{"id":"c291cfff-62c8-4130-916a-304d6131a86d","arxiv_id":"2411.08557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"An unsupervised pipeline that learns a density-aware Riemannian metric by mapping data to a uniform sphere with a normalizing flow and measuring distances through the flow's Jacobian.","lead":"LAMINAR is a new machine learning pipeline that learns a locally adaptive distance metric from data using normalizing flows, without requiring the user to specify a metric in advance. It aims to help scientists in fields like astrophysics find structure in complex point data, and the authors demonstrate improved clustering on toy examples compared with the standard Euclidean metric.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The kNN graph that defines LAMINAR distances is built on an unproven locality-preservation assumption: the normalizing flow is trained only by likelihood, so Euclidean neighborhoods in the pseudo-cdf need not correspond to structure in the data.","rationale":"I agree with the reader's weakest-assumption identification: the locality-preservation statement is the load-bearing bridge between the learned flow and the kNN graph that defines all subsequent distances. My review adds that this assumption is not merely unproved; the training objective permits maximal distortion, so the burden is on showing that locality holds for the datasets used. The paper is positioned as a proof-of-concept and provides code, so the deficiency is addressable with a simple numerical check. No part of the argument requires rejecting the method outright, and the empirical demonstrations are suggestive but not decisive because they do not isolate the graph topology from the metric-tensor computation. Therefore the recommended verdict remains CONDITIONAL, and the reader's conditional verdict is unchanged.","tokens_in":5021,"tokens_out":7761,"duration_ms":74941,"concrete_test":"On a 2D synthetic dataset with known structure (e.g., two elongated clusters separated by a low-density gap), train LAMINAR. For each data point, compute its k nearest neighbors in the pseudo-cdf (the KDTree input used by the method) and its k nearest neighbors under Euclidean distance in the original data space, for k = 5, 10, 20. Report the mean Jaccard overlap of the two neighbor sets; if overlap is not high (e.g., below 0.5), the locality-preservation assumption fails and the graph feeding Eq. 4 is not structurally representative. This directly tests the unproven bridge between the learned transformation and the kNN graph.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction in Section 2 states that the pseudo-cdf transformation 'behaves like a cumulative distribution function ... in the sense that it is uniform and that locality is preserved.' The entire kNN graph, and therefore every shortest-path distance from Eq. 4, depends on this assertion. However, the continuous planar flow is trained by maximum likelihood only (Eq. 2); nothing in the loss controls the Jacobian's singular values or the map's bi-Lipschitz constant. A likelihood-trained flow can stretch low-density regions and compress high-density ones, so points that are far apart in the data can land close together in the uniform sphere and vice versa. The radial transform in Eq. 3 is monotone and locally well-behaved, but it cannot undo an arbitrarily anisotropic flow Jacobian. Consequently, Euclidean kNN search in the pseudo-cdf may connect points that are not local structural neighbors, and the 'structurally-representative' graph topology asserted in Section 2 is exactly what needs proof. The paper's own Fig. 2 shows discrepancies with ground-truth metrics (attributed to noise and edge effects), but no quantitative test isolates whether the graph topology itself is correct. If locality fails, the metric tensor in Eq. 5 may be locally correct while the global distances in Eq. 4 are built on the wrong pairings, and the downstream clustering improvement in Section 5 would be accidental rather than guaranteed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"LAMINAR is an unsupervised algorithm that learns a locally adaptive Riemannian metric on d-dimensional point data. It trains a continuous planar normalizing flow (Eq. 2) to map the data to a standard normal, then applies the radial inverse-CDF transform of Eq. (3) so that the images fill the unit ball uniformly, in what the authors call a pseudo-cdf sense. A Euclidean k-nearest-neighbour graph is built in this transformed space; each edge is weighted by the Mahalanobis distance of Eq. (4) using the Jacobian-derived tensor Sigma = (J^T J)^(-1) of Eq. (5); and pairwise distances are computed as Dijkstra shortest paths. Validation is threefold: comparison of learned metric tensors with ground-truth pullback metrics for four transformations of a uniform circle (Figs. 1-2), qualitative distance-field comparisons (Figs. 4-5), and a k-medoids clustering benchmark (Fig. 6) in which LAMINAR distances improve Jaccard scores over Euclidean distances in 'almost all' cases. The paper is explicitly a proof-of-concept for physical-science applications, and