Introduction to the classical DeGiorgi method and its application in the regularity theory of
incompressible Navier-Stokes equations
Yi-Hsuan Lin
Abstract
The note is mainly for personal record, if you want to read it, please be careful. This note is given by Prof. Chi Hin Chan’s lecture during the summer course in NCTS, 2015. We will introduce the De-Giorgi’s method in the classical setting of elliptic PDE’s in the divergence form with rough diffusive coefficients. It was in this setting that De-Giorgi used his method to establish the regularity of solutions to general linear elliptic PDE’s in the framework of Holder’s continuity.
After this preparation in the classical elliptic setting, we go ahead to discuss a piece of recent work of Alexis Vasseur in which he applied the De-Giorgi’s method to yield an alternative proof of the key proposition in the famous partial regularity theory of Caffarelli, Kohn, Nirenberg.
Further, we will introduce notions such as the Serrin’s regularity criteria, and the notion of suitable weak solutions in the theory of weak solutions to the incompressible Navier-Stokes equations.
1 Introduction to DeGiorgi’s ideas
Let B(1) ⊂ RN (N ≥ 2). We consider a function u ∈ H1(B(1)) which is a weak solution to
∂α{aαβ∂βu} = 0 in B(1). (1) Here (aαβ) : B(1) → RN2 satisfies the following constraint:
1
Λ|ξ|2≤ aαβ(x)ξαξβ≤ Λ|ξ|2, ∀x ∈ B(1), ξ ⊂ RN. Or regard ξ ∈ Tx∗(B(1)). Relation with Riemannian geometry,
g = gαβdxα⊗ dxβ and
∆f = 1
√G∂α{gαβ√ G∂βf }, where gαβ= d(dxα, dxβ).
In this note, we focus on the DeGiorgi method, 1957-1960. Compare to the Schauder’s estimate, we need the coefficients lying in Cα(Ω), then the solution will lie in C2,α(Ω). The idea is that if we rescale uλ(x) = u(λx) and aαβλ (x) = aαβ(λx), if we make λ smaller and smaller, then aαβλ will behave like a piecewise
constant function. This scaling just “flatten” the graph of solutions, it will not change the height of functions. Therefore, the regularity will similar as the Laplacian case.
1.1 First DeGiorgi lemma
The first step is from the L2 to L∞. If the L2energy of the solution u in B(1) is small enough, then
kukL∞(B(12))≤ 1.
Lemma 1.1. There exists an absolute constant 0 = 0(N, Λ) > 0 such that the following statement holds for any weak solution u ∈ H1(B(1)) to (1). If ku+k2L2(B(1))< 0, then
u+≤ 1 in B(1 2).
Remark 1.2. For the whole information of u, just multiply the negative sign and consider the previous lemma again. What if kukL2(B(1)) is not small ? Rescale
V = 0
2kukL2(B(1))
u, then kV kL2(B(1))< 0. This will imply
kukL∞(B(12))≤ 2
0kukL2(B(1)). Set up the Game: Let Rk= 1
2(1 + 1
2k), k ≥ 0. Let ϕk ∈ C0∞(B(Rk)) with
ϕk=
(1 in B(Rk+1), 0 outside B(Rk),
and |∇ϕk| ≤ 2 Rk− Rk+1
χB(Rk)= 2k+3χB(Rk). Cutting level. Let V0 = u+, Vk = [u − (1 − 1
2k)]+ with Vk ≤ Vk−1 and
∇Vk = ∇u · χ{vk>0}. Uk=´
B(Rk)|Vk|2. Rough idea: Build up the following
Uk ≤ C0kUk−1β , β > 1. (2) If U0 = ku+k2L2(B(1)) < 0, then Uk → 0 as k → ∞. ´
B(12)[u − 1]+ = limk→∞Uk= 0 which implies u ≤ 1 in B(1
2).
Test against (1) by ϕ2kVk, Vk∈ H/ 01(B(1)) but ϕ2kVk ∈ H01(B(1)), which will be the test function. Then
0 = ˆ
B(1)
aαβ∂βu∂α{ϕ2kVk}
= ˆ
B(1)
aαβ∂βu · 2ϕk· ∂αϕk· Vk+ ˆ
B(1)
aαβ∂βu · ϕ2k∂αVk
= ˆ
B(1)
aαβ∂βVk· 2ϕk· ∂αϕk· Vk+ ˆ
B(1)
aαβ∂βVkϕ2k∂αVk,
and let gx(A, B) = aαβ(x) · AαBβ, ˆ
B(1)
aαβ∂αVk∂βVkϕ2k= −2 ˆ
B(1)
aαβ∂βVk· ϕk· ∂αϕk· Vk
= −2 ˆ
B(1)
gx(ϕk∇Vk, Vk∇ϕk)|xdx
≤ ˆ
B(1)
gx(ϕk∇Vk, ϕk∇Vk)|xdx
+1
ˆ
B(1)
gx(Vk∇ϕk, Vk∇ϕk)|xdx.
