Simple Algorithm
Quantum Computing · Excited State · 22 min

Generalized Measurements (POVMs)

Measurements beyond projectors: the most general quantum measurement.

Can a single qubit have a measurement with three outcomes? Or one that sometimes announces "definitely not ∣0⟩\ket0" and is never wrong? Projective measurements cannot do either. Generalized measurements, POVMs, can, and they are what real detectors actually implement.

What you’ll learn

  • Define a POVM and compute outcome probabilities Tr⁡(Ekρ)\Tr(E_k\rho).
  • Derive POVMs from measurement operators, and explain why a POVM does not fix the post-measurement state.
  • Characterize qubit POVMs geometrically and analyze the trine and tetrahedral POVMs.
  • Design the optimal unambiguous measurement for ∣0⟩\ket0 versus ∣+⟩\ket+ and compare it with projective strategies.
  • Implement a POVM by coupling the system to an ancilla and measuring the ancilla projectively.

Measurements beyond projectors

A projective measurement on a qubit has at most two outcomes, one per basis vector. Yet many tasks call for something else:

  • Noisy detectors. A readout that reports 1 with probability 0.980.98 on ∣1⟩\ket1 and 0.030.03 on ∣0⟩\ket0 is not a projector onto anything.
  • Unambiguous identification. To tell ∣0⟩\ket0 from ∣+⟩\ket+ without ever erring, you want three outcomes: "it was ∣0⟩\ket0", "it was ∣+⟩\ket+" and "don't know".
  • More outcomes than dimensions. Tomography and communication often favor measurements with more outcomes than the dimension of the system.

Positivity makes every pk≥0p_k\ge0, and completeness makes them sum to 1 for every state. Projective measurements are the special case of orthogonal projectors, Ek=PkE_k = P_k with PjPk=δjkPkP_jP_k = \delta_{jk}P_k. In general the effects need not be projectors, need not commute, and can outnumber the dimension.

Check your understanding

Which set is a valid POVM on a qubit?

  1. {∣0⟩ ⁣⟨0∣, ∣+⟩ ⁣⟨+∣}\{\dyad0,\ \dyad{+}\}
  2. {12∣0⟩ ⁣⟨0∣, 12∣1⟩ ⁣⟨1∣, 12∣+⟩ ⁣⟨+∣, 12∣−⟩ ⁣⟨−∣}\{\tfrac12\dyad0,\ \tfrac12\dyad1,\ \tfrac12\dyad{+},\ \tfrac12\dyad{-}\}
  3. {12(I+X+12Z), 12(I−X−12Z)}\{\tfrac12(\id+X+\tfrac12Z),\ \tfrac12(\id-X-\tfrac12Z)\}
  4. {23∣0⟩ ⁣⟨0∣, 23∣1⟩ ⁣⟨1∣, 23∣+⟩ ⁣⟨+∣}\{\tfrac23\dyad0,\ \tfrac23\dyad1,\ \tfrac23\dyad{+}\}
Show the answer

B. {12∣0⟩ ⁣⟨0∣, 12∣1⟩ ⁣⟨1∣, 12∣+⟩ ⁣⟨+∣, 12∣−⟩ ⁣⟨−∣}\{\tfrac12\dyad0,\ \tfrac12\dyad1,\ \tfrac12\dyad{+},\ \tfrac12\dyad{-}\}

All four effects are positive, and they sum to 12I+12I=I\tfrac12\id+\tfrac12\id = \id. Operationally: flip a fair coin, then measure ZZ or XX.

Measurement operators and post-measurement states

Where do POVMs come from? The general measurement postulate describes a measurement by measurement operators {Mk}\{M_k\} satisfying ∑kMk†Mk=I\sum_kM_k^\dagger M_k = \id.

DerivationFrom measurement operators to effects
  1. pk=⟨ψ∣Mk†Mk∣ψ⟩,∣ψ⟩ ↦ Mk∣ψ⟩pkp_k = \bra\psi M_k^\dagger M_k\ket\psi,\qquad \ket\psi\ \mapsto\ \frac{M_k\ket\psi}{\sqrt{p_k}}

    The general measurement postulate (Nielsen & Chuang §2.2.3).

  2. Ek:=Mk†Mk ≥0,∑kEk=IE_k := M_k^\dagger M_k\ \ge0,\qquad \sum_kE_k = \id

    Positivity because ⟨v∣Mk†Mk∣v⟩=∥Mk∣v⟩∥2≥0\bra vM_k^\dagger M_k\ket v = \norm{M_k\ket v}^2\ge0; completeness is assumed.

  3. ρ ↦ MkρMk†Tr⁡(Ekρ),pk=Tr⁡(MkρMk†)=Tr⁡(Ekρ)\rho\ \mapsto\ \frac{M_k\rho M_k^\dagger}{\Tr(E_k\rho)},\qquad p_k = \Tr\big(M_k\rho M_k^\dagger\big) = \Tr(E_k\rho)

    For mixed states, by linearity over the ensemble and cyclicity of the trace.

  4. Mk=UkEk  for any unitaries Uk⟹Mk†Mk=EkM_k = U_k\sqrt{E_k}\ \text{ for any unitaries } U_k\quad\Longrightarrow\quad M_k^\dagger M_k = E_k

    Conversely, every POVM arises this way, but not uniquely: the probabilities fix EkE_k, not MkM_k. The post-measurement state depends on how the measurement is implemented.

Think first Suppose the noisy readout of step 1 were implemented as the Lüders instrument, Mk=EkM_k = \sqrt{E_k}. The input is ∣+⟩\ket+ and the outcome is 1. What is the post-measurement state?

M1=diag⁡(0.03,0.98)M_1 = \diag\big(\sqrt{0.03},\sqrt{0.98}\big), so the state becomes (0.03 ∣0⟩+0.98 ∣1⟩)/1.01\big(\sqrt{0.03}\,\ket0+\sqrt{0.98}\,\ket1\big)/\sqrt{1.01}: mostly ∣1⟩\ket1, but still a coherent superposition. An unsharp Lüders measurement only partially collapses the state. Real readouts typically destroy the remaining coherence, or the system itself, which is exactly why the POVM alone cannot tell you the post-measurement state.

The geometry of qubit POVMs

Write each qubit effect in Bloch form. Positivity and completeness then become simple geometry:

Ek=ak(I+m⃗k⋅σ⃗),ak≥0,∣m⃗k∣≤1,∑kak=1,∑kakm⃗k=0⃗E_k = a_k\big(\id+\vec m_k\cdot\vec\sigma\big),\qquad a_k\ge0,\quad \abs{\vec m_k}\le1,\qquad \sum_ka_k = 1,\quad \sum_ka_k\vec m_k = \vec0
Ek≥0E_k\ge0 iff ∣m⃗k∣≤1\abs{\vec m_k}\le1, since its eigenvalues are ak(1±∣m⃗k∣)a_k(1\pm\abs{\vec m_k}). ∑kEk=I\sum_kE_k = \id iff the weights sum to one and the weighted vectors balance at the origin. Outcome probabilities are pk=Tr⁡(Ekρ)=ak(1+m⃗k⋅r⃗)p_k = \Tr(E_k\rho) = a_k\big(1+\vec m_k\cdot\vec r\big).
POVMEffectsBloch picture
Projective ZZ∣0⟩ ⁣⟨0∣, ∣1⟩ ⁣⟨1∣\dyad0,\ \dyad1two antipodal unit vectors, weights 12\tfrac12
Trine23∣ψk⟩ ⁣⟨ψk∣\tfrac23\dyad{\psi_k}, k=0,1,2k = 0,1,2three unit vectors 120∘120^\circ apart in a plane, weights 13\tfrac13
Tetrahedron (SIC)12∣ψk⟩ ⁣⟨ψk∣\tfrac12\dyad{\psi_k}, k=0,…,3k = 0,\dots,3the vertices of a regular tetrahedron, weights 14\tfrac14
Unsharp ZZ12(I±ηZ)\tfrac12(\id\pm\eta Z), 0≤η≤10\le\eta\le1two antipodal vectors of length η\eta, weights 12\tfrac12
Predict before you look

You measure ∣0⟩\ket0 with the trine POVM whose Bloch vectors are z^\hat z and two vectors at 120∘120^\circ from it. What is the probability of one of the two other outcomes (not the z^\hat z one)?

