{"id":"1cd5ea4e-2c5b-40b0-8a14-474459c07424","arxiv_id":"2412.05195","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A piecewise-linear gauge function with explicit volume computation enables fast semi-parametric inference for multivariate geometric extremes, plus KDE-based radial quantile estimation.","lead":"This paper introduces a piecewise-linear function to describe the shape of extreme events in multivariate data, making the recent geometric approach to extremes simpler, faster, and interpretable. The method aims to extend extreme-event modeling to higher dimensions, with a demonstration on four air pollutants.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fixed-shape truncated-gamma likelihood (Eq. 2) is load-bearing: at the application threshold τ=0.70, angle-dependent shape (acknowledged for Gaussian dependence) is unassessed, so extrapolation bias could escape detection by the τ=0.95 simulations.","rationale":"I read the paper as proposing a semiparametric piecewise-linear gauge with two intended advances: an interpretable, low-cost approximation of g, and a way to do full radial-angular inference. The piecewise-linear volume formula (Prop. 1) is correct, the KDE quantile estimator is reasonable and compared to alternatives in Supplement C, and the d=2,3 simulations show the method competes with a parametric fit that uses the true gauge. I do not object to the construction itself. The reader's weakest_assumption points to Eq. (2), and I agree that is where the argument is least secure. The paper explicitly acknowledges the shape issue for Gaussian dependence, and claims without showing that estimating the shape makes little difference. Because the application threshold is 0.70 and the dimension is 4, the finite-threshold misspecification could be larger there than in any simulation. The application diagnostics are not able to detect a systematic bias in extrapolated probabilities: return-level checks are essentially counting exceedances of the fitted boundary, and χC plots cover only u in [0.85,1]. Thus a targeted simulation at τ=0.70 with Gaussian dependence would settle the matter. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":37811,"tokens_out":8636,"duration_ms":100119,"concrete_test":"Extend the Section 5 simulation to Gaussian dependence (the case where shape variation is known) at the application threshold τ=0.70: e.g., d=2 with ρ=0.8 and a new d=3 equicorrelated ρ=0.6 example in exponential margins. For each replicated dataset, fit the PWL model twice — once with the shape fixed at d and once with the gamma shape estimated (per angular bin or as a smooth function of w) — and compare the RMSE/bias of the resulting probability estimates for extremal regions B1–B3 against true probabilities. If fixed-shape estimates are materially worse (e.g., more than 25% RMSE increase, or bias outside the parametric baseline's uncertainty), the concern lands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim of Section 5 is that the piecewise-linear gauge gives probability estimates comparable to a parametric method that knows the true gauge. Everything downstream of that claim inherits the truncated-gamma likelihood in Section 4.1, which treats f_{R|W}(r|w) as exactly Gamma(shape=d, rate=g(w)) above rτ(w). The only asymptotic justification is Eq. (2), and the paper itself notes in Section 1.1 that for Gaussian dependence the gamma shape also depends on w, while Section 4.1 says estimating the shape 'lead to little difference in terms of bias' without reporting the experiment. The simulations that support the headline claim use τ=0.95; the d=4 application uses τ=0.70, where asymptotic shape stabilization is weaker and d=4 has no simulation backing. In-sample PP/QQ and χC agreement in the application is not a sharp check of extrapolation bias, because those diagnostics are largely driven by data near the threshold. If shape variation with w is substantial at τ=0.70, the MLE of gpwl is biased, and the Monte Carlo probability estimates in Section 4.3 inherit that bias. This is the load-bearing weak point, not the piecewise-linear construction itself.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a semiparametric estimator of the gauge function in multivariate geometric extremes. The gauge is piecewise-linear on a Delaunay triangulation of the simplex (Eq. 5), with an explicit limit-set volume formula (Proposition 1), a KDE-based radial quantile estimator (Section 3), and penalized maximum likelihood fitting (Section 4). Simulation studies in Section 5 compare six model-fitting settings against parametric estimation that uses the true gauge for d = 2, 3, and Section 6 applies the method to four air pollutants. The central claim is that the semiparametric piecewise-linear approach is comparable in accuracy to parametric methods while requiring no knowledge of the underlying distribution.","tokens_in":37963,"tokens_out":4330,"duration_ms":49322,"significance":"If the inferential claim holds, this is a practical and interpretable default tool for geometric extremes in d = 2, 3, with a credible d = 4 application. The construction is elegant: the piecewise-linear gauge is continuous and 1-homogeneous, the volume of its unit level set is computed exactly through determinants, and optimization is inexpensive. The paper provides reproducible code, detailed supplementary comparisons across logistic, Gaussian, inverted logistic, asymmetric logistic, and mixture dependence structures, and reports probability estimates against true values in simulations. The mathematical core, including Proposition 1 and the triangulation construction, is clean; the main uncertainties concern the finite-sample validity of the truncated-gamma likelihood used for inference, not the piecewise-linear construction itself.","major_comments":[{"comment":"The fixed-shape truncated-gamma likelihood is load-bearing for every downstream probability estimate, but its validity at the application's operating point is not demonstrated. The paper notes in Section 1.1 that for Gaussian dependence the gamma shape parameter also depends on w, and Section 4.1 states that estimating the shape \"lead to little difference in terms of bias\" without reporting the supporting experiment. The simulations that support the headline claim use tau = 0.95 and d = 2, 3, whereas the application in Section 6 uses tau = 0.70 and d = 4, where asymptotic shape stabilization is weaker and no simulation evidence is provided. Please report the shape-parameter experiment and add a sensitivity analysis at tau = 0.70 (and, if feasible, in d = 4) comparing fixed-shape and estimated-shape fits; otherwise the extrapolation bias remains uncontrolled.","section":"Section 4.1, Eq. (2)"},{"comment":"The claim that the semiparametric approach is \"comparable\" to the parametric method is not uniformly supported by the reported RMSEs. For example, in Table S.3, distribution (V), region B1, the parametric RMSE is 0.4718 while SS4 reports 2.0826, a factor of about 4.4; in Table S.2, distribution (II), region B1, the parametric RMSE is 0.2524 versus 1.008 for SS4. These settings show substantially larger log-probability errors. The text should either define \"comparable\" quantitatively (e.g., within a factor of 2 in RMSE) and discuss the settings where that definition fails, or qualify the headline claim accordingly.","section":"Section 5, Tables S.2 and S.3"},{"comment":"The application diagnostics are in-sample and largely reflect behavior near or above the threshold, so they do not sharply test extrapolation bias induced by the truncated-gamma approximation. The chi_C(u) curves, return-period checks, and PP/QQ plots all use the same data that informed the fitted gauge and angular model, and their agreement with empirical values at moderate u does not directly validate probability estimates far beyond the threshold. Please add a threshold-stability check (e.g., refitting at tau = 0.80 and 0.90 and comparing estimated probabilities or return periods) or a genuinely out-of-sample temporal validation, and report whether the conclusions change.","section":"Section 6, Figures 11 and S.28-S.30"}],"minor_comments":[{"comment":"The phrase \"weather or not\" should be \"whether or not\".","section":"Section 4.4"},{"comment":"The sentence \"A similar conclusion can me made\" should read \"A similar conclusion can be made.