Pith. sign in

REVIEW 2 major objections 4 minor 47 references

Robust Learnability of Sample-Compressible Distributions under Noisy or Adversarial Perturbations

T0 review · 2 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Sample-compressible distribution families remain PAC-learnable from noisy or adversarially perturbed samples, with explicit sample-complexity inflation terms.

desk verdict A promising perturbation-quantization framework undermined by a false inequality in the additive-noise main theorem; the adversarial results may survive but need independent verification. read the letter →

arxiv 2506.06613 v1 pith:ZZYBSUFJ submitted 2025-06-07 stat.ML cs.LG

Arefe Boushehrian , Amir Najafi This is my paper · ORCID
classification stat.MLcs.LG
keywords samplecompressionPAClearnabilitydistributionlearningadditivenoiseadversarialcorruptiontotalvariationdistanceGaussianmixturemodelsuniform
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that sample compressibility—the structural property behind many clean-sample PAC learning bounds—also guarantees learnability when samples are noisy or adversarially corrupted, provided the compression decoder is stable under small changes to its input samples and, for additive noise, the family's densities keep enough energy in low frequencies. Under those conditions it gives sample complexity bounds of roughly the clean-sample cost plus an explicit perturbation term: polylogarithmic in the noise level or corruption budget, and scaling with the dimension, the compression description length \(\tau\), and \(1/\$epsilon^{2}$\). The results reduce robust distribution learning to two checkable structural properties rather than a case-by-case analysis. As concrete consequences, the paper derives first-time bounds for finite mixtures of high-dimensional uniform distributions under both perturbation models and for Gaussian mixture models under adversarial corruption.

What carries the argument

The mechanism is the perturbation-quantization step layered on top of the sample-compression decoder. Given the \(\tau\) representative samples that the compression scheme would select, the algorithm approximates each perturbation vector coordinate-wise by a point on a finite grid whose spacing \(\eta\) controls the denoising error, and it pays \(d\tau\log|I|\) bits to encode which grid point was used; the Local Lipschitz Decodability assumption then shows that any denoised representative sequence \(L'\) produces a distribution within \((r/2)\eta\sqrt{d\tau}\) of the clean decoder's output. In the additive case this makes \(\mathcal{F}*G\) sample-compressible so the clean PAC theorem applies, and in the adversarial case the same grid is combined with \(s\log(n/s)\) bits to locate corrupted samples while a multi-group clique procedure selects a candidate. The second supporting object is the Low-Frequency Property, a Fourier-tail bound ensuring that a pair of densities that differ in \(\ell_2\) keep a fraction \(1-\xi\) of their energy below frequency \(\$\alpha$\), so that convolution with noise, which multiplies Fourier coefficients by \(G(\omega)\), cannot erase the difference.

What would settle it

Run the two-point hypothesis test behind Claim 2.5: take two Gaussians with means separated by \(10\epsilon\$\sigma$\) and variance \(\$sigma^{2}$\), add Gaussian noise with variance \(\$sigma_0^{2}$\), and compute the \(n\)-sample total variation between the two noisy hypotheses, which is at most \(5\epsilon\$\sigma$\sqrt{n}/\sigma_0\); letting \(\$\sigma$\to 0\) pushes the minimax error toward \(1/2\) for any \(n\), so any claim that sample compressibility alone suffices for robust learnability would fail this calculation.

