Estimation.Orthogonal­Learning

Generic orthogonal statistical-learning substrate: population derivative bundles, cross-fitted plug-in ERM, local moduli, oracle inequalities, and sparse specializations.

Local­Emp­Process 21 core · 6 supporting · 5 submodules Local empirical-process modulus machinery for orthogonal learning, derived from Rademacher, localized, fixed-nuisance, and random-nuisance regimes. Population 10 core · 0 supporting · 3 submodules Population-side orthogonal-learning definitions: directional derivatives, Neyman orthogonality, and nuisance-induced second-order bias. Sparse 11 core · 1 supporting · 3 submodules Sparse finite-dimensional plug-in ERM theory: restricted strong convexity, restricted cones, one-norm geometry, and lasso-style oracle guarantees.
Setup 4 core · 0 supporting This file defines the basic population system for orthogonal statistical learning: a measurable loss, a convex target class in a Hilbert space, an admissible nuisance class, and distinguished target and nuisance elements

Orthogonal Statistical Learning Setup

This file defines the basic population system for orthogonal statistical learning: a measurable loss, a convex target class in a Hilbert space, an admissible nuisance class, and distinguished target and nuisance elements. It is the loss-based analogue of an orthogonal-moment model.

The central structure is LearningSystem. Its namespace provides the population risk LearningSystem.L, and the file also exposes the segment-closure predicates Θ_PerturbClosed and G_PerturbClosed used by later directional-derivative modules.

structure LearningSystem reviewed
Causalean.Estimation.OrthogonalLearning

Orthogonal statistical-learning system. Bundles a convex target class Θ_set containing a distinguished target θ₀, a nuisance class G_set containing a distinguished nuisance g₀, a jointly measurable loss , and the condition that θ₀ minimizes the population risk over the target class at the true nuisance g₀.

Definition (Lean source)
Ω :
μ :
Z :
P_Z :
G :
Type*
Module ℝ G
Θ_set :
Set Θ
Θ_convex :
Convex ℝ Θ_set
θ₀ :
Θ
θ₀_mem :
θ₀ ∈ Θ_set
G_set :
Set G
g₀ :
G
g₀_mem :
g₀ ∈ G_set
:
Z → Θ → G → ℝ
ℓ_meas :
∀ θ g, Measurable (fun z => ℓ z θ g)
θ₀_minimizes :
∀ θ ∈ Θ_set, ∫ z, ℓ z θ₀ g₀ ∂P_Z ≤ ∫ z, ℓ z θ g₀ ∂P_Z
Causalean.Estimation.OrthogonalLearning.LearningSystem · Causalean/Estimation/OrthogonalLearning/Setup.lean:43
def L reviewed
Causalean.Estimation.OrthogonalLearning.LearningSystem

Population risk: L(θ, g) := ∫ ℓ(z; θ, g) dP_Z.

Definition (Lean source)
noncomputable def L (S : LearningSystem Ω μ Z P_Z Θ G) (θ : Θ) (g : G) : ℝ := ∫ z, S.ℓ z θ g ∂P_Z
Causalean.Estimation.OrthogonalLearning.LearningSystem.L · Causalean/Estimation/OrthogonalLearning/Setup.lean:88 · uses LearningSystem
def Θ_PerturbClosed reviewed
Causalean.Estimation.OrthogonalLearning

Closure-under-perturbation predicate for Θ_set: every line segment between θ₀ and θ ∈ Θ_set stays in Θ_set. Strictly weaker than convexity; useful for stating directional-derivative hypotheses without requiring the whole set to be convex.

Definition (Lean source)
def Θ_PerturbClosed (S : LearningSystem Ω μ Z P_Z Θ G) : Prop := ∀ θ ∈ S.Θ_set, ∀ t ∈ Icc (0 : ℝ) 1, S.θ₀ + t • (θ - S.θ₀) ∈ S.Θ_set
Causalean.Estimation.OrthogonalLearning.Θ_PerturbClosed · Causalean/Estimation/OrthogonalLearning/Setup.lean:99 · uses LearningSystem
def G_PerturbClosed reviewed
Causalean.Estimation.OrthogonalLearning

Closure-under-perturbation predicate for G_set.

