Pith. sign in

REVIEW 2 major objections 5 minor 35 references

Sampling and Identity-Testing Without Approximate Tensorization of Entropy

T0 review · 2 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper shows that mixtures of ATE distributions, which do not themselves tensorize entropy, still admit fast Glauber-dynamics sampling from a data-based initialization and an efficient coordinate-conditional identity tester, resolving…

desk verdict Solid sampling results and a plausible identity-testing answer, but Theorem 5 as printed has a real gap in Lemma 30 that needs a fix. read the letter →

arxiv 2506.23456 v1 pith:PCD2KJ4E submitted 2025-06-30 math.ST cs.DScs.LGstat.MLstat.TH

William Gay , William He , Nicholas Kocurek , Ryan O'Donnell This is my paper · ORCID
classification math.STcs.DScs.LGstat.MLstat.TH MSC 62B1062H1560J2268W20
keywords approximatetensorizationofentropymixturedistributionsGlauberdynamicsidentitytestingcoordinateconditionalsamplingmodifiedlog-SobolevinequalityKLdivergencePhi-entropy
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 studies distributions formed by mixing a small number of components, each of which satisfies approximate tensorization of entropy (ATE), a local-to-global property that makes sampling and identity testing easy. A mixture of ATE distributions typically does not satisfy ATE, so the standard tools no longer apply. The main claim is that two tasks survive this failure: Glauber dynamics started from an empirical distribution based on $O(k/\varepsilon+\log(1/\delta)/\varepsilon)$ samples mixes to KL error at most $\varepsilon$ in time roughly $c_* n(\log\log(1/\min_x\mu(x))+\log(1/\varepsilon))$, and there is an identity tester in the coordinate-conditional model with sample complexity roughly $(c_* n/\varepsilon)\log^2(c_* n/\varepsilon)\sqrt{|\Sigma|}\log(1/\eta)\log\log(1/\eta)$ plus $O(\sqrt{k}\log(1/\rho_*)/\varepsilon)$. The proof hinges on the chain rule for entropy, which splits the failure of ATE into entropy inside components and entropy across components; the two parts are handled by local coordinate tests and by estimating the posterior distribution over components. This answers an open question from earlier work and extends prior data-initialization sampling results from Poincaré inequalities to modified log-Sobolev inequalities.

What carries the argument

The central object is the chain rule for $\Phi$-entropy (Lemma 11), specialized to $\Phi(u)=u\log u$, which reads $\mathrm{Ent}_\mu[f]=E_{a\sim\rho}\mathrm{Ent}_{\mu_a}[f]+\mathrm{Ent}_{a\sim\rho}[E_{\mu_a}[f]]$. This identity carries the argument by separating every obstacle into an intra-component part, which the local coordinate functional can see, and an inter-component part, which reduces to the KL divergence of the posterior component weights. Around this identity the paper places three supporting mechanisms: a lemma that lower-bounds local entropy by the intra-component entropy when each component satisfies $c_*$-ATE; a weak $\Phi$-Sobolev inequality for mixtures whose error term is exactly the inter-component entropy; and the multinomial KL concentration bound of the cited literature, which controls the inter-component term for empirical distributions. In the identity tester, the equivalent mechanism is the algorithm's two-step structure: repeated KL tests on random coordinate conditionals with a random threshold, followed by a KL test on the posterior $\rho_\pi$.

What would settle it

Recompute the averaging in the proof of Lemma 30 for the extremal case where the local entropy $Y$ equals $\log(1/\eta)$ with probability $0.5\varepsilon/(c_* n)$ and $0$ otherwise, with threshold $\theta$ uniform on $[0.05\varepsilon/c_*,\log(1/\eta)]$. The exact probability that a single draw has $Y\ge\theta$ is $0.5\varepsilon/(c_* n)$, and comparing the proof's lower bound of $0.45\varepsilon/(c_* n)$ against $\mathbb{E}[Y]/\log(1/\eta)$ decides whether $T_1$ needs an extra factor of $\log(1/\eta)$; that fixes whether the oracle complexity in Theorem 5 is correct.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that the missing ATE inequality can be replaced by the chain rule for entropy. For a mixture $\mu=\sum_a\rho(a)\mu_a$ whose components satisfy $c_*$-ATE, the entropy $\mathrm{Ent}_\mu[f]$ of any density $f=\pi/\mu$ is the sum of $E_{a\sim\rho}\mathrm{Ent}_{\mu_a}[f]$ and $\mathrm{Ent}_{a\sim\rho}[E_{\mu_a}[f]]$. The first, intra-component term is controlled by the coordinate-wise local entropy via Lemma 12, and the second, inter-component term is identified with $\mathrm{D}_{\mathrm{KL}}(\rho_\pi\|\rho)$, the divergence between the posterior distribution over components under $\pi$ and the true mixing weights. Theorem 4 turns this split into a sampling guarantee for Glauber dynamics from a data-based initialization, using a weak modified log-Sobolev inequality for mixtures together with a moment-generating-function bound for the empirical posterior. Theorem 5 turns the same split into an algorithm that runs random coordinate-conditional KL tests and then a KL test on the component posterior, achieving the stated oracle complexity and constant error probabilities of rejecting with probability at least 0.6 when $\mathrm{D}_{\mathrm{KL}}(\pi\|\mu)\ge\varepsilon$ and at most 0.4 when $\pi=\mu$. The two regimes cover all alternatives: large intra-component entropy triggers the first step, and large inter-component entropy triggers the second.

Load-bearing premise

The load-bearing premise is that a bounded number of random single-coordinate comparisons can catch any large divergence that sits inside one mixture component; the estimate of how often one such comparison succeeds is the step that sets the sample count, because the detection threshold is drawn from a range wide enough to dilute the probability.

Editorial extensions

If this is right

  • Theorem 4 makes data-based initialization work at the modified log-Sobolev level: for mixtures of MLSI components, $m=O(k/\varepsilon+\log(1/\delta)/\varepsilon)$ samples put Glauber dynamics within KL error $\varepsilon$ in time $c_* n\,O(\log\log(1/\min_x\mu(x))+\log(1/\varepsilon))$.
  • Theorem 5 gives the first efficient identity tester for mixtures of ATE distributions in the coordinate-conditional sampling model, with oracle complexity $O((c_* n/\varepsilon)\log^2(c_* n/\varepsilon)\sqrt{|\Sigma|}\log(1/\eta)\log\log(1/\eta)) + O(\sqrt{k}\log(1/\rho_*)/\varepsilon)$.
  • The tester achieves rejection probability at least 0.6 at KL distance $\varepsilon$ and at most 0.4 under the null hypothesis, so standard repetition amplifies it to any desired confidence.
  • The tester does not need adaptive coordinate queries: the set of coordinate-conditional queries can be fixed in advance as random pairs $(x,i)$ with $x\sim\pi$ and $i\in[n]$ uniform.
  • When $k=1$ the results recover the known ATE and MLSI sampling and testing guarantees, confirming that the mixture statements are strict extensions rather than new phenomena.

Reading between the lines

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

  • A general recipe suggested by these proofs is that a mixture's failure of a local-to-global inequality can be diagnosed by separating within-component and across-component contributions; the same decomposition may give testers for other divergences if a substitute for KL's 1-homogeneity is found, since the paper notes that this homogeneity is the only obstacle in the relevant lemma.
  • The sampling theorem has a practical reading not spelled out in the paper: one can run Glauber dynamics on a multimodal target using only a modest pilot sample as the warm start, without knowing the components in advance, and still expect fast mixing.
  • An immediate computational check of the identity-testing claim would run the proposed algorithm on simple mixtures, such as two biased product distributions, and compare rejection rates against $\varepsilon$-separated alternatives to see how the threshold range $\log(1/\eta)$ enters the required number of local checks.
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 / 5 minor

