Throughout, U={x∈X:∥x∥≤1} is the closed unit ball of X, and every F∈BM∗ satisfies
∣F(x)∣≤∥F∥∗∥x∥≤M∥x∥(x∈X).
The proofs of (i) and (ii) use no choice principle; the proof of (iii) uses the axiom of dependent choice, a standing hypothesis by Radon Measures on Locally Compact Metric Spaces: Standing Notation (iv), both directly and through the compactness lemma it cites.
Proof of (i): such a sequence exists. The proof of (i) needs a sequence (xk) in U whose range is dense in U; we construct one. By Separable Metric Space and Separable Real Normed Space (ii) and Separable Metric Space and Separable Real Normed Space (i) there is a countable subset D⊆X that is dense in (X,dX), dX the norm metric. It is nonempty, since the open ball B(0,1) meets it, so by Countable Set (i) there is a surjective function N→D, k↦yk, whose range D is dense in (X,dX). Let
xk=max{1,∥yk∥}1yk(k∈N).
Then ∥xk∥=∥yk∥/max{1,∥yk∥}≤1, so xk∈U.
The range of (xk) is dense in U. Let x∈U and δ>0; we find k with ∥xk−x∥<δ. By density choose k with ∥yk−x∥<δ/2. If ∥yk∥≤1, then xk=yk. Otherwise xk=yk/∥yk∥ and
∥yk−xk∥=(1−∥yk∥1)∥yk∥=∥yk∥−1≤∥yk∥−∥x∥≤∥yk−x∥<2δ.
In both cases ∥xk−x∥≤∥xk−yk∥+∥yk−x∥<δ. So (xk) has the properties the statement requires of the sequence in (ii); the proof of (i) uses this sequence, and the proof of (ii) the one of the statement.
An approximation estimate. Let (xk) be a sequence in U whose range is dense in U. For F,G∈BM∗, x∈U and k∈N, linearity gives F(x)−G(x)=(F(xk)−G(xk))+F(x−xk)−G(x−xk), and hence
∣F(x)−G(x)∣≤∣F(xk)−G(xk)∣+2M∥x−xk∥.(1)
Moreover, for x∈X with x=0 the vector x/∥x∥ lies in U and F(x)=∥x∥F(x/∥x∥); so a statement about the values at the points of U that is preserved under multiplication by a constant passes to all of X.
Proof of (i): the diagonal subsequence. Let (xk) be the sequence constructed above, let (Fm) be a sequence in BM∗, and put am,k=Fm(xk). Since ∣am,k∣≤M∥xk∥≤M, every sequence (am,k)m∈N is bounded, and The Diagonal Subsequence Lemma for Bounded Real Arrays (i) gives a strictly increasing sequence (mj)j∈N in N such that (Fmj(xk))j∈N converges for every k. Write Gj=Fmj; then (Gj) is a subsequence of (Fm), and each (Gj(xk))j is a Cauchy sequence by Completeness of the Real Numbers for Sequences: Monotone Convergence, the Bolzano-Weierstrass Theorem and Cauchy Sequences (v).
Convergence at every point. Let x∈U and ε>0. First choose k with ∥x−xk∥<ε/(4M+1), so that 2M∥x−xk∥<ε/2; then, with this k fixed, choose N∈N with ∣Gj(xk)−Gl(xk)∣<ε/2 for all j,l≥N. By (1), ∣Gj(x)−Gl(x)∣<ε for all j,l≥N. Thus (Gj(x))j is a Cauchy sequence, and so convergent by Completeness of the Real Numbers for Sequences: Monotone Convergence, the Bolzano-Weierstrass Theorem and Cauchy Sequences (v), for every x∈U; for x∈/U the sequence Gj(x)=∥x∥Gj(x/∥x∥) converges as well, by Limits of Sequences of Real Numbers: Uniqueness, Boundedness, Constants, Tails, Arithmetic, Quotients, Absolute Values, Finite Sums, Order, Squeezing, Domination and Subsequences (vi).
The limit functional. Define F:X→R by F(x)=limj→∞Gj(x). Passing to the limit in Gj(u+v)=Gj(u)+Gj(v) and Gj(λu)=λGj(u) shows, by Limits of Sequences of Real Numbers: Uniqueness, Boundedness, Constants, Tails, Arithmetic, Quotients, Absolute Values, Finite Sums, Order, Squeezing, Domination and Subsequences (vi) and Limits of Sequences of Real Numbers: Uniqueness, Boundedness, Constants, Tails, Arithmetic, Quotients, Absolute Values, Finite Sums, Order, Squeezing, Domination and Subsequences (ii), that F is a linear functional on X; passing to the limit in ∣Gj(x)∣≤M∥x∥ gives ∣F(x)∣≤M∥x∥ by Limits of Sequences of Real Numbers: Uniqueness, Boundedness, Constants, Tails, Arithmetic, Quotients, Absolute Values, Finite Sums, Order, Squeezing, Domination and Subsequences (viii) and Limits of Sequences of Real Numbers: Uniqueness, Boundedness, Constants, Tails, Arithmetic, Quotients, Absolute Values, Finite Sums, Order, Squeezing, Domination and Subsequences (x). So F∈X∗ and, ∥F∥∗ being the infimum of the bounds of F by The Dual Norm (i), ∥F∥∗≤M. Hence F∈BM∗, and Gj(x)→F(x) for every x∈X, which is Gj⇀∗F by Weak-Star Convergence in the Dual of a Real Normed Space (i).
Proof of (ii): d is a metric. From here on, (xk) and d are as in the statement. Let F,G,H∈BM∗. For every k, ∣F(xk)−G(xk)∣≤2M∥xk∥≤2M and 2−k≤1/2, so the numbers 2−k∣F(xk)−G(xk)∣ form a nonempty subset of [0,M]; its supremum d(F,G) exists and lies in [0,M]. Clearly d(F,G)=d(G,F) and d(F,F)=0. For every k,
2−k∣F(xk)−H(xk)∣≤2−k∣F(xk)−G(xk)∣+2−k∣G(xk)−H(xk)∣≤d(F,G)+d(G,H),
and taking the supremum over k gives the triangle inequality. Finally suppose d(F,G)=0, so that F(xk)=G(xk) for every k. For x∈U and ε>0 choose k with 2M∥x−xk∥<ε; then (1) gives ∣F(x)−G(x)∣<ε. Hence F(x)=G(x) for x∈U, and by homogeneity for every x∈X, so F=G. By Metric Space (i), d is a metric on BM∗.
Convergence for d implies weak-star convergence. Let (Gm) be a sequence in BM∗ and G∈BM∗, and suppose Gm→G in (BM∗,d). For every k, ∣Gm(xk)−G(xk)∣≤2kd(Gm,G), so Gm(xk)→G(xk). Let x∈U and ε>0. First choose k with 2M∥x−xk∥<ε/2; then, with k fixed, choose N with ∣Gm(xk)−G(xk)∣<ε/2 for every m≥N. By (1), ∣Gm(x)−G(x)∣<ε for every m≥N. So Gm(x)→G(x) for every x∈U, hence, by homogeneity and Limits of Sequences of Real Numbers: Uniqueness, Boundedness, Constants, Tails, Arithmetic, Quotients, Absolute Values, Finite Sums, Order, Squeezing, Domination and Subsequences (vi), for every x∈X: Gm⇀∗G.
Weak-star convergence implies convergence for d. Conversely, suppose Gm⇀∗G, and let ε>0. First choose K∈N with 2−K(2M+1)<ε/2, which is possible since (1/2)n→0 by Completeness of the Real Numbers for Sequences: Monotone Convergence, the Bolzano-Weierstrass Theorem and Cauchy Sequences (ii). Then, with K fixed, choose for each k≤K an index Nk with ∣Gm(xk)−G(xk)∣<ε/2 for every m≥Nk, and let N=max{N1,…,NK}. For m≥N and every k: if k≤K, then 2−k∣Gm(xk)−G(xk)∣<ε/2; if k>K, then 2−k∣Gm(xk)−G(xk)∣≤2−K⋅2M<ε/2. Hence d(Gm,G)≤ε/2<ε for every m≥N, that is, Gm→G in (BM∗,d) by Convergent Sequence in a Metric Space (i).
Proof of (iii): sequential compactness. Every sequence in BM∗ has, by (i), a subsequence converging weak-star to some G∈BM∗, and by (ii) this subsequence converges to G in (BM∗,d). So BM∗ is sequentially compact in (BM∗,d).
Completeness. Let (Gm) be a Cauchy sequence in (BM∗,d). By sequential compactness it has a subsequence (Gσ(j))j converging to some G∈BM∗. Given ε>0, choose N with d(Gm,Gl)<ε/2 for all m,l≥N, and then j≥N with d(Gσ(j),G)<ε/2; since σ is strictly increasing, σ(j)≥j≥N. For every m≥N,
d(Gm,G)≤d(Gm,Gσ(j))+d(Gσ(j),G)<ε.
So (Gm) converges to G, and (BM∗,d) is complete.
Total boundedness. Suppose that BM∗ is not totally bounded in (BM∗,d). Then there is ε>0 such that for every finite subset S⊆BM∗ some G∈BM∗ satisfies d(G,H)≥ε for every H∈S (for S=∅, any G, such as 0∈BM∗). Applying dependent choice to the relation that extends a finite sequence (G1,…,Gn) in BM∗ by a Gn+1∈BM∗ with d(Gn+1,Gl)≥ε for every l≤n, we obtain a sequence (Gm) in BM∗ with d(Gm,Gl)≥ε whenever l<m. No subsequence of (Gm) converges in (BM∗,d): if Gσ(j)→G, then d(Gσ(j),Gσ(i))≤d(Gσ(j),G)+d(G,Gσ(i))<ε for all large i<j, a contradiction. This contradicts sequential compactness, so BM∗ is totally bounded in (BM∗,d).
Compactness. By The Open Subsets of a Metric Space Form a Topology (i), Td is a topology on BM∗, so ∅=BM∗∖BM∗ lies in Td and BM∗ is closed in (BM∗,Td). Thus BM∗ is a closed, totally bounded subset of the complete metric space (BM∗,d), and A Closed Totally Bounded Subset of a Complete Metric Space is Compact (ii) shows that it is compact in (BM∗,Td).