27
$\begingroup$

How to evaluate this integral: $$\int_0^1 \int_0^1 \cdots \int_0^1\frac{x_{1}^2+x_{2}^2+\cdots+x_{n}^2}{x_{1}+x_{2}+\cdots+x_{n}}dx_{1}\, dx_{2}\cdots \, dx_{n}=?$$ I'm making use of the integral identity: $$\int_{0}^{+\infty }e^{-t(x_{1}+x_{2}\cdots +x_{n})}dt=\frac{1}{x_{1}+x_{2}\cdots +x_{n}}$$ and then reversing the order of integration with respect to time and space variables. But for $n=1$, then such that, $$\int_{0}^{\infty }dt\int_{0}^{1}x^{2}e^{-tx}dx=\int_{0}^{\infty }\frac{2 - e^{-t}(2 + 2t+t^2)}{t^3}dt=\int_{0}^{1}x\,dx=\frac{1}{2},$$ and $$\int_0^1 \int_0^1 \cdots \int_0^1\frac{x_{1}^2+x_{2}^2+\cdots+x_{n}^2}{x_{1}+x_{2}+\cdots+x_{n}}dx_{1}\, dx_{2}\cdots \, dx_{n}\\=n\int_{0}^{+\infty }\frac{2 - e^{-t}\left ( 2 + 2t+t^2 \right )}{t^3}\left ( \frac{1-e^{-t}}{t} \right )^{n-1}dt.$$

$\endgroup$
4

3 Answers 3

13
$\begingroup$

Here is another approach, which also gives the rational term.

(I) To see how it works let $n\geq 2$ and consider first the simpler case \begin{align*} \mathbb{E}\bigg(\frac{1}{X_1+\ldots+X_n}\bigg)=\int_0^\infty \bigg(\frac{1-e^{-t}}{t}\bigg)^n\,dt \end{align*} Using $\frac{1}{t^n}=\int_0^\infty \frac{z^{n-1}}{(n-1)!} e^{-zt}\,dz$ we find \begin{align*} \int_0^\infty \bigg(\frac{1-e^{-t}}{t}\bigg)^n\,dt&=\int_0^\infty\int_0^\infty \frac{z^{n-1}}{(n-1)!} e^{-zt}(1-e^{-t})^n\, dz\,dt\\ &=\int_0^\infty\int_0^\infty \frac{z^{n-1}}{(n-1)!} e^{-zt}(1-e^{-t})^n\, dt\,dz\\ &=\int_0^\infty \frac{z^{n-1}}{(n-1)!}\, \mathrm{Beta}(z,n+1)\,dz\\ &=n\,\int_0^\infty \frac{z^{n-1}}{z(z+1)\cdots(z+n)}\,dz \end{align*} The following observation will be the key:

Lemma: let $x_0,\ldots,x_n$ be distinct positive numbers and $k\leq n-1$. Then $$I_{k,n}(x_0,\ldots,x_n):=\int_0^\infty \frac{z^{k}}{(z+x_0)(z+x_1)\cdots(z+x_n)}\,dz=(-1)^{n+k+1}\Delta^n(x^k\log(x);x_0,\ldots,x_n)$$ where (for a real function $f$) $\Delta^n(f;x_0,\ldots,x_n)$ denotes the divided difference of $f$ corresponding to $x_0,\ldots,x_n$.

Recall that (Newton-interpolation)

(1) the divided differences are for $f$ and mutually distinct $x_0,\ldots,x_n$ are defined recursively by $\Delta^0(f;x_0)=f(x_0)$, $\Delta^n(f;x_0,\ldots,x_n)=\frac{\Delta^{n-1}(f; x_1,\ldots,x_n)-\Delta^{n-1}(f; x_0,\ldots,x_{n-1})}{x_n-x_0}$

(2) they are explicitly given by $$\Delta^n(f;x_0,\ldots,x_n)=\sum_{i=0}^n \frac{f(x_i)}{\prod_{j\neq i} (x_i-x_j)}\;\;(**)$$ (3) $$\Delta^n(f;x,x+1,\ldots,x+n)=\frac{1}{n!} \sum_{i=0}^n {n \choose i} (-1)^{n-i} f(x+i)$$

Proof of the lemma: For $k=0, n=1$ we have $$\int_0^\infty \frac{1}{(z+x_0)(z+x_1)}=\frac{\log(x_1)-\log(x_0)}{x_1-x_0}=\Delta^1(\log(x);,x_0,x_1)$$ For $k=0,n>1$ the repeated use of $\frac{1}{(z+a)(z+b)}=\frac{-1}{b-a}\left(\frac{1}{z+b}-\frac{1}{z+a}\right)$ shows that $(-1)^{n+1}I_{0,n}(x)=\Delta^n(\log(x),x)$. The validity for $k>0,n=k+1$ follows from $\frac{z}{z+b}=1-\frac{b}{z+b}$ and $(**)$. End of proof.