Summary. The paper studies sampling and identity-testing for distributions that are mixtures of components satisfying approximate tensorization of entropy (ATE) or a modified log-Sobolev inequality (MLSI). The main sampling result (Theorem 4/23) shows that Glauber dynamics for such a mixture, when warm-started at the empirical distribution of m = O(k/epsilon + log(1/delta)/epsilon) samples, mixes to KL-divergence eps in time c* n · O(log log(1/min_x mu(x)) + log(1/epsilon)). The main testing result (Theorem 5/26) gives an identity tester in the coordinate-conditional sampling model with sample complexity O((c* n / epsilon) log^2(c* n / epsilon) sqrt(|Sigma|) log(1/eta) log log(1/eta)) + O(sqrt(k) log(1/rho*)/epsilon), answering an open question of Blanca et al. The sampling proof proceeds via a Phi-entropic weak Sobolev inequality for mixtures plus a convexity argument reducing the empirical posterior concentration to Agrawal's multinomial concentration bound; the testing proof splits the KL divergence into intra-component and inter-component parts and handles each with a local tester and a posterior-weight tester, respectively.

Significance. The results are significant if the testing theorem can be proved as stated. The sampling theorem cleanly extends prior Poincare-based mixture sampling results to MLSI components and achieves optimal sample complexity, and the use of Agrawal's m.g.f. bound rather than standard Chernoff bounds is a genuine technical improvement. The identity-testing result answers an open question and introduces a useful simplification (Theorem 24) of the KL tester by combining existing Hellinger/KL inequalities with the DKW tester. The proofs are transparent and rely on standard external tools; there are no fitted parameters, no circular claims, and no normalization forced by the target results. However, the identity-testing theorem as printed has a concrete gap in the analysis of the local-testing step, so the main testing claim is not yet established. The gap appears local and fixable, and the stated final complexity is consistent with a corrected threshold count, so the paper merits a major revision rather than rejection.

major comments (2)
  1. [4.1.1, Lemma 30] The calculation of Pr_{theta,Y}[Y >= theta] drops a factor of 1/log(1/eta). The displayed identity in Lemma 30 reads Pr_{theta,Y}[Y >= theta] >= (1/log(1/eta)) * integral_0^{log(1/eta)} Pr[Y >= theta] dtheta - (1/log(1/eta)) * integral_0^{0.05 eps/(c* n)} Pr[Y >= theta] dtheta = E[Y] - (1/log(1/eta))*..., but the first term is E[Y]/log(1/eta), not E[Y], because the prefactor applies to the whole integral. Consequently the lower bound is approximately 0.45 eps / (c* n log(1/eta)), not 0.45 eps / (c* n). With the printed T1 = O(c* n / eps), the probability that none of the T1 threshold tests fires is exp(-Omega(1/log(1/eta))), which for eta = 2^{-n} is 1 - O(1/n), not at most 0.04. The rejection guarantee in Theorem 26 therefore does not follow as written. Fixing the issue requires either setting T1 = O((c* n / eps) log(1/eta)) or choosing theta from a different distribution; either fix preserves the stated final complexity in Equation (2). Additionally, Algorithm 1 defines theta as uniform on [0.05 eps/c*, log(1/eta)], but Lemma 30 uses the interval [0.05 eps/(c* n), log(1/eta)] in the denominator and the lower limit; these must be reconciled.
  2. [4.1.1, Lemma 30, event bookkeeping] The union-bound argument in Lemma 30 contains an inconsistency in the definition of event A. The text first shows that the probability that none of the threshold checks fires is at most 0.04, then says 'Let A be the event that this occurs so that Pr[A] <= 0.01', which is numerically wrong and does not identify whether A is the threshold-failure event or the sample-limit event from Lemma 29. The sample-limit event is invoked earlier in the proof but is not explicitly included in the final union bound. This needs to be rewritten so that the three failure events (threshold misses, sample limit exceeded, and local tester false acceptance) are properly defined and union-bounded.
minor comments (5)
  1. [Remark 9] There is a typo in 'defiend'; it should be 'defined'.
  2. [Lemma 21] The proof states 'For all lambda > 0', but the application of Theorem 20 requires 0 <= lambda < m; the statement and proof should say 'for 0 <= lambda < m'.
  3. [Algorithm 1 and Theorem 26] The algorithm is named Product-Set-KL-Test(pi, mu, epsilon), but Theorem 26 and its proof refer to Coordinate-Oracle-Test(mu, pi, epsilon). The notation should be unified.
  4. [Section 4.1.1, Lemma 29] In the displayed computation of E[1/theta], the denominator and the lower limit of integration are inconsistent: the text writes '1/(log(1/eta) - 0.05 eps/c*)' but the lower limit of the integral is '0.05 eps/(c* n)'. Since theta is defined in Algorithm 1 with lower endpoint 0.05 eps/c*, this needs correction.
  5. [Section 4.1.1, Lemma 30] The sentence 'Let A be the event that this occurs so that Pr[A] <= 0.01' appears to conflate two bounds; the preceding display gives probability at most 0.04, so the intended numerical value and the event's definition should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's steps invoke external concentration, tester, and divergence inequalities; the Lemma 30 arithmetic slip is a correctness issue, not a circular derivation.