This implies ˆ
B(1)
aαβ∂αVk∂βVkϕ2k ≤ 4 ˆ
B(1)
aαβ∂αϕk∂βϕk· Vk2
≤ 4Λ ˆ
B(1)
|∇ϕk|2Vk2
≤ 4Λ22k+6 ˆ
B(Rk)
Vk2
= 128Λ4kUk and use ellipticity condition, we have
ˆ
B(1)
|∇(ϕkVk)|2≤ 2 ˆ
B(1)
|∇ϕk|2Vk2+ 2 ˆ
B(1)
ϕ2k|∇Vk|2
≤ 256 · 4k ˆ
B(Rk)
|Vk|2+ 256Λ4kUk
= 256(1 + Λ)4kUk. For N ≥ 3, use the Sobolev embedding theorem,
kϕkVkk
LN −22N (B(1))≤ C0( ˆ
B(1)
|∇(ϕkVk)|2)1/2.
For N = 2, use the Ladyzhenskaya inequality,
kϕkVkkL4(B(1))≤ kϕkVkk1/2L2(B(1))· k∇(ϕkVk)k1/2L2(B(1)).
Note that, {x ∈ B(1) : Vk > 0} ⊂ {x ∈ B(1) : Vk−1(x) > 1
2k}, χ{Vk>0} ≤ (2kVk−1)N −24 , and χB(Rk)≤ ϕ
2N N −2
k−1, Uk =
ˆ
B(Rk)
Vk2χ{Vk>0}≤ ˆ
B(Rk)
Vk2(2N −24 )kV
4 N −2
k−1
≤ (2N −24 )k ˆ
B(Rk)
V
2N N −2
k−1 ≤ (2N −24 )kkϕk−1Vk−1k
2N N −2
LN −22N (B(1))
,
which raised up the index. Finally, Uk≤ (2N −24 )kC0( ˆ
B(1)
|∇(ϕk−1Vk−1)|2)N −2N
≤ (2N −24 )kC0(256(1 + Λ)4kUk)N −2N
≤ [C0(N, Λ)]kU
N N −2
k−1 , k ≥ 1.
Thinking: {Uk}∞k=1 ⊂ [0, ∞), Uk+1 ≤ (C0)k+1Ukβ, k ≥ 0, C0 > 1, β > 1.
Ul+1∼ (C0k+1Ukβ−1)Ukand Uk+1∼ 1
AUk, where A > 1. It means C0k+1Ukβ∼ 1 A, then Uk ∼ 1
(AC0k+1)β−11
and Uk+1 ∼ 1
AUk ∼ 1 A
1 (AC0k+1)β−11
,..., A ∼ C
1 β−1
0 . Want C0k+1Uβ−1≤ 1
C
1 β−1
0
for all k ≥ 0. ku+k2L2(B(1)) ≤ 1 (C0C
1 β−1
0 )β−11 .
1.2 A detour: DeGiorgi’s isoperimetric inequality
Take u ∈ H1(B(1)) ∩ C∞(B(1)), A = {x ∈ B(1) : u(x) ≤ 0}, B = {x ∈ B(1) : u(x) ≥ 1}, C = {x ∈ B(1) : 0 < u(x) < 1}.
Proposition 1.3. The following estimate holds:
|A| · |B| ≤ CNk∇u+kL2(B(1))|C|12.
Takeu = max{0, inf{1, u}}, x ∈ A, y ∈ B, (Trick: Think of |∇e eu| = |∇eu|χC as a function on RN.)
eu(y) −u(x) = 1 =e ˆ |y−x|
0
d
dτ{u(x + τe y − x
|y − x|)}dτ
= ˆ |y−x|
0
∇u|e(x+τ y−x
|y−x|)· y − x
|y − x|dτ.
Then 1 ≤
ˆ |y−x|
0
|∇u(x + τe y − x
|y − x|)|dτ ≤ ˆ ∞
0
|∇eu(x + τ y − x
|y − x|)|dτ, and since |x| ≤ 1, if y ∈ B, |y − x| ≤ 2,
|B| ≤ ˆ
y∈B
ˆ ∞ 0
|∇eu(x + τ y − x
|y − x|)|dτ dy
≤ ˆ
|y−x|≤2
ˆ ∞ 0
|∇u(x + τe y − x
|y − x|)|dτ
≤ ˆ 2
0
rN −1 ˆ
SN −1
ˆ ∞ 0
|∇u(x + τ ω)|dτ dS(ω)dre
= 2N N
ˆ ∞ 0
ˆ
SN −1
τN −1|∇eu(x + τ ω)|dS(ω)dτ
= 2N N
ˆ
RN
|∇eu(x + y)
|y|N −1 dy = 2N N
ˆ
C
|∇u(y)|e
|x − y|N −1dy,
then
|A| · |B| ≤ 2N N
ˆ
y∈C
ˆ
x∈A
|∇u(y)|e
|x − y|N −1dxdy
=2N N
ˆ
y∈C
|∇eu(y)| · ˆ
x∈A
1
|x − y|N −1dxdy, x ∈ A and |x − y| ≤ 2. A ⊂ {x ∈ |x − y| ≤ 2} provided |y| ≤ 1,
ˆ
A
1
|x − y|N −1dy ≤ ˆ
|x−y|≤2
1
|x − y|N −1dy = CN. Therefore,
|A| · |B| ≤ CN ˆ
y∈C
|∇eu(y)|dy.
2 Oscillating lemma
Suppose we have a solution u : B(2) → R to (1) which satisfies 1. u ≤ 2 in B(2),
2. |{x ∈ B(1) : u(x) ≤ 0}| ≥ |B(1)|
2 . The second DeGiorgi lemma shows that
Lemma 2.1. ∃ another absolute constant λ = λ(Λ, N ) ∈ (0, 1) if the solution u : B(2) → R satisfies 1,2, then
u ≤ 2 − λ in B(1 2).
Is it possible that ˆ
B(1)
(2(u − 1)+)2< 0,
we care since u1 = 2(u − 1) is another solution to (1). If this happens, which means´
B(1)(u1)2< 0, by the first DeGiorgi lemma, we have u1≤ 1 in B(1
2), equivalently, u ≤ 3 2.
This may be false (do successive cuttings...) and treat u1 just in the same manner as u and let u2= 2(u1− 1), and iterate these steps until we find some uk≤ 3
2. Note that u2= 2(u1− 1) = 22u − 22−2, then uk= 2ku − (2k+ 2k−1+ · · · + 2)
= 2k{u − 2(1 − 1 2k)}.
Look at u = u0 and so on, set uk ∈ H1(B(1)), Ak = {x ∈ B(1) : uk(x) ≤ 0}, Bk = {x ∈ B(1) : uk(x) ≥ 1}, Ck = {x ∈ B(1) : 0 < uk(x) < 1}. |A1| =
|A0| + |C0|, |Ak| = |Ak−1| + |Ck−1| and property 2 gives that |Ak| ≥ |B(1)|
2 .
Remark 2.2. The structure of (1) allows us to obtain: For each solution ukto (1) in B(2),
ˆ
B(1)
|∇(uk)+|2≤ CΛ,N
ˆ
B(2)
|(uk)+|2≤ 4CΛ,N|B(2)|,
i.e.,
k∇(uk)+kL2(B(1))≤ CΛ,N.
|Ak| · |Bk| ≤ CNk∇(uk)+kL2(B(1))|Ck|12 ≤ CΛ,N|Ck|12.
Remember: We want |Bk| to be small eventually. We want 4|Bk| < 0, i.e., ˆ
B(1)
(uk+1)2+= 4 ˆ
B(1)
(uk− 1)2+χBk≤ 4|Bk| < 0.
If in the worst case
|Bk| ≤ 0
4 holds for all k, then
|Ck| ≥ (|B(1)|
2
0
4 1 CΛ,N
)2:= αN,Λ, ∀k ≥ 0.
But |B(1)| ≥ |Ak| ≥ kαN,Λ+|B(1)|
2 or k ≤ |B(1)|
2αN,Λ
, choose K0:= |B(1)|
2αN,Λ
+ 1.
In other words, the propagation on the property |Bk| ≥ 0
4 would stop at some 0 ≤ k ≤ K0, i.e., |BK0| < 0
4 will imply´
B(1)(uK0+1)2+ < 0 with uK0+1 solves (1). Apply the first DeGiorgi lemma to uK0+1 and get uK0+1≤ 1 in B(1
2), i.e., {u − 2(1 − 1
2K0+1)} ≤ 1
2K0+1 in B(1 2), then u ≤ 2 − 1
2K0+1.
