I've encountered the following problem that involves alternating sums of multinomial coefficients. Let $$f(k)=\sum_{i=0}^{n-k}(-1)^i\binom{n}{k,i,n-k-i}(k+i)^\alpha$$ where $\binom{n}{k,i,n-k-i}=\frac{n!}{k!i!(n-k-i)!}$ is the multinomial coefficient, and $\alpha\in\mathbb{R}$ is a constant that satisfies $-1\leqslant\alpha<0$

Problem: I'm trying to prove that $$\frac{f(k)}{f(k+1)}\leqslant\frac{k+1}{k}$$, and equality happens when $\alpha=-1$.

Background: this problem stems from my study of expected order statistics of Weibull distribution. When a Weibull distribution has shape parameter larger than 1 (assuming scale parameter equals 1), the expected difference between the $k$-th largest order statistic and the $k+1$-th is exactly $f(k)$. Numerical computations showed the results hold, but I cannot prove it analytically. Any suggestions or pointers are much appreciated!

  • 1
    $\begingroup$ Did you try the usual trick $x^{\alpha}=\Gamma(-\alpha)^{-1}\int e^{-tx} t^{-\alpha-1}dt$? $\endgroup$ – Abdelmalek Abdesselam Jun 2 '17 at 16:54

The case $\alpha=-1$, i.e. we need to prove that $kf(k)=(k+1)f(k+1)$.

Let $m=n-k$ and $g(k)=kf(k)$ so that $g(k)=\sum_{j=0}^mF_k(m,j)$ where $$F_k(m,j):=(-1)^j\binom{m}j\binom{m+k}m\frac{k}{k+j}.$$ Define another function $$G_k(m,j):=(-1)^{j-1}\binom{m+1}j\binom{m+k}m\frac{kj}{(m+1)^2}.$$ Now, check this identity $F_k(m+1,j)-F_k(m,j)=G_k(m,j+1)-G_k(m,j)$, for instance routinely divide through by $F_k(m,j)$ and simplifying. As a next step, sum over all integers $j$: $$\sum_{j\in\mathbb{Z}}F_k(m+1,j)-\sum_{j\in\mathbb{Z}}F_k(m,k) =\sum_{j\in\mathbb{Z}}G_k(m,j+1)-\sum_{j\in\mathbb{Z}}G_k(m,j).$$ Since the RHS vanishes (equal sums), the outcome is: $g(k+1)-g(k)=0$. It follows that $g(k)=1$ because $g(0)=1$. Hence, $kf(k)=(k+1)f(k+1)$ as desired.

This approach uses the Wilf-Zeilberger method together with Zeilberger's algorithm.

  • $\begingroup$ Thanks for this answer and pointers to the paper. I will try and see if this approach also works for the case of $-1<\alpha<0$. The case of $\alpha=-1$ actually reduces to order statistics of a standard exponential distribution, whereas the general case of Weibull distribution seems somewhat harder. $\endgroup$ – user3026001 Jun 2 '17 at 21:59
  • $\begingroup$ I expect the case $\alpha\neq-1$ to be harder. Warning: the above method does not apply when $\alpha\not\in\mathbb{Z}$. $\endgroup$ – T. Amdeberhan Jun 2 '17 at 22:01

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.