Arzelà-Ascoli Theorem

Let’s conclude our contemplations on compactness with function spaces that will hopefully motivate future topological discussions on function spaces.

Theorem 1. \mathcal C([0, 1], \mathbb R) equipped with the supremum metric is not compact.

Proof of Theorem 1. The sequence \{ f_n : n \in \mathbb N \} defined by f_n = n has no convergent subsequence.

Theorem 2. For any n \in \mathbb N_0 and M > 0, define

\mathcal D_{n,M} := \{f \in \mathcal C^{(n)}([0, 1], \mathbb R) : \| f^{(n)} \|_\infty \leq M\}.

Then \mathcal D_{0,M} and \mathcal D_{1,M} are not compact, but \mathcal D_{0,M} \cap \mathcal D_{1,M} is.

Partial Proof of Theorem 2. The sequences \{ f_n \} and \{g_n\} defined by f_n(x) = M \cdot \sin(nx) and g_n = n have no convergent subsequence. We delay the proof of the third claim.

The third claim is in fact one of the many key applications of the Arzelà-Ascoli theorem, which is the content of this post. Fix \mathcal F \subseteq \mathcal C([0, 1], \mathbb R), equipped with the uniform norm.

Lemma 1. Suppose \mathcal F is compact. Then the following properties hold.

  • \mathcal F is closed in the following sense: for any \{f_n\} \subseteq \mathcal F, if f_n \to f in the uniform metric, then f \in \mathcal F.
  • \mathcal F is bounded in the following sense: there exists M > 0 such that f \in \mathcal D_{0,M}.
  • \mathcal F is equicontinuous in the following sense: for any \epsilon > 0, there exists \delta > 0 such that for any x,y \in [0, 1], |x-y| < \delta implies that for any f \in \mathcal F, |f(x) - f(y)| < \epsilon.

Proof. Compactness implies closedness and boundedness by metric space arguments discussed before. For any open U \subseteq [0, 1], define

\displaystyle \mathcal U(\epsilon, U) := \left\{f \in \mathcal F : \sup_{t \in U}f(t) - \inf_{t \in U} f(t) < \epsilon \right\}.

Fix x \in [0, 1] and \epsilon > 0. For each x \in [0, 1], let \mathcal U_x denote the collection of neighborhoods of x. Then \{U(\epsilon, V) : V \in \mathcal U_x\} forms an open cover for \mathcal F. To verify this, since each f \in \mathcal F is continuous at x, there exists \delta > 0 such that for y \in (x-\delta,x+\delta) =: V_\delta,

\displaystyle |f(x) - f(y)| < \frac{\epsilon}3 \quad \Rightarrow \quad \sup_{t \in V_\delta} f(t) - \inf_{t \in V_\delta} f(t) \leq \frac{ 2\epsilon }3<  \epsilon,

so that f \in U(\epsilon, V_\delta). Since \mathcal F is compact, we can extract a finite sub cover \{U(\epsilon, V_{\delta_i}) : i = 1,\dots,n\}. Selecting \delta := \min\{\delta_1,\dots,\delta_n\} yields equicontinuity near x.

Finally, we aim to prove uniform equicontinuity. Let k_1,k_2,k_3 > 0 be constants to be tuned. For each x \in [0, 1], use the pointwise equicontinuity of \mathcal F to extract a constant \delta_x > 0 such that

|x-y| < \delta_x \quad \Rightarrow \quad |f(x) - f(y)| < k_1 \cdot \epsilon,\quad f \in \mathcal F.

Since \{(x- k_2 \cdot \delta_x, x+k_2 \cdot \delta_x) : x \in [0, 1]\} yields an open cover for [0, 1], we can extract a finite sub-cover

\{(x_i -k_2 \cdot \delta_{x_i}, x_i +k_2 \cdot \delta_{x_i}) : i = 1,\dots, n\}.

Define \delta := k_3 \cdot \min \{\delta_{x_1},\dots,\delta_{x_n}\} and x, y \in [0, 1]. Suppose |x-y | < \delta. Find x_i such that |x - x_i| < k_2 \cdot \delta_{x_i}, so that |y - x_i| < (k_2 + k_3) \cdot \delta_{x_i}. Setting k_2 = k_3 = 1/2,

|f(x) - f(y)| \leq |f(x) - f(x_i)| + |f(x_i) - f(y)| < 2k_1 \cdot \epsilon.

Setting k_1 = 1/2 yields the desired result.

The simplest form of the Arzelà-Ascoli theorem claims that the converse holds.

Theorem 3 (Arzelà-Ascoli Theorem). \mathcal F is compact if and only if it is closed, bounded, and equicontinuous.

