Pith. sign in

REVIEW 2 major objections 5 minor 68 references

Robust Tangent Space Estimation via Laplacian Eigenvector Gradient Orthogonalization

T0 review · 2 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that tangent spaces on noisy data manifolds can be estimated robustly by orthogonalizing the gradients of low-frequency graph-Laplacian eigenvectors, sidestepping the neighborhood-size trade-off that defeats local PCA.

desk verdict Genuinely new spectral method for tangent estimation, but the central recovery claim has an unproven spanning premise that the authors' own Figure 7 shows can fail. read the letter →

arxiv 2510.02308 v2 pith:52VUOJ6S submitted 2025-10-02 cs.LG math.DG

Dhruv Kohli , Sawyer J. Robertson , Gal Mishne , Alexander Cloninger This is my paper
classification cs.LGmath.DG
keywords tangentspaceestimationgraphLaplacianeigenvectorgradientstubularneighborhoodlocalPCAmanifoldlearningintrinsicdimensionspectralmethods
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Tangent spaces of data manifolds are the foundation of geometric data analysis, and the usual local-PCA estimator breaks down under noise because a good neighborhood size depends on unknown curvature and noise level. The paper proposes LEGO, which estimates each tangent space by orthogonalizing the gradients of low-frequency eigenvectors of the graph Laplacian built from the noisy data. It argues, by a differential-geometric tube analysis, that low-frequency eigenfunctions have small gradient normal to the manifold, and, by a random-matrix analysis, that low-frequency eigenvectors are stable under noise. Experiments on a Swiss roll, a truncated torus, and an image dataset show LEGO tangent estimates are more accurate than LPCA's and improve downstream manifold embeddings, boundary detection, and local intrinsic dimension estimation. A sympathetic reader would care because the method sidesteps the neighborhood-size trade-off without requiring prior knowledge of noise or curvature.

What carries the argument

The central object is the gradient of a low-frequency eigenvector of the graph Laplacian. The primary mechanism is a tube-energy inequality: on a tubular neighborhood of the manifold, an eigenfunction's eigenvalue is bounded below by roughly the reciprocal of the square of the tube width times its vertical energy (gradient across the tube cross-sections), while horizontal energy enters only linearly. This forces low-eigenvalue eigenfunctions to have small gradient in the normal/noise directions. The second mechanism is Laplacian stability: the noisy graph Laplacian differs from the clean one by O(n^{-1/2}) in Frobenius norm with high probability under sub-Gaussian noise, and classical eigenv

What would settle it

The cleanest check is to take a manifold on which the first m low-frequency Laplacian gradients are known to span only a subspace—for example, a high-aspect-ratio Swiss roll at the m values the paper uses—and show the orthogonalized gradients miss a tangent direction; the paper's Figure 7 already exhibits that failure. More generally, a synthetic dataset where LPCA with an oracle-selected neighborhood size beats LEGO would settle the robustness claim.

Watch

Extended reading notes

Core claim

The paper's core claim is that the span of gradients of the first few low-frequency eigenvectors of a graph Laplacian built from noisy data, after orthogonalization, yields a tangent space estimate that is robust to noise. Formally, on a tubular neighborhood of a smooth submanifold, any eigenfunction whose eigenvalue is small has small vertical energy—its gradient component normal to the manifold is tiny (Theorem 3 and Corollary 4 give that the eigenvalue scales as a large negative power of the tube-width parameter times vertical energy, while horizontal energy only enters linearly). Separately, for sub-Gaussian noise, the random-walk graph Laplacian of the noisy data is O(n^{-1/2}) close to

Load-bearing premise

The load-bearing premise is that the gradients of the first m low-frequency eigenvectors actually span the full d-dimensional tangent space at every point; the paper proves their normal components are small and that the eigenvectors are stable, but it never derives spanning, and its own high-aspect-ratio Swiss roll shows the first 20 gradients collapsing to a line.

