This notebook is an element of the risk-engineering.org courseware. It can be distributed under the terms of the Creative Commons Attribution-ShareAlike licence.
Author: Eric Marsden eric.marsden@risk-engineering.org
This notebook contains an introduction to the use of SciPy library's support for various probability distributions. The library documentation is available online. We also show how the SymPy library for symbolic mathematics can be used to calculate various statistical properties analytically. Visit the associated course materials for background material and to download this content as a Python notebook.
import numpy
import scipy.stats
import matplotlib.pyplot as plt
plt.style.use("bmh")
%config InlineBackend.figure_formats=["svg"]
Let’s generate 5 random variates from a continuous uniform distribution between 90 and 100 (the first argument to the scipy.stats.uniform
function is the lower bound, and the second argument is the width). The object u
will contain the “frozen distribution”.
u = scipy.stats.uniform(90, 10)
u.rvs(5)
array([91.85688457, 90.87813946, 95.49755923, 94.21517598, 95.77316387])
Let’s check that the expected value of the distribution is around 95.
u.rvs(1000).mean()
94.96971154289815
Let’s check that around 20% of the variates are less than 92.
(u.rvs(1000) < 92).sum() / 1000.0
0.227
We can also use the stats
module of the SymPy library to obtain the same information using an analytical (rather than stochastic) method.
import sympy.stats
u = sympy.Symbol("u")
u = sympy.stats.Uniform(u, 90, 100)
# generate one random variate
sympy.stats.sample(u)
Check that the expected value (the mean of the distribution) is 95.
sympy.stats.E(u)
The probability of a random variate being less than 92:
sympy.stats.P(u < 92)
Consider a Gaussian (normal) distribution centered in 5, with a standard deviation of 1.
norm = scipy.stats.norm(5, 1)
x = numpy.linspace(1, 9, 100)
plt.plot(x, norm.pdf(x));
norm.mean()
5.0
Check that half the distribution is located to the left of 5.
norm.cdf(5)
0.5
norm.ppf(0.5)
5.0
Find the first percentile of the distribution (the value of $x$ which has 1% of realizations to the left). Check that it is also equal to the 99% survival quantile.
norm.ppf(0.01)
2.6736521259591592
norm.isf(0.99)
2.6736521259591592
norm.cdf(norm.isf(0.99))
0.01
We can also execute some calculations on the normal distribution analytically, using the SymPy library. This provides computer algebra capabilities similar to (though currently less powerful than) the commercial tools Mathematica and Maple.
z = sympy.Symbol("z")
Norm = sympy.stats.Normal(z, 5, 1)
sympy.stats.density(Norm)(z)
# we ask for the evaluation to 30 significant figures
sympy.stats.density(Norm)(4).evalf(40)
norm.pdf(4)
0.24197072451914337
# check that half of our distribution is greater than 5 (this is equivalent to the CDF(5))
sympy.stats.P(Norm > 5)
# check that the expected value (the mean) is 5
sympy.stats.E(Norm)
# check that the standard deviation is 1
sympy.stats.std(Norm)
# generate one random variate from the distribution
sympy.stats.sample(Norm)
# calculate the first percentile (the 0.01 quantile) -- this requires Sympy version 1.5
sympy.stats.quantile(Norm)(0.01).evalf(30)
# compare with the value calculated by numpy
norm.ppf(0.01)
2.6736521259591592
The central limit theorem states that the mean of a set of random measurements will tend to a normal distribution, no matter the shape of the original measurement distribution. The property is also true of the sum of a set of random measurements. Let's test that in Python, simulating measurements from a uniform distribution between 30 and 40.
Procedure: take 100 measurements from the $U(30, 40)$ distribution, and calculate their mean. Repeat this 10000 times and plot a histogram of the means, which should be normally distributed.
N = 10000
sim = numpy.zeros(N)
for i in range(N):
sim[i] = numpy.random.uniform(30, 40, 100).mean()
plt.hist(sim, bins=20, alpha=0.5, density=True);
Exercise: try this with other probability distributions.
The exponential distribution is often used in reliability engineering to represent failure of equipment which is not exposed to wear. The hazard function, or failure rate, of the exponential distribution is constant, equal to $\lambda$. Let's check the property that the expected value (or mean) of an exponential random variable is $\frac{1}{\lambda}$.
lbda = 25
obs = scipy.stats.expon(scale=1/float(lbda)).rvs(size=1000)
# the mean of 1000 random variates from the exponential distribution
obs.mean()
0.041653180277278996
1/float(lbda)
0.04
Indeed, those are quite close! Let’s check another property of the exponential distribution: that the variance is equal to $\lambda^{-2}$.
obs.var()
0.0018636918057768205
1/float(lbda)**2
0.0016
And of course since the standard deviation is the square root of the variance, it should be equal to the expected value.
obs.std()
0.043170496936876
As previously, we can also check these properties analytically, using the SymPy symbolic mathematics library, starting with the expected value:
exp = sympy.Symbol("exp")
exp = sympy.stats.Exponential(exp, lbda)
sympy.stats.E(exp)
# now check that the variance is lambda ^ -2
sympy.stats.variance(exp) - lbda**-2
# check that the standard deviation is equal to the expectation
sympy.stats.std(exp) - sympy.stats.E(exp)