Proof of Theorem 3. We have proven (\Rightarrow) in Lemma 1, and turn our attention to (\Leftarrow).

Let \mathcal F be closed, bounded, and equicontinuous in the sense of Lemma 1. Since \mathcal F is a metric space, it suffices to prove that it is sequentially compact. Fix any sequence \{f_n\} of \mathcal F. We wish to extract a convergent subsequence \{f_{n_k}\}.

Let \mathbf x : \mathbb N \to \mathbb Q \cap [0, 1] be an enumeration of the rational numbers in [0, 1] (so that \mathbf x is bijective) and denote x_i := \mathbf x(i). Since \mathcal F_0 := \mathcal F is bounded, \{f(x_1) : f \in \mathcal F_0\} is bounded and hence contains a convergent subsequence \{f_{1, i}(x_1) : i \in \mathbb N\}. Define \mathcal F_1 := \{f_{1, i} : i \in \mathbb N\}.

Inductively, given each bounded collection \mathcal F_k, \{f(x_{k+1}) : f \in \mathcal F_k\} is bounded and hence contains a convergent subsequence \{f_{k+1, i}(x_1) : i \in \mathbb N\}. Define \mathcal F_{k+1} := \{f_{k+1, i} : i \in \mathbb N\}, so that inductively we can construct the decreasing chain of subsequences

\mathcal F = \mathcal F_0 \supseteq \mathcal F_1 \supseteq \mathcal F_2 \supseteq \cdots.

Now define the sub-sequence \{g_n\} of \{f_n\} by g_k := f_{k,k}. We claim that \{g_n\} converges. By our construction, it is clear that for any x_i, \{ g_n(x_i) : n \in \mathbb N\} converges and is thus Cauchy. Now, fix \epsilon > 0 and x_i \in \mathbb Q \cap [0, 1]. For any k_1 > 0, find N \in \mathbb N such that for m > n \geq N,

|g_m(x_i) - g_n(x_i)| < k_1 \cdot \epsilon.

Since \mathcal F is equicontinuous, for any k_2 > 0, we have that for any x \in [0, 1], there exists a neighbourhood U_x such that for any y \in [0, 1] and f \in \mathcal F,

y \in U_x \quad \Rightarrow \quad |f(x) - f(y)| < k_2 \cdot \epsilon.

Since \{U_x : x \in [0, 1]\} forms an open cover for [0, 1], we can extract a finite sub-cover \{U_{x_1'},\dots, U_{x_J'}\}. For each 1 \leq j \leq J, find x_{n_j} \in U_{x_j'}.

Now, for any x \in [0, 1], x \in U_{x_j'} for some j. Hence, for sufficiently large m,n,

\begin{aligned} |g_m(x) - g_n(x)| &\leq |g_m(x) - g_m(x_{n_j})| + |g_m(x_{n_j}) - g_n(x_{n_j})| + |g_n(x_{n_j}) - g_n(x)| \\ &\leq k_2 \cdot \epsilon + k_1 \cdot \epsilon + k_2 \cdot \epsilon = (k_1 + 2k_2) \cdot \epsilon. \end{aligned}

Setting k_1 = 1/2, k_2 = 1/4 proves that \{g_n(x)\} is Cauchy. A compactness argument shows that \{g_n\} is Cauchy, and since \mathcal C([0,1],\mathbb R) is complete, converges to some g \in \mathcal C([0, 1],\mathbb R). Since \mathcal F is closed, g \in \mathcal F, so that g_n \to g, as required.

The Arzelà-Ascoli theorem is incredibly useful in real analysis. For now, we complete the proof of Theorem 2.

Complete Proof of Theorem 2. By the Arzelà-Ascoli theorem, we need to prove that \mathcal F := \mathcal D_{0,M} \cap \mathcal D_{1,M} is closed, bounded, and equicontinuous. It is bounded since \mathcal D_{0,M} is. For any f \in \mathcal F, the mean value theorem yields the estimate

|f(x) - f(y)| = |f'(t)| \cdot |x-y| \leq M \cdot |x-y|.

Thus, given \epsilon > 0, the choice \delta := \epsilon/(2M) yields the desired equicontinuity. To check that \mathcal F is closed, fix a sequence \{f_n\} \subseteq \mathcal F such that f_n \to f. Then

\| f \|_\infty \leq \|f_n - f\|_\infty + M \quad \Rightarrow \quad f \in \mathcal D_{0, M}

and by the differentiable limit theorem, f is differentiable at any x \in [0, 1] with a similar estimate

\|f'\|_\infty \leq \|f_n' - f' \|_\infty + M \quad \Rightarrow \quad f \in \mathcal D_{1, M}.

—Joel Kindiak, 29 May 2025, 1439H

,

Published by


Leave a comment