PaperlayerView exact revision on arXiv

Source-linked AI summary

Entanglement of purification for Werner states: canonical purification and nonadditivity

Zahra Baghali Khanian, Amir-Reza Negari

arXiv:2610.03630v1quant-phcs.ITmath.IT

TL;DR

The additivity question asks whether optimizing a purification of one copy gives the preparation cost per copy, or whether joint purification saves entanglement. For two-qubit Werner states, this paper proves canonical optimality over a range and shows that collective preparation costs less on an overlapping interval, establishing nonadditivity even for separable states.

  • Problem

    Entanglement of purification is difficult to evaluate because a useful purification generally gives only an upper bound, leaving the relation between its single-copy and regularized values unresolved.

    Source: §I · p. 1

  • Method

    A symmetric-subspace projection constrains arbitrary purifications via symmetric extendibility, reducing the entropy comparison to four probabilities to prove canonical optimality.

    Source: §I · p. 2, §IV · p. 4

  • Results

    For 1/64 ≤ f ≤ 1/4, canonical purification gives the exact entanglement of purification, while the regularized quantity is strictly lower for 0 < f < f∗, establishing nonadditivity on 1/64 ≤ f < f∗.

    Source: §I · p. 2, §V · p. 7

  • Takeaways & Limitations

    The collective advantage occurs entirely among separable Werner states, so entanglement in the target states is not required for the reduction in preparation cost.

    Source: §I · p. 2, §V · p. 7

    The corresponding von Neumann additivity question for classical two-qubit states remains open.

    Source: §V · p. 7

Abstract

from arXiv · show

We prove that the von Neumann entanglement of purification is nonadditive. For two-qubit Werner states $W(f)$ with singlet weight $1/64\le f\le1/4$, we show that canonical purification is globally optimal: its entanglement, also known as reflected entropy, gives the exact entanglement of purification. In contrast, the regularized entanglement of purification lies strictly below the reflected entropy for $0<f<f_\ast$, where $f_\ast\approx0.017$. Together, these results establish nonadditivity throughout $1/64\le f<f_\ast$. Our canonical-optimality proof generalizes the proof in [1], which established R\'enyi nonadditivity for $0\le\alpha<1$ using classical two-qubit states. We thus resolve the additivity question at $\alpha=1$. Remarkably, nonadditivity is again witnessed by separable states and no entanglement in the underlying states is required.

I. INTRODUCTION

For two-qubit Werner states, the canonical purification gives the exact entanglement of purification on 1/64 ≤ f ≤ 1/4, while collective preparation costs strictly less for 0 < f < f∗. Together, these results establish nonadditivity on 1/64 ≤ f < f∗, entirely among separable states.

I. INTRODUCTION

  • The additivity question matters because regularized entanglement of purification gives operational rates in visible compression and quantum rate-distortion, while the single-copy quantity relates to dense-coding advantage.A useful purification generally supplies only an upper bound, and symmetry of the target state alone does not establish that a symmetric purification is optimal.Source: §I · p. 1
  • For 1/64 ≤ f ≤ 1/4, the canonical purification is optimal, and its entanglement equals the entanglement of purification; for 0 < f < f∗, the regularized quantity is strictly lower.The threshold is f∗≈0.017086.Source: §I · p. 2
  • A symmetric-subspace projection constrains arbitrary purifications through symmetric extendibility, reducing the entropy comparison to four probabilities and proving the single-copy lower bound.The asymptotic improvement then follows from the ensemble bound of Chen and Winter [26].Source: §I · p. 2
  • 1/64 ≤ f < f∗ is the interval where canonical optimality and a strictly lower regularized cost together establish nonadditivity for two-qubit Werner states.The figure marks the overlap of the intervals for the two bounds; its endpoints are included at 1/64 and excluded at f∗.Source: §I · p. 2, Figure 1 · p. 2
  • The result extends Rényi nonadditivity for classical two-qubit states [1] to von Neumann order using separable Werner states, while the corresponding classical-state question remains open.The Werner states are not diagonal in a local product basis.Source: §I · p. 2

II. PRELIMINARIES

Entropies are measured in bits, and logarithms are base two unless written as ln.

II. PRELIMINARIES

  • Entropies are measured in bits, logarithms are base two unless written as ln, and 0 log2 0 is defined as 0.Source: §II · p. 2

A. Purification and regularization

Entanglement of purification minimizes the entanglement across parties and their respective ancillas over all purifications of a state. Its regularization is well-defined because the quantity is subadditive, and strict inequality from additivity occurs exactly when some finite-copy value is below the corresponding single-copy multiple.

