{"id":"52d96722-1439-419f-ade8-295a19f8d11f","arxiv_id":"1909.02556","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For stationary AR(1) segments of length 2, 3 and 4, the variance of the number of sign changes is computed exactly; the independent interval approximation matches exactly for n=2,3 but deviates slightly for n=4, especially for negative correlation.","lead":"This math note derives the exact variance of the number of sign changes in a short segment of an autoregressive time series for up to four observations. It compares this exact result with an approximate model and finds a close but imperfect match, with an unresolved asymmetry for negative correlations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Exact V(S4) relies on Cheng's I(h,x) for negative arguments outside the stated domain 0<x<h^2<1; the paper does not justify this extension, so the central formula is conditional on an unproved analytic continuation.","rationale":"The paper does substantial correct work: it reproduces the known n=2 and n=3 variances, derives E(S4) from Rice's theorem, and gives an explicit orthant-probability decomposition for V(S4). The IIA comparison is honestly presented, including the unexplained asymmetry for negative ρ. The main risk is not in the combinatorics but in the unproved external formula for I(h,x) and its application outside the quoted domain. The reader's weakest assumption already identifies Cheng's formula and negative ρ; my stress-test sharpens it to the specific negative-x calls a^2 b = −ρ^3 in f(ρ,−ρ), g(ρ,−ρ), and f(−ρ,−ρ). This is a conditional gap: it is very likely fixable by a short evenness argument, but until that argument is supplied or the numerical check is performed across a range of ρ, including negative ρ, the exactness of the headline formula is not fully demonstrated. Since the reader's CONDITIONAL verdict already expresses exactly this conditionality, I recommend no change to the verdict.","tokens_in":8369,"tokens_out":10825,"duration_ms":111863,"concrete_test":"Use an independent quadrature routine (e.g., R's mvtnorm with high accuracy settings) to recompute the six probabilities f(ρ,ρ), g(ρ,ρ), f(ρ,−ρ), g(ρ,−ρ), f(−ρ,ρ), f(−ρ,−ρ) for ρ = 0.5, ρ = 0.8, ρ = −0.5, and ρ = −0.8, and compare every value to the paper's formulas evaluated from the displayed Li₂ expression. Pay particular attention to the negative-x entries f(ρ,−ρ), g(ρ,−ρ), and f(−ρ,−ρ); if any of these disagrees beyond quadrature tolerance, the V(S4) formula is invalid. A complementary check is to verify I(h,−x)=I(h,x) by substituting t→−t in the integral and then comparing the Cheng expression at x=0.4 with x=−0.4 for h=0.8.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the closed-form V(S4) displayed at the end of Section 1. It is built from the orthant probabilities f(a,b) and g(a,b), each containing the term I(a, a^2 b)/(4π^2), where I(h,x) is Cheng's integral quoted at the start of Section 1. As stated, I(h,x) is valid for 0 < x < h^2 < 1. But the V(S4) formula requires evaluating f(ρ,−ρ), g(ρ,−ρ), f(−ρ,−ρ), and g(−ρ,−ρ); in each of these terms a^2 b = −ρ^3 < 0. The paper only says 'We assume that |a| < 1 and |b| < 1'; it never gives a domain condition on x or proves that the quoted Li₂ expression extends to negative x. The integral itself is even in x because the arcsin integrand is odd, so I(h,−x)=I(h,x) is probably the needed continuation, but that is not recorded and the complex dilogarithm branch structure could make the naive extension fail. Therefore, as written, exactness of V(S4) for all ρ in (−1,1) is not established; the visible ρ=1/2 numerical check does not independently exercise the negative-x terms. This is a proof gap rather than a witnessed contradiction, but it is the load-bearing step: if the extension is invalid, the central expression fails.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript derives exact formulas for the mean and variance of the number of sign changes in a short segment (n ≤ 4) of a stationary Gaussian AR(1) process with correlation ρ. For n = 2 and n =  ít3 the paper recovers known results; for n = 4 it presents a closed-form expression for V(S4) in terms of quadrivariate normal orthant probabilities f and g, which are built from Cheng's dilogarithm evaluation of a one-dimensional integral I(h,x). The paper then compares V(S4) with the independent-interval approximation (IIA) of Nyberg et al., finds close agreement for positive ρ and a small discrepancy for negative ρ, and explains that the IIA asymmetry for negative