Projective Measurements

A projective measurement is quantum theory's exact account of the everyday act of reading a meter: you point an apparatus at a system, a pointer lands on one of several marks, and a re-reading lands on the same mark. What makes the quantum version subtle is that the reading is genuinely random, it changes the thing being read, and only a privileged family — the projective, or von Neumann, measurements — is this sharp and repeatable. The whole structure is built from one geometric object: a set of orthogonal projectors $\{P_m\}$ , one per outcome $m$ , that carve the Hilbert space into perpendicular subspaces filling it entirely. Three algebraic conditions do all the work. The projectors are Hermitian ( $P_m^\dagger=P_m$ ), idempotent and mutually orthogonal ( $P_mP_{m'}=\delta_{mm'}P_m$ ), and complete ( $\sum_m P_m=I$ ) — exactly what is required for the probabilities $p(m)=\langle\psi|P_m|\psi\rangle$ to be real, nonnegative, and to sum to one for every input. The bridge to physics is the spectral theorem: any observable $A=A^\dagger$ decomposes as $A=\sum_m m\,P_m$ , so measuring the quantity $A$ is performing the projective measurement $\{P_m\}$ and recording the eigenvalue $m$ — energy, spin, polarization, or a qubit's logical bit. Two properties set projective measurements apart. They are repeatable — the second of two immediate measurements reproduces the first with certainty — and sharp — the number of distinguishable outcomes never exceeds the dimension, and the post-measurement state is fixed by the outcome. General (POVM) measurements relax both, yet Naimark's theorem says every one is a projective measurement hiding on a larger space.

Key takeaways

A projective (von Neumann) measurement is a set of orthogonal, complete projectors: $P_mP_{m'}=\delta_{mm'}P_m$ and $\sum_m P_m=I$ .
Outcome $m$ occurs with probability $p(m)=\langle\psi|P_m|\psi\rangle=\lVert P_m|\psi\rangle\rVert^2$ , and the state collapses to $P_m|\psi\rangle/\sqrt{p(m)}$ .
Any Hermitian observable $A=\sum_m m\,P_m$ is a projective measurement: eigenvalues are the outcomes, eigenspaces the projectors, and $\langle A\rangle=\sum_m m\,p(m)$ .
They are the uniquely repeatable measurements: measure twice in a row and the second result equals the first with certainty, since the collapsed state already lies in one eigenspace.
Discarding the outcome enacts the Lüders dephasing channel $\rho\mapsto\sum_m P_m\rho P_m$ — the operational face of decoherence.
Key idea: A projective measurement resolves the identity into orthogonal projectors $\{P_m\}$ ; outcome probabilities are the squared shadows $p(m)=\langle\psi|P_m|\psi\rangle$ , and the state collapses to the normalized projection $P_m|\psi\rangle/\sqrt{p(m)}$ .

The core insight

A projective measurement is a resolution of the identity into orthogonal projectors $\{P_m\}$ ; outcome probabilities are $p(m)=\langle\psi|P_m|\psi\rangle$ and the state collapses to the normalized projection $P_m|\psi\rangle/\sqrt{p(m)}$ .

Where this lives today

Projective measurement is the last operation in almost every quantum experiment and the backbone of several subfields. In gate-model computing, a circuit ends by measuring each qubit in the computational basis — $P_0=|0\rangle\langle0|,\ P_1=|1\rangle\langle1|$ — and a rotated-basis measurement applies a unitary (a Hadamard for the $X$ basis) first, since $UP_mU^\dagger$ are the rotated projectors. In superconducting hardware this readout is dispersive: the qubit shifts a coupled microwave cavity's frequency, and homodyne detection realizes an approximately projective, increasingly quantum non-demolition (repeatable) measurement with single-shot fidelity now above 99%. Quantum error correction runs on degenerate projective measurements: each stabilizer generator $S$ (a commuting Pauli with rank- $2^{n-1}$ projectors $\tfrac12(I\pm S)$ ) is measured every cycle, and the syndrome locates errors without collapsing the logical qubit — the surface-code experiments now running below threshold are, at bottom, millions of repeated parity measurements. And the non-selective Lüders channel $\rho\mapsto\sum_m P_m\rho P_m$ is the operational model of environment-induced decoherence into a classical pointer basis. Wherever a quantum device outputs a number, a projector did the sorting.

