{"id":"61880532-5516-48ae-9ac8-7a45c4f8d975","arxiv_id":"1908.07521","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Distributed binary hypothesis testing over a discrete memoryless channel is studied, with an exact error-exponent tradeoff for remote testing and two inner bounds for the general correlated case.","lead":"Two parties observe correlated samples and one sends a message over a noisy channel; this paper derives the best tradeoff between the two kinds of mistakes in the resulting hypothesis test. For the simpler case where the receiver has no side data, the tradeoff is characterized exactly and shown to be separable into testing and channel coding.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 1 excludes noiseless channels from the exact RHT characterization, yet the advertised noiseless recovery in Remark 1 relies on an unproved limiting argument; the boundary claim is not fully established.","rationale":"The reader's weakest assumption is the same one I regard as load-bearing: Assumption 1. I checked the main steps of Proposition 1 and Theorem 3; under Assumption 1 the union expression (23), the two-message achievability, and the converse using (72) and (77) are internally coherent. The noiseless channel is a natural and advertised special case, but it lies outside Assumption 1, and the paper supplies no limiting argument. This is a real gap, but it is localized: it does not invalidate the exact characterization for absolutely-continuous channels nor the TAD example with p=0.35. I therefore do not move the verdict; CONDITIONAL remains appropriate. The typo in the last line of the proof of Theorem 5 (“κ(κα)≥κ∗h(κα)” should refer to κ∗u(κα)) is non-substantive. No stronger concern about the central noisy-channel result emerged from this pass.","tokens_in":49010,"tokens_out":18858,"duration_ms":209054,"concrete_test":"Fix P_U,Q_U and take a BSC with crossover ε>0, which satisfies Assumption 1. Compute the RHS of (23) as ε→0 and check whether it converges, for every θ0, to the direct HT region R_D of Theorem 1, i.e., whether the channel term in (23) diverges to +∞ while the min with the local-NP term yields ψ*_{P_U,Π}(θ0) and ψ*_{P_U,Π}(θ0)−θ0. Repeat this limit for Remark 1 by substituting a regularized noiseless channel into Theorem 4 and comparing the limit with [13, Theorem 1]. If the limit depends on the chosen ε-family or does not reproduce the noiseless inner bound, the recovery claim in Remark 1 and the boundary version of Theorem 3 fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption 1 in Section II-C requires PY|X(·|x̃) ≪ PY|X(·|x′) for all ordered input pairs. This is what makes every log-MGF ψ_{PY|X(·|x̃),\\barΠ_{x̃,x′}}(λ) in Proposition 1 and Theorem 3 finite on λ∈R. A deterministic noiseless channel violates the assumption: outputs of distinct inputs have disjoint supports, so \\barΠ in (15) is ±∞ and ψ* in (7) is not finite. The proof of the exact RHT region (Theorem 3, achievability via (70)–(71) and converse via (72)–(77)) is built on these finite-log-MGF estimates. Remark 1 then asserts that Theorem 4 recovers the noiseless Han-Kobayashi bound by “setting Ex(R,PSX), Em(PSX,θ) and Em(PSX,θ)−θ to ∞, which hold when the channel is noiseless”, but no regularizing family of channels satisfying Assumption 1 is given and the limit of the ψ* terms is not analyzed. Consequently, the recovery of [13] and the exactness claim for boundary noiseless channels are unsupported as written. The inner bounds for genuinely noisy channels (e.g., BSC with 0<p<1) are not threatened; the gap is specifically in the advertised limiting recovery and in the scope of the exact characterization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a two-terminal distributed binary hypothesis testing (DHT) problem in which an observer transmits over a discrete memoryless channel to a decision maker who also holds independent samples. The goal is the trade-off between type I and type II error exponents. The main results are: (i) an exact single-letter characterization of the error-exponent region for the remote hypothesis testing (RHT) special case in which the decision maker has no side information V (Theorem 3); (ii) an inner bound for general DHT based on separation of type-based source coding and unequal error-protection channel coding (Theorem 4); and (iii) an inner bound based on joint hybrid coding (Theorem 5), with an example showing that the joint scheme strictly outperforms the separation-based scheme. The paper also claims that Theorem 4 recovers the noiseless Han–Kobayashi bound and prior bounds of the authors.","tokens_in":49400,"tokens_out":23707,"duration_ms":255095,"significance":"If correct, Theorem 3 is a substantial