Pith. sign in

REVIEW 3 major objections 4 minor 38 references

Near-optimal algorithms for private estimation and sequential testing of collision probability

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read The paper claims support-size-free near-optimal sample complexities for private collision-probability estimation and for sequential testing with unknown separation, achieved by counting collisions among all pairs of samples.

desk verdict The private estimator is a real advance and appears sound; the sequential tester's proof has a load-bearing algebra error in the centering of the test statistic, so the sequential-testing claims are false as written. read the letter →

arxiv 2504.13804 v1 pith:MVYBH3L3 submitted 2025-04-18 stat.ML cs.AIcs.LG

Robert Busa-Fekete , Umar Syed This is my paper · ORCID
classification stat.MLcs.AIcs.LG MSC 62L1062G0568W2094A17
keywords collisionprobabilitylocaldifferentialprivacysequentialhypothesistestingsamplecomplexityfrequencymomentsRényientropyU-statisticsmartingaleconcentration
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 sets out to show that collision probability — the chance two independent draws from a discrete distribution coincide — can be both privately estimated and sequentially tested with sample counts that are nearly the best possible and do not depend on the size of the distribution's support. For estimation under $(\alpha,\beta)$-local differential privacy, it proposes a one-bit-per-user mechanism whose stated sample complexity is $\tilde O(\log(1/\beta)\log(1/\delta)/(\alpha^2 \epsilon^2))$ for $\alpha \le 1$, a factor $1/\alpha^2$ better than prior support-size-free work. For testing, it proposes an adaptive sequential algorithm that is claimed to reject the null $C(p)=c_0$ within $\tilde O(\epsilon^{-2} \log\log(1/\epsilon) \log(1/\delta))$ samples whenever $|C(p)-c_0| \ge \epsilon$, even when $\epsilon$ is unknown. If both claims hold, these are near-minimax algorithms, support-size independent, and a private sequential tester follows by plugging the hashing into the sequential procedure.

What carries the argument

The load-bearing object is the all-pairs collision frequency $U_m = \frac{2}{m(m-1)}\sum_{i<j}\mathbf{1}\{X_i=X_j\}$, which estimates $C(p)$ directly. Because the pairs overlap, the paper decouples the kernel into a zero-mean martingale via a standard decoupling theorem; the sequential tester's statistic $T_i = \sum_{j<i}\mathbf{1}\{X_i=X_j\} - 2(i-1)c_0$ is meant to be controlled by a time-uniform confidence bound whose threshold shrinks like $\sqrt{(\log\log i)/i}$. For privacy, the mechanism uses a random one-bit hash with per-user salts; the salt count $r = 6((e^{\alpha}+1)/(e^{\alpha}-1))^2 \log(4/\beta)$ trades privacy against estimation noise.

What would settle it

Evaluate the identity for a two-point uniform distribution with $c_0 = 1/2$: direct expansion gives $\frac{2}{m(m-1)}\sum_{i=1}^m T_i = U_m - 2C(p)$, not $U_m - C(p)$. Running Algorithm 2 on synthetic data from this null distribution with $\delta = 0.05$ and checking whether the empirical rejection rate exceeds the promised confidence level would settle whether the sequential claim holds.

Watch

Extended reading notes

Core claim

In the paper's own terms, the central claim is that counting collisions among all $\Theta(n^2)$ pairs of samples — rather than only disjoint pairs — is both privacy-compatible and statistically efficient. Mechanism 1 hashes each user's sample together with a private random salt and has the server count hash collisions; Theorem 2 and Corollary 1 assert that the median of bias-corrected group estimates achieves absolute error $\epsilon$ with $\tilde O(\log(1/\beta)\log(1/\delta)/(\alpha^2 \epsilon^2))$ samples when $\alpha \le 1$. Algorithm 2 maintains a running all-pairs collision statistic and rejects as soon as it crosses a threshold shrinking like $\sqrt{(\log\log i)/i}$; Theorem 5 asserts that under the null it never rejects with probability $1-\delta$, and under the alternative it stops after $\tilde O(\epsilon^{-2}\log\log(1/\epsilon)\log(1/\delta))$ samples. Theorem 6 asserts matching lower bounds up to log-log factors.

Load-bearing premise

The sequential tester's type-I error guarantee depends on the displayed identity $\frac{2}{m(m-1)}\sum_{i=1}^m T_i(p) = U_m - C(p)$, which the proof uses to claim the running statistic is centered; the proof's own expansion yields a different center, so this is the premise most worth checking.

Editorial extensions

If this is right

  • Mechanism 1 estimates $C(p)$ with additive error $\epsilon$ under $(\alpha,\beta)$-local differential privacy using $\tilde O(\log(1/\beta)\log(1/\delta)/(\alpha^2 \epsilon^2))$ samples when $\alpha \le 1$, with only one bit sent per user and no dependence on support size.
  • Algorithm 2 distinguishes $C(p)=c_0$ from $|C(p)-c_0| \ge \epsilon$ using $\tilde O(\epsilon^{-2}\log\log(1/\epsilon)\log(1/\delta))$ samples with probability $1-\delta$, even though $\epsilon$ is not given to the algorithm.
  • Theorem 6 shows any sequential tester needs $\Omega(\log(1/\delta)/\epsilon^2)$ samples in the worst case, so Algorithm 2 is optimal up to log-log factors.
  • Theorem 3 shows any $(\alpha,0)$-locally differentially private estimator requires $\omega(1/(\alpha^2 \epsilon^2))$ samples for small $\alpha$, matching Corollary 1 up to logarithmic factors.
  • Theorem 4 shows that reducing private collision-probability estimation to private distribution estimation cannot give support-size-independent guarantees, so direct all-pairs collision counting is necessary for the claimed improvement.

Reading between the lines

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

  • Even if the algebraic centering in the current sequential proof needs adjustment, the all-pairs decoupling recipe is reusable: a tester built on the directly centered statistic $U_m - c_0$ could be evaluated against the paper's power-law and exponential benchmarks.
  • The same time-uniform martingale approach could in principle yield sequential tests for other degenerate $U$-statistics, such as Gini mean difference or Kendall's tau; the paper does not draw this connection.
  • Because Corollary 1's stated bound carries a factor $C(p)$, the absolute-error guarantee is easiest when $C(p)$ is not extremely small; for distributions over very large alphabets, the relative-error version of Theorem 2 may be the more honest benchmark, a regime the experiments do not stress.
  • If the support-size-free results extend to infinite alphabets, collision probability becomes a practical, communication-efficient test statistic over very large item spaces; deployment questions of that kind are not discussed in the paper.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes two contributions: (i) a locally differentially private estimator for collision probability C(p) of a discrete distribution (Mechanism 1, Theorems 1-2, Corollary 1), with sample complexity O~(log(1/beta)/(alpha^2 epsilon^2)) for alpha <= 1 and support-size-independent guarantees, together with a lower bound (Theorem 3) and a separation result for reduction to private distribution estimation (Theorem 4); and (ii) a sequential testing algorithm (Algorithm 2, Theorem 5) that claims to distinguish C(p)=c0 from |C(p)-c0| >= epsilon using O~(epsilon^{-2} log log 1/epsilon log 1/delta) samples even when epsilon is unknown, plus a private variant (Algorithm 3, Theorem 8). The paper also reports experiments comparing the proposed methods with prior work.

