Let $\mathscr B$ be the Borel $\sigma$-algebra on $\mathbb R$ and let $\mathscr P$ denote the set of all probability measures on the measurable space $(\mathbb R,\mathscr B)$.

Lévy’s metric on $\mathscr P$ is defined as $$d(\mathbb P_1,\mathbb P_2)\equiv\inf\{\varepsilon>0\,|\,\mathbb P_2(-\infty,x-\varepsilon]-\varepsilon\leq\mathbb P_1(-\infty,x]\leq\mathbb P_2(-\infty,x+\varepsilon]+\varepsilon\text{ for each $x\in\mathbb R$}\}$$ for $\mathbb P_1,\mathbb P_2\in\mathscr P$.

Now define another metric on $\mathscr P$ as follows: $$\rho(\mathbb P_1,\mathbb P_2)\equiv\sup_{t\in\mathbb R}\left\{\frac{|\varphi_1(t)-\varphi_2(t)|}{1+|t|}\right\}$$ for $\mathbb P_1,\mathbb P_2\in\mathscr P$, where $\varphi_i:\mathbb R\to\mathbb C$ is the characteristic function of $\mathbb P_i$: $$\varphi_i(t)\equiv\int_{x\in\mathbb R}\cos(tx)\,\mathrm d\mathbb P_i(x)+\mathsf{i}\int_{x\in\mathbb R}\sin(tx)\,\mathrm d\mathbb P_i(x)\quad\text{for $t\in\mathbb R$ and $i\in\{1,2\}$}.$$

It is not difficult to show that $\rho$ is a bona fide metric on $\mathscr P$. I can also prove that $d$ and $\rho$ are equivalent; that is, they generate the same topology on $\mathscr P$.

What I am curious about is whether $d$ and $\rho$ are strongly equivalent. That is, does there exist positive constants $A$ and $B$ such that $$A\rho(\mathbb P_1,\mathbb P_2)\leq d(\mathbb P_1,\mathbb P_2)\leq B\rho(\mathbb P_1,\mathbb P_2)\quad\text{ for any $\mathbb P_1,\mathbb P_2\in\mathscr P$}?$$

Any suggestions on proof strategies or on ways to construct a counterexample are much appreciated.


I conducted some numerical experimentation, and while I have been unable to provide a rigorous disproof, the results I have obtained seem to suggest that the metrics $d$ and $\rho$ are not strongly equivalent.

To see this, let, for each $n\in\mathbb N$, $\mathbb P_n$ denote the probability measure corresponding to the normal distribution with mean $0$ and variance $1/n^2$. Then, the corresponding characteristic function is $$\varphi_n(t)=\exp\left(-\frac{t^2}{2n^2}\right)\quad\text{for each $t\in\mathbb R$}.$$ Furthermore, let $\mathbb P$ denote the unit mass at $0$, which has characteristic function $\varphi(t)=1$ for each $t\in\mathbb R$. Let $F_n$ and $F$ denote the respective corresponding distribution functions; note that $F(x)=\mathsf I_{[0,\infty)}(x)$ for $x\in\mathbb R$.

