Bayesian Posterior Contraction

• Posterior contraction in Bayesian settings is the process where the posterior distribution increasingly concentrates around the true parameter as more data is observed.
• In high-dimensional and nonparametric models, contraction rates are crucial for achieving adaptive, minimax-optimal inference using methods like spike-and-slab and Gaussian process priors.
• Advanced techniques such as Wasserstein metrics and SPDE representations facilitate rigorous uncertainty quantification and efficient computation in complex Bayesian problems.

Posterior contraction in Bayesian settings refers to the phenomenon where the posterior distribution, given increasing amounts of data, becomes increasingly concentrated (in a suitable sense) around the true value of the parameter or function generating the data. Posterior contraction—quantified via contraction rates—is fundamental to Bayesian nonparametric, high-dimensional, and inverse-problem methodologies, determining the strength and reliability of Bayesian inference under various model, prior, and observational regimes.

1. Definition and Foundational Principles

A posterior contraction rate (PCR) is a sequence $\epsilon_n \downarrow 0$ such that, for any diverging sequence $M_n \to \infty$, the posterior probability assigned to parameters $||\theta - \theta_0|| > M_n \epsilon_n$ converges to zero under the truth $P_{\theta_0}$. For example, for a parameter space $\Theta$ equipped with norm $||\cdot||$, a prior $\Pi$ and posterior $\Pi_n$, one requires

$$\Pi_n\{\theta: ||\theta - \theta_0|| > M_n \epsilon_n\} \to 0 \quad \text{in } P_{\theta_0}\text{-probability}.$$

For infinite-dimensional settings, the same rate governs contraction in strong norms (e.g. Sobolev or $L^2$), and, equivalently, in the $M_n \to \infty$0-Wasserstein distance on $M_n \to \infty$1, $M_n \to \infty$2 (Dolera et al., 2022).

PCRs generalize Bayesian consistency, quantifying the speed at which the posterior "learns" as more data are observed, and are defined analogously for finite- and infinite-dimensional, parametric and nonparametric, as well as function/process-valued parameters.

2. Posterior Contraction in High-Dimensional and Sparse Models

In sparse high-dimensional models, posterior contraction is both minimax-optimal and adaptive if the prior, often constructed via continuous shrinkage or spike-and-slab families, is sufficiently concentrated around the low-dimensional (sparse) structure. For the sparse normal means model with $M_n \to \infty$3 nonzero signals among $M_n \to \infty$4 components, the minimax $M_n \to \infty$5-contraction rate is $M_n \to \infty$6.

Key sufficient conditions for contraction at the minimax rate include:

• The prior must assign enough mass near zero (for "spike") to force most coefficients close to zero, and heavy enough tails (at least Laplace, not heavier than Cauchy) for large signals.
• Horseshoe, horseshoe+, normal-gamma, inverse-Gaussian, and spike-and-slab Lasso all satisfy these, yielding contraction at rate $M_n \to \infty$7; see (Pas et al., 2015).
• Both empirical Bayes (e.g., MMLE for global shrinkage parameter in horseshoe priors) and hierarchical Bayes (placing a hyper-prior) deliver adaptive contraction rates, automatically matching the unknown sparsity $M_n \to \infty$8 (Pas et al., 2017).

Credible sets constructed from such posteriors can be shown to adaptively cover the true sparse vector, with radii also of the minimax order (Pas et al., 2017).

3. Bayesian Posterior Contraction in Nonparametric and Inverse Problems

In nonparametric models, contraction rates depend critically on the interplay between data-generating process smoothness and prior regularity.

• For Gaussian process (GP) priors on Sobolev/Besov smoothness spaces $M_n \to \infty$9 and the true function $||\theta - \theta_0|| > M_n \epsilon_n$0, the minimax rate in $||\theta - \theta_0|| > M_n \epsilon_n$1 or strong Sobolev norm is $||\theta - \theta_0|| > M_n \epsilon_n$2 ($||\theta - \theta_0|| > M_n \epsilon_n$3 data dimension), attained for suitably chosen prior smoothness and sample-size scaling (Rosa, 23 Dec 2025, Dolera et al., 2022).
• Random series/sieve priors (e.g., B-spline and wavelet expansions) yield similar polynomial contraction rates, with the potential for adaptation to unknown $||\theta - \theta_0|| > M_n \epsilon_n$4 via random truncation or hierarchical construction (Rosa, 23 Dec 2025).
• In severely ill-posed inverse problems (e.g., compact operators with exponentially decaying singular values), contraction is logarithmic, $||\theta - \theta_0|| > M_n \epsilon_n$5, where $||\theta - \theta_0|| > M_n \epsilon_n$6 is the truth's smoothness and $||\theta - \theta_0|| > M_n \epsilon_n$7 the exponent of the singular value decay (Agapiou et al., 2012). Mildly ill-posed settings (algebraic singular value decay) admit polynomial rates (Agapiou et al., 2012, Jia et al., 2018).

Adaptive and non-diagonal-structure contraction rates can be achieved by empirical Bayes estimation of regularity hyperparameters even when a common eigenbasis is not shared by the forward operator, prior, and noise, up to a slowly varying (e.g., $||\theta - \theta_0|| > M_n \epsilon_n$8) factor (Jia et al., 2018).

