The Gaussian mixture of an exponential tilt, in the extended nonnegative reals
ProvedBanditAlgorithm.lintegral_gaussian_mixture_exp_tiltreference
The method of mixtures, in the -valued form that Tonelli's theorem applies to: for and ,
The Bochner-integral version is the same computation, but every downstream use sits inside an interchange with the trajectory measure of a bandit, and Tonelli's theorem needs the integrand to be -valued. Transporting between the two costs the integrability of the Gaussian with a linear term. Applied to the bandit exponential martingale with and , this converts a family of fixed-tilt martingales into a single martingale whose exponent is the self-normalised deviation .
Preamble
import Mathlib.MeasureTheory.Measure.Prod import Mathlib.Analysis.SpecialFunctions.Gaussian.GaussianIntegral open MeasureTheory Real
Formal statement
theorem BanditAlgorithm.lintegral_gaussian_mixture_exp_tilt {t : ℝ} (ht : 0 ≤ t) (z : ℝ) :
∫⁻ lam : ℝ, ENNReal.ofReal ((Real.sqrt (2 * Real.pi))⁻¹ * Real.exp (-(lam ^ 2) / 2)
* Real.exp (lam * z - lam ^ 2 * t / 2))
= ENNReal.ofReal ((Real.sqrt (1 + t))⁻¹ * Real.exp (z ^ 2 / (2 * (1 + t)))) := by
sorrySource
Standard Gaussian mixture computation (Robbins & Siegmund 1970; Kaufmann & Koolen, JMLR 22, 2021, Section 3). Stated in the extended nonnegative reals so that Tonelli's theorem applies inside the bandit trajectory integral.