The planar Gaussian mixture of a pair of exponential tilts
ProvedBanditAlgorithm.lintegral_gaussian_mixture_exp_tilt_pairreference
The two-dimensional method of mixtures: for and , integrating a pair of independent exponential tilts against the standard Gaussian density of the plane gives
where is the standard Gaussian density.
The statistic of Chernoff's stopping rule compares the threshold against a pair of arms at a time, so the martingale controlling it mixes over a tilt supported on two coordinates, and this is the resulting weight. The two prefactors , are the source of the polynomial term in the threshold of Lattimore--Szepesv'ari Lemma 33.7.
Preamble
import Mathlib.MeasureTheory.Measure.Prod import Mathlib.Analysis.SpecialFunctions.Gaussian.GaussianIntegral open MeasureTheory Real
Formal statement
theorem BanditAlgorithm.lintegral_gaussian_mixture_exp_tilt_pair {s t : ℝ} (hs : 0 ≤ s)
(ht : 0 ≤ t) (y z : ℝ) :
∫⁻ p : ℝ × ℝ, ENNReal.ofReal
(((Real.sqrt (2 * Real.pi))⁻¹ * Real.exp (-(p.1 ^ 2) / 2)
* ((Real.sqrt (2 * Real.pi))⁻¹ * Real.exp (-(p.2 ^ 2) / 2)))
* Real.exp ((p.1 * y - p.1 ^ 2 * s / 2) + (p.2 * z - p.2 ^ 2 * t / 2)))
= ENNReal.ofReal ((Real.sqrt (1 + s))⁻¹ * Real.exp (y ^ 2 / (2 * (1 + s))))
* ENNReal.ofReal ((Real.sqrt (1 + t))⁻¹ * Real.exp (z ^ 2 / (2 * (1 + t)))) := by
sorrySource
Standard Gaussian mixture computation in two coordinates (Kaufmann & Koolen, Mixture martingales revisited, JMLR 22, 2021, Section 3). The two-arm case is what the pairwise statistic of Chernoff's stopping rule requires.