{"id":"8f0ee367-517e-446a-8c11-63bea3ea7f9f","arxiv_id":"2412.00785","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Proper latent decomposition (PLD) is introduced as a nonlinear analogue of proper orthogonal decomposition: it computes principal geodesics on the autoencoder-inferred manifold of a fluid flow, demonstrated on a laminar wake and turbulent Kolmogorov flow.","lead":"This paper proposes a method to find a few meaningful modes (principal geodesics) in a curved, low-dimensional description of a fluid flow, using an autoencoder to learn the curved space and differential geometry to navigate it. If it works, turbulent flows could be compressed and interpreted with a handful of physically meaningful coordinates instead of thousands of grid points.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim relies on a well-conditioned autoencoder-inferred metric, but it never verifies conditioning for the exact models used in the results, so the reported geodesic modes may be numerical artifacts.","rationale":"The central claim is that PLD computes energy-ordered principal geodesics on the autoencoder-inferred manifold, and that for the Kolmogorov flow the leading mode is a physical realization. For this to hold, the geodesic computations must be accurate. The paper's own Section 7.3 states that standard autoencoder training produces metrics with exponential magnification factors and poor condition numbers, making distance and geodesic computations challenging. The authors say they introduced regularization to alleviate this, but they never demonstrate that the actual trained autoencoders used in Sections 7.1 and 7.2 have well-conditioned metrics on the data support. The only conditioning evidence, Figure 6, is for a 'prototypical' autoencoder, not the production models. This gap is load-bearing because the Eikonal solver (Section 5), the log-map shooting (Section 6), and the Frechet mean (Section 3.1) all depend on g and its inverse; an ill-conditioned g corrupts every downstream quantity, so the SVD modes may not correspond to true geometric variance. The reader's weakest assumption identifies exactly this issue, and I agree. Other concerns, such as the overclaimed 'first manifold-aware ROM' status and the absence of code, are real but secondary: they affect interpretation and reproducibility, not the internal correctness of the algorithm. The proposed concrete test, computing the condition-number distribution for the actual autoencoders and recomputing modes under a spectrally truncated metric, would settle whether the reported 'physical' mode is stable. If the mode is stable, the central claim survives; if not, the results are numerical artifacts. Therefore, the appropriate verdict remains CONDITIONAL pending this verification, and my read does not change the reader's verdict.","tokens_in":8552,"tokens_out":5366,"duration_ms":46969,"concrete_test":"Using the exact trained autoencoder from Section 7.2 (and if possible Section 7.1), compute the metric g(z)=J_D(z)^T J_D(z) at all encoded training and validation points, where J_D is the decoder Jacobian. Report the distribution of condition numbers kappa(g)=lambda_max/lambda_min and of the magnification factor sqrt(det g). If the median kappa exceeds 10^2 or the 95th percentile exceeds 10^4 on the data support, the Eikonal and log-map solvers are likely unreliable. Then recompute the leading PLD mode after spectrally truncating g at each point to enforce kappa <= 10^2, and compare the decoded geodesic field to the reported one. A qualitative change in the mode would confirm that the reported 'physical' mode is an artifact of metric ill-conditioning, whereas a stable mode would support the central claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PLD's validity requires the autoencoder-inferred metric g to be well-conditioned on the data support. The Eikonal solver in Section 5 computes distances using g and its inverse; the log-map shooting in Section 6 integrates the gradient of the distance function, which involves g^{-1}; and the Frechet mean in Section 3.1 solves distance minimization problems. If g has a large condition number, these computations become inaccurate or unstable, so the tangent-space SVD modes and the resulting principal geodesics are not faithful to the data manifold. Section 7.3 explicitly concedes that standard autoencoder training produces metrics with exponential magnification factors and poor condition numbers, and states that regularization was introduced for stability. However, the paper provides no quantitative conditioning or magnification-factor statistics for the actual trained autoencoders used in Sections 7.1 and 7.2; Figure 6 shows only a 'prototypical' autoencoder, not the production model. Without this check, the leading 'physical' mode in the Kolmogorov case and the semi-analytical fit in the laminar case could be artifacts of the numerical scheme rather than true geometric modes. This is the single most load-bearing concern because every downstream step (mean, log map, SVD, exp map) feeds on the metric.