{"id":"57d3b124-9c0c-4179-bbfa-b97170f846d4","arxiv_id":"2506.04480","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The authors formulate geodesic PCA in Wasserstein space through Otto's fiber bundle, solve it exactly for Gaussians, and approximate it for general measures with neural geodesic parameterizations.","lead":"This paper computes principal modes of variation for collections of probability distributions using Wasserstein geometry, both for Gaussian distributions and for general absolutely continuous distributions. It introduces a neural-network parameterization of Wasserstein geodesics and tests it on synthetic data, images, and 3D point clouds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"General a.c. GPCA does not certify that output curves are true Wasserstein geodesics: Hessian convexity is checked only on finite minibatches and φ is unconstrained.","rationale":"The reader's weakest_assumption identifies exactly the gap that I consider most load-bearing: in the general a.c. case, the computed curves are not certified to be true Wasserstein geodesics. My analysis sharpens the point: the key mathematical condition is not merely that φ is a diffeomorphism but that φθ#ρ is absolutely continuous and that Id + t Hfψ is positive semidefinite everywhere on R^d. The paper checks the Hessian condition only on finite minibatches, and φ is an unconstrained MLP with no a.c. guarantee. Appendix E explicitly acknowledges the φ issue, which is an unusual self-flagged limitation and should carry weight in the verdict. The Gaussian section is mathematically sound and supports the paper's Gaussian claims, so a rejection would be too harsh; however, the general a.c. 'exact' claim is not yet substantiated. The proposed concrete test would directly probe whether the failure actually occurs in representative experiments. If the check passes, the concern would be mitigated; if it fails, the paper would need to either constrain the parameterization (e.g., using normalizing flows for φ and rigorous Hessian bounds) or soften the exactness claim. The reader's CONDITIONAL verdict is therefore appropriate, and no adjustment is needed.","tokens_in":23098,"tokens_out":8750,"duration_ms":105050,"concrete_test":"Using the released code or a reimplementation, train GPCA GEN on one of the paper's synthetic experiments. Then draw 10^6 points from a heavy-tailed distribution (e.g., Student t with 3 degrees of freedom) and, for every learned projection time t_i, compute the minimal eigenvalue of Id + t_i Hfψ(x). If any eigenvalue is negative, the corresponding curve is not a true Wasserstein geodesic and the exactness claim is falsified. As a complement, repeat the same eigenvalue check on a dense grid over a ball of radius 5 around the origin to see whether the original Gaussian minibatch covered the relevant region; if eigenvalues fail outside that ball, the minibatch-based interval estimation is insufficient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GPCA GEN returns true Wasserstein geodesics minimizing (13) requires that every evaluated curve t ↦ (id + t∇fψ)#(φθ#ρ) actually is a geodesic. This requires (i) φθ#ρ to be absolutely continuous and (ii) x ↦ ‖x‖²/2 + t fψ(x) to be convex on all of R^d, equivalently Id + t Hfψ(x) ⪰ 0 for every x. Algorithm 1 line 5 estimates tmin and tmax using Hessian eigenvalues on a finite minibatch {x_k} drawn from the standard Gaussian ρ. With batch size 1024, this minibatch covers only a bounded region, while fψ is an MLP whose Hessian can take arbitrary values far outside that region. The paper's own Appendix E admits that φθ is not constrained to be a diffeomorphism and that a degenerate φ can make the computed curve degenerate, but no certificate is provided that φθ#ρ is a.c. or that the Hessian condition holds globally. Consequently, the optimized curve is only verified to satisfy the geodesic property on the sampled points; objective (14) is minimized over a superset of the true geodesics, so the returned components are not established to be true GPCA components. This is the single most load-bearing weakness because the paper's novelty rests on this 'exact' geodesic parameterization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper develops two geodesic PCA procedures in Wasserstein space. In the Gaussian case, it uses the Otto bundle over SPD matrices to lift the geodesic PCA problem to GL_d with horizontal lines, proves existence and equivalence, and compares with tangent PCA, including a second-order distortion formula. In the general absolutely continuous case, it parameterizes geodesics as t ↦ (id+t∇f_ψ)#(φ_θ#ρ) using MLPs, optimizes a Sinkhorn-based residual loss with soft orthogonality/intersection penalties, and demonstrates the method on synthetic MNIST geodesics, 3D point clouds, and image color distributions.","tokens_in":23431,"tokens_out":10448,"duration_ms":110627,"significance":"The Gaussian contribution is a valuable exact finite-dimensional formulation of Wasserstein GPCA; if the proofs are completed, it gives practitioners a principled alternative to tangent PCA and a quantitative account of when linearization fails. The neural parameterization of Otto geodesics is also a novel and promising idea, and the experiments are a useful sanity check. However, the paper's central claim of exactness for general a.c. measures is not supported by the implemented algorithm: the geodesic conditions are only checked on minibatches, the loss is a Sinkhorn approximation, and the constraints for later components are soft. With appropriate certification or reframing, the paper would be a solid contribution; as it stands, the general-case claims outrun the evidence.","major_comments":[{"comment":"The abstract and Section 1 state that in the a.c. case the components are 'true geodesics that minimize (1)' and that the method is 'exact'. The implemented procedure does not support this claim. First, the objective in Algorithm 1 line 7 is the Sinkhorn divergence S_ε, not W_2^2, so (14) is minimized only approximately; no convergence or bias result is provided for ε→0 or as the batch size grows. Second, the interval [tmin,tmax] that should guarantee that id+t∇f_ψ is a diffeomorphism is estimated from Hessian eigenvalues on a finite minibatch (line 5), which cannot certify global positive-definiteness for an MLP Hessian on all of R^d. Third, φ_θ is an unconstrained MLP, whereas the geodesic representation requires φ to be a diffeomorphism and φ_θ#ρ to be absolutely continuous; Appendix E explicitly notes that a non-diffeomorphic φ can make the computed curve degenerate. Consequently, the optimized curves are not certified to be Wasserstein geodesics, and the minimization is effectively over a superset of the true geodesics. The exactness claims should either be accompanied by certificates (for example, invertibility-by-construction architectures and global Hessian bounds) or be restated as an approximate algorithm.","section":"Section 4, Eq. (14), Algorithm 1, Appendix E"},{"comment":"The proof of Proposition 4 shows only that the Gaussian projection of any candidate geodesic has cost no larger than the candidate, and hence that the projected geodesic is also optimal; it does not imply that a given optimal geodesic coincides with its projection. Equality of optimal costs does not force equality of minimizers without a uniqueness or strict-convexity argument. As written, the statement that 'the first principal geodesic component ... remains in the space of Gaussian distributions' is not established. Please provide a uniqueness argument or weaken the proposition to the existence of a Gaussian optimal component.","section":"Section 3 / Proposition 4 (Appendix D.1, Proposition 13)"},{"comment":"The definitions of successive GPCA components require exact intersection and orthogonality, but the implementation uses soft penalties λ_I I and λ_O O with λ_I=λ_O=1.0 (Appendix E, Table 1). No constraint-satisfaction guarantee or post-hoc verification is given; the text only states that the regularization terms are monitored to decrease. Thus even if each estimated curve were a true geodesic, the output pair need not intersect orthogonally, and the result need not satisfy the recursive definition of GPCA components used in Section 3 and in the narrative of Section 4. This further weakens the claim that the method computes exact GPCA in the a.c. case.","section":"Section 4, second and higher-order components"}],"minor_comments":[{"comment":"The coercivity argument