ρ may indicate an assumption ρ > 0 in the model. An appendix discusses the additional integral needed for n = 5 and reports that it resists closed-form evaluation except in a special case.","tokens_in":8666,"tokens_out":11916,"duration_ms":114432,"significance":"If correct, the V(S4) formula is a useful exact benchmark for zero-crossing statistics of AR(1) processes, extending known results beyond n = 3. The derivation is transparent and parameter-free: it uses only standard orthant probability formulas and Cheng's integral, with no fitting and no use of the model to define theory. The paper is commendably explicit about the assumptions behind the IIA comparison and about numerical verification using R's pmvnorm. However, the exactness of V(S4) rests on a domain extension that is not justified in the manuscript, so the central claim is not fully established as written.","major_comments":[{"comment":"The quoted evaluation of I(h,x) is stated for 0 < x < h^2 < 1, but the displayed formula for V(S4) immediately applies it to f(ρ,−ρ), g(ρ,−ρ), and f(−ρ,−ρ), which involve I(ρ,−ρ^3) and I(−ρ,−ρ^3), i.e., negative upper limits. The paper does not justify the analytic continuation to negative x, nor does it give a domain condition for f and g beyond |a| < 1 and |b| < 1. The numerical check at ρ = 1/2 involves one negative argument (x = −1/8) and is reassuring, but it is not a proof for all ρ ∈ (−1,1). Because V(S4) is the central new result, this is a load-bearing gap. The author should add a short argument (e.g., the integrand is odd, so I(h,−x) = I(h,x) if the right-hand side is continued accordingly) or cite a source that covers −h^2 < x < 0, and state the branch of Li2 used for negative x.","section":"Section 1, display for I(h,x) and display for V(S4)"}],"minor_comments":[{"comment":"The figure is not included in the manuscript text, so the quantitative claims about the largest separation between theory and IIA cannot be verified by the reader; providing the figure or a data table would improve reproducibility.","section":"Section 2, Figure 1"},{"comment":"The sentence 'Closed-form variance expressions become impossible for n ≥ 5 (see the appendix)' overstates what is shown; the appendix demonstrates only that one particular integral J(ρ, ρ^2, ρ^4) is not evaluated in closed form. Suggest rewording to 'we were unable to obtain' or 'no closed form is known to us'.","section":"Introduction, paragraph 1"},{"comment":"There are minor typographical errors, e.g., 'mom ents' in the abstract, and some covariance matrices are typeset with misalignments; these should be corrected.","section":"Abstract and body"},{"comment":"The acknowledgments state that R's pmvnorm was used for verification, but no direct numerical comparison is shown in the text. A small table comparing the closed-form V(S4) with pmvnorm over a grid of ρ values, including negative ρ, would substantially strengthen the evidence for the extension of I(h,x).","section":"Section 1 and Acknowledgments"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short research note with a potentially correct new variance formula for n = 4. The main obstacle is the unstated analytic continuation of Cheng's integral to negative arguments; this is fixable with a short parity argument or an additional citation. I would not require a full proof of Cheng's formula itself, but the domain issue must be settled for the central claim to be rigorous. The numerical verification at a single ρ is not sufficient by itself. The paper fits the journal's scope as a mathematical note, but the 'impossible' claim for n ≥ 5 should also be softened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Steve, read this one. It's a small but honest paper: exact variance of the number of sign changes in n=4 consecutive AR(1) observations, using Cheng's quadrivariate orthant probability formulas. The n=2 and n=3 variances were known; the n=4 expression is new and, as far as I can tell, correct. The paper shows its work, checks ρ=0 and ρ=1/2, and computes values to 25 digits with pmvnorm verification. The IIA comparison is a side dish: it reproduces theory exactly for n=2,3, and for n=4 the fit is close except for an unexplained asymmetry at negative ρ. The author flags that himself and doesn't oversell.