0
$\begingroup$

Let $I$ be a finite set, and $A_{ij},B_{ij},x_i,y_j\ge0$. I want to find the choice of $x_i,y_j$ maximizing $$\sum_{i\in I}\sum_{j\in J}A_{ij}\min\left(x_i,B_{ij}y_j\right)\tag1$$ subject to $$\sum_{i\in I}x_i=\sum_{j\in I}y_j=1\tag2.$$ As discussed on Mathematics, the feasible region is divided into polygons such that the objective function is linear on each interval. With this approach we end with a finite number of candidate points (the vertices) we need to compute the maximum of.

However, I need to find a closed form expression for $x_i,y_j$. Are we able to achieve this? If not, are we at least able to find a (sharp) lower bound for $(1)$ for which it is easier to obtain a closed form maximizer?

$\endgroup$
8
  • $\begingroup$ Standard trick is to introduce variables $z_{ij}$ with $z_{ij}\leq x_i$ and $z_{ij}\leq B_{ij}y_j$, and to maximize:$$\sum_{i\in I}\sum_{j\in J}A_{ij}z_{ij}.$$ $\endgroup$ Sep 6, 2019 at 11:33
  • $\begingroup$ @MaxAlekseyev Thank you for your comment. I guess we're enforcing $z_{ij}\ge0$, right? It's been a while since I thought about linear optimization, but I need to implement a solver (with as less overhead as possible) for the problem. Do you have a good reference at hand? (I guess there's no hope for a closed form solution, right?) $\endgroup$
    – 0xbadf00d
    Sep 6, 2019 at 12:45
  • $\begingroup$ @MaxAlekseyev Oh, and while I see the formal point in bringing the problem into some kind of "normal" or "standard" form by introducing $z_{ij}$, this should produce a significant overhead. Is there a more direct approach for the special shape of the problem? $\endgroup$
    – 0xbadf00d
    Sep 6, 2019 at 14:17
  • 1
    $\begingroup$ Since $A_{ij}\geq 0$, you don't need to enforce $z_{ij}\geq 0$. $\endgroup$ Sep 6, 2019 at 14:19
  • $\begingroup$ @MaxAlekseyev Would be great if you could (briefly) elaborate on the other questions in my comments. $\endgroup$
    – 0xbadf00d
    Sep 8, 2019 at 19:01

0

Your Answer

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