Simple Algorithm
Quantum Computing · Rydberg State · 26 min

Naimark Dilation

Every POVM is a projective measurement on a bigger system.

Generalized measurements look like a new postulate, but they are not: every POVM is an ordinary projective measurement on a larger system, seen through a window. Naimark's theorem makes this precise, tells you how large the window must be, and shows how to wire it up with an ancilla.

What you’ll learn

  • State and prove Naimark's dilation theorem in its ancilla (tensor-product) and direct-sum forms.
  • Compute the minimal dilation dimension ∑krank⁡Ek\sum_k\rank E_k, and construct explicit dilations for the trine and for optimal unambiguous discrimination.
  • Relate Naimark's theorem to Stinespring's theorem and to Ozawa's measurement models for instruments.
  • Characterize extremal POVMs and prove that they satisfy ∑k(rank⁡Ek)2≤d2\sum_k(\rank E_k)^2\le d^2.

The dilation theorem

DerivationProof 1: the ancilla (tensor-product) dilation
  1. K=H⊗Cm,V∣ψ⟩=∑k=1mEk∣ψ⟩⊗∣k⟩,Pk=I⊗∣k⟩ ⁣⟨k∣\mathcal K = \Hil\otimes\C^m,\qquad V\ket\psi = \sum_{k=1}^m\sqrt{E_k}\ket\psi\otimes\ket k,\qquad P_k = \id\otimes\dyad k

    Attach an mm-level ancilla, one level per outcome.

  2. V†V=∑k,lEkEl ⟨k∣l⟩=∑kEk=IV^\dagger V = \sum_{k,l}\sqrt{E_k}\sqrt{E_l}\,\braket{k|l} = \sum_kE_k = \id

    VV is an isometry because the effects sum to the identity.

  3. V†PkV=Ek ⟨k∣k⟩ Ek=Ek□V^\dagger P_kV = \sqrt{E_k}\,\braket{k|k}\,\sqrt{E_k} = E_k\quad\square

    Physically: prepare the ancilla in ∣0⟩\ket0, apply any unitary UU with U(∣ψ⟩∣0⟩)=V∣ψ⟩U(\ket\psi\ket0) = V\ket\psi, and measure the ancilla. The system is left in Ek∣ψ⟩/pk\sqrt{E_k}\ket\psi/\sqrt{p_k}: this dilation implements the Lüders instrument.

DerivationProof 2: the direct-sum (minimal) dilation
  1. AA†=∑k,j∣e~kj⟩ ⁣⟨e~kj∣=∑kEk=IdAA^\dagger = \sum_{k,j}\dyad{\tilde e_{kj}} = \sum_kE_k = \id_d

    Split each effect into rank-one pieces, Ek=∑j∣e~kj⟩ ⁣⟨e~kj∣E_k = \sum_j\dyad{\tilde e_{kj}} (for instance from its spectral decomposition), for n=∑krank⁡Ekn = \sum_k\rank E_k vectors in total. Make them the columns of a d×nd\times n matrix AA:

  2. ⟨wkj∣ψ⟩=⟨e~kj∣ψ⟩for all ∣ψ⟩∈Cd⊂Cn\braket{w_{kj}|\psi} = \braket{\tilde e_{kj}|\psi}\qquad\text{for all }\ket\psi\in\C^d\subset\C^n

    So the dd rows of AA are orthonormal vectors in Cn\C^n. Complete them by n−dn-d further rows to an n×nn\times n unitary WW. Its columns ∣wkj⟩\ket{w_{kj}} form an orthonormal basis of K=Cn\mathcal K = \C^n, and if H=Cd\Hil = \C^d is embedded as the first dd coordinates,

  3. Hence Pk=∑j∣wkj⟩ ⁣⟨wkj∣P_k = \sum_j\dyad{w_{kj}} satisfies V†PkV=∑j∣e~kj⟩ ⁣⟨e~kj∣=EkV^\dagger P_kV = \sum_j\dyad{\tilde e_{kj}} = E_k, with VV the inclusion and dim⁡K=∑krank⁡Ek\dim\mathcal K = \sum_k\rank E_k. □\square

Examples: a qutrit for the trine, two qubits for discrimination

The trine POVM Ek=23∣ψk⟩ ⁣⟨ψk∣E_k = \tfrac23\dyad{\psi_k}, with real states ∣ψk⟩=cos⁡θk2∣0⟩+sin⁡θk2∣1⟩\ket{\psi_k} = \cos\tfrac{\theta_k}2\ket0+\sin\tfrac{\theta_k}2\ket1 and θk=0, 2π3, 4π3\theta_k = 0,\ \tfrac{2\pi}3,\ \tfrac{4\pi}3, has three rank-one effects, so its minimal dilation is a qutrit. The vectors ∣e~k⟩=2/3 ∣ψk⟩\ket{\tilde e_k} = \sqrt{2/3}\,\ket{\psi_k} form the first two rows of

W=(2/31/6−1/601/21/21/3−1/31/3)W = \begin{pmatrix}\sqrt{2/3} & \sqrt{1/6} & -\sqrt{1/6}\\ 0 & \sqrt{1/2} & \sqrt{1/2}\\ \sqrt{1/3} & -\sqrt{1/3} & \sqrt{1/3}\end{pmatrix}
The third row is the unique (up to sign) completion to an orthogonal matrix. Embed the qubit as the first two levels of a qutrit and measure in the basis of columns of WW: that measurement is the trine.

Two qubits span four dimensions, enough to dilate any rank-one qubit POVM with up to four outcomes. Below is a dilation of the optimal unambiguous POVM for ∣0⟩\ket0 versus ∣+⟩\ket+, with the system on the top wire and one ancilla qubit below. Outcome bits are written system first.

  1. RY(3π/4)R_Y(3\pi/4) rotates the two states symmetrically about ∣1⟩\ket1: ∣0⟩↦sin⁡π8∣0⟩+cos⁡π8∣1⟩\ket0\mapsto\sin\tfrac\pi8\ket0+\cos\tfrac\pi8\ket1 and ∣+⟩↦−sin⁡π8∣0⟩+cos⁡π8∣1⟩\ket+\mapsto-\sin\tfrac\pi8\ket0+\cos\tfrac\pi8\ket1.
  2. A controlled RY(β)R_Y(\beta) with cos⁡β2=tan⁡π8\cos\tfrac\beta2 = \tan\tfrac\pi8 (so β≈131∘\beta\approx131^\circ) moves part of the common ∣1⟩\ket1 component into the ancilla's ∣1⟩\ket1. In the ancilla-∣0⟩\ket0 branch the system is left in ±sin⁡π8∣0⟩+sin⁡π8∣1⟩\pm\sin\tfrac\pi8\ket0+\sin\tfrac\pi8\ket1, proportional to ∣+⟩\ket+ or ∣−⟩\ket-: now orthogonal.
  3. A final HH maps these to ∣0⟩\ket0 and ∣1⟩\ket1. Outcome 0000 certifies the input ∣0⟩\ket0, outcome 1010 certifies ∣+⟩\ket+, and any outcome with the ancilla at 11 is inconclusive.
Check your understanding

What is the minimal dimension of a Naimark dilation for the qubit SIC-POVM, whose four effects are 12∣ψk⟩ ⁣⟨ψk∣\tfrac12\dyad{\psi_k}?

  1. 22
  2. 33
  3. 44
  4. 88
Show the answer

C. 44

∑krank⁡Ek=4\sum_k\rank E_k = 4, so the minimal dilation space is C4\C^4: for example the qubit plus one ancilla qubit, measured in an orthonormal basis of two-qubit states.

Instruments, Stinespring and measurement models

Naimark's theorem is the measurement half of a bigger picture. A quantum instrument is a family of completely positive maps {Φk}\{\Phi_k\} whose sum is trace preserving. Outcome kk occurs with probability Tr⁡Φk(ρ)\Tr\Phi_k(\rho) and leaves the state Φk(ρ)/Tr⁡Φk(ρ)\Phi_k(\rho)/\Tr\Phi_k(\rho); the instrument's POVM is Ek=Φk†(I)E_k = \Phi_k^\dagger(\id).

For finitely many outcomes this follows from Stinespring's theorem. Write each Φk(ρ)=∑jKkjρKkj†\Phi_k(\rho) = \sum_jK_{kj}\rho K_{kj}^\dagger in Kraus form and use the isometry V∣ψ⟩=∑k,jKkj∣ψ⟩⊗∣k⟩∣j⟩V\ket\psi = \sum_{k,j}K_{kj}\ket\psi\otimes\ket{k}\ket{j}, which satisfies V†V=∑k,jKkj†Kkj=IV^\dagger V = \sum_{k,j}K_{kj}^\dagger K_{kj} = \id. Measuring the ∣k⟩\ket k register realizes the instrument, while the ∣j⟩\ket j register is discarded.

Naimark is the special case in which only the outcome probabilities matter. In operator-algebra language, Naimark's theorem is Stinespring's 1955 dilation theorem applied to the positive (hence completely positive) map that a POVM defines on the commutative algebra of functions on the outcome set.

