pith. sign in

arxiv: 2604.18402 · v1 · submitted 2026-04-20 · 📊 stat.ML · math.DS

Adaptive Kernel Selection for Kernelized Diffusion Maps

Othmane Aboussaad , Adam Miraoui , Boumediene Hamzi , Houman Owhadi This is my paper

Pith reviewed 2026-05-10 04:00 UTC · model grok-4.3

classification 📊 stat.ML math.DS
keywords adaptive kernel selectionkernelized diffusion mapsspectral methodscross-validationvariational optimizationrandom Fourier featureseigenfunction recovery
0
0 comments X

The pith

Adaptive kernel selection for Kernelized Diffusion Maps improves eigenfunction recovery with variational and cross-validation methods.

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

The paper addresses the challenge of choosing suitable kernels in Kernelized Diffusion Maps, where the kernel controls the accuracy of the RKHS estimator for a diffusion-type operator and the stability of recovered eigenfunctions. It proposes two approaches: a variational outer loop that optimizes continuous kernel parameters like bandwidths and mixture weights by differentiating through the Cholesky-reduced eigenproblem, and an unsupervised cross-validation pipeline that selects kernel families and bandwidths via an eigenvalue-sum criterion with random Fourier features for scalability. Both rest on a shared theory proving Lipschitz dependence of KDM operators on kernel weights, continuity of spectral projectors when a gap condition holds, a residual-control theorem for proximity to the target eigenspace, and exponential consistency of the cross-validation selector over a finite dictionary of kernels.

Core claim

The central claim is that adaptive kernel selection can be performed reliably for Kernelized Diffusion Maps through either a differentiable variational optimization of kernel parameters or an eigenvalue-based cross-validation procedure, both backed by proofs of Lipschitz continuity with respect to kernel weights, continuity of projectors under spectral gaps, residual control to the desired eigenspace, and exponential consistency of the selector on finite kernel dictionaries.

What carries the argument

The variational outer loop that learns kernel parameters (bandwidths and mixture weights) by differentiating through the Cholesky-reduced KDM eigenproblem using an objective of eigenvalue maximization, subspace orthonormality, and RKHS regularization; together with the unsupervised cross-validation that uses an eigenvalue-sum criterion and random Fourier features.

If this is right

  • The recovered eigenfunctions become more accurate and stable because the kernel is tuned to maximize the relevant eigenvalues while preserving orthonormality.
  • The cross-validation procedure scales to large dictionaries by replacing exact kernel evaluations with random Fourier features.
  • Exponential consistency guarantees that, for sufficiently large samples, the selector will pick a kernel whose KDM operator is arbitrarily close to the best kernel in the dictionary with high probability.
  • The Lipschitz dependence result allows small changes in kernel weights to produce only bounded changes in the diffusion operator, supporting stable optimization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same variational differentiation trick could be applied to other kernel spectral methods such as kernel PCA or Laplacian eigenmaps to automate bandwidth selection.
  • When the gap condition fails on a given dataset, one could fall back to the residual-control theorem to still certify that the obtained subspace is close to some nearby eigenspace even if it is not the exact target one.
  • The finite-dictionary consistency result suggests that enlarging the dictionary with more kernel families would still yield reliable selection provided the sample size grows appropriately.

Load-bearing premise

A spectral gap condition must hold so that the spectral projectors remain continuous with respect to perturbations in the kernel weights.

What would settle it

Apply the cross-validation selector to synthetic or real datasets where the underlying operator lacks a clear spectral gap and check whether the selected kernel still produces eigenfunctions measurably closer to a known ground-truth subspace than a fixed baseline kernel.

Figures

Figures reproduced from arXiv: 2604.18402 by Adam Miraoui, Boumediene Hamzi, Houman Owhadi, Othmane Aboussaad.

