Nonparametric Empirical Bayes Confidence Intervals

• NP-EBCIs are a family of methods that construct confidence intervals by nonparametrically estimating unknown priors or mixing distributions for various inferential targets.
• Feasible constructions replace unknown posterior quantiles with deconvolution-kernel estimates or smooth NPMLE approaches, achieving near-oracle performance under different settings.
• Applications span normal and Poisson mean models, sparse signal inference, and empirical Bayes rebiasing, often reducing interval lengths while maintaining targeted coverage.

Searching arXiv for recent and foundational papers on nonparametric empirical Bayes confidence intervals and closely related methods. Nonparametric empirical Bayes confidence intervals (NP-EBCIs) are confidence intervals or, more generally, confidence sets obtained by estimating an unknown prior or mixing distribution nonparametrically and then using the resulting empirical Bayes posterior to perform uncertainty quantification. In the normal means formulation, the oracle intervals are posterior-quantile intervals under a point-identified, fully nonparametric prior, while feasible intervals replace those quantiles with nonparametric estimates (Xie, 8 May 2026). Closely related strands include optimal selection-adjusted confidence sets for sparse signals, bias-aware intervals for empirical Bayes functionals, and smooth-NPMLE marginal coverage sets in Gaussian and Poisson mixture models (Woody et al., 2018, Ignatiadis et al., 2019, Kim et al., 29 Mar 2026, Kim, 5 May 2026).

1. Conceptual scope

A common abstract formulation is the one-parameter empirical Bayes model

$\mu \sim G,\qquad Z \mid \mu \sim p(z\mid \mu),$

where $G$ is unknown and is estimated from repeated observations $Z_1,\dots,Z_n$. In this literature, the inferential target is not unique. Some papers study intervals for latent unit-specific effects, such as $\theta_i$ in a normal means model; others study intervals for empirical Bayes functionals such as the posterior mean $\theta_G(z)=E_G[h(\mu)\mid Z=z]$, posterior quantiles, or the local false sign rate (Ignatiadis et al., 2019).

The phrase NP-EBCI therefore names a family of procedures rather than a single algorithm. The literature also distinguishes between across-group inference and group-specific inference: standard empirical Bayes interval procedures often target across-group average coverage, whereas other constructions pursue group-specific or conditional coverage (Hoff, 2022).

Paper	Primary setting	Main interval target
(Xie, 8 May 2026)	Normal means	Unobservable individual effects
(Woody et al., 2018)	Sparse signals after selection	Selection-adjusted confidence sets
(Ignatiadis et al., 2019)	General empirical Bayes model	Posterior mean or local false sign rate
(Kim et al., 29 Mar 2026)	Normal means with smooth NPMLE	Marginal coverage sets
(Kim, 5 May 2026)	Poisson means with Gamma smoothing	Plug-in empirical Bayes confidence sets
(Armstrong et al., 2020)	Normal means with moment restrictions	Robust average-coverage EBCIs

This variety matters because validity claims depend on the target. An interval calibrated for the posterior quantile $q_G(\tau;z)$ is conceptually different from an interval calibrated for a latent $\theta_i$, and both differ from intervals designed only to achieve average coverage under moment restrictions.

2. Oracle constructions

In the normal means model studied by Xie,

$$Y_i\mid \theta_i \sim N(\theta_i,\sigma_i^2),\qquad \theta_i\sim G,$$

with known $\sigma_i^2$ and unknown mixing distribution $G$, the oracle posterior of $G$0 induces posterior quantiles

$G$1

The equal-tailed oracle $G$2 interval is

$G$3

By construction, this interval has conditional coverage

$G$4

for each $G$5, and hence also marginal coverage $G$6 (Xie, 8 May 2026).

Related oracle constructions appear in other models. In the Poisson means problem, the oracle $G$7 marginal-coverage set is a posterior level set

$G$8

where the single threshold $G$9 is chosen so that $Z_1,\dots,Z_n$0 (Kim, 5 May 2026). In the sparse-signal post-selection setting, the oracle object is a selection-adjusted confidence set that is as short as possible on average while adjusting for adaptive selection and maintaining exact frequentist coverage uniformly over the parameter space (Woody et al., 2018).

These oracle constructions establish the benchmark that empirical Bayes procedures try to emulate. The core question is not whether a Bayesian interval exists, but whether a data-driven nonparametric substitute can approximate the oracle interval without losing the intended coverage property.

