Probability distributions based on SciPy

In this example, we seek to create a probability distribution based on the SciPy library.

from __future__ import annotations

from gemseo.api import configure_logger
from gemseo.uncertainty.api import create_distribution
from gemseo.uncertainty.api import get_available_distributions

configure_logger()

<RootLogger root (INFO)>

First of all, we can access the names of the available probability distributions from the API:

all_distributions = get_available_distributions()
print(all_distributions)

['OTComposedDistribution', 'OTDiracDistribution', 'OTDistribution', 'OTExponentialDistribution', 'OTNormalDistribution', 'OTTriangularDistribution', 'OTUniformDistribution', 'SPComposedDistribution', 'SPDistribution', 'SPExponentialDistribution', 'SPNormalDistribution', 'SPTriangularDistribution', 'SPUniformDistribution']

and filter the ones based on the SciPy library (their names start with the acronym ‘SP’):

sp_distributions = [dist for dist in all_distributions if dist.startswith("SP")]
print(sp_distributions)

['SPComposedDistribution', 'SPDistribution', 'SPExponentialDistribution', 'SPNormalDistribution', 'SPTriangularDistribution', 'SPUniformDistribution']

Create a distribution

Then, we can create a probability distribution for a two-dimensional random variable whose components are independent and distributed as the standard normal distribution (mean = 0 and standard deviation = 1):

distribution_0_1 = create_distribution("x", "SPNormalDistribution", 2)
print(distribution_0_1)

norm(mu=0.0, sigma=1.0)

or create another distribution with mean = 1 and standard deviation = 2 for the marginal distributions:

.. GENERATED FROM PYTHON SOURCE LINES 61-66

distribution_1_2 = create_distribution(
    "x", "SPNormalDistribution", 2, mu=1.0, sigma=2.0
)
print(distribution_1_2)

norm(mu=1.0, sigma=2.0)

We could also use the generic SPDistribution which allows access to all the SciPy distributions but this requires to know the signature of the methods of this library:

distribution_1_2 = create_distribution(
    "x",
    "SPDistribution",
    2,
    interfaced_distribution="norm",
    parameters={"loc": 1.0, "scale": 2.0},
)
print(distribution_1_2)

norm(loc=1.0, scale=2.0)

Plot the distribution

We can plot both cumulative and probability density functions for the first marginal:

distribution_0_1.plot()

<Figure size 640x320 with 2 Axes>

Note

We can provide a marginal index as first argument of the Distribution.plot() method but in the current version of GEMSEO, all components have the same distributions and so the plot will be the same.

Get mean

We can access the mean of the distribution:

print(distribution_0_1.mean)

[0. 0.]

Get standard deviation

We can access the standard deviation of the distribution:

print(distribution_0_1.standard_deviation)

[1. 1.]

Get numerical range

We can access the range, ie. the difference between the numerical minimum and maximum, of the distribution:

print(distribution_0_1.range)

[array([-7.03448383,  7.03448691]), array([-7.03448383,  7.03448691])]

Get mathematical support

We can access the range, ie. the difference between the minimum and maximum, of the distribution:

print(distribution_0_1.support)

[array([-inf,  inf]), array([-inf,  inf])]

Generate samples

We can generate 10 samples of the distribution:

print(distribution_0_1.compute_samples(10))

[[-0.44712856 -1.37311732]
 [ 1.2245077   0.31515939]
 [ 0.40349164  0.84616065]
 [ 0.59357852 -0.85951594]
 [-1.09491185  0.35054598]
 [ 0.16938243 -1.31228341]
 [ 0.74055645 -0.03869551]
 [-0.9537006  -1.61577235]
 [-0.26621851  1.12141771]
 [ 0.03261455  0.40890054]]

Compute CDF

We can compute the cumulative density function component per component (here the probability that the first component is lower than 0. and that the second one is lower than 1.):

.. GENERATED FROM PYTHON SOURCE LINES 132-134

print(distribution_0_1.compute_cdf([0.0, 1.0]))

[0.5        0.84134475]

Compute inverse CDF

We can compute the inverse cumulative density function component per component (here the quantile at 50% for the first component and the quantile at 97.5% for the second one):

print(distribution_0_1.compute_inverse_cdf([0.5, 0.975]))

[0.         1.95996398]