What Must Be Checked in a Function-Space Compactness Claim?
A compactness argument for a family of continuous functions can fail even when every function takes values in a bounded set. The first question is not merely whether the pointwise values stay bounded: it is which topology is being used on the function space. The domain also matters. On a compact metric domain, compact-open convergence agrees with uniform convergence; on a noncompact domain, it generally does not. Finally, pointwise relative compactness does not replace equicontinuity in the Arzela-Ascoli criterion.
We will use \(C(K,\mathbb{R}^m)\) for the set of continuous maps from a compact metric space \(K\) to Euclidean space \(\mathbb{R}^m\). A family \(A\subseteq C(K,\mathbb{R}^m)\) is pointwise relatively compact if, for every \(x\in K\), the set \(\{f(x):f\in A\}\) has compact closure in \(\mathbb{R}^m\). It is equicontinuous if, for every \(x\in K\) and every \(\varepsilon>0\), there is \(\delta>0\) such that, for every \(f\in A\) and every \(y\in K\), \(d(x,y)<\delta\) implies \(\|f(x)-f(y)\|<\varepsilon\). The same \(\delta\) must work for all \(f\in A\).
The supremum in this formula is finite: each continuous function on compact \(K\) is bounded, so \(f-g\) is bounded. To use a compactness theorem correctly, one must establish that the theorem's topology is the topology of the claim. The next result justifies replacing compact-open topology by the uniform metric on this particular domain.
Compact-Open and Uniform Topologies on a Compact Metric Domain
Proof. First consider a compact-open subbasic set \([C,U]\), and \(f\in[C,U]\). If \(\mathbb{R}^m\setminus U\) is empty, \([C,U]\) is the whole function space. Otherwise \(f(C)\) is compact and disjoint from the closed set \(\mathbb{R}^m\setminus U\), so their distance is positive. Let that distance be \(r>0\). If \(d_\infty(f,g)<r\), then every \(g(x)\), \(x\in C\), remains within distance \(r\) of \(f(x)\) and cannot lie in \(\mathbb{R}^m\setminus U\). Thus \(g\in[C,U]\). This gives a uniform-metric neighborhood of \(f\) inside \([C,U]\), so every compact-open subbasic set is open in the uniform topology.
For the reverse inclusion, fix \(f\) and \(\varepsilon>0\). By continuity, each \(x\in K\) has an open neighborhood \(W_x\) whose closure satisfies \(\|f(z)-f(x)\|<\varepsilon/3\) for every \(z\in\overline{W_x}\). Such neighborhoods can be chosen by taking sufficiently small metric balls. Compactness gives a finite subcover \(W_{x_1},\ldots,W_{x_q}\) of \(K\). Consider the compact-open neighborhood $$ \bigcap_{i=1}^{q}[\overline{W_{x_i}},B(f(x_i),2\varepsilon/3)]. $$ It contains \(f\). If \(g\) belongs to this neighborhood and \(z\in K\), choose \(i\) with \(z\in W_{x_i}\). Then \(\|g(z)-f(x_i)\|<2\varepsilon/3\) and \(\|f(z)-f(x_i)\|<\varepsilon/3\), so the triangle inequality gives \(\|g(z)-f(z)\|<\varepsilon\). This holds for every \(z\), hence \(d_\infty(f,g)<\varepsilon\). Every uniform ball therefore contains a compact-open neighborhood of its center. The two topologies agree. \(\square\)
This equality is useful because uniform distance makes the role of equicontinuity precise. It is not a license to identify compact-open and uniform topologies when the domain is noncompact.
Worked Example: Pointwise Bounded but Not Relatively Compact
Let \(K=[0,1]\) with its usual metric and let \(A=\{f_n:n\in\mathbb{N}\}\subseteq C(K,\mathbb{R})\), where \(f_n(x)=x^n\). At each fixed \(x\in[0,1]\), all values \(f_n(x)\) lie in \([0,1]\), so the pointwise value set has compact closure. But \(A\) is not equicontinuous at \(1\).
Indeed, fix any \(\delta>0\). Choose \(x=1-\min(\delta/2,1/2)\). Then \(x\in[0,1)\) and \(|1-x|<\delta\). Since \(0\le x<1\), choose \(n\) large enough that \(x^n<1/2\). But \(f_n(1)=1\), so \(|f_n(1)-f_n(x)|>1/2\). Thus \(\varepsilon=1/2\) defeats every proposed common \(\delta\).
In fact, \(A\) is not relatively compact in the uniform topology. If it were, its uniform closure would be compact. A compact family of continuous functions on a compact metric domain is equicontinuous, as proved below, contradicting the calculation above. Equivalently, any uniformly convergent subsequence would have a continuous uniform limit, but pointwise its limit must be \(0\) on \([0,1)\) and \(1\) at \(1\), which is discontinuous. The pointwise bounds alone do not establish function-space compactness.
What Compactness Forces
Proof. For each \(x\in K\), the evaluation map \(e_x(f)=f(x)\) is continuous, since \(\|e_x(f)-e_x(g)\|\le d_\infty(f,g)\). Its image \(e_x(A)\) is compact, so the pointwise values have compact closure. This proves pointwise relative compactness.
Now fix \(\varepsilon>0\). The open uniform balls of radius \(\varepsilon/3\), centered at members of \(A\), cover \(A\). Compactness gives a finite subcover with centers \(h_1,\ldots,h_s\in A\). Each \(h_j\) is uniformly continuous on compact \(K\). For completeness, this follows by choosing, at each point of \(K\), a neighborhood on which the variation of \(h_j\) is less than \(\varepsilon/3\), and then taking a finite subcover and a sufficiently small common distance. Taking the minimum of the resulting finitely many positive distances gives \(\delta>0\) such that \(d(x,y)<\delta\) implies \(\|h_j(x)-h_j(y)\|<\varepsilon/3\) for every \(j\).
Given \(f\in A\), choose a center \(h_j\) with \(d_\infty(f,h_j)<\varepsilon/3\). For \(d(x,y)<\delta\), the triangle inequality gives $$ \|f(x)-f(y)\| \le \|f(x)-h_j(x)\|+\|h_j(x)-h_j(y)\|+\|h_j(y)-f(y)\| <\varepsilon. $$ The same \(\delta\) works for every \(f\in A\), proving equicontinuity. \(\square\)
This proposition gives a reliable direction for auditing a proposed proof: compactness implies both conditions. For the reverse direction, the relevant result is the Arzela-Ascoli theorem, not pointwise boundedness by itself.
Thus, on compact metric \(K\), a proposed proof of relative compactness must verify pointwise relative compactness and equicontinuity. If it proves only pointwise boundedness, it has not yet supplied the needed hypotheses. On a compact domain, pointwise boundedness in \(\mathbb{R}^m\) does imply pointwise relative compactness, by the Heine-Borel theorem; equicontinuity remains a separate requirement.
Worked Example: A Family That Passes the Test
For \(a\in[0,1]\), define \(g_a(x)=ax\) on \(K=[0,1]\), and set \(G=\{g_a:a\in[0,1]\}\). For each \(x\), the pointwise values are \(\{ax:a\in[0,1]\}=[0,x]\), which is compact (and is \(\{0\}\) when \(x=0\)). Also, for all \(a\in[0,1]\) and \(x,y\in[0,1]\), $$ |g_a(x)-g_a(y)|=a|x-y|\le |x-y|. $$ Given \(\varepsilon>0\), the common choice \(\delta=\varepsilon\) proves equicontinuity.
Arzela-Ascoli therefore says \(G\) is relatively compact. In this example we can check more: it is compact. The map \(a\mapsto g_a\) from \([0,1]\) to the uniform function space is continuous, because \(d_\infty(g_a,g_b)=|a-b|\). The continuous image of compact \([0,1]\) is compact, so \(G\) is compact. The explicit metric calculation verifies the continuity claim.
Why the Domain Hypothesis Cannot Be Ignored
On a noncompact domain, compact-open convergence tests behavior on each compact subset, while uniform convergence tests the supremum over the entire domain. These tests can differ substantially. The following example makes the distinction explicit without requiring any subtle properties of the compact-open topology.
Worked Example: Compact-Open Compactness Need Not Be Uniform Compactness
Give \(\mathbb{R}\) its usual topology and define \(u_n(x)=\max(0,1-|x-n|)\), for \(n\in\mathbb{N}\). Each \(u_n\) is continuous, is zero outside \([n-1,n+1]\), and satisfies \(u_n(n)=1\). Let \(H=\{0\}\cup\{u_n:n\in\mathbb{N}\}\), where \(0\) is the zero function.
For every compact \(C\subseteq\mathbb{R}\), there is \(M\) with \(C\subseteq[-M,M]\). If \(n>M+1\), then every \(x\in C\) satisfies \(x<n-1\), so \(u_n(x)=0\). Thus \(u_n|_C=0|_C\) for all sufficiently large \(n\). In particular, the sequence \(u_n\) converges to \(0\) in the compact-open topology: for a subbasic neighborhood \([C,U]\) of \(0\), where \(0\in U\), all sufficiently late \(u_n\) map \(C\) to \(0\in U\); finite intersections have the same eventual property.
The set \(H\) is compact in this topology. Given any open cover of \(H\), select one member containing \(0\). Convergence to \(0\) means this member contains all but finitely many \(u_n\). Choose one cover member for each of the remaining finitely many functions. These finitely many choices form a finite subcover.
But \(H\) is not compact in the uniform metric on bounded functions: for every \(n\), \(d_\infty(u_n,0)=1\), since \(u_n(n)=1\) and \(0\le u_n(x)\le1\). For distinct \(n,k\), the intervals \([n-1,n+1]\) and \([k-1,k+1]\) may touch, but at \(x=n\), \(u_n(n)=1\) and \(u_k(n)=0\) when \(|n-k|\ge2\). In particular the subsequence with even indices has pairwise uniform distances \(1\). It has no uniformly convergent subsequence, so the family cannot be relatively compact in the uniform metric. Compact-open compactness on a noncompact domain therefore does not assert uniform compactness.
A Practical Audit and a Common Mistake
When checking a function-space compactness argument, write down the function space and topology first. Then check the domain hypotheses of the proposed criterion. If \(K\) is compact metric and the topology is compact-open, the proved topology equality permits use of the uniform form of Arzela-Ascoli. Verify pointwise relative compactness at every \(x\), and verify equicontinuity with a single \(\delta\) for all functions in the family. If the domain is noncompact, do not silently replace compact-open convergence by uniform convergence or apply the compact-domain criterion as stated.
Common mistake. For \(f_n(x)=x^n\) on \([0,1]\), one might write: “For every \(x\), the values lie in the compact interval \([0,1]\); hence \(\{f_n\}\) is relatively compact.” This establishes pointwise relative compactness, not equicontinuity. The calculation at \(1\) above shows that for every \(\delta>0\), some \(x\) with \(|x-1|<\delta\) and some \(n\) satisfy \(|f_n(1)-f_n(x)|>1/2\). The corrected conclusion is that Arzela-Ascoli cannot be applied to infer relative compactness: its equicontinuity hypothesis fails, and the family is indeed not relatively compact in the uniform topology.
Check Your Understanding
For each question, identify the topology and every hypothesis needed before deciding what compactness conclusion follows.
- On \(C([0,1],\mathbb{R})\), why is it legitimate to use the uniform form of Arzela-Ascoli for the compact-open topology?
- For \(f_n(x)=x^n\) on \([0,1]\), which Arzela-Ascoli condition fails? Give the fixed positive \(\varepsilon\) that witnesses the failure.
- Does pointwise relative compactness alone imply relative compactness for a family of continuous functions on a compact metric domain? Explain which additional condition the criterion requires.
- For the functions \(u_n\) on \(\mathbb{R}\), why does convergence on every compact subset not imply uniform convergence?
- State the common-\(\delta\) requirement in the definition of equicontinuity, and explain why choosing a different \(\delta\) for each function is insufficient.