18.

103 Lecture Notes

Jonathan Campbell May 23, 2004

This class is really about Fourier Series/ Fourier Integrals, that is harmonic analysis dealing with things like f (x) =
k=

cn exp(ikn),

f (x)eixt dx

the thing is we rst have to do some measure theory, because we want to know about integration (Lebesgue integration). Measure theory itself is boring, because its a stripped down compact theory, so we learn probability mixed in with measure theory.

Simple Probability and Denition of Measure

First consider the Coin Toss. Toss a fair coin and record the outcomes of the unbiased coin. We notice that there are not an equal number of heads and tails (even though we would expect as such in the long run). What is the relationship between this and measure theory? Start by saying that H = 1, T = 0, so the sequence of heads and tails can be encoded as a sequence of 1 and 0. Our state space is the space of all possible outcomes B = {all possible sequences of 0, 1} = {an , n {1 . . . }, an = 0, 1} Bernoulli (hence the B) thought of these as binary expansions of real numbers, that is as a map an {an }n=1 B = 2n n=1 and [0, 1], obviously. The problem with this map is that it is not onetoone and onto. For example 1 .10000 = .01111 = 2 1

map to the same number. So this mapping is not quite 11, but it is close enough. The problem is recurring 1s. For (0, 1] consider the binary expansion which ends in innite 1 s. If we have such an expansion, we do the following a1 . . . ak 0111 a1 . . . ak 1 we kill all sequences that end in innite number of 1s. This denes unique maps e : (0, 1] B\Bdeg , b : B\Bdeg (0, 1]

where Bdeg sequences that end in an innite number of 1s, e is the binary expansion, b is the binary sum, b and e are inverses (so in particular they are both 11 and onto). Observe two things 1. Bdeg is countable 2. B is not countable We can see that Bdeg is countable by the following argument. Dene
N Bdeg {{an } : an = 0, n N } n=1

this is countable, and now Bdeg =

N N Bdeg

this is a countable union of countable sets, and so it is countable. So we can throw Bdeg away. Question. What is the probability of a certain outcome? The outcome being identied with a subset of the state space, S B. Now we would like to map S B S (0, 1], that is we want to say the probability of the outcome is (S ), where denotes measure, where measure, for now, is length. Example. a1 = H is an outcome S = 1 . . . , then S = this set is simply (1/2, 1]. Example. a2 = H. Then 11 . . . S= 01 . . . ( 3 , 1] 4 S = (1, 1] 4 2

an n=1

1 an = + 2n 2 n=2 2n

so the measure is the sum of those (S) = 1/2. In general aN = H, then the interval (0, 1) is split up into parts of size 1/2N . 2

Now we can ask complicated questions, like On average how many heads occur in the rst N tosses as N If we say that N number of heads = ai = S N
i=1

we would expect that 1 SN = N N 2 this is the weak law of large numbers. However when we are trying to take the measure of something we can see that S can be nasty. lim Problem. Can we measure the length of any subset of (0, 1]? Can we do the same with volume in Rn ? Our rst job is to construct Lebesgue measure. We know lengths of simple subsets of R. We can agree that ((a, b)) = b a where b a. We can agree on the lengths of nite unions of disjoint intervals. That is N N
Ii = (bi ai ), Ii = (ai , bi )
i=1 i=1

Idea We abstract this denition and just work in the abstract picture. Our abstract setting is a set, instead of (0, 1]. Let X be our base set (Rn ). We have the set theoretic operations A B, A B, A B = A\B = {x A, x B} and sometimes we have A B / c the symmetric dierence A\B B\A. And also A = X\A. Denition. Ring A ring of subsets is a collection of subsets of X closed under unions and symmetric dierence, that is A, B R A B R, A B R Question Is the intersection in the ring R. The answer is yes since (A B) (A B) = A B so it contains intersection. Example. 2X = P, the collection of all subsets of X Example. Our nite unions of disjoint intervals in R is a ring. I1 , . . . , IN , and
I1 , . . . , Im

  and A = Ii , B = Ii , then A B R, since we can always split up A B into non overlapping chunks. 3

Now say that X is a set, R is a ring of subsets and we have a function : R [0, ). This is the measure, what were looking for. One of the properties we need for this function is additivity : (A B) = (A) + (B), if A, B R, A B = From just this property we can derive a number of properties about . Simple Properties of Measure 1. () = 0 [ = A A] Proof:
(A ) = (A) + () = (A)
2. Monotonicity If A, B R, A B then (A) (B).
Proof:
B = A (B A) (B) = (A) + (B A) (A)
So monotonicity is a consequence of additivity and positivity.
3. Finite Additivity If A1 , . . . , An R are pairwise disjoint then N N
(Ai ). Ai =
i=1 i=1

Proof: Use induction on N . Assume that this statement is true for N 1 disjoint sets.  Then look at AN , B = N 1 Ai R and AN and B are disjoint, so (AN B) = (AN ) + (B) (AN ) + by the inductive hypothesis and we are done. 4. Lattice Property For A, B R
(A B) + (A B) = (A) + (B)
Proof: To prove this we break the sets up into disjoint unions and reassemble.
A = (A\B) (A B), and
A B = (A\B) (B\A) (A B)
we can easily see from this breakdown that
(A B) + (A B) = (A) + (B).
B = (B\A) (A B)

N 1 i=1

(Ai ) =

N i=1

(Ai )

5. Finite Subadditivity For Ai R, i = 1, . . . , N N N

Ai (Ai )
i=1 i=1

Proof: Looks exactly the same as the nite additivity proof. Denition. Countable Subadditivity A measure on a ring R is said to be  countably subadditive if given a countable collection of Ai R, Ai Aj = , i = j and Ai R if we then have
(Ai ) Ai
i=1 i=1

(note that this makes sense even if the right diverges). Theorem. Measures 1. Our length measure is countably subadditive.(Once we have shown this it follows that the measure is countably additive, and thus technically a measure) 2. (Cartheodory) We can extend our current measure to a countably additive measure. This is Lebesgue measure.  Theorem. If Ai R, i = 1, . . . , and A = Ai R by assumption, and Ai Aj = i=1 Then
Ai (Ai ), Ai Aj =
i=1 i=1

Proof. Let

N

Ai R and (
N

Ai ) =

N i=1

(Ai )

and

N

Ai A implies that N N

Ai Ai = (Ai )N = (A) (Ai )
i=1 i=1 i=1 i=1

this coupled with subadditivity means that (A) = (Ai )

i=1

and this is what it means to be a measure. Idea: Caratheodory Try to replace R by the biggest possible collection of subsets on which (extends to be) countably additive  We want to extend R so that Ai R, so its not just an assumption. This is the core of measure theory. 5

Example. length in R for today we refer to volume in Rn . I Rn is a multiinterval I = I1 I2 In then

n L (I) = (bi ai ) i=1

now RL is the nite unions of disjoint multiintervals so the volume in RL is L =

N k=1

(I

(k)

),

A=

N
k=1

I (k)

But what about countable additivity? when all the multiintervals can be small and nasty? Proposition. L is countably additive.  Proof. Given Ai R
Ai Aj = , we know that Ai = A RL we must show that
, i=1 L (A) i=1
(Ai ), and then we are done, since we already know by the previous theorem L that L (A) L (Ai ). i=1 Lemma. If A RL , (A) > 0 1. F RL such that F is closed and F A and L (F ) L (A) 2. If A RL then G RL , G open and G A and (G) (A) + . Proof. It suces to show this for multiintervals. Let I = (a1 , b1 ) (an , bn ) (assume bi > ai ), I have denoted the interval here as open, but it may be open closed or semiopen. Dene F as follows F = [a1 + , b1 ] [an + , bn ] I then (F ) =
n n (bj aj 2) = (bj aj ) F () < L (A). j=1 j=1

if we choose small enough (F () is a polynomial in and vanishes as 0. We can dene G similarly by

G = (a1 , b1 + ) (an , bn + ) I

We continue with our proof. Given > 0 apply the above lemma to get F A, L (F ) (A) , given > 0 apply lemma to each Ai to get Gi open such that Ai Gi and (Gi ) (Ai ) + /2i . Now  F A = Ai and so F is closed and bounded and thus compact, by HeineBorel. So there is a nite subcover. This implies that F
N
i=1

Ai

for some N (since we have a nite subcover, we might as well just take the rst N ). But we know that we have nite additivity of , so (F ) and (A) (F ) then we have (A) so L (A)
i=1 N i=1

(Gi )

N i=1

(Gi )

N i=1 i=1

(Ai ) + i (Ai ) + 2 i=1

(Ai ) +

L (Ai )

We try to enlarge R. First we try to measure every subset of X. Suppose B X. Try to cover B  a countable collection of elements of R. by Is B Ai , Ai R i=1 Denition. Outer Measure: : 2X [0, ] has two conditions 1. If it is not possible to cover the set then we dene (B) = 2. If it is possible we say = inf Lemma. If A R then (A) = (A) Proof. A is a cover of itself, so automatically, (A) (A). Suppose {Ai } is a cover of i=1 A, and Ai R. We would like that (A) Because then, in particular (A) inf
i=1 covers i=1

(Ai ) [0, ]

(Ai )

(Ai ) = (A) and we would be done. 7

 Set Ai = Ai A R, then A = Ai . So the Ai are in R, but they are not disjoint. i=1 So to make them disjoint we do the following A = A1 1 A2 = A2 \ A1 . . . N 1
Ai R A = AN \ N
i=1

Then AN

N

i=1

A and this construction means that Ak Al = , h = l and now i

i=1

A=

A , A i j

A k

= (A) =

i=1

(A ) i

i=1

(Ai )

the last part follows since A Ai and since Ai Ai then i (A) and we are done. Caratheodorys Idea What subsets of X should be measurable. Denition. Measureable Set Call M the set of measurable sets. Then () A M (C) = (C A) + (C Ac ), C 2X
j=1

(Aj )

Denition. Set of Measure 0 Consider countably additive, B X has measure 0 if  (B) = 0, i.e. > 0 Ai R such that B Ai such that i=1
i=1

(A) <

Theorem. A set of measure 0 has the Caratheodory property. Proof. Assume (A) = 0. Then (C A) = 0C, because a cover of A is a cover of C A. So then we try to show (C) = (C Ac ). It is clear that (C) (C Ac ). And by subadditivity (C A)+ (C Ac ) (C), so () holds.

1.1

Applications to Probability

B is the bernoulli space of outcome of coin toss. Where as usual H 1, T 0, and B \ Bdeg (0, 1]. Let E be the set of outcomes in which an event occurs. Then the probability is L (E) Dene Rk , the Radamacher function Rk () = 2ak 1, = .a1 a2 . . .

(sort of a heads you win, tails you lose function) The function looks like a square wave thingy. We would like to measure how many heads in rst N tosses. Let SN be the number of heads in the rst N tosses, then SN () = a1 + + an , And we see that 1 SN = 2 so N 1 = Ri () SN () 2 2 i=1 Let E be the number of heads in the rst N tosses from N/2. The distance from N/2
that is.
N E = B : SN () >

2 the question is what is limN L (EN ) This is the same as saying that EN = Before we proceed, Theorem. Special Chebyshev f is piecewise constant, nonnegative on [0, 1] then
1 {f () > } f [0,1] Proof. If f is piecewise constant then we have intervals such that
(xi , xi+1 ) = [0, 1], (xi , xi+1 ) disjoint
i N

SN : (0, 1] N

N
i=1

Ri () + N

2
N 2
SN 1 1
> 2 = ; B; > 2 Ri
N

4N 2 2
i=1

so f (x) = ci , x (xi , xi+1 ), ci 0 and then f= ci (xi+1 xi ) ci (xi+1 xi ) (xi+1 xi ) = {f () > }

[0,1] i ci ci >

and so

1 {f () > }

f
[0,1]

Theorem. Weak Law of Large Numbers

N

lim L (EN ) = 0

Proof. EN is equivalent to {; fN () > 42 N 2 }, fN () = N

i=1

2 Ri ()

then we apply Chebyshev with = 42 N 2 , fn = ( Ri )2 then N N 2 f= Ri + Ri Rj = N

[0,1] [0,1] i=1 i=1

So (EN ) this tends to 0 as N

N 1 = 2 2N 2 4 4 N

1.2

Continue Extending Measures

(C) = (C A) + (C Ac ), C 2X

One of our methods was to dene a measurable set, A, as one in which

A second approach is to dene a set MF as follows MF = {B 2x | a sequence Ai R, such that (B Ai ) = 0} then we dene M as the set of countable unions of sets of MF On 2X consider the function d (A, B) = (B A) [0, ]

10

Idea This is almost a metric on the power set. We see its, symmetric and nonnegative, but d(A, B) = 0 does not mean that A = B, but thats alright. However, we have to check that it has the triangle inequality: (A B) (A C) + (B C) this follows from the fact that (A B) (A C) (B C) (check above relation for self) Lemma. MF is closed under (X X). Operations are continuous with respect to this metric. We say An A in outer measure means that (An A) 0 as n Proposition. If An A, Bn B then An Bn , An Bn , An \Bn , Ac converge to where n you would think they do
Proof. Long and tedious, its in the book.
Proposition. If A, B 2X and (A) < then
| (A) (B)| (A B) = d(A, B) Proof. Assume that (A), (A) < , all of the other cases are easy. And assume without loss of generality that (A) (B), then (A) (B) (A B) (A) d(A, B) + (B) and this is true since A (A B) B. Before we proceed, a useful fact (A \ B ) = (A B c ) Denition. The sets of measurable sets are dened as
M = A X|A = Bi , Bi MF
i=1

We would like to show that is a measure on MF . We need to prove that MF is a ring and that is countably additive. Lemma. MF is a ring, i.e. A, B MF , then A B, A B MF . 11

Proof. A, B MF , by assumption there exists approximating sequences Ai , Bi R such that (Ai A) 0 and (Bi B) 0. We compute ((A B) (Ai Bi )) = (A B) \ (Ai Bi ) (Ai Bi ) \ (A B). We just compute one side of this union,because the argument would be symmetric, and we already know that the union will be in the ring. (A B) \ (Ai Bi ) = (A B) (Ai Bi )c = (A B) (Ac Bic ) i = (A Ac Bic ) (B Ac Bic ) (A \ Ai ) (B \ Bi ) i i and also (Ai Bi )\(AB) (Ai \A)(Bi \B). So the entire set contains (AAi )(B Bi ). So then (A B, Ai Bi ) (A Ai , B Bi ) 0 so A B MF . Also like to show that A B MF . Since (A B) = (A \ B) (B \ A). It suces to show that A \ B MF , since we know that the union is in MF already. We again compute half of ((A \ B) (Ai \ Bi )), that is ((A \ B) \ (Ai \ Bi )) (A \ B) \ (Ai \ Bi ) = (A B c ) (Ai Bic )c = (A B c ) (Ac Bi ) i c c c = ((A B Ai ) (A B Bi )) (A \ Ai ) (Bi \ B) A Ai B Bi if we work out the other have we get the same thing, so ((A \ B) \ (B \ Bi )) (A Ai ) (B Bi ) and so ((A \ B) \ (B \ Bi )) (A Ai ) + (B Bi ) 0 In order that is a measure, it must be nite, so we want to show that < for all A MF . Lemma. < Proof. | (A) (Ai )| (A Ai ) = d(A, Ai ) 0, So (A) (Ai ) + 1 for some i Lemma. (A B) + (A B) = (A) + (B) Proof. We know this for Ai , Bi R that is (Ai Bi ) (A B), and we know that | (A B) (Ai Bi )| ((A B) (Ai Bi )) So now we have a nitely additive measure. 12 (Ai Bi ) (A B) (Ai ) = (Ai )

to nish our proof we must show that is a countably additive measure Lemma. is countably additive on MF Proof. Want to show that Bi MF , Bi Bj = , i = j (B) = (Bi )
i=1 We can easily prove inequality one way.
BN =
N
i=1

i=1

Bi = B MF , then

Bi = BN B

and we get the following strings of inequalities (B) (BN ) =

N i=1

(Bi ) = (B)

N i=1

(Bi )

 now we need to show countably subadditiviy, that is ( Ai ) (Ai ). Given i=1 i=1 > 0, i, {Ai,j } R a cover of Ai such that j=1 (Ai,j ) (Ai ) + i 2 j=1

 the collection {Ai,j } covers Ai , then i,j=1 i=1

Ai (Ai,j ) = (Ai,j ) (Ai ) + i = + (Ai ) 2 i=1 i,j=1 i=1 j=1 i=1 i=1  this is true , so ( Ai ) (Ai ), and so is countably additive on MF i=1 i=1 Lemma. MF = {A M| (A) < } Proof. We would like to prove that if (A) < and A M then A MF . Since A M  we know that A = Bj , Bj MF . We can replace the Bj s by a disjoint sequence in j=1 MF , but with the same measure. A= Then

j=1 Bj ,
BN

= BN \

N 1
i=1

Bi MF , Bi disjoint

N
j=1

Bj A =

j=1

Bj

(A) < ,

and we have that and so N

j=1

Bj

N i=1

(Bj ) (A)

13

let CN =

N

j=1

Bj then CN MF and so

(A CN ) =

j>N Bj

j>N (Bj ) <

for any for N large enough. And so (A BN ) 0, A MF since MF is closed under (X X). Denition. A collection of subsets N 2X is a ring if 1. N is a ring 2. If Ai N then 
i=1

Ai N .[closure under countable unions]

Theorem. M is a ring. Proof. Suppose Ai M we need to show that Ai =

i=1

i=1

Ai M. Well,

i=1

Bij , Bij MF =

Ai =

i,j=1

Bij

countable union of countable unions is still countable, so this is indeed in M. We also need to prove that its a ring. We have already proved the condition for unions. So we have to prove that A B M if A, B M, it suces to show that A \ B M, since we already know closure under unions. Now, c



Bi = A\B = Ai Ai Bic = (Ai \ B), B= Bic
i=1 i=1 i=1 i=1 i=1 i=1

Ai \ B is in MF (Melrose says think about it) Theorem. If Ai M and Ai Aj = , i = j then

Ai = (Ai )
i=1 i=1

i.e. is a measure on M  Proof. Let A = Ai there are two cases i=1 (A) = . Then (A) so (Ai ) = 14
i=1

(Ai )

 (A) = then A MF and N Ai A and so i=1 N N

Ai (A) = Ai M F
i=1 i=1

So we have extended from R to MF to M.

1.3

Extending Lebesgue Measure

Lebesgue measures (X, R, ) is countably additive and we extended it to (X, M, ) where M is a ring. Apply these constructions to X = Rn , R = RLeb (disjoint union of a nite number of rectangles). So we can generate (Rn , M, ). M is closed under countable unions.  E M then E = Ei , Ei MF . And F MF , there exist Ai RLeb such that i=1 (E A) 0. Theorem. Every open set in Rn is in M (i.e. is Lebesgue measurable) Proof. Consider elements of Rn with rational endpoints, (a1 , b1 ) (an , bn ). Consider U Rn , open. And consider all multisets of the type above in U . Then
U= R, R the rational sets
RU

So U is open and measurable. Rn is itself measurable, so closed sets are in M, since for E open Rn \ E M is open and so E c M. Recall that M is a ring, closed under countable unions and intersections. F1 , F2 are ring subsets of Rn , then F1 F2 is a ring. Denition. The Borel sets of Rn are the elements of the smallest ring containing all open (and closed) sets of Rn . There is a smallest ring because we can take intersections of all such rings, and get one contained in all of them. Theorem. If A MLeb , then there exists B B (the borel ring) such that B A and (A \ B) = 0 Proof. Do this in three steps: 1. If A MLeb , then there exists a borel set G B, such that G A and (G \ A) < (this is not quite what we want, since G contains A)  NB RLeb B, so if A F , then Ai RLeb (Ai s are multiintervals), A Ai M i=1  implies that (A) + > (Ai ), so B = i=1 Ai B, then (A \ B) < (so if the measure is nite the denition automatically gives the theorem). 15

 In general A = Ei , Ei MF . Apply the previous step to each Ei nd Bi B, i=1  Bi Ei such that (Ei \ Bi ) < /2i . and then B = Bi B, (A \ B) = i=1 (Ei \ Bi ) < . i=1 2. If A M, B B, B A with (A \ B) < . Apply (1) to Rn \ A then we get E B such that E Rn \ A, (E \ (Rn \ A)) < , but if we set B = Rn \ E then B A and (A \ B) < . 3. Finally, use (2) to construct FN B, FN A, such that (A \ FN ) < 1/N , then  B = =1 FN B, B A and (A\B), A\B A\FN , (A\B) (A\FN ) < 1/N , N so (A \ B) = 0. We are done so the Lebesgue sets are trapped between Borel sets. MEASURE THEORY IS DONE.

Measurable Functions

Were going to deal with measure spaces (X, F, ), : F [0, ] is countably additive. Now we are interested in functions on X: f : X [, ] = {} R {} (using the extended real numbers allows us to take about infs and sups) Dene B , it is a ring of subsets of [, ]. That is B = {A, A {}, A {}, A {, }|A B} this is the extended Borel ring. And for X our good sets (measurable) are F. Denition. f : X [, ] is measurable i f 1 (, a) F, that is {x|f (x) < a} F, a M

Proposition. The following conditions on f : X [, ] are equivalent 1. {x|f (x) > a} F, a 2. {x|f (x) < a} F, a 3. {x|f (x) a} F, a 4. {x|f (x) a} F, a

16

Proof. Compare (1) and (3), (2) and (4), they are complements and so one in the ring implies the other is in the ring as well. Prove (4) implies (1)
1 x X |f (x) a + = {x X |f (x) > a} n n=1 and 1 = {x X |f (x) a} x X |f (x) > a n n=1

prove these makes good use of the eld properties (closed under complements, intersection and union). Theorem. f : X [, ] i f 1 (B) F, B B .(kind of like continuity, note that this makes continuous functions automatically measurable) Proof. (1) above implies that f 1 ((a, ]) F, a. Note that (a, b) = [, b) (a, ], so we can get any Borel set as some union and intersection of (a, ], [, a),
[a, ). Consider
C = {C [, ] such that f 1 (C) F} Note that C is a eld, since

1 Ci = f f 1 (Ci ),
i=1 i=1

i=1

Ci =

i=1

f 1 (Ci )

and f 1 ([, ]) is in C. Since f is measurable, then C contains all open sets, and so C B, because C contains the open sets. Then {x|f (x) > a} F implies that f 1 (B) F for all B B . The converse is easy. Theorem. If f : Rn R is continuous then it is measurable (Lebesgue) Proof. f 1 (B) F. We know that f 1 (A) Rn then A R is open, so f 1 (A) is Lebesgue measurable, because all open sets are. C = {B R|f 1 (B) M} is a eld containing the open sets. Theorem. Suppose f and g are measurable functions then max(f, g), min(f, g) are measur able. Proof. F 1 ([, a)) = {x X |F (x) < a} = {x x| (x) < a, g(x) < a} = {x X |f (x) < f a} {x X |g(x) < a} F . Notice that if f is measurable then f+ is measurable because f+ = max(f, 0) Theorem. Suppose fi is a sequence of measurable functions, fi : E [, ] and let f = supi fi , f (x) = sup fi (x) then f is measurable. (same thing for inf) i=1 17

Proof. {x|f (x) > a} =

j=1 {x|fj (x)

> a} F .

Theorem. If fi is a a sequence of measurable functions on (X, F), then lim sup and lim inf fi are measurable. Proof. f (x) = lim supi fi (x) = limj supji fi (x). This is a decreasing set of numbers, conver gent if bounded below. Then this is inf i supji fj (x). Apply the above theorem twice. Theorem. If fi is a sequence of measurable functions such that fi (x) f (x) (pointwise convergent), then f is measurable. Proof. Since fi (x) converges, lim sup fi = lim inf fi = f , so its measurable This very dierent from normal analysis, where we normally do not know much about pointwise convergent functions. Note that for some of the above properties, its the algebra that is allowing us to do what we are, by taking unions and intersections of an innite number of sets. Theorem. If f : X R is a measurable (note: X need not have a topology) and g : R R is continuous, then g f is measurable. Proof. Consider {x X |g f > a} = f 1 {t R|g(t) > a} since {t R|g(t) > a} is open, then f 1 of it is open, since f is measurable and so the inverse contains Borel sets.

The integral

Start with simple functions, f : X [, ] is simple if it takes only nitely many values. Denition. If f is a nonnegative (initially, just to be careful about cancellation), simple, measurable functions on (X, F, ) and E F, then IE (f ) = ci (f 1 (ci ) E)
ci (0,]

Note that its a nite sum, and that the integral could be +. Since f is measurable, f 1 (ci ) F and IE (f ) is well dened. Properties of the integral. Theorem. If si : X R are simple and measurable then 1. IE (csi ) = cIE (si ), c 0 2. IE (s1 + s2 ) = IE (s1 ) + IE (s2 ) 18

3. s1 s2 IE (s1 ) IE (s2 ). Proof. We just prove the second one. Let Ei = s1 (ci ) (the sets on which s1 is constant, and 1   Fi = s1 (ci ) (the sets on which s2 is constant. Then X =
i Ei = i Fi , and on Ei Fi , 2 s1 + s2 is constant. So IE (s1 + s2 ) = (ci + cj )(Ei Fj E) = d((s1 + s2 )1 (d) E)
i,j d(0,)

= = 

i,j

ci (Ei Fj E) + cj (Ei Fj E)
i,j

(Ei Fj E) ci (Ei Fj E) + cj
j j i

by j Ei Fj E = Ei E, because they are disjoint unions, and so IE (s1 + s2 ) = IE (s1 ) + IE (s2 ). For Riemann integrals we forced f 1 (ci ) Eto be intervals, so the Lebesgue integral is very dierent. Denition. f : X [0, ] is a nonnegative and measurable function, then we dene
f d = sup{IE (s)|0 s f simple, nonnegative, measurable}
E

We pay for such an easy denition. We need to show that there exist such simple functions Proposition. If f is measurable, a sequence of simple measurable functions 0 si si+1 f , such that si (x) f (x) and on any subset of X on which f is bounded, si (x) f (x) uniformly Proof. Suppose we take [0, n) R. Divide this into n2n subintervals of length 2n .
i 1 i Ii = t R n t n ,
1 i n2n 2 2 Then we dene the sets Ei = f 1 (Ii ) and Fn = f 1 ([n, ]). Essentially, each Ei is the set on which f attains a value between (i 1)/2n and i/2n , i.e., x Ei (i 1)/2n f (x) i/2n .  And Fn is the set on which f attains a value in [n, ]. Then X = i Ei Fn . Dene
n2
i 1
sn (x) = Ei (x) + nFn (x) 2n i=1
n

