acceptodds
Under review as a conference paper at ICLR 2027

Challenging the Manifold Hypothesis: A Geometric Diagnostic for Data and Representation Spaces

Abstract

The manifold hypothesis is widely invoked to motivate representation learning, dimensionality reduction, and geometric operations in high-dimensional data spaces. Yet it is often stated informally: data are said to lie near a low-dimensional “manifold” without specifying the geometric properties this object should satisfy or whether finite datasets exhibit them. We introduce the Geometric Manifoldness Index (GMI), a continuous diagnostic for two necessary signatures of smooth embedded manifold structure in point clouds: local intrinsic-dimension consistency and positive Federer reach. GMI is scalable to modern high-dimensional representations and separates dimension inconsistency from reach-related failures such as cusps and self-intersections. We validate GMI on synthetic manifolds, controlled singular examples, and image-transformation manifolds, then apply it to raw images, autoencoder latents, token embeddings, and single-cell transcriptomics data. Across these modalities, many data and representation spaces remain far from the smooth-manifold regime, even when low-dimensional structure is visible in projections or induced by compression. Learned embeddings become more manifold-like mainly when training objectives impose geometric regularity, and manifoldness shows no consistent positive correlation with reconstruction quality. Our results challenge common informal uses of the manifold hypothesis and motivate diagnosing manifoldness as an explicit geometric assumption rather than presuming it.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
manifoldness, manifold, manifold hypothesis, intrinsic dimension, reach, autoencoders

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_UFs3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manifold hypothesis\textit{manifold hypothesis} is a well-known conjecture that is often raised to explain why learning models in general are able to learn from finite samples even over high-dimensional data. Nonetheless, in many works this hypothesis is invoked without much theoretical rigor, and without an analysis of whether the space in which the data lie – or its (learned) representation – satisfies the geometric properties characteristic of manifolds. Hence, the authors propose the Geometric Manifoldness Index (GMI), an index to measure the "manifoldness" of point clouds based on two features of smooth embedded manifolds\textit{smooth embedded manifolds}: the local intrinsic dimension and (Federer) reach. Combining these two into a single continuous metric, the authors are able to diagnose when and to what degree such conditions may not be fully satisfied. Experiments on synthetic datasets show the usefulness of the metric, though on so-called real-world data the results are more nuanced. Across raw images, autoencoder latents, token embeddings, and single-cell transcriptomics, many spaces score well below synthetic manifolds, per-class subsets score higher than pooled datasets, latent spaces become more manifold-like mainly under distributional regularization, and GMI shows no consistent relation with reconstruction quality. That is an interesting and important step toward a more informed use of the manifold hypothesis.

Critical strengths

S1. The authors seek to partially address one of the field's long-standing hypotheses, i.e., the manifold hypothesis. Their proposal is based on the analysis of how close data or (latent) representations are to a subclass of smooth embedded manifolds, through a fine-grained metric that captures two properties of such objects. Thus, the method is well-motivated and the authors do a good job in explaining it effectively. Moreover, the decomposition into sconsts_{\mathrm{const}} and sreachs_{\mathrm{reach}} is useful in itself, as it separates dimension inconsistency from reach failures such as cusps and self-intersections.

S2. The metric aligns well with its expected behavior on data known to be coming from a manifold. Specifically, the synthetic dataset experiments show that it achieves higher values when samples come from a smooth embedded manifold, and generally lower values otherwise. For instance, it separates the Klein bottle embedded in R4\mathbb{R}^4 from its self-intersecting immersion in R3\mathbb{R}^3, flags the junction of the lollipop, and degrades continuously under additive noise. The breadth of the study across modalities and the ablations on the main design choices are also valuable.

S3. The authors are transparent about the scope of the metric. As they put it, GMI is "an operational diagnostic rather than a certificate of manifoldness" (or lack thereof). The same care applies to its components: they state that sconsts_{\mathrm{const}} measures the consistency, not the correctness, of the dimension estimates, they acknowledge that the reach score is a biased proxy, and they report the cases where sreachs_{\mathrm{reach}} is undefined or not invariant to reparametrization.

Critical weaknesses

W1. The paper's promise is bold: challenge informal uses of the manifold hypothesis. Nonetheless, the proposed method has a set of limitations, raised at the beginning, that make it more of an overpromise. (i) It is based on properties of a subclass of submanifolds embedded in Euclidean spaces, whose notion of smoothness is itself imprecise. Although "smooth" usually means C∞C^\infty, as in Definition 7, the paper also uses it for C1,1C^{1,1} manifolds (Definition 1) and for loosely "well-behaved" structures (e.g., the "nearly-smooth" controls of Section 5). (ii) The second property (positive reach) is neither sufficient nor necessary for smooth manifolds, and smoothness implies positive reach only under further assumptions such as compactness. Even jointly, the two signatures cannot detect boundaries, which the introduction lists among the departures from a smooth manifold.

W2. Though the reach proxy helps with computational tractability, it is biased, and the bias does not seem to go in the direction the paper relies on. Section 4.2 argues that both relaxations can only raise τ^\hat{\tau}. This holds relative to the plug-in minimum over the sampled pairs, but not relative to the true reach, since tangent-estimation errors and off-manifold noise typically enlarge the normal component in the denominator of Eq. (1). Indeed, Figure 18 places τ^\hat{\tau} below the analytic reach for almost every shape, by about an order of magnitude for the circle and the ellipse, and an underestimated reach favors the paper's main claim. Moreover, since equation 7 normalizes τ^\hat{\tau} by the median first-neighbor distance, sreachs_{\mathrm{reach}} also depends on the sampling density, and clean high-dimensional manifolds in Table 1 obtain values comparable to those of reach-zero shapes and not far from that of pure noise.