Watch

Extended reading notes

Core claim

The central discovery is that sample compressibility is not just a clean-sample guarantee: it is a robust-learning guarantee once two stability conditions hold. For a family \(\mathcal{F}\) with a \((\tau,t,m)\)-compression scheme, the paper proves a perturbation-quantization lemma: noisy or adversarial samples can be denoised onto a finite grid of quantization points, and the decoder's local Lipschitz property guarantees that the resulting candidate distributions are still close to the target. With additive independent noise \(G\), the class \(\mathcal{F}*G\) is shown to be sample-compressible with the same \(\tau\), an extra \(d\tau\log(1/\eta)\) bits for a grid of spacing \(\eta\), and the same \(m\); Theorem 2.13 then converts TV closeness of the convolved densities into \(\ell_2\) closeness of the original densities using the Low-Frequency Property and the noise's Fourier lower envelope \(B_G(\$\alpha$)\), yielding sample complexity \(N_{\mathrm{clean}} + \tilde{O}(d\tau/\$epsilon^{2}$)\log(1+\$\sigma$)\) for Gaussian or Laplace noise. In the adversarial model, Proposition 3.1 adds \(s\log(n/s) + ds\log(1+Cr\sqrt{ds}/\epsilon)\) bits to describe the corruption pattern and quantized corrections, and Theorem 3.2 obtains TV learnability with \(n \ge \tilde{O}(m(\epsilon)+(s(t+\tau)+$ds^{2}$)/\$epsilon^{2}$)\log(1+C)\) without any low-frequency assumption, because uncorrupted samples still exist and are found by a clique-recovery argument among \(2s+1\) groups.

Load-bearing premise

The load-bearing premise is Assumption 2.1: the family must have a compression decoder that is locally Lipschitz, meaning that moving the few representative samples by a tiny amount changes the reconstructed distribution by at most a proportional total-variation amount; if no such stable decoder exists, both main theorems lose their engine.

Editorial extensions

If this is right

  • Under the two assumptions, any family admitting clean sample compression remains PAC-learnable from samples corrupted by independent product noise, with sample complexity \(N_{\mathrm{clean}}(\epsilon,\delta)+\tilde{O}(d\tau(\epsilon)/\epsilon^2)\log(1+\sigma)\) for Gaussian or Laplace noise.
  • Under adversarial corruption of at most \(s\) samples with an \(\ell_\infty\) budget \(C\), the same families remain TV-learnable with \(n \ge \tilde{O}(m(\epsilon)+(s(t(\epsilon)+\tau(\epsilon))+ds^2)/\epsilon^2)\log(1+C)\), so the budget \(C\) enters only logarithmically.
  • Both structural conditions are minimax necessary in the paper's sense: unbounded-variance Gaussians fail local Lipschitz decodability and are unlearnable from Gaussian-noise-corrupted samples, and high-frequency sine densities fail the low-frequency property and become impossible to recover after convolution with noise.
  • The new bounds apply to \(k\)-mixtures of uniform distributions over axis-aligned hyperrectangles under both perturbation models, and to \(k\)-component Gaussian mixtures with minimum eigenvalue \(\sigma_0^2\) under adversarial corruption, at rates \(\tilde{O}(skd^2/\epsilon^2)\log(1/\delta)+\tilde{O}(ds^2/\epsilon^2)\log(1+C/(\delta\sigma_0))\) for the Gaussian mixture case.

Reading between the lines

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

  • If the sample-compression conjecture eventually holds (every learnable class admits a tight compression scheme), then the two assumptions in this paper become the only extra gap between plain learnability and robust learnability; a useful next step would be to catalogue which common parametric families satisfy local Lipschitz decodability with small constants.
  • The Low-Frequency Property is essentially a prohibition on densities whose pairwise differences live at high frequencies. A testable extension is to check whether log-concave or bounded-density families satisfy it with \(\xi(\varepsilon)<1\), which would immediately import the additive-noise bounds to broader families.
  • In the adversarial bound, the \(ds^2/\epsilon^2\) term comes from quantizing the \(\ell_\infty\) budget in \(d\) dimensions. A parallel analysis with an \(\ell_2\) budget should replace this by roughly \(s^2/\epsilon^2\) without the explicit factor \(d\), a concrete calculation one could carry out with the same proof technique.
  • The guarantees are information-theoretic and the decoder is allowed to be expensive. A natural follow-up is to ask which of these rates are achievable in polynomial time, particularly for Gaussian mixtures, since the multi-group clique-selection step is the main algorithmic bottleneck.
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 / 4 minor

Summary. The paper develops a perturbation-quantization framework to extend sample-compression-based PAC learnability of distribution families to two corrupted-data settings: additive independent noise and adversarial corruption with an l_infinity budget. For the additive setting it introduces Local Lipschitz Decodability (Assumption 2.1) and a Low-Frequency Property (Assumption 2.9), and claims l2-learnability of any sample-compressible family satisfying these conditions. For the adversarial setting it claims TV-learnability under sample compressibility plus Local Lipschitz Decodability, with no low-frequency assumption. Concrete applications are given for mixtures of uniform distributions over hyperrectangles and for Gaussian mixture models under adversarial corruption, stated as resolving two open problems.

