EW Navier-Stokes Equations
EW Navier-Stokes Equations
Chapter 1. Introduction 4
1.1. Physical derivation (sketch) 4
1.2. Elementary mathematical properties 5
1.3. Related Models 7
Bibliography 44
3
CHAPTER 1
Introduction
The Navier-Stokes equations are
u∶ Ω × [0, T ] → Rd ;
the unknown pressure eld
p∶ Ω × [0, T ] → R;
and the given constant viscosity ν > 0. It can be written in components, for i = 1, . . . , d:
d d
∂t ui + ∑ uj ∂j ui + ∂i p = ν ∑ ∂x2j ui
j=1 j=1
d
∑ ∂j uj = 0.
j=1
The Navier-Stokes Equations (NSE) describe the time evolution of the velocity and
pressure of a viscous incompressible uid (e.g. water) without external forces.
where n(x) is the outer unit normal of the surface ∂ Ω̃ at the point x. For regular enough
boundary ∂ Ω̃ and u by the Gauss-Green theorem, the surface integral is equal to
ˆ
div u(x, t) dx,
Ω̃
and since this should equal zero for every Ω̃, we conclude div u = 0 everywhere in Ω.
Conservation of momentum/Newton's 2nd law: Consider a uid particle initially lo-
cated at x ∈ Ω and denote its location at time t by X(x, t) (Lagrangian description).
Newton's 2nd law for this particle (point) reads F = ma, and by assuming constant
density ( m = 1) we obtain
4
1.2. ELEMENTARY MATHEMATICAL PROPERTIES 5
because the particle moves according to the ow of u. Therefore, by the chain rule,
1.2.1. Energy balance. If u is smooth we can multiply the (NSE) by u and integrate
over (space) Ω to obtain
ˆ ˆ ˆ ˆ
∂t u ⋅ u dx + (u ⋅ ∇)u ⋅ u dx + ∇p ⋅ u dx = ν ∆u ⋅ u dx. (1.1)
Ω Ω Ω Ω
The rst term of (1.1) becomes
ˆ
1d 1d
∣u∣2 dx = ∥u(⋅, t)∥2L2 (Ω) .
2 dt Ω 2 dt
Further, we note, by integrating by parts, that
ˆ ˆ ˆ ˆ
(u ⋅ ∇)u ⋅ u dx = ∑ ui uj ∂j ui dx = − ∑ ∂j ui uj ui dx − ∑ ui ∂j uj ui dx.
Ω Ω i,j Ω i,j Ω i,j
For the term involving the pressure in (1.1) we can also integrate by parts to obtain
ˆ ˆ
∇p ⋅ u dx = − p div u dx = 0
Ω Ω
again by incompressibility. For the last term on the RHS of (1.1) we integrate by parts
and see that
ˆ ˆ
ν ∆u ⋅ u dx = −ν ∣∇u∣2 dx.
Ω Ω
This suggests that u ∈ L2t Hx1 (Ω) ∩ t Lx (Ω) is a suitable function space for NSE (the
L∞ 2
∂t u + div(u ⊗ u) + ∇p =ν∆u,
div u =0.
Take the divergence of the NSE and we obtain
∂t u + div(u ⊗ u) + ∇p = 0
div u = 0.
Here without the parabolic term from the Laplacian, the mathematical theory is very
dierent.
1.3.2. Compressible uids. We can study compressible uids (like air) where we
have an extra non-negative scalar eld ρ modelling the density:
We only work with real-valued functions u, which implies that ûk = û−k , for k ∈ Z3 .
By Plancherel's Theorem, for u ∈ L (T3 ) we have that
2
ˆ
∣u(x)∣2 dx = (2π)3 ∑ ∣ûk ∣2 ,
T3 k∈Z3
∑ ∣ûk ∣ < ∞.
2
k∈Z3
⎡ ⎤
⎢ ⎥
∥u∥2H s (T3 ) ∶= ∥u∥2L2 (T3 ) + ∑ ∥∂ α u∥2L2 (T3 ) = (2π)3 ⎢⎢ ∑ ∣ûk ∣2 + ∑ ∑ ∣∂̂ α u ∣2 ⎥ ,
k ⎥
⎢k∈Z3 ⎥
∣α∣≤s ⎣ ∣α∣≤s k∈Z3 ⎦
where we used Plancherel's Theorem in the last equality. Note that the derivatives are
taken in the weak (distributional) sense.
Further, we can integrate by parts to see that
ˆ ˆ
̂ 1 1
j uk = ∂j u(x) dx = ikj e−ik⋅x u(x) dx = ikj ûk
−ik⋅x
∂ e
(2π)3 T3 (2π)3 T3
and hence
∣∂̂
α u ∣2 = ∣k∣2∣α∣ ∣û ∣2
k k
and thus
8
2.2. HELMHOLTZ DECOMPOSITION 9
When s ∈ N, this denition coincides with the denition by weak derivatives. It will be
useful to consider homogeneous Sobolev spaces, where the zero-th Fourier mode is zero.
Definition 2.2. (1) The homogeneous space L̇2 (T3 ) consists of all u ∈ L2 (T3 )
such that
ˆ
u(x)dx = 0 (i.e. û0 = 0).
T3
(2) The homogeneous Sobolev space Ḣ s (T3 ) is dened as H s (T3 ) ∩ L̇2 (T3 ), with the
norm
with
{u ∈ L̇ (T ; R )∶
2 3 3
k ⋅ uk = 0 ∀k ≠ 0} .
Note that H may contain vector elds that are not in H 1 and hence div is not well
2
dened simply by taking derivatives. H is equipped with the L norm.
10 2. FUNCTION SPACES AND WEAK SOLUTIONS
Lemma 2.4. Every u ∈ H(T3 ) is weakly divergence free in the sense that
ˆ
u(x) ⋅ ∇φ(x) dx = 0
T3
for all φ ∈ Ḣ 1 (T3 ).
Proof. We can write u as a Fourier series as u(x) = ∑j≠0 ûj eij⋅x , then using orthogo-
´
nality in the form e dx = δjk , we obtain
ij⋅x −ik⋅x
e
ˆ ˆ ˆ
u(x) ⋅ ∇(e−ik⋅x
) dx = ∑ ûj e ⋅ (−ik)e
ij⋅x −ik⋅x
dx = −ik ⋅ ûk dx = 0
T3 T3 j≠0 T3
since u ∈ H(T3 ). Further, since {e−ik⋅x }k∈Z3 ∖{0} form an orthonormal basis (ONB) of
G = {g ∈ L̇ (T ; R )∶
2 3 3
g = ∇φ for some φ ∈ Ḣ 1 (T3 )} .
Hence Lemma 2.4 says that G and H are orthogonal subspaces of L̇2 .
Theorem 2.6 (Helmholtz decomposition). L̇2 = G⊕H , i.e. for all u ∈ L̇2 (T3 ; R3 ) there
exist unique g ∈ G and h ∈ H such that
ˆ
u=g+h and g ⋅ h dx = 0.
T3
Moreover, if u ∈ Ḣ s (T3 ) then g = ∇φ for φ ∈ Ḣ s+1 (T3 ) and h ∈ Ḣ s (T3 ).
Proof. We can write u as a Fourier series and see that (ûk )k∈Z3 ∈ l2 . We can then
write each ûk as a linear combination of k and a vector wk perpendicular to k . Thus for
all k ≠ 0, let ûk = αk k + wk with k ⋅ wk = 0 and αk ∈ C. Note that by orthogonality
∣ûk ∣2 = ∣αk ∣2 ∣k∣2 + ∣wk ∣2 (2.1)
and hence
u(x) = ∑ ûk eik⋅x = ∑ (αk k +wk )eik⋅x = ∑ −iαk ∇eik⋅x + ∑ wk eik⋅x = ∇φ(x)+h(x),
k≠0 k≠0 k≠0 k≠0
where
hence φ ∈ Ḣ 1
and h ∈ L̇ 2
and thus g ∈ G and h ∈ H .
Further, suppose that u ∈ Ḣ s (T3 ) and multiply (2.1) by ∣k∣2s to conclude that φ ∈ Ḣ s+1
and h ∈ Ḣ s .
Finally, we must show uniqueness. Suppose that u = h1 + ∇φ1 = h2 + ∇φ2 for h1 , h2 ∈ H
and ∇φ1 , ∇φ2 ∈ G, then
Definition 2.7. The orthogonal projection from L̇2 (T3 ) onto H is called the Leray
projection : If u=h+g with h∈H and g ∈ G, then Pu = h.
In Fourier series (exercise!)
ûk ⋅ k
Pu(x) = ∑ (ûk − k) eik⋅x .
k≠0 ∣k∣2
Note that the following useful lemma only holds true when the spatial domain has no
physical boundaries.
̂
∂ j uk ⋅ k ûk ⋅ k ûk ⋅ k
̂
P∂ ̂
j uk = ∂j uk − k = ikj ûk − ikj k = ikj (ûk − k) = ∂̂
j Puk .
∣k∣2 ∣k∣ 2 ∣k∣2
That is, V consists of weakly divergence-free vector elds with extra regularity H 1.
Definition 2.10 (Stokes operator). The Stokes operator is dened as −P∆, in the
domain V ∩ H 2 (T3 ; R3 ).
We notice that from Lemma 2.8 that if u ∈ V ∩ H 2, then
● mk ⊥ k , m−k ⊥ k , mk ⊥ m−k ,
● ∥mk cos(k ⋅ x)∥L2 (T3 ) = ∥mk sin(k ⋅ x)∥L2 (T3 ) = 1,
● m(−k) = m−k .
Then {mk cos(k ⋅ x)} ∪ {mk sin(k ⋅ x)} ⊂ H : Indeed,
ˆ ˆ
mk cos(k ⋅ x) dx = 0, mk sin(k ⋅ x) dx = 0,
T3 T3
and k ⋅ mk cos(k ⋅ x), k ⋅ mk sin(k ⋅ x) = 0 by choice of mk . Hence, we have an orthonormal
family in H whose members are also in the domain of the Stokes operator (because they
are smooth). Moreover,
−∆mk cos(k ⋅ x) = −mk ∑ ∂j2 cos(k ⋅ x) = +mk ∣k∣2 cos(k ⋅ x) = ∣k∣2 mk cos(k ⋅ x),
12 2. FUNCTION SPACES AND WEAK SOLUTIONS
so we can dene ∣k∣2 ∶= λk , and similar for sin(k ⋅ x), whence (2) is proved after re-labelling
indices from Z ∖ {0} to N. To show that these functions form a Hilbert basis of H , let
3
1 1
u(x) = ∑ ûk eik⋅x = ∑ ûk e + ∑ û−k e
ik⋅x −ik⋅x
k≠0 2 k≠0 2 k≠0
1 1
= ∑ (ûk + û−k ) cos(k ⋅ x) + ∑ i(ûk − û−k ) sin(k ⋅ x),
2 k≠0 2 k≠0
and we see that both (ûk + û−k ) and i(ûk − û−k ) are in R3 and perpendicular to k. This
becomes, for some ak , bk , ck , dk , αk , βk ∈ R,
∑ (ak mk + bk m−k ) cos(k ⋅ x) + ∑ (ck mk + dk m−k ) sin(k ⋅ x)
k≠0 k≠0
= ∑ (ak mk + b−k mk ) cos(k ⋅ x) + ∑ (ck mk − d−k mk ) sin(k ⋅ x)
k≠0 k≠0
= ∑ αk mk cos(k ⋅ x) + ∑ βk mk sin(k ⋅ x).
k≠0 k≠0
It will be convenient to check that this denition can actually be tested on a smaller
class of test functions than Dσ . Therefore, set
N
D̃σ ∶= {φ = ∑ dk (t)wk (x)∶ dk ∈ Cc∞ ([0, ∞))} ,
k=1
where {wk } is the eigenbasis of the Stokes operator from Theorem 2.11. Clearly we have
that D̃σ ⊂ Dσ .
Lemma 2.14. If u ∈ L∞ (0, ∞; H) ∩ L2 (0, ∞; V ) satises (2.2) for all φ ∈ D̃σ , then it
even satises (2.2) for all φ ∈ Dσ , i.e. it is a weak solution.
Proof. Let φ ∈ Dσ , then for every t ≥ 0, φ(t) ∈ H , and we can write
∞
φ(x, t) = ∑ dk (t)wk (x)
k=1
since {wk } form a Hilbert basis of H. Set
N
φN ∶= ∑ dk (t)wk (x) ∈ D̃σ .
k=1
Then φN → φ in C([0, ∞); V ). Indeed,
∞ ∞
sup ∥φ(t) − φN (t)∥2V = sup ∥ ∑ dk (t)wk (⋅)∥2V = sup ∑ λk d2k (t),
t t k=N +1 t k=N +1
√
as {wk } are orthogonal in V and ∥wk ∥V = λk (from remark after Theorem 2.11). This
then becomes, as λk increases to ∞,
∞ λ2k d2k (t) 1 ∞
sup ∑ ≤ sup ∑ λ2k d2k (t)
t k=N +1 λk λN t k=N +1
1 ∞
= sup ∑ (−∆wk (x)dk (t), −∆wk (x)dk (t))L2
λN t k=N +1
by orthogonality. We then see that we can bound this above by
1 1
sup ∥ − ∆φ∥L2 (T3 ) ≤ sup ∥φ(t)∥H 2 (T3 ) → 0
λN t λN t
as N → ∞, since supt ∥φ(t)∥H 2 (T3 ) is independent of N .
Furthermore, ∂t φN → ∂t φ in L (0, T ; L (T )) because
2 2 3
ˆ ∞ ˆ ∞ ∞
∥∂t φ − ∂t φN ∥L2 (T3 ) dt =
2
∥ ∑ d′k (t)wk (x)∥2L2 (T3 ) dt
0 0 k=N +1
ˆ ∞ ∞
= ∑ (dk (t)) dt → 0
′ 2
0 k=N +1
as N → ∞, where the latter convergence follows from
and hence
∞
sup ∑ (d′k (t))2 = sup ∥∂t φ∥2L2 (T3 ) < ∞.
t k=1 t
It follows that
ˆ ∞ˆ ˆ ∞ˆ
∂t φN ⋅ u dx dt →
∂t φ ⋅ u dx dt,
0 0
ˆ ∞ˆ ˆ ∞ˆ
∇φN ∶ ∇u dx dt → ∇φ ∶ ∇u dx dt,
ˆ0 ∞ ˆ 0
ˆ ∞ˆ
φN (0) ⋅ u0 (x)∇u dx dt → φ(0) ⋅ u0 (x) dx dt.
0 ´´ 0
For the remaining term (u⋅∇u)⋅φN dx dt, we will use the Sobolev embedding H 1 (T3 ) ⊂
L6 (T3 ), so it follows that
sup ∥φN − φ∥L6 (T3 ) ≤ C sup ∥φN − φ∥V → 0
t t
as N → ∞. Thus, by Hölders's inequality,
ˆ ˆ ˆ ˆ
∣ (u ⋅ ∇u) ⋅ (φN − φ) dx dt∣ ≤ ∣u∣∣∇u∣∣φN − φ∣ dx dt
≤ ∥u∥L2t L3x ∥∇u∥L2t L2x ∥φN − φ∥L∞ 6 → 0
t Lx
asN → ∞. (We note that using again the same embedding theorem we have ∥u∥L2t L3x ≤
C∥u∥L2t L6x ≤ C∥u∥L2t Hx1 ∥u∥L2t L3x < ∞.) So if we consider the equation for a weak
and so
solution to the NSE (2.2) with φN used as a test function, then we see that every term
will converge to the corresponding one with φ ∈ Dσ and so (2.2) follows for φ ∈ Dσ .
For later reference, we prove another lemma which allows us to test a weak solution
with functions of the form χ[t1 ,t2 ] φ for φ ∈ Dσ , for almost every t1 < t2 , where χ denotes the
indicator function of a set. This is a consequence of the Lebesgue dierentiation theorem,
which we recall without proof:
A point x for which the statement of the dierentiation theorem is true is called a
Lebesgue point of f; the theorem thus says that, given a locally integrable function on a
domain, almost every point in that domain is a Lebesgue point.
Proof. We prove only the case t1 = 0 in detail. So let t2 > 0 and consider a smooth
(cut-o ) function ζ∶R→R such that
● ζ ≥ 0,
2.4. WEAK SOLUTIONS 15
t − t2
ζ (t) ∶= ζ ( ).
Thus, ζ (restricted to t ≥ 0) is a smooth approximation of the indicator function χ[0;t2 ] .
If φ ∈ Dσ , then the product ζφ is still in Dσ , so using it as a test function in the weak
formulation of NSE gives
ˆ ∞ˆ ˆ ∞ˆ
− u ⋅ ∂t (ζ φ) dx dt + (u ⋅ ∇)u ⋅ (ζ φ) dx dt
T3 T3
0 0
ˆ ∞ˆ ˆ
+ν ∇u ∶ ∇(ζ φ) dx dt = u0 ⋅ φ(0) dx
0 T3 T3
(for the term involving u0 , note that ζ (0) = 1
). The two integralsfor suciently small
including space derivatives are easily seen to converge as ζ converges almost → 0: Indeed,
everywhere to χ[0,t2 ] , and the integrand (u⋅∇)u⋅(ζ φ) is bounded pointwise by ∣(u⋅∇)u∣∣φ∣,
uniformly in , which is of course integrable. Hence, by the dominated convergence theorem,
ˆ ∞ˆ ˆ ∞ˆ ˆ t2 ˆ
(u⋅∇)u⋅(ζ φ) dx dt → (u⋅∇)u⋅(χ[0,t2 ] φ) dx dt = (u⋅∇)u⋅φ dx dt
0 T3 0 T3 0 T3
as → 0, and likewise
ˆ ∞ˆ ˆ t2 ˆ
∇u ∶ ∇(ζ φ) dx dt → ∇u ∶ ∇φ dx dt
0 T3 0 T3
(of course the space derivative does not hit ζ , which depends only on time).
The rst integral, which contains the time derivative, is a bit more delicate. We
compute
ˆ ∞ˆ ˆ ∞ˆ ˆ ∞ˆ
u ⋅ ∂t (ζ φ) dx dt = u ⋅ ζ′ φ dx dt + u ⋅ ζ ∂t φ dx dt,
0 T3 0 T3 0 T3
and the latter integral is seen, as before, to converge to
ˆ t2 ˆ
u ⋅ ∂t φ dx dt.
0 T3
For the integral involving ζ′ , observe that by denition,
1 t − t2
ζ′ (t) = ζ ′ ( ),
´∞ ′
which also implies
0 ζ (t) dt = −1 for every > 0. Note also that ζ′ is supported in B (t2 ).
Therefore,
ˆ ∞ˆ ˆ
∣ u ⋅ ζ′ φ dx dt + u(t2 ) ⋅ φ(t2 ) dx∣
0 T3 T3
ˆ t2 + ˆ
≤ ∣ζ (t)∣ ∣
′
u(t) ⋅ φ(t) − u(t2 ) ⋅ φ(t2 ) dx∣ dt
t2 − T3
ˆ t2 + ˆ
1
≤ ∥ζ∥C 1 ∣ u(t) ⋅ φ(t) − u(t2 ) ⋅ φ(t2 ) dx∣ dt → 0
t2 − T3
as → 0, provided t2 is a Lebesgue point of the map
ˆ
t↦ u(t) ⋅ φ(t) dx.
T3
Since, by Lebesgue's theorem, this is the case for almost every t2 > 0, by collecting all
terms we nally arrive at (2.3) in the case t1 = 0.
16 2. FUNCTION SPACES AND WEAK SOLUTIONS
In the general case, we would use the test function ζ (t)ξ (t)φ(x, t), where ζ is as
before and
t1 − t
ξ (t) ∶= ζ ( ).
The passage to the limit → 0 can then be achieved in exactly the same way as above.
CHAPTER 3
∂t u = ∆u on T3 ,
(3.1)
u(t = 0) = u0 .
Let {wk }k∈N be an eigenbasis of −∆ and project the problem to the nite dimensional
subspace PN H ∶= span{w1 , . . . , wN }: If u(t) is in this space for every t, then so is ∂t u(t),
and thanks to the eigenfunction property also −∆u(t) ∈ PN H . The projected version of
the heat equation therefore simply reads
∂t uN = ∆uN on T3 ,
(3.2)
uN (t = 0) = PN u0 .
This equation is known as the Galerkin equation of order N , and we want to solve it in
PN H . To this end, take the ansatz uN (x, t) = ∑N l=1 dl (t)wl (x), insert it into (3.2), multiply
N
(dN
k ) (t) + λk dk (t) = 0,
′ N
k (0) = (u , wk )L2 ,
dN 0
where we used orthonormality and the eigenfunction property of the wl . This is a system
(actually a decoupled one in this simple case) of linear ordinary dierential equations,
which has a global smooth solution by standard ODE theory.
We wish to let N →∞ and hope to obtain a solution to the original problem in the
limit. To this end, observe that multiplication of (3.2) with its solution uN and integration
in space yields (in analogy to NSE) the energy equality
ˆ ˆ tˆ ˆ ˆ
1 1 1
∣uN (x, t)∣ dx+
2
∣∇uN (x, s)∣ dx ds =
2
∣PN u (x)∣ dx ≤
0
∣u0 (x)∣ dx,
2 T3 0 T 3 2 T3 2 T3
and thus a uniform (in N ) bound of the Galerkin sequence in L∞ L2 ∩L2 H 1 . By the Banach-
Alaoglu Theorem, we may therefore a weakly*-convergent subsequence (not relabelled), so
ˆ ∞ˆ ˆ ∞ˆ
uN ∆φ dx dt → u∆φ dx dt,
0 T3 0 T3
and nally
ˆ ˆ
PN u0 φ(t = 0)dx → u0 φ(t = 0)dx,
T3 T3
17
18 3. EXISTENCE OF WEAK SOLUTIONS
3.1.2. Galerkin for NSE. Recall the basis of eigenfunctions of −∆ (now viewed as
the Stokes Operator) from Theorem 2.11. Let PN H ∶= span{w1 , . . . , wN } and consider the
projection operator PN ∶ L̇2 → PN H given by
N
PN (u) = ∑ (u, wj )wj
j=1
using the L2 inner product. Clearly, for all u∈H we have PN u → u in H (i.e. in the
2
L -norm.) Indeed,
∞ ∞
∥PN u − u∥2L2 = ∥ ∑ (u, wj )wj ∥2 = ∑ ∣(u, wj )∣2 → 0
j=N +1 j=N +1
as N → ∞.
Definition 3.1. The N -th order Galerkin approximation of the NSE with initial data
u0 ∈ H is the solution of the equation
uN (0) = PN u0 .
Note we have not yet proved that such uN exists! Again we have projected away the
pressure. Like for the heat equation, we want to take N → ∞.
From the energy estimate, we expect a uniform bound of the form
ˆ ˆ tˆ
sup ∣uN (x, t)∣ dx + 2ν
2
∣∇uN (x, s)∣2 dx ds ≤ C < ∞, (3.4)
t Ω 0 Ω
so by the Banach-Alaoglu Theorem we will be able to pass to weak limits:
uN ⇀ u, ∇uN ⇀ ∇u
2
weakly in L .
So far, the general strategy seems similar as for the heat equation. However, for NSE,
there are two main issues to solve:
First, the Galerkin approximation (3.3) and thus the resulting system of ODEs now
feature a quadratic term, so that the ODE solution is prima facie only obtained up to
a possibly nite blow-up time; indeed, the simplest quadratic ODE, ẋ = x2 , does exhibit
nite-time blow-up. Even worse, the existence interval [0, TN ) might depend on N and
could therefore, in the worst case, converge to zero as N → ∞, so that in the limit, we
would be left with nothing. It turns out, luckily, that such blow-up scenarios can rather
easily be ruled out by virtue of the nite dimensional energy equality (3.4).
Once we have globally existing Galerkin approximants {uN }N ∈N which satisfy the uni-
form bound (3.4), we need to establish that the weak limit u is a weak solution. For the
heat equation (and more generally for linear equations), this was trivial. However, for the
NSE, in the weak formulation we have the nonlinear term
ˆ ∞ˆ
(uN ⋅ ∇)uN ⋅ φ dx dt
0 T3
and it is not clear whether
meaning that weak limits and nonlinearities do, in general, not converge. However, if
we knew uN ⇀ u strongly and ∇uN ⇀ ∇u weakly, then it would follow (uN ⋅ ∇)uN ⇀
(u ⋅ ∇)u weakly. The strong convergence of uN will be obtained by means of the following
compactness result, which can be seen as a time-dependent version of the classical Rellich
compactness theorem:
Lemma 3.2 (Aubin-Lions). Let 0 < T < ∞ and assume for some 1 < p, q ≤ ∞ that
∥uN ∥Lq (0,T ;V ) + ∥∂t uN ∥Lp (0,T ;V ′ ) ≤ C,
for a constant C ≠ C(N ). Here, V ′ is the dual space of V . Then there exists a subsequence
{uNj }j∈N such that uNj → u strongly in Lq (0, T ; H), for some u ∈ Lq (0, T ; H).
Proof. Consider for k ∈ N the map t ↦ (uN (t), wk )L2 . It is not dicult to see
(exercise!) that this map is weakly dierentiable with weak derivative (∂t uN (t), wk ), which
is a well-dened Lp function since ∂t uN ∈ Lpt V ′ and wk ∈ V , and so (cf. exercise) s ↦
(uN (s), wk ) is (absolutely) continuous and
ˆ s
(uN (s), wk ) = (uN (s ), wk ) +
∗
(∂t uN , wk ) dt
s∗
for all s, s∗ ∈ [0, T ]. Again by continuity we may invoke the mean value theorem for
integrals to conclude there exists an s∗ ∈ [0, T ] such that
ˆ T
1
(uN (s ), wk ) =
∗
(uN , wk ) dt.
T 0
Hence,
ˆ s
sup ∣(uN (s), wk )∣ ≤ ∣(uN (s ), wk )∣ + ∣ (∂t uN , wk ) dt∣
∗
0≤s≤T s∗
ˆ T ˆ T
1
≤ ∥uN ∥L2 (T3 ) ∥wk ∥L2 (T3 ) dt + ∥∂t uN ∥V ′ ∥wk ∥V dt.
T 0 0
We can now use Hölder's inequality combined with the facts that ∥wk ∥L2 (T3 ) = 1 and
√
∥wk ∥V = λk to obtain the bound
1 1− 1q 1− 1
√ √
≤ T ∥uN ∥Lq (0,T ;H) + T p ∥∂t uN ∥Lp (0,T ;V ′ ) λk ≤ C1 + λk C2
T
where we have used that ∥∂t uN ∥Lp (0,T ;V ′ ) is uniformly bounded and by Poincaré's inequality
we have ∥u∥H ≤ C∥u∥V , which gives a uniform bound on ∥uN ∥Lq (0,T ;H) . It follows that
Pk uN ∶= ∑kj=1 (uN , wj )wj is in C([0, T ]; H) and
k √ √
sup ∥Pk uN (s)∥H ≤ ∑ (C1 + C2 λj ) ≤ k(C1 + C2 λk )
s∈(0,T ) j=1
as λj is increasing.
Claim 1: For every k , {Pk uN }N ∈N has a subsequence converging in C([0, T ]; H). For
this, we will use the Arzelà-Ascoli Theorem. As Pk H is nite-dimensional, it suces to
check that the sequence is uniformly bounded and equicontinuous. From the previous
estimate we already have uniform boundedness, so all we need to check is equicontinuity.
20 3. EXISTENCE OF WEAK SOLUTIONS
for all N ≥ k . Indeed, we know that C ≥ ∥uN ∥Lq (0,T ;V ) and as (∇wj , ∇wk ) = (−∆wj , wk ) =
λj (wj , wk ) = λj δjk , then
ˆ ˆ q
T T ⎛∞ 2⎞
2
ˆ q
ˆ
q T ⎛ ∞ 2⎞
2 q T
≥ λk+1
2
∑ ∣(uN (s), wj )L2 ∣ ds = λk+1
2
∥Pk uN (s) − uN (s)∥qL2 ds,
0 ⎝j=k+1 ⎠ 0
∥uN − uM ∥Lq (0,T ;H) ≤ ∥uN − Pk uN ∥Lq (0,T ;H) + ∥Pk uN − Pk uM ∥Lq (0,T ;H)
1
+ ∥uM − Pk uM ∥Lq (0,T ;H) ≤ 3δ q
and since δ>0 was arbitrary, Claim 2 follows and we are done.
Only one ingredient is missing before we can embark on the existence proof for weak
solutions:
3.2. THE EXISTENCE PROOF 21
Lemma 3.3 (L
p
Let Ω be a measure space and u ∈ Lp (Ω) ∩ Lq (Ω) for
interpolation).
some 1 ≤ p ≤ q ≤ ∞. Then u ∈ L (Ω) for all p ≤ r ≤ q, and ∥u∥Lr ≤ ∥u∥αLp ∥u∥1−α
r
Lq , where
1
r = α
p + 1−α
q .
p q
Proof. Using Hölder's inequality with
αr , (1−α)r we see that
ˆ ˆ ˆ αr ˆ (1−α)r
p p
∣u∣ dx =
r
∣u∣ ∣u∣
αr (1−α)r
dx ≤ ( ∣u∣ dx) ( ∣u∣ dx)
p q
Ω Ω Ω Ω
r
= [∥u∥αLp ∥u∥1−α
Lq ]
and so we are done.
lution of the NSE with initial data u0 . Moreover, this solution satises ∂t u ∈ L4/3
loc (0, T ; V ).
′
N
k ) (t) + νλk dj (t) + ∑ dj (t)dl (t)Bkjl = 0,
(dN ′ N N N
j,l=1
with k = 1, . . . , N and
ˆ
Bkjl ∶= (wj (x) ⋅ ∇)wl (x) ⋅ wk (x) dx.
T3
Indeed, (PN v, wk )L2 = (v, wk )L2 for all v ∈ L2 by a simple linear algebra argument.
This is a system of N ODEs for the N unknown functions dk (k = 1, . . . , N ),
N
with
initial condition
ˆ
k (0) =
dN u0 (x)wk (x) dx.
T3
(The latter is obtained by multiplying
N
uN (x, 0) = ∑ dN
j (0)wj (x)
j=1
by wk and employing the initial condition uN (x, 0) = PN u0 (x). Note again (PN u0 (x), wk (x)) =
(u0 , wk (x)).)
22 3. EXISTENCE OF WEAK SOLUTIONS
Step 2: Show TN = ∞ for all N , via energy estimates. Let s ∈ (0, TN ). Multiply the
Galerkin equation at time uN (s) and integrate in x to get
s with
ˆ ˆ
∂t uN (s) ⋅ uN (s) dx + PN (uN (s) ⋅ ∇)uN (s) ⋅ uN (s) dx
T3 T3
ˆ
=ν ∆uN (s) ⋅ uN (s) dx.
T3
Observe each integral in order: for the rst we see that
ˆ ˆ
1 d
∂t uN (s) ⋅ uN (s) dx = ∣uN (s)∣2 dx,
T3 2 dt T3
for the second,
ˆ ˆ
PN (uN (s) ⋅ ∇)uN (s) ⋅ uN (s) dx = (uN (s) ⋅ ∇)uN (s) ⋅ uN (s) dx
T3 T3
3 ˆ
= ∑ ujN (s)∂xj ulN (s)ulN (s) dx
j,l=1 T3
3 ˆ
=− ∑ ∂xj ujN (s)(ulN (s))2 dx
j,l=1 T3
3 ˆ
− ∑ ∂xj ujN (s)ulN (s)∂xj ulN (s) dx = 0
j,l=1 T3
1 d
∥uN (s)∥2L2 (T3 ) + ν∥∇uN (s)∥2L2 (T3 ) = 0.
