Pith. sign in

REVIEW 3 major objections 4 minor 41 references

This paper proves an end-to-end total-variation bound for score-based diffusion models, replacing the oracle-score assumption with explicit finite-sample, discretization, and optimization terms.

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 · deepseek-v4-flash

2026-08-01 00:03 UTC pith:GMRPA43M

load-bearing objection A genuinely synthetic end-to-end TV bound for ResNet score networks, undermined by assumptions that exclude the OU-type processes it claims to cover. the 3 major comments →

arxiv 2607.23226 v1 pith:GMRPA43M submitted 2026-07-25 cs.LG math.STstat.TH

From Score Learning to Discretized Sampling: An End-to-End Generalization Analysis of Diffusion Models

Jinshu Huang , Yiming Jiang , Chunlin Wu This is my paper
classification cs.LG math.STstat.TH MSC 65R2068T0737N9945G10
keywords score-based diffusion modelsend-to-end generalizationtotal variation distancedenoising score matchingResNet score networksEuler–Maruyama discretizationRademacher complexityGirsanov path measure
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.

The paper aims to close the gap between score learning and sampling theory in diffusion models. It claims that for a ResNet-type score network trained on S i.i.d. samples using a discretized denoising score-matching loss on an N-step time grid, the total-variation distance between the distribution generated by Euler–Maruyama sampling and the true data distribution is bounded by four explicit, interpretable terms: forward-process truncation, reverse-time discretization, finite-sample generalization, and optimization error. The bound is end-to-end in the sense that no oracle score function is assumed; every term is either an explicit function of S, N, network parameters, and SDE coefficients, or the empirical training loss itself. If the theorem is correct, it turns theoretical diffusion guarantees into a concrete diagnostic: practitioners can see how sample size, grid resolution, and training accuracy trade off against final sample quality.

Core claim

The central claim, Theorem 4.9, is an inequality of the form TV(ˇp(T,·), p_data) ≤ TV(π, p(T,·)) + C8(...) + C9 h_N^{1/4} + O(S^{-1/4}) + sqrt(T/2 (ℓ^DSM_{N,S}(Θ)-ℓ*_{N,S})) + sqrt(T ℓ*_{N,S}/2), where h_N is the reverse-time step, S is the sample size, ℓ^DSM is the empirical discrete-time denoising score-matching loss, and ℓ* is its optimum over the parameter class. The proof obtains this by first bounding KL between the true reverse path measure and the learned continuous-time path measure by T/2 times the explicit score-matching risk, relating that to the denoising objective, controlling the Euler–Maruyama discretization with a Girsanov-type argument, and controlling the finite-sample gap

What carries the argument

The chain TV ≤ sqrt(KL/2) is the engine: KL(µ∥ν) ≤ T/2 ℓ_ESM(Θ) via Girsanov; KL(ν∥ν_N,S) is bounded by the Euler–Maruyama drift mismatch via Girsanov; ℓ_ESM is compared to ℓ_DSM through the denoising score-matching identity; and ℓ_DSM − ℓ* is controlled by a Rademacher-complexity generalization bound for ResNet score networks. The total-variation bound then follows from Pinsker's inequality and the triangle inequality on path measures.

Load-bearing premise

The argument assumes the conditional score ∇ln p(t,·|x) is Lipschitz in space and time with a finite constant uniformly for all t∈[0,T] and that the joint density p(0,x)p(t,x_t|x) admits an integrable dominating function; for the OU-type forward processes highlighted in the paper (VP, VE, CDPM), the conditional score is (r(t)x - x_t)/(r^2(t)v^2(t)) with v(t)→0, so that constant is infinite at t=0, making this assumption fail for exactly the standard examples, and Lemma A.1's

What would settle it

For the VP-SDE, choose any smooth non-atomic data distribution and evaluate the conditional score near t=0: ∇ln p(t,x|y) ≈ -(x - r(t)y)/(r^2(t)v^2(t)) with v(t)=√(∫_0^t g^2/r^2) ~ √t, so the norm is O(1/√t) and the Lipschitz constant in x is O(1/(r^2 v^2)) → ∞ as t→0; no finite C_p satisfies (13)-(14) on [0,T]. Also, Lemma A.1's bounding function B_grad(t)=(2π r^2(t)v^2(t))^{-d/2}(r(t)v(t))^{-1}e^{-1/2} diverges at t=0, so it cannot supply the integrable dominating function h required by (29).

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

If this is right

  • If the bound holds, sample quality improves at rate S^{-1/4} in total variation with more training data, and at rate h_N^{1/4} with finer reverse-time grids; the slower rate comes from converting KL control into TV.
  • Near-optimal training matters: the optimization gap enters as sqrt(ℓ^DSM(Θ)-ℓ*), so the benefit of extra data is lost unless the optimizer is close to the empirical optimum.
  • The prior-mismatch term TV(π∥p(T,·)) is the only term independent of learning; for contractive forward SDEs it decays with T, so the framework recovers the standard long-horizon truncation trade-off.
  • The moduli of continuity of the forward drift, diffusion coefficient, and conditional score enter directly, so smoothing these coefficients or the weighting λ reduces the discretization component.
  • The generalization bound is uniform over the whole bounded-parameter ResNet class, so the guarantee holds simultaneously for all score networks in that class, not just for the trained optimum.

Where Pith is reading between the lines

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

  • Editorial inference: the theorem's assumption (A4)/(29) fails for the standard VP, VE, and CDPM forward processes near t=0, where the conditional score diverges as v(t)→0; the theorem should be read as applying to forward processes with a positive noise floor or to a t≥ε analysis, a repair the paper does not make explicitly.
  • Editorial inference: the O(S^{-1/4}) rate is an artifact of converting KL control into a TV bound; a direct L2-score-to-TV contraction argument, if available, would likely restore S^{-1/2}, so the displayed rate is not necessarily an information-theoretic limit.
  • Editorial inference: because the optimization gap remains an unmodeled external term, the end-to-end guarantee is conditional on the optimizer reaching low empirical loss; analyzing non-convex training dynamics is the paper's acknowledged next step.
  • Editorial inference: a practical diagnostic suggested by the decomposition is to measure the four terms separately on a given dataset and grid — for example, estimate TV(π∥p(T,·)) by Monte Carlo and the score generalization gap by hold-out data — to decide whether to add data, refine the grid, or train longer.

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

3 major / 4 minor

Summary. The paper develops an end-to-end theoretical analysis of score-based diffusion models with ResNet-type score parameterizations. It defines population, finite-sample, and fully discrete denoising score matching problems, proves uniform generalization bounds for the score network class, establishes consistency of empirical minimizers, and then propagates the resulting score error through an Euler–Maruyama discretization of the reverse SDE. The main result, Theorem 4.9, bounds the total variation distance between the generated terminal distribution and the data distribution by a sum of four interpretable terms: prior truncation, discretization/regularity, generalization, and optimization errors. The proof is a self-contained chain of lemmas with explicit constants, and it does not rely on an oracle score. However, the hypotheses under which the main theorem is proved are incompatible with the OU-type forward processes highlighted in Example 3.1, and the manuscript's own Lemma A.1 does not repair this mismatch.

Significance. If the result held under its stated hypotheses for the advertised examples, it would be a useful contribution: it would quantify, in one bound, the interaction of sample size, time discretization, network architecture, and optimization gap, and it would go beyond oracle-score analyses. The paper has real strengths: the proof structure is transparent, the constants are explicit, the generalization analysis covers a practical ResNet parameterization, and the decomposition into truncation/discretization/generalization/optimization errors is conceptually clean. The central problem is that the main theorem's assumptions exclude the canonical variance-preserving, variance-exploding, and CDPM forward processes that the paper explicitly claims as instantiations. This is not merely a technical nuisance; it affects the well-posedness of the training objective and the validity of the key inequality (29). The significance is therefore conditional until the assumptions are repaired or the theorem is restated with an epsilon-truncation.