Significance. If the results were correct, the paper would provide a broad and elegant reduction: robust learnability from perturbed samples would follow from sample compressibility plus two structural conditions, with sample complexity degrading gracefully in the noise level or corruption budget. The framework is clearly presented, the assumptions are explicit, and the minimax-necessity examples are a useful contribution. The adversarial section, which avoids the Fourier-based step of the additive section, appears plausible and addresses genuine open problems for UMMs and GMMs. However, the central additive-noise theorem rests on a demonstrably false inequality in Lemma A.4; as written, the headline claim that sample-compressible families remain l2-learnable under additive noise is not established. This is a load-bearing error, not a presentation issue, and it also invalidates the additive applications.

major comments (2)
  1. [Appendix A, Lemma A.4 (proof, equations (34)–(35))] The proof uses the chain 576ε² ≥ (∫|f∗G−bf∗G|)² ≥ ∫|f∗G−bf∗G|², and the second inequality is false for general L¹∩L² functions on R^d; for example h = 2·1_{[0,1/2]} − 1_{[1/2,1]} gives (∫|h|)² = 2.25 < 2.5 = ∫|h|². The correct inequality goes in the opposite direction and requires a bounded-support factor, which is absent from the theorem's assumptions. This step is the only bridge from the TV bound (34) to the L² bound (35), so Lemma A.4 is not established; consequently Theorem 2.13, Corollary 2.14, and Proposition 4.2, which all rely on it, are unsupported. The adversarial results in Section 3 do not use this lemma and may be correct.
  2. [Theorem 2.13, equation (14)] The statement advertises PAC learnability 'within ℓ2 error ϵ', but the bound (14) is ∥bf−f∗∥2 ≤ ϵ · inf_{...} 24/(√{BG(α)(1−ξ)}). The multiplicative factor can be large, for Gaussian noise it contains e^{(σα)²/2}, so to guarantee an actual ℓ2 error of ϵ the accuracy parameter inside Nclean and τ,m in (13) must be replaced by ϵ/C(α,ξ), not by ϵ. As written, the sample complexity displayed in (13) and Corollary 2.14 corresponds to an error that is a constant multiple of ϵ, which is a different guarantee from the one stated in the abstract and Section 1.1.
minor comments (4)
  1. [Theorem 3.2, proof (Appendix B)] The proof says 'By Proposition 2.7' when it should refer to Proposition 3.1; the adversarial compression result is the one being used.
  2. [Section 4.1, Proposition 4.2] The proposition restricts to 'sufficiently large σ', while the abstract and Section 1.1 claim learnability under additive noise generally; the small-σ regime is only discussed in the proof, which makes the proposition's scope unclear.
  3. [Section 1.2, Figure 1 reference] The caption labels the figure as 'Figure 1' but the text refers to it as 'Figure 1.2'.
  4. [Throughout] There are typographical errors such as 'throgh' before Proposition 4.2 and 'corrputed' in Proposition 4.3; a proofreading pass is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the derivation is self-contained, with only minor non-load-bearing self-citations.

full rationale

The paper's main claims are derived from explicit assumptions (sample compressibility, Assumption 2.1 Local Lipschitz Decodability, and Assumption 2.9 Low-Frequency Property) through constructive proofs, rather than by fitting parameters or importing conclusions from self-citations. Proposition 2.7 builds the noisy sample-compression scheme from the original clean compression scheme plus a quantization argument; Theorem 2.13 combines this with Parseval's theorem and Assumption 2.9; Theorem 3.2 uses the same stable-decoder assumption with a clique-recovery selection routine. Each step is stated with its own equations and proofs in the appendices. The cited works [NIS+21] and [SNMK23] include overlapping authors, but they are used for context, motivation, and prior building blocks, not to force the paper's new conclusions; the load-bearing lemmas are proved in-paper or cite the external [ABDH+18] compression-to-learnability machinery. No quantity is fitted to data and then renamed a prediction, and no uniqueness or structural condition is imported from the authors' own prior work as an unexamined postulate. Any concern about the validity of Lemma A.4 is a correctness issue, not circularity, because the argument is explicit and does not reduce to its own output.

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

The paper is a pure information-theoretic analysis with no data fitting. Constants in the bounds are universal or depend on problem parameters (e.g., Lipschitz constant r, frequency cutoff α, ξ) which are existential properties of the class, not fitted values. The listed axioms are the stated assumptions that the central claims rest on.

assumptions (6)
  • domain assumption Assumption 1.5: every density in F is square-integrable on X
    Invoked to ensure Fourier transforms exist and Parseval's identity applies (Section 1.4, used in Lemma A.4).
  • domain assumption Assumption 1.6: F admits (τ,t,m)-sample compression
    Core structural assumption inherited from ABDH+18; the entire framework builds on it.
  • ad hoc to paper Assumption 2.1: Local Lipschitz Decodability
    New in this paper; the decoder must be Lipschitz around good sample sequences. It is not implied by sample compressibility and is load-bearing for both main theorems.
  • ad hoc to paper Assumption 2.9: Low-Frequency Property
    New spectral condition requiring non-negligible low-frequency energy in any density difference; needed to recover f* from f**G in the noise model. Not needed in the adversarial model.
  • domain assumption Noise distribution G is known, symmetric, with independent coordinates and known component CDF
    The robust decoder and bounds in Theorem 2.13 require exact knowledge of G and quantile functions Φ_G; the paper states this in Section 1.4.
  • domain assumption Adversarial model: at most s corruptions, each of ℓ∞ norm at most C, adversary knows f* and the algorithm
    Defines the threat model in Section 1.4; bounds depend on s and C.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Learnability of Sample-Compressible Distributions under Noisy or Adversarial Perturbations." pith.science (2026). https://pith.science/paper/ZZYBSUFJ

@misc{pith2026250606613,
  author       = {Pith},
  title        = {Pith review of: Robust Learnability of Sample-Compressible Distributions under Noisy or Adversarial Perturbations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZYBSUFJ}},
  note         = {Machine review of arXiv:2506.06613}
}
abstract

Learning distribution families over $\mathbb{R}^d$ is a fundamental problem in unsupervised learning and statistics. A central question in this setting is whether a given family of distributions possesses sufficient structure to be (at least) information-theoretically learnable and, if so, to characterize its sample complexity. In 2018, Ashtiani et al. reframed \emph{sample compressibility}, originally due to Littlestone and Warmuth (1986), as a structural property of distribution classes, proving that it guarantees PAC-learnability. This discovery subsequently enabled a series of recent advancements in deriving nearly tight sample complexity bounds for various high-dimensional open problems. It has been further conjectured that the converse also holds: every learnable class admits a tight sample compression scheme. In this work, we establish that sample compressible families remain learnable even from perturbed samples, subject to a set of necessary and sufficient conditions. We analyze two models of data perturbation: (i) an additive independent noise model, and (ii) an adversarial corruption model, where an adversary manipulates a limited subset of the samples unknown to the learner. Our results are general and rely on as minimal assumptions as possible. We develop a perturbation-quantization framework that interfaces naturally with the compression scheme and leads to sample complexity bounds that scale gracefully with the noise level and corruption budget. As concrete applications, we establish new sample complexity bounds for learning finite mixtures of high-dimensional uniform distributions under both noise and adversarial perturbations, as well as for learning Gaussian mixture models from adversarially corrupted samples, resolving two open problems in the literature.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 43 canonical work pages

  1. [1]

