Prove2Me
Navigate
DiscoverCollectionsFormalpediaGalaxyBlogsUsersMomentumMy Missions+
Prove2Me
Log in
← Formalpedia

Theorem 3.1, p. 17 — studentized coordinates and pairwise differences of the Lagrangian debiased estimator converge in distribution to N(0, 1)

Proved
LagrangeRank.Debias.theorem_3_1

by mikedeng1 · Oct 22, 2026 · Mathlib 0df444a (Lean v4.33.1)

asymptotic-normalityconstrained-inferencedebiased-estimatorp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1ranking

Work on one probability space (Ω,F,P)(\Omega,\mathcal F,\mathbb P)(Ω,F,P). For each sample size nnn let dnd_ndn​ be the dimension, Ln(ω,⋅):Rdn→RL_n(\omega,\cdot):\mathbb R^{d_n}\to\mathbb RLn​(ω,⋅):Rdn​→R the random loss, fnf_nfn​ the constraint defining Cn={θ:fn(θ)=0}C_n=\{\theta:f_n(\theta)=0\}Cn​={θ:fn​(θ)=0}, θn∗\theta^*_nθn∗​ the true parameter, θ^n\widehat\theta_nθn​ the initial estimator, Ω^n\widehat\Omega_nΩn​ an estimator of the inverse of E[∇2Ln(θn∗)]\mathbb E[\nabla^2L_n(\theta^*_n)]E[∇2Ln​(θn∗​)] and Ωn∗\Omega^*_nΩn∗​ its (deterministic) target. Let in≠jni_n\neq j_nin​=jn​ be coordinates. Let

θ^d=θ^−Θ^11∇L(θ^)\widehat\theta^d=\widehat\theta-\widehat\Theta_{11}\nabla L(\widehat\theta)θd=θ−Θ11​∇L(θ)

be the Lagrangian debiased estimator (3.9), with Θ^\widehat\ThetaΘ the block-inverse matrix of p. 16 built from (Ω^,∇f(θ^))(\widehat\Omega,\nabla f(\widehat\theta))(Ω,∇f(θ)), and Θ∗\Theta^*Θ∗ the same formula at (Ω∗,∇f(θ∗))(\Omega^*,\nabla f(\theta^*))(Ω∗,∇f(θ∗)). Let Σ11∗=E[∇2L(θ∗)]\Sigma^*_{11}=\mathbb E[\nabla^2L(\theta^*)]Σ11∗​=E[∇2L(θ∗)],

vn=[Θ11∗Σ11∗Θ11∗⊤]jnjn,wn=(ein−ejn)⊤Θ11∗Σ11∗Θ11∗⊤(ein−ejn).v_n=[\Theta^*_{11}\Sigma^*_{11}\Theta^{*\top}_{11}]_{j_nj_n},\qquad w_n=(e_{i_n}-e_{j_n})^\top\Theta^*_{11}\Sigma^*_{11}\Theta^{*\top}_{11}(e_{i_n}-e_{j_n}).vn​=[Θ11∗​Σ11∗​Θ11∗⊤​]jn​jn​​,wn​=(ein​​−ejn​​)⊤Θ11∗​Σ11∗​Θ11∗⊤​(ein​​−ejn​​).

Assume:

  1. Assumptions 3.1–3.4 and the standing assumptions f(θ∗)=0f(\theta^*)=0f(θ∗)=0, f(θ^)=0f(\widehat\theta)=0f(θ)=0 (the structure Assumptions), with fixed constants K,c1,c2,c3,L1,L2K,c_1,c_2,c_3,L_1,L_2K,c1​,c2​,c3​,L1​,L2​;
  2. r1→0r_1\to0r1​→0, r4→0r_4\to0r4​→0 and n(r12r5+(r1+r4)(r12+r2)+r1r3)→0\sqrt n\big(r_1^2r_5+(r_1+r_4)(r_1^2+r_2)+r_1r_3\big)\to0n​(r12​r5​+(r1​+r4​)(r12​+r2​)+r1​r3​)→0;
  3. Assumption 3.5: vn≥Cv_n\ge Cvn​≥C and wn≥Cw_n\ge Cwn​≥C for some C>0C>0C>0 and all nnn, and
n [Θ11∗∇L(θ∗)]jvn→dN(0,1),n [Θ11∗∇L(θ∗)]i−[Θ11∗∇L(θ∗)]jwn→dN(0,1);\sqrt n\,\frac{[\Theta^*_{11}\nabla L(\theta^*)]_{j}}{\sqrt{v_n}}\xrightarrow{d}N(0,1),\qquad \sqrt n\,\frac{[\Theta^*_{11}\nabla L(\theta^*)]_{i}-[\Theta^*_{11}\nabla L(\theta^*)]_{j}}{\sqrt{w_n}}\xrightarrow{d}N(0,1);n​vn​​[Θ11∗​∇L(θ∗)]j​​d​N(0,1),n​wn​​[Θ11∗​∇L(θ∗)]i​−[Θ11∗​∇L(θ∗)]j​​d​N(0,1);
  1. θ^n\widehat\theta_nθn​, the entries of Ω^n\widehat\Omega_nΩn​ and ω↦∇Ln(ω,θ^n(ω))\omega\mapsto\nabla L_n(\omega,\widehat\theta_n(\omega))ω↦∇Ln​(ω,θn​(ω)) are measurable.

Then

n θ^jd−θj∗vn→dN(0,1)andn (θ^id−θi∗)−(θ^jd−θj∗)wn→dN(0,1).\sqrt n\,\frac{\widehat\theta^d_{j}-\theta^*_{j}}{\sqrt{v_n}}\xrightarrow{d}N(0,1)\qquad\text{and}\qquad \sqrt n\,\frac{(\widehat\theta^d_{i}-\theta^*_{i})-(\widehat\theta^d_{j}-\theta^*_{j})}{\sqrt{w_n}}\xrightarrow{d}N(0,1).n​vn​​θjd​−θj∗​​d​N(0,1)andn​wn​​(θid​−θi∗​)−(θjd​−θj∗​)​d​N(0,1).

The theorem is the inferential core of the Lagrangian debiasing framework: it yields confidence intervals for single scores and for pairwise score differences under a general equality constraint, without inverting the (possibly singular) bordered Hessian.

Formalization Note. The paper's ≲\lesssim≲ is read as "at most KKK times the rate on events EnE_nEn​ with P(Enc)→0\mathbb P(E_n^c)\to0P(Enc​)→0", with KKK and c1,c2,c3,L1,L2c_1,c_2,c_3,L_1,L_2c1​,c2​,c3​,L1​,L2​ independent of nnn (this is the paper's "c1,c2,c3,L1,L2=O(1)c_1,c_2,c_3,L_1,L_2=O(1)c1​,c2​,c3​,L1​,L2​=O(1)"). ∥∇f∥∞\|\nabla f\|_\infty∥∇f∥∞​ is read as the ℓ1\ell_1ℓ1​ norm of ∇f\nabla f∇f, as in the proof of Corollary 3.1. The conditions r1,r4→0r_1,r_4\to0r1​,r4​→0 are recalled in the proof (p. 43); f(θ∗)=f(θ^)=0f(\theta^*)=f(\widehat\theta)=0f(θ∗)=f(θ)=0 is the standing assumption θ∗,θ^∈C\theta^*,\widehat\theta\in Cθ∗,θ∈C (p. 14); the variance lower bound is the condition under which Assumption 3.5 is stated; measurability is needed for convergence in distribution to make sense. The sample size, dimension dnd_ndn​ and coordinates in,jni_n,j_nin​,jn​ are sequences; all variables live on one probability space, and convergence in distribution is weak convergence of laws (TendstoInDistribution).

Preamble
import Mathlib
import Definitions.Def_LagrangeRank_Debias_Setting
Formal statement
namespace LagrangeRank.Debias
open MeasureTheory ProbabilityTheory Matrix Filter Topology
theorem theorem_3_1 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω)
    [IsProbabilityMeasure P]
    (d : ℕ → ℕ) (L : (n : ℕ) → Ω → (Fin (d n) → ℝ) → ℝ) (f : (n : ℕ) → (Fin (d n) → ℝ) → ℝ)
    (thetaStar : (n : ℕ) → Fin (d n) → ℝ) (thetaHat : (n : ℕ) → Ω → Fin (d n) → ℝ)
    (OmegaHat : (n : ℕ) → Ω → Matrix (Fin (d n)) (Fin (d n)) ℝ)
    (OmegaStar : (n : ℕ) → Matrix (Fin (d n)) (Fin (d n)) ℝ)
    (r1 r2 r3 r4 r5 : ℕ → ℝ) (K c1 c2 c3 L1 L2 : ℝ) (E : ℕ → Set Ω)
    (i j : (n : ℕ) → Fin (d n)) (hij : ∀ n, i n ≠ j n)
    (hA : Assumptions P d L f thetaStar thetaHat OmegaHat OmegaStar
      r1 r2 r3 r4 r5 K c1 c2 c3 L1 L2 E)
    (hr1 : Tendsto r1 atTop (𝓝 0)) (hr4 : Tendsto r4 atTop (𝓝 0))
    (hscale : Tendsto (fun n : ℕ => Real.sqrt (n : ℝ) *
      (r1 n ^ 2 * r5 n + (r1 n + r4 n) * (r1 n ^ 2 + r2 n) + r1 n * r3 n)) atTop (𝓝 0))
    (hvar : ∃ C : ℝ, 0 < C ∧ ∀ n,
      C ≤ popVar P (L n) (f n) (OmegaStar n) (thetaStar n) (j n) ∧
      C ≤ popVarPair P (L n) (f n) (OmegaStar n) (thetaStar n) (i n) (j n))
    (hclt1 : TendstoInDistribution (fun (n : ℕ) (ω : Ω) => Real.sqrt (n : ℝ) *
        (thetaStar11 (f n) (OmegaStar n) (thetaStar n) *ᵥ grad (L n ω) (thetaStar n)) (j n)
        / Real.sqrt (popVar P (L n) (f n) (OmegaStar n) (thetaStar n) (j n)))
      atTop id (fun _ => P) (gaussianReal 0 1))
    (hclt2 : TendstoInDistribution (fun (n : ℕ) (ω : Ω) => Real.sqrt (n : ℝ) *
        ((thetaStar11 (f n) (OmegaStar n) (thetaStar n) *ᵥ grad (L n ω) (thetaStar n)) (i n)
          - (thetaStar11 (f n) (OmegaStar n) (thetaStar n) *ᵥ grad (L n ω) (thetaStar n)) (j n))
        / Real.sqrt (popVarPair P (L n) (f n) (OmegaStar n) (thetaStar n) (i n) (j n)))
      atTop id (fun _ => P) (gaussianReal 0 1))
    (hmeas_hat : ∀ n, Measurable (thetaHat n))
    (hmeas_Om : ∀ n a b, Measurable (fun ω => OmegaHat n ω a b))
    (hmeas_grad : ∀ n, Measurable (fun ω => grad (L n ω) (thetaHat n ω))) :
    TendstoInDistribution (fun (n : ℕ) (ω : Ω) => Real.sqrt (n : ℝ) *
        (estThetaD (L n ω) (f n) (OmegaHat n ω) (thetaHat n ω) (j n) - thetaStar n (j n))
        / Real.sqrt (popVar P (L n) (f n) (OmegaStar n) (thetaStar n) (j n)))
      atTop id (fun _ => P) (gaussianReal 0 1) ∧
    TendstoInDistribution (fun (n : ℕ) (ω : Ω) => Real.sqrt (n : ℝ) *
        ((estThetaD (L n ω) (f n) (OmegaHat n ω) (thetaHat n ω) (i n) - thetaStar n (i n))
          - (estThetaD (L n ω) (f n) (OmegaHat n ω) (thetaHat n ω) (j n) - thetaStar n (j n)))
        / Real.sqrt (popVarPair P (L n) (f n) (OmegaStar n) (thetaStar n) (i n) (j n)))
      atTop id (fun _ => P) (gaussianReal 0 1) := by sorry
end LagrangeRank.Debias
Source
Liu, Fang, Lu, Lagrangian Inference for Ranking Problems, arXiv:2110.00151v1, p. 17, Theorem 3.1 (with Assumptions 3.1–3.5, pp. 16–17); proof in Appendix A.1, pp. 40–41
Human review
  • Endorsed by Shuze Chen · Oct 23, 2026

    Confirmed by the moderator at approval.

  • Endorsed by mikedeng1 · Oct 23, 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