The physical picture

Picture the projectors $\{P_m\}$ as bins that slice the Hilbert space into mutually perpendicular subspaces filling it. Measuring sorts the input vector $|\psi\rangle$ into one bin: the probability of landing in bin $m$ is the squared length of the shadow $|\psi\rangle$ casts into that subspace, and the surviving state is that shadow, rescaled to unit length. Because the subspaces are orthogonal and exhaust the space, the squared shadow-lengths add up to the full squared length of the vector — which is one. The first figure makes "probability = squared shadow" a single geometric statement with real numbers attached.

Figure

A qubit with amplitudes 0.8 and 0.6 on the two outcomes — the Born probabilities are the squared shadows, 0.64 and 0.36, which sum to exactly 1

Repeatability is the geometric signature of projectivity and what earns these measurements the name sharp. Because $P_m^2=P_m$ , once the vector is projected into a subspace it already lies inside it; projecting again does nothing, so an immediate repeat reproduces the first outcome with probability one — a needle that stays put. This fingerprint separates a clean von Neumann measurement from a noisy POVM. The second figure traces it through a two-round experiment.

Figure

Repeatability: measuring Z sorts the state into outcome 0 (probability 0.64) or 1 (probability 0.36), and a second Z measurement returns the same outcome with probability 1

Finally, measuring $A=\sum_m m\,P_m$ is a relabeling of the same sorting: the apparatus projects and prints the eigenvalue $m$ . Observables built from the same projectors — $A$ , $A^2$ , any $f(A)$ — implement the identical sort and are measured simultaneously. The eigenvalues are ink on the dial; the projectors are the machine.

The core equations

A projective measurement is defined not by a state but by its projectors, and three conditions pin them down: Hermiticity makes the values real, orthogonality makes outcomes exclusive and repeatable, and completeness makes the probabilities exhaust unity.

P_m P_{m'}=\delta_{mm'}P_m,\qquad \sum_m P_m = I,\qquad P_m^\dagger=P_m \qquad (1)

where: $P_m$ — projector for outcome $m$ (Hermitian and idempotent, $P_m^2=P_m$ ); $\delta_{mm'}$ — Kronecker delta enforcing orthogonality; $I$ — identity on $\mathcal H$ (completeness). Read it as: perpendicular bins that tile the whole space. The Born rule assigns each outcome a probability equal to the squared length of the state's projection. The three forms below use idempotence and Hermiticity to write it as a squared norm — manifestly nonnegative and real.

p(m)=\langle\psi|P_m|\psi\rangle = \langle\psi|P_m^\dagger P_m|\psi\rangle = \lVert P_m|\psi\rangle\rVert^2 \qquad (2)

where: $|\psi\rangle$ — normalized pure state, $\langle\psi|\psi\rangle=1$ ; $p(m)$ — probability of outcome $m$ ; $\lVert\cdot\rVert$ — norm from the inner product. Read it as: probability is the squared shadow. When outcome $m$ is observed, the state is replaced by its normalized projection into the $m$ -eigenspace: the numerator projects, the denominator restores unit norm.

|\psi_m\rangle=\frac{P_m|\psi\rangle}{\sqrt{\langle\psi|P_m|\psi\rangle}}=\frac{P_m|\psi\rangle}{\sqrt{p(m)}} \qquad (3)

where: $|\psi_m\rangle$ — normalized post-measurement state given outcome $m$ ; $\sqrt{p(m)}$ — renormalization factor. Read it as: keep the shadow, rescale. The link between the abstract projectors and a physical quantity is the spectral theorem: every Hermitian observable is a weighted sum of orthogonal projectors, the weights being its real eigenvalues.

