, ∞ for a square

domain subject to Dirichlet boundary conditions and where the forcing f = 0, see
Fig. 1. For the boundary conditions g, we have taken the piecewise linear interpolant
of the trace (1 X (x, y))|∂ of the characteristic function 1 X (x, y), where X is the set

123
392 S. Loisel

Fig. 1 Solving the p-Laplacian for p = 1, 2, ∞ with the same boundary conditions g and zero forcing
f = 0 on a 200 × 200 grid. Because of the zero forcing, the minimum and maximum principles hold, which
provides some protection against the near-discontinuities in the boundary data, e.g. when p = ∞

X = ({0}×[0.25, 0.75])∪([0.6, 1]×[0.25, 1]), which we approximate on the discrete

grid by piecewise linear elements. Note that this creates very challenging numerical
and functional analytical problems, e.g. the trace of W 1,∞ functions are also W 1,∞ so
the ∞−Laplacian here is solving a problem approximating one outside the usual trace
space. The forcing f = 0 means that solutions must satisfy minimum and maximum
principles, and so the solution is always between the extremal 0 and 1 boundary values
for all values of p and all x ∈ . The zero forcing provides some “protection” against
the “bad” boundary data.
We have varied the number n of grid points from n = 16 (a 4 × 4 grid) up to
n = 40, 000 (200 × 200) and in all cases, solved to a tolerance of  ≈ 10−6 . We have
reported the number of Newton iterations required for convergence in Table 1. This
detailed table reveals those values of κ, n, p that failed to converge within five minutes.
Most of these convergence failures are due to purely numerical problems. Indeed, we
have noted in the introduction that when p is large, minimizing J (u) is intrinsically
challenging because it exhausts the accuracy of double precision floating point. Thus,
the difficulty in solving p-Laplacians accurately for large p is not particular to our
algorithm but indeeds affects all algorithms for solving p-Laplacians. MATLAB has
also issued warnings that the Hessian was singular to machine precision, for large
values of p and n.
The scaling properties of our algorithms are not immediately obvious from Table 1.
In order to visualize the scaling properties of our algorithms, we have sketched the
iterations counts of Table 1 in Fig. 2. Note that both axes are in logarithmic scale,
so straight lines of slope α correspond to O(n α ) scaling. We see that the short step
algorithm of Sect. 2.3 requires the largest number of Newton iterations to converge
(blue lines). This is consistent with experience in convex optimization. For this reason,
we were not able to solve larger problems with the short-step algorithm. The √ scaling
of the short-step algorithm is consistent with the theoretical prediction O( n log n)
of Theorems 1 and 3.
The long step algorithms (black lines) all require fewer Newton steps than the
short step algorithm, even though the theoretical estimate O(n log n) for long step
algorithms is worse than for short step algorithms. This is a well-known phenomenon,
and in practice, long step algorithms perform better, as is the case here.

123
 Table 1 Newton iteration counts for various problem sizes n, various step strategies κ and various values of p

κ n = 16 36 64 100 169 400 900 2500 5625 10,000 22,500 40,000

p = 1.0
Short 985 2165 3163 4392 6086 10, 512 17, 192 — — — — —
κ=2 128 141 164 171 190 211 243 289 357 392 496 533
κ=3 88 98 136 118 148 165 194 237 343 535 825 —
κ=4 60 68 114 96 124 138 160 340 496 720 877 —
κ=7 68 67 136 99 252 255 384 464 833 1113 — —
Adaptive 70 76 181 106 163 169 216 270 275 315 405 449
p = 1.1
Short 954 1844 2677 3674 5094 8803 14, 510 — — — — —
Efficient algorithms for solving the p-Laplacian

κ=2 128 141 156 166 170 195 225 247 272 279 312 327
κ=3 89 96 103 113 116 130 157 185 211 229 244 274
κ=4 61 77 84 97 102 115 145 164 182 195 227 256
κ=7 68 69 74 91 88 104 143 172 190 237 314 391
Adaptive 60 81 68 97 80 103 190 186 243 210 231 274
p = 1.2
Short 938 1779 2626 3566 4991 8589 14, 099 — — — — —
κ=2 128 141 156 162 169 193 218 244 259 265 291 295
κ=3 87 102 107 110 116 126 138 162 176 189 199 210
κ=4 60 67 74 86 89 98 114 124 139 152 168 176
κ=7 60 71 72 80 83 97 110 118 129 136 140 170
Adaptive 54 57 64 70 74 84 105 113 127 198 147 204

123
393
 Table 1 continued
394

κ n = 16 36 64 100 169 400 900 2500 5625 10,000 22,500 40,000

123
p = 1.5
Short 922 1731 2579 3484 4896 8405 13, 754 — — — — —
κ=2 129 146 154 163 174 193 220 249 266 273 287 294
κ=3 86 102 106 111 116 132 142 161 175 186 194 201
κ=4 58 71 73 84 88 102 116 126 131 148 156 162
κ=7 59 64 77 80 83 98 108 117 120 128 126 139
Adaptive 52 58 66 69 72 85 96 111 110 120 123 130
p = 2.0
Short 829 1543 2305 3102 4356 7440 12, 123 22,082 — — — —
κ=2 135 151 158 165 174 190 220 256 274 287 296 309
κ=3 93 100 108 113 120 129 146 169 184 192 201 209
κ=4 51 64 70 71 84 89 99 109 119 129 139 142
κ=7 62 74 80 81 98 113 125 132 145 148 158 160
Adaptive 59 63 69 73 79 86 99 107 113 118 126 134
p = 3.0
Short 857 1624 2451 3329 4701 8127 13, 372 — — — — —
κ=2 142 161 172 184 195 208 249 299 315 331 348 363
κ=3 99 113 122 129 136 144 164 193 212 223 239 249
κ=4 56 65 74 79 87 102 110 128 136 157 163 163
κ=7 67 75 84 90 102 116 127 138 154 156 165 175
Adaptive 84 70 76 112 88 94 109 118 129 135 145 147
S. Loisel
 Table 1 continued

κ n = 16 36 64 100 169 400 900 2500 5625 10,000 22,500 40,000

p = 4.0
Short 907 1765 2697 3696 5260 9197 15, 265 — — — — —
κ=2 153 179 192 205 217 236 273 346 373 413 433 —
κ=3 105 122 132 144 153 167 189 224 243 289 — —
κ=4 57 75 82 84 97 115 127 142 158 219 186 212
κ=7 69 84 94 99 108 128 141 159 174 178 193 —
Adaptive 66 77 80 91 100 110 125 137 148 157 512 —
p = 5.0
Short 970 1927 2974 4102 5872 10, 352 17, 280 — — — — —
Efficient algorithms for solving the p-Laplacian

κ=2 162 194 213 226 245 268 303 414 — — — —

κ=3 112 132 145 159 169 188 214 — — — — —
κ=4 63 78 90 95 111 132 143 176 — — — —
κ=7 74 91 102 109 119 146 164 — — — — —
Adaptive 67 82 91 100 109 123 136 430 — — — —
p=∞
Short 560 841 1191 1560 2145 3614 5884 10,809 — — — —
κ=2 96 103 111 120 130 137 160 771 309 489 — —
κ=3 69 73 78 86 91 95 113 451 536 — — —
κ=4 55 53 52 64 58 97 78 2606 3489 — — —
κ=7 54 48 57 62 63 75 1272 6989 — — — —
Adaptive 50 62 66 62 73 83 120 150 276 191 302 341
The symbol — indicates that the algorithm failed to converge by exceeding the 5 minutes time limit

123
395
396 S. Loisel

Fig. 2 The number of Newton iterations for various grid sizes n and parameters p and step sizes κ

In Fig. 2, most of the black curves are approximately straight lines, indicating
O(n α ) scaling, but there are notable exceptions when p = 1 or p = ∞, especially
when κ is also large. By contrast, the adaptive step size algorithm (red lines), with
κ0 = 10, is seen to be the best algorithm in most cases, and these red lines are much
straighter than the black lines. We denote by N p (n) the number of iterations required
for a certain value of p and problem size n for the adaptive step size algorithm. We
have fitted straight lines to the red curves of Fig. 2 in the least-squares sense and
obtained the following approximations:

p = 1.0 1.1 1.2 1.5 2.0 3.0 4.0 5.0 ∞

N p (n) ≈ 62n 0.18 33n 0.21 31n 0.17 43n 0.11 47n 0.10 60n 0.09 30n 0.22 17n 0.36 18n 0.28

1
Thus, it seems like the adaptive scheme requires about O(n 4 ) Newton iterations,
regardless of the value of p.
Note that the case p = 2 is a linear Laplacian that can be computed by solving
a single linear problem. When we embed this linear problem into the machinery of
convex optimization, the overall algorithm is very inefficient since it may require
hundreds of linear solves. We are including this test case for completeness, not as a
recommendation.

123
Efficient algorithms for solving the p-Laplacian 397

Fig. 3 Solving the 1-Laplacian (top row) and ∞-Laplacian (bottom row) in 3d. The left column shows
the solutions on the whole volumetric domain  with transparency, while the right column shows a slice
through  of the same solutions with opaque colors

4.1 3d experiments

Consider the following function:

7
 
9    3 e−x 2
φ= − x2 + y2 1/10 + (|x − cos (y)|) + z +
2
. (85)
20 25

We define the “spaceship domain”  ˜ = {(x, y, z) ∈ R3 : φ > 0}; this domain

˜ is
is slightly rescaled so that it is aesthetically pleasing. In practice, the domain 
approximated on a discrete grid with a tetrahedral mesh Th and the corresponding
polyhedral approximation  of . ˜ On this tetrahedral mesh, we solve the p−Laplacian
with forcing f = 1 with p ∈ {1, ∞}. The boundary values g are the indicating
function of the set {y > 0.45}, as approximated by a piecewise linear function on the
finite element grid. This problem features n = 11, 224 unknowns and m = 47, 956
elements. The solutions are displayed in Fig. 3.
For these problems, the solution of the 1-Laplacian seems to approximate the indi-
cating function of {y > 0.45}, as expected. However, the solution of the ∞-Laplacian
is very large (exceeding 2, 000 somewhere in the middle of the spaceship). This is
because the traces of W 1,∞ () functions are in W 1,∞ (∂) but our boundary data g
is a piecewise linear approximation of a discontinuous trace with jumps (an indicating

123
398 S. Loisel

function), an hence g X ∞ is very large and so is the solution u + g. The 1-Laplacian
is better able to tolerate the boundary data g with (near)-jumps because the trace of a
W 1,1 () function is merely L 1 (∂), thus allowing jumps.
The solution for the p = 1-Laplacian seems very close to what one would obtain
if one were to put f = 0 instead of f = 1. This is not surprising, because the 1-
Laplacian is a linear program and the solutions of linear programs change in discrete
steps when the forcing changes continuously. For example, the unique minimizer of
J˜(x) = |x| + f x (x ∈ R) is x = 0 whenever | f | < 1, and switches to “undefined”
(or ±∞) when | f | > 1 because then J˜ is unbounded below.
For the p = ∞-Laplacian, the solution u + g is a large positive bump because
f > 0 and there is a minimum principle stating that the minimum must of u + g be
on the boundary ∂. When one takes f < 0 instead, the solution u + g is a large
negative bump because in that scenario, u + g satisfies a maximum principle. In the
2d experiments, the ∞-Laplacian did not develop large bumps because the boundary
data was between 0 and 1 and the forcing was 0. This meant that u + g had to satisfy
both minimum and maximum principles, and u was constrained by 0 ≤ u + g ≤ 1,
preventing the formation of large bumps in the solution.

