{"id":"108b4493-d7fb-4fc9-91d7-3bf5bdfe52ab","arxiv_id":"2608.11000","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For every quantum channel, the identification capacity is at most the entanglement-assisted transmission capacity, with equality for low-noise channels and strict inequality for a family of transpose-depolarizing channels.","lead":"This paper proves a new upper bound on how many messages can be reliably identified, rather than transmitted, through a noisy quantum channel, and shows the bound is exact for low-noise channels. It also finds a channel where identification capacity is strictly smaller than the entanglement-assisted transmission capacity, and where identification capacity is superadditive.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Universal converse is not self-contained: Theorem 3.3 rests on the unproved one-shot geometric bound (2.51)/Theorem 2.10 from the author's companion preprint [25]; if that bound has a wrong dimension or fails for non-tensor-product channels, C_ID <= C_E is unsupported.","rationale":"Reading in good faith: the paper's internal chain from Lemma 3.2 through smoothing and the AEP is coherent, and the Proposition 3.5/Corollary 3.6 strictness and superadditivity arguments are valid given Theorem 3.3. The identified factor-of-2 error in Eqs. (3.30)-(3.31) does not change the final theorem, because the error budget can absorb 2 epsilon. Thus I do not see an internal inconsistency that would reject the core result. The single load-bearing weakness is that the universal converse is outsourced to Theorem 2.10 and Eq. (2.51) from the author's own unpublished preprint [25]. That theorem is the only place where the Gaussian-width-to-packing reduction is established; it is not reproduced, machine-checked, or independently verified here. Since the one-shot bound is applied to the smoothed channel tilde N_n, which is not a tensor-product channel, even a subtle inapplicability of [25] to non-product channels would invalidate Theorem 3.3. This is exactly the condition identified by the reader as the weakest assumption, so the conditional verdict is appropriate. An independent derivation plus targeted numerical checks would settle whether the concern lands.","tokens_in":19864,"tokens_out":28702,"duration_ms":260405,"concrete_test":"Independently derive (2.51) and Theorem 2.10 from the definition of w_{G,W}, an epsilon-net on the state space, and Sudakov's inequality, tracking all constants and dimension factors. Then test the derived bound on the identity channel id_d with the Bures structure (3.10): for d=2,4,8 and n=1,2, numerically sample w_{G,W}(I_n(id_d)) and compare with sqrt(2n log d_A ||Q_{id_d^{⊗n},W}||_∞). If the bound is violated, if it scales with the wrong power of d, or if the resulting double-log rate is not 2 log d = C_ID(id_d), Theorem 3.3 is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every step in Theorem 3.3 after Lemma 3.2 is a reduction to the one-shot Euclidean-Gaussian converse. An arbitrary diamond-close smoothed channel is fed into Theorem 2.10, yielding log N <= (2nC log d_A / zeta^2) mu*(tilde N_n), and Theorem 2.10 is imported verbatim from [25, Thm 5.2]. The paper states but does not prove the key input-width estimate w_{G,W}(I_n(N)) <= sqrt(2n log d_A ||Q_{N^{⊗n},W}||_∞). The central claim therefore inherits any hidden assumption, missing constant, or dimension error in [25]. In particular, the bound must hold for the non-tensor-product one-shot channel tilde N_n after smoothing; a proof that only works for tensor powers would break the argument at exactly this step. If the dimension in (2.51) should be log d_B instead of log d_A, or if the Sudakov step has a superlinear n factor, the limiting rate would change and the strong converse would not follow. The reader's factor-of-2 slip in (3.30)-(3.31) is real but nonfatal: the correct effect bound is 2 epsilon, not epsilon; choosing epsilon < Delta/4 still keeps the smoothed code valid.