Tight Long-Term Tail Decay of (Clipped) SGD in Non-Convex Optimization
Published in arXiv preprint, 2026
Recommended citation: Armacki, A., Bajović, D., Jakovetić, D., Kar, S., & Sayed, A. H. (2025). Tight Long-Term Tail Decay of (Clipped) SGD in Non-Convex Optimization. In arXiv:2602.05657 https://arxiv.org/abs/2602.05657
TLDR: We show that the tail probability induced by SGD-type methods in non-convex optimization decays at an order of magnitude faster rate than previously known and show that our results are tight, by providing matching lower bounds.