The lemma and (3) now give that \begin{align*} \int_0^\infty \bigg(\frac{1-e^{-t}}{t}\bigg)^n\,dt &=n \Delta^{n-1}(x^{n-2}\log(x);x+1,x+2,\ldots,x+n)\\ &=\frac{n}{(n-1)!}\sum_{i=0}^{n-1}{ n-1 \choose i} (-1)^{n-1-i} (i+1)^{n-2}\log(i+1) \end{align*}

(II) Now let $n\geq 1$ and consider $$ Q_{n+1}:=\mathbb{E}\bigg(\frac{X_1^2+\ldots + X_{n+1}^2}{X_1+\ldots+X_{n+1}}\bigg)$$ Write $$Q_{n+1}=(n+1)\int_0^1 u^2 \bigg(\int_0^\infty e^{-ut}\bigg(\frac{1-e^{-t}}{t}\bigg)^n\,dt\bigg)\,du$$ Proceeding as above shows that for $u>0$ \begin{align*} \int_0^\infty e^{-ut}\bigg(\frac{1-e^{-t}}{t}\bigg)^n\,dt &=\frac{n}{n!} \sum_{i=0}^n {n \choose i} (-1)^{n-i} (u+i)^{n-1}\log(u+i) \end{align*} Thus $$q_{n+1}:=\frac{Q_{n+1}}{n(n+1)}=\frac{1}{n!} \sum_{i=0}^n {n \choose i} (-1)^{n-i} \int_0^1 u^2(u+i)^{n-1}\log(u+i)\,du$$ Now expand $u^2=(u+i-i)^2$ and integrate partial to find that $q_{n+1}=L-R$ where $L=\frac{1}{n!}\sum_{i=0}^n{n\choose i}(-1)^{n-i} \ell(i),\;R=\frac{1}{n!}\sum_{i=0}^n {n \choose i}(-1)^{n-i} r(i)$ with \begin{align*} \ell(i)=&\frac{1}{n+2}\left((i+1)^{n+2}\log(i+1)-i^{n+2}\log(i)\right)\\ &-\frac{2i}{n+1}\left((i+1)^{n+1}\log(i+1)-i^{n+1}\log(i)\right)\\ &+\frac{i^2}{n}\left((i+1)^{n}\log(i+1)-i^n\log(i)\right)\\[0.2cm] r(i)=&\frac{1}{(n+2)^2}\left((i+1)^{n+2}-i^{n+2}\right)\\ &-\frac{2i}{(n+1)^2}\left((i+1)^{n+1}-i^{n+1}\right)\\ &+\frac{i^2}{n^2}\left((i+1)^{n}-i^n\right) \end{align*} Collecting terms in the logarithmic part $L$ shows that the coefficient of $\log(i)$ in $q_{n+1}$ is $c_{i,n+1} =(-1)^{n-i} \frac{i^n}{n}{ n \choose i-1}\left(\frac{n^2+3n+2-2i}{(n+2)!}\right)$, matching Fedor Petrov's answer.

The rational part $R$ can readily be summed (note that only the contributions of the powers $n$ and $n+1$ (with coefficients $-\frac{1}{n(n+1)(n+2)}$ and $\frac{2}{n(n+1)(n+2)}$) need to be considered) to give $$R=\frac{1}{n+2}-\frac{1}{n(n+1)(n+2)}\;\;,$$ so that the rational term of $Q_{n+1}$ is $-\frac{n(n+1)-1}{n+2}=-\frac{(n+2)(n-1)+1}{n+2}$, confirming Sylvain JULIEN's guess.

