Pith. sign in

REVIEW 4 minor 3 cited by

Minimax Rates for the Estimation of Eigenpairs of Weighted Laplace-Beltrami Operators on Manifolds

T0 review · 0 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proves that no estimator built from $n$ data points can recover eigenpairs of a weighted Laplace-Beltrami operator faster than $n^{-2/(d+4)}$, and that graph Laplacians match this rate up to logarithmic factors, making them…

desk verdict First real minimax result for estimating weighted Laplace-Beltrami eigenpairs from samples, with a clean Fano lower bound and a serious graph-Laplacian upper bound; the one substantive caveat is the C^{2,alpha}-vs-C^2 regularity gap in the upper bound. read the letter →

arxiv 2506.00171 v1 pith:PG2BF6B6 submitted 2025-05-30 stat.ML cs.LGmath.AP

Nicolás García Trillos , Chenghui Li , Raghavendra Venkatraman This is my paper
classification stat.MLcs.LGmath.AP MSC 62G0562G2060D0558J5035P15
keywords minimaxrateseigenpairestimationgraphLaplacianweightedLaplace-BeltramioperatorspectralconvergencemanifoldlearningH1normdensity
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

This paper asks a basic statistical question about spectral methods in unsupervised learning: given $n$ points drawn from an unknown density on an unknown manifold, how accurately can one recover an eigenpair (eigenvalue plus eigenfunction) of the weighted Laplace-Beltrami operator $\Delta_\rho f = -\frac{1}{\rho}\,\mathrm{div}(\rho^2\nabla f)$? Its central claim, stated as a minimax theorem, is that the worst-case error measured by $|\lambda_l - \hat\lambda_l| + \|f_l - \hat f_l\|_{H^1(M)}$ cannot decay faster than $n^{-2/(d+4)}$ in dimension $d$, the same rate as the classical minimax rate for density estimation. The paper further proves that the familiar graph Laplacian estimator, with connectivity parameter $\varepsilon_n \sim (\log n/n)^{1/(d+4)}$ and a data-driven extension operator, attains this rate up to logarithmic factors, making it essentially minimax optimal while remaining completely manifold-agnostic. This matters because spectral clustering, diffusion maps, and related manifold learning tools all rely on graph-Laplacian eigenpairs, and the paper gives what it presents as the first statistical lower bounds for this estimation problem.

What carries the argument

The lower bound is carried by Fano's method over a family of densities on the flat torus: each density is the uniform base plus many small localized perturbations $\frac{1}{m^2} a_i$, placed so that a perturbation in a region where the eigenfunction gradient is large moves the eigenpair by a definite amount, while any two members of the family stay at Kullback-Leibler distance $\sim m^{-4}$; choosing $m \sim n^{1/(d+4)}$ yields the $n^{-2/(d+4)}$ rate. The upper bound is carried by the discrete $H^{-1}(X_n)$ semi-norm, the dual of the graph energy semi-norm $\|u\|_{H^1(X_n)}$, reduced by a multiscale Poincaré inequality to finitely many inner products against rescaled indicator functions of cubes at all scales between $\varepsilon_n$ and the manifold diameter. A bias-variance analysis of $L_{\varepsilon_n,n} u - \Delta_\rho u$ in this weak norm, built on a second-order Taylor expansion along geodesics with exact remainder and a symmetrization that cancels the dominant linear term, produces $\varepsilon_n^2$ rates; the extension operator $\Lambda_r$, defined through the kernel $\psi(t) = \int_t^\infty \eta(s)s\,ds$, then transfers discrete $H^1$ control to $H^1(M)$.

What would settle it

Estimate the first nontrivial eigenpair on the flat torus (or a sphere) in dimension $d=2$, where the predicted rate is $n^{-1/3}$: if the empirical error of any estimator decays strictly faster than $n^{-1/3}$ across the hardest densities in $\mathcal{P}_{M,l}$, the lower-bound exponent is wrong, and if the graph-Laplacian error decays strictly slower than $n^{-1/3}\log n$, the upper bound is. A sharper probe: build a density with bounded second derivatives but borderline $\alpha=0$ regularity whose eigenfunction is not $C^3$ (as in the counterexamples cited in Remark 1.7) and test whether the graph-Laplacian rate degrades; if it does not, Assumption 3 is not needed for the theorem's conclusion.

Watch

Extended reading notes

Core claim

The paper establishes two matching statements that together assert a sharp rate correspondence between an information-theoretic limit and a practical algorithm. First (Theorem 1.4), for any $l \geq 2$, over the class of $d$-dimensional manifolds $M$ with bounded geometry and densities $\rho$ in the class $\mathcal{P}_{M,l}$ (bounded first and second derivatives plus a spectral gap), the minimax risk with the metric $|\lambda_l - \hat\lambda_l| + \|f_l - \hat f_l\|_{H^1(M)}$ is at least $c\,\lambda_l(\mathbb{T}^d,1)\,n^{-2/(d+4)}$; the lower bound is proved by Fano's method on the flat torus, using a packing of local density perturbations that separate eigenpairs while keeping Kullback-Leibler divergences small. Second (Theorems 1.6 and 1.10), under the slightly stronger assumption that $\rho$ has $C^{2,\alpha}$ second derivatives with $\alpha > 0$, the graph Laplacian built with bandwidth $\varepsilon_n \sim (\log n/n)^{1/(d+4)}$, followed by the extension $\Lambda_{\varepsilon_n/2}$, satisfies $\mathbb{E}[|\lambda_{n,l} - \lambda_l| + \|\hat\phi_{n,l} - f_l\|_{H^1(M)}] \leq C\,\lambda_l\, n^{-2/(d+4)} \log n / \log\log n$. Taken together, these theorems claim that graph Laplacian eigenpairs are essentially minimax optimal for estimating eigenpairs of weighted Laplace-Beltrami operators, uniformly over families of smooth densities and without knowing the manifold.

Load-bearing premise

For the graph-Laplacian upper bound, the density's second derivatives must be Hölder continuous with a strictly positive exponent $\alpha$; at the borderline $\alpha = 0$ the eigenfunctions of $\Delta_\rho$ may fail to be $C^3$, and the third-order Taylor-expansion argument that yields the $\varepsilon_n^2$ rates collapses. The $n^{-2/(d+4)}$ lower bound itself needs only bounded second derivatives, so the information-theoretic rate is more robust than the achievability proof.

Editorial extensions

If this is right

  • No estimator, graph-based or otherwise, can recover eigenpairs faster than $n^{-2/(d+4)}$ in the worst case over the class $\mathcal{P}_{M,l}$: the rate is an information-theoretic limit, not an artifact of the graph construction.
  • Graph Laplacians with bandwidth $\varepsilon_n \sim n^{-1/(d+4)}$, the same scaling as optimal kernel density estimators, are essentially minimax optimal, so staying well above the connectivity threshold is statistically justified for eigenpair estimation.
  • Manifold-agnostic graph estimators match the rate of the plug-in estimator that knows $M$ and solves a PDE, up to logarithms: knowing the manifold adds no asymptotic statistical power for this problem.
  • Measuring eigenfunction error in the $H^1(M)$ norm, which captures gradient information and therefore the tangent-plane content of spectral embeddings, comes at no extra statistical cost relative to $L^2$-type rates.
  • The arguments extend directly to estimating a fixed finite collection of eigenpairs, so the same rate governs the spectral objects used in spectral clustering and diffusion maps.

Reading between the lines

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

  • If the logarithmic factors in the upper bound are removable, as the paper expects, the upper and lower bounds coincide exactly and graph Laplacians are literally minimax optimal rather than optimal up to logs; this is a gap the paper leaves open for future work.
  • By analogy with the homogenization results the paper cites, the $n^{-2/(d+4)}$-type rates may extend to sparser graphs down toward the percolation threshold, with $\varepsilon_n\sqrt{\lambda_l} < 1$ as the natural validity range, but that regime is explicitly beyond this paper's scope.
  • Remark 2.9 leaves open the possibility that estimating eigenvalues alone has a strictly faster minimax rate than $n^{-2/(d+4)}$; if true, practitioners who only need eigenvalues could beat eigenpair estimators, a concrete and testable question.
  • The lower-bound construction is a template that likely transfers to normalized and random-walk graph Laplacians, $k$-NN graphs, and other elliptic operators arising as graph scaling limits, giving a general method for minimax bounds in operator learning.
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

0 major / 4 minor

Summary. The paper studies the minimax estimation of eigenpairs (λ_l, f_l) of weighted Laplace-Beltrami operators Δ_ρ = -ρ^{-1} div(ρ^2 ∇·) from n i.i.d. samples drawn from an unknown density ρ on an unknown d-dimensional manifold M. The central lower bound (Theorem 1.4) states that, uniformly over bounded-geometry manifolds and densities in the class P_{M,l} with controlled second derivatives and an eigengap, the risk |λ̂_l - λ_l| + ||f̂_l - f_l||_{H^1(M)} is at least of order n^{-2/(d+4)}. The main upper-bound contribution is a detailed analysis of graph Laplacians: under a slightly stronger C^{2,α} regularity assumption (Assumption 3), the graph Laplacian estimator with ε_n ~ (log n / n)^{1/(d+4)} and the extension Λ_{ε/2} achieves the same rate up to logarithmic factors (Theorems 1.6 and 1.10), and a known-manifold kernel-density plug-in estimator achieves the rate over the original C^2 class (Appendix C).

