Definition 13.1 (Separable metric space). A metric space (X,d) is said to be separable if there exists D⊂X such that D is dense and countable, as in Definition 12.8 and Definition 1.3.
Example 13.2. The set Q⊂R is dense by Example 12.10, and it is countable by Example 1.4. Hence R, with the distance d(x,y)=∣x−y∣, is separable.
Remark 13.3. Let (X,d) be a metric space and let D⊂X. Then D is dense if and only if U∩D=∅ for every nonempty open set U⊂X.
Proof (sketch). Only the implication from left to right is proved. Let D be dense, let U⊂X be open with U=∅, and take x∈U. Since U is open, there exists r>0 such that Br(x)⊂U, by Definition 1.19. Since D is dense, by Definition 12.8 there exists y∈Br(x)∩D, and Br(x)∩D⊂U∩D. Hence U∩D=∅.■
Theorem 13.4. Let −∞<a<b<+∞, let N be a positive integer and let U⊂RN be open, with the Lebesgue measure as in Example 10.8 (i). (1) The space (C([a,b]),∥⋅∥∞) of Example 10.5 (iii) is separable. The idea is the Stone–Weierstrass theorem in Remark 12.15, by which the polynomials are dense in C([a,b]): the polynomials a0+a1x+⋯+akxk with ai∈Q form a dense and countable subset of C([a,b]). (2) For every p∈[1,+∞), the space (Lp(U),∥⋅∥p) is separable. (3) The space (L∞(U),∥⋅∥∞) is not separable.
Proof (sketch). Only (3) is proved, in the case N=1 and U=R, that is, for L∞(R). For α>0 let
Figure 13.1. The drawing of the proof of Theorem 13.4 (3), for the α and β of the sliders. On the left, the graphs of fα and of fβ, dashed, which differ on the hatched set; on the right, the balls Uα and Uβ, drawn as discs, whose centres are at distance ∥fα−fβ∥∞=1.
Hence Uα∩Uβ=∅ for all α=β, and F={Uα:α>0} is an uncountable family of pairwise disjoint nonempty open balls.
Suppose, by contradiction, that there exists D={um:m∈N}⊂L∞(R) dense and countable. By Remark 13.3, D∩Uα=∅ for every α>0. That is, for every α>0 there exists mα∈N such that umα∈Uα, so that there is a function
ϕ:(0,+∞)→N,α↦mα.
The function ϕ is injective. Let α=β and suppose that mα=mβ. Then umα=umβ belongs to Uα and to Uβ, so that Uα∩Uβ=∅, which is impossible. Hence mα=mβ.
Therefore #(0,+∞)≤#N by Definition 1.1, a contradiction. ■
Definition 13.5 (Linear operator, bounded operator). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be normed spaces.
(i) A linear operator is a map T:X→Y such that
T(αx)T(x+y)=αTx=Tx+Tyfor all α∈R and x∈X,for all x,y∈X.
(ii) A linear operator T is said to be bounded if there exists C>0 such that ∥Tx∥Y≤C∥x∥X for every x∈X.
Remark 13.6. Let N and M be positive integers. If T:RN→RM is linear, then T is represented by a matrix, and T is continuous.
Example 13.7. Let −∞<a<b<+∞. The linear operator T from (C1([a,b]),∥⋅∥∞) to (C([a,b]),∥⋅∥∞) given by Tf=f′ is not continuous.
For Example 13.7, recall that, by Lemma 1.21, a map T:X→Y between normed spaces is continuous if and only if Txm→Tx0 whenever xm→x0 in X. We find (fm)⊂C1([a,b]) such that fm→0 with respect to ∥⋅∥∞, while the derivatives Tfm=fm′ do not converge to the derivative of the function 0, which is 0, in ∥⋅∥∞.
Suppose that a<0<b, as in Figure 13.2, and let m be an integer with 1/m<b. Let fm∈C1([a,b]) vanish outside (0,1/m), with 0≤fm≤1/m and fm(1/(2m))=1/m.
