For a positive vector $\alpha\in\mathbb{R}^n$ ($n\geq 1$), denote by $\text{Dir}(\alpha)$ the Dirichlet distribution with parameter $\alpha$. In terms of weak convergence, is it true that, if $\sum\limits_{i=1}^n\alpha_i=1$, then $\lim\limits_{\varepsilon\rightarrow 0^+}\text{Dir}(\varepsilon\alpha)\longrightarrow \sum\limits_{i=1}^n \alpha_i \delta_{\lbrace e_i\rbrace}$ (where $(e_i)_{1\leq i\leq n}$ is the canonical base of $\mathbb{R}^n$)?


$\newcommand\Ga\Gamma\newcommand\R{\mathbb R}$For any $a=(a_1,\dots,a_n)\in(0,\infty)^n$ and any real $t\in(0,1/2)$, let $X=(X_1,\dots,X_n)$ have the Dirichlet distribution with parameter $ta$. Then $X_1$ has the beta distribution with parameters $ta_1$ and $tb_1$, where $b_1:=s-a_1$ and $$s:=a_1+\dots+a_n.$$

Let $t\downarrow0$. Then $\Ga(t)=\Ga(1+t)/t\sim1/t$ and hence
$$P(X_1>1-t)=\frac{\Ga(ts)}{\Ga(ta_1)\Ga(tb_1)}\,J \sim\frac{ta_1b_1}s\,J,$$ where $$J:=\int_{1-t}^1 x^{ta_1-1}(1-x)^{tb_1-1}\,dx \sim\int_{1-t}^1 (1-x)^{tb_1-1}\,dx=\frac{t^{tb_1}}{tb_1}\sim\frac1{tb_1},$$ so that $P(X_1>1-t)\to\dfrac{a_1}s$. Similarly, for each $j\in[n]:=\{1,\dots,n\}$, $$P(X_j>1-t)\to\dfrac{a_j}s.$$ Hence, $$P(X_j\le 1-t\ \forall j\in[n])\to1-\sum_{j=1}^n\dfrac{a_j}s=0.$$

So, for any continuous function $f\colon\R^n\to\R$, $$Ef(X)=\sum_{j=1}^n Ef(X)1(X_j>1-t)+Ef(X)1(X_j\le 1-t\ \forall j\in[n]) \to\sum_{j=1}^n f(e_j)\dfrac{a_j}s+0,$$ where $e_j$ is the $j$th standard basis vector of $\R^n$; here we used the implications $X_j>1-t\iff1>X_j>1-t\implies0<X_i<t<1-t\ \forall i\in[n]\setminus\{j\}$.

Thus, the Dirichlet distribution with parameter $ta$ converges to $\sum\limits_{j=1}^n \dfrac{a_j}s \delta_{\{e_i\}}$ as $t\downarrow0$. That is, your conjecture holds iff $s=1$.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.