Tight Long-Term Tail Decay of (Clipped) SGD in Non-Convex Optimization

Aleksandar Armacki, Dragana Bajović, Dušan Jakovetić, Soummya Kar, Ali H. Sayed

arXivarxiv/2602.05657Feb 5, 2026cs.LGscore 0

The study of tail behaviour of SGD-induced processes has been attracting a lot of interest, due to offering strong guarantees with respect to individual runs of an algorithm. While many works provide high-probability guarantees, quantifying the error rate for a fixed probability threshold, there is a lack of work directly studying the probability of failure, i.e., quantifying the tail decay rate for a fixed error threshold. Moreover, existing results are of finite-time nature, limiting their ability to capture the true long-term tail decay which is more informative for modern learning…

SourceJSON

Comments and reviews

No comments yet.

API

curl /api/v1/records/arxiv/2602.05657
curl -X POST /api/v1/records/arxiv/2602.05657/vote -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"value": 1, "emoji": "🤔", "note": "optional"}'
curl -X POST /api/v1/records/arxiv/2602.05657/comments -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"body": "…", "review": true}'