1: \begin{abstract}

2: This paper presents an approach for efficiently approximating the inverse of Fisher information, a key component in variational Bayes inference. A notable aspect of our approach is the avoidance of analytically computing the Fisher information matrix and its explicit inversion. Instead, we introduce an iterative procedure for generating a sequence of matrices that converge to the inverse of Fisher information. 

3: The natural gradient variational Bayes algorithm without matrix inversion is provably convergent and achieves a convergence rate of order $\mathcal O(\log s/s)$, with $s$ the number of iterations.

4: We also obtain a central limit theorem for the iterates.

5: Our algorithm exhibits versatility, making it applicable across a diverse array of variational Bayes domains, including Gaussian approximation and normalizing flow Variational Bayes. 

6: We offer a range of numerical examples to demonstrate the efficiency and reliability of the proposed variational Bayes method.

7: \end{abstract}

8: