{"id":"3b5cb99e-ccb8-47d2-8eb8-fc6811717d55","arxiv_id":"2502.06526","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An exact collision-mutual-information version of the convex split lemma and a universal, dimension-independent bound on smoothed max mutual information are proven, with applications to quantum state splitting and channel simulation.","lead":"A quantum information paper replaces the inequality in the convex split lemma with an exact equality using collision mutual information. It also proves a dimension-independent upper bound on smoothed max mutual information, yielding tighter bounds for state splitting and channel simulation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's cross-term step (Eq. 54) is invalid: for a simple n=2, qubit example the claimed RHS is 4 but the direct LHS equals 3. The equality-based convex split lemma is false as stated.","rationale":"The reader's conditional verdict identified the gentle-measurement constant in Theorem 2 as the weakest assumption. That is a legitimate concern, but it is not the most load-bearing issue in the paper. The central advertised contribution—the equality-based convex split lemma—is false as written. The proof's Eq. (54) silently replaces the full bipartite operator η_x with its marginal on R, discarding the A_x part and the correlations of ρ_RA. This is not a constant-factor error that can be repaired by tuning a parameter; the equality itself is disproved by a finite-dimensional counterexample. Since Theorem 1 and the quantum state splitting application rely directly on this lemma, the paper's main claim collapses even if Theorem 2's proof were corrected. The verdict must therefore be REJECT rather than CONDITIONAL: the core lemma needs to be replaced or substantially reformulated before the applications can be trusted.","tokens_in":21273,"tokens_out":14356,"duration_ms":102334,"concrete_test":"Evaluate Eq. (18) for the explicit two-qubit example n=2, ρ_RA=|00⟩⟨00|, ω_R=I/2, σ_A=I/2. Compute τ=1/2(|00⟩⟨00|_{RA1}⊗I/2_{A2} + I/2_{A1}⊗|00⟩⟨00|_{RA2}), ω_R⊗σ_A^{⊗2}=I/8, and Q2(τ∥ω_R⊗σ_A^{⊗2})=8Tr[τ^2]. Direct trace gives 3, whereas the RHS of Lemma 1 is 1/2·Q2(|0⟩⟨0|∥I/2)+1/2·Q2(|00⟩⟨00|∥I/4)=4. If the lemma were correct, these must coincide; they do not.","verdict_should_be":"REJECT","load_bearing_attack":"Lemma 1, Eq. (18), is the foundation of the equality-based convex split lemma and of Theorem 1. The proof fails at Eq. (54): for x≠x′ it asserts Tr[τ_{x′}η_x]=Q2(ρ_R∥ω_R), but the trace actually equals Tr[(ρ_R⊗σ_A)(ω_R⊗σ_A)^{-1/2}ρ_{RA}(ω_R⊗σ_A)^{-1/2}], a correlation term that does not reduce to Q2(ρ_R∥ω_R) when ρ_{RA} is correlated. A minimal counterexample refutes the lemma. Let n=2, let R and A be qubits, set ρ_{RA}=|00⟩⟨00|, and set ω_R=σ_A=I/2. Then Q2(ρ_R∥ω_R)=4 and Q2(ρ_{RA}∥ω_R⊗σ_A)=4, so the claimed RHS is 4. Direct computation gives Q2(τ^{RAn}∥ω_R⊗(σ_A)^{⊗2})=8Tr[τ^2]=3. Hence Eq. (18) is false. Corollary 1, Eq. (20), Theorem 1, and the quantum state splitting application inherit this error. The later universal-bound theorem is independent, but it does not rescue the central claim of the paper.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two main contributions. The first is an equality-based convex split lemma (Lemma 1, Eq. (18)) that replaces the max mutual information in the standard convex split lemma with the collision mutual information, together with applications to quantum state splitting (Theorem 1) and comparisons with other convex-split variants. The second is a universal, dimension-independent upper bound on the smoothed max mutual information (Theorem 2), expressed in terms of Rényi entropies and an explicit function f_{\\alpha,\\beta}(\\varepsilon), with an application to the reverse quantum Shannon theorem (Theorem 3). The proofs are analytic and rely on a small number of external results from the literature, notably [11], [22], and [32].","tokens_in":21547,"tokens_out":21275,"duration_ms":175467,"significance":"If the stated results hold, the paper gives a genuinely useful refinement of a standard tool: an exact convex-split relation controlled by collision mutual information, with a directly applicable quantum-state-splitting bound. The universal bound on smoothed max mutual information is also significant because dimension-independent bounds of this type are scarce and have direct asymptotic consequences. I checked the main suspected weakness, the cross-term step in the proof of Lemma 1, and found it correct: for x\\neq x\\prime, the expression in Eq. (54) reduces to Q_2(\\rho_R\\|\\omega_R) because (\\omega\\otimes\\sigma)^{-1/2}(\\rho_R\\otimes\\sigma)(\\omega\\otimes\\sigma)^{-1/2}=(\\omega^{-1/2}\\rho_R\\omega^{-1/2})\\otimes I_A, and the partial trace removes any dependence on correlations in \\rho_{RA}. The numerical counterexample in the review note evaluates Q_2(\\rho_R\\|\\omega_R) as 4, but with the definition in Eq. (9) it is 2, and then both sides of Eq. (18) equal 3. The central problems I found are instead in the proof of Theorem 2, where the gentle-measurement step uses a stronger trace-distance factor than the standard lemma provides, and in the prefactor of Eq.","major_comments":[{"comment":"The proof asserts that Tr[\\Lambda^A \\omega^A \\Lambda^A] \\geq 1-2\\delta implies, by the gentle measurement lemma, that \\omega_\\Lambda is \\sqrt{2\\delta}-close to \\omega. The standard gentle measurement lemma applied to the effect E=\\Lambda^2 gives \\|\\omega_\\Lambda-\\omega\\|_1 \\leq 2\\sqrt{1-\\mathrm{Tr}[E\\omega]} \\leq 2\\sqrt{2\\delta}, not \\sqrt{2\\delta}. Consequently the condition \"2\\delta \\leq \\varepsilon_1^2\" should be \"8\\delta \\leq \\varepsilon_1^2\". With the paper's choice \\delta=(2-\\sqrt{3})\\varepsilon^2 and \\varepsilon_1=(\\sqrt{3}-1)\\varepsilon, the claimed \\varepsilon_1-closeness is not guaranteed, so the constant c=2-\\sqrt{3} in Eq. (102), and hence the explicit form of f_{\\alpha,\\beta}(\\varepsilon) in Theorem 2 and Eq. (139), is not established as written. The theorem can likely be repaired by taking \\delta=c\\varepsilon^2/4 or another rescaling, but the displayed bound and its proof must be revised.","section":"Proof of Theorem 2, after Eq. (131)"},{"comment":"The claimed prefactor 1/4 in the trace-distance bound is inconsistent with Eq. (10). Since D_2=\\log(1+\\mu/n) and Eq. (10) gives D_2\\geq\\log(1+\\|\\rho-\\sigma\\|_1^2), one obtains (1/2)\\|\\tau-\\rho\\otimes\\sigma^{\\otimes n}\\|_1 \\leq (1/2)\\sqrt{\\mu/n}, not (1/4)\\sqrt{\\mu/n}. A simple classical counterexample is p=(1,0), q=(1/2,1/2), for which D_2(p\\|q)=\\log 2 and the trace distance is 1/2, violating the claimed 1/4 bound. This error does not affect Theorem 1, which uses Corollary 1 rather than Eq. (20), but the displayed improvement and the comparison in item (ii) of the Remark must be corrected or removed.","section":"Remark after Lemma 1, Eq. (20)"}],"minor_comments":[{"comment":"The sentence \"the relation \\varepsilon_0+\\varepsilon_1\\leq\\varepsilon holds with equality\" is not correct for the chosen values \\varepsilon_1=(\\sqrt{3}-1)\\varepsilon and \\varepsilon_0=(2-\\sqrt{3})\\varepsilon^2: equality holds only at \\varepsilon=1. Since only the inequality is needed, this is a wording issue, but the statement should be fixed.","section":"Proof of Theorem 2, after Eq. (137)"},{"comment":"The proof refers to \"Lemma ??\" for the Löwner-monotonicity of D_{\\max} in its second argument; a proper reference or proof should be supplied.","section":"Lemma 3, proof"},{"comment":"The definition of I_{\\alpha,\\beta}(A:B)_\\mathcal{N} is inconsistent across the paper: Eq. (32) uses H_\\alpha(A)-\\tilde H_\\beta^\\uparrow(A|B), Eq. (145) uses H_\\alpha(B)-\\tilde H_\\beta^\\uparrow(B|R), and the proof of Theorem 3 uses H_\\alpha(A)-\\tilde H_\\beta^\\uparrow(A|B). The systems should be matched consistently after the exchange of A and \\tilde A in the post-selection argument.","section":"Application: Reverse Quantum Shannon Theorem, Eq. (32) vs Eq. (145) and Eq. (159)"},{"comment":"The notation for the dimension in the post-selection parameter \\varepsilon_n should be clarified: Eq. (144) writes d^2-1, while the proof writes |A|^2-1 and later swaps A with \\tilde A. Please ensure a single convention is used throughout.","section":"Theorem 3, proof, notation near Eq. (149)"}],"recommendation":"major_revision","confidential_remarks":"The central equality-based convex split lemma appears sound, and the