PL conditions let gradient descent converge linearly on highly non-convex functions

MarkSchmidtUBC · x · 2026-11-17

UBC professor Mark Schmidt shared a summary table expressing all the classical conditions for linear convergence of gradient descent (SC, ESC, SQC, WSC, QSC, RSI, EB, PL, QG) as lower bounds on particular directional derivatives — all except QG, which only lower-bounds function values.

The highlight is the PL (not RSI) 'toilet flush' function: a highly non-convex function where the negative gradient can point away from the minimizer, yet gradient descent still achieves a linear convergence rate. Schmidt argues this shows the power of the PL (and equivalent EB) conditions, since none of the other conditions allow this.

Related event: UBC's Mark Schmidt Uses AI to Map Optimization Convergence Assumptions(4 posts)→

Original post →

More from Research

Research channel →