Model-free Rank Aggregation in the Presence of Rater Heterogeneity: A Maximum Score Approach

Haoran Zhang; Yunxiao Chen

REVIEW 3 major objections 4 minor 22 references

Model-free Rank Aggregation in the Presence of Rater Heterogeneity: A Maximum Score Approach

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper establishes that a maximum-score objective for rank aggregation is consistent and nearly minimax optimal under only weak stochastic transitivity, allowing sparse and nonuniform pairwise comparison data.

desk verdict Genuinely new batch-setting rank aggregation under WST with a sound main theorem, but the algorithm gap and a fixable proof error mean the paper needs revision before it can be trusted as stated. read the letter →

arxiv 2510.06789 v2 pith:S6IQSQBW submitted 2025-10-08 stat.ME

Haoran Zhang , Yunxiao Chen This is my paper

classification stat.ME MSC 62F0762G0562G20

keywords rankaggregationweakstochastictransitivitymaximumscoreestimatorKendall'stauminimaxoptimalitypairwisecomparisonssparsemissingdatanonparametricranking

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

Rank aggregation usually assumes strong stochastic transitivity (SST): if i usually beats j and j usually beats k, then i beats k at least as decisively. This paper shows that a maximum-score objective can recover the true global ranking under the much weaker condition that i beats k merely more often than not. MASTER maximizes a sum of signed pairwise win margins over permutations, and the paper proves that, under sparse pair sampling within a logarithmic density threshold, its modified Kendall distance to the truth is O_p(n log n/u_n), with a matching n/u_n lower bound. This makes the estimator nearly minimax optimal and consistent, meaning the proportion of discordant pairs tends to zero as the number of players grows. The weaker assumption matters because real pairwise-comparison data, such as tennis head-to-head records, often violates SST while still admitting a global ranking.

What carries the argument

The central object is the modified Kendall loss tau_m(pi,pi*)=tau(pi,pi*) Q(tau(pi,pi*))^2, where tau counts discordant pairs and Q(s) is the average of the s smallest values of q_ij=|2p*_ij-1| across all pairs. This weighting captures the difficulty of the ranking task: pairs with p*_ij near 1/2 are hard, so their contribution to the loss is small. The objective L(pi)=sum(2y_ij-n_ij)I(pi_i>pi_j) is the maximum-score engine; its expectation is maximized exactly at the true ranking. The technical work is a U-empirical-process argument: Bernstein concentration plus a union bound over permutations gives the upper rate, and an epsilon-packing/Fano argument over WST-constrained probability matric

What would settle it

Construct a WST probability matrix with many hard pairs at q_ij approximately c/sqrt(n), run the exact maximizer of L(pi) for moderate n with T and xi within the paper's range, and record the modified Kendall loss over many replications. If tau_m exceeds a constant multiple of n log n/u_n with non-negligible frequency, Theorem 2.1 fails. Alternatively, exhibit an instance where the K=3 search returns a ranking with L < L(pi*) with non-negligible probability, showing Proposition 2.1 does not apply to the delivered estimates.

Watch

Extended reading notes

Core claim

Under weak stochastic transitivity—if i beats j more often than not and j beats k more often than not, then i beats k more often than not—a global ranking exists, but the monotone constraints of strong stochastic transitivity are absent. MASTER is the permutation maximizer of L(pi)=sum_{i<j}(2y_ij-n_ij)I(pi_i>pi_j). The central result is that, provided each pair is sampled with probability between l_n and u_n satisfying l_n as u_n at least c log n/n, any ranking with L(pi_tilde) >= L(pi*) has modified Kendall loss tau_m = tau Q(tau)^2 = O_p(n log n/u_n), and no estimator can achieve a uniformly better rate than n/u_n over the WST parameter space. Therefore the proportion of discordant pairs

Load-bearing premise

The whole guarantee rests on the assumptions that all pairs are sampled with comparable probabilities at least c log n/n, and that one can actually find a ranking whose objective value is at least as high as the true ranking's; neither is proven for the algorithm exactly as implemented.

Editorial extensions

If this is right

If the central claim is right, practitioners no longer need strong stochastic transitivity to obtain consistent global rankings from pairwise comparisons; Borda-style and Bradley-Terry-based rankings can be biased when WST holds but SST fails.
Sparse data with nonuniform missingness—where pairs are more likely observed when players are close in strength—can still yield near-optimal rankings, as long as sampling probabilities are comparable and at least order log n/n.
The modified Kendall loss provides a principled accuracy statement: error is counted in discordant pairs weighted by how close the relevant win probabilities are to 0.5, and this weighted error converges at the near-minimax rate.
Under SST the method remains consistent, with only a modest efficiency loss relative to model-based estimators designed for SST, according to the paper's simulations.
Because WST is invariant to random flips of comparison outcomes, the guarantees carry over to privatized or randomized pairwise data without a debiasing correction.

Reading between the lines

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

Editorial: The theory covers any ranking whose objective value reaches or exceeds the true ranking's; Algorithm 1's K=3 tuple search has no proof of achieving that, and the paper itself notes its behavior in more general settings is unclear. The minimax guarantee thus attaches to the idealized optimizer, with the search algorithm as a heuristic bridge.
Editorial: The comparability assumption l_n ≈ u_n may fail in exactly the tennis setting, since top players meet top players more often than they meet lower-ranked players. Extending the bounds to strongly non-comparable sampling probabilities, perhaps with a reweighted objective, is the natural next step.
Editorial: Because the objective decomposes over pairs, the WST framework suggests a direct extension to k-way comparison data by pairwise decomposition; the paper notes this possibility but does not prove the rate preservation.

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

Here is my take for your internal file.

The paper deserves a serious referee. The core contribution—batch-setting rank aggregation under weak stochastic transitivity with nonuniform missingness, plus a near-minimax rate for a modified Kendall loss—is genuinely new and technically substantial. The upper-bound proof is sound: the expected-objective identity, the variance bound, and the Bernstein/union-bound argument all check out, and the monotone threshold trick that converts the bound to the modified loss is clean. The lower bound is likely true, but as printed it contains an algebraic mistake in the appendix that reverses the inequality; it is fixable with a sharper counting bound, but the derivation needs to be rewritten.

The bigger mismatch is between the theory and the algorithm. The theory covers any ranking whose objective is at least that of the truth, which includes the global maximizer of (1). Algorithm 1 is a K-tuple local search with no proof that it reaches such a point. Section 4 honestly discloses this, but the abstract and introduction still sell consistency and minimaxity without the caveat. Either provide a guarantee for a tractable algorithm, or reframe the theory as applying to the oracle maximizer and present the algorithm as a heuristic demonstrated in simulations. The simulation evidence is strong when SST is violated, but it doesn't compensate for the missing guarantee.

Two smaller things. The consistency claim in the abstract—that the proportion of discordant pairs converges to zero—does not follow from WST alone under the stated assumptions; Theorem 2.1 gives a modified loss, and the standard Kendall proportion requires extra conditions on how many pairs have win probabilities near 0.5. Remark 2.4 notes cases, but the abstract oversells. Also, the reported Kendall proportions in Table 1 seem to be off by a factor of four relative to the formula in the text; check the normalization.

The authors clearly know the literature and have not overclaimed the novelty relative to sequential WST work. The paper would benefit from a revision, but the main idea is solid. I'd send it to referees.

Referee Report

3 major / 4 minor

Summary. This paper studies rank aggregation from pairwise comparison data under weak stochastic transitivity (WST). The proposed MASTER estimator maximizes the objective L(π) = ∑_{i<j} (2y_ij − n_ij) I(π_i > π_j) over permutations. Under Assumption 2.1 (pair-sampling probabilities l_n ≍ u_n ≳ log n/n and bounded per-pair trials), Theorem 2.1 establishes that the modified Kendall loss τ_m(π̂,π*) = τ(π̂,π*)·Q(τ(π̂,π*))² of the global maximizer is O_p((n log n)/u_n); Proposition 2.1 extends this to any ranking π̃ with L(π̃) ≥ L(π*) with probability approaching 1. Theorem 2.2 gives a lower bound n/u_n over the WST parameter space, so the bounds match up to a logarithmic factor. Algorithm 1, a K-tuple local search initialized by a smoothed maximum-score fit, is proposed as the practical solver; simulations and a tennis application illustrate empirical performance.

