美文网首页
Notes for "A variational Bayesia

Notes for "A variational Bayesia

作者: jjx323 | 来源:发表于2019-01-14 12:58 被阅读0次

Proof of Theorem 3.3 

Theorem 3.3. Suppose that D_{KL} has a multivariate Taylor expansion with a positive definite Hessian at \Theta_{0}. Then we have \Theta_{k} \rightarrow \Theta_{0}.

Proof: Given \Theta_{0} there exists \epsilon > 0 such that 

D_{KL}(\Theta)=D_{KL}(\Theta_{0})+\frac{1}{2}(\Theta - \Theta_0)^{T}H(\Theta_0)(\Theta-\Theta_0) + o(\|\Theta - \Theta_0\|^2)

with H being the Hessian with respect to \Theta and \|\Theta - \Theta_0\|^2 < \epsilon. Considering the positive definiteness of the Hessian at the stationary point \Theta_0 and the smoothness of D_{KL}, we know that if \|\Theta - \Theta_0\|^2<\epsilon, then D_{KL}(\Theta) - D_{KL}(\Theta_0) < \epsilon_1implies that \|\Theta - \Theta_0\|^2<c\epsilon_1 for some c>0

Since \{\Theta_k\}_{k=1}^{\infty} is a bounded sequence, there is a subsequence \{\Theta_{n_k}\}_{k=1}^{\infty} such that \Theta_{n_k}\rightarrow \Theta_0 as k\rightarrow \infty. For every \epsilon>0, choosing \epsilon_1>0 such that c\epsilon_1 <\epsilon, we can take K>0 large enough such that D_{KL}(\Theta_{n_K}) - D_{KL}(\Theta_0) \leq \epsilon_1. Recalling the above paragraph, we have \|\Theta_{n_K} - \Theta_0\|^2<c\epsilon_1 < \epsilon. By taking \ell > n_{K}, we have D_{KL}(\Theta_{\ell}) - D_{KL}(\Theta_0) \leq \epsilon_1which implies \|\Theta_{\ell} - \Theta_0\|^2 < \epsilon. At this stage, the conclusion obviously holds true. 

相关文章

网友评论

      本文标题:Notes for "A variational Bayesia

      本文链接:https://www.haomeiwen.com/subject/ilkhqqtx.html