High-dimensional Bernstein–von Mises theorem for the Diaconis–Ylvisaker prior

We study the asymptotic normality of the posterior distribution of canonical parameter in the exponential family under the Diaconis–Ylvisaker prior which is a conjugate prior when the dimension of parameter space increases with the sample size. We prove under mild conditions on the true parameter value ${\theta }_0$ and hyperparameters of priors, the difference between the posterior distribution and a normal distribution centered at the maximum likelihood estimator, and variance equal to the inverse of the Fisher information matrix goes to 0 in the expected total variation distance. The proof assumes dimension of parameter space $d$ grows linearly with sample size $n$ only requiring $d=o\left(n\right)$. En route, we derive a concentration inequality of the quadratic form of the maximum likelihood estimator without any specific assumption such as sub-Gaussianity. A specific illustration is provided for the Multinomial-Dirichlet model with an extension to the density estimation and Normal mean estimation problems.