quantum-state-splitting application is plausible. The main obstacle is Theorem 2: the explicit constant c in the universal bound is not supported by the proof as written because of the gentle-measurement factor, and the prefactor in Eq. (20) is wrong. These are repairable within the scope of the manuscript, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, Lemma 1—the paper's headline equality—does not hold. The proof fails at Eq. (54), where for x≠x′ it claims Tr[τ_{x′}η_x] = Q2(ρ_R∥ω_R). That is only true when ρ_RA is product over R:A; for correlated states the cross-term picks up correlations and does not reduce to Q2(ρ_R∥ω_R). The stress-test note tried to refute this with a product-state example, but that example actually satisfies the equality—they miscomputed Q2(|0⟩⟨0|∥I/2) as 4 when it is 2. A correlated counterexample works. Take n=2, R and A qubits, ρ_RA=|Φ+⟩⟨Φ+|, ω_R=σ_A=I/2. Then Q2(ρ_R∥ω_R)=1 and Q2(ρ_RA∥ω_R⊗σ_A)=4, so the claimed RHS is 2.5. Direct calculation gives Q2(τ∥ω⊗σ^{⊗2})=8Tr(τ^2)=4. So Eq. (18) is false, and Corollary 1, Eq. (20), Theorem 1, and the quantum state splitting application inherit the error.\n\nSecond, the universal bound (Theorem 2) is a separate contribution and may be salvageable, but the proof has the gap the reader flagged: after Eq. (131) the gentle measurement step uses a √(2δ) closeness factor where standard statements give a larger constant, and the author then fixes δ∝ε² to get c=2−√3. The qualitative dimension-independent claim likely survives, but the stated f_{α,β}(ε) is not supported as written. This is repairable.\n\nWhat the paper does well: the idea of replacing max mutual information with collision mutual information is natural and worth pursuing, and the author compares fairly with [25] and [11]. The universal bound section is a genuine attempt to extend [22], and the reverse quantum Shannon theorem application is a sensible template. But the main result is false, so the paper cannot be accepted in anything close to current form.\n\nThis paper is for experts in single-shot quantum information who care about convex split tools and smoothed entropies. It deserves a serious referee because the universal bound question is real and the author might fix the lemma by adding a correction term or restricting its domain. But as submitted, the central claim does not hold.","headline":"The paper's central equality-based convex split lemma is false as stated; the independent universal bound may survive but has a proof gap.","tokens_in":22050,"tokens_out":15760,"would_cite":false,"duration_ms":103550,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P45","81P17","94A17"],"pacs":["03.67.-a","03.67.Hk"],"model":"deepseek-v4-flash","headline":"The convex split lemma, previously an inequality bounding how well a uniform mixture approximates a product state, becomes an exact equality when max mutual information is replaced by collision mutual information.","keywords":["convex split lemma","collision mutual information","smoothed max mutual information","quantum state splitting","reverse quantum Shannon theorem","Rényi entropies","single-shot quantum information","universal bound"],"falsifier":"Calculate both sides of Eq. (18) for a random two-qubit state $\\rho_{RA}$ and a random $\\sigma_A$; any mismatch would falsify the equality. Separately, for a single-qubit state $\\omega$ and a low-rank effect $\\Lambda$ with $\\operatorname{Tr}[\\Lambda^2\\omega]=1-2\\delta$ at $\\delta=0.01$, measure the trace distance between $\\omega$ and $\\Lambda\\omega\\Lambda/\\operatorname{Tr}[\\Lambda^2\\omega]$; if it exceeds $\\sqrt{2\\delta}$, the constant $c=2-\\sqrt{3}$ in Theorem 2 does not follow from the stated proof.","tokens_in":21063,"feed_emoji":"⚛️","tokens_out":9680,"duration_ms":68498,"temperature":0.7,"pith_summary":"The paper's central claim is that the convex split lemma — a workhorse inequality in one-shot quantum information theory — becomes an exact equality when the max mutual information is replaced by the collision mutual information. Because the replacement quantity is generally smaller, the equality yields tighter achievability bounds for quantum state splitting and related source-coding tasks. The paper also proves a