i

MAS3314/8314:
Multivariate Data Analysis
Prof D J Wilkinson

Module description:
In the 21st Century, statisticians and data analysts typically work with data sets containing
a large number of observations and many variables. This course will consider methods
for making sense of data of this kind, with an emphasis on practical techniques. Consider,
for example, a medical database containing records on a large number of people. Each
person has a height, a weight, an age, a blood type, a blood pressure, and a host of
other attributes, some quantitative, and others categorical. We will look at graphical and
descriptive techniques which help us to visualise a multi-dimensional data set and at
inferential methods which help us to answer more specific questions about the population
from which the individuals were sampled. We will also consider new problems which
arise in the analysis of large and/or multi-dimensional data, such as variable selection
and multiple testing.

Course texts:
T. Hastie, R. Tibshirani, J. Friedman: The Elements of Statistical Learning: Data mining,
inference, and prediction, 2nd Edition (Springer-Verlag, 2009).

B. Everitt: An R and S-Plus Companion to Multivariate Analysis (Springer-Verlag, 2005).

I will refer frequently to these texts in the notes, especially the former, which I will cite
as [HTF]. I will refer to the latter as [Everitt], mainly for R-related information. Note that
the PDF of the full text of [HTF] is available freely on-line, and that [Everitt] should be
available electronically to Newcastle University students via the University Library reading
list web site.

WWW page:
http://www.staff.ncl.ac.uk/d.j.wilkinson/teaching/mas3314/

Last update:
May 29, 2012
These notes correspond roughly to the course as delivered in the Spring of 2012. They
will be revised before I deliver the course again in the first half of 2013.
Use the date above to check when this file was generated.

c 2012, Darren J Wilkinson
Contents

1 Introduction to multivariate data 1

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 A few quotes... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Data in the internet age . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1.3 Module outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Multivariate data and basic visualisation . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Working with data frames . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Representing and summarising multivariate data . . . . . . . . . . . . . . . 15
1.3.1 The sample mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.3.2 Sample variance and covariance . . . . . . . . . . . . . . . . . . . . 18
1.3.3 Sample correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.4 Multivariate random quantities . . . . . . . . . . . . . . . . . . . . . . . . . 30
1.5 Transformation and manipulation of multivariate random quantities and data 31
1.5.1 Linear and affine transformations . . . . . . . . . . . . . . . . . . . . 32
1.5.2 Transforming multivariate random quantities . . . . . . . . . . . . . . 33
1.5.3 Transforming multivariate data . . . . . . . . . . . . . . . . . . . . . 37

2 PCA and matrix factorisations 44

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.1.1 Factorisation, inversion and linear systems . . . . . . . . . . . . . . 44
2.2 Triangular matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2.1 Upper and lower triangular matrices . . . . . . . . . . . . . . . . . . 45
2.2.2 Unit triangular matrices . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.2.3 Forward and backward substitution . . . . . . . . . . . . . . . . . . . 47
2.3 Triangular matrix decompositions . . . . . . . . . . . . . . . . . . . . . . . . 50
2.3.1 LU decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.3.2 LDMT decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3.3 LDLT decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
2.3.4 The Cholesky decomposition . . . . . . . . . . . . . . . . . . . . . . 53
2.4 Other matrix factorisations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.4.1 QR factorisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
2.4.2 Least squares problems . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4.3 Spectral decomposition . . . . . . . . . . . . . . . . . . . . . . . . . 63
2.4.4 Mahalanobis transformation and distance . . . . . . . . . . . . . . . 66
2.4.5 The singular value decomposition (SVD) . . . . . . . . . . . . . . . . 71
2.5 Principal components analysis (PCA) . . . . . . . . . . . . . . . . . . . . . 73

ii
CONTENTS iii

2.5.1 Derivation from the spectral decomposition . . . . . . . . . . . . . . 73

2.5.2 Total variation and variance explained . . . . . . . . . . . . . . . . . 75
2.5.3 Principal components from a sample variance matrix . . . . . . . . . 76
2.5.4 Construction from the SVD . . . . . . . . . . . . . . . . . . . . . . . 79
2.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

3 Inference, the MVN and multivariate regression 84

3.1 Inference and estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
3.2 Multivariate regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
3.2.1 Univariate multiple linear regression . . . . . . . . . . . . . . . . . . 88
3.2.2 The general linear model . . . . . . . . . . . . . . . . . . . . . . . . 91
3.2.3 Weighted errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
3.3 The multivariate normal (MVN) distribution . . . . . . . . . . . . . . . . . . . 94
3.3.1 Evaluation of the MVN density . . . . . . . . . . . . . . . . . . . . . 95
3.3.2 Properties of the MVN . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3.3.3 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . . . 98
3.3.4 MLE for the general linear model . . . . . . . . . . . . . . . . . . . . 100

4 Cluster analysis and unsupervised learning 102

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
4.1.2 Dissimilarity and distance . . . . . . . . . . . . . . . . . . . . . . . . 102
4.2 Clustering methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2.1 K-means clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
4.2.2 Hierarchical clustering . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4.2.3 Model-based clustering . . . . . . . . . . . . . . . . . . . . . . . . . 122

5 Discrimination and classification 123

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.2 Heuristic classifiers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
5.2.1 Closest group mean classifier . . . . . . . . . . . . . . . . . . . . . . 123
5.2.2 Linear discriminant analysis (LDA) . . . . . . . . . . . . . . . . . . . 125
5.2.3 Quadratic discrimination . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.2.4 Discrimination functions . . . . . . . . . . . . . . . . . . . . . . . . . 128
5.3 Maximum likelihood discrimination . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.1 LDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
5.3.2 Quadratic discriminant analysis (QDA) . . . . . . . . . . . . . . . . . 130
5.3.3 Estimation from data . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
5.4 Misclassification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
5.5 Bayesian classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

6 Graphical modelling 138

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.2 Independence, conditional independence and factorisation . . . . . . . . . 138
6.3 Undirected graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.3.1 Graph theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
6.3.2 Graphical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
6.4 Gaussian graphical models (GGMs) . . . . . . . . . . . . . . . . . . . . . . 145
CONTENTS iv

6.4.1 Partial covariance and correlation . . . . . . . . . . . . . . . . . . . 147

6.5 Directed acyclic graph (DAG) models . . . . . . . . . . . . . . . . . . . . . . 152
6.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.5.2 Directed graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.5.3 DAG models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153
6.5.4 Fitting to data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159

7 Variable selection and multiple testing 160

7.1 Regularisation and variable selection . . . . . . . . . . . . . . . . . . . . . . 160
7.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
7.1.2 Ridge regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
7.1.3 The LASSO and variable selection . . . . . . . . . . . . . . . . . . . 164
7.1.4 The elastic net . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
7.1.5 p >> n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
7.2 Multiple testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
7.2.2 The multiple testing problem . . . . . . . . . . . . . . . . . . . . . . 169
7.2.3 Bonferroni correction . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7.2.4 False discovery rate (FDR) . . . . . . . . . . . . . . . . . . . . . . . 170
Chapter 1

Introduction to multivariate data and

random quantities

1.1 Introduction
1.1.1 A few quotes...
Google’s Chief Economist Hal Varian on Statistics and Data:

I keep saying the sexy job in the next ten years will be statisticians. People
think I’m joking, but who would’ve guessed that computer engineers would’ve
been the sexy job of the 1990s?

Varian then goes on to say:

The ability to take data - to be able to understand it, to process it, to extract
value from it, to visualize it, to communicate it’s going to be a hugely important
skill in the next decades, not only at the professional level but even at the
educational level for elementary school kids, for high school kids, for college
kids. Because now we really do have essentially free and ubiquitous data.
So the complimentary scarce factor is the ability to understand that data and
extract value from it.

Source: FlowingData.com
The big data revolution’s “lovely” and “lousy” jobs:

The lovely jobs are why we should all enroll our children immediately in statis-
tics courses. Big data can only be unlocked by shamans with tremendous
mathematical aptitude and training. McKinsey estimates that by 2018 in the
United States alone, there will be a shortfall of between 140,000 and 190,000
graduates with “deep analytical talent”. If you are one of them, you will surely
have a “lovely” well-paying job.

Source: The Globe and Mail

1
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 2

The Search For Analysts To Make Sense Of ’Big Data’ (an article on an NPR programme)
begins:

Businesses keep vast troves of data about things like online shopping behav-
ior, or millions of changes in weather patterns, or trillions of financial transac-
tions — information that goes by the generic name of big data.
Now, more companies are trying to make sense of what the data can tell them
about how to do business better. That, in turn, is fueling demand for people
who can make sense of the information — mathematicians — and creating
something of a recruiting war.

Source: NPR.org
Also see this article in the NYT: For Today’s Graduate, Just One Word: Statistics

1.1.2 Data in the internet age

It is clear from the above quotes and articles that technology is currently having a dra-
matic impact on the way that data is being collected and analysed in the 21st Century.
Until recently, the cost of collecting and measuring data for statistical analysis and in-
terpretation meant that most analyses of statistical data concerned small data sets or
carefully designed experiments that were relatively straightforward to model. Today, most
organisations (commercial businesses and companies, as well as other institutions, such
as research institutes and public sector bodies) record and conduct most aspects of their
operation electronically, by default. Furthermore, technology for measuring, recording,
analysing, and archiving data is rapidly evolving, and this is leading to a so-called “data
explosion”. Most organisations have vast “data warehouses” of information which contain
vital information that could allow the organisation to work more effectively. The business
of extracting value from such data goes by various names in different contexts, including
“business intelligence”, “business analytics”, “predictive analytics”, “predictive modelling”,
“informatics”, “machine learning”, “data science”, “data mining”, as well as the more con-
ventional “data analysis and modelling”. Occasionally, even as “statistics”...
Similarly, many areas of scientific research are currently being similarly transformed.
Within physics, the CERN Large Hadron Collider is generating terabytes of data. In biol-
ogy, new technologies such as high throughput sequencing are resulting in routine gener-
ation of massive data sets. Here a typical (processed) datafile will contain many millions of
“reads”, each of which will consist of around 100 base pairs (letters), and a typical exper-
iment will generate several such data files. Such experiments are being conducted every
day in research institutions all over the world, and “bioinformaticians” skilled in statistics
and data analysis are in short supply.
Analysing and interpreting large and complex data sets is a significant challenge re-
quiring many skills, including those of statistical data analysis and modelling. It would be
unrealistic to attempt in a single module to provide all of the knowledge and skills neces-
sary to become a real “data scientist”. Here we will concentrate on some of the key statis-
tical concepts and techniques necessary for modern data analysis. Typical characteristics
of modern data analysis include working with data sets that are large, multivariate, and
highly structured, but with a non-trivial structure inconsistent with classical experimental
design ideas.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 3

1.1.3 Module outline

Chapter 1: Introduction to multivariate data and random quantities
This first chapter will lay the foundation for the rest of the course, providing an intro-
duction to multivariate data, mainly using R “data frames” as a canonical example,
and introducing some basic methods for visualisation and summarisation, illustrated
with plenty of worked examples.

Chapter 2: PCA and matrix factorisations

In this chapter we will examine how techniques from linear algebra can be used to
transform data in a way that better reveals its underlying structure, and the role that
matrix factorisations play in this. Principal components analysis (PCA) will be used
to illustrate the ideas.

Chapter 3: Inference, the MVN distribution, and multivariate regression

In this chapter we will introduce the simplest and most useful multivariate probability
distribution: the multivariate normal (MVN) distribution. We will construct it, exam-
ine some of its key properties, and look at some applications. We will also briefly
examine linear regression for multivariate outputs.

Chapter 4: Cluster analysis and unsupervised learning

In this chapter we will look at how it is possible to uncover latent structure in multi-
variate data without any “training data”. Clustering is the canonical example of what
data miners refer to as “unsupervised learning”.

Chapter 5: Discrimination and classification

In this chapter we will briefly examine the related problems of discrimination and
classification.

Chapter 6: Graphical modelling

In this chapter we will look how notions of conditional independence can be used
to understand multivariate relationships, and how these naturally lead to graphical
models of variable dependencies.

Chapter 7: Variable selection and multiple testing

In this final chapter we will look briefly at some of the issues that arise when working
with data sets that are large or high dimensional.

1.2 Multivariate data and basic visualisation

1.2.1 Tables
Many interesting data sets start life as a (view of a) table from a relational database. Not
at all coincidentally, database tables have a structure very similar to an R “data frame” (R
data frames are modelled on relational database tables). Therefore for this course we will
restrict our attention to the analysis and interpretation of data sets contained in R data
frames. Conceptually a data frame consists as a collection of n “rows”, each of which is a
p-tuple, xi = (xi1 , xi2 , . . . , xip ) from some set of the form S1 × S2 × · · · × Sp , where each Sj
is typically a finite discrete set (representing a factor ) or Sj ⊆ R. In the important special
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 4

case where we have Sj ⊆ R, j = 1, 2, . . . , p, we may embed each Sj in R and regard

each p-tuple xi as a vector in (some subset of) Rp . If we now imagine our collection of
p-tuples stacked one on top of the other, we have a table with n rows and p columns,
where everything in the jth column is an element of Sj . In fact, R internally stores data
frames as a list of columns of the same length, where everything in the column is a vector
of a particular type.∗ It is important to make a clear distinction between the rows and
the columns. The rows, i = 1, . . . , n typically represent cases, or individuals, or items,
or objects, or replicates, or observations, or realisations (from a probability model). The
columns j = 1, . . . , p represent attributes, or variables, or types, or factors associated with
a particular case, and each contains data that are necessarily of the same type.

Example: Insect sprays

To make things concrete, let’s look at a few simple data frames bundled as standard
data frames in R. We can get a list of R data sets using the R command data(). The
data set InsectSprays is a data frame. We can get information about this object with
?InsectSprays, and we can look at the first few rows with
> head(InsectSprays)
count spray
1 10 A
2 7 A
3 20 A
4 14 A
5 14 A
6 12 A
>
The data represents counts of numbers of insects in a given agricultural area for different
kinds of insect sprays. We can get a simple plot of the data with boxplot(count∼spray
,data=InsectSprays) (which shows that sprays C, D and E seem to be most effective),
and understand the content with
> class(InsectSprays)
[1] "data.frame"
> dim(InsectSprays)
[1] 72 2
> levels(InsectSprays$spray)
[1] "A" "B" "C" "D" "E" "F"
> table(InsectSprays$count)

0 1 2 3 4 5 6 7 9 10 11 12 13 14 15 16 17 19 20 21 22 23 24 26
2 6 4 8 4 7 3 3 1 3 3 2 4 4 2 2 4 1 2 2 1 1 1 2
> str(InsectSprays)
’data.frame’: 72 obs. of 2 variables:
$ count: num 10 7 20 14 14 12 10 23 17 20 ...
$ spray: Factor w/ 6 levels "A","B","C","D",..: 1 1 1 1 1 1 1 1 1 1
...
∗
This internal storage is different to the way that relational databases typically store tables — database tables are
typically stored internally as collections of rows. However, since R provides mechanisms for accessing data by rows
as well as columns, this difference will not trouble us.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 5

>
Consequently the data frame has n = 72 and p = 2. Formally, we can regard each row as
being a 2-tuple from Z × {A, B, C, D, E, F }. Note that although we can easily embed Z in
R, there is no natural embedding of the unordered factor {A, B, C, D, E, F } in R, and so it
will probably not be sensible to attempt to regard the rows as vectors in R2 .

1.2.2 Working with data frames

The previous example R session shows some useful methods for finding out some basic
information about a data frame. However, we very often wish to extract information from
data frames for subsequent analysis. We will look at methods for transforming data later.
Here we will just look at methods for extracting and subsetting rows and columns from
a data frame ready for further analysis. First, we can get the names of the variables
using names(), and the names of the cases with rownames(). We can illustrate with the
InsectSprays data frame that we have already introduced:
> names(InsectSprays)
[1] "count" "spray"
> rownames(InsectSprays)
[1] "1" "2" "3" "4" "5" "6" "7" "8" "9" "10" "11" "12" "13"
[14] "14" "15" "16" "17" "18" "19" "20" "21" "22" "23" "24" "25" "26"
[27] "27" "28" "29" "30" "31" "32" "33" "34" "35" "36" "37" "38" "39"
[40] "40" "41" "42" "43" "44" "45" "46" "47" "48" "49" "50" "51" "52"
[53] "53" "54" "55" "56" "57" "58" "59" "60" "61" "62" "63" "64" "65"
[66] "66" "67" "68" "69" "70" "71" "72"
Here the row names are not interesting, and just correspond to row numbers, but in some
cases the row names are useful. We can access individual elements of the data frame by
row and column number, so that InsectSprays[3,1] returns 20. Alternatively, we can
access using names, so that InsectSprays[3,"count"] is equivalent.
We can column slice a data frame, to get a data frame with fewer columns, using single
bracket notation, so that InsectSprays[1] and InsectSprays["count"] both return a
data frame containing a single column.
A vector representing a single column of a data frame can be obtained in a variety of
ways — the following are all equivalent:
InsectSprays[[1]]
InsectSprays[["count"]]
InsectSprays$count
InsectSprays[,1]
InsectSprays[,"count"]
Note the subtle but important distinction between a vector and a data frame containing a
single column.
We can also access and subset data frames by rows. A row can be accessed by
name or number, so that here InsectSprays[3,] and InsectSprays["3",] both return
a data frame consisting of just one row, corresponding to the third row of the original
InsectSprays data frame. Note that there is no direct way to return a vector correspond-
ing to a row, since the elements of the row will typically not all be of the same type. We
will look at some of the things that can be done with a data frame where all of the values
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 6

are numeric in the next example. We can create a data frame containing a subset of rows
from the data frame by using a vector of row numbers or names. eg.
> InsectSprays[3:5,]
count spray
3 20 A
4 14 A
5 14 A
>
We can also subset the rows using a boolean vector of length n which has TRUE elements
corresponding the the required rows. eg. InsectSprays[InsectSprays$spray=="B",]
will extract the rows where the spray factor is B. If the vector of counts is required, this
can be extracted from the resulting data frame in the usual way, but this could also be
obtained more directly using
> InsectSprays$count[InsectSprays$spray=="B"]
[1] 11 17 21 11 16 14 17 17 19 21 7 13
>

Example: Motor Trend Cars

We will illustrate a few more important R techniques for working with data frames using
another simple built-in data set, mtcars, before moving on to look at some more interest-
ing examples. Again, the str() command is very useful.
> str(mtcars)
’data.frame’: 32 obs. of 11 variables:
$ mpg : num 21 21 22.8 21.4 18.7 18.1 14.3 24.4 22.8 19.2 ...
$ cyl : num 6 6 4 6 8 6 8 4 4 6 ...
$ disp: num 160 160 108 258 360 ...
$ hp : num 110 110 93 110 175 105 245 62 95 123 ...
$ drat: num 3.9 3.9 3.85 3.08 3.15 2.76 3.21 3.69 3.92 3.92 ...
$ wt : num 2.62 2.88 2.32 3.21 3.44 ...
$ qsec: num 16.5 17 18.6 19.4 17 ...
$ vs : num 0 0 1 1 0 1 0 1 1 1 ...
$ am : num 1 1 1 0 0 0 0 0 0 0 ...
$ gear: num 4 4 4 3 3 3 3 4 4 4 ...
$ carb: num 4 4 1 1 2 1 4 2 2 4 ...
>
So we see that we have n = 32 and p = 11, and here also that all 11 variables are numeric
, so we could regard the rows as points in R11 if we felt that was sensible. However, closer
inspection reveals that some columns are counts, and some are numeric encodings of
binary variables, so direct embedding into R11 may not be such a great idea. This data
frame has sensible names for both the rows and the columns
> names(mtcars)
[1] "mpg" "cyl" "disp" "hp" "drat" "wt" "qsec" "vs" "am" "
gear"
[11] "carb"
> rownames(mtcars)
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 7

[1] "Mazda RX4" "Mazda RX4 Wag" "Datsun 710"

[4] "Hornet 4 Drive" "Hornet Sportabout" "Valiant"
[7] "Duster 360" "Merc 240D" "Merc 230"
[10] "Merc 280" "Merc 280C" "Merc 450SE"
[13] "Merc 450SL" "Merc 450SLC" "Cadillac Fleetwood"
[16] "Lincoln Continental" "Chrysler Imperial" "Fiat 128"
[19] "Honda Civic" "Toyota Corolla" "Toyota Corona"
[22] "Dodge Challenger" "AMC Javelin" "Camaro Z28"
[25] "Pontiac Firebird" "Fiat X1-9" "Porsche 914-2"
[28] "Lotus Europa" "Ford Pantera L" "Ferrari Dino"
[31] "Maserati Bora" "Volvo 142E"
>
Let us suppose that we are only interested in the first 7 variables. A new data frame
containing just those columns can be created with
myCars=mtcars[c("mpg","cyl","disp","hp","drat","wt","qsec")]
or more consisely with
myCars=mtcars[1:7]
A particular row can be extracted with
> myCars["Fiat 128",]
mpg cyl disp hp drat wt qsec
Fiat 128 32.4 4 78.7 66 4.08 2.2 19.47
>
And a subset of rows (say, economical cars) can be extracted with
> myCars[myCars$mpg>30,]
mpg cyl disp hp drat wt qsec
Fiat 128 32.4 4 78.7 66 4.08 2.200 19.47
Honda Civic 30.4 4 75.7 52 4.93 1.615 18.52
Toyota Corolla 33.9 4 71.1 65 4.22 1.835 19.90
Lotus Europa 30.4 4 95.1 113 3.77 1.513 16.90
>
There are obviously many simple plots one can do to investigate the relationship between
pairs of variables, so boxplot(mpg∼cyl,data=myCars) shows how fuel economy varies
with the number of cylinders, and plot(myCars$disp,myCars$mpg) shows fuel economy
as a function of engine size. Plots such as these are useful, but multivariate analysis is
concerned with developing more sophisticated methods for understanding the relation-
ships between larger groups of variables.
Before moving on to look at more interesting examples, it is worth noting that in cases
such as this, where all of the variables are numeric, we can use the function as.matrix()
to coerce the data frame into an n × p real-valued matrix, which can be used as a regular
numeric matrix in other R functions. This is very useful!
> class(myCars)
[1] "data.frame"
> carsMat=as.matrix(myCars)
> class(carsMat)
[1] "matrix"
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 8

> dim(carsMat)
[1] 32 7
>

Example: Galaxy data

Several of the example data sets that we will consider in this course are associated with
the book [HTF]. There is a CRAN R package containing data sets from the book called
ElemStatLearn. Typing library(ElemStatLearn) or require(ElemStatLearn) will
load the package provided that it is installed. If these commands give an error, the most
likely explanation is that the package is not installed. It should be possible to install on
a machine with an internet connection using install.packages("ElemStatLearn"). A
list of data sets in the package can be obtained with
require(ElemStatLearn)
data(package="ElemStatLearn")
help(package="ElemStatLearn")
We will look first at the galaxy data frame (use ?galaxy for background information about
the data).
> str(galaxy)
’data.frame’: 323 obs. of 5 variables:
$ east.west : num 8.46 7.96 7.47 6.97 6.47 ...
$ north.south : num -38.2 -35.9 -33.7 -31.4 -29.2 ...
$ angle : num 102 102 102 102 102 ...
$ radial.position: num 39.1 36.8 34.5 32.2 29.9 ...
$ velocity : int 1769 1749 1749 1758 1750 1745 1750 1753 1734
1710 ...
>
Clearly here we have n = 323 and p = 5, and it appears at first as though all of the
variables are numerical, making it sensible to regard the rows as vectors in R5 . For a
numerical data frame such as this, we can produce a set of all pairs of scatterplots of one
variable against another using the pairs() command. So pairs(galaxy) gives such
a scatterplot matrix, as shown in Figure 1.1. In fact, the plot() command defaults to
pairs() for a data frame, so plot(galaxy) works just as well, but it is probably better to
be explicit about exactly what kind of plot is required.
Scatterplot matrices such as this are one of the simplest yet most useful methods for
beginning to understand a multivariate dataset, provided that the number of variables is
not too large. All pairwise plots of one variable against another are shown in the matrix.
The row determines the variable plotted on the vertical (y) axis, and the column deter-
mines the variable plotted on the horizontal (x) axis. For example, the plot in the last
column of the first row shows a plot of east.west against velocity. Close inspection of
these plots reveals that angle in fact looks as though it takes on only a small number of
discrete values. We can confirm this with
> table(galaxy$angle)

12.5 43 63.5 92.5 102.5 111 133

51 28 46 38 80 45 35
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 9

−40 0 40 −40 0 40
● ●
● ●
● ● ●●● ● ●●
●● ●
● ● ●
●
● ●●
●● ●
● ● ●
● ● ●●●
●●● ●
●● ●●●
●● ●●● ●● ●●
●
●●●● ●●
●●
●
●
●
● ●
●
●
● ●●●●● ●
●
● ●
●
●
●
●
● ●
● ●
●
● ●●
●
●●●
●●●●
●●●
●
●●
● ●
●●●
●● ●
●
●●
● ●●
● ●●
●

10
●● ●
●●
●●● ●
●● ●
●
● ●
●
● ●●●
●● ●
●
● ●
● ●
● ● ●
● ● ●●
●
●●●
●●●
●
●
●●
●●
●● ●● ●●
● ●● ●●
●
● ●●● ●
●● ●
●●
●●
●●●
●●
●●
●●
●
●●●
● ●
●●●●●
●●●
●● ●
● ● ●
● ●
●
● ●
● ●
● ●●
●
●●
●
●●●
●●
●●●
●●●
●● ●●●
● ●
●
● ●●●
●
●●
●
●●
●●
●
east.west ●●●
●●●
●●●
●●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●●●
●●●●
●●● ●
●
● ●
●
●
●
●
● ●●
● ●
●
●
●
●
● ●
● ●●
●●
●●●
●●
●●●
●
●●
●
●●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●●
●●
●●●
●●●● ●
●
●●●
●●
●●
●●
●●
●●
●
●●●●●
●
●
●●
●●
●●
●
●
●●
●
●●
●
●
●●
●
●●●
●
●●
●●
●
● ●
●●
●●
●●
●
●●●
●●●●●
●● ●●
●●
●●
●●● ● ● ● ● ● ● ●●●
●●●
●●
●● ●●
●●●
● ● ● ●
●● ●
●●
●●
●●
●●●●
●
●
●●
●●●
●
●
●●
●●
●
●
●●●
●●
●●
●
●●
●
●
●●
● ●
●●● ●
●
●
●
●
●
●
●
●
●
●
●●
● ●
● ●
●
● ●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
● ●
●●
●
●
●
●●
●
●●
●●
●● ●
●●●●●●
● ●●●●
●
●●● ●●
●●
●●
●●
●●
● ●● ● ●
●● ●●●
●●
● ●
● ● ●
●
● ●
● ● ●
●●
●●
●
●●●
●●●●
●●● ●
●● ●●●
●● ●
●●
● ●
●●●●●
●● ● ● ● ● ● ●● ● ● ● ●
●

−30
●
● ● ●
● ●● ●
●● ●
● ● ● ● ●● ●
●● ●
● ●●
●●●● ●
● ●●
● ●
● ●
● ●
●● ●● ●●● ● ●