Significance. If the algorithmic gap is closed, this is a significant theoretical contribution: it relaxes strong stochastic transitivity (SST) to WST in non-sequential rank aggregation, handles sparse and mildly nonuniform sampling, and gives matching upper and lower bounds up to a log factor. The U-empirical-process proof of Theorem 2.1 is largely sound, and Proposition 2.1 usefully isolates the required optimization property. However, the paper's central practical claim—that the implemented MASTER method is consistent and nearly minimax optimal—rests on an unproven property of Algorithm 1. The authors are candid about this in Section 4, but the gap is load-bearing. With the algorithmic issue addressed or the claims appropriately scoped, the paper would be a solid addition to the literature.

major comments (3)

[§2.3, §4, Proposition 2.1] Proposition 2.1 and Theorems 2.1–2.2 are stated for the global maximizer of (1), or for any π̃ with L(π̃) ≥ L(π*). The implemented estimator, however, is the output of Algorithm 1. The paper states in §2.3 that 'There is no guarantee that Algorithm 1 finds the global optimal solution' and in §4 that 'its performance under more general settings is unclear.' The only support is the simulation statement that 'more than 95% of the experiments' satisfy the condition. This is not an asymptotic guarantee. Consequently, the consistency and near-minimax optimality of the delivered MASTER procedure are not established. The authors should either prove that a computationally feasible variant reaches the Proposition 2.1 condition with probability approaching 1, or explicitly separate the oracle theory from the heuristic algorithm and avoid claiming that the implemented method is minimax optimal.
[Appendix, proof of Theorem 2.2, Eq. (11)] The displayed inequality chain in (11) contains an algebraic error. The step log(r/n)! + (n − r/n)log(r/n) − n − n log(1 + ǫ/n) = n log(r/n) + ǫ − r/n + O(log r/n) replaces −n log(1 + ǫ/n) by +ǫ and drops the −n term. This is invalid when ǫ/n = ⌊1/u_n⌋ can diverge; in that regime −n log(1 + ǫ/n) ≈ −n log(1/u_n), which cancels the leading term of n log(r/n). A correct lower bound would be log M ≥ n log(r/(n+ǫ)) − r/n − n + O(log(r/n)), and the claimed 'log 5' constant is not a consequence (the relevant log ratio tends to log 10 as u_n → 0). The Fano conclusion may still follow with a corrected packing bound and sufficiently small δ, but the proof as written is incomplete.
[Assumption 2.1 and §1/§3.2] The paper advertises nonuniform missingness: the Introduction says missingness probabilities 'are now allowed to depend on the ranks of the players' and the abstract says 'possibly nonuniform missingness.' But Assumption 2.1 requires l_n ≍ u_n, so all pair-sampling probabilities must be within a constant factor of each other. This is a mild inhomogeneity only. The tennis data, where pair-meeting probabilities are strongly rank-dependent, likely violates the condition. This does not invalidate the theorems, but the scope claims should be calibrated to what the assumption actually permits.

minor comments (4)

[§3.1] The formula for the proportion of discordant pairs is given as τ(π̂,π*)/(2n(n−1)); the correct proportion is τ(π̂,π*)/(n(n−1)/2) = 2τ(π̂,π*)/(n(n−1)). As written, the formula is inconsistent with the Table 1 values.
[Algorithm 1] There is a typo: 'updata t by t + 1' should be 'update t by t + 1.' Also, the reset rule 'set t = K' after an improvement could be described more precisely.
[Remark 2.4] The integral approximation for Q(s) under uniform q_ij gives Q(s) ≈ s/[n(n−1)], not 2s/[n(n−1)]. The subsequent rate τ = O_p(n^{5/3} u_n^{-1/3} log^{1/3} n) is unaffected.
[Various] Theorem statements refer to 'Assumption 1' but the assumption is numbered 2.1. Also 'del Petro' in Table 2 should be 'del Potro.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical derivation is self-contained; the main caveat is a computational/algorithmic gap, which is not a circularity.

