Beruflich Dokumente
Kultur Dokumente
S. Kumaresan
School of Math. and Stat.
University of Hyderabad
Hyderabad 500046
kumaresa@gmail.com
Introduction
Abstract
This article has some very annoying typographical mistakes which may lead you to
think they are mathematical in nature! Beware! I did not proof-read them carefully. I
would appreciate anybody who send me the list of corrections. This exposition is accessible
to all those who need the result in their domain but are not experts in analysis.
Let be an open set in Rn and L be a linear partial differential operator on . That is,
X
L=
a D ,
||k
1
where = (1 , . . . , n ), i Z+ , D = ( x 1 ) ( xn )
order of L is k if a 6= 0 for some with || = k.
The main question of these lectures is the following question: Consider an equation of the
type Lu = f in . Assume u C k () and f C r (). Then is u C k+r ()?
Regularity theorem essentially says that under certain conditions on L, the answer is in
the affirmative. (Not quite in these words however for the time being this will do!)
We are already familiar with many cases where such a thing happens:
(1) Let f : R R be C k . Let u : R R be a C 1 -solution of the equation
u is in fact C k+1 .
2
du
dx
= f . Then
2
(2) = R2 and L = x
2 + y 2 . u C () is said to be harmonic if Lu = 0. Here 0 is
a C function and so we would expect that if u satisfies Lu = 0, then u C (R2 ). Indeed
this is true since if v is a harmonic conjugate of u, then f = u + iv is an analytic function of
a complex variable and hence infinitely differentiable. So its real part, which is u, is infinitely
differentiable.
u
z
But the following example shows that for regularity theorem we need some conditions on
L.
1
Y = x y, the equation (
2
x2
2
y 2
2
x2
2
.
y 2
By the transformation
Weak Derivatives
u
dx = f (x)(x) dx
x1
Similarly, if L = D we get
(1)||
uD dx =
Z
f (x)(x) dx
(2)
Conversely, let Eq. 2 be satisfied for all Cc (). Then hD u, i = hf, i for all Cc ().
That is, hD u f, i = 0 for all Cc (). One then easily shows that D u = f . (See
exercise below).
R
Ex. 1. Let C() such that = 0 for all Cc (). Then show that = 0.
Now we make the following definition:
R
Definition 2. Let u L1loc () (i.e., for all K K compact, we have K |u| < ). We say
R
R
that D u = f in , in the weak sense if (1)|| uD = f (x)(x) dx for all Cc ().
Note that in the case where u C || (), D u = f in the weak sense iff D u = f in the
calculus or classical sense.
(
x if 0 < x 1
Example 3. Let u : R R be defined by u(x) =
. We claim that v :=
1 if 1 < x < 2
(
1 if 0 < x 1
is the weak derivative of u. To prove this, we need to show that
0 if 0 < x < 2
Z
u =
0
v,
0
x +
=
=
[x(x)]10
+ (2) (1)
Z 2
(x) (1),
= (1)
0
Z 2
v.
=
since (2) = 0
The simplicity of the above argument should not lull the reader into false confidence. So,
let us look at another example.
(
x if 0 < x 1
Example 4. Let u : R R be defined by u(x) =
. You may be tempted
2 if 1 < x < 2
to guess a weak derivative of u . We claim that u has no weak derivatives. Assume that v is
a weak derivative of u. Hence we have
Z 2
Z 2
u0 =
v,
for Cc ((0, 2)).
(3)
0
=
=
=
=
x +
20
0
1
Z 1
[x]10
+ 2((2) (1))
0
Z 1
(1)
+ 2(2) 2(1)
0
Z 1
(2) (1)
0
Z 1
(1)
.
(4)
From Eq. 3 and Eqellip:eq4, it follows that v must satisfy the equation
Z 2
Z 1
v = (1) +
,
for all Cc (0, 2).
0
(5)
We choose n such that (1) = 1 and support of n shrinks to the point 1. Then we have
Z 1
Z 2
1+
n =
vn .
0
Using the dominated convergence theorem, we get 1 = 0. This contradiction proves our claim.
For every L there exists a unique differential operator L (called the P
formal adjoint of L)
i for all u, C (). If L =
defined by the equation
hLu,
i
=
hu,
L
a (x)D then L
c
P
||
P
Note that for each k and K, pk,K () = ||k supK |D | defines a seminorm on Cc ()
and the above condition means that u is continuous with respect to the family of seminorms
pk,K .
R
Example 7. Let f L1loc (). Define uf () = f (x)(x) dx for Cc (). Then uf is a
distribution with k = 0 in the definition. We say that uf is represented by f .
R
Example 8. Let be a regular Borel measure. Define u() = d. Then u is a distribution
with k = 0.
Example 9. Assume that 0 . Define () = (0). Then is a distribution with k = 0.
But cannot be represented by any f i.e., 6= uf for any f L1loc (). is called the Dirac
distribution.
Notation. We denote the space of distributions over by D0 (). Example 1 shows that
D0 () L1loc (). If u is a distribution and Cc () we denote u() by (u, ).
Definition 10. The i-th partial derivative
by
u
( x
, )
i
:= (u,
xi ).
u
xi
u
xi
Ex.
= Ruf , where f C k (). Then D f makes sense for all with || k and
R 12. Let u ||
D f = (1)
uD , by integration by parts, since f has compact support. This shows
that D uf = uD f .
(
0 if x 0
d
uH is
Ex. 13. Define the Heaveside function H L1loc (R) by H(x) =
. Then dx
1 if x > 0
the Dirac distribution .
Definition 14. Let L be a linear partial differential operator on . We say that Lu = f on
in the distribution sense if (u, L ) = hf, i for all Cc ().
Ex. 15. If u = ug for g L1loc , then Lu = f in the distribution sense if and only if Lg = f
in the weak sense.
4
Fourier Series
The following observation is the cornerstone for our analysis of the problem.
Lemma 16. Assume f C 1 (S 1 ). The partial sums of the Fourier series of f converges to
f uniformly on S 1 , i.e., in the supremum norm on C(S 1 ).
Proof. Let sn (f ) :=
inx
|k|n fn e
f
(t)e
dt
= infb(n).
fb0 (n) :=
0
2 0
2
0
Hence
X
|fb(n)| =
n6=0
X |fb0 (n)|
n6=0
|n|
2
|fb0 (n)| ) 2 (
n6=0
X 1 1
)2
n2
n6=0
0
n (x) := einx and fn := hf, n i, the L2 -inner product of the functions. Let f L2 (T). Then
f has the Fourier series representation
X
f=
fn einx
in L2 (T).
nZr
The following exercise, though trivial, is repeatedly used in the sequel, often without
explicit mention.
Ex. 19. Let f : Rn \{0} R be defined by f (x) := kxk for R. Then f L1 (Rn \B(0, ))
for any > 0 iff < n. Hint: Use the spherical polar coordinates and the change of variable
formula.
Ex. 20. Let k > r/2. If f C k (T), then as in the one dimensional case show that the
Fourier series of f converges uniformly to f . Hint: Last exercise may be of use.
P
inx then D f has as its Fourier series
Ex.
21. If f C (T) and if f (x) =
nZr fn e
P
inx
r
1
fn (in) e , where (in) = (in1 ) (inr ) .
Pr
2
2
The Laplacian =
i=1 xi 2 maps C (T) to C(T). We would like to extend this to an
operator
on L2 (T). We would
like to define : L2 (T) L2 (T) as follows: Let f L2 (T), f =
P
P
fn einx . Define f =
fn (|n|2 )einx . But there
P is one problem this series may not
converge in the L2 norm. For example take f = nZ\{0} n1 einx . On the other hand, we
P
can define : C (T) C (T) by f =
fn (|n|2 )einx and since C (T) is dense in
2
L (T) we can ask whether can be extended to the whole of L2 (T) continuously. But this
is not possible either. For : C (T) C (T) is not continuous, as the bounded set
{einx | n Zr } in L2 (T) is taken to the unbounded set {|n|2 einx | n Zr } in L2 (T).
So in order to define
way, we proceed as follows. ConsiderP
the collection
P in a meaningful
inx
|n|2s |vn |2 <
of formal series H = { nZr vn e }. Fix s Z. Look at those vn s for which
P
P
1
. We have a natural seminorm on this space given by
vn einx 7 ( |n|2s |vn |2 ) 2 . We
P
1
modify this seminorm to a norm ((1 + |n|2 )s |vn |2 ) 2 . We call this normed linear space the
s-th Sobolev space and denote it by Hs .
Definition 22. Let s Z. The s-th Sobolev space Hs is defined by
X
X
Hs = {v =
vn einx | kv k2s :=
(1 + |n|2 )s |vn |2 < },
nZr
(formal sum)
nZr
un einx , v =
vn einx , we
(6)
inx
D f formally by D f =
fn (in) e , where = (1 , . . . , r ), i non-negative. Then f
has L2 -derivatives of order less than or equal to s, i.e., for || s, D f defines an element
of L2 . For,
X
X
|(in) |2 |fn |2 =
|n1 |21 . . . |nr |2r |fn |2
X
if || s.
Now since the formal derivative D f L2 , it is in L1loc and hence one would expect that it
is the th derivative of f in the weak sense also. Indeed, this is true. All we have to prove is
that
Z
Z
D f dx = (1)|| f D dx for all C (T).
That is, we have to prove that
Z
Z
inx
||
fn (in) e dx = (1)
f D dx
(from the continuity of the L2 inner product). But this follows by integration by parts.
7
un einx , D u =
)t u
un einx Hs . Then
k(1 )s uk2s =
|un |2 (1 + |n|2 )s
= kuk2s .
C(T)?
Let
un e P
Hs .
s
P
P
Now
un einx C(T) if the series
un einx converges uniformly. This is true if
|un |
converges. Now
X
X
s
s
|un | =
|un |(1 + |n|2 ) 2 (1 + |n|2 ) 2
X
1 X
1
1
2
(
|un |2 (1 + |n|2 )s ) 2 (
2 s)
(1 + |n| )
X
1
1
2
= kuks (
2 s)
(1 + |n| )
1
(1+|n|2 )s
P
1
converges
|n|2s
un einx of Hs is in
r
2
C(T) if s >
r
2
+ k, then Hs C k (T).
r
2
kD u(x)kL2 .
||k
|n|2 |un |2
|un |2 (1 + |n|2 )k .
P
k
To get the other way we note that both (1 + |x|2 ) 2 and k |x | are nonzero for nonzero x
and grow like |x|k at infinity so that we have
X
k
(1 + |x|2 ) 2 C
|x |,
k
Distributions on T
xT
D0 (T) =
Hs .
sZ
||k
X X
|n ||n|||
||k n
C0
= C0
|n ||n|k
n
r 1
r 1
|n ||n|k+[ 2 ] 2 + 2 |n|[ 2 ] 2 2
C 0(
|n |2 |n|2k+[ 2 ]+1 ) 2 (
C 00 (
|n|6=0
r
[ r2 ]+1
|n|
)2
r
+ 1.
2
This shows that u : C , Hs C is continuous. But C (T) is dense in Hs . So we can
extend u to a continuous functional on Hs . So we can consider u as an element of Hs . Thus
we have proved that D0 (T) sZ Hs .
00
= C kks where s = k +
||s
kuks c
sup |D |.
||s
10
Thus u is a distribution.
Remark 40. For distributions we have already defined distributional derivatives. Also, for
u L1loc , we defined weak derivatives and for u Hs we defined formal derivative. Now if
u C k (T) we have the usual derivative. We also showed that these definitions coincide.
Now we have identified D0 (T) with sZ Hs . We have to show that under this identification
both definitions of derivatives, viz., formal and distributional, coincide.
P
P
Suppose u Ht , where u = un einx P
and let C (T) and =
n einx . We have the
inx
formal derivative of u given by DF (u) = un (in) e
and also the distributional derivative
||
a D , where C (T).
Proof. As we have earlier proved that D : Hs Hs|| ( > 0) is continuous, the result
follows from the next lemma.
Lemma 42. Let C (T). Let (u) := u for u Hs . Then there exists C > 0 such that
k(u)ks C kuks for u C (T).
Proof. One may perhaps multiply the Fourier series of and u and then try to estimate
kuks . But this is an extremely tedious process. So instead we make use of the fact that for
u Hk , k 0,
X
kukk
kD uk0 .
||k
||s
kk kD uk0 +
X
||<s
11
ci kD uk0
||<s
Elliptic Operators
P
Let L =
a D , where a C (T) and the order of L is k. We have seen that L takes
Hs into Hsk and L : Hs Hsk is continuous. Now one may naturally ask the following
question: Let u Ht for some t and Lu = f , where f Hs . Then does this imply that
u Hs+k ?
The answer is in the affirmative if one is able to show that kuks+k kLuks . But this is
not possible. For example, this is not true for constant functions. So let us modify this and
ask whether
kuks+k C(kLuks + kukt )
(8)
can be true, where C is some constant depending on s, k and t. We shall see that under
certain conditions on L, Eq. 8 holds. Then Eq. 8 is called Gardings inequality.
Now let us see what should be the condition imposed on L such that Eq. 8 holds. Let us
consider
the case of a homogeneous differential operator with constant coefficients, i.e.,L =
P
a
||=k D , where a C.
12
We then have
Lu =
un (L)(n)(i)|| einx .
Lu Hs implies that
kLuk2s =
Now if |(L)(n)|2 (1+|n|2 )k , then we see that kuks+k ckLuks , where c = c(s, k). But this
may not be possible. First, we note that if n Zr is such that (L)(n) = 0 then (L)(ln) = 0
for all l Z. Hence |(L)(n)|2 (1 + |n|2 )k for n 0 only if (L)(n) 6= 0 for any non-zero
n Zr . Thus we are led to require that (L)() 6= 0 for Rr \ {0}. This condition is also
sufficient in the sense that if (L)() 6= 0 for 6= 0 then |(L)()|2 (1 + ||2 )k since we can
find constants c1 and c2 such that
c1 |(L)()|2 (1 + ||2 )k c2 |(L)()|2 .
Note however that (L)(0) = 0, since L is homogeneous. Hence if un = 0 for all n 6= 0, then
Lu = 0 and we can never get kuks+k ckLuks . So we need an error term to take care of
this situation. What we have done above shows that if (L)() 6= 0 for Rn \ {0} then
there exists a constant c such that
kuks+k c(kLuks + kukt )
Now we make the definition of ellipticity.
Definition 44. Let L be a homogeneous differential operator of order k with constant coefficients. L is said to be elliptic if (L)() 6= 0 for Rn \ {0}.
P
Lemma 47. [Rellichs Lemma] Let s < t. The natural inclusion Ht , Hs is compact.
That is, the image of the closed unit ball in Ht is totally bounded in Hs .
P
P
P
Proof. Let u := un einx be in the unit ball of Ht . Write u = |n|N un einx + |n|>N un einx
for an N to be chosen later. We compute the norm of the second term in Hs :
X
X
|un |2 (1 + |n|2 )t (1 + |n|2 )st
|un |2 (1 + |n|2 )s
|n|>N
|n|>N
|n|>N
< ,
if N is chosen sufficiently large.
The first term lies in a finite dimensional space of vs such that if vn 6= 0 then |n| N .
Since any finite dimensional normed linear space is locally compact, we can find an -net for
such a set. Thus the full unit ball in Ht is totally bounded in Hs .
In the course of the proof we have proved the following
Corollary 48. Let t < s. Given > 0, there exists N0 0 such that if u Hs with un = 0
for all |n| N0 , then kukt kuks .
Proposition 49 (Fredohlm Property). Let L be an elliptic operator of order k with C
coefficients so that L : Hk H0 H0 . Then ker L and coker L are finite dimensional.
Proof. We use the corollary to prove that ker L is finite dimensional. For that, for n N,
define Hs (N ) = {u Hs | un = 0 for |n| N }. We have L : Hs (N ) Hsk , with L elliptic
operator of order k. By Gardings inequality,
kuks c(kLuksk + kukt )
for u Ht .
Now let u Hs (N ). We can also consider u Ht for any t < s. So fix some t < s. Given
> 0, we can choose N 0 such that for u Hs (N ), kukt kuks (by the above corollary).
Therefore
kuks ckLuksk + ckuks .
Now choose such that c 21 and choose N accordingly. Hence kuks ckLuksk . This
means that L is a topological isomorphism of Hs (N ) LHs (N ) Hsk . Now if Y , Z, M
are linear spaces and T : Y Z M is a homomorphism such that T |Y is one-one, then
dim ker T dim Z. Therefore dim ker L dim Hs (N ) is finite.
Alternate Proof. Let L : Hk H0 . Suppose ker L is not finite dimensional. Then
there exists a countable collection of linearly independent vectors {ul } in Hk H0 such that
14
L(ul ) = 0 for all l. We can find an orthonormal family {el }lN in H0 and L(el ) = 0 for all
l N. By Gardings inequality,
l
l
l
+
c(
Le
e
e
) = c
0
That is, {
el
k }lN is bounded. Since Hk , H0 is a compact operator, this means that
{el }lN has a convergent subsequence in H0 . But this is absurd since
el em
0 = 2 for
l 6= m. Thus ker L is finite dimensional.
coker L is isomorphic to ker L . Now if L is elliptic then we can show that L is elliptic.
Therefore ker L is finite dimensional and hence coker L is finite dimensional.
We now use these results to prove the regularity of solutions of elliptic partial differential
operators on T. We shall indicate two proofs of this result.
Theorem 50 (Regularity on T). Let L be an elliptic partail differential operator on T of
order k. Let f Hs (T). Let u Ht for some t be such that Lu = f . Then u Hs+k .
Proof. We may assume that t < s + k. Let Hs (N ) denote the space of v Hs such that
vn = 0 for |n| > N . For N 0 we have
kks kL ksk ,
Hs (N ).
We now compute
hu, L i
hf, i
kf ks kks
kf ks kL ksk .
Thus u defines a bounded linear map on L C (T) Hsk . The closure of L C (T)
in Hsk is of finite codimension. Thus the space generated by it and a finite number of
Gardings Inequality
15
P
Theorem 51. Let L be an elliptic operator of order k, L = ||k a D , a C (T). If
u Ht and Lu Hs , then there exists a constant c which depend only on L, s and t such
that kuks+k c(kLuks + kukt ).
Proof. Case 1: L =
||=k
a D , a C.
Case
2:
Step
1.
Let
L
=
||=k a D , where a C (T). Let x0 T and L0 =
P
(9)
1
2
then kuks+k
Case 2: Proceeding exactly as in Case 2 we get kuks+k c(c0 kuks+k1 +kLuks +kukt ).
Step 2 Since a s are uniformly continuous on T, given > 0, there exists = (L), such
that the coefficients of L oscillates within on any ball of radius . So take a corresponding
to the given in Case 2 and cover T with balls. Since T is compact we can extract a
N
subcover. Choose a partition
of unity
covering. Let u Ht
P
P {i }i=1 subordinate to the above
P
for some t. Then u =
i u since
i = 1. Therefore kuks+k
ki uks+k . Since supp i
is contained in a -ball and sup |a (x) a (x0 )| < for x, x0 in a -ball from Step 1 we get
that
ki uks+k (c0i ki uks+k1 + kLi uks + ki ukt ).
Now
kL(i u)ks ki L(u)ks + c(i )kuks+k1 .
Therefore
kuks+k C(c0 kuks+k1 + kLuks + kukt ).
(10)
(11)
Then if we choose such that Cc0 21 , we get Gardings inequality from Eq. 11. So let
us prove Eq. 11.
We want to show that for r s t, kuk2s kuk2t + C()kuk2r . That is to prove that
X
X
|un |2 (1 + |n|2 )t
|un |2 ((1 + |n|2 )t + C()(1 + |n|2 )r )
16
which is true if
(1 + |n|2 )s (1 + |n|2 )t + C()(1 + |n|2 )r for all n.
That is, if
1 (1 + |n|2 )ts + C()(1 + |n|2 )rs .
Now for x > 0, we have the identity
xs xt + xr or 1 xts + xrs .
Putting x = x, we have 1 ts xts + rs xrs . Choose such that ts = . Hence
rs
rs = ts = C() (say), so that 1 xts + C()xrs . Putting x = 1 + |n|2 we get
1 (1 + |n|2 )ts + C()(1 + |n|2 )rs .
10
Theorem 52. Let be an open set in Rn and L an elliptic linear partial differential operator
of order k. Let u D0 () and Lu = f where f C (). Then u C ().
P
Proof. Let L = ||k a D , a C (). To prove that given x0 , u is C around x0 .
Choose neighbourhoods V V U Q of x0 such that Q is one of the fundamental cubes.
In fact we choose r small enough so that B(x0 , 2r) Q and put V = B(x0 , r), U = B(x0 , 2r).
Let g C () be such
P that g takes values in [0, 1], g = 1 inside B(x0 , r) and g = 0 outside
B(x0 , 2r). Let L0 = ||k a (x0 )D . Let l = gL + (1 g)L0 . Then l L inside B(x0 , r)
and l L0 outside B(x0 , 2r). Since the coefficients of l are constant on the boundary of
Q, l defines a differential operator A on the torus T. Also we can show that if the support
of g is sufficiently small, then l is elliptic and hence A is elliptic. Choose a C function
g1 : [0, 1] such that
(
1 inside B(x0 , 2r )
g1 =
0 outside B(x0 , 3r
4 ).
Now we observe that inside B(x0 , 2r ), g1 u as a distribution on T is the same as u as
a distribution on . That is, if h Cc (), with support of h a subset of B(x0 , 2r ) and
C (T) is got by extending h periodically outside Q, then (g1 u, h)
= (g1 u, h). Now for
h
17
all C (T),
(A(g1 u), ) = (g1 u, A )
= (g1 u, A (g)) + (g1 u, A (1 g))
= (g1 u, A (g))
= (g1 u, L (g))
= (u, L (g1 g)) + (u, P )
= (u, L (g1 )) + (u, P )
= (Lu, g1 ) + (u, P )
= (f, g1 ) + (P u, )
= (f g1 , ) + (P u, )
= (g1 f + P u, ).
In the above, P is a differential operator on of order less than or equal to k 1 with
coefficients having support inside the support of g.
This is true for all C (T). Therefore
A(g1 u) = g1 f + P u.
(12)
Now let u Ht . Then P u Htk+1 . Therefore Eq. 12 gives A(gu) Htk+1 . Hence
the elliptic regularity for the torus gives g1 u Htk+1+k = Ht+1 . But g1 u = u inside
B(x0 , 2r ) T. Therefore u Ht+1 inside B(x0 , 2r ). Repeating the above process we see that
u Ht for all t, inside B(x0 , 2r ). Therefore u and hence g1 u is C inside B(x0 , 2r ) T.
Therefore u as a distribution onR is C inside B(x0 , 2r ). That is, for all Cc () with
support B(x0 , 2r ), u() = g where g Cc (). Now we use a partition of unity
argument to show that u is globally given by a C function. Cover by balls of radius
r
with
2 and choose a partition of unity subordinate to the above covering. Let be a C
support
P inside a compact set K. Then there exists
P an open set W K and 1 , . . . m such
that
1 (x) = 1 for all x W . Therefore = m
i=1 i and
u() =
m
X
u(i )
(13)
i=1
18