Simple Algorithm
Quantum Computing · Rydberg State · 30 min

Ensembles, Purification & the HJW Theorem

Different mixtures, same density matrix, and why it matters.

The same density matrix can be built from infinitely many ensembles, and no experiment can tell them apart. The Hughston–Jozsa–Wootters theorem says exactly how these ensembles are related, and shows that whoever holds a purification of your system can remotely choose which ensemble you have, without ever changing your ρ\rho.

What you’ll learn

  • State and prove the unitary freedom in the ensemble for a density operator.
  • Construct purifications and prove that they are unique up to an isometry on the purifying system.
  • Prove the Schrödinger–HJW steering theorem and derive no-signalling from it.
  • Characterize which pure states, with which weights, can appear in an ensemble for a given ρ\rho (support condition, pmax⁡p_{\max}, majorization).

Unitary freedom in the ensemble

Work with unnormalized vectors ∣ψ~i⟩=pi∣ψi⟩\ket{\tilde\psi_i} = \sqrt{p_i}\ket{\psi_i}, so that an ensemble is just a list of vectors with ρ=∑i∣ψ~i⟩ ⁣⟨ψ~i∣\rho = \sum_i\dyad{\tilde\psi_i}, and the weights are recovered as pi=⟨ψ~i∣ψ~i⟩p_i = \braket{\tilde\psi_i|\tilde\psi_i}. Pad shorter lists with zero vectors whenever two lists are compared.

DerivationProof
  1. ∑i∣ψ~i⟩ ⁣⟨ψ~i∣=∑j,k(∑iuijuik∗)∣ϕ~j⟩ ⁣⟨ϕ~k∣=∑j∣ϕ~j⟩ ⁣⟨ϕ~j∣\sum_i\dyad{\tilde\psi_i} = \sum_{j,k}\Big(\sum_iu_{ij}u_{ik}^*\Big)\ketbra{\tilde\phi_j}{\tilde\phi_k} = \sum_j\dyad{\tilde\phi_j}

    (⇐) Direct computation, using ∑iuijuik∗=(u†u)kj=δjk\sum_iu_{ij}u^*_{ik} = (u^\dagger u)_{kj} = \delta_{jk}:

  2. ∣ψ~i⟩=∑k=1rcik∣k~⟩,cik=⟨k∣ψ~i⟩λk\ket{\tilde\psi_i} = \sum_{k=1}^rc_{ik}\ket{\tilde k},\qquad c_{ik} = \frac{\braket{k|\tilde\psi_i}}{\sqrt{\lambda_k}}

    (⇒) Let ρ=∑k=1rλk∣k⟩ ⁣⟨k∣\rho = \sum_{k=1}^r\lambda_k\dyad k with λk>0\lambda_k>0, and set ∣k~⟩=λk∣k⟩\ket{\tilde k} = \sqrt{\lambda_k}\ket k. Every ∣ψ~i⟩\ket{\tilde\psi_i} lies in supp⁡ρ\supp\rho: if ρ∣v⟩=0\rho\ket v = 0 then 0=⟨v∣ρ∣v⟩=∑i∣⟨v∣ψ~i⟩∣20 = \bra v\rho\ket v = \sum_i\abs{\braket{v|\tilde\psi_i}}^2. So we can expand

  3. ∑k,l(∑icikcil∗)∣k~⟩ ⁣⟨l~∣=∑k∣k~⟩ ⁣⟨k~∣⟹c†c=Ir\sum_{k,l}\Big(\sum_ic_{ik}c_{il}^*\Big)\ketbra{\tilde k}{\tilde l} = \sum_k\dyad{\tilde k}\quad\Longrightarrow\quad c^\dagger c = \id_r

    Substitute into ρ=∑i∣ψ~i⟩ ⁣⟨ψ~i∣\rho = \sum_i\dyad{\tilde\psi_i} and compare coefficients of the linearly independent operators ∣k~⟩ ⁣⟨l~∣\ketbra{\tilde k}{\tilde l}:

  4. ψ~=U U′† ϕ~,u:=UU′† is unitary.□\tilde\psi = U\,U'^{\dagger}\,\tilde\phi,\qquad u := UU'^{\dagger}\ \text{is unitary.}\quad\square

    So the N×rN\times r matrix cc (zero rows for padding) has orthonormal columns; complete it to an N×NN\times N unitary UU. Viewing each list as a column of kets and padding the eigen-ensemble k~\tilde k with zeros, ψ~=Uk~\tilde\psi = U\tilde k. Likewise ϕ~=U′k~\tilde\phi = U'\tilde k, hence

Example: I/2\id/2 arises from {∣0⟩,∣1⟩}\{\ket0,\ket1\}, from {∣+⟩,∣−⟩}\{\ket+,\ket-\} (related by u=Hu = H), and from the trine {13:∣ψk⟩}k=0,1,2\{\tfrac13:\ket{\psi_k}\}_{k=0,1,2} whose Bloch vectors are 120∘120^\circ apart on a great circle. Comparing the trine with {∣0⟩,∣1⟩}\{\ket0,\ket1\} needs one zero vector of padding and a 3×33\times3 unitary.

