1: \begin{abstract}

2:   We study mean-field variational Bayesian inference using the TAP approach, for

3:     $\Z_2$-synchronization as a prototypical example of a high-dimensional Bayesian

4:     model. We show that for any signal strength $\lambda>1$ (the weak-recovery

5:     threshold), there exists a unique local minimizer of the TAP free energy functional 

6:     near the mean of the Bayes posterior law. Furthermore, the TAP free energy

7: in a local neighborhood of this minimizer is strongly convex. Consequently, a

8: natural-gradient/mirror-descent algorithm achieves linear convergence to this

9: minimizer from a local initialization, which may be obtained by a constant number

10: of iterations of Approximate Message Passing (AMP). This provides a rigorous foundation for variational inference in high dimensions via minimization of the TAP free energy.

11: 

12:     We also analyze the finite-sample convergence of AMP, showing that AMP

13:     is asymptotically stable at the TAP minimizer for any $\lambda>1$, and is

14:     linearly convergent to this minimizer from a spectral initialization

15:     for sufficiently large $\lambda$. Such a guarantee is stronger than

16:     results obtainable by state evolution analyses, which only describe a fixed

17:     number of AMP iterations in the infinite-sample limit.

18: 

19:     Our proofs combine the Kac-Rice formula and Sudakov-Fernique Gaussian comparison inequality to analyze the complexity of critical points that satisfy strong convexity and stability conditions within their local neighborhoods.

20: \end{abstract}

21: