{"id":"36322f0e-5e21-4364-953e-e41c3fb86d73","arxiv_id":"2510.20424","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A closed-form divergence between conditional-extremes models yields a new clustering method for multivariate tail dependence that works in arbitrary dimensions.","lead":"The paper proposes a clustering method for multivariate extreme events that compares fitted conditional-extremes models with a closed-form, skew-geometric Jensen-Shannon divergence, then applies k-medoids. It is a new, computationally cheap way to group locations or variables with similar tail dependence, including in dimensions beyond the bivariate case.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6)/(10) define the CE conditional covariance as a scalar for d>2, invalidating the claimed closed-form multivariate JSG divergence as printed.","rationale":"The reader's weakest assumption was the Gaussian residual working assumption. That is a real limitation, but it is explicitly acknowledged in the Discussion and partially validated by simulations from Gaussian and t-copulas. The dimensional inconsistency in Eq. (6)/(10) is a more fundamental, unacknowledged internal error: it affects the very definition of the object from which the closed-form divergence is computed. Because the paper's headline novelty is the extension of the closed-form JSG divergence to arbitrary dimensions, and because the reader's verdict was already CONDITIONAL, the same verdict remains appropriate: the method may be salvageable with a corrected covariance formula and better reproducibility, but the central claim cannot be fully endorsed as written.","tokens_in":18576,"tokens_out":6083,"duration_ms":58516,"concrete_test":"Re-derive Eq. (6) from Eq. (4) for a trivariate example (d=3). Using component-wise multiplication, show that Var(Y_{-i} | Y_i=y) = diag(y^{β_{|i}}) Σ_{|i} diag(y^{β_{|i}}), not (y^{β_{|i}})^T Σ_{|i} y^{β_{|i}}. Then implement cJSG_λ with the corrected matrix for the Section 3.2.2 simulation setting and compare the resulting eJSG dissimilarity matrix with one computed from the printed scalar formula. If the printed formula cannot be evaluated as a matrix, or the two matrices differ substantially, the closed-form multivariate claim requires correction. If code is made available, inspect the covariance construction directly.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central contribution is a closed-form dissimilarity measure for multivariate tails, but Eq. (6) writes the conditional covariance of (Y_{-i,s} | Y_{i,s}=y) as (y^{β_{|i,s}})^T Σ_{|i,s} y^{β_{|i,s}}. For d>2, β_{|i,s} is a (d−1)-vector, so this expression is a scalar, not a (d−1)×(d−1) covariance matrix. The same scalar is inverted in Ω_λ(y) and m_λ(y) in Eq. (10), making the closed-form application of Eq. (9) undefined in the multivariate setting. The bivariate case (d=2) hides this because y^β is scalar, which may explain why the derivations appear to work there. No code or data are provided to check whether the implementation actually uses diag(y^β) Σ diag(y^β) or some other matrix. Since the paper explicitly claims 'applicable in arbitrary dimensions' and presents the multivariate extension as its key novelty over Vignotto et al. (2021), this is not a cosmetic typo: the central derivation is internally inconsistent for the multivariate case as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a clustering method for multivariate extremes based on the conditional extremes (CE) framework of Heffernan and Tawn (2004). For each location/vector, a CE model is fitted for each conditioning variable under the working assumption of Gaussian residuals. The paper then defines a skew-geometric Jensen-Shannon (JSG) divergence between the fitted conditional Gaussian distributions, computes its expectation conditional on threshold exceedance, builds a pairwise dissimilarity matrix, and applies PAM clustering. Claims include a closed-form divergence, applicability in arbitrary dimensions, superior performance over Vignotto et al. (2021) in bivariate simulations, and a meaningful application to Irish precipitation/wind-speed data. The paper also reports extensions to d=3 and d=5 simulations and discusses limitations around the Gaussian residual assumption and uncertainty propagation.","tokens_in":18915,"tokens_out":7389,"duration_ms":66618,"significance":"If the multivariate derivation is correct, the method is a useful and computationally efficient tool for clustering tail-dependence structures, particularly because it handles both asymptotic dependence and asymptotic independence and avoids the bivariate restriction of existing nonparametric approaches. The paper is honest about the Gaussian working assumption and about the lack of uncertainty propagation, and the simulation studies are reasonably extensive. The main methodological novelty is the use of a closed-form divergence between fitted CE models to obtain a dissimilarity matrix. However, the central multivariate derivation as printed contains