You have already met Hilbert spaces as inner-product spaces where lengths and angles make sense. Now we return to them from the functional-analysis viewpoint, sitting one rung above Banach spaces in the hierarchy of spaces. The question that organises this whole page is sharp and worth holding in mind:

Of all the complete normed spaces, what makes a Hilbert space so much better behaved?

The one-word answer is geometry. A Banach space knows how long its vectors are; a Hilbert space also knows the angle between any two of them, and in particular when two vectors are at right angles. That single extra structure — an inner product — buys three theorems that have no counterpart in a general Banach space, and those three theorems are the backbone of least-squares fitting, Fourier analysis, quantum mechanics and signal processing.

A real-world hook to keep in view. A noisy signal or a data cloud is a vector x in some enormous space; the "clean" signals you are willing to accept form a subspace M. The engineer's daily task — find the best approximation of x inside M — is, word for word, the problem of dropping a perpendicular from a point to a plane. In a Hilbert space that perpendicular always exists, is unique, and is computable coefficient by coefficient. That is the payoff, and the rest of the page earns it.

The definition, and the norm behind it

Let H be a vector space over \mathbb{K} (read \mathbb{R} or \mathbb{C}) carrying an inner product \langle \cdot, \cdot \rangle — linear in the first slot, conjugate-symmetric, and positive-definite. Every inner product manufactures a norm,

and a norm manufactures the distance d(x, y) = \|x - y\|. So an inner-product space is automatically a normed space, hence a metric space. The last ingredient is completeness: no Cauchy sequence is allowed to escape.

This raises an obvious question: given a Banach space, can you tell whether its norm secretly came from an inner product? Remarkably, yes — and by a single identity you can check with two vectors.

Expand \|x \pm y\|^2 = \langle x \pm y, x \pm y\rangle = \|x\|^2 \pm 2\,\mathrm{Re}\langle x, y\rangle + \|y\|^2 and add the two versions. The cross terms cancel, leaving the parallelogram law:

Geometrically it says the two diagonals of a parallelogram determine its four sides — a fact that holds for genuine Euclidean lengths and, it turns out, only for them. The Jordan–von Neumann theorem makes this a perfect characterisation: a norm satisfies the parallelogram law if and only if it arises from an inner product, in which case the product is recovered from the norm by the polarisation identity (real case)

So "is this Banach space a Hilbert space?" is settled by a one-line test on the norm. Fail the parallelogram law at even a single pair (x, y), and no inner product can exist.

Theorem 1 — orthogonal projection: drop a perpendicular

This is the theorem the real-world hook was pointing at. In a general metric space there is no reason a set should contain a nearest point to a given x — the infimum of the distances need not be attained, and even if it is it might be attained twice. In a Hilbert space, convexity plus completeness rescue both existence and uniqueness.

Why is the nearest point unique? Suppose p and q both realise the minimal distance \delta. Their midpoint \tfrac{1}{2}(p + q) lies in K by convexity, so it is no closer than \delta either. Now feed x - p and x - q into the parallelogram law:

So \|p - q\| = 0, i.e. p = q. Notice what did the work: the parallelogram law — the very identity that separates Hilbert spaces from Banach spaces. Uniqueness of best approximation is a Hilbert-space privilege, and it is exactly the least-squares "normal equations" x - p \perp M in disguise.

Drag the sliders below. The blue line is a subspace M through the origin; x is the point being approximated; p = P_M x is its shadow on M; and the dashed residual x - p always meets M at a right angle — that right angle is the reason p is the closest point.

Theorem 2 — orthonormal bases and generalised Fourier series

Projection onto a line spanned by a unit vector e is a one-liner: the shadow is \langle x, e\rangle\,e, and the scalar \langle x, e\rangle is the coordinate of x along e. An orthonormal system \{e_n\} — mutually perpendicular unit vectors, \langle e_m, e_n\rangle = \delta_{mn} — lets us do this in every direction at once and assemble the projections into a series.

The numbers \hat{x}_n = \langle x, e_n\rangle are the generalised Fourier coefficients of x. The partial sum s_N = \sum_{n=1}^{N}\langle x, e_n\rangle\,e_n is precisely the orthogonal projection of x onto \mathrm{span}\{e_1, \dots, e_N\} — the best approximation of x by those N directions, by Theorem 1. Because x - s_N \perp s_N, Pythagoras gives \|x\|^2 = \|s_N\|^2 + \|x - s_N\|^2 \ge \|s_N\|^2 = \sum_{n=1}^N |\langle x, e_n\rangle|^2, and letting N \to \infty yields Bessel's inequality.

The classical Fourier series is the special case H = L^2[-\pi, \pi] with the orthonormal system \big\{\tfrac{1}{\sqrt{2\pi}}e^{inx}\big\}_{n \in \mathbb{Z}}. Parseval's identity then reads \int_{-\pi}^{\pi} |f|^2 = 2\pi\sum_n |c_n|^2 — the statement that a function and its spectrum carry the same total energy, the mathematical heart of every equaliser, MP3 encoder and spectrum analyser. Abstractly, Parseval says a separable Hilbert space is isometrically the coordinate space \ell^2: choose a basis and every vector becomes its sequence of coefficients.

Theorem 3 — Riesz representation: a Hilbert space is its own dual

A bounded linear functional is a continuous linear "measurement" f : H \to \mathbb{K}; together they form the dual space H^{*}. In a Hilbert space every such measurement has a startlingly concrete form: it is just taking the inner product with one fixed vector.

The vector y is built by Theorem 1. If f = 0 take y = 0. Otherwise the kernel M = \ker f is a closed subspace with M \ne H, so M^{\perp} contains a unit vector e — that is where completeness and the projection theorem enter — and one checks that y = \overline{f(e)}\,e does the job. Uniqueness is immediate: if \langle x, y_1\rangle = \langle x, y_2\rangle for all x, put x = y_1 - y_2 to get \|y_1 - y_2\|^2 = 0.

This is a striking contrast with the wider world of Banach spaces, where the dual is usually a different space (the dual of \ell^1 is \ell^\infty, and the dual of \ell^\infty is bigger still). Only in a Hilbert space does H^{*} fold back onto H itself. Self-duality is what lets us define adjoint operators \langle Tx, y\rangle = \langle x, T^{*}y\rangle, and it is the reason the bra \langle \psi| and the ket |\psi\rangle of quantum mechanics are two faces of the same state.

The three theorems, in one breath

Everything above rests on the single fact that a Hilbert-space norm knows angles. From it:

• Projection — closed convex sets have unique nearest points, and H = M \oplus M^{\perp}. (Best approximation, least squares.)
• Fourier series — x = \sum_n \langle x, e_n\rangle e_n with Bessel's inequality \sum|\hat{x}_n|^2 \le \|x\|^2 and, for a basis, Parseval \|x\|^2 = \sum|\hat{x}_n|^2. (Signal decomposition.)
• Riesz — f(x) = \langle x, y\rangle, \|f\| = \|y\|, so H^{*} \cong H. (Self-duality, adjoints, quantum bra–ket.)

A Banach space gives you none of these for free. That gap — closed by the parallelogram law — is the whole reason Hilbert spaces sit in a class of their own.