Figure 1
Figure 1. Figure 1: 1D potentials: reference (black) vs CV-selected kernel (red dashed) vs Uniform Gaussian [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: OU processes: CV-selected eigenfunctions (scatter colored by eigenfunction value). OU2D [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Circle S 1 (σnoise = 0.05): eigenfunctions as functions of θ [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Residuals ϕˆ k − ϕ ⋆ k for DW1D (top) and OU2D αy = 16 (bottom). CV kernel (red) produces smaller residuals than Uniform (blue), especially on higher modes. only eigenvalue maximization and the mild RKHS penalty (L = − P k µk + 0.001P k ∥ak∥ 2 ), with no orthonormality term to avoid triggering the σ-collapse [PITH_FULL_IMAGE:figures/full_fig_p020_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Summary: SubR2 across all examples. Red bars (CV-selected kernel + RFF) consistently exceed blue bars (Uniform Gaussian + Nyström), with percentage errors labeled. Example CV+RFF VarRFF-bounded Learned σj (seed 42) OU2D (αy = 4) 0.977 ± 0.013 0.991 ± 0.008 [163, 163] OU2D (αy = 16) 0.939 ± 0.032 0.960 ± 0.026 [135, 135] OU 3D 0.980 ± 0.007 0.963 ± 0.044 [171, 171, 171] [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 6
Figure 6. Figure 6: CV eigenvalue-sum score across kernel families and bandwidths ( [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: SubR2 vs sample size N on OU2D (αy = 4). CV+RFF (Matérn-3/2) achieves near-perfect recovery at all sample sizes; Uniform improves slowly with N. Config Eigsum-CV Gap-CV Uniform d= 6 (2s+4f) 0.512 (σ = 107) 0.789 (σ = 1.0) 0.822 d= 10 (2s+8f) 0.497 ± 0.017 (σ ≈ 110) 0.778 ± 0.011 (σ ≈ 1.1) 0.805 d= 20 (2s+18f) 0.468 (σ = 120) 0.777 (σ = 2.0) 0.785 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
read the original abstract

Selecting an appropriate kernel is a central challenge in kernel-based spectral methods. In \emph{Kernelized Diffusion Maps} (KDM), the kernel determines the accuracy of the RKHS estimator of a diffusion-type operator and hence the quality and stability of the recovered eigenfunctions. We introduce two complementary approaches to adaptive kernel selection for KDM. First, we develop a variational outer loop that learns continuous kernel parameters, including bandwidths and mixture weights, by differentiating through the Cholesky-reduced KDM eigenproblem with an objective combining eigenvalue maximization, subspace orthonormality, and RKHS regularization. Second, we propose an unsupervised cross-validation pipeline that selects kernel families and bandwidths using an eigenvalue-sum criterion together with random Fourier features for scalability. Both methods share a common theoretical foundation: we prove Lipschitz dependence of KDM operators on kernel weights, continuity of spectral projectors under a gap condition, a residual-control theorem certifying proximity to the target eigenspace, and exponential consistency of the cross-validation selector over a finite kernel dictionary.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces two complementary adaptive kernel selection methods for Kernelized Diffusion Maps (KDM): (1) a variational outer loop that learns continuous kernel parameters (bandwidths, mixture weights) by differentiating through the Cholesky-reduced KDM eigenproblem with an objective combining eigenvalue maximization, subspace orthonormality, and RKHS regularization; and (2) an unsupervised cross-validation pipeline that selects kernel families and bandwidths via an eigenvalue-sum criterion, using random Fourier features for scalability. Both rest on a shared theoretical foundation consisting of Lipschitz dependence of KDM operators on kernel weights, continuity of spectral projectors under a gap condition, a residual-control theorem for proximity to the target eigenspace, and exponential consistency of the CV selector over a finite kernel dictionary.

Significance. If the theoretical results hold, the work addresses a practically important limitation of kernel-based spectral methods by automating kernel choice while supplying stability and consistency guarantees. The dual continuous/discrete strategy, the use of differentiable eigenproblems, and the scalability via RFF are strengths; the explicit residual-control and consistency theorems could be useful for downstream applications if the gap assumption can be managed.

major comments (2)
  1. [theoretical foundation (as summarized in the abstract)] The continuity of spectral projectors (and the subsequent residual-control and consistency theorems) is proved only under an explicit gap condition on the diffusion operator. Neither the variational outer loop (which optimizes eigenvalue sums and orthonormality) nor the eigenvalue-sum CV criterion enforces or verifies that the selected kernel produces a positive spectral gap. This assumption can fail for data near lower-dimensional manifolds or for bandwidths that collapse, rendering the continuity statement inapplicable to the output of the adaptive procedure.
  2. [abstract and theoretical claims] The abstract states the four main theorems (Lipschitz dependence, gap-conditioned continuity, residual control, exponential consistency) but supplies no proof sketches, key intermediate lemmas, or derivation outlines. Without these, it is impossible to assess whether the Lipschitz bound on the KDM operator is derived in a manner that survives the adaptive selection or whether the exponential consistency rate depends on the gap in a way that the CV procedure controls.
minor comments (2)
  1. [variational outer loop] The description of the Cholesky reduction used to enable differentiation through the eigenproblem is too brief; a short algorithmic outline or pseudocode would clarify how the gradient is obtained without explicit eigendecomposition.
  2. [cross-validation pipeline] The manuscript should state the precise form of the eigenvalue-sum CV criterion and the size of the finite kernel dictionary used for the consistency theorem.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful and constructive review. The comments highlight important aspects of the gap assumption and the presentation of theoretical results. We respond point by point below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: The continuity of spectral projectors (and the subsequent residual-control and consistency theorems) is proved only under an explicit gap condition on the diffusion operator. Neither the variational outer loop (which optimizes eigenvalue sums and orthonormality) nor the eigenvalue-sum CV criterion enforces or verifies that the selected kernel produces a positive spectral gap. This assumption can fail for data near lower-dimensional manifolds or for bandwidths that collapse, rendering the continuity statement inapplicable to the output of the adaptive procedure.

    Authors: We agree that the gap condition is required for continuity of the spectral projectors and for the residual-control and consistency theorems to apply. The variational objective maximizes eigenvalue sums and the CV criterion likewise selects on eigenvalue sums; these choices tend to favor larger gaps but do not explicitly enforce or verify a positive gap. Consequently the continuity statement is conditional and may not hold for every output of the adaptive procedures, especially on lower-dimensional data or with collapsing bandwidths. In the revised manuscript we will add a dedicated discussion of this assumption, its possible violations, and a practical post-selection check that computes the estimated gap from the chosen kernel. We will also note that the Lipschitz dependence result holds independently of the gap. revision: partial

  2. Referee: The abstract states the four main theorems (Lipschitz dependence, gap-conditioned continuity, residual control, exponential consistency) but supplies no proof sketches, key intermediate lemmas, or derivation outlines. Without these, it is impossible to assess whether the Lipschitz bound on the KDM operator is derived in a manner that survives the adaptive selection or whether the exponential consistency rate depends on the gap in a way that the CV procedure controls.

    Authors: The abstract is deliberately concise. The full proofs of Lipschitz dependence of the KDM operator on kernel weights, gap-conditioned continuity of projectors, the residual-control theorem, and exponential consistency of the CV selector appear in Sections 3.1–3.4 together with the supporting lemmas in the appendix. To improve transparency we will insert a short subsection in the introduction that outlines the proof strategy for each theorem, explicitly indicating that the Lipschitz bound is established before adaptive selection and that the consistency rate depends on the gap size (which the CV procedure influences indirectly through eigenvalue-sum selection). revision: yes

Circularity Check

0 steps flagged

No significant circularity; theoretical results are independent proofs

full rationale

The paper establishes Lipschitz dependence of KDM operators on kernel weights, continuity of spectral projectors under an explicit gap condition, a residual-control theorem, and exponential consistency of the CV selector as separate mathematical results. These are not obtained by fitting parameters to data and renaming the fit as a prediction, nor by defining quantities in terms of each other, nor by load-bearing self-citations whose content reduces to the present claims. The gap condition is stated as an assumption required for the continuity result and is not enforced or verified by the adaptive procedure, but this is a limitation on applicability rather than a circular reduction in the derivation. The variational outer loop and eigenvalue-sum CV criterion are presented as practical methods whose supporting theory is derived independently.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on a domain assumption about eigenvalue gaps and standard properties of RKHS and spectral projectors; no free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Existence of a spectral gap condition ensuring continuity of spectral projectors
    Explicitly invoked to obtain continuity of spectral projectors in the theoretical foundation section of the abstract.

pith-pipeline@v0.9.0 · 5483 in / 1248 out tokens · 32935 ms · 2026-05-10T04:00:06.064754+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

300 extracted references · 300 canonical work pages

  1. [1]

    The Annals of Probability , keywords =

    Amarjit Budhiraja and Paul Dupuis and Vasileios Maroulas , doi =. The Annals of Probability , keywords =

    work page

  2. [2]

    Balanced reduction of nonlinear control systems in reproducing kernel

    Bouvrie, Jake and Hamzi, Boumediene , booktitle =. Balanced reduction of nonlinear control systems in reproducing kernel. doi:10.1109/ALLERTON.2010.5706920 , keywords =

    work page doi:10.1109/allerton.2010.5706920 2010

  3. [3]

    and Magnus, Jan R

    Abadir, Karim M. and Magnus, Jan R. , doi =. Matrix

    work page

  4. [4]

    , isbn =

    Abarbanel, H. , isbn =. Analysis of Observed Chaotic Data , year =

    work page

  5. [5]

    and Moskowitz, M

    Abbaspour, H. and Moskowitz, M. A. , owner =. Basic Lie Theory , year =

    work page

  6. [6]

    and Weinan, E

    Abdulle, A. and Weinan, E. and Engquist, B. and Vanden-Eijnden, E. , journal =. The heterogeneous multiscale method , volume =

    work page

  7. [7]

    and Cadre, B

    Abraham, C. and Cadre, B. , doi =. Asymptotic properties of posterior distributions derived from misspecified models , url =. C. R. Math. Acad. Sci. Paris , mrclass =

    work page

  8. [8]

    and Cadre, B

    Abraham, C. and Cadre, B. , fjournal =. Concentration of posterior distributions with misspecified models , volume =. Ann. I.S.U.P. , mrclass =

    work page

  9. [9]

    and Marsden, J

    Abraham, R. and Marsden, J. E. and Ratiu, T. S. and Cushman, R. , owner =. Foundations of Mechanics , year =

    work page

  10. [10]

    Abramovich, Y. A. and Aliprantis, C. D. , owner =. An Invitation to Operator Theory , volume =

    work page

  11. [11]

    Ackerman, N. L. and Freer, C. E. and Roy, D. M. , journal =. On the computability of conditional probability , year =

    work page

  12. [12]

    and Hensman, J

    Adam, V. and Hensman, J. and Sahani, M. , booktitle =. Scalable transformed additive signal decomposition by non-conjugate

    work page

  13. [13]

    Adams, R. A. and Fournier, J. J. F. , publisher =. Sobolev Spaces , volume =

    work page

  14. [14]

    , journal =

    Adamski, W. , journal =. Complete spaces and zero-one measures , volume =

    work page

  15. [15]

    and Hamilton, J

    Adler, A. and Hamilton, J. , journal =. Invariant means via the ultrapower , volume =

    work page

  16. [16]

    Time-Series Clustering

    Aghabozorgi, Saeed and Seyed Shirkhorshidi, Ali and Ying Wah, Teh , doi =. Time-Series Clustering. Information Systems , langid =

    work page

  17. [17]

    , note =

    Agterberg, F. , note =. Georges

    work page

  18. [18]

    Ahlberg, J. H. and Nilson, E. N. , journal =. Convergence properties of the spline fit , volume =

    work page

  19. [19]

    Linearizing nonlinear dynamics using deep learning , volume =

    Akhil Ahmed and Ehecatl Antonio del Rio Chanona and Mehmet Mercangöz , journal =. Linearizing nonlinear dynamics using deep learning , volume =

    work page

  20. [20]

    Learning ``Best'' Kernels from Data in

    Akian, Jean-Luc and Bonnet, Luc and Owhadi, Houman and Savin,. Learning ``Best'' Kernels from Data in. Journal of Computational Physics , langid =. doi:10.1016/j.jcp.2022.111595 , issn =

    work page doi:10.1016/j.jcp.2022.111595 2022

  21. [21]

    , journal =

    Allaire, G. , journal =. Homogenization and two-scale convergence , volume =

    work page

  22. [22]

    , booktitle =

    Allaire, G. , booktitle =. Two-scale convergence: a new method in periodic homogenization. Nonlinear partial differential equations and their applications , volume =

    work page

  23. [23]

    and Kalton, N

    Albiac, F. and Kalton, N. J. , publisher =. Topics in

    work page

  24. [24]

    Operator-theoretic framework for forecasting nonlinear time series with kernel analog techniques , url =

    Romeo Alexander and Dimitrios Giannakis , doi =. Operator-theoretic framework for forecasting nonlinear time series with kernel analog techniques , url =. Physica D: Nonlinear Phenomena , keywords =

    work page

  25. [25]

    Alfsen, E. M. , owner =. Compact Convex Sets and Boundary Integrals , volume =

    work page

  26. [26]

    Alfsen, E. M. and Andersen, T. B. , journal =. Split faces of compact convex sets , volume =

    work page

  27. [27]

    Aliprantis, C. D. and Border, K. C. , edition =. Infinite

    work page

  28. [28]

    Aliprantis, C. D. and Burkinshaw, O. , number =. Locally solid

    work page

  29. [29]

    Aliprantis, C. D. and Burkinshaw, O. , owner =. Positive operators , volume =

    work page

  30. [30]

    Geodesic shooting and diffeomorphic matching via textured meshes , year =

    Allassonni. Geodesic shooting and diffeomorphic matching via textured meshes , year =. International Workshop on Energy Minimization Methods in Computer Vision and Pattern Recognition , organization =

    work page

  31. [31]

    Sampling rare switching events in biochemical networks , volume =

    Allen, Rosalind J and Warren, Patrick B and Ten Wolde, Pieter Rein , journal =. Sampling rare switching events in biochemical networks , volume =

    work page

  32. [32]

    Simulating rare events in equilibrium or nonequilibrium stochastic systems , volume =

    Allen, Rosalind J and Frenkel, Daan and ten Wolde, Pieter Rein , journal =. Simulating rare events in equilibrium or nonequilibrium stochastic systems , volume =

    work page

  33. [33]

    and Beylkin, G

    Alpert, B. and Beylkin, G. and Coifman, R. and Rokhlin, V. , journal =. Wavelet-like bases for the fast solution of second-kind integral equations , volume =

    work page

  34. [34]

    Kernels for vector-valued functions: A review , volume =

    Alvarez, Mauricio A and Rosasco, Lorenzo and Lawrence, Neil D and others , journal =. Kernels for vector-valued functions: A review , volume =

    work page

  35. [35]

    The effects of adding noise during backpropagation training on a generalization performance , volume =

    An, Guozhong , journal =. The effects of adding noise during backpropagation training on a generalization performance , volume =

    work page

  36. [36]

    , journal =

    Ancona, A. , journal =. Some results and examples about the behavior of harmonic functions and

    work page

  37. [37]

    Anderson, E. J. , journal =. A review of duality theory for linear programming over topological vector spaces , volume =

    work page

  38. [38]

    Anderson, T. W. , doi =. The integral of a symmetric unimodal function over a symmetric convex set and some probability inequalities , volume =. Proc. Amer. Math. Soc. , mrclass =

    work page

  39. [39]

    Anderson, Jr., W. N. and Trapp, G. E. , journal =. Shorted operators

    work page

  40. [40]

    Anderson, E. J. and Nash, P. , owner =. Linear programming in infinite-dimensional spaces: theory and applications , year =

    work page

  41. [41]

    , journal =

    Andersson, S. , journal =. Distributions of maximal invariants using quotient measures , year =

    work page

  42. [42]

    , booktitle =

    Ando, T. , booktitle =. Schur complements and matrix inequalities:

    work page

  43. [43]

    , journal =

    Gloria, A. , journal =. An analytical framework for the numerical homogenization of monotone elliptic operators and quasiconvex energies , volume =

    work page

  44. [44]

    and Speckbacher, M

    Antoine, J.-P. and Speckbacher, M. and Trapani, C. , journal =. Reproducing pairs of measurable functions , volume =

    work page

  45. [45]

    L\'evy Processes and Stochastic Calculus, second ed , year =

    work page

  46. [46]

    and Burke, James V

    Aravkin, Aleksandr Y. and Burke, James V. and Pillonetto, Gianluigi , booktitle =. Optimization. doi:10.1007/978-3-642-38398-4_8 , editor =

    work page doi:10.1007/978-3-642-38398-4_8

  47. [47]

    , owner =

    Arendt, W. , owner =

    work page

  48. [48]

    Arens, R. F. , fjournal =. A topology for spaces of transformations , volume =. Ann. of Math. (2) , mrclass =

    work page

  49. [49]

    Arfken, G. B. and Weber, H. J. , publisher =. Mathematical Methods for Physicists , year =

    work page

  50. [50]

    Fundamental theory of ordinary differential equations , year =

    Arino, Julien , journal =. Fundamental theory of ordinary differential equations , year =

    work page

  51. [51]

    Arnborg, S. and Sj. On the foundations of. Bayesian inference and maximum entropy methods in science and engineering (

    work page

  52. [52]

    Arnold, V. I. , publisher =. Mathematical Methods of Classical Mechanics , year =

    work page

  53. [53]

    Geometrical Methods in the Theory of Ordinary Differential Equations , year =

    Arnold, Vladimir , publisher =. Geometrical Methods in the Theory of Ordinary Differential Equations , year =

    work page

  54. [54]

    , description =

    Aronszajn, N. , description =. Theory of Reproducing Kernels , url =. Transactions of the American Mathematical Society , keywords =

    work page

  55. [55]

    Arrow, K. J. , doi =. A difficulty in the concept of social welfare , volume =. J. Pol. Econ. , month =

    work page

  56. [56]

    and Blackwell, D

    Arrow, J. and Blackwell, D. and Girshick, M. A. , journal =. Bayes and minimax solutions of sequential decision problems , volume =

    work page

  57. [57]

    , journal =

    Artstein, Z. , journal =. Discrete and continuous bang-bang and facial spaces or: look for the extreme points , volume =

    work page

  58. [58]

    , owner =

    Arveson, W. , owner =. An Invitation to C

    work page

  59. [59]

    and Lehn, J

    Ascherl, A. and Lehn, J. , journal =. Two principles for extending probability measures , volume =

    work page

  60. [60]

    Ash, R. B. , mrclass =. Real

    work page

  61. [61]

    Ashwin and S

    P. Ashwin and S. Wieczorek and R. Vitolo and P. Cox , journal =. Tipping points in open systems: Bifurcation, noise-induced and rate-dependent examples in the climate system , volume =

    work page

  62. [62]

    and Dalibard, J

    Aspect, A. and Dalibard, J. and Roger, G. , journal =. Experimental test of Bell's inequalities using time-varying analyzers , volume =

    work page

  63. [63]

    and Grangier, P

    Aspect, A. and Grangier, P. and Roger, G. , journal =. Experimental realization of Einstein-Podolsky-Rosen-Bohm Gedankenexperiment: a new violation of Bell's inequalities , volume =

    work page

  64. [64]

    and Grangier, P

    Aspect, A. and Grangier, P. , journal =. Wave-particle duality for single photons , volume =

    work page

  65. [65]

    Data-driven discovery of free-form governing differential equations.arXiv preprint arXiv:1910.05117, 2019

    Data-Driven Discovery of Free-Form Governing Differential Equations , year =. arXiv:1910.05117 [physics, stat] , langid =. arXiv , author =:1910.05117 , eprinttype =

    work page arXiv 1910

  66. [66]

    Proximal alternating minimization and projection methods for nonconvex problems: An approach based on the

    Attouch, H. Proximal alternating minimization and projection methods for nonconvex problems: An approach based on the. Mathematics of Operations Research , number =

    work page

  67. [67]

    Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward-backward splitting, and regularized

    Attouch, H\'. Convergence of descent methods for semi-algebraic and tame problems: proximal algorithms, forward-backward splitting, and regularized. Mathematical Programming , pages =

    work page

  68. [68]

    Reproducing Kernels for Hilbert Spaces of Real Harmonic Functions , url =

    Giles Auchmuty , doi =. Reproducing Kernels for Hilbert Spaces of Real Harmonic Functions , url =. SIAM Journal on Mathematical Analysis , number =

    work page

  69. [69]

    , booktitle =

    Augustin, T. , booktitle =. On the Suboptimality of the Generalized

    work page

  70. [70]

    Aumann, R. J. , journal =. Measurable utility and measurable choice theorem , year =

    work page

  71. [71]

    ???????? , journal =. Sparse

    work page

  72. [72]

    , journal =

    Averna, D. , journal =. Lusin type theorems for multifunctions,

    work page

  73. [73]

    and Fiacca, A

    Averna, D. and Fiacca, A. , journal =. Sulla proprieta di

    work page

  74. [74]

    org/abs/2309.04522

    Avidan, Yehonatan and Li, Qianyi and Sompolinsky, Haim , date =. Connecting. arXiv , pubstate =. doi:10.48550/arXiv.2309.04522 , eprint =

    work page doi:10.48550/arxiv.2309.04522

  75. [75]

    and Kolotilina, L

    Axelsson, O. and Kolotilina, L. , journal =. Monotonicity and discretization error estimates , volume =

    work page

  76. [76]

    A stochastic collocation method for elliptic partial differential equations with random input data , url =

    Babu. A stochastic collocation method for elliptic partial differential equations with random input data , url =. SIAM Rev. , mrclass =. doi:10.1137/100786356 , fjournal =

    work page doi:10.1137/100786356

  77. [77]

    Backus, G. E. , journal =. Bayesian inference in geomagnetism , volume =

    work page

  78. [78]

    Backus, G. E. , journal =. Trimming and procrastination as inversion techniques , volume =

    work page

  79. [79]

    and Herrmann, Benjamin and McKeon, Beverley J

    Baddoo, Peter J. and Herrmann, Benjamin and McKeon, Beverley J. and Brunton, Steven L. , doi =. Kernel Learning for Robust Dynamic Mode Decomposition: Linear and Nonlinear Disambiguation Optimization , volume =. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , langid =

    work page

  80. [80]

    , journal =

    Baer, R. , journal =. Crossed isomorphisms , volume =

    work page

Showing first 80 references.