Quadratic integer program with positive definite matrix

91 Views Asked by At

Given a positive definite matrix $A \in M_n$ and a vector $b \in \mathbb{R}^n$,

$$\begin{array}{ll} \text{minimize} & \frac{1}{2} x^T A x + b^T x\\ \text{subject to} & x \in \mathbb{Z}^n\end{array}$$

That is the program I am currently writing in Java. I succeeded in finding a solution which has real values. However, I hit a wall when it comes to finding integer solutions.

I considered genetic algorithms, checking every integer vector "around" real solution.

I would like to understand how to solve this problem. It doesn't matter if the solution is the optimal one, I wish it was an easy to understand one (and if it wasn't perhaps the worst case in terms of being optimal) so that I can get a better understanding of what is going on and keep researching in that area.

2

There are 2 best solutions below

2
On

On $\mathbb{R}^n$, the minimum is reached for $x_0=-A^{-1}b$. We research a solution in the form $x_1=x_0+h\in\mathbb{Z}^n$.

EDIT. Correction. Then $F(x_1)-F(x_0)=1/2h^TAh$. It's more difficult! We are looking for some $h$ pretty much in the direction of the eigenvector of $A$ associated to the smallest eigenvalue of $A$.

0
On

I came across this document: https://pdfs.semanticscholar.org/5f9f/d8dadd4b6efd4d8549c204ce14f1e339c645.pdf It seems to be what I'm looking for. It seems a bit hard to understand at my level but i will eventually get through with it.