The Measurement Postulate

The measurement postulate is the hardware interface of quantum technology. Every superconducting or trapped-ion processor ends its circuit with a projective readout in the computational basis, and the art of algorithm design — Grover, Shor, variational eigensolvers — is engineering interference so the wanted answer carries near-unit Born probability before that readout fires. POVMs are the working language of quantum key distribution, where minimum-error and unambiguous discrimination (Example 2's zero-error trick, run on the BB84 or B92 states) set the security proofs and achievable rates. Quantum error correction leans on repeatable projective measurements of stabilizer operators: the syndrome is read without collapsing the protected logical information, because the stabilizer projectors commute with the logical operators. At the frontier, weak-measurement operators near the identity drive continuous feedback: labs monitor a superconducting qubit shot by shot, watch a single quantum trajectory unravel in real time, and steer it with the back-action $M_m\rho M_m^\dagger$. Meanwhile the interpretation of process 1 — objective collapse, decoherence, many-worlds — remains a live program, with tabletop tests probing whether collapse is a real physical event or, as the postulate assumes, an update of the observer's description.

Picture the operators $\{M_m\}$ as a bank of detector channels wired in parallel. Feed in $|\psi\rangle$, and channel $m$ emits the unnormalized vector $M_m|\psi\rangle$. Two numbers read off it answer everything: its squared length is the probability the apparatus reports outcome $m$, and its direction, renormalized, is the state left behind. Completeness is a conservation law for that squared length — summed over channels it equals the input's own, so no probability leaks away and none is manufactured. "The wavefunction collapses" is not an extra axiom but the second number the operators already carry.

The renormalization is where irreversibility lives. Before the measurement, the relative phase between the parts of $|\psi\rangle$ feeding different outcomes is genuine quantum information that steers any later interference experiment; conditioning on the outcome you saw keeps one branch and throws that cross-branch phase away for good. A unitary can never do this — it preserves inner products and is invertible — which is why measurement is the non-unitary primitive and the origin of quantum randomness: from a fully specified pure state, a single outcome cannot be predicted, only its distribution. Keep two jobs separate. The effect $E_m=M_m^\dagger M_m$ fixes what you learn — it alone sets $p(m)$. The full operator $M_m$ also fixes what you do to the system — the back-action. Two experiments with identical effects but different $M_m$ (a left unitary apart, $M_m'=U_mM_m$) are statistically indistinguishable on any single shot yet leave the system in different states. When the system is destroyed on detection — a photon absorbed at a photodiode — the back-action is irrelevant and you need only the effects: a POVM, a set $\{E_m\}$ with $E_m\ge0$, $\sum_m E_m=I$. That freedom is what lets a three-outcome measurement live on a two-dimensional qubit and beat any projective measurement.

The postulate names its data first: not a Hermitian "observable," but a set of operators indexed by outcomes. Observables come later as a special case; everything else here reads a number off these operators or constrains them.

where: $M_m$ — a linear operator on the system Hilbert space $\mathcal{H}$, attached to outcome $m$; $m$ — the classical label recorded by the apparatus; $\mathcal{H}$ — the state space. One operator per distinguishable result; the index set may be finite, countable, or (with an integral) continuous. Given the operators, the probability of each outcome is a quadratic form in the state — the formula from which the Born rule, expectation values, and all measurement statistics descend.

where: $|\psi\rangle$ — the normalized pre-measurement state, $\langle\psi|\psi\rangle=1$; $p(m)$ — probability of outcome $m$, automatically real and $\ge0$ because $M_m^\dagger M_m$ is positive semidefinite; $\|\cdot\|$ — the Hilbert-space norm. Positivity is free; normalization is imposed by the next line. Completeness is the entire consistency requirement, chosen for one reason — to make the probabilities in (2) sum to one for every input — and Derivation A shows this "if and only if" is tight.

where: $M_m^\dagger$ — the Hermitian adjoint (conjugate transpose) of $M_m$; $I$ — the identity on $\mathcal{H}$. It is $\sum_m M_m^\dagger M_m$ that must equal $I$, not $\sum_m M_m$; the $M_m$ themselves need be neither Hermitian nor positive. Conditional on the outcome, the state is transformed by $M_m$ and renormalized — the division by $\sqrt{p(m)}$ is mandatory because $M_m$ is generally not norm-preserving. This "collapse" is a consequence of (2), not an independent axiom.

where: $|\psi_m\rangle$ — the normalized state given outcome $m$; the denominator $\sqrt{p(m)}$ restores unit norm and is defined only when $p(m)>0$ (zero-probability outcomes never occur, so their post-states are irrelevant). When only the statistics matter — destructive detection, or a discarded system — the probabilities depend on the operators solely through $E_m=M_m^\dagger M_m$. Collecting these effects gives a POVM, the most general description of "what an experiment can learn."