Show the answer

0.17

pk=13(1+m⃗k⋅z^)=13(1+cos⁡120∘)=13⋅12=16≈0.17p_k = \tfrac13\big(1+\vec m_k\cdot\hat z\big) = \tfrac13\big(1+\cos120^\circ\big) = \tfrac13\cdot\tfrac12 = \tfrac16\approx0.17. The z^\hat z outcome has 13(1+1)=23\tfrac13(1+1) = \tfrac23, and indeed 23+16+16=1\tfrac23+\tfrac16+\tfrac16 = 1.

Notice what the trine buys: three outcomes from a two-dimensional system, and outcome kk never occurs on the state orthogonal to ∣ψk⟩\ket{\psi_k}. Such exclusion events are the key to the next step.

Unambiguous discrimination: never wrong, sometimes silent

You receive ∣0⟩\ket0 or ∣+⟩\ket+ with equal probability. Since ⟨0∣+⟩=1/2≠0\braket{0|+} = 1/\sqrt2\ne0, no measurement identifies them perfectly. Suppose you insist on never being wrong, at the price of sometimes answering "inconclusive".

Projective strategy. Measure ZZ. Outcome 11 is impossible for ∣0⟩\ket0, so it certifies ∣+⟩\ket+; outcome 00 is inconclusive. This succeeds with probability 12⋅∣⟨1∣+⟩∣2=14\tfrac12\cdot\abs{\braket{1|+}}^2 = \tfrac14, and measuring XX instead does no better.

Predict before you look

Can a POVM beat 25%? Predict the optimal probability of a conclusive (and therefore correct) answer.

Show the answer

0.293

The optimum is 1−∣⟨0∣+⟩∣=1−1/2≈0.2931-\abs{\braket{0|+}} = 1-1/\sqrt2\approx0.293, the Ivanovic–Dieks–Peres limit, derived below.

DerivationThe optimal unambiguous POVM for |0⟩ versus |+⟩
  1. E+=c ∣1⟩ ⁣⟨1∣,E0=c ∣−⟩ ⁣⟨−∣,E?=I−E+−E0E_{+} = c\,\dyad{1},\qquad E_{0} = c\,\dyad{-},\qquad E_{?} = \id-E_+-E_0

    E+E_+ (announce "∣+⟩\ket+") must never fire on ∣0⟩\ket0, so it is proportional to ∣1⟩ ⁣⟨1∣\dyad1; likewise E0E_0 must vanish on ∣+⟩\ket+.

  2. ∣1⟩ ⁣⟨1∣+∣−⟩ ⁣⟨−∣=(12−12−1232),λmax⁡=1+12\dyad{1}+\dyad{-} = \begin{pmatrix}\tfrac12 & -\tfrac12\\ -\tfrac12 & \tfrac32\end{pmatrix},\qquad \lambda_{\max} = 1+\tfrac{1}{\sqrt2}

    Trace 2 and determinant 34−14=12\tfrac34-\tfrac14 = \tfrac12, so the eigenvalues are 1±121\pm\tfrac{1}{\sqrt2}.

  3. E?≥0  ⟺  c λmax⁡≤1  ⟺  c≤11+1/2=2−2E_?\ge0\iff c\,\lambda_{\max}\le1\iff c\le\frac{1}{1+1/\sqrt2} = 2-\sqrt2

    Make the conclusive effects as large as positivity of E?E_? allows.