(i) The norms satisfy ∥fm∥∞≤1/m, which tends to 0, so that fm→0 in ∥⋅∥∞. (ii) By the mean value theorem, there exists xm∈(0,1/(2m)) such that fm′(xm)=2. Hence ∥fm′∥∞≥2 for every m, so that ∥fm′∥∞ does not tend to 0, and T is not continuous.
Figure 13.2. The drawing of Example 13.7, an instance on [a,b]=[−1/2,3/2] for the m of the slider: the graph of fm, the dashed level 1/m and the points 1/(2m) and 1/m. The tangent at xm is parallel to the dotted segment from (0,0) to (1/(2m),1/m), of slope 2.
A map T from a normed space X to a normed space Y is Lipschitz if there exists L>0 such that ∥Tx−Ty∥Y≤L∥x−y∥X for all x,y∈X. It is continuous at0 if for every ε>0 there exists δ>0 such that
∥x∥X<δ⟹∥Tx−T0∥Y<ε.
Theorem 13.8 (Characterization of continuity). Let T:(X,∥⋅∥X)→(Y,∥⋅∥Y) be linear. The following are equivalent. (1)T is bounded. (2)T is Lipschitz. (3)T is continuous. (4)T is continuous at 0.
Proof (complete). We prove that (1) implies (2), that (2) implies (3), that (3) implies (4) and that (4) implies (1). Suppose (1), with the constant C of Definition 13.5 (ii). Then, for all x,y∈X,
∥Tx−Ty∥Y=∥T(x−y)∥Y≤C∥x−y∥X,
where the equality is the linearity of T and the inequality is (1) applied to x−y. Hence (2) holds with L=C.
Suppose (2), and let (xm)⊂X be such that xm→x0 in X. Then, by (2),
∥Txm−Tx0∥Y≤L∥xm−x0∥X⟶0,
so that Txm→Tx0, and T is continuous by Lemma 1.21. Next, (3) implies (4) by Lemma 1.21 (ii) at the point x0=0.
Suppose (4). Since T0=0 by linearity, (4) means that for every ε>0 there exists δ>0 such that
∥x∥X≤δ⟹∥Tx∥Y≤ε.
In detail, the strict inequalities of the definition of continuity at 0 give this with δ/2 in place of δ. Take ε=1. There is δ>0 such that ∥x∥X≤δ implies ∥Tx∥Y≤1. If x=0, then ∥Tx∥Y=∥T0∥Y=∥0∥Y=0=∥x∥X. If x=0, the vector δx/∥x∥X has norm δ, hence, by (4),
where the equality in the first line is the linearity of T. The last inequality holds for every x∈X, the case x=0 included, that is, T is bounded. ■
Definition 13.9 (Dual space). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be normed spaces. (i) The set of the linear and continuous operators T:X→Y is denoted L(X,Y). We observe that L(X,Y) is a vector space, with (T+S)(x)=Tx+Sx and (αT)(x)=αTx for α∈R. (ii) We write L(X):=L(X,X). (iii) The set X∗:=L(X,R), where R has the norm ∣⋅∣, is the set of the linear continuous operators T:X→R, the functionals onX. It is also denoted X′, and it is called the dual space of X.
(iv) On L(X,Y) we define
∥T∥L(X,Y):=∥x∥X≤1sup∥Tx∥Y,
also written ∥T∥L or ∥T∥. Notice that (L(X,Y),∥⋅∥L(X,Y)) is a normed space, and that, if T∈X∗, then
∥T∥L=∥x∥X≤1sup∣Tx∣.
Proposition 13.10 (Characterization of ∥⋅∥L). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be normed spaces with X={0}, and let T∈L(X,Y). Then
∥T∥L=∥x∥X=1sup∥Tx∥Y=x∈Xx=0sup∥x∥X∥Tx∥Y=inf{M>0:∥Tx∥Y≤M∥x∥Xfor every x∈X}.
