# Is there an optimization scheme/algorithm that converges, for this non-convex scenario but with some special properties

I have a smooth function $$f(x) = \frac{g(x)}{h(x)}$$ that is the ratio of two smooth convex functions $$g(x)$$ and $$h(x)$$. It is known that $$f(x)$$ has a global minimum, achieved at the unique point $$x_0$$. Also is known that $$f(x)$$ has countably many local minima, and function $$f(x)$$ value at no two local minima, is the same.

Question: Is there any optimization algorithm for this scenario, which converges to the global minimum. I am hoping there is some variant of gradient descent or the likes. Appreciate any references/solutions.

Despite your claim that these functions have "special properties", the properties you've supplied still leave f and g extremely general. This means the answers you're going to get must be correspondingly general. If, for instance, you know that g is always positive or f is quadratic, or g is some kind of positive semi-definite matrix, and so on, then you should ask a new question specifying this critical information from the start.

What you're trying to do is convex-convex fractional programming. Schaible's 1976 paper "Minimization of ratios" explains how such a problem can be transformed into a quasi-convex problem

So you can make this into a quasi-convex problem. That means that you can use bisection to solve the problem to $$\epsilon$$ tolerance in $$\lceil log_2((u-l)/\epsilon)\rceil$$ steps where $$l$$ and $$u$$ bound the range of the $$t$$ variable shown above.

Depending on the functions involved, you may be able to use cvxpy to solve the problem. Some examples showing bisection of quasiconvex functions are here and here.

• Sorry I forgot to mention : f,g are strictly positive. Hope that helps : in case in this light, if it would deserve another answer, please advice accordingly. Also both$f$ and $g$ are quadratic. – Rajesh Dachiraju Feb 23 at 3:45
• need your permission to proceed if I can post another question, or if you could help on this post itself. Thanks – Rajesh Dachiraju Feb 23 at 3:50
• sorry both $g$ and $h$ are quadratic and strictly positive. not $f$ as I had mistakenly put in previous comment. – Rajesh Dachiraju Feb 23 at 3:56
• @RajeshDachiraju: Does your problem have any constraints? – Richard Feb 23 at 4:50
• None except one parameter in parameter vector is always zero. – Rajesh Dachiraju Feb 23 at 5:06