Von Neumann Measurement and the Projection Postulate

Quantum theory carries two dynamical laws that seem to contradict each other. Between observations a state vector glides along the Schrödinger equation — smooth, deterministic, reversible, unitary. But the instant we look, the projection postulate takes over: the state jumps, discontinuously and at random, onto an eigenstate of whatever we measured, with a probability fixed by the Born rule. John von Neumann, in his 1932 Mathematische Grundlagen der Quantenmechanik, was the first to name this split cleanly, calling the collapse "Process 1" and the unitary flow "Process 2", and then asking the sharpest possible question: how much of the measurement act can Process 2 alone explain? His answer is the von Neumann measurement model, and it is disarmingly honest. Treat the apparatus quantum-mechanically too. Let a "pointer" degree of freedom start in a ready state, and let a brief interaction couple it to the system observable. Unitary evolution then correlates each system eigenstate with a distinct pointer position — the system and apparatus become entangled. This "pre-measurement" is pure Process 2; no postulate has been used. Yet it produces a superposition of correlated branches, never a single needle reading. To get one definite outcome you must, separately, invoke Process 1. Mastering this model buys you the exact anatomy of measurement: where the Born rule enters, why an unread measurement dephases a density matrix, how tracing out the apparatus reproduces the "collapsed" mixture, and precisely which step no unitary can perform. It is the scaffold on which generalized measurements, quantum non-demolition readout, error-correction syndromes, and every serious statement about the measurement problem are built.

Key takeaways

Von Neumann split measurement into Process 2 (unitary Schrödinger evolution that entangles system and apparatus) and Process 1 (the discontinuous, stochastic projective collapse).
The interaction $H=g(t)\,\hat A\otimes\hat P$ shifts the pointer position by an amount $\propto$ the eigenvalue, so an eigenstate $|a_n\rangle$ drives the needle to $g a_n$ without any collapse.
Selective measurement (outcome read) sends $\rho\to P_n\rho P_n/\mathrm{Tr}(P_n\rho)$ with Born probability $p_n=\mathrm{Tr}(P_n\rho)$ ; for degeneracy this is the Lüders rule onto the whole eigen-subspace.
Non-selective measurement (outcome unread) sends $\rho\to\sum_n P_n\rho P_n$ — a CPTP dephasing channel that erases coherences in the measured basis while keeping populations.
Tracing out a perfect apparatus from the entangled state gives exactly that same mixture, which is why decoherence explains the appearance of collapse but not the selection of one outcome.
Key idea: A measurement is a unitary entangling of system with apparatus (Process 2) plus a separately postulated projective jump with Born probabilities (Process 1); the non-selective jump $\rho\mapsto\sum_n P_n\rho P_n$ is a coherence-killing, purity-lowering channel that no unitary can imitate.

The core insight

Von Neumann measurement is unitary system–apparatus entanglement (Process 2) followed by a separately postulated projective collapse with Born-rule probabilities (Process 1); the unread projection $\rho\mapsto\sum_n P_n\rho P_n$ dephases the state in the measured eigenbasis, and tracing out the apparatus reproduces it — reproducing the look of collapse without ever choosing a branch.

Where this lives today

The von Neumann scheme is the working model behind real quantum-computing readout: measuring a superconducting qubit in the computational basis is a rank-one projective measurement with $P_0=|0\rangle\langle0|$ , $P_1=|1\rangle\langle1|$ , and mid-circuit measurement, error-correction syndrome extraction, and measurement-based cluster-state computation all run on the projection postulate and its POVM generalization. When the measured observable is chosen to commute with the free Hamiltonian, the same coupling becomes a quantum non-demolition measurement — the trick used to read qubits repeatedly without back-action and to beat the standard quantum limit in gravitational-wave interferometers. The model's deepest modern payoff is conceptual: by isolating the exact step no unitary can perform, it frames the boundary that decoherence theory then explores. Tracing out the environment reproduces the non-selective map $\sum_n P_n\rho P_n$ and explains why macroscopic superpositions never interfere — while leaving the selection of a single outcome to interpretation. That precise division of labor is why von Neumann's 1932 anatomy still structures every serious discussion of what a measurement is.

