Pith. sign in

REVIEW 2 major objections 5 minor 160 references

When functional predictors are seen only through noisy points, minimax prediction risk splits into a dense-FLR term plus design-specific discretization costs.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 03:40 UTC pith:44HC6J6H

load-bearing objection Sharp joint (n,m) minimax rates for discretely observed FLR, with a genuine four-term common-design rate and matching adaptive estimators. the 2 major comments →

arxiv 2607.09350 v1 pith:44HC6J6H submitted 2026-07-10 math.ST stat.TH

The Cost of Discretization in Functional Linear Regression: Minimax Rates and Adaptation

T. Tony Cai , Yicheng Li This is my paper
classification math.ST stat.TH MSC 62G0862M2062R10
keywords functional linear regressionminimax ratesdiscrete noisy observationsindependent designcommon designadaptive estimationphase transitioneigenvalue identification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Functional linear regression is usually studied as if each predictor curve is fully observed. In practice one sees only m noisy point samples on each of n curves. This paper gives exact minimax rates for prediction (and estimation) as joint functions of n and m under two standard designs. Under independent random grids the rate is the familiar fully observed benchmark plus a second term that reflects noisy sampling after amplification by the inverse covariance; the phase transition occurs when m scales like n to a power set by eigenvalue decay and slope smoothness. Under a shared equispaced grid two further unavoidable terms appear: fixed-grid approximation error and the cost of identifying unknown eigenvalues from the common lattice. Adaptive estimators that screen covariance scale and threshold blockwise prediction energy attain the rates without knowing the spectrum or the smoothness indices. The message is that discretization is not a mere technical nuisance: sampling geometry itself changes the statistical experiment and forces new rate terms that cannot be averaged away by collecting more subjects.

Core claim

Under independent design the minimax prediction risk is of exact order n to the power -(2α+2s)/(2α+2s+1) plus (nm) to the power -(2α+2s)/(4α+2s+1). Under common design with unknown eigenvalues the same two terms remain and are joined by m to the power -(2α+2s) and m to the power -4α, with matching adaptive upper bounds. The first term recovers the fully observed functional-linear-regression benchmark; the second is the statistical price of noisy point evaluations after inverse-covariance amplification; the last two are geometric obstructions created by a shared grid.

What carries the argument

Matching upper and lower bounds obtained by combining a plug-in Fourier estimator (with eigenvalue screening and blockwise prediction-energy thresholding) against van Trees lower bounds that control Fisher information for noisy point evaluations, plus two indistinguishability constructions that isolate fixed-grid aliasing and unknown-eigenvalue identification on a common lattice.

Load-bearing premise

The covariance operator is assumed to be diagonalized by a fixed, known trigonometric basis, so the theory does not yet pay for estimating an unknown eigenbasis or for misalignment between covariance geometry and slope smoothness.

What would settle it

If, for independent design, the excess prediction risk fails to obey the two-term rate (or the stated phase transition at m ~ n^{2α/(2α+2s+1)}) under the paper's Gaussian trigonometric model, or if common-design risk remains free of the m^{-4α} term when eigenvalues are unknown, the claimed minimax characterization is false.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper derives matching minimax rates for prediction (and sketches estimation) in scalar-on-function linear regression when each trajectory is observed only at m noisy points. Working in a fixed trigonometric eigenbasis with eigenvalue decay α and slope Sobolev smoothness s, it treats independent random design and common equispaced design separately. Under independent design the minimax prediction risk is of order n^{-(2α+2s)/(2α+2s+1)}+(nm)^{-(2α+2s)/(4α+2s+1)}; under common design with unknown eigenvalues the rate adds the grid terms m^{-(2α+2s)}+m^{-4α}. Adaptive estimators that screen covariance scale and threshold blockwise prediction energy attain these rates without knowledge of (λ_r), α, or s. Phase diagrams, simulations, and a wheat-spectra example illustrate the theory.