Editorial extensions

If this is right

  • Tangent space estimates stop depending on tuning neighborhood size to unknown noise and curvature; the global spectrum supplies the local geometry.
  • Downstream algorithms that consume tangent estimates—bottom-up manifold learning alignment, boundary detection, and local intrinsic dimension—inherit the noise robustness and produce markedly better outputs.
  • The spectral principle that vertical energy inflates eigenvalue by roughly the inverse squared tube width means noise-contaminated eigenmodes can be identified by their spectral depth rather than by local density.
  • The random-matrix stability result establishes that a fixed set of low eigenvector gradients is a reliable surrogate for the clean ones as the number of points grows, provided the low eigengaps of the clean Laplacian do not decay faster than n^{-1/2}.
  • A practitioner can keep the number of eigenvectors whose gradients are orthogonalized small; the experiments show stable tangent estimates across a broad range of that hyperparameter.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Reading the theory strictly, the rate assumption ε=O(1/√(n log n)) models vanishing noise, while all experiments use fixed noise; the theorems should be read as qualitative support for the demonstrated regime rather than a quantitative cover.
  • The proof that low-frequency eigenvector gradients span the full tangent space is imported from earlier local-parameterization results rather than re-derived; for a high-aspect-ratio Swiss roll the paper's own Figure 7 shows the first 20 gradients confined to a line, so the guarantee is conditional on using enough eigenvectors and on the manifold's aspect ratio.
  • A direct extension would be a data-driven rule for choosing the number of eigenvectors from the functional-variance plateau of the gradient Gram matrix; the paper shows stability across values but leaves the choice to the user.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper proposes LEGO, a spectral method for tangent-space estimation from noisy point clouds. The algorithm computes m0 low-frequency eigenvectors of a graph Laplacian, estimates their gradients by local least-squares fits on k_nn neighborhoods, and at each point takes the top d left singular vectors of the resulting p-by-m gradient matrix. Theoretical support is offered in two parts: (i) a tubular-neighborhood analysis (Theorem 3 and Corollary 7) showing that eigenfunctions with small eigenvalue have small normal ('vertical') energy, so their gradients are nearly tangent to the clean manifold; and (ii) a random-matrix concentration analysis (Theorems 10--13 and Corollary 14) showing that, when the noise proxy epsilon = O(1/sqrt(n log n)), the noisy graph Laplacian is within O(n^{-1/2}) of the clean one, implying eigenvector stability under eigengap conditions. Experiments on Swiss-roll, truncated-torus, and image data compare LEGO with LPCA and report improved tangent accuracy and downstream performance in manifold learning, boundary detection, and local intrinsic-dimension estimation.

Significance. If the claims held, the method would be practically valuable: it promises tangent-space estimates that are robust to noise without knowing the noise level or curvature in advance, and the paper demonstrates clear empirical gains in several downstream tasks. The analytical contributions are substantial and mostly carefully executed: the tube eigenvalue--energy bounds (Theorem 3) and the explicit concentration chain leading to O(n^{-1/2}) Laplacian stability are nontrivial, with constants that are not fitted to the data. However, the central algorithmic claim requires that the gradients of the first m low-frequency global eigenvectors span the d-dimensional tangent space at every point. This spanning property is not proven, and the paper's own Figure 7 documents a failure exactly of this kind on a high-aspect-ratio Swiss roll. Thus the theoretical scope is narrower than the abstract suggests, and the paper needs substantial revision before the central claim can be accepted.

major comments (2)
  1. [Section 3, Theorem 3 and Corollary 7; Figure 7 caption] The results in Section 3 bound the vertical energy of low-eigenvalue tube eigenfunctions, but they do not establish that the gradients of the first m global low-frequency eigenvectors span the d-dimensional tangent space at each point. The algorithm (Algorithm 1, lines 5--7) uses the top d left singular vectors of the p-by-m gradient matrix; if those gradients lie in a proper subspace, the estimate is incomplete. The citations to [25] and [4] are not sufficient, because those works select point-dependent subsets of eigenfunctions, whereas LEGO uses the first m global modes. The paper itself documents the failure: in the caption of Figure 7, 'm=10 and 20 result in high tangent space discrepancy for the Swiss roll. This is because of its high aspect ratio which makes the gradients of the first 20 eigenvectors to be restricted to a one-dimensional subspace.' A flat model is instructive: for
  2. [Section 4, Theorems 10--13; Section 5.1 and Figure 6] The random-matrix analysis assumes epsilon = O(1/sqrt(n log n)) with n tending to infinity, and the proofs use this decay explicitly in the union-bound estimates. The experiments, however, use fixed noise levels (sigma = 1 in the ablation of Figure 6, and epsilon = 0.0175 for the Swiss roll in Section 5.1) and fixed n. Consequently, the stated O(n^{-1/2}) concentration bounds do not apply to the experimental regime. The tube analysis (Theorem 3) does cover fixed epsilon, but it only bounds normal energy and does not, by itself, give the eigenvector stability needed for gradient estimation. The authors should either provide fixed-epsilon, non-asymptotic concentration bounds (with constants depending on epsilon), or explicitly state that the experiments are heuristic demonstrations outside the proven regime.
minor comments (5)
  1. [Section 5.1 vs Figure 6] There is an inconsistency in the reported noise level: Section 5.1 states epsilon = 0.0175 for the Swiss roll, while the Figure 6 caption says experiments use the maximum noise level sigma = 1. Please reconcile these.
  2. [Equations (2)--(6)] The notation bnabla-phi_i is overloaded: in Eq. (2) it denotes a p-by-n matrix of estimated gradients, while in the optimization (3)--(4) it is used for local quantities. Using distinct symbols would improve readability.
  3. [Section 3.1, Eq. (12)] The constant C_k is said to depend on the dimension k and the boundary conditions, but its value or defining eigenvalue problem is never stated explicitly. Please give the precise definition or a reference.
  4. [Proof of Theorem 13] In the second-half bound, the expression 'D^{-1}K - D^{-1}K' is easy to misread as involving D^{-1}D = I. Since D and D are different diagonal matrices, the displayed factorization is correct, but a brief explanation or parentheses would prevent confusion.
  5. [Algorithm 1] The pseudocode requires either d or f_var, but no default or guidance is given for choosing f_var when d is unknown. A sentence on typical values would be helpful.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LEGO's derivation is self-contained; the main gap is an unproved spanning premise, not a circular reduction.

