Finite-Dimensional Hilbert Spaces

A finite-dimensional Hilbert space is a finite-dimensional complex vector space equipped with an inner product. Completeness, the property that distinguishes Hilbert spaces from incomplete inner-product spaces in general, is automatic in finite dimension.

Why Quantum Mechanics Needs This

Qubits, qudits, spin multiplets, finite-level effective systems, finite registers, and many numerical approximations use finite-dimensional Hilbert spaces. They provide the cleanest setting in which to learn state vectors, orthogonal alternatives, Hermitian operators, projectors, spectral decompositions, and unitary evolution.

Finite dimension removes several analytic complications:

• every linear operator is bounded and continuous;
• every operator can be defined on the whole space;
• every subspace is closed;
• every Cauchy sequence converges;
• spectra contain only finitely many eigenvalues;
• sums replace convergence-sensitive spectral integrals.

It does not supply the physical postulates. Linear algebra describes the available structures; the Finite-Dimensional Postulates state how quantum theory uses them.

Definition

A Hilbert space is a complete inner-product space. Thus a finite-dimensional complex Hilbert space consists of:

1. a complex vector space $\mathcal H$ of finite dimension $d$;
2. an inner product $\langle\cdot\vert\cdot\rangle$;
3. completeness in the induced norm

$$\lVert\psi\rVert = \sqrt{\langle\psi\vert\psi\rangle}.$$

The model example is

$$\mathcal H=\mathbb C^d$$

with the standard inner product

$$\langle v\vert w\rangle = \sum_{k=1}^{d}v_k^*w_k.$$

Other positive-definite inner products on $\mathbb C^d$ are possible. After an orthonormal basis is chosen, however, every finite-dimensional Hilbert space has this standard coordinate form.

Why Completeness Is Automatic

Let $(\psi_m)$ be a Cauchy sequence in a $d$-dimensional inner-product space, and choose an orthonormal basis $(e_1,\ldots,e_d)$. Expand

$$\psi_m = \sum_{k=1}^{d}c_k^{(m)}e_k, \qquad c_k^{(m)} = \langle e_k\vert\psi_m\rangle.$$

Cauchy–Schwarz gives

$$\left\lvert c_k^{(m)}-c_k^{(n)} \right\rvert \leq \lVert\psi_m-\psi_n\rVert.$$

Each coordinate sequence $(c_k^{(m)})$ is therefore Cauchy in $\mathbb C$ and has a limit $c_k$. Define

$$\psi = \sum_{k=1}^{d}c_ke_k.$$

Because the sum has only finitely many terms,

$$\begin{aligned} \lVert\psi_m-\psi\rVert^2 &= \sum_{k=1}^{d} \left\lvert c_k^{(m)}-c_k \right\rvert^2 \\ &\longrightarrow 0. \end{aligned}$$

Thus $\psi_m\to\psi$ inside the space. The finiteness of the coordinate sum is the decisive step. Infinite-dimensional inner-product spaces may have Cauchy sequences whose limits are missing until the space is completed.

The Standard Coordinate Model

Every $d$-dimensional complex Hilbert space is unitarily isomorphic to $\mathbb C^d$. Choose an orthonormal basis $\mathcal E=(e_1,\ldots,e_d)$ and define

$$U_{\mathcal E}:\mathcal H\longrightarrow\mathbb C^d$$

by

$$U_{\mathcal E}\psi = \begin{pmatrix} \langle e_1\vert\psi\rangle\\ \vdots\\ \langle e_d\vert\psi\rangle \end{pmatrix}.$$

This map is linear, bijective, and preserves inner products:

$$\langle U_{\mathcal E}\phi \vert U_{\mathcal E}\psi\rangle_{\mathbb C^d} = \langle\phi\vert\psi\rangle_{\mathcal H}.$$

The isomorphism depends on the chosen orthonormal basis. The abstract Hilbert space does not come with a preferred coordinate column unless the physics selects one through an observable, symmetry, preparation, measurement, or computational choice.

The statement “all $d$-dimensional Hilbert spaces are isomorphic” means their bare Hilbert-space geometry is the same. It does not mean that all $d$-level physical systems are equivalent. Their Hamiltonians, observables, symmetries, tensor factorizations, and accessible operations may differ.

Orthonormal Coordinates

An orthonormal basis satisfies

$$\langle e_j\vert e_k\rangle = \delta_{jk}.$$

Any vector can be expanded as

$$\lvert\psi\rangle = \sum_{k=1}^{d} \lvert e_k\rangle \langle e_k\vert\psi\rangle.$$

Writing

$$c_k = \langle e_k\vert\psi\rangle,$$

the norm becomes

$$\lVert\psi\rVert^2 = \sum_{k=1}^{d}\lvert c_k\rvert^2.$$

The identity has the finite resolution

$$I = \sum_{k=1}^{d} \lvert e_k\rangle\langle e_k\vert.$$

These formulas are developed in Orthonormal Bases. Their role here is to show how Hilbert-space geometry becomes ordinary complex Euclidean geometry after a suitable basis choice.

Vectors, Rays, and Physical States

The Hilbert space contains vectors, including the zero vector and vectors of arbitrary norm. A normalized vector satisfies

$$\langle\psi\vert\psi\rangle=1.$$

Normalized vectors form the unit sphere, not a vector subspace. For pure-state quantum mechanics, normalized vectors related by a global phase,

$$\lvert\psi\rangle \sim e^{i\alpha}\lvert\psi\rangle,$$

represent the same physical ray. The pure-state space is therefore the complex projective space

$$\mathbb{CP}^{d-1},$$

not the Hilbert space itself.

A normalized vector in $\mathbb C^d$ has $2d$ real components. Normalization removes one real parameter and global phase removes another, leaving $2d-2$ real parameters for a generic pure state. This count describes the state manifold, not the number of measurement outcomes.

The physical interpretation and its limits belong to State Vectors and Rays and Global Phase.

Example: A Qubit

A qubit uses $\mathbb C^2$. A normalized state vector can be written

$$\lvert\psi\rangle = \alpha\lvert0\rangle+\beta\lvert1\rangle, \qquad \lvert\alpha\rvert^2+\lvert\beta\rvert^2=1.$$

After removing a global phase, every pure qubit state can be represented as

$$\lvert\psi\rangle = \cos\frac{\theta}{2}\lvert0\rangle +e^{i\varphi} \sin\frac{\theta}{2}\lvert1\rangle,$$

where

$$0\leq\theta\leq\pi, \qquad 0\leq\varphi<2\pi.$$

The two real parameters agree with the count $2d-2=2$. Their Bloch-sphere geometry is developed in Bloch Sphere Geometry.

Worked Example: The Same Qubit in Two Bases

In the computational basis, consider

$$\lvert\psi\rangle = \frac{ \lvert0\rangle+i\lvert1\rangle }{\sqrt2}.$$

Define the orthonormal basis

$$\lvert+\rangle = \frac{\lvert0\rangle+\lvert1\rangle}{\sqrt2}, \qquad \lvert-\rangle = \frac{\lvert0\rangle-\lvert1\rangle}{\sqrt2}.$$

The new coordinates are inner products:

$$\begin{aligned} c_+ &= \langle+\vert\psi\rangle = \frac{1+i}{2}, \\ c_- &= \langle-\vert\psi\rangle = \frac{1-i}{2}. \end{aligned}$$

Hence

$$\lvert\psi\rangle = \frac{1+i}{2}\lvert+\rangle +\frac{1-i}{2}\lvert-\rangle.$$

Normalization is unchanged:

$$\lvert c_+\rvert^2+\lvert c_-\rvert^2 = \frac12+\frac12 =1.$$

The coordinate column changed, but the vector and its norm did not. A physical probability interpretation requires specifying which basis is associated with a measurement and applying the Born rule.

Finite-Dimensional Subspaces

Every subspace $S\subseteq\mathcal H$ is finite-dimensional and therefore closed. It has an orthogonal complement

$$S^\perp = \lbrace \psi\in\mathcal H: \langle s\vert\psi\rangle=0 \text{ for all }s\in S \rbrace,$$

and the space decomposes as

$$\mathcal H = S\oplus S^\perp.$$

The orthogonal projector $P_S$ is defined on all of $\mathcal H$ and satisfies

$$P_S^2=P_S=P_S^\dagger.$$

These facts support degenerate eigenspaces, coarse-grained alternatives, and finite-dimensional measurement subspaces. Their canonical algebraic treatment is Projectors.

Operators Are Especially Well Behaved

Let $A:\mathcal H\to\mathcal H$ be linear. In finite dimension:

• $A$ is automatically bounded and continuous;
• its adjoint $A^\dagger$ exists on the whole space;
• Hermitian and self-adjoint mean the same thing;
• its spectrum is a finite set of eigenvalues;
• every normal operator has an orthonormal eigenbasis;
• traces, determinants, matrix exponentials, and operator polynomials are finite matrix constructions.

For any orthonormal basis, $A$ has matrix elements

$$A_{jk} = \langle e_j\vert A\vert e_k\rangle,$$

and acts through

$$A\lvert\psi\rangle = \sum_{j,k=1}^{d} \lvert e_j\rangle A_{jk} \langle e_k\vert\psi\rangle.$$

The operator is not the matrix; the matrix is its representation in the chosen basis. See Matrices as Linear Maps.

Important distinctions remain. A general finite matrix need not be diagonalizable, and eigenvectors of a nonnormal matrix need not be orthogonal. Hermitian Operators, Unitary Operators, and Spectral Decomposition state the conditions that guarantee the familiar quantum structure.

Finite Spectral Structure

For a Hermitian operator $A$, the finite-dimensional spectral theorem gives

$$A = \sum_a aP_a,$$

where the distinct eigenvalues $a$ are real and the orthogonal spectral projectors satisfy

$$\sum_aP_a=I, \qquad P_aP_b=\delta_{ab}P_a.$$

Degeneracy changes the rank of $P_a$, not the form of the theorem. There is no continuous spectrum and no need for generalized eigenvectors in a genuinely finite-dimensional space.

The theorem is mathematical. Interpreting $A$ as an observable and $\langle\psi\vert P_a\vert\psi\rangle$ as an outcome probability requires the quantum postulates.

Composite Dimensions

If finite systems have Hilbert spaces $\mathcal H_A$ and $\mathcal H_B$ with dimensions $d_A$ and $d_B$, then

$$\dim( \mathcal H_A\otimes\mathcal H_B ) = d_Ad_B.$$

For $N$ qubits,

$$\dim\mathcal H=2^N.$$

This exponential dimension is one reason finite many-body quantum mechanics becomes computationally difficult even though each local factor is only two-dimensional. The tensor-product construction belongs to Tensor Products; its physical interpretation belongs to Composite Systems and Entanglement Basics.

Exact Finite Systems and Truncations

Distinguish two uses of finite-dimensional Hilbert spaces.

Genuinely finite model. A qubit or ideal spin-$j$ degree of freedom is modeled with a Hilbert space of fixed finite dimension.

Finite approximation. A basis cutoff, finite grid, finite box, or selected energy subspace approximates an underlying infinite-dimensional problem. If $P_N$ projects onto the retained subspace, a common compressed Hamiltonian is

$$H_N = P_NHP_N$$

acting on $P_N\mathcal H$. For an unbounded $H$, the retained basis vectors must lie in the operator domain before this expression is meaningful.

Diagonalizing $H_N$ accurately solves the truncated matrix problem. It does not by itself prove convergence to the continuum or infinite-basis problem. One must vary $N$, the spatial domain, grid spacing, basis family, or other cutoffs and monitor the physical quantities of interest. See Discretization and Convergence Tests.

What Does Not Transfer Automatically

The finite-dimensional setting is an exact model for its own systems and a powerful guide to broader quantum mechanics. The following extensions require new analysis:

• unbounded position, momentum, and Hamiltonian operators;
• domains and boundary conditions;
• continuous or mixed spectra;
• generalized eigenvectors and delta normalization;
• infinite series and spectral integrals;
• trace-class conditions for density operators;
• convergence of finite-dimensional truncations.

