A recurring challenge throughout analysis is the mismatch between the uncountable nature of the objects we study and the countable nature of the tools we use to study them. The points of a metric space, the elements of a function space, the open sets in a topology — all are typically uncountable collections. Yet our most powerful techniques — sequential arguments, diagonal extractions, series expansions, exhaustion by countable families — are fundamentally countable operations. This tension surfaces immediately in elementary questions. In a metric space $(X, d)$, a set $F \subset X$ is closed if and only if every convergent sequence in $F$ has its limit in $F$. This sequential characterization is so natural that one might expect it to hold in any topological space. It does not. In a general topological space, closedness is defined by complements of open sets, and sequences may be too "thin" to detect all limit points — nets or filters are required instead. The sequential characterization works in metric spaces precisely because every metric space admits a countable local base at each point (take balls of radius $1/m$ for $m \in \mathbb{N}$), and this countable structure is enough to ensure that sequences capture the full topology. The property that makes these countable reductions possible at the global level is **separability**: the existence of a countable dense subset. When a space is separable, every element can be approximated — to arbitrary precision — by elements drawn from a single countable collection. This transforms uncountable problems into countable ones. Conversely, when separability fails, many of the standard tools of analysis break down in surprising and consequential ways. [example: The Failure of Separability in $\ell^\infty$] Consider the [Banach space](/page/Banach%20Spaces) $\ell^\infty$ of bounded real sequences equipped with the supremum norm $\|x\|_{\ell^\infty} = \sup_{k \in \mathbb{N}} |x_k|$. We show that $\ell^\infty$ is **not** separable by constructing an uncountable family of elements that are mutually "far apart." For each subset $S \subset \mathbb{N}$, define the indicator sequence $\mathbb{1}_S \in \ell^\infty$ by $(\mathbb{1}_S)_k = 1$ if $k \in S$ and $(\mathbb{1}_S)_k = 0$ if $k \notin S$. The collection $\{\mathbb{1}_S : S \subset \mathbb{N}\}$ has cardinality $2^{\aleph_0}$ (the cardinality of the power set of $\mathbb{N}$), which is uncountable. For any two distinct subsets $S, T \subset \mathbb{N}$, there exists an index $k$ belonging to exactly one of $S$ and $T$, so: \begin{align*} \|\mathbb{1}_S - \mathbb{1}_T\|_{\ell^\infty} = 1. \end{align*} Now suppose, for contradiction, that $\ell^\infty$ contained a countable dense subset $D = \{d_1, d_2, \ldots\}$. For each subset $S \subset \mathbb{N}$, the open ball $B(\mathbb{1}_S, 1/3)$ would need to contain at least one element $d_{n(S)}$ of $D$. Since the balls $B(\mathbb{1}_S, 1/3)$ and $B(\mathbb{1}_T, 1/3)$ are disjoint whenever $S \neq T$ (their centres are distance $1$ apart and the radii sum to $2/3 < 1$), the map $S \mapsto n(S)$ is injective from the uncountable power set $\mathcal{P}(\mathbb{N})$ into the countable set $\mathbb{N}$. This is impossible. The non-separability of $\ell^\infty$ has far-reaching consequences: the weak* topology on $(\ell^\infty)^* = (\ell^1)^{**}$ restricted to bounded sets is not metrizable, diagonal arguments cannot be applied to bounded sequences in $\ell^\infty$ to extract weak* convergent subsequences, and the space resists the kind of sequential analysis that works seamlessly in $\ell^p$ for $1 \le p < \infty$. [/example] Before stating the formal definition of separability, we recall the underlying notion of density. Let $(X, \tau)$ be a topological space and let $A \subset X$. The set $A$ is **dense** in $X$ if its closure equals the entire space: $\overline{A} = X$. Equivalently, $A$ is dense if and only if every nonempty open set $G \in \tau$ satisfies $G \cap A \neq \varnothing$. In a metric space $(X, d)$, this is further equivalent to: for every $x \in X$ and every $\varepsilon > 0$, there exists $a \in A$ with $d(x, a) < \varepsilon$. ## Definition With density in hand, separability is a single additional requirement. [definition: Separable Space] A topological space $(X, \tau)$ is **separable** if it contains a countable dense subset. That is, there exists a set $D \subset X$ with $D$ countable (finite or countably infinite) such that $\overline{D} = X$. [/definition] Separability depends on the topology, not on the set-theoretic structure of $X$ alone. The same underlying set can be separable in one topology and non-separable in another. For instance, $\mathbb{R}$ is separable in its usual (Euclidean) topology — $\mathbb{Q}$ is a countable dense subset — but non-separable in the discrete topology, where every subset is open and so no proper subset can be dense. The following examples establish separability for the spaces that appear most frequently in analysis. [example: Separability of Euclidean Space] The space $\mathbb{R}^n$ with the Euclidean metric is separable. The countable dense subset is $\mathbb{Q}^n = \{(q_1, \ldots, q_n) : q_i \in \mathbb{Q}\}$, which is countable as a finite product of countable sets. For any $x = (x_1, \ldots, x_n) \in \mathbb{R}^n$ and $\varepsilon > 0$, choose $q_i \in \mathbb{Q}$ with $|x_i - q_i| < \varepsilon / \sqrt{n}$ for each $i$. Then: \begin{align*} |x - q| = \left(\sum_{i=1}^n |x_i - q_i|^2\right)^{1/2} < \left(\sum_{i=1}^n \frac{\varepsilon^2}{n}\right)^{1/2} = \varepsilon. \end{align*} [/example] The same "rational approximation" strategy extends to infinite-dimensional sequence spaces, but with an essential difference: the tail of a sequence must be controlled, and only $p$-summability (not mere boundedness) makes this possible. [example: Separability of $\ell^p$ for $1 \le p < \infty$] The sequence space $\ell^p$ consists of all real sequences $x = (x_1, x_2, \ldots)$ with finite $\ell^p$-norm: \begin{align*} \|x\|_{\ell^p} = \left(\sum_{k=1}^\infty |x_k|^p\right)^{1/p}. \end{align*} Consider the set $D$ of sequences that are eventually zero and have rational entries: \begin{align*} D = \{(q_1, q_2, \ldots, q_N, 0, 0, \ldots) : N \in \mathbb{N}, \, q_i \in \mathbb{Q}\}. \end{align*} This is a countable union of countable sets: $D = \bigcup_{N=1}^\infty \mathbb{Q}^N$, hence countable. To see that $D$ is dense, fix $x \in \ell^p$ and $\varepsilon > 0$. Since $\sum_{k=1}^\infty |x_k|^p < \infty$, there exists $N \in \mathbb{N}$ such that $\sum_{k=N+1}^\infty |x_k|^p < (\varepsilon/2)^p$. For each $1 \le k \le N$, choose $q_k \in \mathbb{Q}$ with $|x_k - q_k| < \varepsilon/(2N^{1/p})$. Define $d = (q_1, \ldots, q_N, 0, 0, \ldots) \in D$. Then: \begin{align*} \|x - d\|_{\ell^p}^p &= \sum_{k=1}^N |x_k - q_k|^p + \sum_{k=N+1}^\infty |x_k|^p \\ &< N \cdot \frac{\varepsilon^p}{2^p N} + \frac{\varepsilon^p}{2^p} = \frac{\varepsilon^p}{2^p} + \frac{\varepsilon^p}{2^p} = \frac{\varepsilon^p}{2^{p-1}}. \end{align*} Therefore $\|x - d\|_{\ell^p} < \varepsilon / 2^{(p-1)/p} \le \varepsilon$, confirming that $D$ is dense in $\ell^p$. [/example] The contrast with $\ell^\infty$ is striking: the "tail truncation" argument above relies on the $p$-summability condition $\sum |x_k|^p < \infty$, which guarantees that truncating the tail introduces only a small error. In $\ell^\infty$, the supremum norm provides no such decay — a bounded sequence can maintain unit amplitude indefinitely, and no finite truncation captures the full structure. ## Separability and Second Countability Separability is a condition on the points of a space: the existence of a countable set that "fills out" the whole space. A closely related — but stronger — condition concerns the open sets: the existence of a countable family that generates the entire topology. Understanding the relationship between these two conditions is essential, because in metrizable spaces they turn out to be equivalent, while in general topological spaces they diverge. [definition: Second-Countable Space] A topological space $(X, \tau)$ is **second-countable** if it admits a countable base for its topology. That is, there exists a countable collection $\mathcal{B} = \{B_1, B_2, \ldots\} \subset \tau$ such that every open set $G \in \tau$ can be written as a union of members of $\mathcal{B}$. [/definition] Every second-countable space is separable: simply choose one point from each nonempty member of the countable base $\mathcal{B}$ to form a countable dense subset. The converse, however, fails dramatically in general topological spaces. [example: A Separable Space That Is Not Second-Countable] Let $X = \mathbb{R}$ with the **lower limit topology** (Sorgenfrey line), in which the basic open sets are the half-open intervals $[a, b)$ for $a < b$. The rationals $\mathbb{Q}$ form a countable dense subset (for any $x \in \mathbb{R}$ and $\varepsilon > 0$, the interval $[x, x + \varepsilon)$ contains a rational), so the Sorgenfrey line is separable. However, it is not second-countable. To see this, observe that for each $x \in \mathbb{R}$, the set $\{x\}$ can be distinguished from any point to its left: the open set $[x, x+1)$ contains $x$ but excludes every $y < x$. Any base $\mathcal{B}$ must therefore contain, for each $x \in \mathbb{R}$, a basic open set $B_x \in \mathcal{B}$ with $x \in B_x$ and $\inf B_x = x$. Since distinct real numbers $x \neq y$ produce distinct sets $B_x \neq B_y$ (they have different infima), the base $\mathcal{B}$ must be uncountable. [/example] The situation simplifies dramatically in the presence of a metric. In a metrizable space, the three notions — separability, second countability, and the Lindelof property — collapse into one. [quotetheorem:545] This equivalence is the reason separability is the "right" notion of smallness for metrizable spaces: establishing any one of the three properties automatically provides the other two. [explanation: Why Metrizability Collapses the Hierarchy] The key implication is $(1) \Rightarrow (2)$: given a countable dense subset $D = \{d_1, d_2, \ldots\}$, we construct a countable base by taking all rational-radius balls centred at elements of $D$: \begin{align*} \mathcal{B} = \{B(d_k, 1/m) : k, m \in \mathbb{N}\}. \end{align*} This is countable as a product $\mathbb{N} \times \mathbb{N}$. To verify it is a base: for any open $G \subset X$ and $x \in G$, there exists $r > 0$ with $B(x, r) \subset G$. By density, choose $d_k \in D$ with $d(x, d_k) < r/3$, and choose $m \in \mathbb{N}$ with $1/m < r/3$. Then $x \in B(d_k, 1/m)$ and $B(d_k, 1/m) \subset B(x, r) \subset G$. The implication $(2) \Rightarrow (3)$ holds in all topological spaces (not just metrizable ones), and $(3) \Rightarrow (1)$ uses the metric: cover $X$ by balls $\{B(x, 1/m) : x \in X\}$ for each $m$, extract a countable subcover by Lindelof, and collect the centres. [/explanation] An immediate and frequently used consequence concerns subspaces. [quotetheorem:942] This result depends critically on metrizability. In general topological spaces, subspaces of separable spaces need not be separable — for instance, the discrete topology on an uncountable set can be embedded as a subspace of certain separable spaces. In metric spaces, however, second countability passes to subspaces (a countable base for $X$ restricts to a countable base for any subspace), and second countability implies separability. ## Separability in Function Spaces The function spaces that arise in analysis — [Lebesgue spaces](/page/Lp%20Spaces), [Sobolev spaces](/page/Sobolev%20Spaces), spaces of continuous functions — are the primary setting where separability has practical consequences. Whether a given function space is separable determines which approximation and extraction techniques are available, and the answer depends sensitively on the norm. ### $L^p$ Spaces for $1 \le p < \infty$ The most fundamental approximation result in measure theory — that simple functions are dense in $L^p$ — provides the starting point for establishing separability. The challenge is to refine this approximation from simple functions (which form an uncountable family, since the coefficients are arbitrary reals) to a countable subfamily. [quotetheorem:548] The two hypotheses — $\sigma$-finiteness and countable generation of $\mathcal{F}$ — are both essential. If $\mathcal{F}$ is not countably generated, the "Layer 1" approximation below fails: one cannot reduce to simple functions on a countable generating family of sets, and the resulting family of approximants is uncountable. A concrete example: let $X = [0,1]$ equipped with the $\sigma$-algebra $\mathcal{F} = \mathcal{P}([0,1])$ (all subsets) and counting measure. The functions $\mathbb{1}_{\{t\}}$ for $t \in [0,1]$ are mutually distance $1$ in $L^p$, so $L^p(X, \mu)$ is not separable. In contrast, the Borel $\sigma$-algebra on $\mathbb{R}^n$ is countably generated (by rectangles with rational endpoints), which is why $L^p(\mathbb{R}^n)$ is separable. [explanation: The Construction for $L^p(\mathbb{R}^n)$] The countable dense subset is built in three layers of approximation. **Layer 1: Simple functions with rational coefficients on measurable rectangles.** Every $f \in L^p(\mathbb{R}^n)$ can be approximated by simple functions $\sum_{j=1}^N c_j \mathbb{1}_{E_j}$. By $\sigma$-finiteness, we may assume each $E_j$ has finite measure, and then approximate each $E_j$ by a finite union of rectangles $\prod_{i=1}^n (a_i, b_i]$ with rational endpoints (this uses the fact that such rectangles generate the Borel $\sigma$-algebra). Finally, replace each coefficient $c_j$ by a nearby rational $q_j$. **Layer 2: Counting the resulting family.** The collection of all functions of the form $\sum_{j=1}^N q_j \mathbb{1}_{R_j}$, where $N \in \mathbb{N}$, $q_j \in \mathbb{Q}$, and $R_j$ is a rectangle with rational endpoints, is a countable union of countable sets — hence countable. **Layer 3: Density in $L^p$.** Given $f \in L^p$ and $\varepsilon > 0$, first approximate $f$ by a simple function $g = \sum c_j \mathbb{1}_{E_j}$ with $\|f - g\|_{L^p} < \varepsilon/3$. Then approximate each $E_j$ by a union of rational rectangles $R_j$ with $\mathcal{L}^n(E_j \triangle R_j)$ small enough to ensure $\|g - g'\|_{L^p} < \varepsilon/3$, where $g' = \sum c_j \mathbb{1}_{R_j}$. Finally, replace each $c_j$ by $q_j \in \mathbb{Q}$ close enough to give $\|g' - g''\|_{L^p} < \varepsilon/3$. The triangle inequality yields $\|f - g''\|_{L^p} < \varepsilon$, and $g''$ belongs to the countable family. [/explanation] ### The Non-Separability of $L^\infty$ The space $L^\infty$ stands apart from the rest of the $L^p$ scale. The essential supremum norm, unlike the integral norms for $p < \infty$, does not allow errors to be "spread thin" over a large set — a function that differs from its approximation on a set of positive measure, no matter how small, incurs a fixed penalty in the $L^\infty$ norm. [example: Non-Separability of $L^\infty([0,1])$] For each $t \in (0,1)$, define $f_t = \mathbb{1}_{[0,t]} \in L^\infty([0,1])$. For distinct $s, t \in (0,1)$ with $s < t$, the function $f_t - f_s = \mathbb{1}_{(s,t]}$ satisfies: \begin{align*} \|f_t - f_s\|_{L^\infty} = 1. \end{align*} The collection $\{f_t : t \in (0,1)\}$ is an uncountable family with mutual distance $1$. The same argument as in the $\ell^\infty$ example — that the disjoint balls $B(f_t, 1/3)$ each require a distinct element of any dense subset — shows that no countable subset of $L^\infty([0,1])$ is dense. [/example] The non-separability of $L^\infty$ is not merely a curiosity; it is the source of genuine analytical difficulties. In many PDE and variational arguments, one works with bounded sequences in $L^\infty$ and seeks to extract convergent subsequences. Since $L^\infty$ is the dual of $L^1$ (when the measure space is $\sigma$-finite), the [Banach-Alaoglu theorem](/page/Banach-Alaoglu%20Theorem) guarantees weak* compactness of bounded sets. Because $L^1$ is separable (as established above), the weak* topology on bounded subsets of $L^\infty$ is metrizable, so sequential extraction of weak*-convergent subsequences does work. The subtlety, however, is that the weak* limit lies in $L^\infty$ but the convergence is only in the weak* topology — much weaker than norm convergence. Passing nonlinear operations (like products or compositions) through weak* limits requires additional structure, such as compensated compactness or Young measures. ### Spaces of Continuous Functions For spaces of continuous functions, separability is governed by the topology of the underlying domain. The question of which domains $K$ make $C(K)$ separable reduces, via the [Stone-Weierstrass theorem](/page/Weierstrass%20Approximation), to whether $K$ admits enough countable structure. [quotetheorem:943] Both hypotheses are necessary. Without compactness, the conclusion can fail: the space $C_b(\mathbb{R})$ of bounded continuous functions on $\mathbb{R}$ with the supremum norm is not separable (the uncountable family $\{x \mapsto \sin(\lambda x) : \lambda > 0\}$ can be used to produce elements with mutual distance bounded below). Without metrizability, the underlying space may be "too large" for a countable family to separate points: if $K$ is compact Hausdorff but not metrizable, then $K$ is not second-countable, and the Stone-Weierstrass construction cannot be reduced to a countable family of generators. For $K = [a,b] \subset \mathbb{R}$, the theorem follows from the classical [Weierstrass Approximation Theorem](/page/Weierstrass%20Approximation): polynomials with rational coefficients form a countable dense subset. The general case for compact metrizable $K$ uses the Stone-Weierstrass theorem applied to a countable family of functions that separates points (which exists because $K$, being compact and metrizable, is second-countable). [example: A Countable Dense Subset of $C([0,1])$] The polynomials with rational coefficients, \begin{align*} D = \left\{\sum_{k=0}^N q_k x^k : N \in \mathbb{N}_0, \, q_k \in \mathbb{Q}\right\}, \end{align*} form a countable dense subset of $C([0,1])$. To verify density: given $f \in C([0,1])$ and $\varepsilon > 0$, the Weierstrass Approximation Theorem provides a polynomial $p(x) = \sum_{k=0}^N a_k x^k$ with $\|f - p\|_{C([0,1])} < \varepsilon/2$. Choosing $q_k \in \mathbb{Q}$ with $|a_k - q_k| < \varepsilon/(2(N+1))$ for each $k$, the rational polynomial $p_\mathbb{Q}(x) = \sum_{k=0}^N q_k x^k$ satisfies: \begin{align*} \|p - p_\mathbb{Q}\|_{C([0,1])} \le \sum_{k=0}^N |a_k - q_k| \cdot \max_{x \in [0,1]} |x^k| \le \sum_{k=0}^N \frac{\varepsilon}{2(N+1)} = \frac{\varepsilon}{2}. \end{align*} The triangle inequality gives $\|f - p_\mathbb{Q}\|_{C([0,1])} < \varepsilon$. [/example] ### Sobolev Spaces The [Sobolev spaces](/page/Sobolev%20Spaces) $W^{k,p}(U)$ inherit separability from $L^p$ in a straightforward manner, via an isometric embedding into a product of $L^p$ spaces. [quotetheorem:944] The idea is to exploit the fact that the Sobolev norm controls the function and all its weak derivatives simultaneously, so the space embeds isometrically into a finite product of $L^p$ spaces — each of which is already known to be separable. [explanation: Separability of Sobolev Spaces via Isometric Embedding] The Sobolev norm controls both the function and all its weak derivatives up to order $k$: \begin{align*} \|u\|_{W^{k,p}(U)} = \left(\sum_{|\alpha| \le k} \|D^\alpha u\|_{L^p(U)}^p\right)^{1/p}. \end{align*} Define the map \begin{align*} \Phi: W^{k,p}(U) &\to \prod_{|\alpha| \le k} L^p(U) \\ u &\mapsto (D^\alpha u)_{|\alpha| \le k}. \end{align*} This is a linear isometry onto its image when the product is equipped with the $\ell^p$-sum norm. Since each factor $L^p(U)$ is separable (for $p < \infty$) and the product of finitely many separable metric spaces is separable, the product is separable. The image of $\Phi$ is a subspace of a separable metric space, hence separable. Since $\Phi$ is an isometry, $W^{k,p}(U)$ is separable. Analogously, the spaces $W^{k,p}_0(U)$ and $H^k(U) = W^{k,2}(U)$ are separable for $1 \le p < \infty$. [/explanation] ## Metrizability of Weak Topologies One of the most consequential applications of separability in functional analysis is its role in making [weak topologies](/page/Weak%20Convergence) manageable. The weak and weak* topologies on infinite-dimensional [Banach spaces](/page/Banach%20Spaces) are never metrizable on the full space (unless the space is finite-dimensional), but separability of the underlying space or its predual rescues metrizability on bounded subsets — which is precisely where compactness arguments take place. ### The Problem of Extracting Subsequences In variational problems and PDE theory, existence proofs typically follow a pattern: construct a bounded sequence (often a minimizing sequence for a functional), extract a convergent subsequence, and show the limit is a solution. The [Banach-Alaoglu theorem](/page/Banach-Alaoglu%20Theorem) guarantees that the closed unit ball of a dual space $X^*$ is compact in the weak* topology. However, compactness alone does not guarantee that **sequential** extraction is possible — in a general compact topological space, sequences may not suffice to describe the topology. Sequential compactness and compactness coincide in metrizable spaces. This is the critical link: if we can show the weak* topology is metrizable on bounded sets, then the Banach-Alaoglu compactness becomes sequential compactness, and the standard diagonal argument applies. [quotetheorem:547] The explicit formula for $\rho$ reveals exactly where separability enters: the metric is built from evaluations at a countable dense subset of $B_X$, which exists precisely because $X$ is separable. [explanation: Why Separability Is Needed for This Metric] The series converges because $|f(x_k) - g(x_k)| \le \|f - g\|_{X^*} \cdot \|x_k\|_X \le 2$, so each term is bounded by $2/2^k$. To verify that $\rho$ metrizes the weak* topology on $B_{X^*}$: convergence $f_m \to f$ in $\rho$ means $f_m(x_k) \to f(x_k)$ for every $k$. Since the $\{x_k\}$ are dense in $B_X$ and the $f_m$ are uniformly bounded (they lie in $B_{X^*}$), a standard $\varepsilon/3$ argument extends the convergence to all $x \in X$. This is exactly weak* convergence. Conversely, weak* convergence $f_m \overset{*}{\rightharpoonup} f$ in $B_{X^*}$ means $f_m(x) \to f(x)$ for all $x \in X$, which implies $f_m(x_k) \to f(x_k)$ for each $k$, and dominated convergence for the series (each term bounded by $2/2^k$) gives $\rho(f_m, f) \to 0$. Without separability of $X$, no countable family $\{x_k\}$ can be dense, and no such countable metric can capture the full weak* topology. In fact, if $X$ is a non-separable Banach space, the weak* topology on $B_{X^*}$ is **not** metrizable. [/explanation] The following theorem, which combines Banach-Alaoglu with the metrizability result, is the workhorse of sequential extraction in reflexive spaces. [quotetheorem:496] In practice, this is most often applied when $X$ is a separable reflexive Banach space — for instance, $X = L^p(U)$ with $1 < p < \infty$, or $X = W^{1,p}_0(U)$ with $1 < p < \infty$. In this case, $X$ is isometrically isomorphic to $X^{**}$, and the weak topology on $X$ coincides with the weak* topology under this identification. The theorem then yields: every bounded sequence in $X$ has a weakly convergent subsequence. ### The Role of Separability in Weak Convergence Arguments To illustrate how separability enters a concrete analytical argument, consider the standard approach to proving existence of minimizers in the calculus of variations. Suppose $I: W^{1,p}_0(U) \to \mathbb{R}$ is a functional of the form: \begin{align*} I[u] = \int_U F(x, u(x), \nabla u(x)) \, d\mathcal{L}^n \end{align*} and we seek to minimize $I$ over $W^{1,p}_0(U)$. The **direct method** proceeds as follows: 1. Take a minimizing sequence $\{u_m\}$ with $I[u_m] \to \inf I$. 