Professional Documents
Culture Documents
November 6, 2011
1 (H0 )
Contents
Contents 1 Linear & Nonlinear Static NSE 1.1 Physical properties remarks . . . . . . . . . 1.2 Derivation . . . . . . . . . . . . . . . . . . . . 1.3 Shear ow . . . . . . . . . . . . . . . . . . . . 1.4 Steady-state Stokes - Overview . . . . . . . 1.5 Steady-Stokes - Variational form . . . . . . 1.6 W k,p , H k , H 1 2 . . . . . . . . . . . . . . . . . 1.7 The Auxiliary space E() . . . . . . . . . . 1.8 Extension map . . . . . . . . . . . . . . . . . 1.9 Trace and Stokes formula . . . . . . . . . . . 1.10 Preliminaries for Stokes Equation Existence 1.11 Leray projection operator . . . . . . . . . . . 1.12 Existence for Stokes Equation . . . . . . . . 1.13 Energy approach . . . . . . . . . . . . . . . . 1.14 Energy Method proof. . . . . . . . . . . . . . 1.15 Existence for Inhomogeous Stokes . . . . . . 1.16 Regularity . . . . . . . . . . . . . . . . . . . . 1.17 Navier-Stokes Existence . . . . . . . . . . . . 1.18 Existence - Galerkin . . . . . . . . . . . . . . 1
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
CONTENTS 1.19 1.20 1.21 1.22 1.23 Variational Form for Inhomogeneous Navier-Stokes Uniqueness for Inhomogeneous Navier-Stokes . . . . Inhomogeneous NS Uniqueness . . . . . . . . . . . . Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . Regularity for NSE for n = 2, 3, 4, 5, 6, 7 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 32 33 34 38 43 55 55 56 59 62 66 66 78
2 Evolution NSE 2.1 Stokes operator . . . 2.2 Function spaces . . . 2.3 Galerkin . . . . . . . 2.4 Pass to limits . . . . 2.5 Uniqueness . . . . . 2.6 Stability Properties 2.7 Uniqueness . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
Denn nichts ist fur den Menschen als Menschen etwas wert, was er nicht mit Leidenschaft tun kann.
1.1
These equations establish that changes in momentum (acceleration) of uid particles are simply the product of changes in pressure and dissipative viscous forces (similar to friction) acting inside the uid. These viscous forces originate in molecular interactions and dictate how sticky (viscous) a uid is. Thus, the Navier-Stokes equations are a dynamical statement of the balance of forces acting at any given region of the uid.
1.2
Derivation
F = Fint =
(ij ) = =
1 3
normalize
p 0 0 0 p 0 +T 0 0 p
tensor; material dependent
1 (pI3 ) + 3
gravity
u u + y y
where is proportional to viscosity, which varies with temperature. For u = (u1 , u2 , u3 ) and p constant, ut + u u = pI3 + (u + (u)T ) ??? u
incompressibility
=0
Calculating, (u + (u)T ) or = ui + uj xi xj ui uj + xj xj xi
take transpose div = 0 uid incompressible (NSE)
ut + u u + p = u.
This is formally a coupled system in four unknowns (u1 , u2 , u3 ), p and 2 eqns, and is therefore determinable. To uncouple, write (?) Difculty: p is presumably a
function of 4 variables, but my eqn doesnt tell me; this makes p difcult to handle.For uid have boundary conds., but for p initial condition
p = div (u )u
elliptic eqn
1.3
Shear ow
An element of solid has a preferred shape, to which it relaxes when the external forces on it are withdrawn (whereas a uid does not). Looking at a rectangular element ABCD, under the action of a shear force F , the element assumes the shape ABC D . If the solid is per-
Figure 1.2: Deformation of solid (top) and uid elements (bottom) fectly elastic it goes back to its preferred shape ABCD when F is withdrawn.
In contrast a uid deforms continuously under the action of a shear force (however small). Thus the element of the uid ABCD conned between parallel plates deforms to shapes such as ABC D and ABC D as long as the force F is maintained on the upper place.
du dy
1.4
u
t=0 =
u0 , u0 = 0. In NSE (or just Stokes) the following boundary conditions are commonplace:
u = 0 no slip boundary - uid will have zero velocity relative to the boundary.
u=0 or ( u) v = 0
vorticity - here no tang. component
A. Steady-State Stokes
u + p = 0 u=0 u = 0
drop t. p is harmonic,so p can be wild on boundary we have this from h o l o m o r p h i c functions
B. Evolution Stokes
ut u + p = 0 u=0 u t=0 = u0 , u = 0
C. Steady NSE
u u u + p = 0 u=0 u = 0
1.5
Compare this to Lax-Milgram-type arguments for elliptic eqns. In fact, NS is almost elliptic:
but have this termso we must nd a different solution space.
u
2nd order function with divergent form
=0
+ p, = 0
u + p, = 0,
div (p) p ( )
trace=0
so
u,
p, = 0
Reverse direction:
u, = 0 V
or min u
2
uV .
The space V = Cc () div = 0 ) is not a complete Vector Space ! ! ! We need to complete this So how will we do this? depends 1 on the space we want to close, i.e. H0 , L2 .
Monday, August 29
Function spaces
Once we have the variational formulation of u + p = 0 in u = 0 on we have to set up the corresponding function space. The usual Sobolev space is decient because it requires information about u, whereas here only u is known. In the following, we consider only vector elds that are divergence-free.
1.6
W k,p, H k , H 1 2
The Inner Product (, ) and the Pairing ,
(u, v) we can use the L2 inner product for two functions. u, v use for those arguments that arent functions, for example (u, v) . W k,p . Dene W k,p () = f Lp () D df Lp () k, k 0, 1 p f where W k,p (),
W k,p W k,p
=
k
D f
p Lp
1 p
is a Banach space.
Density theorem
We denote the space C () as the space of smooth functions up to the boundary. For p = 2, 1 p < , this space is dense in W k,p (). k,p Note that Cc W k,p = W0 and u = D u = 0 k 1 on . ?
