The Gromov-Wasserstein distance compares metric spaces equipped with probability measures by transporting
their internal pairwise distances (Mémoli 2011). For compact metric spaces and
with Borel probability
measures
and
,
the quadratic convention is
|
(1)
|
where
consists of probability measures on
with marginals
and
.
The factor
is a normalization convention. Unlike the Wasserstein
distance, this definition requires no shared ambient metric
space.
For finite spaces with masses and
,
a coupling is a nonnegative matrix
with row sums
and column sums
.
Writing
and
gives the squared-loss objective
|
(2)
|
This quadratic form can also be studied for symmetric dissimilarity matrices that are not distance matrices of metric spaces. For uniform masses on equally sized sets, the extreme couplings are scaled permutation matrices, but a minimum of this quadratic objective need not occur at an extreme coupling (Vayer 2026).
For example, Vayer (2026, Example 4.3) gives
|
(3)
| |||
|
(4)
|
Let
have entry
in position
and zero elsewhere. Among these 24 couplings the
minimum of
is
, whereas
|
(5)
|
Both matrices are nonnegative and have zero diagonals, but they fail the triangle inequality; this is not a counterexample for genuine metric spaces. Vayer (2026) credits Claude with finding this family of examples.