{"id":"ceefe4f0-af12-465e-92ad-adc819a8e744","arxiv_id":"2504.18585","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"Summing N independent draws from a biased m-sided source and reducing modulo m drives the output toward uniform exponentially fast; the paper's stronger non-stationary claim requires an extra decay condition it does not state in its theorem.","lead":"The paper proposes a post-processing rule for biased random sources: sum N outcomes and reduce the sum modulo the number of possible outcomes, claiming the result becomes uniform as N grows. The target use is cheap, robust debiasing inside quantum random number generators.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Non-stationary robustness theorem is false as stated: positivity of p_j(k) does not force product decay of Fourier factors; an explicit m=2 counterexample violates Eq. 13.","rationale":"The stationary part of the paper is correct: root-of-unity filtering and Eq. 17 are standard and valid. The advertised novelty and the QRNG motivation, however, center on robustness to time-dependent probabilities. That claim depends entirely on Section 4's product-decay step. The reader's weakest assumption names exactly the missing summability condition and gives the same m=2 counterexample; my independent check confirms it. I therefore agree with the REJECT verdict: the false non-stationary theorem is a headline result, not a typo or a removable proof gap, although adding a summability or uniform-positivity condition and restating the theorem would make the paper acceptable as a conditional contribution.","tokens_in":12523,"tokens_out":4457,"duration_ms":38742,"concrete_test":"Compute Eq. 11 exactly for the m=2 sequence p_j(0)=1-2^{-(j+1)}, p_j(1)=2^{-(j+1)}. The closed form is P(r_f=0)=(1+∏_{j=1}^N(1-2^{-j}))/2. Taking N→∞ gives a limit of (1+C)/2≈0.644, not 0.5, where C=∏_{j=1}^∞(1-2^{-j})≈0.2888. A corroborating numerical check: simulate a large number of independent blocks at N=1000; the sample proportion of residue 0 will be sharply concentrated near 0.644. Either check settles that Eqs. 12-13 are false under the stated assumptions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4's proof of non-stationary robustness rests on an invalid inference. From Eq. 11, P(r_f=k) is 1/m times a sum over r of ω^{-rk}∏_j G_j(ω^r). The argument notes |G_j(ω^r)|<1 for each j and then takes the limit of the product ∏_j G_j(ω^r) to 0 in Eqs. 12-13. Pointwise strict inequality does not imply product decay: if 1-|G_j(ω^r)| is summable, the product can tend to a nonzero constant. The paper even states a sufficient condition, ∑_j (1-|G_j(ω^r)|) diverges, but that condition is not assumed in the theorem and is not implied by p_j(k)>0. This is not a minor gap; it invalidates the abstract's headline robustness claim. Concretely, for m=2 let p_j(0)=1-2^{-(j+1)} and p_j(1)=2^{-(j+1)}. Then G_j(-1)=1-2^{-j}, so P(r_f=0)=1/2+(1/2)∏_{j=1}^N(1-2^{-j}), which tends to (1+C)/2≈0.644 with C=∏_{j=1}^∞(1-2^{-j})≈0.2888, not 1/2. All p_j(k)>0, so the theorem's hypotheses are met and Eq. 13's conclusion fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modular debiasing method for discrete m-ary sources: the outcomes of N independent trials are summed modulo m, and the resulting residue is used as a supposedly uniform output. The authors claim that, provided every outcome has positive probability, the distribution of the modular sum converges exponentially to the uniform distribution, and that the guarantee persists under non-stationary or time-dependent biases. The derivation uses probability generating functions and roots of unity, derives an IID error bound controlled by the spectral radius rho, and validates the claims with numerical simulations for stationary and non-stationary examples. A practical application to spatial photon-detection QRNGs is also discussed.","tokens_in":12836,"tokens_out":4154,"duration_ms":41177,"significance":"The IID part of the paper is both correct and useful: Eq. (9) gives an exact formula for the modular-sum distribution, and the bound in Eq. (17) is a clean exponential convergence statement for any full-support stationary source. The method is conceptually simple and the numerical experiments are extensive, including entropy and TVD convergence tests. However, the paper's headline contribution is the claimed robustness to non-stationary sources, and that claim is false as stated. Since the QRNG motivation in Section 7 and the abstract's central promise rest on this non-stationary robustness, the current manuscript cannot be accepted without a fundamental reformulation and restriction of the theorem.","major_comments":[{"comment":"The inference that positivity of p_j(k) forces the product of Fourier factors to vanish is invalid. From |G_j(omega^r)| < 1 for each j it does not follow that the infinite product tends to 0; this requires a