Projective measurement is only the first layer of quantum measurement. The more general language is a positive-operator valued measure, or POVM, which separates outcome probabilities from one special choice of orthogonal projectors.

Guide 2 introduced measurement through eigenvectors and projectors. Guide 5 rewrote probabilities as trace formulas. Guide 6 introduced channels and Kraus operators. Those ideas now combine into a broader measurement theory that can describe noisy detectors, indirect measurements, state-discrimination strategies and measurement procedures with more outcomes than the system dimension.

Projective measurement is a special case. Generalised measurement keeps the probability rules while allowing a wider family of positive operators.

1. The POVM definition

A POVM with outcomes labelled k is a collection of positive semidefinite operators {E_k} satisfying

E_k ≥ 0 for every k

and

Σ_k E_k = I.

If the quantum state is ρ, the probability of outcome k is

p(k)=Tr(ρE_k).

Positivity guarantees non-negative probabilities. The completeness relation guarantees that the probabilities sum to one.

2. Projective measurements are POVMs

An orthogonal projective measurement uses projectors {P_k} satisfying

P_k=P_k†
P_k²=P_k
P_jP_k=0 for distinct outcomes
Σ_kP_k=I

Every projector is positive semidefinite, so the set is also a POVM. But a general POVM element need not be a projector, and distinct POVM elements need not be orthogonal.

Worked example: computational-basis measurement

For a qubit, let E_0=|0⟩⟨0| and E_1=|1⟩⟨1|. Then E_0+E_1=I. For ρ=[[a,c],[c*,1-a]],

p(0)=Tr(ρE_0)=a and p(1)=1-a.

3. A non-projective two-outcome POVM

Take a parameter 0≤η≤1 and define

E_0=(1/2)(I+ηZ), E_1=(1/2)(I-ηZ).

The operators are positive because their eigenvalues are (1±η)/2, and they sum to I. When η=1, the measurement becomes the sharp Z projective measurement. When η=0, both effects equal I/2 and the output is a fair random label independent of the state.

The parameter therefore interpolates between a completely uninformative detector and an ideal sharp measurement.

4. POVM elements determine probabilities, not the whole state update

The operators E_k are sometimes called effects. They determine outcome probabilities, but they do not by themselves uniquely specify the post-measurement state.

To describe both the outcome and state update, introduce measurement operators M_k satisfying

E_k=M_k†M_k.

Then

p(k)=Tr(M_kρM_k†)=Tr(ρE_k)

and, conditional on outcome k,

ρ_k'=M_kρM_k†/p(k).

Different measurement instruments can share the same POVM while disturbing the state differently.

5. Measurement as a channel with a classical record

A useful way to unify Guides 6 and 7 is to model measurement as a quantum operation whose output includes a classical label. Schematically,

ρ → Σ_k |k⟩⟨k|_C ⊗ M_kρM_k†.

The classical register C stores the outcome. If we ignore the outcome, the system undergoes the channel

ρ → Σ_k M_kρM_k†.

This makes measurement mathematically part of the same operator-sum framework as open-system dynamics.

6. Why generalised measurements are useful

Generalised measurements are useful whenever orthogonal projectors are too restrictive. Examples include:

distinguishing non-orthogonal quantum states as well as possible;
detectors with inefficiency or finite resolution;
measurements implemented indirectly through an ancilla;
informationally complete measurements for tomography;
measurements with more outcomes than the Hilbert-space dimension.

7. Naimark dilation

A powerful theorem says that a POVM on a system can be realised as a projective measurement on a larger Hilbert space after coupling the system to an auxiliary degree of freedom.

This mirrors Stinespring dilation for channels. Generalised local behaviour can arise from ordinary unitary dynamics and projective measurement on a suitably enlarged system.

Generalised measurement downstairs can be projective measurement upstairs in a larger space.

8. The state-discrimination problem

Suppose a source prepares either state ρ_0 with prior probability p_0 or state ρ_1 with prior probability p_1. We receive one copy and want to guess which state was sent.

A two-outcome POVM {E_0,E_1} produces a decision rule. The success probability is

P_success=p_0Tr(ρ_0E_0)+p_1Tr(ρ_1E_1).

The optimisation question is: which POVM maximises this expression?

