The Bregman update yields both gradient-difference identities
ProvedPGDMap20261031.entropy_update_identitybregman-geometryconvex-optimizationperformance-estimation
Statement
A mixture consists of an integer , real numbers indexed by , and a real number , with , for every , and ; no condition is imposed. For every real , write , using the total real logarithm of the code. Define for real . For real , define , where powers are the code's total real-power operation. For real , define , the oriented interval integral with respect to Lebesgue measure. For every such mixture and every real , if , then both and hold. There is no admissibility condition on , no condition , and no nonzero-gradient assumption.
Preamble
import Definitions.Def_PGDMap20261031_Entropy open PGDMap20261031
Formal statement
theorem PGDMap20261031.entropy_update_identity : ∀ (M : EntropyMixture) (c p ell x y : ℝ),
p * geometrySlope c y + M.lambda * artanh y =
p * geometrySlope c x - mixtureSlope M x - ell →
(mixtureSlope M x + M.lambda * artanh x + ell =
p * (geometrySlope c x - geometrySlope c y) + M.lambda * (artanh x - artanh y)) ∧
(mixtureSlope M y + M.lambda * artanh y + ell =
p * (geometrySlope c x - geometrySlope c y) - (mixtureSlope M x - mixtureSlope M y)) := by sorrySource
Jointly Optimal Bregman Geometry and Stepsize for Positive Entropy Mixtures, manuscript dated 2026-10-30, Lemma 1 proof, p. 2; local source: 03_Jointly_Optimal_Bregman_Map_and_Stepsize.pdf