W3. The conclusions on real data are based on absolute GMI values and on their trend in NN, but neither is calibrated. Even for clean manifolds, GMI depends on the intrinsic dimension and on the sample size. For example, in Figure 12b, the tori T4T^4 and T8T^8 end at about the level of the self-intersecting shapes of Table 1, and GMI on T8T^8 decreases steadily with NN. Likewise, the less noisy sphere of Table 1 scores below the cone and both self-intersecting shapes, although Definition 1 allows data to lie in the vicinity of a manifold. Hence, neither the decline in Figure 5 nor per-class scores below the one-dimensional rotation orbit used as positive control are, by themselves, evidence against the manifold hypothesis.

W4. The related work and the paper's positioning are lacking. It would be interesting to have a more in-depth discussion of what has been proposed in the literature on the analysis of the conditions under which the manifold hypothesis might hold. The main text discusses only HADES (Lim et al., 2025) and the test of Robinson et al. (2025), while the literature seems to be much richer, e.g., the local-homology approaches of Bendich et al. (2012) and Stolz et al. (2020), the segmentation by local intrinsic dimension of Allegra et al. (2020), the multiscale local PCA of Little et al. (2017), and the reach estimator of Berenfeld et al. (2022). Since the building blocks of GMI are standard, this positioning is important.

References

  • Allegra, M., Facco, E., Denti, F., Laio, A., and Mira, A. Data segmentation based on the local intrinsic dimension. Scientific Reports, 10, 2020.
  • Bendich, P., Wang, B., and Mukherjee, S. Local homology transfer and stratification learning. In Proceedings of the ACM-SIAM Symposium on Discrete Algorithms (SODA), 2012.
  • Berenfeld, C., Harvey, J., Hoffmann, M., and Shankar, K. Estimating the reach of a manifold via its convexity defect function. Discrete & Computational Geometry, 67, 2022.
  • Little, A. V., Maggioni, M., and Rosasco, L. Multiscale geometric methods for data sets I: Multiscale SVD, noise and curvature. Applied and Computational Harmonic Analysis, 43(3), 2017.
  • Stolz, B. J., Tanner, J., Harrington, H. A., and Nanda, V. Geometric anomaly detection in data. Proceedings of the National Academy of Sciences, 117(33), 2020.
Questions for authors

Q1. How does GMI behave as NN grows for noisy samples of a smooth manifold whose intrinsic dimension, ambient dimension, sample sizes, and noise level match those of the real datasets (e.g., with dd in the range of the published estimates in Table 3)? If GMI also declines with NN in that setting, how can the decline in Figure 5 be distinguished from data that lie in the vicinity of a manifold, which Definition 1 allows? A reference distribution of GMI under such matched controls would make the conclusions of Section 5.1 considerably more convincing.

Q2. Could the authors specify the synthetic shapes of Table 1 (construction, intrinsic dimension, embedding in R128\mathbb{R}^{128}, and noise model)? Figure 11 refers to a full catalogue in Appendix D that I could not find (possibly Table 1, but there is no reference).

I will closely follow the authors' response and update my assessment in case my concerns are addressed.

3 comments
AuthorsOpenReview ↗

response 1

We thank the reviewer for their careful and constructive assessment, particularly for recognising the value of combining local dimensional consistency with reach-related geometric regularity, and for highlighting the importance of explicitly characterising the limitations of GMI.

We agree that the original presentation placed excessive emphasis on challenging the manifold hypothesis, leading to conclusions stronger than those supported by our experiments. Our intended contribution was to develop, systematically characterise, and demonstrate the usefulness of GMI as a finite-sample diagnostic of geometric regularity, rather than to establish whether the manifold hypothesis holds or fails. We have substantially revised the manuscript to better communicate this objective.

The revised paper now separates the controlled characterisation of GMI (Section 5) from its practical applications (Section 6), while making its mathematical assumptions and limitations explicit.

W1: Geometric assumptions and scope of the diagnostic.

We agree that the original manuscript did not sufficiently distinguish between different notions of manifold smoothness and geometric regularity.

Section 3 now explicitly defines the reference setting as C1,1C^{1,1} embedded submanifolds with positive Federer reach and intrinsic dimension smaller than the ambient dimension. This is motivated by the framework of Fefferman et al. (2013), although we do not impose their volume constraint or claim to test membership in their manifold class.

We also agree that dimensional consistency and reach-related regularity do not constitute a complete characterisation of manifold structure. In particular, GMI does not reliably identify all boundaries or singularities, and a high score cannot certify that the underlying support is a regular embedded manifold. We now define manifoldness operationally, referring specifically to the two geometric properties assessed by GMI.

AuthorsOpenReview ↗

response 2

W2: Bias and interpretation of the reach proxy.

We agree with the reviewer that our original discussion did not adequately distinguish between the effects of our computational approximations and the discrepancy between estimated and true Federer reach.

Restricting evaluated pairs and applying mean-then-min aggregation increases the estimate relative to the corresponding sampled pairwise minimum over the same estimated tangent spaces. However, this does not establish an upper bound on true Federer reach, since tangent-estimation errors, noise, and finite sampling may introduce bias in either direction.

We have clarified this distinction in Section 4.2. The resulting quantity is explicitly described as a computational proxy, rather than a statistically consistent estimator or bound on Federer reach.

We have also expanded the empirical analysis of this approximation in Section 5.2. Figure 6 compares the proxy against analytic Federer reach on geometries with known reach, illustrating its sensitivity to the underlying geometric scale while revealing discrepancies from the true values. Figure 7 further examines a sphere progressively flattened into an ellipsoid, showing that GMI decreases as the true reach approaches zero, even though the surface remains smooth for every positive thickness.

Additionally, Figure 8 compares alternative Federer-ratio aggregation strategies. The pairwise minimum is particularly sensitive to individual extreme estimates, whereas mean-then-min reduces this sensitivity but may obscure localised singularities. These experiments illustrate the trade-off underlying our design choice and motivate interpreting the reach component as an empirical geometric diagnostic rather than a precise reach estimator.

W3 and Q1: Calibration, sampling, and interpretation of real-world data.

We agree that absolute GMI values and their variation with sample size are insufficient to establish whether real-world observations are compatible with an underlying manifold. Instead, we have expanded Section 5 to systematically investigate how the diagnostic responds to controlled changes in geometry and sampling.

Section 5.3 examines increasing additive noise on synthetic manifolds, demonstrating that GMI provides a graded assessment of departures from regular geometry in the evaluated settings. Section 5.4 investigates sampling effects by comparing equally spaced observations with randomly selected subsets of the same image-rotation manifold. These experiments demonstrate that sampling nonuniformity alone can substantially affect GMI, even when the underlying geometry remains unchanged.

Regarding Q1, we agree that noisy-manifold reference distributions matched to real-world datasets in intrinsic dimension, ambient dimension, noise level, and sample size would be required for stronger conclusions about manifold compatibility. Our current experiments characterise these effects individually but do not provide fully matched reference distributions. Accordingly, the revised manuscript does not attempt such hypothesis-testing conclusions.

W4: Relation to existing geometric diagnostics.

We agree that GMI should be situated within the broader literature on geometric regularity, rather than being presented primarily as an alternative to manifold-hypothesis testing.

We have now expanded the Related Work section to incorporate the references suggested by the reviewer, including Allegra et al. (2020) on local intrinsic-dimension segmentation, Bendich et al. (2012) on local homology and stratification, Little et al. (2017) on multiscale geometric analysis, Stolz et al. (2020) on geometric anomaly detection, and Berenfeld et al. (2022) on reach estimation through convexity defects.

These works provide important complementary perspectives on identifying geometric irregularities and estimating properties of underlying data structures. Our objective differs in providing a single continuous diagnostic that combines local dimensional consistency with reach-related regularity, rather than performing statistical manifold-membership testing or estimating a particular geometric quantity in isolation.

We do not claim that GMI replaces these methods, but rather that it offers an interpretable and computationally tractable complementary tool for investigating geometric properties of finite point clouds.

Q2: Synthetic constructions and reproducibility.

We agree that the synthetic constructions should be sufficiently specified to permit reproduction and interpretation of the reported results.

Section 5.1 now clarifies the general sampling procedure: synthetic point clouds are generated by independently sampling the parameter spaces of known geometric shapes and mapping these samples into their ambient Euclidean embeddings.

We also provide representative visualisations (Figure 3), component-level results (Table 1), and additional investigations of local dimension and reach-related estimates (Figures 4–5).

AuthorsOpenReview ↗

response 3

Practical applications of GMI.

Beyond addressing the mathematical limitations, an important change to the manuscript is the explicit investigation of what GMI can be used for in practice.

Section 6 now presents four applications:

  • Diagnosing dataset heterogeneity: Comparing pooled datasets with semantic subsets reveals differences in geometric regularity and identifies whether these arise primarily from dimensional inconsistency or reach-related effects. For example, pooled MNIST exhibits greater variability in local dimension estimates than individual digit classes, illustrating how GMI can help diagnose geometric heterogeneity.

  • Analysing learned representations: Comparing autoencoder families reveals how different training objectives are associated with distinct geometric structures in latent space, including differences between class-specific and pooled representations.

  • Evaluating preprocessing choices: In single-cell transcriptomics, GMI reveals substantial sensitivity to PCA dimensionality in Puram, compared with greater stability in Baron2. We complement this analysis with quantitative neighbourhood-preservation measurements, illustrating how GMI can help assess the sensitivity of geometric analyses to representation choices.

  • Investigating representation quality: We examine whether greater geometric regularity corresponds to better reconstruction performance across autoencoder families. The relationship varies across training objectives, demonstrating that higher manifoldness does not necessarily accompany better reconstruction. GMI thus provides complementary information about representation geometry rather than a universal proxy for downstream performance.

We hope these revisions address the reviewer's concerns about the soundness and interpretation of GMI, while better communicating the diagnostic contribution originally intended.

Reviewer_RcSqOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors propose to study the widely present Manifold Hypothesis, which is a handwaving loose wording saying that natural data lies on some low-dimensional structure inside a high dimensional space. In the community, the word manifold is used loosely and figuratively, but the authors decide to challenge the wording using a more rigorist interpretation of the word. Not every low-dimensional structure is a proper manifold, as they do not allow various phenomena such as self-intersections or cusps. The authors geometrically average two existing scores and argue that this score is correlated with the "manifoldness", another loose handwaving description somewhat quantifying how close to a manifold the data structure is. The authors try to evaluate some synthetic and basic datasets with their score to decide on the manifoldness of data or embeddings.

Critical strengths
  • The authors investigate an intellectual question about data. In the world of data, it is always a good idea to try and understand it better.

  • The "method", if I can call it that, to compute the GMI (Geometric Manifold Index) score is simple: just geometrically average two existing scores: local PCA dimension estimates and reach estimates.

Critical weaknesses
  • The paper is void of new knowledge and essentially starts on a false premise. The Manifold Hypothesis is well known to be a loose handwavy hypothesis and it is not believed in the community that it is a rigorous manifold that cannot have say self-intersections. As such, attacking the hypothesis is not discovering something new, it is just stating the obvious that is widely accepted by the community. Debunking the manifold hypothesis, in the sense of unfairly interpreting the term manifold way too rigorously for what it is in this context, is not a contribution in any form.

  • There is no method in this work, solely the construction of a score, which just consist in geometrically averaging two existing scores. As such, this can hardly be considered a novelty. In addition, the authors did not investigate alternative constructions to create a score. This is important, as there are many ways one can combine existing scores to create a new score that would be low for most likely non-manifold structure and large for probably manifold-like structures, for example by applying non-linear operations to each score that would further emphasise/squash scales of interest and better reveal the global manifoldness property.

  • Thus, if the great lesson to take from the score is not a new lesson (data manifolds are not necessarily rigorous manifolds) and a good construction of the score has not been investigated, this begs the following question: what is the score good for? What is the practicality of this specific naive construction and how can it be beneficial downstream? The authors did not investigate downstream applications in any meaningful sense. As such, as far as I am concerned after reading this submission, it is not clear if the proposed score, which is the core "contribution" of the work, is of any use at all. Based on this observation, and the previous ones, I don't really see the point for the submission to exist at least a top conference ML venue.

  • Experiments do not compensate for the earlier problems and are not of high interest.

    • The authors claim to show image embeddings are non-perfectly manifold and instead union of class manifolds, which is expected especially in classification/clustering data which is what the author tested (MNIST, FMNIST, CIFAR10, SVHN) because images of different digits (MNIST) might obviously not lie on same dimensional sub-manifolds.
    • The authors claim that PCA artificially creates manifoldness on scRNA-seq data, due to much larger scores on low-dimensional compressions than on the original sparse high-dimensional data. Yet, this claim is very doubtful, and what these results actually show is that the proposed score is very brittle to the embedding space, thus poorly accounting for the intrinsic geometry of the data.
    • The authors claim stability by looking at 3 images of t-SNE/UMAP embeddings, each in a different dimension. However, the visual stability is far-fetched and based on a strand of a couple outlier points in UMAP. Note that most of the data is large blobs of mixed points and nothing meaningful can be viewed there. As the original high-dimensional data is unknown and incomprehensible, arriving to such a claim based on the visualisation on a single data where this sort of line ordering occurs is abusive. If the authors want to claim something for curvy-like UMAP embeddings, they should first reveal it in constructed synthetic data where the structure (manifold or not) is known and would give a controlled result. Additionally, using t-SNE and UMAP is not a good choice. In fact, t-SNE and UMAP are among the worst embedders to visualise manifolds, even though they are very popular due to their speed (thanks to sparsity). Instead, Isomap and MDS are much more appropriate manifold learning techniques that should be used instead, and they are non-random methods unlike the ones from the authors. t-SNE and UMAP are artificial clustering methods, not manifold perserving embeddings. In fact, I invite the authors to view a talk by McInnes, the first author of UMAP, presenting his work, and how UMAP terribly fails at accurately representing in any meaningful way a simple sinusoidal curve https://www.youtube.com/watch?v=nq6iPZVUxZU&t=958s. By the way, the authors introduce UMAP at the top of the related works l111 but it is not cited then, but 3 other works that have nothing to do with it are. The authors should fix their citations.
    • The authors claims that similar to PCA autoencoders create manifoldness. In fact, they write l372 "manifold-like geometry is imposed by construction rather than discovered in the data". Again, this claim is somewhat abusive and needs further experiments to be able to properly accurately claim this. Crucially, there is no baseline experiment. For example, should we embed non-structured and non-manifold data (perhaps white noise, although that could be argued as manifold as living near the unit sphere in high dimensions) to embedders, would that significantly increase the manifoldness of the non-manifold or non-structured data? If so, by how much and would that increase be of the same order of magnitude that what is observed on MNIST data? Beyond this specific case, the authors make large general claims yet provide very little evidence to actually support them in a legitimate scientific way, falling short of the rigorous scientific standards of top conference venues.
    • The authors claim that manifolds is a bad thing, whether it be in data or latent space, is a bad thing as it correlates with worse reconstruction quality for autoencoders. However, it is far-fetched to claim this based on so little evidence (just MNIST data) with other concepts varying beyond the manifoldness, in the case the size of the latent dimension. I encourage the authors to perform apples-to-apples evaluations: modify the manifoldness only (fully controllable synthetic in the way authors see fit, controllable semi-synthetic via generative models, barely controllable via given datasets and subsamplings of them) without modifying important other factors like the latent dimension that could dominate the change of results. Additionally, the authors should investigate other tasks beyond auto-encoding reconstruction. And even more importantly, the authors should consider an actual useful application for their score, typically on some downstream task, whether it be as a meaningful and useful evaluation score (inference-time only), or to design methods (train/preprocessing-time).
  • Reading the submission, especially the conclusion, it seems that the authors wish to argue that since data manifolds are not exactly manifolds due to technicalities on manifolds, then new geometric tools must be introduced. While this can sound good, on paper, it is highly subjective and is not an objective conclusion to be drawn from the presented work: it should thus be removed (this is not a position paper submission).

    • The first point is that regardless of whether the data is actually perfectly manifold or not, treating it with manifold tools works well or well enough in most cases. This is common in many geometrical aspects of modern machine learning: while the theory on paper does not rigorously hold, applying it in practice works already. As a matter of fact, the theoretically justified fixes often struggle to even compare with the naive theory-violating baselines.
    • The second point is that non-geometric methods tend to work, and work quite well, as scale increases. This is an example of the Richard Sutton's Bitter Lesson (2019). Thus even if motivated by theory, and even if it works, fancy geometric methods will struggle to compare to naive non-geometric ideas at scale.
    • Combinging both frustrating facts, which are outside the control of the authors, with the score created by the authors and the final positioning of the submission as using this score to switch to more fancy non-manifold techniques if detected by the score transforms this idea into very wishful thinking. Unfortunately, it is very hard to believe in 2026. To be more believable, the authors must provide in this paper concrete application examples where their score is useful and would solve something or help in a problem that no other existing criteria or scores etc would have allowed to solve or help in such a way. Should the authors provide these concrete convincing examples, then we hope that the score can be useful elsewhere in other applications (whether that holds or not is something else).

