Stat.MEstimation
M- and Z-estimation: consistency, asymptotic normality, and influence-function representations.
FinitePoisson 4 core · 6 supporting This module defines a finite Poisson criterion composed with a linear design and proves its continuity, existence and uniqueness of the pseudo-true maximizer, and the equivalence between global optimality and vanishing d ★ finitePoissonObjective_exists_unique_max
Finite positive-mean Poisson projections
This module defines a finite Poisson criterion composed with a linear design and proves its continuity, existence and uniqueness of the pseudo-true maximizer, and the equivalence between global optimality and vanishing directional scores.
A point is the unique global maximizer of a real-valued criterion.
Definition (Lean source)
A total choice of a global maximizer, defaulting to zero when none exists.
Definition (Lean source)
A finite positive-mean Poisson criterion composed with a linear design.
Definition (Lean source)
Existence and uniqueness of the finite Poisson pseudo-true parameter. If every cell weight is strictly positive, every cell mean is strictly positive, and the linear design map is injective, then the finite Poisson criterion attains its supremum over the parameter space at a unique point.
Formal statement
Proof (Lean source)
6 supporting declarations (lemmas, instances)
-
poissonCell_linear_coercive_boundlemma — A positive-mean Poisson cell log likelihood is bounded above by a linearly coercive function. This is the elementary tail estimate used in the finite Poisson maximizer existence argument.Proof (Lean source)
lemma poissonCell_linear_coercive_bound (m x : ℝ) (hm : 0 < m) : m * x - exp x ≤ (m + 1) ^ 2 / 2 - min m 1 * |x| := by have hc0 : 0 < min m 1 := lt_min hm zero_lt_one by_cases hx : x ≤ 0 · rw [abs_of_nonpos hx] have he : 0 ≤ exp x := (Real.exp_pos x).le have hc : min m 1 ≤ m := min_le_left _ _ have hK : 0 ≤ (m + 1) ^ 2 / 2 := div_nonneg (sq_nonneg _) (by norm_num) nlinarith · have hx0 : 0 ≤ x := le_of_not_ge hx have he := Real.pow_div_factorial_le_exp x hx0 2 norm_num [factorial] at he have hc : min m 1 ≤ 1 := min_le_right _ _ have hs : 0 ≤ (x - (m + 1)) ^ 2 := sq_nonneg _ rw [abs_of_nonneg hx0] nlinarith -
poissonCell_strictConcave_midpointlemma — Strict midpoint concavity of a positive-weight Poisson cell whenever the two linear predictors differ.hypothesesq m x y :ℝhq :0 < qhxy :x ≠ yProof (Lean source)
lemma poissonCell_strictConcave_midpoint (q m x y : ℝ) (hq : 0 < q) (hxy : x ≠ y) : (q * (m * x - exp x) + q * (m * y - exp y)) / 2 < q * (m * ((x + y) / 2) - exp ((x + y) / 2)) := by have hconv := strictConvexOn_exp.2 (Set.mem_univ x) (Set.mem_univ y) hxy (show 0 < (1 / 2 : ℝ) by norm_num) (show 0 < (1 / 2 : ℝ) by norm_num) (by norm_num : (1 / 2 : ℝ) + 1 / 2 = 1) simp only [smul_eq_mul] at hconv have := mul_lt_mul_of_pos_left hconv hq have harg : (1 / 2 : ℝ) * x + (1 / 2 : ℝ) * y = (x + y) / 2 := by ring rw [harg] at this calc _ = q * (m * ((x + y) / 2) - (exp x + exp y) / 2) := by ring _ < q * (m * ((x + y) / 2) - exp ((x + y) / 2)) := by nlinarith -
finitePoissonObjective_continuouslemma — The finite Poisson criterion is continuous in its parameter whenever the linear design acts on a finite-dimensional normed space.hypothesesE I :q m :I → ℝA :E →ₗ[ℝ] (I → ℝ)conclusionContinuous (finitePoissonObjective q m A)Proof (Lean source)
lemma finitePoissonObjective_continuous {E I : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E] [Fintype I] (q m : I → ℝ) (A : E →ₗ[ℝ] (I → ℝ)) : Continuous (finitePoissonObjective q m A) := by unfold finitePoissonObjective fun_prop -
finitePoissonObjective_scorelemma — Every directional score vanishes at a global maximizer.hypothesesE I :q m :I → ℝA :E →ₗ[ℝ] (I → ℝ)x d :Ehx :∀ y, finitePoissonObjective q m A y ≤ finitePoissonObjective q m A xconclusion∑ i, q i * A d i * (m i - exp (A x i)) = 0Proof (Lean source)
lemma finitePoissonObjective_score {E I : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [Fintype I] (q m : I → ℝ) (A : E →ₗ[ℝ] (I → ℝ)) (x d : E) (hx : ∀ y, finitePoissonObjective q m A y ≤ finitePoissonObjective q m A x) : ∑ i, q i * A d i * (m i - exp (A x i)) = 0 := by let path : ℝ → E := fun s => x + s • d have heta (i : I) : HasDerivAt (fun s => A (path s) i) (A d i) 0 := by have hfun : (fun s : ℝ => A (path s) i) = fun s : ℝ => A x i + s * A d i := by funext s simp [path, map_add, map_smul, smul_eq_mul] rw [hfun] have h : HasDerivAt (fun s : ℝ => A x i + s * A d i) (0 + 1 * A d i) 0 := (hasDerivAt_const (𝕜 := ℝ) 0 (A x i)).fun_add ((hasDerivAt_id (𝕜 := ℝ) 0).mul_const (A d i)) simpa using h have hderiv : HasDerivAt (fun s => finitePoissonObjective q m A (path s)) (∑ i, q i * A d i * (m i - exp (A x i))) 0 := by have hraw : HasDerivAt (fun s : ℝ => ∑ i, q i * (m i * A (path s) i - exp (A (path s) i))) (∑ i, q i * (m i * A d i - exp (A (path 0) i) * A d i)) 0 := HasDerivAt.fun_sum (u := univ) (fun i _ => (((heta i).const_mul (m i)).sub (heta i).exp).const_mul (q i)) have hpath0 : path 0 = x := by simp [path] have hval : (∑ i, q i * (m i * A d i - exp (A (path 0) i) * A d i)) = ∑ i, q i * A d i * (m i - exp (A x i)) := by rw [hpath0] exact Finset.sum_congr rfl fun i _ => by ring have hfe : (fun s : ℝ => finitePoissonObjective q m A (path s)) = fun s : ℝ => ∑ i, q i * (m i * A (path s) i - exp (A (path s) i)) := rfl rw [hfe, ← hval] exact hraw have hlocal : IsLocalMax (fun s => finitePoissonObjective q m A (path s)) 0 := Filter.Eventually.of_forall fun s => by simpa [path] using hx (path s) exact hlocal.hasDerivAt_eq_zero hderiv -
finitePoissonObjective_isMax_of_scorelemma — For a nonnegative weighted Poisson objective, vanishing of every directional score is sufficient for global maximality.hypothesesE I :q m :I → ℝA :E →ₗ[ℝ] (I → ℝ)x :Ehq :∀ i, 0 ≤ q ihscore :∀ d : E, ∑ i, q i * A d i * (m i - exp (A x i)) = 0conclusion∀ y, finitePoissonObjective q m A y ≤ finitePoissonObjective q m A xProof (Lean source)
lemma finitePoissonObjective_isMax_of_score {E I : Type*} [AddCommGroup E] [Module ℝ E] [Fintype I] (q m : I → ℝ) (A : E →ₗ[ℝ] (I → ℝ)) (x : E) (hq : ∀ i, 0 ≤ q i) (hscore : ∀ d : E, ∑ i, q i * A d i * (m i - exp (A x i)) = 0) : ∀ y, finitePoissonObjective q m A y ≤ finitePoissonObjective q m A x := by intro y have hterm (i : I) : q i * (m i * A y i - exp (A y i)) - q i * (m i * A x i - exp (A x i)) ≤ q i * A (y - x) i * (m i - exp (A x i)) := by have he : 1 + (A y i - A x i) ≤ exp (A y i - A x i) := by simpa [add_comm] using Real.add_one_le_exp (A y i - A x i) have hexp : exp (A x i) * (1 + (A y i - A x i)) ≤ exp (A y i) := by calc exp (A x i) * (1 + (A y i - A x i)) ≤ exp (A x i) * exp (A y i - A x i) := mul_le_mul_of_nonneg_left he (Real.exp_pos _).le _ = exp (A y i) := by rw [← Real.exp_add] congr 1 ring have hqexp := mul_le_mul_of_nonneg_left hexp (hq i) simp only [map_sub, Pi.sub_apply] nlinarith unfold finitePoissonObjective have hsum := Finset.sum_le_sum fun i (_ : i ∈ univ) => hterm i rw [hscore (y - x), Finset.sum_sub_distrib] at hsum linarith -
uniqueGlobalMax_maximizerOrZerolemma — A criterion with a unique global maximum is maximized uniquely by its total maximizer selector.hypothesesconclusionProof (Lean source)
lemma uniqueGlobalMax_maximizerOrZero {E : Type*} [Zero E] (f : E → ℝ) (h : ∃! x : E, ∀ y, f y ≤ f x) : IsUniqueGlobalMax f (maximizerOrZero f) := by obtain ⟨x, hx, huniq⟩ := h have hex : ∃ z, ∀ y, f y ≤ f z := ⟨x, hx⟩ have hsel : maximizerOrZero f = x := by rw [maximizerOrZero, dif_pos hex] exact huniq _ (Classical.choose_spec hex) rw [hsel] exact ⟨hx, fun y hy => huniq y (fun z => (hx z).trans_eq hy.symm)⟩
ZEstimator 1 core · 0 supporting This module records the public regularity bundle ZEstimatorRegularity for parametric Z-estimator and M-estimator central limit theorems. ★ ZEstimatorRegularity
Z-estimator regularity
This module records the public regularity bundle ZEstimatorRegularity for
parametric Z-estimator and M-estimator central limit theorems. The structure
collects population identification, derivative invertibility, finite variance,
measurability, local integrability, continuity of the population Jacobian, and
an integrable L2 score envelope used by the empirical-expansion and
asymptotic-linearity layers.
Regularity conditions for the Z-estimator central limit theorem. Bundles, for a score function and target parameter θ₀ under a sampling law, the population identification condition that the score vanishes in mean at the truth, a Jacobian of the population score at θ₀ together with a witnessed inverse, finite variance of the score at the truth, measurability of the score at every parameter value, continuity of the population score map at θ₀, local integrability of the score on a neighborhood of θ₀, and an integrable L² envelope bounding local score differences.
Definition (Lean source)
ArgmaxStability 1 core · 0 supporting This module proves convergence of exact maximizers from uniform convergence on an eventually common compact set and uniqueness of the limiting maximizer. ★ tendsto_argmax_of_eventually_mem_compact
Deterministic argmax stability
This module proves convergence of exact maximizers from uniform convergence on an eventually common compact set and uniqueness of the limiting maximizer.
Deterministic argmax stability. Let criterion : ℕ → E → ℝ be a sequence of objective functions with pointwise limit limitCriterion, over a compact set K. Suppose the limiting maximizer limitArgmax lies in K, limitCriterion is continuous on K, limitArgmax maximizes limitCriterion over K, and it is the unique maximizer over K, up to equal criterion values. If criterion converges to limitCriterion uniformly on K, each argmax N eventually lies in K, and argmax N eventually maximizes criterion N over K, then the sequence argmax converges to limitArgmax.
Formal statement
Proof (Lean source)
EmpiricalExpansion 1 core · 3 supporting This module supplies the empirical-process and population-expansion layer used by zEstimator_clt. ★ localStochasticExpansion
Empirical expansion for Z-estimators
This module supplies the empirical-process and population-expansion layer used
by zEstimator_clt. It proves the mean-square score control
score_diff_L2_isLittleOp_sqrt, packages local stochastic equicontinuity as
StochEquicontAt, derives empiricalScoreDiff_isLittleOp_sqrt from that
package and consistency, restates the population derivative as
populationScoreDiff_eq_jacobian_plus_remainder, and combines the pieces in
localStochasticExpansion.
Local stochastic expansion for Z-estimators. For a criterion function ψ with population Fréchet-Jacobian reg.J₀ at θ₀, an i.i.d. sample S, and an estimator sequence θn, suppose θn is consistent for θ₀, the centered empirical process of ψ is stochastically equicontinuous at θ₀ along θn, and θn approaches θ₀ at the parametric rate, ‖θn − θ₀‖ = O_p(1/√n). Then the empirical average score difference (1/√n) Σᵢ (ψ(θn,Zᵢ) − ψ(θ₀,Zᵢ)) agrees with its linearization √n · reg.J₀(θn − θ₀) up to an o_p(1) remainder in norm.
Formal statement
Proof (Lean source)
3 supporting declarations (lemmas, instances)
-
score_diff_L2_isLittleOp_sqrttheorem — (a) Score difference is o_p(1) in mean square.hypothesesψ :E → X → Eθ₀ :EP :Measure Xreg :ZEstimatorRegularity ψ θ₀ PS :IIDSample Ω X μ Pθn :ℕ → Ω → EconclusionIsLittleOp (fun n ω => (n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖^2) (fun _ => (1 : ℝ)) μProof (Lean source)
theorem score_diff_L2_isLittleOp_sqrt (ψ : E → X → E) (θ₀ : E) (P : Measure X) (reg : ZEstimatorRegularity ψ θ₀ P) [IsProbabilityMeasure μ] (S : IIDSample Ω X μ P) (θn : ℕ → Ω → E) (hConsistent : ∀ ε > 0, Tendsto (fun n => μ {ω | ε < ‖θn n ω - θ₀‖}) atTop (𝓝 0)) : IsLittleOp (fun n ω => (n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖^2) (fun _ => (1 : ℝ)) μ := by intro ε hε rw [ENNReal.tendsto_nhds_zero] intro γ hγ by_cases hγtop : γ = ⊤ · filter_upwards with n simp [hγtop] have hγpos : 0 < γ.toReal := ENNReal.toReal_pos (ne_of_gt hγ) hγtop let α : ℝ := γ.toReal / 4 have hαpos : 0 < α := by dsimp [α] linarith rcases reg.score_envelope with ⟨δenv, hδenv, F, hFmeas, hFnonneg, hFint, hFbound⟩ let K : ℝ := ∫ z, F z ^ 2 ∂P have hK_nonneg : 0 ≤ K := by dsimp [K] exact integral_nonneg fun z => sq_nonneg (F z) let M : ℝ := (K + 1) / α have hK1pos : 0 < K + 1 := by linarith have hMpos : 0 < M := div_pos hK1pos hαpos have hKM_nonneg : 0 ≤ K / M := div_nonneg hK_nonneg (le_of_lt hMpos) have hKM_le : K / M ≤ α := by dsimp [M] field_simp [hαpos.ne', hK1pos.ne'] nlinarith [hK_nonneg, hαpos] let τ : ℝ := min δenv (sqrt (ε / (M + 1))) have hM1pos : 0 < M + 1 := by linarith have hτpos : 0 < τ := by dsimp [τ] exact lt_min hδenv (Real.sqrt_pos.mpr (div_pos hε hM1pos)) have hτhalf_pos : 0 < τ / 2 := by linarith have hτ_le_env : τ ≤ δenv := by dsimp [τ] exact min_le_left _ _ have hτ_le_sqrt : τ ≤ sqrt (ε / (M + 1)) := by dsimp [τ] exact min_le_right _ _ have hτsq_le : τ ^ 2 ≤ ε / (M + 1) := by have hs := pow_le_pow_left₀ (le_of_lt hτpos) hτ_le_sqrt 2 simpa [Real.sq_sqrt (div_nonneg (le_of_lt hε) (le_of_lt hM1pos))] using hs have hτsqM_lt : τ ^ 2 * M < ε := by have hmul := mul_le_mul_of_nonneg_right hτsq_le (le_of_lt hMpos) have hfrac : ε / (M + 1) * M < ε := by rw [div_mul_eq_mul_div] rw [div_lt_iff₀ hM1pos] nlinarith [hε, hMpos] exact lt_of_le_of_lt hmul hfrac let Mn : ℕ → Ω → ℝ := fun n ω => (n : ℝ)⁻¹ * ∑ i ∈ range n, F (S.Z i ω) ^ 2 have hFpow_meas : Measurable fun z => F z ^ 2 := hFmeas.pow_const 2 have hmap_i : ∀ i, μ.map (S.Z i) = P := by intro i calc μ.map (S.Z i) = μ.map (S.Z 0) := (S.identDist i).map_eq.symm _ = P := S.law have hint_i : ∀ i, Integrable (fun ω => F (S.Z i ω) ^ 2) μ := by intro i have hi : Integrable (fun z => F z ^ 2) (μ.map (S.Z i)) := by simpa [hmap_i i] using hFint simpa [Function.comp_def] using hi.comp_measurable (S.meas i) have hint_eq_i : ∀ i, (∫ ω, F (S.Z i ω) ^ 2 ∂μ) = K := by intro i calc (∫ ω, F (S.Z i ω) ^ 2 ∂μ) = ∫ z, F z ^ 2 ∂(μ.map (S.Z i)) := by exact (integral_map (S.meas i).aemeasurable hFpow_meas.aestronglyMeasurable).symm _ = K := by rw [hmap_i i] have hMn_nonneg_point : ∀ n ω, 0 ≤ Mn n ω := by intro n ω dsimp [Mn] exact mul_nonneg (inv_nonneg.mpr (Nat.cast_nonneg n)) (sum_nonneg fun i _hi => sq_nonneg (F (S.Z i ω))) have hMn_nonneg : ∀ n, 0 ≤ᵐ[μ] Mn n := by intro n exact Eventually.of_forall (hMn_nonneg_point n) have hMn_int : ∀ n, Integrable (Mn n) μ := by intro n dsimp [Mn] exact (integrable_finset_sum (range n) (fun i _hi => hint_i i)).const_mul _ have hMn_integral_le : ∀ n, (∫ ω, Mn n ω ∂μ) ≤ K := by intro n by_cases hn : n = 0 · subst n dsimp [Mn, K] simpa using hK_nonneg · have hnpos : 0 < n := Nat.pos_of_ne_zero hn have hnR : (n : ℝ) ≠ 0 := by exact_mod_cast hn dsimp [Mn] rw [integral_const_mul] rw [MeasureTheory.integral_finset_sum (range n)] · simp_rw [hint_eq_i] simp [Finset.sum_const, hnR] · intro i _hi exact hint_i i have hMn_markov : ∀ n, μ {ω | M < Mn n ω} ≤ ofReal (K / M) := by intro n have hge_real : (μ {ω | M ≤ Mn n ω}).toReal ≤ K / M := by have hmark := MeasureTheory.mul_meas_ge_le_integral_of_nonneg (hMn_nonneg n) (hMn_int n) M have hdiv : μ.real {ω | M ≤ Mn n ω} ≤ (∫ ω, Mn n ω ∂μ) / M := by rw [measureReal_def] exact (le_div_iff₀ hMpos).mpr (by simpa [measureReal_def, mul_comm] using hmark) exact le_trans hdiv (div_le_div_of_nonneg_right (hMn_integral_le n) (le_of_lt hMpos)) exact le_trans (measure_mono (by intro ω hω change M ≤ Mn n ω exact le_of_lt hω)) ((ENNReal.le_ofReal_iff_toReal_le (measure_ne_top