pith. sign in

arxiv: 2607.02340 · v1 · pith:3AXIMZPBnew · submitted 2026-07-02 · 📊 stat.ME · math.ST· stat.OT· stat.TH

Merging of Bayes and quasi-Bayes empirical Bayes procedures for Poisson compound decisions

Stefano Favaro , Sandra Fortini This is my paper

Pith reviewed 2026-07-03 07:24 UTC · model grok-4.3

classification 📊 stat.ME math.STstat.OTstat.TH
keywords empirical BayesPoisson compound decisionDirichlet processNewton's algorithmfrequentist mergingmixing distributionregret analysismarginal PMF concentration
0
0 comments X

The pith

Bayesian and quasi-Bayesian empirical Bayes procedures merge in frequentist sense for Poisson compound decisions.

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

The paper studies the Poisson compound decision problem by comparing two ways to estimate an unknown mixing distribution: a Bayesian method that uses the Dirichlet process posterior and a quasi-Bayesian method that uses Newton's algorithm. Under a Poisson mixture model that has a fixed true mixing distribution, both methods produce marginal probability mass functions that concentrate at certain rates. These concentration rates are shown to produce matching rates of decay in the regrets of the two procedures. The matching decay rates establish a frequentist merging result, so that the two strategies become asymptotically equivalent in performance. The same merging is shown to hold when the problem is extended to multiple dimensions.

Core claim

Under a Poisson mixture model with a true or oracle mixing distribution, concentration rates for the marginal probability mass functions induced by the Bayesian and quasi-Bayesian estimates translate into rates of decay for the corresponding regrets and prove a frequentist merging result between the Bayesian and quasi-Bayesian empirical Bayes strategies. The analysis is extended to the multidimensional Poisson compound decision problem.

What carries the argument

Concentration rates of the marginal probability mass functions induced by the Dirichlet process posterior and by Newton's algorithm estimates, translated into rates of decay for the corresponding regrets.

If this is right

  • The quasi-Bayesian strategy achieves accuracy comparable to the Bayesian strategy while using substantially fewer computational resources.
  • The frequentist merging result extends directly to the multidimensional Poisson compound decision problem.
  • Regret decay rates are determined by the concentration behavior of the induced marginal probability mass functions.
  • Numerical experiments on synthetic data confirm that the quasi-Bayesian strategy matches the Bayesian one in practice.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the merging holds, the faster quasi-Bayesian method can replace the Dirichlet process approach for large-scale Poisson compound decisions without asymptotic performance loss.
  • The same concentration-to-regret translation may apply to other mixture models once comparable concentration rates are available.
  • The result supplies a theoretical justification for preferring Newton's algorithm in settings where computational cost grows with dimension.

Load-bearing premise

The data are generated from a Poisson mixture model that possesses a fixed true or oracle mixing distribution.

What would settle it

Repeated simulations drawn from a known fixed mixing distribution in which the difference between the two regrets fails to decay at the established rate would falsify the merging result.

Figures

Figures reproduced from arXiv: 2607.02340 by Sandra Fortini, Stefano Favaro.

Figure 1
Figure 1. Figure 1: Weibull prior, n ∈ {50, 100, 200, 400}: data points plotted against the “true” parameters (grey), together with the corresponding oracle Bayes (black), Bayes (red), and quasi-Bayes (blue) estimates [PITH_FULL_IMAGE:figures/full_fig_p018_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Weibull prior, n ∈ {1, 000, 2, 000, 4, 000, 8, 000}: data points plotted against the “true” pa￾rameters (grey), together with the corresponding oracle Bayes (black), Bayes (red), and quasi-Bayes (blue) estimates. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Weibull prior: quasi-Bayes (blue) and Bayes (red) estimates compared by E-regret (top panels), computational units (middle panels), and CPU time (bottom panels). 19 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Weibull prior: E-regret incurred by using the quasi-Bayes estimate in place of the Bayes estimate. is defined as E-mse(Gˆ n) = 1 2n Xn i=1 X 2 ℓ=1 n ˆθn,ℓ(yi) − θi,ℓo2 . For the oracle Bayes estimate θˆ∗ , the E-mse is referred to as the empirical minimum mean squared error (E-mmse), i.e. E-mmse = 1 2n Xn i=1 X 2 ℓ=1 n ˆθ ∗ ℓ (yi) − θi,ℓo2 . Finally, we define the empirical regret (E-regret) as E-regret(Gˆ… view at source ↗
Figure 5
Figure 5. Figure 5: Weibull product prior, n ∈ {50, 100, 200, 400}: data points plotted against the “true” parameters (grey), together with the corresponding oracle Bayes (black), Bayes (red), and quasi-Bayes (blue) estimates. multivariate likelihood under the auxiliary clusters drawn from H. Conditionally on the allocations, the occupied atoms are updated coordinate-wise from their Gamma full conditional distributions. We ru… view at source ↗
Figure 6
Figure 6. Figure 6: Weibull product prior, n ∈ {1, 000, 2, 000, 4, 000, 8, 000}: data points plotted against the “true” parameters (grey), together with the corresponding oracle Bayes (black), Bayes (red), and quasi-Bayes (blue) estimates. in Martin and Tokdar (2009) and Ignatiadis and Kankanala (2026). These results suggest that Gaussian analogue of our merging results could be obtained by means of an appropriate stability a… view at source ↗
Figure 7
Figure 7. Figure 7: Weibull product prior: quasi-Bayes (blue) and Bayes (red) estimates compared by E-regret (top panels), computational units (middle panels), and CPU time (bottom panels). 23 [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Weibull product prior: E-regret incurred by using the quasi-Bayes estimate in place of the Bayes estimate. Finally, an important extension concerns the estimation of sums of random variables and re￾lated functionals (Zhang, 2005) The present paper focuses on estimating the individual Poisson means through plug-in empirical Bayes rules. In many applications, however, the target is not the collection of indi… view at source ↗
read the original abstract

The Poisson compound decision problem is a long-standing problem in statistics, in which empirical Bayes methods are used to estimate Poisson means under a mixture model. We study this problem from the viewpoint of $g$-modeling, comparing two nonparametric strategies for estimating the unknown mixing distribution: a Bayesian empirical Bayes strategy, based on the Dirichlet process posterior, and a quasi-Bayesian empirical Bayes strategy, based on Newton's algorithm. The latter is computationally attractive, but its relationship with the Bayesian strategy requires theoretical justification. Under a Poisson mixture model with a ``true'', or oracle, mixing distribution, we establish concentration rates for the marginal probability mass functions induced by the Bayesian and quasi-Bayesian estimates. These rates are then translated into rates of decay for the corresponding regrets, interpreted as excess Bayes risks, and used to prove a frequentist merging result between the Bayesian and quasi-Bayesian empirical Bayes strategies. We also extend the analysis to the multidimensional Poisson compound decision problem. Numerical experiments on synthetic data illustrate that the quasi-Bayesian strategy achieves accuracy comparable to the Bayesian strategy, while requiring substantially fewer computational resources, especially in the multidimensional setting.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper studies the Poisson compound decision problem under g-modeling, comparing a Bayesian empirical Bayes estimator based on the Dirichlet process posterior with a quasi-Bayesian estimator based on Newton's algorithm. Under a Poisson mixture model with a fixed true mixing distribution, it derives concentration rates for the induced marginal PMFs, translates these into rates for the corresponding regrets (excess Bayes risk), and proves a frequentist merging result between the two procedures. The analysis is extended to the multidimensional Poisson compound decision problem, with numerical experiments on synthetic data showing comparable accuracy for the quasi-Bayesian method at lower computational cost.

Significance. If the concentration and merging results hold, the work supplies a rigorous frequentist justification for preferring the computationally lighter quasi-Bayesian procedure in empirical Bayes applications to Poisson data. The multidimensional extension and the explicit translation from PMF concentration to regret decay are valuable contributions. The manuscript includes reproducible numerical experiments that directly illustrate the practical performance comparison.

major comments (3)
  1. [§3.2] §3.2 (concentration rates for the quasi-Bayesian estimator): the claimed rate for the marginal PMF induced by Newton's algorithm is stated without an explicit derivation of the n^{-1/4} term or the dependence on the smoothness of the true mixing distribution; this step is load-bearing for the subsequent regret and merging claims.
  2. [§4] §4 (translation to regret decay): the argument invokes standard risk bounds for Bayes rules under Poisson mixtures, but does not verify that the identifiability constants remain uniform over the class of mixing distributions considered; this uniformity is required for the merging rate to hold uniformly.
  3. [§5.1] §5.1 (multidimensional extension): the extension assumes that the concentration rates carry over without additional logarithmic factors in dimension d, yet no explicit dependence on d is derived or bounded; this affects the practical scope of the merging result.
minor comments (3)
  1. [Table 1] Table 1: the reported CPU times lack standard errors or number of replications, making it difficult to assess variability of the computational advantage.
  2. [§2] Notation: the symbol G_n for the quasi-Bayesian estimate is introduced without an explicit algorithmic definition in the main text (only referenced to Newton 2002); a short pseudocode box would improve readability.
  3. [References] References: the citation list omits recent work on concentration of Dirichlet process mixtures for discrete data (e.g., papers post-2018 on posterior contraction for Poisson mixtures).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments, which help clarify the presentation of our concentration and merging results. We address each major comment below and will incorporate the suggested clarifications and derivations in a revised manuscript.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (concentration rates for the quasi-Bayesian estimator): the claimed rate for the marginal PMF induced by Newton's algorithm is stated without an explicit derivation of the n^{-1/4} term or the dependence on the smoothness of the true mixing distribution; this step is load-bearing for the subsequent regret and merging claims.

    Authors: We agree that an explicit derivation strengthens the argument. In the revision we will add a self-contained lemma deriving the n^{-1/4} rate for the marginal PMF under Newton's algorithm, making the dependence on the Hölder smoothness index of the true mixing distribution fully explicit and citing the relevant empirical-process bounds used. revision: yes

  2. Referee: [§4] §4 (translation to regret decay): the argument invokes standard risk bounds for Bayes rules under Poisson mixtures, but does not verify that the identifiability constants remain uniform over the class of mixing distributions considered; this uniformity is required for the merging rate to hold uniformly.

    Authors: We will add a short paragraph in Section 4 verifying uniformity of the identifiability constants over the compact class of mixing distributions used in the paper (or over the slightly restricted subclass needed for uniformity). This will be stated as a standing assumption with a brief justification based on the continuity of the Poisson mixture map. revision: yes

  3. Referee: [§5.1] §5.1 (multidimensional extension): the extension assumes that the concentration rates carry over without additional logarithmic factors in dimension d, yet no explicit dependence on d is derived or bounded; this affects the practical scope of the merging result.

    Authors: We acknowledge the omission. The revision will include an explicit bound on the d-dependence (including possible logarithmic factors) for the multidimensional concentration rates, together with a brief discussion of the resulting restriction on the merging rate when d grows with n. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper derives concentration rates for marginal PMFs induced by Dirichlet process and Newton's algorithm estimators under an assumed Poisson mixture model with fixed oracle mixing distribution, then translates those rates into regret decay and a frequentist merging result. This follows standard frequentist asymptotic arguments relying on identifiability of Poisson mixtures and risk bounds for Bayes rules; no parameters are fitted to the same data used for evaluation, no self-definitional loops appear in the equations, and no load-bearing self-citations or ansatzes are invoked. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption of an oracle mixing distribution; no free parameters or invented entities are indicated in the abstract.

axioms (1)
  • domain assumption Existence of a true or oracle mixing distribution under the Poisson mixture model
    The concentration rates and merging result are derived conditional on this fixed mixing distribution.

pith-pipeline@v0.9.1-grok · 5733 in / 1200 out tokens · 33495 ms · 2026-07-03T07:24:39.524216+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    and Jordan, M.I

    Blei, D.M. and Jordan, M.I. (2006). Variational inference for Dirichlet process mixtures. Bayesian Anal. 1, 121--144

    2006

  2. [2]

    and Farrell, R.H

    Brown, L.D. and Farrell, R.H. (1985). Complete class theorems for estimation of multivariate Poisson means and related problems. Ann. Statist. 13, 706--726

    1985

  3. [3]

    and Ritov, Y

    Brown, L.D., Greenshtein, E. and Ritov, Y. (2013). The Poisson compound decision problem revisited. J. Am. Statist. Assoc. 108, 741--749

    2013

  4. [4]

    Universal priors: solving empirical Bayes via Bayesian inference and pretraining

    Cannella, N., Teh, A., Han Y. and Polyanskiy, Y. (2026). Universal priors: solving empirical Bayes via Bayesian inference and pretraining. Preprint arXiv:2602.15136

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  5. [5]

    and Lindley, D.V

    Deely, J.J. and Lindley, D.V. (1981). Bayes empirical Bayes. J. Am. Statist. Assoc. 76, 833--841

    1981

  6. [6]

    Efron, B. (2014). Two modeling strategies for empirical Bayes estimation. Statist. Sci. 29, 285--301

    2014

  7. [7]

    Efron, B. (2019). Bayes, oracle Bayes and empirical Bayes. Statist. Sci. 34, 177--201

    2019

  8. [8]

    and Fortini, S

    Favaro, S. and Fortini, S. (2026). Quasi-Bayes empirical Bayes: a sequential approach to the Poisson compound decision problem. Biometrika, to appear

    2026

  9. [9]

    Ferguson, T.S. (1973). A Bayesian analysis of some nonparametric problems. Ann. Statist. 1, 209--230

    1973

  10. [10]

    and Petrone, S

    Fortini, S. and Petrone, S. (2020). Quasi-Bayesian properties of a procedure for sequential learning in mixture models. J. R. Statist. Soc. B 82, 1087--1114

