Pith. sign in

REVIEW 2 major objections 6 minor 48 references

Learning convolution operators on compact Abelian groups

T0 review · 2 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read Ridge regression on translation-invariant Hilbert spaces learns convolution operators on compact Abelian groups at the same optimal-type rates as kernel ridge regression, with regularity read as space/frequency localization.

desk verdict A sound, carefully scoped re-derivation of ridge regression rates for convolution operators; the localization interpretation is the genuine contribution, and the input boundedness assumption is the honest price. read the letter →

arxiv 2501.05279 v3 pith:I4CGEEIU submitted 2025-01-09 cs.LG stat.ML

classification cs.LGstat.ML MSC 68T0547A5242B1062J07
keywords compactabeliangroupsconvolutionoperatorsridgeregressionlearningratestranslation-invariantHilbertspacesspace/frequencylocalizationoperatornonparametric
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 studies learning an unknown convolution operator from random input-output pairs, where the outputs are noisy convolutions of the inputs with an unknown kernel. It analyzes a ridge regression estimator over a translation-invariant Hilbert space of kernels and proves finite-sample high-probability bounds: with a suitable choice of the regularization parameter, the prediction error decays as $n^{-(2r+1)/(2r+1+b^{-1})}$ and the kernel error as $n^{-r/(2r+1+b^{-1})}$, where $r$ measures kernel smoothness and $b$ the eigenvalue decay of the input covariance. These are the same optimal-type rates as for classical kernel ridge regression, and the paper gives the standard regularity assumptions a new reading: they encode whether the input signals are localized in space or in frequency. The argument runs through Fourier analysis on compact Abelian groups, where convolution becomes pointwise multiplication and the estimator decouples into one scalar ridge problem per frequency. A sympathetic reader would care because convolution operators are ubiquitous (blur, heat kernels, Green's functions, linear time-invariant systems), and this provides a unified statistical guarantee with a concrete design implication: choose input localization according to the error metric you care about.

What carries the argument

The load-bearing identity is the Fourier diagonalization of convolution: on a compact Abelian group $G$, the convolution operator $C_w$ becomes multiplication by $\widehat w(\xi)$ on the dual group $\widehat G$. The hypothesis space $H$ is the translation-invariant Hilbert space of kernels with weight $\widehat K_\xi$, so the covariance operator $\Sigma$ is diagonal on the Fourier basis with eigenvalues $\sigma_\xi = \widehat K_\xi \mathbb E |\widehat X(\xi)|^2$. Ridge regression then separates frequency-by-frequency: the estimator (31) is a scalar ridge solution $\widehat w_\lambda(\xi) = \frac{\frac1n\sum \widehat Y_i(\xi)\overline{\widehat X_i(\xi)}}{\frac1n\sum |\widehat X_i(\xi)|^2+\lambda \widehat K_\xi^{-1}}$ at each character. The proof transfers the Caponnetto–De Vito error decomposition for kernel ridge regression to this setting, controls the sample covariance concentration by a matrix Bernstein inequality (Lemma A.7), and controls the noise term by a Hilbert-space Bernstein bound (Prop. A.8), with the trace term estimated via the eigenvalue decay (Prop. A.10). The operator norm of the convolution operator equals the $L^2$ norm of its kernel (Lemma A.11), which makes the two error norms in Theorem 4.1 comparable.

What would settle it

On the torus, draw inputs from a Gaussian random field (so $\|X\|_1$ is unbounded with positive probability), keep the same ridge estimator with the theoretical $\lambda$, and measure $\|\Sigma^{1/2}(w_{\lambda_n}^n-w^*)\|_H^2$ versus $n$: if the almost-sure bound is essential, the empirical tail probability will exceed $3e^{-\tau}$ and the rate will degrade. More directly, the proof's event $\Omega_{n,\lambda}$ in Lemma A.7 requires $\|\Phi(X)\|\le \kappa$ almost surely; a single unbounded input makes the event fail.

Watch

Extended reading notes

Core claim

The central claim is that ridge regression on a translation-invariant Hilbert space learns convolution operators on any compact Abelian group at the same rates as nonparametric regression, without assuming the hypothesis space is a reproducing kernel Hilbert space. Theorem 4.1 states that if the target kernel $w^*$ satisfies a source condition with exponent $r\in[0,1/2]$ and the covariance operator $\Sigma$ has eigenvalues decaying like $\ell^{-b}$, then with probability at least $1-3e^{-\tau}$ the estimator (29) with $\lambda$ as in (36) satisfies $\|\Sigma^{1/2}(w_{\lambda_n}^n-w^*)\|_H^2 = O(n^{-(2r+1)/(2r+1+b^{-1})})$ and $\|w_{\lambda_n}^n-w^*\|_H = O(n^{-r/(2r+1+b^{-1})})$. The same source and capacity assumptions that are classical in ridge regression theory translate, in this setting, into statements about space/frequency localization of the random inputs: frequency-localized inputs give faster decay of the prediction error, while space-localized inputs give faster recovery of the kernel in $H$-norm.

Load-bearing premise

The proof needs the random input signals to be almost surely bounded in $L^1$-norm (that is, $\|X\|_1\le D_X$); many natural signal distributions, like Gaussian random fields or heavy-tailed signals, fail this, and without it the matrix-Bernstein concentration argument that delivers the high-probability rates collapses.

Editorial extensions

If this is right

  • For the torus with periodic Sobolev hypothesis space $H^1$ and target $w^*\in H^2$, the rates become explicit: frequency-localized inputs give prediction error decaying as $n^{-5/6}$, while space-localized inputs give $n^{-4/5}$; the ordering reverses for the $H$-norm kernel error.
  • The Fourier-domain formula (31) means the estimator can be computed by FFT and per-frequency scalar ridge solves, making the method scalable to large grids.
  • Because the analysis needs only translation invariance, not an RKHS structure, it covers hypothesis spaces like $L^2$ itself and non-RKHS Sobolev spaces.
  • On finite cyclic groups the estimator is the inverse of a regularized circulant matrix, which connects the rates to the classical problem of identifying linear time-invariant systems from random impulse-like inputs.

Reading between the lines

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

  • The paper's interpretation suggests a design principle the authors state only implicitly: for predicting future outputs, sample with frequency-localized inputs, and for identifying the kernel or system, sample with space-localized inputs; this could be turned into an active-learning strategy for convolution operators.
  • The rates are proven under an almost-sure $L^1$ bound on inputs; a plausible extension to sub-Gaussian processes would need a different concentration argument, and testing whether the same exponents survive would settle how essential the boundedness assumption is.
  • The authors note their heat-kernel experiment is beyond the theory because the group is $\mathbb R$; a direct generalization to locally compact abelian groups would replace the counting measure on $\widehat G$ by a Haar measure and likely keep the same rate formula with the trace condition interpreted as Plancherel weights.
  • The theory says $\lambda \asymp n^{-1/(2r+1+b^{-1})}$, and since $r,b$ are unknown in practice, the simulation's grid-search heuristic could be replaced by a data-driven rule, such as a Lepski-type or marginal-likelihood choice, whose guarantees follow from the same bounds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 6 minor

Summary. The paper studies the statistical problem of learning a convolution operator C_w* from L1(G) to L2(G) on a compact Abelian group G, given i.i.d. noisy input-output pairs (X_i, Y_i) with Y_i = X_i * w* + noise, where w* lies in a translation-invariant Hilbert space H. The estimator is the ridge regression estimator (29), which has an explicit Fourier-domain form (Prop. A.4). The main result, Theorem 4.1, gives high-probability non-asymptotic bounds: under an almost-sure L1-boundedness assumption on X (eq. (11)), a source condition (34), and an eigenvalue decay condition (35), it states that the prediction error E[||C_{λ_n} X - C* X||_2^2 | Z] = ||Σ^{1/2}(w_{λ_n} - w*)||_H^2 is bounded by n^{-(2r+1)/(2r+1+b^{-1})} (up to log factors in a special case), and that the H-norm error ||w_{λ_n} - w*||_H is bounded by n^{-r/(2r+1+b^{-1})}. Section 4.2 interprets r and b as space/frequency localization parameters of the input signals, and Section 5 contains numerical simulations, including a heat-kernel example.

