狄利克雷分布是关于一组 d d d个连续变量 x i ∈ [ 0 , 1 ] x_i\in[0, 1] xi∈[0,1]的概率分布, ∑ i x i = 1 \sum_ix_i=1 ∑ixi=1。令 μ = ( μ 1 , μ 2 , ⋯ , μ d ) \mu=(\mu_1, \mu_2, \cdots, \mu_d) μ=(μ1,μ2,⋯,μd),参数 α = ( α 1 , α 2 , ⋯ , α d ) \alpha=(\alpha_1, \alpha_2, \cdots, \alpha_d) α=(α1,α2,⋯,αd),其中 α i > 0 \alpha_i>0 αi>0且 α ^ = ∑ i α i \hat{\alpha}=\sum_i\alpha_i α^=∑iαi。
D i r ( x ∣ α ) = Γ ( α ^ ) Γ ( α 1 ) Γ ( α 2 ) ⋯ Γ ( α i ) ∏ i = 1 d x i α i − 1 Dir(x|\alpha)=\frac{\Gamma(\hat{\alpha})}{\Gamma(\alpha_1)\Gamma(\alpha_2)\cdots\Gamma(\alpha_i)}\prod_{i=1}^dx_i^{\alpha_i-1} Dir(x∣α)=Γ(α1)Γ(α2)⋯Γ(αi)Γ(α^)i=1∏dxiαi−1
狄利克雷分布有如下性质:
- E [ x i ] = α i α ^ E[x_i]=\frac{\alpha_i}{\hat{\alpha}} E[xi]=α^αi
- V a r ( x i ) = α i ( α ^ − α i ) α ^ 2 ( α ^ + 1 ) Var(x_i)=\frac{\alpha_i(\hat{\alpha}-\alpha_i)}{\hat{\alpha}^2(\hat{\alpha}+1)} Var(xi)=α^2(α^+1)αi(α^−αi)
- C o v ( x i , x j ) = α i α j α ^ 2 ( α ^ + 1 ) Cov(x_i, x_j)=\frac{\alpha_i\alpha_j}{\hat{\alpha}^2(\hat{\alpha}+1)} Cov(xi,xj)=α^2(α^+1)αiαj
当 d = 2 d=2 d=2时,狄利克雷分布退化为贝塔分布。