    Polynomial time and private learning of unbounded gaussian mixture models

    Jamil Arbas, Hassan Ashtiani, and Christopher Liaw. Polynomial time and private learning of unbounded gaussian mixture models. In International Conference on Machine Learning , pages 1018--1040. PMLR, 2023

    work page 2023

  2. [2]

    Mixtures of gaussians are privately learnable with a polynomial number of samples

    Mohammad Afzali, Hassan Ashtiani, and Christopher Liaw. Mixtures of gaussians are privately learnable with a polynomial number of samples. In International Conference on Algorithmic Learning Theory , pages 47--73. PMLR, 2024

    work page 2024

  3. [3]

    Nearly tight sample complexity bounds for learning mixtures of gaussians via sample compression schemes

    Hassan Ashtiani, Shai Ben-David, Nicholas Harvey, Christopher Liaw, Abbas Mehrabian, and Yaniv Plan. Nearly tight sample complexity bounds for learning mixtures of gaussians via sample compression schemes. Advances in Neural Information Processing Systems , 31, 2018

    work page 2018

  4. [4]

    Near-optimal sample complexity bounds for robust learning of gaussian mixtures via compression schemes

    Hassan Ashtiani, Shai Ben-David, Nicholas Harvey, Christopher Liaw, Abbas Mehrabian, and Yaniv Plan. Near-optimal sample complexity bounds for robust learning of gaussian mixtures via compression schemes. Journal of the ACM (JACM) , 67, 2020

    work page 2020

  5. [5]

    Efficient learning of simplices

    Joseph Anderson, Navin Goyal, and Luis Rademacher. Efficient learning of simplices. In Conference on Learning Theory , pages 1020--1045. PMLR, 2013

    work page 2013

  6. [6]

    Private and polynomial time algorithms for learning gaussians and beyond

    Hassan Ashtiani and Christopher Liaw. Private and polynomial time algorithms for learning gaussians and beyond. In Conference on Learning Theory , pages 1075--1076. PMLR, 2022

    work page 2022

  7. [7]

    Private distribution learning with public data: The view from sample compression

    Shai Ben-David, Alex Bie, Cl \'e ment L Canonne, Gautam Kamath, and Vikrant Singhal. Private distribution learning with public data: The view from sample compression. Advances in Neural Information Processing Systems , 36:7184--7215, 2023

    work page 2023

  8. [8]

    Minimax rates for conditional density estimation via empirical entropy

    Blair Bilodeau, Dylan J Foster, and Daniel M Roy. Minimax rates for conditional density estimation via empirical entropy. The Annals of Statistics , 51(2):762--790, 2023

    work page 2023

Show all 47 references
  1. [9]

    Learning smooth shapes by probing

    Jean-Daniel Boissonnat, Leonidas J Guibas, and Steve Oudot. Learning smooth shapes by probing. Computational Geometry , 37(1):38--58, 2007

    2007

  2. [10]

    Density estimation on an unknown submanifold

    Cl \'e ment Berenfeld and Marc Hoffmann. Density estimation on an unknown submanifold. Electronic Journal of Statistics , 15:2179--2223, 2021

    2021

  3. [11]

    Sharp rate of average decay of the fourier transform of a bounded set

    Luca Brandolini, Steve Hofmann, and Alex Iosevich. Sharp rate of average decay of the fourier transform of a bounded set. Geometric & Functional Analysis GAFA , 13:671--680, 2003

    2003

  4. [12]

    Not all learnable distribution classes are privately learnable

    Mark Bun, Gautam Kamath, Argyris Mouzakis, and Vikrant Singhal. Not all learnable distribution classes are privately learnable. In International Conference on Algorithmic Learning Theory , pages 390--401. PMLR, 2024

    2024

  5. [13]

    Model-based learning using a mixture of mixtures of gaussian and uniform distributions

    Ryan P Browne, Paul D McNicholas, and Matthew D Sparling. Model-based learning using a mixture of mixtures of gaussian and uniform distributions. IEEE Transactions on Pattern Analysis and Machine Intelligence , 34(4):814--817, 2011

