Proves ELBO solution existence for elliptic location-scale families and convergence guarantees for preconditioned dynamic-batched projected SGD under Blum-Gladyshev conditions in BBVI.
Classical convergence analyses of SGD typically rely on the assumption of uniformly bounded variance [33, 15, 6], or more recently, on the Expected Smoothness (ABC) condition [25]
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
SGD for Variational Inference: Tackling Unbounded Variance via Preconditioning and Dynamic Batching
Proves ELBO solution existence for elliptic location-scale families and convergence guarantees for preconditioned dynamic-batched projected SGD under Blum-Gladyshev conditions in BBVI.