a dimensional error in the conditional covariance expression, which affects the claimed general-d applicability. The paper does not provide code or data, and the simulation evidence for d>2 cannot be checked against the printed equations.","major_comments":[{"comment":"Equation (6) states that the conditional distribution of Y_{-i,s} | Y_{i,s}=y is MVN with covariance matrix written as (y^{β_{|i,s}})^T Σ_{|i,s} y^{β_{|i,s}}. For d>2, β_{|i,s} is a (d-1)-vector and y^{β_{|i,s}} is also a (d-1)-vector, so this expression is a scalar, not a (d-1)×(d-1) matrix. Equation (10) then substitutes this scalar into the matrix formula (9), making the multivariate closed-form derivation invalid as printed. The bivariate case hides the problem because y^{β} is a scalar when d=2. The manuscript claims applicability in arbitrary dimensions, and the d=3 and d=5 simulations cannot be reproduced from Eq. (6). Please correct Eq. (6) to use diag(y^{β_{|i,s}}) Σ_{|i,s} diag(y^{β_{|i,s}}), and update Eq. (10) and the surrounding text accordingly. This is load-bearing for the paper's central claim.","section":"Section 2.3, Eq. (6) and Eq. (10)"},{"comment":"The closed-form divergence is between fitted Gaussian conditional models, not directly between the true tail dependence structures. The paper acknowledges this, and the simulations show that clustering based on the fitted Gaussian models works for Gaussian and t-copula mixtures. However, the t-copula residuals are not Gaussian, so the simulation results indicate robustness rather than exactness. This is an acceptable limitation if clearly stated, but the phrase 'dissimilarity measure for multivariate tails' in the abstract should be tightened to 'dissimilarity between fitted CE models under the Gaussian working assumption', especially since the application and simulations all rely on that assumption.","section":"Sections 2.3 and 3.2.2"},{"comment":"The d=3 and d=5 simulation results are presented as evidence for multivariate applicability, but the printed equations do not specify how the covariance matrix is computed in those settings. The paper says the same ρ_t is used for all variable pairs, which makes the simulation favourable to aggregation across conditioning variables, but this does not resolve the dimensional inconsistency. Please either provide the corrected multivariate formula and a derivation, or include code so that the implementation can be verified.","section":"Section 3.2.2 and Eq. (6)"}],"minor_comments":[{"comment":"The definition of JSGλ in Eq. (7) has no factor 1/2, while the closed-form formula in Eq. (9) has a leading 1/2. This is likely a convention difference, but it should be reconciled so the equations are consistent. Since a constant multiple does not affect clustering, this is not a substantive issue.","section":"Section 2.3, Eq. (7) vs Eq. (9)"},{"comment":"The text says 500 replicated experiments are performed, but Figure 2's caption refers to 'bootstrap samples'. Please make the terminology consistent; these are simulation replications, not bootstrap resamples.","section":"Section 3.2 and Figure caption 2"},{"comment":"The empirical χ(0.95) values are useful, but it would be clearer to report their uncertainty, even briefly. Many sites have very low χ estimates, and it is unclear whether those are distinguishable from zero.","section":"Section 4.1 and Figure 6"},{"comment":"The supplement says 'Data and R code: Available upon request from the corresponding author.' For reproducibility, please deposit code and processed data in a public repository or add a detailed pseudo-code appendix, particularly because the multivariate divergence formula needs to be verified.","section":"Supplementary material"},{"comment":"The TWGSS elbow criterion is standard, but the paper says 'Unreported experiments showed that the TWGSS method correctly identified k' in Section 3. This is an unsupported claim; either report those experiments or soften the statement.","section":"Section 2.4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is interesting and likely correct in intent, but the printed central equation for the multivariate covariance has a dimensional error that must be fixed. The d>2 simulations suggest the authors know the correct form, but without code or a corrected derivation, the main claim cannot be verified. I recommend major revision with a request for a corrected derivation and reproducibility materials."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe one thing to know: this paper has a good, useful idea — a parametric, closed-form dissimilarity for CE models, computed from the skew-geometric Jensen-Shannon divergence, then fed into PAM clustering — but the central derivation as printed has a dimensional error that invalidates the multivariate closed-form. Eq (6) writes the conditional covariance of the (d−1)-vector Y_{−i,s} as (y^{β_{|i,s}})^T Σ_{|i,s} y^{β_{|i,s}}, which is a scalar for d>2, not a matrix. The same scalar is inverted in Eq (10). So the closed-form application of the JSG formula is undefined in the multivariate setting. This only works in the bivariate case, where y^β is a scalar. It looks like a fixable typo — presumably they intended diag(y^β) Σ diag(y^β) or an equivalent — but as it stands the paper's main claim is not supported by its equations. No code or data are provided to check what was actually implemented.