指数型分布族

ベルヌーイ・ポアソン・正規分布を共通形式で捉える指数型分布族を、統一的な視点で解説します。

ProbabilityExponential FamilyGLM

はじめに

指数型分布族は、多くの確率分布を一つの枠組みで表す方法です。

一般形は

p(xθ)=h(x)exp{η(θ)T(x)A(θ)}p(x\mid\theta)=h(x)\exp\{\eta(\theta)T(x)-A(\theta)\}

で、主に次を表します。

  • T(x)T(x):十分統計量
  • η(θ)\eta(\theta):自然パラメータ
  • A(θ)A(\theta):対数分配関数(正規化)

1. なぜ便利か

指数型分布族で書けると、

  • 最尤推定が整理しやすい
  • 十分統計量が明確になる
  • GLM(一般化線形モデル)と自然に接続できる

という利点があります。

2. 代表例

ベルヌーイ分布

P(X=x)=px(1p)1xP(X=x)=p^x(1-p)^{1-x}

は指数型分布族の形に書き直せます。T(x)=xT(x)=x、自然パラメータは logp1p\log\frac{p}{1-p} です。

ポアソン分布

P(X=x)=λxeλx!P(X=x)=\frac{\lambda^x e^{-\lambda}}{x!}

も同様に指数型分布族に入り、T(x)=xT(x)=x、自然パラメータは logλ\log\lambda です。

正規分布(分散既知)

平均をパラメータにした正規分布も指数型分布族で表せます。

3. 対数分配関数の役割

A(θ)A(\theta) は分布を正規化するだけでなく、微分によって平均や分散などのモーメント情報を与えます。

  • A(θ)A'(\theta):平均に対応
  • A(θ)A''(\theta):分散に対応

4. GLMとの関係

GLMでは、応答分布として指数型分布族を仮定し、リンク関数で平均構造を結びます。

  • ロジスティック回帰(ベルヌーイ)
  • ポアソン回帰(ポアソン)

などはこの統一的枠組みの具体例です。

インタラクティブデモ

下のデモで分布ごとのパラメータを変え、指数型分布族としての共通構造と違いを比較できます。

Exponential Family Builder

Explore exponential family distributions interactively

Distribution

-33

Distribution Visualization

Exponential Family Form

Standard Form

p(x) = pˣ(1-p)¹⁻ˣ

Maximum Probability

0

Distribution Type

Discrete

Components

h(x) - Base Measure

1

Constant (same for all x)

T(x) - Sufficient Statistic

x

Identity function (x itself)

η - Natural Parameter

0.500

Controls the distribution shape

A(θ) - Log-partition

A(θ)

Ensures normalization

Parameter Relationship

η = log(p/(1-p)), where p = 1/(1+e⁻η)

Understanding Exponential Families

General Form

All exponential family distributions can be written as:

p(x) = h(x) × exp(η×T(x) - A(θ))

Component Roles

h(x): Determines the support and basic shape of the distribution
T(x): Captures sufficient information about the data
η: Natural parameter that controls the distribution's characteristics
A(θ): Log-partition function ensuring proper normalization

まとめ

  • 指数型分布族は多くの分布を共通形式で表す枠組み。
  • 十分統計量・自然パラメータ・対数分配関数が中心。
  • 推定理論やGLMを理解する土台になる。
← 図鑑に戻る