I read this post How is the minimum of a set of IID random variables distributed? where I can find how to compute the density distribution of the minimum between $N$ positive random variables. If the cumulative of these random variables is $F(x)$ the cumulative of the minimum between $N$ of these random variables is:

$$ \mathbb{P}(\text{min}\le x)=1-\big(1-F(x)\big)^N$$

From this result it is easy to compute the mean value of the minimum (simply using integration by parts): $$ \overline{\text{min}}=\int_{0}^{\infty}x\,\frac{d \mathbb{P}(\text{min}\le x)}{dx}dx=\int_{0}^{\infty}\big(1-F(x)\big)^Ndx $$ At this point I can compute $\overline{\text{min}}$ for a general cumulative distribution $F(x)$.

I want now to get the scaling of $\overline{\text{min}}$ for $N\rightarrow\infty$ when $F(x)=\frac {1}{r!}\int_{0}^{x}l^re^{-l}dl$ for $r\in\mathbb{Z}$ and $r>-1$. (A simple case can be the one for $r=0$, that is the density distribution for the variables is exponential, in this case the average min value scale as $1/N$). In the general case I can arrive to: $$ \overline{\text{min}}=\int_{0}^{\infty}\left(1-\frac{1}{r!}\int_0^{x}l^re^{-l}dl\right)^Ndx $$ but I don't know how to approximate it. For general $r$ it can be useful to approximate $\frac{l^re^{-l}}{r!}\simeq \frac{l^r}{r!}\propto l^r$, since the important contribution for the average minimum value will be $\ll1$. In this approximation my result is: $$ \overline{\text{min}}=\int_{0}^{\infty}\left(1-\int_0^{x}l^rdl\right)^Ndx $$ but still can't figure out which is the scaling with large $N$.

The result should be: $$ \overline{\text{min}}\simeq N^{-\frac{1}{r+1}} $$

I am studying the following paper: https://hal.science/jpa-00232897/document .

Here the authors find the scaling of the mean value of what I suppose is the minimum of the random variables, which in this case are $l_{ij}$, called “distances”. I cannot find the scaling found after Eq.(5), can anyone help me?

  • $\begingroup$ Can you clarify the exact mathematical meaning of the operator "$\simeq$": does $a_N \simeq b_N$ means $a_N/b_N \to 1$ as $N \to \infty$? $\endgroup$
    – Zhanxiong
    May 28, 2023 at 15:38
  • $\begingroup$ By "$a_N\simeq b_N$" I mean $a_N/b_N\rightarrow c$ as $N\rightarrow\infty$ where $c$ is a constant (independent on $N$). $\endgroup$ May 28, 2023 at 17:19
  • $\begingroup$ The result is a simple one and wholly unrelated to how you frame the question. The energy in (2) is a sum over neighbors of $2N$ terms. The "short links" assumption implies there are $O(N)$ such terms altogether. Each term $l_{ij}$ is approximately $N^{-1/(r+1)}$ following equation (5), whence upon multiplication by $O(N)$ you obtain the stated result. $\endgroup$
    – whuber
    May 28, 2023 at 18:29
  • $\begingroup$ My question is why equation (5) implies that scaling ($N^{-1/(r+1)}$) $\endgroup$ May 29, 2023 at 12:58
  • $\begingroup$ And the answer is that you multiply by $N$. $\endgroup$
    – whuber
    May 29, 2023 at 15:35

1 Answer 1


Here is a way to arrive at the result

  • With $f(l) = \frac{l^re^{-l}}{r!}\simeq \frac{l^r}{r!}\propto l^r$ we have cdf $F(l) \propto l^{r+1}$ and quantile function $Q(p) \propto p^{\frac{1}{r+1}}$
  • The minimum statistic can be expressed with by the transform of a beta distributed variable $$min \sim Q(p) \quad \text{where $p \sim Beta(1,n)$ or $f_P(p) = n(1-p)^{n-1}$}$$
  • The mean can then be approximated with an integral $$\begin{array}{} mean(min)& \approx &\int_0^1 Q(p) f_P(p)\, \text{d}p \\&\approx& n \int_0^1 p^{\frac{1}{r+1}} (1-p)^{n-1}\, \text{d}p\\ &\approx&n Beta(1+\frac{1}{r+1},n) \\&\approx& n \Gamma\left(1+\frac{1}{r+1}\right) n^{-(1+\frac{1}{r+1})}\\& \approx &\Gamma\left(1+\frac{1}{r+1}\right) n^{-\frac{1}{r+1}} \end{array} $$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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