μ _) hKM_nonneg).mpr hge_real) have hCons_le := (ENNReal.tendsto_nhds_zero.mp (hConsistent (τ / 2) hτhalf_pos)) (ofReal α) (ENNReal.ofReal_pos.mpr hαpos) have htwo_alpha_lt_gamma : ofReal (2 * α) < γ := by rw [ENNReal.ofReal_lt_iff_lt_toReal] · dsimp [α] linarith · dsimp [α] linarith [le_of_lt hγpos] · exact hγtop filter_upwards [hCons_le] with n hnCons exact le_of_lt <| calc μ {ω | ε * (fun _ => (1 : ℝ)) n < |(n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2|} ≤ μ ({ω | τ / 2 < ‖θn n ω - θ₀‖} ∪ {ω | M < Mn n ω}) := by apply measure_mono_ae have hFbound_i : ∀ i : ℕ, ∀ᵐ ω ∂μ, ∀ θ : E, ‖θ - θ₀‖ < δenv → ‖ψ θ (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ≤ ‖θ - θ₀‖ * F (S.Z i ω) := by intro i have hbound_map : ∀ᵐ z ∂(μ.map (S.Z i)), ∀ θ : E, ‖θ - θ₀‖ < δenv → ‖ψ θ z - ψ θ₀ z‖ ≤ ‖θ - θ₀‖ * F z := by simpa [hmap_i i] using hFbound exact ae_of_ae_map (S.meas i).aemeasurable hbound_map have hFbound_all : ∀ᵐ ω ∂μ, ∀ i ∈ range n, ∀ θ : E, ‖θ - θ₀‖ < δenv → ‖ψ θ (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ≤ ‖θ - θ₀‖ * F (S.Z i ω) := by rw [Finset.eventually_all] intro i _hi exact hFbound_i i filter_upwards [hFbound_all] with ω hFboundω hω replace hω : ε * (fun _ => (1 : ℝ)) n < |(n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2| := hω by_cases hn0 : n = 0 · subst n have hω0 : ε < 0 := by simpa using hω exact elim (not_lt_of_ge hε.le hω0) · have hnR_nonneg : 0 ≤ (n : ℝ)⁻¹ := inv_nonneg.mpr (Nat.cast_nonneg n) have hT_nonneg : 0 ≤ (n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2 := by exact mul_nonneg hnR_nonneg (sum_nonneg fun i _hi => sq_nonneg _) have hTgt : ε < (n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2 := by have hω' : ε < |(n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2| := by simpa using hω simpa [abs_of_nonneg hT_nonneg] using hω' by_cases hθsmall : ‖θn n ω - θ₀‖ ≤ τ / 2 · by_cases hMsmall : Mn n ω ≤ M · have hθ_lt_env : ‖θn n ω - θ₀‖ < δenv := by have hθ_lt_τ : ‖θn n ω - θ₀‖ < τ := by linarith [hθsmall, hτpos] exact lt_of_lt_of_le hθ_lt_τ hτ_le_env have hsum_le : (∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2) ≤ ‖θn n ω - θ₀‖ ^ 2 * ∑ i ∈ range n, F (S.Z i ω) ^ 2 := by rw [Finset.mul_sum] refine Finset.sum_le_sum ?_ intro i _hi have henv := hFboundω i _hi (θn n ω) hθ_lt_env have hrhs_nonneg : 0 ≤ ‖θn n ω - θ₀‖ * F (S.Z i ω) := mul_nonneg (norm_nonneg _) (hFnonneg _) have hsquare : ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2 ≤ (‖θn n ω - θ₀‖ * F (S.Z i ω)) ^ 2 := by have hnorm_nonneg : 0 ≤ ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ := norm_nonneg _ exact sq_le_sq' (by linarith) henv simpa [mul_pow] using hsquare have hT_le : (n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2 ≤ ‖θn n ω - θ₀‖ ^ 2 * Mn n ω := by have hmul := mul_le_mul_of_nonneg_left hsum_le hnR_nonneg dsimp [Mn] nlinarith have hθsq_le : ‖θn n ω - θ₀‖ ^ 2 ≤ τ ^ 2 := by exact pow_le_pow_left₀ (norm_nonneg _) (by linarith) 2 have hprod_le : ‖θn n ω - θ₀‖ ^ 2 * Mn n ω ≤ τ ^ 2 * M := by have hMn_nonnegω : 0 ≤ Mn n ω := hMn_nonneg_point n ω exact mul_le_mul hθsq_le hMsmall hMn_nonnegω (sq_nonneg τ) have hT_lt : (n : ℝ)⁻¹ * ∑ i ∈ range n, ‖ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω)‖ ^ 2 < ε := lt_of_le_of_lt (le_trans hT_le hprod_le) hτsqM_lt exact elim (not_lt_of_ge hT_lt.le hTgt) · exact inr (lt_of_not_ge hMsmall) · exact inl (lt_of_not_ge hθsmall) _ ≤ μ {ω | τ / 2 < ‖θn n ω - θ₀‖} + μ {ω | M < Mn n ω} := MeasureTheory.measure_union_le _ _ _ ≤ ofReal α + ofReal (K / M) := add_le_add hnCons (hMn_markov n) _ ≤ ofReal α + ofReal α := add_le_add le_rfl (ENNReal.ofReal_le_ofReal hKM_le) _ = ofReal (2 * α) := by rw [← ENNReal.ofReal_add] · congr 1 -- … truncated; follow the source link for the rest … -
empiricalScoreDiff_isLittleOp_sqrttheorem — (b) Empirical-process score difference is o_p(√n).hypothesesψ :E → X → Eθ₀ :EP :Measure X_reg :ZEstimatorRegularity ψ θ₀ PS :IIDSample Ω X μ Pθn :ℕ → Ω → EhStochEquicont :StochEquicontAt ψ θ₀ P μ S θnconclusionIsLittleOp (fun n ω => ‖(sqrt (n : ℝ))⁻¹ • (∑ i ∈ range n, (ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω))) - sqrt (n : ℝ) • ∫ z, (ψ (θn n ω) z - ψ θ₀ z) ∂P‖) (fun _ => (1 : ℝ)) μProof (Lean source)
