Assume I have a particular function $u(x)= \sqrt x $ (note: $u'(x)\geq0, u'(x) \leq 0, \forall x$)

Assume that I have two continuous, probability distributions over $x$, $F(x)$ and $G(x)$ but that they are unknown (we know they exist but we dont know what they look like)

Assume the following about the expectations given the two distributions: $\int u(x)dF(x) \geq \int u(x)dG(x)$


Given the information above, based on one specific functional form of $u(x)$, can I conclude: $F(x)\leq G(x), \forall x$?


If I assume $u(x)$ such that $\quad u'(x)\geq0, \quad u'(x) \leq 0 \quad \forall x$ (but do not specify what functional form $u(x)$ takes)

and I know $\int u(x)dF(x) \geq \int u(x)dG(x)$

can I conclude: $F(x)\leq G(x), \forall x$ ?


For (1), the answer is clearly no. For example, let $F(x)$ be the uniform distribution over $[0,1]$ and $G(x)$ be the cdf of the degenerate distribution $$\mathbf P(X=1/2)=1.$$ Then it is easy to see that $$F(x)=\begin{cases}0, &\text{if }x<0\\x,&\text{if }0\le x<1\\1,&\text{if } x\ge 1\end{cases},\,\,\,G(x)=\begin{cases}0,&\text{if }x<1/2\\1,&\text{if }x\ge 1/2\end{cases}.$$ Clearly, (Note that although $G$ is not continuous here, it can be arbitrarily approximated by a distribution function that is continuous) $$\int_{-\infty}^\infty u(x)\,dF(x)=\frac 23< \int_{-\infty}^\infty u(x)\,dG(x)=\sqrt{\frac 12}.$$ But clearly, neither "$F(x)\ge G(x)$ for all $x$" nor "$G(x)\ge F(x)$ for all $x$" holds.

For (2), if you mean "for SOME $u(x)$ satisfying $u'\ge 0$ and $u''\le 0$, the inequality holds", then the answer is clearly the same as (1), since $u(x)$ could rightly be the function you specified in (1). But if you mean "for ALL $u(x)$ satisfying $u'\ge 0$ and $u''\le 0$, the inequality holds", then the answer is yes, because $F$ first-order stochastically dominates $G$ if and only if $$\int u(x)\,dF(x)\ge\int u(x)\,dG(x)$$ for EACH concave and increasing function $u(x)$, while each concave and increasing function can be approximated by a function which is twice differentiable, increasing, and concave.

  • $\begingroup$ Hi Onol, Thank you for responding. Well then I am not sure what to say any more. If I know $F(x)\leq G(x), \forall x$ I can always conclude $\int u(x)dF(x) \geq \int u(x)dG(x)$, for all nondecreasing $u$, I dont need to check what possible functions satify it as long as I know they are nondecreasing. but based on what your saying and what I suspected I cannot say if $\int u(x)dF(x) \geq \int u(x)dG(x)$ then $F(x)\leq G(x), \forall x$ unless I check every possible $u$ in the universe. Just because I have $\int u(x)dF(x) \geq \int u(x)dG(x)$ for 1 specific function $u(x)$ I cannot assume FOSD. $\endgroup$
    – jessica
    Jun 13 '17 at 19:36
  • $\begingroup$ For case (1) mentioned above. Could I at least conclude SOSD? Adding the assumption that $u$ is concave. For some that is under consideration? ie I dont have to check for all $u$'s. $\endgroup$
    – jessica
    Jun 13 '17 at 19:53
  • $\begingroup$ @jessica For your first comment, yes, you cannot conclude $F(x)\le G(x)$ for all $x$ just based on testing the integral inequality for SOME function $u$. You know, that property is often useful in the converse way: if you know that $F$ FOSD $G$ then for any function nondecreasing you can obtain the integral inequality. $\endgroup$
    – OnoL
    Jun 13 '17 at 21:54
  • $\begingroup$ @jessica For your second comment, unfortunately, the answer is still no. To illustrate, let us shift the degenerate distribution a little bit to $\mathbf P(X=1/2-\epsilon)=1$ for some sufficiently small $\epsilon>0$. Then you can still get the integral inequality for $u(x)=\sqrt x$, but it is easy to see that $\int_0^1 xdF(x)=1/2>\int_0^1xdG(x)=1/2-\epsilon$, which means that $G$ fails to second order stochastically dominate $F$. $\endgroup$
    – OnoL
    Jun 13 '17 at 22:03
  • $\begingroup$ Onol, I really appreciate your help. The reason I am asking these questions is because lets say you have a "big U" $U(L)$ expected utility function that is VNM, with a concave bernuli utility, and it satisfies all the necessary axioms of EUH. Since $U(L)$ is a utility representing preferences of lotteries, we can say: $U(L_1)>U(L_2)$ IFF agent prefers lottery $L_1$ to $L_2$. From what I gather from our discussion is even if an agent, with a VNM function that satifies EUH, and has concave bernulli utility, a preference lottery $L_1$ to $L_2$ does not imply either $L_1$ FOSD or SOSD $L_2$ $\endgroup$
    – jessica
    Jun 13 '17 at 22:28

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.