{"id":"c896e8c5-0cd3-434f-aad5-741f79f2c2d5","arxiv_id":"1908.04179","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a stationary AR(1) process, the expected value of the maximum of short contiguous segments is maximized at negative serial correlation, and the variance of the maximum increases monotonically with correlation.","lead":"This note computes the first two moments of the maximum of a short segment of a stationary Gaussian AR(1) process using formulas from Afonja (1972). It finds that the expected maximum peaks at a negative lag-one correlation, while the variance of the maximum grows throughout the correlation range.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ℓ=5 maximizer and variance monotonicity rest on unverified Afonja formulas; the paper omits the full E(M5) expression yet reports a six-decimal ρ, and its own Section 3 contains a typo in E(M3).","rationale":"The reader's weakest assumption is the dependence of the ℓ≥4 numerical results on Afonja's formulas without independent derivation; I agree. The paper provides good support for ℓ=3 (first-principles proof and limiting checks) and gives credible corrections to Afonja's constants, but the ℓ=5 maximizer is asserted without the underlying formula, and the variance monotonicity is asserted as fact while admitted to be unproved. This is a genuine reproducibility and verification gap, but not an internal contradiction. The conditional verdict is appropriate: the paper is likely correct, but should be revised to either prove the monotonicity or label it as a conjecture, and to provide the full ℓ=5 (and ℓ=6) formulas or reproducible code. My read does not change the reader's verdict.","tokens_in":9721,"tokens_out":26463,"duration_ms":209603,"concrete_test":"Independently implement the full Afonja formula for E(M5) using the corrected constants from Section 2 and the partial-correlation definitions from Section 1, evaluate E(M5(ρ)) on a dense grid (step ≤ 1e-5) over (−1,1), and recompute the maximizer. Also compute Var(M4(ρ)) = E(M4^2) − E(M4)^2 from the Section 3 closed forms and check that its derivative is positive throughout. If the recomputed ℓ=5 maximizer differs from −0.4336476843... by more than 1e-6, or if any derivative of Var(M4) is negative, the central claims fail and require correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claims for ℓ=4 and ℓ=5 are not reproducible from the manuscript. Section 3 reports a six-decimal maximizer for ℓ=5 (−0.4336476843...) but explicitly declines to record E(M5), stating it is 'too lengthy.' The ℓ=4 maximizer is based on a closed-form expression printed in a compressed form, and the paper's own Section 3 display for E(M3) contains a typo (the second term is written √(1−ρ^2)/(4π) instead of (1/2)√((1−ρ^2)/π)), demonstrating that transcription errors are a live risk in this note. The variance-monotonicity claim is likewise an unproved numerical observation, deferred with 'An intuitive reason... would be good to establish someday.' Because E(M5) is not given and no independent derivation is supplied for ℓ≥4, the reported maximizer and the monotonicity claim could both be artifacts of an undetected error in Afonja's 1972 formulas beyond the two corrections the author lists. The limiting checks at ρ→0 are necessary but not sufficient to validate the ρ-dependence of the formulas.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This note studies the maximum M_ℓ of ℓ consecutive observations of a stationary AR(1) process with standard normal white noise. It specializes Afonja's (1972) formulas for the moments of the maximum of correlated normal variables to the Toeplitz correlation structure ρ_{ij}=ρ^{|j−i|}, obtaining explicit expressions for E(M_ℓ) and E(M_ℓ^2) for ℓ≤5 (with ℓ=6 for the second moment). The paper reports that E(M_ℓ), for ℓ≥3, is maximized at an interior negative value of ρ, with closed form ρ=(1−√5)/2 for ℓ=3 and numerical values for ℓ=4,5; that E(M_ℓ^2) also has interior maxima; and that Var(M_ℓ) is strictly increasing in ρ∈(−1,1) for ℓ≥3. It includes an algebraic proof of the equivalence of its E(M^2) formula to Afonja's, a first-principles derivation