[convex optimization]Does there exist a function h such that its composition with a non-convex function g, h(g(x)), is a convex function?

47 Views Asked by At

I have a question regarding convex functions and compositions that I'm hoping someone could provide some insight on.

For a given non-convex function g(x), I'm wondering if it's possible to find another function h such that the composition h(g(x)) results in a convex function. In other words, does there exist a function h that can 'convexify' a non-convex function g when composed together?

I've been studying convex analysis and optimization, but I'm still a bit uncertain about the conditions and techniques for constructing such a function h, if it exists. Any guidance, references, or insights from the knowledgeable members here would be greatly appreciated.

Thank you in advance for your time and help!

1

There are 1 best solutions below

0
On

The answer is, technically, yes. As I hinted at in my comment, making $h$ constant satisfies the conditions. Given any $g$, and any constant function $h$ (which is convex), $h \circ g$ is constant and hence convex.

However, in the realm of , this is likely unhelpful. The unstated spirit of this question is one of convexification of optimisation problems. Can we take a non-convex objective function, compose it with a convex function, and get a nice convex problem, the solution of which tells us something about the solution of the original problem?

Zim's comment is a reasonable attempt to solidify this problem: what if we insisted that the minimisers of $h \circ g$ are precisely the minimisers of $g$? This, unfortunately, doesn't work either: the minimisers of a convex function is a convex set, so if $g$ has a non-convex set of minimisers, then $h \circ g$ must have a different set of minimisers.

Perhaps, instead of insisting that $\operatorname*{argmin} g = \operatorname*{argmin} (h \circ g)$, we insist that $\operatorname{\overline{conv}} \operatorname*{argmin} g = \operatorname*{argmin} (h \circ g)$? Again, this is not going to work. Consider, for example, $$g(x) = x^2(x - 1)^2,$$ a quartic polynomial with precisely two minima at $x = 0$ and $x = 1$. At $x = 1/2$, there is a local maximum of $1/16$. Note that the equation $g(x) = 1/32$ has exactly four solutions $x_1, x_2, x_3, x_4$ such that $$x_1 < 0 < x_2 < \frac{1}{2} < x_3 < 1 < x_4.$$ Note that, for any $h$, $h \circ g$ is constantly $h(1/32)$ on those four points. Since $h \circ g$ is convex, we would have to have $h \circ g$ be constant on $[x_1, x_4]$, which strictly contains $\operatorname{\overline{conv}} \operatorname*{argmin} g = [0, 1]$. The fact that this interval is non-trivial and horizontal implies $h$ achieves its minimum all along this interval, as must $h \circ g$, hence $$\operatorname*{argmin} (h \circ g) \supseteq [x_1, x_4] \supsetneq [0, 1] = \operatorname{\overline{conv}} \operatorname*{argmin} g.$$

If we're happy with strict containment as above, then we might as well set $h$ to be constant once again.

All of this is to say, convexification via composition appears to be a bit of a dead end. You're better off convexifying by the second Fenchel dual instead, which is guaranteed to produce a lower-semicontinuous, convex function, whose infimum agrees with the infimum of whatever function you feed into it.