There exists a function satisfying two-sided PL conditions for which GDA dynamics fail to converge by circling the saddle point.
Solving a class of non-convex min-max games using iterative first order methods
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
math.OC 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
PL conditions do not guarantee convergence of gradient descent-ascent dynamics
There exists a function satisfying two-sided PL conditions for which GDA dynamics fail to converge by circling the saddle point.