$\endgroup$
5
  • 3
    $\begingroup$ Very nice method! $\endgroup$ Dec 22, 2017 at 21:12
  • $\begingroup$ For integrating a rational function like $f(z)=z^{n-2}/\prod_{k=1}^n(z+k)$ against $[0,+\infty)$ we may simply expand it via elementary fractions $f(z)=\sum c_k/(z+k)$, where $c_k={\rm Res}_{z=-k}f(z)=(f(z)(z+k))|_{z=-k}=k^{n-2}(-1)^{n-k}/k!(n-k)!$. Now $\sum c_k=0$ (it is clear a priori from the asymptotics of $f$ for large $z$: $\sum c_k=\lim_{z\to \infty} zf(z)=0$); and we get $\int_0^\infty f=\int_0^\infty \sum c_k (1/(z+k)-1/z)=-\sum c_k\log k$. $\endgroup$ Dec 23, 2017 at 11:12
  • $\begingroup$ Thank you. (1) I also like your method (but have been unable to extend it to cover the rational term) (2) The rational integral: you're right, that's the most natural way to do it. In the present context I was probably led to explain it via divided differences because pdf/cdf of sums of iid uniform variates arise via a differencing process. $\endgroup$
    – esg
    Dec 23, 2017 at 18:33
  • $\begingroup$ nice solution @esg. $\endgroup$
    – Shahrooz
    Dec 24, 2017 at 6:54
  • $\begingroup$ @esg my method actually also allows to get an answer, I simply missed something obvious (that there are no off-integral terms in integrating by parts here), see my updated answer $\endgroup$ Jan 16, 2018 at 7:04
12
$\begingroup$

Ah, we may simply integrate by parts!

Denote $h(t)=(2 - e^{-t}( 2 + 2t+t^2))(1-e^{-t})^{n-1}$. Integrating by parts $(n+1)$ times we get $$\int_0^\infty h(t)t^{-n-2}dt=\frac1{(n+1)!}\int_0^\infty h^{(n+1)}(t)t^{-1}dt,$$ all off-integral terms vanish since $h(0)=h'(0)=\dots=h^{(n+1)}(0)=0$ and at infinity we have $h(t)=O(1)$, $h^{(i)}(t)=o(1)$ for $i>0$. We have $h^{(n+1)}(t)=\sum_{k=1}^n a_ke^{-kt}+u(t)$, where $u(t)=t\times \text{polynomial}(e^{-t},t)$. Note that $\sum a_k=h^{(n+1)}(0)=0$, thus $$\int_0^\infty \sum a_ke^{-kt}t^{-1}dt=\int_0^\infty \sum a_k(e^{-kt}-e^{-t})t^{-1}dt=-\sum a_k \log k$$ by Frullani integrals. It is easy to see that $$a_k=(-1)^{k+n+1}\binom{n-1}{k-1}k^{n-1}(n^2+n-2k).$$ It remains to evaluate $\int_0^\infty u(t) t^{-1}dt$. We have $$u(t)=-(2t+t^2)(g(t))^{(n+1)}-2(n+1)t(g(t))^{(n)},\,g(t)=e^{-t}(1-e^{-t})^{n-1}.$$ Therefore $$\int_0^\infty u(t) t^{-1}dt=-2\int g^{(n+1)}(t)dt-2(n+1)\int g^{(n)}(t)dt-\int_0^{\infty}tg^{(n+1)}(t)dt.$$ We get $\int (g^{(n)}+tg^{(n+1)})=tg^{(n)}$, and the definite integral against $(0,\infty)$ equals 0. It remains $\int_0^\infty -2g^{(n+1)}-(2n+1)g^{(n)}=2g^{(n)}(0)+(2n+1)g^{(n-1)}(0)$. We have $$g(t)=(1-t+\dots)t^{n-1}(1-t/2+\dots)^{n-1}=t^{n-1}-\frac{n+1}2t^n+\dots,$$ $g^{(n-1)}(0)=(n-1)!$, $g^{n}(0)=-\frac{(n+1)!}2$, $2g^{(n)}(0)+(2n+1)g^{(n-1)}(0)=-(n^2-n-1)(n-1)!$, confirming the guess of Sylvain JULIEN.

PREVIOUS NON COMPLETE VERSION

Still not a complete answer, but a method to be completed or improved.