H and V
1 For u Cc (), u = 0 H0 L2 , dene H = {u Cc u = 0}. This space is not closed. After mollifying, dene
10
1 V = u Cc () u = 0 = H0 Here u H0 and u = 0 preserve their denitions. H = u Cc u = 0 and p H 1 () = L2 Here we have uk = Cc and u L2 uk Cc and uk = 0.
What is u = 0 here? In this space, u is = 0, remember though we are taking distributional derivatives. Since u = 0, can we talk about u . If u is a L2 function whose distributional derivative is zero, then do we have enough information? No: If only u L2 , then the boundary cant be dened. We have partial information with u . This expression is well dened when u H 1 2 ()
1.7
Dene the auxiliary space E() = u L2 () Then H 1 E() L2 . Dene (u, v)E() = and u
E()
L2 () .
u + (div u)(div v)
(u, u)E() .
Note that (E(), )E() is Hilbert (check). Density theorem for E(). Assume C 0,1 . Then
Cc ()
this function is onto but still not 1-1: 1 ker 0 = H0 () = whole space
is dense in E().
11
Proof. (. . . ) (see s.31) Trace theorem for E() . A trace theorem allows us to dene u for u E(). To do this, we ask where the trace operator lives. The denition 0 H 1 () L2 ()
has the deciency that not every L2 function can be the trace of a H 1 function: this mapping is not onto. Replace this denition with 0 H 1 () H 2 () Dene H s (Rn ) f + Consider
1 f (x) H 2 (Rn1 ) = f (x, (x)) = Rn1 R 1
f L2 = f Rn R
1+
s 2
f L2 .
D 2 . cant dene fractional derivatives pointwise. use fourier transform this is for at case. not if domain is nonat. If have Lipschitz graph = (x, (x) )
Lipschitz
x Rn1
where f H (). locally Lipschitz Use Partition of unity and dene globally.
1 2
1.8
Extension map
1
Dene H 2 () H 1 (). Find f F by solving F = 0 F = f. The solution of this Laplace equation gives F , the denition of the extension map F . Note that f and F are reciprocals of each other: F
H1 n
c f =
H2
0 = id
12
1.9
For C 1,1 there exists a linear continuous operator E() H 2 () require the (stronger) C 1,1 regularity here and 1 1 dual space. H 2 () = H 2 () Furthermore the following hold: 1. u = u u C 1 ().
div (u) + u u = (u ) .
0 restrict
3. ker( ) = E0 () and
E0 () = Cc ()E()
For Sobolev spaces we had that the whole component on the boundary was zero; here, we have no information about the tangent component, just the normal component. For the function u the normal component is well dened, but grad u is not. Dx1 , Dx2 , . . . are not dened.
which I need for the matrix representing the tangential components.
13
Friday, September 2
1.10 Preliminaries for Stokes Equation Existence
L2 () = H H1 H2 where H = u L2 () div u = 0, = 0
normal trace
Hodge Decomposition
H1 = u L2 () u = grad p, p H 1 (), p = 0
1 H2 = u L2 () u = grad p, p H0 ()
Stokes theorem can be relaxed, 1 only the gradient need be in H0 with the space E().
(1.1) (1.2)
(u, v) = (
q)
1 well dened, q H0
+ (div p, q)
= (0 (p), 0 q)
0 Eq 1.2
=0
14
p = div u in p = 0 on By Lax-Milgram, there exists a solution. div (p u) = 0 u = (up) + p Claim: The term p u is divergence-free but perhaps not trace free ( 0). To see this, write
div -free
u p E()
(up) H 1 2 ()
well-dened
Let q H 1 solve q = 0 q = (u p) where must satisfy (u p), 1 = 0 (1.3) (1.4) div (up) (the
=0
above equation is solvable up to a constant. We write u H 1 R to denote this (modulo the constant). Apply Lax-Milgram to solve (1.3,4) . Thus
H 1
Note: When applying LaxMilgram with Neumann BC you have to be careful (?)
u = u p q + q + p
H H2
15
1.11
which has the apriori estimate Du L2 u L2 where u u p q. The Projection operator is used to interpret Stokes: u + p = f in div u = 0 in u = 0 on . Apply op on both sides of the equation:
L2
Pu + P(p) = P f
H 2 L2
Pu is Stokes operator, denoted by A. Later: Mild solutions Kato, Giga, Fujita, (cf. Du Hamel form, where we use P to kill pressure p). Avantage: P eliminates the pressure term p. Disadvantage:
or
Au = f
1.12
Rn is bounded Lipschitz
CHAPTER 1. LINEAR & NONLINEAR STATIC NSE Use Leray projection operator to make p drop out: u + p, v = f, v
int. by parts
16
v V
u v
= (u, u)V .
We look for u V that satises (u, v)V = (f , v) v V where f V and f H 1 . This implies (u + f, v) = 0 v V.
D () L2
+ f = p
H 1
for some p D ()
Summarize:
Fact u V solves (u, v)V = (f, v) for f V if and only if u + p = f in div u = 0 in u = 0 on for some p L2
V is read in the dual V
17
1.13
Energy approach
2
cf. Laplace Energy functional: Dene the Energy functional E(u) = u V 2(f , u). 2 1 The variational problem u V solves (u, v)V = (f, v) for v V . is u = f 2 u f , where we minimize the functional over equivalent to
1 H0 ; Here we minimize over V .
18
Wednesday, September 7
Existence Regularity
There exists a unique u V and unique p L2 () R solving homogeneous system with estimate u
V
+ p
L2 R
H 1
u, v V. uV vV
2 V
boundedness coersivity
a(u, v) u
for some p D
p L2 p + f + u H 1
19
1.14
The energy functional E is given by E(u) = u L (f , v) v V . Claim u V solves (*) if and only if E(u) = minvV E(v). We know the minimization is always obtained. Pf. (u, v)L2 = (f , v) v V . Let v = u w where w V. Then (u, u)L2 = (f , u) (f , w) + (u, w)L2 Using that
cs
u 2
2 L2
w 2
2 L2
(u, u)L2 = (f , u) (fw) + (u, w)L2 2 (f , u) (f , w) + u L2 + w 2 2 1 2 E(u) = u L2 (fu) 2 2 1 2 w L2 (f , w) = E(v) 2 2 0= d E(u + tv) = 2(u, v)L2 2(f, v). dt
2 L2
(1.5)
20
f constant L2 () R means (f, 1)L2 = f 1 = f = 0
1.15
g=
u=
+ p
L2
H 1
+ g
L2
+ f
H2
1 Remember: A function f H0 2 div f L
The approach for the inhomogeneous is cook up new function that is divergence-free and has zero trace. Note that f changes.
