Exercises for Chapter 2
3.5 Reflexive Spaces
K1= &
x∈E
[−x,+x].
Let us recall that (arbitrary)products of compact spaces are compact—a deep theo- rem due to Tychonoff; see, e.g., H. L. Royden [1], G. B. Folland [2], J. R. Munkres [1], A. Knapp [1], or J. Dixmier [1]. ThereforeK1is compact. On the other hand, K2is closed inY; indeed, for eachfixedλ∈R,x, y∈Ethe sets
Ax,y= {ω∈Y;ωx+y−ωx−ωy=0}, Bλ,x = {ω∈Y;ωλx−λωx=0},
are closed inY (since the mapsω→ ωx+y−ωx−ωy andω →ωλx−λωxare continuous onY) and we may writeK2as
K2=' #
x,y∈E
Ax,y
(∩' #
x∈E λ∈R
Bλ,x (
.
Finally,Kis compact since it is the intersection of a compact set (K1) and a closed set (K2).
3.5 Reflexive Spaces
Definition. LetEbe a Banach space and letJ :E→Ebe the canonical injection fromEintoE(see Section 1.3). The spaceEis said to bereflexiveifJis surjective, i.e.,J (E)=E.
WhenEis reflexive,Eis usually identified withE.
Remark13.Many important spaces in analysis are reflexive. Clearly,finite-dimen- sionalspaces are reflexive (since dimE =dimE =dimE). As we shall see in Chapter 4 (see also Chapter 11),Lp(andp) spaces are reflexive for 1< p <∞. In Chapter 5 we shall see that Hilbert spaces are reflexive. However, equally important spaces in analysis arenotreflexive; for example:
• L1andL∞(and1,∞) are not reflexive (see Chapters 4 and 11);
• C(K),the space of continuous functions on an infinite compact metric spaceK, is not reflexive (see Exercise 3.25).
Remark14.It is essential to use J in the above definition. R. C. James [1] has constructed a striking example of a nonreflexive space with the property that there exists a surjective isometry fromEontoE.
Our next result describes a basic property of reflexive spaces:
•Theorem 3.17 (Kakutani).LetEbe a Banach space. ThenEis reflexive if and only if
BE= {x∈E; x ≤1}
is compact in the weak topologyσ (E, E).
Proof. Assume first thatE is reflexive, so thatJ (BE) = BE. We already know (by Theorem 3.16) thatBE is compact in the topologyσ (E, E). Therefore, it suffices to check thatJ−1is continuous fromEequipped withσ (E, E)with values inEequipped withσ (E, E). In view of Proposition 3.2, we have only to prove that for every fixedf ∈ E the mapξ → f, J−1ξ is continuous onE equipped withσ (E, E). Butf, J−1ξ = ξ, f, and the mapξ → ξ, fis indeed continuous onEfor the topologyσ (E, E). Hence we have proved that BEis compact inσ (E, E).
The converse is more delicate and relies on the following two lemmas:
Lemma 3.3 (Helly).LetE be a Banach space. Letf1, f2, . . . , fk be given inE and letγ1, γ2, . . . , γk be given inR. The following properties are equivalent:
(i)∀ε >0∃xε∈Esuch thatxε ≤1and
|fi, xε −γi|< ε ∀i=1,2, . . . , k, (ii)|k
i=1βiγi| ≤ k
i=1βifi ∀β1, β2, . . . , βk ∈R.
Proof. (i)⇒(ii). Fixβ1, β2, . . . , βkinRand letS=k
i=1|βi|. It follows from (i)
that
k
i=1
βifi, xε − k
i=1
βiγi ≤εS and therefore
k i=1
βiγi ≤
k i=1
βifi
xε +εS ≤
k i=1
βifi +εS.
Since this holds for everyε >0, we obtain (ii).
(ii)⇒(i). Setγ =(γ1, γ2, . . . , γk)∈ Rk and consider the mapϕ : E → Rk defined by
ϕ(x)=(f1, x, . . . ,fk, x).
Property (i) says precisely thatγ ∈ ϕ(BE). Suppose, by contradiction, that (i) fails, so thatγ /∈ϕ(BE). Hence{γ}andϕ(BE)may be strictly separated inRk by some hyperplane; i.e., there exists someβ=(β1, β2, . . . , βk)∈Rkand someα∈R such that
β·ϕ(x) < α < β·γ ∀x ∈BE. It follows that k
i=1
βifi, x < α <
k i=1
βiγi ∀x∈BE,
3.5 Reflexive Spaces 69
and therefore
k i=1
βifi ≤α <
k i=1
βiγi, which contradicts (ii).