Significance. If correct, the private estimation result would be a genuine improvement over Bravo-Hermsdorff et al. (2022) by a factor of 1/alpha^2, and the sequential tester would be a near-optimal, support-size-free procedure with a time-uniform guarantee. The private estimation half appears technically sound: the bias correction E[C_j]=C(p), the variance bound of Proposition 1, the median-of-means concentration, and the salt-based privacy argument are all checkable and internally consistent. The sequential testing half, however, contains a load-bearing algebraic error that invalidates both Theorem 5 and its private analogue, so the paper's central claimed contribution is not established as written.

major comments (3)
  1. [Section 13, identity after Eq. (9)] The proof of Theorem 5 asserts that (2/(m(m-1))) sum_{i=1}^m T_i(p) = U_m - C(p), where T_i(p) = sum_{j<i} 1{X_i=X_j} - 2(i-1)C(p). Direct expansion gives sum_i T_i(p) = sum_i sum_{j<i} 1{X_i=X_j} - 2C(p) sum_i (i-1) = (m(m-1)/2) U_m - 2C(p) (m(m-1)/2), so the normalized statistic equals U_m - 2C(p), not U_m - C(p). This is not a cosmetic constant error: under the null hypothesis C(p)=c0, the test statistic in Algorithm 2, line 5, fluctuates around -c0 instead of 0, while the rejection threshold 3.2 sqrt((log log i + 0.72 log(20.8/delta))/i) tends to 0. Consequently, for any fixed c0>0 and delta, the probability that Algorithm 2 eventually rejects a true null hypothesis tends to 1, directly contradicting the first claim of Theorem 5.
  2. [Section 13, decomposition of Y_i(p) into T_i(p) and E_i(p)] The subsequent martingale argument relies on the false identity to conclude that '2/(m(m-1)) T_i(p) stays close to zero' under the null and that the difference between the statistics under p and under p0 equals |C(p)-C(p0)|. Because the correct centered statistic is U_m - 2C(p), the alternative hypothesis is also mishandled: when C(p) < c0, the mean of the statistic under the alternative is -C(p), which is smaller in absolute value than the null mean -c0, so the threshold test cannot reliably separate the two hypotheses even apart from the false-rejection problem. The proof therefore does not establish either the type-I or the type-II guarantee claimed in Theorem 5.
  3. [Section 16 and Algorithm 3, line 7] The Private Sequential Tester inherits the same centering error. With c = c0/(2r) + 1/2, the statistic T_i in Algorithm 3 subtracts 2(i-1)c, so the normalized sum equals the hashed-space U-statistic minus 2c, not minus c. Under the null hypothesis the statistic fluctuates around -c, and since the threshold in line 8 shrinks to zero, the algorithm rejects a true null with probability approaching 1. The proof of Theorem 8 in Section 16 explicitly invokes Theorem 5 for the confidence interval, so the null guarantee of Theorem 8 fails for the same algebraic reason.
minor comments (4)
  1. [Section 13, equations (10) and (11)] The summation indices in the probability statements appear to be typos: 'Pm i-1 Yi(p)' and 'Pm i-1 Ei(p)' should presumably read 'sum_{i=1}^m' rather than 'sum_{i-1}'; the intended statement is clear from context but should be corrected.
  2. [Algorithm 3, line 7] The indicator reads '1{vi = xj}', which compares a hash value to a raw sample; based on the surrounding text and Mechanism 1, the intended comparison is between hash values, i.e., '1{vi = vj}'.
  3. [Section 6.1] The text says the private sequential experiments use 'the same problem instance that were used for evaluating the non-private testers in Figure 4', but Figure 4 displays estimation error; the reference should likely be to Figure 3, where the non-private testers' sample complexities are plotted.
  4. [Section 13, text following Eq. (9)] The notation is inconsistent between C(p) and F2(p) throughout the proof; using a single symbol for collision probability would improve readability.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: core bounds are proved from external concentration inequalities and direct moment calculations; self-citation is only a baseline.

