Nonparametric & Functional BvM Theorems

• Nonparametric and functional BvM theorems extend classical Gaussian approximations to infinite-dimensional settings, ensuring credible set validity under strict regularity.
• They utilize local quadratic (LAN) expansions and precise posterior contraction rates to quantify the Gaussian behavior of complex Bayesian models.
• Applications include density estimation, regression, and covariance functionals, where carefully controlled conditions lead to sound Bayesian uncertainty quantification.

Nonparametric and Functional Bernstein–von Mises Theorems (BvM) describe the asymptotic Gaussianity of posterior distributions for infinite-dimensional parameters and their functionals in nonparametric and semiparametric statistical models. Extending the classical parametric BvM phenomenon, these results form the theoretical foundation for justifying Bayesian uncertainty quantification and credible sets in complex models, highlighting regimes in which Bayesian credible sets attain frequentist validity, and revealing the influence (or lack thereof) of the prior in infinite-dimensional contexts.

1. Classical and Nonparametric Bernstein–von Mises Phenomenon

Classically, the BvM theorem asserts that the posterior distribution of a finite-dimensional parameter, when centered at an efficient estimator and rescaled by the square root of the Fisher information, converges to a standard normal law. In nonparametric and high-dimensional settings, this phenomenon undergoes substantial modifications due to the infinite-dimensional structure and the nontrivial role of the prior.

In the nonparametric case, the impact of the prior cannot be fully neglected. The posterior can concentrate at suboptimal rates or its limiting shape can remain non-Gaussian, depending on the interplay between the prior and model complexity. The accuracy of the Gaussian approximation is then quantified via explicit metrics (total variation, Kolmogorov, or bounded-Lipschitz), and requires conditions on posterior contraction and model effective dimension, as detailed in (Spokoiny et al., 2019) and (Castillo et al., 2012).

2. Weak and Strong Nonparametric BvM: Posterior Distribution Convergence

There are two principal levels of BvM results in nonparametric contexts:

• Weak Nonparametric BvM: The posterior law, shifted and rescaled, converges in distribution to a Gaussian measure on an appropriate infinite-dimensional Hilbert space (e.g., Sobolev spaces), typically in bounded-Lipschitz (BL) or Prokhorov metrics. In the Gaussian white noise model

$$dX^{(n)}(t)=f(t)\,dt + n^{-1/2}\,dW(t),\quad t\in[0,1],$$

the weak BvM states that the posterior around the data-driven shift $\mathbb{X}^{(n)} = f + n^{-1/2}W$ converges to a canonical Gaussian $\mathcal{N}$ on $H^{-1/2,\delta}_2$ for $\delta > 1/2$ (Castillo et al., 2012).

• Strong/Total Variation BvM: Under further assumptions (notably on prior regularity and effective dimension $d_{\text{eff}}$), the posterior can be approximated in total variation by a Gaussian law with centering at the penalized MLE and covariance related to the Fisher information:

$$\Bigl\|\Pi(\cdot \mid X) - N(\hat\theta, (nI+\Sigma^{-1})^{-1})\Bigr\|_{\mathrm{TV}} \le C\,\frac{d_{\rm eff}^{3/2}}{n}$$

with $d_{\text{eff}} = \mathrm{tr}(I\Sigma) = o(n^{1/3})$ (Spokoiny et al., 2019).

A key methodological ingredient is the local quadratic expansion (LAN), allowing the application of Gaussian comparison arguments and Laplace transform calculations.

3. Functional BvM Theorems: Linear and Nonlinear Functionals

Functional BvM theorems assert asymptotic Gaussianity for smooth (typically, Fréchet-differentiable) functionals of the parameter $\theta$ or $f$:

• For linear functionals $L(\theta)$, the posterior of $n^{1/2} (L(\theta) - L(\hat\theta))$ converges to a normal law with variance $\sigma_L^2 = \langle L, I^{-1} L \rangle$ (Spokoiny et al., 2019, Castillo et al., 2012).
• For nonlinear, smooth functionals $\psi(\eta)$, with a second-order Taylor expansion around the truth and suitable bias control, the posterior (after appropriate centering and bias correction) converges to a normal law at rate $n^{-1/2}$, with limiting variance given by the semi-parametric efficiency bound (Castillo et al., 2013, Castillo et al., 2012).

In semiparametric and nonparametric regression (e.g., Bayesian Additive Regression Trees/BART), for smooth linear functionals of the regression function, BvM theorems hold under compatibility ("no-bias") and regularity (self-similarity) conditions, with centering at efficient estimators and variance given by empirical $L^2$ norms of the test function (Rockova, 2019).

