Stochastic gradient descent (SGD) is a workhorse of optimization, and recent theory has shown that it converges even in its final iterate, not just on average. But those guarantees come with a catch: they typically assume gradient noise with exponentially decaying tails. Under heavy-tailed noise, which is common in real-world data, SGD can behave badly or fail to converge.

Clipped-SGD, which clips gradient updates to a bounded norm, was introduced to handle such heavy-tailed noise. It has performed well in practice, but its last-iterate behavior remained largely unstudied. In a new paper, Zijian Liu proves the first optimal last-iterate convergence rates for Clipped-SGD under heavy-tailed noise, with high-probability guarantees.

The result closes a gap between theory and practice for a simple and widely used algorithm. The paper, accepted to NeurIPS 2026, is available on arXiv.