Significance. This is the first statistical lower bound for the eigenpair estimation problem in the stronger H^1-type norm, and the paper convincingly shows that the n^{-2/(d+4)} density-estimation rate is intrinsic to eigenpair estimation rather than being inherited by a circular argument. The lower bound is a clean Fano construction on the flat torus with explicit packing and KL estimates, and the upper-bound analysis, based on H^{-1}(X_n) estimates, multiscale Poincaré inequalities, and a careful bias-variance decomposition, is a substantial technical advance over pointwise consistency arguments. The paper is also honest about its main limitation: the graph-Laplacian optimality statements require C^{2,α} regularity and thus do not cover the full C^2 density class, although the information-theoretic rate and the known-manifold plug-in upper bound do. I found the central claims sound and the presentation unusually careful about scoping assumptions.

minor comments (4)
  1. [Theorem 1.4 and Remark 1.12] The proof in §2.2 establishes the lower bound for fixed l, and the constants in Lemma 2.7 and Step 4 may depend on λ_l through the choices in (2.10)-(2.12). As written, the displayed lower bound c λ_l(T^d,1) n^{-2/(d+4)} and especially the 'in particular' statement c l^{2/d} n^{-2/(d+4)} in Theorem 1.4, together with the l-scaling comparison in Remark 1.12, are stronger than what the proof shows. Please either prove the stated l-dependence explicitly or reformulate the lower bound with a l-dependent constant.
  2. [Section 2 notation] The square-root-of-integral notation rendered as 'd∫' in equations such as (2.1) and (2.19) is nonstandard and is not defined in the notation list; please introduce explicit notation such as (∫ |·|^2)^{1/2} or define the symbol before first use.
  3. [Assumption 3 and Remark 1.7] The C^{2,α}-versus-C^2 gap is a real limitation of the graph-Laplacian upper bound, but it is explicitly acknowledged and is not an error: Propositions 3.8-3.10 require C^3 eigenfunctions, and the counterexample reference [27, Section 2.2] justifies the need for Assumption 3. For readers, it would be helpful to state in the introduction that the matching upper bound over the full C^2 class is the (manifold-dependent) plug-in estimator of Appendix C, so that the graph-Laplacian optimality claim is clearly understood to be conditional on the stronger regularity.
  4. [Theorem 1.6, display (1.20)] The high-probability bound contains the prefactor C n ε_n^{-d} exp(-c n ε_n^{d+4}); after substituting ε_n ~ (log n/n)^{1/(d+4)}, this tends to zero only when the constant c is larger than the exponent coming from the prefactor. It would be useful to state explicitly that c can be chosen to absorb the logarithmic factors, since the reader otherwise has to verify this from the constants.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the minimax lower bound and the graph-Laplacian upper bound are derived independently, and the C^{2,\alpha} regularity gap is an explicitly scoped limitation, not a circular step.

full rationale

The paper's central claims are self-contained rather than circular. The minimax lower bound (Theorem 1.4) is proved directly by Fano's method over an explicit packing of densities near the uniform density on the flat torus (Section 2.2). The eigenpair separation in Lemma 2.7 is obtained from first-order perturbation identities (2.24)-(2.25) and the analytic bounds (C.4) and (C.6), which are deterministic elliptic estimates and not the target minimax statement. The paper explicitly notes that the density-estimation lower bound does not by itself imply the eigenpair lower bound, so the rate n^{-2/(d+4)} is not imported by renaming a known result. The upper bound (Theorems 1.6 and 1.10) is established by direct bias-variance and concentration analysis: the Taylor expansion (3.10)-(3.11), the multiscale Poincare inequality (Proposition 3.6), the concentration bounds (Propositions 3.8-3.10), and the graph Poisson energy estimate (Proposition 3.11). The bandwidth epsilon_n ~ (log n / n)^{1/(d+4)} is chosen from the tradeoff between bias and variance in these estimates, not fitted to the lower bound or to data. The regularity assumption on C^{2,\alpha} densities (Assumption 3) is a genuine, openly acknowledged limitation (Remark 1.7, Remark 3.7) that affects the scope of the graph-Laplacian upper bound but does not make the argument circular. Citations to the authors' prior work, such as [4,6,15,28], are used for auxiliary a priori bounds, pointwise consistency facts, and proof inspiration; these are parameter-free results with stated assumptions and are not the unverified load-bearing premise of the main theorems. No step reduces, by the paper's own equations or by self-citation, to its own inputs.