full rationale

We walked the claimed derivation chains for Theorems 4 and 5. In the sampling proof, the central concentration input is Agrawal's finite-sample m.g.f. bound (Theorems 19 and 20), which is external, parameter-free with respect to the target claims, and used as a black box. The chain rule (Lemma 11), the mixture ATE/MLSI decomposition (Lemma 12), and the weak Sobolev inequality (Lemma 18) are proved in the appendix from standard functional inequalities; they do not assume the conclusion. In the identity-testing proof, the rejection rules are analyzed via the posterior-weight identity (Fact 13), the chain rule, and a KL-tester (Theorem 24) built from DKW18's Hellinger tester plus the Hellinger-KL inequality. The cited self-references ([FO24], [BOW19]) are not load-bearing: they are published, externally proved inequalities or algorithms, and alternate sources are cited (e.g., [BM98]) for the key inequality. No fitted parameter is renamed as a prediction, no normalized quantity is forced by the target theorem, and no uniqueness claim is imported from the authors' prior work. The only notable defect we found is an arithmetic slip in Lemma 30: the displayed identity relating Pr[Y >= theta] to E[Y] drops a 1/log(1/eta) factor, so the printed rejection probability bound for Step 1 does not follow as written. This is a correctness or proof bug, not circularity: it does not make the conclusion an input to the derivation. Thus the paper shows no significant circularity and receives score 0.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

No numbers are fitted to data and no target result is assumed as an input. The central results stand on standard entropy inequalities, Agrawal's concentration bound, and the DKW and BM98 testing machinery. The only ad hoc modeling choices are the eta-balanced condition and the coordinate-conditional oracle, both explicit problem assumptions.

assumptions (7)
  • standard math c*-ATE implies a (1/(c* n))-MLSI for the associated Glauber dynamics (Remark 17, citing CMT15).
    Bridges the ATE assumption to the Phi-Sobolev framework used in Theorem 4 and Theorem 23.
  • standard math Entropy chain rule for mixtures (Lemma 11) holds for Phi-entropy.
    Decomposes global entropy into intra- and inter-component parts; proved in Appendix A.1.
  • standard math Weak Phi-Sobolev inequality for mixtures follows because the Dirichlet form is concave in the mixture weights (Lemma 18).
    Uses concavity of (a,b) -> ab/(a+b) and monotonicity of Phi'; load-bearing for the sampling result.
  • standard math Agrawal's m.g.f. bound for the empirical distribution's KL divergence to rho (Theorem 20) applies to averaged posterior distributions via Jensen (Lemma 21).
    Controls the data-based initialization's inter-component entropy; requires rho to have full support on [k].
  • standard math The Hellinger-KL inequality D_H^2 >= D_KL / log(e^2/eta) lets the H2-Test of DKW18 become a KL tester (Theorem 24).
    External inequality from BM98 and FO24 used as a black box in the identity tester.
  • domain assumption The Coordinate Oracle gives samples from mu|x\i whose minimum probability is at least eta when mu is eta-balanced (Definition 27).
    All the threshold tests and the boundedness Y <= log(1/eta) rely on this balancedness condition.
  • domain assumption Posterior samples a ~ rho|x can be simulated from the General Oracle for pi (Algorithm 1, Step 2).
    The algorithm assumes one can draw x ~ pi and then a with probability rho(a) mu_a(x) / mu(x); this is an explicit modeling assumption about the access model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sampling and Identity-Testing Without Approximate Tensorization of Entropy." pith.science (2026). https://pith.science/paper/PCD2KJ4E

@misc{pith2026250623456,
  author       = {Pith},
  title        = {Pith review of: Sampling and Identity-Testing Without Approximate Tensorization of Entropy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PCD2KJ4E}},
  note         = {Machine review of arXiv:2506.23456}
}
read the original abstract

Certain tasks in high-dimensional statistics become easier when the underlying distribution satisfies a local-to-global property called approximate tensorization of entropy (ATE). For example, the Glauber dynamics Markov chain of an ATE distribution mixes fast and can produce approximate samples in a small amount of time, since such a distribution satisfies a modified log-Sobolev inequality. Moreover, identity-testing for an ATE distribution requires few samples if the tester is given coordinate conditional access to the unknown distribution, as shown by Blanca, Chen, \v{S}tefankovi\v{c}, and Vigoda (COLT 2023). A natural class of distributions that do not satisfy ATE consists of mixtures of (few) distributions that do satisfy ATE. We study the complexity of identity-testing and sampling for these distributions. Our main results are the following: 1. We show fast mixing of Glauber dynamics from a data-based initialization, with optimal sample complexity, for mixtures of distributions satisfying modified log-Sobolev inequalities. This extends work of Huang, Koehler, Lee, Mohanty, Rajaraman, Vuong, and Wu (STOC 2025, COLT 2025) for mixtures of distributions satisfying Poincar\'e inequalities. 2. Answering an open question posed by Blanca et al., we give efficient identity-testers for mixtures of ATE distributions in the coordinate-conditional sampling access model. We also give some simplifications and improvements to the original algorithm of Blanca et al.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

35 extracted references · 31 canonical work pages

  1. [1]

    Optimal testing for properties of distributions

    Jayadev Acharya, Constantinos Daskalakis, and Gautam Kamath. Optimal testing for properties of distributions. Advances in Neural Information Processing Systems , 28, 2015

    work page 2015

  2. [2]

    Finite-sample concentration of the multinomial in relative entropy

    Rohit Agrawal. Finite-sample concentration of the multinomial in relative entropy. IEEE Transactions on Information Theory , 66(10):6297--6302, 2020

    work page 2020

  3. [3]

    Entropic independence I : Modified log- Sobolev inequalities for fractionally log-concave distributions and high-temperature Ising models

    Nima Anari, Vishesh Jain, Frederic Koehler, Huy Tuan Pham, and Thuy-Duong Vuong. Entropic independence I : Modified log- Sobolev inequalities for fractionally log-concave distributions and high-temperature Ising models. arXiv preprint arXiv:2106.04105 , 2021

    arXiv 2021

  4. [4]

    Property testing of joint distributions using conditional samples

    Rishiraj Bhattacharyya and Sourav Chakraborty. Property testing of joint distributions using conditional samples. ACM Transactions on Computation Theory (TOCT) , 10(4):1--20, 2018

    work page 2018

  5. [5]

    On mixing of Markov chains: Coupling, spectral independence, and entropy factorization

    Antonio Blanca, Pietro Caputo, Zongchen Chen, Daniel Parisi, Daniel v S tefankovi v c , and Eric Vigoda. On mixing of Markov chains: Coupling, spectral independence, and entropy factorization. In Proceedings of the 2022 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages 3670--3692. SIAM, 2022

    work page 2022

  6. [6]

    Complexity of high-dimensional identity testing with coordinate conditional sampling

    Antonio Blanca, Zongchen Chen, Daniel Stefankovic, and Eric Vigoda. Complexity of high-dimensional identity testing with coordinate conditional sampling. In The Thirty Sixth Annual Conference on Learning Theory , pages 1774--1790. PMLR, 2023

    work page 2023

  7. [7]

