Corollary 15.1. Let (X,d) be a nonempty complete metric space and let (Cm) be a sequence of closed sets in X such that ⋃mCm=X. Then there exists m0 such that int(Cm0)=∅.
Theorem 15.2 (Banach–Steinhaus). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let F⊂L(X,Y) be a family of linear continuous operators from X to Y. Suppose that for every x∈X there exists Mx>0 such that
∥Tx∥Y≤Mxfor every T∈F.
Then there exists M>0 such that ∥T∥L≤M for every T∈F. That is, ∥Tx∥Y≤M∥x∥X for every T∈F and every x∈X.
Proof (complete). For every m∈N let Cm={x∈X:∥Tx∥Y≤mfor every T∈F}. (i) The set Cm is closed. Let (xk)⊂Cm be such that xk→x in X. Then, for every T∈F,
∥Tx∥Y=k→∞lim∥Txk∥Y≤m,
since ∥Txk∥Y≤m for every k. Indeed, Txk→Tx because T is continuous, and the norm is continuous by Remark 10.3. Hence x∈Cm, and Cm is closed by Lemma 1.20 (i).
(ii) The union of the sets Cm is X. Let x∈X. By assumption there exists Mx>0 such that ∥Tx∥Y≤Mx for every T∈F. Then x∈Cm for every integer m≥Mx, since ∥Tx∥Y≤Mx≤m for all the operators of F.
By Corollary 15.1, applied in X, there exists m0∈N such that int(Cm0)=∅. That is, there exist a point x0 of X and r>0 such that Br(x0)⊂Cm0.
We want to estimate the supremum of ∥Tz∥Y over z∈B1(0). Let z∈B1(0). Then x0+rz∈Br(x0)⊂Cm0, hence ∥T(x0+rz)∥Y≤m0 for every T∈F. On the other hand, by the linearity of T and the triangle inequality,
where the second inequality is the assumption at the point x0. Therefore ∥T∥L≤M for every T∈F. In detail, for t∈(0,1) and x in the closed unit ball of X, the vector tx lies in B1(0), so that t∥Tx∥Y≤M, and letting t tend to 1 gives ∥Tx∥Y≤M.■
Corollary 15.3. Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let (Tm)⊂L(X,Y). Suppose that for every x∈X the limit of the sequence (Tmx) of points of Y exists in Y, that is, (Tm) converges pointwise on X. Then the map T:X→Y given by
Tx:=m→∞limTmx
belongs to L(X,Y). In particular, T is continuous even though the convergence is only pointwise.
Proof (complete). The map T is linear: for all x1,x2∈X and α,β∈R,
For every x∈X, the sequence (Tmx) converges in Y, hence it is bounded in Y. That is, there exists Mx>0 such that ∥Tmx∥Y≤Mx for every m. By the Banach–Steinhaus theorem, Theorem 15.2, applied to F={Tm:m∈N}, there exists M>0 such that ∥Tm∥L≤M for every m. Hence, for every z∈B1(0),
∥Tz∥Y=m→∞lim∥Tmz∥Y≤M,
where the equality is the continuity of the norm, Remark 10.3, and the inequality holds because ∥Tmz∥Y≤∥Tm∥L∥z∥X≤M for every m. Therefore ∥T∥L≤M, as at the end of the proof of Theorem 15.2, and the linear map T is bounded, hence continuous: T∈L(X,Y).■
Definition 15.4 (Open function). Let (X,dX) and (Y,dY) be metric spaces. A function f:(X,dX)→(Y,dY) is said to be open if f(U) is open in Y for every open set U⊂X.
Theorem 15.5 (Open mapping). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let T∈L(X,Y). If T is surjective, then T is open.
Proof (sketch). For r>0 and points x of X and y of Y, let BrX(x) and BrY(y) be the open balls of radius r centred at x in X and at y in Y. (i) It is sufficient to prove that T(B1X(0))⊃BrY(y0) for some y0∈Y and r>0. (ii) We can prove that T(B1X(0))⊃BrY(y0) for some y0∈Y and r>0. For (ii), the sets T(BmX(0)) with m≥1 are closed and, since T is surjective,
m≥1⋃T(BmX(0))=R(T)=Y.
Hence, by Corollary 15.1, applied in Y, there exist an integer m0≥1, a point z0∈Y and ρ>0 such that
BρY(z0)⊂T(Bm0X(0)).
In detail, T(Bm0X(0))=m0T(B1X(0)) by linearity, so that (ii) holds with y0=z0/m0 and r=ρ/m0.■
Corollary 15.6. Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let T∈L(X,Y). Suppose that T is bijective. Then T−1∈L(Y,X).
Proof (complete). Consider T−1:Y→X. We want to prove that (T−1)−1(A) is open in Y for every open set A⊂X, that is, that T−1 is continuous, as in Definition 1.17. In detail, T−1 is linear, as the inverse of a bijective linear map, so that its continuity gives T−1∈L(Y,X). Let A⊂X be open. Then
(T−1)−1(A)=T(A),
which is open, because T is an open map by Theorem 15.5, the open mapping theorem. ■
Corollary 15.7 (Embeddings). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let T∈L(X,Y) be injective, with R(T) closed. Then T:(X,∥⋅∥X)→(R(T),∥⋅∥Y) is an isomorphism, that is, T−1:R(T)→X is linear and continuous.
Proof (complete). The space (R(T),∥⋅∥Y) is a Banach space. Indeed, R(T) is a subspace of Y, closed by assumption, and a closed subspace of a Banach space is a Banach space. Hence T:(X,∥⋅∥X)→(R(T),∥⋅∥Y) is a bijective map between two Banach spaces, and T−1:R(T)→X is continuous by Corollary 15.6. ■
Corollary 15.8 (Equivalent norms). Let X be a vector space with two norms ∥⋅∥a and ∥⋅∥b such that (X,∥⋅∥a) and (X,∥⋅∥b) are both Banach spaces. Suppose that there exists C1>0 such that ∥x∥b≤C1∥x∥a for every x∈X. Then there exists C2>0 such that ∥x∥a≤C2∥x∥b for every x∈X, that is, ∥⋅∥a and ∥⋅∥b are equivalent.
Proof (complete). The assumption says that the map id:(X,∥⋅∥a)→(X,∥⋅∥b),x↦x, is continuous: ∥idx∥b=∥x∥b≤C1∥x∥a for every x in X. The map id is also bijective. Hence, by Corollary 15.6, the map id:(X,∥⋅∥b)→(X,∥⋅∥a) is also continuous. That is, there exists C2>0 such that ∥idx∥a=∥x∥a≤C2∥x∥b for every x∈X.■
Corollary 15.9 (Isometries). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let T∈L(X,Y) be an isometry, that is, ∥Tx∥Y=∥x∥X for every x∈X. Then: (1)R(T) is closed; (2)T:(X,∥⋅∥X)→(R(T),∥⋅∥Y) is an isometric isomorphism.
Proof (complete). (1) Let (ym)⊂R(T) be such that ym→y in Y. Since ym∈R(T), there exists xm∈X such that ym=Txm. Now, for all m and n,
Since (ym) converges, (ym) is Cauchy in Y. Hence (xm) is Cauchy in X, and, since X is a Banach space, xm→x in X for some x∈X. Since T is continuous, y=limmym=limmTxm=Tx, hence y∈R(T), and R(T) is closed by Lemma 1.20 (i).
(2) The range R(T) is closed by (1), and T is injective. Indeed, if Tx=0, then ∥x∥X=∥Tx∥Y=0. By Corollary 15.7, T:(X,∥⋅∥X)→(R(T),∥⋅∥Y) is then an isomorphism, and it is also an isometry, that is, an isometric isomorphism. ■
Corollary 15.10 (Maps bounded above and below). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let T∈L(X,Y). Suppose that there exist C1,C2>0 such that C1∥x∥X≤∥Tx∥Y≤C2∥x∥X for every x∈X. Then: (0)T is injective; (1)R(T) is closed; (2)T:(X,∥⋅∥X)→(R(T),∥⋅∥Y) is an isomorphism.
Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be normed spaces and let T:X→Y be linear. The graph of T is
graph(T)={(x,Tx):x∈X}⊂X×Y.
We say that graph(T) is closed if, for every sequence (xm,Txm)⊂graph(T) that converges to some (x,y), we have y=Tx. Here (xm,Txm) converges to (x,y) when xm→x in X and Txm→y in Y.
Theorem 15.11 (Closed graph theorem). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be Banach spaces and let T:X→Y be linear. Then T is continuous if and only if graph(T)⊂X×Y is closed.
Remark 15.12.(i) In Theorem 15.11, the implication from left to right follows from the continuity of T: if (xm,Txm) converges to (x,y), then y=limmTxm=Tx. (ii) The converse is false without linearity: the function f:R→R whose graph is drawn in Figure 15.1 is discontinuous, and graph(f) is closed.
Figure 15.1. The drawing of Remark 15.12 (ii): the graph of f, made of a branch over the negative half-line, the dot at the origin and a branch over the positive half-line.