A=\sum_m m\,P_m,\qquad A=A^\dagger,\qquad A|m,k\rangle = m|m,k\rangle \qquad (4)

where: $A$ — observable whose measurement is $\{P_m\}$ ; $m$ — real eigenvalue (the recorded number); $|m,k\rangle$ — orthonormal eigenvector $k$ inside the eigenspace of $m$ ( $k$ runs over any degeneracy). Read it as: the observable is the machine plus its dial markings. The expectation value is the probability-weighted average of eigenvalues — the operator sandwich and the statistical mean are the same number.

\langle A\rangle=\langle\psi|A|\psi\rangle=\sum_m m\,p(m) \qquad (5)

where: $\langle A\rangle$ — mean of the observable in state $|\psi\rangle$ . Read it as: the Born rule averaged over outcomes. For mixed states the same physics is carried by the density operator $\rho$ , with sandwiches promoted to traces — the form used whenever the input is not a single pure vector.

p(m)=\mathrm{Tr}(P_m\,\rho),\qquad \rho_m=\frac{P_m\,\rho\,P_m}{\mathrm{Tr}(P_m\,\rho)} \qquad (6)

where: $\rho$ — density operator of the (possibly mixed) input; $\rho_m$ — post-measurement state given outcome $m$ ; $\mathrm{Tr}$ — trace over $\mathcal H$ . Read it as: the pure-state rules with $|\psi\rangle\langle\psi|\to\rho$ . If the outcome is measured but discarded — a non-selective measurement — the state becomes the probability-weighted average of the conditional states, the Lüders channel. It keeps the diagonal blocks and erases coherences between distinct eigenspaces.

\rho'=\sum_m P_m\,\rho\,P_m \qquad (7)

where: $\rho'$ — state after an unread measurement. Read it as: measuring without looking is dephasing. Finally, the workhorse special case: a single-qubit measurement along any Bloch axis $\hat n$ . Its two projectors are built from the Pauli vector and are orthogonal, summing to $I$ .

P_{\hat n}^{\pm}=\tfrac12\big(I\pm \hat n\cdot\vec\sigma\big),\qquad \hat n\cdot\vec\sigma=n_x X+n_y Y+n_z Z \qquad (8)

where: $\hat n$ — unit vector fixing the measurement axis on the Bloch sphere; $\vec\sigma=(X,Y,Z)$ — Pauli operators; $P_{\hat n}^{\pm}$ — projectors onto the $\pm1$ eigenstates of $\hat n\cdot\vec\sigma$ . Read it as: pick a direction, read out spin along it.

Derivation ARepeatability of a projective measurement

The goal is to prove the property that defines a von Neumann measurement: if you measure $\{P_m\}$ , obtain outcome $m$ , and then immediately measure the same $\{P_m\}$ again, the second reading returns $m$ with certainty and never anything else. We start from the collapsed state, apply the Born rule to it, and let the algebra of orthogonal idempotent projectors deliver the conditional distribution $p(m'\,|\,m)=\delta_{mm'}$ .

Step 1. After the first measurement yields $m$ , the state is the normalized projection $|\psi_m\rangle=P_m|\psi\rangle/\sqrt{p(m)}$ , lying entirely inside the eigenspace of $P_m$ . Why: the collapse rule (3) is the input to the second round, and its whole content is that the surviving vector already sits in one bin.

Step 2. Apply the Born rule (2) to this new state: $p(m'\,|\,m)=\langle\psi_m|P_{m'}|\psi_m\rangle=\langle\psi|P_mP_{m'}P_m|\psi\rangle/p(m)$ . Why: the second measurement knows nothing of history except through the state handed to it, so we feed $|\psi_m\rangle$ back into the same formula.

Step 3. Collapse the sandwich using orthogonality and idempotence: $P_mP_{m'}P_m=\delta_{mm'}P_m$ , since for $m'\neq m$ the middle factor annihilates ( $P_mP_{m'}=0$ ) and for $m'=m$ the three copies fuse. Why: this identity is where "perpendicular, exhaustive bins" becomes "no leakage between outcomes" — the algebraic heart of sharpness.

