The [Bolzano-Weierstrass theorem](/theorems/628) is one of the most powerful results in elementary analysis: every bounded [sequence](/page/Sequence) in $\mathbb{R}^n$ has a convergent subsequence. This single statement underpins the Extreme Value Theorem, the existence of minimisers in optimisation, and the convergence of approximation schemes throughout classical analysis. The natural question is: in which spaces does this property persist?

The answer depends entirely on the topology of the ambient space. In [metric spaces](/page/Metric%20Space), the ability to extract convergent subsequences from arbitrary sequences turns out to be equivalent to [compactness](/page/Compact%20Space) --- the open-cover property that, at first glance, appears to have nothing to do with sequences. This equivalence is the reason analysts in $\mathbb{R}^n$ can move freely between "every open cover has a finite subcover" and "every sequence has a convergent subsequence" without comment.

But this equivalence is a theorem about metric spaces, not a tautology. In general topological spaces, the two properties diverge: a space can be compact without being sequentially compact, and sequentially compact without being compact. Understanding where and why this divergence occurs --- and what additional structure restores the equivalence --- is essential for working with the weak and weak-$*$ topologies that dominate modern functional analysis.

[example: Where Sequences Fail to Detect Compactness] Consider the product space $X = \{0, 1\}^{[0,1]}$, where each factor $\{0, 1\}$ carries the discrete topology and the index set is the uncountable interval $[0, 1]$. Elements of $X$ are functions $f: [0, 1] \to \{0, 1\}$, and convergence in the [product topology](/page/Product%20Topology) means pointwise convergence: $f_n \to f$ if and only if $f_n(t) \to f(t)$ for every $t \in [0, 1]$. By [Tychonoff's theorem](/theorems/953), $X$ is [compact](/page/Compact%20Space) (it is a product of compact spaces). However, $X$ is *not* sequentially compact. For each $n \in \mathbb{N}$, write $n$ in binary as $n = \sum_{k=0}^\infty b_k(n) \, 2^k$ (where $b_k(n) \in \{0,1\}$ and only finitely many digits are nonzero), and define $f_n: [0, 1] \to \{0, 1\}$ by \begin{align*} f_n(t) = b_k(n) \quad \text{where } t \in [k \cdot 2^{-N}, (k+1) \cdot 2^{-N}) \end{align*} for an appropriate dyadic decomposition (more precisely, define $f_n$ to encode the binary digits of $n$ across the coordinates). A cleaner construction uses the following observation: since $|[0,1]| = \mathfrak{c} = |2^{\mathbb{N}}|$, there exists a bijection $\varphi: [0,1] \to 2^{\mathbb{N}}$, where $2^{\mathbb{N}}$ denotes the set of all subsets of $\mathbb{N}$. Define $f_n: [0,1] \to \{0,1\}$ by $f_n(t) = \mathbb{1}_{\{n \in \varphi(t)\}}$, so $f_n(t) = 1$ if $n$ belongs to the subset $\varphi(t)$, and $f_n(t) = 0$ otherwise. We claim no subsequence of $\{f_n\}$ converges pointwise. Suppose for contradiction that $\{f_{n_j}\}_{j=1}^\infty$ converges pointwise to some $f: [0, 1] \to \{0, 1\}$. Since each $f_{n_j}(t) \in \{0, 1\}$ and the limit $f(t)$ must also lie in $\{0, 1\}$, convergence at $t$ means $f_{n_j}(t)$ is eventually constant: either eventually $1$ or eventually $0$. Define the set $A = \{n_j : j \text{ is odd}\} \subset \mathbb{N}$. Since $\varphi$ is a bijection from $[0,1]$ onto $2^{\mathbb{N}}$, there exists $t_0 \in [0,1]$ with $\varphi(t_0) = A$. Then $f_{n_j}(t_0) = \mathbb{1}_{\{n_j \in A\}}$, which equals $1$ when $j$ is odd and $0$ when $j$ is even. The sequence $\{f_{n_j}(t_0)\}$ alternates between $0$ and $1$ and does not converge --- contradicting the assumption that $\{f_{n_j}\}$ converges pointwise at every $t$. This space is compact (by Tychonoff) but not sequentially compact. The open-cover definition detects compactness; sequences do not. [/example]

This example reveals a fundamental limitation: sequences, being indexed by $\mathbb{N}$, are a *countable* probe into the topology. When the topology is determined by uncountably many coordinates (as in uncountable products) or when it lacks a countable local base (as in non-first-countable spaces), sequences may fail to capture the full topological structure. The correct generalisation of sequential compactness to arbitrary topological spaces uses *nets* (Moore-Smith sequences) or *filters*, which can be indexed by arbitrary directed sets. But in the settings that matter most for analysis --- metric spaces and the weak topology on separable Banach spaces --- sequences suffice, and understanding precisely when they suffice is the subject of this article.