    2011

  6. [14]

    Efficiently learning ising models on arbitrary graphs

    Guy Bresler. Efficiently learning ising models on arbitrary graphs. In Proceedings of the forty-seventh annual ACM symposium on Theory of computing , pages 771--782, 2015

    2015

  7. [15]

    A gaussian uniform mixture model for robust kalman filtering

    Mathieu Brunot. A gaussian uniform mixture model for robust kalman filtering. IEEE Transactions on Aerospace and Electronic Systems , 56(4):2656--2665, 2019

    2019

  8. [16]

    Convex Optimization

    Stephen Boyd and Lieven Vandenberghe. Convex Optimization . Cambridge University Press, Cambridge, UK, 2004

    2004

  9. [17]

    A functional approach to data structures and its use in multidimensional searching

    Bernard Chazelle. A functional approach to data structures and its use in multidimensional searching. SIAM Journal on Computing , 17(3):427--462, 1988

    1988

  10. [18]

    Learning from untrusted data

    Moses Charikar, Jacob Steinhardt, and Gregory Valiant. Learning from untrusted data. In Proceedings of the 49th annual ACM SIGACT symposium on theory of computing , pages 47--60, 2017

    2017

  11. [19]

    Computational Geometry: Algorithms and Applications

    Mark de Berg, Otfried Cheong, Marc van Kreveld, and Mark Overmars. Computational Geometry: Algorithms and Applications . Springer, 3rd edition, 2008

    2008

  12. [20]

    A probabilistic theory of pattern recognition , volume 31

    Luc Devroye, L \'a szl \'o Gy \"o rfi, and G \'a bor Lugosi. A probabilistic theory of pattern recognition , volume 31. Springer Science & Business Media, 2013

    2013

  13. [21]

    Combinatorial Methods in Density Estimation

    Luc Devroye and G \'a bor Lugosi. Combinatorial Methods in Density Estimation . Springer-Verlag, New York, 2001

    2001

  14. [22]

    The total variation distance between high-dimensional gaussians

    Luc Devroye, Abbas Mehrabian, and Tommy Reddad. The total variation distance between high-dimensional gaussians. arXiv: Statistics Theory , 2018

    2018

  15. [23]

    The minimax learning rates of normal and ising undirected graphical models

    Luc Devroye, Abbas Mehrabian, and Tommy Reddad. The minimax learning rates of normal and ising undirected graphical models. Electronic Journal of Statistics , 14:2338--2361, 2020

    2020

  16. [24]

    Differential privacy

    Cynthia Dwork. Differential privacy. In International colloquium on automata, languages, and programming , pages 1--12. Springer, 2006

    2006

  17. [25]

    On the sample complexity of adversarial multi-source pac learning

    Nikola Konstantinov, Elias Frantar, Dan Alistarh, and Christoph Lampert. On the sample complexity of adversarial multi-source pac learning. In International Conference on Machine Learning , pages 5416--5425. PMLR, 2020

    2020

  18. [26]

    Learning in the presence of malicious errors

    Michael Kearns and Ming Li. Learning in the presence of malicious errors. In Proceedings of the twentieth annual ACM symposium on Theory of computing , pages 267--280, 1988

    1988

  19. [27]

    Relating data compression and learnability

    Nick Littlestone and Manfred Warmuth. Relating data compression and learnability. Unpublished article , 1986

    1986

  20. [28]

    The curse of concentration in robust learning: Evasion and poisoning attacks from concentration of measure

    Saeed Mahloujifar, Dimitrios I Diochnos, and Mohammad Mahmoody. The curse of concentration in robust learning: Evasion and poisoning attacks from concentration of measure. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 33, pages 4536--4543, 2019

    2019

  21. [29]

    A brief history of generative models for power law and lognormal distributions

    Michael Mitzenmacher. A brief history of generative models for power law and lognormal distributions. Internet mathematics , 1(2):226--251, 2004

    2004

  22. [30]