Remark 2.3. You can always choose d1with |d1| ≤ λ
2 such that |u − d1| ≤ 2 −λ 2. Let
W(1)= 2 2 −λ
2 [u(1
2x) − d1] be a rescaled solution. Finally, we will see
oscB(1
2k)u ≤ (2 −λ2
2 )koscB(1)u or (|u| ≤ 2)
oscB(rk)u ≤ 4δk. Thinking: |x| ∼ rkwhich means k ∼ ln |x|
ln r , then δk ∼ δln |x|ln r = (δlogδ|x|)ln r1 logδ e1 =
|x|ln r1 logδ e1 ∼ |x|γ. Then |u(x) − u(0)| ∼ |x|γ.
3 Introduction to the elements of incompress- ible Navier-Stokes (NS) equations
3.1 Historical view
Starting point: The linear theory. Let Ω be a bounded domain with smooth boundary in RN, N = 2, 3. Consider
(∂tu − ∆u + ∇P = 0
∇ · u = 0 in [0, T ] × Ω. (3)
To write down the weak formulation, we need the right functional space for the
“test-vector fields” on Ω. Let
Λ1c,σ(Ω) = the space of al smooth, compactly supported divergence free v.f on Ω, V (Ω) = Λ1c,σk·kH1 (Ω), and HΩ= Λ1c,σL
2(Ω)
. If Ω is bounded, kϕkH1
0(Ω)= kϕkV = k∇ϕkL2(Ω)since kϕkL2(Ω)≤ C(Ω)k∇ϕkL2(Ω). If Ω = RN, kϕkV = kϕkL2(RN)+ k∇ϕkL2(RN) .
Let eV (Ω) = {u ∈ H01(Ω)|∇ · u = 0}, then V (Ω) ⊂ eV (Ω). Are they the same ? If Ω is either RN, RN+ or any bounded domain with smooth boundary in RN for N = 2, 3 (usual domains), then V (Ω) = eV (Ω). The example when V (Ω) eV (Ω) is to take
Ω = R3− {this (x1, x3)-plane with a disc removed}.
Heywood (1970) proved eV (Ω) = finite dimesional space VΩ, which has some- thing to do non-uniqueness, in general situation.
Theorem 3.1. Let Ω be a “usual domain” in RN(N = 2, 3). There exists a unique element u ∈ C0([0, T ]; HΩ) ∩ L2(0, T ; VΩ) which satisfies
A. ∂tv ∈ L2(0, T ; VΩ0)
B. For a.e. t ∈ [0, T ], and ϕ ∈ VΩ, we have h∂tv|t, ϕiV0
Ω⊗VΩ+ ˆ
Ω
∇v|t: ∇ϕ = (f (t), ϕ)V0
Ω⊗VΩ (weak formulation of (3)), C. u(0) = a.
The following properties are fundamental fact to further build the theory of weak solutions to NS equations. Why do we care about f ? Back to the original NS equations
(∂tu − ∆u + u · ∇u + ∇P = 0
∇ · u = 0 in [0, T ] × Ω (4)
Weak formulation for (4). For a.e. t ∈ [0, T ], ∀ϕ ∈ VΩ, we have h∂tu, ϕiV0
Ω⊗VΩ+ ˆ
Ω
∇u : ∇ϕ = − ˆ
Ω
(u · ∇u) · ϕ
= ˆ
Ω
(u ⊗ u) : ∇ϕ = ˆ
Ω
uαuβ∂αϕβ.
Reference: Seregin, Lecture notes on Regularity theory to NS equations, Oxford (Ch5) and Leray-Schauder fixed point method Twice. Big Glory (Leray-Hopf, 1930-1940).
Theorem 3.2. Take a ∈ HRN. There exists u ∈ L∞(0, T ; HRN) ∩ L2(0, T ; VΩ) which satisfies
D. For any ϕ ∈ L2(RN), the function t → ˆ
RN
u(t)ϕ is continuous, E. ∂tu ∈ L1(0, T ; V0
RN)
F. For a.e. t ∈ [0, T ], and any ϕ ∈ VRN, we have h∂tu, ϕi +
ˆ
RN
∇u : ∇ϕ = ˆ
RN
(u ⊗ u) : ∇ϕ, G.ku(t) − akL2(RN)→ 0 as t → 0.
H.For a.e. t ∈ [0, T ], we have 1
2ku(t)k2L2(RN)+ ˆ T
0
ˆ
RN
|∇u(t, x)|2dxdt ≤ 1
2kak2L2(RN).
The result is not easy by any standard method. Teman (abstract): either bounded or RN. Seregin (concrete): bounded domain with smooth boundary.
Moreover, in 2-dimesional space, the condition H will be an equality, and note that we only discuss the partial regularity properties for higher dimensional (N ≥ 3) spaces.
