InformationTheory.Shannon.FisherConvBound
Stam convolution Fisher information bound J(pX ∗ g_s) ≤ 1/s #
For any probability density pX and Gaussian kernel g_s of variance s,
the Fisher information of the convolution pX ∗ g_s satisfies J(pX ∗ g_s) ≤ 1/s.
Main statements #
gaussianConv_fisher_le_inv_var—fisherInfoOfDensity (pX ∗ g_s) ≤ ENNReal.ofReal (1/s).
Implementation notes #
The proof follows a pointwise Cauchy-Schwarz route: the derivative formula
convDensityAdd_deriv1_gaussian_eq gives (deriv p_s x) as an integral, Hölder's
inequality with p = q = 2 bounds (logDeriv p_s x)² · p_s x pointwise, and
Tonelli's theorem plus the Gaussian second moment ∫ u² g_s u du = s closes the bound.
InformationTheory.Shannon.FisherInfo.integral_sq_mul_gaussianPDFReal
source∫ u, u ^ 2 * gaussianPDFReal 0 ⟨s, _⟩ u ∂volume = s.
@audit:ok
Used by
InformationTheory.Shannon.FisherInfo.integrable_sq_mul_gaussianPDFReal
sourceu ↦ u ^ 2 * gaussianPDFReal 0 ⟨s, _⟩ u is integrable over volume.
@audit:ok
Used by
InformationTheory.Shannon.FisherInfo.convSecondMoment_integrand_integrable
sourcey ↦ (x - y) ^ 2 * (pX y * g_s(x - y)) is integrable over volume for each x.
@audit:ok
Used by
InformationTheory.Shannon.FisherInfo.convScore_sq_le_pointwise
sourcePointwise Cauchy-Schwarz: (∫ pX y (x-y) g_s(x-y))² ≤ p_s(x) · ∫ pX y (x-y)² g_s(x-y).
@audit:ok
Used by
InformationTheory.Shannon.FisherInfo.convDensityAdd_deriv_eq
sourcederiv (convDensityAdd pX g_s) x = ∫ y, pX y * g_s(x-y) * (-(x-y)/s) ∂volume.
@audit:ok
Used by
InformationTheory.Shannon.FisherInfo.convLogDeriv_sq_mul_le
source(logDeriv p_s x) ^ 2 * p_s x ≤ (1/s²) * ∫ (x-y)² * (pX y * g_s(x-y)) ∂volume.
@audit:ok
Used by
InformationTheory.Shannon.FisherInfo.gaussianConv_fisher_le_inv_var
sourceStam convolution Fisher information bound: J(pX ∗ g_s) ≤ 1/s for any probability
density pX and Gaussian kernel g_s of variance s > 0.
@audit:ok