Lemma 3.4 (Goldstine).LetEbe any Banach space. ThenJ (BE)is dense inBE with respect to the topologyσ (E, E), and consequentlyJ (E)is dense inEin the topologyσ (E, E).
Proof. Letξ ∈BE and letV be a neighborhood ofξ for the topologyσ (E, E).
We must prove thatV∩J (BE) = ∅. As usual, we may assume thatV is of the form V =$
η∈E; |η−ξ, fi|< ε ∀i=1,2, . . . , k%
for some given elementsf1, f2, . . . , fkinEand someε >0. We have to find some x∈BEsuch thatJ (x)∈V, i.e.,
|fi, x − ξ, fi|< ε ∀i=1,2, . . . , k.
Setγi = ξ, fi. In view of Lemma 3.3 it suffices to check that
k i=1
βiγi ≤
k
i=1
βifi ,
which is clear sincek
i=1βiγi =) ξ,k
i=1βifi
*
andξ ≤1.
Remark15.Note that J (BE)is closedin BE in the strong topology. Indeed, if ξn = J (xn) → ξ we see that (xn)is a Cauchy sequence in BE (sinceJ is an isometry) and thereforexn→x, so thatξ =J x. It follows thatJ (BE)is not dense inBE in the strong topology, unlessJ (BE)=BE, i.e.,Eis reflexive.
Remark16.See Problem 9 for an alternative proof of Lemma 3.4 (based on a variant of Hahn–Banach inE).
Proof of Theorem3.17,concluded.The canonical injectionJ :E→Eis always continuous fromσ (E, E)intoσ (E, E), since for every fixedf ∈Ethe map x → J x, f = f, xis continuous with respect toσ (E, E). Assuming thatBE is compact in the topologyσ (E, E), we deduce thatJ (BE)is compact—and thus closed—in E with respect to the topology σ (E, E). On the other hand, by Lemma 3.4,J (BE)is dense inBE for the same topology. It follows thatJ (BE)= BE and thusJ (E)=E.
In connection with the compactness properties of reflexive spaces we also have the following two results:
•Theorem 3.18.Assume thatEis a reflexive Banach space and let(xn)be a bounded sequence in E. Then there exists a subsequence(xnk)that converges in the weak topologyσ (E, E).
The converse is also true, namely the following.
Theorem 3.19 (Eberlein–Smulian).ˇ Assume that E is a Banach space such that every bounded sequence in E admits a weakly convergent subsequence (in σ (E, E)). ThenEis reflexive.
The proof of Theorem 3.18 requires a little excursion through separable spaces and will be given in Section 3.6. The proof of Theorem 3.19 is rather delicate and is omitted; see, e.g., R. Holmes [1], K. Yosida [1], N. Dunford–J. T. Schwartz [1], J. Diestel [2], or Problem 10.
Remark17.In order to clarify the connection between Theorems 3.17, 3.18, and 3.19 it is useful to recall the following facts:
(i) IfXis ametricspace, then
[Xis compact] ⇔ [every sequence inXadmits a convergent subsequence]. (ii) There existcompact topologicalspacesXand some sequences inXwithoutany
convergent subsequence. A typical example isX =BE, which is compact in the topologyσ (E, E); whenE=∞it is easy to construct a sequence inX withoutany convergent subsequence (see Exercise 3.18).
(iii) If X is a topological space with the property that every sequence admits a convergent subsequence, thenXneednotbe compact.
Here are some further properties of reflexive spaces.
•Proposition 3.20.Assume thatEis a reflexive Banach space and letM⊂Ebe a closed linear subspace ofE. ThenMis reflexive.
Proof. The spaceM—equipped with the norm ofE—has a priori two distinct weak topologies:
(a) the topology induced byσ (E, E), (b) its own weak topologyσ (M, M).
In fact, these two topologies are the same (since, by Hahn–Banach, every continu- ous linear functional onMis the restriction toMof a continuous linear functional on E). In view of Theorem 3.17, we have to check thatBM is compact in the topology σ (M, M)or equivalently in the topologyσ (E, E). However,BE is compact in the topologyσ (E, E)andMis closed in the topologyσ (E, E)(by Theorem 3.7).
ThereforeBMis compact in the topologyσ (E, E).
Corollary 3.21.A Banach spaceEis reflexive if and only if its dual spaceE is reflexive.
Proof. Ereflexive⇒E reflexive. The idea of the proof is simple, since, roughly speaking, we have thatE = E ⇒ E = E. More precisely, let J be the canonical isomorphism fromEinto E. Letϕ ∈ E be given. The map x → ϕ, J xis a continuous linear functional onE. Call itf ∈E, so that
3.5 Reflexive Spaces 71
ϕ, J x = f, x ∀x∈E.
But we also have
ϕ, J x = J x, f ∀x∈E.
SinceJ is surjective, we infer that
ϕ, ξ = ξ, f ∀ξ ∈E,
which means precisely that the canonical injection fromEintoEis surjective.
E reflexive⇒ E reflexive. From the step above we already know thatEis reflexive. SinceJ (E)is a closed subspace ofEin the strong topology, we conclude (by Proposition 3.20) thatJ (E)is reflexive. Therefore,Eis reflexive.3
•Corollary 3.22.LetE be a reflexive Banach space. LetK ⊂ Ebe a bounded, closed, and convex subset ofE. ThenKis compact in the topologyσ (E, E).
Proof. Kis closed for the topologyσ (E, E)(by Theorem 3.7). On the other hand, there exists a constantmsuch thatK⊂mBE, andmBEis compact inσ (E, E)(by Theorem 3.17).
•Corollary 3.23.LetEbe a reflexive Banach space and letA⊂Ebe a nonempty, closed, convex subset ofE. Letϕ:A→(−∞,+∞]be a convexl.s.c.function such thatϕ ≡ +∞and
(5) lim
x∈A x→∞
ϕ(x)= +∞ (no assumption ifAis bounded).
Thenϕachieves its minimum onA, i.e., there exists somex0∈Asuch that ϕ(x0)=min
A ϕ.
Proof. Fix anya∈Asuch thatϕ(a) <+∞and consider the set A˜= {x ∈A; ϕ(x)≤ϕ(a)}.
ThenA˜is closed, convex, and bounded (by (5)) and thus it iscompactin the topology σ (E, E)(by Corollary 3.22). On the other hand,ϕ is also l.s.c. in the topology σ (E, E) (by Corollary 3.9). It follows that ϕ achieves its minimum on A˜ (see property 5 following the definition of l.s.c. in Chapter 1), i.e., there existsx0 ∈ ˜A such that
ϕ(x0)≤ϕ(x) ∀x ∈ ˜A.
Ifx∈A\ ˜A, we haveϕ(x0)≤ϕ(a) < ϕ(x); therefore ϕ(x0)≤ϕ(x) ∀x∈A.
3It is clear that ifEandFare Banach spaces, andTis a linear surjective isometry fromEontoF, thenEis reflexive iffFis reflexive. Of course, there is no contradiction with Remark 14!
Remark18.Corollary 3.23 is the main reason whyreflexive spacesandconvex func- tionsare so important in many problems occurring in thecalculus of variationsand inoptimization.
Theorem 3.24.LetEandF be two reflexive Banach spaces. LetA:D(A)⊂E→ F be an unbounded linear operator that is densely defined and closed. ThenD(A) is dense inF. ThusAis well defined(A:D(A)⊂E →F)and it may also be viewed as an unbounded operator fromEintoF. Then we have
A=A.
Proof.
1. D(A)is dense inF. Let ϕ be a continuous linear functional on F that vanishes onD(A). In view of Corollary 1.8 it suffices to prove thatϕ ≡0 onF. SinceF is reflexive,ϕ∈F and we have
(6) w, ϕ =0 ∀w∈D(A).
Ifϕ =0 then[0, ϕ]∈/ G(A)inE×F. Thus, one may strictly separate[0, ϕ]and G(A)by a closed hyperplane inE×F; i.e., there exist some[f, v] ∈E×Fand someα∈Rsuch that
f, u + v, Au< α <v, ϕ ∀u∈D(A).
It follows that
f, u + v, Au =0 ∀u∈D(A) and
v, ϕ =0.
Thusv∈D(A), and we are led to a contradiction by choosingw=vin (6).
2.A=A. We recall (see Section 2.6) that I[G(A)] =G(A)⊥ and
I[G(A)] =G(A)⊥. It follows that
G(A)=G(A)⊥⊥=G(A), sinceAis closed.