of E(M_3), and limiting checks at ρ=0.","tokens_in":9895,"tokens_out":9606,"duration_ms":86787,"significance":"If the reported results are correct, the paper gives a complete answer to a natural question about maxima of short AR(1) segments and highlights an interesting contrast: the mean has an interior maximum at a negative serial correlation, while the variance is monotone. The independent derivation of E(M_3) in Section 5 and the algebraic equivalence proof in Section 4 are valuable, as are the checks against known independent-normal limits. However, the results for ℓ=4 and ℓ=5 depend on external formulas that are only partially reproduced, and the variance-monotonicity claim is unsupported. The paper is a useful note but currently falls short of being self-contained.","major_comments":[{"comment":"The displayed formula for E(M_3) is incorrect as printed: it reads √(1−ρ)/π + √(1−ρ^2)/(4π), but the correct value, obtained from the Section 2 formula by substituting ρ12=ρ23=ρ and ρ13=ρ^2, is √((1−ρ)/π) + (1/2)√((1−ρ^2)/π). This is a load-bearing typo in a central formula, even though the limiting value 3/(2√π) at ρ=0 is consistent with the corrected expression.","section":"Section 3, displayed formula for E(M_3)"},{"comment":"The paper reports the ℓ=5 maximizer of E(M_5) to 25 decimal places but never records E(M_5) or E(M_5^2), stating only that they are 'too lengthy.' Since the claimed maximizer is computed from Afonja's formulas, and since the paper itself corrects two typographical errors in Afonja's paper, the reader cannot verify the ℓ=5 result or rule out further undetected errors. To make the central numerical claim reproducible, the full formulas for ℓ=5 (or an evaluation script with clear definitions of the partial correlations) must be included.","section":"Section 3, paragraph 'E(M5) is too lengthy to record here'"},{"comment":"The claim that Var(M_ℓ) is strictly increasing in ρ for all ℓ≥3 is one of the two central questions in the abstract, but it is supported only by visual inspection of Figure 2. The sentence 'An intuitive reason for such behavior would be good to establish someday' acknowledges that no proof is offered. Since closed-form formulas for ℓ=3 and ℓ=4 are available in Section 3, the authors should either provide a proof (at least for these cases) or clearly label the monotonicity as a numerical conjecture.","section":"Section 3, Figure 2 and the paragraph following it"}],"minor_comments":[{"comment":"The identity 'max {X1, X2, X2} = max {max {X1, X2}, max {X2, X3}}' contains an obvious typo: the left-hand side should be max {X1, X2, X3}.","section":"Section 5, first displayed identity"},{"comment":"Several set-builder notations use mismatched parentheses, e.g., '( i, j, m, n} = {1, 2, 3, 4}' and '( i, j, m, n, o} = {1, 2, 3, 4, 5}'. These should be written with matching braces.","section":"Section 1, definitions of sets"},{"comment":"The notation 'Φ ℓ−3(Ri,,jk)' contains a doubled comma and should be 'Φ ℓ−3(R_{i,jk})'.","section":"Section 2, general formulas"},{"comment":"The expression '1/(2π) ri,ji · ri,ki − ri,jk ri,ji/√(1−r_{i,jk}^2)' is ambiguous because a bracket is missing; the intended form, as used in the following line, is '1/(2π) [r_{i,ji}(r_{i,ki} − r_{i,jk} r_{i,ji}) / √(1−r_{i,jk}^2)]'.","section":"Section 4, displayed identity"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a small but honest piece of cleanup work on Afonja's 1972 formulas for the moments of the maximum of correlated normals. Finch decodes the formulas, corrects several typos, applies them to short AR(1) segments, and records the specific ρ values that maximize E(M_ℓ) for ℓ = 3, 4, 5. It also notes — without proof — that Var(M_ℓ) increases in ρ for ℓ ≥ 3.\n\nThe genuinely useful parts are the corrections to Afonja (√(2π) in E(M), a missing factor in E(M) for ℓ = 4, a missing constant 1 and a misstated summation in E(M^2) for ℓ = 5/6), the algebraic equivalence check in Section 4, and the first-principles derivation of E(M_3) in Section 5 using the max-of-maxes identity. The ρ → 0 limits match known independence values, which is a good sanity check. The specific maximizer values and the variance monotonicity observation are new, even if they are direct evaluations of known formulas.\n\nThe soft spots are real, and they are mostly about what is not shown. Section 3's displayed formula for E(M_3) has a typo — the second term is written √(1−ρ^2)/(4π) where it should be √(1−ρ^2)/(2√π). For a note whose main service is correcting other people's typos, that is unfortunate. More substantively, the ℓ = 5 maximizer is reported to six decimals without ever writing E(M_5), which is dismissed as 'too lengthy.' That means the central numerical claim is not independently checkable from the manuscript. The same is true of E(M^2_5) and E(M^2_6). The variance monotonicity is drawn from a figure and explicitly left unproved; it should be labeled a numerical conjecture. The stress-test concern is warranted: limiting checks at ρ → 0 do not validate the ρ-dependence of the formulas, and since the paper does not derive the ℓ = 4, 5 formulas from first principles, an undetected error in Afonja's orthant probabilities would shift the reported maximizers.\n\nNone of this kills the paper. It is a modest note, clearly written, with an honest separation of what is proved and what is observed. The fix is simple: put the full ℓ = 5 and ℓ = 6 formulas in an appendix or a supplementary file, fix the E(M_3) typo, and mark the variance claim as a conjecture. After that, I would be comfortable with it as a published short note. I would send it to a referee, not because the result is revolutionary, but because someone should check the arithmetic and the transcription before these numbers get cited as benchmarks.","headline":"Useful decoding-and-correction note on Afonja's Gaussian-maximum moment formulas, with real benchmark value for short AR(1) segments; the unproved variance monotonicity and the omitted E(M5) need to be flagged before the numbers are used.","tokens_in":10496,"tokens_out":4551,"would_cite":false,"duration_ms":42266,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62E15","60G70","62M10"],"pacs":[],"model":"deepseek-v4-flash","headline":"For segments of length at least three from a stationary first-order autoregressive process, the expected maximum has an interior peak in the lag-one correlation, and the variance is strictly increasing.","keywords":["AR(1) process","expected maximum","variance monotonicity","lag-one correlation","multivariate normal","partial correlation","golden ratio conjugate","order statistics"],"falsifier":"For $\\ell = 4$, evaluate $\\mathbb{E}(\\max\\{X_1,\\ldots,X_4\\})$ for the AR(1) covariance by direct four-dimensional numerical integration on a fine grid of $\\rho$, and check whether the maximum occurs at $\\rho = -0.4973597615161907364022217\\ldots$; repeat for $\\ell = 5$ and check whether $\\mathbb{V}(M_5)$ increases monotonically over the whole interval. A shift in the maximizing $\\rho$ or any local decrease in the variance would refute the central claim.","tokens_in":9441,"feed_emoji":"📈","tokens_out":8156,"duration_ms":69180,"temperature":0.7,"pith_summary":"The paper studies five or fewer consecutive observations of a stationary first-order autoregressive Gaussian process and asks how the moments of their maximum $M_\\ell$ depend on the lag-one correlation $\\rho$. It shows that the expected maximum $\\mathbb{E}(M_\\ell)$ is not monotone in $\\rho$: for $\\ell \\ge 3$ it has an interior maximum at a negative value, and for $\\ell = 3$ the maximizer is exactly $\\rho = (1-\\sqrt{5})/2$, with the $\\ell = 4$ and $\\ell = 5$ maximizers given numerically. The second moment $\\mathbb{E}(M_\\ell^2)$ also attains an interior maximum, but the variance $\\mathbb{V}(M_\\ell)$ is strictly increasing for all $\\rho \\in (-1,1)$ when $\\ell \\ge 3$. These facts matter because they show that short blocks of autocorrelated normal data have their largest expected extremes at negative serial correlation, contrary to the usual worst-case assumption of independence or positive correlation.","feed_headline":"Sample max of an AR(1) segment peaks at negative ρ","feed_subtitle":"For three to five observations the expected maximum is highest with negative lag-one