Significance. If the result holds, the paper extends classical optimal-type ridge-regression rates to a structured operator-learning problem, with a new and insightful interpretation of source and capacity conditions in terms of space/frequency localization of the input signals. The proofs are detailed and self-contained, relying on standard matrix Bernstein and Pinelis concentration inequalities, and the estimator has a simple closed form in the Fourier domain. The paper ships reproducible code and works through instructive examples (torus, circulant matrices, Sobolev spaces). The main caveat is that the almost-sure L1-boundedness of the inputs (Assumption (11)) is essential to the concentration argument and substantially restricts the applicability; this should be stated prominently. On balance, the contribution is solid and publishable after a careful revision that addresses scope and presentation issues.

major comments (2)
  1. [Abstract and Section 3.1, Eq. (11)] The abstract and introduction present the results as finite-sample guarantees under 'natural regularity conditions on the convolution kernel,' but the almost-sure L1-boundedness of the inputs (Assumption (11), ||X||_1 ≤ D_X) is a key structural assumption that is not mentioned there. This bound is what makes the feature map uniformly bounded (Theorem A.3, eq. (54)) and thereby enables the matrix Bernstein concentration argument in Lemma A.7; without it, the stated high-probability rates do not follow. Many natural input models, such as Gaussian random fields or heavy-tailed signals, violate (11). The authors should state this assumption in the abstract, and either discuss possible relaxations or temper the claim that the results apply under natural regularity conditions.
  2. [Section 4.2 and Eq. (38)] The bound (38) on the H-norm error is not tight as stated. In the proof, using eq. (73) together with Prop. A.8, the variance term (1/√λ)||(Σ+λI)^{-1/2}S_n^*ε|| gives a term of order τ/(λn), which for λ as in (36) decays as n^{-(2r+b^{-1})/(2r+1+b^{-1})}, faster than the stated rate n^{-r/(2r+1+b^{-1})} whenever r>0 or b<∞. For example, in the finite-dimensional case (b=∞, r=1/2), the proof yields a rate n^{-1/2} for ||w_λ - w*||_H, whereas (38) states n^{-1/4}. Since Section 4.3 refers to 'sharp bounds,' the authors should either sharpen (38) to the bound actually proved, or explicitly describe it as a simplified upper bound and avoid the sharpness claim for this norm.
minor comments (6)
  1. [Theorem 4.1, Eq. (36)] The definition of λ_n is corrupted: the line 'Cn = Cwλnn' appears to be a typesetting error, and the display alignment of the cases is broken. Please restate the theorem with clean notation, distinguishing λ_n from the operator C_λ^n.
  2. [Lemma A.7 and Theorem A.12] The notation σ_supp(W) ≤ M is nonstandard; since W can have large negative eigenvalues, the matrix Bernstein condition should be stated as λ_max(W) ≤ M. The proof in Lemma A.7 uses the correct reasoning (W ≤ (Σ+λI)^{-1}Σ), but Theorem A.12's statement should be clarified to avoid confusion between spectral radius and largest eigenvalue.
  3. [Section 5.1, Figure 1] The text says 'Dotted lines indicate the theoretical convergence rates for reference,' but it is not specified whether the dotted lines correspond to eq. (37), eq. (38), or their square roots, nor what constants are used. Please clarify the figure legend and caption.
  4. [Section 5.2] The heat-equation example uses G = R, which the paper correctly notes is outside the compact-group theory. This is acceptable as a motivating illustration, but the deviation from the theoretical setting should be flagged even more prominently to avoid overclaiming.
  5. [Section 4.1, Eq. (35)] For 1 < b < ∞, the notation σ_ℓ ≲ 1/ℓ^b presumes an ordering of the eigenvalues; please state explicitly that the eigenvalues are arranged in nonincreasing order.
  6. [Section 2, Notation] There are several typographical issues in the notation, including 'A˚ : F˚Ñ A˚' (the arrow direction is wrong) and 'sequilinear duality pairing' (should be 'sesquilinear'). These should be corrected in a final pass.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Theorem 4.1 is a specialization of established ridge-regression bounds to convolution operators, with all assumptions explicit and rates not fitted.

