8

I encountered the following problem and was hinted to consider the edge case.

Determine all possible values of $$S = \frac{a}{a+b+d}+\frac{b}{a+b+c}+\frac{c}{b+c+d}+\frac{d}{a+c+d}$$ where $a$, $b$, $c$ and $d$ are arbitrary positive numbers.

Hans
  • 9,804
  • 1
    Little hint: the function is continuous on the connected set $(0,+\infty)^4$, so the image is an interval. – Julien Mar 03 '13 at 18:35

1 Answers1

3

The expression is homogeneous of order zero, meaning it does not change if the vector $(a,b,c,d)$ is multiplied by a constant. So you might as well assume $a+b+c+d=1$, and we are looking at the interior points of the simplex with corners at $(1,0,0,0)$, $(0,1,0,0)$, $(0,0,1,0)$ and $(0,0,0,1)$. Further, note we can now write $$ S=\frac{a}{1-c}+\frac{b}{1-d}+\frac{c}{1-a}+\frac{d}{1-b}.$$ If you fix $b$ and $d$ then $a+c$ is fixed, and it is easy to see that the first and third terms are each concave functions of $a$ (or, equivalently, of $c$). Hence so is their sum, so the minimum appears at the ends of the allowable interval, i.e., where either $a$ or $c$ vanishes. By symmetry, we may look at just the case $c=0$. A similar argument applies the the second and fourth terms and the variables $b$ and $d$, so we may set $d=0$. But with $c=d=0$ the whole expression reduces to $a+b=1$, so $S$ has the minimum value $1$ on the closed simplex. Of course, this will not be achieved in the interior of the simplex, but the infimum is still $1$.

Edit: Three years have passed, and now a request has surfaced to finish this. So here goes:

It remains to find the supremum. If we fix $b$ and $d$, as noted the sum of the first and third terms $a/(1-c)+c/(1-a)$ is a concave function on the interval $a+c=1-b-d$, $a>0$, $c>0$. Moreover, this function is symmetric in $a$, $c$, so it has its maximum at the midpoint of the interval, where $a=c$. The same goes for the sum of the second and fourth terms, in the variables $b$, $d$. In other words, for any point in the simplex, there is another point with $a=c$ and $b=d$ with a greater value for the sum $S$. For such a point, we get $$ \frac S2=\frac a{1-a}+\frac b{1-b},$$ and now we have the constraint $a+b=\frac12$, $a>0$, $b>0$. But this function is convex, so its maximum at the endpoints of the interval. Putting $a=0$, $b=\frac12$ or $a=\frac12$, $b=0$, we get $S=2$. Of course, these values aren't achieved, but still, the supremum is $2$.

  • One too many 0's in vertex $(1,0,0,0,0)$. But a good suggestion. +1 – coffeemath Mar 03 '13 at 18:48
  • Thank you, Zev, for the posting advice. And thank you, Harald, for the solution. I expressed S in the way Harald does, but failed to see the convexity in each fraction. Thanks, Harald. – Hans Mar 03 '13 at 19:07
  • As I was too hasty in posting the answer, you should not be too hasty in thanking me. Better put the thinking cap back on. – Harald Hanche-Olsen Mar 03 '13 at 19:17
  • I am also leaving the wrong answer up because it has already attracted some attention, and having it disappear without an explanation will just confuse everybody. – Harald Hanche-Olsen Mar 03 '13 at 19:36
  • The unaltered $S$ of the OP is $S=4/3$ when $a=b=c=d=1$, so how is the supremum $1$? – coffeemath Mar 03 '13 at 22:43
  • @HaraldHanche-Olsen: Indeed I was too excited to see your solution, but you deserve the credit for the main idea any way. Your current solution would be completely correct had you swapped concave with convex and minimum with maximum. So the minimum of $1$ is achieved on the boundary of the simplex while the maximum of $\frac{4}{3}$ is achieve in the interior at $a=b=c=d$. – Hans Mar 04 '13 at 00:39
  • 1
    The maximum is not $4/3$, since if $a=c=1$ while $b=d=1/2$ we have $S=7/5=1.4>4/3=1.33..$ – coffeemath Mar 04 '13 at 05:45
  • @Hans: Indeed. Simple sign mistake. I will correct my answer when I have the time, which will not be for several hours unfortunately. I guess I have too many other things going on right now to post answers here. 8-) – Harald Hanche-Olsen Mar 04 '13 at 09:01
  • @coffeemath Hmm, interesting. I'll look into that, but see my previous comment. – Harald Hanche-Olsen Mar 04 '13 at 09:02
  • @coffeemath is Right and maximum is not $\frac{4}{3}$ at $a=b=c=d$ but $2$ at $a=c,,b=d$ and either $a=0,,b=\frac{1}{2}$ or swap $a$ and $b$ values. I will put up the justification later when I have time as well. :-) – Hans Mar 04 '13 at 13:59
  • Fixed the convex/concave maximum/minimum silliness. As for the maximum, I had a super long workday today and sleep gets priority. – Harald Hanche-Olsen Mar 04 '13 at 21:47
  • @HaraldHanche-Olsen: I am revisiting this problem. Do you have the solution? – Hans Jun 26 '16 at 19:32
  • @Hans Okay, here goes. I must have been really tired that time three years ago, for there wasn't much left. – Harald Hanche-Olsen Jun 27 '16 at 16:29
  • Thank you, @HaraldHanche-Olsen, for your excellent solution. You must have been tired that time. :-) However there is a minor confusion in the wording of the solution. You wrote $a+c<1-b-d$. I initially took it as meaning the domain $(a,c)$ where $a+c<1-b-d$. The sum $\frac{a}{1-c}+\frac{c}{1-a}$ is neither convex nor concave in that domain. So I edited that part of the wording in your solution, mainly setting $a+c=1-b-d$. Please see if you agree with the edit. You can revert back if you do not agree. – Hans Jun 28 '16 at 16:24
  • @Hans You had changed the strict inequality to non-strict, where the correct fix (as you say in your comment) is equality. (I notice that your edit is already rejected by other reviewers, so I am just letting my version stand. It is at least not wrong, though there is always room for improvement.) – Harald Hanche-Olsen Jun 28 '16 at 16:59
  • From strict to non-strict inequality was the first edit (although the main aim of that change was correcting a typo from $c$ to $b$). The latest was changing the inequality to equality, which you mentioned was correct. You did not see that change or maybe I forgot to change the most important part :-P ? I did change other wording slightly in the way I thought would make it even more unequivocal to me. That latter part was not as important as equality though. You current version is fine. I have accepted your solution. Thank you again, @HaraldHanche-Olsen! – Hans Jun 28 '16 at 18:27