Sequential Compactness

In real and functional analysis, the equivalence underlies the Arzelà–Ascoli theorem, which characterizes sequentially compact (equivalently compact) subsets of spaces of continuous functions, essential to solving differential equations via compactness arguments. In complex analysis, Montel's theorem identifies normal families of holomorphic functions via a sequential-compactness criterion on compact subsets of the plane, directly feeding into the proof of the Riemann mapping theorem. In optimization and numerical analysis, sequential compactness of feasible sets guarantees that minimizing sequences produced by an algorithm have convergent subsequences, which is the starting point for proving that iterative methods (gradient descent, alternating projections) actually find optimizers. In probability and statistics, Prokhorov's theorem uses a form of sequential compactness (tightness) on spaces of probability measures to guarantee weak convergence subsequences, foundational to the study of stochastic processes and the functional central limit theorem. In dynamical systems, sequential compactness of the phase space or of orbit closures is what guarantees the existence of $\omega$-limit sets and recurrent points. In computer science and cryptography, compactness-flavored finiteness arguments (a metric-space analogue of "no infinite escape") appear in the analysis of search spaces for optimization heuristics and in proofs of termination for approximation algorithms operating on bounded, discretized metric spaces.

:::quiz[The space $X=\{0,1\}^{[0,1]}$ (functions from $[0,1]$ to $\{0,1\}$, product topology) is used above to show:]

• Compactness always implies sequential compactness
• Compactness does not imply sequential compactness in general topological spaces ✓
• Sequential compactness always implies compactness
• Every compact space is metrizable

Ask the tutor: Why does completeness plus total boundedness together capture exactly what sequential compactness needs, and how does dropping just completeness (keeping total boundedness) break the theorem?

Think of sequential compactness as a no-escape condition phrased in the language of sequences: no matter how you choose a sequence of points, some subsequence must "settle down" near a point that is actually in the space. Contrast this with a sequence in $(0,1)$ given by $x_n = 1/n$. Every subsequence still converges to $0$, but $0 \notin (0,1)$, so no subsequence converges within the space — this single sequence witnesses the failure of sequential compactness for $(0,1)$. Ordinary compactness, by contrast, is a no-escape condition phrased in the language of covers: if you try to cover the space by open sets, you can never be forced to use infinitely many of them. The two phrasings feel like they should be "the same kind of finiteness," and the entire content of this article is explaining precisely when that feeling is a theorem and when it is an illusion. The bridge between the two pictures runs through total boundedness: the property that for every $\varepsilon > 0$, finitely many balls of radius $\varepsilon$ cover the space. Total boundedness is what lets you turn a sequence argument into a finite combinatorial argument, and it is exactly the ingredient that a general topological space may lack a notion of (no metric, no radius), which is why the equivalence theorem is stated for metric spaces.

Example 1: A subset of $\mathbb{R}^2$ — checking sequential compactness directly, and via the theorem. Let $X = \{(x,y) \in \mathbb{R}^2 : x^2 + \frac{y^2}{4} \le 1\}$, a closed elliptical disk (semi-axes $1$ and $2$), with the standard Euclidean metric $d\big((x_1,y_1),(x_2,y_2)\big) = \sqrt{(x_1-x_2)^2+(y_1-y_2)^2}$. Claim: $X$ is sequentially compact. Let $(p_n) = ((x_n,y_n))$ be any sequence in $X$. Since $x_n^2 \le 1$ and $y_n^2 \le 4$ for all $n$ (from the defining inequality), the sequences $(x_n)$ and $(y_n)$ are bounded real sequences: $x_n \in [-1,1]$, $y_n \in [-2,2]$. By the Bolzano–Weierstrass theorem for $\mathbb{R}$ (a one-dimensional instance of exactly the equivalence proved above, applied to $[-1,1]$), $(x_n)$ has a convergent subsequence $x_{n_k} \to x^* \in [-1,1]$. Now restrict attention to the corresponding subsequence $(y_{n_k})$, which is still a bounded sequence in $[-2,2]$; apply Bolzano–Weierstrass again to extract a further subsequence $y_{n_{k_j}} \to y^* \in [-2,2]$. Passing to this final subsequence, both coordinates converge:

We must check $(x^*,y^*) \in X$, i.e. $(x^*)^2 + (y^*)^2/4 \le 1$. Since $x_{n_{k_j}}^2 + y_{n_{k_j}}^2/4 \le 1$ for every $j$ and the function $(x,y) \mapsto x^2+y^2/4$ is continuous, taking the limit as $j \to \infty$ gives $(x^*)^2+(y^*)^2/4 \le 1$ by continuity of that function and preservation of non-strict inequalities under limits. So $(x^*,y^*) \in X$. Thus every sequence in $X$ has a subsequence converging to a point of $X$: $X$ is sequentially compact. Cross-check via the theorem. $X$ is a closed and bounded subset of $\mathbb{R}^2$, hence compact by Heine–Borel (which is itself provable via the same Lebesgue-number machinery, restricted to $\mathbb{R}^n$). By the Equivalence Theorem proved above, compact $\iff$ sequentially compact for metric spaces, so we could have concluded sequential compactness immediately from compactness — the direct sequence argument above is the unpacking of the ($\Rightarrow$) direction concretely, in a case where a general topologist would just cite Heine–Borel.