So on Ei , sn (x) f (x) and on Fn , n f and sn = n, so sn (x) f . Now to show that sn sn+1 . For sn+1 , each Ei from sn gets partitioned into two parts: i 1 2i 1 2i 1 i , n+1 , , 2n 2 2n+1 2n 19

then sn+1 is (i 1)/2n and (2i 1)/2n on the same area where sn is just (i 1)/2n . Further, if we look at [n, n + 1], sn+1 is n on this interval, as it was for sn . So we have shown that sn f and sn sn+1 . Now we need to show that sn f . If f (x) = , x Fn n. So sn (x) = n, n, so sn . For the nite case there is some n0 such that f (x) < n0 . If f (x) < n0 , then n > n0 the intervals Ii range from [0, n] at least, so f (x) is in some Ii since f (x) < n, so i1 i f (x) n , n 2 2 sn (x) = i1 2n

so |f (x) sn (x)| < 1/2n for n > n0 , so sn f . If f is bounded, then f (x) < n0 , x X, then |f (x) sn (x)| < 1/2n , x X, so sn f uniformly. Theorem. f, g measurable, E, F F, then 1. f g E f d E gd 2. E F E f d F f d 3. (E) = 0 E f d = 0 Proof. 1. Let s be simple, s f then s g so supsf IE (s) supsg IE (s)

2. This is true for simple functions,by denitions of the integrals in terms of measures, so true for their supremum. 3. IE (s) = ci (s1 (ci ) E) which is 0 by measure theory Theorem. (Chebychev) f 0 measurable, E F, c > 0 and Ec = {x E |f (x) c} then 1 (Ec ) f d c E Proof. Rewrite as c(Ec ) E f d Dene c, on Ec s= c 0, on Ec so 0 s f and IE (s)
E

f d,

IE (s) = c(Ec )

Lemma. f 0 measurable and

f d < , E F then

({x E, f (x) = }) = 0 20

Proof.
{x E |f (x) = } = and 1 (En ) n

n=1

{x X |f (x) n} =

n=1

En

f d 0 as n
E

Theorem. If f 0, measurable, {Ai } , Ai F, Ai Aj = , then i=1

f d = f d, E= Ai
E i=1 Ai i=1

Proof. As usual, we proceed by proving an inequality in both directions. Given > 0, 0 s f , s simple, measurable (called a test function) such that f d IE (s) +
E

And we already know additivity for simple functions, IE (s) = IAi (s) by the countable i=1 additivity of measures. But IE (s) f d f d f d + ,
i=1 Ai E i=1 Ai

Since is arbitrary, we have the inequality in one direction. To get other direction. Let E = A1 A2 , s1 , s2 simple measurable such that 0 si f the such that Ai si d Ai f d /2. But then set s = max(s1 , s2 ), so sd si d f d /2
Ai Ai Ai

But we know for simple functions that f d sd = sd +

E E A1 A2

sd
A1

f d +
A2

f d

since is arbitrary, we have established the inequality in the other direction (and the rest  of the nite cases follow by induction). So E = N Ai , Ai Ak = then E f d = i=1 N f d. i=1 Ai  In general if E =
Ai then i=1 f d
E EN

f d =

N i=1 Ai

f d,

Since this is true for all N , the innite inequality follows. 21

This is sort of the same as for measures, once we have nite case, we can prove innite case by hedging it in somehow.  Corollary. If Ei measurable and E1 E2 EN , E = Ei then i=1 lim f d = f d
n En E

Proof. Apply the above theorem with A1 = E1 , A2 = E2 \ E1 , Ai = Ei \ Ei1 . Then E =  i=1 A1 , Ai Aj = and f d = lim
E n n i=1 Ai

f d = lim
n

f d
En

Countable additivity of the integral under decomposition of the domain. Corollary. If f 0 measurable, we can dene f (E) for E F as f (E) = f d
E

This is a countably additive measure, because of countable additivity under decomposition of the domain. So we can construct new measures from old. Example. For Borel subsets on the real line E B(R) we can dene a probability measure 1 x2 G (E) = exp( )dL 2 2 E so that G (R) = 1. We can get measures from any positive function. An important question in analysis is how measures are related (RadonNikodym) Proposition. (Monotone Convergence Theorem) We are in the measure space (X, F, ). Suppose fj is a sequence of measurable functions fj : X [0, ], fj (x) increasing for each x. So limi fi (x) = f (x) is a measurable function, then lim fj d = lim fj d
E j j E

Proof. First, fj f
E

fj d
E

fj+1 d
E

f d lim
j

fj d f d
E

(sort of follows because our integral is one from below, but we have to show that the integral somehow works the other ways as well) 22

Suppose 0 s f is simple measurable. Suppose 0 < c < 1 is a constant. Dene  En = {x E |fn (x) cs(x)}. So then En+1 En , and En = E.Then n=1 lim fj d fn d fn d csd
j E E En En

since n is independent of j we can take limits as n , so lim fj d c sd lim fj d sd

j E E j E E

but this is true for any s which approximates f , so in fact lim fj d f d

j E E

There exists a sequence 0 sj of simple functions such that sj f (x), so now weve just shown that f d = lim sn d
E n E

This is stronger than our sup denition, because you need just one sequence approaching f (increasing). Theorem. Linearity f and g nonnegative and measurable and c 0, then cf d = c f d (f + g)d = f d + gd
E E

Proof. We already know that rst one. Now we choose a sequence of simple functions 0 sj f (x) such that sj (x) f (x), x X and similarly tj with tj g, and we are done, because we know linearity for simple functions and we can take limits Corollary. If fj is a sequence of measurable nonnegative functions, and fi (x) = f (x). i=1 Then fi (x) = fi d
E i=1 i=1 E

Proof. Let FN =

i=1

fi , then FN d =
E N i=1 E

fi d
E

f d =

i=1 E

fi d

23

Denition. A measurable function f : X [, ] is integrable, f L(, E) if f+ d < f d <

E E

that is f L(E, ) if and only if f is measurable and

|f |d < .

If f is integrable, f L(E, ), then |f |d = f+ d f d

E E E

But now we have to go back and check that L is a linear space. So we check that for f L(E, ), then cf L(E, ). Also, we need to show that h = f + g, f, g L(E, ) then h d = f d + g d
E E E

do this by checking where f , g and h have a xed sign, and then sum up those regions. Now we discuss more theorems on integrals. Theorem. Fatous Lemma If fj is a sequence of nonnegative measurable functions (dont have to be increasing) then lim inf fj d lim inf fj d
E j E

Proof. Set gk (x) = inf jk fj (x), nonnegative and measurable, but also increasing, i.e. gk+1 (x) gk (x). So we can apply monotone convergence theorem lim gk (x) = lim inf fj = lim inf fj d = lim gk d
k j E k E

For each j k, gk (x) = inf jk fj fj (x) implies that gk (x)d fj d, k = gk (x)d inf fj d
E E E jk E

thus

lim inf fj d = lim

E k

gk (x)d lim inf

E j

fj d
E

Theorem. Lebesgue Dominated Convergence We have (X, F, ). If fj is a sequence of measurable functions, which converge pointwise fj (x) f (x) AND g integrable, non negative such that |fj | g, j. Then lim fj d = lim fj d
E j j E

24

Proof. There is a trick. Apply Fatous Lemma to g + fj 0, measurable. Then lim inf(g + fj )d lim inf (g + fj )d
E j E

Since fj (x) f (x) pointwise, lim inf j (g + fj ) = g + f . gd + f d = (g + f )d lim inf gd + fj d

j E E

and so

f d lim inf
E j

fj d
E

Similarly, apply Fatous Lemma to g fj 0, we get (g f )d lim inf gd fj d = f d lim inf fj d

E E E E j E

We know that lim inf(aj ) = lim sup(aj ), so f d lim sup fj d = f d lim sup fj d.
E j E E j E

combining the two inequalities that we got yields lim inf fj d f d lim sup fj d
j E E j E

Some Functional Analysis

We dene L(, F) = {f : X R|measurable and E |f |d < }. Then we dene L1 (X, ) = {x L(, F)}/a.e.. f L1 (X, ) is an equivalence class of elements of L(, F), that is f = [g] = {g L(, F)|g g = 0 o a set of measure 0}. We know |f |d = 0 |f | = 0 except on a set of measure 0. We want to show that the equivalence E class dened above is in fact an equivalence class. We check that f g = 0 on X \ E for some set E, (E) = 0 is an equivalence relation. We check transitivity f g = 0 on X \ E, (E) = 0 and g h = 0 on X \ F , (F ) = 0, then f h = 0 on X \ (E F ), (E F ) = 0. So in fact L1 (X, ) = {[f ]a.e. |f L(, F)}. Theorem. L1 (X, ) is a metric space with the metric d(f, g) = E |f g|d. Proof. 1. d : L L [0, ), d([f ], [g]) = E |f g|d, f [f ], g [g]. It is clear that working with equivalence classes doesnt matter so its well dened (also d(f, g) = 0 [f ] = [g]) 25

2. d(f, g) = d(g, f ), clear from denitions. 3. d(f, g) d(f, h) + d(h, g), f, g, h L1 (X, ). We write f L1 = X |f |d. So we would like to show that f gL1 f hL1 + h gL1 . Set f = f h, g = h g so f + g L1 f L1 + g L1 . This is true by monotonicity and linearity of the integral. Denition. Normed Space. Conditions on a norm 1. L1 (X, ) is a vector space (over R), 2. : L1 (X, ) [0, ) satisfying (a) f = 0 f = 0 (b) cf = |c|f (c) f + g f + g We go ahead and do the same sort of construction but for complex functions. We have (X, F, ). Now f : X C. We have the following denitions 1. f : X C is measurable f = u + iv, u, v are measurable. 2. f integrable i |f | = (|u|2 + |v |2 )1/2 is integrable over E, i.e.
E

|f |d < .

If f is integrable and complex valued we say f L(X, ). Its a linear space for all of the same reasons as L1 . Theorem. f L f L and E f d = E f d, and | E f d| E |f |d. Proof. Let a = E f d C, then a a |a| = f d = f d |a| E E |a| Write this in the form ud + i vd, but i vd = 0 (because its an absolute value), so ud |u|d |g |d = |f |d
E E E E

We dened the normed space L1 (X, ). This is also a metric with metric d(f, g) = f g. We can ask questions like is this complete? Denition. A Banach Space is a complete normed space. i.e. every Cauchy sequence in L1 (X, ) is convergent. Theorem. L1 (X, ) is a Banach Space. 26

Proof. A Cauchy sequence is a sequence with its limit point stolen, so we want to construct a limit for the sequence fj L1 (X, ) to show that it exists. For fj and , N such that n, m N then fn fm < . Lemma. If a convergent subsequence of a Cauchy sequence then the Cauchy sequence converges. Proof. trivial First a corollary of the Lebesgue dominated convergence theorem. Corollary. If E |fn |d < then fn converges absolutely a.e. on E. n=1 n=1 Proof. Let g = n=1 |fn | Then gd =
E n=1 E

|fn |d <

so g L1 and g is nite a.e. on E. So

n=1

fn

converges absolutely on E. Suppose we have a sequence fi such that fn fm < for n, m N . Choose a n subsequence fk such that fn+1 fn 2 . Set g1 = f1 , g2 = f2 f1 and in general n gn = fn fn1 and fn = k=1 gk (x). Now, n gk 1 converges, since k=1
n k=1

gk 1 g1 1 +

1 2k f1 + , 2 n=2
i=1

Now dene g(x) =

|gk (x)|

this exists as an extension of the real numbers and most importantly |g |d =

X

X k=1

|gk (x)|d =

k=1 X

|gk |d =

gk <

so g L1 . So in fact the sum converges pointwise and o a set of measure 0. Dene a new function g which is just g, but dene it as 0 where g does not exist. Then obviously g L1 . Now the series N gk (x) = fN
k=1

27

 converges pointwise and absolutely. And furthermore, | N gk (x)| g . Suppose the k=1 series converges to f . By the lebesgue dominated convergence theorem lim |fn f |d = lim |fn f | 0
n n

so fn f 0 as n and we are done. Alternate Proof Get to the point where we have fni = then
i r=1

gr ,

gr <

1 2r

r=1 X

|gr |d <

thus r=1 gr converges pointwise almost everywhere. Let f be the limit. It is dened almost everywhere. where it is not dened call it 0. We would like to show that fn f in L1 , in other words, show that the whole sequence converges there. Given > 0, n0 such that n, m n0 then |fm fn |d < If we x n > n0 and let m then lim inf |fm fn |d lim inf |fm fn |d |f fn |d = f fn L1 all of this by fatous lemma. So we are done.