full rationale

The claimed derivation chain is not circular. The differential-geometric justification (Theorem 3, Corollaries 4 and 7) derives lower bounds on eigenvalue in terms of vertical energy and upper bounds in terms of horizontal energy from the pullback metric on the tube; this is a genuine mathematical result, not a restatement of the algorithm's output. The random-matrix part (Theorems 10, 12, 13, Corollary 14) proves ||L−L||_F = O(n^{-1/2}) under epsilon = O(1/sqrt(n log n)) and imports Davis-Kahan; again no fitted quantity is recycled as a prediction. The only fit in the pipeline is the least-squares gradient estimator (Eq. 5) and the subsequent SVD, and the quality measure is an external discrepancy against clean-data tangents (Eq. 44). Quotes in Section 1 invoke the authors' LDLE [4] and Jones-Maggioni-Schul [25] for the premise that eigenfunction gradients span the tangent space, a premise the paper never derives for the first m global eigenvectors; the Figure 7 caption explicitly concedes 'the gradients of the first 20 eigenvectors to be restricted to a one-dimensional subspace' on the high-aspect-ratio Swiss roll. This is a genuine limitation/omitted proof and a coverage mismatch (theory assumes vanishing noise while experiments use fixed noise), but it is not an equation-level circularity: the paper does not define the tangent estimate in terms of the theoretical conclusion, nor fit any parameter to the discrepancy curves. Hence score 0.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

All free parameters are algorithmic hyperparameters; none are fit to the target discrepancy. The theory's geometric constants (ε, r, κ*, κ⊥*, C_k, λ_{B2}) are data dependent or universal, not fitted. Two genuine load-bearing assumptions: the normal-direction noise model (which the tube analysis needs) and the unproven spanning of the tangent space by low-frequency gradients; the vanishing-noise rate ε = O(1/√(n log n)) is also stronger than the experimental fixed-noise setting.

free parameters (5)
  • k_nn (neighborhood size) = 9 (Swiss roll), 14 (torus, puppets)
    User-chosen neighborhood size; LEGO still uses k_nn to estimate gradients, so the LPCA trade-off is reduced but not eliminated.
  • m (number of eigenvectors orthogonalized) = 40 (Swiss roll), 20 (torus, puppets)
    Chosen after the stability analysis in Fig. 7; m=10,20 fail on the Swiss roll case.
  • m0 (spectral basis size) = 100
    Must exceed m; chosen so gradient components are well-approximated in the spectral basis.
  • kernel bandwidth s / kernel type = not stated for experiments
    Experiments say the random-walk Gaussian-kernel Laplacian from §4 is used, but the bandwidth s and its scaling are not given; the §3/§4 theory assumes fixed bandwidth.
  • f_var (intrinsic-dimension threshold) = user-defined
    For dimension estimation; standard in LPCA as well.
assumptions (7)
  • domain assumption Noise in the tube model is confined to directions normal to the manifold; data lies in a tubular neighborhood T_εr of a smooth compact submanifold
    Section 2/3 setup. This is what makes the vertical-gradient eigenvalue penalty (Theorem 3) the relevant noise mechanism; the puppets experiment instead uses isotropic noise, so only one experiment matches this model.
  • domain assumption Noise variance proxy ε = O(1/√(n log n)) as n → ∞, with feature dimension p fixed
    Assumed in Theorems 10–13; also required for the degree lower bound (Lemma 11). The fixed-noise experimental regime is outside this assumption.
  • domain assumption Clean data is bounded in a ball of radius R
    Used in Lemma 11 to bound degrees below; reasonable for the datasets.
  • domain assumption Eigengaps of the clean Laplacian decay slower than n^{-1/2} (ω(n^{-1/2})) for each of the first m eigenvectors
    Corollary 14's eigenvector stability requires this; the paper argues Weyl's law makes it plausible for fixed low-frequency eigenvectors, but it is an assumption, not a consequence.
  • ad hoc to paper Gradients of the first m low-frequency eigenvectors span the d-dimensional tangent space at every point
    Imported from Jones–Maggioni–Schul [25] and LDLE [4]; not derived here, and contradicted for the high-aspect-ratio Swiss roll at m≤20 (Fig. 7).
  • domain assumption Corollary 4 tube-regime conditions: r ≤ √(C_k/λ_{B2}) and (1−ε)^{d+1} ≥ (1+ε)^d ε^{1−t}
    Corollary 4 conditions; rκ* < 1 guaranteed by reach; satisfied for small ε but restrict the scope of the Ω(ε^{-2t})/O(E_B) claim.
  • standard math Standard concentration, perturbation, and spectral-convergence results (sub-Gaussian quadratic-form tails [64], Davis–Kahan [46], Weyl's law [47], graph-Laplacian → tube-Laplacian convergence [49,54])
    These background results are invoked in Sections 3 and 4 and are accepted as standard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Tangent Space Estimation via Laplacian Eigenvector Gradient Orthogonalization." pith.science (2026). https://pith.science/paper/52VUOJ6S

