Upload
others
View
8
Download
0
Embed Size (px)
Citation preview
Functional Analysis: Lecture notes based onFolland
Johan Jonasson ââ âĄ
October 2011
1 Normed vector spacesA vector space (VS) consists of objects (such as vectors or functions) that can beadded and multiplied by scalars in such a way that the commutative and distribu-tive laws hold. The set of scalars can be any field K, but here we will always haveK = R or K = C.
Let X be a VS. A norm on X is a function â · â : X â [0,â) for which
âą âλxâ = |λ|âxâ for all λ â K and x â X ,
âą âx+ yâ †âxâ+ âyâ for all x, y â X (the triangle inequality),
âą âxâ = 0 iff x = 0.
A norm gives rise to a metric, Ï(x, y) = âxâ yâ, and hence to a topology, theso called norm topology. Two norms â · â1 and â · â2 are said to be equivalent ifthere exist 0 < C1 < C2 <â such that
âx â X : C1âxâ1 †âxâ2 †C2âxâ1.
Equivalent norms generate the same topology and the same Cauchy sequences.A VS equipped with a norm is called a normed vector space (NVS).
âChalmers University of Technologyâ Goteborg UniversityâĄ[email protected]
A NVS which is also complete is called a Banach space.The following theorem will provide an important tool for proving complete-
ness of a NVS. Recall that for xn â X , we say that the seriesâ
n xn is abso-lutely convergent if
ân âxnâ <â.
Theorem 1.1 A NVS X is complete iff every absolutely convergent series con-verges (in norm).
Proof. Suppose on one hand that X is complete andâ
n âxnâ <â. WritingSN =
âN1 xn, we have
âSN â SMâ = âMâN+1
xnâ â€MâN+1
âxnâ â€ââN+1
âxnâ â â
as M,N ââ. Thus SN is Cauchy and hence convergent.On the other hand suppose that every absolutely convergent series converges
and that âxn â xmâ â 0 as m,n â â. Pick for each j the index nj so thatm,n â„ nj â âxm â xnâ < 2âj . Setting x0 = 0, we have
xnk =kâ1
(xnj â xnjâ1)
which has norm less than 1 by the triangle inequality. Hence xnk converges tosome limit y. However for n â„ nk and k large enough,
âxn â yâ †âxn â xnkâïžž ïž·ïž· ïžž<2âk
+ âxnk â yâïžž ïž·ïž· ïžž<2âk
< 2â(kâ1).
I.e. xn â y. 2
Example. Let X be a topological space and let B(X) be the space of boundedcontinuous (complex-valued) functions on X . Define the norm â · âu on B(X) by
âfâu = sup|f(x)| : x â X.
Suppose thatâ
n âfnâu <â. Then clearly
âân
fnâu â€ân
âfnâu,
2
so f :=â
n fn â B(X). DoesâN
1 fn converge to f? Yes:
âNâ1
fn â fâu = âââN+1
fnâu â€ââN+1
âfnâu â 0
as N ââ. By Theorem 1.1 we conclude that B(X) is a Banach space. 2
Example. Let (X,M, ”) be a measure space. We claim that L1(X,M, ”) is aBanach space. (Here we have to identify functions that are equal a.e.)
Recall that the norm is given by âfâ1 =â«|f |. Suppose that
ân âfnâ1 <â.
Then by the MCT,â« ân
|fn| =ân
â«|fn| =
ân
âfnâ1 <â
so that f :=â
n fn exists (a.e.) and, using the MCT again,
âNâ1
fn â fâ1 =
â«|ââN+1
fn| â€ââN+1
â«|fn| =
ââN+1
âfnâ1 â 0
as N ââ. Now Theorem 1.1 proves our claim. 2
If X and Y are NVSâs, then the product norm on X ĂY is given by â(x, y)â =max(âxâ, âyâ). This norm is equivalent to âxâ+ âyâ, (âxâ2 + âyâ2)1/2, etc.
Next we consider linear maps from X to Y . A linear map T : X â Y is saidto be bounded if there exists C < â such that âTxâ †Câxâ. Here, of course,the norm notation refers to the norm of X for objects in X and to the norm of Yfor objects in Y . Note that this notion of boundedness is not the same as when wespeak of a bounded function.
Proposition 1.2 The following statements are equivalent
(a) T is bounded.
(b) T is continuous.
(c) T is continuous at 0.
3
Proof. That (b) implies (c) is trivial. If (c) holds, then there exists a ÎŽ > 0such that âxâ < 2ÎŽ â âTxâ < 1. Thus, for any x,
âTxâ = ââxâÎŽT (
ÎŽx
âxâ)â < 1
ÎŽâxâ,
i.e. T is bounded. Finally assume that (a) holds. Fix x and Δ > 0. Then ifâxâČ â xâ < Δ/C, we have
âTxâČ â Txâ = âT (xâČ â x)â †CâxâČ â xâ < Δ.
2
Let L(X ,Y) be the space of bounded linear operators from X to Y . In thisnotation it is also assumed that L(X ,Y) is equipped with the operator norm givenby
âTâ = supâTxâ : âxâ = 1 = supâTxââxâ
: x â X.
The space L(X ,Y) is not always complete, but the following holds
Proposition 1.3 If Y is complete, then so is L(X ,Y).
Proof. Assume that Tn â L(X ,Y) is Cauchy. Then for arbitrary x â X ,
âTnxâ Tmxâ = â(Tn â Tm)xâ †âTn â Tmââxâ.
Hence Tnx is Cauchy and hence convergent, since Y is Cauchy. We can thusdefine
Tx = limnTnx, x â X .
Then the operator T is linear and âTnxâ â âTxâ (see exercise 2), so
âTâ = supâTxâ : âxâ = 1 = suplimnâTnxâ : âxâ = 1 †lim sup
nâTnâ <â
since Tn is Cauchy. Hence T â L(X ,Y). We need to show that âTnâTâ â 0.Pick Δ > 0 and N so that m,n â„ N â âTn â Tmâ < Δ. Then for any x withâxâ = 1,
âTnxâTxâ = âTnxâ limmTmxâ = lim
mâTnxâTmxâ †lim sup
mâTnâTmâ < Δ.
Hence âTn â Tâ < Δ. 2
4
Suppose that T â L(X ,Y) and S â L(Y ,Z). Then
âS(T (x))â †âSââTxâ †âSââTââxâ
so that âS Tâ †âSââTâ and in particular S T â L(X ,Z).If T â L(X ,Y) is bijective and Tâ1 is bounded (i.e. Tâ1 â L(Y ,X )), then we
say that T is an isomorphism. If T also has the property that âTxâ = âxâ for allx, then t is said to be an isometry. Whenever two NVS are isomorphic, the havecorresponding notion of convergence. If they are also isometrically isomorphic,we may think of them as two interpretations of the same space. When X and Yare isometrically isomorphic, one writes for short X âŒ= Y .
2 Lp-spacesFix a measure space (X,M, ”) and 1 †p <â. For f : X â C measurable, let
âfâp =(â«
fp d”)1/p
.
Then Lp(X,M, ”) (or Lp(X) or Lp(”) or simply Lp when there is no risk forconfusion) is the space of measurable f : xâ C for which âfâp <â. A specialcase is when M = P(X) and ” is counting measure, in which case we writelp(X) for Lp(X,M, ”). When also X = N, we write simply lp. (I.e. when wewrite lp without specifying X , it is understood that X = N.)
Clearly âλfâp = |λ|âfâp for any scalar λ. Also
âf + gâpp =
â«|f + g|p â€
â«(2 max(|f |, |g|))p †2p(âfâpp + âgâpp).
Thus Lp is a vector space (with identification of functions that are equal a.e.). Weclaim that â·âp is a norm. What remains to be shown is that the triangle inequalityholds.
Lemma 2.1 Let a, b â„ 0 and λ â (0, 1). Then
aλb1âλ †λa+ (1â λ)b.
Proof. The function g(x) = axb1âx is convex, so g(λ) †λg(1) + (1 â λ)g(0).This is exactly what we wanted to prove. 2
5
Theorem 2.2 (Holderâs inequality) Let 1 < p <â and 1/p+ 1/q = 1. Then
âfgâ1 †âfâpâgâq.
Remark. Below, we will define â · ââ in such a way that it will be easily seenthat Holderâs inequality extends to 1 †p †â.Remark. When 1/p+ 1/q + 1, one says that p and q are conjugate exponents.
Proof. If either âfâp or âgâq is 0 orâ, then the result is trivial, so assumeotherwise. Apply Lemma 2.1 with a = |f(x)|p/âfâpp, b = |g(x)|q/âgâqq andλ = 1/p, so that 1â λ = 1/q, to get
|f(x)g(x)|âfâpâgâq
†|f(x)|p
pâfâpp+|g(x)|q
qâgâqq.
Integrating both sides gives
âfgâ1âfâpâgâq
†1
p+
1
q= 1.
2
We are ready for the triangle inequality for Lp-norm a.k.a. Minkowskiâs in-equality
Theorem 2.3 (Minkowskiâs inequality) Suppose 1 †p <â. Then
âf + gâp †âfâp + âgâp.
Proof. The result is trivial for p = 1 or f + g = 0 a.e. so assume 1 < p <âand âf + gâp > 0. Clearly
|f + g|p †|f + g|pâ1(|f |+ |g|).
Let q = p/(pâ 1), the conjugate exponent of p. Then
âf + gâpp =
â«|f + g|p â€
â«|f ||f + g|pâ1 +
â«|g||f + g|pâ1
†âfâpâ|f + g|pâ1âq + âgâpâ|f + g|pâ1âq = (âfâp + âgâp)(â«|f + g|p
)(pâ1)/p= (âfâp + âgâp)âf + gâpâ1p ,
6
where the second inequality uses Holderâs inequality and the first equality usesthat q(1â p) = p. The result now follows on dividing both sides with âf + gâpâ1p .2
By Minkowskiâs inequality â · âp is indeed a norm and Lp is indeed a NVS.Moreover
Theorem 2.4 Lp is a Banach space.
Proof. We want to show that Lp is complete and to that end we use Theorem1.1. Assume that fn is absolutely convergent, i.e. B :=
ân âfnâp < â.
Write Gn =ân
1 |fk| and G = limnGn =ââ
1 |fn|. By Minkowskiâs inequality,âGnâp â€
ân1 âfkâp †B. By the MCT, âGâp †B. Hence G â Lp and
ââ1 fk
converges a.e. Write F =ââ
1 fk â Lp. It remains to show that âân
1 fkâFâp â0. However F â
ân1 fk =
âân+1 fk â 0 a.e. and
|F ânâ1
fk|p â€(|F |+
kâ1
|fk|)p†(2G)p â Lp
so this follows from the DCT. 2
The next result is an approximation result for Lp.
Proposition 2.5 The family of simple functions, i.e.
