Appendix D, p. 17 — law of the pooled weak feature
ProvedRobustOdds.SVM.z_lawadversarial-robustnessclassificationgaussian-modelp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1
Let and let pool the independent weak features of model (3). Conditional on any real label value , their joint law gives
This identifies the one-dimensional Gaussian used to analyze the SVM minimizer.
Formalization Note The theorem permits any real because the conditional Gaussian family is defined for every real parameter; the model uses .
Preamble
import Mathlib import Definitions.Def_RobustOdds_SVM_Setting
Formal statement
namespace RobustOdds.SVM
open MeasureTheory ProbabilityTheory
theorem z_law (d : ℕ) (hd : 1 ≤ d) (η : ℝ) :
∀ y : ℝ, (RobustOdds.Tradeoff.weakLaw d η y).map (fun z => (1 / Real.sqrt d) * ∑ i, z i) =
gaussianReal (y * η * Real.sqrt d) 1 := by sorry
end RobustOdds.SVM
Source
Tsipras et al., Robustness May Be at Odds with Accuracy, arXiv:1805.12152v5, p. 17, App. D, display defining z
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.