●
●●
● ●
●● ● ●
●●
● ●
● ●●
● ● ●
●
●
●●● ●
●●
● ●●
●●●
●
● ●●●●●
●● ●
● ● ●●
● ●● ●
●● ●
● ● ●
●
●●●●● ● ●
●●●
●● ●
● ●●
● ●
● ● ● ● ●
●●
●●
●
●●
● ●●● ●
●●●
●
●●●●● ●●
●●
●●●●●● ● ● ●
● ● ● ● ● ● ●● ●
● ●● ● ●
●●●
−40 20
● ● ● ● ●
●●●● ● ●
● ● ● ●● ● ●
●
●
●●●●●●
●
●
●●
●
●
●
●
● ●
●●●
●●
●●●●●●●●● ●
●
●
●
●
●
● ● ● ●
● ●
●
●
●
●●●
●
●
●
●●
●
●
●
●●
●
●
●● ●●●
●
●
●●
●●
●
●●●
● ●●
●●
●
●●
●
●
●
●
●●
●
●
●●●
●
●
●
●●
●●●
●●
●●●
●
●●
●
●
●
●●●
●●
●
●
●● ●● ●●
●● ●●
●●●●
●● ● ● ●
● ●●
● ●●● ●
● ●
●●
●
●● ●●
●● ●
●●●
●● ●●
●●●
●
●●
●
●
●●
●●
●●●●●●
● ●●
●●●
●●●●
●
● ●●●●●●
●● ●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●●
●●●●● north.south ●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
● ● ●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●●
●●
●
●●
●●
●●
●●●
●● ● ●
●●●
●●
● ●
●
●
●
●
●
●
●
●●
●●
●●
●●●●●
●● ● ●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
● ●●
●
●●
●
●
●●
● ●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●● ●●● ● ●
●●●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●●
●● ●
● ●● ●●
●● ●● ●
● ● ●
●●● ●●●
● ●
●●
●●● ● ●●● ●● ●●●
●●
●
●
●
●●
●●
●●●
● ●● ● ●● ●● ●
● ●●
● ●●● ●●
●● ●●
●●
●
●●
●
● ●● ●
●
●
●
●●
●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●
●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
● ●
●●
●●
●●●●● ●●●●●●● ●

100
●
●●
●●●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●●●
●●● ●
●●
●●●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●●
●
●●
●
●●● ●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●●
●
●●
●● ●●
●●
●
●●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●●
●● ●
●●
●● ●●●
●● ●●
●●●●●
●
●●
●
●●●
●●
●●
●
●
●
●●
●●
●●
●
●
●
●●
●
●●
●
●
●●
●●
● ●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●● ●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●● ●●
●●
●●● ●
●● ●●● ● ●●●●●●
●●
●●
●●●
●
●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●
●●●
●●●
●●●
●●●●●
●●
● ●
●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●
●●●
●●
●●●●
●●●
●● angle ●
●●
●●
●●
●●●
●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●●
●●
●●
●● ●
●●
●●
●●●
●●
●●● ●
●●●
●●●
●●●
●
●●
●
●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●
●● ●
●
●●
●●
●●
●●
●●
●
●●
●●
●●
●●
●
●●
● ●
●● ●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
● ●●
●
●●
●
●●
●
●●
●
●●
●●●

20
●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●●●
●● ●
●
●
●●
●
●●
●
●
●●
●
● ●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●
●●
●
●●
●●
●
●●
●●
●●●

●● ●● ●
●●
● ●
●
● ●●
●● ●
● ● ● ●●
●● ●
●
●
40

●● ●●● ●
●
● ●
● ● ● ● ● ●● ● ●
●
●
● ●●●
●●
●
●
●
●
●●●●●●●●● ●●
●
●
●●●●●
● ●●
●●●
●●● ● ●
●
●
●
● ●
●
● ●●
● ●
● ●
●
●
●●
●
●●●●
● ●●
●●
●
●
●●●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●●●
●
●●
●●●
●
●
●●
●
●●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●
●
●
● ●●
●
● ●
●●
● ●
●
●
●
● ●
●
●●
●
● ●
●●
●
●
●
●●●● ●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●●
● ●●
● ● ● ● ● ● ● ● ● ● ●●● ●
● ●●●
●
●●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
radial.position ●●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●●●●●
●●●
●
●●
●
●●●
●●
●
●●
●
● ● ●
●●
●●
● ● ●●
●
●
●●
●
●●
●
● ●
● ● ●
● ●●
● ● ●
● ● ●●
● ●
●
●
●● ●
●● ●
● ●●●●●
−40

●● ● ●●
● ● ● ● ●●● ● ●● ● ●●
● ● ● ●
●●● ●●●●●●
●
●
●
●
●●●● ● ●
●●●
● ●
● ●●
●●
●●
●
●
●●
●● ● ● ●
● ●
●●● ● ●
● ●
●●●
●
● ●
●●● ● ●
● ●●● ●●●
●●
●
●●
●●●●●●●
●●
●
●
●
●
●
●●
●
●
●
●
●
● ●
● ●
●●
●●●
● ●●
●●●
●
●●
●●
●
●●●
●
●
●
●
● ●●
● ●
●
●
● ●
●●
●
●
●
●●
●●
●
●●●
● ●●
● ● ●

1400 1700
●●●
●
●●●
●
●●●
●●●●
●● ●●
●●
●
●
●●
●●
●
●●
●●● ● ●
●
● ●
● ●
●●●
●
●●
●●
●
●●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●●
●●●●●●
●●●●●
●
●
●
●●
●●
●●●●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●●
●● ● ●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●●
●
●●●
●
●●
●
●
●●●●●
●●●
●●●● ●
●
●
●
●
●
●
●●
●●●●
●
●●●●●●
●●
●●
●
●
●
●
●●
●
●●
●
● ●● ●
●
●
●
●
●
●●● ●
● ●
●
● ●●●
●
●●
●●●
●●
●● ●
●
●●
●
●
●
●
●●
●
●● ●●
●
● ●● ●●● ●●●
●●
●● ●●● ●●
●●●● ●
● ● ●● ●
●
●●●
●
●●
●●● ● ● ●
● ● ● ●
● ●● ●● ●● ●●
●●●
●●●
●●●● ● ●
●
● ●●
●●●●
●●●
●●● ●●
●● ●●
●●●●●
●
●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●●●●●
●●
●●
●
●
●●
●
●
●
●●
●●●
●
●●● ● ●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●●
● ● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●●
●
●
●
● ●●●
●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●● ●●
velocity
●●●●●●●●●● ●●●●
●
●●
●● ● ●
●●●●●●●
● ●
●●
● ●●
●
●
●●●
●
●●
●●
●
●
●
●
●
●●
●●●
● ●●
●● ●
● ●●
● ● ●
●
●
●
● ●●
●
●●
● ●●
●
●
●
●
●
●
●●●
●●
●●●●● ●
●●
●
●●●●●
●●
●●
●
●
●
●
●●
●● ●
●●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●●
●●
●
●●● ●●●●●
●
● ●
●
●
●
●●
●
●
●●●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●● ●
●
●●
●●●
● ●●
●
●●
●
●●
●●●
● ●
●● ● ●●
●●
●●
●●●
●
●●
● ●●
● ● ●●
●● ●●
● ●
●
●●● ●

−30 0 20 20 60 100 1400 1600

Figure 1.1: Scatterplot matrix for the Galaxy data

>
which shows that angle does indeed take on just 7 distinct values, and hence perhaps
has more in common with an ordered categorical variable than a real-valued variable. We
can use angle as a factor to colour the scatterplot as follows
pairs(galaxy,col=galaxy$angle)
to obtain the plot shown in Figure 1.2.
This new plot indicates that there appear to be rather simple relationships between the
variables, but that those relationships change according to the level of the angle “factor”.
We can focus in on one particular value of angle in the following way
pairs(galaxy[galaxy$angle==102.5,-3])
Note the use of -3 to drop the third (angle) column from the data frame. The new plot is
shown in Figure 1.3.
This plot reveals that (for a given fixed value of angle) there is a simple deterministic
linear relationship between the first three variables (and so these could be reduced to
just one variable), and that there is a largely monotonic “S”-shaped relationship between
velocity and (say) radial.position. So with just a few simple scatterplot matrices and
some simple R commands we have reduced this multivariate data frame with apparently
5 quantitative variables to just 2 quantitative variables plus an ordered categorical factor.
It is not quite as simple as this, as the relationships between the three positional variables
varies with the level of angle, but we have nevertheless found a lot of simple structure
within the data set with some very basic investigation.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 10

−40 0 40 −40 0 40
●● ● ●
● ● ●●● ● ●●
●● ●
● ●● ●
● ●
●●
● ●
● ● ●
● ● ●●●
●●● ●
●● ●●●
●● ●●● ●● ●
●
●
●●●● ●●
●●
●
●
●
● ●●
●● ●●●●
●
●
●
● ●
●
●
●
●
● ●
● ●
●
● ●●
●
●
●●●
●●●●
●●●
●
●●
● ●●
●
●●● ●
●
●●
● ●●
● ●●●

10
●● ●
●●
●●● ●
●● ●
●●●●
● ●●
●●
● ●
●
● ●
● ●
● ● ●
● ● ●●
●
●●●●●
●
●
●●
●●
●● ●● ●●● ●● ●●
●
● ●●● ●●●
● ●●●
●●
●●●
●●
●●
●●
●
●●●
●●●
● ●
●●●●
●●● ●
● ● ●
● ●
●
● ●
● ●
● ●●
●
●●
●
●●●
●●
●●●
●●●
●● ●●●
● ●
●
● ●●●
●
●●
●
●
●●
●●
east.west ●●●
●●●
●●●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●●
●●●
●●●
●●●
● ●
● ●
●
●
●
● ●●
● ●
● ●
● ●
● ●●
●●
●●●
●●
●●●
●●
●
●●
●
●
●●
●
●●
●
●
●●
●
●●
●●
●●
●●●
●●●● ●
● ●●
●
●●
●●
● ●
●
●●
●
●●
●●
●●●
●●
●
●●
●
●●
●●●
●
●●
●●
●●●
● ●
●●
●
●
●●
●●
●
●●
●●
●
●●
●●
●●
●● ●
●
●●
●
●●
●
●●
●
●
●●
●●
●●
●●●
●●●
● ●
●
● ● ●
●
● ●●
●
● ● ● ●●
●●
●
●
●●
●●●
●●
●
●●
●
●●
●
●
●●
●
● ●
●●
●●●
●
●●●
●●●●
●
●●●● ●
●●
●●
●●●
●●● ●
●
●●●
●●
●●●
●
●
●●●
●
●●
●●
●
●
●
●●●
●●
●
●
●
●
●●
●
●
●
●●
● ●
●
●
●
●
●
●
●
●
● ●●
● ●
● ●
●
● ●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●●
●
●
●● ●
●●
●
●
●
●
●●
●●
●
●● ●
●●●
● ●●●●
●
●●● ●
●●●
●
●●
●●
●●
● ● ●
●● ●
●● ●●●
● ●
●
● ●
●
● ●
● ●
● ●
●●
●
●●
●●●●●●●
●● ●
●● ●●●
●●
● ●●
●
●
● ●●●●

−30
●
● ● ●
● ●
● ●●
● ●
● ● ● ● ●● ●
●● ●
● ●●
●●●● ●
● ●
●● ●
● ●
● ●
●● ●● ●●● ● ●

●
●●
● ●
●● ● ●
●●
● ●
● ● ●●
● ● ●
●
●
●●● ●
●●
● ●●
●●●
●
● ●●●●●●
●● ● ●
● ● ●●●● ●
●● ●
● ●
● ●
●
●●●●● ● ●
●●●● ●
● ●●
● ●
● ● ● ● ●
●
●
●●●
●
●●
● ●●● ●
●●●
●●●●● ●
●●
●●●●●● ● ● ●● ● ●● ● ● ●● ●
● ●● ● ●
●●●
−40 20

● ● ● ● ● ●● ● ●
● ● ● ●● ● ●
●
●
●●●●●●
●
●
●●
●
●
●
●
● ●
●●●
●●
●●● ●●●●●● ●
●
●
●
●
●
● ●
● ● ●
● ●
●
●
●
●
●●●
●
●
●
●●
●
●
●
●
●●
●
●● ●●●
●
●
●●
●●
●
●●●
● ●●
●●
●
●●
●
●
●
●
●●
●
●
●●●
●
●
●
●●
●●●
●●
●●●
●
●●
●
●
●
●●●
●●
●
●
●● ●● ●●
●● ●●
●●●●
●● ● ● ●
● ●●●●● ● ●
●
●●
●
●● ●●
●● ●
●●●
●● ●●
●●●
●●
●
●
●●
●●
●●●●●●
● ●●
●● ●●
●●●●
●●●
●● ●●●●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●●●●●
●
●
●
●●●●●
north.south ●
●
● ●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
● ●●●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●●
●●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●●
●
●●
●
●
●
●●●
●●
●
●●
●
●●
●●
● ●
●● ● ●●
●●
●
●● ●
●
●
●
●
●
●
●
●●
●●
●●
●●
●●● ● ●
●
●
●
●
●●
●
●
●●●
●
●
●
● ●
●
●
●
●●
●
●
●●
● ●
●●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●● ●●● ● ●
●
●
●●●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●●
●
●
● ● ● ● ● ● ● ●●● ●
●●
●
●●● ●
●● ●
●●● ●● ●
● ●
●●●●
● ● ●●
●● ●
●●
●●
● ●●●●●●
●●
●●
●
● ●● ●● ● ● ●●● ●● ●●
●
●●
●
● ●● ●
●
● ●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●
●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
● ●
●●
●●
●●●●● ●●●●●●● ●

100
●
●●
●●●
●●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●●●
●●● ●
●●
●●●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●●
●●
●
●●
●
●●● ●
●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●
●
●●
●
●●
●
●
●●●
●
●●
●● ●●
●●
●
●●
●●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●●
●●●
●● ●
●●
●● ●●●
●● ●●
●●●●●
●
●●
●
●●●
●●
●●
●
●
●
●●
●●
●●
●
●
●
●●
●
●●
●
●
●●
●●
● ●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●● ●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●● ●●
●●
●●● ●
●● ●●● ● ●●●●●●
●●
●●
●●●
●
●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●●
●●
●●●
●●●
●●●
●●●●●
●●
● ●
●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●
●●●
●●
●●●●
●●●
●● angle ●
●●
●●
●●
●●●
●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●●
●●
●●
●● ●
●●
●●
●●●
●●
●●● ●
●●●
●●●
●●●
●
●●
●
●● ●●●●●●●●●●●●●●●●●●●●●●●●●● ●
●● ●
●
●●
●●
●●
●●
●●
●
●●
●●
●●
●●
●
●●
● ●
●● ●●
●●
●●●
●●
●●●
●●
●●●
●●
●●●
●●
● ●●
●
●●
●
●●
●
●●
●
●●
●●●

20
●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●●●
●● ●
●
●
●●
●
●●
●
●
●●
●
● ●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●●
●● ●
●●
●
●●
●●
●
●●
●●
●●●

● ●●
● ● ●
●● ●
● ● ●
●● ●
●●● ● ● ●
● ● ● ● ●
●
40

●● ●
●●● ●
●
● ●
● ● ● ● ● ●● ● ●
●
● ●●●●
●
●
●
●
●●●●●●●●● ●●
●
●
●●
●●●●
● ●●
●●●
●●● ● ●
●
●
●
● ●
●
● ●
●●
● ●
● ●
●
●
●●
●
●●●●
● ●●
●●
●
●
●●●●
●
●
●●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●●●●
●
●●
●●●
●
●
●●
●
●●
●
●●
●●
●
●●●
●
●
●
●
●
●
●
●
●●● ●
●
●
●
●
●
●
●
●
●
●
●●
● ●●
● ●
●
●
●
● ●
●
●●
●
● ●
●●
●
●
●
●●●● ●
● ●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●●●
●●
●
●●● ●●
● ● ● ● ● ● ● ● ● ● ●●● ●
● ●●●
●
●●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
radial.position ●●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●● ●
●●●●●
●●●
●
●●
●
●●●
●●
●
●●
●
● ●
● ●
●●
●●
● ● ●●
●
●
●●
●
●●
●
● ●
● ● ●
● ●
● ●
● ●
● ● ●●
● ●
●
●
●● ●
●● ●
● ●●●●●
−40

●● ● ●●
● ● ● ●●● ● ● ●● ● ●●
● ● ● ●
●●● ●●●●●●
●
●
●
●
●●●● ●
● ●
●●●
● ●
● ●●
●●
●●
●
●
●●
●● ● ● ●
● ●
●●● ● ●
● ●
●●
●●
● ●
●●● ● ●
● ●●● ●●●
●●
●
●●
●●●●●●●
●●
●
●
●
●
●
●●
●
●
●
●
●
● ●
● ●
●●
●●●
● ●●
●●●
●
●●
●●
●
●●●
●
●
●
●
● ●●
● ●
●
●
● ●
●●
●
●
●
●
●●
●
●●
●●
● ●●
● ● ●

1400 1700
●●●
●
●●●
●
●●●
●●●●
●● ●●
●●
●
●
●●
●●
●
●●
●●● ● ●
●
● ●
● ●
●●●
●
●●
●
●●
●●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●●
●●●●●●
●●●●●
●
●
●
●●
●●
●●●●●
●
●
●
●●
●
●
●●
●
●●
●
●
●
●
●●
●●
●● ● ●●
●
●
●
●●
●
●
●
●
●
●
●●
●
●●●
●
●
● ●
●
●●
●
●
●●●●●
●●●
●●●● ●
●
●
●
●
●
●
●●
●●●●
●
●●●●●●
●●
●●
●
●
●
●
●●
●
●●
●
● ●● ●
●
●
●
●
●●● ●
● ●
●
● ●●●
●
●●
●●●
●●
●● ●
●
●●
●
●
●
●
●●
●
●● ●●
●
● ●● ●●● ●●●
●●
●● ●●● ●●
●●●● ●
● ● ●● ●
●
●●●
●
●●
●●● ● ● ● ● ● ●
● ●● ●● ●● ●●
●●●
●●●
●●●● ● ●
●
● ●●
●●●●
●●●
●●● ●●
●● ●●
●●●●●
●
●●
●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●
●●●●●●
●●
●●
●
●
●●
●
●
●●
●●
●●
●
●●●
●
● ●●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●●
● ● ●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
● ●●●
●●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●● ●●●
velocity
●●●●●●●●●●●●●●●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●●●●
● ●●
● ●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●
●● ●
● ●●
●
●●
●●●
●
●
●
●
● ●
●
●●
●
●●
●
●●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●●
●●●●● ●●
●
● ●
●
●●
●
●●
●
●●
● ●
●●
●●
●●
●●
●●
●●
●
●● ●
● ● ●
●
●●
●
●●
●●
●●
●●
●●●
●
● ●
●●● ●
●
● ●●●● ●
●● ●● ●●
●●● ●
● ●●
● ● ●●
●● ● ● ●
●
●●● ●

−30 0 20 20 60 100 1400 1600

Figure 1.2: Scatterplot matrix for the Galaxy data, coloured according to the levels of the angle
variable

−40 0 20 1400 1600 10

● ● ●
●●● ● ● ●●●
●●● ●●● ●●●●
●●● ●●● ●
●●
●●● ●●● ●
●
●●
●●● ●●● ●●
●
●●● ●●●● ●●●
●●
east.west ●●● ●
●●● ●●●●●●●
0

●●● ●
●
●●● ●●● ●●●
●●●
● ●
●●● ●●● ●●●●
●●● ●●● ●●
●●● ●●● ●
●
●
●
●●● ●
−10

●●●●● ●●● ●
●
●●●
●●
● ●●● ●
●●●

●●● ●●● ●●
● ●
40

●●● ●●● ●
● ●●
●●
●●● ●●● ●
●●
●●● ●●● ●
●●
●●● ●●● ●●●●●
●●● ●●● ● ●●●
●●
●●● ●●● ●●
● ●●●●
●●●
north.south ●●● ●
−40 0

●●● ●●● ●●●●

●●●
●●● ●●● ●●
●●● ●●● ●
●●●●
●●● ●●● ●
●
●
●●● ●●● ●●
●●●●
●●● ●●● ● ●●●
● ● ●

● ● ●
40

● ●●● ● ●●●
●●● ●●● ●●●●
●●● ●●● ●
●●
●●● ●●● ●
●
●●
●●● ●●● ●●
●
●●●● ●●● ●●●
●●
●
●●● ●●● radial.position ●●●●●●●
−40 0

●●● ●
●
●●● ●●● ●●●
●●●
● ●
●●● ●●● ●●●●
●●● ●●● ●●
●●● ●●● ●
●
●
●
●●●●
● ●● ●
●
●●●
●●
● ●●● ●
●●● ● ●
●●●

●●●●
● ●●● ● ●●●●
●
1700

●●●●●●●● ●
● ● ●●●●●●●●●● ●●●●●●●● ●
●
●●●
● ● ●
●●●
● ●●●
● ●
●●
●●● ●●
●●
●● ●●
●●●
●●●
●
●
●●
●● ●●●
●
●
●●
●●●
●●●
●●●
●●
●●●
●●●
●●●
●●
velocity
●●● ●●● ●●●●
●● ●●●●● ●●●●
●●
●●●●● ●●●● ●●● ●●●
● ●●●●● ● ●●●●●
1400

●●●●●●● ●
●●●●●● ●● ●●●●●●

−10 0 5 10 −40 0 20

Figure 1.3: Scatterplot matrix for the subset of the Galaxy data corresponding to an angle vari-
able of 102.5
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 11

Example: Microarray data

The dataset nci represents measurements of the expression levels of 6,830 genes in
64 different tissue samples taken from cancerous tumors. The first thing we need to
think about here is what the rows and columns should be. Which is n and which is p?
The R command class(nci) reveals that this data set is actually stored as a matrix
and not a data frame, so that is no help. The command dim(nci) indicates that the
matrix is stored with 6,830 rows and 64 columns, but matrices can be easily transposed
using the t() command, so that doesn’t really tell us anything other than how the data
happens to be stored. This problem can arise whenever a multivariate data set consists of
measurements of exactly the same type in all rows and columns. Here the measurements
are expression levels (basically, a measure of the amount of a particular kind of mRNA in a
sample). Since all measurements have essentially the same units, and can be stored in a
matrix, it is technically arbitrary which should be considered rows and columns. However,
from the viewpoint of multivariate data analysis, it matters a great deal.
Here, fundamentally, the observations are 64 different tissue samples, and the vari-
ables are the 6,830 genes, so n = 64 and p = 6, 830, so in some sense the nci matrix
should be transposed. However, it is common practice to store microarray data this way,
and sometimes for certain analyses people really do take the genes to be the observa-
tions and the samples to be the variables. Generally with multivariate analysis we wish
to use observations to learn about the relationship between the variables. This data set
is no exception, in that we wish to use the samples in order to learn something about
the relationship between the genes. Sometimes people work with the transposed data
in order to use the genes in order to learn something about the relationship between the
samples. However, at a fundamental level, the measurements are of the amount of a
particular gene, in units specific to that gene, and in (essentially) identical units for each
sample. So it is really the genes that represent the variables, as columns of the same
type, in some sense...
There is actually something very special about this data in that p > n, and in fact,
p >> n (p is much bigger than n). Multivariate data with this property is often referred to
as “wide data” (as opposed to the more normal “tall data”), for obvious reasons. This is
quite different to the other examples we have looked at, and rather atypical of classical
multivariate data. Indeed, many of the classical statistical techniques we shall examine in
this course actually require n > p, and ideally need n >> p in order to work well. Data
with p > n will cause many classical algorithms to fail (and this is one reason why people
sometimes work with the transposed data). We will revisit this issue at appropriate points
in the course, and look at some simple techniques for analysing wide data near the end
of the course.
It is obviously impractical to produce a 6, 380 × 6, 380 scatterplot matrix for this data! In
fact, even if we work with the transposed data, a 64 × 64 scatterplot matrix is also unlikely
to be helpful. So we must use other strategies for visualisation here. The standard way
to visualise high-dimensional real-valued matrices is as an image, or heat-map. We can
start with a very simple image plot, using
image(nci,axes=FALSE,xlab="genes",ylab="samples")
giving rise to the plot in Figure 1.4.
Each individual rectangular area, or pixel, of the image represents a single value from
the input matrix. By default the origin is at the bottom left, so the matrix is transposed
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 12

Figure 1.4: Image plot of the nci cancer tumor microarray data, using the default colour scheme

(or rather, rotated anti-clockwise). Using the default “heat” colour scheme, low values
are represented as “cold” red colours, and “hot” bright yellow/white colours represent the
high values. Although many people find this colour scheme intuitive, it is not without
problems, and in particular, can be especially problematic for colour-blind individuals. We
can produce a simple greyscale image using
image(nci,axes=FALSE,xlab="genes",ylab="samples",col=grey((0:31)/31))
which has the lightest colours for the highest values, or using
image(nci,axes=FALSE,xlab="genes",ylab="samples",col=grey((31:0)/31))
to make the highest values coloured dark. Another popular colouring scheme is provided
by cm.colors(), which uses cyan for low values, magenta for high values, and middle
values white. It is good for picking out extreme values, and can be used as
image(nci,axes=FALSE,xlab="genes",ylab="samples",col=cm.colors(32))
The image() function is a fairly low-level function for creating images, which makes it
very flexible, but relatively difficult to produce attractive plots. For imaging multivariate
data there is a higher-level function called heatmap() which produces attractive plots
very simply. A basic plot can be obtained with
heatmap(nci,Rowv=NA,Colv=NA,labRow=NA,col=grey((31:0)/31))
leading to the plot shown in Figure 1.5.
Note that this function keeps the natural orientation of the supplied matrix, and by
default will label rows and columns with the row and column names associated with the
matrix. Note that the options Rowv=NA,Colv=NA are critical, as they disable the default
behaviour of the function to use clustering to reorder the rows and columns to reveal
interesting structure present in the data. We will look again at this behaviour when we
study clustering later in the course. This is a computationally intensive operation, so we
don’t want to attempt this on a full microarray matrix at this stage.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 13

Figure 1.5: Heatmap of the nci cancer tumor microarray data, using a greyscale colour scheme
with darker shades for higher values

Example: Handwritten digit images

The next example we consider is a dataset representing images of handwritten digits. The
underlying motivation is automated recognition of US ZIP codes written on envelopes, for
automated routing of US mail. A large number of manually classified examples are in the
data set zip.train. A sample of images is shown in Figure 1.6.
The dataset is stored as a 7, 291 × 257 matrix. Each row represents a single classified
image, so that n = 7, 291 and p = 257. Each image is actually a greyscale image of 16 by
16 pixels. These 256 pixels can be “unwrapped” into a 256-dimensional vector, which can
be regarded as a vector in R256 . The first number in each row is the manually classified
digit for the image. Therefore each row is a 257-tuple from {0, 1, 2, 3, 4, 5, 6, 7, 8, 9} × R256 .
Although it is tempting to regard the digit as an integer or ordered categorical factor,
it should really be considered an unordered factor, as we do not expect the images to
change gradually as the digit value increases. The ElemStatLearn package includes
a convenience function, zip2image() for converting a row back into a 16 × 16 matrix
oriented correctly for passing into the image() function. For example, the image corre-
sponding to the fifth row can be displayed as simply as
image(zip2image(zip.train,5))
This can be tidied up and imaged in greyscale using
image(zip2image(zip.train,5), col=gray(15:0/15), zlim=c(0,1), xlab="",
ylab="" )
giving the image shown in Figure 1.7.
For exploratory analysis of the image data, we can easily strip off the first column
corresponding to the classification using
myDigits=zip.train[,-1]
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 14

1.0
0.8
0.6
0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 1.6: A sample of images from the zip.train dataset, generated with the command
example(zip.train)
1.0
0.8
0.6
0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 1.7: Image corresponding to the fifth row of the zip.train dataset — the digit shown is
a “3”
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 15

to give a multivariate dataset with n = 7, 291 and p = 256, where each row represents
a vector in R256 . It is perhaps concerning that representing the images as a vector in
R256 loses some important information about the structure of the data — namely that it is
actually a 16 × 16 matrix. So, for example, we know that the 2nd and 18th elements of
the 256-dimensional vector actually correspond to adjacent pixels, and hence are likely to
be highly correlated. This is clearly valuable information that we aren’t explicitly including
into our framework. The idea behind “data mining” is that given a sufficiently large number
of example images, it should be possible to “learn” that the 2nd and 18th elements are
highly correlated without explicitly building this in to our modelling framework in the first
place. We will look at how we can use data on observations to learn about the relationship
between variables once we have the appropriate mathematical framework set up.

