ott.solvers.linear.discrete_barycenter.FixedBarycenter

class ott.solvers.linear.discrete_barycenter.FixedBarycenter(threshold=0.01, norm_error=1, inner_iterations=10, min_iterations=0, max_iterations=2000, lse_mode=True, debiased=False)

A Wasserstein barycenter solver for histograms on a common geometry.

This solver uses a variant of the Sinkhorn algorithm proposed in [Janati et al., 2020] to compute the barycenter of various measures supported on the same (common to all) geometry. The geometry is assumed to be either symmetric, or to describe costs between a set of points and another. In that case all reference measures have support on the first measure, whereas the barycenter is supported on the second.

Parameters:

• threshold (float) – convergence threshold. The algorithm stops when the marginal violations of all transport plans computed for that barycenter go below that threshold.

• norm_error (int) – norm used to compute marginal deviation.

• inner_iterations (float) – number of iterations run before recomputing errors.

• min_iterations (int) – number of iterations run without checking whether termination criterion is true.

• max_iterations (int) – maximal number of iterations.

• lse_mode (bool) – sets computations in kernel (False) or log-sum-exp mode.

• debiased (bool) – uses debiasing correction to avoid blur due to entropic regularization.

Methods