Unitary Operators and Transformations

Quantum computing is applied unitarity: every gate is a unitary operator, algorithms are choreographed products of them, and the reversibility guaranteed by $\hat U^{-1} = \hat U^\dagger$ is why quantum computation is in principle dissipation-free. Compilers decompose a target unitary into hardware-native two-level unitaries; benchmarking measures how far a real gate drifts from the unitary ideal, and error correction polices the gap. On the physics side, the generator dictionary is the daily language of quantum simulation — engineering an effective Hamiltonian is engineering the generator of the unitary you want. Wigner's theorem still earns its keep: testing a proposed "symmetry" of a quantum device means checking preservation of transition probabilities.

Think of Hilbert space as an enormous complex-Euclidean space and of a unitary operator as a rigid rotation of it: vectors get spun around and mixed, but no length changes and no angle changes. Since lengths are norms (total probability) and angles are overlaps (transition amplitudes), a unitary moves the furniture without breaking any physics. Contrast the generic operator of Linear Operators on Hilbert Space, which stretches some directions and crushes others: unitaries have zero distortion.

There are two ways to read $|\psi\rangle \to \hat U|\psi\rangle$. Active: the state really moved — you rotated the spin, translated the atom, let time pass. Passive: nothing moved; you re-described the same state in new axes — new components, same arrow. The same algebra does both jobs; which reading applies is bookkeeping. This double life is why unitaries saturate the theory: every change of viewpoint and every symmetry operation passes through them.

The third face: unitarity is probability conservation. If the Born weights $\sum_n|\langle n|\psi\rangle|^2 = 1$ must stay 1, the transformation cannot leak or manufacture probability — and "no leaks, ever, for every state" is precisely $\hat U^\dagger\hat U = \hat 1$. When the dynamics chapter derives $\hat U(t) = e^{-i\hat Ht/\hbar}$, conservation of probability in time will be this theorem wearing a clock.

An operator is unitary when its adjoint is its inverse, both ways:

where: the dagger is the previous article's adjoint, and demanding both orders matters in infinite dimensions (the limiting-cases table has the one-sided counterexample). Condition (1) is the cheapest invertibility certificate in the theory: the inverse is handed to you — conjugate-transpose and done. That is why time evolution can always be run backwards. What (1) buys is the preservation of every inner product:

where: the first equality moves $\hat U$ across the inner product as $\hat U^\dagger$ (the defining relation of Hermitian Conjugate and Adjoint Operators), the second collapses $\hat U^\dagger\hat U$ by (1). Setting $\phi = \psi$ gives norm preservation; keeping them distinct preserves every overlap, hence every transition probability and every basis's orthonormality.

A rigid motion cannot stretch any direction — so no eigenvalue can have modulus different from 1. Take $\hat U|u\rangle = \lambda|u\rangle$ with $|u\rangle$ normalized; preservation of the norm gives $\|\hat U u\| = |\lambda|\,\|u\| = \|u\|$, forcing

where: every eigenvalue is a pure phase. Where a Hermitian spectrum lies on the real line (measured values), a unitary's lies on the unit circle (accumulated phases); the exponential map below is the wrapping of the line onto the circle. Eigenvectors of distinct eigenvalues are orthogonal — unitaries are normal ($[\hat U, \hat U^\dagger] = 0$, immediate from (1)) — so a complete orthonormal eigenbasis exists and $\hat U = \sum_n e^{i\theta_n}|u_n\rangle\langle u_n|$ by the assembly machinery of Completeness Relation and Resolution of Identity.

The completeness article noted its change-of-basis matrix was unitary, "as developed in Unitary Operators and Transformations." Here is the development. Given two orthonormal bases $\{|a_n\rangle\}$ and $\{|b_n\rangle\}$, the operator carrying one onto the other is

where: $\hat U|a_k\rangle = |b_k\rangle$ (orthonormality of the $a$'s collapses the sum) — each old axis maps to the corresponding new axis: "rotate the frame." Unitarity is a two-line check: $\hat U^\dagger\hat U = \sum_{n,m}|a_n\rangle\langle b_n|b_m\rangle\langle a_m| = \sum_n |a_n\rangle\langle a_n| = \hat 1$, using orthonormality of the $b$'s and completeness of the $a$'s — both bases must be complete, or the "rotation" quietly discards part of the space. States keep their identity, components change: $\psi_n^{(b)} = \langle b_n|\psi\rangle$. Operators ride along by similarity:

where: the sandwich structure is forced by consistency — transformed operator acting on transformed states must reproduce transformed results. What survives the ride is everything measurable, and that is a theorem:

The deepest fact about unitaries is where they come from. Exponentiate $i$ times any Hermitian operator and the result is unitary:

where: the dagger of the series $\sum (i\hat G)^n/n!$ conjugates each $i$ and daggers each $\hat G$ — Hermiticity keeps $\hat G$ fixed, so only the exponent's sign flips, and opposite-sign exponentials of the same operator multiply to $\hat 1$. $\hat G$ is the generator. The converse holds too: every unitary is $e^{i\hat G}$ — take $\hat G = \sum_n \theta_n |u_n\rangle\langle u_n|$ from the eigenphases of the spectral form. This is the promise Linear Operators on Hilbert Space made ("exponentials built and certified here") and the division of labor flagged by Hermitian Operators and Observables: Hermitian operators are simultaneously observables and generators — measurement menus and motion engines, one cast in two roles. Physical transformations come with a dial — translate by $a$, rotate by $\theta$, wait $t$ — and the exponential organizes each family into a one-parameter group:

where: dial settings add because same-generator exponentials commute, and the infinitesimal form exhibits the generator as the family's "velocity" at the identity. Stone's theorem is the rigorous capstone: every strongly continuous one-parameter unitary group is $e^{i\lambda\hat G}$ for a unique self-adjoint generator, and vice versa. Stated here as the license; the dynamics chapter's Time-Evolution Operator puts it to work on the family it was born for, $\hat U(t) = e^{-i\hat Ht/\hbar}$. The marquee families, with generators and owners:

The translation family deserves its worked moment, because it cashes a specific IOU. Canonical Commutation Relation proved — by Taylor-expanding the series on wavefunctions — that $e^{-ia\hat p/\hbar}$ genuinely shifts; that proof and the Weyl exchange phase stay its property, recalled by name. What this article certifies is the family's symmetry machinery: $\hat T(a)$ is unitary because $\hat p$ is Hermitian (equation (6) with $\hat G = -a\hat p/\hbar$); the group law $\hat T(a)\hat T(b) = \hat T(a+b)$ holds because the dial adds; and the adjoint is the reverse shift, $\hat T(a)^\dagger = \hat T(-a)$. All three are verified symbolically on explicit wavefunctions. One more theorem closes the circle. Equations (6)–(7) say Hermitian generators produce unitary symmetries; Wigner's theorem says there was never another choice: any transformation preserving all transition probabilities $|\langle\phi|\psi\rangle|^2$ must be implemented unitarily or antiunitarily (unitary composed with complex conjugation). The antiunitary loophole is used exactly once in nature's rulebook — time reversal, the dynamics chapter's Time Reversal in Quantum Dynamics. Every other symmetry you will meet is this article's kind.

Example 1 — certify a rotation and find its generator (confidence builder). The matrix $R = \bigl(\begin{smallmatrix}\cos\theta & -\sin\theta\\ \sin\theta & \cos\theta\end{smallmatrix}\bigr)$ with $\theta = 30°$ acts on a two-state system. Certify unitarity, locate the spectrum, exhibit the Hermitian generator.

Step 1. Unitarity: $R^\dagger R = \hat 1$ by direct multiplication (real matrix: dagger = transpose; rows orthonormal). Verified to $10^{-13}$.

Step 2. Spectrum: eigenvalues $e^{\pm i\pi/6}$ — on the unit circle as (3) demands, not real: a generic unitary is not an observable. The eigenvectors $(1, \mp i)/\sqrt2$ are orthogonal.

Step 3. Generator: guess the family $R(\theta) = e^{-i\theta\sigma_y}$ and check — $\sigma_y = \bigl(\begin{smallmatrix}0 & -i\\ i & 0\end{smallmatrix}\bigr)$ is Hermitian, $\sigma_y^2 = \hat 1$, so the series splits into $\cos\theta\,\hat 1 - i\sin\theta\,\sigma_y$, which is exactly $R$. So $R = e^{i\hat G}$ with $\hat G = -\theta\sigma_y$: the rotation family is generated by a Hermitian operator, equation (6) in the flesh (machine-checked to $10^{-13}$).

Example 2 — invariance audit across a basis change (exam level). A spin-½ state has $z$-basis components $(0.6,\ 0.8)$. Transform to the $x$-basis $|\pm x\rangle = (|{+z}\rangle \pm |{-z}\rangle)/\sqrt2$ and audit what changes and what cannot.

Step 1. New components: $\langle +x|\psi\rangle = (0.6+0.8)/\sqrt2 = 0.9899$, $\langle -x|\psi\rangle = (0.6-0.8)/\sqrt2 = -0.1414$. Probabilities: $\mathbf{0.98}$ and $\mathbf{0.02}$ — sum exactly 1 (norm preserved; components changed dramatically).

Step 2. Transform the operator: $\sigma_z' = \hat U\sigma_z\hat U^\dagger$ works out to $\bigl(\begin{smallmatrix}0 & 1\\ 1 & 0\end{smallmatrix}\bigr)$ — the matrix of $\sigma_z$ looks completely different in the new basis (it became off-diagonal). Its eigenvalues: still $\pm 1$. The menu of outcomes survived the rotation, per the similarity derivation.

Step 3. Expectation audit: old basis $\langle\sigma_z\rangle = 0.36 - 0.64 = -0.28$; new basis $\langle\psi'|\sigma_z'|\psi'\rangle = \mathbf{-0.28}$ — identical to $10^{-13}$. Components, matrices, appearances all changed; every physical number stayed. That is unitarity doing its one job.

References. Sakurai & Napolitano, Modern Quantum Mechanics, 2nd ed., Ch. 1.5–1.6, 2.1 · Shankar, Principles of Quantum Mechanics, 2nd ed., Ch. 1.7–1.9 · Cohen-Tannoudji, Quantum Mechanics, Vol. 1, Complement B-II · Griffiths & Schroeter, Introduction to Quantum Mechanics, 3rd ed., Appendix A · Townsend, A Modern Approach to Quantum Mechanics, Ch. 2–3.

Provenance. All equations are standard textbook results. A verifier should: (1) confirm the equivalence of inner-product preservation and $U^\dagger U=\mathbb{1}$ (Sakurai 1.5, Shankar 1.7); (2) verify $U=e^{iG}$ with Hermitian $G$ via the spectral theorem (Cohen-Tannoudji Complement B-II); (3) check that $|\lambda|=1$ for unitary eigenvalues; (4) confirm $U(t)=e^{-iHt/\hbar}$ solves the Schrödinger equation and conserves the norm (Griffiths Appendix A, Townsend Ch. 4). For this edition all were machine-executed (numpy/scipy/sympy, 34 checks): inner-product/norm preservation and polarization on random unitaries, unit-circle eigenvalues + spectral rebuild, $e^{iH}$ unitarity with group law and $O(d\lambda^2)$ infinitesimal error, the basis-change construction with all four invariance checks, the translation group law and adjoint on explicit wavefunctions, the 2×2 unitary form symbolically, the isometry-not-unitary shift, and both worked examples ($e^{\pm i\pi/6}$; 0.98/0.02, $-0.28$ invariant). Per the chapter ownership map: the adjoint calculus lives in Hermitian Conjugate and Adjoint Operators; the proof that $e^{-ia\hat p/\hbar}$ translates plus the Weyl exchange phase and Stone–von Neumann statement live in Canonical Commutation Relation (recalled by name); basis-insertion machinery lives in Completeness Relation and Resolution of Identity; rotation operators are developed in the angular-momentum chapter; $\hat U(t)$ and Stone's theorem at work are the dynamics chapter's opening article.

Builds on: Hilbert Space and State Vectors · Inner Product and Norm of States · Linear Operators on Hilbert Space · Hermitian Operators and Observables · Dirac Bra-Ket Notation

Leads to: Completeness Relation and Resolution of Identity · The Postulates of Quantum Mechanics · Born Rule and Measurement Probabilities · Eigenvalues and Eigenstates