Pith. sign in

REVIEW 1 major objections 5 minor 22 references

Spectral convergence of graph Laplacians with Ricci curvature bounds and in non-collapsed Ricci limit spaces

T0 review · 1 major / 5 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 establishes quantitative, high-probability bounds on the eigenvalues and eigenfunctions of epsilon-neighborhood graph Laplacians constructed from i.i.d.

desk verdict Solid manifold result; the limit-space extension as stated is not supported because the error term need not vanish under the stated Cheeger–Colding hypotheses. read the letter →

arxiv 2506.07427 v2 pith:N6D5PCPY submitted 2025-06-09 math.DG math.MGmath.SP

Masato Inagaki This is my paper
classification math.DGmath.MGmath.SP MSC 58J5053C2153C23
keywords graphLaplacianspectralconvergenceeigenvalueboundsRiccicurvaturelowerboundnon-collapsedlimitspacesmeasuredGromov-Hausdorffrandomgeometricgraphshigh-probabilityestimates
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

The paper proves that, for a fixed eigenvalue index k, the k-th eigenvalue of a normalized random-walk graph Laplacian built from n i.i.d. points approximates the k-th eigenvalue of the weighted Laplacian on the underlying manifold, with an explicit error that converges to zero as the number of points grows. The result holds under only Ricci curvature lower bounds, a volume lower bound, and a diameter upper bound, without the sectional-curvature or injectivity-radius assumptions used in earlier work. The same approximation holds verbatim on non-collapsed Ricci limit spaces, which are metric limits of such manifolds and may have mild singularities. The error is controlled by two integrals: one measuring in L^p the deviation of small geodesic balls from the constant-curvature model, and one measuring how much the ambient Euclidean distance distorts geodesic balls. A sympathetic reader should care because this significantly widens the class of data manifolds for which Laplacian Eigenmaps and similar spectral methods are provably accurate.

What carries the argument

Two comparison maps carry the argument. A discretization map bounds the graph eigenvalues from above by manifold Rayleigh quotients, and an interpolation map Lambda_epsilon pulls discrete functions back to Lipschitz functions on the manifold and bounds the manifold eigenvalues from above. The key innovation is that the interpolation kernel is controlled only integrally, through sums of its gradients over the sample points, rather than pointwise, which avoids the need for sectional-curvature or injectivity-radius bounds. On the discrete side, the eigenfunctions are controlled by a rough Nash-type inequality derived from a rough volume-doubling property and a Poincare inequality that hold with high probability on random epsilon-neighborhood graphs at scales above epsilon. The integrals V_{p,epsilon}(M) and S_epsilon(M) enter precisely where earlier works used pointwise comparison estimates.

What would settle it

Exhibit a non-collapsed Ricci limit space for which V_{p,epsilon}(M) $epsilon^{{-2/p}}$ does not tend to zero for some p > 2; then the right-hand side of the main estimate would not vanish and the claimed spectral convergence would fail on that space. No such space is currently known, so this is a hypothetical counterexample rather than an observed one.

Watch

Extended reading notes

Core claim

The central claim is a quantitative comparison between the continuous weighted Laplacian $\Delta$^N_rho and the discrete graph Laplacian L_n: for each fixed k, with probability at least 1 - C $n^{{-beta}}$ (provided epsilon $\sqrt$($\beta$) C <= 1), the difference |lambda_k($\Delta$^N_rho) - (m+2) lambda_k(L_n)| is bounded by C times ($epsilon^{{m/(m+2)}}$ + V_{m+2,epsilon}(M) $epsilon^{{-2/(m+2)}}$ + S_epsilon(M) $epsilon^{{-m}}$), where V_{p,epsilon}(M) and S_epsilon(M) are the two distortion integrals. A corresponding $L^{2}$ approximation of eigenfunctions is also proved. The same bounds hold when the samples are drawn from the measure on a non-collapsed Ricci limit space that is the limit of manifolds satisfying the Ricci, diameter, and volume conditions. The error terms V and S vanish in the limit for the manifolds under consideration, so the approximation error tends to zero as n goes to infinity.

Load-bearing premise

The load-bearing premise is that, on the manifold or Ricci limit space, the two distortion integrals V_{p,epsilon}(M) and S_epsilon(M) tend to zero as epsilon goes to zero; for Ricci limit spaces this requires that almost every point has a Euclidean tangent cone so that small balls have nearly Euclidean volume, a property that is imported from the Cheeger-Colding regularity theory but is not proved in this paper.

Editorial extensions

If this is right

  • For data sampled from a density rho on a closed m-manifold with Ric >= -(m-1)K, bounded diameter, and bounded-below volume, the eigenvalues of the normalized epsilon-neighborhood graph Laplacian approximate those of Delta^N_rho to order epsilon^{m/(m+2)} plus the two distortion terms, with failure probability at most C n^{-beta}.
  • Choosing epsilon = (log n / n)^{1/(m+2)} turns the bound into an explicit high-probability rate in the sample size n.
  • Data manifolds with mild singularities that arise as non-collapsed Ricci limit spaces, such as the spindle (0,pi) x S^{m-1} for m >= 3, are covered even though they do not have a smooth Riemannian metric.
  • Eigenfunctions and eigenspaces are also recovered in L^2, with error controlled by the same distortion terms and by the spectral gaps of the limit Laplacian.
  • The previous dependence on sectional-curvature bounds and positive injectivity radius is removed; the rate is slower than the previous O(epsilon) rate but the assumptions are weaker.

Reading between the lines

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

  • Since V_{p,epsilon}(M) and S_epsilon(M) are defined purely from intrinsic and ambient metric data, they could in principle be estimated from the data cloud, giving data-dependent error certificates that do not require knowing the Ricci bound in advance.
  • The discrete estimates only use rough volume-doubling and Poincare inequalities, so the same proof route should transfer to other graph kernels (k-NN graphs, heat-kernel weights) and possibly to nonlinear operators such as graph p-Laplacians.
  • The dimension-two spindle case, where the theorems give uniform but not vanishing bounds, suggests that the integral V_{p,epsilon} epsilon^{-2/p} is a sharp obstacle; if a dimension-two example had vanishing V terms, convergence might still hold, but that is not proved here.
  • Testing the bound on a flat cone (a non-collapsed Ricci limit space with a point singularity) would be a concrete stress test: the theory predicts the distortion integrals vanish and the bound holds, so any violation would point to a missing geometric hypothesis.
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

1 major / 5 minor

Summary. The paper proves quantitative, high-probability bounds for the eigenvalues and eigenfunctions of epsilon-neighborhood graph Laplacians built from i.i.d. samples on closed Riemannian manifolds with a uniform lower Ricci curvature bound, a diameter bound, and a volume lower bound. The main manifold results, Theorems 7.1 and 7.2, compare the weighted Laplacians Δ_ρ and Δ^N_ρ with the unnormalized and random-walk graph Laplacians Γ_{m,ε} and Γ^N_ε, with error terms expressed through the distortion integrals V_{p,ε}(M) and S_ε(M). The proof combines discrete L^p estimates under rough volume doubling and Poincaré inequalities (Section 3), high-probability regularity of data graphs (Section 4), and matching lower and upper eigenvalue comparisons via discretization and interpolation maps (Sections 5 and 6). Theorems 7.3 and 7.4 extend the statements to non-collapsed Ricci limit spaces using Cheeger-Colding regularity. The quantitative manifold results are the central contribution of the paper.

Significance. If the manifold results are correct, they represent a substantial generalization of earlier graph-Laplacian convergence results, replacing assumptions such as an upper sectional-curvature bound and an injectivity-radius lower bound with integral control of small-ball volume distortion. The discrete Moser iteration under rough volume doubling, the systematic use of the integrals V_{p,ε} and S_ε, and the explicit high-probability bounds are valuable and should be of interest to both the spectral geometry and manifold-learning communities. The paper is careful and detailed, and the main manifold theorems are supported by a coherent chain of estimates. However, the claimed extension to all non-collapsed Ricci limit spaces, presented in the abstract and in Section 7, is not justified by the error terms appearing in Theorem 7.3.