Beyond the general points raised above, many other major concerns exist.

  • Both scores that are then combined to estimate the manifoldness are unreliable in their raw estimate of the dimensionality and the reach respectively. For instance, the dimensionality is over estimated l208. As such, these scores can be seen as mere suggestions ("suggests" l209 and "proxy" l220). So their value is not amazing. It also motivates exploring modifications of these scores, via non-linear transforms, to better "suggest" manifoldness or not. I highly encourage authors to investigate this direction to strenghten the submission.

  • For local dimensionality estimation, the authors perform Furthest Point Sampling (FPS) followed by local PCA. However, it is well-known that FPS followed by local patch analysis is not good enough. Indeed, beyond 1-dimensional manifolds (ie curves), as soon as manifolds are 2-dimensional or more, there is no guarantee that the local manifold around FPS samples actually looks like a clean local patch. For instance, Voronoi cells around FPS samples can easily have holes in cases of high curvature. For instance, consider FPS samples on the Stanford bunny. Then each tip of the ears could have an FPS sample and the face could have one too. When computing the voronoi cell of the point on the face, then it will topologically have 2 holes, corresponding to the Voronoi cells of the ears. See [1] Figure 2. This is problem for the authors as if the neighbourhood is non-uniform due to holes, then this will bias the PCA estimate and can create issues. I highly recommend cleaning up this construction, for instance by iterating at finer scales the FPS when needed to avoid this issue, see [1].

  • The local PCA for local dimensionality estimation is not new and is actually quite common. In additiona to Fukunaga & Olsen, I believe the authors should cite other works, for instance TICE [2] in manifold learning where it is used to find boundaries. In general, the authors should cite more works that find specific locations of rigorous manifold assumptions violation, and not just global score estimates.

  • The presentation of the local dimensionality consistency score needs to be reworked. Almost none of the mathematical notations are defined in the text kpatchk_{patch}, d\boldsymbol{d}, dLPCA,id_{LPCA,i}, ncentresn_{centres}, kpatch,mink_{patch,min}, kpatch,maxk_{patch,max}, DactiveD_{active}... This makes the reading of the section very unclear (if not non-understandable) and poorly reproducible.

  • Since authors are essentially playing on words, the definition of "manifold", it would be most welcome to actually define the term in the main paper, and not solely in the appendix. Indeed, there are a lot of different concepts that are coined as manifolds. The author also seem to focus on the more usual differential manifold, so switching to that terminology in the main paper, or discussing about this very clearly, would be most welcome.

  • The authors should tone down the criticism of the community for the Manifold Hypothesis. For instance, it is not "quoted to justify techniques for manifold learning", it is often loosely referred to justify embedding data in lower dimensional spaces. The techniques themselves rarely have anything to do with manifolds, even a method like UMAP, which has manifold in the name, has very little to do with manifolds (just each point is connected to its nearest neighbour with probability 1, and less so with further neighbours). See for example [3] for a modern overview of manifold learning techniques and an other example of how there are always gaps between theory and practice in the manifold learning field (which can be exploited but should be phrased fairly). I am all for exploiting bugs and inconsistencies in the community's assumptions, but the community should not be ridiculed by making it look like it does not know already what is claimed when this is not the case, and the authors should reveal how beneficial the exploit is. It has not been "forgotten by the literature" l128 as claimed by the authors, everyone in the field knows it. There is just not much evidence showing that purely non-manifold ideas are highly beneficial compared to manifold ideas (or dare I even say non-geometric ideas). This is where I encourage authors to put emphasis if they want to argue a strong case about this technicality on the non-manifoldness.

  • The authors motivate the work by mentioning many methods can miss shapes with singularities (from synthetic shapes) l136. While breaking the mathematical rigorous definition of a manifold, singularities can in practice not break them. If they have 0 measure for instance they are virtually non-existent in practice for real data. Even if they do exist, manifold-based ideas can still be completely valid. For instance physicists can still derive the law of physics on a cone even though the tip is non smooth: they will probably argue that if one could zoom enough then what looks from far away like a singularity is actually round and smooth so everything is fine. The same could go for data, especially for highly sparsely-sampled data in very high dimensions. Thus in general, I highly recommend toning down the language and criticism.

  • To better validate the score, I strongly encourage the work on well-known and well-understood data. For instance, a classical manifold (or not?) is the set of 3x3 pixels [4], which is known to be topologically equivalent to three intersecting circles, and itself topologically equivalent to a Klein bottle. This would be a more convincing experiment that would reinforce the synthetic Klein bottle experimetn l252.

  • Limitations are not sufficiently discussed and emphasised. For instance:

    • The score solely considers uniform dimensionality and reach, yet other factors are of importance.
    • As mentioned by the authors, the proposed score breaks down when the considered structure is the ambient space, the most trivial manifold. This is a bit worrying that the concept breaks down in simple cases, but could be fine if the authors insisted more it should hold only in low-dimensional manifolds. In general, I advise to tone down the language of all claims.
    • Again as mentioned by authors, the score is not parametrisation invariant, and I would add not representation invariant (huge discrepancies based on the input space or the latent representation space).

[1] Maggioli & Melzi, UV Parametrization via Topological Disk Segmentation of Surfaces, STAG, 2025

[2] Rosman et al., Nonlinear Dimensionality Reduction by Topologically Constrained Isometric Embedding, IJCV, 2010

[3] Dagès et al., Harnessing Data Asymmetry: Manifold Learning in the Finsler World, arxiv, 2026

[4] Carlsson, Topology and data, Bulletin of the AMS, 2009

Questions for authors

See comments above.

I would strongly encourage finding a way to motivate the score, from design choices to practical usefulness. This has to be motivated with convincing examples and experiments. In general, I would really enjoy if the authors could enlighten me on what the purpose of their submission is.