The detailed comparison is Finite vs Infinite-Dimensional Quantum Mechanics, and the mathematical continuation is Hilbert Spaces.

Common Mistakes

• Calling any complex vector space a Hilbert space. An inner product is required; completeness is then automatic only because the dimension is finite.
• Confusing Hilbert-space completeness with a basis completeness relation. The first concerns Cauchy sequences; the second resolves the identity.
• Treating a coordinate column as the abstract state. Columns depend on basis.
• Treating the unit sphere as a vector subspace. It is not closed under superposition or arbitrary scaling.
• Identifying vectors with physical pure states one-to-one. Nonzero scalar multiples define the same ray after normalization and phase identification.
• Assuming all matrices are unitarily diagonalizable. Normal matrices are; arbitrary matrices need not even be diagonalizable.
• Reading a truncation as an exact finite system. Convergence to the intended infinite-dimensional problem must be demonstrated.
• Transferring operator-domain conclusions to wave mechanics. Finite matrices hide the domain questions of unbounded operators.

Exercises

1. Complete the coordinate proof that every Cauchy sequence in a $d$-dimensional inner-product space converges.

Solution

Choose an orthonormal basis $(e_k)_{k=1}^d$. If $(\psi_m)$ is Cauchy, then

$$c_k^{(m)} = \langle e_k\vert\psi_m\rangle$$

is Cauchy for every fixed $k$ because

$$\left\lvert c_k^{(m)}-c_k^{(n)} \right\rvert \leq \lVert\psi_m-\psi_n\rVert.$$

Let $c_k$ be the coordinate limit and define $\psi=\sum_{k=1}^d c_ke_k$. Then

$$\lVert\psi_m-\psi\rVert^2 = \sum_{k=1}^{d} \left\lvert c_k^{(m)}-c_k \right\rvert^2.$$

There are finitely many terms and every term tends to zero, so the norm tends to zero. Thus $\psi_m\to\psi$ in the original space.

2. Count the real parameters of a normalized pure state in $\mathbb C^d$ and explain why the answer is not $2d$.

Solution

A vector in $\mathbb C^d$ has $2d$ real coordinates. Normalization imposes one real condition,

$$\sum_{k=1}^{d}\lvert c_k\rvert^2=1,$$

leaving $2d-1$. Global phase identifies $c$ with $e^{i\alpha}c$, removing one more real parameter. The pure-state ray therefore has

$$2d-2$$

real parameters. For $d=2$, this gives the two angles of the Bloch sphere.

3. A system consists of a spin-$1$ degree of freedom and two qubits. What is the dimension of the joint Hilbert space, and how many complex coefficients appear in a general state before normalization?

Solution

A spin-$1$ Hilbert space has dimension $2j+1=3$, and two qubits have joint dimension $2^2=4$. Therefore

$$\dim\mathcal H = 3\times4 =12.$$

A general vector has 12 complex coefficients in a product basis. Normalization and global phase reduce the number of independent real pure-state parameters, but they do not change the Hilbert-space dimension.

4. A truncated oscillator calculation produces a normalized eigenvector with matrix residual $10^{-13}$. State what this proves and what must still be checked before quoting 12-digit continuum accuracy.

Solution

The small residual proves that the vector is an accurate eigenvector of the retained finite matrix, up to the scale and conditioning of that matrix. It does not prove that the finite subspace accurately represents the oscillator state or eigenvalue.

Increase the basis cutoff, monitor the target eigenvalue and observables, check that the state has negligible weight near the cutoff, and compare with an exact result or an independently controlled method when available. Solver error and truncation error are separate.

References

• S. Axler, Linear Algebra Done Right, 3rd ed., Springer, 2015.
• B. C. Hall, Quantum Theory for Mathematicians, Springer, 2013.
• M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, 10th anniversary ed., Cambridge University Press, 2010.
• M. Reed and B. Simon, Methods of Modern Mathematical Physics, Volume I: Functional Analysis, revised and enlarged ed., Academic Press, 1980.