I would like to optimize (minimize) the following expression in order to find the functional minimizer $g$ (which should be at least once differentiable): $$ \int_0^\infty g'(x) f(x) \ dx $$ where $f(x)$ is the density of the log-normal distribution, and also subject to the following constraints: $$\forall x>0: g'(x)>0, g(x) > 0, g(0)=0, g(1)=1, \lim\limits_{x\to\infty} g(x) = +\infty, \lim\limits_{x\to\infty} g'(x) = +\infty$$

Any ideas for that? It seems that the standard Euler-Lagrange equation cannot be used (maybe this means there is no minimizer?).


1 Answer 1


Proposition. There is no minimiser.

Proof: An equivalent version of your problem is as follows.

Statement. We wish to minimise $$\int_{-\infty}^\infty xe^{q(x)-x^2}\,dx$$ where $q(x)$ is a strictly increasing function with $q(a)=0$ for a given $a\in\Bbb R$ and $$\lim_{x\to\pm\infty}q(x)=\pm\infty,\quad\lim_{x\to+\infty}q'(x)e^{q(x)-x}=+\infty.$$

The proof for the equivalence is provided at the end of this post.

Now, since $\int_{-\infty}^\infty xe^{q(x)-x^2}\,dx$ is positive by the original integral $\int_0^\infty g'(x)f(x)\,dx>0$, it suffices to choose a sequence of functions $q_A(x)$ such that the integral becomes arbitrarily close to zero.

This is shown by Iosif Pinelis on MathOverflow where we take $$q_A(x):=k(x-a)\,1_{x\le a+A}+(kA+2(x-A))\,1_{x>a+A}$$ for some $k,A\in(0,\infty)$. By the dominated convergence theorem, we have $$\lim_{k\to0^+}\lim_{A\to+\infty}\int_{-\infty}^\infty xe^{q_A(x)-x^2}\,dx=\lim_{k\to0^+}\int_{-\infty}^\infty xe^{k(x-a)-x^2}\,dx=\int_{-\infty}^\infty xe^{-x^2}=0$$ and so we have a zero infimum. $\quad\square$

Proof of Statement: Let $g=e^h$ so that $g'=h'e^h$. Then the constraints become $$h'(x)>0,h(1)=0,\lim_{x\to0}h(x)=-\infty,\lim_{x\to\infty}h(x)=+\infty,\lim_{x\to\infty}h'(x)e^{h(x)}=+\infty$$ and the functional is \begin{align}\int_0^\infty g'f\,dx&=[gf]_0^\infty-\int_0^\infty gf'\,dx\\&=\frac1{\sigma\sqrt{2\pi}}\left(\lim_{x\to\infty}e^{h(x)-\log x-\frac{(\log x-\mu)^2}{2\sigma^2}}-\int_0^\infty\frac{e^{h(x)-\frac{(\log x-\mu)^2}{2\sigma^2}}(\mu-\sigma^2-\log x)}{\sigma^2 x^2}\,dx\right)\end{align} where we assume the limit exists. Let $h(x)=p(u)$ where $x=e^u$. Then $$p'(u)>0,p(0)=0,\lim_{u\to\pm\infty}p(u)=\pm\infty,\lim_{u\to+\infty}p'(u)e^{p(u)-u}=+\infty$$ and \begin{align}\sigma\sqrt{2\pi}\int_0^\infty g'f\,dx&=\lim_{u\to\infty}e^{p(u)-u-\frac{(u-\mu)^2}{2\sigma^2}}+\frac1{\sigma^2}\int_{-\infty}^\infty e^{p(u)-u-\frac{(u-\mu)^2}{2\sigma^2}}(u+\sigma^2-\mu)\,du.\end{align} For both $\int_0^\infty$ and $\int_{-\infty}^0$ to be finite, we must have $$\lim_{u\to\pm\infty}p(u)-u-\frac{(u-\mu)^2}{2\sigma^2}=-\infty\implies p(u)=o(u^2)$$ since $p'(u)>0$ for all $u$ means $p$ cannot have a quadratic term. So \begin{align}\sigma^3\sqrt{2\pi}\int_0^\infty g'f\,dx&=\int_{-\infty}^\infty e^{p(u)-u-\frac{(u-\mu)^2}{2\sigma^2}}(u+\sigma^2-\mu)\,du\\&=\frac{e^{\sigma^2/2-\mu}}{2\pi\sigma^2}\int_{-\infty}^\infty ve^{q(v)-v^2}\,dv\end{align} where $q(u)=p(u/\sqrt{2\sigma^2}-\sigma^2+\mu)$ under the same constraints as $p(u)$ except $q((\sigma^2-\mu)\sqrt{2\sigma^2})=0$. In the statement, we simply define $a:=(\sigma^2-\mu)\sqrt{2\sigma^2}$. $\quad\square$


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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