{"id":"6368fe73-9be7-44bf-ba00-f5ae4a27fed1","arxiv_id":"2506.16581","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The covert capacity region of physically-degraded alarm two-way channels is exactly a region in which secret coordination between users costs nothing asymptotically.","lead":"Two users can exchange covert messages over a two-way alarm channel, where simultaneous transmission would alert an eavesdropper. The paper shows secret coordination between the users strictly improves covert throughput, and this coordination costs nothing asymptotically.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Converse key lemma 6.13 uses an unjustified replacement of per-symbol weights by time-averaged weights in its error term; covertness alone does not force uniform per-symbol vanishing, though an L2 bound from (69) may repair the argument.","rationale":"The reader correctly identified that Lemma 6.13's per-symbol smallness assumption is not a consequence of the covertness constraint. The most load-bearing aspect is more precise: Appendix H replaces a sum of squared per-symbol weights by n times the squared time-averaged weight, and this replacement is invalid in the wrong direction, so the proof of the outer bound is incomplete as written. However, the threat is repairable. Equation (69) supplies a per-symbol sum constraint, and for alarm channels the per-symbol relative entropy is quadratic in the non-innocent weights, giving sum_i(mu_i)^2 = O(1). With that correction, the error term is O(1), which is negligible after dividing by sqrt(n), and the outer bound goes through. I also checked the block-Markov accounting: if M1 is per-block and the total message count is M1^B, the factor B cancels in the throughput normalization, so the displayed rate region is not endangered by the notational ambiguity. Lemma 6.8 is omitted but its role is analogous to a known result in [37]; it is a completeness issue rather than the main correctness risk. Because the identified gap is concrete but plausibly patchable without changing the result, the appropriate disposition remains CONDITIONAL, i.e., the reader's verdict is unchanged.","tokens_in":48740,"tokens_out":23427,"duration_ms":245638,"concrete_test":"Independently re-derive Appendix H with the error term written as O(sum_i (mu_{10}^{(n,i)})^2). Then prove from (69) that sum_i (mu_{10}^{(n,i)})^2 = O(1) using the lower bound D(bQ_i||Q00) >= c(mu_{10,i}^2+mu_{01,i}^2), valid for alarm channels. Separately construct a sequence with mu_1^{(n)} = 1 and mu_i^{(n)} = 1/sqrt(n) for i>1, satisfying D(bQ^n||Q00^n) <= delta; verify numerically for a concrete alarm channel that the stated O(n(mu_10)^2) bound is false in that case, while the corrected O(sum_i mu_i^2) bound holds. If the corrected derivation yields (74) with O(1) error, the converse stands; if not, the outer bound needs a genuinely new argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 5.2 depends on the converse Proposition 6.10, whose crucial step is Lemma 6.13. That lemma assumes mu_kl^{(n,i)} -> 0 uniformly in i and concludes sum_i I(X_{1i};Y_{2i}|X_{2i}) <= n mu_n [ ... ] + O(n(mu_10)^2). In Appendix H, the error is obtained by Taylor-expanding a per-symbol expression (around equations (198)-(200)), producing O((mu_10^{(n,i)})^2), and then replacing sum_i (mu_10^{(n,i)})^2 by n(mu_10)^2. This replacement is not justified: by Cauchy one has sum_i mu_i^2 >= n mu_bar^2, so the stated n(mu_bar)^2 can be an underestimate when the non-innocent mass is concentrated in a few coordinates. A code with one coordinate at mu=1 and the remaining mass spread thinly satisfies D(bQ^n||Q00^n) <= delta but violates uniform per-symbol vanishing; the Taylor step fails on that coordinate. The concern is not fatal because equation (69) gives sum_i D(bQ_i^n||Q00) <= delta, and for alarm channels D(bQ_i||Q00) >= c(mu_{10,i}^2+mu_{01,i}^2) with c>0, forcing sum_i(mu_{10,i})^2 = O(1). With the error written as O(sum_i(mu_i)^2) instead of O(n(mu_bar)^2), the conclusion of Lemma 6.13 survives and the outer bound follows. As written, however, the proof contains a real gap: the uniform per-symbol hypothesis is neither derived from the covertness constraint nor replaced by the correct L2 control.