7

Fix $n>k>1$. Define $\mathcal {A}(i,j)$ be the set of all sets $A\subset \{1,\ldots,n\}$ such that:

  1. $A$ has $k-1$ elements,
  2. $i\not\in A$ and also
  3. $j\not\in A$.

Also, for $A$ in $\mathcal {A}(i,j)$ let $B_{i,j}(A)= \{1,\ldots,n\}\backslash \left(A\cup\{i,j\}\right)$.

Consider the following matrix: $$M=\begin{cases} m_{i,j}=0&\text{ if } i=j, \text{ and }\\ m_{i,j}=\sum_{A\in \mathcal {A}(i,j)}\left( \prod_{l\in A} y_l \prod_{l\in B_{i,j}(A)} x_l\right)&\text{ if } i\neq j\end{cases},$$

where $x_l,y_l>0$ for all $l$.

Conjecture: $\det(M)\neq 0$ or more precisely, $\mathrm{sign}(\det(M))=(-1)^{n-1}$.

For cases that I can compute with 11G memory and 8 nodes in a Linux cluster, the determinant above has sign $(-1)^{n-1}$.

In the case $\mathbf{n=4}$, it is easy to check that the even permutations whose corresponding addends are nonzero (i.e., that have no fixed points) have two cycles of length two. For example, (12)(34) corresponds to the term $+m_{1,2}^2\,m_{3,4}^2$ in the expansion of the det. The odd permutations always come in pairs. For example, both (1234) and (1432) correspond to the term $-m_{1,2}\,m_{2,3}\,m_{3,4}\,m_{1,4}$ so we have $-2m_{1,2}\,m_{2,3}\,m_{3,4}\,m_{1,4}$ in the expansion of the det. Now, by definition of $m_{i,j}$, it is easy to check that: $$ m_{1,2}^2\,m_{3,4}^2<m_{1,2}\,m_{2,3}\,m_{3,4}\,m_{1,4}+m_{1,2}\,m_{1,3}\,m_{3,4}\,m_{2,4}.$$ Thus, using symmetry, we can show that all positive terms are dominated by the negative terms. For higher $n$, a similar argument should work. I'm trying to 1) identify which are the even permutations (when $n$ is even) and 2) show that there are lots of odd permutations that allow us to kill the terms associated to the even permuations. For odd $n$, it is the opposite.

Mathematica code to count the signs of the terms in the expansion of the determinant (it does not work when $k-1=n$ but this case I proved by hand):

Dy[k_, n_] := 
 Module[{f, i, j, a, b, L, c}, 
  f[i_, j_] := 
   Boole[i != j]*
    Sum[Product[x[a], {a, s}]*
      Product[y[b], {b, 
        Complement[Delete[Range[n], {{i}, {j}}], s]}], {s, 
      Subsets[Delete[Range[n], {{i}, {j}}], {k - 1}]}]; 
  L = List @@ Expand[Det[Array[f, {n, n}]]];
  c = Count[L, _?Internal`SyntacticNegativeQ];
  {c, Length[L]}]
  • Thanks! Also, you do not need to bound the variables by $1$, since homogeneity would immediately render such a bound unnecessary, right? – darij grinberg Oct 19 '16 at 18:49
  • Am I seeing it right that this is supposed to be the same problem as (the general version of) http://math.stackexchange.com/questions/1942231/prove-that-determinant-of-a-matrix-with-polynomial-entries-is-non-zero/1969997#1969997 ? In particular, "do not contain the set ${i,j}$" should really be "do not contain any of $i$ and $j$" ? – darij grinberg Oct 19 '16 at 21:20
  • Yes, no need to bound. 2) Yes, this is a generalized version of the previous question. In that particular problem $y_k=1-x_k$. I think it would be easier to solve this more general version. 3) $A\in \mathcal A(i,j) \Leftrightarrow {i,j} \not\subset A$ and $|A|=k-1$ so I think the wording is correct.
  • – Sergio Parreiras Oct 20 '16 at 16:56
  • @darijgrinberg: A related problem,which is harder but has a concrete interpretation, is here http://math.stackexchange.com/questions/1972166/is-there-an-inversion-formula The matrix in this question (with $y_k=1-x_k$) is the Jacobian of the function posted in the harder problem. – Sergio Parreiras Oct 20 '16 at 17:07
  • I'm not sure I understand. $\left{i,j\right}\not\subset A$ is not equivalent to $A \subset \left{1,2,\ldots,n\right}\setminus\left{i,j\right}$, or am I comparing apples to oranges? – darij grinberg Oct 20 '16 at 17:09
  • ${i,j}\not\subset A$ means that 1. $A\subset {1,2,\ldots , n} \setminus\left{i,j\right}$ or 2. $A\subset {1,2,\ldots , n} \setminus\left{i\right}$ or 3. $A\subset {1,2,\ldots , n} \setminus\left{j\right}$. This needs a clarification. – Sungjin Kim Oct 20 '16 at 17:27
  • @darijgrinberg and i707107: sorry I see now the wording is misleading, fixing it, thanks. – Sergio Parreiras Oct 20 '16 at 18:30
  • 1
    This version is equivalent to the previous one. As explained before, if this is true, substituting $(x_k, y_k) = (t_k, 1-t_k)$ proves the other. Conversely, the determinant is homogenous of degree $n-2$ in each pair of variables $(x_k, y_k)$, so we can rescale each pair $(x_k, y_k)$ independently without changing the sign. Thus, if we know the sign whenever $(x_k, y_k)$ lies on the line segment ${ (t,1-t) : 0 \leq t \leq 1 }$, then we know the sign everywhere. PS Nice problem! – David E Speyer Oct 24 '16 at 15:11
  • @DavidSpeyer Thanks for showing they are equivalent! – Sergio Parreiras Oct 24 '16 at 16:26
  • 1
    This problem seems to scream "Lindström-Gessel-Viennot". However, a direct "planar" application of the principle, a la how Jacobi-Trudi is proven, is impossible, because the terms that survive cannot all come from one single permutation. So, if Lindström-Gessel-Viennot is to be used, then we somehow must argue that the nonintersecting path families all yield permutations of one and the same sign. I have never seen such an argument in the literature, although I believe this to be possible. (Also, nice work @DavidSpeyer !) – darij grinberg Oct 26 '16 at 00:25