In words, ∥T∥L is the least constant that can be put in the definition of a bounded operator, Definition 13.5 (ii). In particular, ∥Tx∥Y≤∥T∥L∥x∥X for every x∈X.
Proof (complete). By the definition of ∥T∥L,
∥T∥L=∥x∥X≤1sup∥Tx∥Y≥∥x∥X=1sup∥Tx∥Y≥x∈Xx=0sup∥x∥X∥Tx∥Y=:M0≥inf{M>0:∥Tx∥Y≤M∥x∥Xfor every x∈X}=:M1.
The first inequality restricts the set where the supremum is taken. For the second, if x=0, then x/∥x∥X has norm 1 and, by linearity,
∥x∥X∥Tx∥Y=T(∥x∥Xx)Y≤∥z∥X=1sup∥Tz∥Y.
For the third, the definition of M0 gives ∥Tx∥Y/∥x∥X≤M0 for every x=0. Hence ∥Tx∥Y≤M0∥x∥X for every x∈X, so M0 is a competitor in the computation of that infimum, and M1≤M0. In detail, if M0=0, then every M>0 is a competitor and M1=0.
The last step is M1≥∥T∥L. Let M>0 be any competitor in the infimum, that is, ∥Tx∥Y≤M∥x∥X for every x∈X. Then, for every x with ∥x∥X≤1,
∥Tx∥Y≤M∥x∥X≤M⟹∥T∥L=∥x∥X≤1sup∥Tx∥Y≤M.
Taking the infimum over M gives ∥T∥L≤M1.■
Theorem 13.11. Let (X,∥⋅∥X) be a normed space and (Y,∥⋅∥Y) a Banach space, as in Definition 10.4 (ii). Then (L(X,Y),∥⋅∥L(X,Y)) is a Banach space. In particular, X∗ is a Banach space.
Definition 13.12 (Kernel, range). Let T:(X,∥⋅∥X)→(Y,∥⋅∥Y) be linear. (i) The kernel of T is Ker(T)={x∈X:Tx=0}. (ii) The range, or image, of T is R(T)={Tx:x∈X}.
Remark 13.13. Let T∈L(X,Y). Then Ker(T) is a closed subspace of X, but R(T) is not always closed in Y.
Remark 13.14. Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be normed spaces and let T:X→Y be linear. (i)T is injective if and only if Ker(T)={0}. (ii) If T is bijective, then its inverse T−1:Y→X exists and is linear. (iii) If T∈L(X,Y) is bijective, it is not always true that T−1 is continuous.
Definition 13.15 (Isomorphism, isometry, isometric isomorphism). Let (X,∥⋅∥X) and (Y,∥⋅∥Y) be normed spaces. (i) A map T:X→Y is an isomorphism of normed spaces if T∈L(X,Y),T is bijective and T−1∈L(Y,X). (ii) A map T∈L(X,Y) is an isometry if ∥Tx∥Y=∥x∥X for every x∈X. Notice that an isometry is not required to be surjective. (iii) A map T:X→Y is an isometric isomorphism if T is an isomorphism and an isometry. Notice that in this case also T−1 is an isometric isomorphism.
Remark 13.16. Let T∈L(X,Y) be an isometry. (i)T is injective. (ii)T is an isometric isomorphism if and only if T is surjective.
Definition 13.17 (Embedding). Let (Y,∥⋅∥Y) be a normed space and let X⊂Y be a subspace such that (X,∥⋅∥X) is a normed space. Consider
J:(X,∥⋅∥X)→(Y,∥⋅∥Y),x↦x,
called the identity, the immersion or the injection of X into Y. The map J is called an embedding if it is continuous.
By Definition 13.17 and Theorem 13.8, J is an embedding if and only if J is bounded. That is, there exists C>0 such that
that is,∥Jx∥Y≤C∥x∥Xfor every x∈X,∥x∥Y≤C∥x∥Xfor every x∈X.
Example 13.18. Let (Ω,M,μ) be a measure space with μ(Ω)<+∞, and let 1≤p≤q≤+∞. By the Hölder inequality, Lq(Ω)⊂Lp(Ω), as in Corollary 10.13, and we can consider
J:Lq(Ω)→Lp(Ω),f↦f.
The Hölder inequality, Proposition 10.12, gives a constant C>0 such that ∥f∥p≤C∥f∥q for every f∈Lq(Ω). That is, J is continuous: Lq(Ω) embeds into Lp(Ω).
A metric space (X,d) is complete if every Cauchy sequence converges in X, where Cauchy sequences are defined as in Definition 10.4 (i) with d(xn,xm) in place of ∥xn−xm∥.
Theorem 13.19 (Baire’s lemma). Let (X,d) be a complete metric space. (1) If (Um)m≥1 is a sequence of open dense sets in X, then their intersection is dense in X. (2) If (Cm)m≥1 is a sequence of closed sets in X such that int(Cm)=∅ for every m, then their union has empty interior.
In Theorem 13.19, (1) and (2) are equivalent: with Cm=X∖Um for every m,
m=1⋃∞Cm=X∖m=1⋂∞Um.
Indeed, by Remark 13.3 a set D⊂X is dense if and only if int(X∖D)=∅, since both mean that every nonempty open set meets D. Moreover, Cm is closed if and only if Um is open.
Corollary 13.20. Let (X,d) be a nonempty complete metric space and let (Cm)m≥1 be a sequence of closed sets whose union is X. Then there exists m0 such that int(Cm0)=∅.
Proof (complete). We have
int(m=1⋃∞Cm)=int(X)=∅.
Hence, by Theorem 13.19 (2), not every element Cm of the sequence has empty interior. ■
Proof (complete). We prove Theorem 13.19 (1); then (2) holds by the equivalence of (1) and (2). Let (Um)m≥1 be open and dense, and take O⊂X open with O=∅. By Remark 13.3, the goal is
m=1⋂∞Um∩O=∅.
(i) The set U1∩O is open, and it is nonempty because U1 is dense, by Remark 13.3. Hence there exist x1∈X and r1∈(0,1/2) such that the closed ball Br1(x1)={y∈X:d(x1,y)≤r1} is contained in O∩U1.
(ii) The set U2∩Br1(x1) is open, and it is nonempty because U2 is dense. Hence, as in Figure 13.3, there exist x2∈X and r2∈(0,1/4) such that
Br2(x2)⊂Br1(x1)∩U2⊂O∩U1∩U2.
(iii) Iterating, we get points xm∈X and radii rm∈(0,1/2m), and finding this sequence relies on the axiom of choice. For every m≥2,
Figure 13.3. The drawing of the proof of Theorem 13.19 (1): the open sets O,U1 and U2, the closed ball Br1(x1)⊂O∩U1 and the closed ball Br2(x2)⊂Br1(x1)∩U2. The button shows the steps; shaded, the open sets O∩U1 and Br1(x1)∩U2 in which the two balls are chosen.
The sequence (xm) is Cauchy. For all n,m≥N we have xn,xm∈BrN(xN), hence
d(xn,xm)≤2rN<2N2.
Since 2/2N→0 as N→∞, the sequence (xm) is Cauchy. By completeness, there exists x∞∈X such that xm→x∞. We claim that x∞ belongs to O and to every Um, so that the set of the goal is nonempty. For every N and every m≥N we have xm∈BrN(xN), as in Figure 13.4. Hence
x∞∈BrN(xN)⊂O∩k=1⋂NUkfor every N.
Figure 13.4. The drawing of the claim of the proof, an instance in the plane for the N of the slider: the closed ball BrN(xN), the points xm with m≥N, all inside it, and their limit x∞. The balls of the other indices are faint.