Large Deviations for Sequential Tests of Statistical Sequence Matching

Jingjing Wang; Lin Zhou; Qianyun Wang; Yun Wei

REVIEW 3 major objections 6 minor 25 references

Large Deviations for Sequential Tests of Statistical Sequence Matching

T0 review · 3 major / 6 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 The paper derives exact mismatch exponents for sequential tests of statistical sequence matching and proves the proposed test achieves them when the number of matches is known.

desk verdict Solid exponent computation for known K, but the stated optimality theorem is overbroad and the unknown-K theorem is missing a condition; worth refereeing after statement-level fixes. read the letter →

arxiv 2506.03976 v1 pith:LHRGU3B2 submitted 2025-06-04 cs.IT eess.SPmath.ITmath.STstat.TH

Lin Zhou , Qianyun Wang , Yun Wei , Jingjing Wang This is my paper

classification cs.ITeess.SPmath.ITmath.STstat.TH MSC 94A1594A1760F10

keywords statisticalsequencematchingsequentialtestmismatchexponentRényidivergenceexpectedstoppingtimeuniversalityfalsealarmrejectde-anonymization

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 the sequential version of statistical sequence matching, where two databases of i.i.d. sequences generated from unknown distributions must be matched. With a known number of matches K, it proves that every sequential test with bounded expected stopping time has mismatch exponent at most the quantity Es(l, K, P, Q) = min_{t != l} sum_{(i,j) in M_t^K \ M_l^K} $\alpha$ D_{$\beta$/($\alpha$+$\beta$)}(Q_j || P_i), and it constructs a non-parametric minimal-scoring test that attains this exponent. This settles the sequential analogue of the fixed-length matching problem and shows that sequentiality strictly improves the exponent over fixed-length tests. For unknown K, the paper proposes a sequential test with bounded expected stopping time under certain parameter conditions and characterizes the tradeoff among the exponential rates of mismatch, false alarm, and false reject probabilities. It also gives a one-step fixed-length test that matches or improves the prior two-step fixed-length test.

What carries the argument

The central object is the scoring function S_t^K($x^{{\xi n}}$, $y^{{\chi n}}$) = \sum_{(i,j) \in M_t^K} GJS(\hat{T}_{x_i}, \hat{T}_{y_j}, \$\alpha$, \$\beta$), where GJS(P, Q, \$\alpha$, \$\beta$) = \$\alpha$ D(P\|R) + \$\beta$ D(Q\|R) with R = (\$\alpha$ P + \$\beta$ Q)/(\$\alpha$ + \$\beta$) is a generalized Jensen-Shannon divergence that vanishes exactly when P = Q. The sequential test stops at the first n when some score is at most f(n), with f(n) decaying like (K+1)|\mathcal{X}| \log(n\$\alpha$+2)/n + K|\mathcal{X}|\log(n\$\beta$+2)/n, and then applies the minimal-scoring decision rule. The proof of the exact exponent uses type-class bounds on the probability that a competitive score is small, together with the variational form of the Renyi divergence, \$\alpha$ D_{\$\beta$/(\$\alpha$+\$\beta$)}(Q\|P) = \min_{V} [\$\alpha$ D(V\|P) + \$\beta$ D(V\|Q)]. The converse uses the data processing inequality for KL divergence and Doob's optional stopping theorem to upper bound any test's mismatch exponent.

What would settle it

Take a binary alphabet, set α = β = 1, M1 = M2 = 2, K = 1, with P1 = Q1 = Bern(0.2) as the true match and the alternative pair Q2 = Bern(0.8). Compute Es = D_{1/2}(Bern(0.8) || Bern(0.2)) and simulate the proposed sequential test with large N; if the empirical decay rate of the mismatch probability falls decisively below Es, the theorem's achievability claim is false.

Watch

Extended reading notes

Core claim

For the case of known number of matches, Theorem 1 establishes the exact large-deviation performance of sequential tests that satisfy expected stopping time universality: any such test has mismatch exponent at most Es(l, K, P, Q), and the sequential test that stops when a generalized Jensen-Shannon score falls below a slowly decaying threshold, then decides by the minimal score, achieves this exponent. The exponent is a min over alternative matchings of a sum of scaled Renyi divergences between the matched distributions that differ between the alternative and the true matching. For the unknown number of matches case, Theorem 3 bounds the achievable exponents of a sequential test in terms of three parameters (lambda1, lambda2, lambda3), showing a tradeoff between mismatch, false reject, and false alarm rates, and Theorem 5 provides a one-step fixed-length test that achieves the same false alarm and Bayesian exponents as the earlier two-step test, with a simpler procedure.

Load-bearing premise

The whole analysis assumes that each sequence is generated i.i.d. from a distinct unknown distribution on a finite alphabet and that a pair is matched if and only if the underlying distributions are identical; if distributions can be close but not equal, or if sequences are not i.i.d., the stated exponents no longer apply.

Editorial extensions

If this is right

Optimal sequential tests strictly improve the mismatch exponent over fixed-length tests, revealing the benefit of sequentiality in statistical sequence matching.
Specializing to statistical classification (M2 = K = 1) yields exact large deviations for sequential classification, recovering the binary classification result as a special case.
In the unknown-parameters setting, the proposed sequential test provides larger Bayesian exponent under each non-null hypothesis while achieving the same false alarm exponent as the two-step fixed-length test.
The exponent formula is monotone in both block-length ratios alpha and beta, approaching alpha D(Q_j || P_i) as beta tends to infinity, so the improvement from additional samples saturates.
The one-step fixed-length test eliminates the separate match-count estimation phase of prior fixed-length tests without hurting the achievable exponents.

Reading between the lines

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

The exact exponent formula Es suggests a natural family of testable predictions: for fixed alpha and beta, the worst-case alternative matching is the one that maximizes the additive Renyi cost, a structure reminiscent of a shortest-path or cost-geometry problem.
One could extend the analysis to continuous alphabets by replacing the empirical type scoring with kernel-based divergence estimates, though the type-counting steps would need reworking.
The parameter tradeoff in the unknown-K setting (lambda1 versus lambda2, lambda3) could be optimized per distribution family, yielding a distribution-dependent operating point that the current theorem only bounds.
A converse for the unknown-K case, which the paper leaves open, might take the form of a matching lower bound on the product of exponents; if it holds, the three-error tradeoff would be tight.

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.

Desk Editor's Note

read the letter

Sanjay,

Quick read of arXiv:2506.03976. The known-K part is genuinely new and largely right, but Theorem 1 as stated is false. The converse claims to bound the exponent of any sequential test with bounded expected stopping time. That class includes the trivial test that always outputs H_l after N samples: E[τ]=N, mismatch probability zero, so the limsup exponent is +∞, exceeding the finite bound Es. The proof itself only works for tests with positive mismatch exponent under every hypothesis – the sentence before (141) says exactly that. So the theorem statement needs that added condition (or a worst-case-exponent phrasing). This is a statement-level gap, not a sign the exponent formula is wrong; the data-processing derivation is standard and the algebra in (132)-(135) checks out.

What is new: the exact sequential mismatch exponent for known K, with a matching converse for the right class; the specialization to binary classification recovers Hsu-Li-Wang; the unknown-K achievability and the one-step fixed-length test are also new relative to Unnikrishnan and Zhou et al. The authors also give concrete numerical examples showing their Λ and κ do not simplify, which is useful.

Soft spots: besides Theorem 1, Theorem 3(ii) states bounded expected stopping time under λ2>0 and λ3<Λ, but the proof (around Eq. (189)) also needs λ2<κ_l^K and strict positivity of the corresponding exponent; that condition is in the proof but not the statement. The claimed strict benefit of sequentiality over fixed-length tests is only proven as an inequality (Ef ≤ Es); 'strictly larger' is not established. The one-step fixed-length test comparison also uses ≥ where a strict claim is made. Minor typos ("ur proposed") throughout.

Bottom line: the core exponent computation is solid and likely to be the right benchmark, and anyone working on sequential detection or matching error exponents will find this useful. But the authors need to correct the statements before publication: add the positive-exponent assumption to Theorem 1, add the λ2 condition to Theorem 3(ii), and soften the strictness claims. I would not cite the theorem in its current form, but I'd be glad to see the revised version.

For peer review: send it out, but with a request for major revision and with a referee who checks statement-versus-proof mismatches.

Best, [Your name]

Referee Report

3 major / 6 minor

Summary. The paper studies sequential tests for statistical sequence matching, where two databases of i.i.d. sequences are generated from unknown distinct distributions and a pair is matched if its distributions coincide. For the case where the number of matches K is known, the authors propose a non-parametric sequential test based on a minimal scoring function, prove that it satisfies an expected-stopping-time universality constraint, and claim an exact mismatch exponent Es(l,K,P,Q) = min_{t != l} sum_{(i,j) in M_t^K \ M_l^K} alpha D_{beta/(alpha+beta)}(Q_j || P_i) for all optimal sequential tests satisfying that constraint. For the case of unknown K, they propose a sequential test with thresholds (lambda1, lambda2, lambda3), prove bounded expected stopping time under certain conditions, and lower-bound the false-alarm, mismatch, and false-reject exponents. They also compare with fixed-length tests from Zhou et al. and propose a one-step fixed-length variant, claiming benefits of sequentiality.

Significance. If the central claims are corrected, the paper makes a substantial contribution: it extends the fixed-length statistical sequence matching results of Unnikrishnan and Zhou et al. to the sequential setting and gives a parameter-free, method-of-types derivation of an exact mismatch exponent for optimal sequential tests when K is known. The unknown-K part provides a nontrivial achievable tradeoff among three error exponents and includes an honest statement that no converse is given. The paper also explicitly localizes a claimed error in [10, Eq. (64)] and provides numerical illustrations. However, the exact-optimality theorem is currently stated for a class of tests that is too broad, and the proof of the unknown-K expected stopping time omits a condition used in its own derivation; these issues are load-bearing and must be fixed before the central claims can be accepted as stated.

major comments (3)

[Theorem 1 / Section V.B] The converse in Theorem 1, Eq. (32), is stated for every sequential test satisfying the expected-stopping-time universality constraint (7), but this class includes the constant-output test tau = N, phi_tau = H_l. For that test, under H_l the mismatch probability is exactly zero, so the limsup in (32) is +infinity, while Es(l,K,P,Q) is finite whenever an alternative t has P_t != Q_1 (e.g., M1 = M2 = 2, K = 1, l = 1, t = 2). The proof does not handle this case: the condition 'ensures positive mismatch exponent' appears only in the sentence immediately before (141), not in the theorem statement. Moreover, inequality (150), d(1 - beta(P_tilde), beta(P)) >= -log beta(P), is not valid as a finite-N inequality; an asymptotic version can be recovered under exponential decay of both error probabilities by tracking the a log(1/b) correction term, but that argument is not supplied. The statement and proof should be aligned by adding a positive-mismatch-exponent condition under every hypothesis, or by phrasing the optimality claim in terms of worst-case exponent.
[Theorem 3(ii) / Section VI.B.1] Theorem 3(ii) states that the expected stopping time is bounded when lambda2 > 0 and lambda3 < Lambda_l^K, but the proof's final bound (189) contains geometric-series denominators involving lambda2, F(lambda3, P, Q), lambda3, and G(lambda2, P, Q). By Lemma 1(iii), G(lambda2, P, Q) = 0 whenever lambda2 >= kappa_l^K (cf. (71)), so the h > K terms in (189) do not decay and the expected stopping time need not be bounded for lambda2 in that range, at least when K < M2. The proof actually concludes boundedness under 0 < lambda2 < kappa_l^K and lambda3 < Lambda_l^K; the theorem statement should include this additional upper bound on lambda2.
[Section V.A, Eq. (137)] The achievability proof correctly derives in (132)-(135) that the limit of Delta(n) is Es(l,K,P,Q) = min_{t != l} sum_{(i,j) in M_t^K \ M_l^K} alpha D_{beta/(alpha+beta)}(Q_j || P_i), but the concluding display (137) replaces this with min over individual pairs (i,j) not in M_l^K. Since the minimum over pairs is generally smaller than the minimum over sums, (137) proves only a weaker lower bound and does not match the theorem's claimed exponent. This appears to be a typographical slip, but as written the proof of the achievability part does not establish the stated exponent; the display should be corrected to the min over t of the sums.

minor comments (6)

[Theorem 5] The test in (92) uses the events B^n_{h,t} defined in (53), which depend on lambda3, but Theorem 5 defines only positive parameters (lambda1, lambda2). The proof in (228) appears to replace the lambda3 terms from (182) and (185) with lambda2; please define lambda3 explicitly or state that lambda3 = lambda2 is intended.
[Theorem 5 displays] In Eqs. (93)-(95) the test is denoted Phi^uk_Zhou, but the theorem statement defines Phi^uk_FL; this inconsistency should be corrected.
[Eq. (102)] The probability expression in (102) is written as a sum over the matched pairs only; after marginalizing over the unmatched sequences this is true, but the notation should be clarified to avoid the appearance that unmatched sequences are omitted from the joint distribution.
[Theorem 2] In Eq. (40) the liminf is written with /n, but the fixed-length test length is N; the variable should be N for consistency with the rest of the section.
[Section III.D] The comparison proves Ef(l,K,P,Q) <= Es(l,K,P,Q), which supports 'no worse' but not necessarily a strictly larger exponent; the wording 'larger mismatch exponent' should be qualified if equality can occur for some distributions.
[Section IV.B] The claim that [10, Eq. (64)] is incorrect is supported only by two numerical examples; if the paper intends to correct a prior published result, a derivation of the claimed non-equality would be more appropriate than numerical evidence alone.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: Theorem 1's exponent is derived in-paper, and self-citations to [10] are auxiliary rather than fitted to the claimed results.

full rationale

The central known-K result (Theorem 1) does not reduce to its inputs. The achievability proof bounds the mismatch probability by a type-class sum and evaluates its exponential rate as the variational Rényi expression in (135), giving exactly Es; the converse upper bounds the mismatch exponent via the data-processing inequality and optional sampling, then minimizes the resulting divergence over alternative hypotheses to recover the same Es in (153). No parameter is fitted to the exponent and no benchmark is used to define the claimed quantity. For unknown K, Theorem 3 is an achievability result; the bounded-expected-stopping-time argument uses Lemma 1(ii),(iii) and Eq. (63) from the authors' prior fixed-length paper [10], but those results are auxiliary properties of the exponent functions F and G, not the target exponents of this paper, and they are not calibrated to any data or to the conclusions they support. The paper itself states in Section VII that the unknown-K test is only an achievability result and that optimality remains open. The only substantive concern is a statement-level gap in the converse of Theorem 1: the theorem quantifies over all expected-stopping-time universal tests, while the proof at (141) adds 'ensures positive mismatch exponent' and uses the decay of beta under every hypothesis at (150)-(151); a constant-output test would violate the stated upper bound. That is a correctness issue about the theorem's quantifier, not a circularity, because the exponent formula itself is not assumed by the proof. Accordingly, no circular step is identified.

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

There are no invented physical entities or new postulates beyond the sequential tests themselves. The free parameters are design thresholds and sample-allocation rates, not fitted constants. The main external load is the i.i.d. distinct-distribution model and a set of prior-work lemmas from [10] used in the unknown-K analysis.

free parameters (5)

alpha (sample-rate parameter for first database) = design parameter, not fitted
Sequence lengths are xi_n = ceil(alpha n); the mismatch exponent and all bounded-stopping conditions depend on alpha. It is chosen by the analyst, not estimated from data.
beta (sample-rate parameter for second database) = design parameter, not fitted
Sequence lengths are chi_n = ceil(beta n); the mismatch exponent and all bounded-stopping conditions depend on beta.
lambda1 (null-hypothesis threshold) = chosen threshold
Defines the event A_n in Eq. (50). Under the null hypothesis, bounded expected stopping time requires lambda1 < G0(P,Q,alpha,beta); the false reject exponent equals lambda1.
lambda2 (overestimation threshold) = chosen threshold
Defines B^1_{h,t} in Eq. (51). The proof requires lambda2 > 0 and lambda2 < kappa_l^K(P,Q,alpha,beta) for the expected stopping time to be bounded under non-null hypotheses.
lambda3 (wrong-set threshold) = chosen threshold
Defines B^2_{h,t} in Eq. (52). Bounded expected stopping time requires lambda3 < Lambda_l^K(P,Q,alpha,beta); the mismatch exponent grows with lambda3.

assumptions (5)

domain assumption Each sequence is i.i.d. from a distinct unknown distribution on a finite alphabet, and matched iff the distributions are identical.
Invoked in Section II.A.1. This defines the inference problem; if distributions are not distinct or sequences are not i.i.d., the matched-pair model and the exponent formulas change.
domain assumption The relevant generating distributions have full support, so KL and Renyi divergences are finite.
Section III.A fixes (P,Q) with full support before defining D(P||Q). The type-based probability bounds would need separate handling at zero-probability symbols.
standard math Method-of-types bounds on type-class probabilities and numbers of types (Cover-Thomas Theorems 11.1.1 and 11.1.4).
Used throughout the achievability proofs, e.g., Eqs. (104)-(110), to convert type sums into exponential bounds.
standard math Data processing inequality for KL divergence and Doob's optional sampling theorem.
Used in the converse for Theorem 1, Eqs. (141)-(143), to upper bound the mismatch exponent of any sequential test with bounded expected stopping time.
domain assumption Lemma 1(ii),(iii) and Eq. (63) from Zhou et al. (TIT 2024) are taken as given.
The unknown-K section relies on these cited results from the authors' prior work, e.g., in the proof of Lemma 1 and in Eq. (70). They are not re-derived in this paper.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Large Deviations for Sequential Tests of Statistical Sequence Matching." pith.science (2026). https://pith.science/paper/LHRGU3B2

@misc{pith2026250603976,
  author       = {Pith},
  title        = {Pith review of: Large Deviations for Sequential Tests of Statistical Sequence Matching},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LHRGU3B2}},
  note         = {Machine review of arXiv:2506.03976}
}

read the original abstract

We revisit the problem of statistical sequence matching initiated by Unnikrishnan (TIT 2015) and derive theoretical performance guarantees for sequential tests that have bounded expected stopping times. Specifically, in this problem, one is given two databases of sequences and the task is to identify all matched pairs of sequences. In each database, each sequence is generated i.i.d. from a distinct distribution and a pair of sequences is said matched if they are generated from the same distribution. The generating distribution of each sequence is \emph{unknown}. We first consider the case where the number of matches is known and derive the exact exponential decay rate of the mismatch (error) probability, a.k.a. the mismatch exponent, under each hypothesis for optimal sequential tests. Our results reveal the benefit of sequentiality by showing that optimal sequential tests have larger mismatch exponent than fixed-length tests by Zhou \emph{et al.} (TIT 2024). Subsequently, we generalize our achievability result to the case of unknown number of matches. In this case, two additional error probabilities arise: false alarm and false reject probabilities. We propose a corresponding sequential test, show that the test has bounded expected stopping time under certain conditions, and characterize the tradeoff among the exponential decay rates of three error probabilities. Furthermore, we reveal the benefit of sequentiality over the two-step fixed-length test by Zhou \emph{et al.} (TIT 2024) and propose an one-step fixed-length test that has no worse performance than the fixed-length test by Zhou \emph{et al.} (TIT 2024). When specialized to the case where either database contains a single sequence, our results specialize to large deviations of sequential tests for statistical classification, the binary case of which was recently studied by Hsu, Li and Wang (ITW 2022).

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

25 extracted references · 24 canonical work pages

[1]

Asymptotically optimal classification for multiple tests with empirically observed statistics,

M. Gutman, “Asymptotically optimal classification for multiple tests with empirically observed statistics,” IEEE Trans. Inf. Theory , vol. 35, no. 2, pp. 401–408, 1989

work page 1989
[2]

Second-order optimal statistical classification,

L. Zhou, V . Y . F. Tan, and M. Motani, “Second-order optimal statistical classification,” Information and Inference: A Journal of the IMA, vol. 9, no. 1, pp. 81–111, 2020

work page 2020
[3]

Sequential tests of statistical hypotheses,

A. Wald, “Sequential tests of statistical hypotheses,” Ann. Math. Statist. , vol. 16, no. 2, pp. 117–186, 1945

work page 1945
[4]

Optimum character of the sequential probability ratio test,

A. Wald and J. Wolfowitz, “Optimum character of the sequential probability ratio test,” Ann. Math. Statist. , pp. 326–339, 1948. 30

work page 1948
[5]

Sequential classification with empirically observed statistics,

M. Haghifam, V . Y . F. Tan, and A. Khisti, “Sequential classification with empirically observed statistics,” IEEE Trans. Inf. Theory , vol. 67, no. 5, pp. 3095–3113, 2021

work page 2021
[6]

On universal sequential classification from sequentially observed empirical statistics,

C. Y . Hsu, C. F. Li, and I. H. Wang, “On universal sequential classification from sequentially observed empirical statistics,” in IEEE ITW, 2022, pp. 642–647

work page 2022
[7]

A unified study on sequentiality in universal classification with empirically observed statistics,

C.-F. Li and I.-H. Wang, “A unified study on sequentiality in universal classification with empirically observed statistics,” IEEE Trans. Inf. Theory, vol. 71, no. 3, pp. 1546–1569, 2025

work page 2025
[8]

De-anonymizing private data by matching statistics,

J. Unnikrishnan and F. M. Naini, “De-anonymizing private data by matching statistics,” in Proc. 51st Annu. Allerton Conf. , 2013, pp. 1616–1623

work page 2013

Show all 25 references

[9]

Asymptotically optimal matching of multiple sequences to source distributions and training sequences,

J. Unnikrishnan, “Asymptotically optimal matching of multiple sequences to source distributions and training sequences,” IEEE Trans. Inf. Theory, vol. 61, no. 1, pp. 452–468, 2015

2015
[10]

Large and small deviations for statistical sequence matching,

L. Zhou, Q. Wang, J. Wang, L. Bai, and A. Hero, “Large and small deviations for statistical sequence matching,” IEEE Trans. Inf. Theory, vol. 70, no. 11, pp. 7532 – 7562, 2024

2024
[11]

Asymptotic estimates in information theory with non-vanishing error probabilities,

V . Y . F. Tan, “Asymptotic estimates in information theory with non-vanishing error probabilities,” Foundations and Trends ® in Communications and Information Theory , vol. 11, no. 1–2, pp. 1–184, 2014

2014
[12]

Finite blocklength lossy source coding for discrete memoryless sources,

L. Zhou and M. Motani, “Finite blocklength lossy source coding for discrete memoryless sources,” Foundations and Trends ® in Communications and Information Theory , vol. 20, no. 3, pp. 157–389, 2023

2023
[13]

Achievable error exponents for two-phase multiple classification,

L. Zhou, J. Diao, and L. Bai, “Achievable error exponents for two-phase multiple classification,” arXiv 2210.12736, 2022

2022 arXiv
[14]

Rényi divergence and Kullback-Leibler divergence,

T. Van Erven and P. Harremos, “Rényi divergence and Kullback-Leibler divergence,” IEEE Trans. Inf. Theory , vol. 60, no. 7, pp. 3797–3820, 2014

2014
[15]

Divergence measures based on the shannon entropy,

J. Lin, “Divergence measures based on the shannon entropy,” IEEE Trans. Inf. Theory , vol. 37, no. 1, pp. 145–151, 1991

1991
[16]

Universal outlier hypothesis testing,

Y . Li, S. Nitinawarat, and V . V . Veeravalli, “Universal outlier hypothesis testing,”IEEE Trans. Inf. Theory , vol. 60, no. 7, pp. 4066–4082, 2014

2014
[17]

Distributed detection with empirically observed statistics,

H. He, L. Zhou, and V . Y . Tan, “Distributed detection with empirically observed statistics,” IEEE Trans. Inf. Theory , vol. 65, no. 7, pp. 4349–4367, 2020