major comments (3)
  1. [Section 4.1, Assumption (A4), and Example 3.1] Assumption (A4), specifically the finite Lipschitz/bound constant C_p in Eqs. (13)-(14), fails for the OU-type processes advertised in Example 3.1. For Eq. (A.4), the conditional score is ∇ln p(t,x_t|x) = -(x_t-r(t)x)/(r^2(t)v^2(t)). For VP and VE schedules, r(t)v(t)→0 as t→0, so the constant C_p would have to be infinite. Consequently, the conditional-score bound in (14) is not available on [0,T]. This is not a cosmetic issue: the squared conditional score has expectation d/(r^2(t)v^2(t)), which is non-integrable at t=0 for VP/VE, so the DSM objectives in (6) and (8) can be infinite for all Θ. Thus the learning problems (P) and (P_N,S) are not well-posed for the paper's flagship examples as stated.
  2. [Lemma A.1(iii) and Assumption (29)] Lemma A.1(iii) does not establish the domination condition (29) needed in Theorem 4.9. The proposed h(t,x)=B_grad(t)p(0,x) has B_grad(t)=O((r(t)v(t))^{-(d+1)}), which diverges as t→0. Hence h is not a finite-valued dominating function on [0,T]×R^d, and the interchange in (A.8) is not justified. Moreover, (29) is used in the proof of Theorem 4.9 Step 2 to justify exchanging the integral over x with the gradient in x_t, but the condition as written bounds p(0,x)p(t,x_t|x) rather than the gradient p(0,x)∇_{x_t}p(t,x_t|x) that actually appears; Lemma A.1(iii) only provides a pointwise-in-t gradient bound. As a result, the proof of the key identity (61) is incomplete for the examples the paper claims to cover.
  3. [Equation (8) and the discrete loss at k=0] The empirical DSM loss (8) contains the k=0 term with t_0^N=0. Under Assumption (A3), λ(0)>0. For OU-type forward processes, the conditional density at t=0 is degenerate, and ∇ln p(0,x|x) is undefined/infinite. Even if one ignored this single point, the integral in (6) over [0,T] diverges because E∥∇ln p(t,x_t|x)∥² = d/(r²(t)v²(t)) ~ 1/t for VE/VP. The paper needs either to restrict the time grid to t≥ε for some ε>0, to assume a positive noise floor v(t)≥δ>0, or to introduce an explicit ε-truncation error term. Without such a modification, Theorem 4.9 and the learning-theoretic results that feed into it do not apply to the standard instantiations in Example 3.1.
minor comments (4)
  1. [Theorem 4.9, Eq. (30) and proof of Lemma 4.8] In the proof of Theorem 4.9, applying Pinsker to Lemma 4.8's KL(ν∥ν_N,S)≤C_5 h_N should produce a √(C_5/2) h_N^{1/2} term, but Eq. (30) and the definition of C_9 contain h_N^{1/4}. If the h_N^{1/4} is intended as a conservative upper bound for h_N^{1/2} when h_N≤1, this should be stated explicitly; otherwise the displayed rate is inconsistent.
  2. [Eq. (30)] The term √(ω_g²(h_N)) in Eq. (30) equals ω_g(h_N) and is redundant. The authors probably intend a modulus of continuity for g² or a separate term; please clarify the notation.
  3. [Notation, Section 1] The notation paragraph defines the Lipschitz constant twice ('For a Lipschitz continuous function g, its Lipschitz constant is denoted Lip g' followed by the same sentence with f). This should be cleaned up.
  4. [Lemma 4.4] The proof of Lemma 4.4 is omitted with 'Here, we omit the details.' Since the existence of minimizers is used in the consistency argument, a short proof or a precise reference should be supplied.

Circularity Check

0 steps flagged

No significant circularity: the end-to-end bound is a self-contained theorem chain whose remaining terms (optimization gap, empirical optimum) are explicit inputs, not hidden fitted predictions; the flagged assumption gap is a validity/scope issue, not circularity.

full rationale

The paper's central derivation is a chain of independent estimates: temporal discretization of the DSM loss (Lemma 4.5), a Rademacher-complexity generalization bound (Lemma 4.6), a Girsanov/DP-type KL bound between the exact and learned reverse processes (Step 1 of Theorem 4.9), and an Euler-Maruyama discretization bound (Lemma 4.8), chained through Pinsker and data-processing inequalities to give the TV bound (30). No fitted parameter is renamed as a prediction: the term sqrt(T ℓ*_N,S/2) is explicitly the optimal value of the empirical problem (P_N,S), and sqrt(T/2 (ℓDSM_N,S(Θ)-ℓ*_N,S)) is explicitly the optimization gap; these are acknowledged decomposition terms, not circular inputs. The only self-citation ([27]) appears in a literature-survey sentence about Γ-convergence and is not load-bearing. The manuscript's own limitation passages—the conclusion that the optimization error is left as a decoupled open challenge, and the Appendix A.1 attempt to verify (29) with a dominating function B_grad(t) that blows up as t→0—point to a real hypothesis/scope mismatch: Assumptions (A4) and (29) are not established for the OU-type processes advertised in Example 3.1. But that is a correctness or validity gap in the proof, not a circularity: Theorem 4.9 does not assume its own conclusion or define its target in terms of its inputs. The score is therefore 0.

