2
$\begingroup$

The original problem that gave rise to this question is to solve $$ \sum_{i=1}^n\frac1{1+e^{-(x-c_i)}}=\frac n2 $$ (it comes from the need to determine difficulty of a polytomous item under the graded model in Item Response Theory).

For $n=2$ the (well, a) solution is $\frac{c_1+c_2}2$, so it should behave as certain generalized mean of the $c_i$. The question is whether there might be some explicit expression for it.

I just tried some obvious transformations, but without success. For example, the equation is equivalent to $$ \sum_{i=1}^n\tanh\frac{c_i-x}2=0. $$ Also, one may switch to polynomials: denoting $e^{-x}$ by $y$ and $e^{-c_i}$ by $a_i$ it becomes $$ \sum_{i=1}^n\frac1{1+y/a_i}=\frac n2 $$ (solution being now the geometric mean of the $a_i$ for $n=2$). Modulo some transformations this can be reformulated as finding solutions of $t\frac{d\log p}{dt}=\frac n2$ in terms of roots of a given polynomial $p(t)$ of degree $n$.

...decided to add yet another reformulation: one has to find an extremum (a root of the derivative) for $$ \prod_{i=1}^n\left(1-\frac z{b_i}\right)^{b_i^2-1} $$

$\endgroup$

1 Answer 1

3
$\begingroup$

Using your $y$ and $a_i$ coordinates, one is seeking a solution to $$\sum_{i=1}^n\frac{1}{1+y/a_i}=\frac{n}{2}.$$ Multiplying all this out this becomes some equation of the form $P(y)/Q(y)=0$ with $P(y)$ and $Q(y)$ polynomials in $y$ with coefficients involving the $a_i$.

So one seeks solutions to $P(y)=0$. Now if we let $\sigma_j$ be the $j$th elementary symmetric function in the $a_i$ (so $\sigma_1=a_1+a_2+\cdots+a_n$ and $\sigma_n=a_1a_2\cdots a_n$ then it seems to me that (perhaps up to a constant -- these things are only defined up to constants) we have $$P(y)=ny^n+(n-2)\sigma_1 y^{n-1}+(n-4)\sigma_2 y^{n-2}+\cdots-n\sigma_n.$$ I didn't check this for all $n$ but I did check it for $n=5$.

This is kind-of bad, because it shows that $y$ is a root of a "random" polynomial. For example if one sets $n=5$ and $a_1,a_2,a_3,a_4,a_5=1,2,3,4,5$ then we get a degree 5 polynomial $5y^5 + 45y^4 + 85y^3 - 225y^2 - 822y - 600$ with rational coefficients whose Galois group over the rationals is $S_5$.

This suggests to me that there will be no easy formula in general (unless you allow your formula to be powerful enough to spit out roots of quintics).

$\endgroup$
2
  • $\begingroup$ Mmm I would not say this is random since $\frac n2$ on the right is a very special thing. E. g. in the $\tanh$ reformulation it gives zero on the right. Accordingly, your $P(y)$ can be written as$$2yF'(y)-nF(y),$$ where $F$ is the polynomial with $a_i$ as roots. This is what I had in mind in the last sentence about logarithmic derivative. Roots of $F$ are known, and one has to express (one of the) roots of $P$ in terms of them. $\endgroup$ Feb 7, 2017 at 21:57
  • 1
    $\begingroup$ Yes I see your point -- "random" is not a good word in the above. On the other hand I think the Galois group computation above is an indication that "simply" expressing the roots of $P$ in terms of the roots of $F$ can't really be done. By the way, in my numerical example four roots are negative (so can't be exp of something) and the other is $2.627581098\ldots$, so there is a unique solution -- it's just not expressible in terms of elementary functions. $\endgroup$ Feb 7, 2017 at 22:05

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.