Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 5, p. 21 — oracle inequalities for the expected prediction error of the BIC-type estimators β̂^B_2 and β̂^B_1

Proved
SubsetShrink.BIC.theorem_5

by mikedeng1 · Oct 15, 2026 · Mathlib 0df444a (Lean v4.33.1)

best-subset-selectionmodel-selectionoracle-inequalityp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1sparse-regression

Work in the model y=f∗+ϵy=f^*+\epsilony=f∗+ϵ, where f∗∈Rnf^*\in\mathbb R^nf∗∈Rn is deterministic, ϵ\epsilonϵ has independent N(0,σ2)N(0,\sigma^2)N(0,σ2) entries with σ>0\sigma>0σ>0, and the design X∈Rn×pX\in\mathbb R^{n\times p}X∈Rn×p (p≥1p\ge1p≥1) has columns of unit Euclidean norm. For a,b≥0a,b\ge0a,b≥0 let λβ=a∥β∥0log⁡(ep/∥β∥0)\lambda_\beta=a\sqrt{\|\beta\|_0\log(ep/\|\beta\|_0)}λβ​=a∥β∥0​log(ep/∥β∥0​)​ and μβ=blog⁡(ep/∥β∥0)\mu_\beta=b\log(ep/\|\beta\|_0)μβ​=blog(ep/∥β∥0​), and define the BIC-type estimators

β^2B∈arg⁡min⁡β ∥y−Xβ∥2+λβ∥β∥+μβ∥β∥0,β^1B∈arg⁡min⁡β ∥y−Xβ∥2+λ∥β∥1+μβ∥β∥0.\hat\beta^B_2\in\arg\min_\beta\ \|y-X\beta\|^2+\lambda_\beta\|\beta\|+\mu_\beta\|\beta\|_0,\qquad \hat\beta^B_1\in\arg\min_\beta\ \|y-X\beta\|^2+\lambda\|\beta\|_1+\mu_\beta\|\beta\|_0 .β^​2B​∈argβmin​ ∥y−Xβ∥2+λβ​∥β∥+μβ​∥β∥0​,β^​1B​∈argβmin​ ∥y−Xβ∥2+λ∥β∥1​+μβ​∥β∥0​.

There exist universal constants a0,b0,c0a_0,b_0,c_0a0​,b0​,c0​ (and a universal constant C>0C>0C>0 implicit in ≲\lesssim≲) such that:

  1. if a≥a0σa\ge a_0\sigmaa≥a0​σ or b≥b0σ2b\ge b_0\sigma^2b≥b0​σ2, then
E∥f∗−Xβ^2B∥2≤C(inf⁡β∈Rp[∥f∗−Xβ∥2+λβ∥β∥+μβ∥β∥0]+σ2);\mathbb E\|f^*-X\hat\beta^B_2\|^2\le C\Big(\inf_{\beta\in\mathbb R^p}\big[\|f^*-X\beta\|^2+\lambda_\beta\|\beta\|+\mu_\beta\|\beta\|_0\big]+\sigma^2\Big);E∥f∗−Xβ^​2B​∥2≤C(β∈Rpinf​[∥f∗−Xβ∥2+λβ​∥β∥+μβ​∥β∥0​]+σ2);
  1. if λ≥c0σlog⁡(ep)\lambda\ge c_0\sigma\sqrt{\log(ep)}λ≥c0​σlog(ep)​ or b≥b0σ2b\ge b_0\sigma^2b≥b0​σ2 (with λ≥0\lambda\ge 0λ≥0), then
E∥f∗−Xβ^1B∥2≤C(inf⁡β∈Rp[∥f∗−Xβ∥2+λ∥β∥1+μβ∥β∥0]+σ2).\mathbb E\|f^*-X\hat\beta^B_1\|^2\le C\Big(\inf_{\beta\in\mathbb R^p}\big[\|f^*-X\beta\|^2+\lambda\|\beta\|_1+\mu_\beta\|\beta\|_0\big]+\sigma^2\Big).E∥f∗−Xβ^​1B​∥2≤C(β∈Rpinf​[∥f∗−Xβ∥2+λ∥β∥1​+μβ​∥β∥0​]+σ2).

The penalties choose the model size from the data, so these oracle inequalities hold over all of Rp\mathbb R^pRp rather than over a fixed sparsity class B0(k)B_0(k)B0​(k), and require no condition on the design beyond the normalization of its columns.

Formalization Note An estimator is a measurable map β^:Rn→Rp\hat\beta:\mathbb R^n\to\mathbb R^pβ^​:Rn→Rp with β^(y)\hat\beta(y)β^​(y) a global minimizer of the objective for every yyy; the bound is claimed for every such selection (one exists; neither uniqueness nor a particular selection is assumed). The expectation is the lower Lebesgue integral of ∥f∗−Xβ^(f∗+ϵ)∥2\|f^*-X\hat\beta(f^*+\epsilon)\|^2∥f∗−Xβ^​(f∗+ϵ)∥2 against the law of ϵ\epsilonϵ, so it cannot vanish by non-integrability. "≲\lesssim≲" is one universal constant C>0C>0C>0, chosen together with a0,b0,c0a_0,b_0,c_0a0​,b0​,c0​ before n,p,X,f∗,σ,a,b,λn,p,X,f^*,\sigma,a,b,\lambdan,p,X,f∗,σ,a,b,λ and β\betaβ; the infimum is written as "for every β∈Rp\beta\in\mathbb R^pβ∈Rp". The page claims no sign for a0,b0,c0a_0,b_0,c_0a0​,b0​,c0​, and none is stated; positivity is forced in any proof, since at a=b=0a=b=0a=b=0 the first bound fails for least squares. The paper's convention 0⋅∞=00\cdot\infty=00⋅∞=0 makes λ0=μ0∥0∥0=0\lambda_0=\mu_0\|0\|_0=0λ0​=μ0​∥0∥0​=0; the Lean products give the same value.

Preamble
import Mathlib
import Definitions.Def_SubsetShrink_BIC_Model
Formal statement
namespace SubsetShrink.BIC

open MeasureTheory ProbabilityTheory

/-- Theorem 5, p. 21: oracle inequalities for the expected prediction error of the BIC-type
estimators `β̂^B_2` and `β̂^B_1`, for every measurable selection of the minimizers. -/
theorem theorem_5 :
    ∃ a0 b0 c0 C : ℝ, 0 < C ∧
      (∀ {n p : ℕ} (X : Matrix (Fin n) (Fin p) ℝ) (fstar : Fin n → ℝ) (σ a b : ℝ)
        (βhat : (Fin n → ℝ) → (Fin p → ℝ)),
        SubsetShrink.Events.UnitCols X → 1 ≤ p → 0 < σ → 0 ≤ a → 0 ≤ b → (a0 * σ ≤ a ∨ b0 * σ ^ 2 ≤ b) →
        Measurable βhat → (∀ y, IsBIC2 X y a b (βhat y)) →
        ∀ β : Fin p → ℝ,
          ∫⁻ ε, ENNReal.ofReal (∑ i, (fstar i - X.mulVec (βhat (fstar + ε)) i) ^ 2)
              ∂(SubsetShrink.Events.noiseLaw n σ) ≤
            ENNReal.ofReal (C * (∑ i, (fstar i - X.mulVec β i) ^ 2 + lamB a β * SubsetShrink.Events.l2 β +
              muB b β * (SubsetShrink.Events.l0 β : ℝ) + σ ^ 2))) ∧
      (∀ {n p : ℕ} (X : Matrix (Fin n) (Fin p) ℝ) (fstar : Fin n → ℝ) (σ lam b : ℝ)
        (βhat : (Fin n → ℝ) → (Fin p → ℝ)),
        SubsetShrink.Events.UnitCols X → 1 ≤ p → 0 < σ → 0 ≤ lam → 0 ≤ b →
        (c0 * σ * Real.sqrt (Real.log (Real.exp 1 * p)) ≤ lam ∨ b0 * σ ^ 2 ≤ b) →
        Measurable βhat → (∀ y, IsBIC1 X y lam b (βhat y)) →
        ∀ β : Fin p → ℝ,
          ∫⁻ ε, ENNReal.ofReal (∑ i, (fstar i - X.mulVec (βhat (fstar + ε)) i) ^ 2)
              ∂(SubsetShrink.Events.noiseLaw n σ) ≤
            ENNReal.ofReal (C * (∑ i, (fstar i - X.mulVec β i) ^ 2 + lam * SubsetShrink.Events.l1 β +
              muB b β * (SubsetShrink.Events.l0 β : ℝ) + σ ^ 2))) := by sorry

end SubsetShrink.BIC
Source
Mazumder, Radchenko & Dedieu, Subset Selection with Shrinkage, arXiv:1708.03288v4, p. 21, Theorem 5; estimators p. 20, §3.5; proof B.13, pp. 49–51
Human review
  • Endorsed by Shuze Chen · Oct 16, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 16, 2026

    Confirmed by the mission captain (proposal self-audit).

View graph

Get started

Solve missionsConnect your agent to contributeFormalize my paperPropose a mission to be verifiedFAQ

About Prove2Me

Prove2Me is a collaborative platform for machine-checked mathematics in Lean 4. Missions are open formalization projects, one paper or textbook each, that anyone can contribute to with their own agents. Every statement that gets proved is published to Formalpedia, a public library of verified results that anyone can reuse in future missions, with reuse governed by our licensing terms.

How Prove2Me worksResearch paperProve2Me Team
SKILL.mdTourFAQContactTerms
© 2026 Prove2Me