Hilbert Space and Completeness

Work in the abstract setting. (X, F, ) a measure space. Dene 2 2 L (X, F, ) = f : X C|measurable and |f | d <
X

and dene L2 as the equivalence classes L2 (X, F, ) = {[f ]|f L2 (X, F, )} Easy to see that L2 vector space. Lemma. f, g L2 then f g L1 28

Proof. 0 (f g)2 = |f |2 2f g + |g |2 2f g |f |2 + |g|2 so |f g|d |f | d +

2

|g |2 d

since both of the integrals on the right are nite, we are done. Lemma. (X) < implies that L2 (X, F, ) L1 (X, F, ). L2 is also a normed space with the norm f L2 =
X

1/2 |f | d
2

its not hard to show that 1. f 2 0 2. cf 2 = |c|f 2 3. f + g f 2 + g2 (Schwarz Inequality) This space is quite dierent from L1 , because the norm comes from an inner product, which we dene by f, g = f gd, L2 (X, ) L2 (X, ) C
X

This inner product satises the following properties 1. f1 + f2 , g = f1 , g + f2 , g 2. cf, g = cf, g 3. f, g = g, f f, cg = cf, g 4. f, f 0 and f, f = 0 f = 0 Denition. A preHilbert space is a linear space V with a sesquilinear inner product V V (v, w) v, w C. This is a specic type of normed space since v = v, v1/2 . Denition. A Hilbert Space is a complete preHilbert space. We claim that v = v, v1/2 . Not hard to prove, I omit it. Lemma. In an inner product space, |v, w| vw. 29

Proof. 0 v + w, v + w = v, v + w, v + v, w + 2 v, w Choose = so |v, w|2 v2 w2 |v, w| vw |v, w|2 |v, w|2 |v, w| v, w v2 2 + = v2 2 2 w, w w w w2

We can use this to prove the triangle inequality by computing v + w, v + w. Theorem. For (X, F , ) a measure space L2 (X, ) is a Hilbert space Basically we want to show that its complete, with the norms f 2 =
X

1/2 |f | d ,
2

f, g =
X

f gd

Proof. Let {fn } be a Cauchy sequence, we will use the L1 result. Set Z = {x X|fn (x) = 0 for some n}. Reason we do this is because if we set Zk = {x X||fn (x)| 1/k} measurable and of nite measure. The point is that
Z = {x X||fn (x)| 1/k}
n,k

is measurable, and of nite measure, because (by Chebyshev) 2 {x X||fn (x)| 1/k} k |fn |d
X

Now Z =

Zk , (Zk ) < , Zk+1 Zk . . . . If f L2 , then f |Zk L1 (Zk , ) since

Zk

1|f |d ((Zk ))

1/2

1/2 |f | d
2

so {fn |Zk } Cauchy on L1 (Zk , ) k. So we know that fn |Zk gk L1 (Zk , ) by completeness of L1 (Zk , ). We showed a subsequence fn(j) |Zk gk pointwise a.e. So we can extract successive subsequence Z1 , . . . , Zk , . . . so that fnk(g) gk pointwise on L1 (Zk , ) implies that we have a measurable function g on Z such that fnk(g) (x) g on Zk pointwise. This converges to g,  a.e. on Z = Zk . Modify on a set of measure 0 to get convergence everywhere, g = 0 on i=1 X \ Z. hk is a subsequence of fn in L2 (X, ). Consider the sequence |hm hn |, n xed, m variable, we know this is 30

1. sequence in L2 (X, )
2. | m (x) hn (x)|2 |g(x) hn (x)|2
pointwise. h Fatous Lemma implies 2 lim inf |hm (x) hn (x)| d lim inf |hm hn |2 d m m X X
2 |g(x) hn (x)| d lim inf |hm hn |2 d
x m X

But given > 0, N such that m, n N hn hm 2 . So if n N we get g hn 2 hn g L2

Still in the setting of abstract Hilbert space, H with inner product , . Denition. Orthogonality u, v = 0 uv. Denition. {i }iI , I countable is orthonormal if i H = 1, and i , j = 0, i = j. Example. Basis vectors in Cn . Denition. Fourier Coecients if f H the (discrete) Fourier Coecients of f are cn (f ) = f, n C,
Assume I = {1, . . . } is enumerated. Consider
SN (f ) =
N n=1

nI

cn (f )n (f )

Theorem. Bessels Inequality If {i }iI is an orthonormal set and f H a Hilbert space, then |cn (f )|2 f 2 <
n=1

Proof. SN =

i=1 ci (f )i ,

ci (f ) = f, i . Recall that

f, g H, f g f + g2 = f + g, f + g = f, f + f, g + g, f + g, g = f 2 + g2 Apply this repeatedly so N 1 2 N N 2 2 SN (f ) = ci (f )i + cN (f )N = ci (f )i = ci (f )2

i=1 i=1 i=1

31

Consider f SN (f )i , 1 i N . Then f SN (f ), = f, j SN (f ), j = cj (f ) So f SN (f ), SN (f ) = 0 then f 2 = f SN (f ) + SN (f )2 = f SN (f )2 + SN (f )2 = SN (f ) =

2 N i=1

N
n=1

c n n , j =0

ci (f )2 f 2 ,

because f Sn 2 0

If M N then 2 M M 2 SN (f ) SM (f ) = cn (f )n = |cn (f )|2

n=N n=N

so it follows that SN (f ) is Cauchy in H. Then the completeness of H implies that SN (f ) S(f ) converges in H.
Does S(f ) = f always? We have a condition below to answer this. Denition. In H {i }iI , I countable is said to be complete if, w H and w, i = 0, i I then w = 0. Proposition. Completeness of i is equivalent to the condition
f= cn (f )n , f H, cn = f, n
nI

i.e. f = limN SN (f ) = S(f ). Proof. Look at f S(f ). Then f SN (f )n , N n. We know by Cauchyness that SN (f ) S(f ) for some S(f ). Now we have to show that 0 = f SN (f ), n f S(f ), n we must simply show that SN (f ), n S(f ), n . But if gj g in H then w H, gj , w g, w since |g gj , w| g gj w 0. So by f S(f ), n = 0 for all n, so by completeness, f S(f ) = 0 and f = S(f ). Our main example of complete orthonormal function are the fourier series. Fourier series, functions on [, ] R. We claim that 1 n (x) = exp(inx), 2 32 nZ

is a complete orthonormal set in L2 ([, ], Leb ). We know that 1 2 2 n = |n | dLeb = dx = 1 [,] 2 Also, 1 n , m = n m dL = 2 [,]
[,]

ei(nm)x dx

assume n = m, then the above is

d i(nm)x e 1 ei(nm)x 1 dx dx =
= 0 2 [,] i(n m) 2 i(n m) but there is a problem: these are Riemann integrals. We have only proved the FTC applies to Riemann integrals. We would like Theorem. If f is continuous on [a, b] then b
Riemann Integral = f dx =
a [a,b]

f dLeb = Lebesgue Integral

This theorem would imply that { 1 einx }n is orthonormal in L2 ([, ]). However, it is 2 hard to prove that n are complete. Theorem. Separable Hilbert space has a complete orthonormal set (called a basis) Proof. H is separable if it has a countable dense subset E H, E = H. We prove it has a basis by using GramSchmidt process. Order the set E = {e1 , e2 , . . . }. We have the following two steps 1. e1 if e1 = 0, then go to step 2. If e1 = 0 then 1 = e1 e1

2. Assume after n steps that we have {1 , . . . , n } orthonormal and ej spanC {1 , . . . , n }, j n. Then consider en+1 . If en+1 spanC {1 , . . . , n } pass to next step.
If en+1 spanC {1 , . . . , n } then /
g = en+1
N i=1

en+1 , i i = 0

(because if en+1 can be written as such a sum then en+1 is in the span). Now gi , 1 i n, and dene g n+1 = g 33

if we continue this indenitely, then we get {i }{iI} orthonormal. Show completeness: Suppose w H, wi , i. The density of E in H implies that given > 0, el E such that w el < . But el is a nite linear combination e1 = f inite ci i . So wel , and 2 w el 2 = w2 + el 2 w < since this holds , w = 0. Now lets go back and prove that the Riemann integral is the same as the Lebesgue integral (in the cases we care about) Theorem. If f C([a, b]) then f L1 ([a, b]) and
a b

f (x)dx =
[a,b]

f dL

is true if f is Riemann integrable Proof. f Riemann integrable means that f is bounded and sup L(f, P ) =
P a b

f dx =
a

f dx = inf U (f, P )
P

where L(f, P ) and U (f, P ) are the familiar upper and lower sums. Then set sL (P ) = inf f [ai ,ai )
(a1,ai ]

a simple measurable function. Just to make sure everything is kosher, replace f by f + c 0, so that we know we are working with positive functions. Then b L(f, P ) = I(sL (P )) f dx f dL f dL
[a,b] a [a,b]

We can prove the inequality the other way by considering (f + c) So now we have justied the computation of n , m where n =
1 einx 2

5.1

The Riesz Representation Theorem

Back into prehilbert space. Lemma. in a preHilbert (no completeness) space the parallelogram law holds: x, y H 2(x2 + y 2 ) = x y2 + x + y2 34

Proof. Just derive from x y, x y and x + y, x + y. Theorem. Conversely, any norm space in which this law holds is a preHilbert space. (i.e. an inner product such that x2 = x, x.) Proof. Dene 4x, y = x + y2 x y2 + ix + iy2 ix iy2 we also have to prove linearity, but thats not hard. Proposition. If C H is a closed convex subset of a Hilbert Space then !x C such that x = inf x
xC

Proof. We rst dene convex. C convex implies that for x, y C then x+y C (a rather 2 weak notion of it, but it works in our case) Now continue with the proof. By denition xn C such that xn 2 inf xC x2 = I 2 . We claim that {xn } is automatically Cauchy. Given > 0, N such that n > N xn 2 = inf x2 +
xC

Now, note that xn + xm 2 = 4 xn +xm 2 . Then 2 xn + xm 2 I2 xn xm = 2xn + 2xm xn + xm 4I + 4 4I 4, since

2

2 2 2 2 2 2

So if n, m N , then xn xm 4 so {xn } is Cauchy Theorem. Riesz Representation Theorem If H is a hilbert space and g : H C (a functional) is linear and continuous then y H such that g(x) = x, y, x H

First, what does continuous mean? g : H C continuous i g 1 (O) H is open O C open, which implies that g 1 (B(0, 1)) H is open. i.e. {x H||g(x)| < 1} H is open. (All of this is because g is linear, i.e. you only need continuity around the origin) or simiarly {x H : |g(x)| 1} is closed. {x H : |g(x)| < 1} is open implies > 0 such that {x H : x < } g(x) < 1. g is linear, so x H x 1 2 x < , so |g( 2 x)| < 1 < 2, so 2 4 g(x) = g x
= c
2
So a constant c such that x 1 |g(x)| c 35

that is, we have boundedness, x H, |g(x)| Cx. Thus, by the boundedness around the origin,

x
g C
1 g(x)
= 1 g(x), so |g(x)| Cx
x

x
x so by continuity of g, C such that |g(x)| Cx, x H. But it works conversely. To get direction: If xn x then xn x 0, i.e. given > 0 N such that xn x < if n N , then |g(xn ) g(x)| = |g(xn x)| Cxn x 0 So for linear functions, boundedness is the same as continuity, |g(x)| Cx Proof. If y H by CauchySchwarz implies |x, y| xy = Cx so this jibes with |g(x)| Cx. Conversely, given g : H C linear and continuous nd y such that g(x) = x, y. g 0 implies that y = 0. g nonzero g(x ) = 1, Set Cg = {x H : g(x) = 1} This is closed (inverse image of a single point) and convex, since

x+y 1 1 g = g(x) + g(y) = 1 2 2 2 This is closed, convex. By the proposition above !y Cg such that y = inf xCg x. Then Cg = {y + y : y H, g(y) = 0} y N , N = {x H : g(x) = 0}, the above can be rewritten as Cg = {y + y : y N }. So then y N y, y = 0, because y + ty2 = y 2 + ty, y + ty, y + t2 y2 We claim x H x = sy + y, Proof : g(x) = 0 x N 36 y N, s C x = x g(x)

so then g(x) = s g then

x s

=1

x Cg x = sy + y s

g(x) = g(sy + y) = sg(y ) + g(y) = s Then x H implies x = g(x)y + y,

g(y) = 0 then

x, y = g(x)y , y g(x) =

y x, 2 y

Example. If g : L2 (X, ) C is continuous and linear, then G L2 such that g(f ) = f Gd

Completeness of Fourier Series

Our main theorem: Theorem. If f L1 ([, ]) and 1 an = 2 then f = 0 a.e.

[,]

f (x)eixn dx = 0,

6.1
Steps

Proof 1

1. Replace f by g(x) =

f (t)dt C

where C is chosen so that [,] gdx = 0. At this step we would like to prove that 1) The Fourier coecients of g also vanish. 2) g is continuous. We have already proved the second one somewhere. 2. Replace g by h: g h =
x

g(t)dt C

h is continuously dierentiable and all fourier coecients vanish. OUR MAIN STEP: 37

Theorem. If f is continuous on [, ] and dierentiable at x0 (, ), then the Fourier series of f cn n (x0 ) = f (x0 )
nZ

(i.e. the series converges to f (x0 )) (Note: We are going to apply this to h, we conclude that h = 0, g is the derivative of h so g = 0. And then the nal step is to show g = 0 f = 0 a.e.) Proof. We want to show that Sn (f )(x0 ) f (x0 ). Now n n 1 Sn (f )(x0 ) = ck k (x0 ) = f (x)eikx+ikx0 dx 2 k=n k=n n 1 iks Sn (f )(x0 ) = f (x)Dn (x0 x)dx, Dn (s) = e 2 k=n so now 1 ei(n+1)s eins Dn (s) = 2 eis 1
n 1 ik(x0 x) Dn (x0 x)dx = e dx = 1 2 n

(so we have shown that the Fourier series for f (x) = 1 converges to 1). Now since we know the convergence of the fourier series for constants Sn (f )(x0 ) f (x0 ) = (f (x) f (x0 ))Dn (x0 x)dx f (x) f (x0 ) x x0 i(n+1)(x0 x) 1 ein(x0 x) = e i(xx0 ) 2 x x0 e Now for the crucial part of the argument. We want to show that Sn (f )(x0 ) = An+1 An where f (x) f (x0 ) x x0 1 ei(n+1)x0 eikx An+1 = x x0 ei(xx0 ) 2 The the boxed function is L2 ([, ]), since it is continuous. Thus Ak is a Fourier coe cients of an L2 function. An is similar, but with an opposite sign. By Bessels Inequality |cj |2 f L2 cj 0 j , f L2
j

so then Sn (f )(x0 ) f (x0 ) 0. Theorem. If (A Aj ) 0 then f d

Aj A

f d

38

Proof. Assume f nonnegative. Pass to a subsequence Ak such that (Ak A) = (A \ Ak ) + (Ak \ A) 2k . Now, it is enough to show that f d 0
A\Ak

Set Bk = A \ Ak . Then F = Bk has nite measure, and so if we dene the remainder  FN = kN Bk then (FN ) 0. Now BN FN , N , so f d f d
BN FN

So if we can prove that FN f d 0 we are done.  Now we write F = j1 Gj , with Gj = Fj \ Fj+1 and Gm Gn = . So we have divided F into a countable number of disjoint subsets. So by countable additivity of the integral f d = f d
F j=1 Gj

since f L1 the left is nite, so the right converges. Thus the tail of the sequence goes to 0: f d = f d 0
FN j=N Gj

Theorem.

h1 (s)
[a,b]

[a,s]

h2 (x)dx ds =

[a,b]

[x,b]

h1 (s)ds dx

Proof. Linear in h1 and h2 . Noting this, we could just prove the above for simple functions and then by linearity again for characteristic functions of a measurable set. Say h1 = A , A [a, b]. By denition of measurability, A can be approximated by a sequence of sets that are a nite union of disjoint intervals, Aj and (A Aj ) 0. So Aj 1d A 1d. And for each Aj we can approximate them by a sequence of intervals. So it suces to prove the above theorem for intervals. Let h1 = [a1 ,b1 ] and h2 = [c1 ,d1 ] then we would like that [c1 ,d1 ] (x)dx ds = ds [c1 ,d1 ] (x)dx
[a1 ,b1 ] [a1 ,s] [a1 ,b1 ] [x,b1 ]

Let a = max(a1 , c1 ) and b = min(b1 , d1 ) then the above becomes dx ds = ds dx

[a,b] [a,s] [a,b] [x,b]

which is clearly true. 39

Apply the above with h1 = einx and h2 = f . Then s ins ins e f dx ds = e ds f dx

x

The inner integral on the right is just a linear combination of exponentials so the right is 0. And if we dene the function g(s) = f dx
[,s]

then its Fourier coecients are 0, n = 0. Then if we adjust by an appropriate constant and let g(s) = [,s] f dx C, where C is such that c0 (g) = 0. There is no problem with passing to another function h, with cn (h) = 0, n, which we dene as h(y) = g C
[,y]

So we know that h(y) = 0 by the pointwise convergence of SN (h). Thus g is at least constant, since its indenite integral is 0. Since g is constant, so [,s] f dx C must be 0 s. Now Theorem. f d = 0,
[a,b]

[a, b] [, ]

Proof. f d = g(b) g(a) = 0

[a,b]

And as a consequence of this, and one of the theorems above, the integral of f is 0 on EVERY measurable set. So f = 0 almost everywhere. So we have proved completeness. Theorem. If f L2 ([, ]) then the Fourier series 1 1 inx 2 f (x)einx dx cn n = cn e f L cn = 2 nZ 2 nZ Proof. Bessels Inequality (n s are orthonormal). Then |cn |2 f L2 = cn n
nZ

is Cauchy in L2 ([, ]) because the series converges (its the denition of convergence). So we know n cn n = g convergent in L2 ([, ]). So cn (g) = cn by convergence since g, n = lim c k k , n = c n
N |k|N

f g L2 ([, ]) L1 , all fourier coecients vanish are 0, so f g = 0 a.e. in L2 . Periodic function period 2 on R can be identied with functions on the circle R/2Z (i.e. x y, x y 2Z). Can identify L2 ([, ]) with L2 (S), where S = R/2Z 40

6.2

Proof 2

Theorem. If f, n = 0, n then f = 0 almost everywhere. Theorem. If f is piecewise dierentiable then SN (f ) converges uniformly to f . Proof. Lemma.


df cn = incn (f ) dx 1 2

Proof. df inx 1
d inx inx e dx = (f e ) + inf (x)e dx 2 dx dx
1
= f ()ein f ()ein + incn (f ) 2
df
if f () = f () (periodicity) then cn dx = incn (f ) and f is in L2 as well. So f fourier series converges, and
df

2
2 cn = n
|cn |2 <
dx

Now if we can prove that |cn (f )|
converges then we are done. But
1/2 1 1/2
|cn (f )| n2 |cn (f )|2 n2 Let be the function (x a)/, a x a + = 1, a + x b

(b x), b x b
(which is basically sort of a function that inclines then stays at 1 and goes back down to 0).
This function is piecewise dierentiable, so then SN ( ) converges uniformly to . Now, inx , so f, SN ( ) = 0, since SN ( ) = 1 |n|N cn ( )e 2 1
f, SN ( ) = f, cn ( )einx = 0
2 |n|N since we are assuming that f, n = 0, n. Now, 0 = f, SN ( ) f, . Now if we let 0, then becomes the characteristic function for [a, b], thus 0 = f, f, [a,b] = f d0
[a,b]

This holds for all closed intervals. 41

6.3

Dierential Equations

Try to solve a dierential equation such as d2 f + f = g dx2 g, given g L2 ([pi, ]) f is to be 2 periodic. We can guess if f L2 ([, ]) then f = cn n . Dierentiate fourier series termby term (its illegal, but who cares) d2 f + f = ( n2 )cn n = dn n dx2 nZ nZ dn the fourier coecients of g. So we expect that dn n2 2 2 so if = n2 , then cn = dn /( n2 ). We have |cn |2 < , |ncn |2 | < , |n cn | < , 2 2 because g L |dn | < . ( n2 )cn = dn , i.e. cn =

Fourier Transforms

(Lecture with S. Helgason) Fourier series are dened as f (x) an einx ,

1 an = 2

f (x)einx dx

for 2 period functions. Now take we can take an arbitrary interval, then our dense expo nentials are 1 e(inx)/A . 2 Then for L2 (A, A) 1 inx A inx 1 e A f (x) f (y) e A dy 2 2A A Let
A

f (y)e
A

iny A

dy = g

n A

1 Then f (x) 2 g(n/A)e(inx)/A (/A), then think of n/A as a new variable u = n/A, du = , and let A dene g(u) = f (x)eixu dx R

and we hope that 1 f (x) 2

R

g(u)eixu du

42

 Denition. f L1 (R) dene the fourier transform f by f (y) =

R

f (x)eixy dx,

yR

f is bounded and uniformly continuous because

ixy (y) f |f (x)||e |dx |f (x)|dx <
R R

f L1 (R)

and to show uniform continuity

ix(y+h) ixy ixy ixh |f (y + h) f (y)| = f (x)(e e )dx =
f (x)e (e 1)dx


R
R |f (x)||eixh 1|dx 2|f (x)|
R

so by Lebesgue dominated convergence theorem, we can take limits ixh lim |f (x)||e 1|dx = lim |f (x)||eixh 1|dx = 0
h0 R R h0

so f is continuous, and in fact uniformly continuous, because of the bound 2|f (x)| (the y dropped out). Denition. Schwarz Space A function f is a Schwarz function if for each m 0, l 0 xm
2

dl f , dxl

bounded on R

Note 3 facts Cc (R) S, ex /2 S and S L1 (R). Also, it is not too hard to see by the denition that if f S(R) then

|f (x)| C(1 + |x|2 )N for some C.

Lemma. f S, g(x) = xf (x), then
(y) = f f (x)eixy dx,
R

d f (y) = dy

f (x)(ix)eixy dx

In fact

f (x)(ix)eixy dx is uniformly convergent, and now

df df (y) = ixf (y), xf (y) = i dy
dy
43

Lemma. f S, h = df /dx Then by integration by parts df df ixy ixy h(y) = = f (x)(iy)(eixy )dx e dx = f (x)e dx dx R R and as x , f (x) 0, so then the expression above winds up being iyf (y). Thus df (y) = iyf (y). dx So sum up what we have done above in the following theorem Theorem. dk tk f (t)() = (i)k k f () d dn f (t) () = (i)n f () dtn

This helps in dierential equations, the fourier transforms interchanges dierentiation and multiplication, and vice versa. But the function we deal with has to be small at innity, however the Schwarz space allows for more general fourier transforms, like polynomials. However we dont know that functions are fourier transforms of L1 , that is L1 (R) =?. However, for Schwarz functions... Theorem. f f is a bijection of S onto S and 1 f (x) = f (y)eixy dy 2 R This is the fourier inversion formula. Before the proof we present an example Example. f the characteristic function of [1, 1]. Then 1 sin y (y) = f eiyx dx = 2 y 1 Then we would like to show that 1 f= 2
R

sin y iyx e dy y

But there is a snag at y = 0. Consider A A sin y sin y (cos yx + i sin yx)dy = cos yx dy A y A y

44

(since i sin yx (sin y/y) is odd). So then to take care of y = 0 A A 1 1 sin(x + 1)y sin(x 1)y sin y 2 2 cos y = 2 + 2 dy y y y 0 0 sin y = sgn y 0 so as A we get sgn(x + 1) sgn(x 1) = [1,1] . We continue with the proof of the theorem Proof. If f S, then f S. And the following are both fourier transforms of Schwarz functions: d df (y) f = iyf (y), = iyf (y) dy dy By iterations of these we can show that y m d fl is bounded, in fact dy ym dn f dy n
l

But

is the fourier transform of a Schwarz function, so it is bounded and thus f is a Schwarz function. Assume the inversion formula is proved. Then f f is onetoone, also f f is surjective because g S, g = f , f = g, since f f is onetoone. Now we need to prove the inversion formula. x2 First we nd the fourier transform of e 2 = f (x). This is the solution to the dierential equation df xf + =0 dx Now if h is a solution then x2 dh d x2 (e 2 h) = e 2 xh + =0 dx dx So e 2 h is constant, then h = Ce 2 , now df df xf + =i + iyf (y) = 0 dx dy
y So f satises the same dierential equation as f , so f (y) = Ce 2 . Calculate this at the origin to nd the constant f (0) = 2, so y2 f (y) = 2e 2 2 x2 x2

We need a few lemmas to nish the proof 45

Lemma. Symmetry Lemma g, f S then f (x)(x)dx = g f (y)g(y)dy

R R

Proof. We can just use Fubinis theorem with ixy f (x)e dx g(y)dy
R R

Lemma. f S, a R, f a (x) = f (x + a) then f (y + a) = f a (y) = eiay f (y) Proof. Just by denition of the transform and change of variables. Lemma. f S, fa (x) = f (x/a), then f (y/a) = fa (x) = af (ay) Proof. Same as above Now a special case of the symmetry law gives 2 a2 x2 y2 f (y)e 2a dy = 2a f (x)e 2 dx
R R

Now if we do a change of variables and let s = ax, then the above becomes y2 s s2 (y)e 2a2 dy = f e 2 ds f a R R Let a . Then the above becomes 2 (y)dy = 2f (0) e s2 ds = 2f (0) f
R R

so the inversion holds at point 0. Use lemma about f a (x), so a f (0) = f (a) = eiay f (y)dy
R

Theorem. Let f S, then 2f (x)L2 = 2

R

|f (x)| dx =
R

|f (y)|2 dy = f (y)L2

Proof. Observe 1 f (x) = 2 1 f (y)eixy dy = (f )(x). 2 R

take g = (f ). Then by symmetry g = f (x), and 2 f (x)f (x)dx = f (y)f (y)dy

R R

46

8 Extending Fourier Transform to L2(R), Proving the Schwartz Functions are Dense in L2
The Schwartz functions

m dl S(R) = f : R C, f C , supx
dxl f (x)
< m, l
xR Here are our main propositions for this section. These should always be in the back of our minds. Proposition. S(R) is dense in L2 (R) Theorem. The Fourier Transform F extends by continuity to an isomorphism F : L2 (R) L2 (R).; We remark that S(R) is a complete metric space (but not a Banach Space). Now we can dene a norm on S by p dq x f
f
= sup
k dxq xR p+qk Each of the terms of the sum is a norm of S(R). So S(R) with k is a normed space (but not complete). We dene (not a norm) a metric d(f, g) =
k=1

f gk k 2 , 1 + f gk

[0, )

This is well dened. To show its a metric we just have to show the triangle inequality. Lemma. If is a norm, then d(f, g) = is a distance. Proof. This is not very hard, its just algebra. Exercise Prove that S(R) is complete with respect to the metric dened above. Now we already know Theorem. F : S(R) S(R). Denition. A tempered distribution generalized function is a linear function u : S(R) C 47
f g 1 + f g

Now look at v L2 (R), dene v(f ) = f (x)v(x)dx,

R

du df (f ) = u dx dx

Then L2 (R) S (R) = {u : S(R) C}. S (R) is the dual space of S(R). Example. 0 : S(R) C. Then 0 (f ) = f (0). Because 1, x 0 H= 0, x < 0 And H(f ) =
R

f Hd =

f (x)dx,

f S(R)

Then in S (R)

df d df H(f ) = H = = f (0) = 0 (f ) dx dx dx 0

To show that F can be extended from a dense subset to a L2 . Now formally we want to show that if fn S(R), fn f in L2 . Then Ffn = fn satises fn fm 2 = 2fn fm 2 0, n

So then fn is Cauchy and thus convergent in L2 (R). So we want to dene Ff = lim fn

n

f = lim fn ,
n

fn S(R)

So that F : S(R) S(R) L2 (R)

8.1

Density of the Schwartz Functions

To prove this we need to show the existence of a sequence of functions fn Sn for each f L2 (R) such that fn f .
Theorem. Uniqueness If fn f , fn S, then

lim fn = lim fn = f
n

so that Ff is well dened.