exactness result: it shows that for remote hypothesis testing the optimal error-exponent trade-off is achieved by a separate local Neyman–Pearson test followed by transmission of a single binary decision over the channel, i.e., a form of separation between hypothesis testing and channel coding. Theorems 4 and 5 provide explicit, computable inner bounds for the noisy-channel DHT problem, improving on the earlier inner bound of Weinberger–Kochman–Wigger and recovering several noiseless and Stein-regime results. The proofs are unusually detailed, use standard method-of-types and log-moment-generating machinery, and do not rely on fitted parameters or equivalent-input derivations. The comparison example in Section III-C is concrete and gives a falsifiable separation between the two inner bounds.","major_comments":[{"comment":"In the converse proof of Theorem 3, Eq. (74) states that β_n(c_n) ≥ P_{Y|X(·|x')}(A_n) for some x'. This inequality is backwards for the argument that follows: β_n is an average over channel inputs, so it is upper bounded by the maximum of P_{Y|X(·|x)}(A_n) over x, not lower bounded by a single term. With the stated lower bound, the later step 'β_n ≤ e^{-n(E-θ)}' does not follow from Proposition 1, because the channel type II error is only a lower bound on the DHT type II error. The converse can be repaired by choosing x' to maximize P_{Y|X(·|x)}(A_n) and replacing (74) with β_n(c_n) ≤ P_{Y|X(·|x')}(A_n), then using the joint type of (argmin_x P_{Y|X(·|x)}(A_n^c), argmax_x P_{Y|X(·|x)}(A_n)) in the application of Proposition 1. As written, the exactness proof of Theorem 3 contains a load-bearing sign error and must be corrected.","section":"IV-B, Eq. (74)"},{"comment":"Assumption 1 requires that PY|X(·|x̃) and PY|X(·|x′) be mutually absolutely continuous for every ordered pair of channel inputs. A deterministic noiseless injective channel violates this assumption because outputs of distinct inputs have disjoint supports, so the log-MGF quantities ψ* in Proposition 1 and Theorem 3 are not finite. Nevertheless, Remark 1 claims that Theorem 4 recovers the noiseless Han–Kobayashi bound by 'setting Ex(R,PSX), Em(PSX,θ) and Em(PSX,θ)−θ to ∞, which hold when the channel is noiseless'. No regularizing family of channels satisfying Assumption 1 is supplied, and the limiting behavior of the ψ*-dependent terms is not analyzed. Thus the advertised recovery of [13] is not established as a theorem of the paper; the authors should either provide a rigorous limiting argument (e.g., through a sequence of noisy channels with common support and vanishing noise) or explicitly state that this recovery is formal and outside the assumptions.","section":"II-C / Remark 1 (after Theorem 4)"}],"minor_comments":[{"comment":"There is a typo in the statement of Theorem 5: 'κ*_u((κα))' should be 'κ*_u(κα)'.","section":"Theorem 5 statement"},{"comment":"The symbol R is used both for the error-exponent region in Definition 4 and for the channel-coding rate in Theorem 4 and its proof. This is confusing in statements such as 'R⊆...' versus 'ζ(κα,ω)−ρ(κα,ω)≤R<I_P(X;Y|S)'; consider renaming the rate variable.","section":"Notation"},{"comment":"The text says 'As we show later in (177), it follows from ...', but Eq. (97) is the relevant bound and its proof appears in Appendix A; the cross-reference should be updated to the correct equation number.","section":"Proof of Theorem 4, around Eq. (97)"},{"comment":"The final paragraph of the proof says 'we show that κ(κα) ≥ κ*_h(κα)' before analyzing uncoded transmission, which gives κ*_u(κα). The displayed symbol should be κ*_u(κα) to match the argument.","section":"Proof of Theorem 5, final paragraph"},{"comment":"The claim that Ex(0) is an upper bound on κ*_D(κα) for all κα should be justified in the text or figure caption, since it uses monotonicity of the expurgated exponent in R.","section":"Section III-C, Example 1 and Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The sign error in the converse of Theorem 3 appears to be a genuine but locally fixable mistake: replacing the lower bound in Eq. (74) with the corresponding upper bound repairs the argument and the central result is then plausible. The noiseless-recovery issue in Remark 1 is a scope/rigor gap rather than a falsification of the noisy-channel results. The paper is likely publishable after these two issues are addressed; no concerns about the provenance or novelty of the results arose."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe thing to know: Theorem 3 is the real contribution—an exact single-letter error-exponent tradeoff for remote hypothesis testing over a DMC, achieved by separate local NP tests plus a two-message channel code. That is new relative to the Han\\u2013Kobayashi line and to the authors\\u2019 earlier corner-point results. The proof machinery is standard (Chernoff bounds, types, log-MGFs), but the paper is unusually careful: achievability and converse are spelled out, no fitted constants appear, and self-citations are used to show recoveries rather than to prop up assumptions.\n\nThe general DHT results (Theorems 4 and 5) are inner bounds, not characterizations. That is honestly stated, and the example showing JHTCC beats SHTCC is clean. Those bounds recover the prior Stein-regime results, which is a useful sanity check.\n\nNow the soft spots.\n\nThe main load-bearing gap is Assumption 1 in Section II-C: mutual absolute continuity of PY|X(\\u00b7|x\\u0303) and PY|X(\\u00b7|x\\u2032) for every ordered pair. Every finite log-MGF in Proposition 1 and Theorem 3 depends on it. A deterministic noiseless channel violates it, because output distributions have disjoint supports. Remark 1 then claims recovery of the noiseless Han\\u2013Kobayashi bound by setting Ex, Em, and Em\\u2212\\u03b8 to infinity, but that is not a limiting argument through a family of channels satisfying Assumption 1. As written, the advertised recovery is not fully established. This does not threaten the noisy-channel inner bounds (a BSC with 0<p<1 is fine), and it is likely patchable via a limiting argument over perturbed full-support channels; but the authors need to supply that argument.\n\nThere is also a small typo in the proof of Theorem 5: after proving \\u03ba(\\u03ba\\u03b1) \\u2265 \\u03ba*h(\\u03ba\\u03b1), the text immediately repeats the same claim where the second occurrence should be \\u03ba*u(\\u03ba\\u03b1). It is obvious from context but should be fixed. The comparison example is a single point set; not a flaw, but the reader should not walk away thinking the general bounds are tight.\n\nThe citation pattern looks fair. The use of [10] and [39] is appropriate, and the comparison to [36] is candid.\n\nWho this is for: information theorists working on distributed hypothesis testing or error exponents. It deserves a serious referee. I would send it out; my expectation is that a referee will ask for the noiseless-limit repair and some proof polish, but the central result is worth engaging with.\n\nRecommendation: engage.","headline":"A genuinely new exact characterization for remote testing over noisy channels, but the advertised noiseless-recovery claim outruns Assumption 1 and needs a proper limiting argument.","tokens_in":49800,"tokens_out":1793,"would_cite":true,"duration_ms":17893,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A17","62B10","62F03"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that remote distributed hypothesis testing over a noisy channel has an exactly characterized error-exponent trade-off, achieved by separate source-side testing and two-message channel coding.","keywords":["distributed hypothesis testing","error exponents","type I and type II errors","noisy channel","single-letter characterization","hybrid coding","separation","log-moment generating function"],"falsifier":"Compute the right-hand side of the remote characterization for a specific instance where $V$ is unavailable, with a channel that satisfies the mutual absolute continuity assumption, and compare it against the converse bound derived from Proposition 1; any achievable error-exponent pair strictly outside the claimed union would falsify the theorem. A more direct check is to take a channel whose output distributions are not mutually absolutely continuous for some input pair, such as a channel with one deterministic input symbol, and see whether the formula's infinities reveal the need for an explicit limiting argument.","tokens_in":48802,"feed_emoji":"📡","tokens_out":8592,"duration_ms":80053,"temperature":0.7,"pith_summary":"Two parties observe independent samples of a joint distribution; the observer sends a message through a noisy channel, and the decision maker must decide which of two joint distributions generated the data. This paper asks how fast the two error probabilities can both decay with sample size, and it gives a single-letter formula for the whole trade-off in the remote case where the decision maker does not have its own samples. The formula says the optimal scheme is to run a likelihood-ratio test on the observer's own data, send only the binary decision through a two-message channel code, and run a second likelihood-ratio test on the channel output. For the general case with side information, the paper provides two inner bounds, one from separate testing and channel coding and one from joint hybrid coding, and shows the joint scheme is strictly better on a testing-against-dependence example. If the remote characterization is right, it settles the optimal asymptotic trade-off in that setting and shows that separation costs nothing there.","feed_headline":"Separation is optimal for remote testing over a noisy channel","feed_subtitle":"A single-letter formula gives the full type I/type II error-exponent trade-off when the decision maker has no side information.","key_machinery":"The log-moment generating function (log-MGF) rate function $\\psi^*_{P,f}(\\theta) = \\sup_{\\lambda\\in\\mathbb{R}}(\\theta\\lambda - \\log\\mathbb{E}_P[e^{\\lambda f(Z)}])$ carries the argument: it converts a likelihood-ratio threshold test into error exponents. For the source part, $f = \\Pi_{P_U,Q_U}$ is the log-likelihood ratio of the two candidate distributions of $U$; for the channel part, $f = \\bar\\Pi_{\\tilde x,x',P_{Y|X}}$ compares the channel output distributions under two input letters. The remote-problem formula takes the componentwise minimum of the two rate functions, one for the local source test and one for the channel-output test, then unions over the joint input type; the separation claim is that this componentwise minimum is the whole region. For the general problem, the SHTCC scheme uses type-based quantization and binning plus an unequal-error-protection channel code, with expurgated exponent $E_x(R,P_{SX})$ and a special-message exponent $E_{sp}(P_{SX},\\theta)$; the JHTCC scheme uses hybrid coding with a joint decoding metric.","core_discovery":"The paper's central result, Theorem 3, characterizes the optimal type I/type II error-exponent trade-off for remote hypothesis testing, where $V$ is unavailable at the decision maker. For any joint input type $P_{X_0X_1}$ and thresholds $\\theta_0,\\theta_1$, the achievable exponent pair is $\\zeta_0 = \\min\\{\\psi^*_{P_U,\\Pi_{P_U,Q_U}}(\\theta_0),\\ \\mathbb{E}_{P_{X_0X_1}}[\\psi^*_{P_{Y|X}(\\cdot|X_0),\\bar\\Pi_{X_0,X_1,P_{Y|X}}}(\\theta_1)]\\}$ and $\\zeta_1 = \\min\\{\\psi^*_{P_U,\\Pi_{P_U,Q_U}}(\\theta_0)-\\theta_0,\\ \\mathbb{E}_{P_{X_0X_1}}[\\psi^*_{P_{Y|X}(\\cdot|X_0),\\bar\\Pi_{X_0,X_1,P_{Y|X}}}(\\theta_1)]-\\theta_1\\}$, with the union taken over all joint input types and threshold intervals. This pair is achieved by the observer running the classical likelihood-ratio test on its own $U$ samples, transmitting the one-bit decision through a two-codeword channel code, and the decision maker running the analogous likelihood-ratio test on the channel output; the converse shows that no other scheme can do better. The paper's secondary results, Theorems 4 and 5, give inner bounds for the general DHT problem using type-based quantization with unequal error protection and hybrid coding respectively, recovering the rate-limited noiseless bound and previous corner-point exponents as special cases, and demonstrating a strict gap on a testing-against-dependence example over a binary symmetric channel.","pith_inferences":["The two-threshold structure of the remote formula suggests a testable extension: at finite blocklength the remote problem should obey a refined trade-off governed by the same two log-MGFs plus channel dispersion terms, so a Berry-Esseen-style analysis would give non-asymptotic bounds.","Because the remote characterization is exact only under mutual absolute continuity, the paper's own remark about recovering the noiseless case hints that a limiting argument is needed but not written out; making that limit explicit would remove the main technical caveat.","The SHTCC bound's recovery of the rate-limited case suggests that quantizing $U$ to the type level is not just an analysis tool but may be necessary for optimality; a possible test is to see whether any scheme with finer quantization improves the exponents.","The demonstrated strict gap between joint and separate schemes is shown at very small type I exponents; optimizing all hybrid-coding parameters might reveal whether the gap persists across the whole trade-off, which would strengthen the case for joint source-channel coding."],"forward_implications":["For the remote setting, the exact trade-off means no scheme can beat the simple two-step rule: test $U$ locally, send the binary decision over a channel code, and test the channel output.","The SHTCC inner bound contains the known rate-limited noiseless bound and the earlier corner-point type II exponent as special cases, so the noisy-channel results reduce cleanly to prior results in the right limits.","For testing against independence in the vanishing type I limit, the optimal type II exponent depends on the channel only through its capacity, preserving the earlier capacity-only characterization.","For testing against dependence over a binary