Step 4. Substitute back: $p(m'\,|\,m)=\delta_{mm'}\,\langle\psi|P_m|\psi\rangle/p(m)=\delta_{mm'}$ , since $\langle\psi|P_m|\psi\rangle=p(m)$ cancels the denominator. Why: the cancellation is exact and state-independent, so repeatability holds for any input — a structural fact about projective measurements, not a property of special states. *Verified: algebraic identity from $P_mP_{m'}=\delta_{mm'}P_m$ ; consistent with Nielsen & Chuang §2.2.5. A generic POVM has no such update rule and is not repeatable.*

Go deeper · Derivation B · 1 minThe variance of an observable and when it vanishes

Derivation BThe variance of an observable and when it vanishes

The aim is to show that the statistical spread of the measured eigenvalues is exactly the operator variance $\langle A^2\rangle-\langle A\rangle^2$ , and that it collapses to zero if and only if the input is an eigenstate of the measured observable — the quantitative content of "an eigenstate gives a definite value" and the input to every uncertainty relation. We treat the outcome $m$ as a random variable distributed as $p(m)$ and compute its variance two ways.

Step 1. Compute the second moment. Since $A=\sum_m m\,P_m$ and the projectors are unchanged under powers, $A^2=\sum_m m^2\,P_m$ , so $\langle A^2\rangle=\langle\psi|A^2|\psi\rangle=\sum_m m^2\,p(m)$ . Why: squaring the observable squares the eigenvalues but reuses the same eigenspaces, which is what lets the operator second moment equal the ordinary statistical second moment of the outcomes.

Step 2. Form the variance in operator language, $(\Delta A)^2=\langle A^2\rangle-\langle A\rangle^2$ , the standard spread of the random outcome. Why: this is what an experiment estimates by repeating the measurement on identically prepared states and taking the sample variance.

Step 3. Expand the centered form and match: $\sum_m m^2 p(m)-\big(\sum_m m\,p(m)\big)^2=\sum_m (m-\langle A\rangle)^2\,p(m)$ , a manifestly nonnegative weighted sum of squared deviations. Why: as an average of squared deviations the sign is obvious, and a sum of nonnegative terms is zero only when every populated term is.

Step 4. Read off the vanishing case: $(\Delta A)^2=0$ iff every outcome with $p(m)>0$ has $m=\langle A\rangle$ , i.e. the state populates a single eigenspace and $|\psi\rangle$ is an eigenstate of $A$ . Why: this is the exact statement of "definite value" — measurement is deterministic precisely on eigenstates, and dispersion witnesses superposition across eigenvalues. *Verified: matches the observable-variance definition in Nielsen & Chuang §2.2.5 and underlies the Robertson uncertainty relation.*

Go deeper · Derivation C · 1 minAny orthonormal-basis readout is a projective measurement

Derivation CAny orthonormal-basis readout is a projective measurement

The goal is to establish that "reading out the state in a basis" — the operation ending essentially every quantum circuit — is nothing more than the projective measurement whose projectors are the rank-one outer products of the basis vectors, and that this reproduces the amplitude-squared Born rule and a collapse to a single basis state up to phase. This closes the loop between the abstract projector formalism and the concrete "amplitudes give probabilities" rule, and explains why a computational-basis readout is the canonical sharp measurement.

Step 1. Take any orthonormal basis $\{|u_m\rangle\}$ of $\mathcal H$ and build the projectors $P_m=|u_m\rangle\langle u_m|$ ; orthonormality gives $P_mP_{m'}=\delta_{mm'}P_m$ . Why: a rank-one outer product of a unit vector is automatically an idempotent Hermitian projector, and basis orthonormality is exactly the orthogonality condition (1).

Step 2. Check completeness: $\sum_m P_m=\sum_m|u_m\rangle\langle u_m|=I$ , the resolution of the identity in any orthonormal basis. Why: completeness of the basis and of the projectors are the same statement, so any full basis is a legitimate measurement with no gaps.

Step 3. Compute the probability via (2): $p(m)=\langle\psi|u_m\rangle\langle u_m|\psi\rangle=|\langle u_m|\psi\rangle|^2$ , the squared amplitude of $|\psi\rangle$ along $|u_m\rangle$ . Why: this is the original Born rule for basis amplitudes, recovered as a special case of the projector formula — the two "Born rules" were always one.