1.3 Representing and summarising multivariate data

Clearly an important special case of (multivariate) data frames arises when the p variables
are all real-valued and so the data can be regarded as an n×p real-valued matrix. We saw
in the previous example that even when the original data is not in exactly this format, it is
often possible to extract a subset of variables (or otherwise transform the original data) to
give a new dataset which does have this form. This is the classical form of a multivariate
dataset, and many classical modelling and analysis techniques of multivariate statistics
are tailored specifically to data of this simple form. Classically, as we have seen, the
rows of the matrix correspond to observations, and the columns correspond to variables.
However, it is often convenient to regard the observations as column vectors, and so care
must be taken when working with multivariate data mathematically to transpose rows from
a data matrix in order to get a column vector representing an observation.

Definition 1 The measurement of the jth variable on the ith observation is denoted xij ,
and is stored in the ith row and jth column of an n × p matrix denoted by X, known as the
data matrix.

However, we use xi to denote a column vector corresponding to the ith observation,

and hence the ith row of X. That is
 
xi1
xi2 
xi = (xi1 , xi2 , . . . , xip )T =  ..  .
 
 . 
xip

We denote the column vector representing the jth variable by x(j) , which we can obviously
define directly as  
x1j
 x2j 
x(j) =  ..  .
 
 . 
xnj
We can begin to think about summarising multivariate data by applying univariate sum-
maries that we already know about to the individual variables.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 16

1.3.1 The sample mean

As a concrete example, it is natural to be interested in the sample mean of each variable.
The sample mean of x(j) is denoted x̄j , and is clearly given by
n
1X
x̄j = xij .
n i=1

We can compute x̄j for j = 1, 2, . . . , p, and then collect these sample means together into
a p-vector that we denote x̄, given by

x̄ = (x̄1 , x̄2 , . . . , x̄p )T .

Note, however, that we could have defined x̄ more directly as follows.

Definition 2 The sample mean of the n × p data matrix, X, is the p-vector given by the
sample mean of the observation vectors,
n
1X
x̄ = xi .
n i=1

A moments thought reveals that this is equivalent to our previous construction. The vec-
tor version is more useful however, as we can use it to build an even more convenient
expression directly in terms of the data matrix, X. For this we need notation for a vector
of ones. For an n-vector of ones, we use the notation 1 n , so that

1 n ≡ (1, 1, . . . , 1)T .

We will sometimes drop the subscript if the dimension is clear from the context. Note that
1 n T 1 n = n (an inner product), and that 1 n 1 p T is an n × p matrix of ones (an outer product),
sometimes denoted Jn×p . Pre- or post-multiplying a matrix by a row or column vector of
ones has the effect of summing over rows or columns. In particular, we can now write the
sample mean of observation vectors as follows.
Proposition 1 The sample mean of a data matrix can be computed as
1 T
x̄ = X 1 n.
n

Example: Galaxy data

The R command summary() applies some univariate summary statistics to each vari-
able in a data frame separately, and this can provide some useful information about the
individual variables that make up the data frame
> summary ( g a l a x y )
e a s t . west n o r t h . south angle radial . position
Min . : −29.66693 Min . : −49.108 Min . : 12.50 Min . : −52.4000
1 s t Qu . : −7.91687 1 s t Qu.: −13.554 1 s t Qu . : 63.50 1 s t Qu.: −21.3500
Median : −0.06493 Median : 0.671 Median : 92.50 Median : −0.8000
Mean : −0.33237 Mean : 1.521 Mean : 80.89 Mean : −0.8427
3 r d Qu . : 6.95053 3 r d Qu . : 18.014 3 r d Qu. : 1 0 2 . 5 0 3 r d Qu . : 19.6500
Max . : 29.48414 Max . : 49.889 Max . :133.00 Max . : 55.7000
velocity
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 17

Min . :1409
1 s t Qu. : 1 5 2 3
Median :1586
Mean :1594
3 r d Qu. : 1 6 6 9
Max . :1775
>

The apply() command can also be used to apply arbitrary functions to the rows or
columns of a matrix. Here we can obtain the mean vector using
> apply ( galaxy , 2 ,mean)
e a s t . west n o r t h . south angle r a d i a l . p o s i t i o n velocity
−0.3323685 1.5210889 80.8900929 −0.8427245 1593.6253870
>

The 2 is used to indicate that we wish to apply the function to each column in turn. If
we had instead used 1, the mean of each row of the matrix would have been computed
(which would have been much less interesting). We can also use our matrix expression
to directly compute the mean from the data matrix
> as . vector ( t ( g a l a x y ) %∗% rep ( 1 , 3 2 3 ) / 3 2 3 )
[1] −0.3323685 1.5210889 80.8900929 −0.8427245 1593.6253870
>

where %*% is the matrix multiplication operator in R. Since R helpfully transposes vectors
as required according to the context, we can actually compute this more neatly using
> rep(1,nrow(galaxy))%*%as.matrix(galaxy)/nrow(galaxy)
east.west north.south angle radial.position velocity
[1,] -0.3323685 1.521089 80.89009 -0.8427245 1593.625
>
It is typically much faster to use matrix operations than the apply() command. Note
however, that R includes a convenience function, colMeans() for computing the sample
mean of a data frame:
> colMeans(galaxy)
east.west north.south angle radial.position
-0.3323685 1.5210889 80.8900929 -0.8427245
velocity
1593.6253870
and so it will usually be most convenient to use that.

Example: toy problem

Before moving on, it is worth working through a very small, simple toy problem by hand,
in order to make sure that the ideas are all clear. Suppose that we have measured the
height (in metres) and weight (in kilograms) of 5 students. We have n = 5 and p = 2, and
the data matrix is as follows  
1.67 65.0
1.78 85.0
 
X=  1.60 54.5

1.83 72.0
1.80 94.5
We can calculate the mean vector, x̄ for this data matrix in three different ways. First start
by calculating the column means.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 18

For x̄1 we have

n
1X 1
x̄1 = xi1 = (1.67 + 1.78 + 1.60 + 1.83 + 1.80)
n i=1 5
8.68
= = 1.736,
5
and similarly
1
x̄2 = (65.0 + 85.0 + 54.5 + 72.0 + 94.5) = 74.2.
5
So our mean vector is
x̄ = (1.736, 74.2)T.
Next we can calculate it as the sample mean of the observa-
tion vectors as
n          
1 X 1 1.67 1.78 1.60 1.83 1.80
x̄ = xi = + + + +
n i=1 5 65.0 85.0 54.5 72.0 94.5
 
1.736
= .
74.2
Finally, we can use our matrix expression for the sample mean
 
1
  1    
1 T 1 1.67 1.78 1.60 1.83 1.80   1 8.68
1 = 1.736
x̄ = X 1 n = = .
n 5 65.0 85.0 54.5 72.0 94.5  1 5 371.00
 74.2
1

This example hopefully makes clear that our three different ways of thinking about com-
puting the sample mean are all equivalent. However, the final method based on a matrix
multiplication operation is the neatest both mathematically and computationally, and so
we will make use of this expression, as well as other similar expressions, throughout the
course.

1.3.2 Sample variance and covariance

Just as for the mean, we can think about other summaries starting from univariate sum-
maries applied to the individual variables of the data matrix. We write the sample variance
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 19

of x(j) as s2j or sjj , and calculate as

n
1 X
sjj = (xij − x̄j )2 .
n − 1 i=1
Similarly, we can calculate the sample covariance between variables j and k as
n
1 X
sjk = (xij − x̄j )(xik − x̄k ),
n − 1 i=1
which clearly reduces to the sample variance when k = j. If we compute the sample
covariances for all j, k = 1, 2, . . . , p, then we can use them to form a p × p matrix,
 
s11 s12 · · · s1p
s21 s22 · · · s2p 
S =  .. ..  ,
 
.. . .
 . . . . 
sp1 sp2 · · · spp
known as the sample covariance matrix, or sample variance matrix, or sometimes as the
sample variance-covariance matrix. Again, a with a little thought, one can see that we
can construct this matrix directly from the observation vectors.
Definition 3 The sample variance matrix is defined by
n
1 X
S= (xi − x̄)(xi − x̄)T .
n − 1 i=1
This expression is much simpler than computing each component individually, but can be
simplified further if we consider the problem of centering the data.
We can obviously write the sample covariance matrix as
n
1 X
S= wi wi T ,
n − 1 i=1
where wi = xi − x̄, i = 1, 2, . . . , n, and we can regard the wi as observations from a new
n × p data matrix W. We can construct W as
 
x̄T
x̄T 
W = X −  .. 
 
 . 
x̄T
= X − 1 n x̄T
 
1 T
= X − 1n X 1n T
n
1
= X − 1 n1 nTX
 n 
1 T
= In×n − 1 n 1 n X
n
= Hn X,
where
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 20

Definition 4
1
Hn ≡ In×n − 1 n 1 n T
n
is known as the centering matrix.
So we can subtract the mean from a data matrix by pre-multiplying by the centering matrix.
This isn’t a numerically efficient way to strip out the mean, but is mathematically elegant
and convenient. The centering matrix has several useful properties that we can exploit.
Proposition 2 The centering matrix Hn has the following properties:
1. Hn is symmetric, Hn T = Hn ,
2. Hn is idempotent, H2n = Hn ,
3. If X is an n × p data matrix, then the n × p matrix W = Hn X has sample mean equal
to the zero p-vector.
Proof
These are trivial exercises in matrix algebra. 1. and 2. are left as exercises. We will
use symmetry to show 3.

1 T
w̄ = W 1n
n
1
= (HnX)T1 n
n
1
= XTHn1 n
n
1 1
= XT(In×n − 1 n1 nT)11n
n n
1 1
= XT(11n − 1 n1 nT1 n)
n n
1 T
= X (11n − 1 n)
n
= 0.

Re-writing the sample variance matrix, S, in terms of the centered data matrix W, is useful,
since we can now simplify the expression further. We first write
 
w1 T
X n  w2 T 
wi wi = (w1 , w2 , . . . , wn )  ..  = WT W,
T  
i=1
 . 
wn T
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 21

using properties of block matrix multiplication. From this we conclude that

n
1 X 1
S= wi wi T = WT W
n − 1 i=1 n−1

We can then substitute back in our definition of W using the centering matrix to get
1 1 1 1
S= WT W = (Hn X)T Hn X = XT Hn T Hn X = XT Hn X,
n−1 n−1 n−1 n−1
using symmetry and idempotency of Hn . This gives us the rather elegant result:
Proposition 3 The sample variance matrix can be written
1
S= XT Hn X.
n−1
We shall make considerable use of this result.

Example: galaxy data

With the galaxy data, we can start by computing the column variances with
> apply(galaxy,2,var)
east.west north.south angle radial.position
144.6609 523.8497 1462.6269 670.2299
velocity
8886.4772
which gives diagonal elements consistent with the built-in var() function:
> var(galaxy)
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
radial.position 1637.78301
velocity 8886.47724
as can be verified with
> diag(var(galaxy))
east.west north.south angle radial.position
144.6609 523.8497 1462.6269 670.2299
velocity
8886.4772
The built-in var() function is very efficient, and should generally be preferred as the
way to compute a sample variance matrix in R. However, we can easily check that our
mathematically elegant matrix result works using
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 22

> H=diag(323)-matrix(1/323,ncol=323,nrow=323)
> t(galaxy)%*%H%*%as.matrix(galaxy)/(323-1)
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
radial.position 1637.78301
velocity 8886.47724
This method is clearly mathematically correct. However, the problem is that building the
n × n centering matrix and carrying out a matrix multiplication using it is a very compu-
tationally inefficient way to simply strip the mean out of a data matrix. If we wanted to
implement our own method more efficiently, we could do it along the following lines
> Wt=t(galaxy)-colMeans(galaxy)
> Wt %*% t(Wt)/(323-1)
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
radial.position 1637.78301
velocity 8886.47724
This uses a couple of R tricks, relying on the fact that R stores matrices in column-major
order and “recycles” short vectors. We can improve on this slightly by directly constructing
the outer product 1 n x̄T .
> W=galaxy-outer(rep(1,323),colMeans(galaxy))
> W=as.matrix(W)
> t(W)%*%W/(323-1)
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 23

radial.position 1637.78301
velocity 8886.47724
In fact, we can do even better than this by exploiting the R function sweep(), which is
intended to be used for exactly this sort of centering procedure, where some statistics are
to be “swept” out of a data frame.
> W=as.matrix(sweep(galaxy,2,colMeans(galaxy)))
> t(W)%*%W/(323-1)
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
radial.position 1637.78301
velocity 8886.47724

Example: ZIP code digits

For our handwritten image data, we can easily compute the sample variance matrix for
the 256 columns corresponding to the image with
> v=var(zip.train[,-1])
> dim(v)
[1] 256 256
> prod(dim(v))
[1] 65536
There probably isn’t much point in printing the resulting 256 × 256 matrix to the screen
and reading through all 65,536 covariances. However, we have already seen that we can
sometimes use images to visualise large matrices, and we can also do that here
image(v[,256:1],col=gray(15:0/15),axes=FALSE)
giving the plot shown in Figure 1.8.
The image shows a narrow dark band down the main diagonal corresponding to the
pixel variances, but alternating between dark and light, due to the pixels near the edge
of each row having relatively low variance (due to usually being blank). Then we see
parallel bands every 16 pixels due to pixels adjacent on consecutive rows being highly
correlated. The bands perpendicular to the leading diagonal are at first a little more
mysterious, but interestingly, these arise from the fact that many digits have a degree of
bilateral symmetry, with digits such as “1” and “8” being classic examples, but others such
as “6”, “5” and “2” also having some degree of bilateral symmetry.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 24

Figure 1.8: An image of the variance matrix for the zip.train dataset

Example: toy example

To work through a trivial example by hand, consider a simple data matrix with n = 3 and
p = 2:  
2 3
X = 4 1 
6 2
First, let’s calculate the sample covariance matrix using the original definition:

For this we first need to compute

x̄ = (4, 2)T
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 25

Then we have
n
1 X
S= (xi − x̄)(xi − x̄)T
n − 1 i=1
       
1 2 4 2 4 T
= − −
2 3 2 3 2
       
4 4 4 4 T
+ − −
1 2 1 2
        
6 4 6 4 T
+ − −
2 2 2 2
      
1 −2 0 2
= (−2, 1) + (0, −1) + (2, 0)
2 1 −1 0
     
4 −2 0 0 4 0
= + +
−2 1 0 1 0 0
   
1 8 −2 4 −1
= =
2 −2 2 −1 1
Next, let’s calculate using our formula based around the centering matrix

First calculate
2 −1 −1
 
1 3 3 3
−1 −1 
H3 = I3 − 1 31 3T = 
3
2
3 3
,
3 −1 −1 2
3 3 3
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 26

and then use this to compute

1
S= XTHnX
n−1
 23 −1 −1
  
 2 3
1 2 4 6  −1 32 −1 3
= 3 3 3
 4 1
2 3 1 2 −1 −1 2
6 2
 3 3 3
  −2 1
1 2 4 6 
= 0 −1
2 3 1 2
2 0
   
1 8 −2 4 −1
= =
2 −2 2 −1 1
Finally, let’s think about how to do this computation a little more efficiently. Let’s start
by constructing  
−2 1
W = X − 1 3 x̄T =  0 −1
2 0
Now
 
 −2 1 
1 1 −2 0 2 
S = WT W = 0 −1
2 2 1 −1 0
2 0
   
1 8 −2 4 −1
= =
2 −2 2 −1 1

So, the variances of the variables are 4 and 1, and the covariance between them is -1,
which indicates that the variables are negatively correlated.

Before moving on, it is worth dwelling a little more on the intermediate result
1
S= WT W,
n−1
which shows us that the matrix S factorises in a very nice way. We will return to matrix
factorisation later, but for now it is worth noting that we can re-write the above as

S = AT A

where
1
A= √ W.
n−1
Matrices which factorise in this way have two very important properties
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 27

Proposition 4 Let S be a p × p matrix of the form

S = AT A,

where A is an n × p matrix. Then S has the following properties:

1. S is symmetric, ST = S,
2. S is positive semi-definite (or non-negative definite), written S ≥ 0. This means that
αT Sα ≥ 0 for all choices of p-vector α.
Proof

1. is trivial. For 2., note that

αTSα = αTATAα
= (Aα)T(Aα)
= kAαk22
≥ 0.

So the sample variance matrix is symmetric and non-negative definite.

1.3.3 Sample correlation

The sample correlation between variables i and j is defined by
sij
rij = .
si sj
Clearly, from the definition of sample covariance we have that rii = 1, i = 1, 2, . . . , p, and
putting w(i) = x(i) − x̄i 1 n ,
w(i)T w(j) w(i) · w(j)
rij = √ √ =
w(i)T w(i) w(j)T w(j) kw(i) kkw(j) k
and so by the Cauchy-Schwarz inequality we have |rij | ≤ 1, i, j = 1, 2, . . . , p. As with the
sample covariance matrix, we can create a p × p matrix of sample correlations, known as
the sample correlation matrix, R. If R = Ip , so that all off-diagonal correlations are zero,
then we say that the variables are uncorrelated.
If we now define D to be the diagonal matrix of sample standard deviations, that is
 
s1 0 · · · 0
 0 s2 · · · 0 
D = diag{s1 , s2 , . . . , sp } =  .. .. . . ..  ,
 
. . . .
0 0 · · · sp

then we get the following result.

CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 28

Proposition 5 The sample correlation matrix may be computed from the sample variance
matrix as
R = D−1 SD−1 ,
and conversely, the sample covariance matrix may be computed from the sample corre-
lation matrix as
S = DRD.

Proof
This follows since pre-multiplying by a diagonal matrix re-scales the rows and post-
multiplying by a diagonal matrix re-scales the columns. 

Proposition 6 The sample correlation matrix is symmetric and positive semi-definite.

Proof
We have already shown that the sample covariance matrix can be written as S = AT A.
Consequently, R can be written as
R = BT B,
where B = AD−1 . We showed earlier that matrices that can be written in this form must
be symmetric and positive semi-definite. 

Example: Galaxy data

Returning to the galaxy data example, we can compute the sample correlation matrix
directly using the built-in R command cor() as follows
> cor(galaxy)
east.west north.south angle radial.position
east.west 1.00000000 -0.11871407 -0.04674954 0.8476414
north.south -0.11871407 1.00000000 0.03011136 -0.4418112
angle -0.04674954 0.03011136 1.00000000 -0.0495015
radial.position 0.84764144 -0.44181124 -0.04950150 1.0000000
velocity 0.39818438 -0.89449554 0.01032294 0.6710883
velocity
east.west 0.39818438
north.south -0.89449554
angle 0.01032294
radial.position 0.67108826
velocity 1.00000000
So, for example, the correlation between radial position and velocity is 0.671, as we can
verify with
> cor(galaxy$radial.position,galaxy$velocity)
[1] 0.6710883
Using the built-in cor() function is the preferred way to compute a sample correlation ma-
trix, as it is computationally efficient. Nevertheless, we can calculate using Proposition 5
as
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 29

> diag(1/sd(galaxy)) %*% var(galaxy) %*% diag(1/sd(galaxy))

[,1] [,2] [,3] [,4] [,5]
[1,] 1.00000000 -0.11871407 -0.04674954 0.8476414 0.39818438
[2,] -0.11871407 1.00000000 0.03011136 -0.4418112 -0.89449554
[3,] -0.04674954 0.03011136 1.00000000 -0.0495015 0.01032294
[4,] 0.84764144 -0.44181124 -0.04950150 1.0000000 0.67108826
[5,] 0.39818438 -0.89449554 0.01032294 0.6710883 1.00000000

Example: ZIP digits

Again, we can easily compute the correlation matrix using
R=cor(zip.train[,-1])
but we cannot easily comprehend all of the numbers in this huge matrix. So we can image
the correlation matrix in exactly the same way that we imaged the sample variance matrix
earlier, viz
image(R[,256:1],col=gray(15:0/15),axes=FALSE)
giving the image shown in Figure 1.9.
We see that the overall pattern looks quite similar to the image of the variance ma-
trix we created earlier. However, whilst this image obviously loses the information about
variance (the diagonal is now solid black), it better highlights the correlation structure,
bringing out a somewhat stronger off-diagonal pattern.

Example: toy problem

Earlier we calculated by hand a 2 × 2 sample variance matrix
 
4 −1
S=
−1 1

We can now use Proposition 5 to calculate the associated correlation matrix as

   
1/2 0 4 −1 1/2 0
R=
0 1 −1 1 0 1
  
1/2 0 2 −1
=
0 1 −1/2 1
 
1 −1/2
= ,
−1/2 1
giving a sample correlation coefficient between the two variables of -0.5.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 30

Figure 1.9: An image of the sample correlation matrix for the zip.train dataset

1.4 Multivariate random quantities

So far everything we have been doing has been purely descriptive, focused on under-
standing the particular dataset we have available. Usually, however, in statistics we want
at some point to go beyond this, and use the data at hand to say something predictive
about the larger population from which the data was sampled. To do this, we need to
make some modelling assumptions. In the context of an n × p data matrix, X, the typical
assumption made is that the observations, xi ∈ Rp have been drawn independently from
some unknown probability distribution on Rp . So the really key distinction between the
rows and columns of a data matrix is that we typically assume that the rows are indepen-
dent but that the columns are not (necessarily). A random vector X takes the form
X = (X1 , X2 , . . . , Xp )T
where each Xj , j = 1, 2, . . . , p is a real-valued random variable. We do not assume that
the components of the random vector are independent of one another.
Definition 5 The expectation of a random p-vector X is defined by
E(X) = [ E(X1 ) , E(X2 ) , . . . , E(Xp )]T .
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 31

So, the expectation of a random vector is just defined to be the vector of expected values.
Similarly, the expected value of a random matrix is the matrix of expected values. Also, if
we recall that the covariance between two scalar random variables X and Y is defined by

Cov(X, Y ) = E([X − E(X)][Y − E(Y )])

(so for independent X, Y we will have Cov(X, Y ) = 0), we see that the variance matrix for
a random matrix is defined to be the matrix of covariances between pairs of elements:
Definition 6 The variance matrix of a random p-vector X is defined by

Var(X) = E [X − E(X)][X − E(X)]T .

Example: toy problem

Suppose that X = (X1 , X2 )T , where X1 ∼ Exp(1) and X2 ∼ Exp(2) are independent
random quantities. What is E(X) and Var(X)?

Calculating everything required one component at a time,

we have E(X1) = 1, Var(X1) = 1, E(X2) = 1/2, Var(X2) =
1/4, and since the variables are independent, we must have
Cov(X1, X2) = 0, giving
   
1 1 0
E(X) = , Var(X) = .
1/2 0 1/4
We will show later that in some appropriate sense, if we regard a multivariate dataset
X as a random sample from a multivariate distribution, then it turns out that the sample
mean of X is a “good” estimator of E(X), and similarly, the sample variance matrix, S, is
a “good” estimator of Var(X).

1.5 Transformation and manipulation of multivariate random

quantities and data
When working with multivariate random quantities and data, we often want to transform
them in some way in order to highlight some aspect of interest, or reduce dimension, etc.
We have already seen some examples of transformations applied to multivariate data.
Subsetting columns from a data matrix is a particularly simple example, and formally can
be regarded as a projection into a lower dimensional space. In particular, picking out two
columns in order to produce a scatter-plot is an example of a projection of the data into
a 2-dimensional space. The centering matrix was another example of a transformation of
the data matrix, corresponding to subtraction of the mean. These simple examples are
all examples of affine transformations, and so it will be helpful to remind ourselves of the
basic idea behind affine transformations before proceeding to see how we can use them
to transform random vectors and multivariate data.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 32

1.5.1 Linear and affine transformations

In Euclidean space a linear transformation (or linear map) is represented by a real-valued
matrix. For example, the map
f : Rp −→ Rq
will have matrix representation
f (x) = Ax
for some q × p matrix A. In the context of multivariate statistics we would expect that q ≤ p,
so that the map will either preserve the dimension of the original space q = p, or project
the data into a lower dimensional space q < p. Since it is clear that

f (ej ) = Aej = a(j) ,

the jth column of A, we can form A by considering the destination of ej , j = 1, 2, . . . , p in

the new space, Rq .
An important special case arises in the case q = 1, so that the map is real valued, and
represents a linear functional. Linear functionals can be represented with a p-vector, a,
as
f (x) = aT x.
It turns out that in many applications in multivariate statistics, we want to combine a linear
transformation with a shift in location. Such maps are referred to as affine transformations,
and the affine map
f : Rp −→ Rq
can be represented as
f (x) = Ax + b,
for some q × p matrix A and q-vector b. Note that f (0) = b, so that b can be identified as
the image of 0, and then
f (ej ) = a(j) + b,
and hence the columns of A are determined by

a(j) = f (ej ) − b.

Note that affine transformations can be represented with a single (q + 1) × (p + 1) matrix

of the form  
A b
A=
0T 1
by augmenting the vectors in both spaces with a “1”, since then
   
y x
=A
1 1
  
A b x
= T
0 1 1
 
Ax + b
= ,
1
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 33

and is therefore equivalent to

y = Ax + b.
This technique of representing an affine transformation as a linear transformation in an
augmented space makes it convenient to combine multiple affine transformations (via
matrix multiplication), and is used extensively in computer graphics.

Example: linear combination

The linear combination of variables

y = a1 x 1 + a2 x 2 + · · · + ap x p + b

can be represented as the affine transformation

y = aT x + b.

Example: 2d projection
We may view the selection of variables to plot in one window of a scatterplot matrix as a
2d projection of the multivariate data. Selection of components xk and xj corresponds to
the projection  T
ek
y= x,
ej T
and hence the linear transformation with A = (ek , ej )T .
More generally, we can project onto the linear subspace spanned by any two orthogo-
nal unit p-vectors, v 1 , v 2 using
A = (v 1 , v 2 )T .
This follows since we can write x as

x = y1 v 1 + y2 v 2 + w,

where w is orthogonal to v1 and v2 . From this it follows that v i · x = yi , and the result
follows. Note that, as above, A has the property AAT = I2 . 2d projections are obviously
very useful for visualisation of high-dimensional data on a computer screen.

