POVMs: First Encounter

A positive-operator-valued measure, or POVM, is the part of a quantum measurement that determines the probabilities of its classical outcomes. It generalizes a projection-valued measure by replacing mutually orthogonal projectors with positive effects.

For a finite outcome set, a POVM is a collection

$$\left\lbrace E_1,\ldots,E_n \right\rbrace$$

of operators on the system Hilbert space such that

$$E_i \geq 0, \qquad \sum_{i=1}^{n}E_i = I.$$

For a state $\rho$, the probability of outcome $i$ is

$$p(i) = \operatorname{Tr}(\rho E_i).$$

These three equations are the finite-outcome POVM formalism. They predict outcome statistics, but not the state left behind. That boundary is essential: a POVM is a probability model, whereas a quantum instrument also specifies backaction.

Effects as Quantum Events

An individual effect $E$ is a positive operator bounded above by the identity:

$$0 \leq E \leq I.$$

The inequality $E\leq I$ means $I-E\geq0$. Every effect can therefore define a two-outcome yes–no POVM,

$$\left\lbrace E,\, I-E \right\rbrace.$$

For state $\rho$,

$$\begin{aligned} p(\mathrm{yes}) &= \operatorname{Tr}(\rho E), \\ p(\mathrm{no}) &= 1-p(\mathrm{yes}). \end{aligned}$$

In finite dimensions, positivity is equivalent to Hermiticity with nonnegative eigenvalues. The upper bound then implies that every effect eigenvalue lies in $[0,1]$. A projector has only eigenvalues $0$ and $1$; a general effect can also have intermediate eigenvalues.

Why every member of a POVM is bounded

If $\{E_i\}$ is a POVM, then

$$I-E_i = \sum_{j\neq i}E_j \geq 0.$$

Hence $E_i\leq I$. The two defining POVM conditions already ensure that each outcome probability lies between zero and one.

Effects are not states

Both density operators and effects are positive, but they play different roles:

• a state satisfies $\operatorname{Tr}\rho=1$ and represents a preparation;
• an effect satisfies $0\leq E\leq I$ and represents a possible measurement event.

An effect need not have trace one. A density operator need not be bounded by a small multiple chosen for a particular outcome. The number

$$\operatorname{Tr}(\rho E)$$

pairs a preparation with an event to produce a probability.

The Probability Rule

For a pure state,

$$\rho = \lvert\psi\rangle\langle\psi\rvert,$$

the POVM rule becomes

$$p(i) = \langle\psi\rvert E_i\lvert\psi\rangle.$$

For a mixed state,

$$\rho = \sum_k w_k \lvert\psi_k\rangle\langle\psi_k\rvert,$$

linearity gives

$$p(i) = \sum_k w_k \langle\psi_k\rvert E_i \lvert\psi_k\rangle.$$

Thus probabilities depend affinely on the density operator. If

$$\rho_\lambda = \lambda\rho_1 + (1-\lambda)\rho_2,$$

then

$$\begin{aligned} p(i\mid\rho_\lambda) &= \lambda p(i\mid\rho_1) \\ &\quad+ (1-\lambda)p(i\mid\rho_2). \end{aligned}$$

Normalization follows from completeness:

$$\begin{aligned} \sum_i p(i) &= \sum_i\operatorname{Tr}(\rho E_i)\\ &= \operatorname{Tr} \left( \rho\sum_iE_i \right)\\ &= \operatorname{Tr}\rho = 1. \end{aligned}$$

This is the state–effect form of the Born Rule.

Outcome Labels and Moment Operators

The index $i$ is only a label until the measurement model assigns a reported value $x_i$. The mean reported value is

$$\mathbb E[X] = \sum_i x_i p(i) = \operatorname{Tr}(\rho A_1),$$

where

$$A_1 = \sum_i x_iE_i$$

is the first-moment operator.