(div u0 g ) = g
Compatibility Condition
g=0
Somehow, u0 , u1 kills g.
CHAPTER 1. LINEAR & NONLINEAR STATIC NSE Trace 0 (V ) = 0 (u) 0 (u0 ) 0 (u1 ) = = 0. The Eq (IS) in v becomes v = p = (u u0 u1 ) + p = f + ( u0 + u1 ) = f
H 1 H 1 H 1
21
L2 ()
f =0
(1.6)
Not 1-1: add divergence-free function and we still have = f (no uniqueness). The term p: The gradient operator L2 () H 1 After taking slash out constants so we can have f + c = f . out constants, this space is 1-1. Adjoint: , g = f, g
integrate by parts
(1.7)
1.16
Regularity
If Rn is bounded and C m+2 and u W2, , p W 1, , 1 < < +. For u W2, , u is pointwise dened and p is well dened If addi- strong solution instead of weak 1 solution. tionally f Wm, , g Wm+1 , and () W m+2 (), u Wm+2, (), p W m+1,
22
Wm+2,
Wm+1, Wm+1
Wm,
+ g
Wm+2 ()
+ u
L2
Question: Why do we have u W2, ? Ans: If f W,1 ?? Theorem 3 If f L2 , g H 1 , H 3 2 , then u W2,2 p W 1,2 .
Why? Suppose g = = 0 for the moment and = 0 u + p = f u=0 u=0 For this eq 1.16, which has only two terms on the lhs, f L2 u L2 then automatically p W 1, . We must show, using nite quotients, that (u, v) = (f , v) Approach: Use the nite dierence quotient as test function, but have to cut out.
()
u(x+he)u(x) H1 use u h uniformly with respect to h. D2 u L2 .
(1.8)
23
Friday, September 9
regularity for n = 2.
Theorem 4 (regularity for n = 2) 1 For f Wm, , g W m+1, , Wm+2 , (), C m+2 and g = Remember that (B) and (C) be then u + p = f in divu = g in (B) u = on (C) has unique solution (modulo constant) such that u Wm+2, , p Wm+1, and
1
Wm+2,
+ p
W m+1, Wm+
1
W m,
+ g
Wm+2 ,
+ u
L2
Approach: We nd auxillary equationto kill (B),(C). In particular: nd gradient potential kill pressure p get formally biharmonic eq coupled 2nd order (like 2nd order Elliptic equation) for which Sobolev W k,p theory holds. Claim v Wm+2, such that divv = g in v = on (Prove later) If Claim 1 holds, then w = u v where w satises ( )
w + p = f + v = f divw =0 w =0
24
Case 1: is 1-connected
divw = 0 can nd can nd extr in 1-form that is closed dW = 0. 1 homotopy cohomology is exact translated such that = , x2 x1 Condition (B)
1 f
rst component of the vector f
2 (D2
W m1,
(1.9)
Note that 1.9, an equation in does not involve p. Further, it is a rst order dierential operator lost 1-order regularity. dV = 0 =0 tangent derivative =0 normal derivative on
For the second order operator need a second condition; this problem as it stands is not closed tangent gradient = 0
Note that the potential is constant, but because is not unique, pick = 0 on u = f Wm,2 u=0 u Wm+2,2
Using nite quotients and integrate by parts, get 2 more derivative estimate. This is possible = 2 For u Wm+2, , Calderon-Zygmund theory of
harmonic analysis does this for higher order .
25
Figure 1.4: Diculty: cannot shrink closed paths to 0. The 1-cohomology of the domain is like copies H 1 () = Zk have k freedoms cannot not specify constants ci 2 u = f u Wm+2, u =0 Wm1+4, w Wm+2, From p
pW m+1,
= + W m, . f
Case 2: is k-connected
Suppose 2-connected for simplicity. We cut k (here: 2) times to make 1-connected
k . R such 1ik
26
Figure 1.5: is dened except on cut. For to be welldened there too extend make well-dened trace For the Multi-Connected case, equation same as in single connected case. The biharmonic equation is uniquely solable; but the dirichlet + neumann boundary conditions are dicult ? Rest of argument is exactly same.
For x
+ = + (x0 ) + d.
Proofs of Case 1
Omitted. (see p7,8 handcopy)
27
Wednesday, September 14
For the Navier-Stokes Equations we introduction the nonlinear advection term u u, i.e. the advection operator is
u=u
+v +w . x y z
Picture advection as transport of salt dumped in a river. If the river is originally fresh water and is owing quickly, the predominant form of transport of the salt in the water will be advective, as the water ow itself would transport the salt. If the river were not owing, the salt would simply disperse outwards from its source in a diusive manner, which is not advection. Very few exact solutions of NSE are known in closed form. In general, exact solutions are possible only when the nonlinear terms vanish identically.
1.17
Navier-Stokes Existence
(N S)
u + u u p = f in divu = 0 in u = 0 on .
L2
Variation form: ((u, v)) + B[u, u, v] = (f, v) v V B[u, vw] = u v w bigger space
2n L n2
u, v V, w V
bounded
We to to extend to the
Ln
V = V.
28
= u
1 H0
+ u
Ln
1 for n = 2, H0 Ln = V .
L2
L2
V
1 H0
replaces
with Ln
c(n) u
1 H0
V =V
2n 1 but in general H0 Ln V n2 n 2 n 2 n 4 In particular, B V V V R is well dened for 2 n 4; for m 5, V V B is not well-dened. We need to understand trilinear form better; It can induce bilinear form B[u, v] V V R replace Lemma
with V
u v v = ui i v
2 2
ui i v j v j i uj v2 =0 2
ibp
29
= 0.
1.18
Existence - Galerkin
((u, v)) + B[u, u, v] = (fv) v V .
replace with V . Let {wk }k=1 be an orthonormal basis (ONB) of L2 () 2 is innite-dimensional vector space cut o nite like The space L n have family of solutions if we can prove apriori estimate, we can R take limit function can solve the limiting equation done. Let X m = span {w1 , . . . , wm } m 1 and let um = m am wi coeff tb determined i=1 i
((um , wj )) + B[um , um , wj ] = (f , wj ) 1 j m
1 N O T E: ONB L2 , but H0
choose wj to be basis
am + Aj am am = (f , wj ) j il i l
We ask if we can solve the term Aj am am , being nonlinear. Formally, this il i l term is not like solving Ax = B, but rather like solving Ax + BxxT = C; does solution exist maybe equal number of equations solution maybe not unique though nonlinear terms. Therefore we need to turn to the following
Theorem 5 (Brouwer) If f BR BR is continuous x0 BR such that f (x0 ) = x0 , where BR are closed balls.
30
Figure 1.7: and P () Proof (contradiction) If f (x) x x Bn F (x) = f (x)x BR ; F BR BR . F (x) = x if x B F is a continuous map and wi is xed on the boundary F BR = Id impossible by degree theorems x = 0 on in BR , but boundary is of degree 1. innite-d versions of Browder Lemma (Variation of the Brouwer Fixpoint Theorem) Let X be a nite-dimensional Hilbert space equipt with inner product [, ] and norm []; P X X is continuous (but not necessarily linear). Suppose [P ()] = k > 0 X with []k such that P () = 0. Proof If p() 0 BR = X [] k , then dene Q() = k and [Q()] = k. Now Q BR BR continuously and Brouwer 0 BR such that Q(0 ) = 0 where automatically Q(x0 ) is on the boundary and [Q(0 ), 0 ] = k. p() [p()] BR
thm very useful for existence of nonlinear pde.
31
= [k
which is impossible because 90degree Dene Pm (u) X m X m by [Pm (u), v] = ((u, v)) + B[u, u, v] (f , v). Need to verify: B[(), ] > 0. ((u, u)) (f , u) = u2 f V
f
= u
u f
>0
if u V = k where m > VV um X m with um V < k such that f pm (um ) = equivalent to am + Aj am am = (f , wj ). In fact um V VV . j il i l Since[pm (um ), um ] um um
V 2 V
(fum ) um
2 V
= (fum ) f
um
which gives the sought after apriori estimate with uniform bound (so that we can now take limits).
Challange
Write down Lax-Milgram proof using a Galerkin method approach
32
Friday, September 16
Variational Form for Inhomogeneous Navier-Stokes Uniqueness for Inhomogeneous Navier-Stokes
1.19
We show that the variational form for inhomogeous Navier-Stokes is well posed. Dene the eigenfunction {X m } of Stokes operator having zero Dirichlet m boundary conditions X m = span {wi }i=1 . Find u V such that ((u, v)) + B[u, u, v]
uuvwell dened for vV
= (f , v) v V , f V
= (f , v) v X m
(1.10) where X m V
((um , um )) = (f , um ) um
Send m .
V 1
This is called an apriori estimate lhs holds for all sequences um , & rhs is independent of the number m.
By weak compactness properties of V , bounded sequence take subsequence so sequence converges weakly:
in V um u in V.
um
u in L2
Rellich compactness um u in L2
33
((um , v)) ((u, v)), (f , v) (f , v) vacuously B[um , um , v] B[u, u, v] The last of these 1.11 is not obvious. Write
skew-sym tensor
(1.11)
B[um , um , v] =
um um v = B[um , v, um ] =
um um v
L1 L
so that =
Note: we dont want solution & test function to vary at the same time: x solution and send test function ; x test function & send solution . we have, then, that For all v V there exists vm X m so that vm v V .
1.20
!u V
Brouwer FPT is great for nonlinear you can have many xpoints; LM is only for 1 xpoint.
34
the solution is unique in a ball ? Suppose u1 , u2 unique solutions. Let w = u1 u2 . Then (ui , vi ) + B[ui , ui , v] = (f , v) replace v with w (w, v) + B[u1 , u1 , v] Note that (u1 u1 u2 u2 ) = (u1 u2 )u2 + u2 (u1 u2 ) v = wu1 v u2 wv (u, v) + B[w, u, v] + B[u2 , w, v] = 0.
2 2
w u w
V
=
switch
v u w w
L4 L4 L2
Sobolev
c(4) w 2 .
Use u
c(n) f
c(4) f
<0
2 2
0.
1.21
Inhomogeneous NS Uniqueness
f H 1 , nd u H 1 such that
not u V not 0 on .
For Rn bounded, C 2 , ()
u + u u + p = f in u = 0 in u = on .
35
Curl
i curl = ( ) = det
x 1
j
y 2
k
z 3
vanishes
n 4, = ( 1 , . . . , n ),
aijk Dj k
j,k=1
We want curl H 1 Ln . Theorem 7 at least one u H 1 and p D () solving 1.21, H 1 Ln such that div = 0 and = on Proof. Let u = u for u=0 u = 0 on In terms of u =f
solve this gives variational form
( + ) + ( + ) ( + ) + p u u u
linear rst order
+ u + + u + p = f + + u u u
difcult quadratic terms f
36
, =
well-dened
L2
Ln
1 H0
Ln
L2
1 H0
H 1 .
37
Monday, September 19
coersivity
For u p + u u = f divu = 0 u= Suppose H 2 Ln satises div = 0, where = f + H 1 . f The corresponding weak form is = 0 on , u = u + +p = f u
(( , v)) + B[ , u, v] + B[ , , v] + B[, u, v] = (, v) v V u u u f We need the coersivity estimate for LM: For > 0 (( , u)) + B[ , u, u] + B[ , , u] + B[, u, u] u u u u 2 ((u, u)) + B[ , , u] u . u It suces to show that B[ , , u] u 2 u . 2
2
Sobolev embedding
B[ , , u] = B[ , u, ] = u u
u D u u
L n2
2n
D u
L2
L2
c(n) u
Ln
38
1.22
Lemmas
Ln
c(n) u
Ln
How: We introduce the function = curl indirectly: we can cut potential, but cant cut curl:
not here
cannot chop o
= curl
cut here
If we cut , the divergence-free condition breaks down! Additionally H 2 W 1,n L and = curl ; H 2 H 1 2 .
bounded
curl () Ln + Ln , c + Ln (2 ) Ln (2 ) Lemma 1 Let (x) = dist (x, ) > 0 there exists C 2 () such that 2 = 1 if (x) = 0 if (x) , = exp 1 (x) if (x)
39
from Proof Dene = (x) . If C 2 0 = 0 () > 0 2 ( ) differential geometry such that C 2 t t 2 Let = log t t 2 Then ? (x) = ((x)); = ((x))(x) by chain rule. 0 t Why is 1? Since
So (?) f +
Ln Ln
curl (2 )
Ln
We need to estimate n . Be careful: is small near by use L Poincaire inequality near the boundary (Poincaire inequality with weight)
v2
2
c1
1 v H0
!
This is like a Poincaire inequality with weight. Cf: v2 c1 (v2 )
B Br
0 0
v2 c {} 2
40
We want to estimate
For
v2 . {}
{>}
1 2
v2
f g =
0 {g=t}
f dH n1 dt
where f L1 , g Lip .
Like a curvilinear Fubini For level surfaces For good Lipschitz surfaces, this is just Fubini; for curvilinear its Federer.
Then
41
v2 = {}
ibp
1 t2
{=t}
v2 d dt =
0
0 {=t}
v2 d d 1 2
1 t
0
1 t
v2
=t
1d t dt
=t
v2 dt =
v2 +
1 t
=t
v v d dt
1 t
v
=t
v
1 2
dt
0 =t 2
=2
1 t2
v ddt
=t
v ddt
1 2
v2 2 2
v2 2
1 2
= { }.
42
Wednesday, September 21
regularity estimate
(Last time) v2 c v
1 v H0 .
(x) = 1 x
v2
B1
1 x
B1
lhs =
1 0
1 (1 r)
2 Br
v2 d dr
u u
L2
L2
uw
L2
= u ui j
boundary data has curl extension; cut o interior part: = curl ( ) H 2 W 1,n L and Use to cut o u + u
L2
L2
43
Figure 1.9: the bump function set = x (x) . This is the tubular neightborhood . Has support in annular region; has support near boundary. u
max est
L2 ( )
+ u
L2 ( )
: we restrict on = exp 1 ; L2 .
Use
u
1 2
+ u
L2 ( ) 1 1 n + 2n
n2
2n L n2
Ln ( )
Ln ( )
Now we can make the coecient as small as possible by absolute continuity: lim0 Ln ( ) = 0.
1.23
n=2 n=3 n = 4 critical - papers n = 5 supercritical - this dimension is the heart of the matter n = 6 solved n = 7 open problem
44
n=2
1 We begin with minimum regularity: u H0 . Assume f = 0.
u u = (u u)
sob emb (crit dim)
1 u H0 () u Lq for some q < +
u uj xj =
xj
uj u
uj u xj 0 divu=0.
as long as nite
u u u Lq
q < +
(Lq ) W 1,q
q =
q q1
Last time, if u + p W 1,q add 2nd order regularity u W1,q , p Lq q. Now go back: u u = (u u) product Lq
Lp Lp
u W2,q
p W 1,q .
Go back: (u u) = u 2 u + u u
Lq
W 1,q
45
n = 3.
u L2 u u L3 2
6 2
W 1,2 embedding
u + p L3 2
n=4
Previous argument gets stuck: W 1,2 (R4 ) uW
2, 4 3
pW
1, 4 3
This is good because we have one more derivative. Its bad because we now have lower integrability (!): u W 1, 3 (R4 )
4
L2
stop!
Idea for n = 4
Consider instead Burgers equation (ignore pressure p):
L2
u = u u 2 u = (u u) cancel 1 derivative u = (u u) u u u u
Sobolev emb
2 2 2 4
= u
2 4
c u
2 4
(1.12)
Note that the integrability does not improve, but the power in 1.12 does
CHAPTER 1. LINEAR & NONLINEAR STATIC NSE Theorem 8 Any solution u L4 (R4 ) of u = (u u) is zero u Then 1c u is not true.
L4 (R4 ) L4 (R4 )
46
c u
2 L4 (R4 ) .
L4 (R4 )
1 c
47
Monday, October 3
Lemma 4 For all m R if f H m,2 (B) & u
B
c f
Hm
+ u
H m+2 2 ,2
(B)
[From last time - needs to be proved]. For m = 1, m Z Lemma has been known via Solonikhovs theorem. For m = 2, by interpolation, it sucies to prove the lemma for m Z .
If two integers true, then between two integers also true
use standard interpolation.
f f
Lp1 Lp2
c f
Lq
m = and =
m
m0 m < 0.
Sobolev space
u H m 2 ,2 (B), B u d = 0 u H m 2 (B)
1
m m<
u,
(S)
=0
1 2 1 2
= H m 2 (S) . . . . so
CHAPTER 1. LINEAR & NONLINEAR STATIC NSE For any m , ! v m such that v + p = 0 v =0 v B = . To prove above, m = k, k Z. We want to estimate v Decompose: v = v1 + s where s satises () We know s
H m,2 (B) H k (B)
48
rk
s
s
=0 on B.
In order for this to be solvable, the average on the boundary must be 0. We know that by Lax-Milgram (?)
=0
is compatible
v1 That is, v1
H k (B)
H k (B)
H k (B)
sup
f H k (B)
v1 , f
H k (B)
49
H k+2 (B)
c f
H k (B)
ibp
v1 ,
v1 v ,v 1
v1 , v
+ v1 , q p,
0
v
0 div p,v pdivv
= sup = 1.
k f H0 f 0
v1 ,
k H0
Continuing, v
1 H m+ 2 (B)
v1
H k 2 (B)
v
1
H k+2 (B)
H k 2
+ s
H k 2 (B)
f Using
k H0 (B)
H m (B)
H m 2 (B)
s =? s =
50
derivatives.
c f
k H0 (B)
H m 2 (B)
H m 2 (B)
S m Z m is an isomorphism V.
v + p = 0 v =0 v=
m (B) .
interpolate true for all whole numbers. duality interpolate negative numbers. Using conclusion,
v + p = 0 v =0 v B = u
H 1 (B).
51
v v
2 L2 (B) 2 L2 (B )
c4 v c4 v
2 L2 (B) 2 L2 (B) .
This estimate is not obvious, because of p term: Let v = v B u Then average on boundary still
8 L 3 (B)
= vu
3 2 (B)
uu
Poincaire H 1 (B)
L2 (B)
L2 (B)
L2 (B)
C u
L2 (B) .
p is harmonic we can control L2 norm, control all high order derivatives of harmonic fnction inside
q By embedding
H m,2 (B)
c u
L2 (B) .
k+1 p
k+2 v
L (B)
L (B)
c u
L2 (B)
52
Sunday, October 28
3d regularity For n = 3, f L (0, T ; H), f L1 (0, T ; H), u0 H 2 V .
If
u=0
assume smooth, carry out argument
u u u + p + (u u) = f t d dt
u + 2 2
u + 2B[u , u, u ] = 2 f , u 2 f B[u , u , u ] u
4 2 4 2 L2
L2
L2
,
2 L2
L2
C u
L2
2 u 4
1 4
poincaire 2
3 4
2 2
d u dt
L
2 2
+ 2 c u
if 0 then ok.
u L2
2 f
L2
L2
If the term is not 0, we impose condition on the initial data: c u0 2 > positive. When we turn on evolution the term c u 2
CHAPTER 1. LINEAR & NONLINEAR STATIC NSE grows, so it could become negative. We estimate: u (u u + p + u u) = f u, u
2 L2
53
Call d1 = f (0) d2 = f
L L2
+ c0 u0
H2
+ c0 u 0
1 2
2 H2
,
2 L2
d2 + 2 u0 d3 = d4 = d2 + 1 + d2 1
T d2 + u0
2 L2
d1 T d2 exp
T 0
f (s)
L2
3 c2
can make true for either large coecient or small data. Clearly d3 d4 u(0)
2 L2
d3 d4 <
3 c2 .
We can show that, even at max time T . Choose T near zero such that 0 < T < T and such that c u(t) 0. We show that T can be extended all the way to T .
L2
>
d u dt d 1 + u dt 1 + u
2 L2 t 0
add
2 L2
2 f
L2
L2
L2
2 L2
1 + u
2 L2
exp
f (s)
ds 0 L2
CHAPTER 1. LINEAR & NONLINEAR STATIC NSE Then 1 + u (t) and u d4 c u geq0. Contradiction cannot stop.
Need to prove that next time L L2 .
2 L2
54
1 + d2 exp 1
T 0
f (s)
L2
ds
((u, v)) = Du Dv u =
u, v V ,
((u, u))
H compact linear
H V
2.1
Stokes operator
The space u V can be viewed as an element in the dual as follows: identify use Riesz v is a bounded linear functional on V , so by RRT another element, Reprentation Theorem 55
56
V ,V
Au v
Du Dv u, v V.
u t
u + p = f .
Stokes
[[u, v]] =
can dene
(uv + Du Dv)
V ,V
2.2
Function spaces
For 1 +,
L (a,b;X)<
1 < + = +
dont forget - the spatial variation X is Banach, i.e. it has the inner product X .
57
- cannot discount measure zero set (!) -Compare with norm having ess sup not dened everywhere, so we can discount. Lemma. Let X Banach space with dual X and u, g L1 (a, b; X). Following are equiv: 1. u a.e. integrable equal to primative function of g u(t) = +
t
g(s) ds
0
u(t) (t) dt =
g(t)(t) dt
a
3.
d dt
-If (i) (iii) hold, u is a.e. equal to a continuous function from [a, b] X after mollifying. Proof
(i) (ii) Easy to see (iii) (ii). For all D((a, b)), true in sense of distributions:
b a
u, (t) dt =
g, (t) dt
a
b a
58
u = 0.
Let 0 D((a, b)) be such that the mass a 0 (t) dt = 1. Then
b a b
)(t) dt = 0.
compact support/smooth
((t) 0 ( )) d
Use (ii).
b a b a
u(t)((t) dt) =
(t) dt
b a
u(t)0 (t) dx
0 = constant
b a (t) dt
b a
u( ) (t) dt = 0
=0 t
u(t) = 0 a.e.
(u v0 ) (t) dt = 0
u v0 =
so v0 = g(t).
If f L1 (a, b; X) satises
b a
Since, actually the spatial variation is the Banach space X, he replaces R with X.
then
59
Friday, October 7
H 1
For u L2 (0, t; V )
u ( 0, T ; V ), then
like L2
u C([0, T ], H) d 2 u = u(t), u (t) V,V L1 (0, T ) dt 2 u (t) W 1,1 (0, T ) C([0, T ]) Stokes system ut u + p =f u = 0 has two cond: u = 0 = u0 u0 = 0 u0 = 0
match initial data. cannot talk a priori about trace. But with regularity above we can. Question: for f L2 (0, T ; V ) , u0 H are given, we seek u L2 (0, T ; V ) such that
compatibility condition
v V
2.3
Galerkin
solve using limiting process adjust nonlinearity to look more linear. Let {wi }i=1 be complete (but not orthogonal) base of V & orthonormal base in H.
for i j
60
m .
not fm
tail disappears.
project m
i=1
(u0 , wi )wi
gim (t) wi , wj
orthon. m j=1
m j=1
Fj (t) = f , wj and
gim (t) =
m j=1
lim
d dt
X + AX = B(t) X(0) = X0 Note: We do not need to use Brouwer xpoint here, because we have time. time is actually nice right?
take limit: Clearly {um } L2 (0, T ; Vm ).
61
how to take limit? Use compactness. Compactness introduces some form of control (trapped in a ball) - use apriori estimate.
d d 2 um = 2 um , um dt dt
d (um , uj ) dt
d = (um , dt uj )
d d dum d 1 2 um = um , um = um , um = , wj dt 2 dt dt dt Using that 2ab a2 + 1 b2 , d 1 2 um + ((um , um )) = f , um dt 2 d 2 2 schwarz 1 2 2 um + 2 um f + um V dt V t 1 t 2 2 2 um (t) + f V d m um ( ) d 0 0 T 1 t 2 2 2 f V d sup um H (t) + um ( ) d um + 0 0 0 0tT so (or where) Then um
pass limit.
