The "toilet flush" function: PL conditions give linear convergence on non-convex landscapes

MarkSchmidtUBC · x · 2026-11-17

Mark Schmidt showcases a spiral "toilet flush" function (PL but not RSI) where gradient descent still achieves linear convergence on a highly non-convex landscape — the negative gradient can even point away from the minimizer. He argues this demonstrates the unique power of PL/equivalent Error Bound conditions, which no other standard condition permits.

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

Original post →

More from Research

Research channel →