\n\nThe main gap is domain. Cheng's integral I(h,x) is quoted for 0<x<h^2<1, but the orthant probabilities f(a,b) and g(a,b) require x=a^2 b, which is negative for several terms in the V(S4) formula (e.g., f(ρ,−ρ) has x=−ρ^3). The integral's integrand is odd, so I(h,−x)=I(h,x) is an obvious continuation, and the numerical checks at ρ=1/2 do exercise negative x. But the paper never says this. As written, the exactness claim for all ρ∈(−1,1) rests on an unstated analytic continuation of the dilogarithm expression. That's fixable—one sentence plus a check for negative x—but a referee should catch it. Also, the paper leans on Cheng's formula without proof; that's acceptable as a citation but the typo correction means the reader should double-check.\n\nThe IIA asymmetry for negative ρ is left unresolved; the paper suggests replacing ρ by |ρ| but isn't sure. That's an honest limitation, not a flaw in the exact result.\n\nOverall: a solid, narrow contribution for people who do zero-crossing statistics or orthant probabilities. It doesn't open new lines, but it's exactly the kind of benchmark that's useful to have. I'd send it to a short-paper journal; with a small revision addressing the domain issue, it's publishable. Bring it to reading group only if your group is into this kind of thing.","headline":"A careful, narrow computation of the variance of sign changes in AR(1) segments; the n=4 formula is new and checkable, but the paper should state the extension of Cheng's integral to negative arguments.","tokens_in":9190,"tokens_out":3918,"would_cite":false,"duration_ms":39350,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G10","62M10","33B30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper derives an exact closed-form formula for the variance of the number of sign changes in four observations of a stationary AR(1) Gaussian process.","keywords":["sign changes","AR(1) process","orthant probabilities","dilogarithm","variance","independent interval approximation","zero crossings","Gaussian time series"],"falsifier":"Compute V(S4) at ρ=-0.5 by direct high-precision numerical integration of the four-dimensional Gaussian orthant probabilities p_{e1e2e3e4}, without using the dilogarithm formula, and compare with the paper's expression; any disagreement beyond integration error refutes it. The same check can be done at ρ=-0.897, where the approximation gap is largest, since a mismatch there would localize the failure.","tokens_in":8154,"feed_emoji":"📉","tokens_out":8204,"duration_ms":82484,"temperature":0.7,"pith_summary":"This paper establishes an exact closed-form formula for the variance of the number of sign changes among four consecutive observations of a stationary first-order autoregressive Gaussian series with correlation parameter ρ. The formula is assembled from quadrivariate normal orthant probabilities, each expressed through a dilogarithm integral, and it supplements the known linear mean formula. The resulting variance is symmetric in ρ, attains its maximum 3/4 at ρ=0, and takes the value 0.721407566... at ρ=1/2. Against the independent-interval approximation, the new expression matches exactly for segments of length two and three and comes very close for length four, with the largest discrepancy near ρ≈-0.897.","feed_headline":"Exact variance formula derived for sign changes in AR(1) segments","feed_subtitle":"Dilogarithm orthant probabilities give the n=4 variance and expose an approximation's weak spot.","key_machinery":"The argument turns on a dilogarithm integral I(h,x) = ∫₀ˣ arcsin((1−h²)t/(h²−t²))/√(1−t²) dt, expressed as a sum of complex dilogarithms. Two orthant-probability functions f(a,b) and g(a,b) are defined from this integral for correlation matrices with entries a, b, ab, and a²b. Every sign pattern in a length-four AR(1) segment, after flipping coordinates, has a covariance matrix of one of those two forms, so every p_{e1 e2 e3 e4} reduces to f or g with arguments ±ρ. Summing the resulting probabilities with the appropriate multiplicities yields the variance.","core_discovery":"The central result is that V(S4) = 4g(ρ,ρ)+2f(ρ,−ρ)+16g(ρ,−ρ)+8f(−ρ,ρ)+18f(−ρ,−ρ) − 9[1/2 − arcsin(ρ)/π]^2, where f and g denote probabilities that four variables with two specific correlation matrices are all positive. These are evaluated in closed form via a dilogarithm integral, so the variance itself is a closed expression in ρ for every |ρ|<1. The paper verifies symmetry, the maximum 3/4 at ρ=0, and numerical agreement at ρ=1/2, and contrasts the formula with the independent-interval approximation: the approximation is exact for n=2 and n=3 variances and nearly exact for n=4, with the biggest gap of about 0.036 near ρ=-0.897 and about 0.002 near ρ=0.763.","pith_inferences":["A similar f/g sign-flip decomposition should extend to