{"id":"8637f8c8-344f-47f3-b30e-4596df477df8","arxiv_id":"2602.11568","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Non-signaling assistance makes causal and non-causal channel-state knowledge at the transmitter have the same capacity, max_{P_{X|S}} I(X;Y|S), though the best finite-blocklength error rate still improves when the receiver also sees the state.","lead":"This paper shows that when classical channels with random states are helped by non-signaling resources (shared correlations that cannot by themselves transmit information), letting the transmitter know the state in real time gives the same capacity as letting it know the whole state sequence in advance. The result pins down an open capacity formula and reveals a finite-blocklength twist: giving the receiver the state data still improves the best error rate under causal state","discovery_kind":"unification","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: Theorem 1 is internally consistent; the C1–C3 model is the least explored premise but no concrete counterexample exists.","rationale":"The reader's weakest assumption was the completeness of the C1–C3 model. I agree that this is the least formally justified premise, since the paper does not prove equivalence with the general sequential NS polytope. However, on close reading the conditions appear to be the correct time-ordered no-signaling constraints for the specified input/output order, and the achievability proof checks out: the authentication construction satisfies C1–C3, the rate analysis is consistent, and the apparent tension between uniform C2 marginals and high success probability is resolved by the channel-induced correlation between X and Y. The abstract's missing proof of the feedback/strictly-causal-CSIT result is a real issue for the paper's completeness but does not threaten the central capacity theorem. Thus no change to the reader's CONDITIONAL verdict is warranted by this stress-test pass.","tokens_in":21495,"tokens_out":34739,"duration_ms":355423,"concrete_test":"Use Fourier–Motzkin elimination to compute the projection of the full time-ordered NS polytope of Gallego et al. [15] for n=2 (inputs W,S1,S2; outputs X1,X2,W-hat; receiver input Y1,Y2) onto the variables Z(x^n,W-hat|w,s^n,y^n), and verify that the projected polytope equals the C1–C3 polytope of Section 2.4. Repeating for n=3 would further settle the model-completeness question; if the projections match, the weakest modeling premise is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"After tracing the achievability proof of Theorem 1, I find no concrete internal flaw. The natural weak point is the C1–C3 model of Section 2.4: if a sequential NS resource with internal memory could satisfy no-signaling but violate these three conditions, the capacity formula could understate the true NS-assisted capacity. However, C1–C3 are exactly the time-ordered no-signaling constraints for this input order: transmitter outputs X^n must be independent of receiver input Y^n (C1), receiver output W-hat must be independent of transmitter inputs (W,S^n) (C2), and prefixes must be independent of future states S_{i+1:n} (C3). Any sequential resource in the TONS framework induces a conditional distribution satisfying these constraints; the converse inclusion is the only unproven direction, and no counterexample is given. The authentication construction verifies C1–C3 directly, and the success-probability analysis is self-consistent: the marginal Z(W-hat|w,s^n,y^n) being uniform does not preclude high success because the channel induces correlation between X and Y that is invisible to the no-signaling test. The abstract's unproved separate claim about feedback plus strictly-causal CSIT is a genuine deficiency, but it is not needed for Theorem 1.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the non-signaling (NS) assisted capacity of a classical discrete memoryless channel with state and causal channel state information at the transmitter (CSIT). Its main result, Theorem 1, states that this capacity equals max_{P_{X|S}} I(X;Y|S). Because the same expression was previously established for NS-assisted non-causal CSIT in [13], the theorem implies that, under NS assistance, causal CSIT is as good as non-causal CSIT, and both coincide with the classical capacity when the state is also known at the receiver. The achievability proof constructs an explicit 'authentication' NS box satisfying conditions C1–C3: Algorithm 1 causally transforms the state sequence into a fixed-type sequence, the transmitter samples X_i from P_{X|\\tilde S_i}, and the decoder performs a joint-typicality authentication before outputting the message; C2 is enforced by choosing M=ceil(μ) and λ=μ/M. The success probability is then shown to tend to 1. Theorem 2 uses the Z0/Z1 channel and a linear-programming dual to show that, for M=2, n=2, CSIR strictly improves the optimal success probability over causal CSIT alone (≥7/8 vs ≤13/16). Appendix B gives a toy example. The abstract additionally claims a separate result on the combined effect of NS assistance, feedback, and strictly causal CSIT, but no statement or proof of this claim appears in the body.","tokens_in":21810,"tokens_out":17985,"duration_ms":168753,"significance":"If the results are correct, they complete a natural gap in the NS-assisted capacity literature: the state-dependent point-to-point channel with causal CSIT. The equality of causal and non-causal CSIT under NS assistance is a clean and somewhat surprising structural result, and the finite-blocklength contrast with the non-causal setting is a useful refinement of the 'virtual teleportation' idea. The paper has notable strengths: the achievability scheme is constructive rather than existential; the C1–C3 model is explicitly stated; the LP reduction in Appendix A is checkable; and the dual feasible point, apart from a subscript typo, gives a verifiable numeric bound. These features make the core of the paper credible and reproducible. The main caveat is that an advertised separate result in the abstract is missing from the body, and one step of the success-probability analysis is too terse for a load-bearing claim.","major_comments":[{"comment":"The abstract states: 'As a separate result we prove that non-signaling assistance, feedback, and strictly causal CSIT ... cannot increase capacity when they are collectively made available to the transmitter.' No theorem, lemma, or proof of this claim appears anywhere in Sections 1–5 or the appendices. This is not a minor omission: it is an advertised contribution. The authors must either include a precise statement with proof, cite a published proof, or delete/qualify the claim so that the abstract matches the content of the paper.","section":"Abstract, final sentence"},{"comment":"The proof that η(Z)→1 is incomplete at the point where, after conditioning on F=1, it is asserted that the last factor in (57) 'approaches 1 as \\tilde n_σ→∞'. The set \\tilde I_σ is data-dependent: it consists of positions where the y-Algorithm did not output φ, and its composition is exactly the budget counts, not an i.i.d. sample. One must justify that for every y^{\\tilde I_σ} with that fixed empirical type, the joint typicality probability with X_{ \\tilde I_σ} tends to 1 uniformly, with a tolerance chosen to accommodate the boundary value (1-ε)P_Y(y). This is likely fixable with a standard typicality argument, but as written the conditioning step is too terse for a central claim. Please expand the argument and make the tolerance bookkeeping explicit.","section":"Section 4.3, Eqs. (52)–(57)"},{"comment":"Related to the previous comment: the text says that conditioned on flag[S^n]=1, '˜S^n is unchanged from S^n at the positions indexed by I_σ'. This is true for positions in I_σ, but flag[S^n]=1 does not imply \\tilde S^n=S^n everywhere; positions where the budget for a symbol was exceeded are changed to φ in the y/s algorithm (see the first example in Fig. 3). The proof uses only the weaker statement, but the wording is likely to mislead. Please restate precisely what F=1 gives for each σ and for the positions in I_σ and \\tilde I_σ, and distinguish it from an identity of the whole sequences.","section":"Section 4.3, around Eq. (57)"}],"minor_comments":[{"comment":"The displayed λ vector has entries with six subscripts, e.g. λ_{0,0,0,1,0,0}, while λ is defined with four subscripts (y1,y2,s1,s2) in Eq. (81). This appears to be a typographical error; the intended entries are λ_{0,0,0,1}, λ_{0,0,1,0}, λ_{0,1,0,1}, λ_{1,0,0,1}, λ_{1,0,1,0}, λ_{1,1,1,0} with the listed values. Please correct the indexing so the dual feasible point can be checked without guesswork.","section":"Appendix A, Eq. (85)"},{"comment":"O2 says 'the capacity gain from NS assistance is evident, as there is a gap between the classical capacity and the NS assisted capacity in both cases.' This is not true for every channel with state (e.g., a state-free channel or a state that does not affect the channel); the intended statement is that there can be an unbounded gap. Please rephrase as 'can be a gap'.","section":"Section 3, observation O2"},{"comment":"The budgets t_α=floor(n(1-ε)P_A(α)) make the output type have counts exactly at the boundary of the strong typicality set with tolerance ε. This is internally consistent, but the paper should state explicitly that the tolerance in the joint typicality arguments may need to be enlarged slightly (e.g., to 2ε) so that the deterministic type produced by Algorithm 1 lies inside the relevant typical set for all large n.","section":"Section 4.1, Algorithm 1"},{"comment":"The abstract and conclusion both describe the missing 'separate result' on feedback and strictly causal CSIT. Even if the claim is to be removed, both places need to be edited consistently. Also, if the result is retained, the relevant references for the individual capacity claims (non-signaling assistance, feedback, strictly causal CSIT) should be cited.","section":"Abstract and Section 5"}],"recommendation":"major_revision","confidential_remarks":"The core results, Theorem 1 and Theorem 2, appear sound: I traced the C1–C3 verification, the LP reduction, and spot-checked constraints of the dual feasible point. The main problems are (i) the abstract advertises a result not present in the manuscript, and (ii) the success-probability proof has a terse conditioning step that needs a rigorous typicality argument. Both are fixable within the scope of a revision. I would not require another full review if the missing abstract claim is removed or properly proved and the proof gap is filled."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main result is real and the proof is substantive. The paper shows that NS-assisted capacity with causal CSIT equals max_{P_{X|S}} I(X;Y|S), matching the non-causal NS-assisted and CSIR capacities. That resolves an open case and collapses a distinction that matters classically. The causal type-stabilizing algorithm is new, the authentication construction is carefully built, and the permutation argument for C2/C3 is explicit. The rate analysis and the success-probability argument are consistent. Theorem 2 also delivers a clean finite-blocklength separation (13/16 vs 7/8) via an LP dual, and the listed dual point is algebraically checkable. This is a real contribution, not a restatement of prior work.\n\nNow the soft spots. The abstract claims a separate result: NS assistance, feedback, and strictly-causal CSIT collectively cannot increase capacity. I could not find a proof of that anywhere in the body. That is a genuine mismatch. Either prove it or cut it from the abstract. Second, the LP in Appendix A has an indexing typo in the lambda vector (six-tuples with an extra coordinate), and the feasibility verification is deferred with an \"it can be verified.\" Spot checks pass, so this is minor, but it should be cleaned up. Third, the C1–C3 model is the least defended premise. The authors assume any sequential NS resource induces a conditional distribution satisfying those three constraints; the converse inclusion is not proved. That is a modeling assumption in the TONS framework, and no counterexample is offered, so I would flag it as a point for the authors to make explicit rather than as a flaw that undermines the argument.\n\nOverall, the central theorem appears correct under the stated model. The deficiencies are fixable and not load-bearing. I would send this to peer review, asking the authors to deal with the abstract claim and the appendix hygiene. For anyone working on NS-assisted information theory, this is worth reading and worth citing.\n\nYes to serious refereeing.","headline":"Solid resolution of an open causal-CSIT case, with one unproved claim in the abstract that should not ship as-is.","tokens_in":22278,"tokens_out":2206,"would_cite":true,"duration_ms":24047,"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 any channel with state, non-signaling assistance lets causal channel-state information achieve the same capacity as non-causal channel-state information: max_{P_{X|S}} I(X;Y|S).","keywords":["non-signaling assistance","causal CSIT","channel with state","capacity","non-causal CSIT","finite blocklength probability of success","non-signaling correlations","CSIR"],"falsifier":"Solve the fully specified linear program LP1/LP2 for the Z0/Z1 channel with M=2, n=2: any violation of the stated optimum, or any non-signaling-assisted causal scheme whose success probability exceeds 13/16, would refute Theorem 2. For Theorem 1, exhibit a channel with state and any causal non-signaling-assisted scheme achieving a rate greater than max_{P_{X|S}} I(X;Y|S); the theorem's converse says this cannot happen.","tokens_in":1400,"feed_emoji":"📡","tokens_out":1550,"duration_ms":74150,"temperature":0.7,"pith_summary":"This paper establishes that for any discrete memoryless channel with state, if the transmitter and receiver may share arbitrary non-signaling resources, causal channel-state information suffices to achieve the same rate as non-causal channel-state information: max_{P_{X|S}} I(X;Y|S). This value is also the capacity when the state is available to the receiver, with or without non-signaling assistance, so the result unifies four capacity settings. At finite blocklength, however, the equivalence breaks: for a simple Z0/Z1 channel with two uses and a one-bit message, giving the state to the receiver strictly improves the optimal success probability beyond what non-signaling-assisted causal channel-state information alone can achieve. The paper also claims that non-signaling assistance, feedback, and strictly causal channel-state information, each individually unable to increase capacity, cannot increase capacity even when made jointly available.","feed_headline":"Causal CSIT matches non-causal CSIT when non-signaling help is allowed","feed_subtitle":"Shared non-signaling aids make future state info unnecessary for rate; short blocks still favor receiver state info.","key_machinery":"The paper models a non-signaling-assisted causal coding scheme as a sequential non-signaling correlation, specified by three conditions: C1 says the transmitter's inputs do not reveal the receiver's outputs; C2 says the receiver's estimate does not reveal the message or state without the channel; C3 says the transmitter's partial input does not reveal future channel states. The achievability construction centers on Algorithm 1, a causal type-fixing map that turns every state sequence into one with a fixed type, plus an authentication step that checks joint typicality of the input-output pairs against the fixed-type state sequence. The authentication solution makes the scheme's internal proba","core_discovery":"The main result, Theorem 1, is that with causal CSIT the non-signaling-assisted capacity equals C^{NS,ca}(N,P_S) = max_{P_{X|S}} I(X;Y|S). Since the non-signaling-assisted non-causal capacity, previously known, is the same expression, the two settings coincide. The value also equals the classical capacity of the channel when the state is known at the receiver, so non-signaling assistance plus causal CSIT matches the rate achievable by state knowledge at the receiver. The proof constructs an explicit achievability scheme that satisfies the three non-signaling and causality constraints C1-C3. The paper then shows, in Theorem 2, that this rate-level equality does not extend to finite-blocklengt","pith_inferences":["Because quantum-assisted capacity is sandwiched between classical and non-signaling-assisted capacity, any channel where max_{P_{X|S}} I(X;Y|S) exceeds the classical causal capacity is a concrete candidate for an entanglement-assisted rate gain, with the non-signaling formula bounding how large that gain can be.","Algorithm 1's causal type-fixing construction is reusable in principle for other sequential non-signaling settings, offering a general trick for making an encoder's distribution depend only on a fixed-type function of the causal state.","Theorem 2's finite-blocklength separation suggests that other capacity equalities under non-signaling assistance may similarly hide finite-blocklength distinctions, and analogous strict gains from receiver state information could be probed in channels with memory or in network settings.","The claimed joint uselessness of individually useless resources, if generalized, points to a broader principle: resources that cannot increase capacity alone may also fail to do so collectively, which is worth testing in other information-theoretic scenarios."],"forward_implications":["Non-signaling assistance makes causal CSIT achieve the same asymptotic rate as non-causal CSIT, so future channel-state knowledge is not needed for capacity.","Providing the state to the receiver offers no asymptotic rate advantage: the same expression is the classical capacity with both CSIT and CSIR.","Non-signaling assistance can strictly beat the classical causal-CSIT capacity whenever max_{P_{X|S}} I(X;Y|S) exceeds the classical causal expression, since the known unbounded gaps in the non-causal setting carry over.","At finite blocklength the equivalence breaks: for the Z0/Z1 channel with two uses and a one-bit message, CSIR strictly improves the non-signaling-assisted success probability from at most 13/16 to at least 7/8.","The separate joint claim says non-signaling assistance, feedback, and strictly causal CSIT, even when available together, do not yield a capacity increase."],"fun_headline_variants":["NS-assisted causal CSIT matches non-causal capacity exactly","Causal and non-causal CSIT equal under non-signaling help","Receiver state info improves short-block error with causal CSIT","Feedback, strict CSIT, and NS aid fail to raise capacity"],"cache_read_input_tokens":23680,"weakest_assumption_plain":"The load-bearing premise is that the sequential non-signaling box conditions C1-C3 capture every non-signaling resource a causal encoder could exploit; if a legitimate non-signaling resource with internal memory across time falls outside this class, the claimed capacity formula would understate the true non-signaling-assisted capacity.","fun_headline_variants_meta":{"raw":{"variants":["NS-assisted causal CSIT matches non-causal capacity exactly","Causal and non-causal CSIT equal under non-signaling help","Receiver state info improves short-block error with causal CSIT","Feedback, strict CSIT, and NS aid fail to raise capacity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00066,"raw_usage":{"total_tokens":2896,"prompt_tokens":829,"completion_tokens":2067,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":2007}},"tokens_in":573,"tokens_out":2067,"duration_ms":13302,"temperature":1.0,"reasoning_tokens":2007,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T00:07:24.749502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the fully specified linear program LP1/LP2 for the Z0/Z1 channel with M=2, n=2: any violation of the stated optimum, or any non-signaling-assisted causal scheme whose success probability exceeds 13/16, would refute Theorem 2. For Theorem 1, exhibit a channel with state and any causal non-signaling-assisted scheme achieving a rate greater than max_{P_{X|S}} I(X;Y|S); the theorem's converse says this cannot happen.","supporting_citations":[],"review_version":1}