@misc{pith2026251002308,
  author       = {Pith},
  title        = {Pith review of: Robust Tangent Space Estimation via Laplacian Eigenvector Gradient Orthogonalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/52VUOJ6S}},
  note         = {Machine review of arXiv:2510.02308}
}
read the original abstract

Estimating the tangent spaces of a data manifold is a fundamental problem in geometric data analysis. The standard approach, Local Principal Component Analysis (LPCA), struggles in high-noise setting due to a critical trade-off in choosing the neighborhood size. Selecting an optimal size requires prior knowledge of the geometric and noise characteristics of the data that are often unavailable. In this paper, we propose a spectral method, Laplacian Eigenvector Gradient Orthogonalization (LEGO), that utilizes the global structure of the data to guide local tangent space estimation. Instead of relying solely on local neighborhoods, LEGO estimates the tangent space at each data point by orthogonalizing the gradients of low-frequency eigenvectors of the graph Laplacian. We provide two theoretical justifications of our method. First, a differential geometric analysis on the tubular neighborhood of a manifold shows that gradients of the low-frequency Neumann eigenfunctions of the tube align closely with the manifold's tangent bundle, while an eigenfunction with high gradient in directions orthogonal to the manifold lie deeper in the spectrum. Second, a random matrix theoretic analysis also demonstrates that low-frequency eigenvectors are robust to sub-Gaussian noise. These results allow us to derive the asymptotic scaling and stability of the estimated eigenvector gradients. Numerical experiments demonstrate that LEGO yields tangent space estimates that are significantly more robust to noise than those from LPCA, resulting in marked improvements in downstream tasks such as manifold learning, boundary detection, and local intrinsic dimension estimation.

Figures

Figures reproduced from arXiv: 2510.02308 by the authors.

Figure 1
Figure 1. Illustration of tangent space estimation using LPCA and LEGO on a noisy point cloud generated by non-uniform sampling of a closed curve—wave on a circle— with heteroskedastic noise added in the normal direction. (a) Clean data points with ground truth tangent vectors, along with tangent vectors estimated from the noisy data using LPCA (knn = 14 and d = 1) and LEGO (knn = 14, m0 = 20, m = 100 and d = 1). (b) Cosine d… view at source ↗
Figure 2
Figure 2. The first non-trivial eigenvector of the Laplacian L is plotted against increas￾ing noise level ε. Here, the noise is independent but non-identically distributed as in Figure 1a (also see Remark 9). The common colorbar represents the range of values, from minimum to maximum, across the eigenvectors. As before, let Y = {Y1, . . . , Yn} be the clean data points sampled from a bounded region in R p . Fixing a bandwidth… view at source ↗
Figure 3
Figure 3. (a) Clean and noisy Swiss roll with high-aspect ratio in R 3 colored by the “roll” parameter. (b) Discrepancy between the true and the estimated tangent spaces due to LPCA (knn = 9) and LEGO (knn = 9, m0 = 100, m = 40), as computed using Eq. 44. (c, d) 2-dimensional parameterization of the noisy data, and the boundary points detected from the noisy data using the estimated and the true tangent spaces (see Section C.… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: (a) Clean and noisy truncated torus in R 3 colored by the noise level. (b) Discrepancy between the true and the estimated tangent spaces due to LPCA (knn = 14) and LEGO (knn = 14, m0 = 100, m = 20), as computed using Eq. 44. (c, d) 2-dimensional parameterization of the…
Figure 5
Figure 5. Figure 5: (a) Sample clean images from the Yoda and Bulldog dataset [66] (first and third columns), along with their noise-perturbed versions (second and fourth columns). (b) Explained variance ratio for the first 30 principal directions obtained via PCA. As the variance saturat…
Figure 6
Figure 6. Figure 6: The discrepancy between true and the estimated tangent spaces from the noisy data Xj = Yj + σηjνj , j ∈ [1, n], as the noise level σ varies between 0 and 1. In our experiments in Section 5, we used the maximum noise level i.e. σ = 1 [PITH_FULL_IMAGE:figures/full_fig_p…
Figure 7
Figure 7. Figure 7: The discrepancy between the true and the estimated tangent spaces on the noisy datasets described in Section 5, against several different values of the hyperparam￾eters m and m0 in LEGO, provided at the top and the bottom of the x-axis, respectively. The noise level co…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 7 linked inside Pith