full rationale

The derivation is self-contained given standard ridge-regression theory. Theorem 4.1 takes source condition (34), capacity condition (35), and bounded-input assumption (11) as explicit inputs, and derives rates (37)-(38) through the error decomposition of Proposition A.5, matrix-Bernstein concentration (Lemma A.7/A.12), and the trace bound (Proposition A.10). The regularization parameter (36) is prescribed in terms of n, r, b, and kappa, and is not fitted to data; the Section 4.2 localization comparisons are closed-form consequences of the spectrum of Sigma in Examples 3 and 4, not fitted outputs. The cited results [7], [13], [36], and [37] are existing published lemmas and theorems used as tools; even where authorship overlaps with the present paper, they are parameter-free results with stated assumptions that do not include the target convolution-operator rates, so they constitute independent support rather than a circular self-citation chain. No equation in the paper reduces to its own input by construction, and the numerical experiments use a separate grid-search lambda and compare to the theoretical exponents only as validation.

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

No new entities or fitted constants are introduced. The theoretical rates are derived from standard ridge-regression theory under explicit source and capacity assumptions; the only tuning parameter lambda is set by formula (36) in the theory and by grid search in experiments.

assumptions (6)
  • domain assumption Input X is bounded in L1: ||X||_1 <= D_X almost surely (eq. 11).
    Assumed in Section 3.1; used to bound the feature map Phi and to obtain concentration inequalities; the rates in Theorem 4.1 depend on kappa = D_X D_K.
  • domain assumption Noise satisfies the moment condition E[||epsilon||^m | X] <= (m!/2) M_epsilon^{m-2} sigma_epsilon^2 (eq. 13).
    Standard sub-Gaussian-type condition used in Prop A.8 for the noise term bound.
  • domain assumption The target w* satisfies the source condition (34) with r in [0, 1/2].
    Assumed in Theorem 4.1; controls the approximation error through Prop A.9.
  • domain assumption The eigenvalues sigma_xi of Sigma satisfy the decay condition (35) with b in [1, infinity].
    Assumed in Theorem 4.1; yields the trace bound in Prop A.10.
  • standard math Standard harmonic analysis facts on compact Abelian groups (Fourier inversion, Young's inequality, convolution theorem).
    Used throughout Section 2 and the proofs.
  • standard math Matrix Bernstein (Tropp) and Pinelis concentration inequalities.
    Used in Lemma A.7 and Prop A.8 for high-probability bounds.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning convolution operators on compact Abelian groups." pith.science (2026). https://pith.science/paper/I4CGEEIU

@misc{pith2026250105279,
  author       = {Pith},
  title        = {Pith review of: Learning convolution operators on compact Abelian groups},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I4CGEEIU}},
  note         = {Machine review of arXiv:2501.05279}
}
read the original abstract

We consider the problem of learning convolution operators associated to compact Abelian groups. We study a regularization-based approach and provide corresponding learning guarantees under natural regularity conditions on the convolution kernel. More precisely, we assume the convolution kernel is a function in a translation invariant Hilbert space and analyze a natural ridge regression (RR) estimator. Building on existing results for RR, we characterize the accuracy of the estimator in terms of finite sample bounds. Interestingly, regularity assumptions which are classical in the analysis of RR, have a novel and natural interpretation in terms of space/frequency localization. Theoretical results are illustrated by numerical simulations.

Figures

Figures reproduced from arXiv: 2501.05279 by the authors.

Figure 1
Figure 1. Error decay. (Left) }w λ n ´ w˚} 2 H. (Right) }Σ 1 2 pw λ n ´ w˚q}2 H. Each curve compares frequency-localized vs. space-localized inputs. Dotted lines indicate the the￾oretical convergence rates for reference. partial differential equations, namely the recovery of a fundamental solution of the heat equation. This example goes beyond the theoretical results we stated in this paper since the group is R, but it furthe… view at source ↗
Figure 2
Figure 2. Example of reconstruction. Comparison of the true w˚ with the esti￾mated w λn n for n “ 50 in both input scenarios. Implementation details. For each scenario (frequency vs. space localization), our numerical work￾flow is as follows: 1. Generate data: Draw tXku n k“1 according to the chosen localization distribution and compute Yk “ C˚Xk ` εk, k “ 1, . . . , n. 2. Estimate w˚: Use the ridge regression formula in the … view at source ↗
Figure 4
Figure 4. Corresponding convo￾lution operators (circulant matri￾ces). function) for a classic PDE, namely the heat equation. Concretely, consider the initial value problem Bu Bt px, tq ´ ∆upx, tq “ 0 px, tq P R D ˆ p0, 8q, upx, 0q “ gpxq x P R D. (41) The function upx, tq represents the temperature at location x P R D and time t ě 0. The function g specifies the temperature at t “ 0. It is well known (see [39]) that the solut… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 44 canonical work pages

