3
$\begingroup$

Let $f_1,f_2$ be two smooth quasiconcave functions defined on a convex subset of $\mathbb{R}^d$.

It is known that $f_1+f_2$ is not necessarily quasiconcave.

Does there always exist monotonically increasing functions $g_1,g_2$ such that the sum $g_1\circ f_1 + g_2\circ f_2$ is quasiconcave?

The motiavation to this question is economic. In economics, it is often assumed that people have quasiconcave utilities: if they are indifferent between $x$ and $y$, then they prefer $(x+y)/2$ to $x$ and to $y$, which is (approximately) the definition of quasiconcavity. Given two or more people, we would like to create a social welfare function by taking the sum of the utility functions of the agents. In general, such a sum is not guaranteed to be quasiconcave. However, a utility function is only a numeric representation of a preference-relation, so we are allowed to transform it using arbitrary monotone transformations and it will still represent the same preference. The question is: can we always transform the utility functions of the agents such that the resulting social-welfare function be quasiconcave?

There are two cases for which I know the answer.

  1. If $f_1$ and $f_2$ are concavifiable, then the answer is "yes", since we can choose $g_1,g_2$ such that $g_1\circ f_1$ and $g_2\circ f_2$ are concave, which makes their sum concave too, and concavity implies quasic

  2. If $f_1$ and $f_2$ are are only weakly quasiconcave, then the answer is "not always", as shown by the example of Taneli Huuskonen below

So the question now becomes: is this always possible to find such $g_1,g_2$ when $f_1,f_2$ are strictly quasiconcave*?

Reminder: $f$ is quasiconcave iff $\forall x,y: \forall r\in(0,1): f(r x + (1-r) y)\geq \min(f(x),f(y))$. It is strictly quasiconcave iff $\forall x,y: \forall r\in(0,1): f(r x + (1-r) y)> \min(f(x),f(y))$.

$\endgroup$
  • $\begingroup$ Are $f_i$ and $g_i$ meant to be continuous, or not? And what is the the domain of $f_i$: a convex subset of $\mathbb{R}^n$, or the whole space? $\endgroup$ – Pietro Majer Feb 12 '18 at 18:43
  • $\begingroup$ @PietroMajer to focus on the quasiconcavity aspect, let's assume that the $f_i$ are smooth, and the $g_i$ can be arbitrarily increasing functions. The domain can be any convex subset of $\mathbb{R}^n$. $\endgroup$ – Erel Segal-Halevi Feb 13 '18 at 8:40
  • $\begingroup$ @PeterHeinig OK, I did not know that. $\endgroup$ – Erel Segal-Halevi Feb 15 '18 at 18:20
3
+50
$\begingroup$

There is an easy counterexample. Let $$ f_1(x)=\begin{cases} x-n,&\mbox{when $2n\le x\le 2n+1$,}\\ n+1,&\mbox{when $2n+1\le x\le 2n+2$,} \end{cases} $$ and let $f_2(x)=f_1(-x)$. Now whenever $g_1$ and $g_2$ are strictly increasing, the function $g_1\circ f_1+g_2\circ f_2$ is strictly increasing on intervals $[2n,2n+1]$ and strictly decreasing between them, hence not quasiconcave.

You can replace $f_1$ and $f_2$ with smooth functions if you wish, but I cannot see any obvious way to obtain a similar counterexample with strictly quasiconcave functions.

$\endgroup$
  • 1
    $\begingroup$ Very interesting. It seems that the main problem is the fact that the functions are constant in some range, since no monotone transformation can change this behavior. [for visualization, here are some plots of this function: mathematica.stackexchange.com/a/165674/49101] $\endgroup$ – Erel Segal-Halevi Feb 12 '18 at 8:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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