Predict before you look

Let ρ=34∣0⟩ ⁣⟨0∣+14∣1⟩ ⁣⟨1∣\rho = \tfrac34\dyad0+\tfrac14\dyad1. What is the largest weight pp with which ∣+⟩\ket+ can appear in an ensemble for ρ\rho?

Show the answer

0.375

ρ−1=diag⁡(43,4)\rho^{-1} = \diag(\tfrac43,4), so ⟨+∣ρ−1∣+⟩=12(43+4)=83\bra+\rho^{-1}\ket+ = \tfrac12\big(\tfrac43+4\big) = \tfrac83 and pmax⁡=38p_{\max} = \tfrac38. Bloch-ball check: with r⃗=(0,0,12)\vec r = (0,0,\tfrac12), the remainder's Bloch vector (r⃗−px^)/(1−p)(\vec r-p\hat x)/(1-p) stays in the ball iff p2+14≤(1−p)2p^2+\tfrac14\le(1-p)^2, i.e. p≤38p\le\tfrac38.

Purification

Every mixed state is the shadow of a pure state on a larger system. A pure state ∣Ψ⟩AR\ket\Psi_{AR} purifies ρA\rho_A if Tr⁡R∣Ψ⟩ ⁣⟨Ψ∣=ρA\Tr_R\dyad\Psi = \rho_A; RR is called the reference or purifying system. Two standard constructions:

∣Ψ⟩=∑k=1rλk ∣k⟩A∣k⟩R,∣Ψρ⟩=(ρ⊗IR)∣Γ⟩,∣Γ⟩=∑i=1d∣i⟩A∣i⟩R\ket\Psi = \sum_{k=1}^{r}\sqrt{\lambda_k}\,\ket{k}_A\ket{k}_R,\qquad \ket{\Psi_\rho} = (\sqrt\rho\otimes\id_R)\ket\Gamma,\quad \ket\Gamma = \sum_{i=1}^d\ket{i}_A\ket{i}_R
The Schmidt form needs dim⁡R≥r=rank⁡ρ\dim R\ge r = \rank\rho. The canonical purification uses dim⁡R=d\dim R = d; since Tr⁡R∣Γ⟩ ⁣⟨Γ∣=I\Tr_R\dyad\Gamma = \id, it gives Tr⁡R∣Ψρ⟩ ⁣⟨Ψρ∣=ρ I ρ=ρ\Tr_R\dyad{\Psi_\rho} = \sqrt\rho\,\id\,\sqrt\rho = \rho.
DerivationProof
  1. ∣rk(a)⟩:=λk−1/2 (⟨k∣⊗I)∣Ψa⟩ ∈Ra\ket{r^{(a)}_k} := \lambda_k^{-1/2}\,\big(\bra{k}\otimes\id\big)\ket{\Psi_a}\ \in R_a

    Fix any eigenbasis {∣k⟩}\{\ket k\} of ρA\rho_A, with λk>0\lambda_k>0 for k≤rk\le r, and define for each purification a=1,2a = 1,2

  2. ⟨rk(a)∣rl(a)⟩=⟨Ψa∣(∣k⟩ ⁣⟨l∣⊗I)∣Ψa⟩λkλl=⟨l∣ρA∣k⟩λkλl=δkl\braket{r^{(a)}_k|r^{(a)}_l} = \frac{\bra{\Psi_a}\big(\ketbra{k}{l}\otimes\id\big)\ket{\Psi_a}}{\sqrt{\lambda_k\lambda_l}} = \frac{\bra{l}\rho_A\ket{k}}{\sqrt{\lambda_k\lambda_l}} = \delta_{kl}

    These vectors are orthonormal, degenerate eigenvalues included:

  3. ∣Ψa⟩=∑k=1rλk ∣k⟩∣rk(a)⟩\ket{\Psi_a} = \sum_{k=1}^r\sqrt{\lambda_k}\,\ket{k}\ket{r^{(a)}_k}

    Expand ∣Ψa⟩\ket{\Psi_a} in the basis {∣k⟩}\{\ket k\} of AA. Components with λk=0\lambda_k = 0 vanish, because ∥(⟨k∣⊗I)∣Ψa⟩∥2=⟨k∣ρA∣k⟩=0\norm{(\bra k\otimes\id)\ket{\Psi_a}}^2 = \bra k\rho_A\ket k = 0. This gives Schmidt forms with the same AA-side vectors:

  4. Define V∣rk(1)⟩=∣rk(2)⟩V\ket{r^{(1)}_k} = \ket{r^{(2)}_k} for k≤rk\le r and extend it isometrically to all of R1R_1, which is possible because dim⁡R1≤dim⁡R2\dim R_1\le\dim R_2. Then (I⊗V)∣Ψ1⟩=∣Ψ2⟩(\id\otimes V)\ket{\Psi_1} = \ket{\Psi_2}. □\square

Steering: the HJW theorem

Now the punchline. Suppose Bob holds the purifying system RR of Alice's ρA\rho_A. Which ensembles can Bob prepare for Alice, just by measuring RR and announcing nothing?

DerivationProof
  1. ∣mi⟩:=∑k=1ruik∗∣k⟩R,∑i∣mi⟩ ⁣⟨mi∣=∑k,l≤r(u†u)kl∣k⟩ ⁣⟨l∣=Πr\ket{m_i} := \sum_{k=1}^ru_{ik}^*\ket{k}_R,\qquad \sum_i\dyad{m_i} = \sum_{k,l\le r}\big(u^\dagger u\big)_{kl}\ketbra{k}{l} = \Pi_r

    Take ∣Ψ⟩=∑k=1r∣k~⟩A∣k⟩R\ket\Psi = \sum_{k=1}^r\ket{\tilde k}_A\ket{k}_R with ∣k~⟩=λk∣k⟩\ket{\tilde k} = \sqrt{\lambda_k}\ket k; any other purification differs by an isometry on RR, which Bob can absorb into his measurement. By unitary freedom, ∣ψ~i⟩=∑kuik∣k~⟩\ket{\tilde\psi_i} = \sum_ku_{ik}\ket{\tilde k} for an N×NN\times N unitary uu (padding with ∣k~⟩=0\ket{\tilde k} = 0 for k>rk>r). Let Bob measure

  2. (IA⊗⟨mi∣)∣Ψ⟩=∑kuik∣k~⟩=∣ψ~i⟩\big(\id_A\otimes\bra{m_i}\big)\ket\Psi = \sum_ku_{ik}\ket{\tilde k} = \ket{\tilde\psi_i}

    This is a rank-one POVM on the support Πr\Pi_r of ρR\rho_R (add I−Πr\id-\Pi_r as an outcome that never occurs). Outcome ii projects AA onto

  3. so it occurs with probability ⟨ψ~i∣ψ~i⟩=pi\braket{\tilde\psi_i|\tilde\psi_i} = p_i and leaves AA in ∣ψi⟩\ket{\psi_i}. If dim⁡R≥N\dim R\ge N, extend {∣k⟩R}\{\ket k_R\} to NN orthonormal vectors and sum kk up to NN: the ∣mi⟩\ket{m_i} become orthonormal, a projective measurement. □\square

So every ensemble decomposition of ρA\rho_A is physically realized by some measurement on the purifying system. Schrödinger, reacting to the EPR paper, called this steering and found it deeply troubling.

Check your understanding

Alice and Bob share ∣Φ+⟩=12(∣00⟩+∣11⟩)\ket{\Phi^+} = \tfrac{1}{\sqrt2}(\ket{00}+\ket{11}). Bob measures YY on his qubit and finds +1+1, i.e. the state ∣+i⟩\ket{+i}. What is Alice's conditional state?

  1. ∣+i⟩\ket{+i}
  2. ∣−i⟩\ket{-i}
  3. ∣+⟩\ket{+}
  4. I/2\id/2
Show the answer

B. ∣−i⟩\ket{-i}

For ∣Φ+⟩\ket{\Phi^+}, (I⊗⟨m∣)∣Φ+⟩=12∣mˉ⟩(\id\otimes\bra m)\ket{\Phi^+} = \tfrac{1}{\sqrt2}\ket{\bar m}, the complex conjugate of Bob's vector. Indeed ∣Φ+⟩=12(∣−i⟩∣+i⟩+∣+i⟩∣−i⟩)\ket{\Phi^+} = \tfrac{1}{\sqrt2}\big(\ket{-i}\ket{+i}+\ket{+i}\ket{-i}\big), so Bob's ∣+i⟩\ket{+i} steers Alice to ∣−i⟩\ket{-i}. For real vectors such as ∣0⟩\ket0 or ∣±⟩\ket\pm the conjugation is invisible, which is why ZZ and XX outcomes are perfectly correlated.