where: $E_m$ — the effect for outcome $m$, positive semidefinite; $\{E_m\}$ is a positive operator-valued measure. Any $M_m=U_m\sqrt{E_m}$ with $U_m$ unitary reproduces the same $E_m$, so a POVM fixes the statistics while leaving the back-action free (Derivation C). For mixed states the pure-state rules lift verbatim: replace $|\psi\rangle\langle\psi|$ with the density operator $\rho$ and inner products with traces. Derivation C shows this is not a new postulate but an average of the old one.

where: $\rho$ — density operator, $\rho\ge0$, $\mathrm{tr}\,\rho=1$; $\rho_m$ — post-measurement density operator given outcome $m$; $\mathrm{tr}$ — trace over $\mathcal{H}$. Setting $\rho=|\psi\rangle\langle\psi|$ returns (2) and (4) exactly. The textbook "measure an observable" language is the projective special case: a Hermitian $A$ supplies its own operators through the spectral theorem, and its mean is read straight off the outcome distribution.

where: $A$ — a Hermitian observable; $P_m$ — the orthogonal projector onto the eigenspace with eigenvalue $m$, obeying $P_m^\dagger=P_m$, $P_mP_n=\delta_{mn}P_m$, $\sum_m P_m=I$; $\langle A\rangle$ — the mean over many identical runs. Here $M_m=P_m$, and Derivation B drops out the Born rule and repeatability at once.

The one general postulate specializes to every standard scheme by constraining $\{M_m\}$. Each row below is a downstream node in its own right — here, the boundary values the master rule must reproduce.

1 · A qubit measured in two bases (confidence builder)

Prepare $|\psi\rangle=\tfrac{\sqrt3}{2}|0\rangle+\tfrac12|1\rangle$ and measure it (i) in the $Z$ basis, (ii) in the $X$ basis $|\pm\rangle=(|0\rangle\pm|1\rangle)/\sqrt2$. Find every probability and both expectation values, checking against $\langle A\rangle=\langle\psi|A|\psi\rangle$.

Step 1. Read $Z$ probabilities off the amplitudes. Check $\tfrac34+\tfrac14=1$, so the state is legal. With $P_0=|0\rangle\langle0|$, $P_1=|1\rangle\langle1|$, the Born rule gives $p(0)=\tfrac34$, $p(1)=\tfrac14$; outcome $0$ collapses to $|0\rangle$, outcome $1$ to $|1\rangle$. p(0) = 3/4, p(1) = 1/4.

Step 2. Get the $Z$ expectation two ways. From the distribution ($\sigma_z$ eigenvalues $\pm1$): $\langle\sigma_z\rangle=(+1)\tfrac34+(-1)\tfrac14=\tfrac12$. Directly, $\langle\psi|\sigma_z|\psi\rangle=\tfrac34-\tfrac14=\tfrac12$ — they agree, as (7) demands. ⟨σ_z⟩ = 1/2.

Step 3. Re-expand the same state in the $X$ basis. The $|+\rangle$ amplitude is $\langle+|\psi\rangle=\tfrac{1}{\sqrt2}\big(\tfrac{\sqrt3}{2}+\tfrac12\big)=\tfrac{\sqrt3+1}{2\sqrt2}$, so $p(+)=\tfrac{(\sqrt3+1)^2}{8}=\tfrac{2+\sqrt3}{4}\approx0.933$ and $p(-)=\tfrac{2-\sqrt3}{4}\approx0.067$, summing to one. p(+) ≈ 0.933, p(−) ≈ 0.067.

Step 4. The $X$ expectation: $\langle\sigma_x\rangle=p(+)-p(-)=\tfrac{2\sqrt3}{4}=\tfrac{\sqrt3}{2}\approx0.866$, matching $\langle\psi|\sigma_x|\psi\rangle=2\,\mathrm{Re}(\alpha^*\beta)=2\cdot\tfrac{\sqrt3}{2}\cdot\tfrac12=\tfrac{\sqrt3}{2}$. The same state gives lopsided odds in one basis and near-certain odds in another — statistics are a statement about the state and the chosen basis, never the state alone. ⟨σ_x⟩ = √3/2 ≈ 0.866.

2 · A three-outcome POVM that beats every projective measurement

Two equally likely non-orthogonal states arrive, $|\psi_1\rangle=|0\rangle$ and $|\psi_2\rangle=\tfrac12|0\rangle+\tfrac{\sqrt3}{2}|1\rangle$, overlap $s=\langle\psi_1|\psi_2\rangle=\tfrac12$. Build a measurement that is <strong>never wrong</strong> — allowed to sometimes answer "don't know" — and find how often it succeeds. A projective measurement cannot; a POVM can.

