{"id":"b1d57102-6220-403f-be0f-db6c329c4dc4","arxiv_id":"2607.18494","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The oscillating corrections to the binomial mean absolute deviation are Bernoulli polynomials evaluated at the fractional part of Np, to all orders, with a rigorous bracketing bound.","lead":"This paper derives complete asymptotic expansions for the binomial distribution's mean absolute deviation, with coefficients that oscillate and are governed by Bernoulli polynomials. It also proves a two-sided bound and an enveloping series at integer means.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniformity in Lemma 2.1 is the load-bearing assumption; the proof asserts the shifted-remainder bound rather than deriving it, and Theorems 3.1/4.2 inherit this dependence.","rationale":"I read the main argument in good faith. The algebra in Theorems 3.1 and 4.2 is internally consistent: the gamma-quotient constants (Θ,U,K) combine correctly with the probability prefactor, the Bernoulli-polynomial identities are applied correctly, and the term-by-term cancellation of the elementary part with log(1+h/(Np)) is exact. The parity split (23)-(24), the Frame-Johnson identification, the p=1/2 parity oscillation, and the Cesàro-mean Proposition 6.1 all check out. The enveloping theorem (5.5) has a valid sign-definite Binet-kernel mechanism, and Proposition 5.1's Robbins-bracket computation is correct. The only point where the paper asks the reader to accept a nontrivial uniformity statement without a fully displayed proof is Lemma 2.1. This is the same point the Reader's weakest_assumption identifies. It is a real soft spot, but it is standard and almost certainly fillable, so it does not overturn the central claim; it justifies keeping the verdict CONDITIONAL rather than ACCEPT. The editorial artifacts noted by the Reader are irrelevant to the mathematics but support the same conditional rating. Hence my recommendation is UNCHANGED: the Reader's verdict already captures the state of the paper.","tokens_in":16890,"tokens_out":26444,"duration_ms":235971,"concrete_test":"Derive Lemma 2.1 via Binet's first formula: log Γ(z) = (z-1/2)log z - z + 1/2 log 2π + ∫_0^∞ φ(t)e^{-zt}dt. For z=λx+u, expand the polynomial prefactor in powers of x^{-1} and expand e^{-ut} in powers of t up to order M; the remainder is then an explicit absolutely convergent integral. Bound this integral uniformly for u∈H (in particular H=[0,2]) and λ∈[c,C] to obtain (11) with an explicit constant C_{M,H,c,C}. If the bound reproduces the claimed O(x^{-M-1}) uniformity, the gap is closed; if extra restrictions on u or λ appear, Theorem 4.2's uniform remainder needs amendment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (Theorem 4.2) inherits its uniform remainder from Lemma 2.1, specifically from estimate (11) for gamma quotients with unequal scalings. The proof in Section 2 derives (11) by applying the unshifted Stirling expansion at w=λ_j x+u_j and then re-expanding log w = log(λ_jx)+log(1+u_j/(λ_jx)) and w^{1-2n} = (λ_jx)^{1-2n}(1+u_j/(λ_jx))^{1-2n}. The statement that 'collecting powers ... yields (6) in truncated form' with a uniform O(x^{-M-1}) remainder is plausible but not demonstrated: the proof does not exhibit the remainder after re-expansion, nor does it track how the constants depend on H and c. Two facts have to be checked: (i) the Taylor remainders in z=u_j/(λ_jx) and (1+z)^{1-2n} are uniform for u_j∈H and λ_j≥c once x≥x_0; (ii) the tail of the unshifted Stirling series can be truncated after finitely many n without introducing x-dependence that spoils the x^{-M-1} order. Since h_N oscillates in [0,1] and p,q vary over a compact subset of (0,1), the entire uniformity statement of Theorem 4.2 — not just the formal coefficient computation — rests on this step. No machine-checked proof or independent derivation is supplied, so this is the weakest load-bearing link.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper derives complete asymptotic expansions, as N→∞, for the binomial local mass at a bounded lattice displacement and for the mean absolute deviation of X∼Bin(N,p). Starting from De Moivre's exact formula, the MAD is reduced to a single local mass at ν=⌈Np⌉. The authors develop a gamma-quotient expansion with unequal scalings (Lemma 2.1) and show, in Theorems 3.1 and 4.2, that the coefficients are built from Bernoulli polynomials evaluated at the oscillating displacement h_N=⌈Np⌉−Np; the elementary part cancels term by term. At integer mean, the expansion collapses to a sign-alternating odd series, and Theorem 5.5 proves that this series is enveloping via a sign-definite Binet-kernel representation. The paper also gives an elementary two-sided bound (Proposition 5.1) and Cesàro means of the oscillating coefficients (Proposition 6.1).","tokens_in":17205,"tokens_out":20923,"duration_ms":156620,"significance":"If the main results are correct, the paper solves a classical problem in full generality: it extends the first-order Frame–Johnson correction to all orders and identifies the Appell-shift mechanism. The exact cancellation, the enveloping property at integer means, and the simple proof of the two-sided bound are attractive. The paper is self-contained and uses only classical tools, and it gives a nice illustration of Bernoulli polynomials in a probabilistic context. The main risk is the uniformity in Lemma 2.1, on which the central Theorems 3.1 and 4.2 rely.","major_comments":[{"comment":"The uniformity statement (11) is load-bearing for Theorems 3.1 and 4.2, but its proof is only sketched. After truncating the unshifted Stirling series, the re-expansion of (1+u_j/(λ_jx))^{1-2n} produces contributions at all orders in x^{-1}. The text asserts that 'collecting the powers' yields the truncated shifted expansion with a uniform O(x^{-M-1}) remainder, but it does not show that each coefficient of x^{-k}, k≤M, involves only finitely many n, nor does it track the dependence of the remainder on H and c. Please expand this argument or give a precise reference for the uniform unequal-scaling version. This is necessary because h_N does not converge and p ranges over a compact set.","section":"§2, Lemma 2.1, Eq. (11)"}],"minor_comments":[{"comment":"The section heading 'A veraging the oscillation' appears to be a typo; it should be 'Averaging the oscillation'.","section":"§6, heading"},{"comment":"The exact identity (39) is stated without derivation. Adding the algebra showing that the +1 shifts in the gamma arguments assemble into the leading factor would improve readability.","section":"§5, Theorem 5.5, Eq. (39)"},{"comment":"In the phrase 'the remainder is O(x^{-M-1}) uniformly', it would be useful to restate explicitly that the implied constant depends on M, H, c, and C.","section":"§2, proof of Lemma 2.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of the journal and the main results appear correct. The only substantive concern is the terse proof of Lemma 2.1; this is addressable in revision. The numerical and historical context is sound, and the exposition is generally clear."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First the punchline. The paper really does find the full coefficient sequence for E|X−Np|: Bernoulli polynomials in the oscillating displacement h_N, with the non-Bernoulli part cancelled term by term against De Moivre's prefactor. That is new, and it is the right framework for this problem. The first correction goes back to Frame and Johnson, and the O(N^{-1}) local mass expansion is in Ouimet, but the Appell-shift mechanism that generates all orders at once is this paper's contribution. The parity collapse at integer mean and the enveloping odd series are also genuinely nice.\n\nThe central derivation is clean and largely self-contained. The gamma-quotient lemma is standard, and the cancellation in Theorem 4.2 is elegant. The check against central binomial coefficients for p=1/2 works. The numerical table is consistent with a true integer-power expansion. The paper is also honest about the prior literature.\n\nThe soft spots are two. The first is the proof of uniformity in Lemma 2.1. The proof applies the unshifted Stirling expansion at w = λ_j x + u_j and then says 'collecting the powers' yields the shifted truncated expansion with uniform O(x^{-M-1}) remainder. That step needs the remainder after Taylor re-expansion and after truncating the infinite Stirling tail to be bounded uniformly in the shift and in the scalings. It is plausible, but as written the constant is not exhibited. Since Theorem 4.2 inherits that uniformity, this is load-bearing. I don't think it is wrong — the ingredients are classical — but it needs to be written out. Second, Theorem 5.5's sign-definite integral argument is good, but the interchange of the Binet integral with the geometric-series representation is only briefly justified; that is minor.\n\nThe manuscript also has editorial artifacts in the gathered version (duplicated abstract, formatting glitches). They do not affect the substance but should be fixed.\n\nAll in all, this is a serious, coherent paper. Send it to a referee. If the referee fills or tames the uniformity gap, it is a solid citable contribution for anyone working on binomial asymptotics. I would not desk-reject it.","headline":"Genuinely new coefficient structure for binomial MAD, with a uniformity gap in Lemma 2.1 that needs a referee's attention; otherwise a solid, citable paper.","tokens_in":17726,"tokens_out":3259,"would_cite":true,"duration_ms":32818,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["41A60","11B68","60C05","33B15"],"pacs":[],"model":"deepseek-v4-flash","headline":"For any binomial distribution, the mean absolute deviation of X from its mean has a complete asymptotic expansion in powers of 1/N, with coefficients that are explicit Bernoulli polynomials of the fractional displacement of the mean.","keywords":["binomial distribution","mean absolute deviation","asymptotic expansion","Bernoulli polynomials","Appell polynomials","gamma quotient","lattice correction","Stirling series"],"falsifier":"Fix a truncation order M and a compact subinterval K of (0,1). Take a sequence of N with h_N tending to 1 and p tending to the left endpoint of K, and compute the exact remainder R_M(N,p) = log(E|X−Np|/√(2Npq/π)) − Σ_{k=1}^M a_k(h_N;p)/N^k. If |R_M(N,p)|·N^{M+1} is unbounded over such a sequence, the uniform statement (26) fails. Separately, for N p integer, checking a single violation of the strict envelope inequality (38) would refute the enveloping claim.","tokens_in":16727,"feed_emoji":"📊","tokens_out":6709,"duration_ms":56548,"temperature":0.7,"pith_summary":"The paper establishes that for X~Bin(N,p), the quantity E|X−Np| has a full asymptotic expansion in powers of N−1 relative to the Gaussian leading term √(2Npq/π), with coefficients that depend on the bounded, non-convergent displacement h_N = ⌈Np⌉−Np through Bernoulli polynomials. The route is exact: De Moivre's identity collapses the expectation to a single binomial mass, which is then written as a ratio of three gamma functions with unequal linear scalings. The central structural cancellation—the elementary part of the shift cancels term by term against the prefactor's logarithm—leaves coefficients that are pure Bernoulli polynomials, so the expansion is both complete and algebraically transparent. This fills a gap that stopped after the first correction, and it supplies rigorous two-sided bounds as well as an enveloping sign-alternating series when the mean is an integer.","feed_headline":"All-order expansion found for binomial mean deviation","feed_subtitle":"The Gaussian guess is only the leading term; higher-order corrections are explicit Bernoulli polynomials of the fractional displacement of t","key_machinery":"The central object is a gamma-quotient expansion for ratios of the form Γ(N+1)/(Γ(Np+h+1)Γ(Nq−h+1)) with arbitrary real scalings p,q, stated with uniformity in a compact set of shifts. It extends the usual Stirling series—where all arguments share the same scale—by weighting each contributing Bernoulli polynomial B_{n+1}(u) by the (−n)-th power of that factor's own scaling λ_j. This expansion is combined with two structural identities of Appell sequences: the difference identity B_m(h+1)=B_m(h)+m h^{m−1} and the reflection identity B_m(1−h)=(−1)^m B_m(h), which reduce the three gamma shifts to a single Bernoulli polynomial B_{k+1}(h) plus an elementary residue. The term-by-term cancellation","core_discovery":"The paper's main theorem states that, as N→∞, uniformly for p in a compact subinterval of (0,1), E|X−Np| = √(2Npq/π) exp(Σ_{k≥1} a_k(h_N;p)/N^k), with a_k(h;p) = [(-1)^{k+1}/(k(k+1))] [ B_{k+1} − (p^{-k} + (-1)^{k+1} q^{-k}) B_{k+1}(h) ]. Here B_{k+1} are the Bernoulli polynomials, h_N = ⌈Np⌉−Np is the bounded, oscillating displacement of the mean above an integer, and all powers of N are integer relative to the leading factor—no half-integer corrections occur at any depth. The derivation reduces the expectation, via the exact identity, to