{"id":"16953953-5c44-4ef8-8f31-07eff043c116","arxiv_id":"2607.27818","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A bias-corrected, selfnormalized statistic based on a log-sum-exp smoothing of the supremum norm yields an asymptotically pivotal test for relevant changes in functional time series.","lead":"This paper builds a test for whether a functional time series has a relevant change, measured by the largest deviation between the two regimes, without estimating the long-run covariance. It replaces the supremum norm with a smooth log-sum-exp approximation and cancels the resulting bias by mixing three smoothing levels, giving a statistic whose limit depends only on the break location.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The bias-cancellation guarantee in Thm 2.7 is not robust to C^2 contrasts with cubic (degenerate) extrema: at the recommended beta=n^{1/3}, the residual bias after cancellation can grow like n^{1/18}, so exact boundary size is not established outside Thm 2.4/2.5.","rationale":"I read the manuscript as a serious attempt to extend selfnormalization to sup-norm relevant inference. Theorems 2.2 and 2.7 proofs are detailed, the cancellation algebra checks out for the stated regimes, and the simulations support the claimed size improvement over the bootstrap. The weakest point is not the proof under the stated assumptions but the breadth of the claim. The reader flagged the geometric dichotomy; my partial disagreement is that the dichotomy is not just 'not covered' — concrete C^2 contrasts with cubic extrema produce a residual bias after cancellation that diverges at the recommended smoothing rate. This is a correctness risk for the method as advertised, since a practitioner cannot verify C^4 nondegeneracy or the plateau expansion from data. The fix (choose beta faster, e.g. n^{0.4}, or extend Theorem 2.5 to subleading near-extreme terms) would make the concern evaporate, which is why I keep the CONDITIONAL verdict rather than moving to REJECT. I recommend no change to the reader's verdict.","tokens_in":22105,"tokens_out":43747,"duration_ms":383795,"concrete_test":"Monte Carlo check: n=400, tau0=1/2, Delta=1, contrast d(t)=1-|t-1/2|^3-0.1|t-1/2|^4, errors iid standard Brownian bridges (or the paper's AR(1) bridge design), 10,000 replications. Apply the bias-corrected test with beta=n^{1/3} and the paper's break estimator; compute the empirical rejection rate at the boundary M=Delta. If it is not within Monte Carlo error of 5% (e.g. <3% or >7%), the residual-bias divergence is confirmed. As a control, rerun with beta=n^{0.4}: if the rate returns to 5%, the failure is the rate condition for the cubic geometry rather than the selfnormalizer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central exactness claim at M=Delta requires sqrt(n){L_{beta n}(d)-M}=o(1). The proof supplies this only under Thm 2.4 (C^4 isolated nondegenerate extrema, sqrt(n) beta^{-2}=o(1)) or Thm 2.5 (positive-measure plateau with F(u)=m+K u^alpha+o(u^alpha), sqrt(n) beta^{-(1+alpha)}=o(1)). For a contrast outside this dichotomy the cancellation can fail at the recommended rate. Consider d(t)=M-|t-1/2|^3-eps|t-1/2|^4, a C^2 function with a degenerate cubic extremum. The near-extreme volume expands as F(u)=2u^{1/3}-(2eps/3)u^{2/3}+o(u^{2/3}), giving Psi_{beta}(d)-M = [-(1/3)log beta + log A]/beta + (B/A) beta^{-4/3} + o(beta^{-4/3}). The two cancellation constraints sum a_i/c_i=0 and sum a_i log c_i/c_i=0 kill the leading term but leave residual (B/A) beta^{-4/3} sum a_i c_i^{-4/3}. With the paper's recommended beta=n^{1/3}, sqrt(n) times this residual is n^{1/18} -> infinity; the statistic drifts and the boundary rejection probability is not alpha0. Boundary extrema (deferred in Remark 2.8) and any subleading near-extreme exponent with alpha'-alpha<1/2 produce the same phenomenon. Since the geometry is unknown and the abstract says 'mild regularity conditions', the advertised asymptotic exactness is not secured for a natural class of continuous contrasts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes selfnormalized tests for relevant hypotheses on the sup-norm magnitude M = ||mu1 - mu2||_infinity in a change-point model for functional time series. The supremum norm is replaced by a log-sum-exp soft maximum Psi_beta, and a projected selfnormalizer based on the derivative of Psi at the estimated contrast is introduced. Explicit bias expansions are derived for two extremal geometries: finitely many interior nondegenerate extrema (Theorem 2.4) and positive-measure extremal sets with a near-extreme-volume expansion (Theorem 2.5). A three-scale linear combination of soft maxima cancels the leading bias terms without estimating geometric constants. The main result (Theorem 2.7) states that the bias-corrected statistic converges to a pivotal law depending only on the break fraction