§5·4, p. 29 — the M-dimensional form of Anscombe's Theorem B
ProvedSmithRegenerative.CLT.theorem_B_vectorLet and be as in Theorem B: is an unbounded, non-decreasing real function, takes values in the positive integers, and in probability. Let
be independent, identically distributed random vectors in with square-integrable components, for all , and covariance matrix . Then the random vector
is asymptotically normally distributed, with mean and covariance matrix , as .
This vector form of Anscombe's theorem is what Smith uses to prove the joint asymptotic normality of several cumulative processes (Theorem 10).
Formalization Note Vectors are elements of the Euclidean space . "Asymptotically normally distributed" is convergence in distribution (weak convergence of laws) to the centred Gaussian law . The matrix is defined as the covariance matrix of , hence positive semidefinite; the limit may be degenerate and no positivity is assumed. No independence between and the , nor between the components of , is assumed.
import Mathlib import Definitions.Def_SmithRegenerative_CLT_CumulativeProcess open MeasureTheory ProbabilityTheory Filter Topology
namespace SmithRegenerative.CLT
/-- Smith (1955), §5·4, p. 29 (unnumbered): the `M`-dimensional form of Anscombe's Theorem B.
Let `y_j = (y_j^{(1)}, …, y_j^{(M)})` (`j = 1, 2, …`) be independent, identically distributed
random vectors with `E y_j^{(l)} = 0` and `cov(y_j^{(k)}, y_j^{(l)}) = a_kl` (5·4·3)–(5·4·4), and let
`ψ`, `m_t` be as in Theorem B. Then "the random vector
`(1/√{ψ(t)}){Σ₁^{m_t} y_j^{(1)}, …, Σ₁^{m_t} y_j^{(M)}}` is asymptotically normally distributed,
with covariance matrix `a_kl = cov y_j^{(k)}, y_j^{(l)}`."
**Formalization Note** Vectors live in `EuclideanSpace ℝ (Fin M)`. "Asymptotically normally
distributed with covariance matrix `a`" is convergence in distribution, along real `t → ∞`, to the
centred Gaussian law `multivariateGaussian 0 a` (written as `TendstoInDistribution` with limit
variable `id` on that law). `a` is defined as the covariance matrix of `y_1`, so it is positive
semidefinite and Mathlib's `multivariateGaussian 0 a` is the genuine Gaussian law with covariance
`a` (possibly degenerate). No independence between `m_t` and the `y_j`, and none between the
components of `y_j`, is assumed. `y 0` is unused. -/
theorem theorem_B_vector {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P]
{M : ℕ} (ψ : ℝ → ℝ) (hψ_mono : Monotone ψ) (hψ_unbdd : ¬ BddAbove (Set.range ψ))
(m : ℝ → Ω → ℕ) (hm_meas : ∀ t, Measurable (m t)) (hm_pos : ∀ t ω, 1 ≤ m t ω)
(hm_ratio : TendstoInMeasure P (fun t ω => (m t ω : ℝ) / ψ t) atTop (fun _ => 1))
(y : ℕ → Ω → EuclideanSpace ℝ (Fin M)) (hy_indep : iIndepFun (fun j => y (j + 1)) P)
(hy_ident : ∀ j, IdentDistrib (y (j + 1)) (y 1) P P)
(hy_L2 : ∀ l, MemLp (fun ω => y 1 ω l) 2 P) (hy_mean : ∀ l, ∫ ω, y 1 ω l ∂P = 0) :
TendstoInDistribution
(fun (t : ℝ) (ω : Ω) => (Real.sqrt (ψ t))⁻¹ • ∑ j ∈ Finset.Icc 1 (m t ω), y j ω)
atTop id (fun _ => P)
(multivariateGaussian 0 (covMatrix P (fun l ω => y 1 ω l))) := by sorry
end SmithRegenerative.CLT
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.