Estimation.MinimaxATE.Achievability
Achievability side of the structure-agnostic program: estimators matching the lower-bound rate.
AIPWEstimator 6 core · 3 supporting This file defines the fixed-center augmented inverse probability weighted estimator for the finite structure-agnostic average treatment effect problem. ★ aipw_bias_identity★ aipw_bias_bound★ aipw_var_bound
Fixed-Center AIPW Achievability
This file defines the fixed-center augmented inverse probability weighted estimator
for the finite structure-agnostic average treatment effect problem. The estimator
estAIPW is the sample average of aipwScoreFin mhat ghat, with the nuisance
center (mhat, ghat) held fixed rather than estimated from the sample.
The main facts are aipw_pop_mean, the finite population mean formula;
aipw_bias_identity and aipw_bias_bound, the doubly robust product-remainder
calculation; exists_center_overlap, which extracts a finite positive overlap
constant for the fixed center; and aipwScore_bound/aipw_var_bound, the score
and sample-average variance bounds used by Optimality.lean to prove the
miss-probability achievability theorem.
The outcome Y : Bool read as a real number {0,1}.
The fixed-center augmented inverse-propensity score combines the fitted treated-versus-control outcome-regression contrast with the residual from the observed treatment arm, weighted by the fitted propensity at the covariate value.
The fixed-center AIPW estimator: the sample average of the AIPW score.
Definition (Lean source)
Doubly-robust bias identity. Assume the fitted propensity mhat takes values strictly between 0 and 1 at every covariate value. Then the population mean of the AIPW score under the true data-generating process (m, g), minus the true average treatment effect, equals a finite doubly-robust remainder built from cell-by-cell products of the propensity error m − mhat and the outcome-regression errors g − ghat on each treatment arm.
Formal statement
Proof (Lean source)
Doubly-robust bias bound. Given a positive slack ε such that the fitted propensity mhat stays at least ε away from both 0 and 1 at every covariate value — the center has uniform overlap, the absolute bias of the population AIPW mean relative to the true ATE is at most ε⁻¹ times the product of the combined treated/control L²(P_X) outcome-regression error and the L²(P_X) propensity error.
Formal statement
Proof (Lean source)
Variance bound. Suppose the true data-generating process (m, g) is valid, the fixed nuisance center (mhat, ghat) is itself a valid data-generating process, and that center has uniform overlap: the fitted propensity mhat stays at least a positive constant ε away from both 0 and 1 at every covariate value. Then on an n-observation i.i.d. sample from the true single-observation law, the variance of the fixed-center AIPW sample-average estimator is at most (1 + 2/ε)²/n.
Formal statement
Proof (Lean source)
3 supporting declarations (lemmas, instances)
-
aipw_pop_meantheorem — Population mean of the AIPW score under the DGP (m, g) is the finite sum Σ_z obsReal m g z · ψ_AIPW(z; m̂, ĝ).hypothesesconclusion∫ z, aipwScoreFin mhat ghat z ∂(obsLaw hv)Proof (Lean source)
theorem aipw_pop_mean [Nonempty C] {m : C → ℝ} {g : Bool → C → ℝ} (hv : ValidDGP m g) (mhat : C → ℝ) (ghat : Bool → C → ℝ) : ∫ z, aipwScoreFin mhat ghat z ∂(obsLaw hv) = ∑ z : Obs C, obsReal m g z * aipwScoreFin mhat ghat z := by unfold obsLaw rw [PMF.integral_eq_sum] refine Finset.sum_congr rfl fun z _ => ?_ rw [smul_eq_mul] congr 1 unfold obsPMF rw [PMF.ofFintype_apply, ENNReal.toReal_ofReal (obsReal_nonneg hv z)] -
exists_center_overlaptheorem — Uniform lower bound on the center weights. With mhat strictly inside (0,1) on the finite type C, there is ε > 0 with ε ≤ mhat x and ε ≤ 1 − mhat x for all x.hypothesesNonempty Cmhat :C → ℝhmhat :∀ x, 0 < mhat xhmhat1 :∀ x, mhat x < 1conclusion∃ ε > 0, ∀ x, ε ≤ mhat x ∧ ε ≤ 1 - mhat xProof (Lean source)
theorem exists_center_overlap [Nonempty C] (mhat : C → ℝ) (hmhat : ∀ x, 0 < mhat x) (hmhat1 : ∀ x, mhat x < 1) : ∃ ε > 0, ∀ x, ε ≤ mhat x ∧ ε ≤ 1 - mhat x := by classical refine ⟨Finset.univ.inf' Finset.univ_nonempty (fun x => min (mhat x) (1 - mhat x)), ?_, ?_⟩ · rw [gt_iff_lt, Finset.lt_inf'_iff] intro x _ exact lt_min (hmhat x) (by linarith [hmhat1 x]) · intro x have h := Finset.inf'_le (s := (Finset.univ : Finset C)) (fun x => min (mhat x) (1 - mhat x)) (Finset.mem_univ x) exact ⟨le_trans h (min_le_left _ _), le_trans h (min_le_right _ _)⟩ -
aipwScore_boundtheorem — Score bound. The AIPW score is bounded by B = 1 + 2/ε whenever the center nuisances are valid and mhat is ε-bounded off {0,1}.hypothesesNonempty Cmhat :C → ℝghat :Bool → C → ℝhghat :ValidDGP mhat ghatε :ℝhε :0 < εhco :∀ x, ε ≤ mhat x ∧ ε ≤ 1 - mhat xz :Obs Cconclusion|aipwScoreFin mhat ghat z| ≤ 1 + 2 / εProof (Lean source)
theorem aipwScore_bound [Nonempty C] (mhat : C → ℝ) (ghat : Bool → C → ℝ) (hghat : ValidDGP mhat ghat) {ε : ℝ} (hε : 0 < ε) (hco : ∀ x, ε ≤ mhat x ∧ ε ≤ 1 - mhat x) (z : Obs C) : |aipwScoreFin mhat ghat z| ≤ 1 + 2 / ε := by obtain ⟨x, d, y⟩ := z obtain ⟨hco1, hco2⟩ := hco x obtain ⟨hg1a, hg1b⟩ := hghat.g_mem true x obtain ⟨hg0a, hg0b⟩ := hghat.g_mem false x have hya : (0:ℝ) ≤ yReal y := by unfold yReal; cases y <;> simp have hyb : yReal y ≤ 1 := by unfold yReal; cases y <;> simp -- bounded-ratio helper: `|(u−v)/w| ≤ 1/ε` for `u,v ∈ [0,1]`, `w ≥ ε > 0`. have hratio : ∀ (u v w : ℝ), 0 ≤ u → u ≤ 1 → 0 ≤ v → v ≤ 1 → ε ≤ w → |(u - v) / w| ≤ 1 / ε := by intro u v w hu hu1 hv hv1 hw have hw0 : 0 < w := lt_of_lt_of_le hε hw rw [abs_div, abs_of_pos hw0, div_le_div_iff₀ hw0 hε] have huv : |u - v| ≤ 1 := by rw [abs_le]; constructor <;> linarith nlinarith [mul_le_mul_of_nonneg_right huv hε.le, hw] have hA : |ghat true x - ghat false x| ≤ 1 := by rw [abs_le]; constructor <;> linarith have hB : |cond d ((yReal y - ghat true x) / mhat x) (-(yReal y - ghat false x) / (1 - mhat x))| ≤ 1 / ε := by cases d · change |(-(yReal y - ghat false x)) / (1 - mhat x)| ≤ 1 / ε rw [neg_div, abs_neg] exact hratio (yReal y) (ghat false x) (1 - mhat x) hya hyb hg0a hg0b hco2 · change |(yReal y - ghat true x) / mhat x| ≤ 1 / ε exact hratio (yReal y) (ghat true x) (mhat x) hya hyb hg1a hg1b hco1 unfold aipwScoreFin calc |(ghat true x - ghat false x) + cond d _ _| ≤ |ghat true x - ghat false x| + |cond d _ _| := abs_add_le _ _ _ ≤ 1 + 1 / ε := add_le_add hA hB _ ≤ 1 + 2 / ε := by gcongr; norm_num