Majorization, entropy, and where ensembles matter

DerivationProof sketch and the entropy bound
  1. pi=⟨ψ~i∣ψ~i⟩=∑k∣uik∣2λkp_i = \braket{\tilde\psi_i|\tilde\psi_i} = \sum_k\abs{u_{ik}}^2\lambda_k

    (⇒) Unitary freedom relative to the eigen-ensemble gives ∣ψ~i⟩=∑kuik∣k~⟩\ket{\tilde\psi_i} = \sum_ku_{ik}\ket{\tilde k}, and the ∣k~⟩\ket{\tilde k} are orthogonal with squared norms λk\lambda_k:

  2. The matrix Bik=∣uik∣2B_{ik} = \abs{u_{ik}}^2 is unistochastic, hence doubly stochastic, so p⃗=Bλ⃗≺λ⃗\vec p = B\vec\lambda\prec\vec\lambda (Hardy–Littlewood–Pólya). (⇐) Conversely, by Horn's theorem every p⃗≺λ⃗\vec p\prec\vec\lambda can be written p⃗=Bλ⃗\vec p = B\vec\lambda with Bik=∣uik∣2B_{ik} = \abs{u_{ik}}^2 for a real orthogonal uu; then ∣ψ~i⟩=∑kuik∣k~⟩\ket{\tilde\psi_i} = \sum_ku_{ik}\ket{\tilde k} is the desired ensemble.

  3. H(p⃗) ≥ H(λ⃗)=S(ρ),with equality for the eigen-ensemble.H(\vec p)\ \ge\ H(\vec\lambda) = S(\rho),\qquad\text{with equality for the eigen-ensemble.}

    Shannon entropy is Schur-concave, so every pure-state ensemble pays at least the von Neumann entropy in classical randomness:

Ensembles are invisible to measurements on AA, yet they are not idle bookkeeping. Quantities defined by optimizing over ensembles, so-called convex roofs, are central to entanglement theory:

EF(ρAB)=min⁡{pi,∣ψi⟩} ∑ipi S(Tr⁡B∣ψi⟩ ⁣⟨ψi∣)E_F(\rho_{AB}) = \min_{\{p_i,\ket{\psi_i}\}}\ \sum_ip_i\,S\big(\Tr_B\dyad{\psi_i}\big)
Entanglement of formation (Bennett, DiVincenzo, Smolin & Wootters, 1996). The minimum runs over all pure-state ensembles of ρAB\rho_{AB}, exactly the set classified by the HJW theorem. Wootters (1998) found a closed form for two qubits.

Ensembles also give meaning to unravelings of open-system dynamics. The same Lindblad equation can be unraveled into quantum-jump or diffusive trajectories, depending on how the environment, which plays the role of the purifying system, is monitored. Each monitoring scheme picks out an ensemble, and all of them average to the same ρ(t)\rho(t).

Key takeaways

  • Ensembles {∣ψ~i⟩}\{\ket{\tilde\psi_i}\} and {∣ϕ~j⟩}\{\ket{\tilde\phi_j}\} give the same ρ\rho iff ∣ψ~i⟩=∑juij∣ϕ~j⟩\ket{\tilde\psi_i} = \sum_ju_{ij}\ket{\tilde\phi_j} for a unitary uu, after zero-padding.
  • Purifications exist with dim⁡R=rank⁡ρ\dim R = \rank\rho and are unique up to an isometry on RR.
  • HJW steering: whoever holds the purification can prepare any ensemble of ρA\rho_A by measuring RR alone, without changing ρA\rho_A, which is precisely why no signal is sent.
  • ∣ψ⟩\ket\psi can appear with weight at most 1/⟨ψ∣ρ−1∣ψ⟩1/\bra\psi\rho^{-1}\ket\psi; the possible weight vectors are exactly those majorized by the spectrum, so H(p⃗)≥S(ρ)H(\vec p)\ge S(\rho).
  • Optimizations over ensembles (convex roofs such as entanglement of formation) are where ensembles regain operational meaning, and they are hard to compute in general.

Check yourself

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

Question 1

Which ensemble has the same density matrix as {12:∣0⟩, 12:∣+⟩}\{\tfrac12:\ket0,\ \tfrac12:\ket+\}?

