6
$\begingroup$

The presented solution of the following problem on Art of Problem Solving using Jensen's inequality is wrong, since the function $f(x):=\sqrt[3]{1+\frac{2t}{x}} $ is convex rather than concave. How would one prove this inequality, correctly?

Let $a, b ,c $ be positive real numbers such that $ a+b+c+abc=4$. Prove that : $$\sum_{\text{cyc}}^{}\sqrt[3] {1+2ac} \le 3\sqrt [3] {3}.$$

$\endgroup$
7
  • $\begingroup$ Can you clarrify what $\sum_{cyc}$ means? $\endgroup$ – Aditya De Saha Dec 14 '19 at 18:50
  • $\begingroup$ @AdityaDeSaha: $\sum_{\text{cyc}}$ stands for cyclic summation, meaning the summing indices run through a cyclic permutation math.stackexchange.com/questions/280372/…. $\endgroup$ – Hans Dec 14 '19 at 18:59
  • $\begingroup$ I can't open the link, says you must be registered user to view the content. (try opening in incognito mode to see exact message if you want) $\endgroup$ – prosinac Dec 14 '19 at 23:24
  • $\begingroup$ @prosinac: Fixed it. How about now? $\endgroup$ – Hans Dec 14 '19 at 23:46
  • $\begingroup$ It's ok now. And yes, I agree solution is wrong, but I can't think of a right one now $\endgroup$ – prosinac Dec 15 '19 at 9:26
2
$\begingroup$

I deleted my previous solution. I give another solution.

WLOG, assume that $c = \min(a,b,c)$.

Since $x\mapsto \sqrt[3]{x}$ is concave on $(0, \infty)$, we have \begin{align} \sqrt[3]{1+2ac} + \sqrt[3]{1+2ba} + \sqrt[3]{1+2cb} &\le 2\sqrt[3]{\frac{1+2ac + 1+2cb}{2}} + \sqrt[3]{1+2ba}\\ &= 2\sqrt[3]{1+ac + cb} + \sqrt[3]{1+2ba}. \end{align} It suffices to prove that $$2\sqrt[3]{1+ac + cb} + \sqrt[3]{1+2ba} \le 3\sqrt[3]{3}.$$

We split into two cases:

1) $ba \le 1$: We have $ac + cb \le 2ba\le 2$ and thus $2\sqrt[3]{1+ac + cb} + \sqrt[3]{1+2ba} \le 3\sqrt[3]{3}$.

2) $ba > 1$: From $a+b+c+abc = 4$, we have $c = \frac{4-a-b}{ab+1}$. Also, we have $a+b \ge 2\sqrt{ab} > 2$. Thus, we have $$ac + cb = \frac{(4-a-b)(a+b)}{ab + 1} = \frac{4 - (a+b - 2)^2}{ab+1} \le \frac{4 - (2\sqrt{ab} - 2)^2}{ab+1}.$$ Thus, it suffices to prove that $$2\sqrt[3]{1+ \frac{4 - (2\sqrt{ab} - 2)^2}{ab+1}} + \sqrt[3]{1+2ba} \le 3\sqrt[3]{3}.$$ Let $x = \sqrt{ba}$. Then $1 < x \le 2$. It suffices to prove that for $1< x\le 2$, $$2\sqrt[3]{1+ \frac{4 - (2x - 2)^2}{x^2+1}} + \sqrt[3]{1+2x^2} \le 3\sqrt[3]{3}$$ or $$2\sqrt[3]{1 + \frac{2(3x-1)(1-x)}{3(x^2+1)}} + \sqrt[3]{\frac{1+2x^2}{3}} \le 3.$$ Note that $\sqrt[3]{1+u} \le 1 + \frac{u}{3}$ for all $u > -1$ since $(1+\frac{u}{3})^3 - (1+u) = \frac{1}{27}u^2(u+9)$. Thus, it suffices to prove that $$2\left(1 + \frac{1}{3} \frac{2(3x-1)(1-x)}{3(x^2+1)}\right) + \sqrt[3]{\frac{1+2x^2}{3}} \le 3 \tag1$$ or $$\sqrt[3]{\frac{1+2x^2}{3}} \le \frac{21x^2-16x+13}{9(x^2+1)}$$ or $$\frac{1+2x^2}{3} \le \left(\frac{21x^2-16x+13}{9(x^2+1)}\right)^3$$ or $$\frac{2(x-1)^2(-243x^6-486x^5+3051x^4-3996x^3+4527x^2-2102x+977)}{729(x^2+1)^3}\ge 0.$$ It is easy to prove that $-243x^6-486x^5+3051x^4-3996x^3+4527x^2-2102x+977 > 0$ for $1 < x\le 2$.
EDIT: Indeed, By letting $x = 1 + v$ for $0 < v \le 1$, we have \begin{align} &-243x^6-486x^5+3051x^4-3996x^3+4527x^2-2102x+977\\ =\ & -243v^6-1944v^5-3024v^4-1512v^3+2340v^2+3280v+1728\\ \ge \ & -243v^2-1944v^2-3024v^2-1512v^2+2340v^2+3280v+1728\\ =\ & -4383v^2+3280v+1728\\ \ge \ & -4383v+3280v+1728\\ =\ & -1103v+1728\\ >\ & 0. \end{align} We are done.