Axiom & Free-Parameter Ledger

1 free parameters · 8 axioms · 0 invented entities

The paper rests on standard diffusion-model regularity assumptions plus a network-parameter bound. No new physical/mathematical entities are introduced. The most fragile inputs are (A4) and (29), which fail for the OU-type forward processes the paper highlights as its main examples; this is an assumption-vs-example mismatch rather than a fitted-parameter issue.

free parameters (1)
  • B_Θ (network parameter norm bound)
    Introduced in (A5) to compactify the parameter space; chosen by hand, controls all constants via exp(L Lip_ψ B_Θ²).
axioms (8)
  • domain assumption (A1) data distribution supported on B_R(0)
    Restricts p_data to a bounded ball; used for moment bounds and a.s. bounds on the network output.
  • domain assumption (A2) drift f ∈ C^1 with linear growth, Lipschitz in x, modulus-of-continuity in t
    Ensures SDE well-posedness and moment bounds; automatic for OU-type f(t)x.
  • domain assumption (A3) λ, g continuous and positive on [0,T]
    Needed for KL discretization bound and weighting; excludes g that vanish on [0,T].
  • domain assumption (A4) marginal and conditional score satisfy uniform-in-t Lipschitz bounds incl. (14)
    Load-bearing: used in Lemmas 4.3, 4.5, 4.6, Theorem 4.9; not satisfied by OU-type processes with v(0)=0 because the conditional score has a 1/v²(t) singularity.
  • ad hoc to paper (A5) parameter norm bound B_Θ
    Bounded parameter space ensures compactness and uniform Lipschitz properties of sΘ; no data-driven justification.
  • standard math (A6) time embedding and activation Lipschitz with ψ(0)=0
    Structural assumptions on the ResNet; mild.
  • domain assumption Condition (29): integrable dominating function for p(0,x)p(t,x_t|x) (or its gradient)
    Used in Step 2 of Theorem 4.9 to justify the DSM/ESM integration-by-parts identity; as stated with L² norm it is insufficient and for OU near t=0 no uniform h exists.
  • standard math Girsanov theorem, McDiarmid inequality, Dudley entropy bound, Pinsker inequality, Borel-Cantelli lemma
    Invoked in proofs; well-established background results.

pith-pipeline@v1.3.0-alltime-deepseek · 31415 in / 28353 out tokens · 248206 ms · 2026-08-01T00:03:19.932718+00:00 · methodology

0 comments
read the original abstract

Despite the empirical success of score-based diffusion models, a complete theoretical understanding of how finite-sample learning, network parameterization, and numerical discretization jointly dictate generative quality remains underdeveloped. Existing sampling analyses often evaluate the generative performance conditional on an oracle score or a pre-specified error threshold. In this work, we establish a unified convergence and generalization framework for score-based diffusion models parameterized by practical ResNet-type architectures. We analyze the generalization and convergence properties from the practical finite-sample, discrete-time learning problem of the score function to the ideal continuous-time, population-level objective. Based on the generalization result of the learning problem of score function, we analyze the sampling process induced by the learned score function and provide an end-to-end total variation distance estimate for the generated terminal distribution. This estimate explicitly decomposes the overall generative error into four interpretable components: the truncation error of the forward process, the reverse-time discretization error, the generalization error incorporating both finite data and forward-time discretization, and the training optimization gap. Our results quantitatively characterize how the training sample size, temporal discretization grids, and optimization accuracy jointly control the final fidelity of samples generated by diffusion models.

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

41 extracted references · 6 linked inside Pith

  1. [1]

    L. Yang, Z. Zhang, Y. Song, S. Hong, R. Xu, Y. Zhao, W. Zhang, B. Cui, M.-H. Yang, Diffusion models: A comprehensive survey of methods and applications, ACM Comput. Surv. 56 (4).doi:10.1145/3626235. URLhttps://doi.org/10.1145/3626235

    doi:10.1145/3626235

  2. [2]

    M. Chen, S. Mei, J. Fan, M. Wang, An overview of diffusion models: Applications, guided generation, statistical rates and optimization, arXiv preprint arXiv:2404.07771

    Pith/arXiv arXiv

  3. [3]

    Huang, J

    Y. Huang, J. Huang, Y. Liu, M. Yan, J. Lv, J. Liu, W. Xiong, H. Zhang, L. Cao, S. Chen, Diffusion model-based image editing: A survey, IEEE Transactions on Pattern Analysis and Machine Intelligence 47 (6) (2025) 4409–4437.doi:10.1109/TPAMI.2025.3541625

    arXiv 2025

  4. [4]

    Ramesh, P

    A. Ramesh, P. Dhariwal, A. Nichol, C. Chu, M. Chen, Hierarchical text-conditional image generation with clip latents, arXiv preprint arXiv:2204.06125 1 (2) (2022) 3

    Pith/arXiv arXiv 2022

  5. [5]

    Saharia, W

    C. Saharia, W. Chan, S. Saxena, L. Li, J. Whang, E. L. Denton, K. Ghasemipour, R. Gontijo Lopes, B. Karagol Ayan, T. Salimans, et al., Photorealistic text-to-image diffusion models with deep language understanding, Advances in neural information processing systems 35 (2022) 36479–36494

    2022

  6. [6]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-resolution image synthesis with la- tent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695

    2022

  7. [7]

    OpenAI, Sora: Creating video from text,https://openai.com/sora, accessed: 2024-11-01 (2024)

    2024

  8. [8]

    Y. Song, S. Ermon, Generative modeling by estimating gradients of the data distribution, Advances in Neural Information Processing Systems 32

  9. [9]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Advances in Neural Information Processing Systems 33 (2020) 6840–6851

    2020

  10. [10]

    J. Song, C. Meng, S. Ermon, Denoising diffusion implicit models, in: International Conference on Learning Representations, 2021

    2021

  11. [11]

    Y. Song, C. Durkan, I. Murray, S. Ermon, Maximum likelihood training of score-based diffusion models, Advances in neural information processing systems 34 (2021) 1415–1428

    2021

  12. [12]

    Y. Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, B. Poole, Score-based generative modeling through stochastic differential equations, in: International Conference on Learning Representations, 2021. 27

    2021

  13. [13]

    W. Tang, H. Zhao, Score-based diffusion models via stochastic differential equations, Statistics Surveys 19 (2025) 28 – 64.doi:10.1214/25-SS152. URLhttps://doi.org/10.1214/25-SS152

    doi:10.1214/25-ss152 2025

  14. [14]

    H. Lee, J. Lu, Y. Tan, Convergence for score-based generative modeling with polynomial complexity, Advances in Neural Information Processing Systems 35 (2022) 22870–22882

    2022

  15. [15]

    De Bortoli, Convergence of denoising diffusion models under the manifold hypothesis, arXiv preprint arXiv:2208.05314

    V. De Bortoli, Convergence of denoising diffusion models under the manifold hypothesis, arXiv preprint arXiv:2208.05314

    Pith/arXiv arXiv

  16. [16]

    S. Chen, S. Chewi, J. Li, Y. Li, A. Salim, A. R. Zhang, Sampling is as easy as learning the score: Theory for diffusion models with minimal data assumptions, in: International Conference on Learning Representations (ICLR), 2023.arXiv:2209.11215. URLhttps://openreview.net/forum?id=zyLVMgsZ0U_

    Pith/arXiv arXiv 2023

  17. [17]

    H. Chen, H. Lee, J. Lu, Improved analysis of score-based generative modeling: User-friendly bounds under minimal smoothness assumptions, in: International Conference on Machine Learning, PMLR, 2023, pp. 4735–4763

    2023

  18. [18]

    H. Lee, J. Lu, Y. Tan, Convergence of score-based generative modeling for general data distributions, in: International Conference on Algorithmic Learning Theory, PMLR, 2023, pp. 946–985

    2023

  19. [19]

    W. Tang, H. Zhao, Contractive diffusion probabilistic models, arXiv preprint arXiv:2401.13115

    Pith/arXiv arXiv

  20. [20]

    Block, Y

    A. Block, Y. Mroueh, A. Rakhlin, Generative modeling with denoising auto-encoders and langevin sam- pling, arXiv preprint arXiv:2002.00107

    Pith/arXiv arXiv 2002

  21. [21]