2. Establish that $\{u_m\}$ is bounded in $W^{1,p}_0(U)$ (this typically uses coercivity of $F$). 3. Since $W^{1,p}_0(U)$ is **separable** and reflexive (for $1 < p < \infty$), the Sequential Banach-Alaoglu theorem extracts a weakly convergent subsequence $u_{m_j} \rightharpoonup u$ in $W^{1,p}_0(U)$. 4. Use lower semicontinuity of $I$ with respect to [weak convergence](/page/Weak%20Convergence) to conclude $I[u] \le \liminf I[u_{m_j}] = \inf I$. Step 3 would fail without separability: the Banach-Alaoglu theorem would still give compactness, but not sequential compactness, and the subsequent analysis — which relies on working with a concrete subsequence — would break down. ## Permanence Properties When a new space is constructed from spaces already known to be separable — by taking subspaces, quotients, products, or images under continuous maps — does the new space inherit separability? This question arises constantly in practice: to prove that a Sobolev space is separable, we embedded it into a product of $L^p$ spaces; to show a quotient Banach space $X/M$ is separable, we need to know that continuous images preserve separability. Without systematic permanence results, each new space would require a separability proof from scratch. ### Continuous Images A continuous function cannot "increase" the complexity of a space in terms of density. Separability is preserved under continuous surjections. [quotetheorem:945] The theorem does not require $f$ to be injective or to have any metric structure. It applies equally to quotient maps, projections, and arbitrary continuous surjections. However, it says nothing about preimages: a separable space can map continuously onto a non-separable one if the map is not surjective, and a non-separable space can certainly map onto a separable one. [explanation: Continuous Images Preserve Separability] Let $D = \{d_1, d_2, \ldots\}$ be a countable dense subset of $X$. The image $f(D) = \{f(d_1), f(d_2), \ldots\}$ is countable. To show it is dense in $Y$: let $V \subset Y$ be a nonempty open set. Since $f$ is surjective and continuous, $f^{-1}(V)$ is a nonempty open subset of $X$. By density of $D$, there exists $d_k \in D \cap f^{-1}(V)$, which gives $f(d_k) \in f(D) \cap V$. Since $V$ was arbitrary, $f(D)$ is dense in $Y$. This has an important corollary: every quotient of a separable space (by an equivalence relation with an open quotient map) is separable. In particular, if $X$ is a separable Banach space and $M \subset X$ is a closed subspace, then the quotient space $X / M$ is separable. [/explanation] ### Countable Products One of the most useful permanence properties is that separability is closed under countable products. This is essential for constructing countable dense subsets in spaces like $\mathbb{R}^\mathbb{N}$ (the space of all real sequences with the product topology). [quotetheorem:946] The restriction to countable products and the product topology are both essential. The result fails for uncountable products (see the remark below), and it also fails if the product topology is replaced by the box topology — where basic open sets may specify constraints on every coordinate simultaneously rather than only finitely many. [explanation: Construction of a Dense Subset in the Product] For each $k \in \mathbb{N}$, let $D_k = \{d_{k,1}, d_{k,2}, \ldots\}$ be a countable dense subset of $X_k$. Fix a "base point" $p_k \in D_k$ for each $k$. Consider the set $E$ of all sequences in $\prod_{k=1}^\infty X_k$ that agree with some element of $D_k$ in finitely many coordinates and equal $p_k$ in all remaining coordinates: \begin{align*} E = \left\{(y_1, y_2, \ldots) \in \prod_{k=1}^\infty X_k : y_k \in D_k \text{ for all } k, \text{ and } y_k = p_k \text{ for all but finitely many } k\right\}. \end{align*} This set is countable: it is the countable union $E = \bigcup_{N=1}^\infty E_N$, where $E_N$ consists of those sequences with $y_k = p_k$ for all $k > N$, and $E_N \cong D_1 \times \cdots \times D_N$ is a finite product of countable sets. To verify density: a basic open set in the product topology has the form $\prod_{k=1}^\infty G_k$ where $G_k = X_k$ for all but finitely many $k$, and $G_k$ is open in $X_k$ for the remaining indices. Say $G_k \neq X_k$ only for $k \in \{k_1, \ldots, k_N\}$. For each such $k_j$, the density of $D_{k_j}$ provides an element $d_{k_j} \in D_{k_j} \cap G_{k_j}$. Setting $y_{k_j} = d_{k_j}$ and $y_k = p_k$ for all other $k$ produces an element of $E$ in the basic open set. [/explanation] [remark: Uncountable Products and the Failure of Separability] The restriction to countable products is sharp. The uncountable product $\{0,1\}^{\mathbb{R}}$ (with the product topology) is **not** separable, despite each factor $\{0,1\}$ being finite. This can be shown by