2020
[18]

On binary statistical classification from mismatched empirically observed statistics,

H.-W. Hsu and I.-H. Wang, “On binary statistical classification from mismatched empirically observed statistics,” in IEEE ISIT, 2020, pp. 2533–2538

2020
[19]

Second-order asymptotically optimal outlier hypothesis testing,

L. Zhou, Y . Wei, and A. O. Hero, “Second-order asymptotically optimal outlier hypothesis testing,” IEEE Trans. Inf. Theory , vol. 68, no. 6, pp. 3585–3607, 2022

2022
[20]

The method of types [information theory],

I. Csiszar, “The method of types [information theory],” IEEE Trans. Inf. Theory , vol. 44, no. 6, pp. 2505–2523, 1998

1998
[21]

T. M. Cover and J. A. Thomas, Elements of Information Theory . John Wiley & Sons, 2012

2012
[22]

Optional sampling theorems,

A. Klenke and A. Klenke, “Optional sampling theorems,” Probability Theory: A Comprehensive Course , pp. 205–215, 2014

2014
[23]

A kernel two-sample test,

A. Gretton, K. Borgwardt, M. Rasch, B. Scholkopf, and A. Smola, “A kernel two-sample test,” J. Mach. Learn. Res. , vol. 13, pp. 723–773, 2012

2012
[24]

Asymptotics of sequential composite hypothesis testing under probabilistic constraints,

J. Pan, Y . Li, and V . Y . F. Tan, “Asymptotics of sequential composite hypothesis testing under probabilistic constraints,” IEEE Trans. Inf. Theory, vol. 68, no. 8, pp. 4998–5012, 2022

2022
[25]

Second-order asymptotics of sequential hypothesis testing,

Y . Li and V . Y . F. Tan, “Second-order asymptotics of sequential hypothesis testing,” arXiv:2001.04598, 2020

2001 arXiv

Pith tools

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

[1] [1]

Asymptotically optimal classification for multiple tests with empirically observed statistics,

M. Gutman, “Asymptotically optimal classification for multiple tests with empirically observed statistics,” IEEE Trans. Inf. Theory , vol. 35, no. 2, pp. 401–408, 1989

work page 1989

[2] [2]

Second-order optimal statistical classification,

L. Zhou, V . Y . F. Tan, and M. Motani, “Second-order optimal statistical classification,” Information and Inference: A Journal of the IMA, vol. 9, no. 1, pp. 81–111, 2020

work page 2020

[3] [3]

Sequential tests of statistical hypotheses,

A. Wald, “Sequential tests of statistical hypotheses,” Ann. Math. Statist. , vol. 16, no. 2, pp. 117–186, 1945

work page 1945

[4] [4]

Optimum character of the sequential probability ratio test,

A. Wald and J. Wolfowitz, “Optimum character of the sequential probability ratio test,” Ann. Math. Statist. , pp. 326–339, 1948. 30

work page 1948

[5] [5]

Sequential classification with empirically observed statistics,

M. Haghifam, V . Y . F. Tan, and A. Khisti, “Sequential classification with empirically observed statistics,” IEEE Trans. Inf. Theory , vol. 67, no. 5, pp. 3095–3113, 2021

work page 2021

[6] [6]

On universal sequential classification from sequentially observed empirical statistics,

C. Y . Hsu, C. F. Li, and I. H. Wang, “On universal sequential classification from sequentially observed empirical statistics,” in IEEE ITW, 2022, pp. 642–647

work page 2022

[7] [7]

A unified study on sequentiality in universal classification with empirically observed statistics,

C.-F. Li and I.-H. Wang, “A unified study on sequentiality in universal classification with empirically observed statistics,” IEEE Trans. Inf. Theory, vol. 71, no. 3, pp. 1546–1569, 2025

work page 2025

[8] [8]

De-anonymizing private data by matching statistics,

J. Unnikrishnan and F. M. Naini, “De-anonymizing private data by matching statistics,” in Proc. 51st Annu. Allerton Conf. , 2013, pp. 1616–1623

work page 2013