    M. Chen, K. Huang, T. Zhao, M. Wang, Score approximation, estimation and distribution recovery of diffusion models on low-dimensional data, in: International Conference on Machine Learning, PMLR, 2023, pp. 4672–4712

    2023

  22. [22]

    P. Li, Z. Li, H. Zhang, J. Bian, On the generalization properties of diffusion models, Advances in Neural Information Processing Systems 36 (2023) 2097–2127

    2023

  23. [23]

    Pidstrigach, Score-based generative models detect manifolds, Advances in Neural Information Process- ing Systems 35 (2022) 35852–35865

    J. Pidstrigach, Score-based generative models detect manifolds, Advances in Neural Information Process- ing Systems 35 (2022) 35852–35865

    2022

  24. [24]

    A. W. Van der Vaart, Asymptotic statistics, Vol. 3, Cambridge University Press, 2000

    2000

  25. [25]

    L. Bo, A. Capponi, H. Liao, Large sample mean-field stochastic optimization, SIAM Journal on Control and Optimization 60 (4) (2022) 2538–2573

    2022

  26. [26]

    Thorpe, Y

    M. Thorpe, Y. van Gennip, Deep limits of residual neural networks, Research in the Mathematical Sciences 10 (1) (2023) 6

    2023

  27. [27]

    Huang, Y

    J. Huang, Y. Gao, C. Wu, On dynamical system modeling of learned primal-dual with a linear operator K: stability and convergence properties, Inverse Problems 40 (7) (2024) 075006

    2024

  28. [28]

    Le Gall, Brownian motion, martingales, and stochastic calculus, Springer, 2016

    J.-F. Le Gall, Brownian motion, martingales, and stochastic calculus, Springer, 2016

    2016

  29. [29]

    Oksendal, Stochastic differential equations: An introduction with applications, Springer Science & Business Media, Berlin, Heidelberg, 2013

    B. Oksendal, Stochastic differential equations: An introduction with applications, Springer Science & Business Media, Berlin, Heidelberg, 2013

    2013

  30. [30]

    B. D. Anderson, Reverse-time diffusion equation models, Stochastic Processes and their Applications 12 (3) (1982) 313–326

    1982

  31. [31]

    Vincent, A connection between score matching and denoising autoencoders, Neural computation 23 (7) (2011) 1661–1674

    P. Vincent, A connection between score matching and denoising autoencoders, Neural computation 23 (7) (2011) 1661–1674

    2011

  32. [32]

    P. C. Paton, Ö. D. Akyıldız, N. Kantas, Investigating generalisation in score-based generative models

  33. [33]

    A. Q. Nichol, P. Dhariwal, Improved denoising diffusion probabilistic models, in: International conference on machine learning, PMLR, 2021, pp. 8162–8171. 28

    2021

  34. [34]

    S.Shalev-Shwartz, S.Ben-David, Understandingmachinelearning: Fromtheorytoalgorithms, Cambridge university press, 2014

    2014

  35. [35]

    Foucart, H

    S. Foucart, H. Rauhut, A mathematical introduction to compressive sensing, Applied and Numerical Harmonic Analysis, Springer, New York, NY, 2013

    2013

  36. [36]

    Schnoor, A

    E. Schnoor, A. Behboodi, H. Rauhut, Generalization error bounds for iterative recovery algorithms un- folded as neural networks, Information and Inference: A Journal of the IMA 12 (3) (2023) 2267–2299

    2023

  37. [37]

    K. L. Chung, A course in probability theory, Elsevier, 2000

    2000

  38. [38]

    L.Ambrosio, N.Gigli, G.Savaré, Gradientflows: inmetricspacesandinthespaceofprobabilitymeasures, Springer, 2005

    2005

  39. [39]

    A. B. Tsybakov, Introduction to nonparametric estimation, Springer Series in Statistics, Springer, New York, NY, 2008

    2008

  40. [40]

    E.P.PeterE.Kloeden, Numericalsolutionofstochasticdifferentialequations, SpringerBerlin, Heidelberg, 1992

    1992

  41. [41]

    Emmrich, Discrete versions of Gronwall’s lemma and their application to the numerical analysis of parabolic problems, Techn

    E. Emmrich, Discrete versions of Gronwall’s lemma and their application to the numerical analysis of parabolic problems, Techn. Univ., 1999. 29

    1999