Step 4. Compute the collapse via (3): $|\psi_m\rangle=|u_m\rangle\langle u_m|\psi\rangle/|\langle u_m|\psi\rangle|=e^{i\theta}|u_m\rangle$ with $\theta=\arg\langle u_m|\psi\rangle$ — the basis vector up to an unobservable global phase. Why: for a rank-one projector the outcome pins the state to a single ray, which is why non-degenerate measurements are maximally informative and degeneracy is the only way to lose that uniqueness. *Verified: standard; the global phase is physically irrelevant, and the computational basis gives the qubit readout $P_0=|0\rangle\langle0|,\ P_1=|1\rangle\langle1|$ .*

Limiting cases

Familiar scenarios are the special shapes the projector set can take — rank-one everywhere is sharpest, a single identity projector emptiest.

Case	Result	Meaning
Rank-one / non-degenerate	$P_m=\vert m\rangle\langle m\vert,\ \ p(m)=\vert\langle m\vert\psi\rangle\vert^2$	Orthonormal-basis measurement; the standard Born rule, and the collapsed state is exactly $\vert m\rangle$ (one ray per outcome).
Computational basis (qubit)	$P_0=\vert0\rangle\langle0\vert,\ P_1=\vert1\rangle\langle1\vert,\ p(0)=\vert\alpha\vert^2$	For $\vert\psi\rangle=\alpha\vert0\rangle+\beta\vert1\rangle$ , the canonical single-qubit readout ending every circuit.
Degenerate eigenvalue	$P_m=\sum_k\vert m,k\rangle\langle m,k\vert,\ \ \mathrm{rank}\,P_m>1$	Eigenspace projector; the outcome fixes only the subspace, not a single vector — the collapse keeps the whole projected component.
Trivial measurement	$P_1=I,\ \ p(1)=1$	One projector equal to the identity: the reading is certain, reveals nothing, and leaves the state untouched.
Non-selective (unread)	$\rho'=\sum_m P_m\rho P_m$	Outcome discarded: a dephasing channel that kills coherences between eigenspaces — decoherence into the pointer basis.

Common trapMeasuring a quantum state just tells you the state it was already in.

It does not, and this is the deepest trap. A measurement returns one eigenvalue, not the state, and for any superposition across eigenspaces the eigenvalue is genuinely random — $|\psi\rangle=\tfrac{1}{\sqrt2}(|0\rangle+|1\rangle)$ gives $Z$ -outcome $0$ or $1$ equally, and from a single reading you cannot reconstruct the amplitudes. Worse, the measurement changes the state: afterward the qubit is $|0\rangle$ or $|1\rangle$ , and the phase that made it $|+\rangle$ rather than $|-\rangle$ is gone forever. Two further pitfalls: the projectors must be both orthogonal and complete — dropping orthogonality gives a POVM (no repeatability, no unique collapse), dropping completeness makes probabilities miss unity — and $P_m$ need not be rank-one, so for a degenerate eigenvalue the outcome fixes only the projection into the eigenspace. "The measurement reveals a pre-existing value" is true only on eigenstates, exactly where the variance vanishes.

Worked examples

1 · Spin measured along a tilted axis (fully worked)

A spin-½ atom is prepared in $|0\rangle$ (spin-up along $z$ ). We measure its spin along the axis $\hat n$ tilted $\theta=60^\circ$ from $z$ in the $x\!-\!z$ plane, i.e. we measure $\hat n\cdot\vec\sigma$ . Find the outcome probabilities, the post-measurement states, and the expectation value.

Step 1. Build the projectors. With $\hat n=(\sin\theta,0,\cos\theta)$ , equation (8) gives $P^\pm=\tfrac12(I\pm\hat n\cdot\vec\sigma)$ with $\hat n\cdot\vec\sigma=\sin\theta\,X+\cos\theta\,Z$ ; sandwiching in $|0\rangle$ uses $\langle0|X|0\rangle=0$ , $\langle0|Z|0\rangle=1$ .