Significance. The contribution is substantial for functional data analysis and nonparametric inverse problems. Fully observed FLR rates and sparse-to-dense transitions for mean/covariance estimation were known; a complete (n,m)-minimax theory for FLR that separates noisy-point cost from fixed-grid aliasing and eigenvalue identification was not. Matching lower bounds (van Trees with refined Fisher control; two distinct common-grid indistinguishability constructions) and adaptive upper bounds make the phase transitions sharp rather than merely sufficient. The fixed trigonometric eigenbasis is an explicit modeling choice that isolates discretization cost; within that model the results are definitive and will serve as a benchmark for subsequent work on unknown eigenbases and alignment.

major comments (2)
  1. Section 6 states L2 estimation rates (independent: n^{-2s/(2α+2s+1)}+(nm)^{-2s/(4α+2s+1)}; common: plus m^{-2s}+m^{-4α}) as following by the same arguments after reweighting by λ_r^2 and adjusting the energy threshold. The sketch is plausible but incomplete: the van Trees block priors, Fisher bounds, and eligible-block risk decompositions all change with the unweighted loss, and the plug-in remainder B^λ_ℓ must be re-controlled. Either supply the full parallel proofs (or a self-contained appendix lemma) or present the L2 rates as conjectured extensions rather than established corollaries.
  2. The fixed trigonometric eigenbasis (Section 2 after (4)–(5)) is load-bearing for every rate. Section 6 correctly flags eigenbasis estimation and alignment as open, but the abstract and introduction still present the rates as the cost of discretization for FLR without always restating the basis restriction. A short, prominent caveat in the abstract and at the start of Corollaries 3.4 and 4.4 would prevent misapplication when the covariance eigenbasis is unknown or misaligned with the Sobolev scale of β.
minor comments (5)
  1. Table 2 (wheat data): CD beats IND and the published benchmarks, contrary to the asymptotic ranking. The finite-sample explanation is reasonable; a one-sentence note that the common subgrid is deterministic while IND and the Zhou et al. numbers use random subsampling would make the comparison protocol fully transparent.
  2. Figure 1 phase diagrams are clear; adding the critical exponents (e.g. ζ_ind = 2α/(ν+1)) as axis annotations would help readers match the figure to Subsection 4.4.1 without flipping back to the text.
  3. Notation: ν := 2α+2s and κ := 4α+2s+1 are introduced late (Table 1 / §4.4). Defining them once in Section 2 or at the first rate display would reduce repetition.
  4. Adaptive constants (M0, CV, CL, cL) are fixed but their practical choice is not discussed. A brief remark in Section 5 on the values used in the simulations would aid reproducibility.
  5. Typos/style: “What left open” (p. 3) → “What is left open”; occasional double spaces and “them −4α” line-break artifacts in the PDF.

Circularity Check

0 steps flagged

No significant circularity: minimax rates are derived from stated model assumptions via Fisher/van Trees lower bounds and constructive oracle/adaptive upper bounds, not by fitting or redefining the target.

full rationale

The paper is a standard minimax theory contribution. Under a fixed trigonometric eigenbasis that diagonalizes the covariance (Section 2), prediction risk is the weighted ℓ2 error ∑ λ_r |θ̂_r − θ_r|². Oracle upper bounds (Theorems 3.1, 4.1) balance truncation bias against variance of cross-covariance estimators; matching lower bounds use van Trees on Gaussian block submodels with Fisher information controlled for noisy point evaluations (Appendix D) and, under common design, two distinct indistinguishability constructions for fixed-grid aliasing and unknown-eigenvalue identification (Appendix G). Adaptive procedures screen eigenvalues and threshold blockwise prediction energy with fixed universal constants (M0, CV, CL), attaining the same rates without knowledge of (λ_r), α, or s (Theorems 3.3, 4.3). Self-citations to Cai–Yuan supply the fully-observed FLR benchmark term and mean/covariance phase-transition context; they do not redefine or force the new (nm) and m-only discretization terms. Simulations and the wheat example illustrate the theory rather than fit the rates. No step reduces the claimed rates to their inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central rates rest on a standard Gaussian FLR model plus a strong but explicit geometric assumption that a fixed trigonometric basis diagonalizes the covariance. Smoothness and eigenvalue-decay classes are classical. Adaptive estimators introduce fixed numerical tuning constants that affect constants, not rates. No new physical entities are postulated.