the binomial mass at the first lattice point above the mean, writes that mass as a gamma quotient, and expands it using a version of the Stirling series w","pith_inferences":["The same gamma-quotient machinery with unequal scalings is likely to extend to the multivariate analogue—the expected total-variation-type statistic Σ_j |X_j − Np_j| for multinomial variables—where no closed form exists; the single-bin expansion may serve as the exactly solvable template for the general lattice Edgeworth expansion.","The cancellation between the prefactor's logarithm and the difference-residue of the Bernoulli shift suggests a general principle for Appell sequences: whenever a prefactor coincides with the argument of a shifted gamma quotient, the non-polynomial residue cancels, leaving pure Appell polynomials; this could be tested on other Appell systems such as Euler polynomials.","The enveloping property is proven only for integer means; for non-integer means the sign of the remainder after two terms is expected to oscillate with h_N, but a stochastic or averaged version might hold—examining the Cesàro means of the truncation errors would be a concrete testable extension."],"forward_implications":["The classical first-order correction to the Gaussian mean deviation, known since the 1940s, is recovered as the coefficient a_1(h_N;p) of the new all-order expansion; all higher-order corrections are now generated by a single recursion.","Because the coefficients carry the oscillating factor h_N, the expansion explains the non-uniformity of the Gaussian approximation near p=0 or 1: coefficients grow like (pq)^{-m} as the endpoints are approached.","At an integer mean (Np ∈ Z) the logarithmic expansion collapses to a sign-alternating series in odd powers of N^{-1}; the paper proves that this series is enveloping, so successive truncations bracket the true expectation, giving rigorous two-sided inequalities.","A universally valid two-sided bound of logarithmic width O(N^{-2}) is obtained by combining the exact reduction with Robbins's sharpened Stirling inequality, holding uniformly in the lattice position for all interior p.","The Cesàro means of the oscillating coefficients are explicit: for irrational p the mean is the smooth part δ_m(p), and for rational p=a/b the mean over a period is obtained by replacing B_j(h) with b^{-j}B_j, a direct consequence of the Bernoulli multiplication theorem."],"fun_headline_variants":["Binomial mean deviation expanded to all orders with Bernoulli polynomials","Complete asymptotic series for binomial mean absolute deviation","No half-integer terms in full binomial deviation expansion","Bernoulli polynomials govern full binomial deviation expansion","Exact all-order expansion for binomial mean absolute deviation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the gamma-quotient expansion is uniform in the bounded displacement h and in the scalings p, q bounded away from zero; if the remainder cannot be uniformly controlled as p approaches 0 or 1, then the expansion is empty because the displacement h_N does not converge.","fun_headline_variants_meta":{"raw":{"variants":["Binomial mean deviation expanded to all orders with Bernoulli polynomials","Complete asymptotic series for binomial mean absolute deviation","No half-integer terms in full binomial deviation expansion","Bernoulli polynomials govern full binomial deviation expansion","Exact all-order expansion for binomial mean absolute deviation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2745,"prompt_tokens":921,"completion_tokens":1824,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":1752}},"tokens_in":665,"tokens_out":1824,"duration_ms":12178,"temperature":1.0,"reasoning_tokens":1752,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T15:12:28.659443+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix a truncation order M and a compact subinterval K of (0,1). Take a sequence of N with h_N tending to 1 and p tending to the left endpoint of K, and compute the exact remainder R_M(N,p) = log(E|X−Np|/√(2Npq/π)) − Σ_{k=1}^M a_k(h_N;p)/N^k. If |R_M(N,p)|·N^{M+1} is unbounded over such a sequence, the uniform statement (26) fails. Separately, for N p integer, checking a single violation of the strict envelope inequality (38) would refute the enveloping claim.","supporting_citations":[],"review_version":1}