  4. Psucc=12⟨0∣E0∣0⟩+12⟨+∣E+∣+⟩=12⋅c2+12⋅c2=c2=1−12≈0.293P_{\text{succ}} = \tfrac12\bra0E_0\ket0+\tfrac12\bra+E_+\ket+ = \tfrac12\cdot\tfrac c2+\tfrac12\cdot\tfrac c2 = \tfrac c2 = 1-\tfrac{1}{\sqrt2}\approx0.293

    Three outcomes on a qubit, so this is a genuine POVM, and it is what beats the projective 25%.

Every POVM is a projective measurement in disguise

How do you build a POVM in the lab? Couple the system to an ancilla and measure the ancilla projectively. Given measurement operators {Mk}\{M_k\}, define

V∣ψ⟩=∑kMk∣ψ⟩⊗∣k⟩anc,V†V=∑kMk†Mk=IV\ket\psi = \sum_kM_k\ket\psi\otimes\ket{k}_{\text{anc}},\qquad V^\dagger V = \sum_kM_k^\dagger M_k = \id
VV is an isometry, so it extends to a unitary on system ⊗ ancilla with the ancilla starting in ∣0⟩\ket0. Measuring the ancilla in its computational basis gives outcome kk with probability ∥Mk∣ψ⟩∥2=⟨ψ∣Ek∣ψ⟩\norm{M_k\ket\psi}^2 = \bra\psi E_k\ket\psi and leaves the system in Mk∣ψ⟩/pkM_k\ket\psi/\sqrt{p_k}.

Here is a noisy detector built this way. The system (top wire) controls a rotation RY(θ)R_Y(\theta) of the ancilla (bottom wire):

∣0⟩∣0⟩↦∣0⟩∣0⟩,∣1⟩∣0⟩↦∣1⟩(cos⁡θ2∣0⟩+sin⁡θ2∣1⟩)\ket0\ket0\mapsto\ket0\ket0,\qquad \ket1\ket0\mapsto\ket1\big(\cos\tfrac\theta2\ket0+\sin\tfrac\theta2\ket1\big)
Reading the ancilla realizes the POVM Eclick=sin⁡2θ2 ∣1⟩ ⁣⟨1∣E_{\text{click}} = \sin^2\tfrac\theta2\,\dyad1, Eno click=∣0⟩ ⁣⟨0∣+cos⁡2θ2 ∣1⟩ ⁣⟨1∣E_{\text{no click}} = \dyad0+\cos^2\tfrac\theta2\,\dyad1: a detector with efficiency η=sin⁡2θ2\eta = \sin^2\tfrac\theta2 and no dark counts.

Key takeaways

  • A POVM is any set of effects Ek≥0E_k\ge0 with ∑kEk=I\sum_kE_k = \id; outcome probabilities are Tr⁡(Ekρ)\Tr(E_k\rho).
  • Measurement operators MkM_k with Mk†Mk=EkM_k^\dagger M_k = E_k determine the post-measurement state; the POVM alone does not.
  • Qubit POVMs are weighted Bloch vectors that balance at the origin; they can have more outcomes than dimensions (trine, tetrahedron).
  • POVMs beat projective measurements at some tasks: unambiguous discrimination of ∣0⟩\ket0 versus ∣+⟩\ket+ reaches 1−1/2≈0.291-1/\sqrt2\approx0.29 instead of 0.250.25.
  • Every POVM is realized by an ancilla, a unitary coupling and a projective measurement of the ancilla.

Check yourself

In the interactive lesson these are scored with your confidence, and what you miss comes back for review.

Question 1

Two labs implement the same POVM {Ek}\{E_k\} with different measurement operators. What can differ between the labs?

  1. The outcome probabilities for some input state
  2. The post-measurement states
  3. Whether the probabilities sum to one
  4. Nothing: the POVM determines everything
Show the answer

B. The post-measurement states

pk=Tr⁡(Ekρ)p_k = \Tr(E_k\rho) depends only on Ek=Mk†MkE_k = M_k^\dagger M_k. But MkM_k and Mk′=UkMkM'_k = U_kM_k produce post-measurement states that differ by the unitary UkU_k, and other implementations may not even preserve the system.