free parameters (2)
  • Adaptive pilot/threshold constants (M0, CV, CL, cL, R)
    Fixed large/small numerical constants in eigenvalue floors and block energy thresholds; chosen sufficiently large/small for theory, not fitted to the wheat or simulation risk surfaces.
  • Noise levels σε, σδ and eigenvalue envelope constants cλ, Cλ, R0
    Model parameters treated as fixed known-order constants in rate statements; simulations pick concrete values (e.g. σε=0.5, σδ=0.1) for illustration only.
axioms (5)
  • domain assumption Predictor trajectories are centered Gaussian processes with continuous paths; response and measurement errors are independent Gaussian.
    Section 2 model (1)–(2); used for Fisher information and concentration throughout lower and upper bounds.
  • domain assumption Covariance operator is diagonalized by the fixed trigonometric basis with eigenvalues in Lα(cλ,Cλ), α>1/2.
    Section 2 after (4)–(9); isolates discretization cost from eigenbasis estimation.
  • domain assumption Slope Fourier coefficients lie in Sobolev ball Θs(R0), s≥0.
    Equation (10); defines the parameter space for minimax rates.
  • domain assumption Independent design: tij iid Unif[0,1]; common design: equal grid tj=(j-1)/m.
    Assumptions 1–2; define the two statistical experiments.
  • standard math Van Trees inequality and standard sub-Weibull/Bernstein concentration tools.
    Lemma D.1 and Appendix H; used for lower and adaptive upper bounds.

pith-pipeline@v1.1.0-grok45 · 76585 in / 2697 out tokens · 30305 ms · 2026-07-13T03:40:28.194287+00:00 · methodology

0 comments
read the original abstract

We study scalar-on-function linear regression when each covariate curve is observed only through finitely many noisy point evaluations. Our goal is to characterize the minimax estimation and prediction risks as joint functions of the number of trajectories $n$ and the within-trajectory resolution $m$. Working in a fixed trigonometric eigenbasis, with covariance eigenvalues decaying at rate $\alpha$ and slope function of Sobolev smoothness $s$, we derive matching minimax upper and lower bounds under two canonical sampling schemes. Under an independent random design, the minimax prediction rate is $n^{-\frac{2\alpha+2s}{2\alpha+2s+1}} + (nm)^{-\frac{2\alpha+2s}{4\alpha+2s+1}}$. The first term is the fully observed functional linear regression benchmark, while the second term captures the cost of noisy point evaluations after amplification by the inverse covariance operator. Under a common design on an equally spaced grid, the shared sampling geometry introduces additional obstructions, and the minimax prediction rate becomes $n^{-\frac{2\alpha+2s}{2\alpha+2s+1}} + (nm)^{-\frac{2\alpha+2s}{4\alpha+2s+1}} + m^{-(2\alpha+2s)} + m^{-4\alpha}$. Here the third term represents discretization error induced by the fixed grid, whereas the fourth reflects the cost of identifying unknown eigenvalues from observations on a common grid. We further construct data-driven adaptive estimators that screen the covariance scale and threshold blockwise prediction energy, attaining these rates without prior knowledge of the eigenvalue sequence or the smoothness indices. The results reveal a sharp phase transition that depends on the sampling resolution under independent design and a richer phase diagram under common design. Numerical simulations and a real data example illustrate the theoretical findings.

Figures

Figures reproduced from arXiv: 2607.09350 by T. Tony Cai, Yicheng Li.

