Acceleration with Inexact Gradients: Convergence Rates and Error Accumulation
Time: 2026-10-21
Published By: He Liu
Speaker(s): Yin Liu (BICMR)
Time: 15:00-16:00 October 23, 2026
Venue: Room 29, Quan Zhai, BICMR
Accelerated gradient methods achieve faster convergence rates than standard gradient descent in smooth convex optimization, but they can be more sensitive to gradient estimation errors. This raises a natural question: how much acceleration can be preserved when only approximate gradients are available?
In this talk, I will explain the mechanism of error accumulation under an absolute error condition. Using the Performance Estimation Problem (PEP) framework, we derive nonasymptotic convergence bounds that quantify the overall impact of these errors. Our results reveal a trade-off between acceleration and error accumulation, providing insights into the choice of algorithmic parameters.