5 Conclusions and outlook

We have presented new algorithms for solving the p-Laplacian efficiently for any given
tolerance and for all 1 ≤ p ≤ ∞. We have proven that our √algorithms compute a solu-
tion to any given tolerance in polynomial time, using √ O( n log n) Newton iterations,
and an adaptive stepping variant converges in O( n log2 n) Newton iterations. We
have confirmed these scalings with numerical experiments. We have further shown by
numerical experiments that the adaptive step variant of the barrier method converges
much faster than the short-step variant for the p-Laplacian and also usually faster than
long-step barrier methods, thus achieving the practical speedup of long-step algo-
rithms while avoiding the O(n log n) worst-case behavior of long-step algorithms. We
1
have numerically estimated that the adaptive step algorithm requires O(n 4 ) Newton
iterations across all values of 1 ≤ p ≤ ∞. We have observed numerical difficulties
for p ≥ 5, which are expected since large powers exhaust the accuracy of double
precision floating point arithmetic; this difficulty is not specific to our algorithm but
is inherent to the p-Laplacian for large values of p. Our algorithms are particularly
attractive when p ≈ 1 and p = ∞, where there are no other algorithms that are
efficient at all tolerances.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which
permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give
appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence,
and indicate if changes were made. The images or other third party material in this article are included
in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If
material is not included in the article’s Creative Commons licence and your intended use is not permitted
by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

123
Efficient algorithms for solving the p-Laplacian 399

References
1. Andersen, K.D., Christiansen, E., Conn, A.R., Overton, M.L.: An efficient primal-dual interior-point
method for minimizing a sum of Euclidean norms. SIAM J. Sci. Comput. 22, 243–262 (1999)
2. Andreu, F., Mazón, J.M., Rossi, J.D., Toledo, J.: A nonlocal p-Laplacian evolution equation with
nonhomogeneous Dirichlet boundary conditions. SIAM J. Math. Anal. 40(5), 1815–1851 (2009)
3. Barrett, J.W., Liu, W.B.: Finite element approximation of the p-Laplacian. Math. Comput. 61(204),
523–537 (1993)
4. Berninger, H., Loisel, S., Sander, O.: The 2-Lagrange multiplier method applied to nonlinear transmis-
sion problems for the Richards equation in heterogeneous soil with cross points. SIAM J. Sci. Comput.
36(5), A2166–A2198 (2014)
5. Boyd, S., Vandenberghe, L.: Convex Optimization. Cambridge University Press, New York (2004)
6. Calamai, P.H., Conn, A.R.: A projected Newton method for l p norm location problems. Math. Program.
38(1), 75–109 (1987). https://doi.org/10.1007/BF02591853
7. Candès, E.J., Romberg, J., Tao, T.: Robust uncertainty principles: exact signal reconstruction from
highly incomplete frequency information. IEEE Trans. Inf. Theory 52(2), 489–509 (2006)
8. Catté, F., Lions, P.-L., Morel, J.-M., Coll, T.: Image selective smoothing and edge detection by nonlinear
diffusion. SIAM J. Numer. Anal. 29(1), 182–193 (1992)
9. Chen, R.-L., Zeng, J.-P.: A new subspace correction method for nonlinear unconstrained convex opti-
mization problems. Acta Mathematicae Applicatae Sinica 28(4), 745–756 (2012). https://doi.org/10.
1007/s10255-012-0185-z. (English Series)
10. Côté, J., Gander, M.J., Laayouni, L., Loisel, S.: Comparison of the Dirichlet-Neumann and optimal
Schwarz method on the sphere. In: Domain Decomposition Methods in Science and Engineering, pp.
235–242. Springer, New York (2005)
11. Drábek, P.: The p-Laplacian-mascot of nonlinear analysis. Acta Mathematica Universitatis Comeni-
anae 76(1), 85–98 (2007)
12. Drury, S.W., Loisel, S.: Sharp condition number estimates for the symmetric 2-Lagrange multiplier
method. In: Domain Decomposition Methods in Science and Engineering, pp. 255–261. Springer, New
York(2013)
13. Ekeland, I., Témam, R.: Convex Analysis and Variational Problems. SIAM, Philadelphia (1999)
14. Gander, M.J., Loisel, S., Szyld, D.B.: An optimal block iterative method and preconditioner for banded
matrices with applications to pdes on irregular domains. SIAM J. Matrix Anal. Appl. 33(2), 653–680
(2012)
15. Greer, N., Loisel, S.: The optimised Schwarz method and the two-Lagrange multiplier method for
heterogeneous problems in general domains with two general subdomains. Numer. Algorithms 69(4),
737–762 (2015)
16. Heinonen, J., Kilperläinen, T., Martio, O.: Nonlinear Potential Theory of Degenerate Elliptic Equations.
Dover Publications, Mineola (2006)
17. Hertog, D., Jarre, F., Roos, C., Terlaky, T.: A sufficient condition for self-concordance, with application
to some classes of structured convex programming problems. Math. Program. 69(1–3), 75–88 (1995).
https://doi.org/10.1007/BF01585553
18. Huang, Y.Q., Li, R., Liu, W.: Preconditioned descent algorithms for p-Laplacian. J. Sci. Comput. 32(2),
343–371 (2007). https://doi.org/10.1007/s10915-007-9134-z
19. Kohn, R., Serfaty, S.: A deterministic-control-based approach motion by curvature. Commun. Pure
Appl. Math. 59(3), 344–407 (2006). https://doi.org/10.1002/cpa.20101
20. Le Gall, F.: Powers of tensors and fast matrix multiplication. In: Proceedings of the 39th International
Symposium on Symbolic and Algebraic Computation, pp. 296–303. ACM, New York (2014)
21. Lindqvist, P.: Notes on the p-Laplace Equation. University of Jyväskylä, Finland (2017)
22. Loisel, S., Nguyen, H.: An optimal Schwarz preconditioner for a class of parallel adaptive finite
elements. J. Comput. Appl. Math. 321, 90–107 (2017)
23. Loisel, S., Szyld, D.B.: On the convergence of optimized Schwarz methods by way of matrix analysis.
In: Domain Decomposition Methods in Science and Engineering, vol. XVIII, pp. 363–370. Springer,
New York (2009a)
24. Loisel, S., Szyld, D.B.: A maximum principle for l 2 -trace norms with an application to optimized
Schwarz methods. In: Domain Decomposition Methods in Science and Engineering, vol. XVIII, pp.
193–200. Springer, New York (2009b)

123
400 S. Loisel

25. Loisel, S., Côté, J., Gander, M.J., Laayouni, L., Qaddouri, A.: Optimized domain decomposition
methods for the spherical Laplacian. SIAM J. Numer. Anal. 48(2), 524–551 (2010)
26. Loisel, S., Nguyen, H., Scheichl, R.: Optimized Schwarz and 2-Lagrange multiplier methods for
multiscale elliptic PDEs. SIAM J. Sci. Comput. 37(6), A2896–A2923 (2015)
27. Nesterov, Y.: Introductory Lectures on Convex Optimization: A Basic Course, vol. 87. Springer, New
York (2013)
28. Nesterov, Y., Nemirovskii, A.: Interior-point polynomial algorithms in convex programming. Soc. Ind.
Appl. Math. (1994). https://doi.org/10.1137/1.9781611970791
29. Nocedal, J., Wächter, A., Waltz, R.A.: Adaptive barrier update strategies for nonlinear interior methods.
SIAM J. Optim. 19(4), 1674–1693 (2009)
30. Oberman, A.M.: Finite difference methods for the infinity Laplace and p-Laplace equations. J. Comput.
Appl. Math. 254, 65–80 (2013)
31. Pinkall, U., Polthier, K.: Computing discrete minimal surfaces and their conjugates. Exp. Math. 2(1),
15–36 (1993). https://doi.org/10.1080/10586458.1993.10504266
32. Quarteroni, A., Valli, A.: Numerical Approximation of Partial Differential Equations, vol. 23. Springer,
New York (2008)
33. Rektorys, K.: Variational Methods in Mathematics, Science and Engineering. Springer, New York
(2012)
34. Ridgway Scott, L., Zhang, S.: Finite element interpolation of nonsmooth functions satisfying boundary
conditions. Math. Comput. 54(190), 483–493 (1990)
35. Subber, W., Loisel, S.: Schwarz preconditioners for stochastic elliptic PDES. Comput. Methods Appl.
Mech. Eng. 272, 34–57 (2014)
36. Tai, X.-C., Jinchao, X.: Global and uniform convergence of subspace correction methods for some
convex optimization problems. Math. Comput. 71, 105–124 (2001)
37. Toselli, A., Widlund, O.: Domain Decomposition Methods-Algorithms and Theory, vol. 34. Springer,
New York (2006)
38. Xue, G., Ye, Y.: An efficient algorithm for minimizing a sum of p-norms. SIAM J. Optim. 10, 551–579
(1997)

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps
and institutional affiliations.

123