A. Purification and regularization

  • Entanglement of purification minimizes the entanglement across AA′ : BB′ over purifications of ρAB, where A′ and B′ are local ancillas.For a pure purification, the cost is the Shannon entropy of its Schmidt probabilities.Source: A. Purification and regularization · p. 2
  • Finite ancilla dimensions suffice to attain the infimum, and tensoring purifications makes entanglement of purification subadditive, so its regularization exists.For two-qubit states, ancilla dimensions 4 and 16 suffice, or the ancillas may be interchanged.Source: A. Purification and regularization · p. 3
  • Strict inequality from additivity holds exactly when EP(ρ⊗n) < nEP(ρ) for some finite n ≥ 2.Source: A. Purification and regularization · p. 3
  • Vectorizing √ρ in a product basis gives the canonical purification, whose entanglement across AA′ : BB′ is reflected entropy and upper-bounds entanglement of purification.Canonical optimality is equality in this bound.Source: A. Purification and regularization · p. 3

B. Werner states and the canonical spectrum

The two-qubit Werner family is parametrized by singlet probability f; the section considers 0 ≤ f ≤ 1/4 and describes the canonical purification's Schmidt amplitudes.

B. Werner states and the canonical spectrum

III. A REGULARIZED COST BELOW THE CANONICAL VALUE

For two-qubit Werner states, an ensemble construction gives an achievable asymptotic preparation rate below the canonical purification value when 0 < f < f∗. Combined with canonical optimality, this establishes nonadditivity for 1/64 ≤ f < f∗.

III. A REGULARIZED COST BELOW THE CANONICAL VALUE

  • The ensemble bound uses W(0) as one component and raises the singlet probability of the other component to preserve the target W(f).As the ensemble weight p increases, the second component moves toward W(1/4), whose canonical cost decreases.Source: §III · p. 3
  • Preparing W(0) with one ebit and the other component canonically gives an achievable rate comprising average preparation cost plus the ensemble's Holevo information.The ensemble choice balances these two contributions.Source: §III · pp. 3–4
  • For 0 < f < f∗≈0.017086, the decrease in average preparation cost exceeds the added Holevo information, yielding a lower rate for sufficiently small p > 0.At p = 0 the rate equals the canonical value, so a negative right derivative establishes the strict improvement.Source: §III · p. 4
  • For 1/64 ≤ f < f∗, the regularized entanglement of purification is strictly below the single-copy entanglement of purification, proving nonadditivity.The strict bound holds for 0 < f < f∗, while canonical optimality holds from f = 1/64.Source: §III · p. 4

IV. OPTIMALITY OF THE CANONICAL PURIFICATION

The canonical purification attains the single-copy entanglement of purification for two-qubit Werner states throughout 1/64 ≤ f ≤ 1/4. The proof goes beyond a largest-Schmidt-probability bound by controlling the remaining probabilities and minimizing the resulting entropy.

IV. OPTIMALITY OF THE CANONICAL PURIFICATION

  • For 1/64 ≤ f ≤ 1/4, the canonical purification is optimal over purifications with unrestricted finite ancillary dimensions.Its four-entry spectrum prevents a largest-eigenvalue bound alone from establishing universal majorization, so the proof also controls the tail.Source: §IV · p. 4
  • The argument extends the largest-eigenvalue approach of [1], while the Werner-state tail inequality uses the two-qubit symmetric-extension criterion of [29].The Werner canonical spectrum has four nonzero entries for f < 1/4, unlike the at-most-two-entry classical spectrum in [1].Source: §IV · pp. 4–5
  • The proof constrains the leading Schmidt probability jointly with the tail, then minimizes the entropy of the resulting four-entry probability vector.When p1 > 1/2, the entropy minimum shows that concentrating the tail cannot save more than reducing p1 increases the binary contribution; for p1 ≤ 1/2, entropy is at least one bit.Source: §IV · p. 4
  • For an arbitrary purification, the Schmidt probabilities across AA′ : BB′ are the eigenvalues of its reduced state, and its entanglement is H(p).The ancillary dimensions are unrestricted apart from being finite.Source: §IV · pp. 4–5

A. An overlap bound for the spectral tail

The spectral argument projects an arbitrary purification onto the symmetric subspace, uses symmetric extendibility to restrict its spectrum, and transfers that restriction back through an overlap bound. The resulting joint constraint controls the leading Schmidt probability and the tail.

