{"id":"f2043b8b-8ce7-4b81-bf97-42db0d0af8db","arxiv_id":"2606.19011","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors show that FPCA on multivariate densities in Bayes space is equivalent to multivariate FPCA on their independent-interactive decomposition, with the variance decomposition being PCA-optimal.","lead":"This paper develops a dimension reduction approach for multivariate probability density functions by placing them in the Bayes space and applying functional principal component analysis with a new orthogonal decomposition. A smart generalist might read it to see how structured variance decomposition can improve interpretation of density data in applied fields like geology or housing analysis.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's weakest_assumption correctly isolates the clr isometry, which is the sole non-trivial prerequisite for all subsequent claims. Because that isometry holds by construction, the optimality and equivalence statements follow without additional load-bearing gaps. The low-confidence UNVERDICTED rating is therefore attributable only to the absence of the full manuscript at the time of the first review.","tokens_in":1768,"tokens_out":287,"duration_ms":21574,"concrete_test":"Apply the clr map and the claimed orthogonal decomposition to a simple product of two independent Beta densities on [0,1]²; recompute the total variance both before and after decomposition and verify that the sum of the independent-part and interaction-part variances equals the original total variance (within floating-point tolerance).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claims (PCA-optimality of the independent/interactive variance decomposition and equivalence of direct vs. decomposed FPCA) rest on the clr map being an isometric isomorphism from the Bayes space of multivariate densities into a subspace of L². This property is standard in the Bayes-space literature and directly supplies the Hilbert-space inner product needed for both the orthogonal decomposition and the subsequent FPCA. No internal inconsistency, hidden regularity assumption, or failure of the isometry in the multivariate setting is visible in the stated results.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript develops dimension reduction techniques for multivariate probability density functions within the Bayes space framework. It utilizes the centred logratio (clr) transformation to establish an isometric isomorphism with a subspace of L², allowing the application of functional principal component analysis (FPCA). The key contributions include an orthogonal decomposition of multivariate densities into independent and interactive components, with the independent part further decomposed into orthogonal geometric marginals. The paper proves that this variance decomposition is optimal in a PCA sense and demonstrates the equivalence of applying FPCA directly to the densities versus to their decomposed form, with corresponding decomposition of eigenfunctions and scores. The theoretical results are illustrated with applications to housing and geological data.","tokens_in":1852,"tokens_out":360,"duration_ms":19189,"significance":"If the results hold, this provides a significant advancement in the analysis of multivariate density data by offering a structured way to decompose and interpret variance sources. The reliance on the standard clr isometry ensures the framework is built on solid Hilbert space foundations, and the optimality and equivalence results could influence how FPCA is applied and interpreted in compositional data analysis. The empirical applications demonstrate practical utility. The use of an established isometric isomorphism and the focus on reproducible theoretical structure are strengths.","major_comments":[],"minor_comments":[{"comment":"Abstract: the phrase 'equivalent in a certain sense' is imprecise; a brief clarification of the precise sense of equivalence (e.g., with respect to the inner product or the resulting scores) would improve readability without altering the claim.","section":null},{"comment":"The manuscript would benefit from an explicit statement early in the introduction of how the geometric marginals are defined and why they are mutually orthogonal under the Bayes-space inner product.","section":null}],"recommendation":"minor_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the positive summary, significance assessment, and recommendation of minor revision. No major comments were listed in the report.","responses":[],"tokens_in":1285,"tokens_out":46,"duration_ms":7053,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main point is that this work takes the clr-based Bayes space framework, already used for univariate densities, and shows how to decompose multivariate ones into independent geometric marginals plus an interactive term, with the split being optimal for PCA and the two FPCA routes being equivalent in the resulting eigenfunctions and scores.\n\nThe authors use the established isometry to get an orthogonal decomposition of total variance that respects the PCA criterion. This should make the eigenfunctions more directly interpretable as sources of variation. They then apply the method to housing and geological density data, where the decomposition separates marginal effects from interactions in a way that standard multivariate FPCA would not highlight as cleanly.\n\nThe equivalence result is the clearest new piece; it is not just a restatement of the univariate case. The optimality claim follows from the Hilbert space structure once the decomposition is in place. Both rest on the clr map being an isometry, which is standard but applied here to the multivariate setting in a non-routine way.\n\nThe soft spot is that the abstract and summary give no explicit proof steps or counter-checks, so it is difficult to judge whether the multivariate extension introduces any extra regularity conditions or edge cases. The data examples are illustrative rather than comparative, so they show usability but do not test whether the optimality translates to better out-of-sample performance.\n\nThis paper is for researchers already working in functional data analysis on compositional or density objects. It refines an existing toolkit rather than opening a new area. The claims are specific enough that a serious referee in the subfield could check them, so it deserves peer review.","headline":"The paper extends Bayes-space FPCA to multivariate densities by proving equivalence between direct and decomposed FPCA plus PCA-optimality of the independent-interactive variance split.","tokens_in":2365,"tokens_out":400,"would_cite":false,"duration_ms":13978,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Multivariate densities in Bayes space decompose orthogonally into independent geometric marginals and an interactive component, making FPCA equivalent to separate multivariate analyses on the parts.","keywords":["Bayes space","multivariate densities","centred logratio transformation","functional principal component analysis","dimension reduction","orthogonal decomposition","variance decomposition","geometric marginals"],"falsifier":"If the eigenfunctions and scores from direct FPCA on multivariate densities fail to match the decomposed versions up to the claimed additive structure, or if the variance explained by the parts is not maximal among all orthogonal splits, the optimality and equivalence would not hold.","tokens_in":2653,"feed_emoji":"","tokens_out":733,"duration_ms":25053,"temperature":0.7,"pith_summary":"The paper shows that the Bayes space structure, via the centred logratio transformation, lets multivariate probability densities be split into mutually orthogonal parts: geometric marginals that capture independent variation and a remaining interactive component. This split decomposes the total variance in a way that is optimal for principal component analysis, so the eigenfunctions and scores from FPCA on the full densities break down cleanly into contributions from each part. The equivalence means one can run FPCA directly on the densities or on the decomposed pieces and obtain matching results. A reader would care because the approach turns the constrained, relative nature of density data into a geometrically natural setting where dimension reduction reveals separate sources of variation rather than mixing them.","feed_headline":"Bayes space splits multivariate densities for optimal PCA","feed_subtitle":"Orthogonal decomposition into independent marginals and interactive part makes FPCA results decompose and interpret naturally.","key_machinery":"The orthogonal decomposition of multivariate densities into independent geometric marginals and interactive component, enabled by the centred logratio (clr) transformation that gives an isometric isomorphism to an L² subspace.","core_discovery":"Embedding multivariate PDFs in the Bayes space enables an orthogonal decomposition into independent and interactive components, with the independent part further split into mutually orthogonal geometric marginals. The centred logratio transformation maps this structure isometrically to a subspace of L², so functional principal component analysis applies directly. The resulting variance decomposition is optimal in the PCA sense, and applying FPCA to the original densities is equivalent to multivariate FPCA on the decomposed form, with eigenfunctions and scores decomposing accordingly.","pith_inferences":["The same orthogonal split could be used with other functional data techniques such as functional regression or clustering on density data.","Fields that routinely work with joint distributions, such as compositional data or spatial statistics, might adopt the geometric marginals as a standard way to separate marginal and dependence effects.","Simulated examples with known independent and dependence structures could be used to check whether the PCA optimality holds numerically beyond the theoretical proof."],"forward_implications":["The decomposition of total variance is optimal in a PCA sense, so eigenfunctions and scores from FPCA have a direct interpretation in terms of independent and interactive contributions.","FPCA applied directly to multivariate densities produces results equivalent to multivariate FPCA performed on the decomposed independent and interactive parts.","Eigenfunctions and scores obtained from the full densities decompose additively according to the independent and interactive split.","The decomposition applied to empirical housing and geological data yields interpretable components that separate sources of variation."],"fun_headline_variants":["Orthogonal Bayes space decomposition for multivariate density FPCA","Bayes space maps densities to L2 for direct FPCA application","Geometric marginals in Bayes space enable optimal variance decomposition","FPCA on Bayes-embedded densities equals multivariate decomposed FPCA"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The centred logratio transformation establishes an isometric isomorphism between the Bayes space and a subspace of L² space.","fun_headline_variants_meta":{"raw":{"variants":["Orthogonal Bayes space decomposition for multivariate density FPCA","Bayes space maps densities to L2 for direct FPCA application","Geometric marginals in Bayes space enable optimal variance decomposition","FPCA on Bayes-embedded densities equals multivariate decomposed FPCA"]},"model":"grok-4.3","cost_usd":0.004794,"raw_usage":{"total_tokens":2366,"prompt_tokens":682,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":47937000,"prompt_tokens_details":{"text_tokens":682,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1620,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":682,"tokens_out":64,"duration_ms":13936,"temperature":1.0,"reasoning_tokens":1620,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T20:03:30.888206+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"If the eigenfunctions and scores from direct FPCA on multivariate densities fail to match the decomposed versions up to the claimed additive structure, or if the variance explained by the parts is not maximal among all orthogonal splits, the optimality and equivalence would not hold.","supporting_citations":[],"review_version":1}