{"id":"9937adf6-0426-400c-8f81-aea020028886","arxiv_id":"1909.00225","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Two MAP-based statistical algorithms reconstruct multiple numbers from unordered noisy residues, claiming stronger robustness than deterministic CRT schemes, especially in heavy noise.","lead":"This paper designs two statistical algorithms to recover multiple unknown numbers from noisy, unlabeled remainders under several moduli, trading deterministic error bounds for probabilistic models. The authors say the approach tolerates much heavier noise than prior deterministic CRT methods, which matters for radar, frequency estimation, and wireless ranging.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MAP derivation fails because Assumption 1 is not independent of the permutation K; Theorem 1 optimizes the wrong posterior and is not the conditional MAP of residue clustering.","rationale":"The reader's REJECT verdict is supported, but the strongest reason is not Eq. (13). Re-examining the derivation, A1 restricts each cluster's residues to lie in an arc of length < Γ/2, and under that restriction the wrapped Gaussian sum for a cluster reduces to a common-shift family; integrating that family over the common residue gives Eq. (13). So the wrapped/unwrapped identity is not the fatal step. The fatal step is the independence of A1 from K. The text between Eqs. (11) and (12) asserts that normality of p(R|K) makes Pr(A1|K) constant; this is a non sequitur because the event A1 is defined through the clusters induced by K. When K is wrong, a cluster can contain residues of several distinct Y_i, and A1 then requires those residues to fit in an arc shorter than Γ/2. For N=3, L=3, a cyclic permutation groups three independent uniform residues; A1 holds only when the three points lie in a semicircle, probability 3/4, while the correct permutation has A1 true with probability 1. This asymmetry means p(K|A1) is not uniform and the posterior of K is not proportional to the likelihood term maximized in Theorem 1. The algorithm may still be a useful heuristic and the simulations are suggestive, but the paper's central claim of a rigorous MAP analysis under Assumption 1 is not established. Hence the REJECT verdict stands, though the justification should rest on the independence failure rather than on Eq. (13).","tokens_in":19301,"tokens_out":36510,"duration_ms":369398,"concrete_test":"Compute, analytically or by Monte Carlo, Pr(A1|K) for Γ=1, N=3, L=3, zero noise, Y_i iid uniform on [0,1). For K_identity, A1 holds with probability 1. For K_cycl defined by K_l(i)=i+l mod 3, each cluster contains one residue from each Y_i, so A1 holds iff the three Y_i lie in some open semicircle, whose probability is exactly 3/4. If these probabilities differ (1 vs 3/4), the independence assertion in Eqs. (11)-(12) is false and Algorithm 1's objective omits the required factor P(A1|K). A 10^5-sample Monte Carlo with σ=0 verifies the probabilities to three decimal places.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The decisive flaw is the asserted independence of Assumption 1 (A1) from the permutation K[1:L], used between Eqs. (11) and (12) to drop p(K|A1) from the posterior. The proof claims that because p(R[1:L]|K[1:L]) is Gaussian, the integral Pr(A1|K) = ∫ p(A1|K,R)p(R|K)dR is constant in K. This does not follow: the integrand p(A1|K,R) itself depends on K because A1 is a property of the clusters induced by K. Concretely, take Γ=1, N=3, L=3, and zero noise. For the identity permutation, each cluster is a singleton, all intervals have length 0, and A1 is true with probability 1. For a cyclic permutation that groups one residue of each of the three Y_i into every cluster, A1 holds only when the three Y_i lie in some open semicircle, which has probability 3/4 by Wendel's theorem. Thus Pr(A1|K) differs between permutations, so p(K|R,A1) ∝ p(R|A1,K)·P(A1|K) contains a K-dependent factor that Theorem 1 drops. Algorithm 1 therefore maximizes p(R|A1,K) (or a profile over τ), not the true conditional MAP of K. Since the paper's headline contribution is the rigorous MAP derivation under Assumption 1, this breaks the central theoretical claim. I do not rest the rejection on Eq. (13): under the A1 restriction the wrapped likelihood for each cluster collapses to the common-shift family, so the wrapped/unwrapped identity is more defensible than the reader suggests. The independence failure is the load-bearing issue.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the reconstruction of N real numbers Y_i from unordered noisy residues modulo m_l = Γ M_l, where the M_l are pairwise coprime and D = Γ∏ M_l is the maximal dynamic range. It proposes two statistical schemes: Algorithm 1, which claims to compute the MAP of residue clustering under Assumption 1 in semi-closed form by enumerating O(NL) cut points, followed by single-number RCRT; and Algorithm 2, an iterative wrapped-Gaussian mixture procedure alternating circular nearest-neighbor matching and common-residue estimation. Residue error-correcting codes and majority voting over modulus subsets are added as robustness enhancements. Simulations compare success rates against the deterministic GRCRT of [25] for N=2,...,10 over SNR -40 to 0 dB.","tokens_in":19713,"tokens_out":24656,"duration_ms":242709,"significance":"If the theoretical claims were correct, the paper would be a substantial step: it promises maximal dynamic range D=lcm(m_l) while simultaneously resolving the correspondence ambiguity and tolerating noise, with polynomial-time algorithms. The algorithmic recipes are concrete, the modular-projection viewpoint is sound, and the simulation comparison is extensive. However, the central theoretical foundation is not established: the likelihood identity used to derive the MAP is not exact in the regime of interest, and the global conditioning on Assumption 1 is not handled correctly. The contribution as a rigorous statistical framework therefore fails, although the empirical heuristics may still be of practical interest.","major_comments":[{"comment":"Lemma 3 replaces the product of wrapped Gaussian likelihoods by a product of real-line Gaussian integrals. This is not an identity for L>1. For L=2 with equal weights, the exact marginal likelihood for two residues r1,r2 assigned to one number is ∫_0^Γ w_σ(r1|μ)w_σ(r2|μ)dμ = (1/Γ)∑_{k∈Z} e^{2πik(r1-r2)/Γ} e^{-4π²σ²k²/Γ²}, a periodic function of r1-r2. Eq. (13) instead gives exp[-(r1-r2)²/(4σ²)] up to constants. When σ is comparable to Γ, the exact likelihood is nearly uniform in the circular difference, whereas the real-line expression still discriminates strongly; for σ=Γ and |r1-r2|=Γ/2 the real-line factor is e^{-1/16}≈0.94, while the wrapped likelihood differs from its maximum by a factor of order e^{-4π²}. Therefore Theorem 1 and Algorithm 1 maximize an approximate objective, not the claimed posterior MAP. The equality in Eq. (7) is special to L=1 and does not extend to the product in Eq. (6).","section":"Section III, Lemma 3"},{"comment":"Eq. (13) also asserts factorization over the N clusters. Assumption 1 is a global event: it requires a single point τ lying outside the union of all intervals I_i, i=1,...,N. The indicator of this event in the numerator p(R,A1|K)=∫ p(R|Y,K) 1_{A1} dY couples all i, so the conditional density does not factorize into per-cluster integrals as the product in Eq. (13) claims. The sentence 'due to the independence of μ_i' does not address the existence of a common τ. Even if one fixes a valid τ, the integration over each μ_i should be restricted to the small arc allowed by the event rather than extended to the whole real line. Thus the semi-closed-form MAP derivation is not established.","section":"Section III, Lemma 3"},{"comment":"The proof that Assumption 1 is independent of the permutation K is invalid as written. The text after Eq. (12) asserts that the integral ∫ p(A1|K,R)p(R|K)dR is constant in K because p(R|K) is a normal distribution. But p(R|K) is a product of wrapped Gaussian densities, not a normal density, and p(A1|K,R) is an indicator that depends on the cluster geometry induced by K. The stated reason does not imply constancy. The conclusion may be salvageable by an exchangeability argument applied to the i.i.d. noise matrix, since K only permutes entries within each column, but the proof given must be replaced if the paper is to claim a rigorous derivation.","section":"Section III, Eqs. (11)-(12)"},{"comment":"Step one of Algorithm 2 replaces the wrapped likelihood in Eq. (23) by the nearest-centroid objective (24), justified only when σ_l² is much smaller than Γ. The simulations in Section V set σ²=10^{-SNR/10}, with Γ=100, so at SNR=-40 dB one has σ=100=Γ. The heavy-noise regime in which the paper claims large performance gains is therefore exactly the regime where the theoretical justification of Eq. (24) fails. The experimental results may reflect a useful heuristic, but they are not backed by the paper's statistical derivation.","section":"Section IV, Eq. (24) and Section V"}],"minor_comments":[{"comment":"There are several typos: 'Methology' in Section II, 'stead state' in Section IV, and 'K[1:N]' in Eq. (19) where the permutation variable is indexed by L, not N.","section":"General"},{"comment":"The input line 'Calculate ril = ⟨\\tilde r_il⟩_{M_l}' appears to be a typo; the algorithm should compute the residue of R_il modulo Γ, i.e., r_il = ⟨R_il⟩_Γ.","section":"Algorithm 2"},{"comment":"The formula in Eq. (29) is garbled: the weighted-average expression should involve the unwrapped values γ_(l)+Γ for l<j, and the displayed index ranges appear to double-count terms. Please correct the expression and its derivation.","section":"Section IV, Eq. (29)"},{"comment":"The paper states that the MAP can be found from O(NL) candidate cut points, but each candidate requires re-sorting the residue sets; if a total complexity claim is intended, it should