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies covert communication over binary-input discrete memoryless two-way channels in which simultaneous non-innocent transmissions by both users trigger an alarm at the eavesdropper. It defines a covert throughput region with the usual sqrt(n) normalization and then proves two main results for physically-degraded alarm two-way channels: Theorem 5.1 characterizes the covert capacity region under public time sharing, and Theorem 5.2 characterizes the full covert capacity region as a union over a time-sharing parameter lambda of bounds involving the chi-square divergence of the eavesdropper mixture and the relative entropies of the legitimate channel outputs. The achievability proof uses a sparse time-sharing distribution with an auxiliary random variable whose probability scales as n^{-1/4}, together with a block-Markov scheme in which secret message parts generated in one block serve as coordination information for the next block. The converse is a specialization of a general outer bound for arbitrary discrete memoryless two-way channels. The paper also claims that the coordination overhead vanishes asymptotically and that cooperation strictly enlarges the public time-sharing region.","tokens_in":49103,"tokens_out":14245,"duration_ms":144068,"significance":"If the technical gaps identified below are closed, this is a valuable contribution to multi-user covert communications. The main result provides the first covert capacity region for a nontrivial two-way channel model and identifies a concrete mechanism by which user cooperation strictly improves covert throughputs while the coordination cost vanishes asymptotically. The paper's analytical strengths include explicit scaling analyses for the auxiliary random variables, a detailed block-Markov achievability scheme, a general converse bound that is of independent interest, and a numerical illustration of the capacity region. The claimed result is falsifiable and the proof is largely self-contained, building on established resolvability and covert-communication tools rather than introducing ad hoc assumptions. The main reservations concern three load-bearing points in the proofs: an unjustified per-symbol weight assumption in the converse, an algebraic inconsistency in the public time-sharing proof, and an omitted proof of a key chaining lemma.","major_comments":[{"comment":"The hypothesis of Lemma 6.13 that the per-symbol non-innocent probabilities mu_{kl}^{(n,i)} vanish uniformly in i is not implied by the covertness constraint D(bQ^n||Q00^n) <= delta; that constraint only forces the time-averaged weight mu_n to be O(1/sqrt(n)). In Appendix H, the Taylor expansion around equations (198)-(200) produces an error O((mu_{10}^{(n,i)})^2), and this is replaced in (206) by O(n(mu_{10}^n)^2) with the time-averaged weight. This replacement is not justified: by Cauchy's inequality one has sum_i (mu_i)^2 >= n (mu_bar)^2, so the stated n(mu_bar)^2 can underestimate the true quadratic error when non-innocent mass is concentrated in a few coordinates. Since Lemma 6.13 is used to derive the outer bound (75) and hence Theorem 5.2, the converse as written has a gap. The gap appears repairable: equation (69) gives sum_i D(bQ_i^n||Q00) <= delta, and for alarm channels D(bQ_i||Q00) is bounded below by a positive constant times (mu_{10,i}^2 + mu_{01,i}^2), so sum_i mu_i^2 = O(1); writing the error as O(sum_i mu_i^2) instead of O(n(mu_bar)^2) would preserve the conclusion. The authors should either prove the uniform vanishing hypothesis or replace it with this L2 control.","section":"VI-C, Lemma 6.13 and Appendix H"},{"comment":"There is a blocklength error in the public time-sharing derivation. User 2 transmits during the second sub-block of length lambdabar n, not lambda n, but the denominators in Propositions 6.1 and 6.2 and in Eq. (32) use sqrt(lambda n (delta - delta_1)) for User 2. With the displayed (32), substituting delta_1 = lambda delta gives r2 <= sqrt(lambda(1-lambda)) times the anticipated constant, not (1-lambda) times it as claimed in Theorem 5.1. The same wrong factor appears in the Lagrangian in Lemma 6.3. The proof of Theorem 5.1 is therefore algebraically inconsistent unless the lambda in the second sub-block is corrected to lambdabar throughout; after that correction, the substitution delta_1 = lambda delta yields the stated region.","section":"VI-A, Propositions 6.1 and 6.2, Eq. (32), Lemma 6.3"},{"comment":"Lemma 6.8 is load-bearing for the achievability proof: it converts per-block resolvability and reliability guarantees into a bound on the total relative entropy D(bQ^{nB}||Q00^{nB}) across the chained blocks, including the effect of estimation errors of the secret messages. The proof is omitted with the note that it is similar to [37]. Given that the present scheme has a two-way block-Markov structure and a different covert process, this is not a purely notational adaptation. The authors should provide the proof or state precisely which result in [37] applies and how the two-way coupling and the estimates of the secret messages are handled.","section":"VI-B, Lemma 6.8"}],"minor_comments":[{"comment":"The displayed inequalities for r2 use lambda instead of lambdabar: the second and fourth displayed bounds should read (1-mu)c_lambda lambdabar