    2020

  11. [11]

    and van der Vaart, A.W

    Ghosal, S. and van der Vaart, A.W. (2001). Entropies and rates of convergence for maximum likelihood and Bayes estimation for mixtures of normal densities. Ann. Statist. 29, 1233--1263

    2001

  12. [12]

    and Kankanala, S

    Ignatiadis, N. and Kankanala, S. (2026). Compound decisions and empirical Bayes via Bayesian nonparametrics. Preprint arXiv:2602.20115

    work page arXiv 2026

  13. [13]

    and James, L.F

    Ishwaran, H. and James, L.F. (2001). Gibbs sampling methods for stick-breaking priors. J. Amer. Statist. Assoc. 96, 161--173

    2001

  14. [14]

    and Wu, Y

    Jana, S., Polyanskiy, Y., Teh, A. and Wu, Y. (2023). Empirical Bayes via ERM and Rademacher complexities: the Poisson model. P. Mach. Learn. Res. 195, 1--37

    2023

  15. [15]

    and Wu, Y

    Jana, S., Polyanskiy, Y. and Wu, Y. (2025). Optimal empirical Bayes estimation for the Poisson model via minimum-distance methods. Inf. Inference 14, 1--42

    2025

  16. [16]

    Johnstone, I. (1986). Admissible estimation, Dirichlet principles and recurrence of birth-death chains on Z _ + ^ p . Probab. Theory Related Fields 71, 231--269

    1986

  17. [17]

    Lo, A.Y. (1984). On a class of Bayesian nonparametric estimates. I. Density estimates Ann. Statist. 12, 351--357

    1984

  18. [18]

    and Ghosh, J.K

    Martin, R. and Ghosh, J.K. (2008). Stochastic approximation and Newton’s estimate of a mixing distribution. Statist. Sci. 23, 365--382

    2008

  19. [19]

    and Tokdar, S.T

    Martin, R. and Tokdar, S.T. (2009) Asymptotic properties of predictive recursion: robustness and rate of convergence. Electron. J. Stat. 3, 1455--1472

    2009

  20. [20]

    Neal, R. M. (2000). Markov chain sampling methods for Dirichlet process mixture models. J. Comput. Graph. Statist. 9, 249--265

    2000

  21. [21]

    and Zhang, Y

    Newton, M.A., Quintana, F.A. and Zhang, Y. (1998). Nonparametric Bayes methods using predictive updating. In Practical Nonparametric and Semiparametric Bayesian Statistics, Springer

    1998

  22. [22]

    and Roberts, G.O

    Papaspiliopoulos, O. and Roberts, G.O. (2008). Retrospective Markov chain Monte Carlo methods for Dirichlet process hierarchical models. Biometrika 95, 169--186

    2008

  23. [23]

    Sharp regret bounds for empirical

    Polyanskiy, Y. and Wu, Y (2021). Sharp regret bounds for empirical Bayes and compound decision problems. Preprint arXiv: 2109.03943

    work page arXiv 2021

  24. [24]

    Robbins, H. (1951). Asymptotically subminimax solutions of compound decision problems. In Proc. Second Berkeley Symp 2, 131--148

    1951

  25. [25]

    Robbins, H. (1956). An empirical Bayes approach to statistics. In Proc. Third Berkeley Symp. 3, 157--164

    1956

  26. [26]

    and Wu, Y

    Shen, Y. and Wu, Y. (2026). Poisson empirical Bayes estimation: When does g -modeling beat f -modeling in theory (and in practice)? Ann. Statist. 54, 146--175

    2026

  27. [27]

    and Makov, U.E

    Smith, A.F.M. and Makov, U.E. (1978). A quasi-Bayes sequential procedure for mixtures. J. R. Statist. Soc. B 40, 106--112

    1978

  28. [28]

    and Polyanskiy, Y

    Teh, A., Jabbour, M. and Polyanskiy, Y. (2025). Solving empirical Bayes via transformers. Preprint arXiv:2502.09844

    work page arXiv 2025

  29. [29]

    Walker, S.G. (2007). Sampling the Dirichlet mixture model with slices. Comm. Statist. Simulation Comput. 36, 45--54

    2007

  30. [30]

    Zhang, C.-H. (2003). Compound decision theory and empirical Bayes methods. Ann. Statist. 31, 379--390

    2003

  31. [31]

    Zhang, C.-H. (2005). Estimation of sums of random variables: examples and information bounds. Ann. Statist. 33, 2022--2041

    2005

  32. [32]

    and Shen, X

    Wong, H.W. and Shen, X. (1995). Probability inequalities for likelihood ratios and convergence rates of sieve MLES Ann. Statist. 23, 339--362

    1995