ott.solvers.quadratic.gromov_wasserstein.solve(geom_xx, geom_yy, geom_xy=None, fused_penalty=1.0, scale_cost=False, a=None, b=None, loss='sqeucl', tau_a=1.0, tau_b=1.0, gw_unbalanced_correction=True, ranks=-1, tolerances=0.01, **kwargs)[source]#

The quadratic loss of a single OT matrix is assumed to have the form given in , eq. 4.

The two geometries below parameterize matrices $$C$$ and $$\bar{C}$$ in that equation. The function $$L$$ (of two real values) in that equation is assumed to match the form given in eq. 5., with our notations:

$L(x, y) = lin1(x) + lin2(y) - quad1(x) * quad2(y)$
Parameters:
Return type:

GWOutput

Returns:

Gromov-Wasserstein output.