D(P_{01}^{(1)}||P_{00}^{(1)}) and (1+mu)c_lambda lambdabar D(Q01||Q00), consistent with equations (60) and (61).","section":"Proposition 6.5"},{"comment":"The term D(Q01||P_{00}^{(2)}) should be D(Q01||Q00); the current notation appears to be a typo.","section":"Eq. (63)"},{"comment":"The specialization of the converse to alarm channels states r2 <= c_lambda lambdabar D(P_{10}^{(2)}||P_{00}^{(2)}); the last factor should be D(P_{01}^{(1)}||P_{00}^{(1)}).","section":"End of Section VI-C"},{"comment":"The mutual information in the sum is written as I(X_{1i};Y_{2i}|X_{1i}); it should be I(X_{1i};Y_{2i}|X_{2i}).","section":"Lemma 6.13 statement"}],"recommendation":"major_revision","confidential_remarks":"The main theorem is likely correct and the identified gaps appear repairable within the scope of the manuscript: the converse gap can be fixed by using the L2 bound obtained from (69), the public time-sharing proof only requires replacing lambda by lambdabar in the second sub-block, and Lemma 6.8 needs a proof or a precise citation. I therefore recommend major revision rather than rejection. The self-citations to [4] and [37] are used as technical tools and do not raise a novelty-disclosure concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper because it opens the first crack in a hard problem: covert capacity for two-way channels. The authors introduce alarm two-way channels, where any simultaneous non-innocent transmission trips an eavesdropper alarm, and they give a capacity region for the physically-degraded case. The interesting phenomenon is that coordination between the two users comes for free in the covert limit: secret coordination enlarges the public time-sharing region without consuming extra covertness. That is a real insight, not a slogan.\n\nWhat is genuinely new is the sparse time-sharing scheme. Earlier multi-user covert results managed to avoid blocklength-dependent auxiliary random variables. Here the auxiliary U has probability mass scaling as n^{-1/4} while the conditional input distributions scale as n^{-1/4}, so that the overall weight is n^{-1/2}. This non-homogeneous scaling is what makes I(U;Z) decay as n^{-3/4}, fast enough for the random-coding concentration argument. The achievability proof is detailed and, as far as I checked, plausible. The general converse in Proposition 6.10 is also a useful stepping stone, even though it is not tight for arbitrary two-way channels.\n\nThe soft spots are real but not fatal. Lemma 6.8, which bounds the end-to-end relative entropy by a sum of per-block terms, is stated without proof; the authors say it follows like [37], and I believe that, but a referee will want it written out. More serious is Lemma 6.13. The lemma assumes that each per-symbol non-innocent weight mu_kl^{(n,i)} goes to zero uniformly in i. The covertness constraint alone only gives a time-averaged weight of order 1/sqrt(n); it does not forbid concentrating all non-innocent uses in a vanishing fraction of positions. The error term in equation (74) is written as O(n(mu_10)^2), which is an underestimate if the mass is concentrated, because Cauchy's inequality gives sum_i mu_i^2 >= n (mu_bar)^2, not the reverse. The stress-test note correctly points out that equation (69) gives sum_i D(bQ_i||Q00) <= delta, and for alarm channels this forces sum_i mu_i^2 = O(1), so writing the error as O(sum_i mu_i^2) repairs the proof. As written, the gap is there, but it is a patchable gap, not a flawed theorem.\n\nThe block Markov rate accounting is a bit loose around the factor B, but I do not see a hidden factor that changes the region. Overall: the result is probably right, the model is new, and the paper deserves a serious referee. It should not be desk-rejected, but it needs a revision that fixes Lemma 6.13 and supplies the missing proof of Lemma 6.8. I would bring it to reading group and would consider citing it once the converse is airtight.","headline":"First covert capacity region for two-way channels, with a genuinely new sparse time-sharing mechanism, but the converse contains a repairable gap in Lemma 6.13.","tokens_in":49638,"tokens_out":1966,"would_cite":true,"duration_ms":22008,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A40"],"pacs":[],"model":"deepseek-v4-flash","headline":"For physically-degraded alarm two-way channels, the paper proves an exact covert capacity region in which secret coordination strictly enlarges what public time-sharing can achieve, with coordination overhead vanishing asymptotically.","keywords":["covert communications","two-way channels","alarm channels","square root law","channel resolvability","capacity region","block Markov coding","low probability of detection"],"falsifier":"Construct a binary alarm two-way code that places all non-innocent inputs in the first $n^{3/4}$ coordinates, each with non-innocent probability $n^{-1/4}$, tuned so the eavesdropper's total relative entropy stays at $\\delta$. If such a code achieves throughputs outside the region of Theorem 5.2, the uniform per-symbol vanishing assumption in Lemma 6.13 fails; if no such code can, the converse's uniformity condition is not the load-bearing gap.","tokens_in":48532,"feed_emoji":"📡","tokens_out":5888,"duration_ms":63142,"temperature":0.7,"pith_summary":"This paper characterizes how much information two users can exchange covertly over a two-way channel in which simultaneous transmissions trigger an alarm at the eavesdropper. It proves that for physically-degraded alarm two-way channels the covert capacity region is the union over $\\lambda\\in[0,1]$ of the throughput pairs bounded by $r_1\\le \\lambda c_\\lambda D(P_{10}^{(2)}\\|P_{00}^{(2)})$ and $r_2\\le \\bar\\lambda c_\\lambda D(P_{01}^{(1)}\\|P_{00}^{(1)})$, where $c_\\lambda=\\sqrt{2/\\chi^2(\\lambda Q_{10}+\\bar\\lambda Q_{01}\\|Q_{00})}$. This region is strictly larger than the public time-sharing region, so secret coordination genuinely improves covert throughput. The coordination itself is generated from the covert traffic, which is why the paper finds that the coordination cost comes asymptotically for free.","feed_headline":"Coordinated covert two-way links beat public time-sharing","feed_subtitle":"For alarm channels, the exact covert capacity region shows secret cooperation is asymptotically free.","key_machinery":"The load-bearing object is the sparse time-sharing auxiliary random variable $U$, with $P_U(0)=1-(q_1+q_2)n^{-1/4}$, $P_U(1)=q_1n^{-1/4}$, $P_U(2)=q_2n^{-1/4}$, and conditional codeword distributions scaled by $n^{-1/4}$. This distribution makes the eavesdropper's induced output differ from $Q_{00}$ at order $n^{-1/2}$, while $I(U;Z)$ decays only as $n^{-3/4}$: fast enough for the resolvability-style concentration arguments to work, but not so fast that the random-coding bounds become trivial. The auxiliary variable indicates which user is allowed to deviate from the innocent symbol, and it is carried across blocks by block-Markov chaining: the secret parts of both users' messages decoded in block $b$ become the common coordination message in block $b+1$. That chaining is the mechanism behind the asymptotic freeness of coordination.","core_discovery":"The central claim is Theorem 5.2: for a binary-input physically-degraded alarm two-way channel, the covert capacity region is exactly the union over $\\lambda\\in[0,1]$ of the sets of throughput pairs satisfying the two inequalities with the common factor $c_\\lambda=\\sqrt{2/\\chi^2(\\lambda Q_{10}+\\bar\\lambda Q_{01}\\|Q_{00})}$. The factor $c_\\lambda$ captures how the total covertness budget is split between the two users' non-innocent symbols, while $\\lambda$ and $\\bar\\lambda$ split the blocklength between the two directions. Achievability is shown by a block-Markov coding scheme with sparse time-sharing, and the matching converse is obtained by specializing a general outer bound for arbitrary discrete memoryless two-way channels to the alarm structure, where covertness forces $\\rho_{11}=0$. The paper also observes that the physical-degradation inequalities are not needed for the achievability of this region, so a user who individually has no channel advantage over the eavesdropper can still communicate covertly with the partner's help.","pith_inferences":[],"forward_implications":["The covert capacity region $C$ is strictly larger than the public time-sharing region $C_{\\mathrm{PTS}}$ for the same alarm channel, so secret coordination is not just a convenience but a genuine throughput gain.","A user with no individual relative-entropy advantage over the eavesdropper can still achieve positive covert throughput, because the partner's coordinated activity helps hide the transmission.","The coordination message size grows only as $O(n^{1/4})$ in the exponent, whereas message throughput grows as $O(\\sqrt{n})$, so the coordination overhead vanishes asymptotically relative to the covert payload.","The general converse Proposition 6.10 provides a single-letter outer bound for any discrete memoryless two-way channel, and it is tight for physically-degraded alarm channels after imposing $\\rho_{11}=0$.","The alarm-channel analysis shows that the square-root law persists under coordination, with the exact throughput governed by a chi-square divergence of the mixture $\\lambda Q_{10}+\\bar\\lambda Q_{01}$ against $Q_{00}$.","The paper's own analysis leaves open whether the full two-way covert capacity for non-alarm or non-degraded channels is also governed by a similar ratio-symmetric chi-square trade-off; the general converse is not tight in those cases.","A testable extension suggested by the proof is a code that concentrates all non-innocent inputs in a vanishing fraction of coordinates: if such a code can satisfy the average covertness constraint while exceeding Theorem 5.2, the uniform per-symbol vanishing assumption in the converse would be the reason the outer bound is not fully general."],"supporting_citations":[{"why":"Supplies the point-to-point covert capacity result and the covert-process/resolvability machinery that both the public time-sharing and the sparse time-sharing schemes build on.","marker":"[4]"},{"why":"Establishes the square-root law and the relative-entropy scaling of the eavesdropper's detection constraint used throughout the paper.","marker":"[3]"},{"why":"Provides the block-Markov chaining of secret coordination messages and the resolvability chaining analysis that the two-way coding scheme adapts.","marker":"[37]"},{"why":"Gives the multi-user covert communication setting and the public time-sharing baseline against which the two-way cooperation gain is measured.","marker":"[13]"},{"why":"Supplies the optimal covertness budget split $\\delta_1(\\lambda)=\\lambda\\delta$ used in the public time-sharing converse and achievability.","marker":"[20]"},{"why":"Converts channel resolvability into strong secrecy of the secret message parts, which is what lets the coordination message stay hidden from the eavesdropper.","marker":"[41]"},{"why":"Demonstrates how cooperative mechanisms can enable covert communication even when an individual user has no channel advantage over the eavesdropper, a phenomenon the paper identifies here as well.","marker":"[22]"}],"fun_headline_variants":["Exact covert capacity for degraded alarm two-way channels","Alarm channels: covert cooperation comes asymptotically free","Two-way covert capacity tight for physically-degraded alarms","Cooperative covert communication solves alarm channel capacity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The converse's per-symbol analysis assumes every coordinate's non-innocent probability vanishes uniformly, while the covertness constraint only forces this on average over coordinates; a code that concentrates non-innocent symbols in a few coordinates could escape the bound.","fun_headline_variants_meta":{"raw":{"variants":["Exact covert capacity for degraded alarm two-way channels","Alarm channels: covert cooperation comes asymptotically free","Two-way covert capacity tight for physically-degraded alarms","Cooperative covert communication solves alarm channel capacity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1512,"prompt_tokens":949,"completion_tokens":563,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":501}},"tokens_in":565,"tokens_out":563,"duration_ms":5490,"temperature":1.0,"reasoning_tokens":501,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:24:10.001768+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a binary alarm two-way code that places all non-innocent inputs in the first $n^{3/4}$ coordinates, each with non-innocent probability $n^{-1/4}$, tuned so the eavesdropper's total relative entropy stays at $\\delta$. If such a code achieves throughputs outside the region of Theorem 5.2, the uniform per-symbol vanishing assumption in Lemma 6.13 fails; if no such code can, the converse's uniformity condition is not the load-bearing gap.","supporting_citations":[{"cited_title":"Cooperative resolvability and secrecy in the cribbing multiple-access channel,","cited_arxiv_id":null,"evidence_quote":"Provides the block-Markov chaining of secret coordination messages and the resolvability chaining analysis that the two-way coding scheme adapts."},{"cited_title":"Covert communication over a k-user multiple access channel,","cited_arxiv_id":null,"evidence_quote":"Gives the multi-user covert communication setting and the public time-sharing baseline against which the two-way cooperation gain is measured."},{"cited_title":"Time-division is optimal for covert communication over some broadcast channels,","cited_arxiv_id":null,"evidence_quote":"Supplies the optimal covertness budget split $\\delta_1(\\lambda)=\\lambda\\delta$ used in the public time-sharing converse and achievability."},{"cited_title":"Strong secrecy from channel resolvability,","cited_arxiv_id":null,"evidence_quote":"Converts channel resolvability into strong secrecy of the secret message parts, which is what lets the coordination message stay hidden from the eavesdropper."},{"cited_title":"Keyless covert communication via channel state information,","cited_arxiv_id":null,"evidence_quote":"Demonstrates how cooperative mechanisms can enable covert communication even when an individual user has no channel advantage over the eavesdropper, a phenomenon the paper identifies here as well."}],"review_version":2}