theorem empiricalScoreDiff_isLittleOp_sqrt (ψ : E → X → E) (θ₀ : E) (P : Measure X) (_reg : ZEstimatorRegularity ψ θ₀ P) [IsProbabilityMeasure μ] (S : IIDSample Ω X μ P) (θn : ℕ → Ω → E) (hConsistent : ∀ ε > 0, Tendsto (fun n => μ {ω | ε < ‖θn n ω - θ₀‖}) atTop (𝓝 0)) (hStochEquicont : StochEquicontAt ψ θ₀ P μ S θn) : IsLittleOp (fun n ω => ‖(sqrt (n : ℝ))⁻¹ • (∑ i ∈ range n, (ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω))) - sqrt (n : ℝ) • ∫ z, (ψ (θn n ω) z - ψ θ₀ z) ∂P‖) (fun _ => (1 : ℝ)) μ := by intro ε hε rw [ENNReal.tendsto_nhds_zero] intro γ hγ by_cases hγtop : γ = ⊤ · filter_upwards with n simp [hγtop] have hγpos : 0 < γ.toReal := ENNReal.toReal_pos (ne_of_gt hγ) hγtop let α : ℝ := γ.toReal / 4 have hαpos : 0 < α := by dsimp [α] linarith rcases hStochEquicont ε hε with ⟨δ, hδ_pos, hStoch⟩ have hδhalf_pos : 0 < δ / 2 := by linarith have hStoch_le := (ENNReal.tendsto_nhds_zero.mp hStoch) (ofReal α) (ENNReal.ofReal_pos.mpr hαpos) have hCons_le := (ENNReal.tendsto_nhds_zero.mp (hConsistent (δ / 2) hδhalf_pos)) (ofReal α) (ENNReal.ofReal_pos.mpr hαpos) have htwo_alpha_lt_gamma : ofReal (2 * α) < γ := by rw [ENNReal.ofReal_lt_iff_lt_toReal] · dsimp [α] linarith · dsimp [α] linarith [le_of_lt hγpos] · exact hγtop filter_upwards [hStoch_le, hCons_le] with n hnStoch hnCons exact le_of_lt <| calc μ {ω | ε * (fun _ => (1 : ℝ)) n < |‖(sqrt (n : ℝ))⁻¹ • (∑ i ∈ range n, (ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω))) - sqrt (n : ℝ) • ∫ z, (ψ (θn n ω) z - ψ θ₀ z) ∂P‖|} ≤ μ ({ω | ‖θn n ω - θ₀‖ < δ ∧ ε < ‖(sqrt (n : ℝ))⁻¹ • (∑ i ∈ range n, (ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω))) - sqrt (n : ℝ) • ∫ z, (ψ (θn n ω) z - ψ θ₀ z) ∂P‖} ∪ {ω | δ / 2 < ‖θn n ω - θ₀‖}) := by apply measure_mono intro ω hω have hR : ε < ‖(sqrt (n : ℝ))⁻¹ • (∑ i ∈ range n, (ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω))) - sqrt (n : ℝ) • ∫ z, (ψ (θn n ω) z - ψ θ₀ z) ∂P‖ := by simpa [abs_of_nonneg (norm_nonneg _)] using hω by_cases hθ : ‖θn n ω - θ₀‖ < δ · exact inl ⟨hθ, hR⟩ · exact inr (by have hδle : δ ≤ ‖θn n ω - θ₀‖ := le_of_not_gt hθ have hhalf_lt_delta : δ / 2 < δ := by linarith [hδ_pos] exact lt_of_lt_of_le hhalf_lt_delta hδle) _ ≤ μ {ω | ‖θn n ω - θ₀‖ < δ ∧ ε < ‖(sqrt (n : ℝ))⁻¹ • (∑ i ∈ range n, (ψ (θn n ω) (S.Z i ω) - ψ θ₀ (S.Z i ω))) - sqrt (n : ℝ) • ∫ z, (ψ (θn n ω) z - ψ θ₀ z) ∂P‖} + μ {ω | δ / 2 < ‖θn n ω - θ₀‖} := MeasureTheory.measure_union_le _ _ _ ≤ ofReal α + ofReal α := add_le_add hnStoch hnCons _ = ofReal (2 * α) := by rw [← ENNReal.ofReal_add] · congr 1 ring · linarith · linarith _ < γ := htwo_alpha_lt_gamma -
populationScoreDiff_eq_jacobian_plus_remaindertheorem — (c) Population Fréchet expansion of the score difference.hypothesesconclusion(fun θ => ∫ z, (ψ θ z - ψ θ₀ z) ∂P - reg.J₀ (θ - θ₀)) =o[𝓝 θ₀] fun θ => ‖θ - θ₀‖Proof (Lean source)
theorem populationScoreDiff_eq_jacobian_plus_remainder (ψ : E → X → E) (θ₀ : E) (P : Measure X) (reg : ZEstimatorRegularity ψ θ₀ P) : (fun θ => ∫ z, (ψ θ z - ψ θ₀ z) ∂P - reg.J₀ (θ - θ₀)) =o[𝓝 θ₀] fun θ => ‖θ - θ₀‖ := by rcases reg.psi_int_neighborhood with ⟨δ, hδ_pos, hδ_int⟩ have hψ₀ : Integrable (ψ θ₀) P := by exact hδ_int θ₀ (by simpa using hδ_pos) have hbase : (fun θ => (∫ z, ψ θ z ∂P) - (∫ z, ψ θ₀ z ∂P) - reg.J₀ (θ - θ₀)) =o[𝓝 θ₀] fun θ => θ - θ₀ := reg.J₀_spec.isLittleO have hbase_norm : (fun θ => (∫ z, ψ θ z ∂P) - (∫ z, ψ θ₀ z ∂P) - reg.J₀ (θ - θ₀)) =o[𝓝 θ₀] fun θ => ‖θ - θ₀‖ := (Asymptotics.isLittleO_norm_right).mpr hbase refine hbase_norm.congr' ?_ EventuallyEq.rfl have hnear : ∀ᶠ θ in 𝓝 θ₀, ‖θ - θ₀‖ < δ := by refine Metric.eventually_nhds_iff.mpr ⟨δ, hδ_pos, ?_⟩ intro θ hθ simpa [dist_eq_norm] using hθ filter_upwards [hnear] with θ hθ rw [MeasureTheory.integral_sub (hδ_int θ (by simpa [dist_eq_norm] using hθ)) hψ₀]
ExtremumConsistency 3 core · 1 supporting This module packages primitive extremum-estimation conditions that imply the consistency assumptions used by Z-estimator and GMM central-limit theorems. ★ zEstimator_clt_of_extremum★ zEstimator_clt_of_donsker★ zEstimator_clt_of_extremum_donsker
Extremum consistency for M-estimation
This module packages primitive extremum-estimation conditions that imply the
consistency assumptions used by Z-estimator and GMM central-limit theorems. The
format bridge consistent_lt_norm_of_le_dist adapts the empirical-process
consistency theorem to the CLT layer, while zEstimator_clt_of_extremum,
zEstimator_clt_of_donsker, and zEstimator_clt_of_extremum_donsker discharge
opaque consistency and equicontinuity hypotheses from Glivenko-Cantelli,
well-separated-optimum, and class-level equicontinuity inputs. (The parallel GMM
reductions live in Causalean/Stat/GMM/AsymptoticNormality.lean.)
Z-estimator CLT from extremum primitives. Identical conclusion to zEstimator_clt, but with consistency derived rather than assumed: given an auxiliary criterion function m, suppose the criterion class indexed by the parameter obeys a uniform law of large numbers, i.e. is Glivenko–Cantelli, the estimator sample-maximises the empirical criterion at every sample size, and the population criterion has a well-separated maximum at the target parameter. Then, provided the score process is stochastically equicontinuous at the target parameter along the estimator sequence, the estimator converges to the target at the parametric -rate, and the estimator solves the empirical estimating equation eventually, almost surely, the estimator is asymptotically linear at the target parameter, with influence function minus the inverse Jacobian applied to the score at the target.
Formal statement
Proof (Lean source)
Z-estimator CLT with the equicontinuity hypothesis discharged. Identical conclusion to zEstimator_clt, but the estimator-sequence-specific equicontinuity modulus is replaced by a class-level Donsker condition: if the estimator converges in probability to the target parameter, the score family is asymptotically equicontinuous at the target parameter — a property of the family alone, independent of the particular estimator sequence, the estimator converges to the target at the parametric -rate, and the estimator solves the empirical estimating equation eventually, almost surely, then the estimator is asymptotically linear at the target parameter, with influence function minus the inverse Jacobian applied to the score at the target.
Formal statement
Proof (Lean source)
Z-estimator CLT from primitive conditions: both opaque hypotheses discharged. Combines zEstimator_clt_of_extremum and zEstimator_clt_of_donsker: if the criterion class indexed by the parameter obeys a uniform law of large numbers, the estimator sample-maximises the empirical criterion at every sample size, the population criterion has a well-separated maximum at the target parameter, the score family is asymptotically equicontinuous at the target parameter, the estimator converges to the target at the parametric -rate, and the estimator solves the empirical estimating equation eventually, almost surely, then the estimator is asymptotically linear at the target parameter, with influence function minus the inverse Jacobian applied to the score at the target.
Formal statement
Proof (Lean source)
1 supporting declaration (lemmas, instances)
-
consistent_lt_norm_of_le_disttheorem — Format bridge. On a normed group dist x θ₀ = ‖x − θ₀‖, so the consistency statement produced by mEstimator_consistent_of_glivenkoCantelli (phrased with ε ≤ dist (θn n) θ₀) implies the strictly-larger-radius form ε < ‖θn n − θ₀‖ consumed by the CLT layer. {ε < ‖·‖} ⊆ {ε ≤ dist}, so the measures are squeezed to 0.hypothesesE :Type*θn :ℕ → Ω → Eθ₀ :EProof (Lean source)
theorem consistent_lt_norm_of_le_dist {E : Type*} [NormedAddCommGroup E] (θn : ℕ → Ω → E) (θ₀ : E) (h : ∀ ε : ℝ, 0 < ε → Tendsto (fun n => μ {ω | ε ≤ dist (θn n ω) θ₀}) atTop (𝓝 0)) : ∀ ε > 0, Tendsto (fun n => μ {ω | ε < ‖θn n ω - θ₀‖}) atTop (𝓝 0) := by intro ε hε refine tendsto_of_tendsto_of_tendsto_of_le_of_le' tendsto_const_nhds (h ε hε) (Eventually.of_forall fun n => zero_le) (Eventually.of_forall fun n => measure_mono ?_) intro ω hω simp only [Set.mem_setOf_eq, dist_eq_norm] at hω ⊢ exact le_of_lt hω
FinitePoissonConsistency 2 core · 2 supporting This module proves uniform convergence of finite Poisson objectives, compact containment and convergence of their maximizers, and continuity of the selected maximizer under perturbations of the finite mean vector. ★ finitePoissonObjective_argmax_tendsto★ finitePoissonObjective_argmax_continuousAt_mean
Consistency and continuity of finite Poisson projections
This module proves uniform convergence of finite Poisson objectives, compact containment and convergence of their maximizers, and continuity of the selected maximizer under perturbations of the finite mean vector.
Stability of finite Poisson maximizers under convergence of weights and means. Given a sequence of finite Poisson criteria with weights qN N and means mN N, suppose the limiting weights are strictly positive, the limiting means are strictly positive, the linear design map is injective, the weights qN N converge cellwise to the limiting weights q, and the means mN N converge cellwise to the limiting means m. If argmax N maximizes the N-th criterion, eventually in N, and limitArgmax is the unique global maximizer of the limiting criterion, then the maximizer sequence argmax converges to limitArgmax.
Formal statement
Proof (Lean source)
Continuity of the finite Poisson maximizer in the means. If every cell weight is strictly positive and the base mean vector m₀ has every entry strictly positive, and the linear design map is injective, then the selected maximizer of the finite Poisson criterion is continuous, as a function of the mean vector, at m₀.
Formal statement
Proof (Lean source)
2 supporting declarations (lemmas, instances)
-
finitePoissonObjective_tendstoUniformlyOnlemma — Pointwise convergence of the finitely many weights and means gives uniform convergence of the finite Poisson objective on every compact parameter set.hypothesesconclusionTendstoUniformlyOn (fun N => finitePoissonObjective (qN N) (mN N) A) (finitePoissonObjective q m A) atTop KProof (Lean source)
lemma finitePoissonObjective_tendstoUniformlyOn {E I : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E] [Fintype I] (qN mN : ℕ → I → ℝ) (q m : I → ℝ) (A : E →ₗ[ℝ] (I → ℝ)) (K : Set E) (hK : IsCompact K) (hq : ∀ i, Tendsto (fun N => qN N i) atTop (nhds (q i))) (hm : ∀ i, Tendsto (fun N => mN N i) atTop (nhds (m i))) : TendstoUniformlyOn (fun N => finitePoissonObjective (qN N) (mN N) A) (finitePoissonObjective q m A) atTop K := by let cN : ℕ → (I → ℝ) × (I → ℝ) := fun N => (qN N, mN N) let c : (I → ℝ) × (I → ℝ) := (q, m) let F : ((I → ℝ) × (I → ℝ)) → E → ℝ := fun p x => finitePoissonObjective p.1 p.2 A x have hc : Tendsto cN atTop (nhds c) := by exact (tendsto_pi_nhds.mpr hq).prodMk_nhds (tendsto_pi_nhds.mpr hm) let U := closedBall c 1 letI : ProperSpace ((I → ℝ) × (I → ℝ)) := FiniteDimensional.proper_real _ have hU : IsCompact U := ProperSpace.isCompact_closedBall c 1 have hcU : c ∈ U := by simp [U] have hevent : ∀ᶠ N in atTop, cN N ∈ U := by have := (Metric.tendsto_nhds.mp hc 1 zero_lt_one) filter_upwards [this] with N hN exact Metric.mem_closedBall.mpr hN.le have hcWithin : Tendsto cN atTop (nhdsWithin c U) := by rw [nhdsWithin] exact tendsto_inf.2 ⟨hc, tendsto_principal.2 hevent⟩ have hcont : Continuous (uncurry F) := by unfold F finitePoissonObjective fun_prop have hUC : UniformContinuousOn (uncurry F) (U ×ˢ K) := (hU.prod hK).uniformContinuousOn_of_continuous hcont.continuousOn have hlocal : TendstoUniformlyOn F (F c) (nhdsWithin c U) K := hUC.tendstoUniformlyOn hcU intro V hV exact hcWithin.eventually (hlocal V hV) -
finitePoissonObjective_eventually_common_compactlemma — Convergent positive coefficients and a fixed injective design put any eventual sequence of global maximizers in one common compact ball.hypothesesE I :qN mN :ℕ → I → ℝq m :I → ℝA :E →ₗ[ℝ] (I → ℝ)argmax :ℕ → Ehqpos :∀ i, 0 < q ihmpos :∀ i, 0 < m ihA :hmax :∀ᶠ N in atTop,∀ y,finitePoissonObjective (qN N) (mN N) A y≤ finitePoissonObjective (qN N) (mN N) A (argmax N)Proof (Lean source)
lemma finitePoissonObjective_eventually_common_compact {E I : Type*} [NormedAddCommGroup E] [NormedSpace ℝ E] [FiniteDimensional ℝ E] [Fintype I] (qN mN : ℕ → I → ℝ) (q m : I → ℝ) (A : E →ₗ[ℝ] (I → ℝ)) (argmax : ℕ → E) (hqpos : ∀ i, 0 < q i) (hmpos : ∀ i, 0 < m i) (hA : Injective A) (hq : ∀ i, Tendsto (fun N => qN N i) atTop (nhds (q i))) (hm : ∀ i, Tendsto (fun N => mN N i) atTop (nhds (m i))) (hmax : ∀ᶠ N in atTop, ∀ y, finitePoissonObjective (qN N) (mN N) A y ≤ finitePoissonObjective (qN N) (mN N) A (argmax N)) : ∃ K : Set E, IsCompact K ∧ ∀ᶠ N in atTop, argmax N ∈ K := by classical let w : I → ℝ := fun i => q i / 2 * min (m i / 2) 1 let B : E →ₗ[ℝ] (I → ℝ) := { toFun := fun x i => w i * A x i map_add' := by intro x y; funext i; simp; ring map_smul' := by intro c x; funext i; simp; ring } have hwpos : ∀ i, 0 < w i := by intro i exact mul_pos (half_pos (hqpos i)) (lt_min (half_pos (hmpos i)) zero_lt_one) have hB : Injective B := by intro x y hxy apply hA funext i have hi := congrFun hxy i dsimp [B] at hi exact (mul_left_cancel₀ (ne_of_gt (hwpos i)) hi) obtain ⟨L, hL, hanti⟩ := (LinearMap.injective_iff_antilipschitz B).mp hB let kappa : ℝ := L⁻¹ have hkappa : 0 < kappa := inv_pos.mpr hL have hBnorm (x : E) : kappa * ‖x‖ ≤ ‖B x‖ := by have h := hanti x 0 have hh := ENNReal.toReal_mono (by finiteness) h have h' : ‖x‖ ≤ (L : ℝ) * ‖B x‖ := by simpa [edist_dist, dist_eq_norm, ENNReal.toReal_mul] using hh change (L : ℝ)⁻¹ * ‖x‖ ≤ ‖B x‖ calc (L : ℝ)⁻¹ * ‖x‖ ≤ (L : ℝ)⁻¹ * ((L : ℝ) * ‖B x‖) := mul_le_mul_of_nonneg_left h' (inv_nonneg.mpr hL.le) _ = ((L : ℝ)⁻¹ * (L : ℝ)) * ‖B x‖ := by ring _ = ‖B x‖ := by rw [inv_mul_cancel₀ (show (L : ℝ) ≠ 0 by exact_mod_cast ne_of_gt hL)] simp have hbounds : ∀ᶠ N in atTop, ∀ i, q i / 2 < qN N i ∧ qN N i < 2 * q i ∧ m i / 2 < mN N i ∧ mN N i < 2 * m i := by rw [Filter.eventually_all] intro i filter_upwards [ (tendsto_order.1 (hq i)).1 (q i / 2) (by linarith [hqpos i]), (tendsto_order.1 (hq i)).2 (2 * q i) (by linarith [hqpos i]), (tendsto_order.1 (hm i)).1 (m i / 2) (by linarith [hmpos i]), (tendsto_order.1 (hm i)).2 (2 * m i) (by linarith [hmpos i])] with N hqlo hqhi hmlo hmhi exact ⟨hqlo, hqhi, hmlo, hmhi⟩ let M : ℝ := ∑ i, (2 * q i) * ((2 * m i + 1) ^ 2 / 2) let Q : ℝ := ∑ i, 2 * q i have hM : 0 ≤ M := sum_nonneg fun i _ => mul_nonneg (mul_nonneg (by norm_num) (hqpos i).le) (div_nonneg (sq_nonneg _) (by norm_num)) have hQ : 0 ≤ Q := sum_nonneg fun i _ => mul_nonneg (by norm_num) (hqpos i).le let R : ℝ := (M + Q) / kappa have hR : 0 ≤ R := div_nonneg (add_nonneg hM hQ) hkappa.le let K : Set E := closedBall 0 R letI : ProperSpace E := proper_real E refine ⟨K, ProperSpace.isCompact_closedBall 0 R, ?_⟩ filter_upwards [hbounds, hmax] with N hb hmaxN have htail (x : E) : finitePoissonObjective (qN N) (mN N) A x ≤ M - ‖B x‖ := by have hterm : ∀ i, qN N i * (mN N i * A x i - exp (A x i)) ≤ (2 * q i) * ((2 * m i + 1) ^ 2 / 2) - |B x i| := by intro i have hc := poissonCell_linear_coercive_bound (mN N i) (A x i) (lt_trans (half_pos (hmpos i)) (hb i).2.2.1) have hqn : 0 ≤ qN N i := (lt_trans (half_pos (hqpos i)) (hb i).1).le have hlower : w i ≤ qN N i * min (mN N i) 1 := by exact mul_le_mul (le_of_lt (hb i).1) (min_le_min (le_of_lt (hb i).2.2.1) le_rfl) (le_min (half_pos (hmpos i)).le zero_le_one) hqn have habs : |B x i| = w i * |A x i| := by dsimp [B] rw [abs_mul, abs_of_pos (hwpos i)] rw [habs] have hcoarse := mul_le_mul_of_nonneg_left hc hqn have hconst : qN N i * ((mN N i + 1) ^ 2 / 2) ≤ (2 * q i) * ((2 * m i + 1) ^ 2 / 2) := by have hmn0 : 0 ≤ mN N i := (lt_trans (half_pos (hmpos i)) (hb i).2.2.1).le have hmupper := (hb i).2.2.2 have hsq : (mN N i + 1) ^ 2 ≤ (2 * m i + 1) ^ 2 := by nlinarith exact mul_le_mul (le_of_lt (hb i).2.1) (div_le_div_of_nonneg_right hsq (by norm_num)) (div_nonneg (sq_nonneg _) (by norm_num)) (mul_nonneg (by norm_num) (hqpos i).le) nlinarith [mul_le_mul_of_nonneg_right hlower (abs_nonneg (A x i))] calc finitePoissonObjective (qN N) (mN N) A x ≤ ∑ i, ((2 * q i) * ((2 * m i + 1) ^ 2 / 2) - |B x i|) := Finset.sum_le_sum fun i _ => hterm i _ = M - ∑ i, |B x i| := by rw [Finset.sum_sub_distrib] _ ≤ M - ‖B x‖ := by gcongr rw [Pi.norm_def] have hnn : (Finset.univ.sup fun i => ‖B x i‖₊) ≤ ∑ i, ‖B x i‖₊ := Finset.sup_le fun i _ => Finset.single_le_sum (f := fun j => ‖B x j‖₊) (fun j _ => zero_le) (Finset.mem_univ i) calc ((Finset.univ.sup fun i => ‖B x i‖₊ : NNReal) : ℝ) ≤ ((∑ i, ‖B x i‖₊ : NNReal) : ℝ) := NNReal.coe_le_coe.mpr hnn _ = ∑ i, |B x i| := by simp [Real.norm_eq_abs] have hzero : -Q ≤ finitePoissonObjective (qN N) (mN N) A 0 := by unfold finitePoissonObjective simp only [map_zero, Pi.zero_apply, mul_zero, Real.exp_zero, zero_sub, mul_neg, mul_one, Finset.sum_neg_distrib, neg_le_neg_iff] exact Finset.sum_le_sum fun i _ => (hb i).2.1.le have hchain := (hzero.trans (hmaxN 0)).trans (htail (argmax N)) have hbn := hBnorm (argmax N) have hnorm : ‖argmax N‖ ≤ R := by dsimp [R] apply (le_div_iff₀ hkappa).2 linarith simpa [K, Metric.mem_closedBall, dist_eq_norm] using hnorm
FinitePoissonDerivative 2 core · 0 supporting This module differentiates the unique maximizer of a full-rank finite Poisson criterion when one cell mean follows a positive exponential path. ★ finitePoissonObjective_expCell_argmax_snd_hasDerivAt
One-cell derivatives of finite Poisson projections
This module differentiates the unique maximizer of a full-rank finite Poisson criterion when one cell mean follows a positive exponential path.
Replace one cell mean by a positive exponential path.
Definition (Lean source)
Derivative of the finite Poisson maximizer under one exponentially perturbed cell mean. Suppose every cell weight is strictly positive, every base cell mean is strictly positive, the perturbation scale B is strictly positive, and the linear design map is injective; consider replacing cell j's mean by the exponential path x ↦ B · exp x. If betaDot is the unique value solving the score equation linearized at the maximizer for base point x₀, for every perturbation direction, then the second (scalar) coordinate of the maximizer, as a function of x, has derivative betaDot at x₀.
Formal statement
Proof (Lean source)
FinitePoissonSign 1 core · 0 supporting This module characterizes the sign of one selected coefficient from the score at a conditional nuisance fit with that coefficient fixed at zero. ★ finitePoissonObjective_snd_sign_of_nuisance_score
Conditional sign characterization for finite Poisson projections
This module characterizes the sign of one selected coefficient from the score at a conditional nuisance fit with that coefficient fixed at zero.
Sign of the selected scalar coefficient from the nuisance score. Suppose every cell weight is strictly positive, every cell mean is strictly positive, the linear design map is injective, and the conditional nuisance fit u₀, with the scalar coordinate held at zero, clears every nuisance-direction score. Then the selected scalar coefficient of the finite Poisson maximizer has exactly the sign of the remaining scalar score: it is negative, zero, or positive exactly when the scalar score is.
Formal statement
Proof (Lean source)
InfluenceFunction 1 core · 0 supporting This module defines InfluenceFunction, the basic data-layer predicate for an influence function under a sampling law. ★ InfluenceFunction
Influence-function data layer
This module defines InfluenceFunction, the basic data-layer predicate for an
influence function under a sampling law. It records exactly the public
requirements needed by downstream asymptotic-linearity wrappers: measurability,
mean zero, and an integrable squared norm.
An influence function for a parameter θ ∈ E under the law P : Measure X is a measurable, mean-zero, square-integrable function ψ : X → E.
Definition (Lean source)
ZEstimatorCLT 1 core · 0 supporting This module proves zEstimator_clt, the theorem layer for parametric estimators that solve empirical estimating equations. ★ zEstimator_clt
Z-estimator asymptotic linearity
This module proves zEstimator_clt, the theorem layer for parametric estimators
that solve empirical estimating equations. It connects the local stochastic
expansion from EmpiricalExpansion.lean to an IsAsymLinearVec
influence-function representation with influence function
fun z => -(reg.J₀_inv (ψ θ₀ z)); a separate vector CLT can then turn this
asymptotic-linear representation into a normal limit.
Z-estimator asymptotic linearity. An estimator sequence that solves the empirical estimating equation eventually, almost surely, is consistent at the target parameter, converges to the target at the parametric -rate, and whose score process is stochastically equicontinuous at the target parameter along the estimator sequence, admits the influence-function representation obtained by applying the negative inverse Jacobian to the target score.