9. Orthogonal states can be distinguished perfectly

If two pure states |ψ⟩ and |φ⟩ are orthogonal, then ⟨ψ|φ⟩=0. A projective measurement onto those two directions identifies the state with certainty.

This provides an operational meaning for orthogonality: orthogonal pure states can be perfectly distinguished in a single ideal measurement.

10. Non-orthogonal pure states cannot be perfectly distinguished deterministically

If 0<|⟨ψ|φ⟩|<1, no measurement can always identify which state was supplied without error while always returning a conclusive guess.

This is not a limitation of poor experimental technique. It is encoded in the geometry of the Hilbert space. The two states overlap.

11. The Helstrom bound

For binary minimum-error discrimination between ρ_0 and ρ_1 with priors p_0,p_1, the optimal success probability is

P_success^* = (1/2)(1 + ||p_0ρ_0-p_1ρ_1||_1),

where ||A||_1=Tr√(A†A) is the trace norm.

The optimal measurement projects onto the positive and negative eigenspaces of the Hermitian operator p_0ρ_0-p_1ρ_1, with suitable conventions for any zero eigenspace.

Equal priors and pure states

For equally likely pure states |ψ⟩ and |φ⟩, the minimum error probability is

P_error^*=(1/2)(1-√(1-|⟨ψ|φ⟩|²)).

If the overlap is zero, the error probability is zero. If the states are identical, the best strategy can do no better than guessing according to the priors.

12. Trace distance

The trace distance between two density matrices is

D(ρ,σ)=(1/2)||ρ-σ||_1.

It has a direct operational interpretation: for equal priors, the optimal probability of distinguishing ρ from σ is

P_success^*=(1+D(ρ,σ))/2.

Thus trace distance measures distinguishability. Zero means the states are identical. One means they have orthogonal support and can be perfectly distinguished.

13. Contractivity under quantum channels

Quantum channels cannot increase trace distance:

D(𝓔(ρ),𝓔(σ))≤D(ρ,σ).

Operationally, processing cannot make two states easier to distinguish than they were before if the same channel is applied to both. Noise usually removes distinguishing information rather than creating it from nowhere.

14. Quantum fidelity

Another way to compare states is fidelity. Using the squared-fidelity convention adopted in this guide,

F(ρ,σ)=[Tr√(√ρ σ √ρ)]².

Fidelity lies between zero and one. Larger values indicate greater similarity. If one state is pure, ρ=|ψ⟩⟨ψ|, then

F(|ψ⟩,σ)=⟨ψ|σ|ψ⟩.

If both states are pure,

F(|ψ⟩,|φ⟩)=|⟨ψ|φ⟩|².

Some books define “fidelity” as the square root of this quantity. Always check the convention before comparing formulas.

Worked example: fidelity with the maximally mixed qubit

Let ρ=|0⟩⟨0| and σ=I/2. Because the first state is pure,

F=⟨0|I/2|0⟩=1/2.

15. Fidelity and channels

Applying the same quantum channel to two states cannot make them less alike according to fidelity:

F(𝓔(ρ),𝓔(σ))≥F(ρ,σ).

This is the counterpart of trace-distance contractivity. Physical processing can erase information that distinguished states, making them operationally closer.

16. Relating fidelity and trace distance

With the squared-fidelity convention, the Fuchs–van de Graaf inequalities give

1-√F(ρ,σ) ≤ D(ρ,σ) ≤ √(1-F(ρ,σ)).

The two measures therefore constrain one another. High fidelity forces small trace distance; large trace distance prevents high fidelity.

17. Worked comparison: two pure qubit states

The overlap is ⟨ψ|φ⟩=cosθ, so

F=cos²θ.

For pure states, the trace distance is

D=√(1-F)=|sinθ|.

When θ is small, the states overlap strongly and are difficult to distinguish. At θ=π/2, they are orthogonal, fidelity is zero and trace distance is one.

18. Unambiguous state discrimination

Minimum-error discrimination always returns a guess and accepts some probability of being wrong. Another strategy is unambiguous state discrimination: allow an inconclusive outcome, but require every conclusive identification to be correct.

For suitable linearly independent non-orthogonal pure states, a POVM with three outcomes can identify state 0, identify state 1, or return “inconclusive.” The price of zero error on conclusive outcomes is a non-zero failure probability.

This demonstrates why POVMs are operationally richer than ordinary two-projector measurements.