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper claims two main results. First, it proves a universal strong converse bound for classical identification over every finite-dimensional quantum channel: C_ID(N) ≤ C_E(N), where C_E is the entanglement-assisted classical transmission capacity. The proof combines a one-shot Euclidean Gaussian mean width converse (Theorem 2.10, imported from the companion preprint [25]) with a new technical lemma (Lemma 3.2) relating the optimized geometric functional to the channel max-information via the Bures metric, followed by smoothing and an asymptotic equipartition property for the smooth max-information. Second, the paper shows that this upper bound can be strict: for the transpose-depolarizing channel at q = d/(d+1), C_ID < C_E. As a corollary, it obtains strict superadditivity of C_ID by tensoring this channel with a sufficiently large noiseless channel. For low-noise channels, the converse combines with Hayden-Winter achievability to yield C_ID = C_E.","tokens_in":20148,"tokens_out":34881,"duration_ms":301093,"significance":"If the proof is correct, this is a substantial advance: it settles an open question raised by Winter about whether C_ID ≥ C_E holds universally, and it does so in the negative direction, while providing the first characterization of identification capacity for a nontrivial class of quantum channels. The separation example is concrete and explicit, and the resulting strict superadditivity of C_ID is a new structural phenomenon. The proof strategy is also interesting in its own right: Lemma 3.2 establishes a clean connection between the geometric converse functional of [25] and the channel max-information, using the Bures metric, and the smoothing step is a natural way to convert one-shot geometric bounds into an asymptotic strong converse. The paper is also commendable for stating precise one-shot comparisons and for including a comparison with existing converse bounds. However, the central theorem is not self-contained: Theorem 2.10 and the key width estimate (2.51) are taken from an unreviewed companion preprint, and the proof of Theorem 3.3 applies that theorem to non-tensor-product smoothed channels, a case not covered by the theorem as stated in the present manuscript.","major_comments":[{"comment":"The diamond-norm smoothing step in the proof of Theorem 3.3 contains a factor-of-two error. From the bound (1/2)‖Ñ_n − N^{⊗n}‖⋄ ≤ ε one obtains d_Tr(ω_i, ω̃_i) ≤ ε, but for every effect 0 ≤ D ≤ 1 the inequality |Tr((ω̃_i − ω_i)D)| ≤ ‖ω̃_i − ω_i‖₁ ≤ 2ε holds, not ε. Consequently (3.30) and (3.31) should read 1 − λ₁ − 2ε and λ₂ + 2ε, respectively, and the packing gap becomes Δ − 4ε rather than Δ − 2ε. The correct choice is ζ = (Δ − 4ε)/2, which is positive under the stated assumption ε < Δ/4. The final rate statement is unaffected because ε → 0, but the displayed proof must be corrected.","section":"§3.1, Eqs. (3.30)–(3.31)"},{"comment":"The central bound C_ID(N) ≤ C_E(N) is obtained by applying Theorem 2.10 to the smoothed channel Ñ_n, which is an arbitrary CPTP map on A^{⊗n} → B^{⊗n}. However, Theorem 2.10 as stated is a theorem about the n-fold tensor power N^{⊗n} of a fixed single-letter channel N, and its right-hand side contains log d_A for the single-letter input dimension. The paper neither states nor proves the one-shot version needed in (3.33), in which the input dimension is d_A^n and the channel need not be a tensor product. Moreover, the load-bearing width estimate (2.51) is imported verbatim from the companion preprint [25] without proof; if that estimate has a hidden assumption, a wrong dimension factor, or a superlinear n dependence, the strong converse would not follow. I therefore ask that the authors either prove the required general one-shot Euclidean converse (including (2.51)) in an appendix, or reproduce the precise statement from [25] and verify that it covers arbitrary block channels. This is not a stylistic point; it is the load-bearing step of the main theorem.","section":"§3.1, Theorem 3.3; §2.3, Eq. (2.51)"}],"minor_comments":[{"comment":"The notation in Theorem 2.10 uses log d_A, which is natural for a single-letter channel, but the proof of Theorem 3.3 later uses the same symbol for the logarithm of the block input dimension d_A^n. The paper should clarify that the one-shot theorem is being used in its block-length-one form with input dimension d_A^n, or restate Theorem 2.10 in a form that covers arbitrary channels.","section":"§2.3, Eq. (2.52)"},{"comment":"The equality C_im(N) = max{C(N), Q_ID,v(N)} is justified by image-monotonicity of C and Q_ID,v, but the displayed argument does not explain why image containment I_1(M) ⊆ I_k(N) at single-block level implies the containment I_m(M) ⊆ I_{km}(N) needed for all tensor powers, especially when the inputs to the m blocks are entangled. This step should be spelled out or the claim weakened.","section":"§4, Theorem 4.2"},{"comment":"The notation g ∼ N(0, 1_m) for a standard Gaussian vector is nonstandard and slightly ambiguous; N(0, I_m) would be clearer.","section":"Definition 2.6"},{"comment":"The sentence 'Motivated by these discussions.' is a fragment and should be merged with the following sentence.","section":"Acknowledgements"}],"recommendation":"major_revision","confidential_remarks":"The main theorem depends on the author's own companion preprint [25], which is not publicly verified and is not reproduced here. If the editor's policy permits substantial dependence on an unpublished companion, the present paper's original contribution (Lemma 3.2, the smoothing argument, and the strict separation/superadditivity examples) is strong. I would encourage the editor to require that the authors make the dependence on [25] precise, either by including the necessary proof or by confirming that the specific one-shot theorem is stated and proved in [25] in exactly the form used here."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know: this paper proves C_ID(N) ≤ C_E(N) for every finite-dimensional quantum channel, which is the strong converse bound that has been missing. If the imported geometric bound is sound, it resolves Winter's question, gives exact equality for low-noise channels, and produces the first strict superadditivity example for C_ID. That is a big deal.\n\nWhat is genuinely new: Lemma 3.2 connecting the Euclidean converse functional µ* to channel max-information via the Bures metric is a clean, self-contained step. The transpose-depolarizing separation and the superadditivity corollary are neat consequences, and the paper is well written. The AI-assistance disclosure is honest and does not raise red flags.\n\nSoft spots, in proportion:\n\n1. The central theorem inherits Theorem 2.10 from the author's own preprint [25]. That theorem is not proved here, and the paper gives no independent check. The stress-test worry about non-tensor-product channels does not actually land: Theorem 2.10 is stated for arbitrary channels at blocklength one, so the application to the smoothed channel is legitimate as written. The real dependency is on the correctness of [25], especially the input-width estimate (2.51). A referee should ask for the companion proof or an independent verification of that bound.\n\n2. The factor-of-2 slip in the diamond-norm smoothing is real but minor. Equations (3.30)-(3.31) give error 2ε, not ε; choosing ε < Δ/4 still works because ε goes to 0, and the final theorem is unaffected.\n\n3. The ordering relative to the fully optimized unsmoothed Euclidean converse is left open, but that is a legitimate open question rather than a flaw.\n\nThe bottom line: this is a substantial paper that deserves a serious referee. The dependency on [25] should be explicitly flagged as a condition of acceptance, but the argument here is coherent and the new ideas are worth engaging. I would bring it to the reading group and would cite it once the companion preprint has been checked.\n\nRecommendation: send to peer review, with a request that the referee audit the import from [25].","headline":"A serious strong-converse result for identification capacity that deserves refereeing, with the main caveat being its dependence on an unproved companion bound.","tokens_in":20699,"tokens_out":2106,"would_cite":true,"duration_ms":20483,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["81P45","81P47","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that for every finite-dimensional quantum channel N, the identification capacity is bounded by the entanglement-assisted transmission capacity, $C_{\\mathrm{ID}}(N)\\leq C_E(N)$, with equality on low-noise channels and…","keywords":["identification capacity","entanglement-assisted classical capacity","quantum channels","strong converse bound","Gaussian mean width","channel max-information","transpose-depolarizing channel","superadditivity"],"falsifier":"Construct any identification code for a fixed finite-dimensional channel whose double-exponential rate exceeds $C_E(\\mathcal{N})$ by a fixed positive amount; the paper proves no such code exists, so one example would falsify Theorem 3.3. A more targeted check is to look for an ID code for the transpose-depolarizing channel at $q=d/(d+1)$ with rate above $\\log d - \\frac{d-1}{d}\\log(d+1)$, which Proposition 3.5 rules out.","tokens_in":19634,"feed_emoji":"📡","tokens_out":13756,"duration_ms":112309,"temperature":0.7,"pith_summary":"Identification asks the