full rationale

The load-bearing derivation is not circular. MASTER's objective (1) is defined from the raw data (nij, yij): L(pi) = sum_{1<=i<j<=n} (2yij - nij) I(pi_i > pi_j), with no fitted or tuned parameter entering the theory. The upper bound (Theorem 2.1) is proved for the global maximizer of L by a U-empirical/union-bound argument in the Appendix; the modified loss tau_m uses the true qij = |2p*_ij - 1| only as an analytical metric, not as an estimated target, and the lower bound (Theorem 2.2) is constructed independently via Fano's lemma over the hard set P_delta. No step reduces a prediction to a fitted input: Proposition 2.1 transfers the rate to any pi_tilde with L(pi_tilde) >= L(pi*) and is a logical corollary of the same proof, not a rename of the conclusion. Self-citations ([4], [10]) are contextual (sequential design, no-WST estimation) and are not load-bearing for Theorems 2.1-2.2. The genuine limitation - Algorithm 1 has no theorem ensuring L(pi_curr) >= L(pi*), with Section 4 conceding 'its performance under more general settings is unclear' - is a computational gap between the estimator analyzed and the delivered algorithm, not a circularity: the analyzed estimator is well-defined even if intractable, and the paper does not claim the search step is forced by the theory. Therefore the paper's open problem is algorithmic, not definitional.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The ledger is light: the theory introduces no fitted constants — T, ξ_ij, p_ij, F_ij are all model inputs, and the only hand-chosen algorithm number is K=3. The real burden sits in three domain assumptions: the binomial sampling scheme that makes the expected-objective identity exact; WST plus strictness p_ij ≠ 1/2 (which excludes exactly-tied pairs and makes the ranking unique); and Assumption 2.1's l_n ≍ u_n, which bounds how non-uniform the missingness can be — in tension with the tennis application. One ad hoc premise carries the practical claims: Algorithm 1's output must beat L(π*) with probability → 1, which is unproven. Standard concentration/information-theoretic lemmas underwrite the proofs; the only nonstandard technical object — the Q-weighted loss — is definitional, not postulated.

free parameters (3)

T (maximum per-pair comparisons) = 5 in simulations; fixed constant in the theory
§2.1: n_ij ~ Binomial(T, ξ_ij) with T constant. Not fitted, but load-bearing: the variance bound (4) and Bernstein bound scale with T, and the per-pair signal is weak (T ~ 5 means ~2 games per pair).
ξ_ij (pair sampling probabilities) = Uniform(0.3, 0.5) in simulations; unobserved in the tennis application
§2.1/Assumption 2.1. The theory takes them as given under l_n ≍ u_n ≳ log n/n; they are not estimated. In the tennis data the meeting probabilities are rank-dependent, which strains l_n ≍ u_n.
K (tuple search length) = 3
§2.3/§3.1 hand-chosen; affects whether the search attains L(π̃) ≥ L(π*); no theory links K to the guarantee, and §4 flags computational behavior as unclear.

assumptions (6)

domain assumption Data: n_ij ~ Binomial(T, ξ_ij) (i>j) independent; y_ij | n_ij ~ Binomial(n_ij, p_ij); p_ij + p_ji = 1 for all pairs
§2.1. The identity EL(π) = Σ Tξ_ij(2p*_ij−1)I(·) and the variance bound (4) both require this binomial sampling model.
domain assumption WST holds and p_ij ≠ 1/2 for all i ≠ j (unique global ranking)
§2.1 ('we assume that data follows WST, and the global ranking implied by the WST assumption is unique'). Defines the parameter class P_WST over which both theorems are stated.
domain assumption Assumption 2.1: l_n = min ξ_ij ≍ max ξ_ij = u_n and u_n ≳ log n/n
§2.2. Needed so the Bernstein/union-bound tail e^{2n log n − ε²/(cT u_n)} vanishes and the ratio l_n/u_n stays constant; excludes strongly rank-dependent missingness of the sort plausibly present in the tennis application.
ad hoc to paper Algorithm 1 returns π̃ with L(π̃) ≥ L(π*) with probability → 1
§2.3/Proposition 2.1. Unproven — supported only by simulation (>95% at K=3); if the search stalls at a local optimum, the rates do not apply to the implemented method.
standard math Standard tools: Bernstein's inequality, union bound, Fano's lemma ([20] Prop. 15.12), inversion-ball cardinality bounds ([22] Lemma 7; [20] Lemma 5.5)
Appendix, proofs of Theorems 2.1–2.2. Cited standard results not re-derived.
standard math Existence of pair-specific CDFs F_ij (strictly monotone, F_ij(0)=0.5) realizing any WST matrix; latent scores β_i
§2.1. Definitional under WST; used for intuition and for the (cited) claim that p_ij cannot be consistently estimated — not used in the main proofs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Model-free Rank Aggregation in the Presence of Rater Heterogeneity: A Maximum Score Approach." pith.science (2026). https://pith.science/paper/S6IQSQBW