Proof. fn f , then fn fn 0 in L2 (R). The Placherel implies fn f 2 = (2)fn f 2 0

48

The proof that S(R) is dense in L2 (R). Comes in 2 steps. First Step Theorem. If f L2 (R) and f (x), fN (x) = 0, then fN f in L2 (R) as N Proof. 0, |x| N f fN 2 = f (x), |x| > N 2 f fN = |f (x)|2 dx 0, as N
|x|N

|x| N |x| > N

because f is square integrable. Second Step Now it suces to show that the functions above can be approximated by in L2 by sequences in S(R). We use Fourier Series. Suppose f vanishes outside radius of 2R. Consider g(y) = f (2Ry). We are sort of retting the function so that g will vanish outside of [1, 1], g L2 . Then by results in Fourier series, g = cn einy converges in L2 , i.e. the C function gm (y) = iny C ([, ]) converges to g in L2 ([, ]).
|n|m cn e We claim the existence of a function C (R) such that 0 1 and
1, |x| 1 (x) = 0, |x| Lemma. gm g in L2 ([, ]) then gm g in L2 ([, ]). Proof.

| m g| dx = g

|gm g | dx

|gm g |2 0

Lemma. We claim that gm and that x f 2R L2 (R) as m

x fm = gm S(R) 2R

49

Proof. fm C (R) and fm (x) 0 with |x| 2R. So this is a smooth function with l dq x Cn,l a constant dxq f Cn,l , so fm S(R), and its easy to see convergence.
Now to nish the proof we need to construct , the cuto function.
Theorem. If a < b in R with > 0, : R [0, 1] which is C such that 1, x [a, b] (x) = 0, x (, a ] [b + , ) Step 1 Observe that
1 e x , x > 0 f (x) = 0, x<0 is in C . We prove this by induction. We want to show that pk (x) 1 dk 1 (e x ) = 2k e x dxk x where pk is a polynomial, this is true because if we take a derivative dk+1 1 x2 pk 2kxpk + pk 1 (e x ) = e x dxk+1 x2k+2 Now we have pk (x) 1 dk f = lim e x 0 k x0 dx x0+ x2k Step 2 We can explicitly construct now: lim + = f (x a + )f (x + b + ) f (x + a + ) + f (x a + ) + f (x + b + ) + f (x b + )

we see that this satises what we want. And we are done. Review Proof We cut o the crap near innity and get a sequence which converges to some function. So now it suces to assume that f = 0 in |x| N . Now approximate in L2 using fourier series in L1 , which we can do since L2 ([N, N ]) L1 ([N, N ]), f n 1 (2N )1/2 f 2

(some stu which I dont know what it means:) f S

ix f () = f (x)e dx, sup |f ()| |f |
R R R

n f in L1 then

sup |f n | f n 1 0
R

50

The Harmonic Oscillator

We know that L2 (R) is Hilbert space and it is separable, so it MUST have a complete orthonormal basis. For L2 ([, ]) we look that the 2 periodic solutions to the equation d2 f = f dx2 and these generated such a basis. Now, for L2 (R) we have the Harmonic Oscillator. This is the operator d2 2 H = 2 +x dx we want to solutions to Hu = u: d2 2 2 + x u = u, dx

u S(R)

To do this, consider the two operators Annihilation and Creation d d A= +x , C = +x dx dx Now, a computation to establish at least one eigenfunction d2 d 2 2 2 2 2 + x ex /2 = (xex /2 ) + x2 ex /2 dx dx = x2 ex
2 2 /2

+ ex

2 /2

+ x2 ex

2 /2

so ex /2 is an eigenfunction for = 1. Now look at the operation of the operators A and C on each other d d d2 u d2 du du 2 2 +x u= 2 +x +ux +x u= 2 +x +1 u ACu = +x dx dx dx dx dx dx So we have the relation (when we include the computation for CAu) CAu = (H 1)u, so [A, C]u (AC CA)u = 2u we already know that Hex
2 /2

ACu = (H + 1)u

= ex

2 /2

, and Aex

2 /2

= 0 and

HCu = (CA + 1)Cu = CAC + Cu = C(AC + 1)u = C(H + 2)u

51

So we now know that if u S(R) then HCu = C(H + 2)u Set u = ex

2 /2

then we get the following relations by induction: (H 1)ex (H 3)(Ce

2 /2

=0

x2 /2

)=0 . . . ) = (2j + 1)C j ex

2 2 /2

H(C j ex Proposition. The functions

2 /2

C j ex /2 j = C j ex2 /2 are a complete orthonormal basis of L2 (R). First C j ex /2 L2 (R) because C j ex Polynomials. And by induction C j+1 ex
2 /2 2 2 /2

= pj (x)ex

2 /2

, pj (unnormalized) Hermite
2 /2

= C(C j ex

2 /2

) = C(pj ex

2 /2

) = (xpj pj )ex

+ (2xpj + pj )ex

2 /2

So pj = 2j xj + lower order terms. NB p0 , . . . , pN span all polynomials of degree N , because for a polynomial qN , qN = c pN + qN 1 2N

and so on (just pick rst term then pick second...) Now we need to show orthonormality, that is j k dx = 0, j , k = j k dx =
R R 1 Hj , 2j+1

j=k

[H (2j + 1)]j = 0, so we write j =

R

and the above becomes

d2 2 2 j + x j k dx dx

Hj 2j + 1

1 k dx = 2j + 1

these are all Schwartz functions, so we cut o at innity and integrate by parts, N d d lim ()k dx, = j N N dx dx N d
= lim k
N N dx 52

(basically, this shifts dierentation to the left)

then we get 1 d2 1 2 j 2 + x k dx = j Hk dx 2j + 1 R dx 2j + 1 R But Hk = (2k + 1)k so the above is just 2k + 1 2j + 1 then n , k = j k dx
R

2k + 1 j , k = j , k = 0 2j + 1 Now we have to prove completeness. i.e. in u L2 (R) and R uj = 0, j, then u = 0. In fact if we have shown this then N 2 u cj j dx = 0 and upN ex /2 dx = 0
R j=0 R

for all polynomials, because we know that combinations of the j are all polynomials. Can you show that this implies u = 0 (a.e. in L2 (R))? Suppose that u has compact support (so its nonzero in a bounded set) then N 2 (uex /2 )p(x)dx = 0, polynomials
N

But the polynomials are dense in L2 . (Think about it, this implies completeness of our problem) So any f L2 (R) can be written as the sum of its Fourier Series 2 C j ex /2 f= cj j , cj = f j dx j = C j ex2 /2 R i=0 There are many consequences of this. For example if f L2 (R) the solution to the equation Hu = f is cj (f ) u= j , cj (f ) = f j dx 2j + 1 R i=0 In some weak sense Hu = f . Its weak because we dont really k now about dierentiation. So our weak formulation is really that d2 2 2 +x u=f u, f L2 (R) dx d2 2 S(R) u 2 + x dx = f , dx 53

Completeness of solutions of the harmonic oscillator. 0 , 1 , 2 , . . . in S(R) with 0 the 1 ground state, which means that H0 = 0 , so 0 = 2 exp(x2 /2), 0 2 = 1, j we dene as C j 0 , Hj = (2j + 1)j j = C j 0 L and span{0 , . . . , n } = {p(x)ex
2 /2

|p poly deg(p) N }

Theorem. 0 , 1 , . . . form a complete set in L2 (R) (scrap the proof from last time). Proof. We use complex variables. Show that if f L2 (R) and f, j = 0 j N then f = 0a.e. 2 Consider the Fourier transform of F (x) = f (x)ex /2 = f (x)0 . F (x) L2 (R) L1 (R), because by CauchySchwarz |f (x)e F (s) =
x2 /2

|f |

1/2

x2 /2

1/2 <

F (x)eixs dx exists in C(R). In fact the integral exists for all s C, 2 f (x)ex2 /2ixs dx |f |ex /2+|x||s| dx
R R

Apply CauchySchwarz again and we get the above bounded by

R

|f |

1/2
R

x2 +2|x||s|

1/2

and

x2 +2|x||s|

dx 2

x2 +|x||s|

dx = e

|s|2 /4

|s|2 |s| 2 e(x+ 2 ) dx 2 e 4

so the integral exists for all complex c. Now we want to show that F (s) is entire if s C, i.e. it is the sum of its Taylor series at s = 0. We claim that 1. That the taylor series at 0 for F (s) converges s C. 2. Formally, F (n) (0) =
R

f (x)ex

2 /2

(1)n xn = (1)n f, xn ex

2 /2

=0

where the nal equality comes from the assumption f should be orthogonal to things 2 of the form pn ex /2

54

2 An outline of our entire argument is F 0 then F (x) = 0 in L2 (R) f (x)ex /2 = 0 almost everywhere, implies f = 0 L2 (R) which implies that {i } is complete We prove the two things above. First plug in the Taylor series for eisx , then we get

e so F (s) =

isx

N 1 n=0

(isx)n + RN (x, s) n!
R

N 1 n=0

(i)n sn n!

f (x)e R

x2 /2 n

f,xn ex

2 /2

x dx +

f (x)ex

2 /2

RN (x, s)dx

we want to prove that the remainder term at 0 decreases quickly. Using the Legendre form for the remainder we get
N 1 (n) f (0) f (1)
n!

n=0

so with our particular case with f (t) = e we have

1 z N 1 z n 1 e |sN 1 z N etzs |
ds
n!
(N 1)! 0 n=0 |rN (z)| e|z| |z |n (N 1)!

tz

in our original equation RN (x, s) = rN (isx) and so our remainder term when z = xs becomes (and then we apply Cauchy Schwartz) |s|N (N 1)!

f (z)e

x2 /2 |x||s| N

x dx

|f | dx

1/2 x

2N x2 +2|x||s|

1/2

C(N !)1/2 (N 1)!

We get the last bound because we can estimate using integration by parts (brings down the N s) (kind of like the function). So we have in fact proved that F (s) = so in fact 0 , 1 , . . . is complete.
sn n=0

n!

F (n) (0)

10

Sobolev Spaces

Let S be the circle, that is S = R/2Z is the equivalence class p = y if p y 0 mod 2Z. A circle in the unique 1dimensional bounded compact manifold. 55

So functions on S can be identied with 2 periodic functions on R. For L1 and L2 we have a simple identication using Fourier series. Because when constructing Fourier series, we did it by thinking about L2 ([, ]) and L1 ([, ]), so L2 ([, ]) L2 (S) {f : R C, period 2} = = L1 ([, ]) L1 (S) =

Fourier Series: f L2 (S) L2 ([, ]) then f is the limit of its Fourier series in L2 (S) = Denition. We dene C (S) as follows C (S) = {u : R C, u innitely dierentiable, period 2} Then we conclude that C (S) L2 (S) is dense (in the L2 topology). Just a consequence of fourier series, and the fact that cn einz C (S).
Denition. Integration on S We dene integration, as would be obvious, by

f dx = f dx, f L1 (S) L1 ([, ]) =
S

Theorem. Integration by Parts Integration by parts works on S, i.e. for f, g C (S) then dg df f dx = gdx S dx S dx Proof. Follows from 2 periodicity

dg df d df df
f dx = (f g)dx gdx = gdx + f g = gdx

dx S dx dx dx S dx

We dont really have a formulation of a derivative of any sort, but we can dene these: Denition. For f L2 (S) we say that df = h L2 dx weakly if h L2 (S) such that
d f dx = hdx, S dx S

C (S)

This is just an application of integration by parts. Note if f is smooth this is true. We also have a strong denition of derivatives Denition. For f L2 (S), h is the strong derivative of f if fn C (S) such that fn f in L2 (S) and dfn h in the L2 (S) norm. dx 56

Theorem. h is unique if it exists. Proof. If we have two derivatives h1 and h2 then h1 = h2 ,

C (S)

but C (S) is dense in L2 (S) so (h1 h2 )dx = 0 = h1 h2 = 0 a.e.

Theorem. Strong derivatives imply existence of a weak derivative. Proof. If C (S) and h is the strong derivative of f L2 (S) then d d dfn f dx = lim fn dx = lim = hdx n S n dx dx S dx S because
dfn dx

converges in L2 .

Using Fourier series we can see that Theorem. If f L2 (R) then f has a strong derivative in L2 if and only if it has a weak derivative in L2 (S) if and only if the Fourier series 1 inx incn e , cn = f einx dx 2 nZ S converges in L2 (S). NOTE: Convergence of the above in L2 happens if and only if .
Proof. If f has a weak derivative in L2 (S) then h L2 (S) such that
d C (S) f = h S S dx If we set = einx then the above becomes inx in f e dx = heinx dx = dn
S S

nZ

n2 |cn |2 <

where dn are the nth fourier coecient of the derivative. Then we see that incn = dn , n. But h L2 (S) then |dn |2 < = n2 |cn |2 <
nZ nZ

57

existence of the weak derivative implies that

n2 |cn |2 < = h = incn einx dx, in L2
nZ nZ

now h is a strong derivative of f since fN = 1

cn einx f, in L2 (S) 2
|n|N

and fN C (S) and because this is a nite sum we can dierentiate term by term 1
dfN = incn einx h, in L2 (S) dx 2
|n|N

So there is no ambiguity in saying f L2 (S) has a derivative df /dx L2 (S). Denition. Sobolev Space H (S) =
1

df 2 L (S) weakly/strong f L (S) : dx

2

and so by the theorem above f H 1 (S) if and only if f= 1 cn einx , 2 nZ

nZ

n2 |cn |2 <

Theorem. H 1 (S) is a Hilbert space with the inner product df dg df dg f, gH1 = , + f, g = + fg dx dx S dx dx S and the norm f 2 1 H 2
df
dx + |f |2 dx =
S dx
S

The extra term just accounts for constant function, whose norm should not be 0. Proof. This is clearly a norm. What about completeness? If fn H 1 (S) is Cauchy then fn
f in H 1 (S). If fn is Cauchy, then dfn /dx and fn are Cauchy in L2 because dfn f H 1 , f L2 f H 1
dx

2 L so fn converges to some f in L2 (S) (by completeness of L2 (S) and dfn /dx converges to some h in L2 (S). But this is why we thought about the weak derivative in the rst place, so d dfn n C (S) = , fn dx dx S S 58

so

df dx 1

d f = S dx

h,
S dfn dx

C (S)
df dx

so h =

for f H (S). So fn
f and

f H 1 f L2

implies that fn f in H 1 (S) since df

+
dx

2 L

10.1

Higher Sobolev Spaces

We know that H 0 (S) = L2 (S). We can iteratively dene f H k (S), k N f H k1 (S) k1 f and d k1 H 1 (S). dx So for f H k (S), dene L2 (S) dk f d dk1 f L2 (S) dxk dx dxk1

Theorem. f H k (S) if and only if f has weak derivatives of all order l k. That is hl L1 (S) such that

dl l f l = (1) hl , C (S) dx S S Proof.
dk f k dx = dx
dk1 f k1 dx

d k1 dx = (1) dx d hk1 = (1)k dx hk

dhk1 /dxk1

Theorem. f H k (S) f = Proof. Just use induction.

1 2

inx nZ cn e

and

nZ

n2k |cn |2 <

Denition. All these H k (S) are Hilbert spaces with the norms
f 2 k = H 1
2k (1 + n2 + n4 + n
)|cn |2 4 2 nZ

This is equivalent to the (Hilbert) Norm f 2 k = H 1

1
(1 + n2k )|cn |2 2 (1 + n2 )k |cn |2 4 2 nZ 4 nZ

This makes sense for k [0, ). 59

Denition. For r R+ dene H r (S) L2 (S) such that f H r (S) if fp f in L2 (S), c fp C (S) and fp 2 r = (1 + n2 )r | n (fp )|2 < i.e. f L2 (S) and Fourier series satises H 2 r 2 nZ (1 + n ) |cn | < . Elementary Properties H r (S) H r (S) if r r. H r has the inner product
f, gr = f H r (S) then
nZ

1
(1 + n2 )r cn dn 2 nZ

|(1 + n2 )r/2 cn |2 <

Completeness (Left as exercise) What this got to do with derivatives? C (S) H r (S) r but in fact H r (S) = C (S)
r>0

this is not too hard to prove. We have to show that H r (S) C (S) C (S) H r (S)
r>0 r>0

neither of which is too hard. Exercise: Prove the two inclusions above.

11

Poisson Summation

This is really a correlation between fourier transforms and fourier series For transforms: 1 itz (z) = eitz f (z)dz f e f (t)dt, f (t) = 2 R R for series (g : R C 2periodic, g L1 ([, ]). 1 itk ck = g(t)eitz dt, g(t) = ck e 2 nN If we assume g is C , then we can integrate by parts in ck : p dp itk d g itk p p p k ck = i g(t) p e dt = (i) e dt = |k p ck | < cp , k dt dtp itk for some constant cp . If g is 2periodic and C then |ck | < cp /k p p and g(t) = 21 nZ ck e converges uniformly with all derivatives. 60

Theorem. If f S(R) then g(t) =

kZ

f (t + 2k)

converges with all derivatives to a 2periodic function and the fourier coecients ck (g) = f (k), k Z. Proof. Formally dp g dp f = (t + 2k) dtp dtp kZ
S(R)

So this converges if we can prove uniform convergence in the Schwartz functions. But we can prove this. We know that suptR |t2 f (t)| < C. Shift this and get suptR |(t + 2k)2 f (t + 2k)| < C. If t [, ], |k | > 2, then |t + 2k | 2 |k 1|. So supt[,] |k 1|2 |f (t + C 2k)| < C implies that |f (t + 2k)| |k1|2 . Now the real meat of Poisson summation: What are the Fourier coecients of g? Well, ck = g(t)eitk dt, since this is a uniformly convergent series we can write ck = f (t + 2p)eitk dt
pZ

Let T = t + 2p, t [, ] be a change of variables, then we have +2p +2p i(T 2p)k ck = f (T )e dT = f (T )eiT k dT
pZ +2p pZ +2p

as p runs through the integers, the integrals domain becomes all of R (the intervals [ + 2p, + 2p] are a countable number of intervals that decompose R). Thus ck = f (T )eiT k dT = f (k)
R

So we get 1 f (t + 2k) = f (k)eikt 2 kZ kZ Now we can do things like g(0) = 1 ck = f (k) 2 kZ kZ

Note: The actual Poisson Summation Formula is usually stated as 1 f (2p) = f (k) 2 kZ pZ In general the left side is a relation of classical geometry and the right is a statement of analysis. 61

12

The Wave Equation

d2 u d2 u 2 = 0, dt2 dx

The wave equation in one space and one time variable is u : (Rx , Rt ) R

Theorem. If f0 , f1 C (Rx ), !u C (R2 ) satisfying the wave equation with the initial x,t conditions u(0, x) = f0 (x) and x u(0, x) = f1 (x), x R. Proof. We can write all solutions of the wave equation as u(t, x) = (t + x) + (t x) where u C (R2 ) and , C (R). If u(t, x) = (t + x) + (t x) obviously 2 2 (t + x) = (t x) t2 x2 Conversely if we have
d2 u dt2

d2 u dx2

= 0,

u : (Rx , Rt ) R then consider u u + t x 2u v 2u = + 2 x tx x v = x . If we dene v(t, x) = (s, r) where

v(t, x) = then v 2u 2u = 2 + , t t tx then by the equality of mixed derivatives, s = t + x and r = t x, then

v t

v s r = + = + t s t r t s r v s r = = x s x r x s r if v/t = v/x then /r = 0. So is only a function of s, so it is a function of t + x. 1 So v = u/t + u/x = (t + x). If = then u = 2 (t + x) satises that equation. So if we let u = 1 (t + x) + u then u /t + u /x = 0. If we apply the same argument with the signs switched then u = (t x), and so u = 1 (t + x) (t x). Now the initial conditions: u(0) = (x) + (x) = f0 u (0, x) = (x) + (x) = f1 t
If we dierentiate the rst equation we get (x) (x) = f0 then

1 (x) = (f1 + f0 ) 2

1 (x) = (f1 (x) f0 (x)) 2 62

If we take the following it solves the equation

x 1 (x) = (f1 + f0 )dx 2 0

Exercise + is not unique, but u is unique.

12.1

Fourier Series solutions

fi (x + 2) = fi (x), x R = u(t, x + 2) = u(t, x), t, x R2

From the uniqueness of u if f0 and f1 are both 2periodic then

Because the equation is translation invariant, u(t, x) = u(t, x + 2) satises 2u 2u (t, x) = 2 (t, x + 2), t2 t 2u 2u (t, x) = (t, x + 2) x2 x2

u(0, x) = f0 (x + 2). Now, we can expand the solution in the fourier series ixk u(t, x) = 21 kZ ck (t)e ck (t) = u(t, x)eixk dx when we combine our wave equation condition with 2 u/t2 = 2 u/x2 with the below, we get 2 2 d2 ck (t) u u ixk = (t, x)e dx = (t, x)eixk dx dt2 t2 x2 Integrate by parts (there are no boundary terms by periodicity) and this is
d2 ixk 2
u(t, x) 2 e u(t, x)eixk dx = k 2 ck (t) dx = k dx So the Fourier coecients of u(t, x) with f0 , f1 are 2periodic (initial conditions/driving system 2 periodic) satisfy 2 d 2 + k ck (t) = 0 dt2 So the general solution for the Fourier coecients are of the form ck (t) = ak eitk + bk eitk . So if the initial data is 2periodic and smooth, then the solutions are 2 periodic in t (as well as space).

63

Now to solve for ak , bk : ixk ck (0) = u(0, x)e dx = f0 (x)eixk dx = ck (f0 )

d du ixk ck (0) = ck (0) = (0, x)e dx = f1 eixk dx = ck (f1 ) dt dt but ck (0) = ak + bk and ck (0) = iak ibk , and so the coecients are 1 1 ak = (ck (0) + ck (0)), 2 ik 1 1 bk = (ck (0) ck (0)) 2 ik

and for k = 0, ck (0) = a0 and ck (0) = b0 . And the solution is 2periodic in t if and only if b0 = f1 (x)dx = 0. So now we have 2 ways of solving the equation for 2period input data 1. By using u(t, x) = (t + x) + (t x). 2. If we set k =

f0 e

ixk

dx, k =

f1 eixk dx

then the Fourier series solution is u(t, x) =

1 1 (k + 1 k ) ei(t+x)k + 1 (k 1
k ) ei(xt)k + 1 (0 + 0 t)
2 2 2 k=0 2 ik
ik



ak bk k=0 term d2 dx2

Note: k 2 are the eigenvalues of

on [, ].

13
13.1

Operators
Preliminaries

Before we discuss operators, we have a few preliminary theorems and lemmas. Lemma. If f L2 (S), i.e. periodic on R, period 2, then f is continuousinthemean, i.e.
|f (x + t) f (x)|2 dx 0

Proof. This is a good exercise to do prove directly using integration theory. But we can do it using Fourier series. f (x) = 1

ikx ck e , 2 kZ f (x + t) = 64
1
ikt ikx (ck e )e 2 kZ

So by plancherel |f (x + t) f (x)|2 dx = C |dk |2

nZ

dk = F.C. of f (x + t) f (x)

=C

kZ

|1 eikt |2 |ck |2

For each k, 1 eikt tends to 0 as t 0. We know that the fourier series for f (x) converges, so |ck |2 < . So given > 0, N such that |k|N |ck |2 < /2, so C
nZ

|dk |2 C

|k|<N

|1 eikt |2 |ck |2 +

|k|N

|1 eikt |2 |ck |2

where the left tends to 0 and the right < . So we are done. We will be discussing compact operators later, so we need some sort of idea of com pactness. This is what the following theorem provides. Theorem. A subset S H of a Hilbert space is compact (here by compact we will mean sequentially compact) if and only if it is 1. Closed 2. Bounded 3. Satises the following condition
C) If {j } is an orthonormal basis, then given > 0, N such that
|j , f |2 < f S.
jN

Note if f xed this is true, but in fact we are trying to say this for all f , so there is a uniformly small tail. Proof. DO THIS. Denition. A sequence {fn } in H is weakly convergent if it is bounded and fn , c converges in C, H This implies: * fn , j cj : The Fourier coecients converge with respect to any complete orthonormal basis Proposition. 1. * and boundedness weak convergence

2. Any closed bounded sequence is weakly compact in the sense that any sequence has a weakly convergence subsequence. 3. fn weakly convergent implies fn , f, , . 65

Proof. Of 1). We want to show that given H and > 0 then fn fm , 0. N fn fm , = fn fm , , j j + fn fm , , j j

j=1 N +1

The second half is bounded by 2 supn fn And the left is bounded by /2 for n, m large.

N +1 , j j ,

which gets small as N .

13.2

Operators

Remember Riesz representation theorem in Hilbert Space. Theorem. If B : H H, H a Hilbert space, B a linear map then B(c1 f1 + c2 f2 ) = c1 Bf1 + c2 Bf2 , then B is continuous is equivalent to saying B is bounded, i.e. C such that Bf H Cf H , f H. Proof. Suppose B is continuous and linear then B is continuous at 0, and B(0) = 0, conti nuity implies that B 1 ({g < }) {f < } for some > 0. That is f H /2 implies f {f < } implies that Bf {g < 1} so Bf < 1, so f H /2 Bf 1. But this is boundedness. For general f H, f = 0

/2f f /2 /2 B 1 = Bf
B f f f
2 2 Bf f Conversely, if Bf Cf , f H, then B(f gj ) Cf gj and Bf Bgj Cf gj so if gj f then f gj 0 Bf Bfj 0 Bgj Bf . So B is continuous. Example. Before we showed that f0 , f1 C (R) is of period 2, then unique solution u(t, x) to the wave equation with u(0, x) = f0 and t u(0, x) = f1 . Claim: if we dene
T : C (R) C (R) C (R) C (R) (f0 , f1 ) (u(T, ), T u(T, ))

so

This is linear. Either directly or from Fourier, this map is bounded t u(T, )2 + u(T, c)2 + x u(T, )2 C(x f0 2 + f0 2 + f1 2 ) So now we know if a linear map B : H H is continuous if and only if its bounded (Bu Cu, u H). Denition. Operator Norm B = sup{Bu; u H, u = 1} 66

so Bu Bu, u H Let B(H) be the space of all bounded operators on H. Then B(H) is a Banach space with norm B. It is clear that is a norm, because (B1 + B2 )uH = B1 u + B2 uH . By the triangle inequality on H, B1 u + B2 u (B1 + B2 )u. B B Note also that that B(H) is an algebra under operator composition H 1 H 2 H. And B2 B1 H, because B2 B1 uH B2 B1 u B2 B1 u. This is called the strong property of bounded operators: B2 B1 B2 B1 . Now we talk about adjoint operators. Denition. Adjoint Operation. If B B(H), B B(H) such that Bu, vH = u, B vH Now we have to dene explicitly what B is. Denition. The Adjoint Given v H dene B v = w H, we want u, w = Bu, v u H. Given v H H u Bu, v C This is a continuous linear map (functional). But remember the Riesz Representation Theorem: If F : H C is continuous and linear, w H such that F (u) = u, w, u H. So Bu, v = u, w for unique w by Riesz. So this is how we dene B v = w. Check 1. Linearity Bu, c1 v1 +c2 v2 = c1 Bu, v1 +c2 Bu, v2 = u, B (c1 v1 +c2 v2 ) = u, c1 B v1 +c2 B v2 . 2. Bounded
|Bu, v| = |u, w| Buv, u
Set u = w, then w2 Bw which implies w = B v Bv so v B B. This is sort of a generalized integration by parts. Question. When does B B(H) map {u H; u 1} into a compact set? Such operators are said to be compact operators. Example. Suppose B has nite rank (i.e. B(H) = V H is nite dimensional). Choose an orthonormal basis 1 , . . . , N , N +1 , . . . for B(H). So f B({u 1}) have equismall
span V

u, v H

tails (in fact no tails at all) 67

Theorem. An element B B(H) is compact if and only if SN a sequence of nite rank operators such that SN B in B(H) (SN B 0, N ). i.e. the space of compact operators is the closure in B(H) of the nite rank operators. Denition. Call C(H) the set of all compact operators. Corollary. C(H) is a closed subalgebra of B(H) and a two sided ideal. That is we have to show that B compact implies B compact. To show its a two sided ideal we must show B1 or B2 compact then B1 B2 compact, and B1 compact and B2 bounded implies that B1 B2 compact. Proof. B1 B2 ({u C}) B1 (B2 (u C)) B1 ({v H; v Bc}) compact set

