Convergence rates of latent topic models under relaxed identifiability conditions (Q1711601)

The paper under review explores the convergence rate for the latent Dirichlet allocation topic models introduced by \textit{D. M. Blei} et al. [J. Mach. Learn. Res. 3, No. 4--5, 993--1022 (2003; Zbl 1112.68379)]. The estimation error between the underlying true model and an estimator is evaluated by the Wasserstein's distance. The convergence rate of the maximum likelihood estimator is obtained in the general case, thus generalizing the previous works of \textit{A. Anandkumar} et al. [Algorithmica 72, No. 1, 193--214 (2015; Zbl 1311.68130); J. Mach. Learn. Res. 15, 2239--2312 (2014; Zbl 1318.68136)].