Figure 1
Figure 1. Figure 1: Phase diagram for the minimax rate under common design. The dominant term in ( [PITH_FULL_IMAGE:figures/full_fig_p015_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Simulation heatmaps under independent design (top) and common design (bottom). [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Matched comparison between common and independent design for ( [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Common design rate comparisons as m varies (top) and along m = n γ (bottom). Each panel uses 50 repetitions and plots the median excess prediction risk with interquartile error bars. The dashed reference lines show the low-resolution term predicted by (35) in the top plots, and show the dominant rate in the bottom plots. To further investigate the rate behavior under common design, we compare the empirical… view at source ↗
Figure 5
Figure 5. Figure 5: Known versus estimated eigenvalues under common design for ( [PITH_FULL_IMAGE:figures/full_fig_p021_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional simulation heatmaps under common design. [PITH_FULL_IMAGE:figures/full_fig_p090_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional simulation heatmaps under independent design. [PITH_FULL_IMAGE:figures/full_fig_p091_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional common-design rate comparisons. [PITH_FULL_IMAGE:figures/full_fig_p092_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Additional independent-design rate comparisons. [PITH_FULL_IMAGE:figures/full_fig_p093_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Common-design rate comparisons with known and unknown eigenvalues. The left [PITH_FULL_IMAGE:figures/full_fig_p094_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Measurement-noise sensitivity under independent design (top) and common design [PITH_FULL_IMAGE:figures/full_fig_p096_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Observation-noise sensitivity under independent design (top) and common design (bot [PITH_FULL_IMAGE:figures/full_fig_p097_12.png] view at source ↗

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

160 extracted references · 27 canonical work pages · 5 internal anchors

  1. [1]

    2003 , publisher =

    Sobolev Spaces , author =. 2003 , publisher =

    2003

  2. [2]

    2008 , journal =

    A Tail Inequality for Suprema of Unbounded Empirical Processes with Applications to Markov Chains , author =. 2008 , journal =

    2008

  3. [3]

    Bousquet, Olivier , year =. A. Comptes Rendus. Math. doi:10.1016/S1631-073X(02)02292-6 , url =

    doi:10.1016/s1631-073x(02)02292-6

  4. [4]

    2013 , publisher =

    Concentration Inequalities: A Nonasymptotic Theory of Independence , author =. 2013 , publisher =. doi:10.1093/acprof:oso/9780199535255.001.0001 , isbn =

    doi:10.1093/acprof:oso/9780199535255.001.0001 2013

  5. [5]

    2019 , journal =

    On the Convergence Rate of Training Recurrent Neural Networks , author =. 2019 , journal =

    2019

  6. [6]

    2019 , month = jun, eprint =

    A Convergence Theory for Deep Learning via Over-Parameterization , author =. 2019 , month = jun, eprint =

    2019

  7. [7]

    2021 , journal =

    Concentration of Kernel Matrices with Application to Kernel Spectral Clustering , author =. 2021 , journal =

    2021

  8. [8]

    2008 , series =

    Support Vector Machines , author =. 2008 , series =. doi:10.1007/978-0-387-77242-4 , isbn =

    doi:10.1007/978-0-387-77242-4 2008

  9. [9]

    and Hu, Wei and Li, Zhiyuan and Salakhutdinov, Russ R and Wang, Ruosong , year =

    Arora, Sanjeev and Du, Simon S. and Hu, Wei and Li, Zhiyuan and Salakhutdinov, Russ R and Wang, Ruosong , year =. On Exact Computation with an Infinitely Wide Neural Net , booktitle =

  10. [10]

    Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks , booktitle =

    Arora, Sanjeev and Du, Simon and Hu, Wei and Li, Zhiyuan and Wang, Ruosong , year =. Fine-Grained Analysis of Optimization and Generalization for Overparameterized Two-Layer Neural Networks , booktitle =

  11. [11]

    2014 , month = jan, journal =

    Sharp Estimates for Eigenvalues of Integral Operators Generated by Dot Product Kernels on the Sphere , author =. 2014 , month = jan, journal =. doi:10.1016/j.jat.2013.10.002 , abstract =

    doi:10.1016/j.jat.2013.10.002 2014

  12. [12]

