Core Concepts
Focusing on closed-form expressions for the Fisher-Rao distance in various probability distributions.
Abstract
確率分布間のFisher-Rao距離の閉形式表現に焦点を当てた調査。統計的なマニホールド上でのFisherメトリックに基づく確率分布間の測地距離。数式や幾何学的概念を用いて、閉形式表現を見つけることが非自明であることが強調されています。
Stats
n = {0, 1, 2, ...}
N∗ = {1, 2, 3, ...}
R+ = [0, ∞[
R∗+ = ]0, ∞[
Université Paris-Saclay, CentraleSupélec, CNRS, France. E-mail: henrique.miyamoto@centralesupelec.fr.
IMECC, Unicamp, Brazil. E-mail: sueli@unicamp.br.
CETEC, UFRB, Brazil. E-mail: julianna.pinele@ufrb.edu.br
arXiv:2304.14885v2 [math.ST] 27 Feb 2024
Binomial distribution:
p(x) = nCx θ^x(1 - θ)^(n-x)
g11(θ) = nθ(1-θ)
dFR(θ1, θ2) = 2√n arcsin √θ1 - arcsin √θ2
Poisson distribution:
p(x) = λ^x e^-λ/x!
g11(λ) = 1/λ
dFR(λ1, λ2) = 2 √λ1 - √λ2
Geometric distribution:
p(x) = θ(1 - θ)^(x-1)
g11(θ) = 1/θ^2(1-θ)
dFR(θ1, θ2) = 2 arctanh √1 - θ1 - arctanh √1 - θ2
Negative binomial distribution:
p(x) = (x+r-1)!/(r-1)! r^x θ^r (1 - θ)^x
g11(θ) = r/θ^2 (1-θ)
dFR(θ1, θ2) = 2√r arctanh √1 - θ_! - arctanh √_!_!
Categorical distribution:
p(x) = P_i p_i * δ_ij + (P_n p_n)^(-i,j)
g_ij(ξ) = δ_ij * p_i + n/(pn + n)*δ_ij
dFR(p_!, p_!)= 2 arccos P_i sqrt(piqi)
Multinomial distribution:
p(x!)=(n!)/(Q_n i=pi x_i /xi!)
g_ij(xi)= xidelta_ijpi+(n/pi+n)*delta_ij
dFR(xi_, xi_)= sqrt(n)*arccos P_i sqrt(piqi)
Negative multinomial distribution:
p(x^n)=xn!/Γ(P_n i=xi x_i)/Γ(x_n)Q_(n-)_i=pixi/xi!
gij(xi)=xideltaij*pipn+xi/p^(i)n+pni,j
dFR(xi, xi_)=sqrt(xi)*arccosh((sqrt(piqi)/sqrt(pnqn))