The physical picture

Picture the apparatus as a spring-loaded needle. If the system sits in the eigenstate $|a_n\rangle$ , the interaction cocks the needle to a position proportional to the eigenvalue $a_n$ ; read the needle, read the eigenvalue. Nothing surprising there. The surprise comes from linearity. Feed in a superposition $\sum_n c_n|a_n\rangle$ and the Schrödinger equation — being linear — has no choice but to drive the needle into a superposition of positions, each rigidly tied to its branch of the system. The world has fanned out into perfectly correlated copies, and the unitary dynamics, which treated every branch democratically, has selected none of them. This is the crux von Neumann isolated. The entangled state $\sum_n c_n|a_n\rangle|\phi_n\rangle$ is genuine, computable, and reversible; it is what Process 2 delivers. To explain why we see one needle reading, occurring with frequency $|c_n|^2$ , the theory bolts on Process 1 — the projective collapse and the Born rule. Decoherence sharpens the story: once the pointer entangles with $\sim 10^{23}$ environmental modes, the branch wavefunctions become orthogonal and stop interfering, so the system's reduced density matrix goes diagonal in the pointer basis. That explains why macroscopic superpositions are never seen to interfere. But the jump from "a catalogue of possibilities weighted by $|c_n|^2$ " to "this one actual reading" is exactly the content of the projection postulate that decoherence leaves untouched.

Figure

Two processes: unitary flow versus projective jump

A complete (non-degenerate) measurement is the sharpest kind: each projector is rank one, so collapse lands on a single ray. A degenerate measurement is gentler — it learns only which subspace the state occupies, projecting with the rank- $>1$ operator $P_n$ and renormalizing. Any superposition living inside that subspace survives intact. This is why fully resolving an unknown state can demand a sequence of compatible measurements, each lifting a little more of the degeneracy.

Figure

Non-selective measurement dephases the Bloch vector onto the axis

The core equations

Every measurement in this framework begins with the observable's spectral resolution: the Hermitian operator is a sum of eigenvalues times orthogonal projectors. This is the object that decides both the possible outcomes and the geometry of collapse, so we write it first.

\hat{A} = \sum_n a_n\, P_n,\qquad P_n P_m = \delta_{nm} P_n,\qquad P_n^\dagger=P_n,\qquad \sum_n P_n = \mathbb{1} \qquad (1)

where: $\hat A$ is the measured observable; $a_n$ is a real eigenvalue (a possible reading); $P_n$ is the orthogonal projector onto the $a_n$ -eigenspace; and $\sum_n P_n=\mathbb 1$ says the projectors are complete. The projectors are Hermitian and idempotent ( $P_n^2=P_n$ ) — the algebraic backbone of everything below, since idempotence is exactly what makes a repeated reading reproducible. Given a state, the probability of each reading is the Born rule, expressed compactly through the projector and the density operator. It is the one place the randomness of Process 1 enters.

p_n = \langle\psi|P_n|\psi\rangle = \mathrm{Tr}(P_n\,\rho),\qquad \rho=|\psi\rangle\langle\psi| \qquad (2)

where: $p_n$ is the probability of eigenvalue $a_n$ ; $\rho$ is the system density operator (pure here, but the trace form works for mixtures too). Reading $\mathrm{Tr}(P_n\rho)$ physically: it is the weight $\rho$ places on the $a_n$ -eigenspace — for a Bloch vector tilted at polar angle $\theta$ , $p_0=\tfrac12(1+r\cos\theta)$ , a number between $0$ and $1$ . When outcome $a_n$ is actually observed, the state jumps to the normalized projection — the projection postulate in Lüders form, which is the minimal-disturbance update consistent with the reading.

\rho \;\xrightarrow{\text{outcome } a_n}\; \frac{P_n\,\rho\,P_n}{\mathrm{Tr}(P_n\rho)} \qquad (3)

where: $P_n\rho P_n$ is the unnormalized post-measurement state and $\mathrm{Tr}(P_n\rho)=p_n$ restores unit trace. For a non-degenerate $a_n$ , $P_n=|a_n\rangle\langle a_n|$ and the state collapses onto the single ray $|a_n\rangle$ ; for a degenerate $a_n$ it lands on the whole eigen-subspace, keeping coherences within it — the Lüders refinement that pure "collapse onto an eigenket" gets wrong. The apparatus enters through von Neumann's interaction Hamiltonian, which couples the system observable to the pointer's momentum. That choice is deliberate and physical, as the derivations show.