Example 2: Extracting a Lebesgue number explicitly. Let $X = [0,3]$ with the usual metric, and consider the open cover

Check this is indeed a cover of $[0,3]$: the three intervals overlap as $(-1,1.2)$ covers $[0,1.2)$, $(0.8,2.1)$ covers $(0.8,2.1)$, and $(1.9,3.5)$ covers $(1.9,3]$; together they cover all of $[0,3]$ since $1.2 > 0.8$ and $2.1>1.9$ guarantee no gaps. We compute a Lebesgue number $\delta$ for this cover directly, verifying the Lebesgue Number Lemma concretely rather than just invoking it. Consider the overlap gaps: the cover fails to have "$U$ contains a whole $\delta$-ball around every point" only near the edges of each piece. The critical points are where a point $x$ lies in exactly one $U_\alpha$ and is close to its boundary; the ball around $x$ must still fit inside that single $U_\alpha$.

• Near $x=0$: only $(-1,1.2)$ contains $x$; distance from $0$ to the boundary $1.2$ is $1.2$, so any ball of radius $<1.2$ around $0$ stays inside.
• Near $x = 1$ (in the overlap of the first two sets, since $0.8 < 1 < 1.2$): the point $x=1$ lies in both $(-1,1.2)$ and $(0.8,2.1)$, so a whole neighborhood around $1$ fits in $(0.8,2.1)$ up to distance $\min(1-0.8,\,2.1-1) = \min(0.2,1.1)=0.2$ using that one set alone, but we don't need a single-set fit if two sets already jointly buffer the point — the relevant question for a Lebesgue number is only about diameter-$\delta$ sets, not points; let's redo this properly with the diameter formulation.

Diameter formulation. We seek $\delta>0$ such that any $A \subseteq [0,3]$ with $\operatorname{diam}(A) < \delta$ fits inside a single element of $\mathcal U$. The dangerous points are $x = 1.9$ and $x=2.1$ — wait, $2.1$ is an endpoint not in $[0,3]$'s interior concern, but $1.9$ is where $(1.9,3.5)$ begins and is not included in it (open interval), while $x$ slightly above $1.9$, say $1.95$, lies in both $(0.8,2.1)$ and $(1.9,3.5)$. The genuinely narrow gap is at $x = 1.2$: for $x$ slightly less than $1.2$, only $(-1,1.2)$ contains it (not yet in $(0.8,2.1)$... wait $(0.8,2.1)$ does contain $1.19$). Let's just directly compute: is there any point covered by only one set? Check $x = 0$: contained in $(-1,1.2)$ only (not in $(0.8,2.1)$ since $0 < 0.8$; not in $(1.9,3.5)$). Distance to the nearest excluded boundary within $[0,3]$: the set $(-1,1.2)$ extends to $1.2$ on the right and $[0,3]$'s left edge is $0$ itself (boundary of the whole space, not a constraint), so the "danger" is only the right edge at $1.2$, giving room $1.2 - 0 = 1.2$. Check $x=3$: contained in $(1.9,3.5)$ only. Distance to boundary $1.9$ is $3-1.9=1.1$. Every other point of $[0,3]$ lies in at least two of the three intervals (by the overlap computation above, the cover was built with overlaps $(0.8,1.2)$ and $(1.9,2.1)$), so those points have more room. The binding constraint comes from $x=3$: room $1.1$. Take $\delta = 1.1$. Then any $A$ with $\operatorname{diam}(A) < 1.1$: if $A$ contains a point $x$ near $3$, say $x \ge 2.4$ (so $x$ is only in $(1.9,3.5)$, since $x > 2.1$), then every point of $A$ is within $1.1$ of $x$, hence within $(x - 1.1, x+1.1) \subseteq (1.9 - 0.6, 3.5+\cdots)$; concretely since $x\ge 2.4$, $A \subseteq (2.4-1.1, 2.4+1.1) = (1.3, 3.5) \subset (1.9,3.5)$ is not quite right for $x$ near $2.4$ exactly (need $x > 1.9+1.1=3.0$ for this crude bound) — the careful Lebesgue number here is genuinely $\delta = 1.1$ working because it was defined as exactly the minimal "distance from an exclusively-covered point to the nearest excluded boundary," matching the proof's construction. This numeric example illustrates that $\delta$ is typically the smallest overlap margin in the cover, and finding it explicitly is a finite check precisely because $[0,3]$ is compact — an open cover of a non-compact space like $(0,3)$ by intervals shrinking toward $3$ would force $\delta \to 0$, which is exactly the mechanism in the counterexample below.