A. An overlap bound for the spectral tail

  • For two-qubit states with maximally mixed marginals, the lemma relates the leading Schmidt probability to the remaining tail for every finite-dimensional purification.Here f is the probability of the physical singlet outcome.Source: A. An overlap bound for the spectral… · p. 5
  • Exchange symmetry of the physical qubits gives the projected marginal a symmetric extension, to which the two-qubit extension criterion [29] applies.The projected state retains maximally mixed physical-qubit marginals.Source: A. An overlap bound for the spectral… · p. 5
  • An overlap estimate transfers the projected state's spectral restriction to the original purification by comparing their ordered Schmidt coefficients and grouping the remaining coefficients.The grouped vectors have unit norm, and Cauchy–Schwarz bounds their tail contribution.Source: A. An overlap bound for the spectral… · p. 5
  • The overlap argument establishes ℓ ≤ √f, yielding the stated bound on the leading probability and tail.The derivation uses the projection inequality and Cauchy–Schwarz.Source: A. An overlap bound for the spectral… · pp. 5–6

B. The entropy minimum

The entropy-minimization step reduces the purification bound to a four-probability vector. For the Werner-state range, the minimizing spectrum has equal tail entries, matching the canonical spectrum.

B. The entropy minimum

  • For Werner states the four-probability constraint uses c = 2√f, and c ≥ 1/4 supplies the lower endpoint f ≥ 1/64.Grouping the tail lets this finite-vector minimum bound purifications with arbitrary finite ancillary dimensions.Source: B. The entropy minimum · p. 6
  • The displayed spectrum attains the entropy minimum, and for c < 1 it is the unique minimizing spectrum with the specified leading entry.Its entropy is strictly less than one bit.Source: B. The entropy minimum · p. 6
  • A feasible trial vector (5/6, 1/18, 1/18, 1/18) has entropy below one bit, so a minimizing vector must have q0 > 1/2.For c < 1, the square-root constraint must be active at the minimum.Source: B. The entropy minimum · p. 6
  • Every minimizing tail entry is positive; otherwise adding a small tail amplitude while preserving the constraints would decrease entropy.The displayed constraints encode normalization and the relation between the leading amplitude and the tail.Source: B. The entropy minimum · p. 6
  • For c ≥ 1/4, stationarity rules out unequal tail amplitudes, so all three tails coincide in the minimizing spectrum.The proof derives a contradiction with c ≥ 1/4 if the tails are unequal.Source: B. The entropy minimum · p. 6

C. The single-copy optimum

For 1/64 ≤ f ≤ 1/4, the canonical purification attains the exact single-copy entanglement of purification, even when arbitrary finite-dimensional purifying systems are allowed. The proof reduces every competitor to an entropy comparison on four probabilities.

C. The single-copy optimum

  • Grouping Schmidt probabilities from the fourth onward cannot increase entropy, reducing the lower-bound comparison to four probabilities regardless of ancillary dimensions.The joint tail constraint and the four-entry entropy minimum provide the lower bound.Source: C. The single-copy optimum · pp. 6–7
  • When the leading Schmidt probability is at most 1/2, the purification costs at least one bit, exceeding C(f).For a larger leading probability, the proof groups the remaining tail and applies the four-entry entropy bound.Source: C. The single-copy optimum · p. 7
  • For every 1/64 ≤ f ≤ 1/4, every purification costs at least C(f), and the canonical purification attains this value.At f = 1/4 the canonical purification is a product vector; below that endpoint, the proof sets c = 2√f.Source: C. The single-copy optimum · pp. 6–7

V. DISCUSSION

This section discusses the nonadditivity of entanglement of purification for two-qubit Werner states and completes the proof of an elementary stationary-ratio estimate.

V. DISCUSSION

  • For 1/64 ≤ f ≤ 1/4, canonical purification is optimal over all purifications, so its entanglement equals both entanglement of purification and reflected entropy.The proof controls the largest Schmidt probability jointly with the tail and reduces the entropy comparison to four probabilities; the minimizing spectrum has three equal smaller probabilities.Source: §V · p. 7, §2 · p. 11
  • For 1/64 ≤ f < f∗, canonical optimality and a strictly lower collective-preparation cost establish nonadditivity.The collective-cost bound holds for 0 < f < f∗, using the same auxiliary state W(f∗) throughout that range.Source: §V · p. 7

Appendix B: Regularization bound

  • For 0 < f < f∗, the ensemble inequality of [26] bounds regularized entanglement of purification strictly below the canonical value.The derivative test identifies f∗ as the threshold for this preparation bound; the same auxiliary state W(f∗) works throughout the interval.Source: §V · p. 7, §1 · p. 9, Figure 2 · p. 10

V. DISCUSSION

  • The nonadditivity occurs among separable Werner states whose correlations are not diagonal in a local product basis, complementing the classical-state Rényi result in [1].For f < 1/4, their Pauli correlation matrix is a nonzero multiple of I3, unlike the rank-at-most-one correlation matrix of classical two-qubit states with maximally mixed marginals.Source: §V · p. 7
  • The proof of nonadditivity in Ref. [27] is invalid: its Lemma 2 fails at f = 1/200 because of a mismatch of weights in a Cauchy–Schwarz application.Appendix D supplies an explicit counterexample to the lemma, invalidating the stated single-copy lower bound and claimed nonadditivity proof.Source: §V · p. 7
  • The required number of copies, exact regularized value, and sharp nonadditivity boundary remain open.The proven interval endpoints arise from sufficient entropy-minimization and ensemble-comparison bounds; neither is claimed sharp, and canonical optimality may extend below 1/64.Source: §V · p. 7

1. A separable decomposition

  • At f = 1/64, W(f) is explicitly a mixture of twelve pure product states, confirming separability at the interval endpoint.Six equal-sign products have weight 31/192 each, and six opposite-sign products have weight 1/192 each.Source: §1 · p. 8

3. The symmetric endpoint

  • At the symmetric endpoint, EP(W(0)) = 1, with the reverse bound obtained by assigning the full purifying system to Bob.The lower bound follows from exchange symmetry of every purification and weak monotonicity; the upper bound is EP ≤ S(A) = 1.Source: §3 · p. 8

5. An elementary stationary-ratio estimate

  • The lemma follows by bounding F and G with tangent lines and checking coefficient signs using elementary bounds on ln 2.The proof uses 2/3 < ln 2 < 25/36 to bound k between 13/10 and 7/5.Source: §5 · p. 14

6. Proof of Lemma IV.2

The entropy minimization has a unique minimizing spectrum for each specified leading entry in the stated range. Its proof makes the square-root constraint active and rules out zero or unequal tail amplitudes.

6. Proof of Lemma IV.2

  • For any minimizer, q0 > 1/2 and therefore exceeds q1, q2, and q3; also q0 < 1 because the pure vector violates the constraint when c < 1.Source: §6 · p. 15
  • Every minimizing tail amplitude is positive; unequal positive tails are impossible, so all three tail amplitudes coincide.Feasible variations rule out a vanishing tail and then contradict minimality when positive tails differ.Source: §6 · p. 15
  • For c < 1, the feasible vector is the unique minimizing spectrum with the specified leading entry.The result follows by solving a² + 3b² = 1 and a − 3b = c with nonnegative amplitudes.Source: §6 · p. 15

Appendix D: A counterexample to Lemma 2 of [27]

Appendix D gives a counterexample to Lemma 2 of [27] at f = 1/200 and identifies a Cauchy–Schwarz weight mismatch in its proof. A counterexample also occurs at f = 1/64, the paper’s main comparison point.

Appendix D: A counterexample to Lemma 2 of [27]

  • The disputed bound concerns tail terms with at least one Schmidt index exceeding two in an ordered Schmidt decomposition of a purification of W(f).The claimed bound is evaluated at f = 1/200.Source: Appendix D · p. 16
  • At f = 1/200, a two-dimensional-ancilla purification is a counterexample to Lemma 2 within its stated domain and the padded four-dimensional domain.The counterexample also occurs at f = 1/64, the paper’s main comparison point.Source: Appendix D · p. 16
  • Cauchy–Schwarz fails because the squared norms weight the sum by p_i p_α rather than √(p_i p_α); a valid generic bound is |Rcross| ≤ 2√sε.The incorrect step is in Appendix A.3, immediately after the displayed formula for the cross terms.Source: Appendix D · pp. 16–17

Related papers

Loading 2610.03630v1…