§9.1.2 — eliminate the nonnegative diagonal in the factor-model loss
ProvedBoydADMM.Nonconvex.factor_model_partial_minimizationadmmfactor-modelp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-bookp2o-v1partial-minimization
Let and let be symmetric. Define . Then
and the infimum is attained by , which is componentwise nonnegative.
This identifies the reduced convex loss that appears in the factor-model ADMM X-update.
Formalization Note The sum over includes both ordered entries and , as in the Frobenius norm. The book treats as a positive matrix dimension; Lean states explicitly.
Preamble
import Mathlib import Definitions.Def_BoydADMM_Nonconvex_FactorModel
Formal statement
namespace BoydADMM.Nonconvex
/-- §9.1.2, p. 75: eliminating the nonnegative diagonal gives the stated
componentwise loss, attained at `d_i = (Σ_ii - X_ii)_+`. -/
theorem factor_model_partial_minimization {n : ℕ} (Sigma X : SqMat n)
(hn : 0 < n) (hSigma : IsSymmetric Sigma) (hX : IsSymmetric X) :
factorLoss Sigma X = factorLossExplicit Sigma X ∧
(∀ i, 0 ≤ optimalDiag Sigma X i) ∧
factorLossWithDiag Sigma X (optimalDiag Sigma X) = factorLoss Sigma X := by sorry
end BoydADMM.Nonconvex
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. 75, §9.1.2, display beginning f(X) = inf
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.