Simple Algorithm
Quantum Computing · Rydberg State · 26 min

Beyond Finite Dimensions

Continuous variables, unbounded operators, and why qubits keep things finite.

A qubit lives in C2\C^2. A particle on a line lives in L2(R)L^2(\R), where position and momentum cannot be matrices, some "eigenvectors" are not vectors, and "Hermitian" no longer guarantees dynamics. Seeing what breaks explains why quantum computers are built from finite pieces.

What you’ll learn

  • Prove that [X,P]=iI\comm{X}{P}=i\id has no solution in finite dimensions or with bounded operators (Wintner–Wielandt).
  • Distinguish symmetric from self-adjoint operators and state Stone's theorem.
  • Explain continuous spectrum, non-normalizable "eigenkets" ∣x⟩\ket x, and the Stone–von Neumann theorem.
  • Bound the error of truncating an energy-constrained oscillator, and describe how GKP codes embed a qubit in one.

Infinite-dimensional Hilbert space

A Hilbert space is a complex inner-product space that is complete: Cauchy sequences converge. Every separable infinite-dimensional one is unitarily isomorphic to ℓ2(N)\ell^2(\N), the square-summable sequences, and so also to L2(R)L^2(\R), the square-integrable wavefunctions. An orthonormal basis {en}\{e_n\} now means ψ=∑n⟨en∣ψ⟩en\psi=\sum_n\braket{e_n|\psi}e_n with convergence in norm: infinite sums, not finite linear combinations.

Finite-dimensional intuition fails quickly. On ℓ2(N)\ell^2(\N) the unilateral shift S∣n⟩=∣n+1⟩S\ket n=\ket{n+1} is an isometry, S†S=IS^\dagger S=\id, yet SS†=I−∣0⟩ ⁣⟨0∣SS^\dagger=\id-\dyad0: it is not unitary, a situation impossible in finite dimensions. The oscillator's annihilation operator a=∑n≥0n+1 ∣n⟩ ⁣⟨n+1∣=S†Na=\sum_{n\ge0}\sqrt{n+1}\,\ketbra{n}{n+1}=S^\dagger\sqrt N is worse: it is unbounded, since ∥a∣n⟩∥=n\norm{a\ket n}=\sqrt n grows without limit.

Check your understanding

What is the set of eigenvalues (point spectrum) of x^\hat x on L2(R)L^2(\R)?

  1. All of R\R
  2. The integers
  3. The empty set
  4. Only 00
Show the answer

C. The empty set

If xψ(x)=λψ(x)x\psi(x)=\lambda\psi(x) almost everywhere, then ψ=0\psi=0 almost everywhere, so there are no eigenvalues. The spectrum, where x^−λ\hat x-\lambda has no bounded inverse, is still all of R\R: it is purely continuous.

Why X and P cannot be matrices

The finite case is one line: Tr⁡[X,P]=Tr⁡(XP)−Tr⁡(PX)=0\Tr\comm{X}{P}=\Tr(XP)-\Tr(PX)=0, while Tr⁡(iId)=id≠0\Tr(i\id_d)=id\neq0. In infinite dimensions the trace is unavailable, but a norm argument works:

DerivationWielandt's proof
  1. XPn−PnX=n c Pn−1XP^n-P^nX=n\,c\,P^{n-1}

    Suppose X,PX,P are bounded with XP−PX=c IXP-PX=c\,\id, c≠0c\neq0. Induction on n≥1n\ge1, using XPn+1−Pn+1X=(XPn−PnX)P+Pn(XP−PX)XP^{n+1}-P^{n+1}X=(XP^n-P^nX)P+P^n(XP-PX), gives

  2. Pn≠0for all nP^n\neq0\quad\text{for all }n

    No power of PP vanishes: if Pn=0P^n=0 then ncPn−1=XPn−PnX=0ncP^{n-1}=XP^n-P^nX=0, so Pn−1=0P^{n-1}=0; descending gives I=P0=0\id=P^0=0, absurd.