  1. [1]

    Solving inverse prob- lems using data-driven models

    Simon Arridge, Peter Maass, Ozan ¨Oktem, and Carola-Bibiane Sch¨ onlieb. Solving inverse prob- lems using data-driven models. Acta Numerica, 28:1–174, 2019

  2. [2]

    Harmonic Analysis on Semigroups: Theory of Positive Definite and Related Functions

    Christian Berg, Jens Peter Reus Christensen, and Paul Ressel. Harmonic Analysis on Semigroups: Theory of Positive Definite and Related Functions. Graduate Texts in Mathematics. Springer, 1984

  3. [3]

    A double regularization approach for inverse problems with noisy data and inexact operator

    Ismael Rodrigo Bleyer and Ronny Ramlau. A double regularization approach for inverse problems with noisy data and inexact operator. Inverse Problems, 29(2):025004, 2013

  4. [4]

    Learning Green’s functions associ- ated with time-dependent partial differential equations

    Nicolas Boull´ e, Seick Kim, Tianyi Shi, and Alex Townsend. Learning Green’s functions associ- ated with time-dependent partial differential equations. Journal of Machine Learning Research, 23(218):1–34, 2022

  5. [5]

    Brunton and J

    Steven L. Brunton and J. Nathan Kutz. Modern Data-Driven Modeling and Control: With Applications to Engineering and Science. Cambridge University Press, 2022

  6. [6]

    Regularization methods for blind deconvolution and blind source separation problems

    Martin Burger and Otmar Scherzer. Regularization methods for blind deconvolution and blind source separation problems. Mathematics of Control, Signals and Systems, 14:358–383, 2001

  7. [7]

    Optimal rates for the regularized least-squares algo- rithm

    Andrea Caponnetto and Ernesto De Vito. Optimal rates for the regularized least-squares algo- rithm. Foundations of Computational Mathematics, 7:331–368, 2007

  8. [8]

    Vector valued reproducing kernel Hilbert spaces of integrable functions and Mercer theorem

    Claudio Carmeli, Ernesto De Vito, and Alessandro Toigo. Vector valued reproducing kernel Hilbert spaces of integrable functions and Mercer theorem. Anal. Appl. (Singap.), 4(4):377–408, 2006

Show all 48 references
  1. [9]

    Fast motion deblurring

    Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Trans. Graph., 28(5), 2009. LEARNING CONVOLUTION OPERATORS ON COMPACT ABELIAN GROUPS 25

  2. [10]

    Asymptotics of prediction in functional linear regression with functional outputs

    Christophe Crambes and Andr´ e Mas. Asymptotics of prediction in functional linear regression with functional outputs. Bernoulli, 19(5B):2627 – 2651, 2013

  3. [11]

    Conver- gence rates for learning linear operators from noisy data

    Maarten V de Hoop, Nikola B Kovachki, Nicholas H Nelsen, and Andrew M Stuart. Conver- gence rates for learning linear operators from noisy data. SIAM/ASA Journal on Uncertainty Quantification, 11(2):480–513, 2023

  4. [12]

    Nonlinear image processing using artificial neural networks

    Dick De Ridder, Robert PW Duin, Michael Egmont-Petersen, Lucas J Van Vliet, and Piet W Verbeek. Nonlinear image processing using artificial neural networks. In Advances in Imaging and Electron Physics, volume 126, pages 351–450. Elsevier, 2003

  5. [13]

    Learning from examples as an inverse problem

    Ernesto De Vito, Lorenzo Rosasco, Andrea Caponnetto, Umberto De Giovannini, Francesca Odone, and Peter Bartlett. Learning from examples as an inverse problem. Journal of Machine Learning Research, 6(5), 2005

  6. [14]

    Egmont-Petersen, D

    M. Egmont-Petersen, D. de Ridder, and H. Handels. Image processing with neural networks—a review. Pattern Recognition, 35(10):2279–2301, 2002