    -Entropic Measures of Correlation

    Salman Beigi and Amin Gohari. -Entropic Measures of Correlation . IEEE Transactions on Information Theory , 64(4):2193--2211, 2018

    work page 2018

  8. [8]

    Minimum contrast estimators on sieves: exponential bounds and rates of convergence

    Lucien Birg\' e and Pascal Massart. Minimum contrast estimators on sieves: exponential bounds and rates of convergence. Bernoulli , 4(3):329--375, 1998

    work page 1998

Show all 35 references
  1. [9]

    Quantum state certification

    Costin B a descu, Ryan O'Donnell, and John Wright. Quantum state certification. In Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing , pages 503--514, 2019

    2019

  2. [10]

    Modified logarithmic Sobolev inequalities in discrete settings

    Sergey G Bobkov and Prasad Tetali. Modified logarithmic Sobolev inequalities in discrete settings. Journal of Theoretical Probability , 19:289--336, 2006

    2006

  3. [11]

    Clément L. Canonne. A short note on learning discrete distributions, 2020

    2020

  4. [12]

    Canonne, Xi Chen, Gautam Kamath, Amit Levi, and Erik Waingarten

    Cl \'e ment L. Canonne, Xi Chen, Gautam Kamath, Amit Levi, and Erik Waingarten. Random restrictions of high dimensional distributions and uniformity testing with subcube conditioning. In Proceedings of the 2021 ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages 321--336....

    2021

  5. [13]

    Localization schemes: A framework for proving mixing bounds for Markov chains

    Yuansi Chen and Ronen Eldan. Localization schemes: A framework for proving mixing bounds for Markov chains. In 2022 IEEE 63rd Annual Symposium on Foundations of Computer Science (FOCS) , pages 110--122. IEEE, 2022

    2022

  6. [14]

    On the power of conditional samples in distribution testing

    Sourav Chakraborty, Eldar Fischer, Yonatan Goldhirsh, and Arie Matsliah. On the power of conditional samples in distribution testing. In Proceedings of the 4th conference on Innovations in Theoretical Computer Science , pages 561--580, 2013

    2013

  7. [15]

    Fast sampling of satisfying assignments from random- SAT with applications to connectivity

    Zongchen Chen, Andreas Galanis, Leslie Ann Goldberg, Heng Guo, Andr \'e s Herrera-Poyatos, Nitya Mani, and Ankur Moitra. Fast sampling of satisfying assignments from random- SAT with applications to connectivity. SIAM Journal on Discrete Mathematics , 38(4):2750--2811, 2024

    2024

  8. [16]

    Entropies, Convexity, and Functional Inequalities, On -Entropies and - Sobolev Inequalities

    Djalil Chafa \" . Entropies, Convexity, and Functional Inequalities, On -Entropies and - Sobolev Inequalities . Journal of Mathematics of Kyoto University , 44(2):325--363, 2004

    2004

  9. [17]

    Learning and testing junta distributions with sub cube conditioning

    Xi Chen, Rajesh Jayaram, Amit Levi, and Erik Waingarten. Learning and testing junta distributions with sub cube conditioning. In Conference on Learning Theory , pages 1060--1113. PMLR, 2021

    2021

  10. [18]

    Cafer Caferov, Bar s Kaya, Ryan O’Donnell, and A.C. Cem Say. Optimal bounds for estimating entropy with pmf queries. In International Symposium on Mathematical Foundations of Computer Science , pages 187--198. Springer, 2015

    2015

  11. [19]

    Optimal mixing of G lauber dynamics: E ntropy factorization via high-dimensional expansion

    Zongchen Chen, Kuikui Liu, and Eric Vigoda. Optimal mixing of G lauber dynamics: E ntropy factorization via high-dimensional expansion. In Proceedings of the 53rd Annual ACM SIGACT Symposium on Theory of Computing , pages 1537--1550, 2021

    2021

  12. [20]

    From algorithms to connectivity and back: finding a giant component in random k- SAT

    Zongchen Chen, Nitya Mani, and Ankur Moitra. From algorithms to connectivity and back: finding a giant component in random k- SAT . In Proceedings of the 2023 Annual ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages 3437--3470. SIAM, 2023