  1. {12:∣1⟩, 12:∣−⟩}\{\tfrac12:\ket1,\ \tfrac12:\ket-\}
  2. {12:∣0⟩, 12:∣−⟩}\{\tfrac12:\ket0,\ \tfrac12:\ket-\}
  3. {cos⁡2π8: cos⁡π8∣0⟩+sin⁡π8∣1⟩,  sin⁡2π8: −sin⁡π8∣0⟩+cos⁡π8∣1⟩}\{\cos^2\tfrac\pi8:\ \cos\tfrac\pi8\ket0+\sin\tfrac\pi8\ket1,\ \ \sin^2\tfrac\pi8:\ -\sin\tfrac\pi8\ket0+\cos\tfrac\pi8\ket1\}
  4. The pure state proportional to ∣0⟩+∣+⟩\ket0+\ket+
Show the answer

C. {cos⁡2π8: cos⁡π8∣0⟩+sin⁡π8∣1⟩,  sin⁡2π8: −sin⁡π8∣0⟩+cos⁡π8∣1⟩}\{\cos^2\tfrac\pi8:\ \cos\tfrac\pi8\ket0+\sin\tfrac\pi8\ket1,\ \ \sin^2\tfrac\pi8:\ -\sin\tfrac\pi8\ket0+\cos\tfrac\pi8\ket1\}

The mixture has Bloch vector 12(x^+z^)\tfrac12(\hat x+\hat z): length 1/21/\sqrt2, direction θ=π/4\theta = \pi/4. Its eigen-ensemble consists of the two antipodal states on that axis with weights 12(1±1/2)=cos⁡2π8, sin⁡2π8\tfrac12(1\pm1/\sqrt2) = \cos^2\tfrac\pi8,\ \sin^2\tfrac\pi8, which reproduces the Bloch vector (cos⁡2π8−sin⁡2π8) n^=12n^(\cos^2\tfrac\pi8-\sin^2\tfrac\pi8)\,\hat n = \tfrac{1}{\sqrt2}\hat n.

Question 2

What is the smallest dimension of a purifying system RR for a qutrit state with eigenvalues (12,12,0)(\tfrac12,\tfrac12,0)?

  1. 11
  2. 22
  3. 33
  4. 99
Show the answer

B. 22

A purification needs dim⁡R≥rank⁡ρ=2\dim R\ge\rank\rho = 2, and the Schmidt form 12(∣k1⟩∣0⟩+∣k2⟩∣1⟩)\tfrac{1}{\sqrt2}\big(\ket{k_1}\ket0+\ket{k_2}\ket1\big) achieves it with a single qubit.

Question 3

Why does the HJW theorem imply that entanglement cannot be used to signal?

  1. Bob's measurement leaves Alice's conditional states unchanged
  2. Purifications are unique, so Bob has only one possible measurement
  3. Steering only works if Bob also sends Alice a quantum system
  4. Every ensemble Bob can steer Alice into averages to the same ρA\rho_A, which alone fixes Alice's statistics
Show the answer

D. Every ensemble Bob can steer Alice into averages to the same ρA\rho_A, which alone fixes Alice's statistics

Bob's choice selects an ensemble {pi,∣ψi⟩}\{p_i,\ket{\psi_i}\}, so it does change Alice's conditional states. But Alice does not learn ii without a classical message, and ∑ipi∣ψi⟩ ⁣⟨ψi∣=ρA\sum_ip_i\dyad{\psi_i} = \rho_A for every choice Bob can make.

Question 4

For ρ=I/d\rho = \id/d, what is the largest weight with which a given pure state can appear in an ensemble for ρ\rho?

  1. 1/d1/d
  2. 1/d21/d^2
  3. 1/21/2
  4. 11
Show the answer

A. 1/d1/d

ρ−1=d I\rho^{-1} = d\,\id, so pmax⁡=1/⟨ψ∣ρ−1∣ψ⟩=1/dp_{\max} = 1/\bra\psi\rho^{-1}\ket\psi = 1/d. It is attained by any orthonormal basis containing ∣ψ⟩\ket\psi, with uniform weights.

Further reading

  • L. P. Hughston, R. Jozsa & W. K. Wootters, "A complete classification of quantum ensembles having a given density matrix", Physics Letters A 183, 14 (1993).
  • E. Schrödinger, "Probability relations between separated systems", Proceedings of the Cambridge Philosophical Society 32, 446 (1936).
  • M. A. Nielsen & I. L. Chuang, Quantum Computation and Quantum Information, Theorem 2.6 and §2.5 (Cambridge, 2010).
  • M. A. Nielsen, "Probability distributions consistent with a mixed state", Physical Review A (2000).
  • H. M. Wiseman, S. J. Jones & A. C. Doherty, "Steering, entanglement, nonlocality, and the Einstein–Podolsky–Rosen paradox", Physical Review Letters 98, 140402 (2007).