\n\nThat said, the paper deserves credit for the construction itself. Combining the CE framework with Nielsen's JSG divergence is a natural and smart move, and the expected divergence, aggregated over conditioning variables, gives a symmetric dissimilarity that handles both asymptotic dependence and independence. The method is cheap, conceptually simple, and fills a real gap left by bivariate nonparametric approaches. The simulation study is extensive, and the Irish application produces spatially coherent clusters without using spatial information, which is a nice result.\n\nThe other soft spots are smaller. The \"uniquely extends\" claim is not backed by any comparison to a multivariate baseline — there may not be one, but then say so explicitly. The Gaussian working assumption for CE residuals is acknowledged, but validation only covers Gaussian and t-copulas, so the dissimilarity could in practice reflect model misfit rather than tail dependence. The Gaussian-copula simulation is also partly in-sample, since CE is nearly the true model there.\n\nFor whom: this is for people working on clustering multivariate or spatial extremes. It would be a useful contribution once the dimensional issue is fixed and the implementation is made available. I'd send it to peer review, but the revision needs to correct the covariance notation, provide code or at least a precise statement of the matrix form, and temper or substantiate the multivariate claim.\n\nBest.","headline":"The method is a good idea, but the printed covariance formula makes the multivariate closed-form a scalar, so the central derivation needs fixing before the claims hold.","tokens_in":19328,"tokens_out":2587,"would_cite":false,"duration_ms":23257,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G32","62H30"],"pacs":[],"model":"deepseek-v4-flash","headline":"A closed-form divergence measure between conditional-extremes models makes clustering of multivariate tail dependence tractable in any dimension and beats bivariate-only methods.","keywords":["conditional extremes","multivariate extremes","tail dependence","clustering","skew-geometric Jensen-Shannon divergence","k-medoids","extreme value analysis","compound weather events"],"falsifier":"Simulate data from conditional-extremes models whose residuals are deliberately non-Gaussian (e.g., skew-t or chi-square distributions) with known cluster labels, compute the Gaussian-assumption eJSGλ dissimilarities, and check whether clustering recovers the labels; if adjusted Rand index collapses while labels are still recoverable from the true conditional distributions, the divergence is tracking model misfit.","tokens_in":18498,"feed_emoji":"🌪️","tokens_out":7017,"duration_ms":57845,"temperature":0.7,"pith_summary":"The paper is trying to establish that the tail-dependence structure of a set of random vectors — for instance, precipitation and wind speed at many weather stations — can be grouped automatically by comparing their fitted conditional-extremes models, rather than by pairwise correlation coefficients alone. The proposed dissimilarity is the expected skew-geometric Jensen-Shannon divergence between the Gaussian conditional distributions implied by the conditional-extremes model, which has a closed form and is cheap to compute. Feeding these pair-wise divergences into a standard k-medoids algorithm yields clusters whose members share homogeneous joint tail behaviour, in arbitrary dimension. The paper argues this approach outperforms the existing nonparametric competitor in bivariate simulations and is the first to extend such clustering to more than two variables. If correct, it gives environmental and financial analysts a fast, model-based way to pool information across sites and to interpret spatially coherent extreme-event regimes.","feed_headline":"Closed-form divergence groups sites by extreme tail dependence","feed_subtitle":"Replaces costly tail comparisons with a formula, then clusters with k-medoids — in any dimension.","key_machinery":"The central object is the expected conditional skew-geometric Jensen-Shannon divergence (eJSGλ). For a chosen conditioning variable i and two sites s and s*, it averages, over extreme values y of the conditioning variable, the skew-geometric Jensen-Shannon divergence between the two fitted Gaussian conditional distributions of the remaining d−1 variables given Yi = y. The skew-geometric Jensen-Shannon divergence is a symmetric, bounded, non-negative divergence built from the weighted geometric mean of two distributions; for Gaussians it reduces to a closed form in the precision-weighted mean and covariance. The expectation exploits the