This requires a lot of work though. While I am always open to change my score in front of a good convincing author response, I have to warn the authors that I currently see so many major issues that I believe it will be challenging to convince me to change my score to acceptance.

4 comments
AuthorsOpenReview ↗

response 1

We thank the reviewer for their detailed assessment, particularly for raising the central question of what practical purpose GMI serves beyond identifying departures from idealised manifold geometry.

We agree that observing that real-world data need not satisfy a strict mathematical definition of a manifold is not, by itself, a novel contribution. This was not the intended focus of our work, although we recognise that the original presentation placed disproportionate emphasis on this interpretation and did not sufficiently communicate the diagnostic's purpose and practical value.

Our intended contribution is the development and systematic characterisation of GMI as an interpretable diagnostic for assessing geometric regularity in finite point clouds. We have substantially revised the manuscript to make this contribution explicit, separating the controlled characterisation of GMI (Section 5) from its practical applications to data and learned representations (Section 6). The manifold hypothesis now serves as a motivation and potential application rather than the central claim being challenged.

1. Practical usefulness of GMI.

We agree that the original submission did not sufficiently establish what GMI could be useful for. A central aspect of the revised manuscript is therefore demonstrating its applicability beyond controlled synthetic geometries. Section 6 presents four practical applications:

  • Diagnosing dataset heterogeneity. Comparing semantic subsets with pooled datasets reveals differences in geometric regularity and allows us to identify which component contributes to these differences. For example, pooled MNIST exhibits greater variability in local dimension estimates than individual digit classes, while the reach-related component changes comparatively little. This illustrates how GMI can identify sources of geometric heterogeneity that a single global intrinsic-dimension estimate would not capture, informing whether geometric analyses should consider the full dataset or more homogeneous subsets.

  • Analysing learned representations. We examine how different autoencoder training objectives are associated with changes in latent-space geometry. Standard AEs and TopoAEs exhibit substantially higher GMI within individual digit classes than in pooled representations, whereas VAEs maintain relatively high GMI across pooled classes, consistent with their distributional regularisation. GMI thus provides a complementary perspective on how training objectives shape representation geometry, beyond reconstruction error or latent dimensionality alone.

  • Evaluating preprocessing choices. We investigate the sensitivity of measured geometry to dimensionality reduction in single-cell transcriptomic data. Puram exhibits substantial variation in GMI across PCA dimensions, whereas Baron2 remains comparatively stable. We further quantify neighbourhood preservation through the Jaccard overlap of nearest-neighbour sets between PCA representations and their 2D UMAP projections. As PCA dimensionality increases from 32 to 1024, this overlap decreases from approximately 0.40 to 0.15 in Puram, compared with 0.46 to 0.41 in Baron2. These results illustrate how GMI can be used alongside neighbourhood-preservation measures to assess the sensitivity of geometric analyses to preprocessing choices, rather than relying solely on qualitative visualisations.

  • Investigating representation quality. We examine whether higher geometric regularity corresponds to better reconstruction performance across five autoencoder families. The relationship varies substantially between training objectives: higher GMI is associated with greater reconstruction error in AE, TopoAE, and WAE-MMD, while the opposite association is observed for VAE. These findings illustrate that greater geometric manifoldness does not necessarily accompany better reconstruction quality. Rather than treating manifoldness as a proxy for downstream performance, GMI provides complementary information about the geometry of learned representations.

These applications demonstrate the practical diagnostic role we intended for GMI: identifying geometric heterogeneity, comparing representation structures, evaluating sensitivity to preprocessing, and investigating relationships between geometry and learning outcomes.

2. Motivation and construction of the score.

We agree that local PCA and Federer reach are established geometric concepts, and we do not claim novelty in either ingredient individually. Our contribution lies in their operational combination into a computational diagnostic and the systematic characterisation of its behaviour, limitations, and applications.

AuthorsOpenReview ↗

response 2

The motivation for combining the two components follows from the joint geometric requirements of regular embedded manifolds, including the class G(d,V,τ)\mathcal{G}(d,V,\tau) considered by Fefferman et al. (2013), where fixed intrinsic dimension and a positive lower bound on reach are required simultaneously.

The geometric mean provides a symmetric soft conjunction of the corresponding diagnostics: a low value in either component reduces the overall score, without prioritising either property. It penalises imbalance more strongly than the arithmetic mean, while remaining less sensitive to the smallest component than the minimum. This provides a practical compromise between compensating for geometric deficiencies and sensitivity to imperfect finite-sample estimates. We acknowledge that this aggregation is a design choice rather than a uniquely determined measure of manifoldness.

We have also strengthened the methodological analysis. Section 5.2 investigates alternative aggregation strategies for the estimated Federer ratios. Appendix C compares intrinsic-dimension estimators, and Section 5.4 evaluates different patch-centre selection strategies. These experiments characterise the consequences of several methodological choices.

3. Systematic characterisation of GMI.

We have substantially expanded the controlled experiments and reorganised them into Section 5, Characterising the Geometric Manifoldness Index.

Beyond comparing regular and singular geometries, we examine the relationship between our reach proxy and analytic Federer reach (Figure 6), sensitivity to decreasing reach on smooth ellipsoids (Figure 7), alternative reach-aggregation strategies (Figure 8), additive noise (Section 5.3), sampling and patch-centre selection (Section 5.4), and computational scalability (Section 5.5).

Importantly, these experiments also reveal limitations of the diagnostic. For example, mean-then-min aggregation reduces sensitivity to individual extreme estimates but can obscure localised singularities, while known regular manifolds of higher intrinsic dimension need not obtain GMI scores approaching one under finite sampling.

The revised manuscript therefore does not present GMI as a universal manifold detector or a statistically calibrated test of manifold compatibility. Instead, it characterises the geometric properties GMI captures, the conditions under which it behaves as expected, and the situations in which its interpretation requires caution.

