Dear contributors,

I discover that OT is (still) able to computer the Wasserstein distance of two discrete measures when the number of weights is lower with the number of spikes.
To be more precise, I can call

With a.size != C.shape[0] and b.size != C.shape[1].
According to the doc, (see https://pot.readthedocs.io/en/latest/all.html#module-ot), it should not be the case.

To Reproduce
Create two discrete distributions with
- n=100 weights
- n+1 spikes

See the following minimal working example:

import numpy as np
import ot

n_points = 101
n_weight = 100

# Draw samples
np.random.seed(24)
samples_1 = np.random.normal(0., 1., n_points)
samples_2 = np.random.normal(0., 1., n_points)

	# b) weights (of spikes)
weights_1 = np.random.exponential(1., n_weight)
weights_1 /= np.sum(weights_1)
weights_2 = np.random.exponential(1., n_weight)
weights_2 /= np.sum(weights_2)

# Compute OT
matC = ot.dist(samples_1.reshape((n_points, 1)), samples_1.reshape((n_points, 1)))
matC /= matC.max()

was = ot.emd2(weights_1, weights_2, matC)
print(was)

Desktop:

• OS: MacOSX Mojave
• Python version: 3.7.5
• POT was installed with pip (and was up to date)

Additional context
Even though the seed is set at the beginning of the MWE, there are cases where POT returns "UserWarning: Problem infeasible. Check that a and b are in the simplex" (because of randomness in the implementation?)