{"id":"84f289a6-02a3-4326-826b-1ef9a23c1378","arxiv_id":"2505.06050","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The strong converse exponents of partially smoothed mutual max-information and conditional min-entropy are determined exactly for classical and pure quantum states.","lead":"This paper computes exact exponential decay rates for two quantum information measures when the problem size grows. It provides formulas for classical and pure quantum states and applies them to data compression and randomness extraction.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sion's minimax is invoked on the open interval s∈(0,1) in Proposition 9 without a compactness or limiting argument, leaving the achievability of the privacy amplification exponent (Theorem 8) and Eq. (36) with an unclosed gap.","rationale":"The reader's weakest assumption correctly identifies the Sion's minimax exchange in Proposition 9 as the most load-bearing step: it underpins the achievability of the classical privacy amplification exponent (Theorem 8), which in turn gives the achievability part of Eq. (36) for the purified-distance conditional min-entropy. My own review confirms that the proof of Proposition 9 applies Sion's theorem to an open interval (0,1) without addressing compactness or providing a limiting argument. I also found the same pattern in Proposition 18 and Proposition 28, which are used for applications and the non-uniformity claim. However, the issue is not a fatal mathematical error: the functions in question are linear in the maximization variable, so the supremum over the open interval equals the maximum over its closure, and the theorem can be applied on the compact interval [0,1] after verifying convexity and semicontinuity. Thus the central claim is not disproven, but the written proof is incomplete and should be amended. For this reason, I concur with the conditional verdict and recommend no change to the reader's assessment.","tokens_in":32296,"tokens_out":62446,"duration_ms":574438,"concrete_test":"Formally re-derive Proposition 9 by replacing s∈(0,1) with s∈[0,1]: verify that for each (t,τ) the function A+sB is linear in s, hence sup_{0<s<1}(A+sB)=max_{s∈[0,1]}(A+sB); then confirm that Q(R)×∏_x S_{ρ_x} is compact and convex, that the objective is continuous in (t,τ) for each s, and that Sion's minimax theorem applies on [0,1]×Q(R)×∏_x S_{ρ_x}. If the equality (54) and the subsequent achievability proof of Theorem 8 hold under this extension, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In the proof of Proposition 9 (arXiv:2505.06050, Section V.A), the variational expression (54) is obtained by exchanging sup_{0<s<1} with inf over t∈Q(R) and {τ_x}∈F, invoking Sion's minimax theorem. The domain of s is the open interval (0,1), which is not compact. The paper only states that convexity/concavity and semicontinuity hold, citing Proposition 3(v)–(viii); it does not explain why the theorem applies on a non-compact domain. This matters because the same exchange is used in Proposition 18 and Proposition 28, and the achievability part of Theorem 8 relies on Proposition 9. If the exchange is invalid, the variational expression (54) and the upper bound on the classical privacy amplification exponent collapse, and consequently the purified-distance exponent in Eq. (36) loses its achievability proof. The gap is likely fixable: for each fixed (t,τ), A(t,τ)+sB(t,τ) is linear in s, so the supremum over (0,1) equals the maximum over [0,1], and one can extend s to the compact interval [0,1] before applying Sion. But as written, the justification is incomplete and the conclusion depends on this step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper determines exact strong converse exponents for the partially smoothed mutual max-information and conditional min-entropy, for classical states (Theorem 4) and pure states (Theorem 5). It shows that the trace-distance exponents are not uniform across state classes (Remark 7, Proposition 28). As applications, it derives strong converse exponents for blind quantum data compression (Theorem 15), intrinsic randomness (Theorem 17), and classical state splitting (Theorem 19). A key technical ingredient is the exact strong converse exponent for classical privacy amplification (Theorem 8), obtained via a variational expression (Proposition 9) and a long achievability proof. The paper is technically dense and builds on several published lemmas from the authors and others.","tokens_in":32557,"tokens_out":15292,"duration_ms":158050,"significance":"The paper addresses a genuine open problem in one-shot quantum information theory. The main results — exact strong converse exponents for partially smoothed information measures in the classical and pure-state cases — are natural and likely useful; the observation that trace-distance exponents are not uniform across state classes is a notable conceptual point. The derivation of the classical privacy amplification exponent is a contribution of independent interest. The proofs are long and mostly well structured, and the paper is honest about its reliance on external results. However, the validity of several minimax swaps, especially in Proposition 9, is not fully justified as written, and since those swaps support the achievability directions of the central theorems, the soundness of the claimed exponents is not yet fully established.","major_comments":[{"comment":"In the proof of Proposition 9 (Eqs. (54)-(55)), the equality between the supremum over 0<s<1 of the infimum over t and {τ_x} and the infimum over t and {τ_x} of the supremum over 0<s<1 is justified by an application of Sion's minimax theorem. The domain of s is the open interval (0,1), which is not compact, and the paper does not supply a limiting argument. This is a load-bearing step: Proposition 9 underlies the achievability of the classical privacy amplification exponent (Theorem 8) and hence Eq. (36). The gap is repairable, because for fixed (t,{τ_x}) the expression is affine in s, so the supremum over (0,1) equals the supremum over the compact interval [0,1]; the authors should extend s to [0,1] and then apply Sion. As written, the proof is incomplete.","section":"Section V.A, Proposition 9"},{"comment":"The same non-compact minimax swap is used in the proof of Eq. (87) in Section VI (swap of sup_{0<s<1} and inf_{t∈Q(X)}) and in Proposition 18 (Eq. (124)) for the intrinsic randomness exponent. In both places the function is linear in the auxiliary parameter for fixed outer variables, so the same extension to the closed interval repairs the argument. The authors should add a remark or lemma stating the compactification once and refer to it at all three occurrences.","section":"Section VI and Section VIII.B"}],"minor_comments":[{"comment":"The introduction names 'an improved Fuchs-van de Graaf inequality (Lemma 23 in Appendix)', but Lemma 23 is a fidelity inequality for projections; the improved Fuchs-van de Graaf inequality used later is Lemma 26. The cross-reference should be corrected.","section":"Introduction"},{"comment":"The derivation in Eq. (61) invokes the Fannes-Audenaert inequality and Hölder's inequality without explicit statements or references; since the resulting constants enter the exponent bound, a citation and a brief statement of the inequalities would help the reader.","section":"Eq. (61)"},{"comment":"The formulas in Eq. (36) and Eq. (51) take a supremum over the closed interval [1/2,1], including α=1, whereas the proofs are written only for 1/2<α<1. The authors should note the continuity of the relevant Rényi quantities in α to justify passage to the endpoint.","section":"Theorem 4 and Theorem 8"},{"comment":"In Eq. (103) and Eq. (104), the notation ϵ^P(ρ^{⊗n},nr) omits the subscripts ˙R^n:A^n; for consistency with the rest of the paper the full notation ϵ^P_{\\dot R^n:A^n}(ρ^{⊗n},nr) should be used.","section":"Proposition 12, Eq. (103)-(104)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically dense and the central claims are likely correct, but the minimax issue in Proposition 9 is load-bearing and must be fixed with a short compactification argument. I recommend major revision rather than rejection because the gap is clearly repairable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nYou should know two things about this paper. It contains genuinely new results: the exact strong converse exponents for partially smoothed conditional min-entropy and mutual max-information for classical states under purified distance, for pure states under both distances, and the classical privacy amplification exponent. And it has a real, fixable technical gap in the justification of a key minimax step. That combination makes it worth engaging with, but not worth taking on faith yet.\n\nThe main contributions are solid. The paper determines the formulas in Theorem 4 (especially Eq. (36)) and Theorem 5, and shows the trace-distance exponents are not uniform across state classes. The privacy amplification exponent in Theorem 8 is a by-product that is independently interesting. The applications to blind quantum data compression, intrinsic randomness, and classical state splitting follow naturally. The proofs are long and technical, but the overall structure is coherent. The paper cites prior work properly, including the authors' own lemmas where they are genuinely needed, and I don't see circularity.\n\nThe soft spot is the one the stress-test note flags. In Proposition 9, the variational expression (54) is obtained by swapping sup_{0<s<1} with an infimum over t and {τx}, invoking Sion's minimax theorem. The domain of s is the open interval (0,1), which is not compact. The paper asserts convexity/concavity and semicontinuity, but gives no limiting argument for why Sion applies. This matters because the same swap is used in Proposition 18 and Proposition 28, and the achievability part of Theorem 8 depends on Proposition 9. The gap is likely easy to close: the expression inside is linear in s for fixed (t,τ), so the supremum over the open interval equals the maximum over [0,1]. But as written, the proof is incomplete.\n\nMinor points: Eq. (34) is a corollary of existing results, so the novelty is concentrated in the purified-distance and pure-state exponents. The paper relies heavily on external lemmas, but those are published and independently proven.