{"id":"b379701e-c0a7-411e-84dc-6d29b6db4eba","arxiv_id":"2502.01749","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For all alpha in [1/2, infinity], the regularized alpha-Renyi divergence from a bipartite state to the set of extensions of a fixed marginal equals the marginal divergence, and the measured version lies between the measured and standard marginal divergences.","lead":"The authors extend Uhlmann's theorem from fidelity to the entire family of Rényi relative entropies, proving that regularization restores a subsystem-to-extension equality and that a measured version gives two-sided bounds. This provides new tools for quantum cryptography and entropy accumulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (41) map E is not trace-preserving as defined, leaving the DPI step in Theorem 3.1's proof unjustified.","rationale":"The paper's main result, the regularized Uhlmann theorem (Theorem 3.1 combined with Theorem 4.1), rests on the proof for alpha in [1/2,1). The DPI step in Eq. (48) is essential, and it requires E to be a quantum channel. The algebra shows that E is only trace-preserving on the support of bar_rho_A, not on the full space. This is a concrete mathematical gap that a careful referee would flag. It is likely fixable by extending E to a TPCP map that leaves the two relevant states unchanged, but the paper does not provide that argument. The reader's identified weakest assumption, the postselection de Finetti reduction (Eq. 37), is imported from a known theorem and is used only for the measured Uhlmann theorem (Theorem 3.2), not for the regularized theorem. The reader's concern about Sion's minimax in Lemma 3.3 is also not a genuine gap: in both alpha ranges the compact convex set is C, and the required quasiconvexity/quasiconcavity conditions are satisfied (for alpha<1, the function is convex in tau and linear in sigma; for alpha>1, it is concave in tau and linear in sigma). Thus the load-bearing weakness is the trace-preservation/DPI step in Theorem 3.1. The proposed concrete test would settle whether the gap is merely cosmetic or actually breaks the proof. Since the paper likely survives with a minor revision, the overall CONDITIONAL verdict remains appropriate.","tokens_in":24002,"tokens_out":61985,"duration_ms":520964,"concrete_test":"Choose a simple example where bar_rho_A is rank-deficient, e.g., A and B qubits, rho_AB = |00⟩⟨00|, sigma_A = |0⟩⟨0|, n=1. Compute E(I_A) using Eq. (41) and check whether tr[E(I_A)] = dim(A) = 2; trace-preservation would require this, but a direct calculation gives tr[E(I_A)] = tr[Π_supp(bar_rho_A)] = 1, confirming E is not TP. Then verify whether the DPI inequality in Eq. (48) still holds for this example, and whether the extension \\tilde E(X) = E(X) + tr[X(I_A − Π_supp(bar_rho_A))] |1⟩⟨1|_A ⊗ |0⟩⟨0|_B is a TPCP map that agrees with E on bar_rho_A and sigma_A^{⊗n}. If the DPI inequality fails for any such example, Theorem 3.1 is invalid; if the extension always works, the gap is only a missing justification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 3.1 for alpha in [1/2,1) relies on the DPI step in Eq. (48): D_alpha(bar_rho || bar_sigma) = D_alpha(E(bar_rho_A) || E(sigma_A^{⊗n})) ≤ D_alpha(bar_rho_A || sigma_A^{⊗n}), where E is defined in Eq. (41) as E(X) = bar_rho^{1/2} bar_rho_A^{-1/2} X bar_rho_A^{-1/2} bar_rho^{1/2}. For E to be trace-preserving, one needs tr_B[V†V] = I_A with V = bar_rho^{1/2}(bar_rho_A^{-1/2} ⊗ I_B). A direct computation gives tr_B[V†V] = bar_rho_A^{-1/2} tr_B[bar_rho] bar_rho_A^{-1/2} = bar_rho_A^{-1/2} bar_rho_A bar_rho_A^{-1/2} = Π_supp(bar_rho_A), the projection onto the support of bar_rho_A, not the identity unless bar_rho_A has full rank. Thus E is not a quantum channel in general. The inequality in Eq. (48) is therefore not justified as written; it would be valid if E can be extended to a TPCP map that agrees with E on bar_rho_A and sigma_A^{⊗n}. Such an extension exists when supp(sigma_A^{⊗n}) ⊆ supp(bar_rho_A), which is forced by the proof that bar_sigma ∈ C (Eqs. 44-47), so the gap is likely repairable. Nevertheless, the paper's claim that E is trace-preserving is false, and the central proof is not fully rigorous as stated. This is internal to Theorem 3.1, unlike the de Finetti reduction (Eq. 37), which is only used for the measured Uhlmann theorem.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper generalizes Uhlmann's theorem for fidelity to sandwiched Rényi relative entropies D_α for α∈[1/2,∞]. The two main results are a regularized Uhlmann theorem, D_α(ρ_A‖σ_A)=D^∞_α(ρ_AB‖C^{σ_A}_{AB}) (Theorem 3.1, extended to α=1 in Theorem 4.1), and a single-letter measured Uhlmann theorem, D_{α,M}(ρ_A‖σ_A)≤D_{α,M}(ρ_AB‖C^{σ_A}_{AB})≤D_α(ρ_A‖σ_A) (Theorem 3.2, extended to α=1 in Theorem 4.2). The proofs use asymptotic spectral pinching, a postselection de Finetti reduction imported from [11], variational formulas for measured Rényi divergences, and Sion's minimax theorem. The paper also gives explicit constructions of extensions for the relative entropy case and discusses why a non-regularized Uhlmann equality cannot hold for intermediate α.","tokens_in":24376,"tokens_out":28566,"duration_ms":224213,"significance":"If correct, these results settle a natural question: Uhlmann-type equalities hold for all α-Rényi relative entropies in the regularized sense, and the measured version provides single-letter bounds. The paper cleanly separates the new regime α∈[1/2,1) and the measured results from previously known α>1 results, and it gives explicit optimizers for α=1. The proof is largely self-contained and clearly cites the one non-elementary input, the postselection de Finetti reduction [11, Cor. 3.2], which underpins the measured Uhlmann theorem via Lemma 3.2. The main technical defect I found is localized to the trace-preservation claim for the map E in Theorem 3.1; this is fixable without changing the structure of the proof.","major_comments":[{"comment":"The map E defined in Eq. (41) as E(X)=ρ̄^{1/2}ρ̄_A^{-1/2}Xρ̄_A^{-1/2}ρ̄^{1/2} is claimed to be trace-preserving, but this is false in general. With V=ρ̄^{1/2}(ρ̄_A^{-1/2}⊗I_B), a direct computation gives tr_B[V†V]=ρ̄_A^{-1/2}tr_B[ρ̄]ρ̄_A^{-1/2}=Π_{supp(ρ̄_A)}, the projection onto the support of ρ̄_A, not the identity unless ρ̄_A has full rank. Consequently the DPI step in Eq. (48) is not justified as written. The gap is repairable: the proof of Eqs. (44)–(47) forces supp(σ_A^{⊗n})⊆supp(ρ̄_A), so one can extend E to a genuine TPCP map E'(X)=VXV†+KXK† with K=(I_A−Π_{supp(ρ̄_A)})⊗|φ⟩_B for a fixed normalized |φ⟩_B. This map agrees with E on ρ̄_A and σ_A^{⊗n}, and the DPI step becomes valid. The manuscript should be revised to either define this extension or otherwise justify the DPI step.","section":"Theorem 3.1, Eqs. (41)–(49)"}],"minor_comments":[{"comment":"The proof applies Sion's minimax theorem with the set of positive definite matrices as one of the convex sets. This set is not compact, but the concern that this invalidates the argument does not land: Sion's theorem only requires compactness of one of the convex sets, and C^{σ_A}_{AB} is compact. I would nonetheless ask the authors to spell out the required semicontinuity and quasi-concavity/quasi-convexity conditions explicitly, especially for the α>1 case where the objective is concave in τ rather than convex.","section":"Lemma 3.3, Eqs. (95)–(100)"},{"comment":"The notation ρ̄_A^{-1/2} is used for the inverse on the support of ρ̄_A; this should be defined explicitly as the pseudoinverse or with a phrase such as 'restricted to the support of ρ̄_A' to avoid ambiguity.","section":"Section 3.1, footnote 12"},{"comment":"The step labeled (⋄) is a scaling argument rather than a pointwise AM-GM comparison at the same τ; the current wording is terse. A short explanation that for every τ one can choose λ so that the sum at λτ equals the geometric term at τ would make the direction of the inequality clear.","section":"Lemma 3.3, Eq. (101)"},{"comment":"The phrase 'a relaxed asymptotic version of the polar assumption introduced in [9, Lemma 8]' is informal and not needed for the proof. Since Lemma 3.2 is a workhorse for the measured Uhlmann theorem, consider removing this aside or explaining the connection in one sentence.","section":"Lemma 3.2, proof"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid contribution and the central claims are likely correct. The only load-bearing defect is the false trace-preservation claim for the map E in Theorem 3.1, and the repair is straightforward. I would be willing to accept after the authors revise the DPI step and address the clarity points listed in the minor comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this for the measured Uhlmann theorem and the alpha in [1/2,1) regularized identity. The alpha>1 regularized equality was already known, and the paper says so. The genuinely new part is the lower range and the measured inequalities for all alpha, plus an explicit optimizer for relative entropy. The paper is honest and well structured.\n\nThe proof strategy is sound in outline: pinch, de Finetti, variational formulas, and the classical case as a guide. The distinction between known and new is clean, and the self-citations are to published theorems. No circularity.\n\nTwo soft spots. First, Lemma 3.3 invokes Sion's minimax with the set of positive definite matrices, which is not compact. This is a minor fix—restrict to a compact sublevel set or use a standard relaxation. Second, and more serious, the stress-test concern about Eq. (41) is correct. The map E satisfies tr_B[V^\\dagger V] = Pi_supp(bar_rho_A), not I, so it is not a quantum channel. The DPI step in Eq. (48) is unjustified as written. The proof needs supp(sigma_A^{\\otimes n}) \\subseteq supp(bar_rho_A); the pinching construction does not force this, and in the typical rank-deficient case bar_sigma is not an extension of sigma_A at all. This gap sits in the central proof of Theorem 3.1 for alpha in [1/2,1). It looks repairable—a full-rank perturbation of rho_AB plus a continuity argument should work, and the theorem is probably true—but the paper as submitted is not fully rigorous there.\n\nWho gets value: quantum information theorists working on entropy accumulation, cryptography, or Renyi divergences generally. The measured Uhlmann bounds are likely to be widely used. This deserves a serious referee. Send it to review, but the referee should require a corrected treatment of E and a proper minimax argument. With those fixed, the paper will be a solid contribution.","headline":"A genuinely new result with an honest proof sketch, but the central proof has a repairable gap around the map E in Theorem 3.1 that the referee will need to see fixed.","tokens_in":24933,"tokens_out":10139,"would_cite":true,"duration_ms":84824,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P45","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"Uhlmann's equality now holds for every α-Rényi relative entropy","keywords":["Uhlmann's theorem","sandwiched Rényi relative entropy","quantum fidelity","max-relative entropy","regularization","measured relative entropy","de Finetti reduction","spectral pinching"],"falsifier":"Compute, for a specific pair of finite-dimensional states such as ρ_AB = |φ⟩⟨φ|_AB with |φ⟩ = √(1/4)|00⟩ + √(3/4)|11⟩ and σ_A = 1/2 I_A, the sequence (1/n) D_{3/2}($ρ_AB^{{⊗n}}$ ‖ $C^{{σ_A^{⊗n}}$}_{A^n B^n}) for n = 1, 2, 3, ... and check whether it converges to D_{3/2}(ρ_A‖σ_A). A limit below D_{3/2}(ρ_A‖σ_A) would disprove the regularized Uhlmann theorem for α = 3/2.","tokens_in":23796,"feed_emoji":"⚛️","tokens_out":6659,"duration_ms":52840,"temperature":0.7,"pith_summary":"Uhlmann's theorem says that for any two quantum states ρ_AB and σ_A, the fidelity between ρ_AB and a suitable extension of σ_A equals the fidelity of the reduced states. This paper extends that equality to the whole family of sandwiched α-Rényi relative entropies, which includes fidelity, relative entropy, and max-relative entropy as special cases. A plain single-copy equality is impossible for intermediate α, so the authors prove two substitutes: a regularized equality that holds asymptotically over many copies, and a measured version in which both sides are preceded by an optimized measurement. The results matter because they let a divergence defined on a small system be evaluated through extensions on a larger system, which is what makes Uhlmann's theorem useful in quantum information.","feed_headline":"Uhlmann's equality now holds for every α-Rényi relative entropy","feed_subtitle":"With a limit over n copies, every Rényi divergence gains the extension property Uhlmann proved for fidelity.","key_machinery":"The proof imports Uhlmann's original identity for fidelity and the known α > 1 case from [21], and then handles α ∈ [1/2, 1) by a two-step lifting of the classical construction. First, asymptotic spectral pinching with respect to $σ_A^{{⊗n}}$ and its pinched reduction ρ'_{A^n} makes the relevant operators commute, producing a feasible extension σ̄_{A^n B^n} of $σ_A^{{⊗n}}$ whose Rényi divergence from $ρ_AB^{{⊗n}}$ is close to D_α(ρ_A‖σ_A). Second, a refined postselection de Finetti reduction with fixed marginal σ_A (Equation (37), imported from [11]) bounds the support function of the extension set: (1/n) log h_{$C^{{σ_A}}$_{AB,n}}($τ^{{⊗n}}$) = log h_{$C^{{σ_A}}$_{AB}}(τ) + o(n)/n, which is the asymptotic multiplicativity that yields the measured inequality. Variational formulas for the measured Rényi divergence and Sion's minimax theorem then convert the support-function bound into single-letter statements.","core_discovery":"The central claim is that for any finite-dimensional ρ_AB, σ_A and any α ∈ [1/2, ∞], D_α(ρ_A‖σ_A) equals the regularized minimum over extensions of σ_A of D_α(ρ_AB‖·), namely D^∞_α(ρ_AB‖$C^{{σ_A}}$_AB) = lim_{n→∞} (1/n) D_α($ρ_AB^{{⊗n}}$‖$C^{{σ_A^{⊗n}}$}_{A^n B^n}). This is Theorem 3.1 for α ≠ 1 and Theorem 4.1 for α = 1, where the divergences are the sandwiched Rényi relative entropies (with the relative entropy and max-relative entropy at the endpoints). Because the unregularized equality fails for α in (1/2, ∞), the paper also proves the measured Uhlmann theorem: D_{α,M}(ρ_A‖σ_A) ≤ D_{α,M}(ρ_AB‖$C^{{σ_A}}$_AB) ≤ D_α(ρ_A‖σ_A), where the measured divergence maximizes over POVMs. Together these results establish that Uhlmann-type identities are not specific to fidelity but hold, in the appropriate sense, for the entire Rényi family.","pith_inferences":["The explicit optimizer in Remark 4.1 is formally a Petz recovery map; if that link is exploitable, the regularized Uhlmann equality may yield new recovery-error bounds for approximate quantum Markov chains.","The de Finetti bound's polynomial overhead depends on d_A^2 d_B^4; tightening this dependence would sharpen the single-letter measured inequality, and the paper's methods suggest testing whether the exponent can be improved.","One could test whether the measured Uhlmann inequality is tight under optimized POVMs for states with non-commuting ρ_A and σ_A, which would give a single-letter converse for the regularization.","The impossibility result in Appendix A suggests that any sufficient divergence obeying a data-processing inequality will need some regularization or measurement before an Uhlmann-type identity can hold; this may guide the search for other divergence families."],"forward_implications":["For every α ∈ [1/2, ∞], the divergence