Extremal POVMs and the d² bound

POVMs with a fixed outcome set form a convex set: mixing two POVMs with probabilities tt and 1−t1-t gives the POVM {tFk+(1−t)Gk}\{tF_k+(1-t)G_k\}. Extremal POVMs cannot be decomposed this way; they are exactly the measurements that contain no classical randomness.

DerivationProof sketch
  1. ∑k(Fk−Ek)=I−I=0\sum_k\big(F_k-E_k\big) = \id-\id = 0

    Suppose Ek=tFk+(1−t)GkE_k = tF_k+(1-t)G_k for POVMs FF, GG and 0<t<10<t<1. Since 0≤tFk≤Ek0\le tF_k\le E_k, each FkF_k is supported in supp⁡Ek\supp E_k, so Fk−Ek=∑i,jcij(k)∣eki⟩ ⁣⟨ekj∣F_k-E_k = \sum_{i,j}c^{(k)}_{ij}\ketbra{e_{ki}}{e_{kj}} for some coefficients cij(k)c^{(k)}_{ij}, and completeness of both POVMs gives

  2. If the operators ∣eki⟩ ⁣⟨ekj∣\ketbra{e_{ki}}{e_{kj}} are linearly independent, every c(k)c^{(k)} vanishes, so F=EF = E and the POVM is extremal. Conversely, a nontrivial relation ∑kDk=0\sum_kD_k = 0 with Dk=∑ijcij(k)∣eki⟩ ⁣⟨ekj∣D_k = \sum_{ij}c^{(k)}_{ij}\ketbra{e_{ki}}{e_{kj}} can be taken Hermitian (use its Hermitian or anti-Hermitian part). Since EkE_k is strictly positive on its support, Ek±ϵDk≥0E_k\pm\epsilon D_k\ge0 for small ϵ\epsilon, which writes EE as the midpoint of two different POVMs.

  3. ∑k(rank⁡Ek)2 ≤ d2\sum_k\big(\rank E_k\big)^2\ \le\ d^2

    The operators ∣eki⟩ ⁣⟨ekj∣\ketbra{e_{ki}}{e_{kj}} live in the d2d^2-dimensional space of operators on Cd\C^d, and there are ∑k(rank⁡Ek)2\sum_k(\rank E_k)^2 of them; linear independence forces the bound. □\square

This structure explains Davies' theorem (1978): the accessible information of an ensemble is attained by a rank-one POVM with at most d2d^2 outcomes, because mutual information is convex in the POVM and so is maximized at an extreme point. It also sorts the qubit examples. The trine (3 outcomes) and the SIC (4 outcomes) have linearly independent rank-one pieces and are extremal. "Random ZZ or XX" is not, because ∣0⟩ ⁣⟨0∣+∣1⟩ ⁣⟨1∣=∣+⟩ ⁣⟨+∣+∣−⟩ ⁣⟨−∣\dyad0+\dyad1 = \dyad{+}+\dyad{-} is a linear dependence.

Key takeaways

  • Naimark: every POVM has the form Ek=V†PkVE_k = V^\dagger P_kV, a projective measurement on a larger space seen through an isometry.
  • The ancilla dilation V∣ψ⟩=∑kEk∣ψ⟩∣k⟩V\ket\psi = \sum_k\sqrt{E_k}\ket\psi\ket k implements the Lüders instrument; the direct-sum dilation achieves the minimal dimension ∑krank⁡Ek\sum_k\rank E_k.
  • Stinespring generalizes Naimark from POVMs to instruments and channels, and Ozawa's theorem gives every instrument an indirect measurement model.
  • Extremal POVMs have linearly independent rank-one pieces, hence ∑k(rank⁡Ek)2≤d2\sum_k(\rank E_k)^2\le d^2; optimal measurements can be sought among them.

Check yourself

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

Question 1

Which statement about the ancilla dilation V∣ψ⟩=∑kEk∣ψ⟩⊗∣k⟩V\ket\psi = \sum_k\sqrt{E_k}\ket\psi\otimes\ket k is true?

  1. It realizes the POVM, and its post-measurement states are those of the Lüders instrument, EkρEk/pk\sqrt{E_k}\rho\sqrt{E_k}/p_k
  2. It realizes the POVM only if the effects EkE_k commute
  3. It is always a minimal dilation
  4. It requires the effects EkE_k to be projectors
Show the answer

A. It realizes the POVM, and its post-measurement states are those of the Lüders instrument, EkρEk/pk\sqrt{E_k}\rho\sqrt{E_k}/p_k

Measuring the ancilla in ∣k⟩\ket k leaves the system in Ek∣ψ⟩\sqrt{E_k}\ket\psi (normalized), so this dilation implements the Lüders instrument. It works for every POVM, but its dimension dmdm usually exceeds the minimum ∑krank⁡Ek\sum_k\rank E_k.

