Multistochastic optimal transport optimizes the integral of a cost over probability measures on a product space whose marginals on every specified collection
of
coordinates are fixed. For
coordinates and
,
it is called the
problem; ordinary optimal
transport corresponds to
(Gladkov et al. 2019). In the
case with all two-coordinate marginals equal to uniform
distributions on
,
the three coordinates are pairwise independent
and uniform.
Write a binary expansion ,
, and let
|
(1)
|
For independent random variables
with uniform distribution, Gvozdev and Kolesnikov
(2026) proved that among all pairwise independent
uniform triples, the choice
minimizes
, while
maximizes it. The symbol
means exclusive OR of corresponding binary digits;
ambiguities at numbers with two binary expansions
have probability zero. The minimizing plan is supported
on the tetrix with vertices
,
,
, and
.
The maximizing plan lies on its reflection with vertices
,
,
, and
.
The case
, whose coefficients satisfy equality instead of
the strict domination above, was established separately by Gladkov et al. (2019).
For the geometric coefficients ,
, Gvozdev and Kolesnikov (2026) also constructed
a dual certificate. Put
,
, and
. On sequences of binary
digits, let
delete the first digit if it is 1, and delete the first digit
and complement all remaining digits if it is 0. The bounded function
|
(2)
|
satisfies
|
(3)
|
with equality for .
Integrating the left side depends only on the fixed two-coordinate marginals, which
certifies minimality for the centered cost. The authors report that GPT-6 Astra contributed
to developing these dual solutions, whose arguments they checked.