Step 2. Probability of the $+1$ outcome: $p(+)=\langle0|P^+|0\rangle=\tfrac12(1+\cos\theta)=\cos^2(\theta/2)$ . For $\theta=60^\circ$ , $\cos^2 30^\circ=\tfrac34=\mathbf{0.75}$ , and $p(-)=\sin^2(\theta/2)=\tfrac14=\mathbf{0.25}$ .

Step 3. Post-measurement states. Outcome $+1$ collapses the atom to the $+1$ eigenstate $|{+\hat n}\rangle=\cos(\theta/2)|0\rangle+\sin(\theta/2)|1\rangle=\tfrac{\sqrt3}{2}|0\rangle+\tfrac12|1\rangle$ ; outcome $-1$ to the orthogonal $|{-\hat n}\rangle$ . Each is rank-one, so the collapsed state is unique up to phase.

Step 4. Expectation value via (5): $\langle\hat n\cdot\vec\sigma\rangle=(+1)(0.75)+(-1)(0.25)=\mathbf{0.5}$ , which equals $\cos\theta=\cos60^\circ$ directly — a clean check. A Stern–Gerlach magnet along $\hat n$ would split the atoms $75\%$ up / $25\%$ down. p(+) = 0.75, p(−) = 0.25, ⟨n·σ⟩ = 0.5 = cos 60°.

2 · A degenerate parity measurement builds a Bell state (fully worked)

Two qubits are in the product state $|++\rangle=\tfrac12(|00\rangle+|01\rangle+|10\rangle+|11\rangle)$ . We measure the parity observable $Z\otimes Z$ , whose eigenvalue $+1$ spans $\{|00\rangle,|11\rangle\}$ and whose eigenvalue $-1$ spans $\{|01\rangle,|10\rangle\}$ . Find the outcome probabilities and post-measurement states.

Step 1. The rank-two projectors are $P_{+}=|00\rangle\langle00|+|11\rangle\langle11|$ and $P_{-}=|01\rangle\langle01|+|10\rangle\langle10|$ ; they are orthogonal and $P_++P_-=I$ . This is the higher-rank case — the outcome fixes a subspace, not a ray.

Step 2. Probability of even parity: $P_+|++\rangle=\tfrac12(|00\rangle+|11\rangle)$ , so $p(+1)=\lVert P_+|++\rangle\rVert^2=\big(\tfrac12\big)^2+\big(\tfrac12\big)^2=\tfrac14+\tfrac14=\mathbf{0.5}$ . By symmetry $p(-1)=\mathbf{0.5}$ .

Step 3. Post-measurement state for $+1$ via (3): $|\psi_{+}\rangle=\tfrac12(|00\rangle+|11\rangle)/\sqrt{1/2}=\tfrac{1}{\sqrt2}(|00\rangle+|11\rangle)$ — the Bell state $|\Phi^+\rangle$ . Outcome $-1$ gives $\tfrac{1}{\sqrt2}(|01\rangle+|10\rangle)=|\Psi^+\rangle$ .

Step 4. Read the lesson. A degenerate projective measurement turned an unentangled state into a maximally entangled one, and the outcome ( $\pm1$ ) did not determine a unique vector — the input's projection into the eigenspace fixed the Bell state. This is exactly the mechanism behind stabilizer/syndrome measurements, which entangle the code block without revealing the logical information. p(+1) = p(−1) = 0.5; even parity collapses |++⟩ onto the Bell state (1/√2)(|00⟩+|11⟩).

3 · The observable, not the state, decides the randomness

The qubit $|\psi\rangle=|+\rangle=\tfrac{1}{\sqrt2}(|0\rangle+|1\rangle)$ is measured once in $Z$ and, on a fresh copy, once in $X$ . For $Z$ (projectors $|0\rangle\langle0|,|1\rangle\langle1|$ ): $p(0)=p(1)=\tfrac12$ , collapsing to $|0\rangle$ or $|1\rangle$ . For $X$ (projectors $|+\rangle\langle+|,|-\rangle\langle-|$ ): $|\psi\rangle$ is the $+1$ eigenstate, so outcome $+1$ occurs with probability $1$ and the state is unchanged. Same state, but $Z$ is maximally random ( $(\Delta Z)^2=1$ ) and $X$ is certain ( $(\Delta X)^2=0$ ): the observable sets the randomness.

