TITLE: A delicate elementary inequality
QUESTION [13 upvotes]: The following "piecewise-quadratic" inequality emerged in a joint work of Rom
Pinchasi and myself. The inequality is surprisingly delicate, and all our
attempts to simplify it made it false. By the end of the day, we were able to
prove the inequality, but the proof is unreasonably sophisticated, totalling
to about 15 pages. We would be happy to have a shorter proof.
The inequality involves the function $G$ of three real variables, defined as
follows: if $(\xi,\eta,\zeta)$ is a non-decreasing rearrangement of
$(x,y,z)$, then we let
  $$ G(x,y,z) := \begin{cases}
                   \xi\eta &\ \text{if}\ \zeta\ge \xi+\eta, \\
                   \xi\eta-\frac14\,(\xi+\eta-\zeta)^2 &\ \text{if}\ \zeta\le \xi+\eta.
                 \end{cases} $$
Thus, for instance, we have $G(9,6,7)=38$, whereas $G(7,14,6)=42$. Now
consider the function $f$ of four variables defined by
$$\begin{align*}
  f(x_0,x_1,y_0,y_1)
    &:= \min \{ 0.15s^2, x_0y_0+x_1y_1 \}  \\
    &\qquad + G(x_0,y_1,1-s) + G(x_1,y_0,1-s)  \\
    &\qquad + 0.25(1-s)^2,
\end{align*}$$
where for brevity I write $s=x_0+x_1+y_0+y_1$, and let
  $$ \Omega := \{ (x_0,x_1,y_0,y_1)\in{\mathbb R}_{\ge 0}^4\colon 1/2 \le s \le 1. \}. $$
All we want to show is that
  $$ \max_\Omega f \le 0.15. $$
(Indeed, the maximum is actually equal to $0.15$: say, we have
$f(0.5,0,0.5,0)=0.15$.)

At Pat Devlin's suggestion, here is the graph of the maximum as a function of the sum $s=x_0+x_1+y_0+y_1$. 

(source: haifa.ac.il) 
It looks nice, but does not seem to be a graph of some "simple" function; and so, there is probably no simple analytic expression for $\max f$ over all quadruples $(x_0,x_1,y_0,y_1)$ adding up to $s$.

REPLY [7 votes]: The paper where the inequality in question emerged is, finally, written (and
uploaded to the arXiv, in case anybody is interested). We were
eventually able to simplify the proof and squeeze it down to just about
six pages; indeed, the whole paper is now shorter than our original proof. I
sketch very briefly the new proof below.
We start with the identity
\begin{multline*}
  G((x+y)/2,(x+y)/2,z) = G(x,y,z) \\
     + \frac14(x-y)^2
                - \frac14 \big(\max\{|x-y|-z, 0 \} \big)^2. \tag{1}
\end{multline*}
Once stated, this is easy to verify by a careful case analysis. An immediate
consequence is that $G(x,y,z)$ can only grow if both $x$ and $y$ are replaced
with their average $(x+y)/2$.
As another preparation step, we exploit the internal symmetries of $f$ to
assume, without loss of generality, that
  $$ x_0+x_1 \ge y_0+y_1 \tag{2} $$
and also
  $$ x_0+y_0 \ge x_1+y_1. \tag{3} $$
Our big plan is to investigate how $f$ changes under the balancing operation
which includes replacing $x_0$ and $y_1$ with their average $(x_0+y_1)/2$
and, simultaneously, $x_1$ and $y_0$ with their average $(x_1+y_0)/2$. Using
the identity (1), we could show that either $f$ is non-decreasing under such
balancing, or, under the assumptions (2) and (3), we have
\begin{align*}
  x_0 &\ge y_1+(1-s), \tag{4} \newline
  y_0 &\ge x_1+(1-s), \tag{5}
\end{align*}
and
  $$ 3(x_0+y_0)+(x_1+y_1) \ge 2. \tag{6} $$
The precise meaning of being non-decreasing under balancing is that
  $$ f(x_0,x_1,y_0,y_1) \le f(z_0,z_1,z_1,z_0), $$
where $z_0=(x_0+y_1)/2$ and $z_1=(x_1+y_0)/2$. Consequently, in this case the
problem reduces to maximizing a function of just two variables, which is a
feasible task.
Now, if $f$ is decreasing under balancing, then, in view of (4) and (5) and
by the definition of the function $G$, we have
  $$ G(x_0,y_1,1-s) = y_1(1-s) \ \text{and}\ G(x_1,y_0,1-s)=x_1(1-s). $$
Hence,
\begin{multline*}
  f(x_0,x_1,y_0,y_1) = \min\{0.15s^2,x_0y_0+x_1y_1\} \newline
                                             + (x_1+y_1)(1-s) + 0.25(1-s)^2.
\end{multline*}
The expression in the right-hand side is can only increase if $x_0$ and $y_0$
are both replaced with their average, and, simultaneously, $x_1$ and $y_1$
are replaced with their average. Consequently, we can assume that $x_0=y_0$
and $x_1=y_1$. This, again, reduces the problem to maximizing a function of
two variables, which takes some two more pages to accomplish.