TITLE: Existence of solution to these inequalities
QUESTION [10 upvotes]: They say that all mathematics problems eventually reduce to linear algebra or combinatorics. I have reduced mine to proving a solution exists for the following set of inequalities but have no idea how to proceed.
Question: For $n \geq 2$, fix once and for all, a permutation $\tau \in S_n, \tau \neq (1,n)(2,n-1)\cdots$, i.e., $\tau$ isn't the long element. Let
$$ \Delta(\tau) = \{ i \in \{1, 2, \cdots, n-1 \} : \tau(\{ 1, 2, \cdots, i \}) \neq \{ n, n-1, \cdots, n-i+1 \} \}$$
Do there exist real numbers $b_1 > b_2 > \cdots > b_{n-1} > b_n = 0$ such that the inequalities
$$ \frac{b_1 + b_2 + \cdots + b_i + b_{\tau(1)} + \cdots + b_{\tau(i)}}{2i} > \frac{b_1 + b_2 + \cdots + b_n}{n} \quad \dots \text{Eq. }(i)$$
hold simultaneously for every $i \in \Delta(\tau)$?
Motiation: My original question is posted here and the above question is the case when $G = SL(n)$, after some simplifications and explicit computation with the Cartan matrix.
Remark: I have explicitly verified this holds for $n \leq 8$. These calculations suggest that the $b_i$'s cannot be independent of $\tau$.


MAJOR EDIT: I made a mistake while posing the question. The $\tau$ in Equation $(i)$ should actually be $\tau^{-1}$. The mistake came
because $\sum b_i \varpi_{\tau(i)}$ was wrongly written as
$\sum b_{\tau(i)} \varpi_i$ instead of $\sum b_{\tau^{-1}(i)} \varpi_i$. I will reward the bounty to the
answer by @Pietro Majer but the question I really want answered is:

Do there exist real numbers $b_1 > b_2 > \cdots > b_{n-1} > b_n = 0$ such that the inequalities
$$ \frac{b_1 + b_2 + \cdots + b_i + b_{\tau^{-1}(1)} + \cdots + b_{\tau^{-1}(i)}}{2i} > \frac{b_1 + b_2 + \cdots + b_n}{n} \quad \dots \text{New Eq. }(i)$$hold simultaneously for every $i \in \Delta(\tau)$?

I can award some more bounty points for proving the new inequality.

REPLY [8 votes]: For a given $\tau\in S_n$ and for any $j\in[n]$ define the numbers
$$a_j:=\chi_{\Delta}(j)-\chi_{\Delta}(j-1)-\chi_{\Delta }(n-j)+\chi_{\Delta}(n-j+1),$$
where $\chi_\Delta:\mathbb{Z}\to\{0,1\}$  denotes the characteristic function of  the set $\Delta:=\Delta(\tau)\subset\mathbb{Z}$. Also, with $c:=5n-a_n$, define
$$b_j:=a_j-5j+c.$$
We may note right away that since $|a_j|\le2$, the $b_j$ are strictly decreasing, and that $b_n=0$, by the choice of the constant $c$.
For any $E\subset[n]$, for simplicity of notation we put
 $$ \alpha(E):=\sum_{j\in E} a_j,\qquad \beta(E):=\sum_{j\in E} b_j$$
(so we may think $ \alpha$ and $\beta$ as discrete signed measures supported in $[n]$).
For  $i\in[n]$, summing over $j=1,\dots i$ we have
$$ \alpha([i])=\chi_\Delta(i)-\chi_\Delta(n-i).$$
Incidentally, for any $i\in[n]$ we have $i\in\Delta$ if and only if, by definition, $\tau([i])\neq[i]+n-i$ thus also, since $\tau$ is bijective,  if and only if $\tau([i]^c)\neq([i]+n-i)^c$, that is  $\tau([n-i]+i)\neq [n-i]$ or $\tau^{-1}([n-i])\neq [n-i]+i$, which means $n-i \in \Delta^{-1}:=\Delta(\tau^{-1})$. Hence the last formula also writes
$$ \alpha([i])=\chi_{\Delta}(i)-\chi_{\Delta^{-1}}(i).$$
Also note that, since $n\not\in\Delta$
$$ \alpha([n])=0,$$
and $$ \alpha([i]+n-i)=- \alpha([n-i]) =-\chi_\Delta(n-i)+\chi_\Delta(i)= \alpha([i]).$$
We proceed showing the inequalities on the arithmetic means. 
Case I. Assume $i\in\Delta\setminus\Delta^{-1}$. Then by definition of $\Delta^{-1}$,  $\tau^{-1}([i])=[i]+n-i$, so that
$$ {\alpha([i])+ \alpha(\tau^{-1}[i])\over 2i}= {\alpha([i])+ \alpha([i]+n-i)\over2i}={\chi_{\Delta}(i)-\chi_{\Delta^{-1}}(i)\over i}={1\over i}>0, $$
and summing the arithmetic means of $-5j+c$ on the same sets we have plainly
$${\beta([i])+ \beta(\tau^{-1}[i])\over 2i}>{\beta([n])\over n}.$$
Case II. Assume $i\in\Delta\cap\Delta^{-1}$. Thus $\tau^{-1}([i])\neq[i]+n-i$ and, just because $b_j$ are strictly decreasing
$${\beta([i])+ \beta(\tau^{-1}[i])\over 2i}>{\beta([i])+ \beta([i]+n-i)\over2i}$$
and since we have $\alpha([i])=\alpha([i]+n-i)=\alpha([n])=0$ because $\chi_{\Delta}(i)=\chi_{\Delta^{-1}}(i)=1$,  summing as before the arithmetic means of the affine part of $b_j$, $${\beta([i])+ \beta([i]+n-i)\over2i}={\beta([n])\over n},$$
concluding the proof.