Question 2

Is the qubit POVM {12∣0⟩ ⁣⟨0∣, 12∣1⟩ ⁣⟨1∣, 12∣+⟩ ⁣⟨+∣, 12∣−⟩ ⁣⟨−∣}\{\tfrac12\dyad0,\ \tfrac12\dyad1,\ \tfrac12\dyad{+},\ \tfrac12\dyad{-}\} extremal?

  1. Yes, because it has d2=4d^2 = 4 outcomes
  2. Yes, because all of its effects have rank one
  3. No: ∣0⟩ ⁣⟨0∣+∣1⟩ ⁣⟨1∣=∣+⟩ ⁣⟨+∣+∣−⟩ ⁣⟨−∣\dyad0+\dyad1 = \dyad{+}+\dyad{-} is a linear dependence, and indeed it is an equal mixture of the ZZ and XX measurements
  4. No, because extremal POVMs must be projective
Show the answer

C. No: ∣0⟩ ⁣⟨0∣+∣1⟩ ⁣⟨1∣=∣+⟩ ⁣⟨+∣+∣−⟩ ⁣⟨−∣\dyad0+\dyad1 = \dyad{+}+\dyad{-} is a linear dependence, and indeed it is an equal mixture of the ZZ and XX measurements

By the D'Ariano–Lo Presti–Perinotti criterion, extremality requires the rank-one pieces to be linearly independent, and here they are not. Explicitly, it is 12(Z measurement)+12(X measurement)\tfrac12(Z\text{ measurement})+\tfrac12(X\text{ measurement}), each padded with two zero effects.

Question 3

Naimark's theorem is a special case of which result?

  1. The spectral theorem
  2. Stinespring's dilation theorem for completely positive maps
  3. The Schmidt decomposition
  4. Gleason's theorem
Show the answer

B. Stinespring's dilation theorem for completely positive maps

A POVM defines a positive map from the commutative algebra of functions on the outcome set into the operators on H\Hil. Such maps are automatically completely positive, and their Stinespring dilation is exactly a Naimark dilation with a projection-valued measure.

Question 4

In the two-qubit dilation for ∣0⟩\ket0 versus ∣+⟩\ket+, why does outcome 0000 never occur for input ∣+⟩\ket+?

  1. Because the controlled rotation acts only when the system is in ∣0⟩\ket0
  2. Because ∣+⟩\ket+ is orthogonal to ∣0⟩\ket0
  3. Because the ancilla always ends in ∣1⟩\ket1 for input ∣+⟩\ket+
  4. Because in the ancilla-∣0⟩\ket0 branch the system is proportional to ∣−⟩\ket-, which the final HH maps to ∣1⟩\ket1
Show the answer

D. Because in the ancilla-∣0⟩\ket0 branch the system is proportional to ∣−⟩\ket-, which the final HH maps to ∣1⟩\ket1

After RY(3π/4)R_Y(3\pi/4) and the controlled rotation, the ancilla-∣0⟩\ket0 branch holds sin⁡π8(∣0⟩+∣1⟩)\sin\tfrac\pi8(\ket0+\ket1) for input ∣0⟩\ket0 and −sin⁡π8(∣0⟩−∣1⟩)-\sin\tfrac\pi8(\ket0-\ket1) for input ∣+⟩\ket+. These are orthogonal, and HH sends them to ∣0⟩\ket0 and ∣1⟩\ket1. The effect for outcome 0000 is therefore proportional to ∣−⟩ ⁣⟨−∣\dyad{-}, which annihilates ∣+⟩\ket+.

Further reading

  • J. Watrous, The Theory of Quantum Information, Ch. 2 (Cambridge, 2018).
  • A. Peres, Quantum Theory: Concepts and Methods (Kluwer, 1993).
  • W. F. Stinespring, "Positive functions on C∗C^*-algebras", Proceedings of the American Mathematical Society 6, 211 (1955).
  • M. Ozawa, "Quantum measuring processes of continuous observables", Journal of Mathematical Physics 25, 79 (1984).
  • G. M. D'Ariano, P. Lo Presti & P. Perinotti, "Classical randomness in quantum measurements", Journal of Physics A 38, 5979 (2005).