    2017 , journal =

    Breaking the Curse of Dimensionality with Convex Neural Networks , author =. 2017 , journal =

    2017

  13. [13]

    Characterization of Sobolev spaces on the sphere

    Barcel. Characterization of. 2019 , month =. doi:10.48550/arXiv.1907.01571 , abstract =. arxiv , keywords =:arXiv:1907.01571 , publisher =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1907.01571 2019

  14. [14]

    A Kernel Perspective of Skip Connections in Convolutional Networks

    A Kernel Perspective of Skip Connections in Convolutional Networks , author =. 2022 , month = nov, number =. doi:10.48550/arXiv.2211.14810 , abstract =. arxiv , keywords =:arXiv:2211.14810 , publisher =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2211.14810 2022

  15. [15]

    2007 , journal =

    On Regularization Algorithms in Learning Theory , author =. 2007 , journal =. doi:10.1016/j.jco.2006.07.001 , abstract =

    doi:10.1016/j.jco.2006.07.001 2007

  16. [16]

    On Deep Learning as a Remedy for the Curse of Dimensionality in Nonparametric Regression , author =

  17. [17]

    2022 , month = jun, number =

    Kernel Ridgeless Regression Is Inconsistent in Low Dimensions , author =. 2022 , month = jun, number =. doi:10.48550/arXiv.2205.13525 , abstract =. arXiv:2205.13525 , publisher =

    Pith/arXiv arXiv doi:10.48550/arxiv.2205.13525 2022

  18. [18]

    Beatson, R. K. and zu Castell, W. and Xu, Y. , year =. A. doi:10.48550/arXiv.1110.2437 , abstract =. arxiv , keywords =:arXiv:1110.2437 , publisher =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1110.2437

  19. [19]

    To Understand Deep Learning We Need to Understand Kernel Learning , booktitle =

    Belkin, Mikhail and Ma, Siyuan and Mandal, Soumik , year =. To Understand Deep Learning We Need to Understand Kernel Learning , booktitle =

  20. [20]

    On the Inductive Bias of Neural Tangent Kernels , booktitle =

    Bietti, Alberto and Mairal, Julien , year =. On the Inductive Bias of Neural Tangent Kernels , booktitle =

  21. [21]

    2020 , journal =

    Deep Equals Shallow for Relu Networks in Kernel Regimes , author =. 2020 , journal =. 2009.14397 , archiveprefix =

    Pith/arXiv arXiv 2020

  22. [22]

    , year =

    Bingham, Nicholas H. , year =. Positive Definite Functions on Spheres , booktitle =

  23. [23]

    2018 , journal =

    Optimal Rates for Regularization of Statistical Inverse Learning Problems , author =. 2018 , journal =. doi:10.1007/s10208-017-9359-7 , abstract =

    work page doi:10.1007/s10208-017-9359-7 2018

  24. [24]

    2003 , journal =

    Spectral Properties of Distance Matrices , author =. 2003 , journal =

    2003

  25. [25]

    Spectrum Dependent Learning Curves in Kernel Regression and Wide Neural Networks , booktitle =

    Bordelon, Blake and Canatar, Abdulkadir and Pehlevan, Cengiz , year =. Spectrum Dependent Learning Curves in Kernel Regression and Wide Neural Networks , booktitle =

  26. [26]

    and Dick, Josef , year =

    Brauchart, Johann S. and Dick, Josef , year =. A Characterization of. Constructive Approximation , volume =. doi:10.1007/s00365-013-9217-z , abstract =

    work page doi:10.1007/s00365-013-9217-z

  27. [27]

    2005 , abstract =

    Spectral Properties of the Kernel Matrix and Their Relation to Kernel Methods in Machine Learning , author =. 2005 , abstract =

    2005

  28. [28]

    What Can Be Learnt With Wide Convolutional Neural Networks?

