Sinkhorn distance [34] is a metric based on optimal transport, which approximates the Wasserstein distance through the introduction of entropy regularization, leading to increased computational efficiency.
The Lipschitz constant L typically depends on the choice of regularization parameters and cost functions.
When the cost function for Sinkhorn distance is Euclidean distance, its Lipschitz constant depends on parameters of the Sinkhorn algorithm used and characteristics of the dataset.
Citations
"Discriminative models for image reasoning offer diverse solutions with multi-dimensional outputs."
"Using distributions to describe human concepts in abstract reasoning problems offers a more comprehensive approach."
"Spectral normalization is a valuable tool for stabilizing deep neural networks."