{"id":"arxiv/2606.17523","url":"https://arxiv.org/abs/2606.17523","tags":["math.OC","cs.LG"],"score":0,"title":"Beyond IGO-Flow: Toward Convergence Analysis of IGO in Continuous Spaces","domain":"math.OC","source":"arxiv","authors":["Ryosuke Kimura","Youhei Akimoto"],"reviews":0,"subject":"math.OC","summary":"Information-Geometric Optimization (IGO) provides a unified framework for black-box optimization by interpreting the adaptation of a search distribution as a natural gradient update. Despite its conceptual importance, the convergence theory of IGO remains limited: most existing results concern continuous-time idealizations such as the IGO flow, rather than discrete-time updates with non-infinitesimal learning rates. In this paper, we study discrete-time IGO in continuous spaces, formulated as natural gradient updates in the expectation-parameter coordinates of an exponential family. In…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-06-16","reactions":{}}