the code is released on GitHub.","tokens_in":5326,"tokens_out":24736,"duration_ms":231876,"significance":"If the results hold, LAMINAR offers a modular, unsupervised way to derive density-adaptive distances for point clouds, avoiding an explicit user-specified metric on the data space; because the continuous planar flow is a diffeomorphism and the local tensor is read off its Jacobian, the construction is conceptually clean and the local metric is machine-checkable (PyTorch autograd for J_flow, analytic J_to_sphere). The synthetic ground-truth tests in Fig. 2 are a genuine strength: the target tensors come from known transformations rather than fitted parameters, so the validation is not circular, and the authors explicitly acknowledge the proof-of-concept status, the restriction to observed points, and the boundary effects. The released code and the honest limitation statements are further strengths. The potential value, if the empirical support is completed, is a simple alternative to recent flow-based metric learning (notably Sorrenson et al. [12]) for physical-science clustering; the missing part is quantitative evidence that the graph topology produced by the pseudo-cdf kNN search is correct, which is what the major comments ask for.","major_comments":[{"comment":"The load-bearing assertion that the pseudo-cdf transformation 'preserves locality' is stated in §2 without proof, and the entire kNN graph — and therefore every global distance from Dijkstra's algorithm — depends on it. The continuous planar flow of Eq. (2) is trained by maximum likelihood only, and nothing in that objective controls the conditioning of J_flow or the curvature of the map, so in the finite-sample regime the Euclidean kNN neighborhoods in the unit ball need not coincide with the pullback-metric neighborhoods in the data space; the radial transform of Eq. (3) cannot undo an arbitrarily anisotropic flow Jacobian. Because the continuous planar flow with Lipschitz h is in fact a diffeomorphism, the locality assertion is plausibly correct in the asymptotic limit, but the paper should (i) state this property together with the needed regularity conditions, and (ii) validate the graph topology directly in the Fig. 2 setting, for instance by reporting the fraction of kNN edges in the pseudo-cdf that connect points that are local neighbors under the ground-truth transformation, as a function of k and sample size. Without such a test, the clustering improvements in §5 cannot be attributed to correct graph topology. Relatedly, the edge weight of Eq. (4) uses a one-point quadrature (the average of the endpoint tensors) for what should be a line integral of the metric; the error of that approximation is controlled by second derivatives of the flow, which are neither bounded nor discussed.","section":"§2, Eqs. (2)–(5)"},{"comment":"The explanatory argument for the discrepancy in Fig. 2 is mathematically incorrect. The paper states that one should not expect a perfect match because the flow could learn 'T(x) vs. T(Rx)' for a rotation R; however, if R is applied in the transformed space, the target distribution on the ball is unchanged and both the metric tensor Sigma of Eq. (5) and the Euclidean kNN graph are invariant under orthogonal post-composition, so T and R∘T yield identical LAMINAR distances. The real sources of deviation are the non-identifiability of the flow beyond rotations, training non-convergence, and boundary effects, and these should be discussed instead. In addition, the Wasserstein discrepancies between ground-truth and LAMINAR tensors are computed for Fig. 2 but never reported numerically, so the claim that 'LAMINAR is able to learn the correct metric' rests on visual inspection of colour wheels; please report the mean and per-transformation discrepancies.","section":"§3"},{"comment":"The manuscript omits the hyperparameters needed to reproduce the results. The value of k in the §2 kNN construction is never stated (k=25 in §4 refers only to the out-of-distribution contour estimator), and the flow architecture (number of layers, hidden width, activation, optimizer, learning rate, number of training steps, ODE solver tolerances) and the toy-data sample sizes are not given. Because a k that is too small disconnects the graph (leaving Dijkstra distances undefined or infinite) and a k that is too large can short-circuit across density modes, the choice of k is load-bearing for the method; the paper should report it, provide a sensitivity analysis, and state how it was chosen. Without these details the results in Figs. 2 and 6 are not reproducible.","section":"§2 and §5"},{"comment":"The downstream evaluation compares LAMINAR only with the Euclidean metric, yet §1 positions the work as improving on existing density-based and flow-based metric-learning implementations, in particular Sorrenson et al. [12], which is credited with practical convergence. To substantiate the claim that LAMINAR 'enhances structural representation,' the clustering benchmark should include at least one density-based baseline (for example Sajama and Orlitsky [9] or [12]); the current