observing that any dense subset must "distinguish" between the uncountably many coordinate projections, which no countable set can do. More precisely, for each $t \in \mathbb{R}$, the projection $\pi_t: \{0,1\}^{\mathbb{R}} \to \{0,1\}$ is continuous, and the preimages $\pi_t^{-1}(\{0\})$ and $\pi_t^{-1}(\{1\})$ are disjoint open sets. Any dense subset must intersect both, providing at least two distinct "choices" for each of uncountably many coordinates — a structure that no countable set can support. [/remark] ### Subspaces The behavior of separability under passing to subspaces depends on the ambient space. In general topological spaces, subspaces of separable spaces need not be separable. The situation is fundamentally different in metrizable spaces, where (as established in the section on second countability) every subspace of a separable metrizable space is separable. This follows from the equivalence between separability and second countability for metrizable spaces, combined with the fact that a subspace of a second-countable space is second-countable. For Banach spaces, this yields the following useful result. [quotetheorem:947] This is an immediate consequence of the general metrizable result, but it is worth stating explicitly because of how frequently it is used: if $L^p(U)$ is separable (for $1 \le p < \infty$), then every closed subspace — such as $W^{k,p}_0(U)$, $L^p_0(U)$ (functions with mean zero), or the range of a bounded projection — is automatically separable. ## Connections to Descriptive Set Theory and Probability Separability plays a foundational role beyond the immediate concerns of functional analysis. Two directions deserve mention: Polish spaces in descriptive set theory, and measurability in probability. A **Polish space** is a separable, completely metrizable topological space. This class — which includes $\mathbb{R}^n$, $\ell^p$ for $1 \le p < \infty$, $C([0,1])$, and all separable Banach spaces — is the natural setting for descriptive set theory, Borel complexity theory, and the theory of analytic sets. The separability hypothesis is not a convenience here; it is structural. Many of the key results — the Kuratowski isomorphism theorem (every uncountable Polish space is Borel isomorphic to $\mathbb{R}$), the Lusin-Suslin theorem, the existence of regular conditional probabilities — depend essentially on separability. In probability theory, separability ensures measurability of essential operations. If $X$ is a separable Banach space, the Borel $\sigma$-algebra $\mathcal{B}(X)$ (generated by open sets) coincides with the $\sigma$-algebra generated by the continuous linear functionals on $X$. This means that an $X$-valued random variable is Borel measurable if and only if $f(X)$ is a real-valued random variable for every $f \in X^*$ — this is Pettis measurability, and it holds automatically in the separable case. Without separability, the Borel $\sigma$-algebra can be strictly larger than the "cylinder" $\sigma$-algebra, and pathological measurability issues arise: there exist non-separably-valued functions that are "weakly measurable" (measurable against every functional) but not Borel measurable. Separability also underpins the disintegration of measures: the existence of regular conditional distributions on a probability space $(\Omega, \mathcal{F}, \mathbb{P})$ given a sub-$\sigma$-algebra requires that the target space be Polish (hence separable). ## Standard Techniques Working with separability involves a small but recurring set of proof techniques. This section collects the patterns that appear most often, so that the reader can recognize and deploy them in new settings. ### Constructing Countable Dense Subsets via Approximation Chains The most common method for proving separability is to build the countable dense subset through successive layers of approximation — each layer replacing one source of uncountability with a countable substitute. The $L^p$ construction above illustrates this: first approximate by simple functions (replacing arbitrary functions by finite sums), then approximate the sets by rational rectangles (replacing arbitrary measurable sets by combinatorial objects), then approximate the coefficients by rationals (replacing arbitrary reals by a countable field). The general pattern is: 1. **Structural approximation:** Use a density theorem (Weierstrass, Meyers-Serrin, density of simple functions) to reduce to a structured class (polynomials, smooth functions, step functions). 2. **Combinatorial reduction:** Reduce the parameter space of the structured class from uncountable to countable — typically by restricting to rational coefficients, rational endpoints, or finite-rank objects. 3. **Verification of density:** Confirm that the countable family is dense by chaining the approximation errors via the triangle inequality. [example: Separability of the Space of Finite Borel Measures] Consider the space $\mathcal{M}([0,1])$ of finite signed Borel measures on $[0,1]$ with the total variation norm $\|\mu\|_{TV} = |\mu|([0,1])$. We show that this space is **not** separable. For each $t \in [0,1]$, the Dirac measure $\delta_t$ has $\|\delta_t\|_{TV} = 1$. For distinct $s, t \in [0,1]$: \begin{align*} \|\delta_s - \delta_t\|_{TV} = |\delta_s - \delta_t|([0,1]) = |\delta_s - \delta_t|(\{s\}) + |\delta_s - \delta_t|(\{t\}) + |\delta_s - \delta_t|([0,1] \setminus \{s,t\}) = 1 + 1 + 0 = 2. \end{align*} The uncountable family $\{\delta_t : t \in [0,1]\}$ has mutual distance $2$, so $\mathcal{M}([0,1])$ is not separable. This is consistent with the Riesz representation theorem identifying $\mathcal{M}([0,1])$ with $C([0,1])^*$: the dual of a separable space need not be separable. [/example] ### The Diagonal Argument for Sequential Extraction When separability provides a countable structure, the diagonal argument converts pointwise convergence on a countable set into convergence on the entire space. The pattern is: 1. **Enumerate** the countable dense subset (or countable base) as $\{d_1, d_2, \ldots\}$. 2. **Extract along $d_1$:** From the bounded sequence $\{f_m\}$, extract a subsequence $\{f_{m}^{(1)}\}$ such that $f_{m}^{(1)}(d_1)$ converges. 3. **Extract along $d_2$:** From $\{f_{m}^{(1)}\}$, extract a further subsequence $\{f_{m}^{(2)}\}$ such that $f_{m}^{(2)}(d_2)$ converges. (Since $\{f_{m}^{(2)}\}$ is a subsequence of $\{f_{m}^{(1)}\}$, convergence at $d_1$ is preserved.) 4. **Iterate and diagonalize:** After extracting along all $d_k$, take the diagonal subsequence $g_m = f_{m}^{(m)}$. This subsequence converges at every $d_k$. 5. **Extend by density:** Use uniform boundedness and the density of $\{d_k\}$ to extend convergence from $D$ to all of $X$. This argument is the engine behind the Sequential Banach-Alaoglu theorem and appears throughout PDE theory, probability, and harmonic analysis whenever one needs to pass from a bounded family to a convergent sequence. Without the countable dense set provided by separability, step 1 fails and the entire argument collapses. ### The Isometric Embedding Technique To prove that a space $X$ is separable, it sometimes suffices to embed $X$ isometrically into a space already known to be separable. The Sobolev space argument above is a prototype: $W^{k,p}(U)$ embeds isometrically into a finite product of $L^p(U)$ spaces, each of which is separable. More generally: [quotetheorem:948] The conclusion follows because $\Phi(X)$ is a subspace of the separable metric space $Y$, and every subspace of a separable metric space is separable (as established earlier). Composing $\Phi^{-1}$ (defined on $\Phi(X)$) with a countable dense subset of $\Phi(X)$ yields a countable dense subset of $X$. Note that the result requires only an isometric embedding, not an isometric surjection — the image $\Phi(X)$ may be a proper subspace of $Y$. This technique is particularly useful for showing separability of spaces defined by norms that control multiple quantities simultaneously — Sobolev norms, Besov norms, Triebel-Lizorkin norms — where the space naturally embeds into a product of simpler spaces. ### Separability and the Existence of Schauder Bases A **Schauder basis** for a [Banach space](/page/Banach%20Spaces) $X$ is a sequence $\{e_k\}_{k=1}^\infty \subset X$ such that every $x \in X$ has a unique representation $x = \sum_{k=1}^\infty a_k e_k$ (convergence in norm). The existence of a Schauder basis immediately implies separability: the countable set of finite linear combinations $\sum_{k=1}^N q_k e_k$ with $q_k \in \mathbb{Q}$ is dense. The converse is false: Per Enflo (1973) constructed a separable Banach space with no Schauder basis. Nevertheless, most concrete separable Banach spaces encountered in analysis do possess natural Schauder bases: - **$\ell^p$ ($1 \le p < \infty$):** The standard unit vectors $e_k = (\underbrace{0, \ldots, 0}_{k-1}, 1, 0, \ldots)$. - **$L^p([0,1])$ ($1 < p < \infty$):** The Haar system. - **$C([0,1])$:** The Schauder system (piecewise linear hat functions). The existence of a basis provides not just separability but a concrete mechanism for approximation: truncate the series at finite $N$ to obtain a sequence of approximations with explicit error control. ## References 1. Brezis, H., *Functional Analysis, Sobolev Spaces and Partial Differential Equations* (2011). 2. Conway, J. B., *A Course in Functional Analysis* (1990). 3. Kechris, A. S., *Classical Descriptive Set Theory* (1995). 4. Munkres, J. R., *Topology* (2000). 5. Rudin, W., *Functional Analysis* (1991). 6. Megginson, R. E., *An Introduction to Banach Space Theory* (1998).