?
1
18.338J/16.394J: The Mathematics of In?nite Random Matrices
Histogramming
Professor Alan Edelman
Handout #2, Tuesday, September 14, 2004
Random Variables and Probability Densities
We assume that the reader is familiar with the most basic of facts concerning continuous random variables
or is willing to settle for the following sketchy description. Samples from a (univariate or multivariate)
experiment can be histogrammed either in practice or as a thought experiment. Histogramming counts
how many samples fall in a certain interval. Underlying is the notion that there is a probability
distribution which precisely represents the probability of falling into an interval.
If x ∈ R is a real random variable with probability density p
x
(t), this means that the probability that x
?
b
may be found in an interval [a,b] is
a
p
x
(t)dt. More generally if S is some subset of R, the probability that
x ∈ S is
S
p(t)dt. Later on, we may be more careful and talk about sets that are Lebesgue measurable,
but this will do for now.
The probability density is roughly the picture you would obtain if you collected many random values of x
and then histogrammed these values. The only problem is that if you have N samples and your bins have
size Δ, then the total area under the boxes is NΔ not 1:
Number
found
between
x?Δ/2
and
x + Δ/2
?
Δ
?
Therefore a normalization must occur for the total area under the boxes to equal to 1 so that the
(normalized) histogram and probability densities can line up.
Normal distribution
The normal distribution with mean 0 and variance 1 (standard normal, Gaussian, “bell shaped curve”) is
the random variable with probability density p
x
(t) =
√
1
2π
e
?t
2
/2
. It deserves its special place in probability
theory because of the central limit theorem which states that if x
1
,...,x
n
,... are iid random variables (iid
= independent and identically distributed) with mean μ and variance σ then
2
?
x
1
+ ···+ x
n
?n·μ
< b
?
= √
1
2π
?
a
b
e
?t /2
dt. lim Prob a <
n→∞
√
nσ
The central limit theorem roughly states that a large collection of identical random variables behaves like
the normal distribution. Many investigations into the eigenvalues of random matrices suggest
experimentally that this statement holds, i.e., the eigenvalues of matrices whose elements are not normal
behave, more or less, like the eigenvalues of normally distributed matrices.
It is of value to note that the normal distribution with mean μ and variance σ
2
has
1
p
x
(t) =
σ
√
2π
e
?(x?μ)
2
/2σ
2
.
2 Univariate Histograms
In Figure 2, we plot the normal distribution as well as a histogram obtained from 5000 samples from the
normal distribution We see in the second line of the code below that we divide the counts n by the total
number times the bin size: 5000*0.2. This guarantees that the total area of the boxes over the whole line
is normalized to 1.
-4 -3 -2 -1 0 1 2 3 4
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Figure 1: This ?gure illustrates the idea that the probability density is a histogram
Code 1 is our MATLAB code to obtain this ?gure.
>> a=randn(1,5000);[n,x]=hist(a,[-3:.2:3]);
>> bar(x,n/(5000*.2));
>> hold on,plot(x,exp(-x.^2/2)/sqrt(2*pi)),hold off
? ?
? ?
%bellcurve.m
%Code 1.1 of Random Eigenvalues by Alan Edelman
%Experiment: Generate random samples from the normal distribution.
%Observation: Histogram the random samples.
%Theory: Falls on a bell curve.
trials=100000; dx=.2;
v=randn(1,trials);[count,x]=hist(v,[-4:dx:4]);
hold off, b=bar(x,count/(trials*dx),’y’); hold on
x=-4:.01:4;
plot(x,exp(-x.^2/2)/sqrt(2*pi),’LineWidth’,2)
axis([-4 4 0 .45]);
Code 1
3 How Accurate Are Histograms?
When playing with Code 1, the reader will happily see that given enough trials the histogram is close to
the true bell curve. One can press further and ask how close? Multiple experiments will show that some of
the bars may be slightly too high while others slightly too low. There are many experiments which we
explore in the exercises to try to understand this more clearly. We will discuss these as the course progresse
4 HISTN: Normalized Histogram
We can incorporate the ideas discussed above into the following MATLAB code.
? ?
? ?
function [h,hn,xspan]=histn(data,x0,binsize,xf);
%HISTN Normalized Histogram.
% [H,HN,XSPAN] = HISTN(DATA,X0,BINSIZE,XF) generates the normalized
% histogram of area 1 from the values in DATA which are binned into
% equally spaced containers that span the region from X0 to XF
% with a bin width specified by BINSIZE.
%
% X0, BINSIZE and XF are all scalars while DATA is a vector.
% H, HN and XSPAN are equally sized vectors.
%
% References:
% [1] Alan Edelman, Handout 2: Histogramming,
% Fall 2004, Course Notes 18.338.
% [2] Alan Edelman, Random Matrix Eigenvalues.
%
% Alan Edelman and Raj Rao, Sept. 2004.
% $Revision: 1.1 $ $Date: 2004/09/10 17:11:18 $
xspan=[x0:binsize:xf];
h=hist(data,xspan); % Generate histogram
hn=h/(length(data)*binsize); % Normalize histogram to have area 1
bar(xspan,hn); % Plot histogram
Code 2
We will use this code throughout the remainder of the course to corroborate theoretical predictions with
experimental data.
