Degree of Polynomial in Centered Moments of Gamma$(n,1)$

89 Views Asked by At

I'm interested in the degree of the polynomial in $n$ of the expression for the $k$-th central moment $$ E((X_n - n)^k) $$ where $X_n$ is a Gamma$(n,1)$ random variable, that is, the sum of $n$ independent exponential variates of parameter $1$. It is easy to show that $$ E((X_n - n)^2) = n; \quad E((X_n - n)^3) = 2n $$

Via Mathematica I verify for $k$ up to $150$ that the resulting polynomial in the variable $n$ is of order $\lfloor k/2 \rfloor$. Alas, I have not found a general demonstration. Developing the expression, and using the expression for the raw moments of this distribution, one easily show that $$ E((X_n - n)^k) = \sum_{i=0}^k \binom{k}{i} (-1)^{k-i} n^{k-i} \frac{(n+i-1)!}{(n-1)!} $$ But I'm stuck there. I tried to express the ratio of factorials via the Stirling numbers expansion of the Pochhammer symbol without success.

Also, Mathematica computes that $$ E((X_n - n)^k) = U(-k, 1-k-n, -n) $$ where $U$ is the Tricomi confluent hypergeometric function.

Any help is welcomed

Thanks in advance!

1

There are 1 best solutions below

1
On BEST ANSWER

To find the degree of this polynomial was Problem 11403 in the Dec. 2008 issue of the American Mathematical Monthly. The function was there called $f_n(x)$ and its degree is indeed $\lfloor n/2\rfloor.$

In your notation we have $E[(X_n-n)^k]=f_k(n).$

Here is the first paragraph of the published solution (March 2011):

The degree of $f_n$ is $\lfloor n/2\rfloor.$ This follows immediately from the stronger statement that the coefficient of $x^r$ in $f_n(x)$ is the number of derangements of $[n]$ with $r$ cycles, since each cycle must have at least two elements. Here $[n]=\{1,\dots,n\}$, and a derangement is a permutation with no fixed points.

The published solution also mentions other interesting combinatorial interpretations of $f_n(x)$.


Here is my never-before-published, inelegant solution.

For $n\geq 0$, the binomial coefficients satisfy ${n+1\choose i}={n\choose i}+{n\choose i-1}$ and ${n\choose i}=0$ unless $0\leq i\leq n$. Thus, for any function $h(i)$ we get, $$\sum_{i=0}^{n+1} {n+1\choose i} h(i) = \sum_{i=0}^n {n\choose i} [h(i)+ h(i+1)].\tag 1$$

For real $x$ and integers $0\leq i\leq n$, define $g(n,i)=(-x)^{n-i}\prod_{j=0}^{i-1}(x+j)$ so that $f_n(x)= \sum_{i=0}^n {n\choose i} g(n,i)$. Note that $f_0(x)\equiv 1$ and $f_1(x)\equiv 0$.

We have $g(n+1,i)+ g(n+1,i+1)=i\,g(n,i)$, so plugging this into (1) gives $$ \sum_{i=0}^{n+1} {n+1\choose i} g(n+1,i)= \sum_{i=0}^n {n\choose i} i\,g(n,i),\quad n\geq0.\tag 2$$

The binomial coefficients satisfy ${n\choose i} i = n {n-1\choose i-1}$, while for $i\geq 1$ we easily check that $g(n,i)=g(n-1,i-1) (x+(i-1))$. Using these and equation (2) gives us, for $n\geq 1$, \begin{eqnarray*} \sum_{i=0}^n {n\choose i}i\, g(n,i) &=& n \sum_{i=1}^{n} {n-1\choose i-1} g(n,i) \\[3pt] &=& n \sum_{i=1}^n {n-1\choose i-1} g(n-1,i-1) (x+(i-1)) \\[3pt] &=& n \sum_{i=0}^{n-1} {n-1\choose i} g(n-1,i) (x+i) \\[3pt] &=& n \left(x f_{n-1}(x) + f_{n}(x)\right). \end{eqnarray*} This shows that the polynomials $f_n$ satisfy the recurrence $$f_{n+1}(x)=n\left(xf_{n-1}(x)+f_{n}(x)\right),\ n\geq 1.\tag 3$$

Here are the first few polynomials. $$f_0(x)=1,\ f_1(x)= 0,\ f_2(x)= x,\ f_3(x)= 2x,\ f_4(x)= 3x^2+6x. $$

Induction now shows that $f_n$ has non-negative coefficients so that there is no cancellation in (3), and hence $$\deg(f_{n+1})=\max(\deg(f_{n-1})+1, \deg(f_{n})).\tag4$$

Using (4) and induction again we see that $\deg(f_n)=\lfloor n/2\rfloor$ for $n\geq 0$.