in the proof of Lemma 1 is difficult to follow: the 'diagonal' A=Σ_i^{1/2}Q_i is discussed before the constraint Q_i∈SO_d is exploited, and the conclusion is stated over (R^{d×d})^n. Coercivity on GL_d × B × SO_d^n follows directly from the compactness of the sets {Σ_i^{1/2}Q_i}; the proof should be reorganized accordingly.","section":"Appendix D.1, Lemma 1"},{"comment":"The projection p_{A,X}(t_i) and the interval [tmin,tmax] are used in Eq. (11) before they are defined; please introduce both before Proposition 3.","section":"Section 3, Eq. (11)"},{"comment":"The Sinkhorn divergence S_ε is used without definition in the main text; either define it and state its relation to W_2^2 in Section 4, or explicitly defer to Appendix E at first use.","section":"Algorithm 1, line 7"},{"comment":"The claim that GPCA reduces the objective by less than 1% relative to TPCA 'on average for 100 trials' is reported without standard deviations or a precise description of the 'uniform distribution on the spectral parameter space'; please specify the sampling procedure and report dispersion.","section":"Section 5.1"},{"comment":"The captions contain the typo 'GPCAG EN' for 'GPCA GEN', and Section 3 contains 'Propositon'; a proofreading pass is needed.","section":"Figures 9 and 10"}],"recommendation":"major_revision","confidential_remarks":"The reader's conditional verdict matches my own: the Gaussian part is essentially sound and publishable after a proof fix for Proposition 4, but the general a.c. part needs either substantial certification or an explicit downgrade from 'exact' to 'approximate'. The self-admitted limitation in Appendix E is central and should not be treated as a minor caveat. There is no code release or reproducibility statement in the manuscript; adding one would strengthen a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the Gaussian half of this paper is a genuine contribution: lifting GPCA to GL_d with horizontal line segments, with an existence proof and an equivalence result, is clean and works. Second, the general-a.c. half does not support the word 'exact' in the abstract; the stress-test note is right that the geodesic certificate is missing.\n\nWhat is new: the Gaussian formulation itself, plus the observation that GPCA generically matches TPCA and the derivation of a curvature-based distortion formula (Prop 5) that predicts when they differ. That is useful practical information. The neural parameterization of Otto geodesics—writing geodesics as (id + t∇fψ)#(φθ#ρ) and optimizing with MLPs—is novel and worth knowing even if the current implementation is approximate.\n\nWhere it softens: the general-a.c. pipeline needs id + t Hess fψ(x) ≻ 0 for every x in order for the curve to be a true Wasserstein geodesic. Algorithm 1 estimates tmin/tmax from Hessian eigenvalues on 1024 samples of the reference Gaussian; an MLP Hessian is unbounded, so this is not a global certificate. φθ is a plain MLP, not a diffeomorphism; Appendix E admits that a degenerate φ can make the curve degenerate. And the objective replaces W2 with Sinkhorn divergence. So the algorithm minimizes over a superset of the true geodesics, and the returned components are not certified to be geodesics. The second component's intersection/orthogonality are soft penalties (λI, λO), which further weakens the exactness claim. These are fixable: constrain φ, verify convexity on the actual support, quantify Sinkhorn bias, release code and data, report run-to-run variability. The Gaussian theory itself looks sound; Prop 4 is a nice univariate result, and the experiments on Gaussians back the claims. The lack of code and the qualitative nature of the a.c. experiments mean the general claims cannot be checked from the preprint. Citation pattern is fine—Otto, McCann, and Huckemann are the right anchors.