Denote $f_n(t)=\frac{2 - e^{-t}\left ( 2 + 2t+t^2 \right )}{t^3}\left ( \frac{1-e^{-t}}{t} \right )^{n-1}$. Assume that we have found the coefficients $c_1,c_2,\dots,c_n$ and polynomials $g_1,\dots,g_n$ such that $$f_n(t)=2t^{-n-2}+\sum_{k=1}^n(g_k(t^{-1})e^{-kt})'+\sum_{k=1}^n c_k \frac{e^{-kt}}t.\,\,(*)$$ Then $\sum c_k=0$ (else $f_n$ would have a non-zero residue at 0, which is absurd). We have $\int_0^\infty \sum_{k=1}^n c_k \frac{e^{-kt}}t dt=-\sum c_k\log k$ by Frullani integrals. The integral of the other part of our sum $(*)$ is minus the limit at zero of the function $-\frac2{n+1}t^{-n-1}+\sum g_k(t^{-1})e^{-kt}$. The limit must exist, since the initial integral converges.

Now I cheat a bit. Note that if we write $f_n(t)=\sum_{k=0}^n q_k(1/t) e^{-kt}$ for polynomials $q_k$, then $c_k$ equals to the residue of the $q_k(1/t) e^{-kt}$, which is pretty computable. If I am not mistaken, $$q_k(t^{-1})=(-1)^kt^{-n-2}\left(2\binom{n-1}k+\binom{n-1}{k-1}(2+2t+t^2)\right)=\\=(-1)^kt^{-n-2}\left(2\binom{n}k+\binom{n-1}{k-1}(2t+t^2)\right).$$ Thus $$c_k=2(-1)^{k+n+1}\binom{n}k\frac{k^{n+1}}{(n+1)!}+2(-1)^{k+n}\binom{n-1}{k-1}\frac{k^{n}}{n!}+(-1)^{k+n+1}\binom{n-1}{k-1}\frac{k^{n-1}}{(n-1)!}=\\ =(-1)^{k+n+1}\frac{k^{n-1}(n^2+n-2k)}{n(n+1)(n-k)!(k-1)!}.$$ This matches a coefficient of $\log 3$ for $n=4$ from Shahrooz Janbaz's answer, you may check others for be sure. It remains to prove Sylvain JULIEN's guess for the rational part.

$\endgroup$
2
  • $\begingroup$ Very nice! I failed to notice that. So we have two independent proofs now, and both are more examples of "technique that should be at our fingertips" (as I've read in this forum a while ago). $\endgroup$
    – esg
    Jan 16, 2018 at 19:25
  • $\begingroup$ My belief is that maybe it is possible to get a generalization of the integral, or maybe study different cases for other positive integer/fractional exponents. I hope don't disturb with my comment. $\endgroup$
    – user142929
    Aug 7, 2019 at 19:36
7
$\begingroup$

A little long as a comment:

Thanks to Maple, some first few terms are as follows:

$n=1:$ $I=\frac{1}{2}$

$n=2:$ $I=\frac{2^2\ln(2)}{3}-\frac{1}{3}$

$n=3:$ $I=-2^3\ln(2)+\frac{3^3\ln(3)}{4}-\frac{5}{4}$

$n=4:$ $I={2^6\ln(2)}-\frac{189\ln(3)}{5}-\frac{11}{5}$

$n=5:$ $I=-\frac{5\times 584\ln(2)}{9}+3^4\ln(3)+\frac{5^4\ln(5)}{36}-\frac{19}{6}$

$n=6:$ $I=\frac{6\times 11696\ln(2)}{63}+\frac{6\times 405\ln(3)}{14}-\frac{6\times 6250\ln(5)}{63}-\frac{29}{7}$

and so on. It seems there is some patterns which may help to obtain closed formula for the integral.

$\endgroup$
2
  • 3
    $\begingroup$ The last constant term seems to be $ -\frac{(n+1)(n-2)+1}{n+1} $ . $\endgroup$ Dec 10, 2017 at 19:44
  • 7
    $\begingroup$ In looking for patterns, it seems plausible that in the $n=4$ and $n=5$ formulas, the $\ln(2)$ term should be split into a multiple of $\ln(2)$ and a multiple of $\ln(4)$, and similarly in the $n=6$ formula some of the $\ln(2)$ and $\ln(3)$ terms should really be a $\ln(6)$ term. Possibly doing the computation by hand, rather than having Maple chunk out the answer, would give a plausible decomposition. Anyway, just a thought. $\endgroup$ Dec 10, 2017 at 23:57

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.