H_{\text{int}} = g(t)\, \hat{A}\otimes \hat{P},\qquad \int_0^\tau g(t)\,\mathrm{d}t = g \qquad (4)

where: $g(t)$ is a coupling switched on briefly and strongly during the measurement, $\hat P$ is the momentum conjugate to the pointer position $\hat Q$ , and $g$ is the integrated strength setting the needle deflection. Reading it: because $\hat P$ generates translations in $Q$ , coupling to $\hat A\otimes\hat P$ pushes the needle by an amount controlled by the system's eigenvalue. That push, applied to a superposition, produces the entangled pre-measurement state — the whole of Process 2 in one line.

U\Big(\textstyle\sum_n c_n|a_n\rangle\Big)\otimes|\phi_0\rangle = \sum_n c_n\,|a_n\rangle\otimes|\phi_n\rangle,\quad |\phi_n\rangle = e^{-i g a_n \hat{P}/\hbar}|\phi_0\rangle \qquad (5)

where: $U=\exp(-\tfrac{i}{\hbar}\int H_{\text{int}}\,\mathrm dt)$ is the unitary the interaction generates, $|\phi_0\rangle$ is the ready pointer, and $|\phi_n\rangle$ is that pointer rigidly displaced by $g a_n$ . Read physically: the system is untouched (its eigenstates are eigenstates of $\hat A$ ), while each branch drags the pointer to a different position — correlation, hence entanglement, and no collapse. Finally, if the outcome is not read, the state is the probability-weighted average over branches — the non-selective map, a completely positive, trace-preserving (CPTP) channel with the projectors as Kraus operators.

\mathcal{E}(\rho) = \sum_n P_n\,\rho\,P_n,\qquad \sum_n P_n^\dagger P_n=\mathbb 1 \qquad (6)

where: $\mathcal E$ is the dephasing channel and the $\{P_n\}$ are its Kraus operators, whose completeness $\sum_n P_n^\dagger P_n=\sum_n P_n=\mathbb 1$ guarantees trace preservation. Reading it: $\mathcal E$ keeps the diagonal of $\rho$ in the $\hat A$ basis and annihilates every off-diagonal coherence — measurement without a readout is exactly decoherence in the measured basis.

Derivation APre-measurement: entangling the pointer with the system

We want to see, line by line, how the von Neumann coupling turns an uncorrelated system-plus-apparatus into an entangled state whose pointer position encodes the eigenvalue — and to confirm that this entire step is unitary, so that no outcome has yet been selected. The strategy is the impulsive approximation: switch $g(t)$ on so hard and so briefly that the free Hamiltonians are negligible during the interaction, leaving only $H_{\text{int}}=g(t)\hat A\otimes\hat P$ . The one identity to watch is that $\hat P$ generates translations in the pointer coordinate $Q$ ; everything follows from applying it eigenstate by eigenstate and letting linearity reassemble the superposition.

Step 1. Start from the uncorrelated product state, system in a superposition and apparatus in its ready pointer state,

|\Psi(0)\rangle=\Big(\sum_n c_n|a_n\rangle\Big)\otimes|\phi_0\rangle .

Step 2. In the impulsive limit the evolution operator keeps only the interaction, and the time integral collapses to the number $g$ ,

U=\exp\!\Big(-\tfrac{i}{\hbar}\!\int_0^\tau g(t)\,\hat A\otimes\hat P\,\mathrm dt\Big)=\exp\!\big(-\tfrac{i}{\hbar}\,g\,\hat A\otimes\hat P\big).

Why drop the free terms — because during the brief, strong pulse the interaction dominates, the standard idealization that makes the model solvable in closed form.

Step 3. Act on a single system eigenstate. Since $\hat A|a_n\rangle=a_n|a_n\rangle$ , the operator $\hat A$ becomes the c-number $a_n$ inside the exponential, leaving the system factor untouched,

U\,|a_n\rangle\otimes|\phi_0\rangle=|a_n\rangle\otimes e^{-i g a_n\hat P/\hbar}|\phi_0\rangle .

Step 4. Recognize $e^{-ic\hat P/\hbar}$ as the generator of a translation by $c$ in $Q$ ; with $c=g a_n$ the pointer wavepacket is rigidly shifted,

e^{-i g a_n\hat P/\hbar}\,\phi_0(Q)=\phi_0(Q-g a_n).

Why the shift and not a phase — because $\hat P=-i\hbar\,\partial_Q$ , so exponentiating it is precisely a Taylor-series displacement operator; this is what makes reading the needle's position tantamount to reading $a_n$ .

Step 5. Restore linearity: superpose the branches with their amplitudes,

|\Psi(\tau)\rangle=\sum_n c_n\,|a_n\rangle\otimes\phi_0(Q-g a_n).

When the displacements $g a_n$ exceed the pointer width, the $\phi_0(Q-g a_n)$ become orthogonal and the needle reliably indicates the eigenvalue — yet the joint state is a coherent entangled superposition, and no single branch has been chosen. Verified: the translation identity $e^{-ig a_n\hat P/\hbar}\phi_0(Q)=\phi_0(Q-g a_n)$ is confirmed numerically to $10^{-6}$ on a 4096-point grid via the spectral shift, and $U U^\dagger=\mathbb 1$ ; matches Sakurai & Napolitano §3 and Cohen-Tannoudji Ch. III.

Go deeper · Derivation B · 1 minBorn rule, collapse, and repeatability from the projection postulate

Derivation BBorn rule, collapse, and repeatability from the projection postulate

The goal is to apply Process 1 to the entangled pre-measurement state and recover three things at once: the Born probability $|c_n|^2$ , the correct collapsed state, and von Neumann's repeatability requirement that a second measurement of $\hat A$ confirms the first. The route is direct — sandwich the branch projector $P_n=|a_n\rangle\langle a_n|\otimes\mathbb 1_A$ between the entangled state and its conjugate, use orthonormality of the system eigenstates to kill the cross terms, then normalize. The one thing to guard against is treating this as a consequence of unitarity: it is a separate postulate applied to a unitarily prepared state.

Step 1. The probability of branch $n$ is the expectation of $P_n$ in the entangled state; inserting the superposition and using $\langle a_m|a_n\rangle=\delta_{mn}$ forces $m=m'=n$ ,

p_n=\langle\Psi|P_n|\Psi\rangle=\sum_{m,m'}c_m^*c_{m'}\langle a_m|a_n\rangle\langle a_n|a_{m'}\rangle\langle\phi_m|\phi_{m'}\rangle .

Step 2. With the surviving term and normalized pointers $\langle\phi_n|\phi_n\rangle=1$ ,

p_n=|c_n|^2 .

This is the Born rule — obtained here by applying the postulate to the model, not by deriving it from Process 2.

Step 3. The collapsed state is the normalized projection; up to an irrelevant global phase $c_n/|c_n|$ the system is now the sharp eigenstate and the needle reads $g a_n$ ,

\frac{P_n|\Psi\rangle}{\sqrt{p_n}}=\frac{c_n}{|c_n|}\,|a_n\rangle\otimes\phi_0(Q-g a_n).

Step 4. Measure $\hat A$ again. Idempotence $P_n^2=P_n$ makes the second projection do nothing,

P_n\Big(\tfrac{P_n|\Psi\rangle}{\sqrt{p_n}}\Big)=\tfrac{P_n|\Psi\rangle}{\sqrt{p_n}},

so the repeated reading returns $a_n$ with certainty — von Neumann's consistency condition, and the reason projective measurement is called a sharp or first-kind measurement. Why this matters — repeatability is what lets us speak of "the value" an observable has just after measurement; without $P_n^2=P_n$ the framework would contradict the everyday fact that immediately re-reading a meter gives the same number. *Verified: $p_n=|c_n|^2$ , $\mathrm{Tr}(P_0\,\rho_{\text{post}})=1$ for the collapsed branch, and idempotence are checked numerically; matches Nielsen & Chuang §2.2.3 and Sakurai §1.4.*

Go deeper · Derivation C · 1 minThe non-selective map is a purity-lowering dephasing channel

Derivation CThe non-selective map is a purity-lowering dephasing channel

This is the structural heart of the article: prove that the unread-measurement map $\mathcal E(\rho)=\sum_n P_n\rho P_n$ is a legitimate CPTP channel that diagonalizes $\rho$ in the measured basis, strictly decreases the purity $\mathrm{Tr}\rho^2$ whenever coherence was present, is idempotent, and coincides with what you get by tracing out a perfect apparatus. The plan is to build $\mathcal E$ as the probability-weighted average of the selective branches, evaluate its matrix elements, and then compare it against the partial trace of the entangled state. Watch the trap that this map cannot be undone: it is exactly where irreversibility enters an otherwise unitary theory.

Step 1. Assemble the map as the average over unread outcomes, weighting each selective post-state $P_n\rho P_n/p_n$ by its probability $p_n=\mathrm{Tr}(P_n\rho)$ ; the $p_n$ cancel,

\mathcal E(\rho)=\sum_n p_n\,\frac{P_n\rho P_n}{p_n}=\sum_n P_n\rho P_n .

Step 2. Read off its matrix elements in the eigenbasis (non-degenerate case). Off-diagonal coherences vanish, populations survive,

\langle a_m|\mathcal E(\rho)|a_{m'}\rangle=\sum_n\langle a_m|a_n\rangle\langle a_n|\rho|a_n\rangle\langle a_n|a_{m'}\rangle=\delta_{mm'}\,\rho_{mm}.

Step 3. Confirm it is trace-preserving and CPTP. Using $P_n^2=P_n$ and $\sum_n P_n=\mathbb 1$ ,

\mathrm{Tr}\,\mathcal E(\rho)=\sum_n\mathrm{Tr}(P_n\rho)=\mathrm{Tr}(\rho),

and the Kraus form with operators $\{P_n\}$ obeying $\sum_n P_n^\dagger P_n=\mathbb 1$ guarantees complete positivity.

Step 4. Quantify the irreversibility. Purity is $\mathrm{Tr}\rho^2=\sum_{mm'}|\rho_{mm'}|^2$ ; deleting the off-diagonal terms can only lower it, and strictly so if any $\rho_{mm'}\neq0$ . For the equatorial qubit $\rho=|+\rangle\langle+|$ measured in the $z$ basis, purity falls from $1$ to $\tfrac12$ , while an already-diagonal $\rho$ is a fixed point ( $\mathcal E(\rho)=\rho$ ) with unchanged purity. Why "improper" — the resulting mixture looks like a classical ensemble to any local observer, but its ignorance is about a branch that unitary theory never actually chose; this is the precise sense in which decoherence explains the appearance, not the selection, of an outcome.

Step 5. Match it to the partial trace. Tracing the apparatus out of the entangled state, and taking the pointer states orthogonal, gives the identical mixture,

\mathrm{Tr}_A|\Psi\rangle\langle\Psi|=\sum_{n,m}c_n c_m^*\langle\phi_m|\phi_n\rangle|a_n\rangle\langle a_m|\;\xrightarrow{\langle\phi_m|\phi_n\rangle=\delta_{nm}}\;\sum_n|c_n|^2|a_n\rangle\langle a_n|=\mathcal E(|\psi\rangle\langle\psi|).

Verified: idempotence $\mathcal E\circ\mathcal E=\mathcal E$ , trace preservation, positive-semidefiniteness of the output, the purity drop $1\to\tfrac12$ for $|+\rangle$ , and equality of $\mathrm{Tr}_A$ (of a maximally entangled state) with $\mathcal E$ are all confirmed numerically, plus a symbolic proof $\mathcal E(\rho)=\mathrm{diag}(\rho_{00},\rho_{11})$ ; matches Zurek, Rev. Mod. Phys. 75, 715 (2003).

Limiting cases

The single model covers a whole spectrum of measurements, tuned by the coupling strength and by whether a result is recorded.

case	result	meaning
Ideal projective, outcome read	$\rho\mapsto \dfrac{P_n\rho P_n}{\mathrm{Tr}(P_n\rho)}$	Strong, sharp coupling with orthogonal pointers and a definite reading — selective (Lüders) collapse.
Non-selective, outcome unread	$\rho\mapsto \sum_n P_n\rho P_n$	Incoherent average over outcomes: dephases $\rho$ in the $\hat A$ basis, kills coherences, keeps populations.
Non-degenerate sharp limit	$P_n=\|a_n\rangle\langle a_n\|$	Rank-one projectors, collapse onto a single ray — a complete or maximal test.
Degenerate measurement	$P_n$ of rank $>1$	Learns only the subspace; superpositions within the $a_n$ -eigenspace survive (Lüders).
Weak-coupling limit	$g\,\delta t\ll\Delta Q$	Pointer shift $\ll$ its spread: little back-action, little information per shot, weak values over many trials.

Common trapCollapse is not derivable from the Schrödinger equation

The tempting mistake is to think the projection postulate falls out of unitary evolution once the apparatus is included. It does not. The entangled state $\sum_n c_n|a_n\rangle|\phi_n\rangle$ is a coherent superposition; tracing out the apparatus gives an improper mixture that is only locally indistinguishable from a real ensemble, and it never selects one branch as actual. A related slip is to forget degeneracy and "collapse onto an eigenket": for a degenerate eigenvalue you must project with the full rank- $>1$ operator $P_n$ (the Lüders rule), preserving coherences inside the subspace — otherwise you overstate the disturbance and get the wrong post-measurement state.

Worked examples

Example 1 — Purity destroyed by an unread measurement. Prepare the qubit in $|+\rangle=(|0\rangle+|1\rangle)/\sqrt2$ and perform an unread $\sigma_z$ measurement. What happens to the coherence and the purity?

Step 1. Write the density matrix. In the $z$ basis $\rho=|+\rangle\langle+|=\tfrac12\begin{pmatrix}1&1\\1&1\end{pmatrix}$ , so the off-diagonal coherence is $\rho_{01}=\tfrac12$ and the purity is $\mathrm{Tr}\rho^2=1$ (pure).

Step 2. Apply the non-selective map with $P_0=\mathrm{diag}(1,0)$ , $P_1=\mathrm{diag}(0,1)$ . Then $\mathcal E(\rho)=P_0\rho P_0+P_1\rho P_1=\tfrac12\begin{pmatrix}1&0\\0&1\end{pmatrix}=\tfrac12\mathbb 1$ .

Step 3. Read off the outcome. The populations are unchanged at $\tfrac12,\tfrac12$ , but the coherence $\rho_{01}$ is gone, and the purity has dropped to $\mathrm{Tr}(\tfrac12\mathbb 1)^2=\tfrac12$ . The state has fallen from a pure superposition to the maximally mixed state — purity $1\to\tfrac12$ — even though no result was recorded. This is decoherence in one line: the mere possibility of a which-path record destroys interference.

Example 2 — Born probabilities and the collapsed state for a tilted mixed input. Take $\rho=\tfrac12(\mathbb 1+r\,\hat n\!\cdot\!\vec\sigma)$ with Bloch radius $r=0.6$ tilted at polar angle $\theta=0.9\ \text{rad}$ ( $\hat n$ in the $x$ – $z$ plane), and measure $\sigma_z$ (eigenvalues $\pm1$ ).

Step 1. Born probabilities. With $n_z=\cos0.9=0.6216$ , $p_0=\tfrac12(1+r\,n_z)=\tfrac12(1+0.6\times0.6216)=\mathbf{0.6865}$ and $p_1=1-p_0=\mathbf{0.3135}$ .

Step 2. Post-measurement states. For outcome $+1$ the selective rule gives $P_0\rho P_0/p_0=|0\rangle\langle0|$ — a pure eigenstate, regardless of the mixedness of the input, because a rank-one projector wipes out everything but the diagonal entry.

Step 3. Statistics of the reading. The expectation is $\langle\sigma_z\rangle=(+1)p_0+(-1)p_1=r\,n_z=0.6\times0.6216=\mathbf{0.373}$ , and the variance is $(\Delta\sigma_z)^2=1-\langle\sigma_z\rangle^2=1-0.373^2=\mathbf{0.861}$ (using $a_n^2=1$ ). The single number $\langle\sigma_z\rangle=rn_z$ is exactly the $z$ -projection of the Bloch vector — the measurement reads off the component of $\vec r$ along the measured axis.

Origins · 1 minJohn von Neumann and how this was discovered

John von Neumann

1903–1957 · Budapest → Princeton; polymath of quantum foundations, computing, and game theory

Von Neumann built the rigorous Hilbert-space formulation of quantum mechanics and, in his 1932 Mathematische Grundlagen der Quantenmechanik, drew the sharp line between the unitary "Process 2" and the projective "Process 1", making the measurement problem statable rather than merely uncomfortable. The same book carried his notorious "impossibility proof" against hidden variables — later shown by Grete Hermann and John Bell to rest on an unjustified additivity assumption, the canonical case of a famous result being famously wrong. His measurement model, apparatus and all, remains the reference frame for everything from decoherence to quantum error correction.

History

1927

— von Neumann introduces the statistical operator (density matrix) and the trace rule for expectation values

1932

— Mathematische Grundlagen der Quantenmechanik: Process 1 versus Process 2, the projection postulate, and the measurement chain

1939

— London and Bauer place the "cut" at the observer in their monograph on quantum observation

1951

— Lüders refines the projection rule for degenerate eigenvalues (minimal-disturbance form)

1963

— Wigner sharpens the projection-postulate critique that grows into the Wigner's-friend debate

2003

— Zurek's decoherence review shows how pointer-state orthogonality makes collapse look automatic

Going further

References. von Neumann, Mathematical Foundations of Quantum Mechanics (1932; Princeton trans. 1955), Ch. V–VI · Sakurai & Napolitano, Modern Quantum Mechanics, §1.4 and §3.4 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, Vol. I, Ch. III · Nielsen & Chuang, Quantum Computation and Quantum Information, §2.2.3 · Lüders, Ann. Phys. (Leipzig) 8, 322 (1951) · Zurek, Rev. Mod. Phys. 75, 715 (2003) · Schlosshauer, Decoherence and the Quantum-to-Classical Transition (2007)

Provenance. All equations are standard and were recomputed: the spectral resolution and completeness of $\{P_n\}$ , the Born rule $p_n=\mathrm{Tr}(P_n\rho)$ , the Lüders projection, trace preservation and CPTP structure of $\sum_n P_n\rho P_n$ , the purity drop $1\to\tfrac12$ for $|+\rangle$ , the pointer-translation identity $e^{-ig a_n\hat P/\hbar}\phi_0(Q)=\phi_0(Q-g a_n)$ (verified on a grid), the entangling unitary producing $(|00\rangle+|11\rangle)/\sqrt2$ , and the equality of $\mathrm{Tr}_A$ with the dephasing map. The Lüders reference is corrected to Ann. Phys. 8, 322 (1951). No empirical constants are quoted; decoherence timescales are order-of-magnitude statements from Zurek (2003). Verifier: `verify_aq-qm-von-neumann-measurement-projection.py` (34 checks).

Prerequisites. The Density Operator and Mixed States · The Reduced Density Matrix and Partial Trace · Quantum Entanglement and Non-Separability · Measurement and Wavefunction Collapse (Lüders update) Leads to. The Quantum Measurement Problem · Generalized Measurements and POVMs · The EPR Paradox and Local Realism · Properties and Time Evolution of the Density Matrix

Is the "collapse" a physical process or just a bookkeeping update?

The von Neumann model is deliberately agnostic. Process 1 is stated as an axiom, not derived, so whether it is a real physical event, an epistemic update of the observer's information, or an artifact of decoherence-plus-interpretation is exactly what the different interpretations of quantum mechanics disagree about. What the model does pin down is that unitary Process 2 alone cannot do the job — it yields entangled branches, never a single result — so something beyond the Schrödinger equation is needed to account for the definite outcome we experience. The model isolates that gap precisely without claiming to close it.

If tracing out the apparatus gives the collapsed mixture, hasn't decoherence solved everything?

No — and this is the single most common overclaim. Tracing out a perfect apparatus (or environment) turns the system's reduced state into $\sum_n|c_n|^2|a_n\rangle\langle a_n|$ , which looks exactly like a classical ensemble of definite outcomes. But this is an "improper" mixture: the global state is still a pure entangled superposition, and the diagonal form only reflects our inability to access the apparatus, not the selection of one branch as real. Decoherence explains why the branches stop interfering and why classicality emerges, but the step from "a weighted list of possibilities" to "this one happened" is untouched. That residual step is the measurement problem.

Why does purity have to drop under a non-selective measurement?

Purity is $\mathrm{Tr}\rho^2=\sum_{mm'}|\rho_{mm'}|^2$ , a sum over the squared magnitudes of every matrix element. The non-selective map $\sum_n P_n\rho P_n$ deletes all off-diagonal entries in the measured basis while leaving the diagonal alone, so it removes non-negative contributions from the sum. Purity therefore cannot increase, and it strictly decreases whenever any coherence $\rho_{mm'}\ (m\neq m')$ was nonzero. Physically, an unread measurement injects classical uncertainty — you now don't know which branch you're in — and mixedness is the quantitative measure of that ignorance. If $\rho$ was already diagonal, there was no coherence to lose, and it is a fixed point.

What exactly is the difference between the Lüders rule and "collapse onto an eigenket"?

For a non-degenerate eigenvalue they coincide: $P_n=|a_n\rangle\langle a_n|$ and the state collapses to the ray $|a_n\rangle$ . The difference appears when $a_n$ is degenerate, so $P_n$ projects onto a multi-dimensional eigen-subspace. The naive "pick an eigenket" prescription is ambiguous and over-disturbs the state; the Lüders rule $\rho\to P_n\rho P_n/\mathrm{Tr}(P_n\rho)$ instead projects onto the whole subspace and renormalizes, which is the minimal-disturbance update consistent with having learned only that the outcome was $a_n$ . Any superposition living inside the subspace is preserved, which is why you sometimes need a second, compatible measurement to fully resolve the state.

Does a projective measurement on one half of an entangled pair let you signal faster than light?

No. When you measure your half of a Bell pair, the projection postulate updates the conditional state of the distant partner, which is the "steering" that troubled Einstein. But your partner's unconditioned state — the reduced density matrix obtained without knowing your outcome — is governed by the non-selective map, which leaves the partial trace over your system unchanged. Averaged over your unknown results, the distant marginal $\rho_B$ is exactly what it was before, so no measurable statistic on the far side depends on whether or when you measured. This no-signalling property is why quantum correlations coexist with relativity; it is developed in the entanglement and EPR articles.

Where is the "Heisenberg cut" between the quantum system and the classical apparatus?

Von Neumann's answer is deliberately unsettling: the cut can be placed anywhere along the chain — system, apparatus, second apparatus, observer — and the predictions for what the final observer sees are identical wherever you draw it. You can push the boundary as far up the chain as you like, treating more of it with Process 2, but at some point you must invoke Process 1 for a definite result. The movability of the cut is the content of the von Neumann chain, and its lack of a natural location is what the measurement problem, in its own article, wrestles with.

Why must the pointer states become orthogonal for the measurement to "work"?

The reduced system state after the entangling interaction is $\rho_S=\sum_{n,m}c_nc_m^*\langle\phi_m|\phi_n\rangle|a_n\rangle\langle a_m|$ . The surviving coherences are weighted by the pointer overlaps $\langle\phi_m|\phi_n\rangle$ . If the pointer wavepackets overlap ( $\langle\phi_m|\phi_n\rangle\neq0$ ), the branches can still interfere and the "measurement" is incomplete — the needle positions are not reliably distinguishable. Only when the displacements $g a_n$ exceed the pointer width do the overlaps vanish, the coherences die, and $\rho_S$ becomes the diagonal mixture $\sum_n|c_n|^2|a_n\rangle\langle a_n|$ . So orthogonal pointer states are the condition for a good, information-extracting, decohering measurement.