  3. n ∣c∣ ∥Pn−1∥=∥XPn−PnX∥≤2∥X∥ ∥P∥ ∥Pn−1∥n\,\abs c\,\norm{P^{n-1}}=\norm{XP^n-P^nX}\le2\norm X\,\norm P\,\norm{P^{n-1}}

    Take norms, using ∥AB∥≤∥A∥∥B∥\norm{AB}\le\norm A\norm B:

  4. n≤2∥X∥∥P∥∣c∣∀nn\le\frac{2\norm X\norm P}{\abs c}\quad\forall n

    Divide by ∥Pn−1∥>0\norm{P^{n-1}}>0. The bound fails for large nn: contradiction. The proof works verbatim in any unital Banach algebra. ■\blacksquare

The Weyl form of the CCR uses bounded unitaries instead: U(t)=eitx^U(t)=e^{it\hat x} and V(s)=eisp^V(s)=e^{is\hat p} satisfy U(t)V(s)=e−istV(s)U(t)U(t)V(s)=e^{-ist}V(s)U(t) (with ℏ=1\hbar=1). These group relations do have finite-dimensional cousins: the clock and shift matrices Z∣j⟩=ωj∣j⟩Z\ket j=\omega^j\ket j, X∣j⟩=∣j+1 mod d⟩X\ket j=\ket{j+1\bmod d}, ω=e2πi/d\omega=e^{2\pi i/d}, satisfy ZX=ωXZZX=\omega XZ. Qudits keep the Heisenberg–Weyl group but lose its Lie algebra.

Domains, self-adjointness and Stone's theorem

An unbounded operator AA comes with a dense domain D(A)⊂HD(A)\subset\Hil. It is symmetric if ⟨ϕ∣Aψ⟩=⟨Aϕ∣ψ⟩\braket{\phi|A\psi}=\braket{A\phi|\psi} for all ϕ,ψ∈D(A)\phi,\psi\in D(A), and self-adjoint if moreover D(A†)=D(A)D(A^\dagger)=D(A). Finite-dimensional "Hermitian" splits into these two notions, and only the second has a spectral theorem. Domains are unavoidable: by the Hellinger–Toeplitz theorem, a symmetric operator defined on all of H\Hil is bounded.

Oscillators, and keeping things finite

Continuous-variable (CV) quantum information uses bosonic modes, oscillators H=ω(a†a+12)H=\omega\big(a^\dagger a+\tfrac12\big) with [a,a†]=I\comm{a}{a^\dagger}=\id and Fock basis ∣n⟩\ket n. Gaussian states and operations (coherent and squeezed states, beam splitters, homodyne detection) are easy to make, but by a CV analog of the Gottesman–Knill theorem they are efficiently classically simulable. Universality needs a non-Gaussian ingredient, such as a cubic phase gate or photon counting.

Energy is what makes an oscillator effectively finite. If ⟨ψ|a†a|ψ⟩=nˉ\mel{\psi}{a^\dagger a}{\psi}=\bar n, Markov's inequality bounds the weight above level KK:

DerivationEnergy-constrained truncation
  1. p=∑n>K∣⟨n∣ψ⟩∣2≤∑n>KnK+1∣⟨n∣ψ⟩∣2≤nˉK+1p=\sum_{n>K}\abs{\braket{n|\psi}}^2\le\sum_{n>K}\frac{n}{K+1}\abs{\braket{n|\psi}}^2\le\frac{\bar n}{K+1}

    Let ΠK=∑n=0K∣n⟩ ⁣⟨n∣\Pi_K=\sum_{n=0}^K\dyad n and p=⟨ψ|I−ΠK|ψ⟩p=\mel{\psi}{\id-\Pi_K}{\psi}, the probability of more than KK quanta. Since n/(K+1)≥1n/(K+1)\ge1 on the tail,

  2. D(ψ,ψK)=p≤nˉK+1D(\psi,\psi_K)=\sqrt p\le\sqrt{\frac{\bar n}{K+1}}

    The normalized truncation ∣ψK⟩=ΠK∣ψ⟩/∥ΠKψ∥\ket{\psi_K}=\Pi_K\ket\psi/\norm{\Pi_K\psi} has ∣⟨ψK∣ψ⟩∣2=∥ΠKψ∥2=1−p\abs{\braket{\psi_K|\psi}}^2=\norm{\Pi_K\psi}^2=1-p, so by the pure-state trace-distance formula

  3. K+1≥nˉϵ2⟹D(ψ,ψK)≤ϵK+1\ge\frac{\bar n}{\epsilon^2}\quad\Longrightarrow\quad D(\psi,\psi_K)\le\epsilon

    Within an energy budget, a (K+1)(K+1)-level qudit captures the state to any accuracy. States with light tails, such as coherent states (Poissonian photon number), need far fewer levels.

Predict before you look

A pure oscillator state has nˉ=0.1\bar n=0.1. You keep only ∣0⟩\ket0 and ∣1⟩\ket1 (a qubit, K=1K=1). By the Markov bound, at most what percentage of the probability lies outside that subspace?

