The parallel-projection dual average vanishes in one step
ProvedBoydADMM.Constrained.average_stepadmmconvex-optimizationp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-bookp2o-v1projections
Let , , let be nonempty closed convex sets, and let be any parallel projection ADMM run from arbitrary . Averaging its second and third updates gives, for every ,
The identities explain when the common iterate equals the average and provide the index base for the simplified algorithm.
Formalization Note The book prints in the intermediate averaged dual update; the displayed iteration on page 35 and the next line on page 36 require . This statement uses the corrected index.
Preamble
import Mathlib import Definitions.Def_BoydADMM_Constrained_Run
Formal statement
namespace BoydADMM.Constrained
/-- §5.1.2, p. 36: averaging the last two updates eliminates the dual average. -/
theorem average_step {N n : ℕ} (hN : 0 < N) (A : Fin N → Set (Vec n))
(hAne : ∀ i, (A i).Nonempty) (hAclosed : ∀ i, IsClosed (A i))
(hAconvex : ∀ i, Convex ℝ (A i))
(ρ : ℝ) (hρ : 0 < ρ)
(x : ℕ → Blocks N n) (z : ℕ → Vec n) (u : ℕ → Blocks N n)
(hrun : IsParallelRun A x z u) :
∀ k, z (k + 1) = avg (x (k + 1)) + avg (u k) ∧ avg (u (k + 1)) = 0 := by sorry
end BoydADMM.Constrained
Source
Boyd, Parikh, Chu, Peleato, Eckstein, Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers, Found. Trends Mach. Learn. 3(1) (2011), p. 36, §5.1.2, averaged updates; https://doi.org/10.1561/2200000016
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.