Definition (Lean source)
def G_PerturbClosed (S : LearningSystem Ω μ Z P_Z Θ G) : Prop := ∀ g ∈ S.G_set, ∀ t ∈ Icc (0 : ℝ) 1, S.g₀ + t • (g - S.g₀) ∈ S.G_set
Causalean.Estimation.OrthogonalLearning.G_PerturbClosed · Causalean/Estimation/OrthogonalLearning/Setup.lean:106 · uses LearningSystem
Plugin­ERM 2 core · 0 supporting This file defines the empirical risk on the estimation fold of a sample split and the predicate that an estimator approximately minimizes that empirical risk after plugging in a nuisance estimate.

Sample-Split Plug-In ERM

This file defines the empirical risk on the estimation fold of a sample split and the predicate that an estimator approximately minimizes that empirical risk after plugging in a nuisance estimate. These objects are the estimation-side inputs to the orthogonal statistical-learning oracle inequality.

The main declarations are empRiskFoldB, the fold-B empirical risk, and SampleSplitPluginERM, the approximate empirical-risk-minimization predicate for a sample-indexed target estimator and plug-in nuisance estimate.

def empRiskFoldB reviewed
Causalean.Estimation.OrthogonalLearning

Empirical risk on the estimation fold B(n): (1 / |B(n)|) · Σ_{i ∈ B(n)} ℓ (Z_i ω) θ g.

Definition (Lean source)
noncomputable def empRiskFoldB (S : LearningSystem Ω μ Z P_Z Θ G) (S_iid : IIDSample Ω Z μ P_Z) (split : OneShotSplit S_iid) (n : ℕ) (ω : Ω) (θ : Θ) (g : G) : ℝ := ((split.foldB n).card : ℝ)⁻¹ * ∑ i ∈ split.foldB n, S.ℓ (S_iid.Z i ω) θ g
structure SampleSplitPluginERM reviewed
Causalean.Estimation.OrthogonalLearning

This predicate says that a sample-size-indexed target estimator, evaluated against a plug-in nuisance estimate on the estimation fold, is an approximate empirical-risk minimizer: the optimization slack is nonnegative at every sample size, the estimator's value always lies in the target parameter set, and its fold-B empirical risk at the plug-in nuisance is within that slack of the risk at every other point in the target set.

Definition (Lean source)
S :
LearningSystem Ω μ Z P_Z Θ G
S_iid :
IIDSample Ω Z μ P_Z
split :
θhat :
ℕ → Ω → Θ
ghat :
ℕ → Ω → G
r_opt :
ℕ → ℝ
r_opt_nonneg :
∀ n, 0 ≤ r_opt n
mem_Θ_set :
∀ n ω, θhat n ω ∈ S.Θ_set
ε-minimizer form: equivalent to `≤ inf + r_opt` and avoids the conditional-completeness pitfall on `ℝ` (where `⨅` defaults to `0` when the set is empty or unbounded below).
approx_min :
∀ n ω,
∀ θ' ∈ S.Θ_set,
empRiskFoldB S S_iid split n ω (θhat n ω) (ghat n ω)
empRiskFoldB S S_iid split n ω θ' (ghat n ω) + r_opt n
Causalean.Estimation.OrthogonalLearning.SampleSplitPluginERM · Causalean/Estimation/OrthogonalLearning/PluginERM.lean:55 · uses LearningSystem , IIDSample , OneShotSplit
Oracle­Inequality 2 core · 0 supporting This file proves deterministic and high-probability conservative oracle inequalities for sample-split plug-in empirical risk minimization with an orthogonal loss. ★ oracle_inequality_plugin_ERM★ oracle_inequality_plugin_ERM_highProb

Orthogonal Statistical Learning Oracle Inequality

This file proves deterministic and high-probability conservative oracle inequalities for sample-split plug-in empirical risk minimization with an orthogonal loss. The bounds convert strong convexity, empirical optimality, a local empirical-process modulus, and second-order bias into squared target-space error control with the displayed 4/σ optimization-slack constant.

The deterministic theorem oracle_inequality_plugin_ERM consumes a realized modulus inequality at one sample point. The high-probability theorem oracle_inequality_plugin_ERM_highProb packages the same bound on the event provided by LocalEmpProcessModulus for a fixed plug-in nuisance.

theorem oracle_inequality_plugin_ERM reviewed
Causalean.Estimation.OrthogonalLearning

Conservative oracle inequality for the orthogonal sample-split plug-in ERM. Assume the realised nuisance at every sample point admits the directional-derivative structure needed for the first-order expansion of the loss, that the estimator is a sample-split plug-in ERM with optimization slack r_opt, and that the strong-convexity modulus σ is strictly positive. Suppose the population risk at the realised nuisance is σ-strongly convex around θ₀ in the chosen norm, and that the first-order inequality holds at the truth (θ₀, g₀). If, at the specific sample point (n, ω), the local empirical-process modulus inequality is realised for the estimator's target value against the realised nuisance, then the squared target-space estimation error is bounded by 4(1+σ)/σ² · (ρ n)² + (4/σ) · Bias_n + (4/σ) · r_opt n.

Formal statement
S :
LearningSystem Ω μ Z P_Z Θ G
S_iid :
IIDSample Ω Z μ P_Z
split :
Dθ_truth :
ghat :
ℕ → Ω → G
Dθ_hat :
∀ n ω, HasDirDerivTheta S (ghat n ω)
θhat :
ℕ → Ω → Θ
r_opt :
ℕ → ℝ
hPluginERM :
SampleSplitPluginERM S S_iid split θhat ghat r_opt
σ :
:
0 < σ
σ-strong convexity of L(·, ghat n ω) around θ₀ w.r.t. ‖·‖.
hSC :
∀ n ω,
∀ θ ∈ S.Θ_set,
S.L θ (ghat n ω) - S.L S.θ₀ (ghat n ω)
≥ (∫ z, (Dθ_hat n ω).dℓ_θ θ z ∂P_Z) + (σ / 2) * ‖θ - S.θ₀‖ ^ 2
First-order inequality at the truth (θ₀, g₀).
hFOI :
ρ :
ℕ → ℝ
n :
ω :
Ω
Modulus inequality realised at the sample point (n, ω) and the realised nuisance ghat n ω. Equivalent (after instantiating the LocalEmpProcessModulus event) to the centred-process bound at the specific θhat n ω.
hRho_holds :
(S.L (θhat n ω) (ghat n ω) - S.L S.θ₀ (ghat n ω))
- (empRiskFoldB S S_iid split n ω (θhat n ω) (ghat n ω) - empRiskFoldB S S_iid split n ω S.θ₀ (ghat n ω))
≤ ρ n * ‖θhat n ω - S.θ₀‖ + (ρ n) ^ 2
‖θhat n ω - S.θ₀‖ ^ 2
≤ (4 * (1 + σ) / σ ^ 2) * (ρ n) ^ 2
+ (4 / σ) * Bias_n S Dθ_truth (Dθ_hat n ω) (θhat n ω)
+ (4 / σ) * r_opt n
Proof (Lean source)
-- TODO(faithfulness): `thm:est-osl-plugin-erm-oracle` displays a `2/σ` -- optimization-slack coefficient; deriving it from the current approximate-ERM -- hypothesis needs an additional sharp basic inequality/slack convention. theorem oracle_inequality_plugin_ERM (S : LearningSystem Ω μ Z P_Z Θ G) (S_iid : IIDSample Ω Z μ P_Z) (split : OneShotSplit S_iid) (Dθ_truth : HasDirDerivTheta S S.g₀) (ghat : ℕ → Ω → G) (Dθ_hat : ∀ n ω, HasDirDerivTheta S (ghat n ω)) (θhat : ℕ → Ω → Θ) (r_opt : ℕ → ℝ) (hPluginERM : SampleSplitPluginERM S S_iid split θhat ghat r_opt) (σ : ℝ) (hσ : 0 < σ) -- σ-strong convexity of L(·, ghat n ω) around θ₀ w.r.t. ‖·‖. (hSC : ∀ n ω, ∀ θ ∈ S.Θ_set, S.L θ (ghat n ω) - S.L S.θ₀ (ghat n ω) ≥ (∫ z, (Dθ_hat n ω).dℓ_θ θ z ∂P_Z) + (σ / 2) * ‖θ - S.θ₀‖ ^ 2) -- First-order inequality at the truth (θ₀, g₀). (hFOI : FirstOrderInequality S Dθ_truth) (ρ : ℕ → ℝ) (n : ℕ) (ω : Ω) -- Modulus inequality realised at the sample point (n, ω) and the -- realised nuisance ghat n ω. Equivalent (after instantiating the -- LocalEmpProcessModulus event) to the centred-process bound at the -- specific θhat n ω. (hRho_holds : (S.L (θhat n ω) (ghat n ω) - S.L S.θ₀ (ghat n ω)) - (empRiskFoldB S S_iid split n ω (θhat n ω) (ghat n ω) - empRiskFoldB S S_iid split n ω S.θ₀ (ghat n ω)) ≤ ρ n * ‖θhat n ω - S.θ₀‖ + (ρ n) ^ 2) : ‖θhat n ω - S.θ₀‖ ^ 2 ≤ (4 * (1 + σ) / σ ^ 2) * (ρ n) ^ 2 + (4 / σ) * Bias_n S Dθ_truth (Dθ_hat n ω) (θhat n ω) + (4 / σ) * r_opt n := by set e : ℝ := ‖θhat n ω - S.θ₀‖ set rho : ℝ := ρ n set ropt : ℝ := r_opt n set B : ℝ := Bias_n S Dθ_truth (Dθ_hat n ω) (θhat n ω) set Dtruth : ℝ := ∫ z, Dθ_truth.dℓ_θ (θhat n ω) z ∂P_Z set Dhat : ℝ := ∫ z, (Dθ_hat n ω).dℓ_θ (θhat n ω) z ∂P_Z set Ldiff : ℝ := S.L (θhat n ω) (ghat n ω) - S.L S.θ₀ (ghat n ω) set Empdiff : ℝ := empRiskFoldB S S_iid split n ω (θhat n ω) (ghat n ω) - empRiskFoldB S S_iid split n ω S.θ₀ (ghat n ω) have hθ_mem : θhat n ω ∈ S.Θ_set := hPluginERM.mem_Θ_set n ω have hθ₀_mem : S.θ₀ ∈ S.Θ_set := S.θ₀_mem have hSC' : Ldiff ≥ Dhat + (σ / 2) * e ^ 2 := by simpa [Ldiff, Dhat, e] using hSC n ω (θhat n ω) hθ_mem have hStrong : (σ / 2) * e ^ 2 ≤ Ldiff - Dhat := by linarith have hFOI' : 0 ≤ Dtruth := by simpa [Dtruth] using hFOI (θhat n ω) hθ_mem have hEmp : Empdiff ≤ ropt := by have happrox := hPluginERM.approx_min n ω S.θ₀ hθ₀_mem simpa [Empdiff, ropt, sub_eq_add_neg, add_comm, add_left_comm, add_assoc] using sub_le_iff_le_add.mpr happrox have hMod : Ldiff - Empdiff ≤ rho * e + rho ^ 2 := by simpa [Ldiff, Empdiff, rho, e] using hRho_holds have hPop : Ldiff ≤ ropt + rho * e + rho ^ 2 := by linarith have hB : B = Dtruth - Dhat := by simp [B, Bias_n, Dtruth, Dhat] have hBasic : (σ / 2) * e ^ 2 ≤ ropt + rho * e + rho ^ 2 + B := by linarith have hYoung : rho * e ≤ (σ / 4) * e ^ 2 + (1 / σ) * rho ^ 2 := by have hσne : σ ≠ 0 := ne_of_gt hσ have hsquare : 0 ≤ (σ / 4) * (e - (2 / σ) * rho) ^ 2 := by exact mul_nonneg (by positivity) (sq_nonneg _) have hident : (σ / 4) * (e - (2 / σ) * rho) ^ 2 = (σ / 4) * e ^ 2 + (1 / σ) * rho ^ 2 - rho * e := by field_simp [hσne] ring linarith have hAbsorb : (σ / 4) * e ^ 2 ≤ ropt + (1 / σ) * rho ^ 2 + rho ^ 2 + B := by linarith calc e ^ 2 = (4 / σ) * ((σ / 4) * e ^ 2) := by field_simp [ne_of_gt hσ] _ ≤ (4 / σ) * (ropt + (1 / σ) * rho ^ 2 + rho ^ 2 + B) := by exact mul_le_mul_of_nonneg_left hAbsorb (by positivity) _ = (4 * (1 + σ) / σ ^ 2) * rho ^ 2 + (4 / σ) * B + (4 / σ) * ropt := by field_simp [ne_of_gt hσ] ring
theorem oracle_inequality_plugin_ERM_highProb reviewed
Causalean.Estimation.OrthogonalLearning

High-probability conservative oracle inequality for the orthogonal sample-split plug-in ERM. Fix a single nuisance value g. Assume the estimator is a sample-split plug-in ERM at this fixed nuisance, with optimization slack r_opt, and that the strong-convexity modulus σ is strictly positive. Suppose the population risk at g is σ-strongly convex around θ₀ in the chosen norm, and that the first-order inequality holds at the truth (θ₀, g₀). If the local empirical-process modulus condition holds at rate ρ and confidence level δ for the fixed nuisance g, then for every sample size n there is an event of probability at least 1 - δ on which the squared target-space estimation error is bounded by 4(1+σ)/σ² · (ρ n)² + (4/σ) · Bias_n + (4/σ) · r_opt n.

Formal statement
S :
LearningSystem Ω μ Z P_Z Θ G
S_iid :
IIDSample Ω Z μ P_Z
split :
Dθ_truth :
g :
G
Dθ_at_g :
θhat :
ℕ → Ω → Θ
r_opt :
ℕ → ℝ
hPluginERM :
SampleSplitPluginERM S S_iid split θhat (fun _ _ => g) r_opt
σ :
:
0 < σ
σ-strong convexity of `L(·, g)` around `θ₀` w.r.t. `‖·‖`.
hSC :
∀ θ ∈ S.Θ_set,
S.L θ g - S.L S.θ₀ g ≥ (∫ z, Dθ_at_g.dℓ_θ θ z ∂P_Z) + (σ / 2) * ‖θ - S.θ₀‖ ^ 2
hFOI :
ρ :
ℕ → ℝ
δ :
hMod :
LocalEmpProcessModulus S S_iid split ρ δ g
∀ n : ℕ,
∃ E : Set Ω,
μ E ≥ 1 - ofReal δ ∧
∀ ω ∈ E,
‖θhat n ω - S.θ₀‖ ^ 2
≤ (4 * (1 + σ) / σ ^ 2) * (ρ n) ^ 2
+ (4 / σ) * Bias_n S Dθ_truth Dθ_at_g (θhat n ω)
+ (4 / σ) * r_opt n
Proof (Lean source)
theorem oracle_inequality_plugin_ERM_highProb (S : LearningSystem Ω μ Z P_Z Θ G) (S_iid : IIDSample Ω Z μ P_Z) (split : OneShotSplit S_iid) (Dθ_truth : HasDirDerivTheta S S.g₀) (g : G) (Dθ_at_g : HasDirDerivTheta S g) (θhat : ℕ → Ω → Θ) (r_opt : ℕ → ℝ) (hPluginERM : SampleSplitPluginERM S S_iid split θhat (fun _ _ => g) r_opt) (σ : ℝ) (hσ : 0 < σ) -- σ-strong convexity of `L(·, g)` around `θ₀` w.r.t. `‖·‖`. (hSC : ∀ θ ∈ S.Θ_set, S.L θ g - S.L S.θ₀ g ≥ (∫ z, Dθ_at_g.dℓ_θ θ z ∂P_Z) + (σ / 2) * ‖θ - S.θ₀‖ ^ 2) (hFOI : FirstOrderInequality S Dθ_truth) (ρ : ℕ → ℝ) (δ : ℝ) (hMod : LocalEmpProcessModulus S S_iid split ρ δ g) : ∀ n : ℕ, ∃ E : Set Ω, MeasurableSet E ∧ μ E ≥ 1 - ofReal δ ∧ ∀ ω ∈ E, ‖θhat n ω - S.θ₀‖ ^ 2 ≤ (4 * (1 + σ) / σ ^ 2) * (ρ n) ^ 2 + (4 / σ) * Bias_n S Dθ_truth Dθ_at_g (θhat n ω) + (4 / σ) * r_opt n := by intro n rcases hMod n with ⟨E, hE_meas, hE_ge, hE_uniform⟩ refine ⟨E, hE_meas, hE_ge, ?_⟩ intro ω hω have hθ_mem : θhat n ω ∈ S.Θ_set := hPluginERM.mem_Θ_set n ω have hRho_holds : (S.L (θhat n ω) ((fun _ _ => g) n ω) - S.L S.θ₀ ((fun _ _ => g) n ω)) - (empRiskFoldB S S_iid split n ω (θhat n ω) ((fun _ _ => g) n ω) - empRiskFoldB S S_iid split n ω S.θ₀ ((fun _ _ => g) n ω)) ≤ ρ n * ‖θhat n ω - S.θ₀‖ + (ρ n) ^ 2 := by simpa using hE_uniform ω hω (θhat n ω) hθ_mem exact oracle_inequality_plugin_ERM S S_iid split Dθ_truth (fun _ _ => g) (fun _ _ => Dθ_at_g) θhat r_opt hPluginERM σ hσ (fun _ _ θ hθ => hSC θ hθ) hFOI ρ n ω hRho_holds