2
$\begingroup$

The Problem

Given $i, n, y$, I am trying to find a closed form solution for $a$ in the equation $$ \sum_{j=i}^n (a -j) \binom{n}{j} y^j (1-y)^{n-j} = 0 $$ Do you have any recommendations on how to solve this problem?

Background/Motivation

Let me explain. I am trying to compute the derivative of a Bernstein polynomial with respect to y. A Bernstein polynomial is defined as $$ B_{i:n}(y) = \sum_{j=i}^n \binom{n}{j} y^j (1-y)^{n-j} $$ and I've worked out the derivative to be $$ \frac{d}{dy} \bigg[ B_{i:n}(y) \bigg] = \sum_{j=i}^n \binom{n}{j} y^{j-1} (1-y)^{n-j} j \\ + \sum_{j=i}^n \binom{n}{j} y^j (1-y)^{n-j - 1} (n-j) $$ However, I want to simplify the derivative so it contains the expression $B_{i:n}$. Using the first sum of the derivative as an example, I want to do something like this: $$ \sum_{j=i}^n \binom{n}{j} y^{j-1} (1-y)^{n-j} j = \frac{1}{y} \sum_{j=i}^n \binom{n}{j} y^j (1-y)^{n-j} j \\ = \frac{a}{y} \sum_{j=i}^n \binom{n}{j} y^j (1-y)^{n-j} = \frac{a}{y} B_{i:n}(y) $$ for some value $a$ that allows us go get rid of the $j$ that we are multiplying each term of the summation by. However, I am unsure how to compute $a$. To compute $a$, I simply have to solve $$ a \sum_{j=i}^n \binom{n}{j} y^j (1-y)^{n-j} = \sum_{j=i}^n \binom{n}{j} y^j (1-y)^{n-j} j \\ \sum_{j=i}^n (a -j) \binom{n}{j} y^j (1-y)^{n-j} = 0 $$ Yet I do not know how to solve this problem for $a$. I know there are various numerical methods for root solvers but I am looking for a closed form solution. Any suggestions would be greatly appreciated.

One more thing: this problem can be generalized to solving: $$ \sum_{j=1}^n f(j) j = a \sum_{j=1}^n f(j) \\ \sum_{j=1}^n (a - j) f(j) = 0 $$ where $a \neq i$. Any recommendations on techniques/methods I can use to solve problems of this form? To me, this generalized problem is trying to remove the $j$ by which we multiply every term in a summation when that summation is defined over $j$. (Perhaps this problem formulation is a more well known problem than my specific problem)

This problem is especially difficult because according to the the Abel–Ruffini theorem, obtaining a closed form solution for the root of a polynomial such as the one we have is incredibly difficult for polynomials of degree five and higher and our polynomial seems to be of degree five or higher.

Update 1

Qiaochu Yuan, thanks your right. We can just write the solution as: $$ a = \frac{\sum_{j=i}^n j \binom{n}{j} (1-y)^{n-j} y^j} {\sum_{j=i}^n \binom{n}{j} (1-y)^{n-j} y^j} \\ = \frac{\sum_{j=i}^n j \binom{n}{j} (1-y)^{n-j} y^j} {B_{i, n}(y)} \\ $$

Any advice on how to simplify this further? Thanks!

$\endgroup$
5
  • 9
    $\begingroup$ I don't understand the question. Isn't this a linear equation in $a$? $\endgroup$ Commented Jun 24, 2022 at 18:39
  • 1
    $\begingroup$ Do you want a theoretical solution, or one that is practical for small $n$? $a$ is the expectation of a binomial random variable with parameters $n,y/(1-y)$, conditioned on being at least $i$. I don't think there is an exact closed form except in terms of special functions like hypergeometric functions. $\endgroup$ Commented Jun 25, 2022 at 3:21
  • $\begingroup$ Possible simplification for $i>0$: $$a = \frac{nyB_{i-1,n-1}(y)}{B_{i,n}(y)}.$$ $\endgroup$ Commented Jun 26, 2022 at 21:27
  • $\begingroup$ @MaxAlekseyev how did you remove the $j$ in the numerator sum? Note that $$ B_{i, n} = \sum_{j=i}^n \binom{n}{j} y^j (1-y)^{n-j} \neq \sum_{j=i}^n j \binom{n}{j} y^j (1-y)^{n-j} $$ $\endgroup$
    – E. Turok
    Commented Jun 26, 2022 at 21:32
  • $\begingroup$ @E.Turok: Use the property $j\binom{n}{j}=n\binom{n-1}{j-1}$. $\endgroup$ Commented Jun 26, 2022 at 21:53

1 Answer 1

1
$\begingroup$

$$\sum_{j=i}^n (a -j) \binom{n}{j} y^j (1-y)^{n-j} =0$$ $$\Rightarrow a=i-\frac{y \binom{n}{i+1} \, _2F_1\left(2,i-n+1;i+2;\frac{y}{y-1}\right)}{(y-1) \binom{n}{i} \, _2F_1\left(1,i-n;i+1;\frac{y}{y-1}\right)}.$$ Check that for $i=1$, this simplifies to $a=n y \left(\frac{1}{(1-y)^{-n}-1}+1\right)$, which is the correct solution.

$\endgroup$
2
  • $\begingroup$ Can you explain how you obtained this solution? Also, can you clarify what $F_1$ is? $\endgroup$
    – E. Turok
    Commented Jun 26, 2022 at 21:30
  • $\begingroup$ $_2F_1$ is the Gauss hypergeometric function; this is how Mathematica evaluates your series. $\endgroup$ Commented Jun 26, 2022 at 22:01

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .