Generalized Measurements and POVMs

The first thing anyone learns about quantum measurement is the projection postulate: an observable is a Hermitian operator, its eigenprojectors label the outcomes, the probabilities are $\langle\psi|P_k|\psi\rangle$ , and the state jumps to the projected eigenvector. That picture — the "von Neumann" or "projective" measurement — is exact, but it is not the most general thing quantum mechanics allows you to do. The moment you let your system touch a second system (a meter, a probe, a stray photon) and then measure that, the statistics you see on the original system are described by operators that are still positive but no longer orthogonal, no longer idempotent, and no longer limited in number by the dimension of the Hilbert space. The object that captures exactly this larger class is the positive operator-valued measure, or POVM. It is deceptively simple: a set of positive operators $\{E_k\}$ that sum to the identity, with outcome probabilities $p_k=\mathrm{Tr}(E_k\rho)$ . That is the whole story at the level of odds. Why should a working physicist care about a "more general" measurement? Because POVMs do things projective measurements provably cannot: a qubit can carry four outcomes, a clever three-outcome measurement can tell two non-orthogonal states apart without ever being wrong, and a single fixed measurement can reconstruct an entire density matrix. Having met the density operator, the partial trace, and system–ancilla coupling, you are now ready for the punchline that ties them together: by Naimark's theorem, every POVM is nothing but an ordinary projective measurement carried out on a larger space and then viewed through the keyhole of your subsystem. Generalized measurement is not exotic — it is what plain measurement looks like from the inside of an open system.

Key takeaways

A POVM is a set of positive operators $\{E_k\}$ (the effects) with $\sum_k E_k=\mathbb{1}$ ; outcome $k$ on state $\rho$ has probability $p_k=\mathrm{Tr}(E_k\rho)$ .
It generalizes projective (von Neumann) measurement: the effects need not be orthogonal, need not be projectors, and there can be more of them than the Hilbert-space dimension.
The effects fix only the odds. The post-measurement state needs Kraus operators $M_k$ with $E_k=M_k^\dagger M_k$ ; the update $\rho\to M_k\rho M_k^\dagger/p_k$ is non-unique ( $M_k\to U_kM_k$ leaves $E_k$ alone).
Naimark's dilation theorem: every POVM is a projective measurement on system-plus-ancilla — the shadow of a sharp measurement on the whole, cast onto the part.
POVMs enable the impossible-for-projectors: unambiguous discrimination of non-orthogonal states with success $1-|\langle\psi_0|\psi_1\rangle|$ , and informationally complete SIC tomography from one setting.
Key idea: a generalized measurement is a set of Kraus operators $\{M_k\}$ with $\sum_k M_k^\dagger M_k=\mathbb{1}$ ; coarse-grain away the post-state and only the effects $E_k=M_k^\dagger M_k$ survive — the POVM — which Naimark's theorem realizes as projective measurement on a dilated space.

The core insight

A generalized measurement is fully specified by Kraus operators $\{M_k\}$ obeying $\sum_k M_k^\dagger M_k=\mathbb{1}$ : outcome $k$ has probability $p_k=\mathrm{Tr}(M_k\rho M_k^\dagger)$ and leaves the state $M_k\rho M_k^\dagger/p_k$ . Coarse-grain away the post-state and only the effects $E_k=M_k^\dagger M_k$ matter — the POVM — and by Naimark dilation any such measurement is a projective measurement on system-plus-ancilla.

Where this lives today

POVMs are the everyday language of experimental quantum information. Every realistic photodetector, with its finite efficiency and dark counts, is described by a POVM rather than a projector, and quantum key distribution security proofs are written entirely in terms of the eavesdropper's optimal POVM. Superconducting-qubit and trapped-ion labs run weak measurements — unsharp effects with $\varepsilon\ll1$ — thousands of times per experiment to track individual quantum trajectories and drive real-time feedback, a capability with no projective-measurement description at all. On the theory side, informationally complete POVMs (SIC-POVMs above all) power gate-set and state tomography, and the SIC existence question in every dimension remains a celebrated open problem in mathematical physics, tied to number theory and Hilbert's twelfth problem. When you read that a device "measured the qubit," the honest operator behind that sentence is almost always an $E_k$ , not a $P_k$ .

The physical picture

Picture a bank of detectors. Each effect $E_k$ is a "chance of clicking": feed it any state and it returns a number between $0$ and $1$ , the probability that detector $k$ fires. Positivity is just the demand that a probability can never be negative; completeness $\sum_k E_k=\mathbb{1}$ is the demand that something always clicks. What makes a POVM richer than a von Neumann measurement is that two detectors can both have a real chance of firing on the same incoming state — their effects overlap, encoding genuinely fuzzy, partially-resolving outcomes. Orthogonal projectors, which carve the Hilbert space into disjoint sectors, can never overlap and can never number more than the dimension $d$ . The deepest way to see where POVMs come from is the dilation picture. You cannot, by any direct cleverness on a single isolated system, extract more than a projective measurement allows. But you can borrow a fresh ancilla, let it interact with your system, and then measure the two together. From inside the system, those borrowed degrees of freedom "smear" the sharp projective measurement on the whole into a soft POVM on the part. POVMs are precisely projective measurements seen through the keyhole of a subsystem — which is why the same reduced-density-matrix and partial-trace machinery you already know reappears here. A useful slogan: the POVM answers "what are the odds?", the Kraus operators answer "and what's left?" The factorization $E_k=M_k^\dagger M_k$ is a kind of quantum square root, and the freedom $M_k\to U_kM_k$ is the freedom to give the state a unitary kick after a click without changing the click statistics. This split is also where the information-versus-disturbance tradeoff lives: a POVM with heavily overlapping effects learns little but can be made gentle (a weak measurement), while a sharp projective POVM learns the most but disturbs the most. A POVM lets you dial continuously between the two.

Figure

A POVM is projective measurement seen through a keyhole

Figure

Projective vertices versus the POVM simplex

The core equations

Everything about a POVM at the level of probabilities is fixed by two axioms — positivity and completeness — and the rest of the framework (the update rule, the dilation, the special cases) is scaffolding built on top of them. We start with the axioms, extract the generalized Born rule, add the Kraus operators for the post-state, and finish with Naimark's theorem, which explains why this larger class is forced on us the moment we allow ancillas. The defining axioms say the effects are positive and resolve the identity. Positivity is what keeps every outcome probability non-negative; completeness is what makes them sum to one.

E_k \ge 0, \qquad \sum_k E_k = \mathbb{1} \tag{1}

where: $E_k$ is the effect operator for outcome $k$ — a positive semidefinite Hermitian operator; $\mathbb{1}$ is the identity on the system Hilbert space; $E_k\ge 0$ means $\langle\phi|E_k|\phi\rangle\ge 0$ for every $|\phi\rangle$ . Read physically: each $E_k$ is a "soft yes/no question" whose answer is a probability rather than a certainty, and the collection of questions is exhaustive. Given the axioms, the generalized Born rule is not an extra assumption — it is the only probability rule that is linear in the state, which is exactly what compatibility with statistical mixtures demands.

p_k = \mathrm{Tr}(E_k\,\rho) \tag{2}

where: $p_k$ is the probability of outcome $k$ ; $\rho$ is the density operator of the measured system; $\mathrm{Tr}$ is the trace over the system Hilbert space. For a pure state $\rho=|\psi\rangle\langle\psi|$ this reduces to $p_k=\langle\psi|E_k|\psi\rangle$ , and when $E_k=P_k$ is a projector it becomes the familiar $|\langle\text{eigvec}|\psi\rangle|^2$ . The effects alone are silent about what the state becomes. To specify the post-measurement state we need Kraus (measurement) operators, one per outcome, whose "squared modulus" is the effect.

E_k = M_k^\dagger M_k, \qquad \sum_k M_k^\dagger M_k = \mathbb{1} \tag{3}

where: $M_k$ is the measurement operator for outcome $k$ and $M_k^\dagger$ its Hermitian conjugate. The completeness relation is now written on the $M_k$ ; it guarantees the probabilities sum to one and, equivalently, that the associated channel preserves the trace. With the Kraus operators in hand, the selective update — the state given that you saw outcome $k$ — is the Kraus operator acting and renormalizing.

\rho \;\longmapsto\; \rho_k = \frac{M_k\,\rho\,M_k^\dagger}{\mathrm{Tr}(M_k\rho M_k^\dagger)} \tag{4}

where: $\rho_k$ is the conditional post-measurement state; the numerator $M_k\rho M_k^\dagger$ is the unnormalized updated state whose trace is exactly $p_k$ . This is the step that goes beyond the POVM: two experimenters realizing the same $\{E_k\}$ can honestly report different $\rho_k$ . If you record the outcome but never look at it — or throw a system through the apparatus and average over results — you get the non-selective evolution, the Kraus sum, which is simultaneously the most general completely-positive trace-preserving (CPTP) map, the object owned as the general evolution class by The Density Operator and Mixed States.

\mathcal{E}(\rho) = \sum_k M_k\,\rho\,M_k^\dagger \tag{5}

where: $\mathcal{E}$ is the quantum channel. Here $p_k$ cancels against the normalization in $(4)$ , leaving a linear map; when all $M_k=P_k$ are orthogonal projectors this is exactly the dephasing/decoherence-by-measurement map. The following table is the "which object when" guide for this family.

you want…	use	what it fixes
outcome odds only	effects $E_k$ , rule $(2)$	probabilities $p_k=\mathrm{Tr}(E_k\rho)$
the conditional state	Kraus $M_k$ , rule $(4)$	post-state $\rho_k$ (non-unique in $M_k$ )
the averaged/unread state	channel $\mathcal{E}$ , rule $(5)$	CPTP evolution
the physical realization	Naimark dilation, eq. $(6)$	ancilla + unitary + projective readout

The Naimark realization turns the abstract effects into a concrete recipe: append an ancilla, apply a joint unitary, measure a standard observable on the composite.

\Pi_k = U^\dagger\big(\mathbb{1}_S\otimes |k\rangle\langle k|_A\big)U, \qquad E_k=\langle 0|_A\,\Pi_k\,|0\rangle_A \tag{6}

where: $|0\rangle_A$ is the fixed initial ancilla state; $U$ is the joint unitary on $\mathcal{H}_S\otimes\mathcal{H}_A$ ; $\{|k\rangle_A\}$ is the orthonormal ancilla pointer basis; $\Pi_k$ are genuine orthogonal projectors on the dilated space, $\Pi_j\Pi_k=\delta_{jk}\Pi_k$ . The projectors are sharp; only their restriction to the system looks soft. A concrete family that lives between "sharp" and "useless" is the unsharp qubit measurement, a noisy spin measurement whose sharpness dial $\varepsilon$ interpolates from projective to uninformative.

E_\pm = \tfrac{1}{2}\big(\mathbb{1} \pm \varepsilon\,\hat n\cdot\boldsymbol{\sigma}\big), \qquad 0<\varepsilon\le 1 \tag{7}

where: $\boldsymbol{\sigma}$ is the vector of Pauli matrices, $\hat n$ the measurement axis, and $\varepsilon$ the sharpness. Its eigenvalues are $\tfrac12(1\pm\varepsilon)$ , so it is positive for all $0<\varepsilon\le1$ ; at $\varepsilon=1$ the eigenvalues are $0$ and $1$ and $E_\pm$ become the sharp projectors $\tfrac12(\mathbb{1}\pm\hat n\cdot\boldsymbol\sigma)$ , while $\varepsilon\to0$ is the "measure nothing" limit.

Derivation AThe IDP POVM: telling non-orthogonal states apart without ever being wrong

Here is the structural payoff that no projective measurement can match. Two pure states $|\psi_0\rangle,|\psi_1\rangle$ with nonzero overlap $\langle\psi_0|\psi_1\rangle=c$ are prepared with equal prior, and we want a measurement that never misidentifies them — at the cost of sometimes answering "don't know." We will build a three-outcome POVM $\{E_0,E_1,E_?\}$ whose no-error conditions force each conclusive effect to be orthogonal to the wrong state, whose positivity of the inconclusive effect caps how often we succeed, and whose optimum reproduces the Ivanovic–Dieks–Peres bound $1-|c|$ . The route is: impose zero error, then push the success rate up until the "don't know" effect is about to go negative. Watch for the binding constraint — it is $E_?\ge0$ , not the effects themselves.

Step 1. Impose the no-error conditions. Outcome $0$ means "it was $|\psi_0\rangle$ ", so it must have zero probability on $|\psi_1\rangle$ : $\mathrm{Tr}(E_0|\psi_1\rangle\langle\psi_1|)=0$ . Since $E_0\ge0$ , this forces $E_0\propto|\psi_1^\perp\rangle\langle\psi_1^\perp|$ with $\langle\psi_1|\psi_1^\perp\rangle=0$ . Symmetrically $E_1\propto|\psi_0^\perp\rangle\langle\psi_0^\perp|$ . Why orthogonality is exactly right: a positive operator gives zero probability on $|\psi_1\rangle$ only if $|\psi_1\rangle$ lies in its kernel, i.e. the rank-one $E_0$ must point along the direction orthogonal to $|\psi_1\rangle$ .

Step 2. Write the symmetric ansatz $E_0=a\,|\psi_1^\perp\rangle\langle\psi_1^\perp|$ , $E_1=a\,|\psi_0^\perp\rangle\langle\psi_0^\perp|$ , and demand positivity of the leftover

E_? = \mathbb{1}-E_0-E_1 \ge 0.

This is the binding constraint: the two conclusive effects, being positive by construction, are fine for any $a>0$ ; it is the inconclusive effect that can go negative if $a$ is too large.

Step 3. Solve the positivity constraint for the largest admissible weight. Taking $|\psi_0\rangle=|0\rangle$ and $|\psi_1\rangle=\cos\alpha|0\rangle+\sin\alpha|1\rangle$ so that $c=\cos\alpha$ , one finds the smaller eigenvalue of $E_?$ vanishes at

a_{\max}=\frac{1}{1+|c|}.

Why this is the optimum: increasing $a$ raises the success probability monotonically, and $a_{\max}$ is the value where $E_?$ first touches the boundary of positive operators.

Step 4. Read off the performance. With $a=a_{\max}$ the success probability at equal priors is

p_{\text{success}} = 1-|\langle\psi_0|\psi_1\rangle| = 1-|c|.

For orthogonal states ( $c=0$ ) it is $1$ ; as the overlap $\to1$ it vanishes — you cannot unambiguously separate states you cannot separate. A projective measurement on a qubit has only two orthogonal outcomes and cannot reserve a distinct "inconclusive" verdict, so it cannot achieve error-free discrimination of non-orthogonal states at all. Verified: the constructed $\{E_0,E_1,E_?\}$ satisfies $E_i\ge0$ , $\sum_i E_i=\mathbb{1}$ , zero error $\mathrm{Tr}(E_0|\psi_1\rangle\langle\psi_1|)=\mathrm{Tr}(E_1|\psi_0\rangle\langle\psi_0|)=0$ , and $p_{\text{success}}=1-|c|$ for $\alpha\in\{30°,45°,60°\}$ — machine-checked in the verifier (Ivanovic 1987, Dieks 1988, Peres 1988).

Go deeper · Derivation B · 1 minNaimark dilation: the POVM is a projective measurement upstairs

Derivation BNaimark dilation: the POVM is a projective measurement upstairs

The conceptual capstone is that we invented nothing — every POVM is an ordinary projective measurement on a larger space. Given any POVM $\{E_k\}_{k=1}^n$ on a $d$ -dimensional $\mathcal{H}_S$ , we choose one valid set of Kraus operators, build an isometry into system-plus-ancilla, and show that projecting onto the ancilla pointer states reproduces the POVM statistics exactly. The strategy is to let the completeness relation do the heavy lifting: it is precisely the condition that makes the dilation map norm-preserving, hence extendable to a unitary. Watch how $\sum_k E_k=\mathbb{1}$ turns into $V^\dagger V=\mathbb{1}_S$ .

Step 1. Pick $M_k=\sqrt{E_k}$ (one legitimate choice) and introduce an ancilla $\mathcal{H}_A=\mathbb{C}^n$ with orthonormal basis $\{|k\rangle_A\}$ and fixed input $|0\rangle_A$ . Define the dilation map

V|\psi\rangle_S = \sum_{k=1}^n \big(M_k|\psi\rangle_S\big)\otimes|k\rangle_A.

Step 2. Check it is an isometry.

V^\dagger V=\sum_{j,k}M_j^\dagger M_k\,\langle j|k\rangle=\sum_k M_k^\dagger M_k=\sum_k E_k=\mathbb{1}_S.

Why completeness is the whole point: without $\sum_k E_k=\mathbb{1}$ the map $V$ would not preserve norms and could not be extended to a unitary $U$ on the joint space with $U(|\psi\rangle_S\otimes|0\rangle_A)=V|\psi\rangle_S$ .

Step 3. Measure the ancilla pointer projectors $\Pi_k=\mathbb{1}_S\otimes|k\rangle\langle k|_A$ , which are genuinely orthogonal, $\Pi_j\Pi_k=\delta_{jk}\Pi_k$ , $\sum_k\Pi_k=\mathbb{1}_{SA}$ . The induced probability on the system is

p_k=\langle\psi|V^\dagger\Pi_kV|\psi\rangle=\langle\psi|M_k^\dagger M_k|\psi\rangle=\langle\psi|E_k|\psi\rangle.

The joint projector collapses, via $\langle j|k\rangle=\delta_{jk}$ , to the single term $M_k^\dagger M_k=E_k$ .

Step 4. Tracing out the ancilla after the joint projection returns the Kraus update, confirming the whole generalized measurement is physically realized by unitary coupling plus projective readout — a sharp measurement on $S\otimes A$ , soft only when restricted to $S$ . This is the same reduce-to-a-subsystem move as The Reduced Density Matrix and Partial Trace, here applied to a measurement rather than a state. Verified: with the tetrahedral SIC effects as a test POVM, $M_k=\sqrt{E_k}$ gives $V^\dagger V=\sum_k M_k^\dagger M_k=\mathbb{1}$ to machine precision (verifier), matching the Naimark/Neumark construction in Nielsen & Chuang §2.2.8 and Peres §9.

Go deeper · Derivation C · 1 minThe tetrahedral SIC-POVM is informationally complete

Derivation CThe tetrahedral SIC-POVM is informationally complete

This card shows the other thing projectors cannot do: reconstruct a whole density matrix from a single fixed measurement. Four rank-one effects whose Bloch vectors point to the corners of a regular tetrahedron form a symmetric, informationally complete POVM on a qubit — four outcomes on a two-dimensional space. We verify completeness from the tetrahedron's zero vector sum, read off the outcome probabilities as an affine map of the unknown Bloch vector, and invert it. The point to watch: a single projective spin measurement gives only two outcomes and pins down only one Bloch component, whereas this one setting pins down all three.

Step 1. Take four pure states $|\phi_k\rangle$ with Bloch vectors $\hat r_k$ at the tetrahedron vertices, $\hat r_j\cdot\hat r_k=-\tfrac13$ for $j\ne k$ , and set $E_k=\tfrac14(\mathbb{1}+\hat r_k\cdot\boldsymbol\sigma)=\tfrac12|\phi_k\rangle\langle\phi_k|$ . Completeness follows from $\sum_k\hat r_k=0$ :

\sum_{k=1}^4 E_k=\tfrac14\Big(4\,\mathbb{1}+\big(\textstyle\sum_k\hat r_k\big)\cdot\boldsymbol\sigma\Big)=\mathbb{1}.

Step 2. For an unknown $\rho=\tfrac12(\mathbb{1}+\mathbf{s}\cdot\boldsymbol\sigma)$ the four probabilities are affine in $\mathbf{s}$ :

p_k=\mathrm{Tr}(E_k\rho)=\tfrac14\big(1+\hat r_k\cdot\mathbf{s}\big).

Step 3. Because the four $\hat r_k$ span $\mathbb{R}^3$ (over-completely), the linear system inverts:

\mathbf{s}=\sum_{k=1}^4\big(3p_k-\tfrac14\big)\,\hat r_k.

Why this is impossible projectively: a two-outcome projective measurement determines a single component $\hat n\cdot\mathbf{s}$ ; you would need three incompatible settings to reconstruct $\mathbf{s}$ , whereas the SIC does it from one. Verified: the SIC inner products $|\langle\phi_j|\phi_k\rangle|^2=\tfrac13=1/(d+1)$ for $d=2$ , $\sum_k E_k=\mathbb{1}$ , and the Bloch inversion recovering a test $\mathbf{s}=(0.3,-0.5,0.2)$ all check in the verifier (Renes, Blume-Kohout, Scott & Caves, J. Math. Phys. 45, 2171 (2004)).

Limiting cases

The framework contains textbook measurement theory as special cases, and its limits are the fastest sanity checks.

case	result	meaning
projective limit $E_k=P_k$	$P_jP_k=\delta_{jk}P_k,\ \sum_k P_k=\mathbb{1}$	orthogonal, idempotent effects recover the von Neumann postulate; with $M_k=P_k$ the update is the Lüders rule $\rho\to P_k\rho P_k/p_k$ (owned by Von Neumann Measurement and the Projection Postulate).
trivial POVM $E_1=\mathbb{1}$	all other $E_k=0$	one outcome that always occurs and reveals nothing; if $M_1=\mathbb{1}$ the channel is $\rho\to\rho$ . The "measure nothing" baseline.
unsharp qubit $\varepsilon\to1$	$E_\pm\to\tfrac12(\mathbb{1}\pm\hat n\cdot\boldsymbol\sigma)$	sharpness dial saturates at the projective spin measurement; $\varepsilon\to0$ gives the uninformative POVM.
SIC / IC-POVM	$E_k=\tfrac1d\|\phi_k\rangle\langle\phi_k\|,\ k=1,\dots,d^2$	$d^2$ rank-one effects; more outcomes than dimensions, impossible projectively, optimal for tomography.

Common trapA POVM does NOT tell you the post-measurement state

The single most common error is to "collapse" a state with an effect: writing $\rho\to E_k\rho E_k/p_k$ is wrong in general. The effects $E_k$ fix only probabilities. The state update is governed by the Kraus operators $M_k$ with $E_k=M_k^\dagger M_k$ , and the same $E_k$ admits infinitely many updates because $M_k\to U_kM_k$ (any unitary $U_k$ ) leaves $E_k$ untouched but changes $M_k\rho M_k^\dagger$ . The naive projective collapse returns only in the special case where $E_k$ is an orthogonal projector and $M_k=E_k$ . Whenever a problem asks for "the state after a generalized measurement," it must hand you the $M_k$ , not just the $E_k$ .

Worked examples

Step 1. Compute the overlap. $\langle\psi_0|\psi_1\rangle=\cos60°=\tfrac12$ , so $|c|=0.5$ . Because $c\ne0$ the states are non-orthogonal and no error-free projective measurement exists.

Step 2. Build the conclusive effects. $E_0=a\,|\psi_1^\perp\rangle\langle\psi_1^\perp|$ with $|\psi_1^\perp\rangle=-\sin60°\,|0\rangle+\cos60°\,|1\rangle$ , and $E_1=a\,|\psi_0^\perp\rangle\langle\psi_0^\perp|$ with $|\psi_0^\perp\rangle=|1\rangle$ , using $a=a_{\max}=1/(1+|c|)=1/1.5=\tfrac23$ .

Step 3. Verify positivity of the leftover. $E_?=\mathbb{1}-E_0-E_1$ has smallest eigenvalue $0$ at this $a$ (it just touches the boundary), so the POVM is valid.

Step 4. Read the success rate. $p_{\text{success}}=1-|c|=1-0.5=\mathbf{0.5}$ : half the time the apparatus names the state with certainty, half the time it honestly says "inconclusive," and it is never wrong. For contrast, the minimum-error (Helstrom) strategy that always guesses would succeed with $\tfrac12(1+\sqrt{1-|c|^2})=\tfrac12(1+\tfrac{\sqrt3}{2})\approx\mathbf{0.933}$ but pays with a $6.7\%$ error rate — a different tradeoff for a different task.

Example 2 — SIC tomography reconstructs a Bloch vector. A qubit in an unknown state $\rho=\tfrac12(\mathbb{1}+\mathbf{s}\cdot\boldsymbol\sigma)$ with $\mathbf{s}=(0.3,-0.5,0.2)$ is measured many times with the tetrahedral SIC-POVM. Recover $\mathbf{s}$ from the outcome frequencies.

Step 1. Predict the four probabilities $p_k=\tfrac14(1+\hat r_k\cdot\mathbf{s})$ . With $\hat r_1=(0,0,1)$ , $\hat r_2=(\tfrac{2\sqrt2}{3},0,-\tfrac13)$ , $\hat r_3=(-\tfrac{\sqrt2}{3},\tfrac{\sqrt6}{3},-\tfrac13)$ , $\hat r_4=(-\tfrac{\sqrt2}{3},-\tfrac{\sqrt6}{3},-\tfrac13)$ , this gives $p_1=0.300,\ p_2=0.320,\ p_3=0.112,\ p_4=0.268$ .

Step 2. Check normalization. $p_1+p_2+p_3+p_4=1.000$ , as required for a valid POVM.

Step 3. Invert. $\mathbf{s}=\sum_k(3p_k-\tfrac14)\hat r_k$ . Plugging in the four probabilities returns $\mathbf{s}=(\mathbf{0.300},\,\mathbf{-0.500},\,\mathbf{0.200})$ — the full state from a single measurement setting.

Step 4. Appreciate the impossibility. A projective spin- $z$ measurement would have given only $p_\uparrow=\tfrac12(1+s_z)=0.6$ , fixing $s_z=0.2$ and nothing about $s_x,s_y$ . Three incompatible projective settings would be needed to match what the one SIC-POVM delivers.

Origins · 1 minPOVM and how this was discovered

POVM

1940s–1970s formalization · Moscow · Princeton · Oxford

The generalized-measurement framework grew from the mathematics of operator dilations. Mark Naimark proved in 1940 that any positive-operator resolution of the identity is the compression of a projection-valued one on a larger space; W. Forrest Stinespring's 1955 theorem gave the completely-positive-map version that underlies the Kraus form. The operational reading — a POVM is the most general observable — was crystallized by Davies and Lewis (1970) and popularized in physics by Kraus, Helstrom, and Peres, whose textbook made effects and instruments standard tools of quantum information.

History

1940

— Naimark proves the dilation theorem: every positive operator-valued measure is a projection-valued measure on a larger Hilbert space

1955

— Stinespring's dilation theorem for completely positive maps supplies the operator-sum (Kraus) representation

1970

— Davies and Lewis formalize measurement as an instrument, framing the POVM as the most general quantum observable

1976

— Helstrom's Quantum Detection and Estimation Theory sets the minimum-error discrimination bound

1987

— Ivanovic constructs unambiguous discrimination of two non-orthogonal states, refined by Dieks (1988) and optimized by Peres (1988) to 1 − |⟨ψ₀|ψ₁⟩|

2004

— Renes, Blume-Kohout, Scott and Caves catalogue symmetric informationally complete POVMs for tomography

Going further

References. Nielsen & Chuang, Quantum Computation and Quantum Information, §2.2.3–2.2.6 (POVMs, Kraus operators) and §2.2.8 (Naimark) · Peres, Quantum Theory: Concepts and Methods, ch. 9 (generalized measurements; §9-5 IDP bound) · Wiseman & Milburn, Quantum Measurement and Control, ch. 1 · Cohen-Tannoudji, Diu & Laloë, Quantum Mechanics, postulates of measurement · Renes, Blume-Kohout, Scott & Caves, J. Math. Phys. 45, 2171 (2004).

Provenance. Axioms and generalized Born rule: Nielsen & Chuang §2.2.6. Kraus/operator-sum and CPTP equivalence: N&C §8.2, Peres §9 — the CPTP/Kraus class is stated and owned as the general evolution in The Density Operator and Mixed States. Naimark dilation: N&C §2.2.8, Peres §9. IDP unambiguous-discrimination bound $1-|\langle\psi_0|\psi_1\rangle|$ : Peres §9-5 (Ivanovic 1987, Dieks 1988, Peres 1988). Tetrahedral SIC inner products $1/(d+1)=1/3$ and Bloch reconstruction: Renes et al. (2004). All numeric claims — completeness and positivity of every POVM example, the IDP success bound, the Naimark isometry, the SIC overlaps and Bloch inversion — are recomputed in `verify_aq-qm-povm-generalized-measurements.py` (39 checks).

Prerequisites: The Density Operator and Mixed States · The Reduced Density Matrix and Partial Trace · Pure versus Mixed States and the Purity Criterion · Properties and Time Evolution of the Density Matrix. Leads to: Quantum Entanglement and Non-Separability · The Schmidt Decomposition · Bell's Theorem and the CHSH Inequality.

Is a POVM an observable in the usual sense?

Not quite — that is the point. A textbook observable is a Hermitian operator whose eigenprojectors form a projection-valued measure: orthogonal, idempotent, at most $d$ of them. A POVM is a positive operator-valued measure: its effects are positive but need not be orthogonal or idempotent, and there can be more than $d$ . Davies and Lewis showed the POVM is the correct general notion of a quantum observable. When you insist on associating a single self-adjoint operator with a measurement, you are silently restricting to the projective special case.

What exactly is the difference between an effect and a Kraus operator?

An effect $E_k$ is a positive operator that gives you the probability of outcome $k$ via $p_k=\mathrm{Tr}(E_k\rho)$ . A Kraus operator $M_k$ is a (generally non-Hermitian) operator that gives you the post-measurement state via $\rho\to M_k\rho M_k^\dagger/p_k$ . They are linked by $E_k=M_k^\dagger M_k$ , so the effect is like the squared modulus of the Kraus operator. Crucially the link is many-to-one: the polar freedom $M_k\to U_kM_k$ produces the same $E_k$ but a different state update.

Why is

E_? \ge 0

the binding constraint in unambiguous discrimination, not

E_0,E_1\ge0

Because $E_0$ and $E_1$ are built as positive multiples of rank-one projectors, so they are positive for any weight $a>0$ — they can never be the obstruction. The leftover $E_?=\mathbb{1}-E_0-E_1$ is where trouble appears: as you crank up $a$ to succeed more often, $E_?$ loses positivity. The optimum $a_{\max}=1/(1+|c|)$ is exactly the value where the smallest eigenvalue of $E_?$ hits zero. Push past it and $E_?$ goes negative and your "POVM" stops being a valid measurement.

Does a POVM let me extract more information than the state actually contains?

No — it lets you extract information in shapes a projective measurement cannot, not more than the state holds. A qubit still carries at most one bit per measurement on average, and no POVM beats the Holevo bound. What the SIC-POVM buys you is doing full tomography from one fixed setting instead of three incompatible ones, and what the IDP POVM buys you is error-free (rather than minimum-error) discrimination at the cost of inconclusive results. Different tasks, same underlying information budget.

Is the ancilla in Naimark's theorem physically real or a mathematical trick?

Both, and that is the beauty of it. Mathematically, Naimark's theorem guarantees an ancilla and joint unitary exist for any POVM. Physically, in the lab the ancilla is usually a real thing — an auxiliary mode, a meter qubit, the electromagnetic field a photon couples to. The theorem says your soft measurement is never magic: somewhere there is a larger system on which a perfectly ordinary sharp measurement is happening, and your POVM is what that sharp measurement looks like once you trace out (in the sense of The Reduced Density Matrix and Partial Trace) the part you don't control.

How does a weak measurement fit into this?

A weak measurement is just a POVM with barely-informative effects, e.g. the unsharp qubit $E_\pm=\tfrac12(\mathbb{1}\pm\varepsilon\,\hat n\cdot\boldsymbol\sigma)$ with $\varepsilon\ll1$ . Because the two effects are nearly equal to $\tfrac12\mathbb{1}$ , a single measurement barely distinguishes the outcomes and barely disturbs the state — the information-versus-disturbance dial turned toward "gentle." Repeating such measurements continuously realizes quantum trajectories and continuous monitoring. The sharp $\varepsilon=1$ limit is the ordinary projective spin measurement; everything in between is genuinely non-projective and needs the POVM language.

Why do people say projective measurement is a "special case" rather than the fundamental one?

Because the moment you take open systems seriously — and every real experiment is open — the projective postulate is not closed under the operations you actually perform. Couple to an ancilla, measure it projectively, and the effective measurement on your system is a POVM, not a projector. So the projective measurement is the idealized, isolated-system limit, while the POVM is what survives contact with the rest of the world. Naimark's theorem reconciles the two: the general case is the special case, viewed from a subsystem.