Assumptions & free parameters 0 free parameters · 9 assumptions · 0 invented entities

The paper introduces no fitted parameters and no invented entities. The central claims depend on standard statistical and geometric assumptions: the manifold class M, the density class P_M, the spectral gap condition, the graph kernel and connectivity assumptions, and the stronger C^{2,alpha} regularity in Assumption 3 for the graph Laplacian upper bounds. Bandwidth choices such as epsilon_n ~ (log n/n)^{1/(d+4)} are deterministic tuning parameters chosen by the analysis, not fitted to data. All other inputs are standard mathematical background.

assumptions (9)
  • domain assumption Manifold class M: smooth, compact, orientable, connected, boundaryless, volume 1, bounded sectional curvature, bounded reach, lower injectivity radius, and Holder regularity of curvature and second fundamental form (Definition 1.1).
    Used throughout to localize the analysis via exponential maps, to control Euclidean versus geodesic distances, and to obtain uniform geometric constants in concentration and PDE estimates (Appendix A).
  • domain assumption Density class P_M: rho in C^2 with rho_min <= rho <= rho_max and bounds on first and second derivatives (Definition 1.2).
    This defines the statistical model and is the source of the n^{-2/(d+4)} rate, matching the L^2 density estimation rate for bounded second derivatives.
  • domain assumption Spectral gap condition: gamma_l >= gamma for the l-th eigenvalue (Definition 1.3 and equation 1.7).
    Needed for identifiability of the l-th eigenpair; the upper bound constants depend on 1/gamma, and the lower bound construction keeps the spectral gap bounded below.
  • domain assumption Assumption 1 on kernel eta: non-increasing, Lipschitz, supported on [0,1], eta(0)=1, eta(1)=0, eta(1/2)>0, and normalized so that the integral of eta(|x|) is 1.
    Defines the graph Laplacian and the extension operator; the support and Lipschitz properties are used in the bias and concentration estimates of Section 3.3.
  • domain assumption Assumption 2 on connectivity: epsilon_n lies above the connectivity threshold C(log n)^{1/d}/n^{1/d} and below a geometric scale factor.
    Ensures the random geometric graph is connected with high probability and that the local geometry is uniformly Euclidean at scale epsilon_n; used in all graph Laplacian upper bound results.
  • domain assumption Assumption 3: M in M with alpha > 0 and rho in P^{2,alpha}_M, so rho has C^{2,alpha} second derivatives.
    Ensures that eigenfunctions of Delta_rho are C^3, which is required for the third-order Taylor expansion and H^{-1} estimates in the proof of Theorems 1.6 and 1.10. The lower bound does not use this assumption.
  • standard math Standard elliptic regularity and spectral theory: existence and completeness of eigenpairs, C^3 regularity under C^{2,alpha} coefficients, and analytic perturbation theory for simple eigenvalues.
    Used in Section 1.1 for the spectral setup and in Appendix C for the plug-in density estimator benchmark.
  • standard math Fano's method and standard concentration inequalities: Bernstein for sums, U-statistic Bernstein bounds, and KL-based mutual information control.
    Used in Section 2 for the lower bound and in Sections 3.3 and 3.4 for the high-probability upper bounds.
  • standard math Transport coupling construction from [28]: existence of a density rho_n comparable to rho and a map T pushing rho_n to the empirical measure with displacement at most r.
    Used in the proof of Theorem 1.10 to establish the extension estimate (3.62); treated as a cited result from the prior literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Minimax Rates for the Estimation of Eigenpairs of Weighted Laplace-Beltrami Operators on Manifolds." pith.science (2026). https://pith.science/paper/PG2BF6B6

@misc{pith2026250600171,
  author       = {Pith},
  title        = {Pith review of: Minimax Rates for the Estimation of Eigenpairs of Weighted Laplace-Beltrami Operators on Manifolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PG2BF6B6}},
  note         = {Machine review of arXiv:2506.00171}
}
abstract

