Scribd
Upload
20 views

Machine Learning and Pattern Recognition Week 2 Univariate Gaussian

The document discusses the Gaussian or normal distribution, which is widely used in probabilistic machine learning. It describes how to generate random values from the standard normal distribution with zero mean and unit variance. It also explains how to shift and scale the normal distribution to have any mean and variance by adding or multiplying the standard normal by constants. Code examples are provided to generate and plot random normal values.

Uploaded by

zeliawillscumberg
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
20 views

Machine Learning and Pattern Recognition Week 2 Univariate Gaussian

The document discusses the Gaussian or normal distribution, which is widely used in probabilistic machine learning. It describes how to generate random values from the standard normal distribution with zero mean and unit variance. It also explains how to shift and scale the normal distribution to have any mean and variance by adding or multiplying the standard normal by constants. Code examples are provided to generate and plot random normal values.

Uploaded by

zeliawillscumberg
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 3

Univariate Gaussians

The Gaussian distribution, also called the normal distribution, is widely used in probabilis-
tic machine learning. This week we’ll see Gaussians in the context of doing some basic
statistics of experimental results. Later in the course we’ll use Gaussians as building blocks
of probabilistic models, and to represent beliefs about unknown quantities in inference
algorithms.
We know that many of you will have seen all of this material before (some of you several
times). However, not everyone has, and large parts of the MLPR course depend on thoroughly
understanding Gaussians. This note is more detailed (slow) than many of the machine
learning text books, and provides some exercises. A later note on multivariate Gaussians
will also be important.

1 The standard normal distribution, the zero-mean unit-variance


Gaussian
I think of a probability distribution as an object, like a black box, that outputs numbers.
A number-producing box for the standard normal distribution has a label, N (0, 1), and a
button. If we press the button, the box displays a number. You can simulate this process
at a python prompt by typing np.random.randn() (after importing numpy as usual). If you
‘press the button’ multiple times, or run randn() multiple times, you’ll get different numbers.
To attach a label to an outcome from the generator we write in mathematics x ∼ N (0, 1), or
in code x = np.random.randn().
Each call to randn() gives a different number, and many calls reveal the distribution encoded
in the generator. A histogram of a million draws from the distribution, reveals a classic ‘bell-
curve’ shape. You should be able to sketch this bell-shaped curve, with points of inflection
at ±1. Try plotting a histogram of a million standard normal samples on a computer now. If
you know how to do it, opening a terminal and creating the plot can be done in about 10
seconds. If you don’t know how to do it, learning to sample test data and to create plots is
an important skill to develop. Example code is at the end of this document.
The observed or empirical mean of the samples will be about 0, and the empirical variance
will be about 1. Check the mean and variance of the samples you’ve just simulated! In
the limit of infinitely many samples, these values become exact, hence R the numbers in
“N (0, 1)”, the description of the distribution. Formally: µ = E [ x ] = x p( x ) dx = 0 and
var[ x ] = E[( x − µ)2 ] = x2 p( x ) dx = 1. If you don’t know these definitions, please work
R
through the review sheet on expectations in the background material section.
The probability density function (PDF) for the standard normal distribution describes the
bell-shaped curve:
1 1 2
p( x ) = N ( x; 0, 1) = √ e− 2 x . (1)
2π
The probability of landing in a narrow range of width δ between x − δ/2 and x + δ/2 is about
p( x )δ. This statement becomes accurate in the limit δ → 0. So with N = 106 samples, we
expect about p( x ) Nδ samples to land in a narrow histogram bin of width δ at position x.
Using this approximation, you should be able to plot a theoretical prediction through the
histogram you produced earlier.

2 Shifting and scaling: general univariate Gaussians


[The separate notes on expectations contain more material on shifting and scaling random variables.]
We can define a process to generate a quantity y that draws a value from a standard normal,
x ∼ N (0, 1), and then adds a constant y = x + µ. The mean of this distribution, is µ.