standard Laplace margins, whose exceedance density is id","core_discovery":"On its own terms, the paper claims that the expected conditional skew-geometric Jensen-Shannon divergence between fitted conditional-extremes models is a one-number summary of multivariate tail dependence that preserves enough information for meaningful clustering. Under the standard working assumption that the residual vector in the conditional-extremes model is multivariate Gaussian, the divergence between two fitted conditional distributions has a closed-form expression, and the expectation over the conditioning variable above the threshold is computed by a cheap Monte Carlo integral. Clustering the resulting dissimilarity matrix with the Partitioning Around Medoids algorithm produces spa","pith_inferences":["The load-bearing Gaussianity of the residuals is a modelling convenience, not a law: with misspecified residuals the closed form measures the divergence between fitted Gaussian approximations, and a numerical divergence computed from non-Gaussian fitted residuals would be the safer alternative in data with clearly skewed or heavy-tailed conditional tails.","The same exponential-family closure that makes the skew-geometric Jensen-Shannon divergence closed-form for Gaussians should extend to other exponential residual families, so the recipe generalizes beyond the particular working assumption used here.","Because the paper does not propagate parameter uncertainty into the dissimilarity matrix, cluster labels are point estimates; a bootstrap over fitted parameters would yield a distribution over clusterings, at a computational cost that the closed form was designed to avoid.","The claim of arbitrary dimension is established for d up to 5 with exchangeable pairwise dependence; the realistic high-dimensional case with heterogeneous, possibly sparse dependence across variable pairs is the natural next stress test."],"forward_implications":["Clustering first, pooling second: data from sites assigned to the same cluster can be pooled and the conditional-extremes model refitted, reducing parameter variance and bias, as the Gaussian-copula simulation demonstrates.","The method extends tail-dependence clustering to any dimension d, where the only existing benchmark is restricted to bivariate data; simulation shows accuracy rising with dimension up to d = 5.","The closed form keeps the method fast enough for realistic applications: 60 sites in simulation, 59 stations in the Irish application, with cluster assignments largely stable as the fitting threshold varies between the 0.85 and 0.90 quantiles.","Applied to weekly Irish precipitation and wind-speed data, the clusters define three spatially coherent regions — east, central, west — without any spatial information being used, and single out known outlier sites such as Malahide Castle.","The method handles both asymptotic dependence and asymptotic independence, unlike several existing extremal-clustering approaches that assume one regime only."],"fun_headline_variants":["Formula-based tail similarity clusters extremes in any dimension","New metric: skew-geometric JS divergence for tail clustering","Clustering extreme dependence via closed-form conditional models","From tails to groups: a fast divergence for multivariate extremes","Tail dependence clustering goes high-dimensional with closed form"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The closed-form divergence assumes the conditional-extremes model residuals are multivariate Gaussian; if the true conditional tails are non-Gaussian, the pairwise distances measure mismatch between fitted Gaussian approximations rather than difference between the actual tail distributions.","fun_headline_variants_meta":{"raw":{"variants":["Formula-based tail similarity clusters extremes in any dimension","New metric: skew-geometric JS divergence for tail clustering","Clustering extreme dependence via closed-form conditional models","From tails to groups: a fast divergence for multivariate extremes","Tail dependence clustering goes high-dimensional with closed form"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000604,"raw_usage":{"total_tokens":2602,"prompt_tokens":637,"completion_tokens":1965,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":381,"completion_tokens_details":{"reasoning_tokens":1890}},"tokens_in":381,"tokens_out":1965,"duration_ms":11619,"temperature":1.0,"reasoning_tokens":1890,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T08:25:37.526831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate data from conditional-extremes models whose residuals are deliberately non-Gaussian (e.g., skew-t or chi-square distributions) with known cluster labels, compute the Gaussian-assumption eJSGλ dissimilarities, and check whether clustering recovers the labels; if adjusted Rand index collapses while labels are still recoverable from the true conditional distributions, the divergence is tracking model misfit.","supporting_citations":[],"review_version":1}