2 L (0,T ;H) 1 um L L2 L2 Hx t x t
Lax-Hopf condition
+ um
2 L2 (0,T ;V )
u0 H +
t 0
f (t)
2 V
dt
cf Heat equation.
62
Monday, October 10
2.4 Pass to limits
in nite dimension, weak convergence & strong convergence same thing take subsequence , we may Without loss of generality v L (0, T ; H) v in L ( 0, T ; H ) & um w and w L2 (0, T ; V) such that u m in L2 (0, T ; V) B1 L is weak star compact 1 . This means:
um , v1 ,
um , um u, or
L
i.e.
um , w, L2 (0, T ; v)
um ,
Claim: v = w.
Note for T < +, L2 (0, T ; V ) L1 (0, T ; H). If I pick an element in the smaller test space, then L2 = L1 - they are the same. Remember that we can cook up a sequence having 2 dierent limits in dierent spaces, for example sin but sin x n x n 0 in L2
where is the Dirac mass in the Radon measure space M([0, 1]).
1
The weak star topolgy is by denition, the weakest one that makes all functionals x x, x
continuous
63
+ ((
um
, wi ) ) = f , wi .
f ,wi
L2 uwi
The term
is converges: d wm , wi w, wi . dt
If function convergent Distribution also convergent. Moreover, the distribution converges in the distributional derivative sense d d fi f dt dt Shift test function to see it converges: d d f = f . dt dt Then the sequence um is convergent, 0 um = 0
Then
m i=0
gim (0)wi u0
in H.
d (u, wi ) + ((u, wi )) = f , wi dt
d (u, v) + ((u, wi )) = f , wi for all v V. dt v satises the equation in the distributional sense and the boundary value u=0 is kosher because of the yellow boxed explanation above.
64
u t = 0 = u0 Show: has well-dened limit in L2 -sp. u C([0, 1], H) H-norm as function of t has well-dened limit. pf Using d = u dt
2
u in
= 2 u , u
V ,V
read in V , V duel
2 u , u
V ,V
L1 (0, T )
2 2
as t t0
u(t) u(t0 ) 0 The last convergence holds because u u(t0 ) < u(t0 ) 2 u(t), u(t0 ) 0
2
u(t) +
0tT
sup u(t) +
T 0
um
dt
u0 +
T 0
f (t)
2 V
dt
lower semi-continuity
65
Lower semicontinuity
weaker than continuity A function f is upper (lower) semi-continuous at a point x0 if, roughly speaking, the function values for arguments near x0 are either close to f (x0 ) or less than f (x0 ). A lower semi-continuous function. The solid red dot indicates f (x0 ) In particular, f is lower semi-continuous at x0 if for every > 0 there exists a neighborhood U of x0 such that f (x) f (x0 ) for all x U . Equivalently, this can be expressed as
A function is lower semi-continuous i x X f (x) > is an open set for every R. A function is continuous at x0 i it is upper and lower semicontinuous there.
66
u = 2 u , u
V ,V
C (0, T ) L2 C (0, T ; V ) L2 (0, T ; V )dense C (0, T ; V ) L2 (0, T ; V )dense. In fact, mollication in t-var um = u {vm } C such that vm u in L2 (0, T ; V ) and Use Hlder o 2 u , vm u , u m using that
d dt
dense in L2
vm u
in L2 (0, T ; V ) in L1 (0, T )
vm
d dt
u .
2.5 2.6
67
Wednesday, October 12
68
Wednesday, October 14
69
Wednesday, October 17
70
Wednesday, October 19
L2 . Then there exists
u=
u
u t [0, T ] 0 otherwise
extension generates
Then g L2 .
71
u( )
2 X1
d < +
2 L2
L1
g (t)
2 X1
+ u(0)
2 X1
+ u( )
2 X1
bounded
cannot integrate; derivatives in L2 , but this term not L2 so sacrice derivatives 2(1 + 2 ) 1 + 2(1) We need to estimate
2
+ 2(1 + 2 )
2 X1 2 X1
2 X1
< 2 + 2.
u(t)
2 X1
d
split
1 + 2 u( ) R 1 + ( 2 )1 1 1 + 2(1)
1
u( )
d +
2 u R 1 + 2(1)
X1
parseval: R u 2 X1
1 2 g R 1 + 2(1)
1
2 X1
+C
R
g ( )
C (1) R 1+
singularity at but0
1 2(1)
L2
d < if <
1 2
compactness follows.
Theorem. For n 4, f L2 ([0, T ]; V ), u0 H, there exists at least one solution u L2 (0, T ; V ), u L2 (0, T ; V ) such that Error. Actually, this is L1 ; see t u u + u u + p = f
nonlinear
below
u=0 u(0) = u0
72 u, v, w V is
L10 3
L2
L10 3
ut Au + Bu = f
V H 1 V V
where it follows that ut V (after looking at where other terms are). Note that V is integrable in t.
Then the regularity of u is improved: u
(0, T ; H), by
f L2 (0, T ; V ),
Bu
2 V
, Bu
= B[u, u]
sup C u u
u
V
is L1 integrable.
not continuous. u C([0, T ; Hweak]. weak continuous can still take weak limit. For all H, u(t), C([0, T ]). nonlinear terms are trouble; take care of Bu , being nonlinear (because T T 2 of square): 0 Bu V 0 u V .
((um , wj ))
Dirichlet inner product
truncate initial data: um (0) = projVm (u0 ) = u0m , where u0m = m (u0 , wi )wi i=1
73
gim (t) +
m i=1
((wi , wj ))g(t) +
m i=1,
where the term is nondegenerate because it forms a base. nonsingular. (wi , wj )1i, jm GL(m).
invert: gim + m ij gjm (t) + m ijk gjm gkm = m ij f , wj where j=1 j,k=1 j=1 ij , ij , ijk R.
nonlinear, cannot solve globally (like X = X 2 .) can solve locally; Fixpoint, X = F (t, X).
fm > 0 & gim (0, fm ) R If tm is the maximum value then gim (t) as tm ; however,
here have cancelation, so even though the equation is nonlinear, the estimate we have is still linear (!)
d 2 um + 2 um dt
um
1 f
2 V
74
um (t) +
bounded
T 0
um ( )
2 V
1 dt um (0) +
2
t 0
f (t)
2 V
dt u0 +
t 0
f ( )
2 V
75
Friday, October 21
Galerkin - nish convergence uniqueness for m = 2. Leray-Hopf solution for n = 3 regularity.