$\endgroup$
4
  • $\begingroup$ Looks nice, but could you please explicitly prove the last inequality? $\endgroup$ – Hans Dec 16 '19 at 18:39
  • $\begingroup$ @Hans I updated it. $\endgroup$ – River Li Dec 17 '19 at 1:16
  • $\begingroup$ +1. Impressive proof, overall! How do you come up with the relaxed inequality $(1)$ especially when the application of $\sqrt[3]{1+u} \le 1 + \frac{u}{3}$ to both cubic roots would have failed? $\endgroup$ – Hans Dec 17 '19 at 3:51
  • $\begingroup$ @Hans First, try simple bounds, if not enough, try better bounds. $\endgroup$ – River Li Dec 17 '19 at 4:04
3
$\begingroup$

TLDR

A standard computer-assisted (but rigorous) proof is to use Lagrange multipliers method together with interval arithmetic libraries such as IntervalRoots.jl.


We are optimizing within a compact set in $\mathbb R^3$ as shown below

enter image description here

So there exist maximum points, either in the interior, or on the boundaries.

We can use Lagrange method for the interior. Let $$ f(a, b, c)=\sqrt[3]{2 a b+1}+\sqrt[3]{2 a c+1}+\sqrt[3]{2 b c+1}-3 \sqrt[3]{3} $$ and $$ g(a,b,c,l) = f(a,b,c)+l (a b c+a+b+c-4). $$ Then we just to find the critical points of $g$, i.e., solve $\nabla g = 0$, i.e., \begin{align} \frac{2 b}{3 (2 a b+1)^{2/3}}+\frac{2 c}{3 (2 a c+1)^{2/3}}+l (b c+1)&=0, \\ \frac{2 a}{3 (2 a b+1)^{2/3}}+l (a c+1)+\frac{2 c}{3 (2 b c+1)^{2/3}}&=0, \\ l (a b+1)+\frac{2 a}{3 (2 a c+1)^{2/3}}+\frac{2 b}{3 (2 b c+1)^{2/3}}&=0, \\ a b c+a+b+c-4&=0. \end{align} Look at this for a while and you will see that one solution is $$ a=b=c=1, l = -\frac{2}{3\ 3^{2/3}} $$ and this should give us the maximum of $f(1,1,1)=0$. To rule out other solutions, we can use rigorous numerics libararies like IntervalRoots.jl.

It's not difficult to see that the solution $(a,b,c,l)$ can only be within $[0,4]^3 \times [-55,0]$. The following Julia code finds all such solutions rigorously.

using IntervalArithmetic, IntervalRootFinding, ForwardDiff

f((a, b, c)) = -3*3^(1/3) + (1 + 2*a*b)^(1/3) + (1 + 2*a*c)^(1/3) + (1 + 2*b*c)^(1/3)
g((a, b, c, l))=f((a, b, c))+l*(a + b + c + a*b*c - 4)
∇g = ∇(g)
box = IntervalBox(0..4,3)×(-55..0)
rts = roots(∇g, box, Krawczyk, 1e-5)
println(rts)

And the result is

Root{IntervalBox{4,Float64}}[Root([0.999999, 1.00001] × [0.999999, 1.00001] × [0.999999, 1.00001] × [-0.3205, -0.320499], :unique)]

To see why only chekcing $l \in [-55,0]$ suffices, note that $$ l = -\frac{2 \left(c (2 a b+1)^{2/3}+b (2 a c+1)^{2/3}\right)}{3 (2 a b+1)^{2/3} (2 a c+1)^{2/3} (b c+1)} $$ Using $a, b, c \ge 0$ at the bottom and $a, b, c \le 4$ at the top shows that $l > -55$.

This in fact proves (not just verifies) our conjecture that there is only one solution of $\nabla g=0$ (if there is no bug in the library).

However, to be sure that maximum points do not appear on the boundary, we still need to check, for instance $a=0$. This reduces to find maximum of $$ \sqrt[3]{2 (4-c) c+1}-3 \sqrt[3]{3}+2, $$ which is $$ 2-3 \sqrt[3]{3}+3^{2/3} < 0 $$ when $c = 2$.

$\endgroup$
7
  • $\begingroup$ By "minimum", you actually mean maximum, right? 1) The solution you found is a stationary point. But why is it the maximum? 2) Could you please show "the solution $(a,b,c,l)$ can only be within $[0,4]^3 \times [-55,0]$"? $\endgroup$ – Hans Dec 15 '19 at 20:51
  • $\begingroup$ @Hans, Thanks. It should be maximum. I choose the range because $a,b,c$ can only be in $[0,4]$ and you can see easily from the equations that $l$ can not be less than $-55$. $\endgroup$ – ablmf Dec 15 '19 at 21:18
  • $\begingroup$ Sure, 𝑎,𝑏,𝑐∈[0,4], but why $l\ge-55$? More essentially, could you please answer question 1) of my last comment? $\endgroup$ – Hans Dec 15 '19 at 22:41
  • $\begingroup$ I like the way you intentionally used bold font for the words that sound suspicious. I am suspicious, but it seems like a great library! Also the approach to solution is nice and automated, although, I'm sure that there is a "nice" solution to this problem. As for questions, it shows it's a stationary point, and to see it's max you only need to verify that at edges, the values are less than at that point. This is easy for $a, b, c$ edges, and for $l$ I can't see it now but hope it is not a big problem. $\endgroup$ – prosinac Dec 15 '19 at 23:15
  • $\begingroup$ @prosinac: If you claim checking the edges is not problem, please show it is not. Otherwise, it does not constitute a proof! $\endgroup$ – Hans Dec 15 '19 at 23:33

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.