\n\nWho should read it: anyone working on PCA for distribution-valued data, Wasserstein geometry, or shape analysis. It deserves a serious referee; I'd send it out and expect a major revision, not a rejection.","headline":"Gaussian GPCA is a solid, citable contribution; the general-a.c. method is promising but its 'exact' claim is not supported by the current geodesic certificate.","tokens_in":23921,"tokens_out":2695,"would_cite":true,"duration_ms":28328,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49Q22","62R30"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper shows that geodesic PCA in the Wasserstein space can be carried out exactly for centered Gaussians and approximated with neural-parameterized geodesics for general absolutely continuous measures.","keywords":["geodesic principal component analysis","Wasserstein space","Otto-Wasserstein geometry","Bures-Wasserstein distance","optimal transport","neural network geodesic parameterization","tangent PCA","probability measures"],"falsifier":"Evaluate the learned maps on a dense grid: if $\\mathrm{Id}+tH_{f_\\psi}$ has a non-positive eigenvalue for any $t$ within the fitted $[t_{\\min}, t_{\\max}]$, then the GPCA GEN curve is not a Wasserstein geodesic and the loss was minimized over an invalid family; in the Gaussian case, run the same-eigenvalue circle experiment with $|a-b|/|a+b|$ near $0.8$ and check whether GPCA's residual cost beats TPCA by the tens of percent that the paper's Figure 4 and equation (17) imply.","tokens_in":2161,"feed_emoji":"📐","tokens_out":3032,"duration_ms":80896,"temperature":0.7,"pith_summary":"The paper attacks a step that has blocked PCA on collections of probability measures: finding the geodesic curve in Wasserstein space that best explains the data's variation, without first flattening the space through a tangent approximation. For centered Gaussians it lifts the problem into the flat space of invertible matrices via Otto's fiber bundle, so the Geodesic PCA objective is optimized over true Wasserstein geodesics rather than linearized ones. For general absolutely continuous measures it parameterizes the same kind of geodesics through a reference diffeomorphism and a scalar potential, both represented by multilayer perceptrons, and fits them by minimizing the sum of squared projection residuals. The method gives curves that are genuine Wasserstein geodesics whenever the parameterization constraint is met, and it makes sampling along components possible. A sympathetic reader would care because tangent PCA, while cheap, can distort data on curved parts of the Wasserstein manifold, and the paper shows where that distortion matters.","feed_headline":"Wasserstein GPCA: exact for Gaussians, neural for all","feed_subtitle":"Finds true Wasserstein geodesics as principal components instead of linearized tangent-space approximations that distort curved data.","key_machinery":"The central object is Otto's fiber bundle over the Wasserstein space: the bottom space is probability measures (or SPD matrices for centered Gaussians), the top space is diffeomorphisms (or $\\mathrm{GL}_d$), and the projection is pushforward by a fixed reference measure (or $A\\mapsto AA^\\top$). A horizontal line segment in the top space projects to a Wasserstein geodesic, so every geodesic writes as $\\mu(t)=(\\mathrm{id}+t\\nabla f)_\\#(\\phi_\\#\\rho)$, with $f$ smooth and $\\mathrm{id}+t\\nabla f$ a diffeomorphism. This parametrization carries the whole argument: it replaces the curved optimization by a flat one, makes projection times explicit, distinguishes horizontal directions so that orthogonality of components has a concrete meaning in $L^2(\\rho)$, and, because the map is explicit, lets the practitioner sample points along any component. In the Gaussian restriction, the same machinery specializes to $A+tX$ with $X^\\top A = A^\\top X$ and gives the Bures-Wasserstein distance as a minimal Frobenius distance between fibers.","core_discovery":"On the paper's own terms, the central discovery is that Otto's Riemannian submersion picture turns Wasserstein GPCA from a geodesic-fitting problem on a curved space into a horizontal-line-fitting problem in a flat total space. In the Gaussian case the total space is $\\mathrm{GL}_d$ with the Frobenius metric, the geodesic is a line $A+tX$ with $X$ horizontal at $A$, and the cost becomes explicit in terms of rotations $Q_i$ representing each covariance in its fiber. The paper proves this reformulation minimizes the original residual objective, and shows that TPCA and GPCA generically agree except near the boundary of the SPD cone, where linearization distorts pairwise distances by a factor quantified in Proposition 5. For absolutely continuous measures, the analogous statement is that every Wasserstein geodesic has the form $(\\mathrm{id}+t\\nabla f)_\\#(\\phi_\\#\\rho)$, so fitting a geodesic component reduces to learning $\\phi$ and $f$; the paper's contribution is a neural parameterization and an algorithm (GPCA GEN) that jointly learns the component, the projection times, and, for the second component, an orthogonal intersection enforced by regularization.","pith_inferences":["Editorial inference: the paper's Gaussian analysis supplies a practical diagnostic for when tangent PCA is acceptable: compare the fitted horizontal line's endpoints with the barycenter and with the SPD cone boundary, as in the paper's Figure 4, and similar diagnostics could guide users of linearized Wasserstein distances elsewhere.","Editorial inference: the same neural parameterization could be compared head-to-head against input-convex neural network geodesic solvers on identical datasets; since Otto's $f$ need not be convex, such a comparison would clarify when convexity is a restrictive assumption.","Editorial inference: a natural stress test would replace the single scalar potential $f$ by a time-dependent potential, which would approximate non-geodesic interpolation structures while still projecting from a flat total space.","Editorial inference: the ratio $|a-b|/|a+b|$ acts as a curvature warning in the Gaussian case, and the paper's Proposition 5 suggests that a similar curvature-based heuristic could be derived for other positively curved manifolds."],"forward_implications":["For Gaussian datasets, practitioners can compute GPCA components by optimizing over $\\mathrm{GL}_d$ and $\\mathrm{SO}_d$, with a certified global minimizer for the first component, instead of relying on tangent-space PCA.","Because the Gaussian formulation is exact, it provides a benchmark to measure how much accuracy tangent PCA loses; Proposition 5 predicts that loss grows near the SPD cone boundary.","For general distributions, GPCA GEN gives geodesic components that can be sampled at any time, enabling interpolations and visualizations that discrete TPCA methods cannot produce directly.","The Otto parameterization avoids convexity constraints, replacing them with Hessian-eigenvalue monitoring, which suggests a new route for parameterizing Wasserstein geodesics without input-convex neural networks.","The univariate result that the first principal geodesic component stays Gaussian (Proposition 4) shows that the known one-dimensional coincidence between GPCA and linearized PCA persists within Gaussian families."],"supporting_citations":[{"why":"Supplies the fiber-bundle geometry of the Wasserstein space and the geodesic parametrization that the whole method is built on.","marker":"[28]"},{"why":"Defines geodesic PCA on Riemannian manifolds modulo isometric group actions, the objective that the paper solves exactly.","marker":"[15]"},{"why":"Gives the Brenier polar factorization underlying the Wasserstein geodesic structure and the Monge map.","marker":"[8]"},{"why":"Provides the Bures-Wasserstein geodesic projection formula for Gaussian measures used in the Gaussian GPCA formulation.","marker":"[36]"},{"why":"Gives the Wasserstein Riemannian geometry of Gaussian densities, including the Monge map between Gaussians.","marker":"[23]"},{"why":"Establishes the Bures-Wasserstein distance formula on SPD matrices that the Gaussian cost function relies on.","marker":"[4]"},{"why":"Shows the one-dimensional Wasserstein GPCA and linearized PCA coincide, grounding the univariate Gaussian result.","marker":"[5]"},{"why":"Documents the difficulty of parametrizing tangent vectors in Wasserstein space, motivating the exact Otto-geometry approach.","marker":"[33]"},{"why":"Defines Wasserstein barycenters, which anchor the tangent PCA baseline used throughout the comparisons.","marker":"[1]"},{"why":"Introduces the linearized optimal transport framework that tangent PCA is based on and that the paper compares against.","marker":"[39]"}],"fun_headline_variants":["True Wasserstein geodesics as principal components","Neural networks find true geodesic components on measure space","From tangent to true: Wasserstein GPCA","Geodesic PCA in measure space, exact for Gaussians, neural