Recall in the Evans’ book, we have ∂tf ∈ L2(0, T ; H−1(Ω)), f ∈ L2(0, T ; H01(Ω)), d
dt|f |2 = 1
2h∂tf (t), f (t)iH−1(Ω)⊗H01(Ω). If we have ∂tu ∈ L2(0, T ; V1
RN), u ∈ L2(0, T ; VRN) (u ∈ C0([0, T ]; HRN)), VRN ⊂ HRN ⊂ V0
RN and ku(t)k2L2(RN)− ku(0)k2L2(RN)=
ˆ T 0
h∂tu|T, u|TiV0
RN⊗VRNdτ.
Look at (4) and define a projection P : L2(RN) → HRN, L2(RN) = HRN {dF ∈ L2(RN) : F ∈ Wloc1,2(RN)}.
| h−∆u, ϕi | = | ˆ
RN
∇u : ∇ϕ| ≤ k∇ukL2(RN)kϕkV
RN, ∀ϕ ∈ VRN, which implies
k(−∆)u|tkV0
RN ≤ k∇u|tkL2(RN)
and
k(−∆)ukL2(0,T ;V0
RN)≤ kukL2(0,T ;V
RN)< ∞.
Moreover, in 2D, we have
| hu · ∇u|t, ϕi = | ˆ
R2
uαuβ∂αϕβ|
≤ ˆ
R2
|u|2|∇ϕ| ≤ ( ˆ
R2
|u|4)12kϕkV
RN.
Note that we have used the Ladyzhenskaya’s inequality, which states ∀f ∈ H01(R2),
kf kL4(R2)≤ C0kf kL122(R2)k∇f kL122(R2), so we have
ku · ∇ukV0
RN ≤ ku(t)k2L4(R2)≤ C0ku(t)kL2(R2)k∇u(t)kL2(R2),
and ˆ T
0
ku · ∇ukV0
RNdt ≤ C0
ˆ T 0
ku(t)kL2(R2)k∇u(t)kL2(R2).
3.2 More weak solutions for the NS equations
In 1980’s, suitable weak solutions which enable us to “localize” the NS equation.
Definition 3.3. (Version 1) A suitable weak solution to (NS) which arises from a ∈ H1
R3 is an element u ∈ L∞(0, T ; HR3)∩L2(0, T ; VR3) which satisfies condition D to H, and for a.e. t ∈ [0, T ], ϕ ∈ Cc∞((0, T ] × R3), we have
1 2 ˆ
R3
ϕ(t, x)|u(t, x)|2+ ˆ t
0
ˆ
R3
|∇u(t, x)|2ϕ(t, x)dxdy
≤1 2
ˆ t 0
ˆ
R3
(∂tϕ + ∆ϕ)|u|2dxdt + ˆ t
0
ˆ
R3
∇ϕ · u(1
2|u|2+ P )dxdt.
We do not know if each Leray-Hopf weak solutions will satisfy this definition.
Remark 3.4. It is a fact that for each initial data a ∈ HR3, there is at least one suitable weak solution to (4) which arises from a.
Key lemma of the Partial Regularity theory.
Lemma 3.5. (Vasseur) For each p > 1, there exists 0 = 0(p) ∈ (0, 1) such that for any suitable weak solution (u, P ) in [−1, 1]×B(1), we have the following implication: If
sup
t∈[−1,1]
ku(t)k2L2(B(1))+ ˆ 1
−1
ˆ
B(1)
|∇u|2dxdt + ( ˆ 1
−1
[ ˆ
B(1)
|P |dx]pdt)2p < 0,
then
|u| ≤ 1 in [−1
2, 1] × B(1 2).
Remark 3.6. For CKN-version: If kukL3([−1,1]×B(1))+ kP k
L32([−1,1]×(B(1))< 0, then
|u| ≤ 1 in [−1
2, 1] × B(1).
From the (4), take divergence on both sides, we have
−∆P = RαRβ{uαuβ},
where R` stands for the Riesz transformation. When u ∈ L3, i.e., |u|2 ∈ L32([−1, 0] × R3), then P ∈ L32([−1, 0] × R3).
Now, (u, P ) satisfies the following: ∀ϕ ∈ Cc∞([−1, 1]×B(1)), a.e. t ∈ (−1, 1], then
1 2
ˆ
B(1)
ϕ(t, ·)|u(t, ·)|2+ ˆ t
−1
ˆ
B(1)
|∇u|2ϕ
≤1 2
ˆ t
−1
(∂tϕ + ∆ϕ)|u|2+ ˆ t
−1
ˆ
B(1)
u · ∇ϕ{1
2|u|2+ P }. (5)
3.3 Do truncation to |u|
vk = [|u|−(1− 1
2k)]+, Bk = B(1+ 1
23k), Bk−1
3 = B(1+ 2
23k), Bk−2
3 = B(1+ 22 23k).
