Arzelà–Ascoli Theorem

When does a family of continuous functions hide a uniformly convergent subsequence? Uniform boundedness plus equicontinuity lets the diagonal argument extract one. Contrast with xnx^n, where it fails.

F⊂C(K) relatively compactF \subset C(K)\ \text{relatively compact}
  ⟺  uniformly bounded+equicontinuous\iff \text{uniformly bounded} + \text{equicontinuous}
∥fak−f∗∥∞=1k+2→0\|f_{a_k} - f^*\|_\infty = \frac{1}{k+2} \to 0

Part of these learning paths

Examples & Exercises

Let KK be a compact metric space and F⊂C(K)F \subset C(K) relatively compact in the sup norm. Prove FF is equicontinuous using a finite ε/3\varepsilon/3-net (hint: each net member is uniformly continuous on the compact KK).

∣f(x)−f(y)∣≤∥f−gj∥∞+∣gj(x)−gj(y)∣+∥gj−f∥∞<ε\lvert f(x) - f(y) \rvert \le \lVert f - g_j \rVert_\infty + \lvert g_j(x) - g_j(y) \rvert + \lVert g_j - f \rVert_\infty < \varepsilon
6 members: sup errors 13→14→15→16→17→18=0.125\frac13 \to \frac14 \to \frac15 \to \frac16 \to \frac17 \to \frac18 = 0.125 — a 32\frac32-Lipschitz sequence with ∣ak∣≤34|a_k| \le \frac34 converging uniformly.
Show proof
  1. Relatively compact implies totally bounded: for any ε>0\varepsilon > 0 there is a finite ε/3\varepsilon/3-net {g1,…,gm}\{g_1, \dots, g_m\} of the closure F‾\overline{F}. That is, every f∈Ff \in F has some gjg_j with ∥f−gj∥∞<ε/3\lVert f - g_j \rVert_\infty < \varepsilon/3. The gjg_j need not lie in FF — continuity is all we need.
  2. Each gjg_j is continuous on the compact metric space KK, hence uniformly continuous (Heine–Cantor): there is δj>0\delta_j > 0 with ∣gj(x)−gj(y)∣<ε/3\lvert g_j(x) - g_j(y) \rvert < \varepsilon/3 whenever d(x,y)<δjd(x, y) < \delta_j. Set δ=min⁡jδj>0\delta = \min_j \delta_j > 0 — this δ\delta does not depend on ff.
  3. Now take f∈Ff \in F and x,yx, y with d(x,y)<δd(x, y) < \delta. Picking gjg_j with ∥f−gj∥∞<ε/3\lVert f - g_j \rVert_\infty < \varepsilon/3, the triangle inequality gives ∣f(x)−f(y)∣≤∣f(x)−gj(x)∣+∣gj(x)−gj(y)∣+∣gj(y)−f(y)∣<ε/3+ε/3+ε/3=ε\lvert f(x) - f(y) \rvert \le \lvert f(x) - g_j(x) \rvert + \lvert g_j(x) - g_j(y) \rvert + \lvert g_j(y) - f(y) \rvert < \varepsilon/3 + \varepsilon/3 + \varepsilon/3 = \varepsilon. Since δ\delta works for every f∈Ff \in F, the family FF is equicontinuous.
  4. In addition, relative compactness gives uniform boundedness (a bounded set has bounded closure): M=sup⁡f∈F∥f∥∞<∞M = \sup_{f \in F} \lVert f \rVert_\infty < \infty. Conversely, on compact KK, pointwise bounded + equicontinuous implies uniformly bounded: cover KK by a finite δ\delta-net {y1,…,yr}\{y_1, \dots, y_r\} for the equicontinuity modulus at ε=1\varepsilon = 1; each xx is δ\delta-close to some yiy_i, so ∣f(x)∣≤∣f(yi)∣+1≤max⁡isup⁡f∣f(yi)∣+1\lvert f(x) \rvert \le \lvert f(y_i) \rvert + 1 \le \max_i \sup_f \lvert f(y_i) \rvert + 1.

Let KK be a compact metric space, {xk}\{x_k\} a countable dense subset of KK, and (fn)(f_n) a sequence in C(K)C(K) with sup⁡n∥fn∥∞≤M\sup_n \lVert f_n \rVert_\infty \le M. Construct, by the diagonal argument, a subsequence converging at every xkx_k.

