F|⊂
i
B(gi, ε) withgi ∈Lp().
Set
¯ gi(x)=
gi(x) in,
0 onRN\.
It is clear thatFis covered by the ballsB(g¯i,2ε)inLp(RN).
Remark13.The converse of Corollary 4.27 is also true (see Exercise 4.34). Therefore we have a complete characterization of compact sets inLp(RN).
We conclude with a useful application of Theorem 4.26:
Corollary 4.28.LetGbe a fixed function inL1(RN)and let F =G B,
whereB is a bounded set inLp(RN)with1 ≤ p < ∞. ThenF| has compact closure inLp()for any measurable setwith finite measure.
Proof. ClearlyFis bounded inLp(RN). On the other hand, if we writef =G u withu∈Bwe have
τhf −fp= (τhG−G) up≤CτhG−G1, and we conclude with the help of the following lemma:
Lemma 4.3.LetG∈Lq(RN)with1≤q <∞. Then
hlim→0τhG−Gq=0.
Proof. Givenε >0, there exists (by Theorem 4.12) a functionG1∈Cc(RN)such thatG−G1q< ε.
We write
τhG−Gq≤ τhG−τhG1q+ τhG1−G1q+ G1−Gq
≤2ε+ τhG1−G1q. Since limh→0τhG1−G1q =0 we see that
lim sup
h→0 τhG−Gq≤2ε ∀ε >0.
4.5 Comments on Chapter 4 115 Theorem 4.29 (Egorov).Assume thatis a measure space with finite measure.
Let(fn)be a sequence of measurable functions onsuch that fn(x)→f (x)a.e. on(with|f (x)|<∞a.e.).
Then∀ε > 0 ∃A⊂measurable such that|\A| < εandfn → f uniformly onA.
For a proof, see Exercise 4.14, P. Halmos [1], G. B. Folland [2], E. Hewitt–
K. Stromberg [1], R. Wheeden–A. Zygmund [1], K. Yosida [1], A. Friedman [3], etc.
2. Weakly compact sets inL1.
SinceL1is not reflexive, bounded sets ofL1do not play an important role with respect to the weak topologyσ (L1, L∞). The following result provides a useful characterization of weakly compact sets ofL1.
Theorem 4.30 (Dunford–Pettis).LetFbe a bounded set inL1(). ThenF has compact closure in the weak topologyσ (L1, L∞)if and only ifFis equi-integrable, that is,
(a)
⎧⎨
⎩
∀ε >0 ∃δ >0 such that
A
|f|< ε ∀A⊂,measurable with|A|< δ, ∀f ∈F and
(b)
⎧⎨
⎩
∀ε >0 ∃ω⊂, measurable with|ω|<∞such that
\ω
|f|< ε ∀f ∈F.
For a proof and discussion of Theorem 4.30 see Problem 23 or N. Dunford–
J. T. Schwartz [1], B. Beauzamy [1], J. Diestel [2], I. Fonseca–G. Leoni [1], and also J. Neveu [1], C. Dellacherie–P. A. Meyer [1] for the probabilistic aspects; see also Exercise 4.36.
3. Radon measures.
As we have just pointed out, bounded sets ofL1enjoy no compactness properties.
To overcome this lack of compactness it is sometimes very usefulto embedL1into a large space: the space of Radon measures.
Assume, for example, thatis a bounded open set ofRN with the Lebesgue measure. Consider the spaceE =C()with its normu = supx∈ |u(x)|. Its dual space, denoted byM(), is called the space ofRadon measureson.The weaktopology onM()is sometimes called the “vague” topology.
We shall identifyL1()with a subspace ofM(). For this purpose we introduce the mappingL1()→M()defined as follows. Givenf ∈L1(), the mapping u∈C()→
f u dxis a continuous linear functional onC(), which we denote Tf, so that
Tf, uE,E=
f u dx ∀u∈E.
ClearlyT is linear, and, moreover,T is anisometry, since TfM()= sup
u∈E u≤1
f u= f1 (see Exercise 4.26).