## Definition

[definition: Sequential Compactness] A topological space $X$ is **sequentially compact** if every sequence $\{x_k\}_{k=1}^\infty$ in $X$ has a convergent subsequence. That is, there exist a strictly increasing function $k: \mathbb{N} \to \mathbb{N}$ (written $k_1 < k_2 < k_3 < \cdots$) and a point $x \in X$ such that $x_{k_j} \to x$ as $j \to \infty$. [/definition]

Several points of clarification are worth recording immediately.

[remark: Convergence of Subsequences Versus Cluster Points] A point $x \in X$ is a **cluster point** (or accumulation point) of a sequence $\{x_k\}_{k=1}^\infty$ if every neighbourhood of $x$ contains $x_k$ for infinitely many values of $k$. In a metric space, $x$ is a cluster point of $\{x_k\}$ if and only if some subsequence of $\{x_k\}$ converges to $x$. In a general topological space, this equivalence can fail: a sequence may have cluster points without having any convergent subsequence (though this pathology does not arise in first-countable spaces). Sequential compactness requires the existence of a *convergent subsequence*, which is a stronger demand than the existence of a cluster point in general, but equivalent in first-countable spaces. [/remark]

[remark: Sequential Compactness Versus Compactness] In full generality, neither sequential compactness nor compactness implies the other: - **Compact but not sequentially compact:** $\{0, 1\}^{\mathfrak{c}}$ (as shown above). - **Sequentially compact but not compact:** The ordinal space $\omega_1 = [0, \omega_1)$ with the order topology, where $\omega_1$ is the first uncountable ordinal. Every sequence in $\omega_1$ is bounded (since a countable union of countable sets is countable in ZFC, and each $\{x_k\}$ has a countable supremum), and hence contained in a closed interval $[0, \alpha]$ for some countable ordinal $\alpha$. This interval is compact (it is order-isomorphic to a well-ordered subset of $\mathbb{R}$), so every sequence has a convergent subsequence. But $\omega_1$ itself is not compact: the open cover $\{[0, \alpha) : \alpha < \omega_1\}$ has no finite subcover. The two notions coincide in metric spaces (and more generally in second-countable spaces), which is the content of the main equivalence theorem below. [/remark]

## The Equivalence in Metric Spaces

The most important fact about sequential compactness is that, in the metric setting, it is simply another name for compactness. This equivalence is not an accident --- it reflects the fact that metric topologies are completely determined by convergent sequences. But the equivalence is non-trivial: it requires linking the "global" open-cover property to the "local" sequential property via an intermediate geometric condition.

The bridge is **total boundedness** --- a quantitative refinement of boundedness that measures how efficiently the space can be approximated by finite sets.

[definition: Total Boundedness] A [metric space](/page/Metric%20Space) $(M, d)$ is **totally bounded** if for every $\varepsilon > 0$, there exists a finite set $\{x_1, \ldots, x_N\} \subset M$ (an **$\varepsilon$-net**) such that \begin{align*} M = \bigcup_{i=1}^N B(x_i, \varepsilon). \end{align*} [/definition]

Total boundedness is strictly stronger than boundedness in general. Every bounded subset of $\mathbb{R}^n$ is totally bounded (one can cover a bounded set by finitely many cubes of any prescribed side length), but in infinite-dimensional normed spaces, the closed unit ball is bounded yet *never* totally bounded --- the standard basis vectors in $\ell^2$ are mutually separated by $\sqrt{2}$, so no finite collection of balls of radius less than $\sqrt{2}/2$ can cover them all.

Total boundedness is also strictly weaker than compactness on its own: the open interval $(0, 1)$ with the standard metric is totally bounded (it inherits total boundedness from the compact set $[0, 1]$) but not compact, because it is not complete. The missing ingredient is completeness, and the following theorem makes this precise.

[quotetheorem:316]

This three-way equivalence is the reason that compactness in metric spaces is such a robust notion. Each characterisation has a distinct practical role:

**Characterisation (1)** (open covers) is needed for proofs that proceed by covering the space with sets having a local property and then extracting a finite subcover to obtain a *global* conclusion. The proof that continuous functions on compact sets are uniformly continuous is the prototype.

**Characterisation (2)** (sequential compactness) is the workhorse of existence theory. The direct method in the [calculus of variations](/page/Calculus%20of%20Variations), Peano's existence theorem for ODEs, and most PDE existence proofs all rely on extracting convergent subsequences from bounded sequences of approximate solutions.