  7. [15]

    Gerald B. Folland. A course in abstract harmonic analysis. Textbooks in Mathematics. CRC Press, Boca Raton, FL, second edition, 2016

  8. [16]

    Physics-informed deep neural operator networks

    Somdatta Goswami, Aniruddha Bora, Yue Yu, and George Em Karniadakis. Physics-informed deep neural operator networks. In Scientific Machine Learning, volume 143 of Lecture Notes in Computational Science and Engineering, pages 223–250. Springer, 2023

  9. [17]

    I. S. Gradshteyn and I. M. Ryzhik. Table of integrals, series, and products. Elsevier/Academic Press, Amsterdam, seventh edition, 2007. Translated from the Russian, Translation edited and with a preface by Alan Jeffrey and Daniel Zwillinger, With one CD-ROM (Windows, Macintosh...

  10. [18]

    C. W. Groetsch. The theory of Tikhonov regularization for Fredholm equations of the first kind, volume 105 of Research Notes in Mathematics. Pitman (Advanced Publishing Program), Boston, MA, 1984

  11. [19]

    A note on estimation in Hilbertian linear models

    Siegfried H¨ ormann and Lukasz Kidzi´ nski. A note on estimation in Hilbertian linear models. Scandinavian journal of statistics, 42(1):43–62, 2015

  12. [20]

    A general framework for soft-shrinkage with applications to blind deconvolution and wavelet denoising

    L Justen and R Ramlau. A general framework for soft-shrinkage with applications to blind deconvolution and wavelet denoising. Applied and Computational Harmonic Analysis, 26(1):43– 63, 2009

  13. [21]

    Operator-valued kernels for learning from functional response data

    Hachem Kadri, Emmanuel Duflos, Philippe Preux, St´ ephane Canu, Alain Rakotomamonjy, and Julien Audiffren. Operator-valued kernels for learning from functional response data. Journal of Machine Learning Research, 17(20):1–54, 2016

  14. [22]

    Neural operator: Learning maps between function spaces

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, An- drew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces. Journal of Machine Learning Research, 24(146):1–64, 2023

  15. [23]

    Kovachki, Samuel Lanthaler, and Andrew M

    Nikola B. Kovachki, Samuel Lanthaler, and Andrew M. Stuart. Operator learning: Algorithms and analysis. arXiv:2402.15715, 2024

  16. [24]

    Kundur and D

    D. Kundur and D. Hatzinakos. Blind image deconvolution. IEEE Signal Processing Magazine, 13(3):43–64, 1996

  17. [25]

    An RKHS framework for functional data analysis

    Ana Kupresanin, Hyejin Shin, David King, and RL Eubank. An RKHS framework for functional data analysis. Journal of statistical planning and inference, 140(12):3627–3637, 2010

  18. [26]

    Minimax prediction for functional linear regression with functional responses in reproducing kernel hilbert spaces

    Heng Lian. Minimax prediction for functional linear regression with functional responses in reproducing kernel hilbert spaces. Journal of Multivariate Analysis, 140:395–402, 2015

  19. [27]

    System Identification: Theory for the User

    Lennart Ljung. System Identification: Theory for the User. Prentice Hall PTR, 1998

  20. [28]

    Linear processes for functional data

    Andr´ e Mas and Besnik Pumo. Linear processes for functional data. Oxford Handbooks Online, 2009

  21. [29]

    On some extensions of Bernstein’s inequality for self-adjoint operators

    Stanislav Minsker. On some extensions of Bernstein’s inequality for self-adjoint operators. Statistics & Probability Letters, 127:111–119, 2017

  22. [30]

    Learning linear operators: Infinite- dimensional regression as a well-behaved non-compact inverse problem

    Mattes Mollenhauer, Nicole M¨ ucke, and TJ Sullivan. Learning linear operators: Infinite- dimensional regression as a well-behaved non-compact inverse problem. arXiv:2211.08875, 2022

  23. [31]

    Optimum bounds for the distributions of martingales in Banach spaces

    Iosif Pinelis. Optimum bounds for the distributions of martingales in Banach spaces. Ann. Probab., 22(4):1679–1706, 1994

  24. [32]

    Optimum bounds for the distributions of martingales in Banach spaces

    Iosif Pinelis. Correction: “Optimum bounds for the distributions of martingales in Banach spaces” [Ann. Probab. 22 (1994), no. 4, 1679–1706; MR1331198 (96b:60010)]. Ann. Probab., 27(4):2119, 1999. 26 EMILIA MAGNANI, ERNESTO DE VITO, PHILIPP HENNIG, AND LORENZO ROSASCO

  25. [33]

    J. O. Ramsay and B. W. Silverman. Functional Data Analysis. Springer Series in Statistics. Springer, New York, second edition, 2005

  26. [34]

    Functional regression with repeated eigenvalues

    Matthew Reimherr. Functional regression with repeated eigenvalues. Statistics and Probability Letters, 107:62–70, 2015

  27. [35]

    Model selection and error esti- mation for regularized least-squares algorithm in learning theory

    Lorenzo Rosasco, Ernesto De Vito, and Alessandro Caponnetto. Model selection and error esti- mation for regularized least-squares algorithm in learning theory. Foundations of Computational Mathematics, 8(5):571–607, 2008

  28. [36]

    Less is more: Nystr¨ om computa- tional regularization

    Alessandro Rudi, Raffaello Camoriano, and Lorenzo Rosasco. Less is more: Nystr¨ om computa- tional regularization. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015

  29. [37]

    On the sample complexity of subspace learning

    Alessandro Rudi, Guille D Canas, and Lorenzo Rosasco. On the sample complexity of subspace learning. arXiv:1408.5032, 2014

  30. [38]

    Fourier Analysis on Groups

    Walter Rudin. Fourier Analysis on Groups. Wiley Classics Library. John Wiley & Sons, New York, 1962

  31. [39]

    Partial differential equations in action: from modelling to theory, volume 99

    Sandro Salsa. Partial differential equations in action: from modelling to theory, volume 99. Springer, 2016

  32. [40]

    Dis- criminative non-blind deblurring

    Uwe Schmidt, Carsten Rother, Sebastian Nowozin, Jeremy Jancsary, and Stefan Roth. Dis- criminative non-blind deblurring. In 2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 604–611, 2013

  33. [41]

    Learning to deblur

    Christian J Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Sch¨ olkopf. Learning to deblur. IEEE transactions on pattern analysis and machine intelligence, 38(7):1439–1451, 2015

  34. [42]

    Sous-espaces hilbertiens d’espaces vectoriels topologiques et noyaux associ´ es (noyaux reproduisants)

    Laurent Schwartz. Sous-espaces hilbertiens d’espaces vectoriels topologiques et noyaux associ´ es (noyaux reproduisants). Journal d’Analyse Math´ ematique, 13:115—-256, 1964

  35. [43]

    Learning Schatten–von Neumann oper- ators

    Puoya Tabaghi, Maarten de Hoop, and Ivan Dokmani´ c. Learning Schatten–von Neumann oper- ators. arXiv:1901.10076, 2019

  36. [44]

    Bayesian inverse problems with unknown operators

    Mathias Trabs. Bayesian inverse problems with unknown operators. Inverse Problems, 34(8), 2018

  37. [45]

    User-friendly tools for random matrices: An introduction

    Joel A Tropp. User-friendly tools for random matrices: An introduction. NIPS Tutorial, 3, 2012

  38. [46]

    Two-phase kernel estimation for robust motion deblurring

    Li Xu and Jiaya Jia. Two-phase kernel estimation for robust motion deblurring. In Kostas Daniilidis, Petros Maragos, and Nikos Paragios, editors, Computer Vision – ECCV 2010, pages 157–170. Springer Berlin Heidelberg, 2010

  39. [47]

    Tony Cai

    Ming Yuan and T. Tony Cai. A reproducing kernel Hilbert space approach to functional linear regression. The Annals of Statistics, 38(6):3412 – 3444, 2010

  40. [48]

    Minimax rate for learning kernels in operators

    Sichong Zhang, Xiong Wang, and Fei Lu. Minimax rate for learning kernels in operators. arXiv:2502.20368, 2025. Emilia Magnani, T ¨ubingen AI Center, University of T ¨ubingen, T¨ubingen, Germany Email address : emilia.magnani@uni-tuebingen.de E. De Vito, MaLGa,, DIMA, Universit...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.