Now, $\rho(\mathbb P_n,\mathbb P)$ can be computed by maximizing the function $$t\mapsto\frac{1-\exp\left[-\large\frac{t^2}{2n^2}\right]}{1+|t|}.$$ Next, I claim that the Lévy metric $d(\mathbb P_n,\mathbb P)$ is given as the solution to the equation $$F_n(\varepsilon_n^*)=1-\varepsilon_n^*.$$ It is easy to see that this solution is unique and positive. To see that this solution gives the Lévy metric, suppose that $\varepsilon>0$ satisfies $$F_n(x-\varepsilon)-\varepsilon\leq F(x)\leq F_n(x+\varepsilon)+\varepsilon\quad\text{for each $x\in\mathbb R$}.$$ If it were the case that $\varepsilon_n^*>\varepsilon$, then, taking $x=\varepsilon_n^*-\varepsilon$, one would have $$1=F(x)\leq F_n(x+\varepsilon)+\varepsilon=F_n(\varepsilon_n^*)+\varepsilon=1-(\varepsilon_n^*-\varepsilon)<1,$$ which is a contradiction. Hence, $\varepsilon_n^*\leq \varepsilon$, so that taking infimum yields $\varepsilon_n^*\leq d(\mathbb P_n,\mathbb P)$. To show that $d(\mathbb P_n,\mathbb P)\leq\varepsilon_n^*$, it is sufficient to check that $\varepsilon_n^*$ is contained in the set over which the infimum is taken. But this is clearly the case, for if $x\geq0$, then $$F_n(x-\varepsilon_n^*)-\varepsilon_n^*\leq 1=F(x)=1-\varepsilon_n^*+\varepsilon_n^*=F_n(\varepsilon_n^*)+\varepsilon_n^*\leq F_n(x+\varepsilon_n^*)+\varepsilon_n^*,$$ whereas if $x<0$, then $$F_n(x-\varepsilon_n^*)-\varepsilon_n^*\leq F_n(-\varepsilon_n^*)-\varepsilon_n^*=0=F(x)\leq F_n(x+\varepsilon_n^*)+\varepsilon_n^*,$$ where I exploited that $F_n(-y)=1-F_n(y)$ for any $y\in\mathbb R$.

Now, numerical calculations show the following (note that $\mathbb P_n\Rightarrow\mathbb P$ weakly, so that the sequence of distances according to either metric converges to $0$): \begin{array}{r|r|r|r} n&\rho(\mathbb P_n,\mathbb P)&d(\mathbb P_n,\mathbb P)&\rho(\mathbb P_n,\mathbb P)/d(\mathbb P_n,\mathbb P)\\\hline 1&0.2883&0.3596&0.8019\\ 10&0.0425&0.1183&0.3591\\ 100&0.0045&0.0204&0.2193\\ 1\mathord,000&[\text{small}]&[\text{small}]&0.1626\\ 10\mathord,000&[\text{small}]&[\text{small}]&0.1328\\ 100\mathord,000&[\text{small}]&[\text{small}]&0.1143\\ 1\mathord,000\mathord,000&[\text{small}]&[\text{small}]&0.0000 \end{array}

I suspect that the sudden drop in the ratio from 100,000 to 1,000,000 is due to rounding error in Wolfram Mathematica, in which the calculations were performed. At any rate, I considered other examples involving more precise numerical estimates, and those, too, suggest that the ratio $\rho/d$ can be made arbitrarily small, contradicting strong equivalence.

For example, if $X$ is a random variable with density function $x\mapsto \exp(-|x|)/2$ (standard Laplace distribution), $\mathbb P_n$ is the distribution of $X/n$ ($n\in\mathbb N$), and $\mathbb P$ is again the unit mass, then I obtain the following table (and these estimates rely less on numerical-precision constraints, and more on partial analytical solutions): \begin{array}{r|r} n&\rho(\mathbb P_n,\mathbb P)/d(\mathbb P_n,\mathbb P)\\\hline 10^0&0.7874\\ 10^1&0.3439\\ 10^2&0.1730\\ 10^3&0.1069\\ 10^4&0.0755\\ 10^5&0.0577\\ 10^6&0.0465\\ 10^7&0.0388\\ 10^8&0.0333\\ 10^9&0.0291\\ 10^{10}&0.0258\\ 10^{20}&0.0120\\ 10^{30}&0.0078\\ 10^{40}&0.0058\\ 10^{50}&0.0046\\ 10^{60}&0.0038\\ 10^{70}&0.0032\\ 10^{80}&0.0028\\ 10^{90}&0.0025\\ 10^{100}&0.0022 \end{array}

Actually, Mathematica could compute the limit analytically, and it turned out to be $0$, indeed (albeit convergence is very slow, as the table shows). However, the expressions involved are so abstruse that I have refrained from double-checking the value of the limit rigorously.


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.