Proof (complete). The implication from left to right is Remark 15.12 (i). Conversely, suppose that graph(T) is closed, and let
∥x∥0:=∥x∥X+∥Tx∥Yfor every x∈X.
The function ∥⋅∥0 is a norm on X. (i) If ∥x∥0=0, then ∥x∥X=0, hence x=0. (ii) For every α∈R and every x∈X,
∥αx∥0=∥αx∥X+∥T(αx)∥Y=∣α∣(∥x∥X+∥Tx∥Y).
(iii) The function ∥⋅∥0 satisfies the triangle inequality. Moreover, (X,∥⋅∥0) is a Banach space. Let (xm)⊂X be Cauchy with respect to ∥⋅∥0: for every ε>0 there exists N such that
∥xm−xn∥0=∥xm−xn∥X+∥Txm−Txn∥Y≤εfor all m,n≥N.
Hence (xm) is Cauchy with respect to ∥⋅∥X and (Txm) is Cauchy with respect to ∥⋅∥Y. Since X and Y are Banach spaces, there exist points x of X and y of Y such that xm→x in X and Txm→y in Y. Since graph(T) is closed, y=Tx. Therefore
where both terms on the right tend to 0. Hence ∥xm−x∥0→0, that is, xm→x with respect to ∥⋅∥0.
Now (X,∥⋅∥X) and (X,∥⋅∥0) are both Banach spaces, and ∥x∥X≤∥x∥X+∥Tx∥Y=∥x∥0 for every x∈X. By Corollary 15.8, with ∥⋅∥0 in place of ∥⋅∥a and ∥⋅∥X in place of ∥⋅∥b, there exists C>0 such that ∥x∥0≤C∥x∥X for every x∈X. Since ∥x∥0=∥x∥X+∥Tx∥Y≥∥Tx∥Y, it follows that ∥Tx∥Y≤C∥x∥X for every x∈X, that is, T is bounded, and therefore continuous. ■
Remark 15.13.(i) Let (X,∥⋅∥X) be a normed space. Then X∗=L(X,R) is a Banach space, whose norm is written ∥⋅∥X∗. (ii) Let V⊂X be a subspace. Then (V,∥⋅∥X) is a normed space, and V∗=L(V,R) is a Banach space, whose norm is written ∥⋅∥V∗. (iii) If T∈X∗, it is possible to restrict T to V: the restriction T∣V:V→R belongs to V∗.
The question raised by Remark 15.13 (iii) is the converse. If L∈V∗, is it possible to extend L to some T∈X∗? The answer is yes.
Theorem 15.14 (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.
The proof of Theorem 15.14 uses Zorn’s lemma, which is equivalent to the axiom of choice. Let S be a set. An order on S is a relation ≤ on S such that, for all x,y,z∈S,
x≤x,x≤yandy≤x⟹x=y,x≤yandy≤z⟹x≤z.
Lemma 15.15 (Zorn). Let S be a set endowed with an order ≤. Suppose that, for every family {xα:α∈A}⊂S such that
xα≤xβ or xβ≤xαfor all α,β∈A,
there exists x∈S such that xα≤x for every α∈A, as in Figure 15.2. Then there exists xˉ∈S such that
xˉ≤y⟹xˉ=yfor every y∈S.
Figure 15.2. The drawing of Lemma 15.15: elements of S, some joined by segments, a family {xα} inside the closed curve, as in the hypothesis, and at its top the element x with xα≤x for every α∈A.
S=⎩⎨⎧R:DR→R:DR⊂X is a subspace,V⊂DR,R∈L(DR,R),R∣V=L,∥R∥DR∗=∥L∥V∗⎭⎬⎫.
Figure 15.3. The drawing of the set S in the proof of Theorem 15.14: in X, a subspace DR, drawn as a plane, that contains the subspace V, drawn as a line.
For two elements R1 and R2 of S we define R1≤R2 if DR1⊂DR2 and R2∣DR1=R1. By Lemma 15.15, S has a maximal element T:DT→R. That is, T≤R implies T=R for every R∈S. The goal is DT=X. Suppose, by contradiction, that DT=X, and take x0∈X∖DT, as in Figure 15.4.
Figure 15.4. The drawing of the step by contradiction in the proof of Theorem 15.14: the subspace DT, drawn as a plane, and a point x0 outside it, joined to a point of DT by a segment.
The idea is to find T~:Span(DT,x0)→R, where
Span(DT,x0)={v+tx0:v∈DT,t∈R},
such that T~∣DT=T and T~∈S. Then T≤T~ and T=T~, against the maximality of T.■