Ï =nâ1
ajÏEj : n â N, Ej disjoint
is dense in Lp.
Proof. Fix an arbitrary f â Lp. A simple function Ï is in Lp if ”(Ej) = 0for every j and it is a fundamental fact of integration theory that one can find suchÏn so that Ïn â f and |Ïn| â f a.e. Since |Ïn â f |p †(2|f |)p, the DCT givesâ«|Ïn â f |p â 0 as desired. 2
The space Lâ. Using the usual convention inf â =â, we define
âfââ = ess supx|f(x)| = infa : ”x : |f(x)| > a > 0.
7
Note that ”x : |f(x)| > âfââ = 0.We now defineLâ(X,M, ”) as the space of measurable functions f : X â C
such that âfââ <â. Note that by the definition of â · ââ, for any f â Lâ, thereexists a bounded function f âČ such that f = f âČ a.e., so Lâ can be regarded as thespace of bounded functions.
A few simple facts follow.
(1) âfgâ1 †âfâ1âgââ (i.e. Holderâs inequality holds also for p â 1,â.)
(2) âf + gââ †âfââ + âgââ (i.e. Minkowskiâs inequality holds also forp =â.)
(3) âfn â fââ â 0â âE such that ”(Ec) = 0 and fn â f uniformly on E.
(4) The family of simple functions is dense in Lâ.
(5) Lâ is a Banach space.
Statements 1-4 are very easy to prove and are left as an exercise. For (5),suppose that fn is absolutely convergent. Then fn(x) is absolutely conver-gent for a.e. x, so there exists a limit f := limn fn a.e. Pick m so large thatn â„ m â âfn â fmââ < 1 and M so large that |fm(x)| < M a.e. Then clearly|f(x)| < M + 1 a.e., i.e. f â Lâ. Also âfn â fââ â€
âân+1 âfââ â 0.
Next we turn to relationships between Lpâs for different pâs. When p < q < r,we have on one hand that Lq â Lp + Lr, which means that any function in Lq
can be written as the sum of one function in Lp and one function in Lr. On theother hand we have that Lq â Lr â© Lp. The proofs of these two facts are left asexercises. Here are two more facts.
Proposition 2.6 Assume that 1 †p < q †â.
(a) If ” is counting measure, then âfâp â„ âfâq. Consequently Lp(”) â Lq(”).
(b) If ” is finite, thenâfâp †âfâq”(X)(qâp)/(pq).
Consequently Lp â Lq.
Remark. Part (b) when ” is a probability measure should be known to a proba-bilistically oriented person: E[Οp]1/p †E[Οq]1/q.
Proof.
8
(a) In case q =â, we have
âfâpâ = (supx|f(x)|)p â€
âx
|f(x)|p = âfâpp.
Now assume q < â and assume without loss of generality that âfâq = 1.Then |f(x)| †1 for every x, so that |f(x)|p â„ |f(x)|q and hence
âfâp =(â
x
|f(x)|p)1/pâ„(â
x
|f(x)|q)1/p
= 1 = âfâq.
(b) If q = â, scale f so that âfââ = 1. Then obviously âfâp †”(X)1/p as
desired. For q < â, observe that1
q/p+
1
q/(q â p)= 1. Hence Holderâs
inequality gives
âfâpp = â|f |p · 1â1 †â|f |pâq/pâ1âq/(qâp) = âfâpq”(X)(qâp)/q.
Now take both sides to the power 1/p to finish the proof.
2
Example. Lp-contraction of Markov chains. Let Xtât=0 be an irreducibleaperiodic Markov chain on the finite state space S. Let Ï denote the stationarydistribution (i.e. P(Xt = s) â Ï(s), s â S.) Let P be the transition matrix andlet f : S â C. Write Es[·] for E[·|X = s]. Then (Pf)(s) = Es[f(X1)]. For theLp(S,P(S), Ï)-norm, this means that
âPfâpp = EÏ[|(Pf)(X0)|p
]= EÏ
[|EX0 [f(X1)]|p
]†EÏ
[EX0 [|f(X1)|]
]= EÏ
[|f(X0)|
]= âfâpp
with equality iff f is constant. Hence âPfâp/âfâp †1. Better estimates of thisratio are sometimes used to bound the mixing time of the Markov chain. 2
3 The dual of Lp
Let X be a NVS. A linear map Ï : X â K is called a (K-valued) linear func-tional. Denote by X â the space L(X , K), the space of bounded linear functionalson X . The space X â is called the dual of X .
9
In this section we will consider the case X = Lp(X,M, ”). The number qwill throughout be assumed to be the conjugate exponent of p. For g â Lq, define
Ïg(f) =
â«fg d”, f â Lp.
Clearly Ïg is linear and
|Ïg(f)| =âŁâŁâŁ â« fg
âŁâŁâŁ †⫠|fg| = âfgâ1 †âfâpâgâq.Hence âÏgâ †âgâq, and in particular Ïg is bounded, i.e. Ïg â (Lp)â.
Proposition 3.1 If p > 1, then âÏgâ = âgâq.
Proof. Since p > 1 we have q < â. By the above, it remains to show thatâgâq †âÏgâ. This is trivial if g = 0 a.e., so we may assume that âgâq > 0. Let
f =|g|qâ1sgng
âgâqâ1q
.
(Here the sgn(z) = ei arg z â z = |z|sgnz â |z| = z sgnz.) Then, since (q â1)p = q,
âfâpp =1
âgâ(qâ1)pq
â«|g|(qâ1)p =
âgâqqâgâqq
= 1.
Hence
âÏgâ â„ |Ïg(f)| =âŁâŁâŁ â« fg
âŁâŁâŁ =
â«|g|qâ1g sgng
âgâqâ1q
=
â«|g|q
âgâqâ1q
= âgâq.
2
Extension: If ” is semifinite, then the result of Proposition 3.1 holds also forp = 1 â q = â. (Recall that ” is semifinite if âE : 0 < ”(E) < â.) We omitthe proof.
Put in other words, Proposition 3.1 says that the map g â Ïg is an isometryfrom Lq to (Lp)â. In fact, it is also an isomorphism:
Theorem 3.2 Let ” be Ï-finite and 1 †p < â. Then for any Ï â (Lp)â, thereexists g â Lq such that Ï = Ïg.
10
Extension: If 1 < p <â, then the result holds without restriction on ”. We willnot prove this.Moral: (Lp)â and Lq are isometrically invariant. In other words these can be seenas two interpretations of the same space. In short this is written (Lp)â âŒ= Lq, butone often, with abuse of notation, writes even (Lp)â = Lq.
Proof. Assume first that ” is finite. Fix an arbitrary Ï â (Lp)â. Since ” isfinite, ÏE â Lp for all E â M. Hence we can define Îœ(E) = Ï(ÏE), E â M.Now suppose that Ej , j = 1, 2, . . . are disjoint and write E = âȘâ1 Ej . Then
âÏE ânâ1
ÏEjâp = ââân+1
Ejâp = ”( âân+1
Ej
)pâ 0.
Henceân
1 ÏEj â Ï(E) in Lp. Since Ï is a continuous functional, this en-tails that Ï(
ân1 ÏEj) â Ï(Ï(E)) in Lp. By the linearity of Ï, it follows that
Îœ(âȘâ1 Ej) =ââ
1 Îœ(Ej), i.e. Îœ is a signed measure. Hence the Radon-NikodymTheorem implies that there exists g â L1 such that Îœ(E) = Ï(ÏE) =
â«Eg d”,
E â M. By linearity of Ï it follows that Ï(f) =â«fg d” for all simple functions
f . For general f , Proposition 2.5 tells us that there exist simple functions fn suchthat âfn â fâp â 0. Hence, by continuity of Ï,
Ï(f) = limnÏ(fn) = lim
n
â«fng.
Also, By Holder,âŁâŁâŁ â« fng ââ«fgâŁâŁâŁ †⫠|(fn â f)g| †âfn â fâpâgâq â 0.
Thus limn
â«fng =
â«fg and hence Ï(f) =
â«fg as desired. (Note however that
there is a gap here, since we have not shown that g â Lq, merely that g â L1. Wewill come back to that immediately after this proof.)
Now let ” be Ï-finite and pick sets En such that En â X and ”(En) < â.By the above there exist gn â Lq such that Ï(f) =
â«fgn, f â Lp(En) and
âgâq = âÏ|Lp(En)â †âÏâ. Note that each gn is unique up to alterations on a nullset. Define g by letting g = gn on En. Then |gn| â |g| a.e. Henceâ«
|g|q = limn|gn|q †âÏâq
11
so that g â Lq. By the DCT, gn â g in Lq. Now for f â Lp, we have fÏEn âLp(En) and by the DCT fÏEn â f in Lp. HenceâŁâŁâŁÏ(f)â
â«fgâŁâŁâŁ =
âŁâŁâŁ limnÏ(fÏEn)â
â«fgâŁâŁâŁ =
âŁâŁâŁ limn
â«fgn â
â«fgâŁâŁâŁ
†lim supn
â«|f(gn â g)| †lim sup
nâfâpâgn â gâq = 0,
by Holderâs inequality. 2
Now we fill in the gap in the above proof. We do so by using the followingresult, which is sometimes referred to as the âReverse of Holderâs inequalityâ.
Theorem 3.3 Assume that ” is semifinite. Let S be the family of simple functionswhose support is of finite measure. Assume that fg â L1 for all f â S and that
Mq(g) := supâŁâŁâŁ â« fg d”
âŁâŁâŁ : f â S, âfâp = 1<â.
Then âgâq = Mq(g). In particular g â Lq.
Proof. That Mq(g) †âgâq follows from Holder, so we focus on the reverseinequality. We will settle for the case when ” is Ï-finite. Assume first that q <â.Let En â X with ”(En) < â. Let hn be simple functions with hn â g and|hn| †|g|. Let gn = hnÏEn . Then gn â S, gn â g and |gn| †|g|. Let
fn =|gn|qâ1sgng
âgnâqâ1q
.
Then, as in the proof of Theorem 3.1, âfnâp = 1. Note also thatâ«|fngn| = âgâq.
Thus, by Fatouâs Lemma,
âgâq †lim infnâgnâq = lim inf
n
â«|fngn|
†lim infn
â«|fng| = lim inf
n
â«fng â€Mq(g)
since fng is nonnegative.
12
Now to the case q =â. Let A = x : |g(x)| > Mâ(g) + Δ and assume that”(A) > 0 for some Δ > 0. Then we can find B â A with 0 < ”(B) < â. Letf = ÏBsgng/”(B). Then f â S and âfâp = âfâ1 = 1. Howeverâ«
fg =
â«B
|g|”(B)
> Mâ(g) + Δ,
a contradiction. 2
Now consider the function g from the proof of Theorem 3.2. For f â S andâfâp = 1, we have |
â«fg| = |Ï(f)| †âÏâ, so Mg(q) <â and hence g â Lq.