Show the answer

5 %

p≤nˉ/(K+1)=0.1/2=0.05p\le\bar n/(K+1)=0.1/2=0.05, i.e. 5%5\%, so the truncation is within trace distance 0.05≈0.22\sqrt{0.05}\approx0.22. For a coherent state with nˉ=0.1\bar n=0.1 the true tail is much smaller: p=1−e−0.1(1+0.1)≈0.47%p=1-e^{-0.1}(1+0.1)\approx0.47\%.

Key takeaways

  • [X,P]=iI\comm{X}{P}=i\id forces infinite dimensions and unbounded operators (trace argument; Wintner–Wielandt).
  • Self-adjoint, not merely symmetric: only self-adjoint operators have a spectral theorem and generate unitary dynamics (Stone).
  • Continuous spectrum has no normalizable eigenvectors; ∣x⟩\ket x is a distribution. Stone–von Neumann makes the Schrödinger representation unique for finitely many modes.
  • An energy budget makes an oscillator effectively finite, with truncation error at most nˉ/(K+1)\sqrt{\bar n/(K+1)}; GKP codes embed a qubit in a mode on purpose.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

Why can no pair of d×dd\times d matrices satisfy [X,P]=iI\comm{X}{P}=i\id?

  1. The commutator of two Hermitian matrices is Hermitian, but iIi\id is not
  2. Tr⁡[X,P]=0\Tr\comm{X}{P}=0, while Tr⁡(iI)=id≠0\Tr(i\id)=id\neq0
  3. Eigenvalues of Hermitian matrices are real
  4. Finite matrices always commute
Show the answer

B. Tr⁡[X,P]=0\Tr\comm{X}{P}=0, while Tr⁡(iI)=id≠0\Tr(i\id)=id\neq0

Cyclicity gives Tr⁡(XP)=Tr⁡(PX)\Tr(XP)=\Tr(PX), so every commutator is traceless; iIdi\id_d has trace idid.

Question 2

On ℓ2(N)\ell^2(\N), let S∣n⟩=∣n+1⟩S\ket n=\ket{n+1}. Which statement is true?

  1. SS is unitary
  2. SS is unbounded
  3. S†S=IS^\dagger S=\id but SS†≠ISS^\dagger\neq\id
  4. SS is self-adjoint
Show the answer

C. S†S=IS^\dagger S=\id but SS†≠ISS^\dagger\neq\id

SS preserves norms (S†S=IS^\dagger S=\id) but nothing maps to ∣0⟩\ket0, so SS†=I−∣0⟩ ⁣⟨0∣SS^\dagger=\id-\dyad0. A non-surjective isometry of a space into itself cannot exist in finite dimensions.

Question 3

Which operators generate strongly continuous one-parameter unitary groups e−itHe^{-itH}?

  1. Exactly the self-adjoint ones
  2. All symmetric operators
  3. Only bounded Hermitian operators
  4. Any densely defined operator with real spectrum
Show the answer

A. Exactly the self-adjoint ones

Stone's theorem is a bijection between strongly continuous unitary groups and self-adjoint generators, bounded or not: p^\hat p on L2(R)L^2(\R) is unbounded and generates translations.

Question 4

What does the Stone–von Neumann theorem assert?

  1. Every symmetric operator has a self-adjoint extension
  2. Position and momentum have finite-matrix representations in a suitable basis
  3. Representations of the CCR are unique even for infinitely many degrees of freedom
  4. Irreducible, strongly continuous representations of the Weyl relations with finitely many degrees of freedom are all unitarily equivalent
Show the answer

D. Irreducible, strongly continuous representations of the Weyl relations with finitely many degrees of freedom are all unitarily equivalent

For finitely many modes, any irreducible representation of the Weyl relations is the Schrödinger representation in disguise, so "the" quantum mechanics of particles is unique up to a change of basis.

Further reading

  • M. Reed & B. Simon, Methods of Modern Mathematical Physics I: Functional Analysis (Academic Press, revised ed., 1980).
  • B. C. Hall, Quantum Theory for Mathematicians, Graduate Texts in Mathematics 267 (Springer, 2013).
  • H. Wielandt, "Über die Unbeschränktheit der Operatoren der Quantenmechanik," Math. Ann. (1949).
  • D. Gottesman, A. Kitaev & J. Preskill, "Encoding a qubit in an oscillator," Phys. Rev. A 64, 012310 (2001).
  • C. Weedbrook et al., "Gaussian quantum information," Rev. Mod. Phys. 84, 621 (2012).