  1. [25]

    Universal local parametrizations via heat kernels and eigenfunctions of the Laplacian

    Peter W Jones, Mauro Maggioni, and Raanan Schul. “Universal local parametrizations via heat kernels and eigenfunctions of the Laplacian”. In:arXiv preprint arXiv:0709.1975(2007)

    arXiv 1975

  2. [4]

    LDLE: Low Distortion Local Eigenmaps

    Dhruv Kohli, Alexander Cloninger, and Gal Mishne. “LDLE: Low Distortion Local Eigenmaps”. In: Journal of Machine Learning Research22.282 (2021), pp. 1–64

    2021

  3. [1]

    Principal manifolds and nonlinear dimensionality reduction via tangent space alignment

    Zhenyue Zhang and Hongyuan Zha. “Principal manifolds and nonlinear dimensionality reduction via tangent space alignment”. In:SIAM journal on scientific computing26.1 (2004), pp. 313–338

    2004

  4. [2]

    Nonlinear dimensionality reduction by locally linear embed- ding

    Sam T Roweis and Lawrence K Saul. “Nonlinear dimensionality reduction by locally linear embed- ding”. In:science290.5500 (2000), pp. 2323–2326. 24 REFERENCES Figure 6.The discrepancy between true and the estimated tangent spaces from the noisy dataX j =Y j +ση jνj,j∈[1, n], as the noise levelσvaries between 0 and 1. In our experiments in Section 5, we use...

    2000

  5. [3]

    Hessian eigenmaps: Locally linear embedding techniques for high-dimensional data

    David L Donoho and Carrie Grimes. “Hessian eigenmaps: Locally linear embedding techniques for high-dimensional data”. In:Proceedings of the National Academy of Sciences100.10 (2003), pp. 5591–5596

    2003

  6. [5]

    RATS: Unsupervised manifold learning using low-distortion alignment of tangent spaces

    Dhruv Kohli et al. “RATS: Unsupervised manifold learning using low-distortion alignment of tangent spaces”. In:bioRxiv(2024)

    2024

  7. [6]

    Manifold learning: What, how, and why

    Marina Meil˘ a and Hanyu Zhang. “Manifold learning: What, how, and why”. In:Annual Review of Statistics and Its Application11 (2024)

    2024

  8. [7]

    Low-dimensional embeddings of high-dimensional data

    Cyril de Bodt et al. “Low-dimensional embeddings of high-dimensional data”. In:arXiv preprint arXiv:2508.15929(2025)

    arXiv 2025

Show all 68 references
  1. [8]

    Locally linear denoising on image manifolds

    Dian Gong, Fei Sha, and G´ erard Medioni. “Locally linear denoising on image manifolds”. In:Pro- ceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics. JMLR Workshop and Conference Proceedings. 2010, pp. 265–272

    2010

  2. [9]

    Large sample spectral analysis of graph-based multi-manifold clustering

    Nicolas Garcia Trillos, Pengfei He, and Chenghui Li. “Large sample spectral analysis of graph-based multi-manifold clustering”. In:Journal of Machine Learning Research24.143 (2023), pp. 1–71

    2023

  3. [10]

    Spectral clustering on multiple manifolds

    Yong Wang et al. “Spectral clustering on multiple manifolds”. In:IEEE Transactions on Neural Networks22.7 (2011), pp. 1149–1161

    2011

  4. [11]

    Robust multiple manifolds structure learning

    Dian Gong, Xuemei Zhao, and G´ erard Medioni. “Robust multiple manifolds structure learning”. In: arXiv preprint arXiv:1206.4624(2012). REFERENCES 25

    2012 arXiv

  5. [12]

    Spectral clustering based on local linear approximations

    Ery Arias-Castro, Guangliang Chen, and Gilad Lerman. “Spectral clustering based on local linear approximations”. In:Electronic Journal of Statistics5.none (2011), pp. 1537–1587

    2011

  6. [13]

    Maximum likelihood estimation of intrinsic dimension

    Elizaveta Levina and Peter Bickel. “Maximum likelihood estimation of intrinsic dimension”. In: Advances in neural information processing systems17 (2004)

    2004

  7. [14]

    Spectral convergence of the connection Laplacian from random samples

    Amit Singer and Hau-Tieng Wu. “Spectral convergence of the connection Laplacian from random samples”. In:Information and Inference: A Journal of the IMA6.1 (2017), pp. 58–123

    2017

  8. [15]

    Vector diffusion maps and the connection Laplacian

    Amit Singer and H-T Wu. “Vector diffusion maps and the connection Laplacian”. In:Communica- tions on pure and applied mathematics65.8 (2012), pp. 1067–1144

    2012

  9. [16]

    On a generalization of Wasserstein distance and the Beckmann problem to connection graphs

    Sawyer Robertson et al. “On a generalization of Wasserstein distance and the Beckmann problem to connection graphs”. In:arXiv:2312.10295(2023)

    2023 arXiv

  10. [17]

    Local linear regression on manifolds and its geometric inter- pretation

    Ming-Yen Cheng and Hau-tieng Wu. “Local linear regression on manifolds and its geometric inter- pretation”. In:Journal of the American Statistical Association108.504 (2013), pp. 1421–1434

    2013

  11. [18]

    Nonasymptotic rates for manifold, tangent space and curva- ture estimation

    Eddie Aamari and Cl´ ement Levrard. “Nonasymptotic rates for manifold, tangent space and curva- ture estimation”. In:The Annals of Statistics47.1 (2019), pp. 177–204

    2019

  12. [19]

    Non-asymptotic analysis of tangent space perturba- tion

    Daniel N Kaslovsky and Fran¸ cois G Meyer. “Non-asymptotic analysis of tangent space perturba- tion”. In:Information and Inference: a Journal of the IMA3.2 (2014), pp. 134–187

    2014

  13. [20]

    Tangent space estimation for smooth embeddings of Riemannian manifolds®

    Hemant Tyagi, Elıf Vural, and Pascal Frossard. “Tangent space estimation for smooth embeddings of Riemannian manifolds®”. In:Information and Inference: A Journal of the IMA2.1 (2013), pp. 69–114

    2013

  14. [23]

    Selecting the independent coordinates of manifolds with large aspect ratios

    Yu-Chia Chen and Marina Meila. “Selecting the independent coordinates of manifolds with large aspect ratios”. In:Advances in Neural Information Processing Systems32 (2019)

    2019

  15. [24]

    The Noisy Laplacian: a Threshold Phenom- enon for Non-Linear Dimension Reduction

    Alex Kokot, Octavian-Vlad Murad, and Marina Meila. “The Noisy Laplacian: a Threshold Phenom- enon for Non-Linear Dimension Reduction”. In:Forty-second International Conference on Machine Learning. 2025

    2025

  16. [26]

    John M Lee.Introduction to Riemannian manifolds. Vol. 2. Springer, 2018

    2018

  17. [27]

    Perturbation of the eigenvectors of the graph Laplacian: Appli- cation to image denoising

    Fran¸ cois G Meyer and Xilin Shen. “Perturbation of the eigenvectors of the graph Laplacian: Appli- cation to image denoising”. In:Applied and Computational Harmonic Analysis36.2 (2014), pp. 326– 334

    2014

  18. [28]

    Impact of signal-to-noise ratio and bandwidth on graph Laplacian spectrum from high-dimensional noisy point cloud

    Xiucai Ding and Hau-Tieng Wu. “Impact of signal-to-noise ratio and bandwidth on graph Laplacian spectrum from high-dimensional noisy point cloud”. In:IEEE Transactions on Information Theory 69.3 (2022), pp. 1899–1931

    2022

  19. [29]

    The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains

    David I Shuman et al. “The emerging field of signal processing on graphs: Extending high-dimensional data analysis to networks and other irregular domains”. In:IEEE signal processing magazine30.3 (2013), pp. 83–98

    2013

  20. [30]

    Wavelets on graphs via spectral graph theory

    David K Hammond, Pierre Vandergheynst, and R´ emi Gribonval. “Wavelets on graphs via spectral graph theory”. In:Applied and Computational Harmonic Analysis30.2 (2011), pp. 129–150

    2011

  21. [31]

    Spectral embedding norm: Looking deep into the spectrum of the graph Laplacian

    Xiuyuan Cheng and Gal Mishne. “Spectral embedding norm: Looking deep into the spectrum of the graph Laplacian”. In:SIAM journal on imaging sciences13.2 (2020), pp. 1015–1048

    2020

  22. [32]

    Elias M Stein and Rami Shakarchi.Fourier analysis: an introduction. Vol. 1. Princeton University Press, 2011

    2011

  23. [33]

    Cambridge University Press, 2004

    Yitzhak Katznelson.An introduction to harmonic analysis. Cambridge University Press, 2004

    2004

  24. [34]

    Elsevier, 1999

    St´ ephane Mallat.A wavelet tour of signal processing. Elsevier, 1999

    1999

  25. [35]

    MLLE: Modified locally linear embedding using multiple weights

    Zhenyue Zhang and Jing Wang. “MLLE: Modified locally linear embedding using multiple weights”. In:Advances in neural information processing systems19 (2006)

    2006

  26. [36]

    Global registration of multiple point clouds using semidefinite programming

    Kunal N Chaudhury, Yuehaw Khoo, and Amit Singer. “Global registration of multiple point clouds using semidefinite programming”. In:SIAM Journal on Optimization25.1 (2015), pp. 468–501. 26 REFERENCES

    2015

  27. [37]

    Robust estimation of boundary using doubly stochastic scaling of Gaussian kernel

    Dhruv Kohli et al. “Robust estimation of boundary using doubly stochastic scaling of Gaussian kernel”. In:arXiv:2411.18942(2024)

    2024

  28. [38]

    Density estimation on manifolds with boundary

    Tyrus Berry and Timothy Sauer. “Density estimation on manifolds with boundary”. In:Computa- tional Statistics & Data Analysis107 (2017), pp. 1–17

    2017

  29. [39]

    Diffusion maps for embedded manifolds with bound- ary with applications to PDEs

    Ryan Vaughn, Tyrus Berry, and Harbir Antil. “Diffusion maps for embedded manifolds with bound- ary with applications to PDEs”. In:Applied and Computational Harmonic Analysis68 (2024), p. 101593.issn: 1063-5203

    2024

  30. [40]

    Curvature measures

    Herbert Federer. “Curvature measures”. In:Transactions of the American Mathematical Society 93.3 (1959), pp. 418–491

    1959

  31. [41]

    Finding the homology of submanifolds with high confidence from random samples

    Partha Niyogi, Stephen Smale, and Shmuel Weinberger. “Finding the homology of submanifolds with high confidence from random samples”. In:Discrete & Computational Geometry39 (2008), pp. 419–441

    2008

  32. [42]

    On information plus noise kernel random matrices

    Noureddine El Karoui. “On information plus noise kernel random matrices”. In:Annals of statistics 38.5 (2010), pp. 3191–3216

    2010

  33. [43]

    Graph Laplacians and their conver- gence on random neighborhood graphs

    Matthias Hein, Jean-Yves Audibert, and Ulrike von Luxburg. “Graph Laplacians and their conver- gence on random neighborhood graphs.” In:Journal of Machine Learning Research8.6 (2007)

    2007

  34. [44]

    Diffusion maps

    Ronald R Coifman and St´ ephane Lafon. “Diffusion maps”. In:Applied and computational harmonic analysis21.1 (2006), pp. 5–30

    2006

  35. [45]

    Strong consistency, graph Laplacians, and the stochastic block model

    Shaofeng Deng, Shuyang Ling, and Thomas Strohmer. “Strong consistency, graph Laplacians, and the stochastic block model”. In:Journal of Machine Learning Research22.117 (2021), pp. 1–44

    2021

  36. [46]

    A Useful Variant of the Davis–Kahan Theorem for Statis- ticians

    Y. Yu, T. Wang, and R. J. Samworth. “A Useful Variant of the Davis–Kahan Theorem for Statis- ticians”. In:Biometrika102.2 (June 2015), pp. 315–323

    2015

  37. [47]

    Isaac Chavel.Eigenvalues in Riemannian geometry. Vol. 115. Academic press, 1984

    1984

  38. [48]

    Self-tuning spectral clustering

    Lihi Zelnik-Manor and Pietro Perona. “Self-tuning spectral clustering”. In:Advances in Neural Information Processing Systems(2005), pp. 1601–1608

    2005

  39. [49]

    Convergence of graph Laplacian with kNN self-tuned kernels

    Xiuyuan Cheng and Hau-Tieng Wu. “Convergence of graph Laplacian with kNN self-tuned kernels”. In:Information and Inference: A Journal of the IMA11.3 (2022), pp. 889–957

    2022

  40. [50]

    Manifold learning with bi-stochastic kernels

    Nicholas F Marshall and Ronald R Coifman. “Manifold learning with bi-stochastic kernels”. In:IMA Journal of Applied Mathematics84.3 (2019), pp. 455–482

    2019

  41. [51]

    Doubly stochastic normalization of the Gauss- ian kernel is robust to heteroskedastic noise

    Boris Landa, Ronald R Coifman, and Yuval Kluger. “Doubly stochastic normalization of the Gauss- ian kernel is robust to heteroskedastic noise”. In:SIAM journal on mathematics of data science3.1 (2021), pp. 388–413

    2021

  42. [52]

    Towards a theoretical foundation for Laplacian-based manifold methods

    Mikhail Belkin and Partha Niyogi. “Towards a theoretical foundation for Laplacian-based manifold methods”. In:Journal of Computer and System Sciences74.8 (2008), pp. 1289–1308

    2008

  43. [53]

    From graph to manifold Laplacian: The convergence rate

    Amit Singer. “From graph to manifold Laplacian: The convergence rate”. In:Applied and Compu- tational Harmonic Analysis21.1 (2006), pp. 128–134

    2006

  44. [54]

    Error estimates for spectral convergence of the graph Laplacian on random geometric graphs toward the Laplace–Beltrami operator