4 The Hahn-Banach TheoremLet X be a NVS and recall that X â = L(X , K). How can we be sure that this isan interesting space in the sense that it really contains any nontrivial objects? Incase X = Lp(X,M, ”) for some concrete X like e.g. X = Rn we know, and sawin the previous section, that this is indeed the case, but in general? As we shallsee, the Hahn-Banach Theorem answers our question with a âyesâ. First however,we need some preliminaries.
Suppose that f : X â C is a linear functional. Then u := <f is a real-valuedlinear functional and since =f(x) = â<(if(x)) = â<f(ix) = âu(ix), we have
f(x) = u(x)â iu(ix).
On the other hand, if u : X â C is a real-valued linear functional, then f(x) :=u(x) â iu(ix), x â X , is a complex-valued linear functional. In any case, theequality f(x) = u(x)â iu(ix), x â X , entails that
âfâ = âuâ.
This follows on one hand from |u(x)| = |<f(x)| †|f(x)|, so that âuâ †âfâ,and on the other hand from
|f(x)| = f(x)sgnf(x) = f(sgnf(x)x) = u(sgnf(x)x) †âuââxâ
so that âfâ †âuâ.Let X be a vector space over R. A Minkowski functional on X is a function
p : X â R such that
13
âą p(λx) = λp(x) for all λ â„ 0 and x â X ,
âą p(x+ y) †p(x) + p(y) for all x, y â X .
In particular all linear functionals and all seminorms are Minkowski functionals.
Theorem 4.1 (Hahn-Banach Theorem, real version) Let X be a vector spaceover R and p a Minkowski functional on X . LetM be a linear subspace of X andf :Mâ R a linear functional such that f †p onM. Then there exists a linearfunctional F : X â R such that F = f onM and F †p.
Proof. The result is trivial if M = X , so assume that there exists an x âX \M. For any y1, y2 âM,
f(y1) + f(y2) = f(y1 + y2) †p(y1 + y2) †p(y1 â x) + p(y2 + x),
i.e.f(y1)â p(y1 â x) †f(y2)â p(y2 + x).
Since y1 and y2 were arbitrary, we get
supyâM
(f(y)â p(y â x)
)†inf
yâM
(f(y)â p(y + x)
).
Fix a number α between these two quantities. Define g : M + Rx â R byg(y + λx) = f(y) + λα. Then g is linear and extends f and for λ > 0,
g(y + λx) = λ(f(y/λ) + α
)†λ
(f(y/λ) + p(x+ y/λ)â f(y/λ)
)= p(y + λx),
where the inequality follows from lower bound in the definition of α. For λ < 0,
g(y + λx) = |λ|(f(y/|λ|)â α
)†|λ|
(f(y/|λ|)â f(y/|λ|) + p(y/|λ| â x)
)= p(y + λx)
where the inequality follows from the upper bound in the definition of α. Henceg †p.
Now let F be the family of linear extensions, g, of p with g †p. Partiallyorder F with respect to inclusion of domain. Then every chain has an upperbound, namely the extension defined on the union of the domains of the elements
14
of the chain. By Zornâs Lemma, F has a maximal element, F . The element Fmust be defined on the whole of X for if not, it could be extended by the aboveprocedure. 2
Now assume that f :Mâ C is a linear functional and that |f | †p, where pis a seminorm, i.e. a Minkowski functional with the extra property that p(λx) =|λ|p(x) for all λ â C and x â X . Then u = <f is a real-valued linear functionaland can hence be extended to U : X â C with |U | †p. Let F (x) = U(x) âiU(ix). Then f extends f to the whole of X and
|F (x)| = F (x)sgnF (x) = F (sgnF (x)x) = U(sgnF (x)x)
†p(sgnF (x)x) = p(x).
In summary:
Theorem 4.2 (Hahn-Banach Theorem, complex version) Let f : M â C bea linear functional on the linear subspaceM of X . Assume that p is a seminormon X and that |f | †p onM. Then there exists F : X â C such that F = f onM and |F | †p.
Here are some applications.
(a) For any x â X there exists an f â X â such that f(x) = âxâ and âfâ = 1.
Proof. Let M = Cx and define f on M by f(λx) = λâxâ. De-fine p by p(y) = âyâ. Then p is a norm andS hence a seminorm and|f(λx)| = |λâxâ| = âλxâ = p(x). Now extend f to the whole of Xusing the Hahn-banach Theorem. The extension satisfies |f | †p, so forany y â X , |f(y)| †âyâ, so âfâ †1. However |f(x)| = âxâ, so âfâ = 1.2
(b) If x1, x2 â X and x1 6= x2, then there exists f â X â such that f(x1) 6=f(x2).
Proof. Use (a) with x = x1 â x2. 2
(c) IfM is a closed subspace of X and x â X \M, then there exists f â X âwith f ⥠0 onM and f(x) 6= 0.
15
Proof. Let ÎŽ = infâxâ yâ : y âM. Then ÎŽ > 0, for if ÎŽ = 0, then therewould exist yn â M with âx â ynâ â 0, i.e. yn â x. SinceM is closedthis entails that x âM, a contradiction.
Define f onM+ Cx by
f(y + λx) = λΎ.
Then f ⥠0 onM and
|f(y + λx)| = |λ|ÎŽ †|λ|âx+ y/λâ = ây + λxâ,
where the inequality follows from the definition of ÎŽ. Thus |f | †â · â, so fcan be extended by Hahn-Banach. 2
(d) For x â X , define x : X â â C by x(f) = f(x). Then x â x is a linearisometry from X to X ââ. In particular, letting X := x : x â X, we haveX â X ââ. A NVS X for which X = X ââ is said to be reflexive.
Proof. Clearly the given map is linear. Also
|x(f)| = |f(x)| †âfââxâ
and hence âxâ †âxâ. On the other hand, by (a) there exists an f â X âsuch that f(x) = âxâ and âfâ = 1, so |x(f)| = âxâ = âxâ and henceâxâ â„ âxâ. 2
5 The Baire Category Theorem and consequencesthereof
The Baire Category Theorem (BCT) is the following topological result.
Theorem 5.1 (Baire Category Theorem) Let X be a complete metric topologi-cal space. Then the following hold.
(a) If for each n = 1, 2, . . ., Un is an open dense set, then â©nUn is dense.
(b) The space X cannot be written as a countable union of nowhere dense sets.
16
Proof. First we show that (a) implies (b). IfEn is nowhere dense, i.e.Eno
= â ,then En
cis dense an open, so by (a) â©nEn
cis dense. Therefore(
âȘn En)c
= â©nEcn â â©nEn
c
and since a dense set cannot be empty, (b) follows.Now we prove (a). We want to show that (â©nUn) â© W 6= â for any open
nonempty set W . Since U1 â©W is open and nonempty, we can find r1 â (0, 1)and x1 â X such thatBr1(x1) â U1â©W . In the same way there exist r2 â (0, 1/2)and x2 â X such thatBr2(x2) â U2â©Br1(x1). Repeating the argument once againgives r3 â (0, 1/4) and x3 â X such that Br3(x3) â U3 â©Br2(x2). Repeating theargument inductively gives a Cauchy sequence of points xn. Since X is complete,xn â x for some x â X . For every n, we have
x â Brn(xn) â Un â©Brnâ1(xnâ1) â Un â©W.
Hence x â (â©nUn) â©W . 2
One should note that the BCT is purely topological, it holds for any topologicalspace that is homeomorphic to a complete metric space.Example. Wienerâs Tauberian Theorem. Let T = R/2ÏZ = R mod 2Ï be theunit circle in the complex plane.Question: For which functions f â L1(T) is it true that the translates fy = f(· ây), y â T, span L1(T ) in the sense that the set
M := Nâj=1
ajfyj : N â N, aj â C, yj â T
is dense in L1(T)?Let ck =
â«T f(x)eâikxdx, f âs kâth Fourier coefficient. Note that the kâth
Fourier coefficient of fy is eâikyck. If ck = 0 for some k, this means that the kâthFourier coefficient of fy is also 0 for all y â T and hence that all functions inM have 0 for their kât Fourier coefficient. Hence e.g. the function eikx â L1(T ),whose kât Fourier coefficient is 1, cannot be approximated by functions in M .
Now assume ck 6= 0 for all k. Assume that M 6= L1(T ) and pick g0 âL1(T ) \M . According to application (c) above of Hahn-Banach, there exists a
17
linear functional Ï on L1(T ) such that Ï âĄ 0 on M and Ï(g0) 6= 0. By Theorem3.2 there exists a function h â Lâ(T) such that
Ï(g) =
â«Tg(x)h(x)dx, g â L1(T).
Since fy âM for all y â T, this entails thatâ«T f(xâ y)h(x)dx = 0 for all y â T.
In other words, the convolution h â f is identically 0 (where h(x) = h(âx)).However the kâth Fourier coefficient of h â f is dkck, where dk is the kâth Fouriercoefficient of h. Since ck 6= 0 for all k, we have dk = 0 for all k. Thus h ⥠0 andconsequently Ï(g0) = 0, a contradiction.
In summary: M spans L1(T) iff ck 6= 0 for all k. 2
Recall that a map f : X â Y is called an open map if f(U) is open in Ywhenever U is open in X . An immediate consequence is that if f is bijective andopen, then fâ1 is continuous. If Y is metric, then openness of f can be expressedas that for all open U in X and all x â U , there exists r > 0 such that
Br(f(x)) â f(U).
If If X and Y are NVSâs and f is linear, then this boils down to
âr > 0 : Br(0) â f(B1(0)).
Theorem 5.2 (The Open Mapping Theorem) Let X and Y be Banach spacesand T â L(X ,Y). Then T surjective implies T open.
Proof. Write Br for Br(0) (in X as well as in Y). We want to show thatâr > 0 : Br â T (B1). Since T is surjective and X = âȘnBn, Y = âȘnT (Bn) =âȘnnT (B1). Hence the BCT implies that T (B1) is not nowhere dense. (Here weuse that Y is complete.) Thus we can find y0 â Y and r > 0 such that
B8r(y0) â T (B1).
Pick y1 â T (B1) such that ây1 â y0â < 4r and y1 = Tx for some x â B1. ThenB4r(y1) â T (B1) and whenever âyâ < 4r, we have y = y1 + (y â y1) â T (B2),since y â y1 â B4r(ây1) â T (B1). By scaling, this generalizes to y â T (B2ân)whenever âyâ < 2r2ân.