receiver only whether the sent message equals a queried one, so the number of identifiable messages can grow doubly exponentially with blocklength. The paper's central claim is that this doubly exponential rate is still capped by the entanglement-assisted transmission capacity: $C_{\\mathrm{ID}}(\\mathcal{N})\\le C_E(\\mathcal{N})$ for every finite-dimensional quantum channel $\\mathcal{N}$. That closes a previously open converse for identification and, for sufficiently low-noise channels, combines with known achievability to give the exact equality $C_{\\mathrm{ID}}(\\mathcal{N}) = C_E(\\mathcal{N})$. For general channels the bound can be strict, and the paper exhibits transpose-depolarizing channels where $C_{\\mathrm{ID}} < C_E$, leading also to strict superadditivity of the identification capacity.","feed_headline":"Identification capacity capped by entanglement-assisted rate","feed_subtitle":"Every quantum channel's ID rate is capped by its entanglement-assisted capacity; low-noise channels meet it.","key_machinery":"The workhorse is the Euclidean Gaussian mean width converse functional $\\mu^*(\\mathcal{N}) = \\inf_W K(W)^2 \\|Q_{\\mathcal{N},W}\\|_\\infty$, where $W$ ranges over Euclidean structures on the output Hermitian space, $K(W)$ is the trace-norm domination cost, and $Q_{\\mathcal{N},W}$ is the $W$-weighted singular operator of the channel. The new technical step is Lemma 3.2, which proves $\\mu^*(\\mathcal{N}) \\le 2^{I_{\\max}(\\mathcal{N})+1}$ by choosing $W$ to be the arithmetic-mean metric induced by a feasible state in the channel max-information optimization, so that the domination cost is one and the singular operator is controlled by the Choi operator inequality. This connects a purely geometric packing bound to the channel smooth max-information, whose asymptotic equipartition property produces the single-letter entanglement-assisted capacity.","core_discovery":"The paper establishes that the entanglement-assisted classical transmission capacity $C_E$ is a strong converse bound for classical identification over quantum channels: for every finite-dimensional channel $\\mathcal{N}$ and every admissible error pair, $\\limsup_{n\\to\\infty} (1/n)\\log\\log N^{(n,\\lambda_1,\\lambda_2)}(\\mathcal{N}) \\le C_E(\\mathcal{N})$. The proof converts any ID code into a packing of the channel's output image, applies the Euclidean Gaussian-mean-width converse to a slightly perturbed copy of the $n$-fold channel, bounds the resulting geometric functional by the channel's smooth max-information, and takes the blocklength limit through the asymptotic equipartition property. For low-noise channels, the known quantum-identification achievability result matches this upper bound, giving $C_{\\mathrm{ID}} = C_E$. The transpose-depolarizing channel at $q=d/(d+1)$ provides strict separation, with $C_{\\mathrm{ID}} \\le \\log d - \\frac{d-1}{d}\\log(d+1) < C_E = 1 + \\log\\frac{d}{d+1}$, and as $d$ grows the gap widens because $C_{\\mathrm{ID}}$ tends to zero while $C_E$ tends to one. The same example yields strict superadditivity of $C_{\\mathrm{ID}}$ when the channel is tensored with a sufficiently large noiseless channel.","pith_inferences":["Extension: because identification capacity depends only on the channel's output image, the image-optimized upper bound $\\bar{C}_{\\mathrm{im}}$ defined in the paper may be tight in regimes where $C_E$ is loose; computing it for specific channels is left open.","Testable extension: Lemma 3.2 shows the arithmetic-mean metric is optimal among monotone metrics, but not among all Euclidean structures; a numerical search for a non-monotone $W$ with $\\mu^*(\\mathcal{N}) < 2^{I_{\\max}(\\mathcal{N})+1}$ would probe whether the smoothing step can be sharpened.","Implicit consequence: since channels with the same output image can have different $C_E$ values, the strict gap identifies a genuinely non-transmission component of identification capacity, so entanglement assistance alone does not account for the power of identification."],"forward_implications":["For every finite-dimensional quantum channel, any identification code has double-exponential rate at most the entanglement-assisted transmission capacity, closing the missing universal converse.","On low-noise channels, identification capacity equals entanglement-assisted capacity: $C_{\\mathrm{ID}} = Q_{\\mathrm{ID},v} = Q_{\\mathrm{ID}} = C_E$.","The bound is not tight in general: for the transpose-depolarizing channel