full rationale

The paper's derivation chain is self-contained against external benchmarks. Mechanism 1's accuracy proof (Section 9, Proposition 1) computes the bias and variance of the salted-hash collision statistic directly from the Poissonization trick and the Alon-Matias-Szegedy second-moment estimator, then applies median-of-means; no parameter is fitted to data and no claimed prediction is defined in terms of the target. The lower bound (Theorem 3) uses the Duchi-Jordan-Wainwright local-privacy Le Cam bound on two explicitly constructed distributions, so it does not import the paper's own upper bound as an assumption. For sequential testing, Section 13 builds a decoupled martingale whose increments are shown to be zero-mean, and then invokes the external time-uniform confidence sequences of Howard et al.; the constants in Algorithm 2 come from those concentration inequalities rather than from any fitted value. The only self-citation is to Bravo-Hermsdorff et al. (2022), which shares two authors with this paper; it appears as a baseline for comparison and as the object of explicit correction (footnote 2), not as the justification of any theorem. Accordingly, no load-bearing self-citation chain, no fitted-input-as-prediction, and no definitional equivalence between input and output occurs. A separate algebraic concern exists in Section 13: the identity (2/(m(m-1))) sum T_i(p) = U_m - C(p) appears to be mis-centered by a factor of two on direct expansion, which would be a correctness flaw in Theorem 5 rather than a circular step. That concern does not change the circularity verdict.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

No free parameters are fitted to data; privacy and threshold constants are analytic. The private estimator relies on standard hashing and Poissonization. The sequential tester additionally relies on time-uniform martingale bounds whose applicability is assumed; the algebraic centering identity in Section 13 is false as written, which is the load-bearing failure.

assumptions (5)
  • domain assumption Uniform random hash function h and user salts.
    Mechanism 1 and the privacy proof in Section 8 rely on h being uniformly random and salts being independent uniform draws from {1,...,r}.
  • standard math Poissonized group sizes N_j ~ Poisson(m) with independent cell counts.
    Proposition 1 in Section 9 uses the Poissonization trick to treat M_{s,x} as independent Poisson(m q_{s,x}) variables.
  • domain assumption Time-uniform confidence sequences of Howard et al. (2021) apply to the martingale sums in Section 13.
    The increments Y_i are only bounded by O(i) and the threshold constants in Algorithm 2 are not reconciled with the proof's constants; this is assumed rather than verified.
  • standard math Decoupling and degenerate kernel construction (de la Pena and Gine 1999) yields zero-mean martingales.
    Lemma 1 uses the centered kernel g_p to build the martingale sequence; this is a standard mathematical result.
  • domain assumption Finite support [k] with positive collision probability.
    The paper studies discrete distributions on [k], so C(p) >= 1/k > 0; this makes the null value c0 positive, which is the regime where the mis-centering error in Algorithm 2 causes false rejections.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Near-optimal algorithms for private estimation and sequential testing of collision probability." pith.science (2026). https://pith.science/paper/MVYBH3L3

@misc{pith2026250413804,
  author       = {Pith},
  title        = {Pith review of: Near-optimal algorithms for private estimation and sequential testing of collision probability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MVYBH3L3}},
  note         = {Machine review of arXiv:2504.13804}
}
abstract

We present new algorithms for estimating and testing \emph{collision probability}, a fundamental measure of the spread of a discrete distribution that is widely used in many scientific fields. We describe an algorithm that satisfies $(\alpha, \beta)$-local differential privacy and estimates collision probability with error at most $\epsilon$ using $\tilde{O}\left(\frac{\log(1/\beta)}{\alpha^2 \epsilon^2}\right)$ samples for $\alpha \le 1$, which improves over previous work by a factor of $\frac{1}{\alpha^2}$. We also present a sequential testing algorithm for collision probability, which can distinguish between collision probability values that are separated by $\epsilon$ using $\tilde{O}(\frac{1}{\epsilon^2})$ samples, even when $\epsilon$ is unknown. Our algorithms have nearly the optimal sample complexity, and in experiments we show that they require significantly fewer samples than previous methods.

Figures

Figures reproduced from arXiv: 2504.13804 by the authors.