    2023

  13. [21]

    Approximate tensorization of entropy at high temperature

    Pietro Caputo, Georg Menz, and Prasad Tetali. Approximate tensorization of entropy at high temperature. In Annales de la Facult \'e des sciences de Toulouse: Math \'e matiques , volume 24, pages 691--716, 2015

    2015

  14. [22]

    Testing probability distributions underlying aggregated data

    Cl \'e ment Canonne and Ronitt Rubinfeld. Testing probability distributions underlying aggregated data. In International Colloquium on Automata, Languages, and Programming , pages 283--295. Springer, 2014

    2014

  15. [23]

    Servedio

    Cl \'e ment Canonne, Dana Ron, and Rocco A. Servedio. Testing equivalence between distributions using conditional samples. In Proceedings of the twenty-fifth annual ACM-SIAM symposium on Discrete algorithms , pages 1174--1192. SIAM, 2014

    2014

  16. [24]

    Which distribution distances are sublinearly testable? In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms , pages 2747--2764

    Constantinos Daskalakis, Gautam Kamath, and John Wright. Which distribution distances are sublinearly testable? In Proceedings of the Twenty-Ninth Annual ACM-SIAM Symposium on Discrete Algorithms , pages 2747--2764. SIAM, 2018

    2018

  17. [25]

    The design of experiments , volume 21

    Ronald Aylmer Fisher. The design of experiments , volume 21. Springer, 1966

    1966

  18. [26]

    Faster algorithms for testing under conditional sampling

    Moein Falahatgar, Ashkan Jafarpour, Alon Orlitsky, Venkatadheeraj Pichapati, and Ananda Theertha Suresh. Faster algorithms for testing under conditional sampling. In Conference on Learning Theory , pages 607--636. PMLR, 2015

    2015

  19. [27]

    Flammia and Ryan O'Donnell

    Steven T. Flammia and Ryan O'Donnell. Quantum chi-squared tomography and mutual information testing. Quantum , 8:1381, 2024

    2024

  20. [28]

    Brice Huang, Sidhanth Mohanty, Amit Rajaraman, and David X. Wu. Weak Poincar \' e inequalities, simulated annealing, and sampling from spherical spin glasses. CoRR , abs/2411.09075, 2024

    2024 arXiv

  21. [29]

    Modified log- S obolev inequalities for strong- R ayleigh measures

    Jonathan Hermon and Justin Salez. Modified log- S obolev inequalities for strong- R ayleigh measures. The Annals of Applied Probability , 33(2):1501--1514, 2023

    2023

  22. [30]

    Efficiently learning and sampling multimodal distributions with data-based initialization

    Frederic Koehler, Holden Lee, and Thuy-Duong Vuong. Efficiently learning and sampling multimodal distributions with data-based initialization. 2024

    2024

  23. [31]

    Logarithmic Sobolev inequality for some models of random walks

    Tzong-Yow Lee and Horng-Tzer Yau. Logarithmic Sobolev inequality for some models of random walks. The Annals of Probability , 26(4):1855--1873, 1998

    1998

  24. [32]

    On tolerant distribution testing in the conditional sampling model

    Shyam Narayanan. On tolerant distribution testing in the conditional sampling model. In Proceedings of the 2021 ACM-SIAM Symposium on Discrete Algorithms (SODA) , pages 357--373. SIAM, 2021

    2021

  25. [33]

    Estimating the effective support size in constant query complexity

    Shyam Narayanan and Jakub Tetek. Estimating the effective support size in constant query complexity. In Symposium on Simplicity in Algorithms (SOSA) , pages 242--252. SIAM, 2023

    2023

  26. [34]

    Karl Pearson. X. On the criterion that a given system of deviations from the probable in the case of a correlated system of variables is such that it can be reasonably supposed to have arisen from random sampling . The London, Edinburgh, and Dublin Philosophical Magazine and J...

    1900

  27. [35]

    A sharp log- Sobolev inequality for the multislice

    Justin Salez. A sharp log- Sobolev inequality for the multislice. Annales Henri Lebesgue , 4:1143--1161, 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.