for all","Wasserstein GPCA: no tangent-space shortcuts"],"cache_read_input_tokens":25984,"weakest_assumption_plain":"In the general absolutely continuous case, the whole construction assumes the fitted maps $\\phi_\\theta$ and $\\mathrm{id}+t\\nabla f_\\psi$ are diffeomorphisms on the chosen time interval; the implementation neither enforces invertibility of $\\phi_\\theta$ nor verifies the Hessian condition outside the minibatches used to estimate $t_{\\min}$ and $t_{\\max}$, so the output curve may not be a true Wasserstein geodesic.","fun_headline_variants_meta":{"raw":{"variants":["True Wasserstein geodesics as principal components","Neural networks find true geodesic components on measure space","From tangent to true: Wasserstein GPCA","Geodesic PCA in measure space, exact for Gaussians, neural for all","Wasserstein GPCA: no tangent-space shortcuts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001219,"raw_usage":{"total_tokens":4990,"prompt_tokens":896,"completion_tokens":4094,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":4014}},"tokens_in":512,"tokens_out":4094,"duration_ms":30110,"temperature":1.0,"reasoning_tokens":4014,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:41:05.443434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the learned maps on a dense grid: if $\\mathrm{Id}+tH_{f_\\psi}$ has a non-positive eigenvalue for any $t$ within the fitted $[t_{\\min}, t_{\\max}]$, then the GPCA GEN curve is not a Wasserstein geodesic and the loss was minimized over an invalid family; in the Gaussian case, run the same-eigenvalue circle experiment with $|a-b|/|a+b|$ near $0.8$ and check whether GPCA's residual cost beats TPCA by the tens of percent that the paper's Figure 4 and equation (17) imply.","supporting_citations":[{"cited_title":"The geometry of dissipative evolution equations: the porous medium equation","cited_arxiv_id":null,"evidence_quote":"Supplies the fiber-bundle geometry of the Wasserstein space and the geodesic parametrization that the whole method is built on."},{"cited_title":"Intrinsic shape analysis: Geodesic PCA for Riemannian manifolds modulo isometric Lie group actions","cited_arxiv_id":null,"evidence_quote":"Defines geodesic PCA on Riemannian manifolds modulo isometric group actions, the objective that the paper solves exactly."},{"cited_title":"Polar factorization and monotone rearrangement of vector-valued functions","cited_arxiv_id":null,"evidence_quote":"Gives the Brenier polar factorization underlying the Wasserstein geodesic structure and the Monge map."},{"cited_title":"Wasserstein geometry of Gaussian measures","cited_arxiv_id":null,"evidence_quote":"Provides the Bures-Wasserstein geodesic projection formula for Gaussian measures used in the Gaussian GPCA formulation."},{"cited_title":"Wasserstein Riemannian geometry of Gaussian densities.Information Geometry, 1:137–179, 2018","cited_arxiv_id":null,"evidence_quote":"Gives the Wasserstein Riemannian geometry of Gaussian densities, including the Monge map between Gaussians."},{"cited_title":"On the Bures–Wasserstein distance between positive definite matrices","cited_arxiv_id":null,"evidence_quote":"Establishes the Bures-Wasserstein distance formula on SPD matrices that the Gaussian cost function relies on."},{"cited_title":"Geodesic PCA in the Wasserstein space by convex PCA","cited_arxiv_id":null,"evidence_quote":"Shows the one-dimensional Wasserstein GPCA and linearized PCA coincide, grounding the univariate Gaussian result."},{"cited_title":"Principal geodesic analysis for probability measures under the optimal transport metric","cited_arxiv_id":null,"evidence_quote":"Documents the difficulty of parametrizing tangent vectors in Wasserstein space, motivating the exact Otto-geometry approach."},{"cited_title":"Barycenters in the Wasserstein space","cited_arxiv_id":null,"evidence_quote":"Defines Wasserstein barycenters, which anchor the tangent PCA baseline used throughout the comparisons."},{"cited_title":"A linear optimal transportation framework for quantifying and visualizing variations in sets of images","cited_arxiv_id":null,"evidence_quote":"Introduces the linearized optimal transport framework that tangent PCA is based on and that the paper compares against."}],"review_version":1}