Step 1. Choose the effect directions from orthogonality. To be never wrong, an "it was $\psi_1$" click must be impossible on input $\psi_2$, so its effect points along $|\psi_2^\perp\rangle=\tfrac{\sqrt3}{2}|0\rangle-\tfrac12|1\rangle$; likewise "it was $\psi_2$" points along $|\psi_1^\perp\rangle=|1\rangle$. Set $E_1=a\,|\psi_2^\perp\rangle\langle\psi_2^\perp|$ and $E_2=a\,|\psi_1^\perp\rangle\langle\psi_1^\perp|$ with shared weight $a$ (symmetric, optimal for equal priors). Since $\langle\psi_2^\perp|\psi_2\rangle=0$, outcome 1 has zero probability on $\psi_2$ — zero error by construction.

Step 2. Fix the weight by positivity of the third effect. The leftover $E_?=I-E_1-E_2$ must stay positive; the optimum pushes $a$ as large as possible until $E_?$ just touches the boundary, at $a=\dfrac{1}{1+s}=\dfrac{2}{3}$. a = 2/3.

Step 3. Verify the POVM closes and $E_?\ge0$. In the $\{|0\rangle,|1\rangle\}$ basis, $E_1=\begin{pmatrix}1/2 & -\sqrt3/6\\ -\sqrt3/6 & 1/6\end{pmatrix}$ and $E_2=\begin{pmatrix}0&0\\0&2/3\end{pmatrix}$, so $E_?=I-E_1-E_2=\begin{pmatrix}1/2 & \sqrt3/6\\ \sqrt3/6 & 1/6\end{pmatrix}$. Its trace $\tfrac23>0$ and determinant $\tfrac1{12}-\big(\tfrac{\sqrt3}{6}\big)^2=0$ give eigenvalues $\{0,\tfrac23\}$: positive semidefinite, exactly on the boundary. All three effects are $\ge0$ and sum to $I$. Σ Eₘ = I, all Eₘ ≥ 0.

Step 4. Compute the success probability. For input $\psi_1$ the conclusive outcome is 1: $p(1|\psi_1)=\langle\psi_1|E_1|\psi_1\rangle=\tfrac23\,|\langle0|\psi_2^\perp\rangle|^2=\tfrac23\cdot\tfrac34=\tfrac12$, with $p(?|\psi_1)=\tfrac12$ and $p(2|\psi_1)=0$; by symmetry the same for $\psi_2$. The device is right half the time, inconclusive the other half, and never wrong — matching the optimum $P_\mathrm{success}=1-|s|=\tfrac12$. P(success) = 1 − s = 1/2, error = 0. The third effect $E_?$ — impossible to add in a projective scheme on a qubit — is what buys zero error, the concrete payoff of the general postulate.

The postulate has one architect above all others. Born supplied the probability in 1926 and Heisenberg the disturbance in 1927, but it was John von Neumann who, in 1932, cut quantum dynamics cleanly into two processes and wrote collapse down as an axiom in its own right — the form every later generalization (Kraus operators, POVMs, instruments) extends rather than replaces.

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.2.3–2.2.6 (general and projective measurements, POVMs) and §2.4 (density operators) · Preskill, Lecture Notes on Quantum Computation, Ch. 2–3 · Watrous, The Theory of Quantum Information, Ch. 2 · Wilde, Quantum Information Theory, Ch. 4 · Kitaev, Shen & Vyalyi, Classical and Quantum Computation, Ch. 8 · Peres, Quantum Theory: Concepts and Methods, Ch. 9 · von Neumann, Mathematical Foundations of Quantum Mechanics, Ch. V–VI.

Provenance. All formulas follow Nielsen & Chuang §2.2.3–2.2.6 and §2.4. A verifier should recompute: (1) Derivation A, $\sum_m p(m)=1$ for all $|\psi\rangle\iff\sum_m M_m^\dagger M_m=I$; (2) Derivation B, the Born rule via $P_m^\dagger P_m=P_m$ and repeatability via $P_kP_mP_k=\delta_{mk}P_k$; (3) Derivation C, $p(m)=\mathrm{tr}(M_m^\dagger M_m\rho)$ and the polar freedom $M_m=U_m\sqrt{E_m}$; (4) Example 1's $p(0)=3/4$, $p(+)=(2+\sqrt3)/4\approx0.933$, $\langle\sigma_z\rangle=1/2$, $\langle\sigma_x\rangle=\sqrt3/2$; (5) Example 2's $a=2/3$, closure $\sum_m E_m=I$ with $E_?$ eigenvalues $\{0,2/3\}$, and $P_\mathrm{success}=1-s=1/2$. Cross-check the POVM formalism against Watrous Ch. 2, Wilde Ch. 4, Helstrom (1976); the process-1/process-2 language against von Neumann Ch. V.

Prerequisites. The Qubit and Its State Space · Dirac Bra-Ket Notation · Inner Product, Norm, and Normalization · Superposition Principle. Leads to. The Born Rule · Projective Measurements · Observables and Hermitian Operators · Density Operators.

Ask the tutor: Ask me to walk through any measurement you like — write down its operators $\{M_m\}$ and I will check completeness, compute the odds, and collapse the state step by step.