    Nicol´ as Garc ´ ıa Trillos et al. “Error estimates for spectral convergence of the graph Laplacian on random geometric graphs toward the Laplace–Beltrami operator”. In:Foundations of Computational Mathematics20.4 (2020), pp. 827–887

    2020

  45. [55]

    Bi-stochastically normalized graph Laplacian: convergence to manifold Laplacian and robustness to outlier noise

    Xiuyuan Cheng and Boris Landa. “Bi-stochastically normalized graph Laplacian: convergence to manifold Laplacian and robustness to outlier noise”. In:Information and Inference: A Journal of the IMA13.4 (2024), iaae026

    2024

  46. [56]

    SIAM, 1998

    Richard B Lehoucq, Danny C Sorensen, and Chao Yang.ARPACK users’ guide: solution of large- scale eigenvalue problems with implicitly restarted Arnoldi methods. SIAM, 1998

    1998

  47. [57]

    Laplacian eigenmaps for dimensionality reduction and data representation

    Mikhail Belkin and Partha Niyogi. “Laplacian eigenmaps for dimensionality reduction and data representation”. In:Neural computation15.6 (2003), pp. 1373–1396

    2003

  48. [58]

    Non-parametric estimation of manifolds from noisy data

    Yariv Aizenbud and Barak Sober. “Non-parametric estimation of manifolds from noisy data”. In: arXiv preprint arXiv:2105.04754(2021)

    2021 arXiv

  49. [59]

    Estimation of Local Geometric Structure on Manifolds from Noisy Data

    Yariv Aizenbud and Barak Sober. “Estimation of Local Geometric Structure on Manifolds from Noisy Data”. In:arXiv preprint arXiv:2503.07220(2025)

    2025 arXiv

  50. [60]

    Estimating the intrinsic dimension of high-dimensional data sets: a multiscale, geometric approach

    Anna V Little. “Estimating the intrinsic dimension of high-dimensional data sets: a multiscale, geometric approach”. PhD thesis. Duke University, 2011. REFERENCES 27

    2011

  51. [61]

    Multiscale geometric methods for data sets I: Multiscale SVD, noise and curvature

    Anna V Little, Mauro Maggioni, and Lorenzo Rosasco. “Multiscale geometric methods for data sets I: Multiscale SVD, noise and curvature”. In:Applied and Computational Harmonic Analysis43.3 (2017), pp. 504–567

    2017

  52. [62]

    Minimax manifold estimation

    Christopher R Genovese et al. “Minimax manifold estimation”. In:The Journal of Machine Learning Research13.1 (2012), pp. 1263–1291

    2012

  53. [63]

    Generalised quantum waveguides

    Stefan Haag, Jonas Lampart, and Stefan Teufel. “Generalised quantum waveguides”. In:Annales Henri Poincar´ e. Vol. 16. Springer. 2015, pp. 2535–2568

    2015

  54. [64]

    A tail inequality for quadratic forms of subgaussian random vectors

    Daniel Hsu, Sham Kakade, and Tong Zhang. “A tail inequality for quadratic forms of subgaussian random vectors”. In:Electronic Communications in Probabilitynone (2012), pp. 1–6.doi:10.1214/ ECP.v17-2079

    2012

  55. [65]

    Principal angles between subspaces in an A-based scalar product: algorithms and perturbation estimates

    Andrew V Knyazev and Merico E Argentati. “Principal angles between subspaces in an A-based scalar product: algorithms and perturbation estimates”. In:SIAM Journal on Scientific Computing 23.6 (2002), pp. 2008–2040

    2002

  56. [66]

    Learning the geometry of common latent variables using alternating-diffusion

    Roy R Lederman and Ronen Talmon. “Learning the geometry of common latent variables using alternating-diffusion”. In:Applied and Computational Harmonic Analysis44.3 (2018), pp. 509–536

    2018

  57. [67]

    Generalized power method for generalized orthogonal Procrustes problem: global convergence and optimization landscape analysis

    Shuyang Ling. “Generalized power method for generalized orthogonal Procrustes problem: global convergence and optimization landscape analysis”. In:arXiv preprint arXiv:2106.15493(2021)

    2021 arXiv

  58. [68]

    Non-degenerate rigid alignment in a patch framework

    Dhruv Kohli, Gal Mishne, and Alexander Cloninger. “Non-degenerate rigid alignment in a patch framework”. In:arXiv:2303.11620(2023)

    2023

  59. [69]

    Global registration of multiple 3D point sets via optimization-on-a- manifold

    Shankar Krishnan et al. “Global registration of multiple 3D point sets via optimization-on-a- manifold.” In:Symposium on Geometry Processing. 2005, pp. 187–196

    2005

  60. [70]

    Robust inference of manifold density and geometry by doubly stochastic scaling

    Boris Landa and Xiuyuan Cheng. “Robust inference of manifold density and geometry by doubly stochastic scaling”. In:SIAM Journal on Mathematics of Data Science5.3 (2023), pp. 589–614

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.