tau0, so that the boundary test has asymptotic rejection probability alpha0. Local-power analysis and simulations comparing the procedure with a bootstrap relevant-change test are also given.","tokens_in":22540,"tokens_out":19800,"duration_ms":187879,"significance":"Within the assumptions of Theorems 2.4 and 2.5, the paper makes a substantial contribution: it is the first selfnormalized relevant-change test in the C([0,1]) sup-norm framework, and the projected selfnormalizer cancels long-run covariance nuisance parameters in a genuinely non-smooth problem. The multiscale bias-cancellation idea is elegant and avoids direct estimation of extremal-set geometry. The paper also provides explicit bias expansions and a finite-sample correction, and the simulation study supports the qualitative claims. However, the advertised 'mild regularity conditions' cover only two geometric regimes, and the boundary exactness claim is not robust to degenerate extremal sets. A key Laplace expansion is also only sketched. These issues do not invalidate the conditional theorems, but they materially limit the broad claims made in the abstract and Section 2.2.","major_comments":[{"comment":"The advertised exactness is not established for natural geometries outside the two regimes. The cancellation identities in (2.6)-(2.7) cover only isolated nondegenerate extrema and plateaus with F(u)=m+Ku^alpha+o(u^alpha). For a C^2 contrast such as d(t)=M-|t-1/2|^3-epsilon|t-1/2|^4, F(u)=2u^{1/3}-(2epsilon/3)u^{2/3}+o(u^{2/3}); after the constraints sum a_i/c_i=0 and sum a_i log c_i/c_i=0 kill the leading term, a residual C beta^{-4/3} sum a_i c_i^{-4/3} remains. With the recommended beta_n=n^{1/3}, sqrt(n) times this residual is O(n^{1/18}), so the statistic drifts and the boundary rejection probability is not alpha0. Smooth degenerate quartic extrema produce a non-vanishing or worse bias at the same rate. The abstract's 'mild regularity conditions' and the claim in Section 2.2 that geometry estimation is avoided are therefore too strong. Either restrict the theorem to the two geometri","section":"Section 2.2 / Theorem 2.7 / Remark 2.8"},{"comment":"The Laplace expansion is not fully verified. After the change of variables y=sqrt(beta lambda_j)x, the proof states 'we omit showing that applying this expansion within the integral is valid in detail' and only lists steps (1)-(3). The expansion (2.6), including the o(beta^{-2}) term, is load-bearing: it is exactly the residual that must be o(n^{-1/2}) after bias correction in Theorem 2.7. Uniform integrability of the expansion terms and the tail estimate for |y|>r_beta need a complete argument. This is a proof gap in a central result, not merely a style issue.","section":"Section 4.3, proof of Theorem 2.4"}],"minor_comments":[{"comment":"The proof of the local-power theorem is abbreviated: the uniform oracle reduction under local alternatives and the uniform validity of the bias expansion along the sequence d_n are asserted rather than written out. Please expand these steps, especially the treatment of the triangular-array remainder terms.","section":"Section 4.1, Theorem 4.1"},{"comment":"The standardization factor z_rho for the AR(1) functional error process should be checked. As written, epsilon_i = z_rho(rho epsilon_{i-1}+eta_i) with eta_i Brownian bridges does not appear to give long-run variance 1/4 at t=1/2 for the stated z_rho=1/(1+rho). Please verify or correct the formula.","section":"Section 3, simulation design"},{"comment":"The finite-sample correction using the constant 0.1 and the maximum pointwise standard deviation is described as conservative, but no sensitivity analysis is reported for this choice. Since the correction subtracts a nonnegligible finite-sample quantity, a short sensitivity table or a formal derivation of the domination claim would be useful.","section":"Remark 2.8(4)"},{"comment":"The hat V notation is used with different bracket styles (e.g., hat V^{(\\beta_n)}_n vs. hat V^{\\beta_n}_n). Unify the notation for the selfnormalizer and its bias-corrected version.","section":"Notation"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is promising and the conditional theorems appear sound. The main risk is that the abstract and Section 2.2 overstate the scope: the bias-cancellation exactness is only proved for isolated nondegenerate extrema and positive-measure plateaus, and the recommended beta_n=n^{1/3} fails for certain degenerate contrasts. If the author can either restrict the claims accordingly or extend the cancellation argument to a broader class of near-extreme-volume expansions, I would support publication. The Laplace expansion proof also needs to be completed before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's the short version. The paper delivers the first selfnormalized relevant-change test for functional time series in the C([0,1]) sup-norm framework, using a log-sum-exp soft maximum and a three-level linear combination to cancel smoothing bias. That is a real advance over the L2-based selfnormalization of Shao–Zhang, Zhang–Shao, and Dette et al. The pivotal limit depending only on the break fraction is attractive, and the simulations show better boundary size than the bootstrap comparator.\n\nThe main theorems are mostly solvent. I checked the Laplace expansion (Thm 2.4) and the plateau expansion (Thm 2.5); the cancellations in Thm 2.7 work when the regularity conditions are met. The proofs do have honest gaps — the Laplace proof says it is omitting the uniform-integrability details, and the local-power theorem (Thm 4.1) is justified by 'minor modifications' without showing the triangular-array uniform bounds. Boundary extrema are explicitly deferred.\n\nBut the real soft spot is a gap in the bias-cancellation itself. The correction is built to eliminate the leading terms that arise from the near-extreme-volume expansion F(u)=K u^α + ... . For a C^2 contrast with a cubic extremum, e.g. d(t)=M-|t-1/2|^3, the expansion has α=1/3 and a subleading u^{2/3} term. That produces a residual bias of order β^{-4/3} after cancellation, and at the recommended β=n^{1/3} the √n-scaled residual is n^{1/18}, which diverges. So the boundary rejection probability is not guaranteed to be α0 outside the two prototypical geometries. The abstract's phrase \"mild regularity conditions\" is too strong; Remark 2.8 only mentions boundary, not degenerate extrema. This is not a fatal flaw — the method works in the cases it actually treats, and the simulations are consistent with the claims — but it is a load-bearing assumption that needs to be either handled or explicitly disclaimed.\n\nBottom line: worth a serious referee. It is a plausible new construction with a real gap that can likely be closed or scoped; I would want the authors to address the degenerate-extrema case, tighten the proof sketches, and recalibrate the abstract. No code is provided, which is a minor reproducibility issue, not a mathematical one.","headline":"Genuinely new sup-norm selfnormalized relevant-change test with a smart bias-cancellation device, but the advertised exactness is over-broad: degenerate (cubic) extrema break the correction at the recommended smoothing rate.","tokens_in":23038,"tokens_out":3932,"would_cite":true,"duration_ms":34950,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G10","62G20","62M10","62R10"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper develops a selfnormalized test for relevant changes in functional time series under the supremum norm, showing that a bias-corrected log-sum-exp statistic converges to a pivotal distribution depending only on the break location.","keywords":["relevant hypothesis testing","selfnormalization","supremum norm","functional time series","change point detection","log-sum-exp smoothing","bias correction","asymptotic pivotal distribution"],"falsifier":"Simulate the bias-corrected statistic at the boundary (M=Δ) for a contrast with a single interior extremum where the second derivative is zero (e.g., d(t)=1−c(t−1/2)^4) and for a contrast with the extremum at the boundary (e.g., d(t)=1−ct^2), using n=400 and β_n=n^{1/3}; if the empirical rejection rate at nominal 0.05 departs from 0.05 by more than sampling error across 10,000 replications, the geometric dichotomy is the load-bearing premise.","tokens_in":21964,"feed_emoji":"🎯","tokens_out":10368,"duration_ms":83218,"temperature":0.7,"pith_summary":"This paper tries to solve a standing problem: tests for whether a change in functional data exceeds a relevance threshold Δ, measured in the supremum norm, could not use selfnormalization because the supremum norm is not smooth and its limiting behavior depends on nuisance geometry and the long-run covariance. The proposed fix is to replace the norm by a log-sum-exp smooth approximation, work out the smoothing bias for two extreme geometries (isolated quadratic peaks and flat plateaus), and cancel that bias with a three-level linear combination of smoothers. The resulting bias-corrected statistic is asymptotically pivotal, with a limit that depends only on the unknown break fraction, so a threshold test based on it is asymptotically exact: rejection probability α0 at M=Δ, 0 inside the null, and 1 outside. If the theorem holds, this is the first selfnormalized relevant-change test in the sup-norm setting, and the smoothing-bias-correction template carries over to other supremum-type relevant hypotheses.","feed_headline":"Sup-norm relevant-change tests now have pivotal limits","feed_subtitle":"A smoothed maximum plus bias cancellation makes the limit depend only on the break fraction, not on error covariance","key_machinery":"The key object is the log-sum-exp (soft maximum) functional Ψ_β(f) = (1/β) log ∫_0^1 (e^{β f(t)}+e^{−β f(t)}) dt, which tends to ||f||∞ as β→∞ and is Fréchet differentiable for finite β. Its derivative DΨ_{β,f}(h) is a weighted average of h with exponential weights, supplying a linear projection whose selfnormalized sums can be handled by a projected invariance principle. The bias corrections come from the linear combination L_β = a_1 Ψ_{c_1 β} + a_2 Ψ_{c_2 β} + a_3 Ψ_{c_3 β} with a≈(1.2509,−2.2555,2.0046) and c≈(1,1.4475,6.5232), chosen to cancel the leading bias in both geometric regimes. The asymptotic expansions are obtained via classical exponential-integral expansions, expressed throug","core_discovery":"The paper's central claim, Theorem 2.7, is that after estimating the break point, computing the post-estimation contrast, and applying the linear combination L_β of three log-sum-exp smoothers, the statistic sqrt(n){L_{β_n}(d̂_n) − M}/√V̂_n converges in distribution to T_{τ0}, a ratio of a standard normal to a weighted sum of squared L2 norms of two independent Brownian bridges, with weights determined by τ0. Because the limit is pivotal and free of long-run covariance, using its (1−α0)-quantile as a critical value gives an asymptotically exact test of the relevant hypotheses. The theorem rests on two explicit bias expansions: for finitely many nondegenerate interior extrema, the bias is (−l","pith_inferences":["The geometric-dichotomy assumption (finite interior peaks vs. plateaus with power-law near-extreme volume) is not verifiable from data; a natural extension would be a data-adaptive choice of β_n that detects which regime holds.","Boundary extrema are deferred in the paper; one could derive analogous one-sided exponential-integral expansions, likely changing the coefficient of (log β)/β and the recommended β_n rate.","The proof reduces to a projected invariance principle for linear functionals ℓ_n, suggesting the construction carries over to other nonsmooth statistics such as suprema of U-statistics or empirical processes.","Because the auxiliary bias-bound test is conservative and can lose power against nonconstant local alternatives, practitioners using the bias-corrected test should be aware that its exactness depends on the extremal geometry in ways the simpler bound does not."],"forward_implications":["Applied at the boundary M=Δ, the test's rejection rate converges to the nominal level α0; for M<Δ it vanishes and for M>Δ it goes to 1.","Users no longer need to estimate long-run covariance or choose bandwidths; the limiting distribution depends only on the estimated break fraction τ0.","The bias-corrected test has non-trivial local power against radial alternatives of size n^{−1/2} in any fixed direction of the contrast.","The smoothing-bias expansions cover both finitely many interior quadratic extrema and positive-measure plateaus, and require no knowledge of which case holds.","The combination of log-sum-exp smoothing and multi-level bias cancellation applies more generally to relevant hypotheses defined by supremum-type functionals."],"fun_headline_variants":["Selfnormalized sup-norm tests now pivotal","Smoothed max yields pivotal relevant-change tests","Bias cancellation makes sup-norm tests pivotal","Pivotal limits for supremum-type relevant inference","Selfnormalization via smoothing for sup-norm tests"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The exactness claim rests on the assumption that the mean contrast's supremum is attained either at finitely many interior points with nonzero curvature or on a plateau whose near-maximal volume grows exactly like u^α; any other geometric shape of the extremal set breaks the bias-cancellation rate and the pivotal limit is not guaranteed.","fun_headline_variants_meta":{"raw":{"variants":["Selfnormalized sup-norm tests now pivotal","Smoothed max yields pivotal relevant-change tests","Bias cancellation makes sup-norm tests pivotal","Pivotal limits for supremum-type relevant inference","Selfnormalization via smoothing for sup-norm tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000579,"raw_usage":{"total_tokens":2565,"prompt_tokens":742,"completion_tokens":1823,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1760}},"tokens_in":486,"tokens_out":1823,"duration_ms":13338,"temperature":1.0,"reasoning_tokens":1760,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T00:48:41.956232+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the bias-corrected statistic at the boundary (M=Δ) for a contrast with a single interior extremum where the second derivative is zero (e.g., d(t)=1−c(t−1/2)^4) and for a contrast with the extremum at the boundary (e.g., d(t)=1−ct^2), using n=400 and β_n=n^{1/3}; if the empirical rejection rate at nominal 0.05 departs from 0.05 by more than sampling error across 10,000 replications, the geometric dichotomy is the load-bearing premise.","supporting_citations":[],"review_version":1}