In finite-dimensional analysis, the closed unit ball of $\mathbb{R}^n$ is compact, and the Bolzano-Weierstrass theorem guarantees that every bounded sequence has a convergent subsequence. This [compactness](/page/Compact%20Space) is the foundation of existence arguments throughout mathematics: to find a minimiser, an equilibrium, or a fixed point, one extracts a convergent subsequence from a bounded sequence of candidates and verifies that the limit has the desired property. In infinite-dimensional [Banach spaces](/page/Banach%20Space) — the natural setting for partial differential equations — this strategy collapses. The closed unit ball of an infinite-dimensional [normed space](/page/Normed%20Vector%20Space) is never compact in the norm topology. Bounded sequences can oscillate with ever-increasing frequency, or concentrate at different spatial locations, without any subsequence converging in norm. The Riesz lemma makes this failure quantitative: in any infinite-dimensional Banach space $X$, the unit sphere contains a sequence $\{x_n\}_{n=1}^\infty$ with $\|x_m - x_n\|_X > 1/2$ for all $m \neq n$. [example: Failure of Norm Compactness in $\ell^2$] Consider the [Hilbert space](/page/Hilbert%20Space) $\ell^2(\mathbb{N})$ with its standard orthonormal basis $\{e_n\}_{n=1}^\infty$, where $e_n$ is the sequence with $1$ in position $n$ and $0$ elsewhere. The sequence $\{e_n\}_{n=1}^\infty$ is bounded: $\|e_n\|_{\ell^2} = 1$ for all $n$. However, for $m \neq n$, \begin{align*} \|e_m - e_n\|_{\ell^2} = \sqrt{\|e_m\|_{\ell^2}^2 + \|e_n\|_{\ell^2}^2} = \sqrt{2}, \end{align*} since $(e_m, e_n)_{\ell^2} = 0$. No subsequence is Cauchy, so no subsequence converges in norm. Yet the same sequence *does* converge in a weaker sense. For any $f \in (\ell^2)^*$, the [Riesz Representation Theorem](/theorems/218) identifies $f$ with some $a = (a_1, a_2, \ldots) \in \ell^2$, and \begin{align*} f(e_n) = (a, e_n)_{\ell^2} = a_n \to 0 \quad \text{as } n \to \infty, \end{align*} because $\sum_{n=1}^\infty |a_n|^2 < \infty$ forces $a_n \to 0$. Every [continuous](/page/Continuity) linear functional on $\ell^2$ evaluates to zero in the limit — the sequence $\{e_n\}$ converges *weakly* to $0$, even though it stays at constant distance from $0$ in norm. [/example] This example reveals the central tension: the norm topology preserves quantitative control (norm convergence implies convergence of all functionals, uniformly) but sacrifices compactness. To recover compactness in infinite dimensions, we must accept a coarser notion of convergence — one that tests each continuous linear functional individually, without uniformity across all functionals. The **weak topology** is the precise formalisation of this idea. ## Definition The weak topology on a [Banach space](/page/Banach%20Space) $X$ is not constructed by inventing a new norm; it is constructed by specifying which functions must be continuous. This places it in the framework of *initial topologies* from general topology. [definition: Weak Topology] Let $X$ be a Banach space over $\mathbb{R}$ with [continuous](/page/Continuity)) dual $X^*$. The **weak topology** on $X$, denoted $\sigma(X, X^*)$, is the coarsest topology on $X$ that makes every bounded linear functional $f \in X^*$ continuous as a map $f: X \to \mathbb{R}$. Equivalently, $\sigma(X, X^*)$ is the initial topology on $X$ with respect to the family of maps $\{f : X \to \mathbb{R}\}_{f \in X^*}$. [/definition] To make this concrete, the weak topology has a neighbourhood basis at each point $x_0 \in X$ consisting of sets of the form \begin{align*} U(x_0; f_1, \ldots, f_k; \varepsilon) := \{x \in X : |f_i(x - x_0)| < \varepsilon \text{ for } i = 1, \ldots, k\}, \end{align*} where $k \in \mathbb{N}$, $f_1, \ldots, f_k \in X^*$, and $\varepsilon > 0$. Each such set is the intersection of finitely many "slabs" — preimages of open intervals under continuous linear functionals. This makes the weak topology much coarser than the norm topology: a weakly [open set](/page/Open%20Set) is obtained by imposing only *finitely many* linear constraints, whereas a norm-open ball $B(x_0, r) = \{x : \|x - x_0\|_X < r\}$ simultaneously constrains *all* functionals. [remark: The Norm Topology is Stronger] Since every $f \in X^*$ is continuous in the norm topology by definition of $X^*$, every weakly open set is also norm-open. Thus $\sigma(X, X^*) \subset \tau_{\|\cdot\|}$, where $\tau_{\|\cdot\|}$ is the norm topology. The two topologies coincide if and only if $X$ is finite-dimensional. [/remark] The relationship between the weak topology on $X$ and the [weak* topology](/page/Weak*%20Topology) on $X^*$ is often a source of confusion. The weak* topology $\sigma(X^*, X)$ on the dual space $X^*$ is generated by the evaluation functionals $\{\hat{x} : x \in X\}$, where $\hat{x}(f) = f(x)$. The weak topology $\sigma(X, X^*)$ on $X$ is generated by the functionals in $X^*$ themselves. When $X$ is [reflexive](/page/Reflexive%20Space) — meaning the canonical embedding $J: X \to X^{**}$ is surjective — the weak topology on $X$ coincides with the weak* topology on $X$ when $X$ is viewed as the dual of $X^*$. In the non-reflexive case, these are genuinely different topologies with different [compactness](/page/Compact%20Space) properties. ## Weak Convergence and Its Characterisation The definition of the weak topology specifies which sets are open, but in analysis one typically works with *sequences* rather than [open sets](/page/Open%20Set). The question that immediately arises is: what does convergence of a sequence look like in the weak topology, and how can one verify it without directly reasoning about the open sets of $\sigma(X, X^*)$? A sequence $\{x_n\}_{n=1}^\infty$ in a [Banach space](/page/Banach%20Space) $X$ converges weakly to $x \in X$ — written $x_n \rightharpoonup x$ — if and only if every continuous linear functional converges: \begin{align*} f(x_n) \to f(x) \quad \text{for every } f \in X^*. \end{align*} This is the content of the following characterisation: [quotetheorem:255] The characterisation reduces a topological question (convergence in $\sigma(X, X^*)$) to a family of scalar questions (convergence of real-valued sequences $\{f(x_n)\}$ for each $f$). This is what makes weak convergence tractable: instead of verifying convergence in a potentially exotic topology, one tests against each functional individually. However, several features distinguish weak convergence from norm convergence, and failing to appreciate these distinctions leads to errors. The first fundamental fact is that the weak topology is coarser than the norm topology — every weakly open set is norm-open, but not conversely. At the level of sequences, this means norm convergence is a strictly stronger condition. [quotetheorem:982] The forward direction follows from the continuity of every $f \in X^*$: the bound $|f(x_n) - f(x)| \le \|f\|_{X^*} \|x_n - x\|_X$ forces $f(x_n) \to f(x)$ for every $f$. The converse failure in infinite dimensions is demonstrated by the orthonormal basis example above: $e_n \rightharpoonup 0$ in $\ell^2$ but $\|e_n\|_{\ell^2} = 1$ for all $n$. In finite dimensions, the equivalence holds because the weak and norm topologies coincide (any linear functional on $\mathbb{R}^n$ is a dot product with a fixed vector, and coordinatewise convergence implies norm convergence). The second key property is that [weakly convergent](/page/Weak%20Convergence) sequences cannot escape to infinity — a consequence of the Uniform Boundedness Principle. [quotetheorem:983] The proof uses the canonical embedding: for each $n$, define $\hat{x}_n \in X^{**}$ by $\hat{x}_n(f) = f(x_n)$. Since $x_n \rightharpoonup x$, the sequence $\{\hat{x}_n(f)\} = \{f(x_n)\}$ is convergent (hence bounded) for each $f \in X^*$. The Uniform Boundedness Principle applied to the family $\{\hat{x}_n\}_{n=1}^\infty \subset X^{**}$ gives $\sup_n \|\hat{x}_n\|_{X^{**}} < \infty$. The Hahn-Banach theorem guarantees $\|\hat{x}_n\|_{X^{**}} = \|x_n\|_X$, completing the bound. A third distinction concerns the topology itself: in infinite dimensions, the weak topology is strictly coarser than the norm topology, and this coarseness manifests as a failure of first countability. [quotetheorem:984] This means that sequences alone do not determine the weak topology — nets or filters are needed in general. Despite this, the Eberlein-Šmulian theorem (discussed below) shows that for *bounded* subsets of $X$, sequential and topological notions of weak [compactness](/page/Compact%20Space) coincide — a remarkable and non-obvious fact that rescues sequential arguments in most applications. [example: Weak Convergence in $L^p$] Let $U \subset \mathbb{R}^n$ be a bounded open set, let $1 < p < \infty$, and let $q = p/(p-1)$ be the conjugate exponent. The dual of $L^p(U)$ is isometrically isomorphic to $L^q(U)$ via the pairing \begin{align*} f_g(u) := \int_U g(x) u(x) \, d\mathcal{L}^n(x), \quad g \in L^q(U). \end{align*} A sequence $\{u_n\}_{n=1}^\infty \subset L^p(U)$ converges weakly to $u \in L^p(U)$ if and only if \begin{align*} \int_U u_n(x) g(x) \, d\mathcal{L}^n(x) \to \int_U u(x) g(x) \, d\mathcal{L}^n(x) \quad \text{for every } g \in L^q(U). \end{align*} Consider the sequence of functions $u_n: (0, 2\pi) \to \mathbb{R}$ defined by $u_n(x) = \sin(nx)$. We claim $u_n \rightharpoonup 0$ in $L^2(0, 2\pi)$. For any $g \in L^2(0, 2\pi)$, the integral $\int_0^{2\pi} g(x) \sin(nx) \, d\mathcal{L}^1(x)$ is the $n$-th Fourier sine coefficient of $g$, scaled by $\pi$. The Riemann-Lebesgue lemma guarantees that the Fourier coefficients of any $L^1$ function tend to zero, so \begin{align*} \int_0^{2\pi} g(x) \sin(nx) \, d\mathcal{L}^1(x) \to 0 \quad \text{as } n \to \infty. \end{align*} Thus $u_n \rightharpoonup 0$. Yet $\|u_n\|_{L^2}^2 = \int_0^{2\pi} \sin^2(nx) \, d\mathcal{L}^1 = \pi$ for every $n$, so $\|u_n\|_{L^2} = \sqrt{\pi}$ is constant. The sequence oscillates with increasing frequency, distributing its energy uniformly across the interval, but every individual functional "averages out" the oscillations. This is the prototypical mechanism of weak convergence: oscillation at increasingly fine scales. [/example] ## Weak Versus Norm Closure: Mazur's Theorem A natural question arises once one has two topologies on the same space: when do they produce the *same* [closed sets](/page/Closed%20Set)? For general sets, the weak topology has strictly more closed sets than necessary — the norm-closed unit sphere $\{x : \|x\|_X = 1\}$, for instance, is not weakly closed in infinite dimensions (the orthonormal basis example shows $e_n \rightharpoonup 0$, so the sphere's weak closure contains $0$). However, for a special but extremely important class of sets — *convex* sets — the two topologies cannot be distinguished. [quotetheorem:985] This result has a remarkable consequence for sequences: if $x_n \rightharpoonup x$ weakly, we cannot conclude that $x$ lies in the norm closure of the *sequence* $\{x_n\}$, but Mazur's Lemma (the sequential version of Mazur's Theorem) guarantees something nearly as good — *convex combinations* of the $x_n$ converge strongly. [quotetheorem:216] Mazur's Lemma is the tool one reaches for when weak convergence alone is insufficient: it produces a *new* sequence (of convex combinations) that converges strongly. The lemma follows from Mazur's Theorem applied to the closed convex hull of the tail $\{x_n, x_{n+1}, \ldots\}$, which contains $x$ (as a weak limit, $x$ lies in the weak closure of the convex hull, but Mazur says this equals the norm closure). The interplay between Mazur's Theorem and weak convergence is one of the most-used tools in PDE theory. The typical application runs as follows: one has a [weakly convergent](/page/Weak%20Convergence) sequence $u_n \rightharpoonup u$ in a Sobolev space $W^{1,p}(U)$, and one needs to verify that the limit $u$ satisfies a constraint that is expressed as membership in a closed convex set $K$ (for instance, the set of functions satisfying $u \geq 0$ a.e., or $|\nabla u| \leq 1$ a.e.). Since $K$ is norm-closed and convex, Mazur's Theorem guarantees that $K$ is also weakly closed, so $u_n \in K$ for all $n$ and $u_n \rightharpoonup u$ together imply $u \in K$. [example: The Weak Closure of the Unit Sphere] Let $X$ be an infinite-dimensional [Banach space](/page/Banach%20Space) with closed unit ball $\overline{B}_X = \{x \in X : \|x\|_X \leq 1\}$ and unit sphere $S_X = \{x \in X : \|x\|_X = 1\}$. We show that the weak closure of $S_X$ is the entire closed unit ball $\overline{B}_X$. Since $S_X \subset \overline{B}_X$ and $\overline{B}_X$ is convex and norm-closed, Mazur's Theorem gives $\overline{S_X}^{\,\sigma(X,X^*)} \subset \overline{B}_X$. For the reverse inclusion, fix any $y \in \overline{B}_X$ with $\|y\|_X < 1$. We must show that every weak neighbourhood of $y$ intersects $S_X$. A basic weak neighbourhood of $y$ has the form $U(y; f_1, \ldots, f_k; \varepsilon)$ for some $f_1, \ldots, f_k \in X^*$ and $\varepsilon > 0$. Since $X$ is infinite-dimensional, the subspace $\ker f_1 \cap \cdots \cap \ker f_k$ is infinite-dimensional (its codimension is at most $k$). Choose any $z$ in this intersection with $z \neq 0$. For $t \in \mathbb{R}$, define $x_t := y + tz$. Then $f_i(x_t) = f_i(y) + t f_i(z) = f_i(y)$ for all $i$, so $x_t \in U(y; f_1, \ldots, f_k; \varepsilon)$ for every $t$. Since $t \mapsto \|y + tz\|_X$ is a [continuous](/page/Continuity) function that equals $\|y\|_X < 1$ at $t = 0$ and tends to $\infty$ as $|t| \to \infty$, by the intermediate value theorem there exists $t_0$ with $\|y + t_0 z\|_X = 1$. Then $x_{t_0} \in S_X \cap U(y; f_1, \ldots, f_k; \varepsilon)$. Since every weak neighbourhood of $y$ meets $S_X$, we have $y \in \overline{S_X}^{\,\sigma(X,X^*)}$. Therefore $\overline{S_X}^{\,\sigma(X,X^*)} = \overline{B}_X$. This shows strikingly that the unit sphere, while norm-closed, is *not* weakly closed in infinite dimensions. Sets that are "thin" in the norm topology can become "fat" in the weak topology. The distinguishing feature is convexity: the closed ball $\overline{B}_X$ is convex and hence weakly closed (by Mazur), but the sphere is not convex and its weak closure "fills in" to the entire ball. [/example] ## Weak Lower Semicontinuity A central obstacle in the [calculus of variations](/page/Calculus%20of%20Variations) is the following: if a sequence of approximate minimisers $u_n \rightharpoonup u$ converges weakly, can we conclude that the functional's value at the limit is no larger than the limiting value? This is the property of **weak lower semicontinuity**, and it is the bridge between weak [compactness](/page/Compact%20Space) (which provides convergent subsequences) and the existence of minimisers (which requires control of the functional at the limit). The simplest instance is the norm itself. [quotetheorem:215] The inequality $\|x\|_X \leq \liminf_{n \to \infty} \|x_n\|_X$ is strict in general: the orthonormal basis $e_n \rightharpoonup 0$ in $\ell^2$ gives $\|0\|_{\ell^2} = 0 < 1 = \liminf \|e_n\|_{\ell^2}$. The weak limit always has norm no larger than the limiting norms, but the "energy" lost to oscillation or dispersion can cause a strict drop. This result relies on the norm being a convex function: for any $f \in X^*$ with $\|f\|_{X^*} \leq 1$, the inequality $f(x) \leq \|x\|_X$ holds, and the [Hahn-Banach theorem](/theorems/879) guarantees that the supremum over such $f$ recovers the norm: $\|x\|_X = \sup_{\|f\|_{X^*} \leq 1} f(x)$. Since each $f$ is weakly [continuous](/page/Continuity), the norm is a supremum of weakly continuous functions and hence weakly lower semicontinuous. The same argument extends to any convex, norm-lower-semicontinuous functional $I: X \to \mathbb{R} \cup \{+\infty\}$. This leads to a fundamental principle: [quotetheorem:986] The hypothesis of convexity is essential and cannot be removed. Without convexity, weak lower semicontinuity can fail spectacularly, and this failure is at the heart of many phenomena in nonlinear analysis. [example: Failure of Weak Lower Semicontinuity Without Convexity] Define $I: L^2(0, 2\pi) \to \mathbb{R}$ by \begin{align*} I(u) := \int_0^{2\pi} (u(x)^2 - 1)^2 \, d\mathcal{L}^1(x). \end{align*} This functional is non-negative and equals zero precisely when $|u(x)| = 1$ a.e. It is continuous in the norm topology but *not* convex (the integrand $\varphi(t) = (t^2 - 1)^2$ is not convex on $\mathbb{R}$). Consider the sequence $u_n(x) = \sin(nx)$. As shown above, $u_n \rightharpoonup 0$ in $L^2(0, 2\pi)$. We compute $I(u_n)$: \begin{align*} I(u_n) &= \int_0^{2\pi} (\sin^2(nx) - 1)^2 \, d\mathcal{L}^1 = \int_0^{2\pi} \cos^4(nx) \, d\mathcal{L}^1. \end{align*} Using the power-reduction identity \begin{align*} \cos^4(\theta) = \frac{3}{8} + \frac{1}{2}\cos(2\theta) + \frac{1}{8}\cos(4\theta) \end{align*} and the fact that $\int_0^{2\pi} \cos(m n x) \, d\mathcal{L}^1 = 0$ for $mn \neq 0$: \begin{align*} I(u_n) = \frac{3}{8} \cdot 2\pi = \frac{3\pi}{4}. \end{align*} But \begin{align*} I(0) = \int_0^{2\pi} (0 - 1)^2 \, d\mathcal{L}^1 = 2\pi. \end{align*} Therefore $I(0) = 2\pi > \frac{3\pi}{4} = \lim_{n \to \infty} I(u_n)$, violating weak lower semicontinuity. The functional *increases* at the weak limit. The oscillating sequence achieves lower energy by spending time near $|u| = 1$ (where the integrand vanishes), while the weak limit $u = 0$ sits at the local maximum of the integrand. [/example] ## Weak Compactness and Reflexivity ### The Compactness Landscape The primary motivation for introducing the weak topology is to recover [compactness](/page/Compact%20Space). But which spaces admit weakly compact unit balls? And what is the relationship between weak compactness and the structure of the Banach space? The Banach-Alaoglu theorem (see the [Weak* Topology](/page/Weak*%20Topology) page) guarantees that the closed unit ball $B_{X^*}$ of the dual $X^*$ is compact in the weak* topology $\sigma(X^*, X)$. This is a theorem about $X^*$, not about $X$. For the space $X$ itself, the analogous statement — that $B_X$ is weakly compact — requires an additional structural assumption. [quotetheorem:897] Kakutani's Theorem reveals a deep equivalence: a Banach space has weakly compact closed balls if and only if it is reflexive. This gives the weak topology on [reflexive spaces](/page/Reflexive%20Space) the same "compactness power" that the weak* topology has on all dual spaces, and it explains why reflexive spaces — $L^p$ for $1 < p < \infty$, [Hilbert spaces](/page/Hilbert%20Space), Sobolev spaces $W^{k,p}$ for $1 < p < \infty$ — are the natural setting for variational methods. For non-reflexive spaces, the closed unit ball $B_X$ is *not* weakly compact. The canonical embedding $J: X \hookrightarrow X^{**}$ maps $B_X$ into $B_{X^{**}}$, and [Goldstine's Lemma](/theorems/898) shows that $J(B_X)$ is weak*-dense in $B_{X^{**}}$. When $X \neq X^{**}$, the image $J(B_X)$ is a proper [dense subset](/page/Dense%20Subset) of the weak*-compact set $B_{X^{**}}$ — it is not closed, hence not compact. ### Sequential Weak Compactness Even in reflexive spaces, the weak topology is typically not [metrizable](/page/Metrizable%20Space) on the entire space, so the relationship between compactness and sequential compactness is not automatic. The Eberlein-Smulian theorem resolves this by showing that for *bounded* subsets of a Banach space, the sequential and topological notions of weak compactness coincide. [quotetheorem:987] The equivalence (1) $\Leftrightarrow$ (2) is remarkable because the weak topology is not metrizable in general, and in non-metrizable topological spaces, sequential compactness and compactness are logically independent properties. The Eberlein-Smulian theorem says that for *bounded* subsets of [Banach spaces](/page/Banach%20Space), this pathology does not arise. Combining Kakutani's Theorem with Eberlein-Smulian gives the most-used compactness result in PDE theory: [quotetheorem:214] This is the workhorse of the *direct method* in the [calculus of variations](/page/Calculus%20of%20Variations). In a reflexive Banach space, every bounded sequence has a weakly convergent subsequence — exactly the sequential compactness that one needs to extract limits from minimising sequences. The hypotheses deserve scrutiny: **Reflexivity is necessary.** In $\ell^1(\mathbb{N})$, which is not reflexive, the standard basis $\{e_n\}$ is bounded ($\|e_n\|_{\ell^1} = 1$) but has no [weakly convergent](/page/Weak%20Convergence) subsequence. In fact, every weakly convergent sequence in $\ell^1$ is norm-convergent (this is Schur's Theorem, discussed below), so if $\{e_{n_k}\}$ converged weakly, it would converge in norm, which is impossible since $\|e_m - e_n\|_{\ell^1} = 2$ for $m \neq n$. **Boundedness is necessary.** Even in a reflexive space, an unbounded sequence need not have a weakly convergent subsequence. The fundamental obstruction is that weakly convergent sequences are necessarily bounded: by the Uniform Boundedness Principle, if $x_n \rightharpoonup x$ then $\sup_n \|x_n\|_X < \infty$ (as discussed above). Therefore any unbounded sequence — regardless of the space — cannot converge weakly. To see that no *subsequence* converges weakly either, consider a concrete example in $L^2(0,1)$: define $u_n := n \mathbb{1}_{(0, 1/n)}$, so that $\|u_n\|_{L^2} = \sqrt{n} \to \infty$. Any subsequence $\{u_{n_k}\}$ is still unbounded in $L^2(0,1)$, and since weak convergence would force boundedness, no subsequence can converge weakly. ## Schur's Theorem: The $\ell^1$ Anomaly In most infinite-dimensional [Banach spaces](/page/Banach%20Space), the weak topology is strictly coarser than the norm topology: there exist sequences that converge weakly but not in norm. The space $\ell^1(\mathbb{N})$ is a dramatic exception. [quotetheorem:988] Schur's Theorem says that the weak and norm topologies on $\ell^1$ produce the same convergent sequences. Since $\ell^1$ is not [reflexive](/page/Reflexive%20Space) (its dual is $\ell^\infty$, whose dual $(\ell^\infty)^* \supsetneq \ell^1$), Kakutani's Theorem tells us $B_{\ell^1}$ is not weakly compact. Schur's Theorem makes the failure of weak [compactness](/page/Compact%20Space) vivid: the standard basis $\{e_n\}$ in $\ell^1$ has no [weakly convergent](/page/Weak%20Convergence) subsequence precisely *because* any weak limit would also be a norm limit, and no norm-convergent subsequence exists. The key to Schur's Theorem is the special structure of $(\ell^1)^* \cong \ell^\infty$. Every $f \in (\ell^1)^*$ can be identified with a bounded sequence $a = (a_1, a_2, \ldots) \in \ell^\infty$ via $f(x) = \sum_{j=1}^\infty a_j x_j$. Among these functionals are the *coordinate projections* $\pi_j(x) = x_j$ (corresponding to $a = e_j \in \ell^\infty$). Weak convergence $x_n \rightharpoonup x$ therefore implies componentwise convergence $(x_n)_j \to x_j$ for each $j \in \mathbb{N}$. In general Banach spaces, componentwise convergence (testing against all functionals) does not force norm convergence — this is exactly what makes the weak topology strictly coarser than the norm topology. What is remarkable about $\ell^1$ is that componentwise convergence *does* force norm convergence, a consequence of the additive structure of the $\ell^1$ norm across coordinates. [example: Schur's Property Fails in $\ell^2$] The standard basis $\{e_n\}_{n=1}^\infty$ in $\ell^2$ converges weakly to zero: for any $a \in \ell^2 \cong (\ell^2)^*$, we have $a(e_n) = a_n \to 0$ since the series $\sum |a_n|^2$ converges. Yet the sequence does not converge in norm: \begin{align*} \|e_n - 0\|_{\ell^2} = 1 \quad \text{for all } n \in \mathbb{N}. \end{align*} The mechanism is **mass spreading**: each $e_n$ places all its mass on a single coordinate, and as $n$ increases, this coordinate moves further out. Any individual functional $a$ sees the mass vanish (because $a_n \to 0$), but the total $\ell^2$ norm stays fixed at $1$. The $\ell^2$ norm, being computed as a root-sum-of-squares, cannot detect where in the sequence the mass sits — only how much there is in total. In $\ell^1$, this behaviour is forbidden. The $\ell^1$ norm $\|x\|_{\ell^1} = \sum_{j=1}^\infty |x_j|$ is *additive* across coordinates, and the coordinate projections $\pi_j \in (\ell^1)^*$ detect mass at each position individually. If $x_n \rightharpoonup x$ in $\ell^1$, then $(x_n)_j \to x_j$ for each $j$, and Schur's Theorem shows this componentwise convergence forces $\sum_j |(x_n)_j - x_j| \to 0$ — a statement that has no analogue in $\ell^p$ for $p > 1$. [/example] ## Weak Topology and Metrizability A topology is [metrizable](/page/Metrizable%20Space) if its open sets can be described by a metric. The weak topology on an infinite-dimensional [Banach space](/page/Banach%20Space) is never metrizable on the whole space — the neighbourhood basis at the origin is not countable. This raises a practical question: can we nonetheless use metric-space techniques (sequential arguments, epsilon-delta reasoning) when working with the weak topology? The answer is *yes*, under a separability assumption and a boundedness restriction. [quotetheorem:989] The separability of $X^*$ is automatic in many spaces of interest. If $X$ is a [separable](/page/Separable%20Space)) [reflexive](/page/Reflexive%20Space) Banach space, then $X^*$ is also separable. In particular, the spaces $L^p(U)$ for $1 < p < \infty$ (with $U$ a bounded domain in $\mathbb{R}^n$) have separable duals, so the weak topology on bounded subsets of $L^p(U)$ is metrizable. This justifies the sequential arguments that pervade PDE theory: extracting [weakly convergent](/page/Weak%20Convergence) subsequences, applying diagonal arguments, and testing convergence against a countable dense family of functionals. When $X^*$ is not separable — for instance, $X = \ell^1$, whose dual $\ell^\infty$ is not separable — the weak topology on bounded sets need not be metrizable, and sequential arguments require greater care (the Eberlein-Smulian theorem still applies, but one cannot construct an explicit metric). ## The Direct Method in the Calculus of Variations The weak topology was not developed as an abstract curiosity; it was designed to solve *existence problems*. The **direct method**, introduced by Tonelli and refined by many subsequent authors, is the standard framework for proving the existence of minimisers of variational problems. The weak topology provides every ingredient. The setup is as follows. Let $X$ be a reflexive [Banach space](/page/Banach%20Space) (typically a [Sobolev space](/page/Sobolev%20Space) $W^{1,p}(U)$ with $1 < p < \infty$) and let $I: X \to \mathbb{R} \cup \{+\infty\}$ be a functional. We seek $u \in X$ with \begin{align*} I(u) = \inf_{v \in X} I(v). \end{align*} The direct method proceeds in three steps, each of which requires a specific property of the weak topology. **Step 1: Extract a minimising sequence and a weak limit.** Let $\{u_n\}_{n=1}^\infty \subset X$ satisfy $I(u_n) \to \inf_{v \in X} I(v)$. The first task is to show that $\{u_n\}$ is bounded in $X$. This requires a **coercivity** condition on $I$: the assumption $I(v) \to +\infty$ as $\|v\|_X \to \infty$ ensures that any sequence with bounded energy has bounded norm. Once boundedness is established, the Weak Sequential [Compactness](/page/Compact%20Space) theorem (which requires [reflexivity](/page/Reflexive%20Space) of $X$) provides a subsequence $u_{n_k} \rightharpoonup u$ for some $u \in X$. **Step 2: Show that the limit is admissible.** If the problem involves a constraint $u \in K$ (e.g., boundary conditions, pointwise bounds, or other side conditions), one must verify that $K$ is weakly closed. By Mazur's Theorem, this holds whenever $K$ is norm-closed and convex — and most constraints arising in applications are of this type. **Step 3: Show that the functional does not increase at the limit.** This requires **weak lower semicontinuity**: $I(u) \leq \liminf_{n \to \infty} I(u_n)$. For convex, norm-lower-semicontinuous functionals, weak lower semicontinuity holds automatically by the theorem stated above. Combining with the fact that $I(u_n) \to \inf I$, we conclude $I(u) = \inf I$. [example: Existence of a Minimiser for the Dirichlet Energy] Let $U \subset \mathbb{R}^n$ be a bounded [open set](/page/Open%20Set) with $C^1$ boundary, let $g \in W^{1,2}(U)$ be a prescribed boundary datum, and define the admissible set \begin{align*} \mathcal{A} := \{v \in W^{1,2}(U) : v - g \in W_0^{1,2}(U)\}. \end{align*} Consider the Dirichlet energy functional $I: \mathcal{A} \to \mathbb{R}$ defined by \begin{align*} I(v) := \frac{1}{2} \int_U |\nabla v|^2 \, d\mathcal{L}^n. \end{align*} We prove that $I$ attains its infimum on $\mathcal{A}$. **Step 1 (Coercivity and boundedness).** Let $m := \inf_{v \in \mathcal{A}} I(v) \geq 0$, and let $\{u_n\} \subset \mathcal{A}$ with $I(u_n) \to m$. Write $w_n := u_n - g \in W_0^{1,2}(U)$. By the Poincare inequality for $W_0^{1,2}(U)$, \begin{align*} \|w_n\|_{L^2(U)} \leq C_P \|\nabla w_n\|_{L^2(U)} \leq C_P(\|\nabla u_n\|_{L^2(U)} + \|\nabla g\|_{L^2(U)}). \end{align*} Since $\|\nabla u_n\|_{L^2}^2 = 2I(u_n) \to 2m$, the sequence $\{w_n\}$ is bounded in $W^{1,2}(U)$, hence $\{u_n\}$ is bounded in $W^{1,2}(U)$. **Step 2 ([Weak convergence](/page/Weak%20Convergence)).** The space $W^{1,2}(U) = H^1(U)$ is a Hilbert space and hence reflexive. By the Weak Sequential Compactness theorem, there exists a subsequence (still denoted $\{u_n\}$) and $u \in W^{1,2}(U)$ with $u_n \rightharpoonup u$ in $W^{1,2}(U)$. **Step 3 (Admissibility of the limit).** We need $u - g \in W_0^{1,2}(U)$. Since $u_n - g = w_n \in W_0^{1,2}(U)$ for each $n$ and $w_n \rightharpoonup u - g$ in $W^{1,2}(U)$, we need $W_0^{1,2}(U)$ to be weakly closed. It is a closed subspace (hence closed and convex), so Mazur's Theorem gives weak closedness. Thus $u \in \mathcal{A}$. **Step 4 (Weak lower semicontinuity).** We claim $I(u) \leq \liminf I(u_n) = m$. Since $u_n \rightharpoonup u$ in $W^{1,2}(U)$, in particular $\nabla u_n \rightharpoonup \nabla u$ in $L^2(U; \mathbb{R}^n)$. By the lower semicontinuity of the norm under weak convergence, \begin{align*} \|\nabla u\|_{L^2(U)} \leq \liminf_{n \to \infty} \|\nabla u_n\|_{L^2(U)}. \end{align*} Squaring and multiplying by $1/2$ preserves the inequality (since $t \mapsto t^2/2$ is increasing on $[0, \infty)$): \begin{align*} I(u) = \frac{1}{2}\|\nabla u\|_{L^2}^2 \leq \frac{1}{2}\left(\liminf_{n \to \infty} \|\nabla u_n\|_{L^2}\right)^2 \leq \liminf_{n \to \infty} \frac{1}{2}\|\nabla u_n\|_{L^2}^2 = \liminf_{n \to \infty} I(u_n) = m. \end{align*} The second inequality uses the fact that $t \mapsto t^2$ is convex and increasing on $[0, \infty)$, so $(\liminf a_n)^2 \leq \liminf a_n^2$ for non-negative sequences. Since $I(u) \geq m$ by definition, we conclude $I(u) = m$. The minimiser exists. [/example] The three theorems used — Weak Sequential Compactness (from reflexivity), Mazur's Theorem (for admissibility), and Weak Lower Semicontinuity of the Norm (for the energy bound) — constitute the complete toolkit for the direct method. The same template applies to far more general functionals, including nonlinear elliptic problems, obstacle problems, and variational inequalities, as long as the functional is convex and coercive on a reflexive Banach space. ## Standard Arguments with the Weak Topology The following techniques appear repeatedly when working with weak convergence in analysis and PDE theory. Each addresses a specific difficulty that arises because weak convergence is strictly weaker than norm convergence. ### Testing Against a Dense Subset Verifying weak convergence $x_n \rightharpoonup x$ requires checking $f(x_n) \to f(x)$ for *every* $f \in X^*$. In practice, this is infeasible — $X^*$ is typically uncountable. The standard shortcut is to verify convergence on a *[dense subset](/page/Dense%20Subset)* of $X^*$, provided one has a uniform bound. [quotetheorem:1001] The argument is a standard $\varepsilon/3$ approximation. Fix $f \in X^*$ and $\varepsilon > 0$. Choose $g \in D$ with $\|f - g\|_{X^*} < \varepsilon/(3M)$, where $M := \sup_n \|x_n\|_X + \|x\|_X < \infty$. Then \begin{align*} |f(x_n) - f(x)| &\leq |f(x_n) - g(x_n)| + |g(x_n) - g(x)| + |g(x) - f(x)| \\ &\leq \|f - g\|_{X^*}(\|x_n\|_X + \|x\|_X) + |g(x_n) - g(x)| \\ &< \frac{2\varepsilon}{3} + |g(x_n) - g(x)|. \end{align*} Since $g(x_n) \to g(x)$, the last term is less than $\varepsilon/3$ for $n$ large enough. In $L^p(U)$ with $1 < p < \infty$, one typically tests against $C_c^\infty(U)$ (which is dense in $L^q(U)$). In $W^{1,p}(U)$, one can test separately against functionals that probe $u$ and $\nabla u$ in the $L^p$ pairing. ### The Weak-Times-Strong Lemma A frequent situation in PDE theory: one has $u_n \rightharpoonup u$ weakly in $L^p(U)$ and $v_n \to v$ strongly in $L^q(U)$ (where $q = p/(p-1)$ is the conjugate exponent), and one needs to pass to the limit in the product $\int_U u_n v_n \, d\mathcal{L}^n$. This is not immediate from weak convergence alone (the product of two [weakly convergent](/page/Weak%20Convergence) sequences need not converge to the product of the limits), but the strong convergence of one factor saves the argument. [quotetheorem:991] The proof is a decomposition: \begin{align*} \left|\int_U u_n v_n \, d\mathcal{L}^n - \int_U u v \, d\mathcal{L}^n\right| &\leq \left|\int_U u_n(v_n - v) \, d\mathcal{L}^n\right| + \left|\int_U (u_n - u) v \, d\mathcal{L}^n\right|. \end{align*} The first term is bounded by $\|u_n\|_{L^p} \|v_n - v\|_{L^q} \leq M \|v_n - v\|_{L^q} \to 0$ (using the boundedness of weakly convergent sequences and the strong convergence of $v_n$). The second term tends to zero because $u_n \rightharpoonup u$ and $v \in L^q(U)$ defines a [continuous](/page/Continuity) linear functional on $L^p(U)$. This lemma is used constantly when passing to the limit in weak formulations of PDEs: the [test function](/page/Test%20Function) $v$ typically converges strongly (or is fixed), while the solution $u_n$ converges only weakly. ### Extracting a Diagonal Subsequence When working with sequences of functions that are bounded in multiple Sobolev or $L^p$ norms simultaneously, one often needs to extract a subsequence that converges weakly in *all* of these spaces at once. The standard tool is Cantor's diagonal argument. Suppose $\{u_n\} \subset W^{1,p}(U)$ is bounded, with $1 < p < \infty$. By Weak Sequential [Compactness](/page/Compact%20Space), there exists a subsequence $\{u_{n_k^{(1)}}\}$ with $u_{n_k^{(1)}} \rightharpoonup u$ in $L^p(U)$. From this subsequence, extract a further subsequence $\{u_{n_k^{(2)}}\}$ with $\partial_{x_1} u_{n_k^{(2)}} \rightharpoonup v_1$ in $L^p(U)$. Continue for each partial derivative $\partial_{x_i}$, obtaining nested subsequences. The diagonal subsequence $u_{n_k^{(k)}}$ converges weakly in $L^p$ for $u$ and for each partial derivative simultaneously. By the closedness of the [weak derivative](/page/Weak%20Derivative), $v_i = \partial_{x_i} u$, and the full diagonal subsequence converges weakly in $W^{1,p}(U)$. This procedure is used at the start of virtually every PDE existence argument: given a bounded sequence in a [Sobolev space](/page/Sobolev%20Space), pass to a subsequence that converges weakly in the Sobolev space and (by Rellich-Kondrachov, if the domain is bounded) strongly in $L^p$. ### Identifying the Weak Limit via the Weak Formulation Once a weakly convergent subsequence $u_n \rightharpoonup u$ has been extracted, one must identify the limit $u$ as a solution of the PDE. The weak formulation provides the mechanism. If $u_n$ solves the equation $Lu_n = f_n$ in the weak sense, meaning \begin{align*} B[u_n, \phi] = f_n(\phi) \quad \text{for all } \phi \in W_0^{1,q}(U), \end{align*} where $q = p/(p-1)$ is the conjugate exponent to $p$ and $f_n \in (W_0^{1,q}(U))^*$, and if the bilinear form $B$ is continuous with respect to weak convergence in its first argument (this holds for linear operators with bounded coefficients), then passing to the limit gives $B[u, \phi] = f(\phi)$ for every test function $\phi$, where $f$ is the limit of $f_n$ in the appropriate [dual space](/page/Dual%20Space). The weak limit $u$ solves the limiting equation. For nonlinear problems, this step is more subtle: the nonlinearity may not be compatible with weak convergence, and techniques such as monotone operator theory, compensated compactness, or the theory of Young measures are required. The failure of weak continuity for nonlinear operations (e.g., $u_n \rightharpoonup u$ does not imply $u_n^2 \rightharpoonup u^2$) is one of the central difficulties in nonlinear PDE theory. ## Comparison of Topologies The three topologies — norm, weak, and weak* — are frequently confused, particularly because in [reflexive spaces](/page/Reflexive%20Space) two of them coincide. The following summary clarifies the distinctions. Let $X$ be a [Banach space](/page/Banach%20Space) with dual $X^*$ and bidual $X^{**}$. **On $X$:** - The **norm topology** $\tau_{\|\cdot\|}$ is the strongest: $x_n \to x$ iff $\|x_n - x\|_X \to 0$. - The **weak topology** $\sigma(X, X^*)$ is strictly coarser in infinite dimensions: $x_n \rightharpoonup x$ iff $f(x_n) \to f(x)$ for all $f \in X^*$. - There is no [weak* topology](/page/Weak*%20Topology) on $X$ unless $X$ is identified with a dual space. If $X = Y^*$ for some Banach space $Y$, then $X$ carries the weak* topology $\sigma(Y^*, Y)$, which is coarser than $\sigma(X, X^*)$. **On $X^*$:** - The norm topology $\tau_{\|\cdot\|_{X^*}}$. - The weak topology $\sigma(X^*, X^{**})$: convergence tested against all elements of $X^{**}$. - The weak* topology $\sigma(X^*, X)$: convergence tested only against the canonical image of $X$ in $X^{**}$. Since $X \subset X^{**}$ (via the canonical embedding), fewer functionals need to converge for weak* convergence than for [weak convergence](/page/Weak%20Convergence) on $X^*$. Hence: \begin{align*} \sigma(X^*, X) \subset \sigma(X^*, X^{**}) \subset \tau_{\|\cdot\|_{X^*}}. \end{align*} **When $X$ is reflexive**, $X = X^{**}$, so $\sigma(X^*, X) = \sigma(X^*, X^{**})$: the weak and weak* topologies on $X^*$ coincide. Moreover, viewing $X$ as $(X^*)^*$, the weak topology on $X$ is $\sigma(X, X^*) = \sigma((X^*)^*, X^*)$, which is the weak* topology on the bidual. Thus in the reflexive case, all the distinctions collapse, and the weak topology on $X$ is simultaneously the weak* topology on $X$ viewed as the dual of $X^*$. [remark: A Practical Diagnostic] When reading a PDE paper, the following heuristic helps identify which topology is at play: - If the argument uses the [Banach-Alaoglu theorem](/theorems/212), the convergence is weak* (on a dual space). - If the argument uses the Weak Sequential [Compactness](/page/Compact%20Space) theorem (requiring reflexivity), the convergence is weak (on the original space). - If the argument uses [Rellich-Kondrachov](/theorems/64), the convergence is strong (in a lower-order norm). In reflexive spaces these distinctions blur, but in non-reflexive spaces (like $L^1$, $L^\infty$, $C(\overline{U})$, spaces of measures) they are crucial. [/remark] ## Weak Convergence in Specific Spaces ### $L^p$ Spaces ($1 < p < \infty$) These are the best-behaved spaces for weak convergence. Since $L^p(U)$ is [reflexive](/page/Reflexive%20Space) for $1 < p < \infty$, every bounded sequence has a [weakly convergent](/page/Weak%20Convergence) subsequence. Weak convergence means \begin{align*} \int_U u_n g \, d\mathcal{L}^n \to \int_U u g \, d\mathcal{L}^n \quad \text{for every } g \in L^q(U), \end{align*} where $q = p/(p-1)$. A useful sufficient condition: $\{u_n\}$ is bounded in $L^p$ and $u_n(x) \to u(x)$ for a.e. $x \in U$ does *not* imply $u_n \rightharpoonup u$ in $L^p$ in general (consider $u_n = n \mathbb{1}_{(0,1/n)}$ in $L^1$), but if additionally $\{u_n\}$ is *uniformly integrable*, then the Dunford-Pettis theorem gives weak convergence. ### $L^1$ and $L^\infty$ The spaces $L^1(U)$ and $L^\infty(U)$ are not reflexive, and weak convergence behaves differently. In $L^1(U)$, the dual is $L^\infty(U)$, and weak convergence $u_n \rightharpoonup u$ means $\int_U u_n g \, d\mathcal{L}^n \to \int_U u g \, d\mathcal{L}^n$ for every $g \in L^\infty(U)$. The Dunford-Pettis theorem characterises weakly compact subsets of $L^1$: a bounded set $K \subset L^1(U)$ is relatively weakly compact if and only if $K$ is uniformly integrable. Since the unit ball of $L^1$ is not uniformly integrable (the sequence $n\mathbb{1}_{(0,1/n)}$ is bounded in $L^1$ but concentrates mass at the origin), $L^1$ is not reflexive, consistent with Kakutani's Theorem. In $L^\infty(U)$, the dual $(L^\infty)^*$ is much larger than $L^1(U)$: it includes finitely additive measures and other singular objects. Weak convergence in $L^\infty$ — convergence tested against *all* of $(L^\infty)^*$ — is rarely useful. What typically arises instead is weak* convergence of $L^\infty$ viewed as the dual of $L^1$: $u_n \overset{*}{\rightharpoonup} u$ means $\int_U u_n g \, d\mathcal{L}^n \to \int_U u g \, d\mathcal{L}^n$ for every $g \in L^1(U)$. This is strictly weaker than weak convergence in $L^\infty$, but the Banach-Alaoglu theorem applies to it, giving [compactness](/page/Compact%20Space) of bounded sets. ### Sobolev Spaces For $1 < p < \infty$, the [Sobolev space](/page/Sobolev%20Space) $W^{k,p}(U)$ is reflexive, and weak convergence $u_n \rightharpoonup u$ in $W^{k,p}(U)$ means \begin{align*} D^\alpha u_n \rightharpoonup D^\alpha u \quad \text{in } L^p(U) \quad \text{for every multi-index } \alpha \text{ with } |\alpha| \leq k. \end{align*} That is, weak convergence in the Sobolev space is equivalent to weak convergence of the function and all its [weak derivatives](/page/Weak%20Derivative) simultaneously. When $U$ is bounded with sufficient regularity, the [Rellich-Kondrachov Theorem](/theorems/64) upgrades weak convergence in $W^{1,p}(U)$ to *strong* convergence in $L^q(U)$ for $q < p^*$ (or in $C(\overline{U})$ if $p > n$). This combination — weak convergence in the Sobolev space plus strong convergence in a lower-order space — is the workhorse of PDE existence theory: the weak convergence preserves the derivative structure (needed for the equation), while the strong convergence in a lower norm handles nonlinear terms. ## References - H. Brezis, *Functional Analysis, [Sobolev Spaces](/page/Sobolev%20Space) and Partial Differential Equations* (2011). - L.C. Evans, *Partial Differential Equations*, 2nd ed. (2010). - J.B. Conway, *A Course in Functional Analysis*, 2nd ed. (1990). - W. Rudin, *Functional Analysis*, 2nd ed. (1991). - K. Yosida, *Functional Analysis*, 6th ed. (1980). - M. Reed and B. Simon, *Methods of Modern Mathematical Physics I: Functional Analysis* (1980).