correlation; variance still climbs.","key_machinery":"The load-bearing object is the partial-correlation machinery inside a 1972 formula for the first two moments of the maximum of $\\ell$ jointly normal variables: the formulas express $\\mathbb{E}(M)$ and $\\mathbb{E}(M^2)$ as finite sums of terms involving the orthant probabilities $\\Phi_{\\ell-2}$ and $\\Phi_{\\ell-3}$ of difference variables with partial correlations, together with arcsin and arccos factors. The paper specifies the matrices $R_{i,j}$ and $R_{i,jk}$ whose entries are those partial correlations, notes that symmetry fails for $\\ell \\ge 4$, corrects two typographical errors in the source formula, and introduces the function $h(x,y,z)$ to organise the many terms. Substituting $\\rho_{ij} = \\rho^{|j-i|}$ reduces the $\\ell = 4,5$ sums to one-dimensional functions of $\\rho$, which are then maximized numerically.","core_discovery":"On the paper's own terms, the central discovery is that the curves $\\rho \\mapsto \\mathbb{E}(M_\\ell)$ and $\\rho \\mapsto \\mathbb{E}(M_\\ell^2)$ each have an interior maximum for every segment length $\\ell \\ge 3$, while $\\rho \\mapsto \\mathbb{V}(M_\\ell)$ is strictly increasing on $(-1,1)$ for $\\ell \\ge 3$. The maximizing $\\rho$ for $\\mathbb{E}(M_3)$ is the golden-ratio conjugate $(1-\\sqrt{5})/2 = -0.6180339887498948482045868\\ldots$; the maximizing values for $\\mathbb{E}(M_4)$ and $\\mathbb{E}(M_5)$ are $-0.4973597615161907364022217\\ldots$ and $-0.4336476843162656141275672\\ldots$. At $\\rho \\to -1$ the mean maximum tends to $\\sqrt{2/\\pi}$, the mean of a standard half-normal variable, and at $\\rho = 1$ it vanishes. The paper achieves this by decoding and correcting a 1972 formula for the moments of the maximum of correlated normal variables, specializing it to the Toeplitz covariance $\\rho^{|j-i|}$ of an AR(1) process, and verifying the $\\ell = 3$ expectation from first principles.","pith_inferences":["A natural numerical extension is to compute the interior maximizer for $\\ell \\ge 6$; the large-segment extreme-value theorem cited in the paper does not put $\\rho$ into its centering constant, so the location of the maximum may be a purely finite-length effect.","The strict monotonicity of the variance invites a coupling or convex-order proof: pairing AR(1) processes with different $\\rho$ values might show that the block maximum for larger $\\rho$ is larger in convex order, which would supply the intuitive explanation the paper leaves open.","For worst-case analyses that use short blocks of autocorrelated normal data, the riskiest serial correlation is negative rather than independence or positive correlation, so scanning $\\rho < 0$ could change reported worst-case maxima in timing or finance applications."],"forward_implications":["For any stationary AR(1) Gaussian series, a block of three to five consecutive observations is expected to be largest when the lag-one correlation is negative, because negative correlation spreads the values apart.","The variance of the block maximum is strictly increasing in $\\rho$ for $\\ell \\ge 3$, so stronger positive serial correlation widens the spread of the maximum even as its mean decreases.","The length-three case yields the exact interior maximizer $\\rho = (1-\\sqrt{5})/2$, a rare closed-form extremum for an expectation over correlated normals.","With the corrected formulas, the first two moments of the block maximum can be computed for $\\ell$ up to six without simulation, and the $\\ell = 3$ expectation is verified from first principles."],"supporting_citations":[{"why":"Supplies the general formulas for the first two moments of the maximum of correlated normal variables that the paper corrects and specializes.","marker":"[1]"},{"why":"Provides the independent-observation expected-maximum values used to check the $\\rho \\to 0$ limits for lengths 2 through 5.","marker":"[5]"},{"why":"Contains the order-statistics exercise $\\max\\{X_1,X_2\\} = (X_1+X_2+|X_1-X_2|)/2$ that motivates the first-principles derivation for length 3.","marker":"[6]"},{"why":"Gives the limit theorem for the maximum of a stationary sequence that anchors the