{"id":"796ab00f-e71e-4ab7-a9a2-d43d3cdf9de4","arxiv_id":"2507.07291","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A VAE decoder's pullback metric, measured by the numerical rank of its Jacobian, estimates dataset intrinsic dimension and guides mixture-VAE atlas construction and pruning monitoring on a CT dataset.","lead":"This paper estimates the true dimension of data that sits on a low-dimensional shape by probing the geometry of a variational autoencoder. The estimate then guides local coordinate charts and network pruning checks in a medical imaging test case.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 3.1 cannot justify the encoder-based metric: its hypothesis DE=I is unsatisfiable for n>m, so J_EJ_E^T is not proved equivalent to the decoder pullback metric; all reported ID estimates use this unproved shortcut.","rationale":"The reader's weakest assumption isolates the same point I find most load-bearing: the only bridge from the decoder pullback metric to the encoder Gram matrix is Proposition 3.1, whose hypothesis DE=I is unsatisfiable in the paper's dimension regime. All experimental ID numbers are produced with the encoder matrix, so a rejection of that bridge removes the theoretical warrant for the headline result. I also considered the sinogram estimate (reported 11 versus the expected 12 from the Radon transform) as an alternative empirical soft spot, but it is less decisive because the paper does not explicitly claim the sinogram ID as ground truth and discretization could play a role; the vacuous proposition is a clean, verifiable flaw. I am not claiming the estimator is necessarily wrong in practice: the rank of J_EJ_E^T equals the rank of J_E, and a correct restricted-rank argument could justify the method. Crediting the public code and toy demonstrations, I therefore do not call for a stronger verdict than the reader's; the preprint as written needs either a corrected theorem or a re-run of Algorithm 1 with J_D^T J_D before the central claim can stand. Thus the verdict remains unchanged at REJECT.","tokens_in":18922,"tokens_out":11013,"duration_ms":144481,"concrete_test":"Use a synthetic benchmark with known d, for example a Swiss roll or a product manifold in R^64 with d=3. Train a beta-VAE with latent dimension m=10 and also m=25 over at least 20 random seeds. On a fixed test set, compute the sorted eigenvalues of J_D^T J_D and J_E J_E^T at each point. Record (i) the position of the largest eigenvalue gap and (ii) the numerical rank at a fixed tolerance such as 1e-3. If the two matrices ever yield different gap positions or ranks, the encoder shortcut is unreliable and the experimental estimates are not backed by the stated theory; if they agree across all seeds, the practical method is robust and the defect is confined to the proof of Proposition 3.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 proposes estimating ID from the numerical rank of the pullback metric g=J_D^T J_D and then states (Proposition 3.1) that, if E:R^n->R^m and D:R^m->R^n satisfy DE=I with n>m, then J_D^T J_D=(J_EJ_E^T)^dagger, concluding that either the decoder or the encoder can be used. The hypothesis is empty: by the chain rule, rank(J_D J_E) is at most min(rank J_E, rank J_D) <= m < n = rank I_n, so DE cannot equal the identity on R^n. Thus the proposition is vacuous, and it provides no support for the switch to J_EJ_E^T. This matters because Algorithm 1 is actually run with J_EJ_E^T (Section 5.2), so the toy and COULE estimates (circle ID=1, paraboloid ID=2, images ID=12, sinograms ID=11) all rest on an equivalence that is neither proved nor stated at the correct level of restriction. The rank of J_EJ_E^T equals the rank of J_E, which may coincide with the manifold dimension when the encoder is a local chart on the data manifold, but the paper does not prove such a statement, and even then the eigenvalue spectra of J_EJ_E^T and J_D^T J_D need not agree. The core idea may be repairable with a restricted-rank statement on the data manifold, but as written the central mathematical justification is missing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method to estimate the intrinsic dimension (ID) of a data manifold under the manifold hypothesis by computing the numerical rank of a singular Riemannian metric induced by a VAE: the pullback metric J_D^T J_D of the decoder or, by a claimed equivalence, the matrix J_E J_E^T of the encoder. The estimated ID is used to set the latent dimension of a mixture of invertible VAEs that builds local charts, and the framework is applied to CT imaging on the COULE dataset, where the image manifold ID is reported as 12 and the sinogram manifold as 11. The paper also studies pruning of the VAE weights and proposes the estimated ID as a proxy for network capacity.","tokens_in":19220,"tokens_out":6284,"duration_ms":63699,"significance":"If the method is valid, it offers a practical, geometry-based ID estimator for generative models, with the notable strengths that the toy-circle and paraboloid examples reproduce known IDs and the COULE image ID=12 matches the known construction of the dataset. The authors provide