Higher moments use different operators:

$$A_k = \sum_i x_i^kE_i, \qquad \mathbb E[X^k] = \operatorname{Tr}(\rho A_k).$$

For a general POVM,

$$A_2 \neq A_1^2$$

in general. Equality is automatic for the spectral measure of a sharp observable, but not for arbitrary effects. Therefore a first-moment operator alone does not specify the full outcome distribution.

This prevents a common shortcut: assigning numbers to POVM outcomes does not automatically make the effects the spectral projectors of the operator $\sum_i x_iE_i$.

Projective Measurements as a Special Case

A projective measurement has effects $\{P_a\}$ satisfying

$$P_a^\dagger = P_a, \qquad P_a^2 = P_a,$$

and

$$P_aP_b = 0 \quad \text{for }a\neq b.$$

Together with $\sum_aP_a=I$, these projectors form both a PVM and a POVM.

A general POVM differs in several possible ways:

• an effect need not be idempotent;
• distinct effects can have overlapping support;
• the number of outcomes can exceed the Hilbert-space dimension;
• no outcome need correspond to a definite eigenspace of a system observable.

Every PVM is a POVM, but not every POVM is a PVM. Projective Measurement remains the canonical home for sharp outcomes and Lüders update.

Coarse Events and Relabeling

If $A$ is a set of fine outcome labels, the effect for the coarse event “the outcome lies in $A$” is

$$E(A) = \sum_{i\in A}E_i.$$

Its probability is

$$p(A) = \operatorname{Tr} \left[ \rho E(A) \right] = \sum_{i\in A}p(i).$$

The complement has effect

$$E(A^{\mathrm c}) = I-E(A).$$

This operator addition is the quantum counterpart of adding probabilities for mutually exclusive classical records. It says nothing by itself about whether a detector physically resolved the fine records and later discarded them; that distinction belongs to the instrument.

Classical postprocessing

Suppose a POVM first produces fine outcome $i$, then a classical channel reports label $j$ with conditional probability $q(j\mid i)$. The reported effects are

$$F_j = \sum_iq(j\mid i)E_i,$$

where

$$q(j\mid i) \geq 0, \qquad \sum_jq(j\mid i) = 1.$$

Indeed,

$$\sum_jF_j = \sum_iE_i = I.$$

Classical relabeling, binning, and readout noise therefore map POVMs to POVMs.

Commuting POVMs as Noisy Sharp Measurements

If all effects in a finite-dimensional POVM commute, they can be diagonalized in a common orthogonal decomposition $\{P_a\}$. Each effect can then be written

$$E_i = \sum_a q(i\mid a)P_a,$$

with

$$q(i\mid a) \geq 0, \qquad \sum_iq(i\mid a) = 1.$$

Operationally, the outcome statistics can be modeled as a sharp PVM result $a$ followed by classical noise $q(i\mid a)$. This statement concerns probabilities. It does not imply that every physical implementation first performs that sharp measurement, nor does it fix the post-measurement state.

A POVM with noncommuting effects cannot be represented as classical postprocessing of one PVM on the same system. It may still arise from a projective measurement on a larger system through an ancilla dilation.

Example: Inefficient Click Detection

Consider a qubit detector intended to click on $\lvert1\rangle$ with efficiency $0\leq\eta\leq1$. A two-outcome model is

$$E_{\mathrm{click}} = \eta P_1,$$

and

$$E_{\mathrm{no}} = P_0 + (1-\eta)P_1.$$

Both effects are positive and

$$E_{\mathrm{click}} + E_{\mathrm{no}} = I.$$

For

$$\lvert\psi\rangle = \alpha\lvert0\rangle + \beta\lvert1\rangle,$$

the probabilities are

$$\begin{aligned} p(\mathrm{click}) &= \eta\lvert\beta\rvert^2,\\ p(\mathrm{no}) &= \lvert\alpha\rvert^2 + (1-\eta)\lvert\beta\rvert^2. \end{aligned}$$

When $0<\eta<1$,

$$E_{\mathrm{click}}^2 = \eta^2P_1 \neq \eta P_1,$$

so the click effect is not a projector. A no-click result is ambiguous: the system could have occupied $\lvert0\rangle$, or it could have occupied $\lvert1\rangle$ and escaped detection. The POVM quantifies that ambiguity at the probability level.