UsingT we may identify L1()with a subspace ofM(). Since M()is the dual space of the separable spaceC(), it has some compactness properties in the weaktopology. In particular, if(fn)is a bounded sequence inL1(),there exist a subsequence(fnk)and a Radon measureμsuch thatfnk μ in the weaktopology σ (E, E), that is,
fnku→ μ, u ∀u∈C().
For example, a sequence inL1 can converge to a Dirac measure with respect to the weak topology. Some futher properties of Radon measures are discussed in Problem 24.
The terminology “measure” is justified by the following result, which connects the above definition with the standard notion of measures in the set-theoretic sense:
Theorem 4.31 (Riesz representation theorem).Letμbe a Radon measure on. Then there is a unique signed Borel measureν on(that is, a measure defined on Borel sets of)such that
μ, u =
udν ∀u∈C().
It is often convenient to replace the spaceE=C()by the subspace E0= {f ∈C();f =0 on the boundary of}.
The dual ofE0 is denoted byM()(as opposed toM()). The Riesz repre- sentation theorem remains valid with the additional condition that|ν|(boundary of )=0.
On this vast and classical subject, see, e.g., H. L. Royden [1], W. Rudin [2], G. B. Folland [2], A. Knapp [1], P. Malliavin [1], P. Halmos [1], I. Fonseca–
G. Leoni [1].
4. The Bochner integral of vector-valued functions.
Letbe a measure space and letEbe a Banach space. The spaceLp(;E)consists of all functionsf defined on with values intoE that are measurable in some appropriate senseand such that
f (x)pdμ <∞(with the usual modification whenp = ∞). Most of the properties described in Sections 4.2 and 4.3 still hold under some additional assumptions onE. For example, ifEis reflexive and 1< p <
∞, thenLp(;E)is reflexive and its dual space isLp(;E). For more details,
4.5 Comments on Chapter 4 117 see K. Yosida [1], D. L. Cohn [1], E. Hille [1], B. Beauzamy [1], L. Schwartz [3].
The spaceLp(;E)is very useful in the study of evolution equations whenis an interval inR(see Chapter 10).
5. Interpolation theory.
The most striking result, which began interpolation theory, is the following.
Theorem 4.32 (Schur, M. Riesz, Thorin).Assume thatis a measure space with
||<∞, and thatT :L1()→L1()is a bounded linear operator with norm M1= TL(L1,L1).
Assume, in addition, thatT : L∞() → L∞()is a bounded linear operator with norm
M∞= TL(L∞,L∞).
ThenT is a bounded operator fromLp()intoLp()for all1< p <∞, and its normMpsatisfies
Mp≤M11/pM∞1/p.
Interpolation theory was originally discovered by I. Schur, M. Riesz, G. O. Thorin, J. Marcinkiewicz, and A. Zygmund. Decisive contributions have been made by a number of authors including J.-L. Lions, J. Peetre, A. P. Calderon, E. Stein, and E. Gagliardo. It has become auseful tool in harmonic analysis(see, e.g., E. Stein–
G. Weiss [1], E. Stein [1], C. Sadosky [1]) and in partial differential equations (see, e.g., J.-L. Lions–E. Magenes [1]). On these questions see also G. B. Folland [2], N. Dunford–J. T. Schwartz [1] (Volume 1 p. 520), J. Bergh–J. Löfström [1], M. Reed–B. Simon [1], (Volume 2, p. 27) and Problem 22.
6. Young’s inequality.
The following is an extension of Theorem 4.15.
Theorem 4.33 (Young).Assumef ∈Lp(RN)andg∈Lq(RN)with1≤p ≤ ∞, 1≤q ≤ ∞and 1r =p1 +q1−1≥0.
Thenf g∈Lr(RN)andf gr ≤ fpgq. For a proof see, e.g., Exercise 4.30.
7. The notion of convolution—extended to distributions (see L. Schwartz [1] or A. Knapp [2])—plays a fundamental role in the theory of partial differential equa- tions. For example, the equationP (D)u=f inRN, whereP (D)is any differential operator with constant coefficients, has a solution of the formu=E f, whereE is thefundamental solutionofP (D) (theorem of Malgrange–Ehrenpreis; see also Comment 2b in Chapter 1). In particular, the equationu=f inR3has a solution of the formu=E f, whereE(x)= −(4π|x|)−1.