other local pattern counts in AR(1) segments, such as peaks or turning points, where the same covariance forms arise, though this is not pursued in the paper.","The exact V(S4) formula gives a cheap benchmark for validating numerical routines for quadrivariate normal probabilities, especially at negative ρ where the approximation is weakest.","The IIA model's unexplained asymmetry at negative ρ could be tested directly by substituting |ρ| in the recursion; if the fit improves, the model's physical derivation likely assumed positive correlation.","Because the two intractable n=5 orthant probabilities depend on a double-parameter integral J(h,k,x), symbolic progress for n≥5 may require evaluating J at h≠k, which the paper leaves open."],"forward_implications":["For n=4, the exact variance of sign changes is a closed function of ρ, so no simulation or quadrature is needed for this sample size.","The variance is symmetric under ρ→−ρ, meaning short sign-change counts cannot by themselves distinguish positive from negative serial correlation.","The independent-interval approximation exactly reproduces the variance for n=2 and n=3 and is nearly exact for n=4, supporting its use for short AR(1) segments.","The appendix identifies the two orthant probabilities needed for n=5 that do not fit the f/g pattern, explaining why closed-form expressions stop at n=4.","The mean formula (n−1) arccos(ρ)/π remains the expectation for all n, consistent with the classical zero-crossing theorem in discrete time."],"supporting_citations":[{"why":"Supplies the dilogarithm evaluation of I(h,x) from which the orthant probabilities f and g are built.","marker":"[6]"},{"why":"Establishes the mean number of sign changes as (n−1) arccos(ρ)/π, the baseline that the variance refines.","marker":"[10]"},{"why":"Provides the independent-interval approximation recursion for E(S_n²) used in the comparison.","marker":"[17]"},{"why":"Provides the inclusion-exclusion framework that shows why n=5 introduces orthant probabilities not covered by f and g.","marker":"[18]"}],"fun_headline_variants":["Closed-form variance for sign changes in AR(1) segments","Exact n=4 sign-change variance via dilogarithms","AR(1) sign-change variance: exact formula for n=4","Approximation gap revealed in AR(1) sign-change variance","Dilogarithm closed form for AR(1) sign-change variance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire variance expression relies on the dilogarithm evaluation of I(h,x) being correct for the parameter ranges needed, including negative argument ρ; the paper cites this evaluation without proving it and even notes a typo in the source, so if the corrected identity fails for some |ρ|<1, the central formula collapses.","fun_headline_variants_meta":{"raw":{"variants":["Closed-form variance for sign changes in AR(1) segments","Exact n=4 sign-change variance via dilogarithms","AR(1) sign-change variance: exact formula for n=4","Approximation gap revealed in AR(1) sign-change variance","Dilogarithm closed form for AR(1) sign-change variance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3119,"prompt_tokens":865,"completion_tokens":2254,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":481,"completion_tokens_details":{"reasoning_tokens":2165}},"tokens_in":481,"tokens_out":2254,"duration_ms":16482,"temperature":1.0,"reasoning_tokens":2165,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:46:58.096743+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute V(S4) at ρ=-0.5 by direct high-precision numerical integration of the four-dimensional Gaussian orthant probabilities p_{e1e2e3e4}, without using the dilogarithm formula, and compare with the paper's expression; any disagreement beyond integration error refutes it. The same check can be done at ρ=-0.897, where the approximation gap is largest, since a mismatch there would localize the failure.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dilogarithm evaluation of I(h,x) from which the orthant probabilities f and g are built."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the mean number of sign changes as (n−1) arccos(ρ)/π, the baseline that the variance refines."},{"cited_title":"Zero-Crossing Statistics for Non-Markovian Time Series","cited_arxiv_id":"1711.02926","evidence_quote":"Provides the independent-interval approximation recursion for E(S_n²) used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the inclusion-exclusion framework that shows why n=5 introduces orthant probabilities not covered by f and g."}],"review_version":1}