D_α(ρ_A‖σ_A) can be expressed as a regularized minimization over extensions, giving a variational handle on Rényi divergences that does not require diagonalizing ρ_A.","For the relative entropy (α = 1), the regularized optimizer has an explicit form, making the equality constructive rather than an abstract limit.","The measured Uhlmann inequalities provide single-letter upper and lower bounds for the regularized quantity in terms of convex optimization problems, since D_{α,M}(ρ_AB‖C^{σ_A}_AB) and D_α(ρ_AB‖C^{σ_A}_AB) are convex.","For α ∈ (1/2, ∞), the single-copy quantity D_α(ρ_AB‖C^{σ_A}_AB) is not additive under tensor products, while at the endpoints α = 1/2 and α = ∞ it is exactly additive.","When [ρ_A, σ_A] = 0, the measured and regularized quantities coincide, recovering the classical case where regularization is unnecessary."],"supporting_citations":[{"why":"The original Uhlmann theorem for fidelity, the statement being generalized.","marker":"[40]"},{"why":"Proved the regularized Uhlmann equality for α > 1 and supplied pinching lemmas and single-copy counterexamples.","marker":"[21]"},{"why":"Supplied the postselection de Finetti reduction with fixed marginal used in Lemma 3.2.","marker":"[11]"},{"why":"Provided variational formulas for measured Rényi relative entropies used throughout the proofs.","marker":"[2]"},{"why":"Gave data-processing inequalities and additivity properties of Rényi relative entropies.","marker":"[39]"},{"why":"Sion's minimax theorem used to exchange infimum and supremum in the support-function arguments.","marker":"[33]"},{"why":"Source for pinching properties, Golden-Thompson type inequalities, and relative entropy variational formulas.","marker":"[34]"}],"fun_headline_variants":["Uhlmann's theorem now holds for every Rényi relative entropy","All Rényi relative entropies gain Uhlmann's extension property","Regularized Uhlmann equality for all α-Rényi divergences","Measured Uhlmann theorem extends Rényi entropies","Uhlmann's identity generalizes to the entire Rényi family"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on a de Finetti-type bound stating that any permutation-symmetric extension of $σ_A^{{⊗n}}$ is dominated, up to a polynomial factor, by a mixture of tensor powers of single-copy extensions of σ_A; if that bound fails, the support-function lemma and the measured Uhlmann inequality lose their single-letter form.","fun_headline_variants_meta":{"raw":{"variants":["Uhlmann's theorem now holds for every Rényi relative entropy","All Rényi relative entropies gain Uhlmann's extension property","Regularized Uhlmann equality for all α-Rényi divergences","Measured Uhlmann theorem extends Rényi entropies","Uhlmann's identity generalizes to the entire Rényi family"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1778,"prompt_tokens":932,"completion_tokens":846,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":750}},"tokens_in":548,"tokens_out":846,"duration_ms":6769,"temperature":1.0,"reasoning_tokens":750,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:37:58.001186+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute, for a specific pair of finite-dimensional states such as ρ_AB = |φ⟩⟨φ|_AB with |φ⟩ = √(1/4)|00⟩ + √(3/4)|11⟩ and σ_A = 1/2 I_A, the sequence (1/n) D_{3/2}($ρ_AB^{{⊗n}}$ ‖ $C^{{σ_A^{⊗n}}$}_{A^n B^n}) for n = 1, 2, 3, ... and check whether it converges to D_{3/2}(ρ_A‖σ_A). A limit below D_{3/2}(ρ_A‖σ_A) would disprove the regularized Uhlmann theorem for α = 3/2.","supporting_citations":[{"cited_title":"Metger, O","cited_arxiv_id":null,"evidence_quote":"Proved the regularized Uhlmann equality for α > 1 and supplied pinching lemmas and single-copy counterexamples."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sion's minimax theorem used to exchange infimum and supremum in the support-function arguments."}],"review_version":1}