The comparison is exact to second order: the Taylor polynomials differ by
ProvedMartingale.taylor_cancellation_cubicAn exact polynomial identity, for real :
The left factor is the third-order Taylor polynomial of ; the right-hand product is times the second-order Taylor polynomial of . Every term of order , and cancels, and what survives is exactly cubic.
This identity is the reason McLeish's comparison works. His proof of the martingale central limit theorem replaces the characteristic function , which cannot be factored because the summands are dependent, by , whose expectation is exactly by the martingale property. The substitution is only useful if the two agree closely enough, and this computation says precisely how closely: the discrepancy between and begins at third order in .
Third order is exactly what the argument needs, and second order would not do. Summing per-factor errors of size gives
because the increments are asymptotically negligible while the sum of squares stays bounded — it converges to . Had the discrepancy been merely quadratic the bound would have been and nothing would vanish. So the exact cancellation through second order recorded here is what makes the whole method work, and it explains why , rather than any other first-order approximation to , is the right comparison factor: it is the one whose product with the Gaussian factor matches to second order.
import Mathlib.Analysis.Complex.Exponential import Mathlib.Analysis.SpecialFunctions.Complex.Circle open Finset
theorem Martingale.taylor_cancellation_cubic (x : ℝ) :
(1 + Complex.I * x - (x:ℂ)^2/2 - Complex.I * (x:ℂ)^3/6)
- (1 + Complex.I * x) * (1 - (x:ℂ)^2/2)
= Complex.I * (x:ℂ)^3 / 3 := by sorry