at $q=d/(d+1)$, $C_{\\mathrm{ID}} < C_E$, and the gap grows with dimension, with $C_{\\mathrm{ID}}$ tending to zero while $C_E$ tends to one.","Identification capacity is strictly superadditive: for $m>(d+1)/2$, $C_{\\mathrm{ID}}(D^T_{d/(d+1)}\\otimes \\mathrm{id}_m) > C_{\\mathrm{ID}}(D^T_{d/(d+1)}) + C_{\\mathrm{ID}}(\\mathrm{id}_m)$.","The $C_E$ converse dominates the previously known capacity-based and state-weighted Gaussian converse bounds, as illustrated for the qubit depolarizing channel."],"supporting_citations":[{"why":"Supplies the Euclidean Gaussian mean width converse (Theorem 2.10) that is the starting inequality of the C_ID ≤ C_E proof.","marker":"[25]"},{"why":"Provides the channel smooth max-information and its asymptotic equipartition property, converting the one-shot geometric bound into the limit C_E.","marker":"[13]"},{"why":"Gives the quantum identification capacity theorem and weak-decoupling achievability used to reach equality on low-noise channels.","marker":"[14]"},{"why":"Defines low-noise channels and records the C_ID ≥ C_E question that the paper resolves negatively in general.","marker":"[31]"},{"why":"Supplies quantum fingerprinting, the construction that lifts quantum identification codes to classical identification codes.","marker":"[10]"},{"why":"Introduces identification codes and the subset-lifting construction that yields the baseline C_ID ≥ C.","marker":"[1]"},{"why":"Establishes the entanglement-assisted classical capacity formula C_E = max I(R:B) that Theorem 3.3 targets.","marker":"[6]"},{"why":"Provides the additivity of C_E and the standard capacity formulas used to evaluate the depolarizing and transpose-depolarizing examples.","marker":"[29]"}],"fun_headline_variants":["Entanglement-assisted capacity caps identification rate","Identification capacity cannot exceed entanglement-assisted rate","Strong converse: ID rate ≤ entanglement-assisted rate","Some channels: identification capacity strictly below entanglement-assisted"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The universal bound inherits the validity of the Gaussian mean width converse from the author's earlier paper, which is cited but not proved here, and also the asymptotic equipartition property for smooth channel max-information; if either fails for the smoothed channels used in the proof, the conclusion $C_{\\mathrm{ID}}\\le C_E$ would not be established.","fun_headline_variants_meta":{"raw":{"variants":["Entanglement-assisted capacity caps identification rate","Identification capacity cannot exceed entanglement-assisted rate","Strong converse: ID rate ≤ entanglement-assisted rate","Some channels: identification capacity strictly below entanglement-assisted"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000392,"raw_usage":{"total_tokens":2109,"prompt_tokens":1045,"completion_tokens":1064,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":1009}},"tokens_in":661,"tokens_out":1064,"duration_ms":9670,"temperature":1.0,"reasoning_tokens":1009,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:25:50.411381+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct any identification code for a fixed finite-dimensional channel whose double-exponential rate exceeds $C_E(\\mathcal{N})$ by a fixed positive amount; the paper proves no such code exists, so one example would falsify Theorem 3.3. A more targeted check is to look for an ID code for the transpose-depolarizing channel at $q=d/(d+1)$ with rate above $\\log d - \\frac{d-1}{d}\\log(d+1)$, which Proposition 3.5 rules out.","supporting_citations":[{"cited_title":"Gaussian mean width strong converse bound on the classical identification capacity of quantum channels","cited_arxiv_id":"2606.05032","evidence_quote":"Supplies the Euclidean Gaussian mean width converse (Theorem 2.10) that is the starting inequality of the C_ID ≤ C_E proof."},{"cited_title":"Quantum Channel Simulation and the Channel's Smooth Max-Information","cited_arxiv_id":"1807.05354","evidence_quote":"Provides the channel smooth max-information and its asymptotic equipartition property, converting the one-shot geometric bound into the limit C_E."},{"cited_title":"Weak decoupling duality and quantum identifi- cation.IEEE Transactions on Information Theory, 58(7):4914–4929, July 2012","cited_arxiv_id":null,"evidence_quote":"Gives the quantum identification capacity theorem and weak-decoupling achievability used to reach equality on low-noise channels."}],"review_version":1}