Physics of the portrait

The projection postulate — the rule that a measurement resets the state to the projected, renormalized branch — is not a theorem inside unitary quantum mechanics; it was added as a second, non-unitary kind of evolution. The person who drew that line was John von Neumann.

Origins · 1 minJohn von Neumann and how this was discovered

John von Neumann

1903–1957 · mathematical foundations of quantum measurement

Von Neumann was the mathematician who turned quantum mechanics from a working recipe into an axiomatic theory. In his 1932 Mathematische Grundlagen der Quantenmechanik he built the subject on abstract Hilbert space, identified observables with self-adjoint operators, and proved the spectral theorem for unbounded self-adjoint operators (1929–1932) — exactly the decomposition that makes "measuring an observable" mean "performing a projective measurement." His most controversial move split quantum dynamics into two irreconcilable processes: continuous, deterministic, reversible unitary evolution (Process 2), and the discontinuous, probabilistic, irreversible state reduction on measurement (Process 1) — the projection postulate, located somewhere along a von Neumann chain from apparatus to observer. He also introduced the von Neumann entropy $S=-\mathrm{Tr}(\rho\log\rho)$ and the density-operator formalism used here. Gerhart Lüders (1951) later corrected the update rule for degenerate outcomes — the Lüders rule of equations (6)–(7) — and Naimark's dilation theorem (1940) showed that sharp measurements are, on a larger space, the whole story: every generalized measurement is a projective one in disguise. His framework remains the standard language of quantum measurement today.

History

1926

— Born gives the probabilistic interpretation of the wave function

1929

— von Neumann proves the spectral theorem for unbounded self-adjoint operators

1932

— von Neumann's Grundlagen states the projection postulate and separates measurement from unitary evolution

1940

— Naimark's dilation theorem: every generalized measurement is a projective one on a larger space

1951

— Lüders gives the correct state-update rule for degenerate projective measurements

Going further

Texts. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.2.3–2.2.6 (postulates, projective measurements, POVMs) · Preskill, Quantum Information lecture notes, 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. 9 · Mermin, Quantum Computer Science, Ch. 1.

Provenance. The defining conditions and probability/collapse rules follow Nielsen & Chuang §2.2.5; the spectral decomposition $A=\sum_m m\,P_m$ and $\sum_m p(m)=1$ are Watrous Ch. 2 and Wilde Ch. 4. A verifier should recompute the repeatability identity $P_mP_{m'}P_m=\delta_{mm'}P_m$ ; the variance $(\Delta A)^2=\langle A^2\rangle-\langle A\rangle^2$ vanishing exactly on eigenstates; the spin example ( $p(\pm)=\cos^2/\sin^2(\theta/2)$ giving $0.75/0.25$ at $\theta=60^\circ$ , $\langle\hat n\cdot\vec\sigma\rangle=\cos\theta=0.5$ ); the parity example ( $|++\rangle$ under $Z\otimes Z$ gives $p(\pm1)=\tfrac12$ and Bell states $|\Phi^+\rangle,|\Psi^+\rangle$ ); and $P^\pm_{\hat n}=\tfrac12(I\pm\hat n\cdot\vec\sigma)$ orthogonal and summing to $I$ . The Lüders channel is standard (Wilde Ch. 4). No nonstandard constants appear.

Prerequisites. The Measurement Postulate · Born Rule · Observables and Hermitian Operators · Inner Product, Norm, and Normalization · Dirac Bra-Ket Notation. Leads to. Expectation Values and Variance · Density Operators · Pure versus Mixed States · Pauli Operators.

Ask the tutor: Want me to walk through why a degenerate stabilizer measurement can entangle a code block while leaving the logical qubit untouched, step by step?

What is the difference between a projective measurement and a general (POVM) measurement?