summability condition such as sum_j (1 - |G_j(omega^r)|) = infinity. The paper itself states such a sufficient condition in passing, but it is not assumed in the theorem and is not implied by p_j(k) > 0. A concrete counterexample with m=2 is p_j(0) = 1 - 2^{-(j+1)}, p_j(1) = 2^{-(j+1)}. Then G_j(-1) = 1 - 2^{-j}, so P(r_f=0) = 1/2 + (1/2) prod_{j=1}^N (1 - 2^{-j}), whose limit is (1+C)/2 with C = prod_{j=1}^infty (1-2^{-j}) > 0, not 1/2. All hypotheses of Eq. (13) are satisfied, yet the conclusion fails.","section":"Section 4, Eqs. (11)-(13)"},{"comment":"The numerical validation does not test the theorem as stated. The non-stationary scenarios (random fluctuations around a fixed mean and cyclic bias with a dominant outcome) are special processes in which the logarithms of |G_j| have a negative drift, so the missing summability condition happens to hold. They do not probe the failure mode exhibited by the Section 4 counterexample, where the probabilities approach a perfectly biased limit fast enough that the Fourier product saturates at a positive constant. The simulations therefore cannot provide evidence for the claimed general robustness.","section":"Section 6 (Non-Stationary simulations)"},{"comment":"The claim that the non-stationary convergence rate is 'also exponential, governed by the average behavior of |G_j|' is not proved and, as written, cannot be true without additional assumptions on the rate at which the sequence approaches degeneracy. The QRNG proposal in Section 7 explicitly rests on robustness to detector-efficiency drift and illumination variation; slow drifts can be exactly the kind of process for which the posited theorem fails, so the practical motivation inherits the Section 4 gap.","section":"Section 5, last paragraph; Section 7"}],"minor_comments":[{"comment":"The sentence 'the complex numbers omega^{jl} for l = 0, ..., m-1 are distinct' is inaccurate when gcd(j,m) > 1; the parenthetical that follows correctly notes periodicity, but the wording should be cleaned up. The strict inequality |A_j| < 1 remains valid because not all nonzero terms have a common phase, but this requires a separate argument in the repeated-root case.","section":"Section 3"},{"comment":"The last bullet under 'The TVD values at specific N' repeats the entropy text ('Shows an increase from 1.0979 bits at N=1 ... to 2.3218 bits at N=100') instead of reporting the TVD value for the Non-Stationary Cyclic Bias scenario.","section":"Section 6, TVD bullet list"},{"comment":"The log-scale y-axis labels appear corrupted (reading like 10^{-203}, 10^{-176}, etc.), and those values are inconsistent with the TVD values quoted for N=100 in the text. Please regenerate the figure or correct the axis formatting.","section":"Figure 3"}],"recommendation":"reject","confidential_remarks":"The core mathematical claim of the paper is false in the non-stationary regime, and that regime is the paper's advertised contribution. A repair is conceivable by adding explicit summability or uniform decay conditions, but this would shrink the main theorem to a much narrower statement and would require a rewrite of the abstract, Sections 4, 5, 6, and 7, including the practical QRNG motivation. In its present form the manuscript does not meet the bar for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. The stationary case is fine: the Fourier/root-of-unity argument for IID sources is correct, and the rho^N error bound is a valid elementary derivation. If you need to debias a known stationary m-ary source by summing mod m, that part works. The numerics for the stationary scenarios are consistent with that, though the tiny TVD values in Figure 3 look impossible with 20,000 blocks — the sample floor for TVD with that many draws is around 10^-3, so either the axes are mislabeled or the simulation is not doing what the text says.\n\nThe problems are the non-stationary claim and the novelty claim. Section 4 states the correct sufficient condition — divergence of sum_j (1-|G_j(omega^r)|) — but then Eqs. 12-13 conclude convergence from positivity alone. That inference is invalid. For m=2, take p_j(0)=1-2^{-(j+1)}, p_j(1)=2^{-(j+1)}; then P(r_f=0) tends to (1+C)/2 with C = product_j (1-2^{-j}) ≈ 0.2888, not 1/2. All probabilities are positive, so the theorem's hypotheses are met and its conclusion fails. This is not a small gap; it is one of the two headline results.\n\nOn novelty: Eq. (9) is the standard Fourier inversion formula for convolution powers on Z_m. The convergence of full-support random walks on finite abelian groups is textbook, and the m=2 case is the well-known XOR trick for biased coins. The paper cites none of this and asserts novelty in Section 1 without engaging the literature. That is a serious citation-pattern problem.\n\nThere is also a copy-paste error in Section 6: the last bullet for Non-Stationary Cyclic Bias (m=5) reports the same entropy values as the previous scenario, not the m=5 numbers.\n\nSo the paper has a correct but elementary IID result, a false non-stationary theorem, and unsupported novelty claims. It deserves rejection rather than conditional acceptance. If the authors repair the non-stationary claim by adding the missing summability condition and honestly position the IID result as a repackaging of known Fourier analysis, there might be a small note in it. As it stands, a serious referee would catch the counterexample and the missing references.","headline":"The IID modular-sum debiasing result is correct but standard, while the advertised non-stationary robustness theorem is false as stated and the novelty claim is unsupported.","tokens_in":13360,"tokens_out":3592,"would_cite":false,"duration_ms":29933,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60E10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For any m-sided source in which every outcome can occur, summing N draws modulo m drives the output distribution to uniform with error bounded by ((m-1)/m) rho^N, and the guarantee persists when probabilities drift across trials.","keywords":["quantum random number generator","randomness extraction","debiasing","modular arithmetic","probability generating functions","roots of unity","non-stationary sources","spatial photon detection"],"falsifier":"Run the modular sum for m = 2 with $p^{{(j)}}$(0) = 1 − $2^{{-(j+1)}}$ and $p^{{(j)}}$(1) = $2^{{-(j+1)}}$. The trial product equals ∏_{j=1}^N (1 − $2^{{-j}}$), which converges to a positive constant c ≈ 0.2888, so P(r_f = 0) tends to (1+c)/2 ≈ 0.644 rather than 1/2. This calculation shows that the divergence condition in Section 4 is doing real work; a simulation of this sequence would settle whether the paper's non-stationary theorem is meant to hold under positivity alone.","tokens_in":12274,"feed_emoji":"🎲","tokens_out":12534,"duration_ms":108024,"temperature":0.7,"pith_summary":"The paper aims to prove that one arithmetic operation—adding the outcomes of N independent draws from a biased m-sided source and reducing the sum modulo m—turns the output into an almost perfectly uniform distribution over {0, 1, ..., m-1}. The argument uses probability generating functions and roots of unity: the exact probability of each residue is 1/m plus a finite sum of contractive factors, each raised to the N-th power. Full support (every outcome has positive probability) forces every nonconstant Fourier coefficient to have magnitude strictly below 1, so the error decays exponentially with N. The paper further claims that convergence survives non-stationary sources whose probabilities change from trial to trial, provided each trial keeps every outcome possible. If correct, this gives spatial-photon quantum random number generators a trivial post-processing step that removes detector and illumination bias without discarding any raw events.","feed_headline":"Adding biased rolls modulo m yields uniform output fast","feed_subtitle":"This integer-only sum trick provably removes bias from any m-sided source—even one whose bias changes over time.","key_machinery":"The load-bearing identity is the root-of-unity extraction formula P(r_f ≡ k mod m) = (1/m) Σ_{r=0}^{m-1} $ω^{{-rk}}$ ∏_{j=1}^N G_j(ω^r), where G_j(t) = Σ_k $p^{{(j)}}$_k t^k is the probability generating function of trial j and ω = $e^{{2πi/m}}$. The r = 0 term contributes exactly 1/m; every r ≠ 0 term is a product of Fourier coefficients G_j(ω^r), each of magnitude ≤ 1, and strictly less than 1 under full support. The rate parameter ρ = max_{r≠0} |Σ_l p_l $ω^{{rl}}$| sets the exponential contraction rate. The machinery converts a question about residue statistics of a sum into a question about whether a product of complex numbers on the unit disk contracts to zero.","core_discovery":"The central discovery is that modular summation is itself a debiasing map. For independent draws with fixed probabilities p_k > 0, the modular sum r_f satisfies P(r_f ≡ k mod m) = (1/m) Σ_{j=0}^{m-1} $ω^{{-jk}}$ (Σ_l p_l $ω^{{jl}}$)^N, and for every j ≠ 0 the base A_j = Σ_l p_l $ω^{{jl}}$ has |A_j| < 1. The deviation from uniform is therefore bounded by ((m-1)/m) ρ^N with ρ = max_{j≠0} |A_j| < 1. The paper extends the statement to time-dependent probabilities, asserting that the product ∏_j G_j(ω^r) vanishes and the output distribution becomes uniform whenever $p^{{(j)}}$_k remains positive for every trial, so arbitrary initial bias and drift both wash out.","pith_inferences":["Going beyond the paper, the bound can be inverted into a practical block-size rule: to guarantee |P - 1/m| ≤ ε, take N ≥ ln(ε m/(m-1)) / ln ρ, which an online extractor could compute on the fly.","Going beyond the paper, the product structure suggests a natural extension where each trial has its own alphabet size but a common final modulus; convergence would still hold as long as the Fourier factors contract.","Going beyond the paper, the method as stated assumes independent trials; whether modular summation remains a debiaser for weakly dependent sources (e.g., fast-mixing Markov chains) is an open testable extension that the current proof does not cover."],"forward_implications":["A biased quantum source with m spatial detection regions can be debiased by a simple running counter that adds and reduces modulo m, removing the need to calibrate per-detector efficiency or illumination.","No raw samples are discarded: every draw contributes to the final block output, so the output rate approaches log2 m bits per block as N grows.","The error bound ((m-1)/m) ρ^N gives a direct way to choose N for a target statistical distance to uniform, even for strongly biased sources.","Under the stated average-decay condition, slow parameter drift or cyclic bias does not prevent convergence to uniformity; only a persistent failure to make every outcome possible would leave a residual bias.","The method is alphabet-agnostic: the same proof works for any discrete source with m outcomes, including non-binary quantum detection maps."],"supporting_citations":[{"why":"Supplies the probability generating function and root-of-unity machinery used to write the exact distribution of the modular sum.","marker":"[18]"},{"why":"The classic rejection-based debiasing baseline against which modular summation is claimed to improve data efficiency and stationarity robustness.","marker":"[10]"},{"why":"Defines the spatial photon-detection QRNG scheme that Section 7 identifies as the main practical application.","marker":"[15]"},{"why":"Earlier spatial-detection QRNG demonstration grounding the proposed application in established experimental practice.","marker":"[16]"},{"why":"Supplies the total variation distance used in Section 6 to measure and validate convergence numerically.","marker":"[20]"}],"fun_headline_variants":["Modular addition of biased trials yields provably uniform randomness","Debias any m-sided source by summing outcomes modulo m","Sum biased rolls, take mod m, and bias decays exponentially","Robust debiasing: modular sum defeats time-dependent noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The non-stationary robustness claim relies on an extra condition beyond positivity: the per-trial deviations from perfect uniformity must accumulate, in the sense that ∑_j (1 − |G_j(ω^r)|) diverges for each nonzero frequency; without that condition, a trial sequence can keep every outcome positive yet fail to converge to uniform.","fun_headline_variants_meta":{"raw":{"variants":["Modular addition of biased trials yields provably uniform randomness","Debias any m-sided source by summing outcomes modulo m","Sum biased rolls, take mod m, and bias decays exponentially","Robust debiasing: modular sum defeats time-dependent noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1532,"prompt_tokens":952,"completion_tokens":580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":511}},"tokens_in":568,"tokens_out":580,"duration_ms":6033,"temperature":1.0,"reasoning_tokens":511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:48:55.234308+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the modular sum for m = 2 with $p^{{(j)}}$(0) = 1 − $2^{{-(j+1)}}$ and $p^{{(j)}}$(1) = $2^{{-(j+1)}}$. The trial product equals ∏_{j=1}^N (1 − $2^{{-j}}$), which converges to a positive constant c ≈ 0.2888, so P(r_f = 0) tends to (1+c)/2 ≈ 0.644 rather than 1/2. This calculation shows that the divergence condition in Section 4 is doing real work; a simulation of this sequence would settle whether the paper's non-stationary theorem is meant to hold under positivity alone.","supporting_citations":[{"cited_title":"Univariate discrete distributions","cited_arxiv_id":null,"evidence_quote":"Supplies the probability generating function and root-of-unity machinery used to write the exact distribution of the modular sum."},{"cited_title":"Various techniques used in connection with random digits","cited_arxiv_id":null,"evidence_quote":"The classic rejection-based debiasing baseline against which modular summation is claimed to improve data efficiency and stationarity robustness."},{"cited_title":"Free-space quantum random number generation based on spatial detection of single photons","cited_arxiv_id":null,"evidence_quote":"Defines the spatial photon-detection QRNG scheme that Section 7 identifies as the main practical application."},{"cited_title":"Quantumrandom-numbergeneration and key sharing","cited_arxiv_id":null,"evidence_quote":"Earlier spatial-detection QRNG demonstration grounding the proposed application in established experimental practice."},{"cited_title":"Introduction to nonparametric estimation","cited_arxiv_id":null,"evidence_quote":"Supplies the total variation distance used in Section 6 to measure and validate convergence numerically."}],"review_version":1}