comparison only shows that the adaptive metric helps on the specific toy geometries of Fig. 6. Furthermore, the Jaccard values in Fig. 6 appear to come from single runs, with no error bars, no stated number of data points, and no table of all cluster scores, and the phrase 'improve cluster extraction for (almost) all cases' is not backed by a complete enumeration of the results.","section":"§5"}],"minor_comments":[{"comment":"The text calls F(·) 'the CDF of the multivariate Gaussian'; the inner term 1 − Γ(d/2, r²/2)/Γ(d/2) is the survival function of the chi (radius) distribution of ||z||, so F is the d-th root of the chi CDF; please correct the terminology.","section":"§2, Eq. (3)"},{"comment":"The four transformations in Fig. 1 are not named in the text or the figure captions; only the linear shear is identified in the body of §3, so the reader cannot tell which panel corresponds to which transformation.","section":"§3 and Fig. 1"},{"comment":"The paper does not state how the number of medoids was chosen for Fig. 6; if it was set to the number of ground-truth clusters, this is a form of supervision that should be acknowledged, and the 'unsupervised' framing should be qualified accordingly.","section":"§5"},{"comment":"Minor typos and infelicities: 'Euclideank-nearest-neighbour' is missing a space, 'Dijkstras's algorithm' has a misplaced apostrophe, and the sentence describing the ODE as a 'continuous transformation from any d-dimensional initial distribution to a d-dimensional standard-normal distribution' overstates what likelihood training guarantees (it holds only within the model family and on the training distribution).","section":"§2"},{"comment":"The statement that 'It is difficult to quantitatively assess LAMINAR's performance' is in tension with the quantitative Jaccard-based assessment in §5; clarifying that clustering quality serves as a proxy for structural meaningfulness would resolve the apparent tension.","section":"§3 and §5"},{"comment":"The claim that 'it is clear that LAMINAR is able to learn and emphasise the structure implicit to the data sets' overstates what qualitative colour maps can establish; a quantitative summary (for example, distance-ratio statistics or summary values for the fields in Fig. 5) would be needed to support the claim.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"For the editor: this is an honest, clearly written proof-of-concept with released code, but the validation depth is currently below what I would expect for a serious journal — the central quantitative claims rest on single-run scores and visual inspection, and the load-bearing locality assumption of §2 is neither proven nor empirically isolated. I recommend major revision rather than rejection because the construction is salvageable: the flow is a diffeomorphism, so an explicit asymptotic argument plus a graph-topology validation experiment would address the central concern, and the missing hyperparameters and baselines are local fixes. One editorial point: the contemporaneous related work by Sorrenson et al. [12] (July 2024) is very close in spirit (normalizing-flow-based learned distances with convergence guarantees), and the paper should either compare against it directly or sharpen the differentiation; the current text cites it in the introduction and outlook but never evaluates against it, which may raise novelty questions. Please also ensure the §3 rotation-based non-uniqueness argument is corrected, as it is simply wrong and a reviewer of the revised version will likely catch it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea in LAMINAR is worth taking seriously. Training a continuous normalizing flow to map data to a uniform sphere and then using the flow Jacobian to define a local Mahalanobis metric is a clean way to make distances data-dependent without requiring a user-specified metric. The specific pipeline—continuous flow, radial transform to the sphere, kNN graph in the pseudo-cdf, shortest paths with Jacobian-based edge weights—is not verbatim in the cited literature. The paper is clearly written, the code is public, and the authors honestly flag the non-uniqueness issue (rotations of the flow give the same density).\n\nThe weak point is exactly what the stress-test note identifies. Section 2 asserts that the pseudo-cdf transformation 'behaves like a cumulative distribution function ... in the sense that it is uniform and that locality is preserved,' but this is not proved. A likelihood-trained continuous flow can compress high-density regions and stretch low-density ones; nothing in the loss controls the Jacobian's singular values or a bi-Lipschitz constant. If Euclidean neighborhoods in the uniform sphere do not correspond to structural neighborhoods in the original data, the kNN edges—and therefore the shortest-path distances—are built on the wrong pairings, even if the local metric tensor is correctly derived. The paper provides no quantitative test that isolates graph topology, and the qualitative comparison in Figure 2 does not resolve this.\n\nThe validation is otherwise thin. The k-medoids clustering comparison in Section 5 shows Jaccard improvements over Euclidean distances, but it does not compare against the closest prior methods: Sorrenson et al.'s flow-based distance or Sharma and Johnston's entropy-based method. The toy examples are useful for intuition but not sufficient to establish that the improvement is robust.\n\nThese soft spots are significant but not disqualifying. The paper is an honest proof-of-concept, and the underlying construction is coherent. To make it publishable, the authors need to provide either a theoretical guarantee or strong empirical evidence that the pseudo-cdf preserves local neighborhoods, and they need a quantitative benchmark against existing density-based metric methods. If those are supplied, the contribution is real.\n\nMy recommendation: send it to peer review, but with the clear expectation of a major revision. The idea deserves referee time; the current validation does not yet earn acceptance.","headline":"LAMINAR's pipeline is a genuine combination of normalizing flows and Jacobian metrics, but the paper's center rests on an unproven locality-preservation assumption and the validation is too qualitative to carry that weight.","tokens_in":5821,"tokens_out":2921,"would_cite":true,"duration_ms":28079,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper presents LAMINAR, an unsupervised method that learns a locally adaptive distance metric from data by transporting the data to a uniform sphere with a continuous normalizing flow and reading off a Riemannian metric from the…","keywords":["locally adaptive metric","normalizing flows","inverse transform sampling","density-based distance","Riemannian metric","metric learning","k-medoids clustering"],"falsifier":"Construct a data set with two high-density clusters connected by a thin low-density bridge, where the Euclidean metric would cut across the gap. If LAMINAR's locality-preservation claim is correct, its shortest paths should stay on the bridge and the ratio of between-cluster to within-cluster distances should be larger than the Euclidean ratio; if instead the learned graph connects points across the gap or the distance ratios do not move in that direction, the central claim is falsified. A quantitative check would compare LAMINAR distances against a kernel-density-estimate geodesic on the same data.","tokens_in":4825,"feed_emoji":"📐","tokens_out":6788,"duration_ms":61703,"temperature":0.7,"pith_summary":"LAMINAR is an unsupervised method that learns a distance metric directly from the shape of a data set, without requiring the user to choose a metric in advance. It transports the data to a uniform distribution inside a unit sphere using a continuous normalizing flow, then uses the Jacobian of that transformation to define a locally varying Riemannian metric tensor at each point. Distances between nearby points are Mahalanobis distances with the averaged local tensor, and longer distances are shortest paths through a $k$-nearest-neighbour graph whose edges carry those local weights. The paper's claim is that these distances are structurally informative: points in the same dense mode end up closer and points in separate modes farther apart than under the Euclidean metric, which would make downstream distance-based analysis more faithful to the data's geometry.","feed_headline":"Locally adaptive metric learned from data improves clusters","feed_subtitle":"LAMINAR maps data to a uniform sphere, derives a local metric, and improves k-medoids clustering on toy sets.","key_machinery":"The machinery is a continuous planar flow (a neural ODE) that maps the input distribution to a standard normal, followed by a radial rescaling that converts it into a uniform distribution inside the $d$-dimensional unit sphere. This combined map acts as a pseudo-cdf: its output is uniform by construction, and the paper assumes it preserves locality. The Jacobian of the combined map, $J_{\\text{total}} = J_{\\text{to sphere}} J_{\\text{flow}}$, defines a Riemannian metric tensor $\\Sigma = (J_{\\text{total}}^T J_{\\text{total}})^{-1}$ at each point. Neighbourhood structure is obtained from a Euclidean $k$-nearest-neighbour graph on the transformed points, edge weights are Mahalanobis distances with the averaged tensor, and global distances are computed as Dijkstra shortest paths on this graph.","core_discovery":"The central claim is that a density-based locally adaptive metric can be learned without a pre-specified metric by using a normalizing flow as a pseudo-cdf. After training, the flow maps the data to a uniform ball, and because the transformation is invertible and locally smooth, Euclidean neighborhoods in the transformed space stand in for structurally meaningful neighborhoods in the original space. LAMINAR defines the metric tensor at each point as $\\Sigma = (J_{\\text{total}}^T J_{\\text{total}})^{-1}$, where $J_{\\text{total}}$ is the Jacobian of the full transformation, and sets the distance between neighbouring points to the Mahalanobis distance using the averaged tensors of the two points. The paper demonstrates on toy data that the learned tensors reproduce