Suppose u L (0, T ; H) L2 (0, T ; V ) solves ut u + u u + p = 0 u=0 u t=0 = u0 . and suppose (mult by u, integrate) u(t) + 2
T 0 R3
R3
u dxdt
R3
u0 .
um :
um +
This says:
T 0 R3
um u0 +
f
0
2 V
um
L (0,T ;H)
+ um
(0,T ;V )
C.
um , where
u in L (0, T ; H)
76 Bum
B[um ,um ]V 2
In fact, B[um , um ] V :
Sobolev
Bum
sup
wV =1 B[um ,um ,w]
= sup
w
V =1
u um w um
R
2 2,
dt
0 um
2 2
dt C. Upshot:
fm L1 (0, T ; V ).
have better control of derivatives,., so um u
in L2 (0, T ; H).
u m where we used um um w um
R 4
L1 (0,T ;V )
um
C um
2 2
= C um
.
apply Albin-Lions, with V H V
Then um u in L2 (0, T ; H)
dum dt , wj
+ ((um , wj )) + B[um , um , wj ] =
((um ,wj ))
f , wj
constantconverge
It suces to show
77
um wj uwj
= (um u)wj
wj
um u
initial data for approx solution pass to limit (will not repeat; see previous Notes).
Lerays idea: ut + u u u + p = f u=0 Mollify the term to u , then equation is linear (Stokes equationwith rst order term). vt + u v v + p = f v =0 so that u u v = T (u).
78
2.7
Uniqueness
2 u
1 2 1 2
Lemma 5 For n = 2, R2 , u
L4
L2
L2
1 u H0
Ladyzhenskaya inequality
1 in 2 dimension, H0 is H 1 (R2 )
Lp
p < +
Proof. Extend by Nirenberg higher dimension. Lemma 6 (better estimate on trilinear form) B[u, v, w] 2u
1 2
1 2
v w
1 2
1 2
L2 (0,T ;V ) .
2 be obtained (idk)
79
replace
by u1 , equality
B[u, u1 , u] C u 4 u 4 u1 so d 2 u +2 u dt
2
C u u u1
2 u
+C u
u1
d 2 u c u1 dt d c t u1 (t)2 e 0 dt
0 0
c 0t u1
2
dt u0 2
u0 2
0
=0
B[u, u, v] = B[u, v, u] u u u
2
1 2
2 2 2 V
u
2 2
1 2
v
T
2 2
1 2
u
2 2
1 2
u
T 0
B[u, u]
2 L (0,T ;L2 ) T 0
u
0
2 2
dt.
80
Monday, October 24
proof of Ladyzhenskaya inequality existence theorem 3d version
Lemma 7 (n=2) u
L4
2u
1 2
L2
1 2
L2
1 u H0 .
even splitting
!
3 4
Lemma 8 (n=3)
L4
1 4
L2
L2
1 u H0
L2 (R2 )
2=
21 211
prove for smooth compact support functions pass limit (which we can do because dense).
holder
L2
c (u )
2
L1
= c 2 u u c L6
1 2
1 2
in 3d, is embedding in L6 .
1 1 1 = + = . 4 2 6 4
1 4 3 4
interpolation inequality
L4
L2
Sobolev inequality
L6
c u
1 4
L2
3 4
L2
W 2,2 (R3 )
L6
81
smooth fc w compact spt; ending pt at is 0 because compact spt
= 2
+ x2
u(x, ) d u(x, ) x2
u(x, ) D2 u(x, ) d
u(, x2 ) D1 u(, x2 ) d.
4 2
R2
u (x1 , x2 )
2
R2 2
D2 u
1 2
R2
u2
R2
D1 u + D2 u
Theorem 10 If n = 3 the solution we obtained via Galerkin method to the NSE satises u L (0, T ; H) L2 (0, T ; V )
uL
8 3
(0, T ; L4 ())
2
8 3
3 4
3 2
=2+
3 6
u L
4 3
(0, T ; V )(0, T ; V )
!
8 3 0 T
2 3
1 Proof For a.e. t [0, T ] we have u(t) H0 . By the 3d interpolation inequality/ Sobolev inequality we have
1 4 3 4
u(t)
L4
u(t)
L2
L2
8 3 L4
1 4
L2
u L2
3 4
L2
u L2 dt
82
L (0,T ;H)
2 L2 (0,T ;V )
L8 3 (0,T ;L4 )
L (0,T ;H)
3 4 L2 (0,T ;V )
proof u = We claim
+ Bu +
f
L2 (0,T ;V )
L2 (0,T ;V ) Bu L4 3 (0, T ; V ).
u L4 u uv
1 2 4 L2
L4
L4
L4
u so
3 42 L2
L2
1 2 L2
Bu B(t) V
1 4 L2
3 2 L2
L4 3 , u
1 2 ; L2 (0,T ;L2 )
L4 3 (0,T )
L (0,T ;L2 )
This is the regularity we have; but this regularity we dont know f unique;
If try to extend 2d proof you will nd we need more If n = 3 and u satises in addition the following
83
u = B[u, u, u] B[u, u2 , u] B[ u , u , u2 ]
0 L4 L2 7 4 L2 L4
int
u(t)
8
L4
4 L2 u2 4
1 4
L2
u2
4 4
3 4
+1=
7 4
Using ab a 7 + 1 b8 where = 2 u
7 4 8 7
L2
u(t) d dt
+ u2 L2 u +
2
1 2
8 L4
2
2
u + u(t)
8 L2
2
2 L2
u2
8 L4
u2
u
L4
u(t)
1 u
dg = c(t) dt, g c( ) d ,
2
ln g ln g(0)
t 0 t 0
t 0
c( ) d,
g( ) g(0) exp
2
c( ) d
8
] u t (t) d 0
4
integrable
so u0 u1 = u2