Now pick y with âyâ < r. Then y â T (B1/2). Hence there exists x1 â B1/2
with ây â Tx1â < r/2. This entails that y â Tx1 â T (B1/4). Thus we can find
18
x2 â B1/4 with ây â Tx1 â Tx2â < r/4. Then y â Tx1 â Tx2 â T (B1/8),etc. Keeping on inductively gives xj â B2âj . Since X is Banach,
âj xj is
convergent by Theorem 1.1. Let x =â
j xj . Then x â B1 and ây â Txâ =
limN ây ââN
1 Txjâ = 0, i.e. y = Tx. In summary there exists x â B1 such thaty = Tx whenever âyâ < r, as desired. 2
Corollary 5.3 If X and Y are Banach and T â L(X ,Y) is bijective, then T is anisomorphism.
Proof. Since T is bijective, then Tâ1 exists. By the Open Mapping Theorem,Tâ1 is continuous and hence bounded. 2
Theorem 5.4 (The Closed Graph Theorem) Assume that X and Y are Banachspaces, T : X â Y linear and that
Î(T ) := (x, Tx) : x â X â X Ă Y
is closed. Then T is bounded.
Proof. Let Ï1 and Ï2 be the projection maps from Î(T ) to X and Y respec-tively, given by Ï1(x, Tx) = x and Ï2(x, Tx) = Tx. By the definition of productnorm, âÏ1â †1 and âÏ2â †1 and hence Ï1 â L(Î(T ),X ) and Ï2 â L(Î(T ),Y).
Note that Î(T ) is complete, since it is a closed subspace of the complete spaceX Ă Y . Thus, since Ï1 is bijective, it follows that Ïâ11 is bounded by the abovecorollary. Since also Ï2 is bounded, Ï2 Ïâ11 = T is bounded. 2
Theorem 5.5 (The Uniform Boundedness Principle) Assume that X is a Ba-nach space and Y a NVS. LetA be a subfamily of L(X ,Y). If supTâA âTxâ <âfor all x â X , then supTâA âTâ <â.
Proof. Let
En = x : supTâAâTxâ †n =
âTâA
x : âTxâ †n.
Observe that by assumption âȘnEn = X and that since âT (·)â is continuous, eachEn is closed. By the BCT, some En must contain a ball, Br(x0). For any x withâxâ †r, we have for any T â A,
âTxâ †âT (xâ x0)â+ âTx0â †2n,
19
since xâ x0 and x0 are both in Br(x0) â En. Hence for any x with âxâ †1 andany T â A, we have âTxâ †2n/r. Hence supTâA âTâ †2n/r. 2
6 Hilbert spacesHilbert spaces are Banach spaces with extra geometric structure. The norm arisesfrom an inner product. Let X be a vector space over C. An operation ă·, ·ă :X Ă X â C is called an inner product if for all x, y, z â X and a, b â C,
(a) ăax+ by, ză = aăx, ză+ băy, ză,
(b) ăy, xă = ăy, xă,
(c) ăx, xă â (0,â) for all x 6= 0.
Note that from (a) and (b), ăx, ay + bză = ăay + bz, xă = aăy, xă+ băz, xă =aăx, yă + băx, ză. It also follows that ă0, xă = ăx, 0ă = 0 for all x. When X isequipped with an inner product, we say that X is a pre-Hilbert space.
Assume now that X is a pre-Hilbert space. Define
âxâ = ăx, xă1/2.
The notation suggests that this is a norm. Let us show that this is indeed the case.For that we will need the following well-known result.
Theorem 6.1 (Schwartzâ inequality) For all x, y â X ,
ăx, yă †âxââyâ,
with equality iff x and y are linearly dependent.
Proof. The result is trivial if y = 0, so we may assume that y 6= 0. Letα = sgnăx, yă, z = αy and pick t â R. Then
0 †ăxâ tz, xâ tză = ăx, xă â tăx, ză â tăz, xă+ t2ăz, ză
= âxâ2 â 2t|ăx, yă|+ t2âyâ2
20
where the last equality follows from the definition of α. Differentiate the righthand side with respect to t and derive that this expression is minimized for t =|ăx, yă|/âyâ2. Plugging this into the inequality yields,
0 †âxâ2 â 2|ăx, yă|2
âyâ2+|ăx, yă|2
âyâ2= âxâ2 â |ăx, yă|
2
âyâ2.
This proves the desired inequality. Also, by part (c) in the definition of innerproduct, equality holds iff xâ tz = xâ αty = 0 as desired. 2
Note. If ă·, ·ă is real-valued, then the t defined in the proof is also real-valued.Hence equality holds in Schwarzâ inequality iff x = cy for a real constant c.
Note that ăx, yă+ ăy, xă = 2<ăx, yă. Therefore
âx+ yâ2 = ăx+ y, x+ yă = âxâ2 + 2<ăx, yă+ âyâ2
†âxâ2 + 2|ăx, yă|âyâ2 †âxâ2 + 2âxââyâ+ âyâ2 = (âxâ+ âyâ)2,
here the second inequality follows from Schwarzâ inequality. This proves thetriangle inequality. Hence â · â is a norm. Thus a pre-Hilbert space X is a NVS.If X is also Banach, then X is said to be a Hilbert space.Example. Consider X = L2(X,M, ”). Let
ăf, gă =
â«fg d”.
This is well defined since |fg| †(|f |2 + |g|2)/2. It is readily checked that thisdefines an inner product, and the resulting norm becomes
âfâ = (
â«ff)1/2 = (
â«|f |2)1/2 = âfâ2.
Hence L2 can be seen as a Hilbert space. (However Lp is not a Hilbert space forany p other than 2.) 2
Proposition 6.2 ă·, ·ă is continuous.
Proof. Let xn â x and yn â y, i.e. âxn â xâ â 0 and âyn â yâ â 0. Then
|ăxn, ynăâăx, yă| = |ăxnâx, ynă+ăx, ynâyă| †âxnâxââynâ+âxââynâyâ â 0,
21
since âynâ converges to âyâ and is hence bounded. 2
Whenever x and y are two elements of the Hilbert space X such that ăx, yă =0, we say that x and y are orthogonal. For short, this is sometimes written asx â„ y. The following version of the Pythagorean Theorem holds for Hilbertspaces. The proof is trivial.
Theorem 6.3 (The Pythagorean Theorem) If xi ℠xj for all 1 †i < j †n,then
ânâ1
xjâ2 =nâ1
âxjâ2.
We also have
Theorem 6.4 (The Parallelogram Law) For all x, y â X ,
âx+ yâ2 + âxâ yâ2 = 2(âxâ2 + âyâ2).
The proof is a straightforward expansion of the left-hand side.The next result states that vectors in a Hilbert space can be projected onto a
closed linear subspace. For a linear subspace M of X , we write Mâ„ := x âX : ây âM : x â„ y.
Theorem 6.5 LetMâ X be a closed linear subspace. Then for any x â X thereexist unique elements y âM and z âMâ„ such that x = z + y.
Proof. Fix x and let ÎŽ = infâx â yâ : y â M. Let yn â M be such thatâxâ ynâ â ÎŽ. By the Parallelogram Law,
2âyn â xâ2 + 2âym â xâ2 = âyn â ymâ2 + âyn + ym â 2xâ2.
Rearranging gives
âyn â ymâ2 = 2âyn â xâ2 + 2âym â xâ2 â 4â1
2(yn + ym)â xâ2
†2(âyn â xâ2 + âym â xâ2)â 4ÎŽ2
by the definition of ÎŽ, since 12(yn + ym) â M. The right-hand side tends to 0 as
m,nââ, so yn is Cauchy and hence convergent. Let y = limn yn. SinceMis closed, y â M. Let z = x â y. We claim that z â Mâ„. To see that this is so,
22
pick u â M. Then there exists c â C with |c| = 1 so that ăz, cuă is real. Sincez + tcu = x â (y â tcu) and y â tcu â M, the function f(t) := âz + tcuâ2 isminimized for t = 0. However
f(t) = âzâ2 + 2tăz, cuă+ t2âuâ2
so that 0 = f âČ(0) = 2ăz, cuă. Hence also ăz, uă = 0 as desired.For the uniqueness part, assume that we also have x = zâČ + yâČ. Then y â yâČ =
zâČ â z so that y â yâČ, zâČ â z âMâ©Mâ„, i.e. y = yâČ, z = zâČ. 2
Fixing an element y â X , we can define the linear functional fy(x) = ăx, yă,x â X . Then, by Schwarz, |fy(x)| = |ăx, yă †âxââyâ with equality when xis a constant times y. Hence âfyâ = âyâ and the map y â fy is a conjugatelinear isometry from X to X â. In fact, the following result shows that it is alsosurjective.
Theorem 6.6 For every f â X â there exists a unique y â X such that f(x) =ăx, yă, x â X .
Proof. If ăx, yă = ăx, yâČă for all x â X , then
ây â yâČâ2 = ăy â yâČ, yă â ăy â yâČ, yâČă = 0,
which proves uniqueness.Now focus on the existence part. The case f ⥠0 it trivial, so assume that
f 6⥠0. Let M = x : f(x) = 0. Then M is a closed linear subspace of X .Hence, by Theorem 6.5, there exists z â Mâ„ such that f(z) 6= 0 and âzâ = 1.Now pick an arbitrary x â X and let u = f(x)z â f(z)x. Then f(u) = 0, sou âM and hence ău, ză = 0. In other words
0 = ăf(x)z â f(z)x, ză = f(x)âzâ2 â f(z)ăx, ză.
Solving for f(x), recalling that âzâ = 1, gives
f(x) = f(z)ăx, ză = ăx, f(z)ză.
Taking y = f(z)z finishes the proof. 2
So all Hilbert spaces X are isomorphic to X â via the conjugate linear isometryx â fy. Hence X âŒ= X â âŒ= Xââ âŒ= . . .. Moreover, recall the definition x(f) =
23
f(x) for a given x â X and X = x : x â X â X ââ. If X is Hilbert, thenx(f) = x(fy) = ăx, yă, i.e. x = ăx, ·ă. This entails that X = X ââ, i.e. X isreflexive.
The family uααâA is said to an orthonormal family if ăuα1 , uα2ă = 0 for allα1 6= α2 and âuαâ = 1 for all α.
Theorem 6.7 (Besselâs inequality) Suppose that uααâA is an orthonormal fam-ily. Then for all x â X , â
뱉A
|ăx, uαă|2 †âxâ2.
In particular α : ăx, uαă 6= 0 is countable.
Proof. Write cα = ăx, uαă. We have, for any finite F â A.
0 †âxââαâF
cαuαâ2 = âxâ2 â 2<ăx,âαâF
cαuαă+ ââαâF
cαuαâ2
âxâ2 â 2âαâF
|cα|2 +âαâF
|cα|2 = âxâ2 ââαâF
|cα|2
where the equality follows from the Pythagorean Theorem. 2
Theorem 6.8 Let uααâA be an orthonormal family. The following three state-ments are equivalent
(a) âα : ăx, uαă = 0â x = 0 (completeness),
(b) âx : âxâ2 =â
α |ăx, uαă|2 (Parsevalâs identity),
(c) âx : x =â
αăx, uαăuα, where the terms are nonzero for at most countablymany α.
