Weak and Weak Star Topologies

Weak convergence watches every continuous readout instead of the distance. Spikes can drift to zero weakly while standing a full unit away in norm.

Four unit spikes stand at full height while their weak limit sits at zero, showing convergence without norm convergence

Why Does This Exist?

Gradient flows, Galerkin approximations and Bayesian posteriors often refuse to converge in norm yet still deliver correct answers for every measurement that matters. Demanding norm convergence then rejects perfectly good algorithms. Weaker topologies salvage them: they declare convergence when every continuous readout converges, which is exactly what downstream decisions consume.

This page defines the weak and weak star topologies and exhibits the canonical non-example (spikes with no norm limit). The readouts doing the testing live in dual spaces; the complete rooms are Banach spaces.

The one idea you need first

A topology decides which sequences converge by declaring open sets. Finer means more opens, harder convergence; coarser means fewer, easier. The norm topology tests distance $\|x_n - x\| \to 0$ . Weak topologies test readouts instead.

Think of It Like This

Election polls versus a census

A census measures every resident (norm convergence): expensive and decisive. Polls ask selected questions of samples (weak convergence): cheap, and enough when every question that matters stabilizes.

A population can poll-converge while the census never settles: sampling noise keeps full counts dancing even as all major-question margins freeze. Where it stops: polls miss what nobody asks. Weak limits can sit outside expected sets unless convexity or closedness restores them (Mazur's lemma territory).

How It Actually Works

The weak topology $\sigma(X, X^\ast)$ on $X$ is the coarsest making every $f \in X^\ast$ continuous: $x_n \rightharpoonup x$ means $f(x_n) \to f(x)$ for all continuous functionals. In plain words: test every readout, demand each converge. The weak star topology $\sigma(X^\ast, X)$ on the dual goes further: $f_n \stackrel{\ast}{\rightharpoonup} f$ means $f_n(x) \to f(x)$ per point $x \in X$ , testing dual elements only against primal points.

Payoffs: Banach-Alaoglu says the dual unit ball is weak star compact (bounded sequences always have weak star cluster points: existence of limits for free). Mazur's lemma says convex combinations of weakly convergent sequences converge in norm (weak plus convexity recovers strength). Reflexive spaces (Hilbert, $L^p$ for $1 < p < \infty$ ) get weak compactness of their own balls.

Worked example

In $\ell^2$ , basis spikes $e_n$ (1 in slot $n$ , 0 elsewhere). Norm: $\|e_n\|_2 = 1$ forever, so no norm convergence to $0$ (distance stays $1$ ). Weak: for $x = (x_k)$ with $\sum x_k^2 < \infty$ , the readout is $\langle e_n, x \rangle = x_n$ . Square-summability forces $x_n \to 0$ (tails of convergent series vanish). Hence every readout tends to $0$ : $e_n \rightharpoonup 0$ weakly while $\|e_n\| = 1$ .

Concretely with $x = (1, 1/2, 1/3, 1/4, \ldots)/C$ : pairings $1/C, 1/2C, 1/3C, \ldots \to 0$ . Readouts die; lengths persist.

Code

Pairings of spikes against a square-summable vector decay to zero while norms stay one.

1import numpy as np2
3
4def pairing(n: int, terms: int = 200) -> tuple[float, float]:5    x = 1.0 / np.arange(1, terms + 1)  # harmonic tail, square summable6    e = np.zeros(terms)7    e[n] = 1.0  # n-th basis spike8    return float(e @ x), float(np.linalg.norm(e))  # readout, norm9
10
11print(pairing(0))  # -> (1.0, 1.0)12print(pairing(99))  # -> (0.01, 1.0)

Readout falls $1.0 \to 0.01$ while norm stays $1.0$ : weak convergence without norm convergence.

Watch Out For

Upgrading weak to norm without convexity

Symptom: weakly convergent iterates assumed close in distance; projections and stopping rules misfire. Fix: take convex combinations (Mazur) or verify norm convergence of lengths too (Radon-Riesz needs $\|x_n\| \to \|x\|$ plus weak for norm in Hilbert).

Expecting weak limits to respect nonlinearity

Symptom: $x_n \rightharpoonup x$ yet $\|x_n\|^2 \not\to \|x\|^2$ or $F(x_n) \not\to F(x)$ for nonlinear $F$ . Fix: pass only linear continuous maps through weak limits freely; nonlinear ones need strong convergence or compensated-compactness structure.

The Quick Version

Weak convergence tests all dual readouts; weak star tests dual elements pointwise on primal points.
Basis spikes $e_n$ tend weakly to $0$ in $\ell^2$ while every norm stays $1$ .
Banach-Alaoglu gives weak star compactness of dual balls: bounded sequences cluster.
Mazur upgrades weak to norm through convex combinations.
Linear maps pass through weak limits; nonlinear ones generally do not.