2
$\begingroup$

Take some $d,C\in\mathbb{C}$ and let

$$A_{n}=\sum_{m=0}^{n}\binom{n}{m}\left(1+\frac{m}{n}\right)^{d}C^{m}.$$

I want to calculate the asymptotics as $n$ tends to $\infty$. Numerically in a few examples and proven for $d\in\mathbb{Z}_{\geq0}$ it has leading order given by

$$A_{n}\sim (C+1)^{n}\left(1+\frac{C}{C+1}\right)^{d}.$$

My proof for $d\in\mathbb{Z}_{\geq0}$ used

$$A_{n}=(C+1)^{n}\sum_{k=0}^{n}\binom{d}{k}\left(\frac{C}{C+1}\right)^{k}\sum_{r=0}^{\infty}\frac{\binom{d-k}{r}(n)_{k}S(r+k,k)}{\binom{r+k}{r}n^{r+k}}$$

where $S(r+k,k)$ are the Stirling numbers of the second kind and as $d\in\mathbb{Z}_{\geq0}$ the sums are just over $k=0,...,d$ and $r=0,...,d-k$.

Any ideas would be great! Thanks in advance!

$\endgroup$
0

2 Answers 2

1
$\begingroup$

If you take out a factor $(C+1)^n$, this is

$$ (C+1)^n\sum_{m=0}^n\binom nm\left(1+\frac mn\right)^d\left(\frac C{C+1}\right)^m\left(\frac1{C+1}\right)^{n-m}\;. $$

Setting $p=\frac C{C+1}$ and $1-p=\frac1{C+1}$ shows that this is an expectation with a “complex-valued binomial distribution”:

$$ (C+1)^n\sum_{m=0}^n\left(1+\frac mn\right)^d\binom nmp^m(1-p)^{n-m}\;. $$

With increasing $n$, the distribution gets concentrated around the “mean” $np$, so we get a first approximation by substituting $p=\frac C{C+1}$ for $\frac mn$:

$$ (C+1)^n\left(1+\frac C{C+1}\right)^d\;. $$

This can be used to systematically derive the asymptotic expansion in inverse powers of $n$ by expanding the “random variable” in powers of the deviation from the mean:

\begin{eqnarray} \left(1+\frac mn\right)^d &=&\left(1+p+\frac mn-p\right)^d \\ &=& (1+p)^d\sum_{k=0}^\infty\binom dk\left(\frac{m-np}{n(1+p)}\right)^k\;, \end{eqnarray}

which yields the sum in terms of the central moments $\mu_k$ of the binomial distribution:

$$ (C+1)^n(1+p)^d\sum_{k=0}^d\binom dk(n(1+p))^{-k}\mu_k(n,p) \\ = (C+1)^n\left(1+\frac C{C+1}\right)^d\sum_{k=0}^d\binom dk\left(n\left(1+\frac C{C+1}\right)\right)^{-k}\mu_k\left(n,\frac C{C+1}\right)\;. $$

The first six central moments of the binomial distribution are given at Wikipedia, and a recurrence relation for them is derived in Closed-Form Expressions for the Moments of the Binomial Probability Distribution by Andreas Knoblauch. As might be expected from the linear increase of the variance with $n$, the $k$-th central moment grows with the $\left\lfloor\frac n2\right\rfloor$-th power of $n$, so that we need two central moments to get one inverse power of $n$ in the expansion. Up to $O \left(n^{-2}\right)$, with the well-known central moments $\mu_0=1$, $\mu_1=0$ and $\mu_2=np(1-p)$, the result is

$$ (C+1)^n\left(1+\frac C{C+1}\right)^d\left(1+\frac1n\binom d2\frac C{(2C+1)^2}+O\left(\frac1{n^2}\right)\right)\;. $$

Note that while the binomial distribution is usually only used for $p\in[0,1]$, the central moments are purely algebraic quantities that have the same algebraic form for any $p\in\mathbb C$.

$\endgroup$
2
  • $\begingroup$ Hey thanks! Does the argument still work when C/(C+1) is not in the interval [0,1]? As I guess you approximate the delta(x-C/(C+1)) and then only integrate over [0,1] right? $\endgroup$
    – Campbell
    May 7, 2020 at 8:57
  • $\begingroup$ @Campbell: Sorry, I'd forgotten that you explicitly wrote that $C\in\mathbb C$. The argument didn't really work for complex values the way I'd formulated it – one would have had to take into account a complex phase in the Gaussian – I completely rewrote the answer based on the central binomial distribution, which not only makes for a cleaner argument but also yields the entire asymptotic expansion in inverse powers of $n$. $\endgroup$
    – joriki
    May 7, 2020 at 18:15
0
$\begingroup$

Just another idea: $$\left(1+\frac{m}{n}\right)^d=\frac{\partial^d}{\partial x^d}\Big[e^{(1+m/n)x}\Big]_{x=0}\implies(1+C)^{-n}A_n=\frac{\partial^d}{\partial x^d}\left[e^x\left(\frac{1+Ce^{x/n}}{1+C}\right)^n\right]_{x=0}.$$

$\endgroup$

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.