3. Feasible nonparametric constructions

The feasible NP-EBCI in the normal means model replaces the unknown posterior quantiles by nonparametric estimates. Xie’s construction estimates the marginal characteristic function by

$Z_1,\dots,Z_n$1

and then forms a deconvolution-kernel estimator of the prior characteristic function,

$Z_1,\dots,Z_n$2

where $Z_1,\dots,Z_n$3 is a compactly supported flat-top spectral kernel and $Z_1,\dots,Z_n$4 is a bandwidth. The posterior quantile is then estimated directly as

$Z_1,\dots,Z_n$5

with

$Z_1,\dots,Z_n$6

The feasible interval is

$Z_1,\dots,Z_n$7

(Xie, 8 May 2026).

That paper also gives a heteroskedastic extension. If $Z_1,\dots,Z_n$8 and $Z_1,\dots,Z_n$9 are allowed to be dependent through $\theta_i$0, one re-normalizes

$\theta_i$1

carries out deconvolution for $\theta_i$2, estimates posterior quantiles for $\theta_i$3, and then back-transforms (Xie, 8 May 2026).

A different line of work replaces the direct deconvolution step by smooth nonparametric maximum likelihood. In the Gaussian case, a hierarchical Gaussian smoothing layer restricts the mixing distribution to a Gaussian location mixture; the resulting smooth NPMLE is computed by solving a convex optimization problem, yields smooth posteriors, and supports plug-in empirical Bayes marginal coverage sets (Kim et al., 29 Mar 2026). In the Poisson case, the analogous construction models the prior as a Gamma mixture, preserves the convex optimization structure of the classical NPMLE, and forms plug-in empirical Bayes confidence sets from the estimated smooth posterior (Kim, 5 May 2026).

Another complementary construction is the F-localization framework for empirical Bayes functionals. There one first constructs a confidence region for the marginal distribution $\theta_i$4, then projects that region to the target functional or forms AMARI bias-aware intervals after linearization. This approach is designed to remain valid even when the empirical Bayes estimand is partially identified or converges slowly (Ignatiadis et al., 2019).

4. Coverage notions and inferential targets

Coverage language in this literature is technically specific. In Xie’s normal means formulation, “conditional coverage” refers to

$\theta_i$5

whereas “marginal coverage” refers to $\theta_i$6 (Xie, 8 May 2026). In the discussion literature, “across-group average coverage” denotes validity after averaging over the empirical Bayes population, while “group-specific coverage” requires validity for each fixed value of the underlying parameter (Hoff, 2022).

This distinction is not cosmetic. Standard empirical Bayes interval procedures focus on controlling the across-group average coverage rate, but if group-specific inferences are primary, confidence intervals with group-specific coverage control may be preferable (Hoff, 2022). The FAB construction discussed there inverts size-$\theta_i$7 tests to obtain intervals with exact group-specific coverage while still using a working prior to reduce width.

Robust empirical Bayes confidence intervals provide another benchmark. In the normal means model, these intervals are centered at the usual linear empirical Bayes estimator, use a critical value that accounts for shrinkage, and guarantee coverage regardless of the means distribution; when the means are treated as fixed, the coverage probability is at least $\theta_i$8 on average across the $\theta_i$9 EBCIs (Armstrong et al., 2020). Their guarantee is therefore average rather than conditional-on-$\theta_G(z)=E_G[h(\mu)\mid Z=z]$0 posterior-quantile exactness.

A common source of confusion is to treat all of these guarantees as interchangeable. They are not. An interval can have exact posterior conditional coverage under an oracle prior, asymptotically exact marginal coverage after nonparametric estimation, or only average coverage under moment restrictions; the papers in this area are explicit that these are different inferential criteria.

5. Ill-posedness, optimality, and oracle efficiency

A central result of the fully nonparametric normal means theory is that posterior quantiles inherit the severe ill-posedness of nonparametric deconvolution. In Xie’s analysis, the minimax optimal estimation rate for posterior quantiles is logarithmic, and this logarithmic rate is also minimax optimal for errors in the conditional coverage probability; the resulting errors in the marginal coverage probability vanish at the same logarithmic rate (Xie, 8 May 2026).

Under homoskedasticity $\theta_G(z)=E_G[h(\mu)\mid Z=z]$1 and a Sobolev class

$\theta_G(z)=E_G[h(\mu)\mid Z=z]$2

the feasible NP-EBCI satisfies

$\theta_G(z)=E_G[h(\mu)\mid Z=z]$3

and, under mild continuity conditions,

$\theta_G(z)=E_G[h(\mu)\mid Z=z]$4

The same paper proves a lower bound showing that no estimator $\theta_G(z)=E_G[h(\mu)\mid Z=z]$5 can achieve squared error $\theta_G(z)=E_G[h(\mu)\mid Z=z]$6, while its Fourier-kernel estimator attains the matching logarithmic upper rate in absolute error (Xie, 8 May 2026).

The smooth-NPMLE literature modifies the model class precisely to alter this rate picture. In the Gaussian case, the classical NPMLE is discrete, which yields discrete posterior credible sets and a logarithmic deconvolution rate; by restricting the prior to a Gaussian location mixture, the smooth NPMLE attains a polynomial convergence rate for deconvolution and posterior estimation, and the plug-in empirical Bayes marginal coverage sets achieve asymptotically exact coverage at a polynomial rate while converging to the oracle optimal set in expected length (Kim et al., 29 Mar 2026). The Poisson analogue establishes polynomial prior and posterior density convergence under compact support and shows that the plug-in sets mimic the oracle optimal marginal coverage sets and converge to the oracle minimal expected length (Kim, 5 May 2026).

In sparse post-selection inference, optimality takes a different form. The nonparametric empirical-Bayes procedure of (Woody et al., 2018) constructs confidence sets that are as short as possible on average while both adjusting for selection and maintaining exact frequentist coverage uniformly over the parameter space; under mild conditions, the procedure asymptotically converges to the results of an oracle-Bayes analysis in which the prior distribution of signal sizes is known exactly. This suggests that “optimality” in NP-EBCIs is typically oracle-relative and tied to the particular coverage notion being enforced.

6. Empirical behavior and applications

Simulation evidence in the fully nonparametric normal means paper uses sample sizes $\theta_G(z)=E_G[h(\mu)\mid Z=z]$7, signal-to-noise ratios $\theta_G(z)=E_G[h(\mu)\mid Z=z]$8, and eight unit-variance mixing distributions, including Gaussian, Laplace, Student-$\theta_G(z)=E_G[h(\mu)\mid Z=z]$9, bimodal mixture, spike-and-slab, and discrete least-favorable cases. The compared methods are the naïve $q_G(\tau;z)$0 interval, the Cox–Morris parametric EBCI, the AKP robust EBCI, and the NP-EBCI. The reported findings are that all empirical Bayes intervals shrink relative to the naïve interval; Cox–Morris is shortest but can undercover severely when $q_G(\tau;z)$1 is non-Gaussian and especially at low SNR; AKP is conservative; and NP-EBCI maintains coverage very close to $q_G(\tau;z)$2 even in small $q_G(\tau;z)$3, low SNR, non-Gaussian settings while still cutting length substantially (Xie, 8 May 2026).

In the Poisson means setting, the Gamma-smoothed NPMLE paper reports four simulated priors, $q_G(\tau;z)$4, 100 repetitions, and nominal coverage $q_G(\tau;z)$5. The competing methods are Garwood exact Poisson intervals and robust empirical Bayes confidence intervals of Armstrong–Kolesár–Plagborg-Møller. The reported empirical coverage of the oracle-optimal plug-in sets is approximately $q_G(\tau;z)$6 up to Monte Carlo error, with average lengths $q_G(\tau;z)$7, Garwood $q_G(\tau;z)$8, and REBCI $q_G(\tau;z)$9. In NHL skater goals data, the average length is $\theta_i$0 for OPT versus $\theta_i$1 for Garwood and $\theta_i$2 for REBCI, while prediction RMSE matches the classical NPMLE (Kim, 5 May 2026).

Related applications broaden the scope of NP-EBCI methodology. Empirical Bayes rebiasing studies intervals for many noisy and biased estimates, with applications to prediction-powered inference and family-based GWAS; in those examples, the reported gain is often $\theta_i$3 interval-length reduction while preserving frequentist calibration at $\theta_i$4 across many tasks (Ling et al., 8 May 2026). Robust empirical Bayes confidence intervals applied to neighborhood-effect estimates achieve $\theta_i$5 length reduction, while the usual parametric EBCI can undercover by $\theta_i$6 percentage points (Armstrong et al., 2020). Beyond mixture models, empirical Bayesian neural-network uncertainty quantification in nonparametric regression yields Bayesian credible sets with frequentist coverage guarantees, although the theory requires $\theta_i$7 inflation and the simulations report that raw $\theta_i$8 $\theta_i$9-bands have zero coverage whereas $$Y_i\mid \theta_i \sim N(\theta_i,\sigma_i^2),\qquad \theta_i\sim G,$$0 inflation yields approximately $$Y_i\mid \theta_i \sim N(\theta_i,\sigma_i^2),\qquad \theta_i\sim G,$$1 coverage (Franssen et al., 2022).

Taken together, these results portray NP-EBCIs as a technically heterogeneous but coherent inferential program: use nonparametric learning of the prior or marginal law to borrow strength across units, then construct intervals whose validity is calibrated to a clearly specified target—conditional, marginal, average, or selection-adjusted. The main unresolved tension, made explicit throughout the literature, is between nonparametric flexibility and statistical difficulty: the more fully nonparametric the target, the more severely deconvolution ill-posedness constrains the attainable rates (Xie, 8 May 2026).