major comments (1)
  1. [Section 7, Theorem 7.3, Eq. (7.8)] The error term δ_{p,ε,a}(M,d_M,\tilde d) contains V_{p,ε}(M)ε^{-2/p}, and for the convergence claim with p=m+2 one needs V_{m+2,ε}(M)ε^{-2/(m+2)} to tend to zero. The Cheeger-Colding regularity imported through Definition 2.6 and Theorem 2.7 gives only H^m-almost-everywhere Euclidean tangent cones, hence pointwise convergence of vol(B(x,ε))/V_K(ε) to 1 and, by dominated convergence, V_{m+2,ε}(M)→0. It provides no rate. Almost-everywhere convergence is compatible with arbitrarily slow decay, e.g. V_{m+2,ε}∼1/|log ε|, for which the product with ε^{-2/(m+2)} diverges. More concretely, on a non-collapsed Ricci limit with a codimension-2 singular stratum modeled on R^{m-2} times a two-dimensional cone of angle <2π, the set of points whose ball-volume ratio is bounded away from 1 has H^m-measure ∼ε², giving V_{m+2,ε}≥cε^{2/(m+2)} and hence V_{m+2,ε}ε^{-2/(m+2)}≥c>0. Thus the right-hand side of (7.8) need not tend to zero, and Theorem 7.3 does not establish spectral convergence for general non-collapsed Ricci limit spaces. The same issue invalidates the convergence interpretation of Theorem 7.4. The theorem remains meaningful as a conditional quantitative bound, but the abstract and introduction overstate the conclusion. The authors should either add explicit hypotheses ensuring V_{p,ε}ε^{-2/p}+S_εε^{-m}→0, prove quantitative singular-set estimates, or restrict the claimed convergence accordingly.
minor comments (5)
  1. [Theorem 1.2] The definition of γ contains the term λ_{k+1}(Δ^N_ρ)−λ_{k+1}(Δ^N_ρ), which is identically zero and forces γ=0; the intended expression is presumably λ_{k+1}(Δ^N_ρ)−λ_k(Δ^N_ρ), matching the convention in Theorem 7.2.
  2. [Theorem 6.8] In statement (ii), the phrase 'with the same probability bound as in(ii)' should refer to part (i), not to itself.
  3. [Definition 2.6] The notation '{ρ_t ⊂ P(M_t:α,L,H)}' in Definition 2.6 and later displays should read ρ_t∈P(M_t:α,L,H), since ρ_t is an element of the class, not a subset.
  4. [Section 5, Theorem 5.7] The application of Lemma 4.1 with \tilde ε=ε/24 and \tilde a=24^{m/2}a requires the condition \tilde a A≤1; this restriction should be stated explicitly among the hypotheses of Theorem 5.7 rather than left implicit.
  5. [Lemma 6.2] In the proof of Lemma 6.2, the Borel frame fields are indexed as e_1,...,e_m, but the sentence 'for all i∈{1,...,n}' should refer to the orthonormal frame index, not the sample index; this is a typographical issue but should be corrected for readability.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: continuous and discrete spectral objects are defined independently, and the error bounds are obtained by Rayleigh-quotient comparisons and concentration inequalities.

full rationale

The paper's central claims compare the eigenvalues of an independently defined weighted Laplacian Delta^N_rho (or Delta_rho) with those of graph Laplacians constructed from i.i.d. samples. The discrete objects (Definitions 2.15) and the continuous operators (Section 2.1) are defined without reference to each other, and no parameter is fitted to the target eigenvalues. The error terms V_{p,epsilon}(M) and S_epsilon(M) (Definitions 2.8 and 2.11) are geometric distortion integrals measuring deviation from the constant-curvature model and metric distortion between d_g and d_R^d; they are inputs quantifying the geometry, not fitted quantities. The proof chain is: Section 3 derives L^p estimates for graph eigenfunctions under rough volume-doubling and Poincare inequalities; Section 4 proves these properties hold for the random data graphs with high probability via concentration estimates; Sections 5 and 6 establish matching lower and upper eigenvalue bounds by comparing Rayleigh quotients through discretization and interpolation maps; Section 7 combines these two-sided bounds. Each step is a genuine comparison of independently defined Rayleigh quotients, and the claimed inequalities do not reduce to their conclusions by construction. For the limit-space theorems, the proof applies the manifold theorem to approximating manifolds and passes to the limit using Cheeger-Colding regularity [12,13] and the independently proven spectral convergence Theorem 2.7; the cited results are external, not self-citations, and the theorem is not invoked to define the target quantity. The paper itself notes that for m=2 the limit-space theorems do not guarantee convergence but only give uniform bounds, which is an honest limitation rather than a circular step. A possible quantitative concern about whether V_{m+2,epsilon}(M) epsilon^{-2/(m+2)} tends to zero without a rate is a correctness or convergence-rate issue, not a circularity, because the error term is defined independently of the eigenvalues and no equation in the paper identifies the claimed bound with its hypotheses.

Assumptions & free parameters 0 free parameters · 6 assumptions · 1 invented entities

The central claim rests on standard Riemannian geometry comparison theorems, Cheeger-Colding regularity theory for non-collapsed Ricci limit spaces, and the explicit density regularity assumption. No free parameters are fitted to data; the constants in the bounds depend only on the fixed geometric parameters and k.

assumptions (6)
  • standard math Bishop-Gromov volume comparison for manifolds with Ric_g >= -(m-1)K (Theorem 2.1)
    Used throughout Sections 4-6 to bound volumes of balls and to justify discretization and interpolation estimates.
  • standard math Spectral stability of weighted Laplacians under measured Gromov-Hausdorff convergence on non-collapsed Ricci limit spaces (Theorem 2.7, from Cheeger-Colding [12,13])
    Provides the operator Delta rho on the limit space and convergence of eigenvalues and eigenfunctions; load-bearing for Theorems 7.3 and 7.4.
  • domain assumption On a non-collapsed Ricci limit space, H^m-almost every point has a Euclidean tangent cone, so vol(B(x,epsilon))/V_K(epsilon) tends to 1 for almost every x
    Imported from Cheeger-Colding; this is the weakest load-bearing regularity assumption for the limit-space theorems and ensures V_{p,epsilon}(M) tends to 0.
  • standard math Sobolev and Nash inequalities on manifolds with lower Ricci bounds (Lemma A.1, citing Saloff-Coste [21])
    Used to obtain L-infinity and gradient bounds for manifold eigenfunctions in Appendix A.
  • standard math Cheng's eigenvalue comparison gives lambda_k(Delta_g) <= C(m,K,D,alpha,k) (Remark 2.4, citing [14])
    Boundedness of continuous eigenvalues is used in Theorem 5.7 and Remark 5.6.
  • domain assumption The sampling density rho lies in P(M:alpha,L,H), meaning it is C^2, L-Lipschitz, with max/min ratio at most alpha and Hess(log rho) <= H
    Used for the L-infinity and Lipschitz bounds of manifold eigenfunctions in Appendix A, which feed into Theorems 5.5 and 7.2.
invented entities (1)
  • None
    purpose: No new particles, forces, dimensions, or mediators are introduced.
    The paper is a purely mathematical comparison theorem; the only new objects are defined integrals V_{p,epsilon} and S_epsilon that measure geometric distortion, not physical or invented entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spectral convergence of graph Laplacians with Ricci curvature bounds and in non-collapsed Ricci limit spaces." pith.science (2026). https://pith.science/paper/N6D5PCPY

@misc{pith2026250607427,
  author       = {Pith},
  title        = {Pith review of: Spectral convergence of graph Laplacians with Ricci curvature bounds and in non-collapsed Ricci limit spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N6D5PCPY}},
  note         = {Machine review of arXiv:2506.07427}
}
abstract

This paper establishes quantitative high-probability bounds on the eigenvalues and eigenfunctions of $\epsilon$-neighborhood graph Laplacians constructed from i.i.d. random variables on $m$-dimensional closed Riemannian manifolds $(M,g)$ that satisfy a uniform lower Ricci curvature bound $\operatorname{Ric}_g\ge -(m-1)K$, a positive lower volume bound, and an upper diameter bound. These results extend to non-collapsed Ricci limit spaces that are measured Gromov-Hausdorff limits of such manifolds, and the bounds give a spectral approximation of weighted Laplacians on manifolds with non-smooth points.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages

  1. [1]

    Geom.72(2024), no

    Masayuki Aino,Convergence of Laplacian eigenmaps and its rate for submanifolds with singularities, Dis- crete Comput. Geom.72(2024), no. 3, 1086–1168. MR4804972

    work page 2024

  2. [2]

    Portegies, and David Tewodrose,Embedding ofRCD ∗(K, N) spaces inL 2 via eigenfunctions, J

    Luigi Ambrosio, Shouhei Honda, Jacobus W. Portegies, and David Tewodrose,Embedding ofRCD ∗(K, N) spaces inL 2 via eigenfunctions, J. Funct. Anal.280(2021), no. 10, Paper No. 108968, 72. MR4224838

    work page 2021

  3. [3]

    Jonathan Bates,The embedding dimension of Laplacian eigenfunction maps, Appl. Comput. Harmon. Anal. 37(2014), no. 3, 516–530. MR3256785

    work page 2014

  4. [4]

    Mikhail Belkin and Partha Niyogi,Laplacian eigenmaps and spectral techniques for embedding and clustering, Advances in neural information processing systems, 2001

    work page 2001

  5. [5]

    ,Convergence of laplacian eigenmaps, Proceedings of the 20th international conference on neural information processing systems, 2006, pp. 129–136

    work page 2006

  6. [6]

    B´ erard, G

    P. B´ erard, G. Besson, and S. Gallot,Embedding Riemannian manifolds by their heat kernel, Geom. Funct. Anal.4(1994), no. 4, 373–398. MR1280119

    work page 1994

  7. [7]

