§8.3.4, pp. 70–71 — the feature-split SVM z̄-update is shifted soft thresholding of v = Ax̄^{k+1} + u^k
ProvedBoydADMM.ModelFit.svm_zbar_shifted_soft_thresholdLet , , and (the averaged partial predictor and the scaled dual variable ); write . The -update of the feature-split support vector machine minimizes
where is the componentwise positive part. A point minimizes this function if and only if, for every component ,
In particular the minimizer is unique.
The -update thus splits into scalar problems, each solved by this shifted soft thresholding.
Formalization Note The objective is exactly the one printed on p. 70, with coefficient on the quadratic; the general feature-split -update of p. 68 carries instead, and the printed thresholds are the ones that match the printed . The book writes ; after the reduction of §8.3 all duals are equal, so .
import Mathlib import Definitions.Def_BoydADMM_ModelFit_Basic open Matrix
namespace BoydADMM.ModelFit
/-- §8.3.4, pp. 70–71: the feature-split SVM `z̄`-update
`argmin_{z̄} (1ᵀ(N z̄ + 1)₊ + (ρ/2)‖z̄ − Ax̄ − u‖₂²)` is the shifted soft thresholding of
`v = Ax̄ + u`, componentwise. -/
theorem svm_zbar_shifted_soft_threshold {N m : ℕ} (hN : 0 < N) (ρ : ℝ) (hρ : 0 < ρ)
(Axbar u zb : EuclideanSpace ℝ (Fin m)) :
IsMinOn (fun z : EuclideanSpace ℝ (Fin m) =>
(∑ j, max ((N : ℝ) * z j + 1) 0) + ρ / 2 * ‖z - Axbar - u‖ ^ 2) Set.univ zb ↔
∀ j, zb j = shiftedSoftThreshold N ρ (Axbar j + u j) := by sorry
end BoydADMM.ModelFit
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.