Example: An Unbiased Binary Qubit POVM

Let $\hat{\mathbf n}$ be a unit vector and define

$$E_\pm = \frac12 \left( I\pm\eta \hat{\mathbf n}\cdot\boldsymbol{\sigma} \right), \qquad 0\leq\eta\leq1.$$

For a state

$$\rho = \frac12 \left( I+\mathbf r\cdot\boldsymbol{\sigma} \right),$$

the outcome probabilities are

$$p(\pm) = \frac12 \left( 1\pm\eta \mathbf r\cdot\hat{\mathbf n} \right).$$

At $\eta=1$, the effects are the sharp spin projectors along $\hat{\mathbf n}$. At $\eta=0$, both effects equal $I/2$ and carry no state information. Intermediate $\eta$ reduces the contrast of the sharp distribution.

The effects do not specify how much the state is disturbed. A square-root instrument is one possible implementation, but other instruments have the same $E_\pm$.

A Qubit Effect in Bloch Form

Every Hermitian qubit operator can be written

$$E = \frac12 \left( \alpha I + \mathbf v\cdot\boldsymbol{\sigma} \right).$$

Its eigenvalues are

$$\lambda_\pm = \frac12 \left( \alpha\pm\lVert\mathbf v\rVert \right).$$

The operator is an effect exactly when

$$0 \leq \alpha-\lVert\mathbf v\rVert$$

and

$$\alpha+\lVert\mathbf v\rVert \leq 2.$$

Equivalently,

$$\lVert\mathbf v\rVert \leq \min(\alpha,2-\alpha).$$

For state $\rho=(I+\mathbf r\cdot\boldsymbol{\sigma})/2$,

$$\operatorname{Tr}(\rho E) = \frac12 \left( \alpha+\mathbf r\cdot\mathbf v \right).$$

This formula turns effect validation and probability calculation into elementary Bloch-vector geometry.

Example: Three Outcomes on a Qubit

Let three unit vectors in the equatorial plane be separated by $120^\circ$. For $k=0,1,2$, define

$$\mathbf n_k = \begin{pmatrix} \cos\frac{2\pi k}{3}\\ \sin\frac{2\pi k}{3}\\ 0 \end{pmatrix}.$$

They satisfy

$$\sum_{k=0}^{2}\mathbf n_k = 0.$$

Define the trine effects

$$E_k = \frac13 \left( I+\mathbf n_k\cdot\boldsymbol{\sigma} \right).$$

Then

$$\sum_{k=0}^{2}E_k = I.$$

Each $E_k$ is positive with eigenvalues $0$ and $2/3$. It is proportional to a rank-one projector but is not itself a projector.

For a qubit state with Bloch vector $\mathbf r$,

$$p(k) = \frac13 \left( 1+\mathbf r\cdot\mathbf n_k \right).$$

This valid qubit measurement has three nonorthogonal outcomes, so it cannot be a PVM on the two-dimensional system. POVMs develops the trine and other nonorthogonal measurements in more detail.

What a POVM Does Not Specify

The effect $E_i$ fixes

$$p(i) = \operatorname{Tr}(\rho E_i)$$

for every input state, but it does not fix the conditional map

$$\rho \longmapsto \rho_i.$$

An instrument with Kraus operators $M_{i\alpha}$ has

$$E_i = \sum_\alpha M_{i\alpha}^\dagger M_{i\alpha}.$$

Many different operator families can produce the same $E_i$. Consequently, a POVM alone cannot determine:

• the selected post-measurement state;
• the unread channel;
• repeatability;
• disturbance to a reference system;
• ordered probabilities for later measurements.

Use Generalized Measurements Overview for the instrument formulas and Quantum Instruments for the detailed theory.

Indirect and Continuous Outcomes

A POVM on a system can be realized by coupling the system to an ancilla and performing a projective measurement on the enlarged space. The system effects are obtained after inserting the ancilla ready state. This is the content of Naimark Dilation; the dilation is not unique and does not by itself choose a unique instrument.

For a continuous outcome space $\Omega$, one uses an operator-valued set function

$$\Delta \longmapsto E(\Delta)$$

such that

$$E(\Omega) = I,$$

and disjoint measurable sets add in the appropriate operator topology. The probability measure is

$$\Pr_\rho(\Delta) = \operatorname{Tr} \left[ \rho E(\Delta) \right].$$

An operator density $E(x)$ can sometimes be written, but it depends on a reference measure and need not exist in every presentation. The finite formulas on this page should not be transferred to continuous outcomes by replacing sums with integrals without checking the measure.

Informational Completeness Preview

Outcome probabilities can reveal a state only to the extent that the effects span operator space. In dimension $d$, Hermitian operators form a real vector space of dimension $d^2$. A finite POVM is informationally complete when its effects span that space.

Such a POVM requires at least $d^2$ outcomes. A single $d$-outcome basis PVM determines only the diagonal of $\rho$ in that basis and is not informationally complete for $d>1$.

Informational completeness does not make finite data exact, remove calibration assumptions, or specify instrument backaction. Those statistical and experimental issues belong in Measurement Tomography.

A Reliable Validation Workflow

Given candidate finite-outcome effects $\{E_i\}$:

1. Verify that every $E_i$ acts on the declared system Hilbert space.
2. Check Hermiticity: $E_i^\dagger=E_i$.
3. Check positivity, for example by computing the smallest eigenvalue.
4. Check completeness: $\sum_iE_i=I$.
5. Compute $p(i)=\operatorname{Tr}(\rho E_i)$.
6. Confirm that probabilities are real, nonnegative, and normalized.
7. Record the physical meaning and numerical value associated with every label.
8. If outcomes are grouped, sum their effects.
9. If later-state predictions are requested, obtain an instrument rather than guessing one from the POVM.

In numerical work, use tolerances appropriate to the matrix scale. A tiny negative eigenvalue can be roundoff; a substantial one means the proposed effect is unphysical.

Common Mistakes

• Calling one positive operator a complete POVM without including its complementary outcomes.
• Checking $E_i\geq0$ but forgetting $\sum_iE_i=I$.
• Assuming that every effect is a projector or a density operator.
• Inferring post-measurement states from effects alone.
• Confusing an outcome label with an eigenvalue.
• Assuming $A_2=A_1^2$ for a general outcome-valued POVM.
• Treating classical postprocessing as a unique physical implementation.
• Assuming that more outcomes than dimension is impossible.
• Forgetting that continuous POVMs are measures, not merely indexed operator lists.
• Declaring a measurement informationally complete because it has many outcomes without checking their operator span.
• Clipping materially negative probabilities instead of diagnosing the state, effects, or numerical procedure.

References

• E. B. Davies, Quantum Theory of Open Systems, Academic Press, 1976.
• K. Kraus, States, Effects, and Operations: Fundamental Notions of Quantum Theory, Springer, 1983.
• A. S. Holevo, Probabilistic and Statistical Aspects of Quantum Theory, 2nd ed., Edizioni della Normale, 2011.
• A. Peres, Quantum Theory: Concepts and Methods, Kluwer Academic Publishers, 1995.
• P. Busch, P. Lahti, J.-P. Pellonpää, and K. Ylinen, Quantum Measurement, Springer, 2016.
• M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information, 10th anniversary ed., Cambridge University Press, 2010.
• J. Watrous, The Theory of Quantum Information, Cambridge University Press, 2018.

Exercises

1. Validate a binary qubit POVM

In the computational basis, let

$$E_0 = \begin{pmatrix} 4/5 & 0\\ 0 & 3/10 \end{pmatrix}, \qquad E_1 = I-E_0.$$

Verify that $\{E_0,E_1\}$ is a POVM. For

$$\rho = \begin{pmatrix} 1/4 & 0\\ 0 & 3/4 \end{pmatrix},$$

compute both probabilities.

Solution

The eigenvalues of $E_0$ are $4/5$ and $3/10$, both in $[0,1]$. The complementary effect is

$$E_1 = \begin{pmatrix} 1/5 & 0\\ 0 & 7/10 \end{pmatrix},$$

which is also positive. By construction, $E_0+E_1=I$.

The first probability is

$$\begin{aligned} p(0) &= \operatorname{Tr}(\rho E_0)\\ &= \frac14\frac45 + \frac34\frac{3}{10}\\ &= \frac15+\frac{9}{40} = \frac{17}{40}. \end{aligned}$$

Therefore

$$p(1) = 1-p(0) = \frac{23}{40}.$$

2. Prove the upper effect bound

Let $\{E_i\}$ be a finite POVM. Prove $E_i\leq I$ for every $i$, and use this to show $p(i)\leq1$ for every state.

Solution

Completeness gives

$$I-E_i = \sum_{j\neq i}E_j.$$

The right side is a sum of positive operators, so $I-E_i\geq0$. Hence $E_i\leq I$.

For a density operator $\rho$,

$$\begin{aligned} 1-p(i) &= \operatorname{Tr}(\rho) - \operatorname{Tr}(\rho E_i)\\ &= \operatorname{Tr} \left[ \rho(I-E_i) \right] \geq 0. \end{aligned}$$

Therefore $p(i)\leq1$.

3. Diagnose an inefficient no-click event

For the inefficient click detector, let $\eta=3/5$ and

$$\lvert\psi\rangle = \frac{ \lvert0\rangle+\sqrt3\,\lvert1\rangle }{2}.$$

Compute click and no-click probabilities. What alternatives contribute to no click?

Solution

The excited-state population is

$$\lvert\beta\rvert^2 = \frac34.$$

Thus

$$p(\mathrm{click}) = \frac35\frac34 = \frac{9}{20},$$

and

$$p(\mathrm{no}) = 1-\frac{9}{20} = \frac{11}{20}.$$

The no-click probability can also be separated as

$$p(\mathrm{no}) = \frac14 + \left( 1-\frac35 \right) \frac34 = \frac14+\frac{3}{10}.$$

The first term comes from the $\lvert0\rangle$ population; the second comes from missed $\lvert1\rangle$ events.

4. Apply classical readout noise

A sharp computational-basis PVM has effects $P_0$ and $P_1$. The reported bit is flipped with probability $\epsilon$. Derive the reported effects and probabilities for arbitrary $\rho$.

Solution

The classical channel gives

$$\begin{aligned} F_0 &= (1-\epsilon)P_0+\epsilon P_1,\\ F_1 &= \epsilon P_0+(1-\epsilon)P_1. \end{aligned}$$

Both effects are positive and $F_0+F_1=I$. Writing

$$p_{\mathrm{true}}(a) = \operatorname{Tr}(\rho P_a),$$

the reported probabilities are

$$\begin{aligned} p_{\mathrm{obs}}(0) &= (1-\epsilon)p_{\mathrm{true}}(0) + \epsilon p_{\mathrm{true}}(1),\\ p_{\mathrm{obs}}(1) &= \epsilon p_{\mathrm{true}}(0) + (1-\epsilon)p_{\mathrm{true}}(1). \end{aligned}$$

They sum to one.

5. Check the trine POVM

For the trine effects

$$E_k = \frac13 \left( I+\mathbf n_k\cdot\boldsymbol{\sigma} \right),$$

verify completeness. Then take the pure state with Bloch vector $\mathbf r=\mathbf n_0$ and compute all three probabilities.

Solution

Because $\sum_k\mathbf n_k=0$,

$$\sum_{k=0}^{2}E_k = \frac13 \left( 3I + \sum_{k=0}^{2} \mathbf n_k\cdot\boldsymbol{\sigma} \right) = I.$$

For $\mathbf r=\mathbf n_0$,

$$\mathbf n_0\cdot\mathbf n_0 = 1,$$

whereas

$$\mathbf n_0\cdot\mathbf n_1 = \mathbf n_0\cdot\mathbf n_2 = -\frac12.$$

Therefore

$$p(0) = \frac23, \qquad p(1) = p(2) = \frac16.$$

6. Compare first and second moments

Assign numerical outcomes $x_\pm=\pm1$ to

$$E_\pm = \frac12 \left( I\pm\eta\sigma_z \right), \qquad 0<\eta<1.$$

Compute $A_1$, $A_2$, and $A_1^2$.

Solution

The first-moment operator is

$$\begin{aligned} A_1 &= (+1)E_+ + (-1)E_-\\ &= \eta\sigma_z. \end{aligned}$$

Because both squared outcome values equal one,

$$A_2 = E_++E_- = I.$$

But

$$A_1^2 = \eta^2\sigma_z^2 = \eta^2I.$$

For $0<\eta<1$, $A_2\neq A_1^2$. The first-moment operator does not encode the full POVM moments.

7. Form a coarse event

A four-outcome POVM has effects $\{E_1,E_2,E_3,E_4\}$. Outcomes $1$ and $3$ are reported together as “success.” Write the binary coarse-grained POVM and prove that its probabilities are normalized.

Solution

The success and failure effects are

$$E_{\mathrm s} = E_1+E_3,$$

and

$$E_{\mathrm f} = E_2+E_4.$$

Both are positive, and

$$E_{\mathrm s}+E_{\mathrm f} = \sum_{i=1}^{4}E_i = I.$$

Thus they form a binary POVM. For any state,

$$\begin{aligned} p(\mathrm s)+p(\mathrm f) &= \operatorname{Tr} \left[ \rho(E_{\mathrm s}+E_{\mathrm f}) \right]\\ &= \operatorname{Tr}\rho = 1. \end{aligned}$$

8. Keep probabilities and backaction separate

For the computational PVM effects $\{P_0,P_1\}$, compare the Lüders operators

$$M_0 = P_0, \qquad M_1 = P_1$$

with

$$N_0 = XP_0, \qquad N_1 = XP_1.$$

Show that both instruments have the same POVM and compare their selected outputs.

Solution

For the first instrument,

$$M_i^\dagger M_i = P_i.$$

For the second,

$$N_i^\dagger N_i = P_iX^\dagger XP_i = P_i.$$

Thus both instruments give

$$p(i) = \operatorname{Tr}(\rho P_i)$$

for every input.

The Lüders instrument outputs $P_0$ after outcome $0$ and $P_1$ after outcome $1$. Because $X\lvert0\rangle=\lvert1\rangle$ and $X\lvert1\rangle=\lvert0\rangle$, the second instrument outputs

$$\rho_0^{(N)} = P_1, \qquad \rho_1^{(N)} = P_0.$$

Identical effects and probabilities do not imply identical post-measurement states.

Summary

A finite POVM is a complete set of effects,

$$E_i \geq 0, \qquad \sum_iE_i = I,$$

with probabilities $p(i)=\operatorname{Tr}(\rho E_i)$. Effects are quantum events satisfying $0\leq E_i\leq I$; they need not be orthogonal projectors, and there can be more outcomes than Hilbert-space dimensions.

POVMs support coarse graining, classical postprocessing, inefficient detection, and nonorthogonal outcome structures. They determine every single-measurement outcome distribution but do not determine state update or sequential statistics. Those require a quantum instrument.