\"","section":"Section 5"},{"comment":"The word \"bootsrap\" should be \"bootstrap\".","section":"Figure 11 caption"},{"comment":"The choices N = 11 for d = 2 and N = 28 for d = 3 are stated without a sensitivity analysis; a brief discussion of how results vary with N would help readers understand the bias-variance trade-off in practice.","section":"Section 4.2.1"},{"comment":"The KDE quantile estimator uses no boundary correction on the simplex, and the application uses a larger bandwidth h_W = 0.075; a sentence quantifying the resulting boundary bias for the radial threshold would be useful.","section":"Supplement C"},{"comment":"The notation w^star_{1,...,N\\i}^j in Algorithm 1 is hard to read; please define the index sets involved more explicitly.","section":"Supplement D, Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a statistics methodology journal and the geometric construction is a genuine contribution. The main issue is that the inferential validation does not yet cover the operating conditions of the application (tau = 0.70, d = 4) and the reported simulation tables do not uniformly support the 'comparable' claim; both are addressable with additional experiments and a more precise statement of the claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This one deserves a serious referee. The piecewise-linear gauge construction, and especially the explicit volume formula in Proposition 1, is a genuine step forward for geometric extremes: it gives the angular model's normalizing constant in closed form, which is exactly the part that forces the competing INLA-based approach into numerical integration at every evaluation. The KDE-based radial quantile estimator in Section 3 is a practical addition, and the d=4 pollution application demonstrates something the existing semiparametric methods cannot do, since they effectively top out at d=3. There is public code on GitHub, which strengthens the empirical claims.\n\nThe simulations back the headline claim. Across the seven dependence structures, the piecewise-linear method's RMSE on log-probability estimates is within a factor of roughly 1-3 of a parametric fit that is handed the true gauge, and it wins in several cells. The paper is also honest about where it fails: the joint SS5/SS6 fits are biased, and the concluding remarks admit that d=5 did not work well. That candor is credible.\n\nThe stress-test note has the right target: the fixed-shape truncated gamma likelihood in Eq. (2) is load-bearing, and the paper's own Section 1.1 concedes that Gaussian dependence has angle-dependent shape. The simulations run at tau=0.95 while the application drops to tau=0.70, with no d=4 simulation backstop, and the Section 4.1 claim that estimating the shape parameter made little difference is stated without the supporting experiment. Still, the Gaussian case is distribution (III), and PWL handles it fine at tau=0.95, so this is a concern about an unexamined regime rather than evidence that the construction is unsound. A sensitivity analysis that varies tau or fits the shape parameter would settle it.\n\nTwo smaller points. There is no head-to-head simulation against Simpson-Tawn or the INLA approach, so the advantage over those competitors is asserted rather than shown. And several hyperparameters (bandwidths, reference-angle count, penalty strengths) are tuned on simulation medians and then fixed in the application; the cross-validation for lambda mitigates this, so it is a modest complaint.\n\nThis paper is for researchers working in geometric extremes or multivariate tail inference, and it gives the field a cheap, interpretable default gauge model. I would accept it for peer review: the construction is sound, the simulations support the central claim in the regime actually tested, and the concerns raised are revision-level rather than rejection-level.","headline":"Deserves a serious referee: the piecewise-linear gauge with explicit volume formula is a genuine, well-tested contribution, and the main open question is whether the d=4 application stretches the fixed-shape gamma assumption past what the tau=0.95 simulations justify.","tokens_in":38638,"tokens_out":10475,"would_cite":true,"duration_ms":89761,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G32","62G05"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that a piecewise-linear gauge function, with parameters placed at reference angles and linear interpolation across a triangulated simplex, delivers semiparametric inference for multivariate geometric extremes with accuracy…","keywords":["geometric extremes","gauge function","piecewise-linear approximation","multivariate extremes","semiparametric inference","limit sets","extremal dependence","radial-angular decomposition"],"falsifier":"Fit the piecewise-linear truncated-gamma model with shape $d$ to $n=5000$ simulated observations from a $d=3$ distribution whose exact conditional radial density is a gamma with shape that varies strongly with $w$ (for example a multivariate $t$ copula or a skew-normal with heavy tail), estimate $\\Pr(X\\in B)$ for $B$ beyond the data range, and compare with a high-accuracy Monte Carlo estimate of the same probability; bias that grows with the shape variation would show the fixed-shape assumption fails at finite thresholds.","tokens_in":37464,"feed_emoji":"📐","tokens_out":5232,"duration_ms":50475,"temperature":0.7,"pith_summary":"The paper proposes a semiparametric estimator of the gauge function that governs the geometry of multivariate extremes, built from a piecewise-linear surface with parameters at selected reference angles on the unit sphere. The claim is that this simple, interpretable estimator reproduces the accuracy of parametric inference that uses the true gauge function, while requiring no knowledge of the underlying distribution and costing little to compute. On simulated data in dimensions two and three, the fitted piecewise-linear model yields extremal probability estimates comparable to the parametric benchmark across logistic, Gaussian, inverted logistic, asymmetric logistic, and mixture dependence structures. The method also scales to a four-dimensional air-pollution application, where it captures simultaneous extremes of all four pollutants and estimates return periods and extremal coefficients.","feed_headline":"Piecewise-linear gauge matches parametric tail accuracy","feed_subtitle":"A linear-tile gauge estimates extreme probabilities as well as the true-model fit, with no distributional assumptions.","key_machinery":"The central object is the piecewise-linear gauge function $g_{\\mathrm{pwl}}(x;\\theta)$, defined by placing reference angles $w^\\star$ on $S^{d-1}$, assigning a parameter $\\theta_k = 1/g_{\\mathrm{pwl}}(w^\\star_k)$ equal to the distance from the origin to the limit-set boundary along that ray, and linearly interpolating across a Delaunay triangulation. It carries the argument in three roles: as the rate function of the truncated gamma model for large radii, as the generator of the angular density $f_W(w) = g(w)^{-d}/\\{d\\,\\mathrm{vol}(G)\\}$ with explicit normalizing constant, and as an object whose gradients are available in closed form, enabling a penalized likelihood that smooths neighboring-cell gradients.","core_discovery":"The central discovery is that the limit set of a light-tailed multivariate distribution can be approximated well enough by a gauge function $g_{\\mathrm{pwl}}$ that is linear on each cell of a Delaunay triangulation of the simplex, with one distance parameter per reference angle (Eq. 5). Because the gauge appears as the rate in the asymptotic truncated gamma density $f_{R\\mid W}(r\\mid w) \\propto r^{d-1}\\exp\\{-r g(w)\\}$, the parameters can be estimated by maximum likelihood from pseudo-polar coordinates above a high radial threshold, and the same $g_{\\mathrm{pwl}}$ supplies an angular density $f_W(w) = g(w)^{-d}/\\{d\\,\\mathrm{vol}(G)\\}$ whose normalizing volume is an explicit sum of determinants. The paper argues that this construction makes semiparametric inference for geometric extremes both flexible and cheap, and demonstrates parity with parametric inference using the true gauge in probability estimation for $d=2,3$, with an extension to $d=4$.","pith_inferences":["If the fixed-shape gamma approximation is as benign in practice as the simulations suggest, the piecewise-linear gauge could serve as a screening tool for tail dependence structure before committing to a parametric model; this extends the paper's claim beyond the copulas it tests.","The gradient penalty suggests an adaptive-knot variant in which reference angles are added where neighboring gradients differ most, which could extend the method beyond the reported $d=4$--$5$ range; the authors note that unreported $d=5$ fits faced difficulties with angle selection.","A natural next test is to apply the estimator to negatively dependent data in Laplace margins, where the paper supplies a bivariate construction but leaves the $d>2$ triangulation of the signed $\\ell^1$ sphere open.","The return-level and $\\chi_C$ diagnostics provide a template for validating gauge estimates against empirical quantities, which could be turned into a formal goodness-of-fit test for geometric extreme models."],"forward_implications":["Practitioners get a default, interpretable