acceptodds
Under review as a conference paper at ICLR 2027

Seemingly Shallow Priors Are Sufficiently Expressive for Asymptotically Bayes Optimal Multiple Testing under Sparsity

Abstract

One-group global-local shrinkage priors, popular alternatives to two-groups spike-and-slab priors, are widely used in Bayesian inference. Regarding the asymptotic optimality of high-dimensional Bayesian multiple testing procedures, the dominant view is that one-group priors should have multiple hyperparameters and heavy tails so that procedures based on these priors are asymptotically Bayes optimal under sparsity (ABOS). Liang et al. (2026) note that virtually all previous procedures based on one-group priors rely solely on the posterior. Instead, by considering both the prior and posterior, Liang et al. (2026) show that relative belief procedures using a simple one-group light-tailed normal prior and a one-group uniform prior, each with a single hyperparameter, are sufficient to be ABOS. All previous ABOS properties of procedures based on one-group priors are established under additive symmetric 0-1 loss. In this paper, we generalize the ABOS properties of relative belief procedures under additive weighted 0-1 loss and broader asymptotic regimes. A rich and flexible scale family of one-group priors, each with a single hyperparameter, can be used in our generalized relative belief procedures, ranging from light-tailed to heavy-tailed priors. The priors can be symmetric or asymmetric, and they can have bounded or unbounded support. Arguably, the ABOS properties of our generalized relative belief procedures are more general and stronger than any existing ABOS properties of procedures based on global-local priors.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Asymptotic optimality, Bayes risk, Bayesian decision theory, Bayesian multiple testing, horseshoe prior, one-group global-local shrinkage priors, two-groups spike-and-slab priors

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Da8FOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies sparse Bayesian multiple testing using generalized relative belief procedures. It extends ABOS results from specific priors to a broad scale family. The family includes light-tailed, heavy-tailed, symmetric, and asymmetric priors. A unified threshold analysis explains why these priors can achieve ABOS under suitable scaling. The paper also establishes empirical Bayes guarantees when sparsity is unknown. The results cover univariate and fixed-dimensional multivariate normal means models under weighted losses.

Critical strengths

A key strength is the unified ABOS theory for a broad prior family. The threshold expansion shows that, under appropriate scaling, prior shape enters only through h(0)h(0), explaining why heavy tails are unnecessary. This elevates earlier examples into a general principle, with uniform control over directions in the multivariate setting further supporting acceptance.

Critical weaknesses

1.The mathematical arguments are clearly organized, but the overall narrative could be more focused. The limited intuitive explanation of the underlying mechanism makes the development harder to follow, and the presentation becomes repetitive across settings. In particular, ABOS is formally defined four times (Definitions 3.3, 3.15, 4.4, and 4.12). A unified definition, with the corresponding oracle benchmark specified for each setting, would make the exposition more concise.

2.The title and introduction emphasize that simple priors can achieve ABOS, but this central observation was already established by Liang et al. (2026) for normal and uniform priors. Although the paper acknowledges this prior work, its framing does not sufficiently distinguish the present contribution. The authors should foreground the generality of the prior family and the new empirical Bayes guarantees, and explain more clearly what these results establish beyond the earlier examples.

Questions for authors

1.Which results require substantially new arguments beyond Liang et al. (2026)? Please highlight the main technical challenges introduced by the general prior family and the multivariate empirical Bayes procedure.

2.Could you explain more intuitively why the prior shape enters the asymptotic threshold only through (h(0)), and how the prescribed scale choices make this approximation valid?

Reviewer_PbxsOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper generalized relative-belief procedures for multiple testing under sparse Gaussian mixture models. Based on previous work (mainly Liang's work), it extended the prior to a broad scale family. First it established an asymptotic rejection-threshold expansion, then derived asymptotic Bayes optimality under sparsity for deterministic and empirical Bayes scales. The paper also considered asymmetric error costs and increasing sample sizes.

Critical strengths

The main contribution of this paper is the general prior-family result together with the multivariate empirical Bayes analysis. The prior-family generalization provides a unified explanation for why a wide range of prior tails and asymmetries lead to the same first-order testing behavior. The whole paper is technically sound.

Critical weaknesses

The core message of the draft is that light-tailed one-group priors can yield asymptotic Bayes optimality under sparsity testing procedures when used through relative belief. Since Liang et al. already established this conceptual point for a normal prior, the author should explain more precisely what becomes possible beyond the earlier normal-prior procedure. The introduction part should distinguish clearly between the inherited conceptual insight, the new uniform prior-family theorem, and the multivariate empirical Bayes result.

Questions for authors

This paper is very theoretical. It would be great if the author can provide/describe some potential applications to help understanding.

Reviewer_tkehOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies Bayesian multiple testing of sparse normal means, H0i:μi=0H_{0i}: \mu_i = 0 against HAi:μi≠0H_{Ai}: \mu_i \neq 0 for i=1,…,mi = 1, \dots, m, using the ABOS (asymptotically Bayes optimal under sparsity) criterion of Bogdan et al. (2011). It builds on Liang et al. (2026), who showed that relative belief procedures with a single-parameter normal or uniform prior are ABOS.

The authors extend this to "generalized relative belief" (GRB) procedures. These use any prior of the scale form πτ(μ)=τ−1/2h(μ/τ)\pi_\tau(\mu) = \tau^{-1/2} h(\mu/\sqrt{\tau}), where hh is bounded, continuous at zero and 0<h(0)<∞0 < h(0) < \infty, and they reject H0iH_{0i} when RB(μi=0∣Xi,τ)<1/δmRB(\mu_i = 0 \mid X_i, \tau) < 1/\delta_m. Here δm\delta_m is the loss ratio of an additive weighted 0–1 loss.

The main results are:

  • Univariate case: the procedure is ABOS when τm≍pm−2log⁡(1/pm)\tau_m \asymp p_m^{-2}\log(1/p_m) (Theorem 3.10), and an empirical Bayes plug-in for τm\tau_m is also ABOS (Theorem 3.12).
  • Increasing sample size: in the framework of Frommlet & Bogdan (2013), the same holds with τm≍pm−2\tau_m \asymp p_m^{-2} (Theorems 3.16–3.17).
  • Multivariate case: analogous results for kk-dimensional means with fixed k≥2k \ge 2 and known Σ\Sigma (Theorems 4.7–4.15).

The work is purely theoretical, and the proofs are in a 57-page appendix.

Critical strengths

S1. The results are general and stated precisely. One rule covers a whole family of priors: light or heavy tailed, symmetric or asymmetric, with bounded or unbounded support. It also covers a weighted 0–1 loss, three asymptotic settings (univariate, increasing sample size, multivariate), and fully data-driven empirical Bayes versions. The assumptions are written out clearly for each theorem. To my knowledge, the result in the Frommlet & Bogdan (2013) setting has not been shown before for procedures based on one-group priors.

S2. The procedure is simple and the proofs are fairly transparent. The rejection rule has an explicit form (Proposition 3.6), and the proof strategy is easy to follow at a high level. The threshold is shown to be unique for each sign of XiX_i (Proposition 3.7), its asymptotic order is derived (Proposition 3.9), and it is compared with fixed-threshold rules already known to be ABOS. Compared with the long arguments needed for global–local priors, this is a clean route and may be useful to others in the area.

Critical weaknesses

W1. I find the contribution beyond Liang et al. (2026) limited, and some claims overstated. By Proposition 3.9, the threshold satisfies cτm2=log⁡τm+2log⁡δm−2log⁡{(2π)1/2h(0)}+o(1)c^2_{\tau_m} = \log\tau_m + 2\log\delta_m - 2\log\{(2\pi)^{1/2}h(0)\} + o(1), so asymptotically the prior enters only through the constant h(0)h(0). Once τm\tau_m is chosen in the right order, the GRB rule behaves like a fixed-threshold rule, and ABOS follows from Theorem 3.4 of Bogdan et al. (2011). The large prior family therefore comes almost for free, and the tails, asymmetry and support of the prior play no role in the result. In my view, this is a natural extension of Liang et al. (2026), from two specific priors to the general scale family and to two more asymptotic settings.

Also, τm\tau_m must grow like pm−2log⁡(1/pm)p_m^{-2}\log(1/p_m), which is much larger than the variance of the true nonzero effects. So the prior acts as a calibration device for the threshold, not as a description of prior belief. This makes the claims that GRB "better reflects the Bayesian spirit" and is "more general and stronger than any existing ABOS properties" hard to accept as written. The claim of being "the first to attain the general ABOS property" also needs a more careful statement.

W2. There is no numerical evidence, and the assumptions are strong. The paper says "This work is theoretical" and contains no simulations, relying on Liang et al. (2026) for numerical results. Since ABOS is an asymptotic property, some evidence on finite-sample behaviour would help a lot, for example:

  • how the choice of hh, which matters only through h(0)h(0) asymptotically, affects the risk for moderate mm;
  • how sensitive the empirical Bayes versions are to the constants d1,d2d_1, d_2 (and the analogous constants elsewhere);
  • a comparison with the horseshoe and with BH.

The theory also assumes a known noise variance or a known Σ\Sigma, a Gaussian slab in the true two-groups model, polynomial sparsity pm∝m−βp_m \propto m^{-\beta} for the empirical Bayes results, and a fixed dimension kk. Finally, the topic is quite specialized for ICLR, and the paper does not explain why it matters for machine learning practice.

Questions for authors

Q1. This first question is the most important one for my assessment. Since the threshold depends on the prior only through h(0)h(0) (Proposition 3.9), what is technically new in the proofs compared with Liang et al. (2026), beyond replacing the normal or uniform density by a general hh? Which step would fail without the conditions of Definition 3.5?

Q2. Could you add a simulation study? It should compare several choices of hh (normal, Cauchy, uniform, an asymmetric density) at moderate mm, and include the horseshoe and BH as baselines. Does the choice of hh matter in practice, even though it does not matter asymptotically?

Q3. The prior scale τm\tau_m has to grow like pm−2log⁡(1/pm)p_m^{-2}\log(1/p_m), far beyond the scale of the true signals. How do you reconcile this with the claim that GRB procedures better reflect the Bayesian spirit and allow prior elicitation?

Q4. How sensitive are the empirical Bayes procedures to the constants d1,d2d_1, d_2 (and d3,…,d8d_3, \dots, d_8)? Is there a practical way to choose them?

Q5. Can the results be extended to an unknown noise variance or Σ\Sigma, to a non-Gaussian slab in the true model, or to kk growing with mm?

Q6. What is the practical value of ABOS under the weighted loss for ML applications? Could you give a concrete example where this kind of test would be used?

Minor:

  • The AI use statement contradicts itself. It first says generative AI was used for "none of the required disclosure tasks", then says it was used "for aiding and polishing writing".
  • There is no conclusion or limitations section. The discussion only lists future directions.
  • The main text has a lot of repeated theorem statements, which could be shortened to make room for intuition or examples.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.