large-segment discussion.","marker":"[13]"},{"why":"Extends the extreme-value theory to stationary Gaussian processes, supporting the claim that the large-segment asymptotics apply to AR(1).","marker":"[14]"},{"why":"Provides the limit distribution for sums and maxima of multivariate Gaussian sequences, used in the large-$\\ell$ context.","marker":"[15]"},{"why":"Identified typographical errors in a related moment computation and supplied code for numerically approximating the first two moments of the maximum.","marker":"[23]"},{"why":"Provides the symbolic integration technique that verified the integrals in the length-3 first-principles derivation.","marker":"[25]"}],"fun_headline_variants":["AR(1) max peaks at negative lag-one correlation","Golden ratio conjugate maximizes AR(1) segment max","Expected max of AR(1) segment has interior peak","Variance of AR(1) max climbs with ρ","Interior optimum for AR(1) mean maximum at negative ρ"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported maximizers for $\\ell = 4$ and $\\ell = 5$, and the monotonicity of the variance for $\\ell \\ge 4$, rest on the accuracy of the 1972 general moment formulas after the paper's two typo corrections, because the paper does not re-derive those formulas from first principles for those lengths.","fun_headline_variants_meta":{"raw":{"variants":["AR(1) max peaks at negative lag-one correlation","Golden ratio conjugate maximizes AR(1) segment max","Expected max of AR(1) segment has interior peak","Variance of AR(1) max climbs with ρ","Interior optimum for AR(1) mean maximum at negative ρ"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000967,"raw_usage":{"total_tokens":4101,"prompt_tokens":919,"completion_tokens":3182,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":3100}},"tokens_in":535,"tokens_out":3182,"duration_ms":24519,"temperature":1.0,"reasoning_tokens":3100,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:48:36.146750+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For $\\ell = 4$, evaluate $\\mathbb{E}(\\max\\{X_1,\\ldots,X_4\\})$ for the AR(1) covariance by direct four-dimensional numerical integration on a fine grid of $\\rho$, and check whether the maximum occurs at $\\rho = -0.4973597615161907364022217\\ldots$; repeat for $\\ell = 5$ and check whether $\\mathbb{V}(M_5)$ increases monotonically over the whole interval. A shift in the maximizing $\\rho$ or any local decrease in the variance would refute the central claim.","supporting_citations":[{"cited_title":"Afonja, The moments of the maximum of correlated normal an d t-variates, J","cited_arxiv_id":null,"evidence_quote":"Supplies the general formulas for the first two moments of the maximum of correlated normal variables that the paper corrects and specializes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the independent-observation expected-maximum values used to check the $\\rho \\to 0$ limits for lengths 2 through 5."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Contains the order-statistics exercise $\\max\\{X_1,X_2\\} = (X_1+X_2+|X_1-X_2|)/2$ that motivates the first-principles derivation for length 3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the limit theorem for the maximum of a stationary sequence that anchors the large-segment discussion."},{"cited_title":"Pickands, Maxima of stationary Gaussian processes, Z","cited_arxiv_id":null,"evidence_quote":"Extends the extreme-value theory to stationary Gaussian processes, supporting the claim that the large-segment asymptotics apply to AR(1)."},{"cited_title":"James, K","cited_arxiv_id":null,"evidence_quote":"Provides the limit distribution for sums and maxima of multivariate Gaussian sequences, used in the large-$\\ell$ context."},{"cited_title":"del Castillo, A","cited_arxiv_id":null,"evidence_quote":"Identified typographical errors in a related moment computation and supplied code for numerically approximating the first two moments of the maximum."},{"cited_title":"Kan and C","cited_arxiv_id":null,"evidence_quote":"Provides the symbolic integration technique that verified the integrals in the length-3 first-principles derivation."}],"review_version":1}