be stated explicitly rather than leaving the reader to infer it from the number of candidates.","section":"Section III, Theorem 1"}],"recommendation":"reject","confidential_remarks":"The central theoretical claim of the paper is not sound: the likelihood identity in Eq. (13) is false in the heavy-noise regime emphasized in the title and abstract, and the global nature of Assumption 1 is not reflected in the factorized posterior. These are not local presentation issues; they invalidate the claimed MAP derivation. If the authors wish to reframe the paper as a heuristic algorithm with empirical evaluation, that would be a different contribution and would require major rewriting. I therefore recommend rejection of the current submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know before you read this one. First, it is a real attempt at something new: a statistical treatment of generalized CRT for multiple numbers with unordered residues, framed as a wrapped Gaussian mixture, and it ships two concrete algorithms with plausible simulations. Second, the central theoretical claim—that Algorithm 1 computes the MAP of residue clustering under Assumption 1—is wrong. The independence argument between Eqs. (11) and (12) fails, so the posterior being maximized is not the conditional MAP.\n\nWhat is good: the problem setup is well motivated. The paper distinguishes correspondence ambiguity from perturbation, and the connection between the common residue µ_i and the clustering of ⟨Ril⟩_Γ is a useful observation. Algorithm 1's O(NL) enumeration via the rearrangement inequality is neat, given a fixed cut point τ. Algorithm 2, the iterative wrapped-Gaussian EM, is sensible as a heuristic, and the experiments show clear gains over the deterministic baseline in heavy noise. The error-correcting-code addition is a decent touch.\n\nWhere it falls: the load-bearing flaw is the claim that Pr(A1|K) is constant across K. The proof asserts that because p(R|K) is Gaussian, the integral with p(A1|K,R) is constant. That does not follow. A concrete counterexample: Γ=1, N=3, L=3, zero noise. Identity permutation puts each cluster as a singleton, so A1 holds with probability 1. A cyclic permutation groups one residue of each Y_i into each cluster; then A1 holds only if the three Y_i lie in some open semicircle, probability 3/4. So P(A1|K) varies, and p(K|R,A1) contains a K-dependent factor that Theorem 1 drops. Algorithm 1 therefore maximizes something close to the likelihood, not the MAP. Eq. (13) is less of a problem than the reader's first pass suggested—under A1 the wrapped/unwrapped likelihood is defensible. The real issue is the independence.\n\nBottom line: the empirical work may be salvageable as a heuristic, and the problem is worth studying, but as written the paper overclaims. It deserves serious refereeing because the algorithms and motivation are substantive; just not acceptance in this form unless the MAP claim is corrected or the paper is repositioned as a heuristic with empirical support.","headline":"A genuinely new statistical framing of GRCRT with promising heuristics, but the advertised MAP derivation is invalid because Assumption 1 is not independent of the permutation K.","tokens_in":20137,"tokens_out":3566,"would_cite":false,"duration_ms":30999,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["11A07","62F15","62H30","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"Statistical robust CRT can be solved by a MAP residue clustering found among O(NL) candidates.","keywords":["robust Chinese remainder theorem","multiple parameter estimation","maximum a posteriori estimation","wrapped Gaussian mixture model","residue clustering","ambiguity resolution","error correcting codes","dynamic range"],"falsifier":"Take $N=2$, $L=2$, $\\Gamma=1$, and noise standard deviation $\\sigma=0.4$; numerically integrate the true posterior in (5) over all residue clusterings and compare the resulting argmax with Algorithm 1’s ranking of the $O(NL)$ candidate cutting points. If equation (13) were exact, the two rankings would coincide; as $\\sigma$ approaches $\\Gamma$, ranking mismatches should grow, showing that the closed form is not the true posterior in heavy noise.","tokens_in":1929,"feed_emoji":"🧮","tokens_out":7548,"duration_ms":115713,"temperature":0.7,"pith_summary":"This paper claims that the robust reconstruction of $N$ unknown numbers from $L$ noisy, unordered residue sets—the generalized robust Chinese remainder problem—can be solved statistically rather than only deterministically. It proposes a statistical model in which the unknown residue-to-number correspondence is a random permutation and the noisy residues wrap onto a circle of length $\\Gamma$. Under an assumption that each number’s noise spreads over less than half that circle, the first algorithm computes the MAP residue clustering from $O(NL)$ candidates, reducing the task to $N$ independent single-number robust CRT problems; the second algorithm iteratively alternates between matching residues to common residues and refining those common residues. If correct, these schemes operate at the full dynamic range $D=\\mathrm{lcm}(m_1,\\ldots,m_L)$ instead of a shrunken range, and they outperform deterministic schemes under heavy noise, while the paper notes that deterministic methods can win at very low noise.","feed_headline":"Statistical CRT recovers noisy residues at full dynamic range","feed_subtitle":"MAP residue clustering narrows the search from factorial to O(NL) candidates and beats deterministic schemes in heavy noise","key_machinery":"The load-bearing object is the common residue $\\mu_i=\\langle Y_i\\rangle_\\Gamma$—the residue every observation of $Y_i$ shares on the small circle of length $\\Gamma$—together with the cutting point $\\tau$ of Assumption 1. Cutting the circle at $\\tau$ straightens the noise intervals into ordered segments, so the ranks of the projected residues $r_{il}$ correspond to the ranks of the noise terms. The MAP objective becomes a quadratic form that, by the rearrangement inequality, is maximized by sorting-matching the projected residues across samplers, reducing the search over $L\\cdot N!$ permutations to a scan of $NL$ candidate cutting points.","core_discovery":"The central discovery is that the correspondence ambiguity—which noisy residue belongs to which underlying number—can be treated as a parameter to be estimated by maximum a posteriori probability, and that under an assumption on how far noise spreads around a small circle of circumference $\\Gamma$, the MAP clustering reduces to an ordering problem. Projecting each observation $R_{il}$ to its residue $r_{il}=\\langle R_{il}\\rangle_\\Gamma$ reveals a common residue $\\mu_i=\\langle Y_i\\rangle_\\Gamma$ shared by all $L$ moduli. For a fixed cutting point $\\tau$, the optimal clustering groups the $i$-th largest elements of each sampler’s projected residues together, a consequence of the rearrangement inequality, and scanning the $NL$ distinct observed projected residues as cutting points gives the global MAP clustering under the assumption. After clustering, reconstruction decomposes into $N$ independent single-number robust CRT problems. The second algorithm replaces the fixed cutting point by an iteration that alternates between circular matching of residues to current common-residue estimates and least-squares updating of those estimates.","pith_inferences":["Editorial inference: the exact-MAP claim depends on the wrapped/unwrapped integral identity in Lemma 3; if that identity is not exact in heavy noise, the $O(NL)$ search produces an approximation, though still a tractable clustering rule.","Editorial inference: the sorting-matching step is a circular assignment problem, so optimal transport or assignment algorithms may extend the method to regimes where Assumption 1 fails or to non-Gaussian circular noise.","Editorial inference: a numerical comparison against exact marginalization for small $N$ and $L$ would settle whether Algorithm 1 is truly MAP or only a high-quality heuristic in heavy noise.","Editorial inference: the same “common residue plus circular sorting” mechanism may apply to other one-dimensional circular registration problems, such as multi-sensor phase unwrapping or circular motif detection."],"forward_implications":["Under Assumption 1, the MAP residue clustering is computed from $O(NL)$ candidates rather than $L\\cdot N!$, so the correspondence bottleneck disappears and each number is reconstructed by a single-number RCRT.","Statistical RCRT can work with dynamic range $D=\\mathrm{lcm}(m_1,\\ldots,m_L)$, the maximal moduli utilization, whereas deterministic generalized robust CRT requires a shrunken range and a noise bound of $\\Gamma/(4N)$.","Algorithm 2’s iterative wrapped-Gaussian mixture refinement gives stronger robustness than Algorithm 1, at the cost of more computation, and in the paper’s simulations it reaches a stationary state within a few iterations, mostly between two and three.","Residue error-correcting codes, used with Theorem 3, tolerate a bounded number of clustering errors; this becomes more beneficial as $N$ grows, when a larger number of moduli can be exploited for correction.","The paper itself states that in the extremely low noise case, deterministic schemes may outperform the proposed schemes, consistent with the theoretical analysis."],"supporting_citations":[{"why":"Supplies the closed-form single-number robust CRT whose quotient-recovery framework and $\\Gamma/4$ noise bound the paper builds on and generalizes.","marker":"[19]"},{"why":"Provides the deterministic generalized robust CRT for arbitrary $N$ that is the primary baseline and the target dynamic-range/noise-bound improvement.","marker":"[25]"},{"why":"The robust generalized CRT for two integers, a direct predecessor whose error condition the paper relaxes and extends to $N$ numbers.","marker":"[7]"},{"why":"The maximum-likelihood single-number RCRT whose small-noise approximation and $O(L)$ common-residue update are used in Algorithm 2’s second step.","marker":"[18]"},{"why":"Supplies Theorem 3 on robust reconstruction with a small number of arbitrary residue errors, which is the foundation of the error-correcting-code enhancement.","marker":"[27]"},{"why":"Proves the optimality of the $\\Gamma/4$ noise bound and modulus selection for single-number RCRT, a benchmark used throughout the paper.","marker":"[26]"},{"why":"The rearrangement inequality is the key inequality in Theorem 1 that proves sorting-matching of projected residues maximizes the objective for a fixed cutting point.","marker":"[2]"},{"why":"Sharpens the dynamic range of the generalized CRT for multiple integers, defining the deterministic utilization limit the statistical schemes aim to surpass.","marker":"[10]"}],"fun_headline_variants":["MAP clustering makes CRT robust to heavy noise","Statistical CRT: first rigorous model for multi-number","CRT residue matching solved by MAP, not brute force","Heavy-noise CRT broken by conditional MAP estimation","Generalized CRT gets a statistical backbone"],"cache_read_input_tokens":22272,"weakest_assumption_plain":"The load-bearing premise is that integrating a product of wrapped Gaussian densities over one period equals integrating the product of unwrapped Gaussians over the whole line; that equality is exact only when the noise is much smaller than the period $\\Gamma$, and the paper applies it precisely in the heavy-noise regime where the two sides diverge.","fun_headline_variants_meta":{"raw":{"variants":["MAP clustering makes CRT robust to heavy noise","Statistical CRT: first rigorous model for multi-number","CRT residue matching solved by MAP, not brute force","Heavy-noise CRT broken by conditional MAP estimation","Generalized CRT gets a statistical backbone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1525,"prompt_tokens":890,"completion_tokens":635,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":575}},"tokens_in":506,"tokens_out":635,"duration_ms":7878,"temperature":1.0,"reasoning_tokens":575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:58:34.038138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $N=2$, $L=2$, $\\Gamma=1$, and noise standard deviation $\\sigma=0.4$; numerically integrate the true posterior in (5) over all residue clusterings and compare the resulting argmax with Algorithm 1’s ranking of the $O(NL)$ candidate cutting points. If equation (13) were exact, the two rankings would coincide; as $\\sigma$ approaches $\\Gamma$, ranking mismatches should grow, showing that the closed form is not the true posterior in heavy noise.","supporting_citations":[{"cited_title":"A closed-form robust chinese remainder theorem and its performance analysis","cited_arxiv_id":null,"evidence_quote":"Supplies the closed-form single-number robust CRT whose quotient-recovery framework and $\\Gamma/4$ noise bound the paper builds on and generalizes."},{"cited_title":"Robustness in chinese remainder theorem for multiple numbers and remainder coding","cited_arxiv_id":null,"evidence_quote":"Provides the deterministic generalized robust CRT for arbitrary $N$ that is the primary baseline and the target dynamic-range/noise-bound improvement."},{"cited_title":"A robust generalized chinese remainder theorem for two integers","cited_arxiv_id":null,"evidence_quote":"The robust generalized CRT for two integers, a direct predecessor whose error condition the paper relaxes and extends to $N$ numbers."},{"cited_title":"Maximum likelihood estimation based robust chinese remainder theorem for real numbers and its fast algorithm","cited_arxiv_id":null,"evidence_quote":"The maximum-likelihood single-number RCRT whose small-noise approximation and $O(L)$ common-residue update are used in Algorithm 2’s second step."},{"cited_title":"On solving ambiguity resolution with robust chinese remainder theorem for multiple numbers","cited_arxiv_id":null,"evidence_quote":"Supplies Theorem 3 on robust reconstruction with a small number of arbitrary residue errors, which is the foundation of the error-correcting-code enhancement."},{"cited_title":"Notes on crt-based robust frequency estimation","cited_arxiv_id":null,"evidence_quote":"Proves the optimality of the $\\Gamma/4$ noise bound and modulus selection for single-number RCRT, a benchmark used throughout the paper."},{"cited_title":"Inequalities","cited_arxiv_id":null,"evidence_quote":"The rearrangement inequality is the key inequality in Theorem 1 that proves sorting-matching of projected residues maximizes the objective for a fixed cutting point."},{"cited_title":"A sharpened dynamic range of a generalized chinese remainder theorem for multiple integers","cited_arxiv_id":null,"evidence_quote":"Sharpens the dynamic range of the generalized CRT for multiple integers, defining the deterministic utilization limit the statistical schemes aim to surpass."}],"review_version":1}