\n\nWho this is for: researchers in one-shot quantum information theory, especially those working on privacy amplification, finite-resource data compression, or state splitting. The paper deserves a serious referee, but I would not accept it without the authors closing the minimax gap. Recommend: send it to peer review with a request for a complete proof of Proposition 9, either by compactification or an explicit approximation argument.","headline":"Genuinely new strong-converse exponents for partially smoothed measures, but the key minimax step in Proposition 9 is not fully justified and needs a fix before I'd rely on the privacy amplification exponent.","tokens_in":33046,"tokens_out":3006,"would_cite":true,"duration_ms":32535,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P45","94A17"],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"The paper determines exact strong converse exponents for partially smoothed quantum information measures, and shows they are not uniform across state classes.","keywords":["partially smoothed information measures","strong converse exponents","conditional min-entropy","mutual max-information","Rényi divergences","privacy amplification","quantum data compression","intrinsic randomness"],"falsifier":"A numerical check could settle the central exchange: choose a two-letter classical state, for instance $R=\\{1,2\\}$ with $p=(1/2,1/2)$ and $\\rho^x_A$ a binary distribution, fix $r$ at an interior value, and compute both sides of (54) directly; any strict gap between the left-hand supremum-over-$\\alpha$ and the right-hand infimum over $t$ and $\\{\\tau^x_A\\}$ would refute Proposition 9.","tokens_in":32090,"feed_emoji":"⚛️","tokens_out":6618,"duration_ms":70516,"temperature":0.7,"pith_summary":"The paper derives the exact exponential rates at which partially smoothed mutual max-information and conditional min-entropy fail—that is, the rates at which their smoothing error converges to 1—for classical states and pure states. These rates, called strong converse exponents, quantify how quickly one-shot quantities deviate from their asymptotic limits when the rate parameter r crosses the mutual information or conditional entropy threshold. The main results give closed-form expressions for both trace and purified distance, and show that the trace-distance exponent takes different forms for classical and pure states, thereby disproving the natural conjecture that one formula covers all quantum states. As applications, the paper obtains exact strong converse exponents for blind quantum data compression, intrinsic randomness extraction, classical state splitting, and classical privacy amplification.","feed_headline":"Smoothing-failure exponents now exact for two quantum state classes","feed_subtitle":"Trace- and purified-distance rates differ by state class; blind compression and randomness exponents follow","key_machinery":"The arguments run on type-theoretic empirical distribution calculus combined with Rényi divergence variational formulas. The central object is the variational expression (54), which rewrites the classical privacy-amplification exponent as an infimum over distributions $t$ and states $\\{\\tau^x_A\\}$ of $2D(t\\|p)+\\mathbb{E}_{x\\sim t}D(\\tau^x_A\\|\\rho^x_A)+|r+\\mathbb{E}_{x\\sim t}D(\\tau^x_A\\|I_A)|_+$, obtained by swapping a supremum over Rényi order via Sion's minimax theorem. The operator Hölder inequality of Lemma 21 supplies the fidelity-to-Rényi bounds used in the optimality parts, while the achievability parts go through privacy amplification and quantum data compression reductions, with pinching and symmetric-state domination controlling the type-class decomposition.","core_discovery":"For classical states with bipartite distribution $p$, the trace-distance strong converse exponents are $\\sup_{0\\leq \\alpha \\leq 1}(1-\\alpha)(r-\\bar{H}_\\alpha(A|R)_p)$ for the partially smoothed conditional min-entropy and $\\sup_{0\\leq \\alpha \\leq 1}(1-\\alpha)(I_\\alpha(R:A)_p-r)$ for the partially smoothed mutual max-information, while the purified-distance conditional min-entropy exponent is $\\sup_{\\frac12\\leq\\alpha\\leq1}\\inf_{t\\in Q(R)}\\{2D(t\\|p)+\\frac{1-\\alpha}{\\alpha}(r-\\mathbb{E}_{x\\sim t}H_\\alpha(p(\\cdot|x)))\\}$. For a pure state with Schmidt coefficients $p$, both trace and purified distance give $\\inf_{t\\in Q(X)}\\{2D(t\\|p)+|r+H(t)|_+\\}$ for the conditional min-entropy and $\\sup_{\\beta>1}\\frac{\\beta-1}{\\beta}(2H_\\beta(R)_\\rho-r)$ for the mutual max-information. The paper proves that the classical trace-distance formulas cannot be converted into the pure-state formulas, so no single trace-distance formula holds for all quantum states.","pith_inferences":["A plausible next target is a mixed-state trace-distance formula that interpolates between the classical and pure expressions; the non-uniformity result suggests the exponent may require optimizing over decompositions of the mixed state into pure or classical pieces.","Because purified distance gives the same exponents as trace distance for pure states, the purified-distance classical formula (36) may be a candidate for a uniform expression across all states, though the paper does not claim this.","The classical privacy-amplification exponent (51) may extend to classical-quantum side information, since Lemma 21 is operator-valued; however, the infimum over types would then lack a classical interpretation, so a different proof would be needed.","The achievability proofs for data compression and intrinsic randomness suggest that matching finite-blocklength refinements beyond exponent order may follow from the same variational formulas."],"forward_implications":["For classical states, the trace-distance exponents become Legendre transforms of Petz Rényi conditional entropy and mutual information, giving exact large-deviation rates for the partially smoothed measures.","For pure states, the