← All problems

Davis–Drusvyatskiy–Jiang stochastic-subgradient convergence conjecture

For a generic class of definable, locally Lipschitz, nonconvex objective functions, stochastic subgradient methods converge to local minimizers even without assuming subdifferential regularity of the objective function.

References

Progress summary

Refreshed
Open · solution reported

An unrefereed preprint claims the conjecture is false by exhibiting stochastic subgradient dynamics that approach nonminimizing attractors.

The conjecture concerns whether broad classes of nonsmooth, nonconvex functions force stochastic subgradient methods to converge only to local minimizers. Davis, Drusvyatskiy, and Jiang formulated this question in their work on generic convergence.

Known results

  • Semialgebraic locally Lipschitz functions yield limit points that are first-order stationary (Davis, Drusvyatskiy, Kakade, and Lee, 2018).
  • Generic convergence to local minimizers was proved for definable, subdifferentially regular functions (Davis, Drusvyatskiy, and Jiang, 2023).
  • The corresponding claim for generic definable, non-weakly-convex functions was withdrawn after a cited theorem was found incorrect; that question remained open.

October 2026 counterexample

Jiajin Li's preprint claims a construction with nonminimizing attractors, refuting removal of subdifferential regularity from the generic convergence guarantee and producing failure for open sets of initial conditions and objective tilts. This is a claimed resolution, not a verified theorem.

Current status (as of October 2026): The conjecture has a claimed counterexample in a new preprint, but its mathematical resolution remains unverified.

Sources

Solutions 0

No solutions have been posted yet.