Gaussian and Rademacher scalar laws; Gaussian-series tail
DefinitionTroppMatrixConcentration_ch4_scalar_lawsFor a real function on a measure space with measure , standardGaussianLaw μ g means its pushforward measure is the real Gaussian probability measure with mean zero and variance one. rademacherLaw μ g means its pushforward is . Measurability is imposed separately in theorem statements. For a natural dimension and real , gaussianSeriesTail d v t is when . At it is if and zero otherwise. Concentration statements use this expression only for and positive dimensions. The piecewise definition avoids totalized division in the zero-series case; it assumes no bound on any random matrix.
import Definitions.Def_TroppMatrixConcentration_probability
import Mathlib.Probability.Distributions.Gaussian.Real
open MeasureTheory ProbabilityTheory
noncomputable section
namespace TroppMatrixConcentration
def standardGaussianLaw {Ω : Type*} [MeasurableSpace Ω]
(μ : Measure Ω) (g : Ω → ℝ) : Prop :=
Measure.map g μ = gaussianReal 0 1
def rademacherLaw {Ω : Type*} [MeasurableSpace Ω]
(μ : Measure Ω) (r : Ω → ℝ) : Prop :=
Measure.map r μ = (1 / 2 : ENNReal) • Measure.dirac (1 : ℝ) +
(1 / 2 : ENNReal) • Measure.dirac (-1 : ℝ)
def gaussianSeriesTail (dimension : ℕ) (v t : ℝ) : ℝ :=
if v = 0 then (if t = 0 then (dimension : ℝ) else 0)
else (dimension : ℝ) * Real.exp (-(t ^ 2) / (2 * v))
end TroppMatrixConcentration
Read-back
What the Lean code literally says, in plain math · gpt-6
TroppMatrixConcentration.standardGaussianLaw: For any type equipped with a measurable space, any measure on , and any function , this predicate means that the pushforward of by equals the Gaussian probability measure on the Borel real line with mean and variance , namely the measure with Lebesgue density . Here pushforward is formed using a measurable representative equal to almost everywhere when such a representative exists, and is defined to be the zero measure otherwise. No probability-measure or measurability assumption is imposed as an input; equality to the specified nonzero probability measure entails that has such a representative and that . The predicate is consequently false when is the zero measure, when is empty, or when is not almost everywhere measurable.
TroppMatrixConcentration.rademacherLaw: For any type equipped with a measurable space, any measure on , and any function , this predicate means that the pushforward of by is on the Borel real line, where is the Dirac probability measure at and both weights are the finite number in the extended nonnegative reals. Thus the specified measure places mass at each of and and no mass elsewhere. Here pushforward is formed using a measurable representative equal to almost everywhere when such a representative exists, and is defined to be the zero measure otherwise. No probability-measure or measurability assumption is imposed as an input; the predicate entails that is almost everywhere measurable and that , and it is false for a zero measure, an empty domain, or a function that is not almost everywhere measurable.
TroppMatrixConcentration.gaussianSeriesTail: For every natural number , including , and every pair of real numbers , this defines a real number equal to if and , equal to if and , and equal to if , with interpreted as a real number. There are no sign restrictions on or : the exponential branch applies also to negative , and both positive and negative are admitted. The value is for every when , and is for every when ; the explicit branch fixes the value there without evaluating the displayed quotient.
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.