We study the problem of estimating eigenpairs of elliptic differential operators from samples of a distribution $\rho$ supported on a manifold $M$. The operators discussed in the paper are relevant in unsupervised learning and in particular are obtained by taking suitable scaling limits of widely used graph Laplacians over data clouds. We study the minimax risk for this eigenpair estimation problem and explore the rates of approximation that can be achieved by commonly used graph Laplacians built from random data. More concretely, assuming that $\rho$ belongs to a certain family of distributions with controlled second derivatives, and assuming that the $d$-dimensional manifold $M$ where $\rho$ is supported has bounded geometry, we prove that the statistical minimax rate for approximating eigenvalues and eigenvectors in the $H^1(M)$-sense is $n^{-2/(d+4)}$, a rate that matches the minimax rate for a closely related density estimation problem. We then revisit the literature studying Laplacians over proximity graphs in the large data limit and prove that, under slightly stronger regularity assumptions on the data generating model, eigenpairs of graph Laplacians induce manifold agnostic estimators with an error of approximation that, up to logarithmic corrections, matches our lower bounds. Our analysis allows us to expand the existing literature on graph-based learning in at least two significant ways: 1) we consider stronger norms to measure the error of approximation than the ones that had been analyzed in the past; 2) our rates of convergence are uniform over a family of smooth distributions and do not just apply to densities with special symmetries, and, as a consequence of our lower bounds, are essentially sharp when the connectivity of the graph is sufficiently high.

Figures

Figures reproduced from arXiv: 2506.00171 by the authors.

Figure 1
Figure 1. An illustration of a cube Mj with j P I. Points in this cube are sufficiently far away from the set of points where the gradient of the eigenfunction fl vanishes. The function aj is also illustrated in this figure. The red concentric circles denote the negative level sets of aj , i.e., the points around the point bj ` 1 m u´, while the green concentric circles represent the region in Mj where aj is positive. 1. The … view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. An indefinite Coulomb interaction from the Steklov spectrum of perforated manifolds

    math.AP 2026-07 conditional novelty 7.0 of 10

    Steklov eigenvalues of a manifold with many small critically-sized holes converge to weighted Laplace–Beltrami eigenvalues at the optimal rate, with the next-order term given by an indefinite Coulomb energy mediated b...

  2. Uniform Sobolev inequalities on geometric graphs

    math.AP 2026-07 conditional novelty 7.0 of 10

    On quasi-uniform geometric graphs, uniform L^q–L^p Sobolev inequalities hold exactly when ε_n |V_n|^{1/p − 1/q} is bounded, allowing ε_n to approach the connectivity threshold.

  3. On the convergence of graph Laplacians with a symmetric divergence

    stat.ML 2026-07 conditional novelty 6.0 of 10

    Graph Laplacians constructed from a smooth nondegenerate symmetric divergence D on a compact Riemannian manifold converge pointwise to the Laplace–Beltrami operator under a fourth-order closeness condition to squared ...

Reference graph

Works this paper leans on

67 extracted references · 60 canonical work pages · cited by 3 Pith papers

  1. [1]

    M. A. Arcones. A bernstein-type inequality for u-statistics and u-processes. Statistics & probability letters, 22(3):239–247, 1995. 71

    work page 1995

  2. [2]

    Armstrong and P

    S. Armstrong and P. Dario. Elliptic regularity and quantitative homogenization on percolation clusters. Comm. Pure Appl. Math. , 71(9):1717–1849, 2018

    work page 2018

  3. [3]

    Armstrong and J

    S. Armstrong and J. Lin. Optimal quantitative estimates in stochastic homogeniza- tion for elliptic equations in nondivergence form. Archive for Rational Mechanics and Analysis, 225:937–991, 2017

    work page 2017

  4. [4]

    Armstrong and R

    S. Armstrong and R. Venkatraman. Quantitative homogenization and large-scale reg- ularity of poisson point clouds, 2023

    work page 2023

  5. [5]

    Armstrong and R

    S. Armstrong and R. Venkatraman. Asymptotic expansion of the spectrum for periodic Schr¨ odinger operators.SIAM J. Math. Anal. , 56(2):1770–1808, 2024

    work page 2024

  6. [6]

    Armstrong and R

    S. Armstrong and R. Venkatraman. Optimal convergence rates for the spectrum of the graph laplacian on poisson point clouds. Foundations of Computational Mathematics , pages 1–26, 2025

    work page 2025

  7. [7]

