BanditAlgorithm.isotropic_gaussian_quadratic_mixture
Disprovedbandit-algorithmsconcentration-inequalitiesgaussianprobability
Let and . Prove that there exists an isotropic Gaussian probability measure on such that, for every vector and every positive-semidefinite matrix , the function
is integrable with respect to and
Equivalently, is the centered Gaussian distribution with precision matrix . This is the Gaussian completion-of-the-square identity used in the method of mixtures.
Source: Tor Lattimore and Csaba Szepesvári, Bandit Algorithms (Cambridge University Press, 2020), Lemma 20.3 and equation (20.8), p. 259.
Preamble
import Mathlib.Probability.Distributions.Gaussian.Multivariate import Mathlib.Analysis.SpecialFunctions.Log.Basic import Mathlib.LinearAlgebra.Matrix.PosDef open MeasureTheory ProbabilityTheory Matrix
Formal statement
theorem BanditAlgorithm.isotropic_gaussian_quadratic_mixture
{d : ℕ} {lam : ℝ} (hlam : 0 < lam) :
∃ h : Measure (Fin d → ℝ),
IsProbabilityMeasure h ∧
∀ (S : Fin d → ℝ) (V : Matrix (Fin d) (Fin d) ℝ),
V.PosSemidef →
Integrable
(fun x => Real.exp
(x ⬝ᵥ S -
1 / 2 * (x ⬝ᵥ (lam • (1 : Matrix (Fin d) (Fin d) ℝ) + V) *ᵥ x))) h ∧
(∫ x, Real.exp
(x ⬝ᵥ S -
1 / 2 * (x ⬝ᵥ (lam • (1 : Matrix (Fin d) (Fin d) ℝ) + V) *ᵥ x)) ∂h) =
Real.exp
(1 / 2 *
(S ⬝ᵥ (lam • (1 : Matrix (Fin d) (Fin d) ℝ) + V)⁻¹ *ᵥ S -
Real.log
((lam • (1 : Matrix (Fin d) (Fin d) ℝ) + V).det / lam ^ d))) := by
sorry
Source
Tor Lattimore and Csaba Szepesvári, Bandit Algorithms, Cambridge University Press, 2020, p. 259, Lemma 20.3 and Eq. (20.8).