MLPR:w2b Iain Murray and Arno Onken, http://www.inf.ed.ac.uk/teaching/courses/mlpr/2020/ 1


If we drew many values from N (0, 1), and added some constant µ to each one, the histogram
of the values keeps the same shape, and simply shifts to be centred at µ. The PDF will shift
in the same way. This distribution has the same shape, just with a different location, and
so is still called a Gaussian, just with mean µ instead of mean 0. To get the PDF of the new
variable, we replace every occurrence of x with (y − µ):
1 1 2
p(y) = N (y; µ, 1) = √ e− 2 (y−µ) . (2)
2π
The choice of variable names, such as x or y, is arbitrary in this note. You will of course also
see x used for Gaussians with non-zero mean.
Similarly, we can define a random variable z that multiplies or ‘scales’ a standard normal
outcome by a constant σ, before adding a constant µ:

z = σx + µ. (3)

If we scale and shift many draws from a standard normal, the histogram of values will be
stretched horizontally, and then shifted. Scaling by σ multiplies the variance by σ2 (see the
notes on expectations), and leaves the mean at zero. Adding µ doesn’t change the width of
the distribution, or its variance, but adds µ to the mean.
The distribution of z maintains the same bell-curve shape, with the points of inflection now
at µ ± σ (note, not ±σ2 ). We still say the variable is Gaussian distributed, but with different
parameters: z ∼ N (µ, σ2 ). By convention, the second parameter of the normal distribution is
usually its variance σ2 , not its width or standard deviation σ. However, if you are reading a
paper, or using a new library routine, it is worth checking the parameterization being used,
just in case. Sometimes people choose to define a Gaussian by its precision, 1/σ2 , instead of
the variance.
For a general univariate Gaussian variable z ∼ N (µ, σ2 ), we can identify a standard normal,
by undoing the shift and scale above:
z−µ
x= . (4)
σ
We now work out the probability density for z, by transforming the density for this x.
[See the further reading if you can’t follow this section, or want more detail.]
Substituting the above expression into the PDF for the standard normal, suggests the shape
of the shifted and scaled distribution that we imagined above is described by the density
− 1 ( z − µ )2
p(z) = N (z; µ, σ2 ) ∝ e 2σ2 . (5)

However, we have to be careful transforming real-valued random variables. Here, stretching


out the bell-curve to be σ times wider would make the area underneath it σ times bigger.
However, PDFs must be normalized: the area under the curve must be one. To conserve
probability mass, if we stretch a region of outcomes, we must also decrease the PDF there
by the same factor. Hence, the PDF of a general (univariate) Gaussian is that for a standard
normal, scaled down by a factor of σ:
1 − 1 ( z − µ )2
p(z) = N (z; µ, σ2 ) = √ e 2σ2 , (6)
σ 2π
which can also be written by replacing the σ in the denominator with σ2 inside the square-
root.

3 Check your understanding


3.1 Notation
[The website version of this note has a question here.]

MLPR:w2b Iain Murray and Arno Onken, http://www.inf.ed.ac.uk/teaching/courses/mlpr/2020/ 2


3.2 Maths background
[The website version of this note has a question here.]

3.3 Code demonstration


[The website version of this note has a question here.]

4 Further reading
https://en.wikipedia.org/wiki/Normal_distribution
If you would like to work through the change of variables more rigorously, or in more detail,
we are using a specific form of the following result:
If an outcome x has probability density p X ( x ), and an invertible, differentiable function g is
used to create a new quantity z = g( x ), then the probability density of the new quantity is
p Z (z) = p X ( g−1 (z))| dx
dz |. In our case the derivative is simple: 1/σ.
This method for transforming densities for a change of variables is in Bishop Section 1.2.1,
with more detail and the multivariate case in Murphy Section 2.6.

5 Python code
To generate a million outcomes from a standard normal and plot a histogram:
import numpy as np
from matplotlib import pyplot as plt

N = int(1e6) # 1e6 is a float, numpy wants int arguments


xx = np.random.randn(N)
hist_stuff = plt.hist(xx, bins=100)
plt.show()
We increase the default number of bins from 10, which is not really enough to see the shape
properly. We also want narrow bins to make an approximation suggested in the notes work
well.
To check the mean and variance:
print('empirical_mean = %g' % np.mean(xx)) # or xx.mean()
print('empirical_var = %g' % np.var(xx)) # or xx.var()
You should understand how to plot a theoretical prediction of this histogram shape. Fill in
the parts marked TODO below. If you don’t know the answers, read the note again: it does
contain both the PDF and how many samples we (approximately) expect to land in a narrow
bin.
bin_centres = 0.5*(hist_stuff[1][1:] + hist_stuff[1][:-1])
# Fill in an expression to evaluate the PDF at the bin_centres.
# To square every element of an array, use **2
pdf = TODO
bin_width = bin_centres[1] - bin_centres[0]
predicted_bin_heights = TODO # pdf needs scaling correctly
# Finally, plot the theoretical prediction over the histogram:
plt.plot(bin_centres, predicted_bin_heights, '-r')
plt.show()

MLPR:w2b Iain Murray and Arno Onken, http://www.inf.ed.ac.uk/teaching/courses/mlpr/2020/ 3

You might also like