Claim that S of nite rank then S has nite rank. Also if B1 B2 B(H), then (B1 B2 ) = B2 B1 . Now B2 B1 = (B1 B2 ) 2 sided ideal. Compact operators behave like matrices. In fact nite rank operators are essentially matrices. Now some more theorems on bounded operators. Example. Suppose {i } is a complete orthonormal basis, then by Bessel, i=1 f= converges. Now consider the following function, which will be important later (N ) = f, i i
iN i=1

f, i i

This is a bounded operator, because (N ) f 2 =

iN 2 and (N ) N = N , and (N ) = 1 and it is not too hard to see that (N ) = (N ) , which makes (N ) idempotent or a projection and (N ) = (N ) . Can restate compactness of S H as S closed, bounded and given > 0, N such that (N ) f < .

|f, i |2 f 2

Theorem. An operator B B(H) is compact, (i.e. B({u 1}) is compact) if and only if there exists a sequence of nite rank operators, Tn such that B Tn 0. 68

Proof. Consider the opposite of (N ) , N f =

jN 2 is of nite rank. Also N = N and Id = N + (N ) and N (N ) = 0. We claim that if B is compact, then N B = Tn B in norm as N . B TN = B N B = (Id N )B = (N ) B. By compactness of B({u; u 1}), if f B({u; u 1}) we know (N ) f < if N large uniformly f B({u 1}). So u H, u = 1, (N ) Bu < . So u H, (N ) Bu < u (N ) B . (We have the converse left to prove).

f, J j

Note about nite rank operators. By denition, B B(H) is of nite rank if B(H) = V is nite dimensional. The following lemma is very important. Lemma. B nite rank ui , vi H, i = 1, . . . , N such that () Bf =
N i=1

f, ui vi

Proof. Clear that () implies B(H) span{v1 , . . . , vn }. Conversely suppose B(H) = V is nite dimensional. Choose an orthonormal basis of V , 1 , . . . , N . f H consider Fi : H f Bf, i C We can see that this is bounded and thus continuous |Fi (f )| = |Bf, i | Bf i Bf Then Fi is a continuous linear functional, then Riesz implies that Fi (f ) = f, ui for some ui H. However, Bf V , so because i is complete and orthonormal Bf =
N i=1

Bf, i i =

N i=1

Fi (f )i =

N i=1

f, ui i

Hence B of nite rank then B of nite rank, because Bu, v = u, B v, u, v H

69

So then

N i=1

u, ui vi , v = u, B v B v=
N i=1

u, v H

v, vi ui

So we can fully justify the statement above that B compact implies that B compact. So we can also justify C(H) is a twosided ideal. Because it is easy to see that B1 B2 is compact if B1 is compact. To show that B2 B1 is compact if B2 is compact, just show that B2 B1 = (B1 B2 ) , which is compact, because the right is compact. Theorem. If A C(H) is a compact operator and A = A (selfadjoint) then a complete orthonormal basis, {i } such that each i is an eigenvector, Ai = i , i, i R. We can also show 1. The set of all i which does this is discrete and can only accumulate at 0 2. And for every = 0, dim{u H; Au = u} < . Proof. First Step Find a nontrivial eigenvalue/eigenvector (if A = 0). Consider the function {u H; u = 1} u Au, u R (Au, u = u, Au = A u, u = Au, u, so Au, u is in fact real). Suppose Au, u assumes a positive value. Dene sup
uH,u=1

Au, u = C

Auu A

Then by denition of sup there exists a sequence uj H, uj = 1 such that Auj , uj C. By diagonalization choose a subsequence ujn such that N ujn N u for some u H, N (weak convergence). Note: the previous is equivalent to ujk , v u, v, v H. Rename ujn as uj . We claim that Au, u = C, i.e. the supremum is attained. This is true because com pactness of A implies Auj Au (in norm). Why? We just prove that Auj is Cauchy. Aun Aum 2 = N (Aun Aum )2 + (N ) (Aun Aum )2 N (Aun Aum )2 + 2 by choice of N large 2 2 + by choice of n, m by weak convergence By completeness Auj w. And we know Auj , v = uj , Av u, Av = Au, v 70

so w = Au. So all weve done is extract a weakly convergent sequence, Auj Au and uj u ( means weak convergence) To show Auj , uj Au, u, |Auj , uj Au, u| = |Auj Au, uj | + |Au, uj u| Aj Au The 1st term converges because of strong convergence of Aj , and second term because uj u and Au xed. So the supremum of Au, u on u = 1 is attained. Now we have found u such that u = 1 and Au, u = supv=1 Av, v. Step 2 Claim: u is an eigenvector for A. Take H, = 1, with u. Consider u + t H. u + t2 = u2 + t2 2 = 1 + t2 Then u + t

(1 + t2 )1/2
= 1
1 1 A(u + t), u + t = [Au, u + tAu, + tA, u + t2 A, ] 2) 2) (1 + t (1 + t

and this is a little curve on the unit sphere(parametically dened), call it ut . Then Aut , ut =

This is a dierentiable function, because it is a polynomial in t. Also Aut , ut Au, u, so ut has a local max at t = 0. So d A(u + t), u + t = Au, + A, u = 0
dt t=0 We can also consider u + it, because i = 1, i, u = 0. The same argument implies that d = Au, A, u = 0 A(u + it), u + it

dt t=0 So when we combine the above two statements we get Au, + A, u = 0 Au, A, u = 0 So then Au, = 0. Thus we have shown that u = Au The above is only true if Au = u for some C, because if u and Au, then u and Au have to be contained in the same 1dimensional subspace. Now, because Au, u is real, R+ . So indeed u is an eigenvector with Au = u, > 0. 71

Step 3: Call u = u1 . Consider the subspace of H, H1 = {v H; vu1 } This is a Hilbert space with the same induced norm and AH1 H1 because v H1 , vu1 implies that Av, u1 = v, Au1 = 0 and so v, u1 = 0. Also A = A is compact. We can repeat the process. We get an orthonormal sequence 1 , . . . , N . . . . And 1. If it terminates, then Au, u = 0, u1 , . . . , N 2. If it does not terminate we claim that Au, u = 0, uj , j N. We state this formally Lemma. 2 above holds, and furthermore if we have such a sequence 1 , 2 , . . . with eigen values 1 , 2 , . . . , then |j | 0 as j . Proof. NB |j | is decreasing (nonincreasing) by construction, since |j | = supu=1 |Au, u|. Now, if it is not true that |j | 0, i.e. c > 0 such that |j | c, j. Then because |j | is increasing, this implies that we can choose a subsequence with ji > 0 or ji < 0, on a subset of {ji }i of {n }n . So then, by the way we have constructed the eigenvalues, |Au, u| > c on span{j1 , . . . , jn , . . . } = G. But this contradicts the fact that |Aj | 0 as j (which follows from the compact ness of A) So Au, u = 0, u1 , . . . , n , . . . . Now, the above result (in particular the last line) implies that H = H1 H2 where H1 = span{1 , 2 , . . . } H2 = {u; uj , j} Because if u H1 then u1 =
j=0

u, j j = u uj j

We know that Au, u = 0 on H2 . Note that A : H2 H2 , since uj Au, j = u, Aj = j u, j = 0 72

So in fact A = 0 on all of H2 , because u, v H2 A(u + v), u + v =0 = Au, v + v, Au A(u + iv), u + iv =0 = Au, v + v, Au then Au, v = 0 implies that A = 0. We can get a corollary of the above for normal operators. Denition. An operator is normal if [A , A] = A A AA = 0 This is slightly weaker than A = A , it just says that A and A commute. Corollary. If A C(H) and A normal, then there exists a complete orthonormal basis of eigenvectors of A. The only dierence between this and the above is that j C, but still |j |
0. Proof. A C(H) then A C(H) which implies that
A + A = B C(H), 2 B = B

apply our theorem above to B. So there exists an orthonormal basis of eigenvectors for B, with Bj = j j , j R, |j | 0. We can also dene A A C(H), D= D = D i Now [A, A ] = 0 implies that [D, B] = 0. For each 0
R dene = H = {u H; Bu = u}
this is nite dimensional (trivially since there can only be a nite number of eigenvalues that are , or else |j | 0 would not be true). By commutivity if u H , then B(Du) = DBu = Du = Du, so in fact D : H H . So we can choose an orthonormal basis of H such that D = j . j j So for A, we get the following eigenvectors and eigenvalues A = B + iD : H H ,
A = ( + ij )j j

We can do this for every which is an eigenvalue of B, and we get a countable complete orthonormal basis. Note that on H2 (B) = {Bu = 0} we can just choose eigenvectors/values from D.

73

13.3

Examples

Example. HilberSchmidt Operators F L2 ([, ] [, ]). The following denes a compact operator:
2 2 A : L ([, ]) L ([, ]) Au(x) = F (x, y)u(y)dy
[,]

For L2 ([, ]) consider

Au, =

(x) F (x, y)u(y)dydx


= F (x, y)(x)u(y)dxdy
[,] [,]
= F (x, y)(x)u(y)dxdy
[,]2

Clearly (x)u(y) L2 ([, ]2 ), so

1/2 1/2 2 2 2 |Au, | = F (x, y)(x)u(y)dxdy
|F (x, y) dxdy |u(y)| |(x)| dxdy

The above is less than Cu2 2 , so |Au, | Cu2 2 Now x u L2 ([, ]) and dene a function L2 ([, ]) = H Au, Now, this mapping is continuous since it is bounded. So we can apply the Riesz Represen tation Theorem and we can nicely dene Au. A is compact, because it is the norm limit of a nite rank operator. To prove this, we need to know the following. Question. Can we use Fourier Series in both variables? Well, dene 1 ikx ily e e L2 ([, ]2 ) k, l Z 2 1 ei(kk )x e(ll )y dxdy kl , k l = (2)2 kl =

Then

is 0 if k = k OR l = l . Essentially is 0 if (k, l) = (k , l ). So kl is an orthonormal set on L2 ([, ]2 ). This is complete as well (Exercise: Prove it). 74

By Bessels inequality (regardless of completeness) () G(x, y) = F, kl kl

k,l

converges in L2 ([, ]). And in fact Lemma. () G(x, y)(x)u(y)dxdy = F (x, y)(y)u(y)dxdy u, L2

Proof. Certainly (*) is true for = j , u = k , for any k, j LHS = G(x, y)eijx eiky = F, jk = RHS Since = , j j u= u, j j LHS = RHS in general. So G gives the same operation as F . The fact that () converges in L2 ([, ]2 ) implies that A A(N ) G G(N ) 0 where A(N ) u = GN (x, y)u(y)dy GN =
|k|N,|l|N

F, kl kl

Observe that GN gives a nite rank operator A(N ) since kl eikx eily u(y)dy A(N ) (u) = GN (x, y)u(y)dy =
|l|,|k|N

so A(N ) (u) span|k|N k . The operators F L2 ([, ]2 ) that we dened are called HilbertSchmidt operators. Example. The Wave Equation and Poisson Summation We go back to the wave equation. If u0 , u1 C (R), 2periodic then there exists a unique u C (R+ R+ ) such that utt uxx = 0 and u(0, x) = u0 , ut (0, x) = u1 . We can dene a map depending on t with (u0 , u1 ) (u(s, ), ut (s, )). This denes a linear map 1 2 1 2 s : H ([, ]) L ([, ]) H ([, ]) L ([, ]) The mapping is continuous (follows from the fact the the problem can also be solved using Fourier Series). It is also adjoint(Exercise), just compute s ut (utt uxx )
0

75

Alternatively we can choose S(R) and consider the map

P : (u0 , u1 ) u(s, )(s)ds, ut (s, )(s)ds =
R R

s (s)ds

Essentially, instead of xing time, we are averaging over it. This is in fact compact on H 1 L2 , it is HilbertSchmidt and normal (we do not have time to prove this) Now how his this related with poisson summation? We have two ways of computing P : 1. Solve the wave equation using the explicit formula (u0 , u1 ) u = f (x + t) g(x t) 2. Fourier Series We claim (but will not prove) the following very useful result Theorem. The eigenvalues of P , j C decrease to 0 and T r(P ) =
i=0

converges. Furthermore the Poisson summation formula gives the relationship between T r(P ) computing with #1 and #2.

76