2 dt
We note that this equality implies (after integration in s) that
ˆ ∞
1 1 1 0 2
sup ∥uN (t)∥2L2 ≤ ∥u0 ∥2L2 and ∥∇uN (s)∥2L2 ds ≤ ∥u ∥L2 ,
t 2 2 0 2ν
thus {uN } are uniformly bounded in L (0, ∞; H) ∩ L2 (0, ∞; V ).
∞
using interpolation (Lemma 3.3). Now we can use the Sobolev embedding ∥v∥L6 ≤ C∥v∥Ḣ 1 =
C∥v∥V = C∥∇v∥L2 and the projection property ∥PN φ∥V ≤ ∥φ∥V to obtain
1 3
∣((uN ⋅ ∇)uN , PN φ)∣ ≤ C∥uN ∥L2 ∥∇uN ∥L2 ∥φ∥V . 2 2
Using Hölder's inequality in time on both terms (rst L3 (on 1), L3/2 second L∞ , L1 ) we
obtain
4
⎡ ˆ T ⎤
1 3
⎢ 2⎥ 2
≤ CνT ⎢⎢( ∥∇uN (s)∥ ds) ⎥⎥ + C∥uN ∥L3 ∞ (0,T ;H) ∥uN (s)∥2L2 (0,T ;V )
1
3
2
⎢ 0 ⎥
⎣ ⎦
which becomes
1 4 2
≤ CνT 3 ∥uN (s)∥L3 2 (0,T ;V ) + C∥uN ∥L3 ∞ (0,T ;H) ∥uN (s)∥2L2 (0,T ;V )
which is bounded uniformly in N (for xed ν and T !) owing to the energy estimates from
Step 2.
Step 4: Extract a convergent subsequence.
By Banach-Alaoglu, there is a subsequence {uNj }j such that
uNj ⇀ u
∗
∂t uN ⇀ ∂t u L4/3 (0, T ; V ′ ).
∗
in
4/3
∂t uN ⇀ ∂t u Lloc (0, ∞; V ′ ).
∗
in
Choosing yet another subsequence and applying again a diagonal argument, we obtain
by Lemma 3.2 (Aubin-Lions)
uN → u
strongly in L2loc (0, ∞; H).
Step 5: Show the limit is a solution of NSE.
By Lemma 2.14, it suces to choose a test function of the form
m
φ(x, t) = ∑ dk (t)wk (x) ∈ D̃σ .
k=1
24 3. EXISTENCE OF WEAK SOLUTIONS
Let T be so large that supp(dk ) ⊂ [0, T ) for all k = 1, . . . , m. By the Galerkin equation,
for any N ≥ m we have
ˆ ∞ˆ ˆ ∞ˆ
− uN ⋅ ∂t φ dx dt + ν ∇uN ∶ ∇φ dx dt
T3 T3
0 0
ˆ ∞ˆ ˆ
+ (uN ⋅ ∇)uN ⋅ φ dx dt = u0 ⋅ φ(0) dx.
0 T3 T3
Strong Solutions
4.1. Some More on Bochner Spaces
For this entire chapter, let 0 < T < ∞ be arbitrary but xed. We collect a few technical
results to be used later.
Proposition 4.1. Let X Banach and suppose u, w ∈ L1 (0, T ; X). Then the following
are equivalent:
(1) ∂t u = win the weak sense;
(2) There exists ξ ∈ X such that, for a.e. t ∈ (0, T ),
ˆ t
u(t) = ξ + w(s) ds;
0
(3) For every v ∈ X ′ , in the weak sense it holds that
d
(u, v) = (w, v).
dt
Moreover, if one (and thus all) of these conditions holds, then u can be altered on a nullset
of times so that it belongs to C([0, T ]; X).
Proof. Exercise.
Proposition 4.2. Let u ∈ L2 (0, T ; Ḣ 1 (T3 )) and ∂t u ∈ L2 (0, T ; H −1 (T3 )). Then,
(1) u ∈ C([0, T ]; L̇2 (T3 ));
(2) the map t ↦ 21 ∥u(t)∥2L̇2 is weakly dierentiable with
d1
∥u(t)∥2L̇2 = (u(t), ∂t u(t)) for a.e. t ∈ (0, T );
dt 2
(3)
´ Proof. Let η ∈ Cc∞ (R) be a standard mollier (in the time variable), that is, η ≥ 0,
R η(θ) dθ = 1, supp η ⊂ B1 (0), and η = η(∣θ∣). Set η (t) ∶= ( ) and, for any f ∈ Lloc (R),
1 t 1
f ∶= f ∗ η , i.e.
ˆ
f (t) = f (t − τ )η (τ ) dτ.
−
Consider now u as given in the statement and extend it to t ∈ R by zero, so that its
mollication is well-dened on all of [0, T ]. For , δ > 0, uδ and u are smooth in time, and
we may thus use the standard Leibniz rule to compute
ˆ
d
∥u (t) − uδ (t)∥2L2 = 2 (u (t) − uδ (t)) ⋅ (∂t u (t) − ∂t uδ (t)) dx. (4.1)
dt T3
By choice of s, the rst expression on the right hand side converges, as , δ → 0, to zero,
and so does the dual pairing under the integral, for almost every τ. But then the integral
itself converges by dominated convergence (exercise).
It follows that {u }>0 is Cauchy in C([0, T ]; L2 (T3 )), and since this space is Banach,
it follows u → u ∈ C([0, T ]; L2 (T3 )), whence (1) is established.
For (2), again by the classical Leibniz rule,
ˆ
d
∥u (t)∥2L2 = 2 u (t) ⋅ ∂t u (t) dx
dt T3
and hence, after integration from s to t,
ˆ tˆ
∥u (t)∥2L2 = ∥u (s)∥2L2 + 2 u (τ ) ⋅ ∂t u (τ ) dτ dx, (4.2)
s T3
and the same equality follows for u instead of u by similar convergence arguments as
before. Application of Proposition 4.1 (2) then gives the desired characterisation of the
weak time derivative.
Finally, for (3), integrate (4.2) in s over (0, T ) to arrive at
ˆ T
T ∥u (t)∥2L2 ≤ ∥u (s)∥2L2 ds + T (∥u∥2L2 Ḣ 1 + ∥∂t u∥2L2 H −1 ),
0
and thus (3) follows because t is arbitrary and the right hand side is independent of t.
Corollary 4.3. Let u, v ∈ L2 (0, T ; Ḣ 1 (T3 )) and ∂t u, ∂t v ∈ L2 (0, T ; H −1 (T3 )). Then
the map t ↦ (u, v) is absolutely continuous with weak derivative
d
(u, v) = (∂t u, v) + (u, ∂t v).
dt
Proof. This follows from the polarisation identity
1
(u, v) = (∥u + v∥2 − ∥u∥2 − ∥v∥2 )
2
and the preceding proposition.
The following can be seen as an extension of Proposition 4.2 to higher order Sobolev
spaces:
Proposition 4.4. Let n ∈ N0 and suppose u ∈ L2 (0, T ; Ḣ n+2 (T3 )) and ∂t u ∈ L2 (0, T ; Ḣ n (T3 )).
Then, u ∈ C([0, T ]; Ḣ n+1 (T3 )), and
max ∥u(t)∥Ḣ n+1 ≤ C (∥u∥L2 Ḣ n+2 + ∥∂t u∥L2 Ḣ n ) .
t∈[0,T ]
Proof. We indicate only the formal argument and remark that the rigorous proof
proceeds exactly as in Proposition 4.2 by time mollication.
So assume u is smooth and take for simplicity n = 0, then we compute
d d
∥u(t)∥2Ḣ 1 = ∥∇u(t)∥2L2
dt dt
ˆ
=2 ∇u ∶ ∂t ∇u dx
T3
ˆ
= −2 ∆u ⋅ ∂t u dx ≤ ∥u(t)∥2H 2 + ∥∂t u(t)∥2L2 ,
T3
and the estimate
Lemma 4.6. Let u be a strong solution, then ∂t u, (u ⋅ ∇)u, ∆u ∈ L2 (0, T ; L2 (T3 )).