4. Interpretation of real-world data and learned representations.

We agree that several conclusions in the original submission were overstated, particularly concerning dimensionality reduction, autoencoder representations, and reconstruction performance.

We no longer interpret differences between pooled datasets and individual semantic classes as evidence establishing a union-of-manifolds model. Instead, these experiments illustrate how GMI can reveal differences in geometric heterogeneity across data partitions.

We agree that GMI depends on the ambient representation and is not invariant under arbitrary changes of coordinates. This is now explicitly acknowledged. GMI measures geometric regularity relative to the Euclidean metric of the supplied representation, and we do not claim that dimensionality reduction necessarily preserves, reveals, or creates the intrinsic geometry of the original data.

We have removed the previous qualitative arguments based on the apparent stability of UMAP and t-SNE visualisations. The revised single-cell analysis instead includes quantitative nearest-neighbour overlap measurements, which are interpreted as complementary evidence about representation sensitivity rather than proof of manifold preservation.

We no longer claim that autoencoders necessarily impose manifold structure or that greater manifoldness is detrimental to reconstruction quality. Our experiments identify associations between GMI and reconstruction error across model configurations, with latent dimension and training stage also varying. We recognise that a controlled analysis isolating manifoldness from these factors would be required to establish a causal relationship.

Finally, we have removed the speculative conclusion suggesting that departures from idealised manifold geometry necessitate replacing existing manifold-based methods with alternative geometric frameworks. We agree that violations of idealised assumptions do not necessarily imply that methods relying on those assumptions are ineffective.

5. Methodological clarity and limitations.

We have clarified the mathematical setting in Section 3, explicitly considering C1,1C^{1,1} embedded submanifolds with positive Federer reach. We also distinguish the operational meaning of manifoldness used by GMI from a complete mathematical characterisation of manifold structure.

AuthorsOpenReview ↗

response 3

Regarding FPS, our implementation uses farthest-point sampling to select patch centres, followed by overlapping k-nearest-neighbour patches constructed from the original point cloud, rather than disjoint Voronoi cells. The specific issue concerning holes in FPS Voronoi regions does not directly describe our neighbourhood construction. Nevertheless, we agree that curvature, sampling density, and neighbourhood scale can affect local PCA estimates, and the revised paper examines the influence of alternative patch-centre selection strategies. We also investigated multi-scale FPS in section 5.4 (figure 12) following your suggestion.

We have further clarified the distinction between dimensional consistency and intrinsic-dimension estimation accuracy. A high consistency score does not imply that the underlying dimension has been correctly estimated. Similarly, we explicitly describe the reach-related component as a computational proxy rather than a statistically consistent estimator or bound on Federer reach.

In summary, the original presentation did not sufficiently communicate the intended scientific contribution or demonstrate its practical relevance. Our goal is not to establish that real-world data fail to satisfy an idealised manifold model, but to provide a computationally tractable and interpretable diagnostic for examining selected geometric properties of finite point clouds.

The revised manuscript now systematically characterises GMI under controlled geometric conditions, explicitly documents its limitations, and demonstrates its applications to data heterogeneity, learned representations, preprocessing sensitivity, and reconstruction quality.

We hope these revisions clarify both the motivation for GMI and its practical utility, addressing the reviewer's central concern regarding the purpose and significance of the proposed diagnostic.

I thank the reviewers for their comments.

Unfortunately, the authors did not mark the changes to their manuscript in colour, making it impossible to quickly view what changed and how. Based on the response, a lot has changed, yet the response came so late, not allowing enough time to properly review the comments and revisions in order to favour a proper discussion.

Quickly reading through the response, the authors seem to want to address my concerns. This is encouraging and I thank the authors for doing so. But even if they did manage to resolve all my concerns, the revised submission would deserve a proper new round of reviewing from scratch. As it stands, I still believe the work has very limited interest for the community, providing a trivial diagnostic, which can barely be called a contribution consisting in averaging two well-known numbers, with little downstream usefulness.

I will thus keep my score, however, I would be happy to add half a point to it to reward the efforts towards completely revising the submission.

Reviewer_8p8kOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper is dedicated to empirical validation of the manifold hypothesis. It proposes the GMI score, a continuous diagnostic for two necessary signatures of smooth embedded manifold structure in point clouds: local intrinsic-dimension consistency and positive Federer reach. GMI is scalable to modern high-dimensional representations and separates dimension inconsistency from reach-related failures such as cusps and self-intersections. An extensive empirical validation with several modalities is done. Authors conclude that 1) real data is far from a manifold structure, but are closer to union of manifolds 2) learned embeddings are far from manifolds.

Critical strengths
  1. The problem setting itself is interesting. In fact, in many papers instead of "manifoldness" of data we mean "low intrinsic dimension" of data. The paper makes a clear distinction here and proposes a score of "manifoldness".
  2. An empirical observation that pre-processing with PCA is critical for "manifoldness" is an interesting observation.
  3. The decomposition of "manifoldness" into 2 term is an interesting idea.
Critical weaknesses
  1. The manifold hypothesis states the data lies in a vicinity of a manifold. So a natural noise is present in every real dataset. GMI decreases with noise (Fig.2). So the manifold hypothesis is not refuted; I consider that low GMI in real datasets is due to a noise. Random rotation of a digit from MNIST yields GMI 0.6-0.8, and the digit subset itself yields GMI ~ 0.6. Other digits have lower GMI, but, probably due to noise since different persons write digits in a different way.

  2. Experiments with autoencoders are not convincing. We don't expect a manifold in a representation space. For example, VAE explicitly enforces representations to be Gaussian, not having low intrinsic dimension in the ambient space. But what is interesting and missing in the paper - what is the GMI of samples from autoencoders or generative models? As deterministic functions of low-dimensional latent data the must be manifold-like.

