An inner-product space already gives you everything geometric — length, angle, orthogonality, projection — for vectors, functions, whatever obeys the axioms. But there is one more property, invisible in finite dimensions and absolutely decisive in infinite ones, that separates a merely useful space from the pristine arena where analysis and quantum mechanics actually live. That property is completeness, and an inner-product space that has it earns a special name: a Hilbert space.

The one-line definition to keep in your pocket: a Hilbert space is an inner-product space with no holes — every sequence that ought to converge really does converge, to a limit that is still inside the space. That "no holes" clause is doing enormous work, and this page is about why. It is the difference between the rationals (full of gaps) and the real numbers (a seamless line), lifted up to spaces of infinitely many dimensions.

The problem with infinity

In finitely many dimensions life is easy: \mathbb{R}^n with the dot product is already complete, so there is nothing to worry about. The trouble starts when a space has infinitely many dimensions — which is precisely the situation for spaces of functions, where you might need an endless supply of independent building blocks (think of all the frequencies \sin x, \sin 2x, \sin 3x, \ldots).

In such a space you constantly build things as limits: a function is approximated by longer and longer sums, a state by better and better estimates. For that to be safe, the limit has to exist, and it has to still be a legal member of the space. If it can escape — if a sequence of perfectly good vectors marches steadily toward something that isn't there — your space has a hole, and the geometry springs a leak.

Completeness: no sequence left behind

The precise tool for "ought to converge" is the Cauchy sequence. A sequence v_1, v_2, v_3, \ldots is Cauchy if its terms get arbitrarily close to each other — for any tolerance \varepsilon > 0 there is a point beyond which every pair of terms is within \varepsilon:

The beauty of the Cauchy condition is that it only mentions the terms themselves — you can check it without knowing the limit in advance. A space is complete when this internal bunching-up is enough to guarantee an actual limit inside the space:

The rational-number analogy

You have already met an incomplete space and a complete one, back on the number line. The rationals \mathbb{Q} are riddled with holes. Consider the sequence

Every term is a perfectly good fraction, and the terms crowd together — it is Cauchy. Yet its limit is \sqrt 2, which is not rational. The sequence heads straight for a gap in \mathbb{Q} and falls through it. The real numbers \mathbb{R} are exactly the repair: you fill in every such limit, and the result is complete — the seamless line. A Hilbert space is this same completion idea, done for whole spaces of functions or states instead of single numbers. Where an inner-product space might have gaps, its Hilbert-space completion plugs them all.

Why quantum mechanics needs completeness

Here is the reason a physicist cares. In an infinite-dimensional space you build the object you want as a series — a limit of partial sums. A quantum state is written as an infinite superposition over an orthonormal basis; a wavefunction is reconstructed from its Fourier components term by term. Each partial sum is a legitimate vector, and the partial sums form a Cauchy sequence. Only in a complete space is the limit guaranteed to be a state that exists. Without completeness, you could write down a convergent-looking superposition whose "sum" is not actually a member of your space — and the whole edifice of expansions, measurements, and probabilities would rest on sand.

The slider below adds more and more terms of a Fourier series. Watch the wiggly partial sum S_n(x) home in on the smooth target curve. Each partial sum is a finite, harmless trigonometric polynomial; the terms bunch together (Cauchy); and because the relevant space is complete, the limit they close in on is a genuine member of it. That guarantee is what makes "sum up infinitely many pieces" a safe operation.

A faint target function and the partial sum of its Fourier series. Adding terms (slide right) pulls the partial sum ever closer to the target — a Cauchy sequence of vectors converging, safely, to a limit inside a complete (Hilbert) space.

The two headline examples

Almost every Hilbert space you meet is one of two prototypes, or is built from them.

1. Square-summable sequences, \ell^2. Its vectors are infinite lists of numbers (x_1, x_2, x_3, \ldots) whose squares add up to something finite, \sum_i |x_i|^2 < \infty. The inner product is the obvious infinite dot product \langle x, y\rangle = \sum_i \overline{x_i}\,y_i. This is the natural home of a state expressed by its coordinates in a countable orthonormal basis.

2. Square-integrable functions, L^2[a,b]. Its vectors are functions with \int_a^b |f(x)|^2\,dx < \infty (finite "energy"), and the inner product is the integral \langle f, g\rangle = \int_a^b \overline{f(x)}\,g(x)\,dx you already know. This is the home of the wavefunction \psi(x): the condition \int |\psi|^2\,dx = 1 that makes total probability equal to one is literally the statement that \psi is a unit vector in L^2.

Remarkably, these two are secretly the same space: choosing an orthonormal basis of L^2 (say the Fourier modes) turns each function into its list of coordinates, a member of \ell^2. Every separable infinite-dimensional Hilbert space is isomorphic to \ell^2 — there is, up to relabelling, only one.

Worked example — is this sequence Cauchy, and where does it live?

Take the sequence of vectors in \ell^2 that switches on one coordinate at a time with shrinking weight: v_n = \left(1, \tfrac12, \tfrac13, \ldots, \tfrac1n, 0, 0, \ldots\right). The gap between the m-th and n-th terms (m > n) is

Because \sum 1/k^2 converges (to \pi^2/6), that tail can be made as small as we like by taking n large — so the sequence is Cauchy. Since \ell^2 is complete, the limit exists in \ell^2: it is the full list \left(1, \tfrac12, \tfrac13, \ldots\right), whose squared norm \sum 1/k^2 = \pi^2/6 is finite, so it is a legal vector. No hole; the completion catches it. (Try the same construction with the harmonic weights 1/\sqrt{k} and the tail \sum 1/k diverges — the sequence is not Cauchy and never settles.)