A POVM is any collection of positive operators $\{E_m\}$ with $\sum_m E_m=I$ ; it gives valid probabilities $p(m)=\langle\psi|E_m|\psi\rangle$ but says nothing on its own about the post-measurement state. A projective measurement is the special case where each $E_m=P_m$ is an orthogonal projector. That extra structure buys three things POVMs lack: repeatability, a canonical collapse rule (3), and a cap on the number of outcomes at the dimension. POVMs can have more outcomes than dimensions and are the right tool for tasks like unambiguous state discrimination. The unifying fact is Naimark's theorem: every POVM is a projective measurement on a larger space (system plus ancilla), so projective measurements are the universal building block.

Why must the projectors be both orthogonal and complete?

You need both, for different reasons. Completeness, $\sum_m P_m=I$ , makes the probabilities sum to one: $\sum_m p(m)=\langle\psi|\big(\sum_m P_m\big)|\psi\rangle=\langle\psi|\psi\rangle=1$ . Without it, some of the state's length would fall into no bin and probability would leak away. Orthogonality, $P_mP_{m'}=0$ for $m\neq m'$ , makes the outcomes mutually exclusive and delivers repeatability and a well-defined collapse — it forces the cross-terms to vanish in the repeatability derivation. Drop orthogonality and you still have a legitimate POVM, but you lose the sharp update rule and the guarantee that a repeat reading confirms the first. Together with Hermiticity and idempotence, the two conditions are exactly what make each $p(m)=\lVert P_m|\psi\rangle\rVert^2$ a genuine, normalized probability.

Does the outcome of a measurement always determine the post-measurement state?

Only when the projector is rank-one. For a non-degenerate eigenvalue, $P_m=|m\rangle\langle m|$ and the collapse (3) sends the state to $|m\rangle$ up to phase — the outcome pins a single ray. For a degenerate eigenvalue the projector has rank $>1$ , and the collapsed state $P_m|\psi\rangle/\sqrt{p(m)}$ still depends on the input: the outcome tells you which eigenspace you are in, not which vector inside it. This is not a defect — it is exactly what stabilizer measurements exploit, learning a syndrome (the subspace) while leaving the logical state (the vector within it) undisturbed. So "the measurement determines the state" is true for non-degenerate readouts and false, usefully, for degenerate ones.

How is measuring a physical observable connected to a set of projectors?

Through the spectral theorem. Any Hermitian operator $A$ can be written $A=\sum_m m\,P_m$ , where the $m$ are its real eigenvalues and the $P_m$ project onto the eigenspaces and satisfy conditions (1) automatically. "Measuring $A$ " means performing the projective measurement $\{P_m\}$ and recording the eigenvalue $m$ that labels the bin you landed in — the projectors are the measurement, the eigenvalues just the numbers on the dial. A consequence is that $A$ , $A^2$ , and any function $f(A)$ share the same projectors and are measured by the same apparatus; only the label changes, and for injective $f$ no information is lost.

What happens if I measure but do not look at the result?

You get a non-selective measurement: the state becomes the probability-weighted average of the conditional branches, the Lüders channel $\rho\mapsto\rho'=\sum_m P_m\rho P_m$ . This keeps the diagonal blocks of $\rho$ within each eigenspace and erases the coherences between distinct eigenspaces — a completely positive, trace-preserving dephasing channel, the same mathematics as environment-induced decoherence. So an unread measurement still has an effect: it does not select a branch, but it destroys interference between the observable's eigenspaces, which is why measurement-like coupling to an environment makes a system behave classically in the pointer basis.

Is state collapse a physical process or just an update of knowledge?

This is the measurement problem, and the formalism above is deliberately neutral on it. Operationally, the projection postulate predicts the statistics of subsequent measurements perfectly, whichever interpretation you hold. In the Copenhagen reading, collapse is a primitive, irreversible process — von Neumann's Process 1 — distinct from unitary evolution. In many-worlds there is no collapse: apparatus and observer simply entangle with the system, and each "outcome" is a branch of one unitary state. In decoherence-based and epistemic accounts, the Lüders update is your revised description conditioned on what has been learned. All reproduce $p(m)=\langle\psi|P_m|\psi\rangle$ and the update (3); they differ only on what the update is. To predict what a lab sees, you can use the projective rules without settling the question.