n(1)⊃n(2)⊃⋯(nested index sequences),mj=nj(j) (diagonal)n^{(1)} \supset n^{(2)} \supset \cdots \quad \text{(nested index sequences)}, \qquad m_j = n^{(j)}_j \ \text{(diagonal)}
6 members: amplitudes ak=12+(−1)kk+2a_k = \frac12 + \frac{(-1)^k}{k+2} oscillate toward 12\frac12 through 16,34,310,23,514,58\frac16, \frac34, \frac{3}{10}, \frac23, \frac{5}{14}, \frac58 — the way diagonal extraction fishes out such an oscillating sequence.
Show proof
  1. A compact metric space is separable (take the union of finite 1/r1/r-nets). Now look at (fn)(f_n). For each kk, the scalar sequence (fn(xk))n(f_n(x_k))_n is trapped in [−M,M][-M, M], so by Bolzano–Weierstrass it has a convergent subsequence.
  2. Build nested index sequences: choose n1(1)<n2(1)<⋯n^{(1)}_1 < n^{(1)}_2 < \cdots so that fnj(1)(x1)f_{n^{(1)}_j}(x_1) converges, and inductively choose n(k)n^{(k)} as a subsequence of n(k−1)n^{(k-1)} with fnj(k)(xk)f_{n^{(k)}_j}(x_k) convergent. Fixing kk, every later stage refines n(k)n^{(k)}, so convergence at x1,…,xkx_1, \dots, x_k is preserved.
  3. Set the diagonal mj=nj(j)m_j = n^{(j)}_j. Fixing kk, the tail (mj)j≥k(m_j)_{j \ge k} is a subsequence of (nj(k))(n^{(k)}_j) (nesting makes mj=nj(j)m_j = n^{(j)}_j a strictly increasing pick from the terms of n(k)n^{(k)}). Hence fmj(xk)f_{m_j}(x_k) converges. Since kk was arbitrary, (fmj)(f_{m_j}) converges at every xkx_k — pointwise convergence on the dense set.
  4. Note: only pointwise boundedness was used here. Upgrading this pointwise convergence to uniform convergence on all of KK is the job of equicontinuity, carried out by the ε/3\varepsilon/3 argument in the next problem.

Let (fn)⊂C(K)(f_n) \subset C(K) be equicontinuous and pointwise bounded (KK compact metric), converging pointwise on a countable dense set {xk}\{x_k\}. Prove by the ε/3\varepsilon/3 argument that the convergence is uniform on all of KK, and deduce as a corollary that (xn)(x^n) has no uniformly convergent subsequence on [0,1][0,1].

∣fmj(x)−fmℓ(x)∣≤ε3+ε3+ε3=ε(uniformly Cauchy)\lvert f_{m_j}(x) - f_{m_\ell}(x) \rvert \le \tfrac{\varepsilon}{3} + \tfrac{\varepsilon}{3} + \tfrac{\varepsilon}{3} = \varepsilon \quad \text{(uniformly Cauchy)}
xn→{0x∈[0,1)1x=1(discontinuous pointwise limit)x^n \to \begin{cases} 0 & x \in [0, 1) \\ 1 & x = 1 \end{cases} \quad \text{(discontinuous pointwise limit)}
6 members: ∥fak−f∗∥∞=1k+2→0\lVert f_{a_k} - f^* \rVert_\infty = \frac{1}{k+2} \to 0 — equicontinuity spreading convergence from the dense set to the whole interval.
Show proof
  1. Given ε>0\varepsilon > 0, equicontinuity yields δ>0\delta > 0: d(x,y)<δd(x, y) < \delta implies ∣fn(x)−fn(y)∣<ε/3\lvert f_n(x) - f_n(y) \rvert < \varepsilon/3 for every nn. Since KK is compact, finitely many points xk1,…,xkrx_{k_1}, \dots, x_{k_r} from the dense set {xk}\{x_k\} form a δ\delta-net of KK.
  2. At the finitely many net points, pointwise convergence becomes "uniform": choose NN so large that j,ℓ≥Nj, \ell \ge N gives ∣fmj(xki)−fmℓ(xki)∣<ε/3\lvert f_{m_j}(x_{k_i}) - f_{m_\ell}(x_{k_i}) \rvert < \varepsilon/3 for every ii (a finite intersection of tails).
  3. For arbitrary x∈Kx \in K, pick xkix_{k_i} with d(x,xki)<δd(x, x_{k_i}) < \delta. By the triangle inequality, ∣fmj(x)−fmℓ(x)∣≤∣fmj(x)−fmj(xki)∣+∣fmj(xki)−fmℓ(xki)∣+∣fmℓ(xki)−fmℓ(x)∣<ε/3+ε/3+ε/3=ε\lvert f_{m_j}(x) - f_{m_\ell}(x) \rvert \le \lvert f_{m_j}(x) - f_{m_j}(x_{k_i}) \rvert + \lvert f_{m_j}(x_{k_i}) - f_{m_\ell}(x_{k_i}) \rvert + \lvert f_{m_\ell}(x_{k_i}) - f_{m_\ell}(x) \rvert < \varepsilon/3 + \varepsilon/3 + \varepsilon/3 = \varepsilon. Hence (fmj)(f_{m_j}) is uniformly Cauchy, so it converges uniformly by completeness of C(K)C(K).
  4. This completes (⟸\Longleftarrow): pointwise bounded + equicontinuous implies uniformly bounded (as noted in the first problem), and every sequence has a uniformly convergent subsequence, so by sequential compactness in metric spaces the set is relatively compact.
  5. Corollary: (xn)(x^n) has no uniformly convergent subsequence on [0,1][0,1]. Its pointwise limit is discontinuous — 0 on [0,1)[0,1), 1 at x=1x = 1. If some subsequence converged uniformly, its limit would be continuous (a uniform limit of continuous functions) and would have to equal the pointwise limit — contradiction. So {xn}\{x^n\} is not equicontinuous at x=1x = 1: as the scene's contrast chart shows, pairwise sup distances between distinct members stay at least 0.47 apart.

5. Approximation