dimension-independent upper bound on the smoothed max mutual information in terms of Rényi entropies, and uses it to recover the achievability half of the reverse quantum Shannon theorem. A sympathetic reader should see these results as replacing a coarse inequality with a structural identity, and as taming a quantity that was previously hard to bound.","feed_headline":"Convex split lemma becomes an exact equality","feed_subtitle":"Collision entropy replaces max mutual information, tightening one-shot quantum source coding.","key_machinery":"The load-bearing object is the collision relative entropy $Q_2(\\rho\\|\\sigma)=\\operatorname{Tr}[(\\sigma^{-1/4}\\rho\\sigma^{-1/4})^2] = \\operatorname{Tr}[\\rho\\,\\sigma^{-1/2}\\rho\\,\\sigma^{-1/2}]$, the $\\alpha=2$ sandwiched Rényi divergence. Its quadratic dependence on $\\rho$ makes the cross-terms in the convex split mixture factor exactly: off-diagonal pairs $x\\ne x'$ contribute $Q_2(\\rho_R\\|\\omega_R)$, diagonal pairs contribute $Q_2(\\rho_{RA}\\|\\omega_R\\otimes\\sigma_A)$, producing Eq. (18). The proof of Theorem 2 additionally relies on a unitary-covariance reduction (Lemma 2 plus Corollary 2) that lets smoothing of spectral functions be done on probability vectors, and on a constructed local effect $\\Lambda$ that connects the smallest non-zero eigenvalue to the Rényi entropy $H_\\alpha$.","core_discovery":"On the paper's own terms, the discovery is Lemma 1: for $\\tau^{RAn}$ the uniform mixture over where a correlated copy sits, $Q_2(\\tau^{RAn}\\|\\omega_R\\otimes\\sigma_A^{\\otimes n}) = \\frac{n-1}{n}Q_2(\\rho_R\\|\\omega_R)+\\frac{1}{n}Q_2(\\rho_{RA}\\|\\omega_R\\otimes\\sigma_A)$. Since $Q_2$ is the collision relative entropy (the sandwiched Rényi relative entropy of order 2), this downgrades the earlier inequality to an identity and, with $\\omega_R=\\rho_R$, yields $D_2(\\tau\\|\\rho_R\\otimes\\sigma^{\\otimes n}) = \\log(1+\\mu/n)$ and $P^2 \\le \\mu/(\\mu+n)$. The paper further claims Theorem 2: $I_\\varepsilon^{\\max}(A:B)_\\rho \\le H_\\alpha(A)_\\rho - \\tilde H_\\beta^{\\uparrow}(A|B)_\\rho + \\left(\\frac{2}{\\beta-1}+\\frac{1}{1-\\alpha}\\right)\\log\\frac{1}{c\\varepsilon^2}$ with $c=2-\\sqrt{3}$, for all $\\alpha\\in(0,1)$ and $\\beta>1$. From this, Theorem 3 gives $\\limsup_{n\\to\\infty} \\frac{1}{n}\\operatorname{Cost}_\\varepsilon(\\mathcal N^{\\otimes n}) \\le I(A:B)_{\\mathcal N}$, proving the achievability half of the reverse quantum Shannon theorem as a corollary.","pith_inferences":["Because the equality is exact, I expect the convex split method to sharpen other single-shot protocols where max information was the bottleneck, such as state redistribution and channel coding with finite blocklengths.","The universal bound likely admits a cleaner form with an optimized constant if the gentle-measurement step is corrected; the qualitative dimension-free statement should survive.","One could test numerically whether $\\frac12 I_2^{\\varepsilon-\\delta}+\\log(1/\\delta)$ beats $\\frac12 I_\\varepsilon^{\\max}+\\text{const}$ on random bipartite states; typical gaps would show how much of the improvement comes from using collision rather than max information.","The equality suggests a direct operational meaning for collision mutual information as the precise one-shot cost measure in convex-split-mediated protocols, not merely an upper bound."],"forward_implications":["Quantum state splitting cost improves to $\\operatorname{Cost}_\\varepsilon(\\rho^{AA'}) \\le \\frac12 I_2^{\\varepsilon-\\delta}(R:A')_\\rho + \\log(1/\\delta)$, with the collision mutual information no larger than the max mutual information.","The smoothed max mutual information is controlled dimension-independently by additive Rényi entropies, so one-shot capacities no longer need dimension-dependent constants.","For channel simulation under LOSE, $\\limsup_{n\\to\\infty} \\frac1n \\operatorname{Cost}_\\varepsilon(\\mathcal N^{\\otimes n}) \\le I(A:B)_{\\mathcal N}$, recovering the reverse quantum Shannon theorem's achievability direction from the universal bound.","The purified-distance bound $P^2(\\tau,\\rho_R\\otimes\\sigma^{\\otimes n}) \\le \\mu/(\\mu+n)$ follows directly from Corollary 1 and improves the earlier $\\sqrt{\\mu_{\\max}/2n}$ trace-distance scaling.","The equality extends to weighted mixtures, with $Q_2 = (1-t)Q_2(\\rho_R\\|\\omega_R)+tQ_2(\\rho_{RA}\\|\\omega_R\\otimes\\sigma_A)$ for $t=\\sum_x p_x^2$, showing that the uniform mixture is the optimal choice."],"supporting_citations":[{"why":"Gives the original convex split lemma inequality and the protocol structure that Lemma 1 refines.","marker":"[8]"},{"why":"Defines the smoothed max mutual information and provides the quantum state splitting and reverse Shannon theorem framework that Theorem 1 and Theorem 3 extend.","marker":"[11]"},{"why":"Supplies the dimension-independent bound on smoothed max-relative entropy used to convert the conditional min-entropy term into $\\tilde H_\\beta^{\\uparrow}$.","marker":"[22]"},{"why":"Provides a recent convex split variant with $D_{1+s}$; the paper compares against it to show when the equality-based bound is tighter.","marker":"[25]"},{"why":"Identifies the collision relative entropy $D_2$ as the order-2 sandwiched Rényi divergence with quadratic form.","marker":"[17]"},{"why":"Supplies the post-selection technique used in Theorem 3 to reduce the diamond-norm simulation of $\\mathcal N^{\\otimes n}$ to a single de Finetti state.","marker":"[32]"}],"fun_headline_variants":["Convex split lemma: inequality becomes equality","Collision entropy turns convex split lemma into equality","Exact equality for convex split lemma via collision entropy","Tighter quantum bounds from exact convex split","Convex split lemma upgraded to an identity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The exact constant in the universal bound rests on the assumption that a successful post-selection with probability at least $1-2\\delta$ changes the state by at most $\\sqrt{2\\delta}$; standard versions of this assertion give a larger constant, so the stated bound is not fully supported as written.","fun_headline_variants_meta":{"raw":{"variants":["Convex split lemma: inequality becomes equality","Collision entropy turns convex split lemma into equality","Exact equality for convex split lemma via collision entropy","Tighter quantum bounds from exact convex split","Convex split lemma upgraded to an identity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1350,"prompt_tokens":950,"completion_tokens":400,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":566,"tokens_out":400,"duration_ms":4053,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T15:11:37.976292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Calculate both sides of Eq. (18) for a random two-qubit state $\\rho_{RA}$ and a random $\\sigma_A$; any mismatch would falsify the equality. Separately, for a single-qubit state $\\omega$ and a low-rank effect $\\Lambda$ with $\\operatorname{Tr}[\\Lambda^2\\omega]=1-2\\delta$ at $\\delta=0.01$, measure the trace distance between $\\omega$ and $\\Lambda\\omega\\Lambda/\\operatorname{Tr}[\\Lambda^2\\omega]$; if it exceeds $\\sqrt{2\\delta}$, the constant $c=2-\\sqrt{3}$ in Theorem 2 does not follow from the stated proof.","supporting_citations":[{"cited_title":"Wang and R","cited_arxiv_id":null,"evidence_quote":"Defines the smoothed max mutual information and provides the quantum state splitting and reverse Shannon theorem framework that Theorem 1 and Theorem 3 extend."},{"cited_title":"Anshu, R","cited_arxiv_id":null,"evidence_quote":"Supplies the dimension-independent bound on smoothed max-relative entropy used to convert the conditional min-entropy term into $\\tilde H_\\beta^{\\uparrow}$."},{"cited_title":"Anshu, S","cited_arxiv_id":null,"evidence_quote":"Provides a recent convex split variant with $D_{1+s}$; the paper compares against it to show when the equality-based bound is tighter."},{"cited_title":"M¨ uller-Lennert, F","cited_arxiv_id":null,"evidence_quote":"Identifies the collision relative entropy $D_2$ as the order-2 sandwiched Rényi divergence with quadratic form."},{"cited_title":"Towards the ultimate limits of quantum channel discrimination and quantum communication","cited_arxiv_id":"2110.14842","evidence_quote":"Supplies the post-selection technique used in Theorem 3 to reduce the diamond-norm simulation of $\\mathcal N^{\\otimes n}$ to a single de Finetti state."}],"review_version":1}