trace- and purified-distance exponents coincide and take simple relative-entropy-plus-positive-part forms.","The trace-distance strong converse exponents are not uniform over all quantum states, since classical and pure states require different formulas.","Blind quantum data compression now has an exact strong converse exponent of $\\sup_{\\beta>1}\\frac{\\beta-1}{\\beta}(2H_\\beta(A)_\\rho-2r)$, filling a gap left by the visible compression result.","Intrinsic randomness extraction, classical state splitting, and classical privacy amplification each acquire exact strong converse exponents as direct consequences."],"supporting_citations":[{"why":"Introduces the partially smoothed information measures and supplies the first-order asymptotics that the strong converse exponents refine.","marker":"[1]"},{"why":"Provides the equality between partially and standard smoothed conditional min-entropy under trace distance, used in the proof of Eq. (34).","marker":"[9]"},{"why":"Its Theorem 2 gives the standard strong converse exponent that Eq. (34) reduces to in the classical case.","marker":"[33]"},{"why":"Lemma 2 supplies the trace-distance expression for the mutual max-information smoothing error used in Eq. (44).","marker":"[34]"},{"why":"Lemma 21, the operator Hölder inequality, carries the purified-distance optimality bounds throughout the paper.","marker":"[52]"},{"why":"Establishes the asymptotic privacy-amplification rate used in the achievability proof of Theorem 8.","marker":"[36]"},{"why":"Provides the secure randomness extraction and strong converse bounds used in the achievability case of Theorem 8.","marker":"[39]"},{"why":"Supplies additivity and duality of Rényi mutual information, used in deriving the classical exponent formulas.","marker":"[13]"},{"why":"Provides the method of types and the classical compression strong converse exponent underlying Eq. (108).","marker":"[32]"}],"fun_headline_variants":["Trace-distance converse exponents split by state class","Pure vs classical: trace-distance exponents diverge","Exact smoothing-failure rates for pure and classical states","State-dependent strong converse exponents pinned down","Smoothing-failure exponents: pure vs classical differ"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the supremum over the Rényi order can be swapped with the infimum over distributions and states in formulas such as (54), even though the order domain is only $(0,1)$ and no compactness argument is supplied; if this exchange fails, the variational expressions and the achievability part of the privacy-amplification exponent lose their foundation.","fun_headline_variants_meta":{"raw":{"variants":["Trace-distance converse exponents split by state class","Pure vs classical: trace-distance exponents diverge","Exact smoothing-failure rates for pure and classical states","State-dependent strong converse exponents pinned down","Smoothing-failure exponents: pure vs classical differ"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000764,"raw_usage":{"total_tokens":3359,"prompt_tokens":887,"completion_tokens":2472,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":2400}},"tokens_in":503,"tokens_out":2472,"duration_ms":18621,"temperature":1.0,"reasoning_tokens":2400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:52:50.510390+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A numerical check could settle the central exchange: choose a two-letter classical state, for instance $R=\\{1,2\\}$ with $p=(1/2,1/2)$ and $\\rho^x_A$ a binary distribution, fix $r$ at an interior value, and compute both sides of (54) directly; any strict gap between the left-hand supremum-over-$\\alpha$ and the right-hand infimum over $t$ and $\\{\\tau^x_A\\}$ would refute Proposition 9.","supporting_citations":[{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Introduces the partially smoothed information measures and supplies the first-order asymptotics that the strong converse exponents refine."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"Provides the equality between partially and standard smoothed conditional min-entropy under trace distance, used in the proof of Eq. (34)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its Theorem 2 gives the standard strong converse exponent that Eq. (34) reduces to in the classical case."},{"cited_title":"S.: Sandwiched R´ enyi divergence satisfies data processing inequality","cited_arxiv_id":null,"evidence_quote":"Lemma 2 supplies the trace-distance expression for the mutual max-information smoothing error used in Eq. (44)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Lemma 21, the operator Hölder inequality, carries the purified-distance optimality bounds throughout the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the asymptotic privacy-amplification rate used in the achievability proof of Theorem 8."},{"cited_title":"Exponents for Shared Randomness-Assisted Channel Simulation","cited_arxiv_id":"2410.07051","evidence_quote":"Provides the secure randomness extraction and strong converse bounds used in the achievability case of Theorem 8."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies additivity and duality of Rényi mutual information, used in deriving the classical exponent formulas."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the method of types and the classical compression strong converse exponent underlying Eq. (108)."}],"review_version":1}