  3. Sometimes GMI decreases when increasing N, this phenomenon is not fully explained. Can it be due to non-uniform sampling?

Questions for authors
  1. What is the difference of your work with : Brown, B. C., Caterini, A. L., Ross, B. L., Cresswell, J. C., & Loaiza-Ganem, G. (2022). Verifying the union of manifolds hypothesis for image data.

Brown et al., have already studied the problem and concluded that the union of manifold hypothesis is right, while the original manifold hypothesis is wrong for real image data.

  1. What does this paragraph means? "A global Two-NN (Facco et al., 2017) query on X_active supplies ε_density, ε_norm"

  2. Why you estimate local dimensionality via local PCA, instead of a local TwoNN ?

  3. Why GMI on real manifolds is only 0.83, not 1?

  4. What is OpenAI 3?

  5. Why GMI is so low for color tokens set, which component of GMI is more influential?

  6. What can be a practical threshold for GMI to conclude "manifoldness" ?

2 comments
AuthorsOpenReview ↗

response 1

We thank the reviewer for their thoughtful assessment and for recognising the value of distinguishing intrinsic dimension from geometric manifoldness.

We would first like to clarify that the original submission did not adequately communicate the intended contribution of our work. Our primary objective was to introduce, characterise, and demonstrate the usefulness of GMI as a continuous diagnostic of geometric regularity in finite point clouds, rather than to refute the manifold hypothesis. We have substantially revised the manuscript to better reflect our intended contribution, separating the systematic characterisation of GMI (Section 5) from its practical applications to data and representation spaces (Section 6).

Practical applications of GMI. A central aspect of the revised manuscript is demonstrating how GMI can be used to investigate geometric properties of real-world data and learned representations. Section 6 presents four applications:

  1. Diagnosing dataset heterogeneity: Comparing semantic subsets with pooled datasets reveals differences in geometric regularity and identifies whether these arise primarily from local dimension inconsistency or reach-related effects. For example, pooled MNIST exhibits greater dimensional heterogeneity than individual digit classes, illustrating how GMI can inform the choice between analysing a dataset globally or through more homogeneous subsets.

  2. Analysing learned representations: Comparing autoencoder families shows how different training objectives are associated with distinct geometric structures in latent space, including differences between class-specific and pooled representations.

  3. Evaluating preprocessing choices: In single-cell transcriptomics, GMI reveals substantial sensitivity to PCA dimensionality in Puram, compared with greater stability in Baron2. We complement this analysis with quantitative neighbourhood-preservation measurements, illustrating how GMI can help assess the sensitivity of geometric analyses to representation choices.

  4. Assessing representation quality: We examine the relationship between GMI and reconstruction error across autoencoder families. The associations differ across training objectives, demonstrating that greater geometric regularity does not necessarily accompany better reconstruction performance. GMI thus provides complementary information about representation geometry rather than a universal proxy for downstream quality.

These applications illustrate the diagnostic value of GMI beyond controlled synthetic geometries, particularly for identifying geometric heterogeneity, understanding representation differences, and assessing sensitivity to preprocessing decisions.

Manifold hypothesis and noise. We agree with the reviewer that the manifold hypothesis allows observations to lie in the vicinity of a manifold, and that noisy data may receive low GMI without contradicting this hypothesis. This distinction is now explicit throughout the manuscript. In particular, Section 5.3 examines how GMI responds to increasing additive noise, demonstrating that it provides a graded assessment of departures from regular geometry rather than a binary test of manifold compatibility.

Sampling and increasing sample size. We also agree that sampling can explain variations in GMI, including decreases with increasing sample size. We have expanded the experimental characterisation to investigate this explicitly. Section 5.4 compares uniformly spaced and randomly selected samples from the same underlying image-rotation manifold (Figure 11), showing that uneven coverage can substantially affect GMI even when the underlying geometry is unchanged. Appendix F further examines finite-sample limitations across intrinsic dimensions.

Relation to Brown et al. (2022). Our work is complementary to Brown et al., who investigate the union-of-manifolds hypothesis for image data. Rather than testing whether observations belong to a particular manifold model or union of manifolds, GMI measures two selected geometric properties without requiring an explicit manifold decomposition. The revised paper uses comparisons between semantic subsets and pooled datasets to diagnose geometric heterogeneity, without claiming that these differences establish a union-of-manifolds model.

Intrinsic-dimension estimation. The global TwoNN calculation provides a coarse dimension estimate and the nearest-neighbour distance scale used for subsequent processing and normalisation. Local PCA is then used to measure dimension consistency across overlapping patches. We have clarified this distinction and included comparisons against local TwoNN, MLE, and caPCA in Appendix C. Our experiments motivate local PCA as a practical compromise between computational cost and stable dimension-consistency measurements.

AuthorsOpenReview ↗

Response 2

Interpretation of GMI values. GMI need not equal one even on known regular manifolds, because local dimensions and tangent spaces are estimated from finite samples and the reach-related component is normalised by sampling resolution. This is particularly evident at higher intrinsic dimensions. Accordingly, we do not propose a universal threshold for concluding that data are manifold-like.

Autoencoder representations. We agree that the latent distribution of a VAE is influenced by its prior and need not concentrate on a lower-dimensional manifold. The revised paper compares the measured geometry of different representations without assuming that all latent spaces should be manifolds. We also appreciate the suggestion to investigate decoder-generated samples. This would be a useful complementary experiment, although the current results concern encoded observations rather than samples generated by the decoder.

Finally, we have removed the token-embedding experiments, including the ambiguously labelled OpenAI 3 and colour-token examples, to focus the paper on the settings for which we provide a more systematic characterisation.

We hope the revised manuscript better communicates both the intended scope and practical utility of GMI: a diagnostic for investigating geometric properties of finite point clouds

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.