Functional Analysis for Quantum Mechanics: What You Need to Know

Functional analysis is the mathematical language behind the standard formulation of quantum mechanics. It explains what a quantum state is, why observables are represented by operators, how measurements produce probabilities, and why time evolution preserves total probability.

You do not need every topic in a graduate functional-analysis course. The most useful path is to learn the theory together with the quantum examples that motivate it.

1. Hilbert spaces: the space of states

A Hilbert space is a complete inner-product space. The inner product of vectors $\psi$ and $\phi$ is written $\langle \psi,\phi\rangle$, and it defines the norm

$$\|\psi\|=\sqrt{\langle \psi,\psi\rangle}.$$

Completeness means that every Cauchy sequence of vectors converges to a vector in the space. This condition matters because quantum calculations commonly involve limits, infinite series, and approximations.

A pure quantum state is represented by a normalized vector $\psi$ in a Hilbert space $\mathcal{H}$:

$$\|\psi\|=1.$$

Strictly, the physical state is a ray: multiplying $\psi$ by a complex number of absolute value one does not change any prediction. Thus $\psi$ and $e^{i\theta}\psi$ represent the same physical pure state.

For a particle moving on a line, the usual Hilbert space is

$$\mathcal{H}=L^2(\mathbb{R}),$$

the space of square-integrable complex functions. Its inner product is commonly

$$\langle \psi,\phi\rangle=\int_{-\infty}^{\infty}\overline{\psi(x)}\phi(x)\,dx.$$

The integral of $|\psi(x)|^2$ is finite, and for a normalized wavefunction it is one. The quantity $|\psi(x)|^2$ gives the position probability density in the position representation.

2. Operators and their domains

An operator $A$ maps vectors in a domain $D(A)\subseteq\mathcal{H}$ to vectors in $\mathcal{H}$. The domain is part of the operator; writing only a formula for $A$ can be incomplete.

A bounded operator satisfies

$$\|A\psi\|\leq C\|\psi\|$$

for some constant $C$ and every $\psi\in\mathcal{H}$. Bounded operators are defined on the whole Hilbert space and are comparatively well behaved.

Many important quantum observables are unbounded. For a particle in one dimension,

$$(X\psi)(x)=x\psi(x),$$

and, in the position representation,

$$(P\psi)(x)=-i\hbar\frac{d\psi}{dx}(x).$$

The position operator is not defined on every element of $L^2(\mathbb{R})$: $x\psi(x)$ must itself be square-integrable. Similarly, the momentum operator requires a suitable differentiability condition. A convenient initial domain for momentum is the Schwartz space of smooth, rapidly decreasing functions, although its full self-adjoint domain is larger.

This domain issue is one of the main reasons elementary statements such as $[X,P]=i\hbar I$ need qualification. The commutator is meaningful only for vectors on which all the relevant products are defined.

3. Symmetric versus self-adjoint

For a densely defined operator $A$, symmetry means

$$\langle \psi,A\phi\rangle=\langle A\psi,\phi\rangle$$

for all $\psi$ and $\phi$ in its domain. A self-adjoint operator satisfies the stronger condition $A=A^*$, including equality of their domains.

This distinction is especially important for unbounded operators. A symmetric operator need not have the mathematically complete properties required of an observable. In the standard formulation, observables correspond to self-adjoint operators, not merely symmetric ones.

Self-adjointness guarantees the appropriate spectral decomposition and ensures that the unitary operator generated by $A$ has the expected behavior. For example, a self-adjoint Hamiltonian $H$ generates time evolution through

$$U(t)=e^{-iHt/\hbar},$$

where $U(t)$ is unitary. Therefore

$$\|U(t)\psi\|=\|\psi\|,$$

so total probability is preserved.

