2
$\begingroup$

I am studying a certain maximisation problem (coming from some sort of likelyhood estimations); after a number of generalisations I need to examine

$$f(x_1, x_2,y) = a_1\ln x_1 +a_2 \ln x_2 + \ln y - \left(x_1 y+ x_2 y + cx_1+\frac {x_2}{c} \right). $$

We have $c>0$, $c\ne 1$, $x_i>0$, $y>0$, $a_i\ge 1$.

It is quite easy to show that the equation $\nabla f=0$ has a unique solution and that in that point $f$ is strictly concave, hence this point is a local maximum. I can give exact analytical expressions; they are a bit cumbersome, however.

One can also show that this function is bounded, because $$f(x_1,x_2,y)= \frac{a_1}{a_1+a_2} \ln (x_1y)-x_1 y\\ +\frac{a_2}{a_1+a_2} \ln (x_2y)-x_2 y\\+ a_1\left(1-\frac{1}{a_1+a_2}\right)\ln x_1 - cx_1\\ +a_2\left(1-\frac{1}{a_1+a_2}\right)\ln x_2 - x_2/c.$$

Each of the four terms in this sum admits a unique global maximum, hence $f$ is bounded.

I'd appreciate any hints and/or links helping to show that $f$ admits a global maximum.

$\endgroup$
4
  • 2
    $\begingroup$ From the decomposition, you see that for every $C$ there are $0 < m_i < M_i$ such that $x_i < m_i \lor x_i > M_i \implies f(x_1,x_2,y) < C$. Then you find $0 < n < N$ so that for $m_i \leqslant x_i \leqslant M_i$ you have $y < n \lor y > N \implies f(x_1,x_2,y) < C$. $\endgroup$ Jun 26 '15 at 13:17
  • $\begingroup$ @DanielFischer can we say that $L_i = \sup_{y} \left(\frac{a_i}{a_1+a_2} \ln (x_iy)-x_i y \right)= \frac{a_i}{a_1+a_2}\ln \frac{a_i}{a_1+a_2}-\frac{a_i}{a_1+a_2}$ and then write $f(x_1,x_2,y)\le L_1+L_2+a_1\left(1-\frac{1}{a_1+a_2}\right)\ln x_1 - cx_1 +a_2\left(1-\frac{1}{a_1+a_2}\right)\ln x_2 - x_2/c$. After that we say that for an arbitrary $C<0$ there exist closed intervals $I_i\subset (0,+\infty)$ such that $f\le C$ whenever $x_i\notin I_i$ and conclude without finding the boundaries for $y$? $\endgroup$ Jun 26 '15 at 14:29
  • 1
    $\begingroup$ If you don't constrain $y$, how would you know that $f$ never exceeds your local maximum for $x_i \in I_i$ and $y$ sufficiently large/sufficiently close to $0$? I'd think it's simpler to get the boundaries for $y$, then you have $f \leqslant C$ for $(x_1,x_2,y) \notin I_1\times I_2 \times I_y$. Choose $C$ small enough, and you can conclude that your local maximum is global, for $f$ attains a maximum in the interior of the cuboid, and that is the global maximum and a local maximum. $\endgroup$ Jun 26 '15 at 14:38
  • $\begingroup$ @DanielFischer but of course! If you post your comment as an answer, I will be able to upvote and accept it. Thank you! $\endgroup$ Jun 26 '15 at 14:49
1
$\begingroup$

If you show the existence of a compact set $K$ such that outside $K$ the values of $f$ are strictly smaller than the value at the critical point, it follows that the (unique) global maximum is attained at the (unique) critical point. For $f$ attains a maximum on $K$ by continuity, and that maximum is then the global (and hence also local) maximum, thus it can only be attained at the critical point.

The shape of $f$ allows to deduce the existence of such a compact set. For $A,B > 0$, the function $g_{A,B}(x) = A\ln x - Bx$ tends to $-\infty$ as $x\to 0$ or $x\to +\infty$, hence is bounded above. You have decomposed $f$ as the sum of four functions of that type.

From the term

$$a_1\biggl(1-\frac{1}{a_1+a_2}\biggr)\ln x_1 - cx_1,$$

by the boundedness from above of the other terms, it follows that for any $C\in \mathbb{R}$ there is a compact interval $I_1 \subset (0,+\infty)$ such that $f(x_1,x_2,y) < C$ for $x_1 \notin I_1$. Ditto we find a compact $I_2 \subset (0,+\infty)$ such that $f(x_x,x_2,y) < C$ for $x_2 \notin I_2$.

Then we can use

$$\frac{a_1}{a_1+a_2}\ln (x_1y) - x_1y \leqslant \frac{a_1}{a_1+a_2}\ln (M_1y) - m_1y$$

where $I_1 = [m_1,M_1]$ to deduce the existence of a compact $I_y \subset (0,+\infty)$ such that $f(x_1,x_2,y) < C$ whenever $x_1\in I_1$ and $y\notin I_y$. The product $I_1\times I_2 \times I_y$ satisfies our requirements if we choose $C$ less than (or equal to) the value at the critical point.

$\endgroup$

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.