4

Let $ a, b, c > 1 $. Prove that $ a^{\log_b(a)} + b^{\log_c(b)} + c^{\log_a(c)} \geq a^{\sqrt{\log_b(a)}} + b^{\sqrt{\log_c(b)}} + c^{\sqrt{\log_a(c)}} \geq a + b + c. $

My approach: I denoted $ \log_b(a) = x , \log_c(b) = y , and \log_a(c) = z$. Then our inequality becomes $ a^x + b^y + c^z \geq a^{\sqrt{x}} + b^{\sqrt{y}} + c^{\sqrt{z}} \geq b^x + c^y + a^z. $ I tried applying rearrangement inequality, but I couldn't find the ordering of $(x, y, z)$.

River Li
  • 49,125

1 Answers1

3

For the left inequality:

Let $x := \frac12 \ln \ln a, y := \frac12 \ln \ln b, z := \frac12 \ln \ln c$. The desired left inequality is written as $$\mathrm{e}^{\mathrm{e}^{4x - 2y}} + \mathrm{e}^{\mathrm{e}^{4y - 2z}} + \mathrm{e}^{\mathrm{e}^{4z - 2x}} \ge \mathrm{e}^{\mathrm{e}^{3x - y}} + \mathrm{e}^{\mathrm{e}^{3y - z}} + \mathrm{e}^{\mathrm{e}^{3z - x}}. \tag{1}$$

Note that $u \mapsto \mathrm{e}^{\mathrm{e}^{u}}$ is convex on $\mathbb{R}$ (easy by checking second derivative).

Let $U := 4x - 2y, V := 4y - 2z, W := 4z - 2x, X := 3x - y, Y := 3y - z, Z := 3z - x$.

Since (1) is cyclic, assume that $x = \max(x, y, z)$.

We split into two cases.

Case 1. $x \ge y \ge z$

Clearly, we have $U + V + W = X + Y + Z$.

We have $U - W = 6x - 2y - 4z \ge 0$, $V - W = 4y - 6z + 2x \ge 0$, $X - Z = 4x - y - 3z \ge 0$, $Y - Z = 3y - 4z + x \ge 0$ which results in $W = \min(U, V, W)$ and $Z = \min(X, Y, Z)$. We have $U + V - (X + Y) = x - z \ge 0$.

Also, we have $U - X = x - y \ge 0$ and $V - Y = y - z \ge 0$ which results in $\max(U, V) \ge \max(X, Y)$.

Thus, $(X, Y, Z)$ is majorized by $(U, V, W)$. By Karamata's inequality, (1) is true.

Case 2. $x \ge z \ge y$

Clearly, $U + V + W = X + Y + Z$.

We have $U - W = 6x - 2y - 4z \ge 0$, $U - V = 4x - 6y + 2z \ge 0$, $X - Z = 4x - y - 3z \ge 0$, $X - Y = 3x - 4y + z \ge 0$ which results in $U = \max(U, V, W)$ and $X = \max(X, Y, Z)$. We have $U - X = x - y \ge 0$.

Also, we have $U + V - (X + Y) = x -z \ge 0$, and $U + W - (X + Z) = z - y \ge 0$ which results in $\max(U + V, U + W) \ge \max(X + Y, X + Z)$.

Thus, $(X, Y, Z)$ is majorized by $(U, V, W)$. By Karamata's inequality, (1) is true.

We are done.

River Li
  • 49,125
  • I understood the second inequality thanks to the link but as in for the first one I have no idea how to approach – math.enthusiast9 Mar 25 '24 at 15:58
  • @DS typo. left inequality. – River Li Mar 26 '24 at 00:04
  • @math.enthusiast9 typo, for left inequality. I will complete it. – River Li Mar 26 '24 at 00:05
  • nice! Also, in a contest, can we directly state (U,V,W) majorizes (X,Y,Z) (your proof is little bit long but obvious)? – D S Mar 26 '24 at 11:05
  • @DS I don't know if there is simple approach for majorization. Another way is to use the fact that $Y$ is majorized by $X$ if there exists a row-stochastic matrix $A$ such that $AX = Y$. We have $$\begin{pmatrix} 11/14 & 1/7 & 1/14 \ 1/14 & 11/14 & 1/7 \ 1/7 & 1/14 & 11/14 \end{pmatrix}\begin{pmatrix} 4x - 2y \ 4y - 2z \ 4z - 2x \end{pmatrix} = \begin{pmatrix} 3x - y \ 3y - z \ 3z - x \end{pmatrix}. $$ Thus, $(3x-y, 3y-z, 3z-x)$ is majorized by $(4x-2y, 4y-2z, 4z - 2x)$. – River Li Mar 26 '24 at 11:17
  • I am not asking for an alternate way of proving majorization. I'm just asking whether the proof of majorization may be skipped in a contest, since it is somewhat obvious. – D S Mar 26 '24 at 11:22
  • @DS The reasoning is easy but I am not sure if we can say directly "(U,V,W) majorizes (X,Y,Z)". – River Li Mar 26 '24 at 11:23
  • Ok thanks. And I know this isn't the right place as it is not much related, but I wanted to ask how you approach inequalities where we can't directly apply well-known theorems; and any other manipulations reverse the inequality? Contests nowadays usually ask such questions (like APMO 2024, none of my friends got the inequality) – D S Mar 26 '24 at 11:30
  • @DS I know little about that. Usually I motivate my solution by computer, although often it is verifiable by hand. – River Li Mar 26 '24 at 12:06