An exact quadratic optimum for a nonconstant loss
ProvedPGDRevision20261113.exact_quadratic_geometry_exampleOn , let the loss curvature be , the penalty curvature be , and the candidate curvature be
For , write
The objective is allowed a free linear tilt, with exact implicit updates and a fixed geometry and inverse stepsize. The finite-horizon criterion is the supremum of the squared objective-gradient ratio over those tilts and nonstationary initial states.
For and , the unique minimizing pair and exact rate are
At this pair the global envelope holds throughout . The errors at and are respectively and .
For every finite , the optimal value is , and the same pair uniquely minimizes this criterion. The admissibility conditions, global envelope, exact contacts, uniqueness, and finite-horizon assertions are all included in the formal statement.
Related results. Geometry classification (Theorem 5.1) · Exact profile performance (Theorem 4.1) · quadratic envelope identities
import Definitions.Def_PGDRevision20261113_Examples open PGDRevision20261113
theorem PGDRevision20261113.exact_quadratic_geometry_example : ExactQuadraticExampleClaim := by sorry