Proof. That (b) implies (a) is trivial. To see that (a) implies (c), pick x and letα1, α2, . . . be an enumeration of the αâs for which ăx, uαă 6= 0. Write uj = uαj .By Besselâs inequality,
âj |ăx, ujă|2 converges, so by the Pythagorean Theorem,
ânâm
ăx, ujăujâ2 =nâm
|ăx, ujă|2 â 0
24
as m,n â â. Since X is complete, this means thatâ
jăx, ujăuj converges.Since ăxâ
âjăx, ujă, uαă = 0 for all α, it follows from (a) that x =
âjăx, ujăuj
as desired.To see that (c) implies (b), by the same calculations as in the proof of Bessel,
âxâ2 ânâ1
|ăx, ujă|2 = âxânâ1
ăx, ujăujâ2 â 0
as nââ by (c). 2
A family that satisfies the three statements in the above theorem, is called anorthonormal basis for X .
Theorem 6.9 Every Hilbert space has an orthonormal basis.
Proof. Order the set of orthonormal families according to inclusion. Thenevery chain has a supremum, namely the union of the families in the chain. ByZornâs Lemma, there is a maximal element, uα. If this family does not span thewhole of X , there exists a z â X such that x := zâ
âαăz, uαăuα 6= 0. However,
then x/âxâ can be added to the uα, contradicting maximality. 2
Theorem 6.10 A Hilbert space has a countable basis iff it is separable.
Remark. This result does not contradict exercise 26, since the notion of a basis isdifferent there.
Proof. For the backwards implication, let xj be countable and dense. Lety1 = x1 and recursively yj = xnj , where nj is the first index k after njâ1 such thatxk is not in the span of y1, . . . , yjâ1. Then yj is linearly independent and dense.Now use the Gram-Schmidt process on the yjâs.
For the forwards direction, suppose that uj is a countable basis. Then it iseasily seen that
âN1 akujk : n â N, ak â Q + iQ is dense. 2
Let uααâA be an orthonormal basis for the Hilbert space X . Then the mapx â ăx, uαăαâA is an isometry from X to l2(A), by Parsevalâs identity. Onecan also show that this map is surjective. Hence X âŒ= l2(A).
25
7 Weak and weak* convergenceLet X be a NVS and X â = L(X ,C). The weak toplogy on X is generated by X â,i.e. xn â x weakly iff f(xn)â f(x) for every f â X â.
The weak* topology on X â is generated by X , i.e. fn â f weakly* ifffn(x)â f(x) for every x â X , i.e. if fn â f pointwise.Remark. When speaking of weak* convergence on a NVS Y , one must alwaysbe able to see Y as the dual of some NVS X , i.e. Y âŒ= X â. One says that X isthen a pre-dual of Y .Remark. Since X â X ââ, weak* convergence on X â is weaker than weak con-vergence on X â. The two notions coincide when X is reflexive.Example. Let 1 < p <â. We may then consider weak and weak* convergenceon Lp. Saying that fn â f weakly in Lp means that Ï(fn) â Ï(f) for everyÏ â (Lp)â, i.e. â«
fng ââ«fg
for every g â Lq, where q is the conjugate exponent of p. Saying that fn â fweakly* on Lp means that we identify Lp with the dual of Lq. Hence fn â fweakly* means Ïfn(g)â Ïf (g), i.e.â«
fng ââ«fg
for every g â Lq. We see that for 1 < p < â, weak and weak* convergence onLp agree.
In fact, this can be seen more directly by noting that Lp is reflexive: (Lp)ââ âŒ=(Lq)â âŒ= Lp.
For p = 1, weak convergence meansâ«fng â
â«fg for every g â Lâ. How-
ever, weak* convergence cannot be similarly characterized, since (Lâ)â 6âŒ= L1.For l1, we have l1 âŒ= câ0
âŒ= câ00 where c0 = g : N â C : limk g(k) = 0 andc00 = g : Nâ C : g(k) = 0 eventually.
For p = â, weak* convergence meansâ«fng â
â«fg for all g â L1. For
weak convergence, no simple characterization can be made. 2
Theorem 7.1 (Alaogluâs Theorem) LetX be a NVS and fn â X â with âfnâ †1,n = 1, 2, . . .. Then there exists a convergent subsequence fnk.
26
Remark. Since X â is a complete metric space, sequential compactness and com-pactness are equivalent. Hence Alaogluâs Theorem can be stated as that the closedunit ball is compact in the weak* topology.
Proof. We will settle for the case when X is separable. Let xk â X bedense. We want to show that there exists an f â X â such that fnk converges tof pointwise. Since |f(x1)| †âx1â, there exists a subsequence n1
j such thatfn1
j(x1) converges. Since |f(x2)| †âx2â, there exists a further subsequence
n2j such that fn2
j(x2) also converges. Continuing this procedure indefinitely
produces subsequences nij such that for each i, fnij(xk) converges for allk †i. Now use diagonalization: fnjj(xk) converges for every k. Let thus
nk = njj . For x 6â xk, fix Δ > 0 and pick xk so that âxâ xkâ < Δ/3. We have
|fni(x)âfnj(x)| †|fni(x)âfni(xk)|+|fni(xk)âfnj(xk)|+|fnj(xk)âfnj(x)| < Δ
for large enough i and j, since the first and the third term are bounded by Δ/3,since âfniâ and âfnjâ are bounded by 1. Hence fnk(x) is Cauchy and henceconvergent. Now set f(x) = limk fnk(x), x â X . 2
Example. The Poisson kernel in the unit disc. Let U = z â C : |z| < 1be the open unit disc in the complex plane and T = âU . Recall that f : Ω â Cis harmonic in Ω if f âČâČxx + f âČâČyy ⥠0 in Ω. The Poisson kernel is the functionP : U Ă Tâ [0,â) given by
P (reiÏ, Ξ) =1â r2
2Ï(1 + r2 â 2r cos(Ξ â Ï))â[ 1â r2
2Ï(1 + r)2,
1â r2
2Ï(1â r)2]
(identifying eiΞ â T with Ξ). It can be shown that if u is harmonic in (1 + Δ)U forsome Δ > 0, then
u(z) =
â«TP (z, Ξ)u(eiΞ)dΞ =: (Pu)(z), z â U.
In particular, with u ⥠1,â«TP (reiÏ, Ξ)dÏ =
â«TP (reiÏ, Ξ)dΞ = 1
for all r < 1, where the first equality follows from the symmetry between Ï and Ξin the definition of P . Now fix 1 < p †â and take f â Lp(T). Define
u(z) := (Pf)(z) =
â«TP (z, Ξ)f(Ξ)dΞ.
27
It can be shown that u thus defined is harmonic. Also, we have
|u(z)|p â€(â«
T|f(Ξ)|P (z, Ξ)dΞ
)pâ€â«T|f(Ξ)|pP (z, Ξ)dΞ
where the last inequality follows from Proposition 2.6 applied to the measured”(Ξ) = P (z, Ξ)dΞ. Hence, for r < 1,â«
T|u(reiÏ)|pdÏ â€
â«T
â«T|f(Ξ)|pP (reiÏ, Ξ)dÏdΞ
=
â«T|f(Ξ)|p
â«TP (reiÏ, Ξ)dÏïžž ïž·ïž· ïžž
=1
dΞ = âfâpp.
In summary, if f â Lp(T), then Pf is harmonic and supr<1
â«T |(Pf)(reiÏ)|pdÏ â€
âfâpp.Conversely, if u is harmonic in U and supr<1
â«T |u(reiÏ)|pdÏ =: M < â,
then there exists an f â Lp(T) such that u(z) = (Pf)(z), z â U .To prove the existence of such an f , let for r < 1, fr(Ξ) = u(reiΞ). Then
âfrâp †M1/p, so by Alaogluâs Theorem, there exist rj â 1 such that frj â fweakly*, i.e. frj â f a.e., for some f â Lp. The function u(rjz) is harmonic onrâ1j U so by the above,
u(rjz) =
â«TP (z, Ξ)frj(Ξ)dΞ.
Since the P (z, Ξ) on the right hand side is bounded by (1â r2j )/(2Ï(1â rj)2) andthe frj âs are bounded in Lp and p > 1, it follows from a version of the DCT, tobe proved below, that the right hand side converges to
â«T P (z, Ξ)f(Ξ)dΞ. The left
hand side converges to u(z), so we are done. 2
Here is the promised version of the DCT.
Theorem 7.2 (The Dominated Convergence Theorem) Let ” be finite. Assumethat fn â f a.e. and that fn is uniformly integrable, i.e. for every Δ > 0, thereexists K <â such that supn
â«|fn|Ïx:|fn(x)|>Kd” < Δ. Then
â«fnd”â
â«fd”.
Proof. Fix Δ > 0 and K as in the theorem. Then
lim supn|â«fn â
â«f | †Δ+
â«|fn(x)|â€K
(fn â f) = Δ
28
by the ordinary DCT. 2
Next we show how this version of the DCT applies to the above. Recall thatp > 1, so that the conjugate exponent q is finite. Suppose that fn is bounded inLp: supn âfnâp â€M <â. By Holder,â«|fn|Ï|fn(x)|>Kd” †âfnâp”|fn(x)| > K1/q = âfnâp”|fn(x)|p > Kp1/q
â€M(Mp
Kp
)1/q< Δ
for large enough K. Note that this line of reasoning does not work for p = 1.
8 The Riesz Representation TheoremWe recall some topological facts. Proofs of these will be given in a later appendix.A topological space X is said to be locally compact if for every x â X , there is acompact set K such that x â Ko. One says that X is Hausdorff if for all distinctx, y â X there are disjoint open sets V and W such that x â V and y â W .WhenX is locally compact and Hausdorff, we write for short thatX is LCH. AnyLCH space X has the property that for any x â X and open set U 3 X , there is acompact set K such that x â Ko â K â U .
Here are some functions spaces that will be extensively considered. Let Y bea topological space (usually either C or R or some subset of one of them).
âą C(X, Y ) = f : X â Y : f continuous.
âą C0(X, Y ) = f â C(X, Y ) : x : |f(x)| ℠Δ compact for all Δ > 0.
âą Cc(X, Y ) = f â C(X, Y ) : âK : f |Kc ⥠0 and K compact.
Clearly Cc(X, Y ) â C0(X, Y ) â C(X, Y ) with equalities if X is compact.Special cases are c00 â c0 â l1 â lp â lâ â C(N,C). These spaces will all beregarded as normed vector spaces, equipped with the uniform norm:
âfâu = supx|f(x)|.
Lemma 8.1 (Urysohnâs Lemma) Let X be LCH. If K is compact, U is open andK â U , then there exists a continuous function f â Cc(X, [0, 1]) such that f ⥠1on K and f ⥠0 on X \ U .
29
Let X be LCH and let ” be a measure on BX . One says that ” is outer regularon E â BX if
”(E) = inf”(U) : U â E,U open.One says that ” is inner regular on E if
”(E) = sup”(K) : K â E,K compact.
If ” is outer regular on all E â BX , one says that ” is outer regular and analo-gously for inner regularity. If ” is both outer and inner regular, then we say that ”is regular. A well-known example of a regular measure is Lebesgue measure.
If ”(K) < â for all compact K, ” is outer regular and ” is inner regular onall open sets, then ” is said to be a Radon measure (RM).
If ” is a RM and f â Cc(X,C), then ”x : |f(x)| > 0 < â. Hencef â L1(”) and we get
Cc(X,C) â L1(”).
Hence we can define the linear functional f ââ«f d”, f â Cc(X,C). The
given linear functional is positive in the sense that the result is positive wheneverf ℠0. Our first version of Riesz theorem will state that every positive linearfunctional, I , on Cc(X,R) can be written this way, i.e. that there exists a RM ”such that I(f) =
â«f d”, f â Cc(X,R). (Note that a positive linear functional I
on Cc(X,R) must be real-valued, since I(f) = I(f+) â I(fâ).) Since we willfor the time being, work with Cc(X,R) rather than Cc(X,C), we write Cc(X) forCc(X,R) until further notice. We will later come back to an extension to all linearfunctionals on C0(X,C).
Recall that the support, supp(f) of the function f is the closure of x : f(x) 6=0. For an open set U â X , we write f < U if 0 †f †1 and supp(f) â U .Note that f < U is a slightly stronger statement than f †ÏU . We will oftenmake use of a similarly stronger version of Urysohnâs Lemma than stated above.Namely for K â U there exists an f â Cc(X, [0, 1]) with f ⥠1 on K andf < U . This follows on, for each x â K, picking a compact set Kx such thatx â Ko
x â Kx â U . Then Kox is an open cover of K and can hence be reduced
to a finite subcover. Now apply Urysohn as above to K and the union of the setsin this subcover.
Theorem 8.2 (The Riesz Representation Theorem) Let I : Cc(X,R) â R bea positive linear functional. Then there is a unique Radon measure ” such that
I(f) =
â«f d”, f â Cc(X,R).
30
Moreover
(*) ”(U) = supI(f) : f < U for all open U ,
(**) ”(K) = infI(f) : f â„ ÏK for all compact K.
The proof will be divided into several parts.Proof of uniqueness and (*). Suppose I(f) =
â«f d”, f â Cc(X), where ” is a
RM. If U is open and f < U , then trivially I(f) †”(U). If K â U is compact,there exists by Urysohn an f â Cc(X) such that f |K ⥠1 and f < U . Hence”(K) †I(f). Now (*) follows from the inner regularity of ” on open sets. From(*) it now follows that ” is unique on opens sets, and hence on all measurable setsby ”âs outer regularity.Proof of existence and (**). The proof will rely on several claims that will beproved afterwards.
Define the set functions
”(U) = supI(f) : f < U, U open
and”â(E) = inf”(U) : U â E,U open, E â X.
Claim I. Ӊ is an outer measure.Claim II. All open sets are Ӊ-measurable.
By these claims, Caratheodoryâs Theorem tells us that ” := ”â|BX is a mea-sure. This measure satisfies (*) and is outer regular by definition. (The measure ”is an extension of the set function ” above, so we allow ourselves to use the samenotation for them.)Claim III. ” satisfies (**).
From this we can quickly show that ” must be a RM. That ” is finite oncompact sets follows immediately from (**). Also, if U is open and a < ”(U),then one can by (*) find f < U with I(f) > a. Let K = supp(f) â U . For anyg â„ ÏK we have g â„ f , so I(g) â„ I(f) > a. Hence ”(K) > a by (**).
At this point it only remains to prove:Claim IV. I(f) =
â«f d”, f â Cc(X).
Before proving the claims, we need a preparatory lemma.
Lemma 8.3 Let K be compact and let Ujnj=1 be an open cover of K. Then onecan find gj , j = 1, . . . , n, such that gj < Uj and
ân1 gj ⥠1 on K.
31
Proof. For each x â K, x â Uj for some j. Pick a compact set Nx such thatx â N o
x â Nx â Uj . Then N oxxâK is an open cover of K. Reduce to a finite
subcover Nx1 , . . . , Nxm . Let
Fj =â
k:NxkâUj
Nxk , j = 1, . . . , n.
Then Fj is compact, so there exists hj â Cc(X) such that ÏFj †hj < Uj . SinceâȘn1Fj contains K, we have
ân1 hj â„ 1 on K. Hence the open set
V := x :nâ1
hj > 0
contains K. Thus we can find f â Cc(X) with ÏK †f < V . Let hn+1 = 1â f .Then
ân+11 hj > 0 everywhere, so that we can define
gj :=hjân+11 hk
, j = 1, . . . , n.
Then supp(gj) = supp(hj) â Uj and since hn+1 = 0 on K,ân
1 gj ⥠1 on K. 2
Proof of Claim I. That ”â(â ) = 0 and A â B â ”â(A) †”â(B) is obvious,so it suffices to show that ”â(âȘâ1 Ej) â€
ââ1 ”â(Ej) for any Ej â X . For this it
suffices, by the definition of ”â, to show that ”(âȘâ1 Uj) â€ââ
1 ”(Uj) for open setsUj . By the definition of ”, this amounts to showing that I(f) â€
ââ1 ”(Uj) for
any f â Cc(X) with f < âȘâ1 Uj . Let K = supp(f). Since K is compact, there isan n such that K â âȘn1Uj . By the lemma above, there exist gj with gj < Uj andân
1 gj ⥠1 on K. However, then f =ân
1 fgj and fgj < Uj , so
I(f) =nâ1
I(fgj) â€nâ1
”(Uj) â€ââ1
”(Uj).
2
Proof of Claim II. We want to show that for an open set U ,
”â(E) ℠”â(E â© U) + ”â(E â© U c).
Fix Δ > 0. If E is open, then E â© U is open, so we can find f < E â© U such that”â(Eâ©U) = ”(Eâ©U) < I(f)+Δ. In the same way there exists g < E \ supp(f)such that ”(E \ supp(f)) < I(g) + Δ. Since f + g < E,
”â(E) = ”(E) â„ I(f) + I(g) > ”(E â© U) + ”(E \ supp(f))â 2Δ
32
℠”â(E â© U) + ”â(E â© U c)â 2Δ.
This proves the desired inequality for open sets. For general E, pick open V with”(V ) < ”â(E) + Δ. Then by the above,
”â(E) ℠”(V )â Δ ℠”â(V â© U) + ”â(V â© U c)â Δ
℠”â(E â© U) + ”â(E â© U c)â Δ.
2
Proof of Claim III. Fix a compact setK, Δ > 0 and f â„ ÏK . Let UΔ = x : f(x) >1 â Δ. Then UΔ is open and contains K. Hence, for any g < UΔ, g †f/(1 â Δ),so that I(g) †I(f)/(1 â Δ). Thus ”(K) †”(UΔ) †I(f)/(1 â Δ). Since Δ isarbitrary, ”(K) †infI(f) : f â„ ÏK.
For the reverse inequality it suffices, by the outer regularity of ” to show thatinfI(f) : f â„ ÏK †”(U) for any open U that contains K. However, byUrysohn, there is an f with ÏK †f < U and for this f , I(f) †”(U). 2
Proof of Claim IV. By linearity, we may assume without loss of generality that 0 â€f †1. Fix a large integer N . Let K0 = supp(f) and Kj = x : f(x) â„ j/N,j = 1, . . . , N . Let
fj(x) =
0, x 6â Kjâ1
f(x)â jâ1N, x â Kjâ1 \Kj
1N, x â Kj
Then f â Cc(X), f =âN
1 fj and ÏKj/N †fj †ÏKjâ1/N . Integrating gives
1
N”(Kj) â€
â«fj d” â€
1
N”(Kjâ1).
On the other hand, if U â Kjâ1 and U is open, then Nfj < U so that NI(fj) â€Â”(U). Since ” is outer regular, NI(fj) †”(Kjâ1). By (**), ”(Kj) †NI(fj).Thus
1
N”(Kj) †I(fj) â€
1
N”(Kjâ1).
We get
âŁâŁâŁI(f)ââ«f d”
âŁâŁâŁ †Nâ1
âŁâŁâŁI(fj)ââ«fj d”
âŁâŁâŁ †Nâ1
1
N
(”(Kjâ1)â ”(Kj)
)
33
†1
N
(”(K0)â ”(KN)
)†1
N”(supp(f))â 0
as N ââ. 2
We now have a characterization of positive linear functionals in Cc(X,R)â. Inthe end we want a characterization of all linear functionals in C0(X,C)â. For thatwe first need some approximation results. From now on the notation Cc(X) willbe understood to mean Cc(X,C). By definition, a RM is inner regular on all opensets. In fact, more is true:
Proposition 8.4 A Radon measure ” is inner regular on all Ï-finite sets.
Proof. Fix Δ > 0. If ”(E) < â, then we can find an open set U â E with”(U) < ”(E) + Δ. We can also pick compact F â U with ”(K) > ”(U) â Δ.Since ”(U \ E) < Δ we can also find open V â U \ E with ”(V ) < Δ. The setK := F \ V is compact and
”(K) ℠”(F )â ”(V ) > ”(E)â 2Δ.
The extension to Ï-finite E is straightforward. 2
Proposition 8.5 If ” is a Ï-finite Radon measure and Δ > 0, then for all E âBX , there exist a closed set F and an open set U such that F â E â U and”(U \ F ) < Δ
Proof. Write E = âȘjEj where ”(Ej) <â for all j. Pick open Uj â Ej suchthat ”(Uj) < ”(Ej) + Δ2â(j+1). With U = âȘjUj , we get ”(U) < ”(E) + Δ/2,so ”(U â© Ec) < Δ/2. Repeat the procedure for Ec to get open V â Ec with”(V ) < ”(Ec) + Δ/2, so ”(V â© E) < Δ/2. Let F = V c. Then
”(U \ F ) = ”(U ⩠V ) †”(U ⩠V ⩠Ec) + ”(U ⩠V ⩠E)
†”(U ⩠Ec) + ”(V ⩠E) < Δ.
2
An immediate consequence or Proposition 8.5 is that for any measurable setE one can find an FÎł-set A and GÎŽ-set B such that A â E â B and ”(A) =”(E) = ”(B).
A set A â X is said to be Ï-compact if it is the union of a countable family ofcompact sets. For example, if X is second countable (on top of being LCH), thenevery open set is Ï-compact.
34
Theorem 8.6 If every open set is Ï-compact, then any measure ” which is finiteon compact sets is regular, and hence a Radon measure.
Proof. Since ” is finite on compact sets, Cc(X) â L1(”) and hence theoperation I : Cc(X)â R given by
I(f) =
â«f d”
is a well defined positive linear functional. Hence there is a RM Μ such thatI(f) =
â«f dÎœ, f â Cc(X).
Let U be open. By assumption, we can write U = âȘjKj , there the Kjâs arecompact and increasing. By Urysohn, there are fj â Cc(X) such that ÏKj †fj <U and we may pick the fjâs so that f1 †f2 †. . .. Then by the MCT
”(U) = limn
â«fn d” = lim
n
â«fn dÎœ = Îœ(U).
Now pick an arbitrary E â BX . Pick, according to Proposition 8.5, open V andclosed F with F â E â V and Îœ(F ) + Δ/2 > Îœ(E) > /Îœ(V )â Δ/2. Then, sinceV \ F is open,
”(V \ F ) = Μ(V \ F ) < Δ.
Hence ”(V ) < ”(F ) + Δ †”(E) + Δ, so ” is outer regular. Also, since X itselfis open, we can write X = âȘjKj for increasing compact Kjâs. Thus ”(F ) =limj ”(F â©Kj), so for large enough j,
”(F â©Kj) > ”(F )â Δ > ”(V )â 2Δ ℠”(E)â 2Δ.
Since F â©Kj is compact, this proves inner regularity. 2
Proposition 8.7 (Approximation of measurable functions.) If ” is a Radonmeasure, then Cc(X,C) is dense in Lp(”) for every p â [1,â).
Proof. By Proposition 2.5, the set of simple functions is dense in Lp, so itsuffices to show that for any measurable E of finite measure and Δ > 0, there isan f â Cc(X,C) with âÏE â fâpp < Δ. By Proposition 8.4, ” is inner regularon E, so we can find a compact K and an open U such that K â E â U and”(U \K) < Δ. By Urysohn, there is an f â Cc(X,C) such that ÏK †f †ÏU .Hence
âÏE â fâpp †”(U \K) < Δ.
2
35
Theorem 8.8 (Luzinâs Theorem) Let ” be a Radon measure, f : X â C mea-surable and Δ > 0. If ”(E) < â, where E = x : f(x) 6= 0, then there existsa Ï â Cc(X) such that ”x : f(x) 6= Ï(x) < Δ. If f is bounded, then Ï can bechosen so that âÏâu †âfâu.
Proof. Recall Egoroffâs Theorem: if ” is a finite measure and fn â f a.e.,then there is a set A with ”(A) < Δ and fn â f uniformly on Ac.
Assume first that f is bounded. Then f â L1, so by the above propositionthere exist functions gn â Cc(X) such that gn â f in L1. Hence, by exercise18, one can find gn â Cc(X) with gn â f a.e. By Egoroff, there is an A â Esuch that ”(E \A) < Δ/3 and gn â f uniformly on Ac. By Proposition 8.4 thereis a compact K and an open U such that K â A â E â U , ”(A \ K) < Δ/3and ”(U \ E) < Δ/3. Since the gnâs are continuous and gn â f uniformly onK, f must be continuous on K. By Tietzeâs Extension Theorem, there exists acontinuous function g with g = f on K and supp(g) â U . Also, ”x : Ï(x) 6=f(x) †”(U \K) < Δ.
Finally let h : C â C be given by h(z) = z when |z| < âfâu and h(z) =âfâusgnz when |z| â„ âfâu. Then h is continuous and Ï = h g satisfies alsoâÏâu †âfâu.
The extension to unbounded f follows on picking M so that ”x : |f(x)| >M < Δ/2 and applying the above to fΟx:|f(x)|â€M. 2
Now we start to build up to the full version of Riesz theorem. First we claimthat C0(X) = Cc(X): if f â C0(X), let Kn = x : |f(x)| â„ 1/n which isa compact set. By Urysohn there is a gn â Cc(X) such that gn â„ ÏKn . Thengnf â Cc(X) and âgnf â fâu †1/n.
Let I be a bounded positive linear functional on Cc(X). For f â C0(X) pickfn â Cc(X) so that âfnâ fâu â 0. Since I is bounded, this implies that I(fn)is Cauchy and hence convergent. We can thus continuously extend I to C0(X):I(f) = limn I(fn). Moreover, by Riesz, I(f) =
â«f d”, f â Cc(X), for a RM
”. Thus, saying that I is bounded, i.e. âIâ < â, is equivalent to supâ«g d” :
g â Cc(X), âgâu †1 < â, which in turn is equivalent to ”(X) < â, by (*)applied to X . We also see that âIâ = ”(X) and since ” is finite it follows fromthe DCT that
I(f) = limn
â«fn d” =
â«f d”.
In summary: every positive bounded linear functional I â C0(X)â corresponds to
36
a finite Radon measure ” via the relation
I(f) =
â«f d”, f â C0(X).
Next we want a complete characterization of C0(X)â and not only for the positivelinear functionals there.
Theorem 8.9 (Jordan decomposition of linear functionals) For any real-valuedI â C0(X,R)â, there are positive linear functionals I+ and Iâ such that I =I+ â Iâ.
Proof. Let for all nonnegative f â C0(X,R)
I+(f) = supI(g) : 0 †g †f.
We first show that I+ is bounded and linear for nonnegative f âs. For any 0 †g â€f , |I(g)| †âIââgâu †âIââfâu, so |I+(f)| †âIââfâu.
Clearly I+(cf) = cI+(f) for c ℠0. We also need to show that I+(f1 + f2) =I+(f1) + I+(f2), f1, f2 ℠0. For this, observe that on the one hand, whenever0 †g1 †f1 and 0 †g2 †f2, I+(f1 + f2) ℠I(g1 + g2) + I(g1) + I(g2) andhence
I+(f1 + f2) â„ I+(f1) + I+(f2).
On the other hand, if 0 †g †f1 + f2, let g1 = min(f1, g) and g2 = g â g1. Then0 †g1 †f1 and 0 †g2 †f2. Hence
I+(f1) + I+(f2) â„ I+(f1 + f2).
Next, extend I+ to all f â C0(X,R) by
I+(f) = I+(f+)â I+(fâ).
To see that this is well defined, suppose that f also can be written as g â h,g, h â„ 0. Then f+ + h = g + fâ so by the above linearity, I+(f+) + I+(h) =I+(g) + I+(fâ), i.e. I+(f+) â I+(fâ) = I+(g) â I+(h). The functional I+ isobviously linear and
|I+(f)| †max(|I+(f+)|, |I+(fâ)|) †âIâmax(âf+âu, âfââu) = âIââfâu
so that âI+â †âIâ. Finally let Iâ = I+ â I , which is positive by the definitionof I+. 2
37
Assume now that I â C0(X)â and write J for the restriction of I to C0(X,R).Then, writing f = g + ih for f â C0(X), we have
I(f) = J(g) + iJ(h).
Let also J1 = <J and J2 = =J . By the decomposition lemma, we can writeJk = J+
k â Jâk for positive functionals J+
k and Jâk , k = 1, 2. By Riesz there areunique RMâs ”1, ”2 ,”3 and ”4 such that J+
1 (f) =â«f d”1, Jâ1 (f) =
â«f d”2,
J+2 =
â«f d”3 and J+
2 (f) =â«f d”4, f â C0(X,R). Putting this together, letting
” denote the complex measure ”1 â ”2 + i(”3 â ”4),
I(f) =
â«f d”, f â C0(X).
This already gives an extended version of Rieszâ theorem, but one can in fact domore. A complex measure ” is said to be Radon if all its parts are Radon measures.Let
M(X) := ” : BX â C : ” finite complex Radon measure.
Make M(X) into a NVS using the norm â”â = |”|(X). Recall that the measure|”|, the total variation of ” is defined as |”| = ”1 + ”2 + ”3 + ”4. Equivalently,for any measure Îœ such that ”k Îœ, k = 1, 2, 3, 4,
|”|(E) =
â«E
|f | dÎœ, E â BX
where f is a Radon-Nikodym derivative d”/dÎœ. From this is follows that for twocomplex measures ” and ”âČ, |”+ ”âČ| †|”|+ |”âČ|. This shows that â · â is indeeda norm.
Clearly |”| is Radon if all its parts are Radon. Conversely, if |”| is Radonand finite, then, for E â BX , there is an open U and a compact K such thatK â E â U and |”|(U \K) < Δ. This follows from combining Propositions 8.4and 8.5. Hence ”k(U \ K) < Δ for all parts of ”. This proves that ”k is regularand in particular Radon. In summary: a finite complex measure ” is Radon iff |”|is Radon.
Theorem 8.10 (Riesz Representation Theorem, complex version) For ” âM(X), let I”(f) =
â«f d”, f â C0(X). Then the map ” â I” is an isomet-
ric isomorphism from M(X) to C0(X)â.
38
Proof. That the map is injective is obvious and surjectivity was shown above.Linearity is also obvious, so it remains to prove that â”â = âI”â. However
|I”(f)| = |â«f d”| â€
â«|f |d|”| †|”|(X)âfâu
so âI”â †â”â. For the reverse inequality, let h = d”/d|”| and fix Δ > 0. ByLuzinâs Theorem there is an f â Cc(X) with âfâu †1 and f = h on Ec where|”|(E) < Δ/2. Since |h| = 1, we get
â”â =
â«|h|2d|”| =
â«hh d|”| =
â«h d”
â€âŁâŁâŁ â« (hâ f)d”
âŁâŁâŁ+âŁâŁâŁ â« f d”
âŁâŁâŁ †2|”|(E) +âŁâŁâŁ â« f d”
âŁâŁâŁ < Δ+ |Iu(f)|.
Hence â”â †âI”â. 2
Corollary 8.11 If X is compact, then M(X) and C(X)â are isometrically iso-morphic.
9 Vague convergence of measuresIn this section we study weak* convergence on M(X) âŒ= C0(X)â in a bit moredetail. Since weak* convergence of measures is usually referred to as weak con-vergence in probability theory, which is a stronger concept here (since C0(X) isusually not reflexive), a common compromise is to refer to weak* convergence asvague convergence. Hence, to say that ”n â ” vaguely means the same as sayingthat ”n â ” weakly*, i.e.
I”n(f)â I”(f), f â C0(X)
i.e. â«f d”n â
â«f d”, f â C0(X).
In probability theory, if ”n is the distribution of the random variable Οn and ” isthe distribution of ”, i.e. ”(E) = P(Ο â E), E â BR and analogously for ”n, and”n â ” vaguely, then this is also referred to as Οn â Ο in distribution.
For a positive measure in M(R), write F (x) = F”(X) = ”(ââ, x], x â R.The following result is important, in particular in probability theory.
39
Proposition 9.1 Assume that ”n, ” â M(R) are positive measures such thatsupn â”nâ <â. Then ”n â ” vaguely iff Fn(x)â F (x) for every x at which Fis continuous.
