Three Things That Stop Being True in Infinite Dimensions — Epoche C2
Three facts that a linear algebra course leaves a student holding — that a closed bounded set is compact, that every vector is a finite combination of basis vectors, and that every linear map is continuous — are all false in infinite dimensions, and functional analysis is largely the record of what replaced each of them. This essay sets out the three in increasing order of damage, where the criterion for damage is what has to be given up: the first failure costs a proof technique and returns a weaker substitute that still works; the second costs a convenience and returns a notion that may not exist at all; the third costs the definition of "operator" itself. The essay closes with what the third did to the definitions of quantum mechanics. 1. The closed unit ball is no longer compact The first failure is the mildest because a replacement is available, but it is worth seeing exactly how total the failure is before looking at the replacement. In $\mathbb{R}^n$ the Heine-Borel theorem says a set is compact if and only if it is closed and bounded, so every bounded sequence has a convergent subsequence. That single fact underwrites most existence proofs in finite dimensions, in a pattern with three steps: build a bounded sequence of approximate solutions, extract a convergent subsequence, and check that the limit solves the problem. The pattern breaks immediately in the sequence space $\ell^2$, the space of square-summable complex sequences with $\|x\|^2 = \sum_k |x_k|^2$. Let $e_n$ be the sequence with a $1$ in position $n$ and zeros elsewhere. Every $e_n$ lies in the closed unit ball, since $\|e_n\| = 1$. But for $n \neq m$ the vectors are orthogonal, so $\|e_n - e_m\|^2 = \|e_n\|^2 + \|e_m\|^2 = 2$, that is $\|e_n - e_m\| = \sqrt{2} \approx 1.414$. No subsequence can be Cauchy when every pair of distinct terms sits a fixed distance apart. The sequence is bounded and has no convergent subsequence whatsoever — not a subsequence that converges slowly, none at all. This is not an accident of $\ell^2$ or of orthogonality. Riesz proved in 1918 that the closed unit ball of a normed space is compact if and only if the space is finite-dimensional, so non-compactness is not a defect of a particular construction but the signature of infinite dimension. The proof of the hard direction runs on a lemma of his that replaces orthogonality in spaces that have no inner product: if $Y$ is a closed proper subspace of a normed space $X$ and $0 0$ — positive because $Y$ is closed — choosing $y_0 \in Y$ with $\|z - y_0\| \leq d/\theta$, and normalising $x = (z-y_0)/\|z-y_0\|$. Given the lemma, one builds a sequence in the unit sphere with mutual distances at least $1/2$: take $x_1$ of norm one, and having chosen $x_1,\dots,x_k$, apply the lemma to the span of those vectors, which is finite-dimensional and hence closed, and proper because the space is infinite-dimensional. That sequence has no Cauchy subsequence, so the ball is not compact. The repair is to weaken the topology until the ball becomes compact in it, at the price of a weaker notion of convergence. The same sequence $e_n$ converges weakly to zero — meaning $f(e_n) \to 0$ for every bounded linear functional $f$ — because in a Hilbert space every such functional is $\langle\,\cdot\,,x\rangle$ for some $x$, and Bessel's inequality $\sum_n |\langle e_n,x\rangle|^2 \leq \|x\|^2$ forces the terms of a convergent series to tend to zero. Note what this costs: $\|e_n\| = 1$ for every $n$ while the weak limit has norm $0$, so the norm is not weakly continuous. It is only weakly lower semicontinuous, and that inequality — the norm of the limit is at most the liminf of the norms — is what variational arguments actually use when they conclude that a weak limit is a minimiser. Two theorems make the substitute usable. The Banach-Alaoglu theorem says the closed unit ball of a dual space $X^*$ is compact in the weak-* topology, the topology of pointwise convergence on $X$; the proof embeds the ball in a product of closed discs, one for each $x \in X$ with radius $\|x\|$, and applies Tychonoff's theorem to that product. The Eberlein-Smulian theorem then says that for the weak topology on a Banach space, compactness and sequential compactness agree — which is what licenses "extract a weakly convergent subsequence" even though the weak topology is not metrisable in general. Whether the ball of the space itself, rather than of its dual, is weakly compact is a further question with a clean answer: exactly when the space is reflexive, which is why reflexivity is a hypothesis in so many existence theorems. Modern existence proofs for partial differential equations run on these substitutes together with compact embedding theorems, of which Rellich-Kondrachov is the standard one: if $\Omega \subset \mathbb{R}^n$ is bounded with Lipschitz boundary and $1 \leq p 2. A basis exists, but not the kind you wanted The second failure is worse than the first because what replaces the finite-dimensional notion is not guaranteed to exist. Every vector space has a Hamel basis, a set such that every vector is a finite linear combination of its members in exactly one way; this follows from Zorn's lemma applied to the partially ordered set of linearly independent subsets. So the existence statement survives. What does not survive is any hope of using it. In an infinite-dimensional Banach space a Hamel basis is never countable, and the argument is a two-line application of the Baire category theorem. Suppose $\{x_1,x_2,\dots\}$ were a countable Hamel basis, and let $E_n$ be the span of the first $n$. Every vector is a finite combination, so $X = \bigcup_n E_n$. Each $E_n$ is finite-dimensional, hence closed, and each is a proper subspace, since $X$ is infinite-dimensional. A proper closed subspace has empty interior — if it contained a ball about some point it would contain a ball about the origin by translation, and then all of $X$ by scaling — so each $E_n$ is nowhere dense. A complete metri