4. Posterior Contraction via Wasserstein and SPDE Dynamics

Recent advances link posterior contraction to Wasserstein geometry and stochastic partial differential equations (SPDE):

• In infinite-dimensional exponential families and GP-prior models, contraction in strong norms (e.g., $||\theta - \theta_0|| > M_n \epsilon_n$9) is proved via bounds on Wasserstein-$P_{\theta_0}$0 distances between the posterior and truth, controlled via Laplace integrals and local Lipschitz properties of the posterior in sufficient statistics. Rates are determined by explicit spectral properties of the prior covariance and the Kullback-Leibler divergence (Dolera et al., 2022).
• The posterior in nonparametric or inverse problems can be represented as the invariant measure of a Langevin SPDE on a Hilbert space, enabling non-asymptotic moment and concentration control in the Hilbert norm. Under suitable curvature and empirical process conditions, this yields contraction rates $P_{\theta_0}$1 (parametric) or slower in the presence of ill-posedness or limited prior regularity (Alberola-Boloix et al., 23 Mar 2026). Such formalisms extend to Laplace approximations and finite-sample Bernstein–von Mises results for Bayesian infinite-dimensional models.

5. Contraction in Structured and Hierarchical Models

Bayesian adaptation to structural dimension or function composition can be achieved:

• In sparse neural networks, optimal contraction rates $P_{\theta_0}$2 are attainable over anisotropic (and hierarchical/composite) Besov spaces, where $P_{\theta_0}$3 quantifies intrinsic smoothness via layerwise or blockwise low-dimensionality. Both spike-and-slab and shrinkage-type priors induce adaptive contraction, matching correct minimax rates even when the smoothness is unknown (Lee et al., 23 Jun 2025).
• Oracle contraction theory using hierarchical (two-step) priors shows that under local Gaussianity, local entropy, and sufficient prior mass, the posterior contracts at the oracle rate $P_{\theta_0}$4—automatically adapting to the unknown complexity of the true model (Han, 2017). This applies to trace regression, shape-restricted regression, sparse covariance estimation, and more.

Posterior contraction for mixture models exhibits different behaviors depending on whether finite mixtures (with explicit prior on number of components) or nonparametric processes (e.g., Dirichlet process mixtures) are used. For finite mixtures under identifiability, the rate is $P_{\theta_0}$5 in the Wasserstein metric for parameters; for misspecified or infinite mixture models, rates are generally slower and can be improved by post-processing algorithms such as merge–truncate–merge (Guha et al., 2019).

6. Prior Regularity and Rate-Determining Factors

In infinite-dimensional and nonparametric models, the achievable posterior contraction rate is dictated by:

• The smoothness of the prior relative to the truth.
• Eigenvalue/spectral decay of prior covariance—faster decay (smoother prior) enables faster contraction up to the parametric rate.
• The interaction between bias (approximation error) and variance (random error) as encoded by the model, prior, and observation operator spectra.
• For models with severe ill-posedness, contraction is fundamentally limited to logarithmic rates by the exponential decay in the forward operator’s singular values, regardless of prior smoothness (Agapiou et al., 2012).

In hierarchical and adaptation contexts, priors which do not undersmooth, and assign sufficient mass to small neighborhoods of the truth (in Kullback-Leibler or strong norm balls), enable the posterior to contract at nearly minimax rates even without knowledge of key structural or smoothness parameters (Pas et al., 2017, Han, 2017).

7. Methodological and Computational Implications

Posterior contraction analysis relies on tools including:

• Construction of powerful tests and entropy/sieve arguments (classical approach).
• Wasserstein dynamic inequalities, local Lipschitz continuity, and infinite-dimensional Laplace or Poincaré–Wirtinger estimates (functional-analytic approach) (Dolera et al., 2022).
• SPDE representations and stochastic process moment control (diffusion process/Langevin approach) (Alberola-Boloix et al., 23 Mar 2026, Mou et al., 2019).
• Data-driven or empirical Bayes tuning (e.g., MMLE for shrinkage parameters or regularity indices) enables adaptation without requiring prior knowledge of signal sparsity or function smoothness (Pas et al., 2017, Jia et al., 2018).

Efficient computational procedures—closed-form posterior updates for conjugate models, scalable MCMC or optimization for hierarchical/empirical Bayes, and fast spectral methods—are facilitated by these contraction analyses.

Posterior contraction theory underpins the reliability of Bayesian credible sets: when the posterior contracts at minimax-optimal rates, credible balls or regions constructed from it provide honest, adaptive uncertainty quantification in both finite- and infinite-dimensional settings (Pas et al., 2017, Yoo et al., 2016).

---

References:

• "Adaptive posterior contraction rates for the horseshoe" (Pas et al., 2017)
• "Conditions for Posterior Contraction in the Sparse Normal Means Problem" (Pas et al., 2015)
• "Posterior contraction for empirical Bayesian approach to inverse problems under non-diagonal assumption" (Jia et al., 2018)
• "Bayesian Posterior Contraction Rates for Linear Severely Ill-posed Inverse Problems" (Agapiou et al., 2012)
• "Strong posterior contraction rates via Wasserstein dynamics" (Dolera et al., 2022)
• "SPDE Methods for Nonparametric Bayesian Posterior Contraction and Laplace Approximation" (Alberola-Boloix et al., 23 Mar 2026)
• "Posterior Contraction for Sparse Neural Networks in Besov Spaces with Intrinsic Dimensionality" (Lee et al., 23 Jun 2025)
• "Oracle posterior contraction rates under hierarchical priors" (Han, 2017)
• "Posterior contraction rates in a sparse non-linear mixed-effects model" (Naveau et al., 2024)
• "Posterior contraction rates of computational methods for Bayesian data assimilation" (Burman et al., 17 Jun 2025)
• "Posterior contraction and uncertainty quantification for the multivariate spike-and-slab LASSO" (Shen et al., 2022)
• "Bayesian mode and maximum estimation and accelerated rates of contraction" (Yoo et al., 2016)
• "Posterior contraction rates in non-dominated Bayesian nonparametric models" (Camerlenghi et al., 2022)
• "Rate-optimal posterior contraction for sparse PCA" (Gao et al., 2013)