4. The spectral theorem and measurement

The finite-dimensional version of the spectral theorem says that a Hermitian matrix can be diagonalized using an orthonormal basis. If

$$A=\sum_n a_n P_n,$$

where $P_n$ projects onto the eigenspace with eigenvalue $a_n$, then the probability of obtaining $a_n$ in a normalized state $\psi$ is

$$\Pr(a_n)=\langle \psi,P_n\psi\rangle.$$

The functional-analytic spectral theorem extends this idea to infinite-dimensional spaces and continuous spectra. For a self-adjoint operator $A$, instead of a sum over eigenvalues, one uses a projection-valued measure $E$:

$$A=\int_{\mathbb{R}} \lambda\,dE(\lambda).$$

For an unbounded $A$, this identity is understood in the spectral-calculus sense: the integral defines $A$ on the domain

$$D(A)=\left\{\psi\in\mathcal{H}:\int_{\mathbb{R}}\lambda^2\,d\langle\psi,E(\lambda)\psi\rangle<\infty\right\}.$$

For a state $\psi$, the probability that a measurement result lies in a set $B$ is

$$\Pr(A\in B)=\langle \psi,E(B)\psi\rangle.$$

This formulation handles both discrete and continuous observables. Position, for example, does not have normalizable eigenvectors in $L^2(\mathbb{R})$, so treating position measurements as an ordinary countable sum of eigenstates is only a formal shorthand.

5. A worked example: a particle in a box

Consider a particle confined to $0<x<L$. The Hilbert space is $L^2([0,L])$, and the Hamiltonian is formally

$$H=-\frac{\hbar^2}{2m}\frac{d^2}{dx^2}.$$

For an infinite potential wall, the boundary conditions are

$$\psi(0)=\psi(L)=0.$$

The corresponding self-adjoint operator has domain

$$D(H)=\left\{\psi\in H^2([0,L]):\psi(0)=\psi(L)=0\right\},$$

where $H^2([0,L])$ denotes the Sobolev space of square-integrable functions with square-integrable first and second weak derivatives. Solving the eigenvalue equation $H\psi=E\psi$ gives normalized energy eigenfunctions

$$\phi_n(x)=\sqrt{\frac{2}{L}}\sin\left(\frac{n\pi x}{L}\right), \qquad n=1,2,3,\ldots$$

with energies

$$E_n=\frac{n^2\pi^2\hbar^2}{2mL^2}.$$

The functions $\{\phi_n\}$ form an orthonormal basis for the relevant Hilbert space. If

$$\psi(x)=\sum_{n=1}^{\infty}c_n\phi_n(x),$$

then normalization means $\sum_n|c_n|^2=1$, and an energy measurement gives $E_n$ with probability $|c_n|^2$.

For example, the state

$$\psi=\frac{1}{\sqrt{2}}\phi_1+\frac{1}{\sqrt{2}}\phi_2$$

produces $E_1$ and $E_2$ with probability $1/2$ each. Its time evolution is

$$\psi(t)=\frac{1}{\sqrt{2}}e^{-iE_1t/\hbar}\phi_1+\frac{1}{\sqrt{2}}e^{-iE_2t/\hbar}\phi_2.$$

The coefficients retain their absolute values, illustrating unitary evolution, while their relative phase changes with time.

6. A practical learning order

A useful sequence is:

1. Review complex vector spaces, inner products, orthonormal bases, and Fourier series.
2. Learn normed spaces, Banach spaces, bounded operators, adjoints, and compact operators.
3. Study Hilbert-space projections and the Riesz representation theorem.
4. Learn unbounded operators, dense domains, closures, adjoints, and self-adjoint extensions.
5. Study the spectral theorem and functional calculus.
6. Apply the theory to position, momentum, angular momentum, and Hamiltonians.
7. Add tensor products for composite systems and rigged Hilbert spaces when distributions such as idealized position eigenstates are useful.

The rigged Hilbert space, or Gelfand triple, is often written

$$\Phi\subset\mathcal{H}\subset\Phi^*.$$

It provides a controlled setting for generalized eigenvectors, but it is not necessary for learning the basic postulates of quantum mechanics.

Common mistakes

• Treating every observable as a matrix: infinite-dimensional observables may be unbounded differential operators.
• Ignoring domains: the expression $-i\hbar\,d/dx$ does not specify a complete momentum operator without a domain.
• Confusing symmetric with self-adjoint: symmetry alone may not give a valid observable.
• Assuming every measurement has eigenvectors in the Hilbert space: continuous spectra require spectral measures and generalized eigenvectors.
• Starting with abstract theorems only: concrete examples such as Fourier transforms and the particle in a box make the definitions easier to retain.

For most physics applications, a strong working knowledge of Hilbert spaces, operators, self-adjointness, spectral theory, and unitary evolution is the essential core. More advanced topics—operator algebras, distributions, and rigged Hilbert spaces—can be added when a particular problem requires them.