Figure 1
Figure 1. Sample complexity of private collision probability estimation mechanisms for [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Sample complexity of our sequential tester (Algorithm 2) compared to the sample complexity of [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Sample complexity of the sequential tester compared to the sample complexity of the batch testers. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Empirical absolute error of plug-in and U-statistic estimators when the data is generated from [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Sample complexity of private sequential testing algorithms with respect to the non-private es [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 27 canonical work pages

  1. [1]

    A competitive test for uniformity of monotone distributions

    Jayadev Acharya, Ashkan Jafarpour, Alon Orlitsky, and Ananda Suresh. A competitive test for uniformity of monotone distributions. In Carlos M. Carvalho and Pradeep Ravikumar, editors, Proceedings of the Sixteenth International Conference on Artificial Intelligence and Statistics, volume 31 of Proceedings of Machine Learning Research, pages 57--65, Scottsd...

    work page 2013

  2. [2]

    The complexity of estimating r \'e nyi entropy

    Jayadev Acharya, Alon Orlitsky, Ananda Theertha Suresh, and Himanshu Tyagi. The complexity of estimating r \'e nyi entropy. In Proceedings of the twenty-sixth annual ACM-SIAM symposium on Discrete algorithms, pages 1855--1869. SIAM, 2014

    work page 2014

  3. [3]

    Test without trust: Optimal locally private distribution testing

    Jayadev Acharya, Clement Canonne, Cody Freitag, and Himanshu Tyagi. Test without trust: Optimal locally private distribution testing. In Kamalika Chaudhuri and Masashi Sugiyama, editors, Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, volume 89 of Proceedings of Machine Learning Research, pages 2067--20...

    work page 2019

  4. [4]

    Hadamard response: Estimating distributions privately, efficiently, and with little communication

    Jayadev Acharya, Ziteng Sun, and Huanyu Zhang. Hadamard response: Estimating distributions privately, efficiently, and with little communication. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 1120--1129. PMLR, 2019 b

    work page 2019

  5. [5]

    The space complexity of approximating the frequency moments

    Noga Alon, Yossi Matias, and Mario Szegedy. The space complexity of approximating the frequency moments. Journal of Computer and System Sciences, 58 0 (1): 0 137--147, 1999

    work page 1999

  6. [6]

    P. Auer, N. Cesa-Bianchi, Y. Freund, and R.E. Schapire. Gambling in a rigged casino: The adversarial multi-armed bandit problem. In Proceedings of IEEE 36th Annual Foundations of Computer Science, pages 322--331, 1995

    work page 1995

  7. [7]

    Private and Communication-Efficient Algorithms for Entropy Estimation

    Gecia Bravo-Hermsdorff, R\' o bert Busa-Fekete, Mohammad Ghavamzadeh, Andres Munoz Medina, and Umar Syed. Private and communication-efficient algorithms for entropy estimation. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems, volume 35, pages 15382--15393. Curran Associates, ...

    work page Pith review arXiv 2022

  8. [8]

    Identity testing for mallows model

    R \' o bert Busa - Fekete, Dimitris Fotakis, Bal \' a zs Sz \" o r \' e nyi, and Emmanouil Zampetakis. Identity testing for mallows model. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 23179--23190, 2021. URL https://proceedings.ne...

    work page 2021

Show all 38 references
  1. [9]

    Locally differentially private estimation of functionals of discrete distributions

    Cristina Butucea and Yann Issartel. Locally differentially private estimation of functionals of discrete distributions. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, volume 34, pages 2...

    2021

  2. [10]

    Cl\' e ment L. Canonne. Topics and techniques in distribution testing: A biased but representative sample. Found. Trends Commun. Inf. Theory, 19 0 (6): 0 1032–1198, nov 2022 a . ISSN 1567-2190. doi:10.1561/0100000114. URL https://doi.org/10.1561/0100000114

    2022 doi

  3. [11]

    Topics and techniques in distribution testing

    Cl \'e ment L Canonne. Topics and techniques in distribution testing. Now Publishers, 2022 b

    2022

  4. [12]

    Prediction, learning, and games

    Nicol\' o Cesa-Bianchi and G\' a bor Lugosi. Prediction, learning, and games. Cambridge University Press, 2006

    2006

  5. [13]

    Join sizes, frequency moments, and applications

    Graham Cormode and Minos Garofalakis. Join sizes, frequency moments, and applications. In Data Stream Management: Processing High-Speed Data Streams, pages 87--102. Springer, 2016

    2016

  6. [14]

    Frequency estimation under local differential privacy

    Graham Cormode, Samuel Maddock, and Carsten Maple. Frequency estimation under local differential privacy. PVLDB Journal Proceedings, 14 0 (11): 0 2046--2058, 2021

    2021

  7. [15]

    Stochastic streams: Sample complexity vs

    Michael Crouch, Andrew McGregor, Gregory Valiant, and David P Woodruff. Stochastic streams: Sample complexity vs. space complexity. In 24th Annual European Symposium on Algorithms (ESA 2016). Schloss Dagstuhl-Leibniz-Zentrum fuer Informatik, 2016

    2016

  8. [16]

    Optimal Stopping Rules for Sequential Hypothesis Testing

    Constantinos Daskalakis and Yasushi Kawase. Optimal Stopping Rules for Sequential Hypothesis Testing . In 25th Annual European Symposium on Algorithms (ESA 2017), volume 87 of Leibniz International Proceedings in Informatics (LIPIcs), pages 32:1--32:14. Schloss Dagstuhl--Leibn...

    2017

  9. [17]

    de la Pe \ n a and E

    V. de la Pe \ n a and E. Gin \'e . Decoupling: From Dependence to Independence. Probability and Its Applications. Springer New York, 1999

    1999

  10. [18]

    Kane, and Vladimir Nikishkin

    Ilias Diakonikolas, Daniel M. Kane, and Vladimir Nikishkin. Testing identity of structured distributions. In Proceedings of the Twenty-Sixth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA 2015, San Diego, CA, USA, January 4-6, 2015 , pages 1841--1854, 2015

    2015

  11. [19]

    Concentration of Measure for the Analysis of Randomized Algorithms

    Devdatt Dubhashi and Alessandro Panconesi. Concentration of Measure for the Analysis of Randomized Algorithms. Cambridge University Press, USA, 1st edition, 2009

    2009

  12. [20]

    Duchi, Martin J

    John C. Duchi, Martin J. Wainwright, and Michael I. Jordan. Minimax optimal procedures for locally private estimation. CoRR, abs/1604.02390, 2016

    2016 arXiv

  13. [21]

    The algorithmic foundations of differential privacy

    Cynthia Dwork, Aaron Roth, et al. The algorithmic foundations of differential privacy. Foundations and Trends in Theoretical Computer Science , 9 0 (3--4): 0 211--407, 2014

    2014

  14. [22]

    Rappor: Randomized aggregatable privacy-preserving ordinal response

    \'U lfar Erlingsson, Vasyl Pihur, and Aleksandra Korolova. Rappor: Randomized aggregatable privacy-preserving ordinal response. In Proceedings of the 2014 ACM SIGSAC conference on computer and communications security, pages 1054--1067, 2014

    2014

  15. [23]

    Variabilit \`a e mutabilit \`a : contributo allo studio delle distribuzioni e delle relazioni statistiche.[Fasc

    Corrado Gini. Variabilit \`a e mutabilit \`a : contributo allo studio delle distribuzioni e delle relazioni statistiche.[Fasc. I.] . Tipogr. di P. Cuppini, 1912

    1912

  16. [24]

    On testing expansion in bounded-degree graphs

    Oded Goldreich and Dana Ron. On testing expansion in bounded-degree graphs. Electron. Colloquium Comput. Complex., 7 0 (20), 2000

    2000

  17. [25]

    Concentration in the steel industry

    Orris C Herfindahl. Concentration in the steel industry. Columbia University, 1997

    1997

  18. [26]

    Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon

    Steven R. Howard, Aaditya Ramdas, Jon McAuliffe, and Jasjeet Sekhon. Time-uniform, nonparametric, nonasymptotic confidence sequences . The Annals of Statistics, 49 0 (2): 0 1055 -- 1080, 2021

    2021

  19. [27]

    E ffective

    Markku Laakso and Rein Taagepera. “ E ffective” number of parties: A measure with application to west europe. Comparative Political Studies, 12 0 (1): 0 3--27, 1979. doi:10.1177/001041407901200101. URL https://doi.org/10.1177/001041407901200101

    1979 doi

  20. [28]

    Entropy and Diversity: The Axiomatic Approach

    Tom Leinster. Entropy and Diversity: The Axiomatic Approach. Cambridge University Press, 2021. doi:10.1017/9781108963558

    2021 doi

  21. [29]

    Mean estimation and regression under heavy-tailed distributions: A survey

    G \'a bor Lugosi and Shahar Mendelson. Mean estimation and regression under heavy-tailed distributions: A survey. Foundations of Computational Mathematics, 19 0 (5): 0 1145--1190, 2019

    2019

  22. [30]

    Information theory of dna shotgun sequencing

    Abolfazl S Motahari, Guy Bresler, and NC David. Information theory of dna shotgun sequencing. IEEE Transactions on Information Theory, 59 0 (10): 0 6273--6289, 2013

    2013

  23. [31]

    Sequential algorithms for testing closeness of distributions

    Aadil Oufkir, Omar Fawzi, Nicolas Flammarion, and Aur\' e lien Garivier. Sequential algorithms for testing closeness of distributions. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, vol...

    2021

  24. [32]

    Estimation of entropy and mutual information

    Liam Paninski. Estimation of entropy and mutual information. Neural Comput., 15 0 (6): 0 1191--1253, 2003

    2003

  25. [33]

    Moment Recurrence Relations for Binomial, Poisson and Hypergeometric Frequency Distributions

    John Riordan. Moment Recurrence Relations for Binomial, Poisson and Hypergeometric Frequency Distributions . The Annals of Mathematical Statistics, 8 0 (2): 0 103 -- 111, 1937. doi:10.1214/aoms/1177732430. URL https://doi.org/10.1214/aoms/1177732430

    1937

  26. [34]

    Measurement of diversity

    Edward H Simpson. Measurement of diversity. nature, 163 0 (4148): 0 688--688, 1949

    1949

  27. [35]

    Evaluating entropy for true random number generators: Efficient, robust and provably secure

    Maciej Skorski. Evaluating entropy for true random number generators: Efficient, robust and provably secure. In International Conference on Information Security and Cryptology, pages 526--541, 03 2017

    2017

  28. [36]

    Possible generalization of boltzmann-gibbs statistics

    Constantino Tsallis. Possible generalization of boltzmann-gibbs statistics. Journal of statistical physics, 52 0 (1): 0 479--487, 1988

    1988

  29. [37]

    Estimating the unseen: An n/log(n)-sample estimator for entropy and support size, shown optimal via new clts

    Gregory Valiant and Paul Valiant. Estimating the unseen: An n/log(n)-sample estimator for entropy and support size, shown optimal via new clts. In Proceedings of the Forty-Third Annual ACM Symposium on Theory of Computing, STOC '11, page 685–694, New York, NY, USA, 2011. Assoc...

    2011

  30. [38]

    Estimating means of bounded random variables by betting, 2020

    Ian Waudby-Smith and Aaditya Ramdas. Estimating means of bounded random variables by betting, 2020. URL https://arxiv.org/abs/2010.09686

    2020 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.