    Differential privacy with higher utility by exploiting coordinate-wise disparity: Laplace mechanism can beat gaussian in high dimensions

    Gokularam Muthukrishnan and Sheetal Kalyani. Differential privacy with higher utility by exploiting coordinate-wise disparity: Laplace mechanism can beat gaussian in high dimensions. IEEE Transactions on Information Forensics and Security , 2025

    2025

  23. [31]

    On statistical learning of simplices: Unmixing problem revisited

    Amir Najafi, Saeed Ilchi, Amir Hossein Saberi, Seyed Abolfazl Motahari, Babak H Khalaj, and Hamid R Rabiee. On statistical learning of simplices: Unmixing problem revisited. The Annals of Statistics , 49(3):1626--1655, 2021

    2021

  24. [32]

    Amir Najafi, Abolfazl Motahari, and Hamid R. Rabiee. Reliable clustering of bernoulli mixture models. Bernoulli , 26(2):1535--1559, 2020

    2020

  25. [33]

    Machine learning for anomaly detection: A systematic review

    Ali Bou Nassif, Manar Abu Talib, Qassim Nasir, and Fatima Mohamad Dakalbab. Machine learning for anomaly detection: A systematic review. Ieee Access , 9:78658--78700, 2021

    2021

  26. [34]

    Minimax estimation of smooth densities in wasserstein distance

    Jonathan Niles-Weed and Quentin Berthet. Minimax estimation of smooth densities in wasserstein distance. The Annals of Statistics , 50(3):1519--1540, 2022

    2022

  27. [35]

    New upper bounds in klee’s measure problem

    Mark H Overmars and Chee-Keng Yap. New upper bounds in klee’s measure problem. SIAM Journal on Computing , 20(6):1034--1045, 1991

    1991

  28. [36]

    A fourier approach to mixture learning

    Mingda Qiao, Guru Guruganesh, Ankit Rawat, Kumar Avinava Dubey, and Manzil Zaheer. A fourier approach to mixture learning. Advances in Neural Information Processing Systems , 35:20850--20861, 2022

    2022

  29. [37]

    Real and Complex Analysis

    Walter Rudin. Real and Complex Analysis . McGraw-Hill, New York, 3 edition, 1987

    1987

  30. [38]

    Certifying some distributional robustness with principled adversarial training

    Aman Sinha, Hongseok Namkoong, and John Duchi. Certifying some distributional robustness with principled adversarial training. In International Conference on Learning Representations , 2018

    2018

  31. [39]

    Sample complexity bounds for learning high-dimensional simplices in noisy regimes

    Seyed Amir Hossein Saberi, Amir Najafi, Abolfazl Motahari, and Babak Khalaj. Sample complexity bounds for learning high-dimensional simplices in noisy regimes. In International Conference on Machine Learning , pages 29514--29541. PMLR, 2023

    2023

  32. [40]

    Special functions: An introduction to the classical functions of mathematical physics

    Nico M Temme. Special functions: An introduction to the classical functions of mathematical physics . John Wiley & Sons, 1996

    1996

  33. [41]

    Tsybakov

    Alexandre B. Tsybakov. Introduction to Nonparametric Estimation . Springer Series in Statistics. Springer, New York, 2009. Revised and extended from the 2004 French original, Translated by Vladimir Zaiats

    2009

  34. [42]

    Minimax rate of distribution estimation on unknown submanifolds under adversarial losses

    Rong Tang and Yun Yang. Minimax rate of distribution estimation on unknown submanifolds under adversarial losses. The Annals of Statistics , 51(3):1282--1308, 2023

    2023

  35. [43]

    A theory of the learnable

    Leslie G Valiant. A theory of the learnable. Communications of the ACM , 27(11):1134--1142, 1984

    1984

  36. [44]

    On minimax density estimation via measure transport

    Sven Wang and Youssef Marzouk. On minimax density estimation via measure transport. arXiv preprint arXiv:2207.10231 , 2022

    2022 arXiv

  37. [45]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  38. [46]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  39. [47]

    water-filling

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    1992 arXiv

Pith tools

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