    Belkin and P

    M. Belkin and P. Niyogi. Towards a theoretical foundation for Laplacian-based manifold methods. In International Conference on Computational Learning Theory , pages 486–

    work page

  8. [8]

    P. J. Bickel and Y. Ritov. Estimating integrated squared density derivatives: sharp best order of convergence estimates. Sankhy¯ a: The Indian Journal of Statistics, Series A, pages 381–393, 1988

    work page 1988

Show all 67 references
  1. [9]

    Birg´ e and P

    L. Birg´ e and P. Massart. Estimation of integral functionals of a density. The Annals of Statistics, 23(1):11–29, 1995

    1995

  2. [10]

    Bo and M

    W. Bo and M. Meil˘ a. How well behaved is finite dimensional diffusion maps? arXiv:2412.03992, 2024

    2024 arXiv

  3. [11]

    Bretagnolle and C

    J. Bretagnolle and C. Huber. Estimation des densit´ es: risque minimax. Zeitschrift f¨ ur Wahrscheinlichkeitstheorie und verwandte Gebiete , 47:119–137, 1979

    1979

  4. [12]

    Br´ ezis

    H. Br´ ezis. Functional analysis, Sobolev spaces and partial differential equations , vol- ume 2. Springer, 2011

    2011

  5. [13]

    Bungert, J

    L. Bungert, J. Calder, M. Mihailescu, K. Houssou, and A. Yuan. Convergence rates for poisson learning to a poisson equation with measure data. arXiv:2407.06783, 2024

    2024 arXiv

  6. [14]

    Burago, S

    D. Burago, S. Ivanov, and Y. Kurylev. A graph discretization of the Laplace-Beltrami operator. Journal of Spectral Theory, 4(4):675–714, 2014

    2014

  7. [15]

    Calder and N

    J. Calder and N. Garc´ ıa Trillos. Improved spectral convergence rates for graph lapla- cians on ε-graphs and k-nn graphs. Applied and Computational Harmonic Analysis , 60:123–175, 2022

    2022

  8. [16]

    Calder, N

    J. Calder, N. Garc´ ıa Trillos, and M. Lewicka. Lipschitz regularity of graph laplacians on random data clouds. SIAM Journal on Mathematical Analysis , 54(1):1169–1222, 2022. 72

    2022

  9. [17]

    N. N. Cencov. Evaluation of an unknown distribution density from observations. Soviet Math., 3:1559–1562, 1962

    1962

  10. [18]

    N. N. Cencov. Statistical decision rules and optimal inference. American Mathematical Soc., 2000

    2000

  11. [19]

    Cheng and N

    X. Cheng and N. Wu. Eigen-convergence of gaussian kernelized graph laplacian by manifold heat interpolation. Applied and Computational Harmonic Analysis , 61:132– 190, 2022

    2022

  12. [20]

    F. Chung. Four proofs for the cheeger inequality and graph partition algorithms. In Proceedings of ICCM, volume 2, page 378. Citeseer, 2007

    2007

  13. [21]

    R. R. Coifman and S. Lafon. Diffusion maps. Applied and computational harmonic analysis, 21(1):5–30, 2006

    2006

  14. [22]

    V. Divol. Measure estimation on manifolds: an optimal transport approach. Probability Theory and Related Fields , 183(1):581–647, 2022

    2022

  15. [23]

    M. P. Do Carmo and F. Flaherty. Riemannian geometry, volume 6. Springer, 1992

    1992

  16. [24]

    D. B. Dunson, H.-T. Wu, and N. Wu. Spectral convergence of graph laplacian and heat kernel reconstruction in l8 from random samples. Applied and Computational Harmonic Analysis, 55:282–336, 2021

    2021

  17. [25]

    L. C. Evans. Partial differential equations , volume 19 of Graduate Studies in Mathe- matics. American Mathematical Society, Providence, RI, 1998

    1998

  18. [26]

    R. H. Farrell. On the best obtainable asymptotic rates of convergence in estimation of a density function at a point. The Annals of Mathematical Statistics , pages 170–180, 1972

    1972

  19. [27]

    Fern´ andez-Real and X

    X. Fern´ andez-Real and X. Ros-Oton.Regularity Theory for Elliptic PDE . EMS Press, Dec. 2022

    2022

  20. [28]

    Garc´ ıa Trillos, M

    N. Garc´ ıa Trillos, M. Gerlach, M. Hein, and D. Slepˇ cev. Error estimates for spectral convergence of the graph laplacian on random geometric graphs toward the laplace– beltrami operator. Foundations of Computational Mathematics , pages 1–61, 2019

    2019

  21. [29]

    Garc´ ıa Trillos, P

    N. Garc´ ıa Trillos, P. He, and C. Li. Large sample spectral analysis of graph-based multi-manifold clustering. Journal of Machine Learning Research, 24(143):1–71, 2023

    2023

  22. [30]

    Garc´ ıa Trillos, Z

    N. Garc´ ıa Trillos, Z. Kaplan, T. Samakhoana, and D. Sanz-Alonso. On the consis- tency of graph-based bayesian semi-supervised learning and the scalability of sampling algorithms. Journal of Machine Learning Research , 21(28):1–47, 2020

    2020

  23. [31]

    Garc´ ıa Trillos, A

    N. Garc´ ıa Trillos, A. Little, D. McKenzie, and J. M. Murphy. Fermat distances: Metric approximation, spectral convergence, and clustering algorithms. Journal of Machine Learning Research, 25(176):1–65, 2024. 73

    2024

  24. [32]

    Garc´ ıa Trillos and D

    N. Garc´ ıa Trillos and D. Slepˇ cev. A variational approach to the consistency of spectral clustering. Applied and Computational Harmonic Analysis , 45(2):239–281, 2018

    2018

  25. [33]

    Garc´ ıa Trillos and M

    N. Garc´ ıa Trillos and M. Weber. Continuum limits of ollivier’s ricci curvature on data clouds: pointwise consistency and global lower bounds. arXiv preprint arXiv:2307.02378, 2023

    2023 arXiv

  26. [34]

    Garc´ ıa Trillos, R

    N. Garc´ ıa Trillos, R. Murray, and M. Thorpe. Rates of convergence for regression with the graph poly-laplacian. Sampling Theory, Signal Processing, and Data Analysis , 21(2), Nov. 2023

    2023

  27. [35]

    C. R. Genovese, M. Perone Pacifico, V. Isabella, L. Wasserman, et al. Minimax mani- fold estimation. Journal of machine learning research , 13:1263–1291, 2012

    2012

  28. [36]

    Gin´ e, R

    E. Gin´ e, R. Latala, and J. Zinn. Exponential and moment inequalities for u-statistics. In High Dimensional Probability II , pages 13–38. Birkh¨ auser Boston, 2000

    2000

  29. [37]

    Green, S

    A. Green, S. Balakrishnan, and R. J. Tibshirani. Minimax optimal regression over sobolev spaces via laplacian eigenmaps on neighbourhood graphs. Information and Inference: A Journal of the IMA , 12(3):2423–2502, 2023

    2023

  30. [38]

    J. Z. HaoChen, C. Wei, A. Gaidon, and T. Ma. Provable guarantees for self-supervised deep learning with spectral contrastive loss.Advances in Neural Information Processing Systems, 34, 2021

    2021

  31. [39]

    Hein, J.-Y

    M. Hein, J.-Y. Audibert, and U. v. Luxburg. Graph laplacians and their convergence on random neighborhood graphs. Journal of Machine Learning Research, 8(Jun):1325– 1368, 2007

    2007

  32. [40]

    Hoffmann, B

    F. Hoffmann, B. Hosseini, A. A. Oberai, and A. M. Stuart. Spectral analysis of weighted laplacians arising in data clustering. Applied and Computational Harmonic Analysis , 56:189–249, 2022

    2022

  33. [41]

    Hoffmann, B

    F. Hoffmann, B. Hosseini, Z. Ren, and A. M. Stuart. Consistency of semi-supervised learning algorithms on graphs: Probit and one-hot methods. The Journal of Machine Learning Research, 21(1):7549–7603, 2020

    2020

  34. [42]

    I. A. Ibragimov and R. Z. Hasminskii. Statistical estimation: asymptotic theory , vol- ume 16. Springer Science & Business Media, 2013

    2013

  35. [43]

    T. Kato. Perturbation theory for linear operators, volume Band 132 of Die Grundlehren der mathematischen Wissenschaften . Springer-Verlag New York Inc., 1966

    1966

  36. [44]

    Kenig, F

    C. Kenig, F. Lin, and Z. Shen. Estimates of eigenvalues and eigenfunctions in periodic homogenization. Journal of the European Mathematical Society, 15(5):1901–1925, 2013

    1901

  37. [45]

    Kerkyacharian and D

    G. Kerkyacharian and D. Picard. Density estimation in besov spaces. Statistics & probability letters, 13:15–24, 1992

    1992

  38. [46]

