How to prove that some combinatorical function has minimum on borders?

75 Views Asked by At

One maximum likelihood estimation is calculated according to equation:

$$\widehat{\ell}=\operatorname*{argmin}_\ell\mathbb{E}f(\ell,k),$$

where $f(\ell,k)$ is a function of two variables:

$$f(\ell,k)=\binom \ell k \binom {n-\ell} {m-k}.$$

One of arguments $k$ is random variable with hypergeometric distribution:

$$P(k)=\frac{\binom {\ell-g_1}{k-g_2}\binom{h_1-\ell+g_1}{h_2-k+g_2}}{\binom {h_1}{h_2}}.$$

This probability has values more then zero on interval $k\in[g_2,g_2+h_2]$.

I want to prove that if $\ell\in[g_1,g_1+h_1]$, then minimum of objective function locates on one of the borders:

$$\min\limits_\ell\mathbb{E}f(\ell,k)=\min\left(\binom {g_1} {g_2}\binom {n-g_1} {m-g_2},\binom {g_1+h_1} {g_2+h_2}\binom {n-g_1-h_1} {m-g_2-h_2}\right),$$

where all variables and constants in binomial cooefs are nonnegative integers.

To prove it I tried to calculate: $$\mathbb{E}f(\ell,k)=\sum\limits_k f(\ell,k)P(k)$$

$$\mathbb{E}f(\ell,k)={\binom {h_1}{h_2}}^{-1} \sum\limits_{k=\max(g_2,g_2+h_2+\ell-g_1-h_1)}^{\min(\ell-g_1,h_2)} \binom \ell k \binom {n-\ell} {m-k}\binom {\ell-g_1}{k-g_2}\binom{h_1-\ell+g_1}{h_2-k+g_2}$$

But when I read "Concrete mathematics" I can not find some formula or trick to use in this sum. It looks like the some modification of Vandermonde's identity, but ...

  1. how to use this to get closed form of this? It is possible that closed form doesn't exist.
  2. Another question is how to find this descrete function extremum?
  3. Am I going right way trying to calculate last expression? May be, there exist another way to prove that minimum is on borders.

Actually, I use log-likelihood instead likelihood (first equation) to estimate $\ell$ beacuse this helps to avoids big integers. I'm not sure about the equation: $$\operatorname*{argmin}\limits_\ell\mathbb{E}\log{f(\ell,k)}=\operatorname*{argmin}\limits_\ell \mathbb{E}f(\ell,k)$$
To ilustrate this problem in terms of log-likelihood I posted this function realization and sample mean over 100 realizations. We can see that hypothesis i want to prove can be statistically accepted. Existing of proof is my question to mathematical community.

Thank you for any help or comments!