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces proper latent decomposition (PLD), a method that generalizes proper orthogonal decomposition (POD) to nonlinear manifolds. The pipeline consists of: (i) training a convolutional autoencoder to map high-dimensional fluid data to a low-dimensional latent manifold and inferring a Riemannian metric from the decoder; (ii) computing geodesic distances and a Fréchet mean using a metric-constrained Eikonal solver; (iii) mapping data to the tangent space at the mean via a log-map that combines gradient-based trajectory interpolation with multiple shooting; and (iv) performing SVD in the tangent space and mapping the principal components back to the manifold via the exponential map. The method is demonstrated on a laminar bluff-body wake and a two-dimensional turbulent Kolmogorov flow. For the laminar case, the tangent-space data are fit with an ellipse, yielding a single-variable description of the flow. For the Kolmogorov flow, the leading geodesic mode is visualized and claimed to be a physical realization of the governing equations.","tokens_in":8895,"tokens_out":4276,"duration_ms":35810,"significance":"If the claims hold, PLD offers a principled, manifold-aware extension of POD that could be valuable for interpretable nonlinear reduced-order modeling. The paper is commendable for being explicit about the numerical challenges (Section 7.3) and for proposing a geometric regularization to stabilize the metric. However, the verification is almost exclusively qualitative: there are no error bars, no quantitative comparison with POD, no conditioning diagnostics for the production models, and no reconstruction-error or energy metrics for the modes. The 'semi-analytical expression' for the laminar wake is a fitted ellipse, not a derived solution. These gaps leave the central claim plausible but not yet established.","major_comments":[{"comment":"The paper's central pipeline (Sections 3–6) relies on the metric tensor g inferred by the autoencoder being well-conditioned on the data support, because the Eikonal solver (Section 5), the log-map shooting (Section 6), and the Fréchet mean (Eq. 3.1) all involve g or g^{-1}. Section 7.3 concedes that standard autoencoder training produces metrics with 'exceedingly large' condition numbers and exponential magnification factors, and states that regularization was introduced to stabilize the metric. Yet no quantitative conditioning or magnification-factor statistics are reported for the actual autoencoders used in Sections 7.1 and 7.2; Figure 6 shows only a 'prototypical' autoencoder trained on the Kolmogorov data. Consequently, the reliability of the distance field, log maps, and hence the principal geodesics displayed in Figures 2–5 is unverified. This is the single most load-bearing issue because every downstream step depends on g.","section":"Section 7.3"},{"comment":"The 'semi-analytical expression' for the laminar wake is an ellipse fitted to the tangent-space coordinates obtained by SVD of the encoded data; its parameters (center, amplitude, angle) are fitted values. The paper claims this expression describes 'the solution of Navier-Stokes' (Abstract) and 'a semi-analytical solution of the Navier-Stokes equations' (Section 8). This is circular and overstated: the expression is a parametric curve in the learned latent space, not a solution or invariant manifold of the governing equations. Please clarify the precise sense in which this is 'semi-analytical' and how the fit differs from a regression.","section":"Section 7.1, Figure 2, Eq. (3.2)"},{"comment":"The claim that the leading Kolmogorov mode is 'a physical realization of the governing equations' is not supported by quantitative evidence. There is no comparison of the geodesic mode with POD in terms of reconstruction error, captured energy, or projection error; no error bars on the mode trajectory; and no check that the decoded mode lies on the attractor (e.g., by evolving it under the Navier-Stokes equations). Without such metrics, the 'physical' characterization is a visual judgment.","section":"Section 7.2, Figure 5"},{"comment":"The log-map computation relies on numerical approximations of the distance function and its gradient. The paper does not report the error of the final geodesic (e.g., endpoint mismatch or residual of the geodesic equation) for either test case, nor how the multiple-shooting refinement improves the initial interpolated trajectory. Since the tangent-space SVD and the resulting modes inherit errors from these approximations, it would be important to quantify these errors, especially for the Kolmogorov case where the manifold is less regular.","section":"Section 6, Eqs. (6.2)–(6.3)"}],"minor_comments":[{"comment":"The Eikonal solver is described only by referring to previous work (Kelshaw & Magri 2024); please state which components are novel in this paper and provide enough detail for reproducibility, or clearly delineate the contributions.","section":"Section 5"},{"comment":"The color scales of the vorticity snapshots are not labeled; please add color bars and state the quantity and units being plotted.","section":"Figures 3 and 5"},{"comment":"The meaning of 'mean residual of 1.734 × 10^{-3}' for the distance function should be defined (e.g., residual of the Eikonal equation or mean absolute error relative to a reference).","section":"Section 7.1"},{"comment":"The sentence 'we propose the proper latent decomposition (PLD) (Magri & Doan 2022)' reads as if the method was proposed in the cited reference; please rephrase to clarify the present contribution.","section":"Section 3"},{"comment":"The statement 'distinct clustering, amenable to singular value decomposition' is vague; state how clustering is measured or why it implies suitability for SVD.","section":"Section 7.2"},{"comment":"The notation dg|b(a) is confusing; consider writing dg(b,a) or dg(b)(a) and explaining the restriction of the distance function.","section":"Eq. (6.2)"}],"recommendation":"major_revision","confidential_remarks":"This is a CTR Summer Program proceedings paper, so the expectation of completeness may be lower than for a full journal paper. Nevertheless, the central methodological claim—that PLD produces reliable geodesic modes—is not yet supported by quantitative validation. The paper should be encouraged to add conditioning diagnostics for the production autoencoders and a quantitative comparison with POD; within a proceedings format, a focused revision addressing the metric-conditioning and circularity issues could make the contribution acceptable. The reader's report is fair."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine integration of existing pieces rather than a new theoretical breakthrough. What is new is the end-to-end pipeline: infer a latent manifold with an autoencoder, compute geodesic distances with a learned Eikonal solver, then extract principal geodesics by SVD in the tangent space at the Fréchet mean. Two fluid cases show the pipeline works well enough to produce plausible modes, and for the laminar wake the latent structure really does look like an ellipse, which is a nice sanity check. That is real work, and the demonstrations are the best part.\n\nBut the paper oversells itself. Calling this the first manifold-aware nonlinear ROM is wrong; principal geodesic analysis has been around for decades and is not cited. The 'semi-analytical solution of Navier-Stokes' is an ellipse fit to the tangent-space data, not a derivation. The claim that the Kolmogorov mode is a 'physical realization of the governing equations' rests on visual resemblance. The central numerical issue—conditioning of the autoencoder-inferred metric—is acknowledged but never quantified for the actual models that produced the results. Figure 6 shows a 'prototypical' autoencoder, not the production one. That matters because every downstream step (mean, log map, SVD, exp map) uses the metric and its inverse. If the metric is badly conditioned, the modes could be numerical artifacts. This is a genuine soft spot, not a manufactured one.\n\nThe lack of code and data makes independent checking harder, but that is a secondary concern. The method is plausible and the demos are suggestive. The paper deserves a serious referee, but it needs major revision: temper the claims, cite the PGA literature, quantify conditioning for the actual trained autoencoders, and ideally release code.\n\nFor a reading group, it is worth discussing as an example of applying geometric statistics to fluid problems, but not as a settled result. I would send it to review.","headline":"A plausible nonlinear extension of POD with nice demos, undercut by overclaims and unquantified metric conditioning; deserves review but needs major revision.","tokens_in":9317,"tokens_out":2101,"would_cite":true,"duration_ms":31159,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","58A05","35Q30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Proper latent decomposition generalizes POD to curved manifolds, recovering physical modes from autoencoder latent spaces.","keywords":["proper latent decomposition","proper orthogonal decomposition","nonlinear reduced-order modeling","autoencoders","Riemannian geometry","Eikonal equation","principal geodesics","Kolmogorov flow"],"falsifier":"Train a standard autoencoder on a dataset lying on a manifold with known geodesics, such as a sphere or torus embedded in three dimensions, compute PLD modes without metric regularization, and compare the decoded principal geodesic trajectories to the ground-truth geodesics; substantial deviation or failure of the log-map solver to converge would confirm the dependence on metric conditioning.","tokens_in":8305,"feed_emoji":"🌀","tokens_out":3964,"duration_ms":37034,"temperature":0.7,"pith_summary":"This paper introduces proper latent decomposition (PLD), a method that extends proper orthogonal decomposition (POD) from flat Euclidean spaces to the curved manifolds that autoencoders learn for high-dimensional data. The paper shows that by treating the latent space as a Riemannian manifold and using its inferred metric tensor, one can compute geodesic distance, find a meaningful mean, and extract energy-ordered principal geodesic modes. These modes, unlike linear POD modes, can correspond to physical structures in the underlying flow. The authors demonstrate this on a laminar bluff-body wake, where they obtain a semi-analytical description of the flow, and on turbulent Kolmogorov flow, where the leading PLD mode decodes to a physical flow field. The work matters because it offers a principled way to make autoencoder latent spaces interpretable for nonlinear reduced-order modeling.","feed_headline":"PLD finds curved modes that recover physical flow patterns","feed_subtitle":"A nonlinear generalization of proper orthogonal decomposition gives energy-ordered, interpretable modes for turbulent flows.","key_machinery":"The central object is the Riemannian manifold structure of the autoencoder latent space, determined by the metric tensor inferred from the decoder. A metric-constrained Eikonal solver yields the distance function on the manifold, whose gradient flow gives geodesics; those geodesics support the Fréchet mean and a robust log map via multiple shooting with a continuous adjoint. Once data are mapped to the tangent space at the mean, standard SVD produces energy-ordered modes, and the exponential map returns them to the manifold as principal geodesics.","core_discovery":"PLD computes, on the latent manifold inferred by an autoencoder, the Fréchet mean of the data, maps the encoded samples to the tangent space at that mean, performs singular value decomposition to obtain an orthonormal, energy-ordered basis, and maps the principal components back to the manifold via the exponential map to obtain principal geodesics. The paper claims these geodesics maximize variance of the data on the manifold, generalize POD to nonlinear settings, and provide interpretable, physical modes. For the laminar wake, the tangent-space data lie on a simple ellipse, yielding a semi-analytical solution of the Navier-Stokes flow with a single angular variable; for the turbulent Kolmogorov flow, the leading PLD mode reveals coherent vortical structures that differ from linear POD modes and resemble a physical realization of the governing equations.","pith_inferences":["A natural extension is to use the geodesic coordinates as a dynamical system, evolving the principal-mode amplitudes to forecast non-periodic turbulent flows, which the paper does not explicitly pursue.","Because PLD modes are energy-ordered by construction, they could serve as a curved-manifold analog of a Galerkin basis for reduced-order models, a connection the paper leaves implicit.","A direct synthetic test on a manifold with known geodesics, such as a sphere or torus embedded in three dimensions, would isolate the error of the Eikonal solver and log map from the error of the autoencoder itself."],"forward_implications":["For periodic or quasi-periodic flows, PLD can produce low-dimensional semi-analytical reduced-order models, as demonstrated by the single-variable ellipse model of the laminar wake.","PLD provides a nonlinear alternative to POD that can represent data lying on curved manifolds more faithfully than linear modes.","Regularizing the autoencoder's inferred metric improves numerical conditioning, making geodesic computation stable and potentially enabling higher-dimensional latent spaces.","Decoded principal geodesic modes give interpretable, physical coordinates, turning autoencoder latent spaces from opaque coordinates into analyzable dynamical features."],"supporting_citations":[{"why":"Introduces the concept of proper latent decomposition and the link between autoencoders and latent manifolds that PLD builds on.","marker":"Magri & Doan 2022"},{"why":"Provides the metric-constrained Eikonal solver used to compute distances and means on the manifold.","marker":"Kelshaw & Magri 2024"},{"why":"Defines intrinsic statistics on Riemannian manifolds, including the Fréchet mean that PLD uses as its central reference point.","marker":"Pennec 2006"},{"why":"Supplies the differential-geometric definitions of Riemannian metrics, geodesics, and exponential/logarithmic maps used throughout the method.","marker":"Lee 2018"},{"why":"Motivates the geometric regularization of the autoencoder loss that improves metric conditioning for stable geodesic computation.","marker":"Nazari et al. 2023"},{"why":"Provides the continuous-adjoint method used to compute gradients efficiently in the multiple-shooting refinement of geodesics.","marker":"Chen et al. 2018"},{"why":"Introduces POD, the linear decomposition that PLD generalizes to nonlinear manifolds.","marker":"Chatterjee 2000"},{"why":"The Adam optimizer used to train the convolutional autoencoders in the experimental sections.","marker":"Kingma & Ba 2017"}],"fun_headline_variants":["PLD maps curved modes onto turbulent flow physics","Curved modes from autoencoders beat POD on turbulence","Nonlinear modes expose hidden flow structures","Latent manifold SVD reveals physical flow patterns","PLD finds energy-ordered modes on manifolds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method rests on the autoencoder's inferred metric tensor being well-conditioned enough to support reliable distance and geodesic computations; if that condition fails, the principal geodesic modes are corrupted.","fun_headline_variants_meta":{"raw":{"variants":["PLD maps curved modes onto turbulent flow physics","Curved modes from autoencoders beat POD on turbulence","Nonlinear modes expose hidden flow structures","Latent manifold SVD reveals physical flow patterns","PLD finds energy-ordered modes on manifolds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1187,"prompt_tokens":916,"completion_tokens":271,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":199}},"tokens_in":532,"tokens_out":271,"duration_ms":3311,"temperature":1.0,"reasoning_tokens":199,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:59:43.291065+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a standard autoencoder on a dataset lying on a manifold with known geodesics, such as a sphere or torus embedded in three dimensions, compute PLD modes without metric regularization, and compare the decoded principal geodesic trajectories to the ground-truth geodesics; substantial deviation or failure of the log-map solver to converge would confirm the dependence on metric conditioning.","supporting_citations":[{"cited_title":"& Doan, A","cited_arxiv_id":null,"evidence_quote":"Introduces the concept of proper latent decomposition and the link between autoencoders and latent manifolds that PLD builds on."},{"cited_title":"2006 Intrinsic statistics on Riemannian manifolds: Basic tools for geometric measurements","cited_arxiv_id":null,"evidence_quote":"Defines intrinsic statistics on Riemannian manifolds, including the Fréchet mean that PLD uses as its central reference point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the differential-geometric definitions of Riemannian metrics, geodesics, and exponential/logarithmic maps used throughout the method."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the continuous-adjoint method used to compute gradients efficiently in the multiple-shooting refinement of geodesics."},{"cited_title":"2000 An introduction to the proper orthogonal decomposition","cited_arxiv_id":null,"evidence_quote":"Introduces POD, the linear decomposition that PLD generalizes to nonlinear manifolds."}],"review_version":1}