Theorem 16.1 (Hahn–Banach). Let (X,∥⋅∥X) be a normed space, let V⊂X be a subspace, with the norm ∥⋅∥X, and let L∈V∗. Then there exists T∈X∗ such that
Tv=Lvfor every v∈V,that is,T∣V=L,∥T∥X∗=∥L∥V∗,that is, the norm does not change.
Corollary 16.2. Let (X,∥⋅∥X) be a normed space and let x0∈X with x0=0. Then there exists T∈X∗ such that ∥T∥X∗=1 and Tx0=∥x0∥X.
Proof (complete). Since x0=0, we can consider the subspace V=Span(x0):={tx0:t∈R}. Let L:V→R be given by L(tx0)=t∥x0∥X. The map L is linear, and it belongs to V∗ with ∥L∥V∗=1, because
By the Hahn–Banach theorem, Theorem 16.1, there exists T∈X∗ such that T∣V=L and ∥T∥X∗=∥L∥V∗=1. Moreover, Tx0=Lx0=∥x0∥X, since x0∈V.■
Definition 16.3 (Hyperplane). Let (X,∥⋅∥X) be a normed space and let T:X→R be linear. A hyperplane is a subset of X given, for some α∈R, by
H={x∈X:Tx=α}.
Notation. We write H={T=α}=T−1({α}).
Remark 16.4.(i) If α=0, then H={T=0}=KerT. (ii) If X=RM for a positive integer M, so that T:RM→R, and T is not identically zero, then H={T=α} is a classical hyperplane.
Corollary 16.5 (Separation of points). Let (X,∥⋅∥X) be a normed space and let x,y∈X with x=y. Then there exists T∈X∗ such that Tx=Ty. That is, the hyperplane H={T=α} with α=Tx contains x and does not contain y, as in Figure 16.1.
Figure 16.1. The drawing of Corollary 16.5: the hyperplane H={T=α}, with α=Tx, drawn as a line, the point x on it and the point y outside it.
Proof (complete). Let x0:=x−y. Then x0=0, because x=y. By Corollary 16.2 there exists T∈X∗ such that Tx0=∥x0∥X, where Tx0=T(x−y)=Tx−Ty. Hence, since ∥x0∥X>0,
Tx=Ty+∥x0∥X>Ty.■
Corollary 16.6 (Separation of point and subspace). Let (X,∥⋅∥X) be a normed space and let V⊂X be a closed subspace with V=X. Let x0∈X∖V. Then there exists T∈X∗ such that Tx0=0 and Tv=0 for every v in V. That is, x0∈/KerT and V⊂KerT, as in Figure 16.2.
Figure 16.2. The drawing of Corollary 16.6: the kernel KerT, drawn as a plane, the subspace V, drawn as a line through 0 inside it, and the point x0 outside it.
Proof (complete). Since V is closed and x0∈/V,
d(x0,V):=v∈Vinf∥x0−v∥X>0,
as in the proof of the Riesz theorem on closed balls, Theorem 12.18. Let
Z:=Span(V,x0)={v+tx0:v∈V,t∈R},
and let L:Z→R be given by L(v+tx0)=t. Indeed, L is well defined, because two writings v+tx0=v′+t′x0 of the same element with t=t′ would give x0=(v′−v)/(t−t′)∈V. Notice that L(x0)=1 and that L(v)=0 for every v∈V. The map L is linear, and it is continuous, that is, L∈Z∗. Let z=v+tx0∈Z, with v∈V and t∈R.
Case 1. If t=0, then ∣Lz∣=∣Lv∣=0.
Case 2. If t=0, then ∣Lz∣=∣L(v+tx0)∣=∣t∣, and
By Cases 1 and 2, ∣Lz∣≤∥z∥X/d(x0,V) for every z∈Z, hence L∈Z∗. By the Hahn–Banach theorem, Theorem 16.1, there exists T∈X∗ such that T∣Z=L. In particular, Tx0=Lx0=1 and Tv=Lv=0 for every v∈V.■
Definition 16.7 (Bidual space). Let (X,∥⋅∥X) be a normed space, and let X∗=L(X,R) be its dual, a Banach space by Remark 15.13 (i). The bidual space of X is X∗∗:=L(X∗,R).
Notation. For T in X∗,x in X and ϕ in X∗∗, we write
Definition 16.8 (Canonical map). Let (X,∥⋅∥X) be a normed space. The canonical map, also called the evaluation map, is the map τ:X→X∗∗,x↦τ(x)∈X∗∗, defined by
∗∗⟨τ(x),T⟩∗:=∗⟨T,x⟩=Tx
for every T∈X∗. That is, τ(x) acts on any T∈X∗ by evaluating T at x.
Theorem 16.9. Let (X,∥⋅∥X) be a Banach space and let τ be its canonical map. Then: (i) the map τ is linear and continuous; (ii) the map τ:(X,∥⋅∥X)→(X∗∗,∥⋅∥X∗∗) is an isometry, that is, ∥τ(x)∥X∗∗=∥x∥X for every x∈X; (iii) the range R(τ) is closed in X∗∗; (iv) the map τ:(X,∥⋅∥X)→(R(τ),∥⋅∥X∗∗) is an isometric isomorphism.
Proof (complete). The map τ is linear: for all x1,x2∈X,α,β∈R and T∈X∗,
In the last equality, ∣Tx∣≤∥T∥X∗∥x∥X≤∥x∥X for every T∈X∗ with ∥T∥X∗≤1. Moreover, if x=0,Corollary 16.2 applied to x gives a functional T of X∗ with ∥T∥X∗=1 and Tx=∥x∥X, so that
In detail, for x=0 both sides of the equality to prove vanish, because τ(0)=0 by linearity. Therefore τ is an isometry, hence bounded, and so continuous.
Corollary 15.9, applied to τ, gives the rest of the theorem. Indeed, X is a Banach space by assumption, and X∗∗=(X∗)∗ is a Banach space by Remark 15.13 (i). ■
Definition 16.10 (Reflexive space). Let (X,∥⋅∥X) be a Banach space. If the canonical map τ:X→X∗∗ is surjective, that is, R(τ)=X∗∗, we say that X is reflexive.
Remark 16.11. Let (X,∥⋅∥X) be a reflexive Banach space. Then τ:(X,∥⋅∥X)→(X∗∗,∥⋅∥X∗∗) is an isometric isomorphism.
Remark 16.12. Let (Ω,M,μ) be a measure space, let p∈(1,∞) and let p′∈(1,∞) be such that 1/p+1/p′=1. Consider Lp=Lp(Ω,M,μ).
(i) For every f∈Lp′ we can define a corresponding functional Tf∈(Lp)∗ by
Tf(g):=∫Ωfgdμfor every g∈Lp,
and ∥Tf∥(Lp)∗=∥f∥p′. (ii) The map F:Lp′→(Lp)∗,f↦Tf, is an isometry.
Proof (complete). By the Hölder inequality, Proposition 10.12, ∣Tf(g)∣≤∥f∥p′∥g∥p for every g∈Lp, hence ∥Tf∥(Lp)∗≤∥f∥p′. Now let g=∣f∣p′−1sgnf, where sgn is the sign function. Notice that g∈Lp, since ∣g∣p=∣f∣p(p′−1)=∣f∣p′ and ∫Ω∣g∣pdμ=∫Ω∣f∣p′dμ<+∞. With this g,
Hence ∥Tf∥(Lp)∗=∥f∥p′. Indeed, if f=0, then ∥g∥p>0 and ∥Tf∥(Lp)∗≥∣Tf(g)∣/∥g∥p=∥f∥p′, while for f equal to 0 both sides vanish. We found that the map F is an isometry, which is (ii). ■
The question raised by Remark 16.12 (ii) is whether F is surjective. The answer is yes.
Theorem 16.13 (Riesz representation in Lp). Let (Ω,M,μ) be a measure space.
(1) If p∈(1,∞), then for every T∈(Lp)∗ there exists a unique f∈Lp′ such that
Tg=∫Ωfgdμfor every g∈Lp.
Moreover, ∥T∥(Lp)∗=∥f∥p′, and Lp is reflexive. The idea of the reflexivity is that (Lp)∗ is identified with Lp′ by the first part of (1), and (Lp′)∗ with Lp by the first part of (1) applied to p′, since (p′)′=p by Remark 10.10 (i):
Lp⇝(Lp)∗∣≀Lp′⇝⇝(Lp)∗∗↓↑(Lp′)∗≃Lp.
(2) If p=1 and (Ω,M,μ) is σ-finite, then for every T∈(L1)∗ there exists a unique f∈L∞ such that
Tg=∫Ωfgdμfor every g∈L1.
Moreover, ∥T∥(L1)∗=∥f∥∞.
Remark 16.14. Let (Ω,M,μ) be a measure space. In general, (L∞)∗ is not isometrically isomorphic to L1, and L1 is not reflexive.
Proof (sketch). Consider ([0,1],L([0,1]),λ). We want to find T∈(L∞)∗ that is not represented by an L1 function, that is, such that T=Tf for every f∈L1, where
Tf(g):=∫01fgdλfor every g∈L∞.
Consider (C([0,1]),∥⋅∥∞)⊂L∞, a closed subspace of L∞, and fix f0∈L∞∖C([0,1]). By Corollary 16.6 there exists T∈(L∞)∗ such that Tf0=0 and T vanishes on C([0,1]).
We claim that there is no u∈L1 such that T=Tu. Suppose, by contradiction, that there exists u∈L1 such that T=Tu, that is,
Tf=Tuf=∫01fudλfor every f∈L∞.
Then Tug=Tg=0 for every g∈C([0,1]), that is,
∫01gudλ=0for every g∈C([0,1]),
hence u=0 almost everywhere. Indeed, every g∈Cc∞(0,1), extended by 0, belongs to C([0,1]), so that (1) of Corollary 5.16 applies. Therefore, as Tf0=0 and u=0 almost everywhere, we reach the contradiction
Theorem 16.15 (Reflexivity versus topology). Let (X,∥⋅∥X) be a Banach space. (1) If X is reflexive and V⊂X is a closed subspace, then V is reflexive. (2) The space X is reflexive if and only if X∗ is reflexive. (3) If X∗ is separable, then X is separable. (4) The space X is reflexive and separable if and only if X∗ is reflexive and separable.
Proof (sketch). Only the idea of (2) is given. If X is reflexive, then, identifying X∗∗ with X,
(X∗)∗∗=X∗∗∗=(X∗∗)∗=X∗.
Conversely, let X∗ be reflexive. Then X∗∗=(X∗)∗ is reflexive, by the first implication applied to the Banach space X∗. The range τ(X) of the canonical map τ:X→X∗∗ is closed in X∗∗ by Theorem 16.9, hence τ(X) is reflexive by (1), and therefore X is reflexive. ■
Example 16.18. Let X=L2((0,∞)), with the Lebesgue measure λ. By Theorem 16.13 (1) with p=p′=2,X∗ is isometrically isomorphic to L2((0,∞)). Notice that, for (fm)⊂L2((0,∞)) and f∈L2((0,∞)),
fm⇀f⟺Tfm→Tffor every T∈(L2)∗⟺∫0∞fmgdλ→∫0∞fgdλfor every g∈L2((0,∞)),
by Definition 16.16 and by Theorem 16.13 (1). Let fm=χ[m,m+1] for every m∈N. Then fm→0 almost everywhere, and fm↛0 in L2. Indeed, ∥fm∥2=1 for every m. Moreover, for every g∈L2((0,∞)), by the Hölder inequality, Proposition 10.12,
Remark 16.19 (Weak convergence in Lp). Let (Ω,M,μ) be a σ-finite measure space, let p∈[1,∞), and let (fm)⊂Lp and f∈Lp. (i) We have fm⇀f w-Lp if and only if ∫Ωfmgdμ→∫Ωfgdμ for every g∈Lp′. (ii) If fm⇀f1 w-Lp and fm→f2 almost everywhere, then f1=f2 almost everywhere.