Take a bump function ηk ∈ Cc∞(Bk−1
3) such that χBk ≤ ηk ≤ χBk− 1
3
,
|∇ηk| ≤ 2 · 23kχB
k− 13
,
|∇2ηk| ≤ 4 · 26kχBk− 1
3
.
Find the inequality satisfied by vk, and (5) can be abstractly phrased as follows:
∂t
|u|2
2 + |∇u|2− ∆(|u|2 2 ) +1
2∇ · {u|u|2} + u · ∇P ≤ 0.
Note that ∂tvk2= vk
|u|u · ∂tu. You know somehow vk should satisfy 1
2∂tv2k+ D2k− ∆(v2k 2 ) +1
2∇ · {u · vk2} + (vk
|u|− 1)u · ∇P + u · ∇P ≤ 0, but this is not the right way to obtain the inequality. The problem shows up on the big part of vk.
3.4 Look at ( v
k|u| − 1)u
1. If vk> 0, then (vk
|u|− 1) = −(1 −21k)
|u| |u|.
2. If vk(x) = 0, |u| ≤ 1 − 1 2k(vk
|u| − 1)u ∈ H1(B(1)) ∩ L∞(B(1)) and |vk
|u|− 1| · |u| ≤ 1 − 1
2k < 1.
Note that the idea is vk
|u| and u themselves maybe worse, but combine them together and after canceling out, the term (vk
|u|− 1)u · ηk will be nice. We use for each t ∈ [−1, 1], (vk
|u|− 1)u · ηk to test against (4) 0 =
(∂tu − ∆u + u · ∇u + ∇P )|t, (vk
|u|− 1)u · ηk
H−1(B(1))⊗H10(B(1))
,
the weak formulation of (4) matters. The idea is let 1
2∇ · {u|u|2} handle the large part of |vk|2 for us.
u · ∇u, (vk
|u|− 1)uηk
= ˆ
B(1)
uα∂αuβ(vk
|u|− 1)uβ· ηk
= ˆ
B(1)
uα· ∂∂{vk2− |u|2 2 } · ηk
= ˆ
B(1)
∇ · {u
2 · (vk2− |u|2)}ηk, since
∂α{v2k− |u|2
2 } = vk∂αvk− u · ∇u = vk∂α|u| − u · ∇u
= vk
u
|u|∇u − u · ∇u = (vk
|u|− 1)u · ∇u,
|∇(vk2− |u|2
2 ) ≤ |∇u| and vk2− |u|2
2 ∈ W1,2(B(1)).
Now, look at ∆u:
−∆u, (vk
|u|− 1)u · ηk
H1(B(1))⊗H10(B(1))
= ˆ
B(1)
∇u : ∇{(vk
|u|− 1)u · ηk}
= ˆ
B(1)
∇u : ∇((vk
|u|− 1)u) · ηk+ ˆ
B(1)
∇u : (vk
|u|− 1)u∇ηk:= I1+ I2. For I2,
I2= ˆ
B(1)
∂αuβ(vk
|u|− 1)uβ∂αηk = ˆ
B(1)
1
2∂α(vk2− |u|2)∂αηk
= −
∆(vk2− |u|2 2 ), ηk
. For I1, note that ∇|u|2= 2u · ∇u = 2|u|∇|u|
I1= ˆ
B(1)
∇u : ∇((vk
|u|− 1)u) · ηk= ˆ
B(1)
∇u : ∇(vk
|u|u) − |∇u|2
= ˆ
B(1)
∇u : ∇(vk
|u|) · u + ∇u : vk
|u|∇u − |∇u|2
= ∇u∇(vk
|u|) · u + (vk
|u|− 1)|∇u|2
= |u|∇|u|2· 1 −21k
|u|2 ∇|u|χ{vk>0}+ (vk
|u|− 1)|∇u|2
= [1 − 21k
|u| ∇|u|2χ{vk>0}+ (vk
|u|− 1)|∇u|2].
Note that vk = [|u| − (1 − 1
2k)]+ and Dk2 = vk
|u||∇|u||2χ{vk>0} + vk
|u||∇u|2. u = (1 − vk
|u|)u + vk
|u|u and |u|vk2
2 ≤ (1 − 1 2k)vk2
2 +v3k 2 .