Proof. We prove only the backwards implication. We want to show thatâ«R f d”n â
â«R f d” for all f â C0(R). Assume first that f has compact support
and is continuously differentiable. Since F is increasing, F is continuous at allbut at most countably many points. In particular Fn â F a.e. By assumption,the Fnâs are uniformly bounded. Hence integration by parts, the DCT and the factthat f has compact support imply thatâ«
Rf d”n =
â«Rf dFn =
â«Rf âČ(x)Fn(x)dx
ââ«Rf âČ(x)F (x)dx =
â«Rf d”.
By the Stone-Weierstrass Theorem, the family, S, of continuously differentiablefunctions with compact support, is dense in C0(X). Fix Δ > 0 and let M =max(â”â, supn â”nâ). For f â C0(X) pick g â S such that âf â gâu < Δ/(3M).By the above |
â«R g d”n â
â«R g d”| < Δ/3 for sufficiently large n. An application
of the triangle inequality then givesâŁâŁâŁ â«Rf d”n â
â«Rf d”
âŁâŁâŁ < Δ.
2
10 Some useful inequalitiesTheorem 10.1 (Markovâs/Chebyshevâs inequality) Let f â Lp(”) and letEα =x : |f(x)| > α. Then
”(Eα) â€âfâppαp
.
Proof.
âfâpp =
â«X
|f |pd” â„â«Eα
|f |pd” ℠αp”(Eα).
2
40
Theorem 10.2 (Boundedness of integral operators) Let (X,M, ”) and (Y,N , Îœ)be two Ï-finite measure spaces and let f : X Ă Y â C beMâN -measurable.Let p â [1,â] and let T : Lp(Îœ)â Lp(”) be the linear operator given by
(Tf)(x) =
â«Y
K(x, y)f(y)Îœ(dy), f â Lp(Îœ).
If there exists a constant C <â such thatâ«X
|K(x, y)|”(dx) †C
for a.e. y â Y and â«Y
|K(x, y)|Μ(dy) †C
for a.e. x â X , then the integral in the definition of T converges absolutely and
âTfâLp(”) †CâfâLp(Îœ).
In particular, T is a well defined bounded linear operator.
Proof. We do the case 1 < p <â, leaving the easier cases p = 1 and p =âas an exercise. Let q be the conjugate exponent of p. We haveâ«
|K(x, y)f(y)|Μ(dy) =
â«|K(x, y)|1/q
(|K(x, y)1/p||f(y)|
)Μ(dy)
â€(â«|K(x, y)|Îœ(dy)
)1/q(â«|K(x, y)||f(y)|pÎœ(dy)
)1/p†C1/q
(â«|K(x, y)||f(y)|pÎœ(dy)
)1/pwhere the first inequality is Holder and the second inequality follows from as-sumption. By Tonelliâs Theoremâ«X
(â«Y
|K(x, y)f(y)|Μ(dy))p”(dx) †Cp/q
â«Y
|f(y)|pâ«X
|K(x, y)|”(dx)Μ(dy)
†Cp/q+1
â«Y
|f(y)|pÎœ(dy) <â
41
where the first inequality follows from the above, the second from assumption andthe third from the fact that f â Lp(Îœ). Hence
â«Y|K(x, y)f(y)|Îœ(dy) converges
for a.e. x as desired. Also,â«X
|(Tf)(x)|p”(dx) â€â«X
(â«Y
|K(x, y)f(y)|Îœ(dy))p”(dx) †Cp/q+1âfâpp
so âTfâp †Câfâp. 2
Theorem 10.3 (Minkowskiâs inequality for integrals) Let (X,M, ”) and (Y,N , Îœ)be Ï-finite, 1 †p <â and f aMâN -measurable nonnegative function. Then[ â«
X
(â«Y
(f(x, y)Μ(dy))p”(dx)
]1/pâ€â«Y
(â«X
f(x, y)p”(dx))1/p
Μ(dy).
In other words
ââ«Y
f(·, y)Îœ(dy)âLp(”) â€â«Y
âf(·, y)âLp(”)Îœ(dy).
Proof. For p = 1 this is an equality that follows immediately from TonelliâsTheorem, so assume p > 1 and let q be the conjugate exponent. If g â Lq(”) andâgâq = 1, thenâ«
X
(â«Y
f(x, y)Μ(dy))|g(x)|”(dx) =
â«Y
â«X
f(x, y)|g(x)|”(dx)Μ(dy)
â€â«Y
(â«X
f(x, y)p”(dx))1/p
Μ(dy)
where the inequality follows from Holder applied to the inner integral, keeping inmind that âgâq = 1. Now use Theorem 3.3, the reverse of Holder, with g as f , qas p and
â«Yf(·, y)Îœ(dy) as g. 2
11 Appendix on topologyThe aim of this appendix is to tie up the âloose endsâ from the section on RieszRepresentation Theorem, by proving the topological facts we used there.
42
Let X and Y be topological spaces and assume that the topology on Y isgenerated by E â P(Y ). Let f : X â Y . Then f is continuous iff fâ1(V ) isopen in X for every V â E . The forward implication is trivial. The backwardimplication follows from that inverse images commute with the elementary setoperations.
A topological space is said to be normal if for any disjoint closed sets A andB, there are disjoint open sets U and V with U â A and V â B.
Lemma 11.1 Let X be normal and A and B closed and disjoint subsets. LetI = k2ân : n â N, k = 1, 2, . . . , 2n. Then there exist open sets Ur, r â I suchthat A â Ur â Bc for all r and Ur â Us whenever r < s.
Proof. Let U1 = Bc. Pick open disjoint V â A andW â B and let U1/2 = V .Then
A â U1/2 â U1/2 â W c â Bc = U1.
Now use induction on n: suppose that we have sets Ur with the desired propertiesfor r â k2ân : n = 1, . . . , N â 1, k = 1, . . . , 2n and let r = i2âN for an oddi, i.e. r = (2j + 1)2âN for some j â 0, . . . , 2Nâ1 â 1. Then by the inductionhypothesis,
Uj2â(Nâ1) â U(j+1)2â(Nâ1) .
By normality, there exist disjoint open sets V â Uj2â(Nâ1) andW â U c(j+1)2â(Nâ1) .
ThenV â W c â U(j+1)2â(Nâ1)
so we can take U(2j+1)2âN = V . 2
Theorem 11.2 (Urysohnâs Lemma for Normal Spaces) Let X be normal andlet A and B be disjoint closed subsets. Then there exists an f â C(X, [0, 1]) suchthat f ⥠0 on A and f ⥠1 on B.
Proof. Let Ur be as in the lemma and let Ur = X for r > 1. Let
f(x) = infr â I : x â Ur.
Trivially f = 0 on A, f = 1 on B and f(x) â [0, 1] for all x. It remains to showthat f is continuous. However
fâ1[0, α) = x : f(x) < α = x : x â Ur for some r < α =âr<α
Ur,
43
an open set. Similarly
fâ1(α, 1] = x : f(x) > α = x 6â Ur for some r > α =âr>α
Urc
is also open. Since the sets [0, α) and (α, 1], α â [0, 1] generate the topology on[0, 1], we are done. 2
If A is a closed subset of a compact space X and Uα is an open cover of A,then Uα âȘ X \ A is an open cover of X . It follows that any closed subset ofa compact space is compact.
Suppose that X is Hausdorff, K a compact subset and x 6â K. Then thereare disjoint open sets U 3 x and V â K. To see this, pick for each y â K,disjoint open sets Uy 3 x and Vy 3 y. Then VyyâK is an open cover of K andcan hence be reduced to a finite subcover Vy1 , . . . , Vyn . Now take U = â©n1Uyk andV = âȘn1Vyk .
Proposition 11.3 Every compact Hausdorff space X is normal.
Proof. Let A and B be closed and disjoint. Since B is compact, we can foreach x â A find disjoint open sets Ux 3 x and Vx â B. Since A is compact,the open cover UxxâA can be reduced to a finite subcover Ux1 , . . . , Uxn . LetU = âȘn1Uxk and V = â©n1Vxk . Then U and V are open and disjoint, U â A andV â B. 2
Proposition 11.4 Let X be a locally compact Hausdorff space. Let U be openand x â U . Then there exists a compact set K such that x â Ko â K â U .
Proof. We may assume that U is compact, since if this is not the case, wecan by definition of local compactness find compact F with x â F o. Then we canreplace U with U â© F o whose closure is compact.
Since U is compact, âU is also compact. Hence there are sets U 3 x andW â âU that are open in U . Since V â U , V is open in X . We also have
V â U \W â U.
Hence V is compact, so we can take K = V . 2
Now suppose X is LCH, U open, K compact and K â U . For each x â K,pick the compact set Kx so that x â Ko
x â Kx â U . Then the Kxâs constitute
44
an open cover of K and can hence be reduced to a finite subcover Kx1 , . . . , Kxn .Let N = âȘn1Kxi . Then N is compact and K â N o â N â U . The subspaceN is compact and hence normal, so by Urysohn for normal spaces there is anf â C(N, [0, 1]) such that f ⥠1 on K and f ⥠0 on âN . Extend to the wholeof X by letting f ⥠0 on X \N . Then f â C(X, [0, 1]) and f < U . This provesUrysohnâs Lemma in the form most used.
Theorem 11.5 (The Tietze Extension Theorem) Let X be LCH, K â X com-pact and f â C(K, [0, 1]). Then there exists an F â C(X, [0, 1]) such that F = fon K and F = 0 outside a compact set.
Proof. Assume first that X is normal and A and B are closed an disjoint.We claim that there exist continuous functions gj with 0 †gj †2jâ1/3j andf â
âk1 gj †(2/3)k. To see this for j = 1, let C = fâ1[0, 1/3] and D =
fâ1[2/3, 1]. Then C and D are closed subsets of A and since A is closed, Cand D are closed in X . Hence, by Urysohn, there is a g1 â C(X, [0, 1/3]) withg1 = 1/3 on D and g1 = 0 on B âȘ C.
Now suppose that we have found g1, . . . , gnâ1 of the desired form. Then, inthe same way, there is a gn â C(X, [0, 2nâ1/3n]) such that gn = 2nâ1/3n on (f âânâ1
1 gj)â1[0, 2nâ1/3n] and gn = 0 on (f â
ânâ11 gj)
â1[(2/3)n, (2/3)nâ1] âȘ B.Since (2/3)nâ1 â 2nâ1/3n = (2/3)n, gn is also of the desired form.
Let F =ââ
1 gj . Clearly F = f on A and F = 0 on B. Since
âF ânâ1
gjâu â€âân+1
(2
3
)jâ 0,
the series converges uniformly and hence F is continuous.For the general statement, apply the above with A = K and B = âN , where
N is compact and K â N o, on the compact and hence normal subspace N of X .Extend by letting F ⥠0 on N c. 2
45