Proof. From the assumption u ∈ L2 (0, T ; H 2 ) it follows immediately that ∆u ∈
L (0, T ; L ).
2 2
For the nonlinear term, note that H
2 ∞
embeds continuously into L , so
that
ˆ T ˆ ˆ T ˆ
∣u∣2 ∣∇u∣2 dx dt ≤ ∥u(t)∥2L∞ ∣∇u(x, t)∣2 dx dt ≤ C∥u∥2L∞ H 1 ∥u∥2L2 H 2 .
0 T3 0 T3
It remains to estimate the time derivative. In view of Lemma 2.16 and the remark after
Proposition 3.5, for every smooth divergence-free vector eld φ ∈ C ∞ (T3 ) and every t∈
[0, T ] we have
ˆ ˆ ˆ tˆ
u(t) ⋅ φ dx = u ⋅ φ dx −
0
∇u ∶ ∇φ + (u ⋅ ∇)u ⋅ φ dx ds
T3 T3 0 T3
ˆ ˆ tˆ
= u ⋅ φ dx +
0
∆u ⋅ φ − (u ⋅ ∇)u ⋅ φ dx ds,
T3 0 T3
as u has weak second space derivatives. Proposition 4.1 allows us to take the time derivative
of this equality to obtain, for every t ∈ [0, T ],
ˆ ˆ
∂t u ⋅ φ dx = ν∆u ⋅ φ − (u ⋅ ∇)u ⋅ φ dx.
T3 T3
Let ψ ∈ C ∞ (T3 ) be any smooth vector eld and ψ = φ + ∇π its Helmholtz decomposition,
so that div φ = 0. Then, as u(t) ∈ H ,
ˆ ˆ
∂t u ⋅ ψ dx = ∂t u ⋅ ψ(t) dx
T3 3
ˆT
= ν∆u ⋅ Pφ − (u ⋅ ∇)u ⋅ Pψ dx
T3
ˆ
= P(ν∆u − (u ⋅ ∇)u) ⋅ ψ dx,
T3
and since ψ was arbitrary, it follows that ∂t u = P(∆u−(u⋅∇)u). By the previous estimates,
this is indeed in L2 (0, T, L2 ).
Lemma 4.7. Let u be a strong solution. Then, for any w ∈ L2 (0, T ; H),
ˆ T ˆ
(∂t u + (u ⋅ ∇)u − ν∆u) ⋅ w dx dt = 0.
0 T3
Proof. By similar arguments as in the proof of Lemma 2.14, the space D̃σ is dense
in L (0, T ; H),
2
and therefore it suces to consider w ∈ D̃σ . But such w may be used as a
28 4. STRONG SOLUTIONS
ˆ T ˆ ˆ tˆ
0= (∂t u+(u⋅∇)u−ν∆u)⋅uχ[s,t] dx dτ = (∂t u+(u⋅∇)u−ν∆u)⋅u dx dτ.
0 T3 s T3
ˆ tˆ ˆ ˆ
1 1
u ⋅ ∂t u dx dτ = ∣u(t)∣ dx − 2
∣u(s)∣2 ds,
s T3 2 T3 2 T3
Lemma 4.10. Let u be a weak solution of NSE and U ∈ L2 (0, T ; H 2 ∩ V ) a vector eld
with ∂t U ∈ L (0, T ; L2 ). Then U is a valid test function in the denition of weak solution
2
meaning that {UN } is Cauchy, and thus convergent, in C([0, T ]; H 1 ). The proof now
proceeds as in Proposition 4.4.
Theorem 4.11 (Weak-strong uniqueness). Let U be a strong solution and u a weak
solution that satises the energy inequality, and assume both solutions share the same initial
data u0 ∈ H . Then u ≡ U almost everywhere.
Proof. The proof relies on an estimate of the relative energy between u and U , dened
as
ˆ
1
Erel (t) = ∣u(x, t) − U (x, t)∣2 dx.
2 T3
In the course of the computation, we use three ingredients:
(1) The weak formulation for u, tested with U, as justied by Lemma 4.10:
ˆ ˆ ˆ tˆ
u(t) ⋅ U (t) dx − u0 ⋅ U (0) dx = u ⋅ ∂t U − ν∇u ∶ ∇U − (u ⋅ ∇)u ⋅ U dx ds;
T3 T3 0 T3
(2) The pointwise solution property of U , integrated against u, as justied by Lemma 4.7:
ˆ T ˆ
(∂t U + (U ⋅ ∇)U − ν∆U ) ⋅ u dx dt = 0;
0 T3
(3) The energy (in)equalities for u and U, as justied by assumption and by Theo-
rem 4.9, respectively:
ˆ ˆ tˆ ˆ
1 1
∣u(t)∣ dx + ν
2
∣∇u(x, τ )∣ dτ ≤ 2
∣u(s)∣2 ds
2 T3 s T3 2 T3
and similar for U.
Using (1), we obtain
ˆ
1
Erel (t) = ∣u(x, t) − U (x, t)∣2 dx
2 T3
ˆ ˆ ˆ
1 1
= ∣u(x, t)∣2 dx + ∣U (x, t)∣2 dx − u(t) ⋅ U (t) dx
2 T3 2 T3 T3
ˆ ˆ
1 1
= ∣u(x, t)∣ dx +
2
∣U (x, t)∣2 dx
2 T3 2 T3
ˆ tˆ ˆ
− u ⋅ ∂t U − ν∇u ∶ ∇U − (u ⋅ ∇)u ⋅ U dx ds − u0 ⋅ U 0 dx.
0 T3 T3
30 4. STRONG SOLUTIONS
By (3) and the assumption u0 = U 0 , the sum of the inital terms is non-positive and can
thus be neglected in the estimate, so that
ˆ tˆ ˆ tˆ ˆ tˆ
Erel (t) ≤ − u⋅∂t U −ν∇u ∶ ∇U −(u⋅∇)u⋅U dx ds−ν ∣∇u∣2 dx ds−ν ∣∇U ∣2 dx ds.
0 T3 0 T3 0 T3
where
ˆ tˆ
R=− (U ⋅ ∇)U ⋅ u + (u ⋅ ∇)u ⋅ U dx ds.
0 T3
whence
ˆ tˆ
R=− (U ⋅ ∇)(U − u) ⋅ u − (u ⋅ ∇)(U − u) ⋅ U dx ds,
0 T3
´
and nally an application of the formula
T3 (U ⋅ ∇)(U − u) ⋅ (U − u) dx = 0 gives
ˆ tˆ
R=− ((U − u) ⋅ ∇)(U − u) ⋅ U dx ds.
0 T3
and since U is a strong solution, ∥U ∥2L∞ ∈ L1 (0, T ), and we may then use Grönwall's
inequality to conclude Erel ≡ 0, which implies the theorem.
Theorem 4.12 (Weak-strong stability). Let U be a strong solution and u a weak solu-
tion that satises the energy inequality, with initial data U 0 ∈ H and u0 ∈ H , respectively.
Then there exists a constant, depending only on the norm of U in L∞ H 1 ∩ L2 H 2 and on
the viscosity ν , such that for all t ∈ [0, T ]
∥u(t) − U (t)∥L2 (T3 ) ≤ eCt ∥u0 − U 0 ∥L2 (T3 ) .
Proof. The proof is almost identical to the one of the preceding theorem, only that
the initial terms do not cancel. It is left as an exercise. The reader might also want to give
an explicit formula for the constant C in terms of U and ν.
4.3. LOCAL EXISTENCE OF STRONG SOLUTIONS 31
1/2 1/2
∥u∥L∞ ≤ C∥u∥H 1 ∥u∥H 2 .
Proof. We split the Fourier series of u into low and high frequencies, with M >0 to
be chosen later:
= ∑ ûk e ik⋅x
∣k∣∣k∣−1 + ∑ ûk eik⋅x ∣k∣2 ∣k∣−2
∣k∣≤M ∣k∣>M
1/2 1/2
⎛ 1 ⎞ ⎛ 1 ⎞
≤ ∥u∥Ḣ 1 ∑ + ∥u∥Ḣ 2 ∑
⎝∣k∣≤M ∣k∣2 ⎠ ⎝∣k∣>M ∣k∣4 ⎠
It is easy to see that
1/2 ˆ 1/2
⎛ 1 ⎞ 1
∑ ≤C( dx)
⎝∣k∣≤M ∣k∣2 ⎠ BM (0)∖B1 (0) ∣x∣2
ˆ M 1/2
r2
=C( dr) ≤ CM 1/2 ,
1 r2
and similarly
1/2 ˆ 1/2
⎛ 1 ⎞ 1
∑ ≤C( dx)
⎝∣k∣>M ∣k∣4 ⎠ R3 ∖BM (0) ∣x∣4
ˆ 1/2
∞
r2
=C( dr) ≤ CM −1/2 ,
M r4
so that in total
M=
∥u∥Ḣ 2
The choice
∥u∥Ḣ 1 now yields the result.
and we wish to estimate the right-hand side. For this, we use Agmon's inequality and the
estimate ∥uN ∥H 2 ≤ C∥∆uN ∥L2 (which is trivial to show in Fourier):
ˆ
∣ (uN ⋅ ∇)uN ⋅ ∆uN dx∣ ≤ ∥uN ∥L∞ ∥∇uN ∥L2 ∥∆uN ∥L2
T3
1/2 1/2
≤ C∥u∥H 1 ∥u∥H 2 ∥∇uN ∥L2 ∥∆uN ∥L2
3/2 3/2
≤ C∥∇uN ∥L2 ∥∆uN ∥L2 .
ap bq
Young's inequality ab ≤ p + 1
q for p + 1
q = 1, applied here with p=4 and q= 4
3 , gives
ˆ
1
∣ (uN ⋅ ∇)uN ⋅ ∆uN dx∣ ≤ C(ν)∥∇uN ∥6L2 + ν∥∆uN ∥2L2 ,
T3 2
so that (4.4) becomes
d
∥∇uN ∥2L2 + ν∥∆uN ∥2L2 ≤ C(ν)∥∇uN ∥6L2 . (4.5)
dt
Setting aside the Laplacian term for the moment, we see that Y ∶= ∥∇uN ∥2L2 satises the
ordinary dierential inequality
Theorem 4.15. Let s > 23 , then H s (T3 ) embeds continuously into L∞ (T3 ) and forms
a Banach algebra, that is,
∥uv∥H s ≤ C∥u∥H s ∥v∥H s ,
where the constant depends only on s.
Proof. We give the proof only for u, v ∈ Ḣ s , as this is the situation we shall need.
However the general statement follows very similarly (replacing ∣k∣s by (1 + ∣k∣)s ). So let
x∈T 3
, then
1/2 1/2
∣u(x)∣ = ∣ ∑ ûk e ik⋅x
∣ ≤ ∑ ∣ûk ∣ = ∑ ∣ûk ∣∣k∣ ∣k∣ s −s
≤ ( ∑ ∣ûk ∣ ∣k∣ )2 2s
( ∑ ∣k∣−2s
) .
k≠0 k≠0 k≠0 k≠0 k≠0
But the rst factor is precisely the homogeneous H s -norm of u, and for the second factor
we compute
ˆ ˆ ∞
1
∑ ∣k∣ ≤C dx = C r2−2s dr,
−2s
k≠0 R3 ∖B0 (1) ∣x∣2s 1
⎛ ⎞ ⎛ ⎞
u(x)v(x) = ( ∑ ûk eik⋅x ) ∑ v̂j eij⋅x = ∑ ∑ ûk−l v̂l eik⋅x , (4.6)
k≠0 ⎝j≠0 ⎠ k∈Z3 ⎝l∈Z3 ⎠
so that ̂
u v k = ∑l∈Z3 ûk−l v̂l .
As another ingredient we recall the inequality ∣k∣s ≤ C(∣k − l∣s + ∣l∣s ), where C depends
only on s > 0.
We can now estimate
∑ (1+∣k−l∣ )ûk−l v̂l = [((1 + ∣∇∣ )u)v]k ∑ (1+∣l∣ )ûk−l v̂l = [u((1 + ∣∇∣ )v)]k ,
s s ∧ s s ∧
and
l∈Z3 l∈Z3
so that (using Plancherel's Theorem)
2 2
∥uv∥2H s ≤ C ∑ ∣[((1 + ∣∇∣s )u)v]∧k ∣ + C ∑ ∣[u((1 + ∣∇∣s )v)]∧k ∣
k≠0 k≠0
=C (∥((1 + ∣∇∣ )u)v∥L̇2 + ∥u((1 + ∣∇∣s )v)∥2L̇2 )
s 2
1 d
∥uN ∥2H m + ν∥∇uN ∥2H m = −((uN ⋅ ∇)uN , uN )H m
2 dt
≤ ∥(uN ⋅ ∇)uN ∥H m ∥uN ∥H m
≤ ∥uN ∥H m ∥∇uN ∥H m ∥uN ∥H m
1
≤ ν∥∇uN ∥2H m + C(ν)∥uN ∥4H m .
2
Note that in the last step, we made use of the Cauchy-Schwarz inequality with (ab ≤
a2 + C()b2 ).
This results in
d
∥uN ∥2H m + ν∥∇uN ∥2H m ≤ (C(ν)∥uN ∥2H m )∥uN ∥2H m = g(t)∥uN ∥2H m , (4.7)
dt
where g ∶= C(ν)∥uN ∥2H m ∈ L1 (0, T ) uniformly in N , by virtue of the induction hypothesis.
Grönwall's inequality therefore yields
ˆ t
∥uN (t)∥2H m ≤ ∥u0 ∥2H m exp ( g(s) ds) .
0
Since the right hand side is nite and independent of N, we obtain the bound
Proof. By denition, u ∈ L2 (0, T ; H 2 ), and therefore for almost every s1 ∈ (0, ),
u(s1 ) ∈ H . Choosing such s1 and using u(s1 ) as initial data, and keeping in mind the
2
we may choose an s2 ∈ (s1 , ) such that u(s2 ) ∈ H . In this way we obtain a sequence
3
0 < s1 < s2 < s3 < . . . < such that u ∈ L∞ (sm , T ; H m+1 ) ∩ L2 (sm , T ; H m+2 ).
Observe that this implies ∂t u ∈ L (, T ; H ), because
2 m−1
u ∈ C([, T ]; H m ).
Theorem 4.18 (Time regularity). Let u be a strong solution, > 0, and j, k ∈ N. Then
∂tj u ∈ L (, T ; H (T )).
∞ k 3
Similarly as in the previous proof, we bound ∥P((u⋅∇)u)∥H k at each time by ∥u∥H k ∥u∥H k+1 ,
and ∥∆u∥H k ≤ ∥u∥H k+2 . But both these are bounded, uniformly in t, in the respective
Sobolev norms by virtue of Theorem 4.17.
For the induction step, dierentiate (4.8) j−1 times with respect to t to obtain
j−1
j−1
∂tj u = − ∑ ( )P((∂ti u ⋅ ∇)∂tj−1−i u) + ν∆∂tj−1 u.
i=0 i
But by induction hypothesis, the rst j −1 time derivatives are in L∞ (0, T ; H m ) for every
m, and so we can apply similar arguments as in the induction base to conclude.
Corollary 4.19. Let u be a strong solution of NSE on [0, T ]. Then, for every > 0,
u ∈ C ∞ (T3 × [, T ]).
4.5.1. Vorticity. Let u be a strong solution on [0, T ], then it is smooth (in space) for
any t ∈ (0, T ]. Dene the curl operator, which acts on smooth vectorelds u ∈ C ∞ (T3 ; R3 )
and yields another such vectoreld, by
It is easy to see that div curl u = 0 for any choice of u, and also curl ∇p = 0 for any scalar
eld p. Taking the curl of NSE and denoting ω = curl u (which is known as the vorticity of
the ow), we obtain
4.5.2. Blowup.
Definition 4.21. Let u0 ∈ V
and u be a corresponding Leray-Hopf solution of NSE. A
time T > 0 is called the blowup time for the solution if u is a strong solution on [0, T ∗ − ]
∗
A few remarks are in order. First, it is possible that no blowup occurs and hence no
(nite) blowup time exists this is trivially so e.g. for the zero solution. Secondly, if there
is a blowup, then the blowup time is uniquely determined by u0 : Indeed, as long as the
strong solution exists, it is unique in the class of Leray-Hopf solutions (Theorem 4.11).
A more substantial remark is that T∗ ∥∇u∥L2 (T ∗ ) = ∞.
is the smallest time at which
Indeed, the proof of Theorem 4.14 shows that u ∈ L (0, T ; H ) as long as u ∈ L (0, T ; H ),
2 2 ∞ 1
so that a solution that exits the former space will thereby also exit the latter.
5.1.1. Dissipative Solutions of the Euler Equations. We consider now the Euler
equations,
∂t u + (u ⋅ ∇)u + ∇p = 0
div u = 0,
´
2 T3 ∣u∣ dx is formally conserved by a similar computation as for NSE.
1 2
whose energy
Therefore the function space L (0, T ; H) appears suitable for the study of solutions.
∞
0
For the following formal computation, suppose u is a smooth solution with data u ,
and let U ∈ C (T × [0, T ]; R ) be any divergence-free eld. Denote
∞ 3 3
(∂t + u ⋅ ∇) (u − U ) + (u − U ) ⋅ ∇U + ∇π = E(U )
for some scalar eld π. Next, multiply this by u−U and integrate to obtain
ˆ ˆ ˆ
1 d
∣u − U ∣ dx +
2
(u ⋅ ∇)(u − U ) ⋅ (u − U ) dx + (u − U ) ⋅ ∇sym U (u − U ) dx
2 dt T3 3 T3
ˆT
= E(U ) ⋅ (u − U ) dx,
T3
where ∇sym = 12 (∇ + ∇t ) denotes the symmetric gradient. Note that the second integral
vanishes by the usual integration by parts argument, so we can estimate
ˆ ˆ ˆ
1 d
∣u − U ∣2 dx ≤ ∥∇sym U ∥L∞ ∣u − U ∣2 dx + E(U ) ⋅ (u − U ) dx.
2 dt T3 T3 T3
38
5.1. THE PERIODIC CASE 39
Recall from the exercises that C([0, T ]; L2w ) is the space of functions in L∞ (0, T ; L2 )
that are weakly continuous in the sense that, for every t ∈ [0, T ], u(s) ⇀ u(t) weakly in
L2 , as s → t.
Our formal computation motivates the following denition:
Remark 5.2. (1) The function spaces in this denition are chosen precisely such
that each term in (5.1) is well-dened.
(2) Choosing U ≡ 0, we obtain simply
ˆ ˆ
1 1
∣u(t)∣ dx ≤
2
∣u0 ∣2 dx ∀t ≥ 0, (5.2)
2 2
meaning that energy is not produced (but preserved or dissipated). This explains
the terminology.
(3) It can be shown that every solution in the sense of distributions that satises the
weak energy inequality (5.2) is also a dissipative solution in the sense of the given
denition. Conversely, there exist dissipative solutions that are not solutions in
the sense of distributions. As we shall see, however, dissipative solutions are
useful regardless of any ontological debates as to whether dissipative solutions are
really solutions of the Euler equations.
1
(U ⋅ ∇)U = 2(∇sym U )U − ∇∣U ∣2 ,
2
so that
On the one hand, E(U ) → E(U ) in L1 L2 as → 0. On the other hand, since ∇sym U ∈
L1 L∞ and U ∈ L∞ L2 , both ((∇sym U )U ) and (∇sym U )U converge to (∇sym U )U in L1 L2 ,
so the second expression in the last line of (5.3) converges to zero in L L . Hence, E(U )
1 2
converges to E(U ) in L L .
1 2
ˆ tˆ ˆ tˆ
− uν ⋅ ∂t v dx ds + (uν ⋅ ∇)uν ⋅ v dx ds
0 T3 0 T3
ˆ tˆ ˆ ˆ
+ν ∇uν ∶ ∇v dx ds = U (0) ⋅ v(0) dx − uν (t) ⋅ v(t) dx.
0 T3 T3 T3
By denition of E(v) and the divergence-free property of both elds (see Lemma 4.8), we
thence get
ˆ tˆ ˆ tˆ
uν ⋅ E(v) dx ds − ((uν − v) ⋅ ∇)v ⋅ (uν − v) dx ds
0 T3 0 T3
ˆ tˆ ˆ ˆ
+ν ∇uν ∶ ∇v dx ds = U (0) ⋅ v(0) dx − uν (t) ⋅ v(t) dx.
0 T3 T3 T3
Definition 5.6. Let Ω ⊂ R3 be a smooth bounded domain, then u ∈ C([0, T ]; L2w (Ω))∩
L 2
(0, T ; H0 (Ω)) is said to be a weak solution of NSE if it is weakly divergence-free and
1
satises
ˆ ∞ˆ ˆ ∞ˆ
− u ⋅ ∂t φ dx dt + (u ⋅ ∇)u ⋅ φ dx dt
0 Ω
ˆ ∞ ˆ0 Ω
ˆ ˆ
+ν ∇u ∶ ∇φ dx dt = u ⋅ φ(0) dx −
0
u(T ) ⋅ φ(T ) dx
0 Ω Ω Ω
for every divergence-free φ ∈ C 1 (Ω × [0, T ]) such that φ=0 on ∂Ω.
The last condition that the test function has to vanish on the boundary is decisive.
such that η(0) = 0, η(x) = 1 for every x ≥ 1, and η is monotone non-decreasing. Set
d(x) ∶= dist(x, ∂Ω) and let
d(x)
ην (x) ∶= η ( ).
ν
It is now tempting to use ην u as a test function, as this now satises the correct boundary
condition. However this function will, in general, not be divergence-free!
By Poincaré's Lemma, however, since div u = 0, there exists a smooth potential φ ∶ Ω →
R 3
such that curl φ = u and φ=0 at ∂Ω. Set
vν ∶= curl (ην φ) ,
then vν is divergence-free, is zero on the boundary, and agrees with u except on Ων .
Moreover one can show various estimates such as
where
ˆ
Rν (t) = (uν ⋅ ∇)(uν − vν ) ⋅ uν + ν∇uν ∶ ∇vν dx.
Ω
´t
It remains to show
0 Rν ds → 0.
We consider only the second term, using (5.6):
ˆ
∣ ν∇uν ∶ ∇vν dx∣ ≤ ν∥∇uν ∥L2 ∥∇u∥L2 + ν∥∇uν ∥L2 (Ων ) ∥∇u − ∇vν ∥L2 (Ων )
Ω
≤ Cν∥∇uν ∥L2 + Cνν −1/2 ∥∇uν ∥L2 (Ων ) .
The time integral of the second term goes to zero by assumption, and so does the integral
of the rst term by virtue of the energy inequality:
ˆ t ˆ t 1/2
ν ∥∇uν ∥L2 ds ≤ CT ν 1/2
(ν ∥∇uν ∥2L2 ds) ≤ CT v 1/2 → 0.
0 0
Bibliography
[1] P. Constantin and C. Foias. Navier-Stokes equations. University of Chicago Press, 1988.
[2] T. Kato. Remarks on zero viscosity limit for nonstationary Navier-Stokes ows with boundary. Seminar
on nonlinear partial dierential equations. Springer, New York, NY, 1984.
[3] P.-L. Lions. Mathematical topics in uid mechanics. Vol. 1. Incompressible models. Oxford Lecture
Series in Mathematics and its Applications, 3. Oxford Science Publications. The Clarendon Press,
Oxford University Press, New York, 1996.
[4] J. C. Robinson, J. L. Rodrigo, and W. Sadowski. The Three-Dimensional NavierStokes equations:
Classical theory. Cambridge Studies in Advanced Mathematics, Vol. 157. Cambridge University Press,
2016.
[5] R. Temam. Navier-Stokes equations. Theory and numerical analysis. Revised edition. Studies in Math-
ematics and its Applications, 2. North-Holland Publishing Co., Amsterdam-New York, 1979.
44