    R. Z. Khasminskii. A lower bound on the risks of non-parametric estimates of densities in the uniform metric. Theory of Probability & Its Applications , 23(4):794–798, 1979. 74

    1979

  39. [47]

    A. K. Kim and H. H. Zhou. Tight minimax rates for manifold estimation under haus- dorff loss. Electronic Journal of Statistics , 9:1562–1582, 2015

    2015

  40. [48]

    J. Kim, A. Rinaldo, and L. Wasserman. Minimax rates for estimating the dimension of a manifold. Journal of Computational Geometry , 10(1), 2019

    2019

  41. [49]

    S. J. Koelle, H. Zhang, O.-V. Murad, and M. Meila. Consistency of dictionary-based manifold learning. In International Conference on Artificial Intelligence and Statistics , pages 4348–4356. PMLR, 2024

    2024

  42. [50]

    C. Li, R. Sonthalia, and N. Garc´ ıa Trillos. Spectral neural networks: Approximation theory and optimization landscape. arXiv:2310.00729, 2023

    2023 arXiv

  43. [51]

    Li and N

    H. Li and N. Saito. Metrics of graph laplacian eigenvectors. In Wavelets and Sparsity XVIII, volume 11138, pages 455–472. SPIE, 2019

    2019

  44. [52]

    J. Lu. Graph approximations to the laplacian spectra. Journal of Topology and Anal- ysis, 14(01):111–145, 2022

    2022

  45. [53]

    A. M. Neuman. Graph laplacians on shared nearest neighbor graphs and graph laplacians on k-nearest neighbor graphs having the same limit. arXiv preprint arXiv:2302.12399, 2023

    2023 arXiv

  46. [54]

    J. W. Peoples and J. Harlim. Spectral convergence of symmetrized graph laplacian on manifolds with boundary. arXiv preprint arXiv:2110.06988 , 2021

    2021 arXiv

  47. [55]

    Rayleigh

    L. Rayleigh. Lvi. on the influence of obstacles arranged in rectangular order upon the properties of a medium. The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science , 34(211):481–502, 1892

  48. [56]

    A. Singer. From graph to manifold laplacian: The convergence rate. Applied and Computational Harmonic Analysis , 21(1):128–134, 2006

    2006

  49. [57]

    C. D. Sogge. Riemannian manifolds with maximal eigenfunction growth. S´ eminaire ´Equations aux d´ eriv´ ees partielles (Polytechnique) dit aussi” S´ eminaire Goulaouic- Schwartz”, pages 1–16, 2001

    2001

  50. [58]

    C. J. Stone. Optimal rates of convergence for nonparametric estimators. The annals of Statistics, pages 1348–1360, 1980

    1980

  51. [59]

    Tan and X

    Y. Tan and X. Cheng. Improved convergence rate of knn graph laplacians. arXiv preprint arXiv:2410.23212, 2024

    2024 arXiv

  52. [60]

    Tao and Z

    W. Tao and Z. Shi. Convergence of laplacian spectra from random samples. Journal of Computational Mathematics , 38(6):952–984, 2020

    2020

  53. [61]

    A. B. Tsybakov and A. B. Tsybakov. Nonparametric estimators. Introduction to Nonparametric Estimation, pages 1–76, 2009

    2009

  54. [62]

    Von Luxburg

    U. Von Luxburg. A tutorial on spectral clustering. Statistics and computing, 17(4):395– 416, 2007. 75

    2007

  55. [63]

    V. Q. Vu and J. Lei. Minimax sparse principal subspace estimation in high dimensions. The Annals of Statistics , 41(6):2905–2947, 2013

    2013

  56. [64]

    M. Wahl. A kernel-based analysis of laplacian eigenmaps. arXiv preprint arXiv:2402.16481, 2024

    2024 arXiv

  57. [65]

    M. J. Wainwright. High-dimensional statistics: A non-asymptotic viewpoint, volume 48. Cambridge university press, 2019

    2019

  58. [66]

    C. L. Wormell and S. Reich. Spectral convergence of diffusion maps: Improved er- ror bounds and an alternative normalization. SIAM Journal on Numerical Analysis , 59(3):1687–1734, 2021

    2021

  59. [67]

    intrinsic curvature

    H.-T. Wu and N. Wu. When locally linear embedding hits boundary. Journal of Machine Learning Research, 24(69):1–80, 2023. A. Background on Riemannian Geometry A.1 Exponential Map and Normal Coordinates Let xP M. The exponential map expx at the point x is the map exp x : TxMÑ M...

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.