Question 2

Two equally likely pure states have ∣⟨ψ0∣ψ1⟩∣=0.8\abs{\braket{\psi_0|\psi_1}} = 0.8. What is the optimal probability of an unambiguous (never wrong) identification?

  1. 0.80.8
  2. 0.360.36
  3. 0.20.2
  4. 0.60.6
Show the answer

C. 0.20.2

The Ivanovic–Dieks–Peres limit is 1−∣⟨ψ0∣ψ1⟩∣=0.21-\abs{\braket{\psi_0|\psi_1}} = 0.2.

Question 3

A detector reports 1 with probability 0.98 when the qubit is ∣1⟩\ket1 and with probability 0.03 when it is ∣0⟩\ket0. Assuming its effects are diagonal in the ZZ basis, what is E1E_1?

  1. 0.03∣0⟩ ⁣⟨0∣+0.98∣1⟩ ⁣⟨1∣0.03\dyad0+0.98\dyad1
  2. 0.98∣1⟩ ⁣⟨1∣0.98\dyad1
  3. 0.97∣0⟩ ⁣⟨0∣+0.02∣1⟩ ⁣⟨1∣0.97\dyad0+0.02\dyad1
  4. ∣1⟩ ⁣⟨1∣\dyad1
Show the answer

A. 0.03∣0⟩ ⁣⟨0∣+0.98∣1⟩ ⁣⟨1∣0.03\dyad0+0.98\dyad1

⟨0∣E1∣0⟩=0.03\bra0E_1\ket0 = 0.03 and ⟨1∣E1∣1⟩=0.98\bra1E_1\ket1 = 0.98 are exactly the two click probabilities.

Question 4

Which statement about POVMs on a qubit is true?

  1. A POVM on a qubit can have at most two outcomes with nonzero effects
  2. The effects of a qubit POVM must commute with each other
  3. Every qubit POVM is a projective measurement in some basis
  4. A qubit POVM can have three or more outcomes, but its weighted Bloch vectors akm⃗ka_k\vec m_k must sum to zero
Show the answer

D. A qubit POVM can have three or more outcomes, but its weighted Bloch vectors akm⃗ka_k\vec m_k must sum to zero

Writing Ek=ak(I+m⃗k⋅σ⃗)E_k = a_k(\id+\vec m_k\cdot\vec\sigma), completeness forces ∑kak=1\sum_ka_k = 1 and ∑kakm⃗k=0⃗\sum_ka_k\vec m_k = \vec0, and any number of outcomes is allowed. The trine and the tetrahedron are examples.

Question 5

In the dilation V∣ψ⟩=∑kMk∣ψ⟩⊗∣k⟩V\ket\psi = \sum_kM_k\ket\psi\otimes\ket k, why is VV an isometry?

  1. Because each MkM_k is unitary
  2. Because the MkM_k commute
  3. Because V†V=∑kMk†Mk=IV^\dagger V = \sum_kM_k^\dagger M_k = \id
  4. Because the ancilla is measured in an orthonormal basis
Show the answer

C. Because V†V=∑kMk†Mk=IV^\dagger V = \sum_kM_k^\dagger M_k = \id

V†V=∑k,lMk†Ml⟨k∣l⟩=∑kMk†Mk=IV^\dagger V = \sum_{k,l}M_k^\dagger M_l\braket{k|l} = \sum_kM_k^\dagger M_k = \id: the completeness relation.

Go deeper

Further reading

  • M. A. Nielsen & I. L. Chuang, Quantum Computation and Quantum Information, §2.2.3 and §2.2.6 (Cambridge, 10th anniv. ed., 2010).
  • J. Preskill, Lecture Notes for Ph219/CS219: Quantum Information and Computation, Ch. 3, "Foundations II: Measurement and Evolution" (Caltech).
  • A. Peres, "How to differentiate between non-orthogonal states", Physics Letters A 128, 19 (1988); see also I. D. Ivanovic (1987) and D. Dieks (1988).