Documentation

LeanPool.PLAcceleratedNesterovLean.Convergence.StateContraction.AuxVarRecursion

Auxiliary Variable Recursion for Arbitrary States #

The one-step recursion for the auxiliary variable u, proved for arbitrary NesterovState (not just nesterovSeq-reachable states). This is the state-based analogue of auxVar_recursion from AuxVar.lean.

Key identity: u' = ((1-a)·P⊥v + √μ'·e - √η·P⊥g) + √μ'·ξ where u' = auxVarOfState at step(s), and all quantities are computed from s.

theorem PLAcceleratedNesterovLean.stepDispOfState_eq {d : } (f : E d) (η ρ : ) (s : NesterovState d) (hη_pos : 0 < η) :
stepDispOfState f η ρ s = ρ η s.v - (1 + ρ) η gradient f (s.lookahead η)

Step displacement formula: h = ρ·√η·v - (1+ρ)·η·g

theorem PLAcceleratedNesterovLean.auxVarOfState_step {d : } (P : E d →L[] E d) (μ' η ρ : ) (π : E dE d) (f : E d) (s : NesterovState d) ( : ρ = (1 - (μ' * η)) / (1 + (μ' * η))) (ha_pos : 0 < (μ' * η)) (hη_pos : 0 < η) (hμ_pos : 0 < μ') :
have gn := gradient f (s.lookahead η); have en := normalDispOfState π η s; have ξn := curvatureErrorOfState (⇑P) π f η ρ s; have a := (μ' * η); auxVarOfState P μ' π η (nesterovStep f η ρ s) = (1 - a) (s.v - P s.v) + μ' en - η (gn - P gn) + μ' ξn

One-step auxVar recursion for arbitrary states.