    Besse,Manifolds all of whose geodesics are closed, Ergebnisse der Mathematik und ihrer Grenzge- biete [Results in Mathematics and Related Areas], vol

    Arthur L. Besse,Manifolds all of whose geodesics are closed, Ergebnisse der Mathematik und ihrer Grenzge- biete [Results in Mathematics and Related Areas], vol. 93, Springer-Verlag, Berlin-New York, 1978. With ap- pendices by D. B. A. Epstein, J.-P. Bourguignon, L. B´ erard-Bergery, M. Berger and J. L. Kazdan. MR496885

    work page 1978

  8. [8]

    33, American Mathematical Society, Providence, RI, 2001

    Dmitri Burago, Yuri Burago, and Sergei Ivanov,A course in metric geometry, Graduate Studies in Mathe- matics, vol. 33, American Mathematical Society, Providence, RI, 2001. MR1835418

    work page 2001

Show all 22 references
  1. [9]

    Dmitri Burago, Sergei Ivanov, and Yaroslav Kurylev,A graph discretization of the Laplace-Beltrami operator, J. Spectr. Theory4(2014), no. 4, 675–714. MR3299811

    2014

  2. [10]

    Jeff Calder and Nicol´ as Garc ´ ıa Trillos,Improved spectral convergence rates for graph Laplacians onε-graphs andk-NN graphs, Appl. Comput. Harmon. Anal.60(2022), 123–175. MR4393800

    2022

  3. [11]

    Jeff Calder, Nicol´ as Garc ´ ıa Trillos, and Marta Lewicka,Lipschitz regularity of graph Laplacians on random data clouds, SIAM J. Math. Anal.54(2022), no. 1, 1169–1222. MR4384039

    2022

  4. [12]

    Colding,On the structure of spaces with Ricci curvature bounded below

    Jeff Cheeger and Tobias H. Colding,On the structure of spaces with Ricci curvature bounded below. I, J. Differential Geom.46(1997), no. 3, 406–480. MR1484888

    1997

  5. [13]

    ,On the structure of spaces with Ricci curvature bounded below. III, J. Differential Geom.54(2000), no. 1, 37–74. MR1815411

    2000

  6. [14]

    Z.143(1975), no

    Shiu Yuen Cheng,Eigenvalue comparison theorems and its geometric applications, Math. Z.143(1975), no. 3, 289–297. MR378001

    1975

  7. [15]

    Mathav Murugan and Laurent Saloff-Coste,Harnack inequalities and Gaussian estimates for random walks on metric measure spaces, Electron. J. Probab.28(2023), Paper No. 64, 81. MR4585410 42

    2023

  8. [16]

    Coifman, and Ioannis G

    Boaz Nadler, St´ ephane Lafon, Ronald R. Coifman, and Ioannis G. Kevrekidis,Difusion maps, spectral clus- tering and reaction coordinates of dynamical systems, Appl. Comput. Harmon. Anal.21(2006), no. 1, 113–

    2006

  9. [17]

    Yukio Otsu,Laplacan of random nets of alexandrov space, preprint (2000)

    2000

  10. [18]

    171, Springer, Cham,

    Peter Petersen,Riemannian geometry, Third, Graduate Texts in Mathematics, vol. 171, Springer, Cham,

  11. [19]

    Portegies,Embeddings of Riemannian manifolds with heat kernels and eigenfunctions, Comm

    Jacobus W. Portegies,Embeddings of Riemannian manifolds with heat kernels and eigenfunctions, Comm. Pure Appl. Math.69(2016), no. 3, 478–518. MR3455592

    2016

  12. [20]

    149, American Mathe- matical Society, Providence, RI, 1996

    Takashi Sakai,Riemannian geometry, Translations of Mathematical Monographs, vol. 149, American Mathe- matical Society, Providence, RI, 1996. Translated from the 1992 Japanese original by the author. MR1390760

    1996

  13. [21]

    Saloff-Coste,A note on Poincar´ e, Sobolev, and Harnack inequalities, Internat

    L. Saloff-Coste,A note on Poincar´ e, Sobolev, and Harnack inequalities, Internat. Math. Res. Notices2 (1992), 27–38. MR1150597

    1992

  14. [22]

    China Ser

    Di Zhao,Eigenvalue estimate on a compact Riemann manifold, Sci. China Ser. A42(1999), no. 9, 897–904. MR1736580 43

    1999

Pith tools

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