4. Sufficient Conditions and Model-Specific Mechanisms

The validity of nonparametric and functional BvM results depends on specific structural and regularity conditions:

• Posterior Contraction: The posterior must concentrate in appropriate balls (Sobolev, $L^2$, operator norm) around the truth at subparametric or parametric rates.
• LAN-like Expansion: The model likelihood admits a local quadratic expansion in the infinite-dimensional norm, enabling local approximation by a Gaussian shift experiment (Castillo et al., 2013).
• No-Bias/Prior-Shift Invariance: The prior must be invariant or nearly invariant under small shifts in the direction of the functional's influence function, ensuring that bias corrections vanish asymptotically.
• Functional Regularity: The functional must be sufficiently smooth (typically, Fréchet differentiable of order up to 2), with influence functions belonging to the relevant RKHS or Sobolev class.
• Self-Similarity: For adaptive procedures (e.g., BART or histogram priors with random partitions), the function/truth must be "self-similar" in Hölder or Sobolev sense to avoid degeneracy ("Gaussian mixtures" as limits), and for posterior mass to concentrate on efficiently regular partitions (Rockova, 2019).

A summary table is provided to relate types of BvM results to key model and functional requirements:

BvM Context	Main Conditions	Limiting Law
Weak Nonparametric (GWN)	Posterior contraction in Hilbert norm	Canonical Gaussian
Strong/TV BvM	Effective dimension $d_{\rm eff}=o(n^{1/3})$, local quadratic expansion	Centered Gaussian
Functional (linear)	Posterior contraction, smooth $L$	$N(0, \sigma_L^2)$
Functional (nonlinear)	Second-order expansion, bias control	$N(0, V)$
BART functionals	Self-similarity, no-bias, posterior contraction	$N(0, \|\ell\|^2)$
Covariance matrix entries	Contracting neighborhoods, Fréchet differentiability	$N(0, \text{asymp var})$

5. Examples and Applications

• Gaussian White Noise Model: BvM for linear/nonlinear functionals, $L^2$-balls, and auto-convolution operators. Minimax Bayesian credible sets with coverage $1-\alpha$ and $L^2$-diameter matching minimax rates (up to logarithmic factors) (Castillo et al., 2012).
• Density Estimation: For random histogram and Gaussian process priors, posterior distributions of smooth functionals (e.g., entropy, moments) satisfy BvM at root-$n$ rate, provided explicit prior shift and bias conditions (Castillo et al., 2013).
• Regression with Gaussian Priors: For generalized regression or log-density estimation, the nonparametric and functional BvM applies with explicit rates in terms of model smoothness and prior regularization (Spokoiny et al., 2019).
• Covariance Matrix Functionals: Entries, quadratic forms, log-determinants, eigenvalues, as well as functionals arising in discriminant analysis admit BvM under explicit bias and prior-compatibility control. The variance and centering formulas are given in terms of influence matrices and sample covariance estimation rates (Gao et al., 2014).
• BART and Regression Trees: Bernstein–von Mises theorems for smooth linear functionals under Bayesian Additive Regression Trees are established by verifying contraction, prior-shift invariance, and self-similarity conditions, with the additional complication that random (adaptive) partitions may otherwise lead to Gaussian mixture limits absent self-similarity (Rockova, 2019).

6. Methodological Frameworks and Proof Techniques

The general proof strategy consists of the following steps:

• Expansion of the log-likelihood around a data-driven estimator (MLE, penalized MLE, or empirical mean), using Local Asymptotic Normality (LAN).
• Taylor expansion of the target functional, with explicit linear and second-order terms.
• Laplace transform techniques: demonstrating convergence of the posterior Laplace/characteristic function to the Gaussian, often via change of measure and control of prior-shift ratios.
• Empirical process and concentration inequalities to quantify remainder and contraction rates.
• Direct total variation or BL-metric comparison for weak/strong convergence, with limiting error rates explicitly given in terms of model complexity (effective dimension, partition size, etc.) (Spokoiny et al., 2019, Castillo et al., 2012, Castillo et al., 2013).

For functionals of covariances and precision matrices, detailed operator and trace bounds, as well as explicit formulas for asymptotic variance in the influence matrix, are required (Gao et al., 2014).

7. Implications, Limitations, and Scope

Nonparametric and functional BvM theorems supply rigorous Bayesian-frequentist calibration for uncertainty quantification in infinite-dimensional models. The achievable rate (typically $n^{-1/2}$, sometimes $O(n^{-1})$ in norm) and limiting law inform whether and in what sense Bayesian credible sets can be used as frequentist confidence sets with asymptotically correct coverage. Deviations from Gaussianity (e.g., Gaussian mixtures as limits) may occur if adaptivity issues or smoothness incompatibility ("curse of adaptivity") are not addressed through structural assumptions like self-similarity.

The necessity of strong regularity, contraction, and prior-shift invariance conditions highlights a core limitation: in the absence of such properties, the prior's influence remains non-negligible, and the posterior may not provide valid frequentist uncertainty quantification (Spokoiny et al., 2019, Rockova, 2019, Castillo et al., 2013). These theorems embody the state-of-the-art extension of the Bernstein–von Mises paradigm to the Bayesian analysis of nonparametric and high-dimensional models, including semiparametric and functional parameters.