open code, compare against three standard ID baselines, and demonstrate a plausible application to atlas construction and inverse problems. However, the paper's central theoretical justification for using the encoder metric is flawed, and the pruning analysis contains an internal inconsistency; these issues must be resolved before the contribution can be assessed reliably.","major_comments":[{"comment":"The proposition is vacuous as stated. For differentiable maps E:R^n→R^m and D:R^m→R^n with n>m, the composition DE cannot equal the identity on R^n because rank(J_D J_E) ≤ min(rank J_D, rank J_E) ≤ m < n = rank(I_n). Hence the hypothesis DE=I is unsatisfiable, and the conclusion J_D^T J_D = (J_E J_E^T)^† does not follow. This is load-bearing because Algorithm 1 is implemented with J_E J_E^T (Section 5.2) and all reported ID estimates (circle, paraboloid, images, sinograms) rely on this step. The authors should replace the proposition with a restricted statement, e.g., assuming D∘E equals the identity on the data manifold M and that E has constant rank on M, in which case the numerical rank of J_E J_E^T equals dim(M), while noting that the eigenvalue spectra of the encoder- and decoder-based matrices need not coincide and that the gap criterion must be justified for J_E J_E^T separately.","section":"Section 3, Proposition 3.1"},{"comment":"The pruning claim is internally inconsistent. The text states that reconstruction quality degrades beyond p=0.996 and that 'this restriction in the network's capacity for pruning ratios greater than p=0.996 is reflected by a change in the estimated intrinsic dimension, illustrated in Figure 11.' However, Figure 11 shows that the estimated ID remains 12 at p=0.996 and only drops to 10 at p=0.9999 and 9 at p=0.99995. The two thresholds do not coincide, so the ID is not acting as an early warning signal for the capacity threshold identified by reconstruction loss. The authors should either rerun the experiments to identify the pruning ratio at which ID first drops, or revise the claim to describe ID as degrading at a later, more aggressive pruning stage.","section":"Section 5.4, Figures 10 and 11"},{"comment":"The eigenvalue-gap criterion is subjective and no statistical support is provided. Table 1 reports single point estimates for each method, with no error bars, no repeated training runs, and no specification of the gap threshold used to declare an eigenvalue 'null'. Since the entire ID estimate hinges on this threshold, the comparison against lPCA, MLE, and CorrID is not meaningful without reporting the threshold and the variability of the estimate across training seeds or dataset subsamples.","section":"Section 3 and Table 1"}],"minor_comments":[{"comment":"'We how this approach' should read 'We show this approach'.","section":"Abstract"},{"comment":"The manifold M is said to be 'embedded in Rd' although the ambient space is later denoted R^D with D>2d+1; the notation should be corrected to R^D.","section":"Hypothesis 2.1"},{"comment":"Step 8 instructs to 'search for the first null eigenvalue' without defining the gap threshold; a precise operationalization (e.g., relative gap > 10^k) is needed to make the algorithm reproducible.","section":"Algorithm 1"},{"comment":"The number of charts K=4 is chosen from the known topology of the image manifold (S^1×S^1×(0,1)^10) rather than from the ID estimate; the text should be clearer that the ID estimate informs the latent dimension, not the number of charts.","section":"Section 5.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds heavily on the authors' related prior work (Refs. [AHSS24, BCQ24]) for the mixture-VAE pipeline; the novelty lies in the ID estimation step, which is currently undermined by the false Proposition 3.1. I believe the approach is repairable, but the revise-and-resubmit should require a corrected proposition, a quantitative gap criterion, and a reconciled pruning analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on Causin and Marta, arXiv:2507.07291. The idea is genuinely useful: estimate the intrinsic dimension of a data manifold from the numerical rank of a VAE pullback metric, then use that estimate to build a mixture-of-VAE atlas for inverse problems. The empirical demonstration on the COULE toy (ID=12 matches construction) and the comparison against lPCA, MLE, CorrID are nice, and the code is public. The circle and paraboloid sanity checks work.\n\nThe problem is Proposition 3.1, which is the paper's advertised justification for computing the metric with the encoder. It states that if E: R^n -> R^m and D: R^m -> R^n satisfy DE = I with n>m, then J_D^T J_D = (J_E J_E^T)^dagger. But DE = I is impossible when n>m: rank(DE) <= m < n = rank(I_n). So the proposition is vacuous, and the proof's algebra is meaningless. All the image and sinogram ID estimates in Section 5.2 are computed from J_E J_E^T, so they rest on an equivalence that is neither proved nor stated at the right level of restriction. This is load-bearing, not cosmetic.\n\nThe rest of the paper has softer issues. The eigenvalue-gap criterion for 'null' eigenvalues is subjective; Table 1 has no error bars or repeated runs. The inverse problem part never actually solves a constrained reconstruction problem—it builds an atlas and shows a reconstruction through the latent bridge, but doesn't demonstrate the claimed regularizing benefit. And the pruning story is internally inconsistent: the text says degradation begins at p=0.996, but the ID plots in Figure 11 show ID=12 at p=0.9960 and only drop to 10 at p=0.9999.\n\nThat said, the core idea is plausible and likely repairable. Replacing the false proposition with a correct restricted-rank statement about the decoder Jacobian on the data manifold (or simply using the decoder pullback metric as in Algorithm 1) would fix the main gap. I'd send this to peer review—it deserves referee time—but it needs major revision: fix the math, add error bars, run a real regularized inverse problem, and reconcile the pruning claims. The paper is for people working on intrinsic dimension estimation or manifold-based priors in imaging; they'll get useful ideas even if they can't cite the theory as-is.","headline":"The core idea—estimating intrinsic dimension from the numerical rank of a VAE pullback metric—is promising and the COULE demo is clean, but the paper's central theorem justifying the encoder-based shortcut is false as stated, so the main estimates rest on an unproved equivalence.","tokens_in":19771,"tokens_out":2686,"would_cite":false,"duration_ms":27662,"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 claims that a dataset's intrinsic dimension is the numerical rank of the VAE decoder's pullback metric, and shows this recovers the known dimension 12 on a CT-image manifold.","keywords":["intrinsic dimension","manifold hypothesis","pullback metric","singular Riemannian metric","variational autoencoders","mixture of VAEs","inverse problems","network pruning"],"falsifier":"Train a VAE on a synthetic manifold of known dimension $d$ with a deliberately weakened encoder, then compare the numerical rank of $J_D^T J_D$ with that of $J_E J_E^T$: if the two ranks disagree or the encoder-based estimate no longer equals $d$, the exact-inverse assumption is doing the work and the method as implemented is not measuring the decoder metric.","tokens_in":18660,"feed_emoji":"📐","tokens_out":11790,"duration_ms":112204,"temperature":0.7,"pith_summary":"This paper tries to show that the intrinsic dimension of a data manifold can be estimated from the geometry a variational autoencoder learns, without any extra machinery. The claim is that the numerical rank of the pullback metric $g = J_D^T J_D$ through the decoder — or of $J_E J_E^T$ through the encoder — counts the manifold's degrees of freedom, with the remaining eigenvalues near zero. On a simplified COULE CT-image dataset this recovers the correct image dimension 12 and gives a sinogram dimension 11, and the estimate is used to fix the latent dimension when building a four-chart atlas of invertible VAEs. The paper also argues that this dimension estimate serves as a warning signal for model capacity: pruning away weights barely changes the estimated ID until a threshold, at which both the ID and reconstruction quality collapse. A sympathetic reader would care because it turns a trained generative model into a tool for the three tasks the manifold hypothesis demands: dimension, coordinates, and mapping back and forth.","feed_headline":"A VAE metric's rank reveals a dataset's true dimension","feed_subtitle":"Eigenvalue gaps in a VAE's metric recover the 12-D CT image manifold and guide constrained reconstruction.","key_machinery":"The load-bearing object is the pullback metric $g = J_D^T J_D$ of the VAE decoder, a possibly degenerate Riemannian metric whose numerical rank is taken as the intrinsic dimension. The method diagonalizes $g$ (or $J_E J_E^T$) at many data points, averages the eigenvalue curves, and cuts at the first order-of-magnitude gap; Proposition 3.1 supplies the encoder-side identity $J_D^T J_D = (J_E J_E^T)^\\dagger$ that justifies using the cheaper encoder Jacobian. The atlas construction then uses mixtures of invertible VAEs built from affine coupling blocks, so each chart is a bijection between a latent patch and a manifold patch, with the estimated ID fixing the latent dimension. The same measured dimension is used as a manifold constraint for the CT inverse problem and as a monitoring signal under pruning.","core_discovery":"The central discovery is that the singular Riemannian metric induced by a VAE carries the dimension of the learned data manifold in its rank. Concretely, for a decoder $D$ with latent dimension smaller than the ambient dimension, the pullback metric $g = J_D^T J_D$ is degenerate exactly along the latent directions the data do not use, and the intrinsic dimension is the number of eigenvalues that are not effectively null. Proposition 3.1 states that when the encoder and decoder are exact inverses ($DE = I$) this metric equals $(J_E J_E^T)^\\dagger$, so the encoder Gram matrix can be used instead, which is cheaper to compute; Algorithm 1 averages the eigenvalue spectra over a sample and looks for the first large gap. Applied to the simplified COULE dataset, the method estimates the image manifold to have dimension 12, matching the known parametrization, and the sinogram manifold dimension 11, while lPCA, MLE, and correlation-dimension baselines miss at least one of the two. The paper further reports that the estimated ID stays at 12 up to a pruning ratio of about 0.996 and then drops (to 10 and 9), matching the visible collapse in reconstruction quality.","pith_inferences":["Vary the number of projection angles in the CT setup and check whether the estimated sinogram dimension tracks the information content of the measurements; if it does, the ID gap between 11 and 12 quantifies the information loss of the discrete Radon transform.","The same rank-of-the-pullback criterion could be applied to other generative architectures by replacing the decoder with a deterministic generator; stochastic maps would require a definition of Jacobian that this paper does not address.","Because the encoder-side form relies on $DE = I$, a direct numerical comparison of the spectra of $J_D^T J_D$ and $(J_E J_E^T)^\\dagger$ on a trained VAE would show whether the practical estimates are stable when the exact-inverse assumption is violated.","The gap-based rank detection is currently visual; formalizing it as an eigenvalue-ratio threshold or a statistical test would turn the reported estimates into an automated decision rule."],"forward_implications":["If the estimate is correct, intrinsic dimension becomes a by-product of training any VAE: one forward pass of Jacobian computation per sample, averaged, is enough.","A correct ID fixes the latent dimension for atlas construction, replacing the heuristic choice of chart count and latent size in mixture-of-VAE manifold learning.","Constraining CT reconstructions to the learned manifold of dimension 12 (images) and 11 (sinograms) should regularize the ill-posed inversion without hand-designed priors.","Because the estimated ID collapses at a pruning ratio of roughly 0.996 while reconstruction loss is still flat, ID can serve as a capacity monitor that flags the onset of manifold degradation before visible artifacts appear."],"supporting_citations":[{"why":"Supplies the mixture-of-invertible-VAEs atlas construction that this paper adopts once the intrinsic dimension is known.","marker":"[AHSS24]"},{"why":"Provides the two-pathway CT architecture with Whitney embedders and latent-space Radon maps that the experiments build on.","marker":"[BCQ24]"},{"why":"Defines the beta-VAE objective whose training quality the ID estimate directly inherits.","marker":"[HMP+17]"},{"why":"Introduces the affine coupling layers used to make the atlas chart networks invertible.","marker":"[DSDB17]"},{"why":"Extends coupling-block invertible networks to inverse problems, the basis for the invertible VAEs in the mixture.","marker":"[AKW+19]"},{"why":"Provides the implementations of lPCA, MLE, and correlation dimension used as comparison baselines in the experiments.","marker":"[BMG+21]"},{"why":"Supplies the correlation-dimension baseline that the proposed method is compared against and outdoes on the image dataset.","marker":"[GP83]"},{"why":"Supplies the multiscale-PCA (lPCA) baseline that overestimates the image intrinsic dimension in the comparison.","marker":"[LMR17]"},{"why":"Defines the dataset whose known parametrization fixes the ground-truth intrinsic dimension 12 used to validate the estimate.","marker":"[COULE]"}],"fun_headline_variants":["Rank of a VAE's metric pinpoints dataset's intrinsic dimension","Find manifold dimension from VAE metric eigenvalue gaps","VAE metric's degenerate eigenvalues expose intrinsic dimension","Use VAE metric rank to estimate data's true dimension"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The estimates rely on the trained encoder and decoder being exact or near-exact inverses of each other, so that the encoder-side computation faithfully stands in for the decoder's geometry — even though a perfect inverse is mathematically impossible when the ambient dimension exceeds the latent dimension.","fun_headline_variants_meta":{"raw":{"variants":["Rank of a VAE's metric pinpoints dataset's intrinsic dimension","Find manifold dimension from VAE metric eigenvalue gaps","VAE metric's degenerate eigenvalues expose intrinsic dimension","Use VAE metric rank to estimate data's true dimension"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000215,"raw_usage":{"total_tokens":1459,"prompt_tokens":1004,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":390}},"tokens_in":620,"tokens_out":455,"duration_ms":5069,"temperature":1.0,"reasoning_tokens":390,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:45:01.827032+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a VAE on a synthetic manifold of known dimension $d$ with a deliberately weakened encoder, then compare the numerical rank of $J_D^T J_D$ with that of $J_E J_E^T$: if the two ranks disagree or the encoder-based estimate no longer equals $d$, the exact-inverse assumption is doing the work and the method as implemented is not measuring the decoder metric.","supporting_citations":[],"review_version":1}