Continuity without asymmetric differentiability permits a continuum of optima
OpenPGDRevision20261113.asymmetric_regularity_counterexamplebregman-geometryconvex-optimizationformalization
On the real line set , , and for . Define the loss curvature by
This function is continuous, lies in , and is not differentiable at either or . Its twice-integrated, zero-normalized primitive is convex, is , has second derivative , and has a -Lipschitz derivative.
With , the entire set of minimizing pairs is
Every displayed pair satisfies the global envelope on the whole real line. This gives distinct optimal exponents and explains the differentiability requirement in the asymmetric classification.
Related results. Geometry classification (Theorem 5.1) · Finite-feature design · regularity band identities
Preamble
import Definitions.Def_PGDRevision20261113_Examples open PGDRevision20261113
Formal statement
theorem PGDRevision20261113.asymmetric_regularity_counterexample : RegularityCounterexampleClaim := by sorry
Source
Author manuscript, Exact Residual Bounds and Optimal Geometry Selection for Proximal Gradient Methods, revised reading edition, 13 November 2026; Section 5, paragraph Why asymmetric regularity matters; sections/05_geometry_classification.tex, eq:regularity_counterexample.