    What Can Be Learnt with Wide Convolutional Neural Networks? , author =. 2022 , month =. doi:10.48550/arXiv.2208.01003 , abstract =. arxiv , keywords =:arXiv:2208.01003 , publisher =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2208.01003 2022

  29. [29]

    2021 , month = may, journal =

    Spectral Bias and Task-Model Alignment Explain Generalization in Kernel Regression and Infinitely Wide Neural Networks , author =. 2021 , month = may, journal =. doi:10.1038/s41467-021-23103-1 , abstract =

    doi:10.1038/s41467-021-23103-1 2021

  30. [30]

    Generalization Error Bounds of Gradient Descent for Learning Over-Parameterized Deep

    Cao, Yuan and Gu, Quanquan , year =. Generalization Error Bounds of Gradient Descent for Learning Over-Parameterized Deep. Proceedings of the

  31. [31]

    2007 , journal =

    Optimal Rates for the Regularized Least-Squares Algorithm , author =. 2007 , journal =. doi:10.1007/s10208-006-0196-8 , abstract =

    doi:10.1007/s10208-006-0196-8 2007

  32. [32]

    2010 , journal =

    Cross-Validation Based Adaptation for Regularization Operators in Learning Theory , author =. 2010 , journal =. doi:10.1142/S0219530510001564 , abstract =

    work page doi:10.1142/s0219530510001564 2010

  33. [33]

    2012 , journal =

    Eigenvalue Decay of Positive Integral Operators on the Sphere , author =. 2012 , journal =

    2012

  34. [34]

    Deep Neural Tangent Kernel and Laplace Kernel Have the Same

    Chen, Lin and Xu, Sheng , year =. Deep Neural Tangent Kernel and Laplace Kernel Have the Same. arXiv preprint arXiv:2009.10683 , eprint =

    Pith/arXiv arXiv 2009

  35. [35]

    Kernel Methods for Deep Learning , booktitle =

    Cho, Youngmin and Saul, Lawrence , editor =. Kernel Methods for Deep Learning , booktitle =. 2009 , volume =

    2009

  36. [36]

    The Loss Surfaces of Multilayer Networks , booktitle =

    Choromanska, Anna and Henaff, Mikael and Mathieu, Michael and Arous, G. The Loss Surfaces of Multilayer Networks , booktitle =. 2015 , pages =

    2015

  37. [37]

    Conference on Learning Theory , author =

    On the Expressive Power of Deep Learning:. Conference on Learning Theory , author =. 2016 , pages =

    2016

  38. [38]

    2013 , series =

    Approximation Theory and Harmonic Analysis on Spheres and Balls , author =. 2013 , series =. doi:10.1007/978-1-4614-6660-4 , isbn =

    doi:10.1007/978-1-4614-6660-4 2013

  39. [39]

    doi:10.48550/arXiv.1810.04805 , abstract =

    Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina , year =. doi:10.48550/arXiv.1810.04805 , abstract =. arxiv , keywords =:arXiv:1810.04805 , publisher =

    Pith/arXiv arXiv doi:10.48550/arxiv.1810.04805

  40. [40]

    and Zhai, Xiyu and Poczos, Barnabas and Singh, Aarti , year =

    Du, Simon S. and Zhai, Xiyu and Poczos, Barnabas and Singh, Aarti , year =. Gradient Descent Provably Optimizes Over-Parameterized Neural Networks , booktitle =

  41. [41]

    and Lee, Jason and Li, Haochuan and Wang, Liwei and Zhai, Xiyu , year =

    Du, Simon S. and Lee, Jason and Li, Haochuan and Wang, Liwei and Zhai, Xiyu , year =. Gradient Descent Finds Global Minima of Deep Neural Networks , booktitle =

  42. [42]

    Deep Neural Networks as Point Estimates for Deep

    Dutordoir, Vincent and Hensman, James and. Deep Neural Networks as Point Estimates for Deep. Advances in. 2021 , volume =

    2021

  43. [43]

    2020 , journal =

    Spectra of the Conjugate Kernel and Neural Tangent Kernel for Linear-Width Neural Networks , author =. 2020 , journal =