tool for geometric extremes in $d=2,3$ that does not require choosing a parametric copula or a neural-network specification.","Extremal probabilities, return-level sets, and extremal coefficients $\\chi_C(u)$ can be produced by simulation from the fitted radial-angular model, with the angular model replacing the empirical angle distribution.","The explicit volume formula makes the angular likelihood cheap enough to maximize, whereas competing semiparametric angular models require numerical integration or latent-variable estimation.","The kernel-density radial quantile estimator evaluates $r_\\tau(w)$ at every angle on the simplex, removing empty-bin failures of empirical binning in higher dimensions.","Dimension 4 becomes feasible with a sparse choice of reference angles, as demonstrated on four pollutants."],"supporting_citations":[{"why":"Supplies the truncated gamma asymptotic density for $R\\mid W$ and the parametric geometric inference framework that the piecewise-linear method extends.","marker":"Wadsworth and Campbell (2024)"},{"why":"Provides the limit-set characterization of scaled sample clouds and the homothetic angular density $f_W(w)=g(w)^{-d}/\\{d\\,\\mathrm{vol}(G)\\}$ used for the angular model.","marker":"Balkema and Nolde (2010)"},{"why":"Establishes the gauge-function limit $g(x)=\\lim_t -\\log f(tx)/t$ and the link between gauge functions and extremal dependence.","marker":"Nolde and Wadsworth (2022)"},{"why":"Is the main Bayesian semiparametric competitor for $d\\le 3$, with a similar angular density and return sets that motivate the return-level diagnostics.","marker":"Papastathopoulos et al. (2025)"},{"why":"Provides the bivariate semiparametric limit-shape alternative against which the piecewise-linear approach is positioned.","marker":"Simpson and Tawn (2024a)"},{"why":"Supplies the triangulation of the simplex used to define linear interpolation of the gauge in $d\\ge 3$.","marker":"Delaunay (1934)"},{"why":"Gives the additive quantile regression approach for $r_\\tau(w)$ in $d=2$ that the KDE estimator is designed to replace in higher dimensions.","marker":"Fasiolo et al. (2021)"},{"why":"Provides the marginal standardization via empirical distribution with generalized Pareto tail used before fitting the gauge.","marker":"Coles and Tawn (1991)"}],"fun_headline_variants":["Linear-tile gauge: cheap and flexible extreme value inference","Piecewise-linear gauge matches parametric tail accuracy","Semiparametric gauge for geometric extremes: simple and fast","Linear geometry for multivariate tails: no neural nets needed","Gauge function made piecewise-linear: accurate and interpretable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the asymptotic density approximation $f_{R\\mid W}(r\\mid w) \\propto r^{d-1}\\exp\\{-r g(w)\\}$ can be used as an exact truncated gamma model with shape parameter fixed at $d$; if the true shape varies substantially with direction at the thresholds used ($\\tau=0.95$ in simulation, $0.70$ in the application), the likelihood is misspecified and probability estimates could be biased.","fun_headline_variants_meta":{"raw":{"variants":["Linear-tile gauge: cheap and flexible extreme value inference","Piecewise-linear gauge matches parametric tail accuracy","Semiparametric gauge for geometric extremes: simple and fast","Linear geometry for multivariate tails: no neural nets needed","Gauge function made piecewise-linear: accurate and interpretable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000921,"raw_usage":{"total_tokens":3934,"prompt_tokens":912,"completion_tokens":3022,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":2943}},"tokens_in":528,"tokens_out":3022,"duration_ms":22447,"temperature":1.0,"reasoning_tokens":2943,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:51:18.689810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fit the piecewise-linear truncated-gamma model with shape $d$ to $n=5000$ simulated observations from a $d=3$ distribution whose exact conditional radial density is a gamma with shape that varies strongly with $w$ (for example a multivariate $t$ copula or a skew-normal with heavy tail), estimate $\\Pr(X\\in B)$ for $B$ beyond the data range, and compare with a high-accuracy Monte Carlo estimate of the same probability; bias that grows with the shape variation would show the fixed-shape assumption fails at finite thresholds.","supporting_citations":[],"review_version":1}