Bernstein-von Mises theorem
Let $ \{ {X _ {j} } : {j \geq 1 } \} $
be independent identically distributed random variables with a probability density depending on a parameter $ \theta $(
cf. Random variable; Probability distribution). Suppose that an a priori distribution for $ \theta $
is chosen. One of the fundamental theorems in the asymptotic theory of Bayesian inference (cf. Bayesian approach) is concerned with the convergence of the a posteriori density of $ \theta $,
given $ X _ {1} \dots X _ {n} $,
to the normal density. In other words, the a posteriori distribution tends to look like a normal distribution asymptotically. This phenomenon was first noted in the case of independent and identically distributed observations by P.S. Laplace. A related, but different, result was proved by S.N. Bernstein [a2], who considered the a posteriori distribution of $ \theta $
given the average $ n ^ {-1 } ( X _ {1} + \dots + X _ {n} ) $.
R. von Mises [a12] extended the result to a posteriori distributions conditioned by a finite number of differentiable functionals of the empirical distribution function. L. Le Cam [a5] studied the problem in his work on asymptotic properties of maximum likelihood and related Bayesian estimates. The Bernstein–von Mises theorem about convergence in the $ L _ {1} $-
mean for the case of independent and identically distributed random variables reads as follows, see [a3].
Let $ X _ {i} $, $ 1 \leq i \leq n $, be independent identically distributed random variables with probability density $ f ( x, \theta ) $, $ \theta \in \Theta \subset \mathbf R $. Suppose $ \Theta $ is open and $ \lambda $ is an a priori probability density on $ \Theta $ which is continuous and positive in an open neighbourhood of the true parameter $ \theta _ {0} $. Let $ h ( x, \theta ) = { \mathop{\rm log} } f ( x, \theta ) $. Suppose that $ { {\partial h } / {\partial \theta } } $ and $ { {\partial ^ {2} h } / {\partial \theta ^ {2} } } $ exist and are continuous in $ \theta $. Further, suppose that $ i ( \theta ) = - {\mathsf E} _ \theta [ { {\partial ^ {2} h } / {\partial \theta ^ {2} } } ] $ is continuous, with $ 0 < i ( \theta ) < \infty $. Let $ K ( \cdot ) $ be a non-negative function satisfying
$$ \int\limits _ {- \infty } ^ \infty {K ( t ) { \mathop{\rm exp} } \left [ - { \frac{( i ( \theta _ {0} ) - \epsilon ) t ^ {2} }{2} } \right ] } {d t } < \infty $$
for some $ 0 < \epsilon < i ( \theta _ {0} ) $. Let $ {\widehat \theta } _ {n} $ be a maximum-likelihood estimator of $ \theta $ based on $ X _ {1} \dots X _ {n} $( cf. Maximum-likelihood method) and let $ L _ {n} ( \theta ) $ be the corresponding likelihood function. It is known that under certain regularity conditions there exists a compact neighbourhood $ U _ {\theta _ {0} } $ of $ \theta _ {0} $ such that:
$ {\widehat \theta } _ {n} \rightarrow \theta _ {0} $ almost surely;
$ ( { {\partial { \mathop{\rm log} } L _ {n} ( \theta ) } / {\partial \theta } } ) \mid _ {\theta = {\widehat \theta } _ {n} } = 0 $ for large $ n $;
$ n ^ {1/2 } ( {\widehat \theta } _ {n} - \theta _ {0} ) $ converges in distribution (cf. Convergence in distribution) to the normal distribution with mean $ 0 $ and variance $ {1 / {i ( \theta _ {0} ) } } $ as $ n \rightarrow \infty $.
Let $ f _ {n} ( \theta \mid x _ {1} \dots x _ {n} ) $ denote the a posteriori density of $ \theta $ given the observation $ ( x _ {1} \dots x _ {n} ) $ and the a priori probability density $ \lambda ( \theta ) $, that is,
$$ f _ {n} ( \theta \mid x _ {1} \dots x _ {n} ) = { \frac{\prod _ {i = 1 } ^ { n } f ( x _ {i} , \theta ) \lambda ( \theta ) }{\int\limits _ \Theta {\prod _ {i = 1 } ^ { n } f ( x _ {i} , \phi ) \lambda ( \phi ) } {d \phi } } } . $$
Let $ f _ {n} ^ {*} ( t \mid x _ {1} \dots x _ {n} ) = n ^ {- 1/2 } f _ {n} ( {\widehat \theta } _ {n} + tn ^ {- 1/2 } ) $. Then $ f _ {n} ^ {*} ( t \mid x _ {1} \dots x _ {n} ) $ is the a posteriori density of $ t = n ^ {1/2 } ( \theta - {\widehat \theta } _ {n} ) $.
A generalized version of the Bernstein–von Mises theorem, under the assumptions stated above and some addition technical conditions, is as follows.
If, for every $ h > 0 $ and $ \delta > 0 $,
$$ e ^ {- n \delta } \int\limits _ {\left | t \right | > h } {K ( n ^ {1/2 } t ) \lambda ( {\widehat \theta } _ {n} + t ) } {d t } \rightarrow 0 \textrm{ a.s. } [ {\mathsf P} _ {\theta _ {0} } ] , $$
then
$$ {\lim\limits } _ {n \rightarrow \infty } \int\limits _ {- \infty } ^ \infty {K ( t ) } \cdot $$
$$ \cdot {\left | {f _ {n} ^ {*} ( t \mid X _ {1} \dots X _ {n} ) - \left ( { \frac{i ( \theta _ {0} ) }{2 \pi } } \right ) ^ { {1 / 2 } } e ^ {- { \frac{1}{2} } i ( \theta _ {0} ) t ^ {2} } } \right | } {d t } = $$
$$ = 0 \textrm{ a.s. } [ {\mathsf P} _ {\theta _ {0} } ] . $$
For $ K ( t ) \equiv 1 $ one finds that the a posteriori density converges to the normal density in $ L _ {1} $- mean convergence. The result can be extended to a multi-dimensional parameter. As an application of the above theorem, it can be shown that the Bayesian estimator is strongly consistent and asymptotically efficient for a suitable class of loss functions (cf. [a11]). For rates of convergence see [a4], [a7], [a8].
B.L.S. Prakasa Rao [a6] has generalized the result to arbitrary discrete-time stochastic processes (cf. [a1]); for extensions to diffusion processes and diffusion fields, see [a9], [a10].
References
[a1] | I.V. Basawa, B.L.S. Prakasa Rao, "Statistical inference for stochastic processes" , Acad. Press (1980) |
[a2] | S.N. Bernstein, "Theory of probability" (1917) (In Russian) |
[a3] | J.D. Borwanker, G. Kallianpur, B.L.S. Prakasa Rao, "The Bernstein–von Mises theorem for Markov processes" Ann. Math. Stat. , 43 (1971) pp. 1241–1253 |
[a4] | C. Hipp, R. Michael, "On the Bernstein–von Mises approximation of posterior distribution" Ann. Stat. , 4 (1976) pp. 972–980 |
[a5] | L. Le Cam, "On some asymptotic properties of maximum likelihood estimates and related Bayes estimates" Univ. California Publ. Stat. , 1 (1953) pp. 277–330 |
[a6] | B.L.S. Prakasa Rao, "Statistical inference for stochastic processes" G. Sankaranarayanan (ed.) , Proc. Advanced Symp. on Probability and its Applications , Annamalai Univ. (1976) pp. 43–150 |
[a7] | B.L.S. Prakasa Rao, "Rate of convergence of Bernstein–von Mises approximation for Markov processes" Serdica , 4 (1978) pp. 36–42 |
[a8] | B.L.S. Prakasa Rao, "The equivalence between (modified) Bayes estimator and maximum likelihood estimator for Markov processes" Ann. Inst. Statist. Math. , 31 (1979) pp. 499–513 |
[a9] | B.L.S. Prakasa Rao, "The Bernstein–von Mises theorem for a class of diffusion processes" Teor. Sluch. Prots. , 9 (1981) pp. 95–104 (In Russian) |
[a10] | B.L.S. Prakasa Rao, "On Bayes estimation for diffusion fields" J.K. Ghosh (ed.) J. Roy (ed.) , Statistics: Applications and New Directions , Statistical Publishing Soc. (1984) pp. 504–511 |
[a11] | B.L.S. Prakasa Rao, "Asymptotic theory of statistical inference" , Wiley (1987) |
[a12] | R. von Mises, "Wahrscheinlichkeitsrechnung" , Springer (1931) |
Bernstein–von Mises theorem. Encyclopedia of Mathematics. URL: http://encyclopediaofmath.org/index.php?title=Bernstein%E2%80%93von_Mises_theorem&oldid=22106