    2020

  44. [44]

    2008 , month = oct, journal =

    Integral Operators on the Sphere Generated by Positive Definite Smooth Kernels , author =. 2008 , month = oct, journal =. doi:10.1016/j.jco.2008.04.001 , abstract =

    work page doi:10.1016/j.jco.2008.04.001 2008

  45. [45]

    2009 , month = may, journal =

    Eigenvalues of Integral Operators Defined by Smooth Positive Definite Kernels , author =. 2009 , month = may, journal =. doi:10.1007/s00020-009-1680-3 , abstract =

    work page doi:10.1007/s00020-009-1680-3 2009

  46. [46]

    2013 , month = dec, journal =

    Eigenvalue Decay Rates for Positive Integral Operators , author =. 2013 , month = dec, journal =. doi:10.1007/s10231-012-0256-z , abstract =

    work page doi:10.1007/s10231-012-0256-z 2013

  47. [47]

    2020 , journal =

    Sobolev Norm Learning Rates for Regularized Least-Squares Algorithms , author =. 2020 , journal =

    2020

  48. [48]

    and Bartlett, Peter , year =

    Frei, Spencer and Chatterji, Niladri S. and Bartlett, Peter , year =. Benign Overfitting without Linearity:. Proceedings of

  49. [49]

    Notes on Spherical Harmonics and Linear Representations of Lie Groups , author =

  50. [50]

    2018 , journal =

    Deep Convolutional Networks as Shallow Gaussian Processes , author =. 2018 , journal =. 1808.05587 , archiveprefix =

    Pith/arXiv arXiv 2018

  51. [51]

    On the Similarity between the

    Geifman, Amnon and Yadav, Abhay and Kasten, Yoni and Galun, Meirav and Jacobs, David and Ronen, Basri , year =. On the Similarity between the. Advances in

  52. [52]

    2008 , journal =

    Spectral Algorithms for Supervised Learning , author =. 2008 , journal =. doi:10.1162/neco.2008.05-07-517 , abstract =

    work page doi:10.1162/neco.2008.05-07-517 2008

  53. [53]

    2020 , journal =

    When Do Neural Networks Outperform Kernel Methods? , author =. 2020 , journal =

    2020

  54. [54]

    Linearized Two-Layers Neural Networks in High Dimension , author =

  55. [55]

    2013 , journal =

    Strictly and Non-Strictly Positive Definite Functions on Spheres , author =. 2013 , journal =

    2013

  56. [56]

    2002 , series =

    A Distribution-Free Theory of Nonparametric Regression , editor =. 2002 , series =

    2002

  57. [57]

    Approximating Continuous Functions by

    Hanin, Boris and Sellke, Mark , year =. Approximating Continuous Functions by. arXiv preprint arXiv:1710.11278 , eprint =

    Pith/arXiv arXiv

  58. [58]

    Random Neural Networks in the Infinite Width Limit as

    Hanin, Boris , year =. Random Neural Networks in the Infinite Width Limit as. arXiv preprint arXiv:2107.01562 , eprint =

    Pith/arXiv arXiv

  59. [59]

    2020 , journal =

    On the Minimax Optimality and Superiority of Deep Neural Network Learning over Sparse Parameter Spaces , author =. 2020 , journal =

    2020

  60. [60]

    The Curse of Depth in Kernel Regime , booktitle =

    Hayou, Soufiane and Doucet, Arnaud and Rousseau, Judith , year =. The Curse of Depth in Kernel Regime , booktitle =

  61. [61]

    Deep Residual Learning for Image Recognition , booktitle =

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , year =. Deep Residual Learning for Image Recognition , booktitle =

  62. [62]

    2017 , edition =

    Matrix Analysis , author =. 2017 , edition =

    2017

  63. [63]

    1989 , journal =

    Multilayer Feedforward Networks Are Universal Approximators , author =. 1989 , journal =

    1989

  64. [64]

    1939 , journal =

    Tubes and Spheres in N-Spaces, and a Class of Statistical Problems , author =. 1939 , journal =

    1939

  65. [65]

    Regularization Matters:

    Hu, Tianyang and Wang, Wenjia and Lin, Cong and Cheng, Guang , year =. Regularization Matters:. International

  66. [66]

    2022 , month = may, number =

    Sharp Asymptotics of Kernel Ridge Regression beyond the Linear Regime , author =. 2022 , month = may, number =. arxiv , langid =:arXiv:2205.06798 , publisher =

    Pith/arXiv arXiv 2022

  67. [67]

    Dynamics of Deep Neural Networks and Neural Tangent Hierarchy , booktitle =

    Huang, Jiaoyang and Yau, Horng-Tzer , year =. Dynamics of Deep Neural Networks and Neural Tangent Hierarchy , booktitle =

  68. [68]

    Advances in Neural Information Processing Systems , author =

    Neural Tangent Kernel:. Advances in Neural Information Processing Systems , author =. 2018 , volume =

    2018

  69. [69]

    2022 , abstract =

    Generalization of Wide Neural Networks from the Perspective of Linearization and Kernel Learning , author =. 2022 , abstract =

    2022

  70. [70]

    , author =

    Distribution of Eigenvalues of Certain Integral Operators. , author =. 1955 , journal =

    1955

  71. [71]

    , year =

    Kanagawa, Motonobu and Hennig, Philipp and Sejdinovic, Dino and Sriperumbudur, Bharath K. , year =. Gaussian Processes and Kernel Methods:. arXiv preprint arXiv:1807.02582 , eprint =

    Pith/arXiv arXiv

  72. [72]

    A Style-Based Generator Architecture for Generative Adversarial Networks , booktitle =

    Karras, Tero and Laine, Samuli and Aila, Timo , year =. A Style-Based Generator Architecture for Generative Adversarial Networks , booktitle =

  73. [73]

    1991 , month = feb, journal =

    Asymptotic Properties of Eigenvalues of Integral Equations , author =. 1991 , month = feb, journal =. doi:10.1137/0151013 , abstract =

    work page doi:10.1137/0151013 1991

  74. [74]

    2000 , journal =

    Random Matrix Approximation of Spectra of Integral Operators , author =. 2000 , journal =. doi:10.2307/3318636 , abstract =

    doi:10.2307/3318636 2000

  75. [75]

    2017 , journal =

    Imagenet Classification with Deep Convolutional Neural Networks , author =. 2017 , journal =

    2017

  76. [76]

    2022 , journal =

    Moving Beyond Sub-Gaussianity in High-Dimensional Statistics: Applications in Covariance Estimation and Linear Regression , author =. 2022 , journal =

    2022

  77. [77]

    Nonparametric Regression with Shallow Overparameterized Neural Networks Trained by

    Kuzborskij, Ilja and Szepesv. Nonparametric Regression with Shallow Overparameterized Neural Networks Trained by. Conference on. 2021 , pages =

    2021

  78. [78]

    Generalization Ability of Wide Neural Networks on

    Lai, Jianfa and Xu, Manyun and Chen, Rui and Lin, Qian , year =. Generalization Ability of Wide Neural Networks on. doi:10.48550/arXiv.2302.05933 , abstract =. arxiv , keywords =:arXiv:2302.05933 , publisher =

    Pith/arXiv arXiv doi:10.48550/arxiv.2302.05933

  79. [79]

    2017 , journal =

    Deep Neural Networks as Gaussian Processes , author =. 2017 , journal =. 1711.00165 , archiveprefix =

    Pith/arXiv arXiv 2017

  80. [80]

    Wide Neural Networks of Any Depth Evolve as Linear Models under Gradient Descent , booktitle =

    Lee, Jaehoon and Xiao, Lechao and Schoenholz, Samuel and Bahri, Yasaman and Novak, Roman and. Wide Neural Networks of Any Depth Evolve as Linear Models under Gradient Descent , booktitle =. 2019 , volume =

    2019

Showing first 80 references.