@misc{pith2026251006789,
  author       = {Pith},
  title        = {Pith review of: Model-free Rank Aggregation in the Presence of Rater Heterogeneity: A Maximum Score Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S6IQSQBW}},
  note         = {Machine review of arXiv:2510.06789}
}

read the original abstract

This paper investigates the rank aggregation problem through the lens of multi-way comparison data derived from rater scores. Departing from traditional parametric frameworks, such as the Bradley-Terry and Plackett-Luce models, we propose a model-free method that accommodates highly heterogeneous preference distributions across raters and encompasses weak stochastic transitivity in pairwise comparisons as a special case. We establish the theoretical foundations of the proposed estimator by proving its consistency, demonstrating that the proportion of discordant pairs (Kendall tau) converges to zero in probability as the number of raters diverges. Furthermore, we derive upper and lower bounds for a performance metric based on Kendall's tau. In certain asymptotic regimes, these bounds coincide up to logarithmic factors, so the estimator is nearly minimax optimal. These results are obtained by analyzing the convergence behavior of a U-empirical process; the novel technical results developed for this analysis may be of independent theoretical interest. The practical utility of our method is validated through extensive simulations and applications to sports player rankings and survey preference aggregation.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 1 linked inside Pith

[1]

Rank analysis of i ncomplete block designs: I

Ralph Allan Bradley and Milton E Terry. Rank analysis of i ncomplete block designs: I. the method of paired comparisons. Biometrika, 39:324–345, 1952

1952
[2]

Estimation in tournaments and graphs under monotonicity constraints

Sabyasachi Chatterjee and Sumit Mukherjee. Estimation in tournaments and graphs under monotonicity constraints. IEEE Transactions on Information Theory , 65:3525–3539, 2019

2019
[3]

Matrix estimation by universal sing ular value thresholding

Sourav Chatterjee. Matrix estimation by universal sing ular value thresholding. The Annals of Statistics, 43:177–214, 2015

2015
[4]

Asymptotically opt imal sequential design for rank aggregation

Xi Chen, Y unxiao Chen, and Xiaoou Li. Asymptotically opt imal sequential design for rank aggregation. Mathematics of Operations Research, 47(3):2310–2332, 2022

2022
[5]

Spe ctral method and regularized mle are both optimal for top-k ranking

Y uxin Chen, Jianqing Fan, Cong Ma, and Kaizheng Wang. Spe ctral method and regularized mle are both optimal for top-k ranking. Annals of Statistics , 47:2204–2235, 2019

2019
[6]

Uncertainty q uantiﬁcation in the Bradley– Terry–Luce model

Chao Gao, Y andi Shen, and Anderson Y Zhang. Uncertainty q uantiﬁcation in the Bradley– Terry–Luce model. Information and Inference: A Journal of the IMA , 12(2):1073–1140, 2023. 19

2023
[7]

Asymptotic theory of sparse bradley– terry model

Ruijian Han, Rougang Y e, Chunxi Tan, and Kani Chen. Asymptotic theory of sparse bradley– terry model. The Annals of Applied Probability , 30(5):2491–2515, 2020

2020
[8]

A general pairwise comparison model for extremely sparse networks

Ruijian Han, Yiming Xu, and Kani Chen. A general pairwise comparison model for extremely sparse networks. Journal of the American Statistical Association , 118:2422–2432, 2023

2023

Show all 22 references

[9]

A smoothed maximum score estimator for t he binary response model

Joel L Horowitz. A smoothed maximum score estimator for t he binary response model. Econometrica: Journal of the Econometric Society , 60:505–531, 1992

1992
[10]

Pairwise comparisons wit hout stochastic transitivity: Model, theory and applications

Sze Ming Lee and Y unxiao Chen. Pairwise comparisons wit hout stochastic transitivity: Model, theory and applications. arXiv preprint arXiv:2501.07437, 2025

2025 arXiv
[11]

Active ranking without strong stochastic transitivity

Hao Lou, Tao Jin, Y ue Wu, Pan Xu, Quanquan Gu, and Farzad F arnoud. Active ranking without strong stochastic transitivity. Advances in Neural Information Processing Systems , 35:297–309, 2022

2022
[12]

Individual choice behavior, volume 4

R Duncan Luce. Individual choice behavior, volume 4. Wiley New Y ork, 1959

1959
[13]

Maximum score estimation of the stoch astic utility model of choice

Charles F Manski. Maximum score estimation of the stoch astic utility model of choice. Journal of Econometrics, 3:205–228, 1975

1975
[14]

The analysis of permutations

Robin L Plackett. The analysis of permutations. Journal of the Royal Statistical Society Series C: Applied Statistics , 24(2):193–202, 1975

1975
[15]

On sample comp lexity upper and lower bounds for exact ranking from noisy comparisons

Wenbo Ren, Jia Kevin Liu, and Ness Shroff. On sample comp lexity upper and lower bounds for exact ranking from noisy comparisons. Advances in Neural Information Processing Sys- tems, 32:10014–10024, 2019

2019
[16]

Simple, robust and optimal ranking from pairwise comparisons

Nihar B Shah and Martin J Wainwright. Simple, robust and optimal ranking from pairwise comparisons. Journal of Machine Learning Research , 18:1–38, 2018. 20

2018
[17]

Stochastically transitive models for pairwise comparison s: Statistical and computational is- sues

Nihar B Shah, Sivaraman Balakrishnan, Adityanand Gunt uboyina, and Martin J Wainwright. Stochastically transitive models for pairwise comparison s: Statistical and computational is- sues. IEEE Transactions on Information Theory , 63(2):934–959, 2016

2016
[18]

Asymptotics when the nu mber of parameters tends to inﬁnity in the Bradley-Terry model for paired comparisons

Gordon Simons and Yi-Ching Y ao. Asymptotics when the nu mber of parameters tends to inﬁnity in the Bradley-Terry model for paired comparisons. The Annals of Statistics , 27: 1041–1060, 1999

1999
[19]

The method of paired comparisons for social values

Louis L Thurstone. The method of paired comparisons for social values. The Journal of Abnormal and Social Psychology , 21:384 – 400, 1927

1927
[20]

High-dimensional statistics: A non-asymptotic viewpoint , volume 48

Martin J Wainwright. High-dimensional statistics: A non-asymptotic viewpoint , volume 48. Cambridge University Press, 2019

2019
[21]

Rate-optimal rank aggregation with private pairwise rankings

Shirong Xu, Will Wei Sun, and Guang Cheng. Rate-optimal rank aggregation with private pairwise rankings. Journal of the American Statistical Association, 120(550):737–750, 2025

2025
[22]

Systematic error-correcting codes for rank modulation

Hongchao Zhou, Moshe Schwartz, Anxiao Andrew Jiang, an d Jehoshua Bruck. Systematic error-correcting codes for rank modulation. IEEE Transactions on Information Theory , 61 (1):17–32, 2014. 21

2014

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

Rank analysis of i ncomplete block designs: I

Ralph Allan Bradley and Milton E Terry. Rank analysis of i ncomplete block designs: I. the method of paired comparisons. Biometrika, 39:324–345, 1952

1952

[2] [2]

Estimation in tournaments and graphs under monotonicity constraints

Sabyasachi Chatterjee and Sumit Mukherjee. Estimation in tournaments and graphs under monotonicity constraints. IEEE Transactions on Information Theory , 65:3525–3539, 2019

2019

[3] [3]

Matrix estimation by universal sing ular value thresholding

Sourav Chatterjee. Matrix estimation by universal sing ular value thresholding. The Annals of Statistics, 43:177–214, 2015

2015

[4] [4]

Asymptotically opt imal sequential design for rank aggregation

Xi Chen, Y unxiao Chen, and Xiaoou Li. Asymptotically opt imal sequential design for rank aggregation. Mathematics of Operations Research, 47(3):2310–2332, 2022

2022

[5] [5]

Spe ctral method and regularized mle are both optimal for top-k ranking

Y uxin Chen, Jianqing Fan, Cong Ma, and Kaizheng Wang. Spe ctral method and regularized mle are both optimal for top-k ranking. Annals of Statistics , 47:2204–2235, 2019

2019

[6] [6]

Uncertainty q uantiﬁcation in the Bradley– Terry–Luce model

Chao Gao, Y andi Shen, and Anderson Y Zhang. Uncertainty q uantiﬁcation in the Bradley– Terry–Luce model. Information and Inference: A Journal of the IMA , 12(2):1073–1140, 2023. 19

2023

[7] [7]

Asymptotic theory of sparse bradley– terry model

Ruijian Han, Rougang Y e, Chunxi Tan, and Kani Chen. Asymptotic theory of sparse bradley– terry model. The Annals of Applied Probability , 30(5):2491–2515, 2020

2020

[8] [8]

A general pairwise comparison model for extremely sparse networks

Ruijian Han, Yiming Xu, and Kani Chen. A general pairwise comparison model for extremely sparse networks. Journal of the American Statistical Association , 118:2422–2432, 2023

2023