I am using scipy's multivariate earth mover distance function wasserstein_distance_nd. I did a quick sanity check that confused me: Given that I draw two Gaussian multivariate samples from the same distribution, I should get an earth mover distance that is close to 0. However, I am getting something large (e.g., 12). Why is this happening? I have tested this with the one dimensional case and I also got something similar (here, the distance produced is always positive).

Code that I used is given as follows:

'''Multi-dimensional'''
import numpy as np
from scipy.stats import wasserstein_distance_nd

mean = np.zeros(100)
cov = np.eye(100)
size = 100
sample1 = np.random.multivariate_normal(mean, cov, size)
sample2 = np.random.multivariate_normal(mean, cov, size)

print("EMD", wasserstein_distance_nd(sample1, sample2))
# output: EMD 12.293968193381374

'''single dimension'''
import numpy as np
from scipy.stats import wasserstein_distance

mean = 0
var = 1
size = 100
sample1 = np.random.normal(mean, np.sqrt(var), size)
sample2 = np.random.normal(mean, np.sqrt(var), size)
dist = wasserstein_distance(sample1, sample2)

print("wasserstein_distance", dist)

I am using scipy's multivariate earth mover distance function wasserstein_distance_nd. I did a quick sanity check that confused me: Given that I draw two Gaussian multivariate samples from the same distribution, I should get an earth mover distance that is close to 0. However, I am getting something large (e.g., 12). Why is this happening? I have tested this with the one dimensional case and I also got something similar (here, the distance produced is always positive).

Code that I used is given as follows:

'''Multi-dimensional'''
import numpy as np
from scipy.stats import wasserstein_distance_nd

mean = np.zeros(100)
cov = np.eye(100)
size = 100
sample1 = np.random.multivariate_normal(mean, cov, size)
sample2 = np.random.multivariate_normal(mean, cov, size)

print("EMD", wasserstein_distance_nd(sample1, sample2))
# output: EMD 12.293968193381374

'''single dimension'''
import numpy as np
from scipy.stats import wasserstein_distance

mean = 0
var = 1
size = 100
sample1 = np.random.normal(mean, np.sqrt(var), size)
sample2 = np.random.normal(mean, np.sqrt(var), size)
dist = wasserstein_distance(sample1, sample2)

print("wasserstein_distance", dist)

• python
• math
• scipy
• statistics
• simulation

1 Answer 1

Your multivariate_normal example is in a 100-dimensional space. You can think of the reason you are getting large distances from an intuitive perspective: there is too much variety possible in a 100-dimensional space for the two samples to be very similar.

Some more motivation:

• In high dimensional spaces, randomly drawn vectors tend to be nearly perpendicular (see e.g. https://math.stackexchange/questions/995623/why-are-randomly-drawn-vectors-nearly-perpendicular-in-high-dimensions). Compare this to a one-dimensional space, in which all random vectors lie on the same line.

• Think also about fixing the error in each coordinate to a given level and increasing the number of dimensions: the distance increases according to the square root of the number of dimensions. For instance, in one dimension, if $p_1 = 0$ and $p_2 = 1$, the distance between them is only 1. But in 100 dimensions, $p_1 = (0, ..., 0)$ and $p_2 = (1, ..., 1)$ have a distance of 10.

For a rigorous explanation, seek help on Math Stack Exchange. My point, though, is just that intuition about magnitudes of distances from a 1 dimensional case just won't work well in 100 dimensions.

If it helps, you can confirm that wasserstein_distance_nd agrees with wasserstein_distance in 1D.

import numpy as np
from scipy.stats import wasserstein_distance

mean = 0
var = 1
size = 100
sample1 = np.random.normal(mean, np.sqrt(var), size)
sample2 = np.random.normal(mean, np.sqrt(var), size)
ref = wasserstein_distance(sample1, sample2)
res = wasserstein_distance_nd(sample1[:, np.newaxis], sample2[:, np.newaxis])
print("wasserstein_distance", res, ref)
# wasserstein_distance 0.2109383092226257 0.21093830922262574