Sum of binomial terms, but with fixed exponents and a variable proportion

65 Views Asked by At

I am working on a problem where I have to calculate the summation of "binomial terms":

$\sum_{t=0}^{N} (t/N)^k (1 - t/N)^{N-k}$

where k and N are fixed (we are indexing over t). Equivalently, we can think of this as

$\sum_{t=0}^{1} (t)^k (1 - t)^{N-k}$ where $t=0, \, \frac{1}{N}, \, \frac{2}{N}, ..., 1$

In plain English, we are summing binomial terms over equally-spaced proportions, e.g. for N=10 we would be summing over t = 0/10, 1/10, 2/10, ..., 9/10, 10/10 (for some known k, N).

Note that the usual binomial expansion handles the summation for indexing by variable k, but in my case k will be fixed.

For context, I am trying to determine the average of a discretely sampled Bezier curve (sampled across consistent intervals). This requires calculating this sum over all k<N, for which there may even be a shortcut that I'm unaware of. The continuous case is actually easy (just average the control points) but I'm struggling with the discrete case.

Thanks in advance.

1

There are 1 best solutions below

0
On

I'm not sure there is an exact solution, but I think I found an approximation: If you rewrite the summand using $x=e^{\log x}$ you get \begin{align} S_n &= \sum_{t=1}^{n}e^{k \log \frac{t}{n}}e^{(n-k)\log (1-\frac{t}{n})} \approx \sum_{t=1}^{n}e^{k \log \frac{t}{n}}e^{-(n-k)\frac{t}{n}} \\ &= \sum_{k=1}^{n}e^{k \log t-t+\frac{kt}{n} - k \log n} = e^{-k \log n}\sum_{k=1}^{n}t^k e^{-c_1 t} \end{align} For some constant $c_1$. I don't know if the last expression exists in the closed form, but it is comparable to the lower incomplete Gamma function: $$ S_n \approx e^{-k \log n}\int_{0}^{n}x^ke^{-c_1 x}dx = \frac{e^{-k \log n}}{c_1^k}\int_{0}^{c_1n}v^ke^{-v}dv = \frac{e^{-k \log n}}{c_1^k}\gamma(k+1, c_1n) $$