Hilbert Spaces: Geometry of Infinite Dimensions

Build a geometric mental model of Hilbert spaces that actually predicts what works. You will see why angles and projections survive in infinite dimensions, why completeness is nonnegotiable, and how this geometry drives bases, operators, and spectra.

Angles and projections do not belong to finite-dimensional linear algebra. They belong to Hilbert spaces, and that is why so many problems in analysis and physics suddenly become clean when you move into $L^2$ or $\ell^2$. The counterintuitive part is that infinite dimension is not the scary feature. The scary feature is losing the inner product or losing completeness. Keep those two, and you get best approximations, orthogonal decompositions, and a version of diagonalization that is still geometric, even when coordinates are no longer the point.

The projection picture is the first stress test. In a Hilbert space, a closed subspace behaves like a flat plane in $\mathbb{R}^n$. Given a point $x$, there is a unique nearest point in that subspace, call it $P_Mx$, and the error $x-P_Mx$ is orthogonal to the subspace. The Pythagorean identity survives verbatim, $\|x\|^2=\|P_Mx\|^2+\|x-P_Mx\|^2$, and that identity is not decoration. It is the mechanism that turns geometry into optimization, and it is exactly what breaks in inner-product spaces that are not complete or in normed spaces that are not Hilbert.

Inner product geometry is a machine

A Hilbert space starts with an inner product $\langle\cdot,\cdot\rangle$ and ends with a full geometry. The induced norm $\|x\|=\sqrt{\langle x,x\rangle}$ gives a distance, and distance gives convergence. The key is that the inner product is not recoverable from an arbitrary norm, but when it exists it controls far more than lengths.

The two identities that make this machine run are:

• Cauchy–Schwarz: $|\langle x,y\rangle|\le\|x\|\|y\|$, which makes $\langle x,\cdot\rangle$ automatically continuous.
• Polarization identity: it reconstructs $\langle x,y\rangle$ from norms of combinations of $x$ and $y$, so inner products and norms are tightly coupled in Hilbert settings.

Orthogonality is where the payoff lands. Saying $x\perp M$ is not just a geometric slogan; it means every perturbation within $M$ increases squared distance. That is why orthogonal projections are linear and contractive, and why normal equations in least squares are really orthogonality conditions in disguise.

Orthogonality pays
In Hilbert spaces, $x-P_Mx\perp M$ is the optimality certificate for best approximation. You do not need coordinates, only the inner product.

Completeness is the load-bearing beam

A pre-Hilbert space has an inner product; a Hilbert space is a pre-Hilbert space that is complete in the induced norm. That extra word complete is what upgrades geometry from local to global.

Featured-snippet version: A Hilbert space is a complete inner-product space. Completeness means every Cauchy sequence converges to a limit inside the space. Without it, best-approximation problems can have infima that are not attained, and orthogonal projection onto closed subspaces can fail to exist.

If you have worked with smooth or compactly supported functions, you have seen the pathology. Many natural inner-product spaces of nice functions are incomplete under the $\|\cdot\|_2$ norm, so sequences that should converge in energy converge to objects that are no longer nice. The Riesz–Fischer theorem viewpoint is that $L^2$ is what you get when you accept all those limits. The completion is not a technicality; it is the smallest setting where Fourier series, projections, and variational arguments stop leaking out of the space.

Riesz representation turns duality into geometry

The dual space is where a lot of infinite-dimensional fear lives. In a general normed space, continuous linear functionals are mysterious and plentiful. In a Hilbert space, the Riesz representation theorem says they are all inner products with a fixed vector.

Concretely, for every continuous linear functional $f$ on a Hilbert space $H$, there exists a unique $y\in H$ such that

$$f(x)=\langle x,y\rangle\quad\text{for all }x\in H.$$

This identifies $H^*$ with $H$ itself, up to the conjugate-linearity convention in the complex case.

That statement is geometric if you read $f(x)=0$ as a hyperplane through the origin. The representing vector $y$ is the normal direction. Distance-to-hyperplane formulas, orthogonal decompositions $H=\ker f\oplus \text{span}\{y\}$, and projection arguments all drop out of the same picture. The theorem is also why adjoints and self-adjointness feel natural in Hilbert spaces. The inner product gives you a canonical way to move linear information from the output side back to the input side.

Orthonormal bases are about error control

An orthonormal basis in a Hilbert space is not mainly a coordinate system. It is a guarantee that expanding in basis elements is the right way to approximate, because orthogonality makes errors add in squares.

Given an orthonormal set $(e_n)$ and coefficients $\langle x,e_n\rangle$, the partial sums

$$s_N=\sum_{n=1}^N\langle x,e_n\rangle e_n$$

are the best approximations to $x$ among all vectors in $\text{span}\{e_1,\dots,e_N\}$. That is the same projection principle from earlier, just iterated along a nested chain of subspaces.

Two bookkeeping identities are the real reason to care:

• Bessel inequality: $\sum_n|\langle x,e_n\rangle|^2\le\|x\|^2$ for any orthonormal set.
• Parseval or Plancherel: equality holds when the set is complete, so $\|x\|^2$ is literally the energy in coefficients.

Separability matters here because it tells you whether countable approximations can be dense. In separable Hilbert spaces, countable orthonormal bases exist, and many analytic arguments become controlled limits of finite-dimensional ones.

Energy view
Parseval is not about Fourier specifically. It is a statement that orthonormal expansions conserve $\|x\|^2$, so approximation error is measurable and monotone.

Operators and the discipline of adjoints

In Hilbert space, linear operators are not just maps; they carry geometry. The organizing concept is the adjoint $T^*$ of a bounded linear operator $T$, defined by

$$\langle Tx,y\rangle=\langle x,T^*y\rangle.$$

This is Riesz representation in action, applied pointwise in $y$.

Several operator classes are really statements about how $T$ interacts with the inner product:

• Self-adjoint: $T=T^*$, the infinite-dimensional analog of symmetric or Hermitian matrices.
• Unitary: $T^*T=TT^*=I$, meaning $T$ preserves inner products and norms.
• Normal: $T^*T=TT^*$, the condition that makes spectral theory behave well.

Boundedness is the quiet constraint that keeps the basic calculus legal. Unbounded operators show up naturally, but then the domain becomes part of the data and naive algebraic manipulations can fail. Starting with bounded operators is not avoidance, it is learning what the geometry wants before domain issues enter.

Spectral theorem as geometric diagonalization

Finite-dimensional diagonalization says you can pick a basis of eigenvectors, and the operator becomes multiplication by scalars. In infinite dimensions, that exact picture holds for some operators, but not all. The spectral theorem is the replacement, and its different forms tell you what diagonalization means in each regime.

For compact self-adjoint operators, the story is closest to matrices. You get an orthonormal basis of eigenvectors for the orthogonal complement of the kernel, real eigenvalues with finite multiplicity, and eigenvalues tending to $0$. The operator is a norm-limit of finite-rank diagonal pieces, a precise version of the idea that compact operators are almost finite-dimensional.

For general bounded self-adjoint operators, you should stop expecting a basis of eigenvectors. The right diagonal object is a projection-valued measure, and the operator becomes an integral

$$A=\int\lambda\,E(d\lambda).$$

That is diagonalization with continuous spectrum allowed. Multiplication operators on $L^2$ are the canonical model. They explain why spectra can be intervals and why eigenvectors can be replaced by generalized modes.

Where the geometry gets used

Least squares is projection onto a model subspace, and the normal equations encode orthogonality to the residual. Fourier methods are orthonormal expansions, where convergence in $L^2$ is the statement that energy controls error even when pointwise behavior is messy. Quantum states live in Hilbert spaces because probabilities are squared norms, global phase is irrelevant, and observables are modeled by self-adjoint operators whose spectra are the possible measurement outcomes.

The common thread is not infinite dimension. It is that the inner product lets you turn analytic questions into geometric ones, and completeness ensures the geometry has actual minimizers and limits instead of mirages.

Pitfalls and expert heuristics

The fastest way to go wrong is to treat every normed space as if it were Hilbert. If your norm does not come from an inner product, orthogonality may be undefined or useless, projections onto closed subspaces can fail, and best approximations may not be unique. In Banach spaces, you can still do a lot with convexity, duality, and weak topology, but the arguments feel different because you lose Pythagoras and the adjoint machinery.

A second trap is forgetting the closedness condition. In Hilbert spaces, orthogonal projection exists onto closed subspaces, not arbitrary subspaces. Many natural subspaces in analysis are dense but not closed, and the projection you want lands in the closure. That distinction is exactly the point of completions like $L^2$.

Third, do not assume eigenvectors exist. Compactness is the friendly hypothesis that forces eigenstructure. Drop it, and continuous spectrum becomes normal, not exceptional.

FAQs for Hilbert spaces