19. Informationally complete POVMs

A POVM is informationally complete if its outcome probabilities contain enough independent information to reconstruct an unknown density matrix in principle from repeated measurements on many identically prepared systems.

A d-dimensional density matrix has d²-1 independent real parameters after Hermiticity and unit trace are imposed. An informationally complete measurement therefore needs enough independent outcome statistics to determine those parameters.

This is the mathematical foundation of quantum state tomography. Tomography does not extract an unknown state from one copy; it estimates a state from repeated preparations and observed frequencies.

20. A symmetric informationally complete idea

One important family under study is the symmetric informationally complete POVM, or SIC-POVM. In dimension d, a SIC construction—when available—uses d² rank-one effects with equal pairwise overlap structure.

The significance for learners is not the existence question in every dimension. It is the design principle: measurement operators can be arranged geometrically so that the resulting probability data cover state space efficiently and symmetrically.

21. Measurement optimisation is geometry plus probability

State discrimination reveals a recurring quantum-mathematical pattern:

states are positive operators;
measurements are positive operators summing to identity;
probabilities are traces of products;
performance becomes an optimisation problem over operator constraints.

This pattern later connects quantum information to convex optimisation and semidefinite programming.

22. Common misconception: every measurement corresponds to an observable with one Hermitian matrix

That is true for standard projective measurements associated with Hermitian observables, but it is too narrow for general measurement theory. A general POVM is specified by a set of positive effects and need not arise as the spectral projectors of one system observable on the original Hilbert space.

23. Common misconception: POVM elements tell us exactly how the system changes

No. The POVM determines probabilities. A measurement instrument, often represented using measurement operators or completely positive maps indexed by outcomes, is needed to specify post-measurement states.

24. Common misconception: fidelity is a probability that two mixed states are secretly the same

Fidelity is a mathematical similarity measure with operational interpretations, not a hidden-state probability. For pure states it reduces to squared overlap, but for mixed states it compares the full density operators through the positive square-root expression.

25. Worked synthesis problem

Two equally likely pure states are

|ψ⟩=|0⟩

and

|φ⟩=(√3/2)|0⟩+(1/2)|1⟩.

Step 1: Compute overlap. |⟨ψ|φ⟩|²=3/4.

Step 2: Fidelity. F=3/4.

Step 3: Trace distance. For pure states, D=√(1-F)=1/2.

Step 4: Optimal equal-prior success probability.

P_success^*=(1+D)/2=3/4.

Step 5: Interpret. The states are not orthogonal, so no deterministic single-copy measurement can distinguish them perfectly. The optimal minimum-error strategy succeeds three quarters of the time.

26. Practice set

State the two defining conditions for a POVM.
Write the probability rule for POVM outcome k.
Explain why a projective measurement is a special POVM.
State the relationship between a measurement operator M_k and POVM effect E_k.
What additional information is needed to determine the post-measurement state?
Define trace distance.
For equal priors, express optimal binary discrimination success in terms of trace distance.
Define fidelity using the squared-fidelity convention.
For pure states, simplify fidelity.
Explain why a quantum channel cannot increase trace distance between two states.

Answers

Each E_k is positive semidefinite and Σ_kE_k=I.
p(k)=Tr(ρE_k).
Orthogonal projectors are positive and sum to identity.
E_k=M_k†M_k.
A measurement instrument or equivalent outcome-conditioned state-update maps.
D(ρ,σ)=||ρ-σ||_1/2.
(1+D)/2.
[Tr√(√ρσ√ρ)]².
|⟨ψ|φ⟩|².
Data processing cannot create extra distinguishability when the same physical channel acts on both candidates.

27. Series navigation

Guide 5: Density Matrices, Mixed States, Trace and Partial Trace
Guide 6: Quantum Channels, Kraus Operators, Noise and Decoherence
Guide 7: Generalised Measurements, POVMs, State Discrimination and Fidelity
Guide 8: Quantum Entropy, Purity, Mutual Information and Correlations
Return to the BTT Mathematics Learning Hub

Educational note: measurement theory in experiments also includes detector calibration, finite statistics, systematic error and physical implementation constraints. The operator formalism describes the ideal mathematical layer, not every engineering detail.

Small Group Tutorials

Quantum Mathematics Learning Guide 7: Generalised Measurements, POVMs, State Discrimination and Fidelity