analytically known ground-truth metrics, that distances travel along dense modes, and that $k$-medoids clustering with LAMINAR distances improves Jaccard similarity over the Euclidean metric for almost all tested cases.","pith_inferences":["Beyond the paper: if the locality-preservation assumption survives a careful test, the sphere-transformation step could become a generic preprocessing layer for any metric-based algorithm, not just $k$-medoids.","Beyond the paper: because infinitely many flows produce the same uniform sphere, different flow architectures will produce different locally adaptive metrics; the choice of flow is thus likely to matter for downstream performance, and the paper's own suggestion of an optimal-transport-based flow points to a concrete comparative experiment.","Beyond the paper: extending the metric to out-of-distribution points by inverting the flow would let LAMINAR serve regression and classification tasks, since the current proof-of-concept can only measure distances between existing data points."],"forward_implications":["Distances computed with LAMINAR make movement along dense modes cheaper and movement across sparse regions more expensive than Euclidean distance does.","Any point-based downstream task that uses distances—clustering, nearest-neighbour classification, visualization—can adopt LAMINAR distances without changing the algorithm.","Because the metric is learned from the data, users do not need to hand-specify a metric or a separate density estimator, removing a common manual step in density-based distance methods.","In the paper's $k$-medoids experiments, Jaccard similarity between recovered and ground-truth clusters is higher with LAMINAR than with Euclidean distances for almost all data sets tested."],"supporting_citations":[{"why":"Supplies the continuous planar flow (neural ODE) architecture that transports the data to a standard normal distribution.","marker":"[3]"},{"why":"Provides the Mahalanobis-distance edge weighting and the locally adaptive metric idea that LAMINAR adapts.","marker":"[10]"},{"why":"Motivates the normalizing-flow-based density metric and supplies the comparison case showing earlier methods do not converge in practice.","marker":"[12]"},{"why":"Provides Dijkstra's shortest-path algorithm used to assemble local edge weights into global distances.","marker":"[5]"},{"why":"Establishes the density-based shortest-path distance paradigm that LAMINAR builds on.","marker":"[9]"}],"fun_headline_variants":["LAMINAR learns data-driven metric, sharpens clustering","Flow-based metric reveals structure, boosts k-medoids","Adaptive distances from normalizing flow improve k-medoids","LAMINAR: density-based metric for better clusters","Local metric from flow mapping enhances structural analysis"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned transformation to the uniform sphere preserves locality, so that Euclidean nearest-neighbour searches in the transformed space identify the same structurally meaningful neighbourhoods as in the original data; the paper asserts this without proof.","fun_headline_variants_meta":{"raw":{"variants":["LAMINAR learns data-driven metric, sharpens clustering","Flow-based metric reveals structure, boosts k-medoids","Adaptive distances from normalizing flow improve k-medoids","LAMINAR: density-based metric for better clusters","Local metric from flow mapping enhances structural analysis"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000158,"raw_usage":{"total_tokens":1191,"prompt_tokens":875,"completion_tokens":316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":238}},"tokens_in":491,"tokens_out":316,"duration_ms":3615,"temperature":1.0,"reasoning_tokens":238,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:31:13.713582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a data set with two high-density clusters connected by a thin low-density bridge, where the Euclidean metric would cut across the gap. If LAMINAR's locality-preservation claim is correct, its shortest paths should stay on the bridge and the ratio of between-cluster to within-cluster distances should be larger than the Euclidean ratio; if instead the learned graph connects points across the gap or the distance ratios do not move in that direction, the central claim is falsified. A quantitative check would compare LAMINAR distances against a kernel-density-estimate geodesic on the same data.","supporting_citations":[{"cited_title":"Johnston","cited_arxiv_id":null,"evidence_quote":"Provides the Mahalanobis-distance edge weighting and the locally adaptive metric idea that LAMINAR adapts."},{"cited_title":"Learning Distances from Data with Normalizing Flows and Score Matching","cited_arxiv_id":"2407.09297","evidence_quote":"Motivates the normalizing-flow-based density metric and supplies the comparison case showing earlier methods do not converge in practice."},{"cited_title":"Estimating and computing density based distance metrics","cited_arxiv_id":null,"evidence_quote":"Establishes the density-based shortest-path distance paradigm that LAMINAR builds on."}],"review_version":1}