A triply line stochastic matrix (t.l.s.m.) of size $N$ is a 3-dimensional array $(a_{ijk})_{i,j,k=1}^N$ with nonnegative entries, whose any row, column or line sums up to $1$. Their set is $TLS_N$.

Consider the following $0$-$1$ t.l.s.m. of size $2^n$ :

Keep the half-cube (of size $2^{n-1}$) containing $(1,1,1)$, and the 3 neighboring half-cubes "in diagonal", and put $0$ everywhere in the other 4 half-cubes. Repeat this operation in each of the 4 kept cubes of size $2^{n-1}$, keeping 4 cubes of size $2^{n-2}$ (the one nearest to $(1,1,1)$ and its 3 neighbors in diagonal), and putting $0$ everywhere else. An so on, until you are left with $2^{2n}$ points, one in each row, column or line, and put $a_{ijk}=1$ there.

By construction, this t.l.s.m. maximizes the total (sum of the entries) in some cubes containing $(1,1,1)$ i.e. $$\sum_{i,j,k=1}^{m}a_{i,j,k}\ge\sum_{i,j,k=1}^{m}b_{i,j,k}$$ for every t.l.s.m. $B:=(b_{ijk})_{i,j,k=1}^{2^n}\in TLS_{2^n}$ , whenever $m=2^p$, $0\le p\le n$ .

Question (motivated by Bound on the joint distribution of three real random variables with given two dimensional marginals) Does this inequality also hold for other values of $m$, $1\le m\le 2^n$ ?

  • 1
    $\begingroup$ I'm not happy about using the word "matrix" to describe a three-dimensional array. $\endgroup$ – Gerry Myerson Jan 29 '17 at 22:08
  • $\begingroup$ I understand. Neither am I. But the notion was given that name (P Fischer, ER Swart - Linear algebra and its applications, 1985 - Elsevier). $\endgroup$ – Jean Duchon Jan 31 '17 at 13:25

Maybe for some other values, but not for all. Let $n=3$ (matrices of size $8$) and $m=3$ (comparing the totals of $A$ and of some other t.l.s.m. $B$ on the cube $[1,3]^3$). The total for $A$ on this cube is $7$. We can have as much as $9$ for $B$ if $b_{ijk}=\frac13$ on $[1,3]^3$, $=0$ on $[1,3]\times[1,3]\times[4,8]$, $[1,3]\times[4,8]\times[1,3]$ and $[4,8]\times[1,3]\times[1,3]$ , $=\frac15$ on $[1,3]\times[4,8]\times[4,8]$, $[4,8]\times[1,3]\times[4,8]$ and $[4,8]\times[4,8]\times[1,3]$, and $=\frac2{25}$ on $[4,8]^3$.

This negative result suggests the following modified question : for which coefficients $c_1>\cdots>c_m>\cdots>c_{2^n}>0$ is it true that $$\sum_{m=1}^{2^n}c_m\sum_{i,j,k=1}^m a_{ijk}\ge\sum_{m=1}^{2^n}c_m\sum_{i,j,k=1}^m b_{ijk}$$ for every t.l.s.m. $B$ ? (Is it true when $c_m$ is a convex function of $m$, for example ?)


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.