By $P(\Omega)$, denote the space of all probability measures on $(\mathbb{R},\mathcal{B})$. Let $F_{\mu}$ denote the distribution function of $\mu\in P(\Omega)$. Let, $$ d_L(\mu,\nu):=\inf\left\{\varepsilon\geq 0: F_{\mu}(x-\varepsilon)-\varepsilon\leq F_{\nu}(x)\leq F_{\mu}(x+\varepsilon)+\varepsilon~\forall x\in\mathbb{R}\right\} $$ This is the so-called Lévy metric.

Show that $\lim_{n\to\infty}F_{\mu_n}(x)=F_{\mu}(x)$ for all points $x\in\mathbb{R}$ in which $F$ is continuous $\Leftrightarrow d_L(\mu_n,\mu)\to 0$

Its a bit confusing to me, did not find a start.

  • $\begingroup$ The wikipedia page on this distance (Lévy–Prokhorov distance) says this is equivalent to weak convergence of measures (congergence in the weak-* topology) if the space is separable $\endgroup$ – Nick Alger May 6 '16 at 16:19
  • $\begingroup$ @NickAlger Is this really the same distance? $\endgroup$ – Rhjg May 6 '16 at 16:22
  • $\begingroup$ @NickAlger This is the en.wikipedia.org/wiki/L%C3%A9vy_metric $\endgroup$ – Rhjg May 6 '16 at 16:23
  • $\begingroup$ I'm not an expert so there could be a subtlety I'm missing, but it looks the same and the linked article says it is the same $\endgroup$ – Nick Alger May 6 '16 at 16:35

"$\Leftarrow$": Suppose that $d_L(\mu_n,\mu) \to 0$ as $n \to \infty$. Fix a sequence $(\varrho_n)_{n \in \mathbb{N}}$ such that $\varrho_n \to 0$ as $n \to \infty$ and $$d_L(\mu_n,\mu) = d_L(\mu,\mu_n) < \varrho_n.$$

By the very definition of the Lévy metric, there exists $\epsilon_n \in [0,\varrho_n]$ such that

$$F_{\mu}(x-\epsilon_n)-\epsilon_n \leq F_{\mu_n}(x) \leq F_{\mu}(x+\epsilon_n) + \epsilon_n, \qquad x \in \mathbb{R}. \tag{1} $$ Note that $\epsilon_n \to 0$ as $n \to \infty$. Now if $x$ is a continuity point of $F_{\mu}$, then

$$F_{\mu}(x-\epsilon_n)-\epsilon_n \to F_{\mu}(x) \quad \text{and} \quad F_{\mu}(x+\epsilon_n)+\epsilon_n \to F_{\mu}(x),$$

and therefore it follows from $(1)$ that $F_{\mu_n}(x) \to F_{\mu}(x)$ as $n \to \infty$.

"$\Rightarrow$": Suppose that $F_{\mu_n}(x) \to F_{\mu}(x)$ for all continuity points of $F_{\mu}$. Fix $\epsilon>0$. Since $F_{\mu}$ is monotone and bounded, it can have at most countably many jumps, and so it is continuous up to a countable set $D$ of discontinuity points. In particular, we can find continuity points $x_1 < \ldots < x_k$ such that $x_{i+1}-x_i < \epsilon$ for all $i$ and

$$F_{\mu}(x_1) < \frac{\epsilon}{2} \qquad \text{and} \qquad F_{\mu}(x_k)>1-\frac{\epsilon}{2}. \tag{2}$$

Choose $N \in \mathbb{N}$ sufficiently large such that

$$|F_{\mu_n}(x_i)-F_{\mu}(x_i)| \leq \frac{\epsilon}{2} \qquad \text{for all $n \geq N$, $i=1,\ldots,k$}. \tag{3}$$

For any $x \in [x_{i-1},x_i]$, we have by the monotonicity of $F_{\mu}$, $F_{\mu_n}$ and $(3)$

$$F_{\mu_n}(x) \leq F_{\mu_n}(x_i) < F_{\mu_n}(x_i)+ \frac{\epsilon}{2} + F_{\mu}(x_i)-F_{\mu}(x_i) \stackrel{(3)}{\leq} F_{\mu}(x_i)+\epsilon \leq F_{\mu}(x+\epsilon)+\epsilon$$

for all $n \geq N$. For $x<x_1$ it holds that

$$F_{\mu_n}(x) \leq F_{\mu_n}(x_1)-F_{\mu}(x_1)+F_{\mu}(x_1) \stackrel{(2),(3)}{<} \frac{\epsilon}{2} + \frac{\epsilon}{2} \leq F_{\mu}(x+\epsilon)+\epsilon$$

and for $x>x_k$

$$F_{\mu_n}(x) \leq 1 \stackrel{(2)}{\leq} F_{\mu}(x+\epsilon)+\epsilon.$$

This proves $F_{\mu_n}(x) \leq F_{\mu}(x+\epsilon)+\epsilon$ for all $x \in \mathbb{R}$ and $n \geq N$. A very similar reasoning shows

$$F_{\mu}(x-\epsilon)-\epsilon \leq F_{\mu_n}(x),$$

and since $\epsilon>0$ is arbitrary, this finishes the proof.

  • $\begingroup$ Hi, in "$\implies$" direction, for $x<x_1$, $F_\mu (x_1)<\frac{\epsilon}{2}$ but why is $F_{\mu_{n}} (x_1)- F_\mu (x_1)<\frac{\epsilon}{2}$ because from (3), $F_{\mu_{n}} (x_1)- F_\mu (x_1)<\epsilon$? Thanks $\endgroup$ – manifolded Mar 13 '19 at 11:08
  • $\begingroup$ @manifolded I changed "$\leq \epsilon$" to "$\leq \epsilon/2$" in (3); this should fix your problem $\endgroup$ – saz Mar 13 '19 at 11:18

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.