3.5 Short summary
∂t(v2k− |u|2
2 ) + ∇ · (u
2(vk2− |u|2)) + D2k− |∇u|2+ (vk
|u|− 1)u · ∇P ≡ 0 (6) hold in D0, whichever
∂t(|u|2
2 ) + |∇u|2−1
2|∆(|u|2) + ∇ · (1
2u|u|2) + u · ∇P ≤ 0 (7) which is (5) to test against nonnegative test functions. From (6) and (7), we obtain
∂t(vk2
2) + D2k− ∆(v2k
2 ) + ∇ · (u
2 − v2k) + (vk
|u|− 1)u · ∇P + u · ∇P ≤ 0. (8) Define
Uk := sup
Tk≤t≤1
kvk(t)k2L2(Bk)+ ˆ 1
Tk
ˆ
Bk
Dk2,
by choosing Tk = −1 2(1 + 1
2k) and derive Uk ≤ C0k[1 + kP kLp
tL1x]Uk−1β to imply the first DeGiorgi lemma holding.
For ϕ ≥ 0 ∈ Cc∞((−1, 1) × B(1)) works for (7) but not for (8). For t > 0, 1
2 ˆ
B(1)
v2k(t, ·)ϕ(t, ·) + ˆ 1
−1
ˆ
B(1)
d2kϕ
≤ ˆ 1
−1
ˆ
B(1)
(∂tϕ + ∆ϕ)vk2+ ˆ 1
−1
ˆ
B(1)
u 2vk2· ∇ϕ +
ˆ 1
−1
ˆ
B(1)
[(vk
|u|− 1)u · ∇P + u∇P ] · ϕdt.
In the final step, let ψ(t) be a bump function with support in (σ − , s + ) (σ < 0, s > 0) and ψ(t) = 1 in (σ + , s − ), and connect 1 and 0 by straight lines. Plug the test function ϕ = ψ(t) · ηk(x) into previous calculations, then we have
ˆ σ+
σ−
ˆ
B(1)
ψ(t)ηk(x)Dk2(t, x)dxd
≤ ˆ 1
−1
ψ0(t) · ˆ
B(1)
ηk(x)vk2(t, x)dxdy + ˆ 1
−1
ˆ
B(1)
ψ(t)∆ηk(x)vk2(t, x)dxdt
+ ˆ σ+
σ−
ψ(t) ˆ
Bk− 13
u
2v2k∇ηk+ ˆ s+
σ−
ˆ
Bk− 13
[(vk
|u|− 1)u · ∇P + u · ∇P ]ψ(t)ηk(x),
and ˆ 1
−1
ψ0(t) · ˆ
B(1)
ηk(x)vk2(t, x)dxdt
= − 1 2
ˆ s+
s−
ˆ
B(1)
ηk(x)u2k(t, x)dxdt + 1 2
ˆ σ+
σ−
ˆ
B(1)
ηk(x)v2kdxdt.
Until now, we have
∂t
vk2
2 + D2k− ∆(vk2
2 ) + ∇ · (uv2k 2 ) + (vk
|u|− 1)u · ∇P + ∇ · (uP ) ≤ 0 in (−1, 1) × B(1).
Select any −1 < σ < t < 1, ˆ
B(1)
ηk|vk(t, ·)|2+ ˆ t
σ
ˆ
B(1)
Dk2· ηkdxdt
≤ ˆ
B(1)
ηk|vk(σ, ·)|2+ ˆ t
σ
v2k 2 ∆ηk+
ˆ t σ
ˆ
B(1)
u ·v2k 2 ∇ηk
+ ˆ t
σ
| ˆ
B(1)
ηk{(vk
|u|− 1)u · ∇P + ∇ · (uP )}dt.
Fix t ∈ [Tk, 1], take average of the above inequality over σ ∈ [Tk−1, Tk], ˆ
B(1)
ηk|vk(t)|2+ ˆ t
Tk
ˆ
B(1)
Dk2ηk≤ Aerage of LHS ≤ Average of LHS
≤ 1
Tk− Tk−1
ˆ Tk Tk−1
ηk|vk(σ, ·)|2dσ + ˆ t
Tk−1
ˆ
B(1)
v2k
2 |∆ηk| + |u|v2k 2 |∇ηk| +
ˆ t Tk−1
| ˆ
B(1)
ηk{(vk
|u|− 1)u · ∇P + ∇ · (uP )|dτ.
Use |∇ηk| ≤ 2 · 23k and |∇2uk| ≤ 426k, then 1
2Uk≤ sup
Tk≤t≤1
ˆ
B(1)
ηk|vk(t)|2
2 +
ˆ 1 Tk
ˆ
B(1)
D2kηk
≤ 2k ˆ 1
Tk
ˆ
BK− 1
3
|vk|2+ ˆ
Qk−1
(4 · 2σk)[|vk|2+ |u|vk2 2 ] +
ˆ t Tk−1
| ˆ
B(1)
ηk{(vk
|u|− 1)u · ∇P + ∇ · (uP )|dτ.
u ∈ L∞(−1, 1; L2(B(1)) ∩ L2(−1, 1; H1(B(1)), vk ∈ L∞(Tk, 1; L2(Bk)). Note that L2(Tk, 1; H1(B(1)) ⊂ L2(Tk, 1; L6(B(1)).
kvkkL6(Bk)≤ C0{kvkkL2(Bk)+ k∇vkkL2(Bk)}, B(1
2) ⊂ Bk⊂ B(1).
For θ ∈ (0, 1), 1 p= θ
2+1 − θ
∞ = θ
6+1 − θ
2 , implying p = 10
3 . Therefore, kvkk
L103 (Qk)≤ kvkk1−θL∞(Tk,1;L2(Bk)· kvkkθL2(Tk,1;H1(Bk))≤ Uk12.
4 Deal with the pressure
Motivation: For Leray Hopf weak solution in R3, −∆P = ∂α∂β{uαuβ}, u ∈ L∞(−1, 1; L2(R3)), |u|(t, ·) ∈ L1(R3) and |u|(t, ·) ∈ L6(R3).
Lemma 4.1. Suppose we have P ∈ Lp(Tk−1, 1; L2(Bk−1)) and some symmetric matrix (Gαβ) ∈ L∞(Tk−1, 1; L1(Bk−1)) such that (−∆)P = ∂α∂β{Gαβ} (Not the whole space setting). Then
P |B
k− 13
= Pk1+ Pk2,
where Pk1 is the nonlocal part and Pk2 = (−∆)−1∂α∂β[Gαβ· ϕk] is the local part. Moreover,
kPk1kLp(Tk−1,1;L∞(Bk− 1
3
))+ k∇Pk1kLp(Tk−1,1;L∞(Bk− 1
3
)). Observe that
−∆(ϕkP ) = −∆ϕk· P − 2∇ϕk· ∇P − ∆P · ϕk
= −P ∆ϕk− 2∇ · {P ∇ϕk} + 2ϕk∆P − ϕk∆P
= −P ∆ϕk− 2∇ · {P ∇ϕk} + ϕk∆P and
−ϕk∆P = ϕk∂α∂β{Gαβ}
= ∂α{ϕk∂βGαβ} − ∂αϕk· ∂βGαβ
= ∂α∂β[ϕkGαβ] − ∂α{∂βϕk· Gαβ} − ∂αϕk∂βGαβ
= (−∆)Pk2− ∂αβϕk· Gαβ− 2∂βϕk· ∂αGαβ, or
(−∆)[ϕkP ] = (−∆)Pk2− ∂αβϕk· Gαβ− 2∂βϕk· ∂αGαβ
− ∆ϕk· P − 2∇ · {P · ∇ϕk}.
Further,
(−∆)Pk1= 1 4π
1
|x|∗ {−∂αβϕk· Gαβ− 2∂βϕk∂αGαβ
− ∆ϕk· P − 2∇ · {P ∇ϕk}.
Since supp∇ϕk, supp∇2ϕk⊂ Bk−1\Bk−2
3, (−∆)Pk1= 0 in Bk−2
3, we estimate kPk1kL∞(Bk− 1
3
) and k∇Pk1kL∞(Bk− 1
3
). For x ∈ Bk−1
3, ˆ
R3
1
|x − y|∇ · {P ∇ϕk}|ydy = ˆ
y∈Bk−Bk− 2
3
1
|x − y|∇y· {P ∇ϕk}dy
= − ˆ
Bk−1\Bk− 2
3
1
|y − x|2( y − x
|y − x|)P (y)∇ϕk|ydy
≤ 2 · 26k· 23k ˆ
Bk−1
|P |.
The conclusion is that
k∇Pk1(t, ·)kL∞(Bk− 1
3
)+ kPk1(t, ·)kL∞(Bk− 1
3
)
≤C0212k ˆ
B(1)
{|{P (t, ·)| + |G|}.
Lemma 4.2. Let p > 1, consider P ∈ Lp(Tk−1, 1; L1(Bk−1)) and (Gαβ) ∈ L∞(Tk−1, 1; L1(Bk−1)) such that (−∆P ) = ∂αβGαβ. Then
P |B
k− 13
= Pk1+ Pk2,
where Pk2= (−∆)∂αβ(Gαβϕk) and Pk1 satisfies the following (−∆)Pk1= 0 in [Tk−1, 1] × Bk−2
3
and
kPk1kLp(Tk−1,1;L∞(B
k− 13
))+ k∇Pk1kLp(Tk−1,1;L∞(B
k− 13
))
≤C0224k{kP kLp(Tk−1,1;L1(Bk−1))+ kGkLp(Tk−1,1;L1(Bk−1))}.
The details of proof, we refer to Seregin’s big paper and Sverak’s work.