symmetric channel, the joint hybrid-coding bound strictly dominates the separation-based bound for small type I exponents, so the trade-off region is not separation-optimal in general.","The two inner bounds give explicit formulas that can be evaluated numerically for finite alphabets, making the error-exponent trade-off computable in principle for small systems."],"supporting_citations":[{"why":"Supplies the log-MGF large-deviation theorems and the converse used to prove Proposition 1 and the remote characterization.","marker":"[38]"},{"why":"Provides the noiseless rate-limited inner bound that the SHTCC scheme recovers as a special case.","marker":"[13]"},{"why":"Unequal error-protection coding scheme used to build the separation-based inner bound.","marker":"[35]"},{"why":"Hybrid coding framework underlying the joint inner bound in Theorem 5.","marker":"[34]"},{"why":"Type-based source-channel error exponent techniques used in the channel coding and uncoded transmission analyses.","marker":"[37]"},{"why":"Prior DHT over noisy channels results that the current bounds extend and whose joint-vs-separate gap example is revisited.","marker":"[10]"},{"why":"Prior single-letter type II exponent for remote hypothesis testing over a noisy channel, recovered as a corner point of Theorem 3.","marker":"[39]"},{"why":"Standard type-counting, covering, and expurgated exponent bounds used throughout the proofs.","marker":"[31]"}],"fun_headline_variants":["Optimal error-exponent trade-off for remote testing","Noisy channel: exact trade-off for remote hypothesis testing","Single-letter formula for remote testing without side info","Remote testing over noisy link: optimal exponents characterized","Tight trade-off for noisy-channel remote hypothesis testing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole calculation assumes that, for every pair of channel inputs, the two possible output distributions overlap completely (are mutually absolutely continuous), so that all log-moment functions stay finite; the noiseless-channel recovery would need a separate limit argument because a deterministic channel has non-overlapping output distributions.","fun_headline_variants_meta":{"raw":{"variants":["Optimal error-exponent trade-off for remote testing","Noisy channel: exact trade-off for remote hypothesis testing","Single-letter formula for remote testing without side info","Remote testing over noisy link: optimal exponents characterized","Tight trade-off for noisy-channel remote hypothesis testing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000332,"raw_usage":{"total_tokens":1945,"prompt_tokens":1142,"completion_tokens":803,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":758,"completion_tokens_details":{"reasoning_tokens":728}},"tokens_in":758,"tokens_out":803,"duration_ms":524823,"temperature":1.0,"reasoning_tokens":728,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:52:53.008818+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the right-hand side of the remote characterization for a specific instance where $V$ is unavailable, with a channel that satisfies the mutual absolute continuity assumption, and compare it against the converse bound derived from Proposition 1; any achievable error-exponent pair strictly outside the claimed union would falsify the theorem. A more direct check is to take a channel whose output distributions are not mutually absolutely continuous for some input pair, such as a channel with one deterministic input symbol, and see whether the formula's infinities reveal the need for an explicit limiting argument.","supporting_citations":[{"cited_title":"Polyanskiy and Y","cited_arxiv_id":null,"evidence_quote":"Supplies the log-MGF large-deviation theorems and the converse used to prove Proposition 1 and the remote characterization."},{"cited_title":"Unequal error protection: An information-theoretic perspective,","cited_arxiv_id":null,"evidence_quote":"Unequal error-protection coding scheme used to build the separation-based inner bound."},{"cited_title":"A unified approach to hybrid coding,","cited_arxiv_id":null,"evidence_quote":"Hybrid coding framework underlying the joint inner bound in Theorem 5."},{"cited_title":"On the error exponent of source-channel transmission with a distortion threshold,","cited_arxiv_id":null,"evidence_quote":"Type-based source-channel error exponent techniques used in the channel coding and uncoded transmission analyses."},{"cited_title":"Distributed hypothesis testing over discrete memoryless channels,","cited_arxiv_id":null,"evidence_quote":"Prior DHT over noisy channels results that the current bounds extend and whose joint-vs-separate gap example is revisited."}],"review_version":1}