1.5.2 Transforming multivariate random quantities

We can apply an affine transformation to a random p-vector X to get a random q-vector
Y . It is natural to wonder how the expectation and variance of Y relate to that of X.
Proposition 7 Let X be a random p-vector with expectation vector E(X) and variance
matrix Var(X). Consider an affine transformation of X from Rp to Rq given by

Y = AX + b,

where A is a q × p matrix and b is a q-vector. Then the expectation and variance of Y are
given by
E(Y ) = A E(X) + b, Var(Y ) = A Var(X) AT .
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 34

Proof
First consider E(Y ),

E(Y ) = E(AX + b)
   
  X1  
a
 11 12 a · · · a 1p b1 
X
 
= E ... ... . . . ...    ..  +  ... 
2
 
 . 
aq1 aq2 · · · aqp bq
 
Xp
 
a11X1 + a12X2 + · · · + a1pXp + b1
= E ... 
aq1X1 + aq2X2 + · · · + aqpXp + bq
 
E(a11X1 + a12X2 + · · · + a1pXp + b1)
= ... 
E(aq1X1 + aq2X2 + · · · + aqpXp + bq )
 
a11 E(X1) + a12 E(X2) + · · · + a1p E(Xp) + b1
= ... 
aq1 E(X1) + aq2 E(X2) + · · · + aqp E(Xp) + bq
= A E(X) + b.
Once we have established linearity of expectation for vector random quantities, the
variance result is relatively straightforward.

Var(Y ) = E [Y − E(Y )][Y − E(Y )]T

T


= E [AX + b − A E(X) − b][AX + b − A E(X) − b]
= E [AX − A E(X)][AX − A E(X)]T


T T


= E A[X − E(X)][X − E(X)] A
= A E [X − E(X)][X − E(X)]T AT

= A Var(X) AT
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 35


We are now in a position to establish that the variance matrix of a random vector
shares two important properties with sample covariance matrices.

Proposition 8 The variance matrix, Var(X) is symmetric and positive semi-definite.

Proof

T
Var(X) = { E [X − E(X)][X − E(X)] }T T

= E {[X − E(X)][X − E(X)]T}T

T


= E [X − E(X)][X − E(X)]
= Var(X) .
Similarly, for any p-vector α we have
T T T


α Var(X) α = α E [X − E(X)][X − E(X)] α
= E αT[X − E(X)][X − E(X)]Tα


T 2


= E {[X − E(X)] α}
≥ 0,
since non-negative random scalars cannot have negative ex-
pectation.


Example: mean and variance of a linear combination

The special case of linear functionals of random quantities turns out to be very important.
So consider the random (scalar) variable Y defined by

Y = αT X + b

where α is a given fixed p-vector and b is a fixed scalar. From our above results it is clear
that
E(Y ) = αT E(X) + b
and
Var(Y ) = αT Var(X) α.
Importantly, we note that the positive semi-definiteness of Var(X) corresponds to the fact
that there are no linear combinations of X with negative variance.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 36

Example: toy problem

Suppose that X = (X1 , X2 , X3 )T , where the components Xi ∼ P o(i), i = 1, 2, 3 are
mutually independent. Consider
   
1 2 3 0
Y = X+ .
4 5 6 1

What are E(Y ) and Var(Y )?

We first need to know that

   
1 1 0 0
E(X) = 2 and Var(X) = 0 2 0 .
3 0 0 3
The we can calculate
 
  1        
1 2 3   0 14 0 14
E(Y ) = 2 + = + = ,
4 5 6 1 32 1 33
3
and
  
  1 0 0 1 4
1 2 3 
Var(Y ) = 0 2 0 2 5
4 5 6
0 0 3 3 6
 
  1 4
1 2 3 
= 4 10
4 5 6
9 18
 
36 78
= .
78 174

Example: transforming standard random vectors

Suppose that we start with a random p-vector Z where the components are mutually
independent with zero expectation and variance 1. Now consider a p-vector Y where

Y = AZ + µ,

where A is a p × p matrix and µ is a p-vector. What is E(Y ) and Var(Y )?

CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 37

We first note that E(Z) = 0 and Var(Z) = Ip , and then compute

E(Y ) = A E(Z) + µ = µ

and
Var(Y ) = A Var(Z) AT = A Ip AT = AAT .
Notes:

1. Suppose we would like to be able to simulate random quantities with a given ex-
pectation µ and variance Σ. The above transformation provides a recipe to do so,
provided that we can simulate iid random quantities with mean zero and variance
one, and we have a method to find a matrix A such that AAT = Σ. We will examine
techniques for factorising Σ in the next chapter.

2. The special case where the elements of Z are iid N (0, 1) is of particular importance,
as then Y is said to have a multivariate normal distribution. We will investigate this
case in more detail later.

1.5.3 Transforming multivariate data

We can apply an affine transformation to each observation of a data matrix, X. Conceptu-
ally, we do this by transposing the data matrix to get column vectors of observations, then
pre-multiply by A, then add b to each column in order to get the transpose of the new data
matrix, Y. In other words
Y = (AXT + b11n T )T ,
leading to

Proposition 9 The affine transformation from Rp to Rq given by

y = Ax + b

can be applied to an n × p data matrix X to get a new n × q data matrix Y given by

Y = XAT + 1 n bT .

Again, we would like to know how the sample mean, ȳ and variance SY of the new data
matrix Y relate to the sample mean x̄ and variance matrix SX of X.

Proposition 10 The sample mean and variance of Y are related to the sample mean and
variance of X by
ȳ = Ax̄ + b, SY = ASX AT .

Proof
Let’s start with the sample mean,
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 38

1 T
ȳ = Y 1n
n
1
= (XAT + 1 nbT)T1 n
n
1
= (AXT1 n + b11nT1 n)
n
1
= AXT1 + b
n
= Ax̄ + b.
The sample variance matrix result is left as an exercise.


Example: Galaxy data

First let’s return to the galaxy data, and just extract two columns:
sub=galaxy[c("radial.position","velocity")]
This is clearly a projection into a 2d space, but there is no point in explicitly constructing
the linear transformation. We can then center the data. This is also an affine transforma-
tion, which we can construct and then plot with
subCentered=sweep(sub,2,colMeans(sub))
plot(subCentered,col=galaxy$angle,pch=19)
giving the figure shown in Figure 1.10.
Now suppose that we would like to apply a further transformation to this data, namely
rotation anti-clockwise through 10 degrees (π/18 radians). We know from elementary
geometry that this can be accomplished as a linear transformation with matrix
 
cos θ − sin θ
A= ,
sin θ cos θ

where θ = π/18. We can then apply this transformation to the data matrix and re-plot with
A=matrix(c(cos(pi/18),-sin(pi/18),sin(pi/18),cos(pi/18)),ncol=2,byrow=
TRUE)
subRotated=as.matrix(subCentered) %*% t(A)
plot(subRotated,col=galaxy$angle,pch=19)
giving the plot shown in Figure 1.11.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 39

●● ●
●●●●
●
●●● ● ● ●
●●●●● ●
● ●●● ●
●●●● ●
●
●
● ●
● ● ●●●
● ●●● ●
100 ●●●●●●● ● ● ● ●●● ●
●
● ● ● ●● ● ●
● ● ●
●●●
●●
●
● ●● ●● ● ● ●● ● ●
● ● ●● ●
●●● ●● ●
●●
velocity

● ● ● ● ●● ● ●●
● ● ● ● ●
●● ●● ● ●
● ●
● ● ●● ● ●●● ●●● ●
●●●●●● ● ●●
●● ●
0

● ●
● ● ● ●
● ● ● ● ●
●
● ●
● ●● ● ● ● ●● ●
●●●●
●● ● ● ●● ● ● ● ● ● ● ●
●●● ●●●● ●
● ●●●●
● ● ●●
● ● ● ●●● ● ● ●●
●●●●●
●● ● ●●●● ● ● ● ●
●● ● ● ● ●●
−100

● ●●● ● ● ● ●● ●
●●●●
● ●
●●●
● ●
● ●●●●● ●
●●●● ●
●● ● ● ●●
● ●● ● ● ● ●●●● ●
● ● ●
●● ● ●● ●● ● ● ●● ●
●● ● ●
●● ●●●●
●●

−40 −20 0 20 40 60

radial.position

Figure 1.10: A plot of velocity against radial position for the centered galaxy data
200

●
●●●● ●● ●
● ●●
●●●● ●●
●●●●
●●● ●●●
● ●●●
●●●●●●●
● ●
● ●● ●● ●●●
●●●● ● ●●
100

●● ●
● ●●● ● ●
subRotated[,2]

● ●● ●●●●●●●● ●●
● ● ● ●●● ● ●●
●●
● ●●● ●
●● ● ● ● ● ● ●● ●●
● ● ●● ●
● ● ● ● ●● ● ● ●● ●●●●●●● ● ●●
● ●●●● ●
● ●● ● ● ● ● ● ●
●●
0

●●●●● ●●●●●●● ●● ●
● ● ● ● ● ● ●●● ●●
●●
●●● ● ●● ● ● ●● ● ●
●●●●●● ●●●●
● ●●
● ●●● ● ● ●●
● ● ● ●● ●
●● ●●●
● ●● ●● ●● ● ● ● ● ●●
● ●
●●●●●●
−200 −100

● ●●● ●●
● ●●● ●●
●
●
●
●●●
●
●●●
● ● ● ●●
● ● ●
●● ●● ● ● ●
● ● ● ● ● ●● ● ● ●
●● ● ●●● ● ●● ●●●●
● ●●
●●●●
●
● ●●
● ●

−40 −20 0 20 40 60

subRotated[,1]

Figure 1.11: A plot of the rotated galaxy data

Example: ZIP digits

Suppose that we are interested in summarising each image by its mean pixel intensity.
This is a linear combination of the image vectors in R256 given by
1
y= 1 256 T x.
256
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 40

Histogram of y

800
600
Frequency

400
200
0

−1.0 −0.8 −0.6 −0.4 −0.2 0.0 0.2 0.4

Figure 1.12: Histogram of the mean pixel intensity for each image in the zip.train data

We could apply this transformation to our data in a variety of ways. First, directly from the
transformation as
y=zip.train[,-1] %*% rep(1/256,256)
We can histogram the resulting vector of mean image intensities with
hist(y,30,col=2)
giving the image shown in Figure 1.12.
Alternatively, we could have used the apply() function to create the vector using y=
apply(zip.train[,-1],1,mean). If we are interested in the variance of this vector, we
can obviously compute it using
> var(y)
[1] 0.0345547
On the other hand, if we still have the 256 × 256 variance matrix for the image data in a
variable v from a previous example, we could have computed this variance directly without
ever explicitly constructing the transformation as
> rep(1/256,256) %*% v %*% rep(1/256,256)
[,1]
[1,] 0.0345547

Example: normal transformation

If we have Z = (Z1 , Z2 )T , where Z1 and Z2 are iid N (0, 1), then the random variable

Y = AZ
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 41

●
● ●● ●

4
● ● ●
● ●● ●●● ●
●●● ● ●●●●●●●●●● ● ●●
● ●
● ● ●●● ●● ● ●
● ●● ● ●●
● ● ●●● ● ●●●
●● ● ● ● ● ● ● ● ●
2
●● ●●● ●
●●●●● ● ● ●● ●
●● ●●
● ● ● ● ●● ●
●
●●
●●
●
●
●●● ●
●
●●●
● ●
●
●●
●●● ●● ● ●
●●● ● ● ● ●
● ● ●● ● ● ●● ●●●● ●●●●
●
●● ●●
●
● ●● ●
● ●
●●● ●
● ●
●●●●●
●
●
●
●● ●●
●●
●
●
●
●●●
●
●
●
●●
●●
●●
●
●●
●
●
●
●
●
●●●
●●●●●
●●● ●
● ● ●
●● ●●● ●●
● ●●●●●●
●
●
●
●●
●
●
● ●
●●
●●
● ●
●
●
●●
●●
● ●
●●
●●
●●
●
●●
●
●
●●
●
●
●●
●●
●
●
●
●●●●●
●
● ●● ● ●●●
Y[,2]

●
● ● ●●●●● ● ●
●●●
●●
● ●
●
●●
●● ●
●
●● ●●●
●
●● ●●
●●●●●●
●●
● ●
●● ●
●●●●
●●
● ● ●●● ● ●
● ● ●●●●
●●
●●
●●
●●●●
●● ●
●●
●
●●●
●●
●●●●
●● ●
●●●●
● ●
● ●●●
●●● ●●
●● ●●●●● ● ● ● ●●●● ●●●●●●● ● ●● ●●
0

● ●● ●●● ●● ● ●●●
●● ● ●● ●
● ●●
●●●
●●
●
● ● ●●●
● ●
●●●
●●●●
●●●
● ●●●●
●
●●●
● ●●●●●●● ●●●● ● ●
●● ●●
●● ●● ●●●●
●●●
●
●
●●
●
●●●●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●●●●
● ●
●●
●●
●●
●●
●●●●●●● ●
●●
● ●●●
●●
●●●
●
● ● ●
● ●
●●
● ●
●●●
● ●
●●●
●
● ●●● ●
●●
●
●
●
●●
●●●
●●●
●
●●●
●
●
●●●●●●
●●●
●
●● ● ●● ● ● ●●
●●●
●●●●●●
●● ●
●
●●●●●●●
●● ● ●
● ● ●●●●
● ● ●
●● ● ● ●● ●
● ●●
●●
● ●
●●
●● ● ●●
●●●● ●
●● ● ● ●
●●●
● ●●●●● ●● ●
● ●●●●●●●
● ●● ●● ●
●●● ●●●
● ●●●●●●●● ●●● ● ● ●●
−2

● ●● ●●●● ●●
● ●● ●● ●●
●●●●● ● ●●
● ●
● ●● ●●●●●●●● ●●
●
●●●● ●
● ●● ●
● ●●●● ● ●
●● ●● ●●● ●● ●
●
● ● ●
● ●●
● ●● ● ●
● ●
−4

● ● ●
● ●● ● ●
●

−4 −3 −2 −1 0 1 2 3

Y[,1]

√
Figure 1.13: Scatterplot of some simulated correlated samples with correlation 1/ 2

where A is a 2 × 2 matrix will have variance matrix AAT . We can investigate this using R
with the matrix  
1 0
A=
1 1
as follows
> Z=matrix(rnorm(2000),ncol=2)
> A=matrix(c(1,0,1,1),ncol=2,byrow=TRUE)
> Y=Z %*% t(A)
> plot(Y,col=2,pch=19)
> A %*% t(A)
[,1] [,2]
[1,] 1 1
[2,] 1 2
> var(Y)
[,1] [,2]
[1,] 1.0194604 0.9926367
[2,] 0.9926367 1.9257760
So we see that the sample variance matrix for 1,000 simulated observations of Y is
very close to the√ true variance matrix. The sample correlation is also close to the true
correlation of 1/ 2. The plot of the samples is shown in Figure 1.13.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 42

Example: toy example

Consider the following data matrix with n = 4 and p = 3:
 
1 2 3
2 3 4
X= 3 2 1 .


1 3 2

Suppose that we wish to construct a new data matrix Y based on the affine transformation
   
1 1 0 0
y= x+ .
0 0 1 1
Construct the 4 × 2 data matrix Y.

   
1 2 3   0 1
 1 0
2 3 4  0 1
  
Y=  1 0 + 
3 2 1 0 1

0 1
1 3 2 0 1
   
3 3 0 1
5 4 0 1
   
= +
5 1 0 1

4 2 0 1
 
3 4
5 5
 
= .
5 2
4 3

When we introduced affine transformations, we discussed the augmented form of the

transformation, where the vector to be transformed was augmented with a “1” so that the
affine map could be accomplished as a linear transformation. If we wish, we can use
this method on a data matrix, X, by augmenting it with a column of “1”s. Then the key
transformation equation is  T 
A 0
(Y, 1 n ) = (X, 1 n ) T .
b 1
Note that this is very much analogous to prepending the design matrix with a column of
“1”s in multiple linear regression, in order to avoid explicit consideration of an intercept
term.
CHAPTER 1. INTRODUCTION TO MULTIVARIATE DATA 43

We now have all of the basic skills we need to be able to begin to think about more
sophisticated methods of multivariate data analysis. See Chapter 1 of [HTF] and Chapters
1 and 2 of [Everitt] for further background and details relating to the material in this
chapter.
Chapter 2

PCA and matrix factorisations

2.1 Introduction
2.1.1 Factorisation, inversion and linear systems
It should be clear by now that linear algebra and matrix computations are central to mul-
tivariate statistics. For the analysis of large and complex datasets it is essential to have
a basic knowledge of computational linear algebra, and this starts with an understanding
of the efficient solution of linear systems, and the role of various different matrix factori-
sations, often known as matrix decompositions, to transform general linear systems into
very special linear systems that are much easier to solve.
We have already seen a couple of instances of matrix factorisation. First, we showed
that the sample covariance matrix, S, could be decomposed as

S = AT A

for some matrix A. We were able to use this fact to deduce important properties of S.
Similarly, we showed that applying a matrix A to a vector of independent random quantities
with unit variance resulted in a vector with covariance matrix AAT . We noted that if we
were interested in simulating random quantities with a given covariance matrix Σ, we
could do this if we were able to find a matrix A such that

Σ = AAT .

This would be an example of a matrix factorisation, or matrix decomposition. Unsurpris-

ingly, there is not a unique solution to this factorisation problem, but in this chapter we
will see how to construct two different factorisations of Σ which both provide an effective
solution to this problem.
The solution of linear systems will also turn out to be a recurring theme. That is,
solving equations of the form
Ax = b
for x, given a specific matrix A and vector b. We know that in principle, if A is invertible,
the unique solution is given by
x = A−1 b.
However, it turns out that solving the problem directly in this way, by first computing A−1
and then pre-multiplying b by this matrix is a very inefficient and numerically unstable way

44
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 45

to approach the problem. It is hardly ever necessary to directly compute the inverse of
a matrix. There are usually far more efficient ways to solve a linear system, sometimes
by solving the linear system directly (say, using Gaussian elimination), or more usually by
solving in multiple steps using matrix factorisations. By way of motivation, suppose that
we have an efficient way to factorise A as

A = LU,

where L and U are lower and upper triangular matrices, respectively.∗ We can then solve
Ax = b by writing the problem as
LUx = b.
If we define v = Ux, we can first solve

Lv = b

for v and then solve

Ux = v
for x. It turns out that triangular linear systems are very easy to solve, so solving two trian-
gular linear systems is much faster than solving one general linear system. To progress
further we need to know more about triangular matrices, linear systems, and factorisa-
tions.

2.2 Triangular matrices

2.2.1 Upper and lower triangular matrices
It is possible to develop a theory of triangular matrices and triangular matrix factorisation
for non-square matrices, but for the purposes of this course this is unnecessary, so we
will restrict ourselves to thinking about square matrices.

Definition 7 An n × n matrix A is lower triangular if all elements above the leading diag-
onal are zero, that is aij = 0, ∀j > i. Similarly, a matrix is upper triangular if all elements
below the leading diagonal are zero, that is aij = 0, ∀i > j.

Clearly only diagonal matrices are both upper and lower triangular. Upper and lower
triangular matrices have very similar properties (since one is just a transpose of the other).
We will focus mainly on lower triangular matrices. It should be assumed that there are
exactly analogous properties for upper triangular matrices unless otherwise stated.

Proposition 11 1. The sum of two lower triangular matrices is also lower triangular

2. The product of two lower triangular matrices is also lower triangular

3. If a lower triangular matrix is invertible, the inverse is also lower triangular

4. The determinant of a lower triangular matrix is the product of the diagonal elements
∗
We will explain what this means in the next section. For now it suffices to know that they are just special matrices.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 46

5. A lower triangular matrix is invertible iff all diagonal elements are non-zero

6. The eigenvalues of a lower triangular matrix are the diagonal elements of the matrix

Proof

1. Clear from definition

2. Left as an exercise

3. Thinking about using the Gaussian elimination method to construct this inverse, it is
clear that it will only be necessary to use row operations which are lower triangular,
leading to a lower triangular inverse

4. Direct expansion of the determinant along the top row in terms of minor determi-
nants, applied recursively, makes this clear

5. The determinant will be non-zero iff all diagonal elements are non-zero, by previous
result

6. The eigenvalues are given by the roots of |A − λ I |. This matrix is lower triangular,
so the determinant is the product of diagonal elements.

2.2.2 Unit triangular matrices

Definition 8 A matrix L is unit lower triangular if it is lower triangular, and has the addi-
tional property that all elements along the leading diagonal are 1, that is, lii = 1, ∀i.

In addition to all of the properties of lower triangular matrices, unit lower triangular matri-
ces have some additional useful properties.

Proposition 12 1. The product of two unit lower triangular matrices is unit lower trian-
gular

2. The inverse of a unit lower triangular matrix is unit lower triangular

3. The determinant of a unit lower triangular matrix is 1, and hence all unit lower trian-
gular matrices are invertible

4. The eigenvalues of a unit lower triangular matrix are all equal to 1

Proof

1. Left as an exercise

2. Again, clear by thinking about Gaussian elimination and row operations

3. The determinant is the product of the diagonal elements

4. |L − λ I | = (1 − λ)n .
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 47


It is clear that a non-singular lower triangular matrix A can be factorised in the form

A = DL

where D is diagonal and L is unit lower triangular, by choosing D = diag {a11 , a22 , . . . , ann }
and L = D−1 A.

Example: Toy problem

Write the 2 × 2 lower triangular matrix
 
2 0
A=
−1 4
in the form A = DL, where D is diagonal and L is unit lower triangular.

If we put
D = diag {2, 4} , D−1 = diag {1/2, 1/4}
we get
    
1/2 0 2 0 1 0
L = D−1A = =
0 1/4 −1 4 −1/4 1
and we are done. The resulting factorisation is
    
2 0 2 0 1 0
= .
−1 4 0 4 −1/4 1

2.2.3 Forward and backward substitution

One of the main reasons for thinking specifically about triangular matrices is that triangular
linear systems turn out to be rather easy to solve. Consider the lower triangular equation

Lx = b,

where n × n L and n-dimensional b are given, and a solution for n-dimensional x is re-
quired. We will assume that L is invertible (no zeros on the diagonal), so that the solution
is unique. If we re-write the equation in component form, a simple solution strategy be-
comes apparent:     
l11 0 · · · 0 x1 b1
 l21 l22 · · · 0   x2   b2 
..   ..  =  ..  .
    
 .. .. . .
 . . . .  .   . 
ln1 ln2 · · · lnn xn bn
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 48

We can see that the first equation is just

l11 x1 = b1 ⇒ x1 = b1 /l11 .
But now we can use x1 to solve the second equation as
l21 x1 + l22 x2 = b2 ⇒ x2 = (b2 − l21 x1 )/l22 .
In general we can solve for xi in terms of x1 , . . . , xi−1 as
i−1
!
X .
x i = bi − lij xj lii .
j=1

Iterative construction of the solution in this way is known as forward substitution. Clearly
the algorithm would fail if any lii = 0, but this corresponds to the rank degenerate case.
For the invertible case we are considering, the algorithm provides a simple and efficient
method for computing a solution.

Example: Toy problem

Solve the triangular system    
2 0 4
x=
1 4 6
for x.

   

2 0 x1 4
= ,
1 4 x2 6
so, starting with the first equation we have
2x1 = 4 ⇒ x1 = 2.
The second equation can then be solved as
6 − x1 6 − 2
x1 + 4x2 = 6 ⇒ x2 = = = 1,
4 4
so our solution is x = (2, 1)T.

Example: forward solve using R

Use R to solve the system    
2 0 0 6
1 3 0 x =  9 
2 3 4 28
for x.
This is easily accomplished as shown in the following R session
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 49

> L=matrix(c(2,0,0,1,3,0,2,3,4),ncol=3,byrow=TRUE)
> L
[,1] [,2] [,3]
[1,] 2 0 0
[2,] 1 3 0
[3,] 2 3 4
> forwardsolve(L,c(6,9,28))
[1] 3 2 4
>
The solution is therefore given by x = (3, 2, 4)T .
The solution of upper triangular systems is analogous, but looks a little different, so
we will consider it explicitly. Suppose that U is an upper triangular matrix, and we want to
solve the system
Ux = b
for x. Again, we write out the system in component form:
    
u11 u12 · · · u1n x1 b1
 0 u22 · · · u2n   x2   b2 
..   ..  =  ..  .
    
 .. .. . .
 . . . .  .   . 
0 0 · · · unn xn bn

We now see that system is easiest to solve starting with the last equation

unn xn = bn ⇒ xn = bn /unn ,

then
un−1,n−1 xn−1 + u(n−1),n xn = bn−1 ⇒ xn−1 = (bn−1 − u(n−1),n xn )/un−1,n−1 ,
and so on, with xi given by
n
!
X .
xi = bi − uij xj uii .
j=i+1

This procedure is known as backward substitution, for obvious reasons.

Example: backward solve using R

Solve the upper triangular system
   
2 2 3 8
0 2 −1 x =  6 
0 0 3 −6

for x using R.
This is easily accomplished as the following R session shows
> R=matrix(c(2,2,3,0,2,-1,0,0,3),ncol=3,byrow=TRUE)
> R
[,1] [,2] [,3]
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 50

[1,] 2 2 3
[2,] 0 2 -1
[3,] 0 0 3
> backsolve(R,c(8,6,-6))
[1] 5 2 -2
So the solution is x = (5, 2, −2)T .

2.3 Triangular matrix decompositions

2.3.1 LU decomposition
We now know how to solve triangular systems of equations, but most linear systems
of interest do not start out in triangular form. However, as already outlined, if we can
factorise or decompose a system into triangular components, we can solve a sequence
of triangular systems in order to get a solution to our original problem. The problem we
will consider now is the following.

Proposition 13 (LU decomposition) Given a non-singular n × n matrix A, we can find

unit lower triangular L and upper triangular U such that

A = LU.

Clearly then, since |A| = |LU| = |L||U| = |U|, we have |A| = ni=1 uii , and this is an efficient
Q
way to compute the determinant in the rare cases where it is really needed.
The factorisation is just a matrix representation of the procedure of using Gaussian
elimination to zero out the lower triangle of A to create U. L represents the (inverse of
the) row operations required to achieve this. We will not give a formal derivation of the
construction, but instead give a rough outline. First we note that the row operations used
in Gaussian elimination can be represented by matrices. So the row operation which adds
λ times row i to row j can be represented by the matrix

Mij (λ) = I +λej ei T ,

and λ is known as the multiplier of the operation. The row operations needed to zero out
the lower diagonal will all have i < j, so Mij (λ) will be unit lower triangular, and therefore
the product of all of the row operations will also be unit lower triangular. Row operations
are easily inverted, as
Mij (λ)−1 = I − λej ei T .
This can be verified by multiplication, noting particularly how the final product term van-
ishes. It will require n(n − 1)/2 such row operations to zero out the lower triangle. If
we number them sequentially, so that M1 is the row operation to zero out a21 , M2 is the
row operation to zero out a31 , etc., then the matrix representing the complete set of row
operations is
M = Mn(n−1)/2 · · · M2 M1 .
Then if we denote the upper triangle that we are left with by U we have

MA = U ⇒ A = M−1 U = LU
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 51

where L = M−1 . Now

L = M1 −1 M2 −1 · · · Mn(n−1)/2 −1 ,
and the individual terms of this product are all trivial to write down. Remarkably, the unit
lower triangular product is also trivial to compute. It turns out to be simply the unit lower
triangular matrix whose off-diagonal elements are (minus) the row operation multipliers
used to zero out that entry. Note that this is due to all of the product terms cancelling,
and this in turn depends crucially on the order in which the operations are applied. The
product terms cancel in the inverse L and not in M because in the inverse, no new row
operation is introduced which uses a row that has already been modified as the “source”
for a new row operation. The factorisation procedure is best illustrated by example.

Example: toy problem

Let A be the 3 × 3 matrix  
1 4 7
A = 2 5 8  .
3 6 10
Find the LU decomposition of A.

Start with  
1 4 7
A = 2 5 8 
3 6 10
and zero out the first column. The required multipliers are -2
and -3 respectively, so we record their “inverses”, 2 and 3, in
the locations we have zeroed out:
 
1 4 7
2 −3 −6  .
3 −6 −11
Now we zero out the second column with the multiplier -2, and
so we record the number 2 in the position we zero out:
 
1 4 7
2 −3 −6 .
3 2 1
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 52

That’s it! The LU factorisation is

    
1 4 7 1 0 0 1 4 7
A = 2 5 8  = 2 1 0 0 −3 −6 .
3 6 10 3 2 1 0 0 1

2.3.2 LDMT decomposition

The LU factorisation has U upper triangular and L unit lower triangular. We have already
seen that we can factorise non-singular U as U = DR where D is diagonal and R is unit
upper triangular, and this leads directly to the LDMT decomposition of a non-singular
matrix A as
A = LDMT ,
where L and M = RT are unit lower triangular, and D is Qndiagonal. It is clear from this
T T
decomposition that |A| = |LDM | = |L||D||M | = |D| = i=1 dii . We can obviously con-
struct this decomposition from the LU factorisation of A. It turns out that there is also a
direct construction which is slightly more efficient, but this is not especially relevant in the
context of this course.

2.3.3 LDLT decomposition

In many statistical applications, the matrix A we want to factorise is symmetric (for exam-
ple, it is often a covariance or correlation matrix). In this case it turns out that the LDMT
decomposition takes the form LDLT , that is L = M. In some sense this is obvious, and
is certainly intuitive, as it has the same form as the LDMT decomposition, and is clearly
symmetric. However, it is not completely obvious that the LDMT decomposition has to be
of this form, but it has.
Proposition 14 The LDMT decomposition of a non-singular symmetric matrix A has L =
M, giving the decomposition
A = LDLT ,
where L is unit lower triangular and D is diagonal.
Proof
Start with A = LDMT , and first post-multiply by M−T and then pre-multiply by M−1 to
get
M−1 AM−T = M−1 LD.
The LHS of this equation is clearly symmetric, and the RHS is clearly lower triangular.
Therefore both sides must be diagonal. But if M−1 LD is diagonal, M−1 L must also be. But
this is also clearly unit lower triangular, so must be the identity. That is, M−1 L = I, and
hence M = L. 
The symmetry in this factorisation opens up the possibility of developing more efficient
algorithms for constructing the decomposition, rather than starting with a basic LU de-
composition. Rather than examine such strategies for this particular factorisation, we
will do so in the context of a very closely related decomposition, usually known as the
Cholesky decomposition.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 53

2.3.4 The Cholesky decomposition

If the symmetric matrix A is not only non-singular, but also positive definite, then we can
construct another factorisation that is very widely used in statistical applications. We will
first construct it from the LDLT decomposition, but then derive a more direct and efficient
algorithm for its construction. We first need a couple of simple results whose proofs are
left as exercises.

Proposition 15 If
A = LDLT
is the LDLT decomposition of symmetric positive definite A, then

1. D is positive definite, and

2. the diagonal elements of D are all strictly positive.

Now since the elements of D are all strictly positive, we can define D1/2 to be the matrix
whose diagonal elements are the square root of the corresponding elements of D, and
then we have

Proposition 16 (Cholesky decomposition) A symmetric positive definite matrix A can

be decomposed as
A = GGT ,
where G is a lower triangular matrix.

Proof
Start with the decomposition A = LDLT and put G = LD1/2 . 
The Cholesky decomposition has many important applications in multivariate statistics
and data analysis, and we will examine a few of these shortly, and more later in the course.
Note that ifQwe are interested in the determinant of A, we have |A| = |G||GT | = |G|2 ,
and |G| = ni=1 gii , so this provides an efficient way to compute the determinant of a
symmetric positive definite matrix. Obviously we can construct the decomposition starting
from a basic LU factorisation, but that is a very inefficient method. It turns out to be quite
straightforward to derive a fairly efficient algorithm for its construction from first principles.

Direct construction
We can write out the Cholesky decomposition in component form as

g11 0 · · · 0 g11 g21 · · · gn1

    
a11 a12 · · · a1n
. .
 g21 g22 . . 0   0 g22 . . gn2   a21 a22 · · · a2n 
    
 . . ..   . . ..  =  .. .. .. ..  .
 .. .. ... .   .. .. ... .   . . . . 
gn1 gn2 · · · gnn 0 0 · · · gnn an1 an2 · · · ann

Somewhat analogous to the technique of forward substitution, we can work through the
equations one by one, starting from the top left, working row by row through the non-zero
elements of G. The first few equations can be solved as follows:
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 54

2 √
g11 = a11 ⇒ g11 = a11
a21
g21g11 = a21 ⇒ g21 =
g
q11
2 2 2
g21 + g22 = a22 ⇒ g22 = a22 − g21
a31
g31g11 = a31 ⇒ g31 =
g11
a32 − g31g21
g31g21 + g32g22 = a32 ⇒ g32 = ,
g22
and so on. We see that if we work through the equations in order, we know enough at
each stage to solve explicitly for the next unknown. In general we have
j−1
X
v
u i−1
aij − gik gjk
u X k=1
2
gii = aii −
t gik , gij = , i > j.
k=1
gii

This algorithm is a bit faster than an LU decomposition as it exploits the symmetry in A. It

is also a bit more stable for the same reason. However, it only works reliably for matrices
that are strictly positive definite. It can be extended to the positive semi-definite case using
pivoting (beyond the scope of this course), but there are numerical stability issues there,
and another decomposition (such as the SVD, to be covered later) may be preferable in
that case.

Example: toy problem

Compute the Cholesky factorisation of the 3 × 3 matrix
 
1 −1 2
A = −1 5 0  .
2 0 14

For a small problem like this, it is simplest to write out the

matrix relations in full as
    
g11 0 0 g11 g21 g31 1 −1 2
g21 g22 0   0 g22 g32 = −1 5 0  .
g31 g32 g33 0 0 g33 2 0 14
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 55

We now solve the equations one at a time as

2
g11 = 1 ⇒ g11 = 1
−1 −1
g21g11 = −1 ⇒ g21 = = = −1
g11 1
2 2
q
2
√
g21 + g22 = 5 ⇒ g22 = 5 − g21 = 5 − 1 = 2
2 2
g31g11 = 2 ⇒ g31 = = =2
g11 1
−g31g21 2 × 1
g31g21 + g32g22 = 0 ⇒ g32 = = =1
g22 2
2 2 2
q
2 2
√
g31 + g32 + g33 = 14 ⇒ g33 = 14 − g31 − g32 = 14 − 4 − 1 = 3.

We therefore have A = GGT, where

 
1 0 0
G = −1 2 0 .
2 1 3
Having computed the decomposition, it is always a good idea
to check the solution as
    
1 0 0 1 −1 2 1 −1 2
−1 2 0 0 2 1 = −1 5 0  .
2 1 3 0 0 3 2 0 14
Having computed the Cholesky factorisation by hand using a relatively efficient method,
it is instructive to, just once, construct the factorisation starting from the LU decomposition.
Let’s begin by directly constructing the LU decomposition of A as
     
1 −1 2 1 −1 2 1 −1 2
−1 5 0  → −1 4 2  → −1 4 2 ,
2 0 14 2 2 10 2 1/2 9

giving the LU decomposition

    
1 −1 2 1 0 0 1 −1 2
−1 5 0  = −1 1 0 0 4 2 .
2 0 14 2 1/2 1 0 0 9
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 56

We now construct the LDMT decomposition by extracting the diagonal of U and transpos-
ing the result to get
     T
1 −1 2 1 0 0 1 0 0 1 0 0
−1 5 0  = −1 1 0 0 22 0  −1 1 0 .
2 0 14 2 1/2 1 0 0 32 2 1/2 1

At this point we note that, due to the symmetry of A, we have L = M, and so this is in
fact the LDLT decomposition. Since the elements of D are all positive, A must be strictly
positive definite, and we can construct the Cholesky factor as LD1/2 to get
    
1 0 0 1 0 0 1 0 0
−1 1 0 0 2 0 = −1 2 0 ,
2 1/2 1 0 0 3 2 1 3

as before.

Example: Galaxy data

We can use R to construct the Cholesky decomposition of the sample variance matrix for
the Galaxy data, and confirm the solution, as illustrated in the following session
> v=var(galaxy)
> v
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
radial.position 1637.78301
velocity 8886.47724
> c=chol(v)
> c
east.west north.south angle radial.position
east.west 12.0275 -2.71710 -1.7879038 21.94441872
north.south 0.0000 22.72591 0.9460287 -8.89575759
angle 0.0000 0.00000 38.1907749 -0.03564306
radial.position 0.0000 0.00000 0.0000000 10.46597448
velocity 0.0000 0.00000 0.0000000 0.00000000
velocity
east.west 37.536090
north.south -80.435144
angle 4.724212
radial.position 9.431724
velocity 29.940461
> g=t(c)
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 57

> g
east.west north.south angle radial.position
east.west 12.027505 0.0000000 0.00000000 0.000000
north.south -2.717100 22.7259122 0.00000000 0.000000
angle -1.787904 0.9460287 38.19077490 0.000000
radial.position 21.944419 -8.8957576 -0.03564306 10.465974
velocity 37.536090 -80.4351441 4.72421244 9.431724
velocity
east.west 0.00000
north.south 0.00000
angle 0.00000
radial.position 0.00000
velocity 29.94046
> g%*%c
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
radial.position 1637.78301
velocity 8886.47724
>
Note that R returns the upper triangle of the Cholesky factorisation, which can easily be
transposed to give the lower triangle if required. We next consider why it might be useful
to be able to compute the Cholesky factorisation of a variance matrix.

Transformations from and to standardised variables

Recall that one of the motivations for thinking about matrix factorisations was to be able
to find a matrix G such that
Σ = GGT . (2.1)
Clearly the Cholesky decomposition provides an efficient solution to this problem for the
case of a symmetric positive definite Σ. Before looking at applications of this decompo-
sition, it is worth thinking briefly about the existence and uniqueness of solutions to this
problem. Since Σ is symmetric, there are only n(n + 1)/2 degrees of freedom. If we al-
lowed an arbitrary choice of G, there would be n2 degrees of freedom, and so it seems
likely that there would be many different G which satisfy the (2.1). This is indeed the case.
By restricting ourselves to looking for a lower triangular matrix, we have ensured that
we have only n(n + 1)/2 degrees of freedom, and this leads to a unique solution (in the
positive definite case). Clearly another obvious strategy would be to look for a symmet-
ric matrix satisfying (2.1). This is also possible, and also leads to an essentially unique
solution. This matrix is known as the symmetric square root, as it is a symmetric matrix
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 58

satisfying the equation

Σ = G2 ,
in addition to (2.1). Note that many mathematicians would not regard the Cholesky factor
as a “true” square root, since it does not have this property. However, since (2.1) is the
most commonly encountered requirement for a “square root” in statistical applications,
most statisticians use (2.1) as the definition of a square root matrix. This could potentially
lead to confusion, so care must be taken when reading literature relating to matrix square
roots.
We now consider the problem stated in Chapter 1 of how to transform standard ran-
dom quantities to give a vector of random quantities with a pre-specified variance matrix.
Suppose that we wish to simulate random quantities with p-dimensional mean vector µ
and p × p variance matrix Σ. Let Z be a vector of independent random quantities each
with mean zero and variance one. Then the random quantity

X = µ + GZ,

where G is a matrix satisfying Σ = GGT , has the required mean and variance.
We can also invert this affine transformation, to get

Z = G−1 (X − µ),

a method for transforming quantities with a given mean and variance to standard uncor-
related form. This kind of standardisation transformation also has many applications, and
we shall look at the symmetric version of this kind of transformation later. Note also that
the above equation does not require us to compute the inverse of G. We are merely using
the equation as shorthand for solving the linear system

GZ = X − µ

for Z using forward substitution.

Example: standardising the Galaxy data

The following R session shows how to standardise the Galaxy data
> W=sweep(galaxy,2,colMeans(galaxy))
> g=chol(var(galaxy))
> Z=as.matrix(W)%*%solve(g)
> var(Z)
east.west north.south angle
east.west 1.000000e+00 9.870280e-18 4.921924e-18
north.south 9.870280e-18 1.000000e+00 -7.548842e-18
angle 4.921924e-18 -7.548842e-18 1.000000e+00
radial.position -6.091896e-16 -3.758545e-17 -3.116896e-17
velocity 2.160761e-17 -1.472481e-16 -1.313282e-17
radial.position velocity
east.west -6.091896e-16 2.160761e-17
north.south -3.758545e-17 -1.472481e-16
angle -3.116896e-17 -1.313282e-17
radial.position 1.000000e+00 1.231192e-16
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 59

velocity 1.231192e-16 1.000000e+00

> colMeans(Z)
east.west north.south angle radial.position
-1.353205e-17 -1.512512e-17 -6.784303e-17 -7.906418e-17
velocity
-3.376896e-15
>
Note that here, to keep things simple, I did actually compute the inverse of the Cholesky
factor (using solve()), and then used this to transform the data matrix. As there are
many more rows than columns in the data matrix, this isn’t actually terrible way to do the
transformation, in this particular case, but we will see a much better method in the next
section. We will examine later why it might be useful to standardise data in this way.

2.4 Other matrix factorisations

2.4.1 QR factorisation
Not all useful matrix factorisations use triangular matrices. Orthogonal matrices also turn
out to be very useful (they are even easier to invert than triangular matrices). We will
start by looking briefly at a factorisation that is part orthogonal and part triangular. The
QR factorisation can be applied to an n × p matrix A, where (typically) n ≥ p. There are
several variants of this factorisation, but the variant most commonly used in practice, and
the version implemented in R, is the so-called “thin” QR factorisation where A is factorised
as
A = QR,
where Q is an n×p matrix with orthogonal columns (QT Q = I, but typically QQT 6= I), and R
is upper triangular. We will not dwell extensively on the construction of this decomposition,
but it is essentially a matrix representation of the Gram-Schmidt orthonormalisation of the
columns of A. Column operations are represented by post-multiplication by a matrix. The
Gram-Schmidt operations are represented by upper triangular matrices, so the combined
set of operations is represented by an upper triangular matrix, C. We then have

AC = Q

but then
A = QR,
where R = C−1 . There are various variations on the precise way that this decomposi-
tion is computed, but these will not concern us. This decomposition has many potential
applications in statistical problems.
We can compute and manipulate QR factorisations using R.
> A=matrix(c(1,2,3,4,5,6),ncol=2)
> A
[,1] [,2]
[1,] 1 4
[2,] 2 5
[3,] 3 6
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 60

> myQR=qr(A)
> qr.Q(myQR)
[,1] [,2]
[1,] -0.2672612 0.8728716
[2,] -0.5345225 0.2182179
[3,] -0.8017837 -0.4364358
> qr.R(myQR)
[,1] [,2]
[1,] -3.741657 -8.552360
[2,] 0.000000 1.963961
> qr.Q(myQR)%*%qr.R(myQR)
[,1] [,2]
[1,] 1 4
[2,] 2 5
[3,] 3 6
> t(qr.Q(myQR))%*%qr.Q(myQR)
[,1] [,2]
[1,] 1.000000e+00 -5.144539e-17
[2,] -5.144539e-17 1.000000e+00
In the case where A is square (n = p), we have
n
Y
|A| = |QR| = |Q||R| = |R| = rii ,
i=1

and this represents an efficient way to compute the determinant.

Uniqueness and relation to Cholesky factorisation

It is clear that we can construct a “thin” QR factorisation via Gram-Schmidt orthonormal-
isation of the columns of A, but it is natural to wonder whether such a factorisation is
unique? In the full rank case it is, and understanding why sheds light on its relationship
with the Cholesky factorisation. Suppose that
A = QR
is the QR factorisation of A, and consider
AT A = (QR)T QR
= RT QT QR
= RT R,
where R is upper triangular. It is therefore clear that R = GT , where G is the Cholesky
factor for AT A. We know that the Cholesky factor is unique (in the positive definite case),
and therefore so is R, and we can construct Q via
Q = AR−1
and is also unique. Note that this construction is almost identical to the way we stan-
dardised a data matrix in the previous Galaxy data example. Note that this construction
gives us insight into the nature of the QR factorisation, but does not represent an effi-
cient method for its construction. Variations on Gram-Schmidt are much faster and more
numerically stable.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 61

Using the QR decomposition for standardisation

The QR factorisation of a centered data matrix therefore gives us a much more effi-
cient way of standardising a data matrix, and also a very efficient way of computing the
Cholesky factor of the sample variance matrix. Let’s just spell out the details to be abso-
lutely clear. Starting from an n × p data matrix X, we form the centered data matrix

W = Hn X,

where, again, we are just using Hn X as shorthand for X − 1 n x̄T — we are not suggesting
actually constructing the centering matrix Hn .† Next construct the QR decomposition of W
as
W = QR.
Putting
1
G= √ RT
n−1
gives lower triangular G as the Cholesky factor of the sample variance matrix for X. We
can verify this as follows:
1
S= WT W
n−1
1
= (QR)T QR
n−1
1
= RT R
n−1
= GGT .

Similarly, if we set √
Z= n − 1Q,
then Z has mean zero and variance the identity, since
1
SZ = ZT Z
n−1
= QT Q
= Ip .

Example: Galaxy data

We can use the QR factorisation to standardise the Galaxy data analysed previously as
illustrated in the following R session:
> myQR=qr(W)
> qr.R(myQR)/sqrt(322)
east.west north.south angle radial.position velocity
3 -12.0275 2.71710 1.7879038 -21.94441872 -37.536090
4 0.0000 22.72591 0.9460287 -8.89575759 -80.435144
†
This is analogous to using A−1 b as shorthand for solving the linear system Ax = b, and not actually suggesting
constructing the inverse of A.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 62

5 0.0000 0.00000 -38.1907749 0.03564306 -4.724212

6 0.0000 0.00000 0.0000000 -10.46597448 -9.431724
7 0.0000 0.00000 0.0000000 0.00000000 -29.940461
> g
east.west north.south angle radial.position
east.west 12.0275 -2.71710 -1.7879038 21.94441872
north.south 0.0000 22.72591 0.9460287 -8.89575759
angle 0.0000 0.00000 38.1907749 -0.03564306
radial.position 0.0000 0.00000 0.0000000 10.46597448
velocity 0.0000 0.00000 0.0000000 0.00000000
velocity
east.west 37.536090
north.south -80.435144
angle 4.724212
radial.position 9.431724
velocity 29.940461
> head(qr.Q(myQR)*sqrt(322))
[,1] [,2] [,3] [,4] [,5]
[1,] -0.7312537 -1.659224 -0.6411756 -0.8750822 -0.10633537
[2,] -0.6898643 -1.565365 -0.6369129 -0.8218676 0.24018028
[3,] -0.6484749 -1.471507 -0.6326503 -0.7686530 -0.08129645
[4,] -0.6070855 -1.377648 -0.6283877 -0.7154384 -0.70337004
[5,] -0.5656961 -1.283789 -0.6241250 -0.6622238 -0.75765010
[6,] -0.5243067 -1.189931 -0.6198624 -0.6090092 -0.91212873
> head(Z)
east.west north.south angle radial.position velocity
3 0.7312537 -1.659224 0.6411756 0.8750822 0.10633537
4 0.6898643 -1.565365 0.6369129 0.8218676 -0.24018028
5 0.6484749 -1.471507 0.6326503 0.7686530 0.08129645
6 0.6070855 -1.377648 0.6283877 0.7154384 0.70337004
7 0.5656961 -1.283789 0.6241250 0.6622238 0.75765010
8 0.5243067 -1.189931 0.6198624 0.6090092 0.91212873
>
From this we see that the results of using the QR factorisation are identical to the results
obtained earlier, except for arbitrary sign issues, which are usually unimportant, and can
be easily corrected if necessary.

2.4.2 Least squares problems

As the previous section suggests, QR factorisations have many potential applications in
statistics, but so called “least squares” problems are the classic example. First note how
we can use the QR factorisation to solve a linear system. Suppose that we want to solve

Ax = b

for x. If A is factorised as A = QR, we have

QRx = b
⇒ Rx = QT b,
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 63

since QT Q = I, and then we can solve for x using backward substitution.

Now consider the classical multiple linear regression problem

y = Xβ + ε.

Given data y and X we want to find the β which minimises kεk22 = εT ε. Differentiating with
respect to the vector β and equating to zero leads to the so-called “normal equations”

XT Xβ = XT y.

The classical method for solving these equations was via a Cholesky decomposition for
XT X and then backward and forward substitution. However, it turns out to be more efficient
and numerically stable to use the QR factorisation of X, as

(QR)T QRβ = (QR)T y

⇒ RT QT QRβ = RT QT y
⇒ RT Rβ = RT QT y
⇒ β = (RT R)−1 RT QT y
= R−1 R−T RT QT y
⇒ β = R−1 QT y.

That is, β is the solution of

Rβ = QT y,
obtained by backward substitution. This is the method of solution used by R’s lm() func-
tion.

2.4.3 Spectral decomposition

Construction for real symmetric matrices
Recall that v is an eigenvector of an n × n square matrix A, and λ is its corresponding
eigenvalue if
Av = λv
Clearly (A − λ I)v = 0, so v is in the null-space of A − λ I, which means that this must be
singular, and so eigenvalues can be identified as roots of the characteristic equation

|A − λ I | = 0.

Since this is a polynomial in λ of degree n, there are n (not necessarily distinct) eigenval-
ues of A. In general, some of these eigenvalues will be complex, but not if A is symmetric.

Proposition 17 The eigenvalues of a real symmetric matrix A are real.

Proof
Suppose that λ is a complex eigenvalue of A. In this case the corresponding eigen-
vector v is also complex. By definition we have Av = λv, and taking complex conjugates
we have
Av̄ = λ̄v̄
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 64

so λ̄ is also an eigenvalue with corresponding eigenvector v̄.‡ But then

v̄ T Av = λ̄v † v = λv † v
⇒ λ̄kvk2 = λkvk2
⇒ λ̄ = λ,
and so λ ∈ R. 
Since we can choose the magnitude of an eigenvector arbitrarily, we will assume that we
always choose eigenvectors to have norm 1. That is v T v = 1.
Proposition 18 The eigenvectors corresponding to distinct eigenvalues of a real sym-
metric matrix A are orthogonal.
Proof

Suppose that (λ1, v 1) and (λ2, v 2) are both eigen-pairs of A.

Then
v 1TAv 2 = λ2v 1Tv 1 = λ1v 1Tv 2
⇒ (λ2 − λ2)v 1 · v 2 = 0,
so either λ1 = λ2 or v 1 · v 2 = 0.

In fact, it turns out that repeated eigenvalues do not complicate the picture, since the
eigen-space corresponding to repeated eigenvalues has the same dimension as the mul-
tiplicity of the root, and hence orthogonal eigenvectors may also be constructed for the
repeated eigenvalues. Hence we can order the eigenvalues in decreasing order as
Av i = λi v i , i = 1, 2, . . . , n
and by stacking the eigenvectors as columns of an n × n orthogonal matrix V we can
re-write the above equations as
AV = VD
where D = diag {λ1 , λ2 , . . . , λn }. This then leads to one of the most important results in
linear algebra.
Proposition 19 (Spectral decomposition) If A is a real symmetric n × n matrix, it can
be factorised in the form
A = VDVT
where V is an orthogonal matrix whose columns are the eigenvectors of A and D =
diag {λ1 , λ2 , . . . , λn }, where the λi are the ordered eigenvalues of A corresponding to the
eigenvectors in the columns of V.
We will often be interested in symmetric matrices which are positive semi-definite. These
matrices clearly have non-negative eigenvalues, and similarly, strictly positive definite ma-
trices have strictly positive eigenvalues.
‡
Note that here we are using a bar (¯) to denote complex conjugation, rather than a sample mean. This proof is the
only place in the whole course where complex numbers occur, and hence the only place where a bar does not represent
a sample mean. Here we also use a dagger for conjugate transpose. That is, v † = v̄ T .
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 65

Symmetric square root

The spectral decomposition can be used to create a symmetric square root matrix, since
if
A = VDVT
then
G = VD1/2 VT
satisfies GT = G and GG = A, and hence also GGT = A. This symmetric square root
is a “true” square root matrix, unlike the Cholesky triangle. It is also well defined for
singular positive semi-definite matrices, where use of the Cholesky decomposition can be
problematic. In the positive definite case, we can also directly construct its inverse as

G−1 = VD−1/2 VT ,

which can be very useful in several contexts. The symmetry of this square root matrix is
also desirable in some applications. However, the Cholesky factorisation is much faster
than an eigen-solver, so in the many applications where the cheaper Cholesky factor will
be adequate, it is usually to be preferred.

Example: toy problem

Construct the spectral decomposition of the matrix
 
9 7
A= ,
7 9

and use it to construct a symmetric square root, G, of A. Verify that G2 = A.

We first find roots of the characteristic equation via

 
 9−λ 7 
|A − λ I | =  
7 9−λ 
= (9 − λ)2 − 49
= (λ − 2)(λ − 16).
So the ordered eigenvalues are λ1 = 16, λ2 = 2. Now
 
−7 7
A − λ1 I = ,
7 −7
and so  
1 1
v1 = √ .
2 1
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 66

Similarly, 
7 7
A − λ2 I = ,
7 7
so  
1 1
v2 = √ .
2 −1
Consequently, our spectral decomposition is A = VDVT, where
   
1 1 1 16 0
V=√ , D= .
2 1 −1 0 2
Now we have the spectral decomposition, it is straightforward
to construct the symmetric square root as
G = VD1/2VT
   
1 1 1 4 √0 1 1
=
2 1 −1 0 2 1 −1
  
1 1 1 √4 √4
=
2 1 −1 2 − 2
 √ √ 
1 4+ 2 4− 2
= √ √ .
2 4− 2 4+ 2
Our proposed square root is clearly symmetric, and we can
verify that it is indeed a square root with
 √ √  √ √   
1 4 + √ 2 4 − √ 2 4 + √ 2 4 − √ 2 1 36 28
G2 = = = A.
4 4− 2 4+ 2 4− 2 4+ 2 4 28 36

2.4.4 Mahalanobis transformation and distance

Mahalanobis transformation
The Mahalanobis transformation is just a standardisation transformation using the (in-
verse) symmetric square root of the variance matrix. So if X is a random vector with
mean µ and variance matrix Σ, the Mahalanobis transformation is

Z = Σ−1/2 (X − µ),
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 67

where Σ−1/2 is the inverse of the symmetric square root of Σ. Z will have mean 0 and
variance I.
Similarly, if X is an n × p data matrix, the Mahalanobis transform is

Z = Hn XS−1/2 ,

though again, it must be emphasised that this is a mathematical description of the trans-
formation, and not a recipe for how to construct it numerically. Z will have sample mean
zero and sample variance Ip .

Example: Galaxy data

We will look now at how to compute the spectral decomposition of the galaxy data set
using R, and how to use it to construct a symmetric square root, and its inverse.
> v=var(galaxy)
> v
east.west north.south angle radial.position
east.west 144.66088 -32.67993 -21.50402 263.93661
north.south -32.67993 523.84971 26.35728 -261.78938
angle -21.50402 26.35728 1462.62686 -49.01139
radial.position 263.93661 -261.78938 -49.01139 670.22991
velocity 451.46551 -1929.95131 37.21646 1637.78301
velocity
east.west 451.46551
north.south -1929.95131
angle 37.21646
radial.position 1637.78301
velocity 8886.47724
> e=eigen(v,symmetric=TRUE)
> e
$values
[1] 9642.07343 1466.48964 487.88910 68.87222 22.52020

$vectors
[,1] [,2] [,3] [,4] [,5]
[1,] 0.051396600 0.02254908 0.45279536 0.1122205 0.88274167
[2,] -0.208493672 -0.01789604 0.31495251 0.8879502 -0.26183862
[3,] 0.002463535 -0.99826125 0.04727651 -0.0346832 0.00551559
[4,] 0.182722226 0.05000384 0.82508240 -0.3634307 -0.38893366
[5,] 0.959424462 -0.01205694 -0.11307185 0.2562542 -0.03013096

> e$vectors%*%t(e$vectors)
[,1] [,2] [,3] [,4]
[1,] 1.000000e+00 -7.224852e-17 -4.978182e-17 1.115915e-16
[2,] -7.224852e-17 1.000000e+00 1.760090e-16 -8.700722e-18
[3,] -4.978182e-17 1.760090e-16 1.000000e+00 2.170533e-16
[4,] 1.115915e-16 -8.700722e-18 2.170533e-16 1.000000e+00
[5,] 2.232271e-17 -1.676939e-16 -1.400287e-16 8.567399e-17
[,5]
[1,] 2.232271e-17
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 68

[2,] -1.676939e-16
[3,] -1.400287e-16
[4,] 8.567399e-17
[5,] 1.000000e+00
> e$vectors%*%diag(e$values)%*%t(e$vectors)
[,1] [,2] [,3] [,4] [,5]
[1,] 144.66088 -32.67993 -21.50402 263.93661 451.46551
[2,] -32.67993 523.84971 26.35728 -261.78938 -1929.95131
[3,] -21.50402 26.35728 1462.62686 -49.01139 37.21646
[4,] 263.93661 -261.78938 -49.01139 670.22991 1637.78301
[5,] 451.46551 -1929.95131 37.21646 1637.78301 8886.47724
> G=e$vectors%*%diag(sqrt(e$values))%*%t(e$vectors)
> G
[,1] [,2] [,3] [,4] [,5]
[1,] 8.6098647 1.8123819 -0.3859399 7.2496158 3.8132008
[2,] 1.8123819 13.3404531 0.7001528 -0.2300872 -18.4947058
[3,] -0.3859399 0.7001528 38.2218104 -0.9113324 0.5003811
[4,] 7.2496158 -0.2300872 -0.9113324 20.2249833 14.4131733
[5,] 3.8132008 -18.4947058 0.5003811 14.4131733 91.2244082
> G%*%G
[,1] [,2] [,3] [,4] [,5]
[1,] 144.66088 -32.67993 -21.50402 263.93661 451.46551
[2,] -32.67993 523.84971 26.35728 -261.78938 -1929.95131
[3,] -21.50402 26.35728 1462.62686 -49.01139 37.21646
[4,] 263.93661 -261.78938 -49.01139 670.22991 1637.78301
[5,] 451.46551 -1929.95131 37.21646 1637.78301 8886.47724
> Ginv=e$vectors%*%diag(1/sqrt(e$values))%*%t(e$vectors)
> G%*%Ginv
[,1] [,2] [,3] [,4]
[1,] 1.000000e+00 -2.478825e-16 -2.061869e-17 1.212003e-16
[2,] -2.428952e-16 1.000000e+00 3.009745e-16 1.170261e-16
[3,] -1.111113e-15 8.248890e-16 1.000000e+00 1.020284e-15
[4,] 6.617055e-16 -3.846478e-16 2.830987e-16 1.000000e+00
[5,] 8.075138e-16 -1.285430e-15 -4.671759e-16 -4.477755e-17
[,5]
[1,] 1.171616e-17
[2,] -2.679606e-16
[3,] -1.481901e-16
[4,] 8.386304e-17
[5,] 1.000000e+00
Now we have Ginv we can use it to construct the Mahalanobis transform of the data
> W=sweep(galaxy,2,colMeans(galaxy))
> Z=as.matrix(W)%*%Ginv
> colMeans(Z)
[1] 4.419302e-16 -2.600506e-15 2.784151e-17 1.075502e-15
[5] -1.821798e-15
> var(Z)
[,1] [,2] [,3] [,4]
[1,] 1.000000e+00 6.811770e-16 -8.034862e-16 1.586732e-15
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 69

[2,] 6.811770e-16 1.000000e+00 1.309490e-15 5.266862e-15

[3,] -8.034862e-16 1.309490e-15 1.000000e+00 4.550018e-16
[4,] 1.586732e-15 5.266862e-15 4.550018e-16 1.000000e+00
[5,] -3.265377e-16 -2.529045e-15 -2.198962e-16 1.568981e-15
[,5]
[1,] -3.265377e-16
[2,] -2.529045e-15
[3,] -2.198962e-16
[4,] 1.568981e-15
[5,] 1.000000e+00

Mahalanobis distance
Given an observation x of a random vector X with E(X) = µ and Var(X) = Σ, it is
natural to wonder how “unusual” it is, in some appropriate sense. We cannot give a very
formal solution to this problem until we develop some multivariate distribution theory, but
even now we can give an intuitive explanation of the most commonly used summary. In
some sense we want to know how “far” the observation is from the mean, but in general
some components are more variable than others, so it doesn’t make sense to look at
“distance” on the scale of the original variables. However, it does make perfect sense for
a standardised observation, z. So the Mahalanobis distance of x from µ is just
√
kzk = z T z
q
= [Σ−1/2 (x − µ)] T Σ−1/2 (x − µ)
p
= (x − µ)T Σ−1 (x − µ).
√
Computationally, the efficient way to compute this distance measure is as z T z where

z = Σ−1/2 (x − µ).

Note that it is not necessary to use a symmetric square root for this computation. Stan-
dardisation using the Cholesky factor leads to exactly the same distance, and is much
cheaper to compute and then solve than using a symmetric square root. Confirmation of
this fact is left as an exercise.
It is also worth noting that if the variance is defined empirically, as the sample variance
matrix S of a data matrix X, then the Mahalanobis distance can be calculated yet more
stably and efficiently using the QR factorisation of X.
Computing the Mahalanobis distance for each observation in a data set is the most
commonly used technique for identifying outlying observations in multivariate data.

Example: Galaxy data

Continuing our previous example, we have already constructed a standardised data matrix
Z using a Mahalanobis transformation. We can use it to construct a Mahalanobis distance
for each observation as follows.
d=sqrt(apply(Z*Z,1,sum))
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 70

●● ●
●●●●
●
●●● ● ● ●
●●●●● ●
● ●●● ●
●●●● ●
●
●
● ●
● ● ●●●
● ●●● ●
1700 ● ●●
●●●
●
●●●● ● ● ● ●●●
● ●
● ●
galaxy$velocity

● ● ● ●
●●●
●●
● ●● ●● ● ●
● ●● ● ●
● ● ●● ●
●●● ●● ●
● ● ● ●● ●●
● ●●
1600

● ● ● ● ● ●
●● ●● ● ●
● ●
● ● ●● ● ●●● ●●● ●
●●●●●● ● ●●
●
● ● ● ●
● ● ●●
● ● ●
●
● ●●
● ● ●● ● ● ● ●● ●
●●●●
●● ● ● ●● ● ● ● ● ● ● ●
●●● ●●●● ●
● ●●●●
● ● ●●
● ● ● ●●● ● ● ●●
●●●●●
●● ● ●●●● ● ● ● ●
● ● ● ● ●●
1500

● ●●● ● ● ● ●
●● ●
●●●●
● ●
●●●
● ●
● ●●●●● ●
●●●● ●
●● ● ● ●●
● ●● ● ● ● ●●●● ●
● ● ●
●● ● ●● ●● ● ● ●● ●
●● ● ●
●● ●●●●
1400

●●

−40 −20 0 20 40 60

galaxy$radial.position

Figure 2.1: Scatterplot for the Galaxy data showing outliers identified by large Mahalanobis dis-
tance

Note that regular element-wise multiplication was used to square the elements in Z, rather
than matrix multiplication, which is not even valid here due to Z being non-square. A
histogram of the distances (not shown) can be obtained with hist(d,20), and this shows
a fairly uninteresting distribution, suggesting no particularly outlying values. There are a
few distances larger than 4, so it may be worth looking at those in more detail. We can
inspect the cases individually using
> galaxy[d>4,]
east.west north.south angle radial.position velocity
286 24.85322 49.84784 63.5 55.7 1531
287 23.82696 47.78950 63.5 53.4 1533
288 22.80071 45.73114 63.5 51.1 1539
330 -21.46211 -43.04634 63.5 -48.1 1642
331 -22.48837 -45.10469 63.5 -50.4 1616
and we can highlight them on a scatterplot using
> plot(galaxy$radial.position,galaxy$velocity,pch=19,col=4)
> points(galaxy$radial.position[d>4],galaxy$velocity[d>4],pch=19,col=2)
resulting in the plot show in Figure 2.1. Although it is difficult to tell much from a single 2d
projection, it does appear as though the 5 values highlighted are indeed some way away
from the centre of the cloud of data points.
Note that once again, R has a built-in function for computing the (squared) Maha-
lanobis distance, so our vector of distances could have been constructed much more
straightforwardly using
d=sqrt(mahalanobis(galaxy,colMeans(galaxy),var(galaxy)))
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 71

2.4.5 The singular value decomposition (SVD)

The final matrix factorisation we will examine here, the singular value decomposition, can
be applied to arbitrary n × p matrices (we typically consider the case n ≥ p), and is
closely related to both the QR factorisation and the spectral decomposition. As for the QR
factorisation, there are a few variants, and we concentrate here on a variant known as the
“thin” SVD, as it is the most commonly used in practice.
Proposition 20 The n × p matrix A can be factorised as

A = UDVT ,

where U is an n × p matrix with orthonormal columns (UT U = Ip ), D is a p × p diagonal

matrix with non-negative entries, and V is a p × p orthogonal matrix. The diagonal entries
of D are known as the singular values of A. The columns of U are known as the left
singular vectors, and the columns of V are known as the right singular vectors.
Proof
Consider the spectral decomposition of AT A, which is clearly p × p, real, symmetric,
and positive semi-definite. We can therefore write the spectral decomposition in the form

AT A = VD2 VT ,

as the eigenvalues are non-negative. By construction, V is an orthogonal matrix. We now

put
U = AVD−1 ,
to get the SVD A = UDVT , and we verify that the columns of U are orthogonal as

UT U = (AVD−1 )T (AVD−1 )
= D−1 VT AT AVD−1
= D−1 VT VD2 VT VD−1
= D−1 D2 D−1
= Ip .


It must be emphasised that this is a mathematical construction to demonstrate the exis-
tence of the factorisation, and its relationship to the spectral decomposition, but this is not
how the factorisation is constructed in practice (and in any case, there are issues with this
construction when D is singular). There are very efficient methods which can be used for
this, often starting from a QR factorisation of A. Typically AT A is not constructed at all, as
this is a numerically unstable computation.

The SVD and spectral decomposition

From our construction it is clear that the columns of V are the eigenvectors of AT A, so the
left singular vectors are eigenvectors of AT A. Similarly, the n × n matrix AAT = UD2 UT .
Now U is n × p, but post-multiplying by U gives

(AAT )U = UD2 ,
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 72

and so the columns of U are eigenvectors of AAT corresponding to the p eigenvalues in

D2 , which includes all of the non-zero eigenvalues of AAT . So the right singular vectors
are (the important) eigenvectors of AAT .
It should be clear that if the SVD is applied directly to a real symmetric positive semi-
definite matrix, that the spectral decomposition is obtained. Numerically, however, the
algorithms used to calculate the SVD are quite different to those typically used for eigen-
solvers. The SVD will typically be slightly slower than a symmetric eigen-solver but faster
than an asymmetric eigen-solver. The SVD may be more numerically stable if the input
matrix is close to singular, and is therefore very commonly used in practice.

SVD for least squares

The SVD provides an alternative approach to solving the normal equations
XT Xβ = XT y
for β.

Let X = UDVT be the SVD of X, and then

XTXβ = XTy
⇒ (UDVT)T(UDVT)β = (UDVT)Ty
⇒ VDUTUDVTβ = VDUTy
⇒ VD2VTβ = VDUTy
⇒ DVTβ = UTy
⇒ β = VD−1UTy.
This turns out to be a numerically stable way of solving (full rank) least squares prob-
lems, but since the SVD is slower than the QR factorisation, the previously described QR
method is usually to be preferred.

Symmetrising a Cholesky factor

Note that a Cholesky factor may be converted into the symmetric square root via the SVD.
Suppose that the Cholesky decomposition of A is
A = LLT .
Then suppose that we form the SVD of L as
L = UDVT ,
so that
A = (UDVT )(UDVT )T = UD2 UT .
If we put
G = UDUT ,
it is clear that G is symmetric, and that G2 = A, and hence is the symmetric square root.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 73

Example: Galaxy data

We can form the SVD of the variance matrix for the galaxy data set using
> svd(var(galaxy))
$d
[1] 9642.07343 1466.48964 487.88910 68.87222 22.52020

$u
[,1] [,2] [,3] [,4] [,5]
[1,] -0.051396600 -0.02254908 -0.45279536 -0.1122205 0.88274167
[2,] 0.208493672 0.01789604 -0.31495251 -0.8879502 -0.26183862
[3,] -0.002463535 0.99826125 -0.04727651 0.0346832 0.00551559
[4,] -0.182722226 -0.05000384 -0.82508240 0.3634307 -0.38893366
[5,] -0.959424462 0.01205694 0.11307185 -0.2562542 -0.03013096

$v
[,1] [,2] [,3] [,4] [,5]
[1,] -0.051396600 -0.02254908 -0.45279536 -0.1122205 0.88274167
[2,] 0.208493672 0.01789604 -0.31495251 -0.8879502 -0.26183862
[3,] -0.002463535 0.99826125 -0.04727651 0.0346832 0.00551559
[4,] -0.182722226 -0.05000384 -0.82508240 0.3634307 -0.38893366
[5,] -0.959424462 0.01205694 0.11307185 -0.2562542 -0.03013096
We see that it is exactly the same as the spectral decomposition obtained earlier, modulo
some sign changes on some of the columns, which are arbitrary.

2.5 Principal components analysis (PCA)

2.5.1 Derivation from the spectral decomposition
Principal components analysis, or PCA as it is more commonly known, is a technique
for finding interesting low dimensional projections of a multivariate random variable or,
more commonly, data matrix. It is very closely related to the spectral decomposition
and SVD. We will begin by understanding principal components in relation to the eigen-
decomposition of a variance matrix. To keep things as simple as possible, we begin by
considering a p-dimensional random vector X with E(X) = µ and Var(X) = Σ. Write
the spectral decomposition of Σ as

Σ = VDVT .

Now consider a linear combination αT X, or one-dimensional projection of X, a scalar

random quantity for given fixed p-dimensional α. We know that the variance of this linear
combination is αT Σα. Let us now consider the linear combination represented by the
eigenvectors in the columns of V. The ith eigenvector, v i , corresponding to the eigenvalue
λi , has variance
v i T Σv i = λi v i T v i = λi .
So the eigenvalues are the variances of the linear combinations represented by the cor-
responding eigenvectors. Similarly, if we consider the correlation between the random
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 74

quantities represented by two distinct eigenvectors, vi , v j , we have

Cov v i T X, v j T X = v i T Cov(X, X) v j = v i T Σv j = λj v i T v j = 0.

So the eigenvectors form a basis for the space of linear combinations of X, and repre-
sent a collection of uncorrelated random quantities whose variances are given by their
respective eigenvalues.
Formally we characterise principal components as linear combinations of variables
that explain the greatest variation in the random quantity or data set. However, we need
to place a constraint on the linear combination for this to be well defined. For example,
if αT X is a linear combination with variance σ 2 , then it is clear that kαT X has variance
k 2 σ 2 , so we can make the variance of a linear combination as large as we wish simply by
multiplying it by a large enough scalar. However, we are really interested in the direction
with largest variation, and so we impose the constraint kαk = 1, or αT α = 1. To maximise
subject to a non-linear constraint we use a Lagrange multiplier, and begin by optimising

f = αT Σα − λ(αT α − 1)
= αT (Σ − λ I)α + λ.

Now, recalling that for a symmetric matrix A, we have

∂ T
∇xT Ax = x Ax = 2Ax,
∂x
we can differentiate and equate to zero to get
∂f
= 2(Σ − λ I)α = 0.
∂x
In other words,
|Σ − λ I | = 0,
and λ is an eigenvalue of Σ, and α is the (normalised) eigenvector corresponding to λ.
Since we typically construct the spectral decomposition with eigenvalues in decreasing
order, we pick the solution with largest variance as

λ = λ1 , α = v1.

So the first eigenvector of the spectral decomposition represents the linear combination
of variables with the largest variance, and its variance is given by its corresponding eigen-
value. This is the first principal component.
The second principal component is defined to be the linear combination uncorrelated
with the first principal component which has the greatest variance. But since we have
seen that the linear combinations corresponding to the eigenvectors are uncorrelated, it
follows that the set of linear combinations corresponding to quantities uncorrelated with
the first principal component is spanned by the 2nd to pth eigenvectors, and hence the
maximum variance will be obtained at λ2 , corresponding to the 2nd eigenvector, v 2 . The
third principal component is the linear combination with greatest variance uncorrelated
with the first two principal components, and so on. So we see that there is a direct corre-
spondence between the spectral decomposition of the variance matrix and the principal
components of the random vector.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 75

We can associate each principal component with a random quantity, Yi = v i T X, i =

1, 2, . . . , p, and use these to construct a new random vector
Y = VT X,
the vector of principal components. The variance matrix of Y is clearly given by
Var(Y ) = Var VT X

= VT Var(X) V
= VT VDVT V
= D.
So the components of Y are uncorrelated, and each has the correct variance, as required.

2.5.2 Total variation and variance explained

We know that the variance matrix is a very rich summary of the variability of a random vec-
tor, X. However, sometimes it is useful to have a single scalar summary which captures
some aspect of the overall variability of the random vector. There are obviously many
ways that this could be done, and none will be perfect, but a commonly used summary of
variation is the sum of the variances of the individual variables,
p
X
t(X) = Var(Xi ) = Tr (Var(X)) ,
i=1

often known as the total variation of the random vector. The total variation has the de-
sirable property that it is preserved by orthogonal rotation. In particular, if we consider
transforming X to its principal components via
Y = VT X
we have
t(X) = Tr (Var(X))
= Tr VDVT

= Tr DVT V

= Tr (D)
= t(Y ),
using the fact that Tr (AB) = Tr (BA) for square matrices.
One motivation for undertaking principal components analysis is to try and find low-
dimensional projections of the original variables which contain most of the information in
the original higher-dimensional vectors. One way of characterising the information is by
its variation. Since the principal components are ordered with most variable first, it makes
sense to consider the first q principal components (q ≤ p). The total variation of the first q
principal components is clearly qi=1 λi . The proportion of variance explained is therefore
P
q
X
λi q
i=1 1 X
p = λi ,
X Tr (D) i=1
λi
i=1
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 76

with the proportion increasing to 1 by q = p. Therefore an obvious strategy for dimension

reduction is to work with the first q principal components, where q is the smallest number
such that the proportion of variance explained exceeds a given pre-specified threshold.
The actual threshold will be problem dependent — there are no hard-and-fast rules, as
principal components are primarily an exploratory tool.

2.5.3 Principal components from a sample variance matrix

To keep the notation simple we have introduced the idea of principal components for a
random vector. However, in practice principal components analysis is mainly used in the
analysis of multivariate data. So now suppose that we have an n × p data matrix, X, and
wish to use principal components analysis to find interesting low-dimensional projections.
Let us suppose that we have the sample variance matrix
1
S= XT Hn X
n−1
and form its spectral decomposition as

S = VDVT .

Then all of the properties we described for random variables easily follow. For example,
the eigenvectors in the columns of V correspond to linear combinations of the variables
whose sample variance is the corresponding eigenvalue, and distinct eigenvectors corre-
spond to linear combinations of variables whose sample covariance is zero. So we can
define the first principal component of X to be the linear combination of variables whose
sample variance is largest (subject to the unit norm constraint), and it trivially follows that
this corresponds to the first eigenvector, v 1 , and its sample variance is λ1 . Similarly the
second principal component corresponds to v 2 , with sample variance λ2 , and so on.
We can obviously transform observations to their principal component representation
using
y = VT x
and so we can transform X to Y, the data rotated to the principal component axes, using

Y = XV.

However, in practice the data are usually centered before rotating, so in practice the trans-
formation
Y = Hn XV
is used.§ In the context of principal components analysis, the matrix of eigenvectors, V,
is often known as the loadings matrix, and the transformed observations in Y are often
known as component scores.
§
The usual caveat about not actually using the centering matrix to center the data applies.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 77

●
●●
●● ●●●
●
● ●●●●
●● ●
●● ●●
●
●●
●
●●● ●
●
● ● ●
● ●

60
● ●●
● ●●
● ●
●●
●●
●
●
●
●●●
●
●● ●●
●●●
20

●●
●●●●
●●
●● ●
●
●●●●
●● ●● ●●● ●
Y[, 2]

●●●● ● ●●●●
●● ● ●●●●
0

●
●● ● ● ●●
●●●●● ● ●● ●● ●● ●
● ●●● ●●● ●●● ●●● ●
●● ●● ●● ●●●
●
● ●●
● ● ●● ● ●
● ● ●● ● ●● ●
●● ●●
● ● ● ● ● ● ● ● ● ● ●
●● ●
● ●●● ●● ●
●
●●
●●●
●●
● ●●●●
●● ● ●●●●
●●●● ●
●● ● ● ●● ● ● ● ● ●● ●● ● ● ● ●●●●●●● ●●●
●●●
−40

●●
● ●● ● ●
●●
●●●●●
● ●● ● ● ● ● ● ● ●● ●●●
●
●●

−200 −100 0 100 200

Y[, 1]

Figure 2.2: Scatterplot for the Galaxy data showing the second principal component plotted against
the first

Example: Galaxy data

We will begin by constructing a PCA “by hand” using the spectral decomposition of the
sample variance matrix. We will use the objects v (the sample variance matrix), W (the
centered data matrix) and e (the spectral decomposition of v) from the previous example.
We can examine the variance explained by the principal components using
> cumsum(e$values)/sum(e$values)
[1] 0.8249659 0.9504373 0.9921806 0.9980732 1.0000000
So we see that the first principal component already explains over 80% of the total varia-
tion, and the first two together explain over 95% of the variation. We should therefore get a
very adequate representation of the data by projection into the first two components, and
essentially all of the variation is explained by the first 3 components. We can construct
the principal component scores with
Y=as.matrix(W)%*%e$vectors
We can plot the second principal component against the first using
plot(Y[,1],Y[,2],pch=19,col=galaxy$angle)
giving the plot shown in Figure 2.2.
As can be verified by inspecting the factor loading matrix, the first principal component
is dominated by velocity and the second is dominated by angle. We have rotated
the data to find a 2d projection where the observations are as “spread out” as possible.
We can also plot the third principal component against the first (not shown) in order to
get information “orthogonal” to the angle dominated principal component. In fact, if the
rgl package is installed, we can produce an interactive 3d plot of the first 3 principal
components as
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 78

Figure 2.3: Screen grab of a 3d interactive plot of the first three principal components for the
galaxy data

require(rgl)
plot3d(Y[,1:3],col=galaxy$angle)
Since the first 3 principal components explain over 99% of the variation in the data, es-
sentially no information is lost in this 3d projection, and the ability to interact with the plot,
and drag it around to look at it from different angles gives a great deal of insight into the
structure of the data. A screen grab of this plot is shown in Figure 2.3.
We have investigated the principal components for this data by hand, but R has built-in
functions for carrying out principal components analysis. The classic function for carrying
out PCA was the princomp() function, which works similarly to how we have just been
doing things “by hand”.
> pca=princomp(galaxy)
> pca
Call:
princomp(x = galaxy)

Standard deviations:
Comp.1 Comp.2 Comp.3 Comp.4 Comp.5
98.041939 38.235447 22.053993 8.286072 4.738194

5 variables and 323 observations.

> pca$loadings

Loadings:
Comp.1 Comp.2 Comp.3 Comp.4 Comp.5
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 79

east.west 0.453 0.112 0.883

north.south -0.208 0.315 0.888 -0.262
angle -0.998
radial.position 0.183 0.825 -0.363 -0.389
velocity 0.959 -0.113 0.256

Comp.1 Comp.2 Comp.3 Comp.4 Comp.5

SS loadings 1.0 1.0 1.0 1.0 1.0
Proportion Var 0.2 0.2 0.2 0.2 0.2
Cumulative Var 0.2 0.4 0.6 0.8 1.0
> dim(pca$scores)
[1] 323 5
> plot(pca$scores[,1],pca$scores[,2],pch=19,col=galaxy$angle)
The final command produces a plot very similar to Figure 2.2. Also note that summary
(pca) will print some additional information and that plot(pca) will produce a barplot
indicating the variances of the principal components.

2.5.4 Construction from the SVD

Although the spectral decomposition of the variance matrix is a conceptually elegant way
to understand principal components, it turns out that constructing them this way from a
sample variance matrix is inefficient and numerically unstable. Starting from a data matrix
X, we first compute the centred data

W = Hn X.

Then the sample variance matrix is given by

1
S= WT W.
n−1
If we were to follow the usual approach, we could go on to form the spectral decomposition
of S and construct principal components as before. However, it turns out that we can avoid
the computation of S completely using a SVD of the centered data W. Putting

W = UDVT

we see that
1
S= WT W
n−1
1
= VDUT UDVT
n−1
1
= VD2 VT
n−1
= VΛVT ,

where
1
Λ= D2 .
n−1
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 80

We have therefore constructed the spectral decomposition of S directly from the SVD
of W, and the loadings matrix of the principal components corresponds precisely to the
right singular vectors of W. We can use this loadings matrix together with the diagonal
matrix of eigenvalues, Λ just as if we had constructed S and then carried out a spectral
decomposition.

Example: galaxy data

First we can confirm that the SVD gives the correct loading matrix, using
> svd(W)$v
[,1] [,2] [,3] [,4] [,5]
[1,] -0.051396600 -0.02254908 0.45279536 -0.1122205 0.88274167
[2,] 0.208493672 0.01789604 0.31495251 -0.8879502 -0.26183862
[3,] -0.002463535 0.99826125 0.04727651 0.0346832 0.00551559
[4,] -0.182722226 -0.05000384 0.82508240 0.3634307 -0.38893366
[5,] -0.959424462 0.01205694 -0.11307185 -0.2562542 -0.03013096
We see that apart from some arbitrary sign changes, everything is as it should be. We
looked previously at the R command princomp() which computes principal components
the classical way, via spectral decomposition of the sample covariance matrix. R also
has a command prcomp() which instead uses the SVD of the data matrix, as described
above. Use of this command is similar to princomp(), but slightly different. The following
R session demonstrates its use.
> pca=prcomp(galaxy)
> pca
Standard deviations:
[1] 98.194060 38.294773 22.088212 8.298929 4.745546

Rotation:
PC1 PC2 PC3 PC4
east.west -0.051396600 -0.02254908 0.45279536 -0.1122205
north.south 0.208493672 0.01789604 0.31495251 -0.8879502
angle -0.002463535 0.99826125 0.04727651 0.0346832
radial.position -0.182722226 -0.05000384 0.82508240 0.3634307
velocity -0.959424462 0.01205694 -0.11307185 -0.2562542
PC5
east.west 0.88274167
north.south -0.26183862
angle 0.00551559
radial.position -0.38893366
velocity -0.03013096
> pca$rotation
PC1 PC2 PC3 PC4
east.west -0.051396600 -0.02254908 0.45279536 -0.1122205
north.south 0.208493672 0.01789604 0.31495251 -0.8879502
angle -0.002463535 0.99826125 0.04727651 0.0346832
radial.position -0.182722226 -0.05000384 0.82508240 0.3634307
velocity -0.959424462 0.01205694 -0.11307185 -0.2562542
PC5
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 81

NSCLC

20
COLON
●RENALRENAL
OVARIAN OVARIAN NSCLC COLON
RENAL
RENAL ● ● UNKNOWN
OVARIAN
PROSTATE ●
●NSCLC
OVARIAN
OVARIAN
● COLON
●COLON MCF7A−repro
COLON
MCF7D−repro
● BREAST ●
BREAST
●RENAL
NSCLC
RENAL NSCLC ● ●
● PROSTATE COLON
OVARIAN
● ● ● ● ●
● ● ●
BREAST
● ●
RENAL
● ●
BREAST ● ● ● COLON

● CNS
NSCLC ● NSCLC
RENAL
● NSCLC●NSCLC ●
CNS
CNS ●
● CNSCNS
●MELANOMA
BREAST ●● ● ● ●
●
●●
0
K562B−repro
● ● LEUKEMIA
K562A−repro
pca$x[, 2]

LEUKEMIA ●
LEUKEMIALEUKEMIA
● LEUKEMIALEUKEMIA
RENAL ● ● ●●
● ●
●
−20

MELANOMA
MELANOMA
MELANOMA
● ●
MELANOMA
●
−40

MELANOMA
●
MELANOMA MELANOMA
BREAST
●● ●
● BREAST

−40 −20 0 20 40 60

pca$x[, 1]

Figure 2.4: Scatterplot of the first two principal components of the nci microarray data

east.west 0.88274167
north.south -0.26183862
angle 0.00551559
radial.position -0.38893366
velocity -0.03013096
> dim(pca$x)
[1] 323 5
> plot(pca$x[,1],pca$x[,2],pch=19,col=galaxy$angle)
The final command again gives a plot similar to Figure 2.2. Since the SVD provides a more
numerically stable approach to PCA, use of the command prcomp() is recommended
unless there is a good reason not to.

Example: microarray data

We can carry out PCA for the nci microarray data using the following commands
pca=prcomp(t(nci))
plot(pca$x[,1],pca$x[,2],pch=19,col=4)
text(pca$x[,1],pca$x[,2],colnames(nci),cex=0.3,pos=3)
This results in the plot shown in Figure 2.4.
Analysing the plot, we see some interesting structure immediately. There seems to be
a cluster of Melanoma samples at the bottom, with two Breast cancers seemingly part of
the same cluster. There also seems to be a cluster of Leukaemia samples on the right
hand side of the plot, and a cluster of Renal cancers towards the left hand side of the
plot. The first two principal components therefore provide us with a 2d projection of the
6,830 dimensional data that immediately allows us to begin to understand some of the
underlying structure.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 82

●
●● ● ●● ●● ●●●●
●
●●●●
●
●●
●
●●●
●●
●
●●
●
●●● ●
●
●●
●●● ●● ●●
●●●
●
● ●●
●
●
●●
●●
●●
●●●
●
●●
●●●
●● ●
●
●●
●
●
● ●
●●●● ●
●● ●
●
●●
●●
●●
●
●●●
●
●
●
●●●
●
●●
●
●●
●
●●
●
●●
●●●
●
●
●●●
●●●●●
●
● ● ● ●●
● ●●
●● ●
● ●
● ● ● ● ●●● ●● ●
●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●●● ●
●
●
●
●
●
●●
●●
●●
●
●●● ● ●●●● ● ● ●
●●
●
●●
●
●
●
●●
●
●●
●
●●
●●
●●
●
●
●
●
●●
●●
●
●●
●
●●
●
●
●
●●
●●●
●●
●●●●
●●
●●●●
●
●●●
●●
●●● ●
● ●●●● ●● ● ●
● ● ● ● ● ● ● ● ●
5
●
●●●
●●●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●●●
●
●
●
●
●●
●●
●●
●●
●
●●
●
●
●●
●●
●
●●
●
●●
●
● ●
●
●●
●
●●●●
●●●
● ●
●●
●●● ●
●
●
● ●
●●
●●
● ●
● ●●●●● ●● ●
●●
●●
●●
●
●●
●●
●
●
●●
●
●
●
●●
●
●●
●
●
●
●●
●
●●
●●
●
●
●
● ●
●
●
●
●
●●
●
●
●●
●
●●
●●
●
●●●
●
●
●●●
●
● ● ● ●●
●●●●
●●●●
●● ●
●●
●●
●●●
●●●
● ● ● ●●
●●●
●●
●
●●●
●
●●
●
●●
●
●●
●●
●
●
●●●
● ●
●
●
●●
●●
●
●●
●●●
●●●
●●
●●●
●
●●
●
●●
●●
●
●●
●●
●●●●
●
●
●●●●●
●
● ●
●
●●● ●
●●
●
●●
●
●
●●● ●●● ●
●● ●●●●●●● ●
● ●
●
●
●●
●●
●
●●
●
●
●●
●
●
●●
●
●●●
●
●
●●●
●
●
●
●●●●
●●
●●●●
●
● ●
●●●
●
●
●●
●
●● ●
●
●
● ●
●
●●
●●●●●
●●●●●●
●
●
●●
●●
●●● ●
●
●
● ●
●●
●●●●
●●●● ● ●
● ●●●
pca$x[, 2]

●●● ●
●●
●●
●●
●●
●
●● ●
●●
●●
●●
●●●●●●●
●●
●●● ●
●●
● ● ●●
● ●●●● ●
●●●●●●
●
●●●●●●
●●● ●
●●
● ● ● ● ●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●●
●●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
● ●
●●
●● ●●
●●●●● ●●
●●
●●
● ●
●
●●
●
●
●●●
●
●●
●
●
●●
●
●●
●●
●
●●
●
●
●●
●●
●
●●●
●●
●
●●
●
●
●●●
●
●●
●●
●
●●
●●
●
●
●●
●●
●
●●
●●●
●
●
● ●
●●
●●
●
●
●●
●●
●●
●
●
●
●●
●●
●●●●
●
●
●
●●
●●
●●
●
●●
●
●●
●●
●
●●
●●
●
●●
●●●
●●●
●
●
●●●
●●
●
●●●
●●●
●●
● ● ●●●● ●● ●●● ● ●
●
●
●
●
●●●
●
●
●●
●
●●
●●●
●●
●
●●● ●
●
● ●●
●
●
●●
●●
●
●●
●●
●
●●
●
●●●
●●
●
●●
●●●
●
●●
●
●●
●●
●
●● ●
●
●
●●
●●●
●●
●
●●
●
●●
●
●●●
●
●●●
●
●●
●●●●●
●●●
●●● ●●●
●
●●
●
●●
●
●●●●● ●● ● ● ●
●●●●●
●●
●
●●
●
●
●●●●●● ●
●● ●●
●● ●
●
●●●●
●
●●●
●
●●●●
● ●
●
●
●●
●
●●
●●●
●●●
●
●●
●●●●●
●
●●
●
●●
●
●●●
●●
●●
●
●●
●●● ● ●●●
●●●
●● ●
●
●●●
●●●●●●●●●
●●● ●● ●●
●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
● ●●
●
●●●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●●●
●●
●
●
●● ●●
●
●● ●●
●●
●●● ●
●●●●●● ● ●● ●
●
● ●●● ●●● ● ● ●●●●●
●
●● ●●● ● ● ●
0

●●
●
●●
●●
●●●●
●●
●●●
●●
●
●●
●
●●
●●●
●●
● ●
●●
●
●●
●●
●●
●
●●
●
●
●●●
●●
●
●●●
●●
●●
●
●●
●
●●●
●●
●●●
●
●
●●
●●●
●●
●●
●●
●●
●●
●
●●●
●●
● ●
●
●●
●●●●
●
●●●●
●
●●
●●
●● ●
●
●●●
●● ●●●● ●●●
●●●
●
●●●●
● ● ●●
●●
●●
●●
●
●●●
●●
●●●
●●
●●
●●
●●●
●●●
●
●●
●●●●●
●●●
●
●●●
●
●●
●
●
●●●
●●●
●●
●●
●●
●
●●●●
●●●
●●
●●
●
●●
●●●
●●
●●
●●
●●
●●●●
●●●●● ●● ● ● ●
●
●● ●●●●
●
● ●●●●
●● ●●●
●
●●
●●●
● ●
●●
●
●
●●
●●
●
●
●●●●
● ●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●●
●
●●
●●●
●●
●●
● ●
●●
●
●
●●
●●
●●●
●
●
●●
● ●●
●●
●
●
●●
●●
●
● ●
● ●
●
●●●●●●
● ●●
● ●●
●● ●
●
●●●
● ●● ● ● ●●
●
●●●●
●●
●●
●●●
●
●
●●●
●●
●
●●●●
●●●●
●
●●●●
●
●
●
●
●
●●●
●●
●
●
●●●
●
●
● ●
●
●
●●
●
●
●●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
●●
●
●
●
●
●●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●●
●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●●
●●
●●●
●
●●
●●●
●●●●
●●
●
●
●●●● ●●●●
● ●
●● ●
●●
●
●
●●
●
●●
●●
●
●●
●
●
● ● ●● ●●●●●
●
●● ●
●
●●
●●
●
●●
●●●
●●
●
●
●●●
●●
●●
●
● ●
●●
●●
● ●
●●
●
●
●●
●
●
●●●
●
●
●
●●
● ●
●●●
●●● ●
●
●●●
●●
●
●●
●●●
●●
●●
●●●●●●● ●●
●●●●
●
●● ● ●
●●●
●● ● ● ● ●
●
●
●
●●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●●●● ●
●●
●
● ●●
●●●●
●● ●
● ●●●
●●●
●● ●
●●
●●●
●
●
●●● ●
●●
●
● ● ●●
●●
●●
●●
●●●
●● ●
● ●●●● ●
●
● ● ●●●
●●
●
●●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●●●
●
●
●●
●
●
●●
●●●●
● ●● ● ●
● ●●●
●
●●●
● ●
●●●
●●
●
●●●
●
●
●●
●
●●
●
●
●
●
●●
●●
●
●
●●
●
●
●●
●●●●●●●●
●●
●●●●
●●●●
● ●
●
●●
●●
● ●●
●●
●● ●● ●●●
●●
●
● ●
●
●●
●●
●●
●
● ● ● ●●
● ● ●● ●
●
●
●●
●
●
●●
●●
●
●●
●●●●
● ● ● ●● ●●
● ●● ●
● ● ● ● ●
● ●
● ● ●
●
●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●●●
●●
●
●●●●●
●● ●●
●●●●●● ●●
●
●
●
●
●
●
●
●
●
●●●
● ●
●
●●●
●
●●
●
●
●●
●
●
●●●
●
●
●
●●
●
● ●
●●●●
●
●●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●●●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
●●●●
●
●
● ●●● ● ●
● ● ●
●●
●●
●
●●●
●
● ● ●
● ●● ●●●●●● ● ● ●
● ● ●●●●●●● ●● ●● ● ●
● ● ●
●
−5

●●●●
●
● ● ●●● ● ● ● ●●
●●● ● ●
● ● ●●●●●●●●
●● ● ●
●●
●●●
● ●
●●●●● ● ●●
●●●●
● ●●●●
●
●●●
● ●● ●●
●
●
●
●
●● ●●
●● ●● ●
●
●
●●
●
●
●
●
●
●
● ●
●
●● ●
●
●●
●●
●
●
●●
●
●
●●●●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●● ●●
●●
●
●
● ●
●●
●●
●●
●●
●● ●●●● ●●●● ●
● ● ● ●
●● ●●
●● ●●●●● ●●●
●
●●●●●●
● ●
●● ●●
●●●
● ●
●●● ●● ● ● ● ●
●● ●●●●●● ●●●● ●●● ● ●● ●● ●
● ● ● ● ●●●●●● ● ●
● ●●
● ●

−5 0 5 10

pca$x[, 1]

Figure 2.5: Scatterplot of the first two principal components of the zip.train digit image data

Example: ZIP code digits

We can carry out a basic PCA for the ZIP digits as follows:
pca=prcomp(zip.train[,-1])
plot(pca$x[,1],pca$x[,2],pch=19,col=zip.train[,1])
This gives the plot shown in Figure 2.5.
Although there is clearly some structure in the plot, the individual digits are not well
separated in this 2d projection. Note that it is not the goal of PCA to find projections to do
this — that is the goal of discrimination, which we shall analyse later in the course. How-
ever, PCA often happens to throw up projections which do discriminate, since they often
correspond to the projections with greatest variation. However, analysis of a 3d plot of the
first 3 principal components (not shown) shows that there are no obvious 2d projections
of the first 3 principal components that discriminate between the digits particularly well.
So far we have not been paying much attention to the loadings for the principal com-
ponents. Note that these are linear combinations of the variables, and hence can be
interpreted as observations. Since here the observations are images, we can visualise
the loadings of the principal components as images, also. We can generate images cor-
responding to the loadings of the first 4 principal components with
op=par(mfrow=c(2,2))
for (i in 1:4) {
image(matrix(pca$rotation[,i],ncol=16)[,16:1],col=grey(15:0/15))
}
par(op)
leading to the plot shown in Figure 2.6.
This provides an effective way to understand the loading vectors in this particular case.
CHAPTER 2. PCA AND MATRIX FACTORISATIONS 83

0.8

0.8
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
0.8

0.8
0.0

0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

Figure 2.6: Images of the first 4 principal component loading vectors for the zip.train digit
data

2.6 Conclusion
Numerical linear algebra and matrix computations are the key tools needed to make sense
of multivariate data sets. The methods covered in this chapter provide the foundation for
much of the rest of the course.
Note that the standard reference on numerical linear algebra is:

Golub, G.H. and Loan, C.F.V. (1996) Matrix computations, Johns Hopkins University
Press.

Some of the examples in this chapter were based on corresponding examples in the
above text. In addition, see section 3.2.3 (p.52) and 3.9 (p.93), as well as section 14.5
(p.534) of [HTF] for further details.
Chapter 3

Inference, the MVN and multivariate

regression

3.1 Inference and estimation

So far we have considered the analysis of data matrices X, including the computation of
the sample mean vector x̄ and variance matrix SX . We have also considered random
vectors X, including the expectation vector E(X) and variance matrix Var(X). How-
ever, although we have informally noted a relationship between sample and population
quantities, we have not yet formally established a connection between the two.

Proposition 21 Suppose that we form an n×p data matrix X with rows X i T , i = 1, 2, . . . , n

where the X i are independent realisations of a multivariate random quantity with E(X i ) =
µ and Var(X i ) = Σ. Assuming that the elements of µ and Σ are all finite, we have the
following:

1. The sample mean X̄ is unbiased for µ, that is

E X̄ = µ.

2. The variance of X̄ is given by

1
Var X̄ = Σ.
n

3. X̄ is a consistent estimator of µ, that is

n
P kX̄ − µk < ε −→ 1, ∀ε > 0.
∞

4. The sample variance SX is unbiased for Σ, that is

E(SX ) = Σ.

Before demonstrating these results, it should be noted that under some additional (fairly
mild) assumptions it can be shown that the sample variance matrix is also a consistent
estimator of Σ, but we will not pursue this here.

84
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 85

Proof

1.
n
!

1 X
E X̄ = E Xi
n i=1
n
1 X
= E(X i)
n i=1
n
1 X
= µ
n i=1
= µ.

2.
n
!

1 X
Var X̄ = Var Xi
n i=1
n
1 X
= Var(X i)
n2 i=1
n
1 X
= Σ
n2 i=1
1
= Σ.
n


3. Consistency

is intuitively obvious, since 1. and 2. tell us that E X̄ → µ and
Var X̄ → 0 as n → ∞. However, we can establish consistency more formally
by applying Markov’s inequality to kX̄ − µk2 . First note that
p
X
2 T
kX̄ − µk = (X̄ − µ) (X̄ − µ) = (X̄i − µi )2 ≥ 0.
i=1
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 86

Now we have
p
!
X
2
(X̄i − µi )2


E kX̄ − µk = E
i=1
p
X
E (X̄i − µi )2


=
i=1
p
X

= Var X̄i
i=1
p
X1
= Var(Xi )
i=1
n
1
= Tr (Σ) .
n
Then Markov’s inequality tells us that for any a > 0 we have

Tr (Σ)
P kX̄ − µk2 ≥ a ≤
na
Tr (Σ)
⇒ P kX̄ − µk2 < a ≥ 1 −


na

and putting a = ε2 we get

Tr (Σ)
P kX̄ − µk2 < ε2 ≥ 1 −


nε2

Tr (Σ)
⇒ P kX̄ − µk < ε ≥ 1 −
nε2
n
−→ 1.
∞

4. First note that

n
!
1 X
E(SX ) = E [X i − X̄][X i − X̄]T
n − 1 i=1
n
1 X
E [X i − X̄][X i − X̄]T


=
n − 1 i=1

Now since n

1X
E X i − X̄ = µ − µ=0
n i=1
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 87

we have

E [X i − X̄][X i − X̄]T = Var X i − X̄

!
n−1 1X
= Var Xi − Xj
n n j6=i
(n − 1)2 1 X
= Var(X i ) + Var(X j )
n2 n2 j6=i
(n − 1)2 n−1
= 2
Σ+ Σ
n n2
n−1
= Σ,
n
and so n
1 Xn−1
E(SX ) = Σ = Σ.
n − 1 i=1 n


These results now make explicit the relationship between the data summaries and
associated population quantities, and give us confidence that, at least in the case of
“large n”, our sample summary statistics should be “good” estimators of the corresponding
summaries for the population from which the data is sampled.

3.2 Multivariate regression

In this section we will examine the problem of regression for a multivariate observation.
We shall begin by reminding ourselves about multiple regression for a univariate output
as a “least squares” procedure, before going on to see how this can be generalised to
the case of multivariate output. Key to understanding regression as a “least squares”
problem is the idea of minimising an error function with respect to a vector (or matrix)
of parameters. There are various ways to approach this problem, but often the simplest
is to use calculus to find a stationary point of the cost function (by differentiating wrt the
parameters and equating to zero). For this purpose it is helpful to use some results on
differentiating various scalar functions with respect to vector and matrix parameters. We
state the following results without proof.∗

Proposition 22 (Derivatives wrt vectors and matrices) In each of the following results,
the derivative is of a scalar function, and is wrt a vector x or matrix X, where the compo-
nents of the vector or matrix are assumed to be algebraically independent.

1.
∂ T ∂ T
a x= x a=a
∂x ∂x
∗
The results can all be found in The Matrix Cookbook, to which you are referred for further details and many other
interesting results and examples. Note that you are not expected to memorise these results, and their derivations are
not examinable in this course.
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 88

2.
∂ T
x Ax = (A + AT )x
∂x
and note that this reduces to 2Ax when A is symmetric.

3.
∂
Tr (XA) = AT
∂X
4.
∂
Tr XT AX = AX + AT X


∂X
and note that this reduces to 2AX when A is symmetric.

5.
∂ ∂
Tr XT AXB = Tr BXT AX = AXB + AT XBT



∂X ∂X
and note that this reduces to 2AXB when both A and B are symmetric.

6.
∂
|X| = |X|X−T
∂X
7.
∂
log |X| = X−T
∂X
8.
∂ T
a Xb = abT
∂X
9.
∂ T −1
a X b = −X−T abT X−T
∂X

Now we have the necessary technical results, we can think about the regression problem.

3.2.1 Univariate multiple linear regression

Let’s start with classical multiple linear regression where we model each output yi as a
linear combination of a corresponding covariate vector xi , subject to error, as

yi = xi T β + εi , i = 1, 2, . . . , n.

We assume that the covariates are the rows of an n × p data matrix X, and so β is a
p-dimensional vector. We can write this in matrix form as

y = Xβ + ε.

Since we typically consider the case n > p, the presence of the error term ε is necessary,
otherwise the system would be over-determined, and then typically no β would exist in
order to solve the linear system
y = Xβ.
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 89

However, by introducing ε, we ensure that every β is a potential solution, for an appropri-

ate choice of ε. The least squares approach is to choose the β which makes the 2-norm
of ε as small as possible. That is, we choose β in order to minimise
n
X
2 T
kεk = ε ε = ε2i .
i=1

We can find such a minimising solution by first noting that

εT ε = (y − Xβ)T (y − Xβ)
= y T y − 2y T Xβ + β T XT Xβ,

and then differentiate wrt β as

∂ T
ε ε = −2XT y + 2XT Xβ.
∂β
Here we have used results 1. and 2. of Proposition 22. Then equating to zero leads to the
“normal equations”
XT Xβ̂ = XT y.
The mathematical solution to this is

β̂ = (XT X)−1 XT y,

but there are very efficient and numerically stable ways to compute this based on matrix
decompositions such as the QR factorisation, as discussed in Section 2.4.2.

Example: galaxy data

Suppose that we would like to use the variables angle and radial.position in order to
predict the variable velocity. We can accomplish this in R using the built-in command
lm(), either using R model notation, as
> lm(velocity ∼ angle+radial.position,data=galaxy)

Call:
lm(formula = velocity ∼ angle + radial.position, data = galaxy)

Coefficients:
(Intercept) angle radial.position
1586.9882 0.1076 2.4515
or just using vectors and matrices, as
> lm(galaxy[,5] ∼ as.matrix(galaxy[,3:4]))

Call:
lm(formula = galaxy[, 5] ∼ as.matrix(galaxy[, 3:4]))

Coefficients:
(Intercept)
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 90

1586.9882
as.matrix(galaxy[, 3:4])angle
0.1076
as.matrix(galaxy[, 3:4])radial.position
2.4515
Note that lm() includes an intercept term by default, as this is usually what we want.
However, we can tell R not to include an intercept term by using 0+ at the beginning of
the covariate list, either using model notation
> lm(velocity ∼ 0+angle+radial.position,data=galaxy)

Call:
lm(formula = velocity ∼ 0 + angle + radial.position, data = galaxy)

Coefficients:
angle radial.position
16.163 3.261
or matrix notation:
> lm(galaxy[,5] ∼ 0+as.matrix(galaxy[,3:4]))

Call:
lm(formula = galaxy[, 5] ∼ 0 + as.matrix(galaxy[, 3:4]))

Coefficients:
as.matrix(galaxy[, 3:4])angle
16.163
as.matrix(galaxy[, 3:4])radial.position
3.261
We can compute the regression coefficients directly using a QR factorisation with
> QR=qr(as.matrix(galaxy[,3:4]))
> backsolve(qr.R(QR),t(qr.Q(QR))%*%galaxy[,5])
[,1]
[1,] 16.163269
[2,] 3.261159
The intercept term can be found by appending 1 to the front of X, as
> QR=qr(cbind(rep(1,323),galaxy[,3:4]))
> backsolve(qr.R(QR),t(qr.Q(QR))%*%galaxy[,5])
[,1]
[1,] 1586.9881822
[2,] 0.1075920
[3,] 2.4514815
So we see that least squares problems can be solved with a QR factorisation, a rotation,
and a backwards solve. This is the method that the lm() function uses.
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 91

3.2.2 The general linear model

Now consider the obvious generalisation of the previous case, where our “output”, y i is
now a q-vector, rather than a scalar. We then have the model

y i = BT xi + εi , i = 1, 2, . . . n,

where B is now a p × q matrix. We can write this in matrix form as

Y = XB + E,

where E is an n × q matrix of “errors”, and this is typically known as the general linear
model. The regression problem here is, given data matrices X and Y, to find the matrix B
which makes the Frobenius norm of the error matrix, kEk, as small as possible. Note that
q
n X
X
2 T
ε2ij ,


kEk = Tr E E =
i=1 j=1

and so this is again a least squares problem, where we choose B to minimise the sum of
squares of the errors.

We proceed as before by noting that

T T



Tr E E = Tr [Y − XB] [Y − XB]
= Tr YTY − YTXB − BTXTY + BTXTXB


T T T T T T





= Tr Y Y − Tr Y XB − Tr B X Y + Tr B X XB
= Tr YTY − 2 Tr YTXB + Tr BTXTXB .

Now differentiating wrt the matrix B we get

∂
Tr ETE = −2XTY + 2XTXB,


∂B
where here we used results 3. and 4. of Proposition 22.
Equating to zero gives a matrix version of the normal equations,

XT XB̂ = XT Y.

This has mathematical solution

B̂ = (XT X)−1 XT Y.
By considering the columns of B̂ and Y, we see that this corresponds to the usual multiple
linear regression solution for each column of Y. That is, the multivariate regression prob-
lem is solved by separately regressing each column of Y on the data matrix X. However,
the full problem will be solved more efficiently and stably by directly solving the above
normal equations using a QR factorisation of X (with multiple RHSs). To be clear, if we
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 92

have the QR factorisation of X as X = QR, substituting in to the normal equations leads

directly to
RB̂ = QT Y.
These equations may be solved for B̂ as a triangular system with multiple RHSs (generally
to be preferred), or explicitly solved as

B̂ = R−1 QT Y.

Example: microarray data

To keep things simple, let us consider using the first 5 genes in the nci dataset in order to
predict the next 3. We therefore wish to compute B̂, which will be a 5 × 3 matrix, if we do
not include an intercept term, and will be 6 × 3 if we do, with the first row corresponding to
a location shift. Recall that the nci data is stored as a matrix with rows representing the
genes. We first solve using lm(), with and without an intercept
> X=t(nci[1:5,])
> Y=t(nci[6:8,])
> lm(Y ∼ X)

Call:
lm(formula = Y ∼ X)

Coefficients:
[,1] [,2] [,3]
(Intercept) 0.005571 0.014090 -0.031221
X1 0.055359 0.011646 0.327704
X2 0.069211 0.163145 -0.031904
X3 -0.059708 -0.028433 -0.025860
X4 0.004978 -0.017083 -0.038912
X5 0.175285 0.160014 0.005762

> lm(Y ∼ 0+X)

Call:
lm(formula = Y ∼ 0 + X)

Coefficients:
[,1] [,2] [,3]
X1 0.0550853 0.0109542 0.3292373
X2 0.0694000 0.1636241 -0.0329654
X3 -0.0595292 -0.0279796 -0.0268643
X4 0.0034643 -0.0209107 -0.0304312
X5 0.1762379 0.1624230 0.0004232
We now solve directly using the QR factorisation, first without an intercept, and then with.
Note how the backsolve() function solves multiple RHSs exactly as we would wish.
> QR=qr(X)
> backsolve(qr.R(QR),t(qr.Q(QR))%*%Y)
[,1] [,2] [,3]
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 93

[1,] 0.055085288 0.01095420 0.3292372739

[2,] 0.069399969 0.16362409 -0.0329654306
[3,] -0.059529200 -0.02797956 -0.0268642629
[4,] 0.003464287 -0.02091073 -0.0304311643
[5,] 0.176237890 0.16242302 0.0004231837
> X=cbind(rep(1,64),X)
> QR=qr(X)
> backsolve(qr.R(QR),t(qr.Q(QR))%*%Y)
[,1] [,2] [,3]
[1,] 0.005570868 0.01408965 -0.031220986
[2,] 0.055358790 0.01164593 0.327704478
[3,] 0.069210662 0.16314530 -0.031904493
[4,] -0.059708338 -0.02843263 -0.025860316
[5,] 0.004977558 -0.01708341 -0.038912041
[6,] 0.175285335 0.16001385 0.005761614

3.2.3 Weighted errors

One possible concern with the above solution could be that it appears to treat all errors
with equal weight, even though some outputs may be measured on different scales, and
some outputs may also be correlated. If we know the variance matrix associated with the
errors,
Var(εi ) = Σ,
we can use the symmetric square root to form the Mahalanobis transformation of the error
matrix
E0 = EΣ−1/2 ,
and instead minimise

Tr E0T E0 = Tr Σ−1/2 ET EΣ−1/2 = Tr Σ−1 ET E .

We now note that

Tr Σ−1 ET E = Tr Σ−1 [Y − XB]T [Y − XB]

= Tr Σ−1 YT Y − Σ−1 YT XB − Σ−1 BT XT Y + Σ−1 BT XT XB

= Tr Σ−1 YT Y − Tr Σ−1 YT XB − Tr Σ−1 BT XT Y + Tr Σ−1 BT XT XB

= Tr Σ−1 YT Y − 2 Tr Σ−1 YT XB + Tr Σ−1 BT XT XB .

Now differentiating wrt the matrix B we get

∂
Tr Σ−1 ET E = −2XT YΣ−1 + 2XT XBΣ−1 ,


∂B
where here we used results 3. and 5. of Proposition 22. Equating to zero gives the same
normal equations as previously, since the variance matrix cancels out. Therefore the
weighted solution is in fact the same as the unweighted solution, and appropriate irre-
spective of any covariance structure associated with the errors.
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 94

3.3 The multivariate normal (MVN) distribution

Many courses on multivariate statistics introduce the multivariate normal distribution very
early, and then make extensive use of the properties of this distribution. The problem with
such an approach is that it gives the impression that almost all of multivariate statistical
theory is dependent on an assumption of multivariate normality, and as we have seen,
this is not the case at all. However, the MVN distribution is the most important non-trivial
multivariate distribution in statistics, and has many interesting applications. It is therefore
appropriate at this point in the course to consider it, and re-consider some of the results
of this chapter in the context of MVN data.
We begin by defining multivariate normal random quantities to be affine transforma-
tions of standard normal random vectors.

Definition 9 Let Z = (Z1 , Z2 , . . . , Zq )T , where the components Zi are iid N (0, 1). Then for
a p × q matrix A and p-vector µ, we say that

X = AZ + µ

has a multivariate normal distribution with mean µ and variance matrix Σ = AAT , and we
write
X ∼ N (µ, Σ).

There are several issues with this definition which require clarification. The first thing to
note is that the expectation and variance of X do indeed correspond to µ and Σ, using
properties of affine transformations that we considered in Chapter 1. The next concern
is whether the distribution is really characterised by its mean and variance, since there
are many possible affine transformations that will lead to the same mean and variance. In
order to keep things simple, we will begin by considering only invertible transformations
A. In this case, we must have p = q.

Proposition 23 The density of a p-dimensional MVN random vector X, with mean µ and
positive definite variance Σ is given by
 
−p/2 −1/2 1 T −1
f (x) = (2π) |Σ| exp − (x − µ) Σ (x − µ) .
2

Proof
We choose any invertible p × p matrix A such that AAT = Σ, and put

X = AZ + µ.
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 95

We first need to know the probability density function for Z, which is clearly given by
p
Y
φ(z) = φ(zi )
i=1
p
Y 1
= √ exp{−zi2 /2}
i=1
2π
p
( )
1 X
= (2π)−p/2 exp − z2
2 i=1 i
 
−p/2 1 T
= (2π) exp − z z .
2

Now since X is an invertible transformation of Z, it’s density is given by

 
 ∂z 
f (x) = φ(z)   ,
∂x
where z = A−1 (x − µ). But now
∂z
= A−1 ,
∂x
and so  
 ∂z 
  = |A−1 | = |A|−1 = |Σ|−1/2 ,
 ∂x 
giving
 
−p/2 1 −1
f (x) = (2π) exp − [A (x − µ)] [A (x − µ)] |Σ|−1/2
T −1
2
 
−p/2 −1/2 1 T −1
= (2π) |Σ| exp − (x − µ) Σ (x − µ) .
2

The crucial thing to note here is that despite the fact that we considered any (invertible) A
leading to a required variance matrix, the density of the resulting random quantity depends
only on the variance matrix (and mean vector), and not on the particular affine transforma-
tion used in the construction. Consideration of non-invertible (and non-square) transfor-
mations complicates the analysis, but essentially the same conclusion holds, which is that
the distribution of the MVN random quantity depends only on its mean and variance, and
not on the particular affine transformation used to construct it. It is therefore reasonable
to characterise MVN random quantities by their mean and variance.

3.3.1 Evaluation of the MVN density

In some applications it is necessary to evaluate the MVN density. However, in most ap-
plications, it is sufficient to evaluate the log of the density, and this is usually safer, as it is
less susceptible to numerical underflow. Clearly the log-density is given by
p 1 1
l(x) = − log 2π − log |Σ| − (x − µ)T Σ−1 (x − µ).
2 2 2
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 96

Note that the final term is essentially the Mahalanobis distance. We can avoid direct
computation of |Σ| and Σ−1 by using an appropriate matrix decomposition. In the positive
definite case, the Cholesky decomposition usually provides the best solution, so suppose
that Σ has Cholesky decomposition

Σ = GGT ,

where G is lower triangular. To see how this helps, first consider the determinant term
1
log |Σ| = log |Σ|1/2 = log |G|.
2
But since G is lower triangular, its determinant is the product of its diagonal elements, and
so the log of this is given by
p
X
log |G| = log gii .
i=1

Similarly, the final quadratic form can be simplified as

(x − µ)T Σ−1 (x − µ) = (x − µ)T (GGT )−1 (x − µ)

= (x − µ)T G−T G−1 (x − µ)
= z T z,

where z = G−1 (x − µ), and hence may be found by forward solving the triangular system

Gz = (x − µ).

Given this solution, we can therefore write the log density as

p p
p X 1X 2
l(x) = − log 2π − log gii − zi .
2 i=1
2 i=1

This is an efficient way to compute the log density, as it just requires a Cholesky decom-
position and a single forward solve.

3.3.2 Properties of the MVN

We now consider some very useful properties of the MVN distribution.

Proposition 24 (Affine transformation) If X ∼ N (µ, Σ), and

Y = AX + b,

then Y is multivariate normal, and Y ∼ N (Aµ + b, AΣAT ).

Proof
The mean and variance of Y are clear from our understanding of affine transforma-
tions of random vectors. The key to this result is understanding why Y is MVN. For this,
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 97

note that we need to show that Y is an affine transformation of a standard normal vector,
but since X is MVN, there must exist a matrix M such that
X = MZ + µ,
where Z is a standard normal vector. Then
Y = AX + b
= A(MZ + µ) + b
= (AM)Z + (Aµ + b),
and so Y is also an affine transformation of a standard normal vector, and hence MVN.

Now since marginal distributions correspond to linear transformations, it follows that all
marginal distributions of the MVN are (multivariate) normal.
Proposition 25 If X ∼ N (µ, Σ) is partitioned as
     
X1 µ1 Σ11 Σ12
∼N , ,
X2 µ2 Σ21 Σ22
then the marginal distribution of X 1 is given by
X 1 ∼ N (µ1 , Σ11 ).
Proof
The result follows simply by noting that
 
X1
X 1 = (I, 0) ,
X2
and using properties of linear transformations. 
Similarly, univariate marginals my be computed using the fact that Xi = ei T X, giving
Xi ∼ N (µi , σii ).

Proposition 26 If X 1 ∼ N (µ1 , Σ1 ) and X 2 ∼ N (µ2 , Σ2 ) are independent, then

X 1 + X 2 ∼ N (µ1 + µ2 , Σ1 + Σ2 ).

Proof

First note that X 1 and X 2 are jointly MVN, with

     
X1 µ1 Σ1 0
∼N , .
X2 µ2 0 Σ2
Then observing that X 1 + X 2 corresponds to the linear trans-
formation  
X1
(I, I)
X2
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 98

leads directly to the result.


Finally, we state without proof the key result concerning conditional distributions of the
MVN. The proof is straightforward, but somewhat involved.

Proposition 27 If X ∼ N (µ, Σ) is partitioned as

     
X1 µ1 Σ11 Σ12
∼N , ,
X2 µ2 Σ21 Σ22

then the conditional distribution of X 1 given that X 2 is observed to be x2 is MVN, with

(X 1 |X 2 = x2 ) ∼ N (µ1|2 , Σ1|2 ),

where

µ1|2 = µ1 + Σ12 Σ22 −1 (x2 − µ2 ),

Σ1|2 = Σ11 − Σ12 Σ22 −1 Σ21 .

Note that the simple version of the result presented here assumes that Σ22 is invertible,
but it generalises straightforwardly to the case of singular Σ22 .

3.3.3 Maximum likelihood estimation

If we consider an n × p data matrix X with iid rows X i ∼ N (µ, Σ), then the likelihood for
an observed data matrix is given by
n
Y
L(µ, Σ; X) = f (xi )
i=1
n  
Y
−p/2 −1/2 1 T −1
= (2π) |Σ| exp − (xi − µ) Σ (xi − µ)
i=1
2
( n
)
1 X
= (2π)−np/2 |Σ|−n/2 exp − (xi − µ)T Σ−1 (xi − µ) .
2 i=1

Consequently, the log-likelihood is given by

n
np n 1X
l(µ, Σ; X) = − log 2π − log |Σ| − (xi − µ)T Σ−1 (xi − µ).
2 2 2 i=1

Note that if we need to numerically compute this log-likelihood, we would do so using

the Cholesky decomposition, as already described for the log-density. We can use the
log-likelihood in order to compute maximum likelihood estimators for the MVN.

Proposition 28 For an MVN model, the maximum likelihood estimator of the mean, µ is
given by
µ̂ = x̄.
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 99

Proof
We proceed by differentiating the log-likelihood function wrt µ and equating to zero.
First note that n
∂l 1X ∂
=− (xi − µ)T Σ−1 (xi − µ).
∂µ 2 i=1 ∂µ
Now
∂ ∂
(xi − µ)T Σ−1 (xi − µ) = (xi T Σ−1 xi − 2µT Σ−1 xi + µT Σ−1 µ)
∂µ ∂µ
= −2Σ−1 xi + 2Σ−1 µ
= 2Σ−1 (µ − xi ),

using properties 1. and 2. of Proposition 22. So now

n
∂l 1 X −1
=− 2Σ (µ − xi )
∂µ 2 i=1
n
X
−1
=Σ (xi − µ)
i=1
−1
= Σ (nx̄ − nµ)
= nΣ−1 (x̄ − µ),

and so equating to zero gives the result. 

Proposition 29 For an MVN model, the maximum likelihood estimator of the variance
matrix, Σ is given by
n−1
Σ̂ = S.
n
Consequently, the maximum likelihood estimator is biased, but asymptotically unbiased,
and consistent.
Proof
We will proceed by using matrix differentiation to find a matrix Σ which maximises the
likelihood.
n
∂l n ∂ 1X ∂
=− log |Σ| − (xi − µ)T Σ−1 (xi − µ)
∂Σ 2 ∂Σ 2 i=1 ∂Σ
n
n 1 X −1
= − Σ−1 + Σ (xi − µ)(xi − µ)T Σ−1
2 2 i=1
" n #
n −1 1 −1 X
=− Σ + Σ (xi − µ)(xi − µ)T Σ−1 ,
2 2 i=1
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 100

using properties 7. and 9. of Proposition 22. Equating to zero gives

" n #
X
nΣ̂−1 = Σ̂−1 (xi − µ̂)(xi − µ̂)T Σ̂−1
i=1
n
1 X
⇒ Σ̂ = (xi − x̄)(xi − x̄)T
n i=1
n−1
= S.
n

The maximum likelihood estimator of Σ therefore uses a divisor of n, rather than n − 1,
as required to obtain an unbiased estimate. Apart from this minor issue, we see that
the sample estimates x̄ and S of the population quantities µ and Σ, which we have al-
ready shown to be effective without making any distributional assumptions, may also be
interpreted as maximum likelihood estimators in cases where an MVN distributional as-
sumption is felt to be appropriate.

3.3.4 MLE for the general linear model

Let us now reconsider the general linear model

Y = XB + E,

where the observations are modelled as

y i = B T x i + εi ,

and we now make the modelling assumption εi ∼ N (0, Σ), and we assume for now that
Σ is known. We have already found the least squares estimate for B, but given the ad-
ditional distribution assumption we are now making, we can now construct the maximum
likelihood estimate of B. We first construct the likelihood as
n  
Y
−p/2 −1/2 1 T −1
L(B) = (2π) |Σ| exp − εi Σ εi
i=1
2
( n
)
1 X
= (2π)−np/2 |Σ|−n/2 exp − εi T Σ−1 εi
2 i=1
 
−np/2 −n/2 1 −1 T


= (2π) |Σ| exp − Tr EΣ E
2
 
−np/2 −n/2 1 −1 T


= (2π) |Σ| exp − Tr Σ E E ,
2

where E = Y − XB. Consequently the log-likelihood is given by

np n 1
log 2π − log |Σ| − Tr Σ−1 ET E .


l(B) = −
2 2 2
CHAPTER 3. INFERENCE, THE MVN AND MULTIVARIATE REGRESSION 101

Since the only dependence on B in this log-likelihood is via E, it is clear that maximising
this function wrt B is equivalent to minimising

Tr Σ−1 ET E .

But we have already seen that minimising this function wrt B leads to the normal equa-
tions, and the usual least squares solution

B̂ = (XT X)−1 XT Y.

So again, the natural estimator that we have already derived without making any distri-
butional assumptions may be re-interpreted as a maximum likelihood estimator in cases
where an assumption of multivariate normality is felt to be appropriate. Note further that
since the solution we obtain is independent of Σ, it also represents the maximum likeli-
hood estimate of B in the case of unknown Σ.

See 3.2 (p.44) and 3.7 (p.84) of [HTF] for further discussion of the topics discussed in
this Chapter.
Chapter 4

Cluster analysis and unsupervised learning

4.1 Introduction
4.1.1 Motivation
When we did PCA for the microarray data at the end of Chapter 2, we saw in Figure 2.4
that some of the samples appeared to “cluster” together into groups. This is something
that we can quite naturally do “by eye” in one or two dimensions, by examining the relative
“closeness” of groups of points, but it is not so obvious how to formalise the process,
or understand how it generalises to arbitrary data sets in p-dimensional space. Cluster
analysis is a formalisation of this process which can be applied to any set of observations
where some measure of “closeness” of observations can be defined. If one associates a
label with each group or cluster, then a clustering algorithm assigns a group label to each
observation in a data set. In the context of data mining and machine learning, cluster
analysis is considered to be an example of unsupervised learning. This is because it is
a mechanism for allocating observations to groups that does not require any training set,
which provides examples of labelled observations which can be used to learn about the
relationship between observations and labels. There are different approaches to cluster
analysis. Some approaches assign observations to a pre-specified number of groups,
and others hierarchically combine observations into small clusters and then successively
into bigger clusters until the number of clusters is reduced to the desired number. Of
course, a priori is is not always clear how many clusters will be appropriate, but methods
exist to address this issue, too.

4.1.2 Dissimilarity and distance

The aim of cluster analysis is to find natural groupings of the data into subsets that are
“close together” in some appropriate sense. For an n×p data matrix, with rows represent-
ing p-dimensional observations, Euclidean distance is the natural “closeness” metric, but
we often want to apply clustering algorithms to data where there is no natural metric, so
we simply require that for every pair of observations xi and xj a corresponding distance,
dij , is defined. These distances are typically considered as an n × n distance matrix, D,
and the distances are expected to have the following properties:

1. dij ≥ 0, ∀i, j;

102
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 103

2. dii = 0, ∀i;

3. dij = 0 ⇒ xi = xj .

In addition to these fundamental properties, it is expected that the dissimilarities measure

“closeness”, so that vectors that are more “similar” have dissimilarities which are smaller
than than those for vectors which are very “different”. In addition, most clustering algo-
rithms require that the dissimilarities are symmetric, so that dij = dji . A distance matrix
which is not symmetric can be symmetrised by instead using D0 = (D + DT )/2, which
clearly is. Note however, that we typically do not require strong properties of norms or
metrics, such as the triangle inequality.
For observations xi ∈ Rp , we typically use a dissimilarity measure derived from a
genuine norm. The most commonly used choices are Euclidean distance,
q
dij = d(xi , xj ) = (xi − xj )T (xi − xj )

or squared Euclidean distance

d(xi , xj ) = (xi − xj )T (xi − xj ).

Occasionally the l1 norm, often known as Manhattan distance or the “taxi cab” norm,
d(xi , xj ) = 1 p T |xi − xj |, is used. In some applications it also makes sense to use the
Mahalanobis distance, q
dij = (xi − xj )T S−1 (xi − xj ),
(or the square of this), where S is the sample variance matrix of X.
If not all variables are real-valued, then usually an overall distance function d(·, ·) is
constructed as a weighted sum of distances defined appropriately for each variable, as
p
X
d(xi , xj ) = wk dk (xik , xjk ).
k=1

For any real valued variables, the obvious choice is

dk (xik , xjk ) = (xik − xjk )2 ,

and for categorical variables, or factors, the simplest choice is given by

(
0 xik = xjk
dk (xik , xjk ) =
1 xik 6= xjk ,

though there are obviously other possibilities. The weights, wk , could be chosen to be
equal, but this probably only makes sense if the distances for the different variables are
comparable, and all variables are felt to be equally important. Otherwise the weights
should be chosen to ensure that each variable is weighted appropriately in the overall
distance function. See section 14.3 (p.501) of [HTF] for further details.
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 104

Example: toy problem

Suppose that we have the data matrix
 
1 2
3 3
 
4
X= 2.
2 1
3 1
We can create this in R using
X=matrix(c(1,2,3,3,4,2,2,1,3,1),ncol=2,byrow=TRUE)
We can compute a distance matrix for X using the R function dist(). For example
> dist(X)
1 2 3 4
2 2.236068
3 3.000000 1.414214
4 1.414214 2.236068 2.236068
5 2.236068 2.000000 1.414214 1.000000
shows that dist() returns the lower triangle of the distance matrix (not including the
diagonal, which is zero). It also defaults to Euclidean distance, but can use other metrics,
as the following session illustrates.
> dist(X,method="manhattan")
1 2 3 4
2 3
3 3 2
4 2 3 3
5 3 2 2 1
See ?dist for further details about the different metrics supported. Note that dist()
returns a dist object, which is not an R matrix object. If a matrix is required, it can be
converted to a matrix using as.matrix(), as follows.
> as.matrix(dist(X,method="manhattan"))
1 2 3 4 5
1 0 3 3 2 3
2 3 0 2 3 2
3 3 2 0 3 2
4 2 3 3 0 1
5 3 2 2 1 0
Similarly, a matrix can be converted into a dist object using as.dist(). dist objects
are used by some of the clustering functions in R.

4.2 Clustering methods

4.2.1 K-means clustering
We begin by examining a clustering method which makes most sense when the observa-
tions are regarded as elements of Euclidean space, Rp and the distance used is Euclidean
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 105

distance (though in fact, it will work for observations belonging to any inner product space
with distance defined using the induced norm), with

d2ij = kxi − xj k2 = (xi − xj )T (xi − xj ).

So suppose we have an n × p data matrix X, and consider a clustering algorithm which

allocates
Pk each of the n observations to one of k clusters, of size n1 , n2 , . . . , nk , where
n = i=1 ni . It is helpful to introduce a new indexing notation for the observations. We
will now use xij to denote the jth observation allocated to the ith cluster, and we will
define the cluster means in the obvious way as
ni
1 X
x̄i = xij , i = 1, 2, . . . k.
ni j=1

These are the k-means referred to in the name of the algorithm. At this point it is helpful
to note that we have a sum-of-squares decomposition for the distances of observations
from their centroid exactly analogous to that used in one-way ANOVA.

Proposition 30 The sum-of-squares decomposition for the clustered observations can

be written in the form
SST OT = SSW + SSB ,
where
ni
k X
X
SST OT = kxij − x̄k2
i=1 j=1
ni
k X
X
SSW = kxij − x̄i k2
i=1 j=1
k
X
SSB = ni kx̄i − x̄k2 .
i=1

The overall sum-of-squares distance for the data can therefore be orthogonally decom-
posed into a “within groups” and “between groups” sum of squares.

Proof
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 106

ni
k X ni
k X
X X
kxij − x̄k2 = kxij − x̄i + x̄i − x̄k2
i=1 j=1 i=1 j=1
k ni
XXh
= kxij − x̄ik2 + kx̄i − x̄k2
i=1 j=1
i
T
+ 2(x̄i − x̄) (xij − x̄i)
ni
k X ni
k X
X X
= kxij − x̄ik2 + kx̄i − x̄k2
i=1 j=1 i=1 j=1
ni
k X
X
+2 (x̄i − x̄)T(xij − x̄i)
i=1 j=1
ni
k X k
X X
= kxij − x̄ik2 + nikx̄i − x̄k2
i=1 j=1 i=1
k ni
X X
T
+2 (x̄i − x̄) (xij − x̄i)
i=1 j=1
ni
k X k
X X
2
= kxij − x̄ik + nikx̄i − x̄k2,
i=1 j=1 i=1
Pni
since j=1 (xij − x̄i) = 0.

An effective clustering algorithm with make the within-group sum-of-squares SSW as
small as possible (conditional on k), and so equivalently, will make SSB as large as pos-
sible (since SST OT is fixed). This, then, is precisely the goal of k-means clustering: to
allocate observations to clusters to minimise SSW . This initially seems like a simple task:
just consider every possible partition of the n observations into k clusters and choose the
allocation resulting in the smallest value of SSW . Unfortunately this is not practical for
any data set of reasonable size, as the number of possible partitions of the data grows
combinatorally. For example, for n = 100 and k = 5 the number of possible partitions is
around 1068 . Clearly a more practical solution is required. The k-means algorithm is one
possible approach.
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 107

The k-means algorithm

1. Start by picking k “means”, m1 , m2 , . . . , mk , randomly (randomly choosing k of the
n observations, without replacement, often works well)

2. Go through each observation, x, in turn, and allocate it to cluster i if mi is the

“closest” mean, that is, kx − mi k < kx − mj k, j 6= i.

3. Set each mi to be the sample mean of observations allocated to that cluster, that is,
set mi = x̄i , i = 1, 2, . . . , k.

4. If the procedure has not converged, return to step 2.

To understand why the k-means algorithm works, first note that once we have com-
pleted step 2. the first time through the loop, we have a set of means and an allocation of
each observation to a cluster, and so we can consider the quantity
ni
k X
X
SS = kxij − mi k2 ,
i=1 j=1

and note that at the end of step 3., each time through the loop, the value of SS coincides
with SSW for the particular cluster allocation. But the k-means algorithm minimises SS
with respect to both the choice of the mi and the allocation of observations to clusters,
and hence minimises SSW .
To see this, first consider step 3. We have a value of SS from the end of step 2., but
in step 3 we alter the mi by replacing them with the cluster means x̄i . Now note that the
expression
Xni
kxij − mi k2
j=1

is minimised with respect to mi by choosing mi = x̄i , and hence step 3 has the effect
of minimising SS with respect to the mi conditional on the allocation of observations to
clusters. So step 3 cannot increase the value of SS, and any change in any mi will
correspond to a decrease in the value of SS.
Now assuming that the algorithm has not converged, the algorithm will return to step
2. Again note that there will be a value of SS from the end of step 3. (corresponding to
SSW for the cluster allocation). Step 2. will have the effect of leaving some observations
assigned to the same cluster, and moving some observations from one cluster to another.
Consider an observation x that is moved from cluster i to cluster j. It is moved because
kx − mj k < kx − mi k, and so thinking about the effect of this move on SS, it is clear that
the increase in SS associated with cluster j is less than the decrease in SS associated
with cluster i, and so the overall effect of the move is to decrease SS. This is true of every
move, and so the overall effect of step 2 is to decrease SS.
Since both steps 2. and 3. have the effect of decreasing SS, and SS is bounded below
by zero, the algorithm must converge to a local minimum, and this corresponds to a local
minimum of SSW . However, it must be emphasised that the algorithm is not guaranteed
to converge to a global minimum of SSW , and there is always the possibility that the
algorithm will converge to a local minimum that is very poor compared to the true global
minimum. This is typically dealt with by running the whole algorithm to convergence
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 108

many times with different random starting means and then choosing the run leading to
the smallest value of SSW .

Example: microarray data

We can carry out a k-means clustering of the microarray data using R very simply using
the kmeans() function. In the simplest case, we can do a clustering (with k = 4) using the
command:
km=kmeans(t(nci),4)
The results of the clustering procedure are stored in an object that we have called km,
and we will examine this shortly. Note that additional arguments may be passed into
the kmeans function to influence the way the algorithm is carried out. For example, the
command
km=kmeans(t(nci),4,iter.max=50,nstart=20)
will allow the algorithm to run for up to 50 iterations before terminating, and will be run 20
times with different random starting points, returning the results from the best of the 20
runs (in terms of smallest SSW ). The following R session shows how to access the results
of the clustering procedure.
> km$tot.withinss
[1] 200105.4
> km$betweenss
[1] 67757.05
> km$totss
[1] 267862.4
> km$cluster
CNS CNS CNS RENAL BREAST
4 4 4 4 4
CNS CNS BREAST NSCLC NSCLC
4 4 4 4 4
RENAL RENAL RENAL RENAL RENAL
4 4 4 4 4
RENAL RENAL BREAST NSCLC RENAL
4 4 4 4 4
UNKNOWN OVARIAN MELANOMA PROSTATE OVARIAN
4 4 4 1 1
OVARIAN OVARIAN OVARIAN OVARIAN PROSTATE
4 1 4 4 4
NSCLC NSCLC NSCLC LEUKEMIA K562B-repro
4 4 4 2 2
K562A-repro LEUKEMIA LEUKEMIA LEUKEMIA LEUKEMIA
2 2 2 2 2
LEUKEMIA COLON COLON COLON COLON
2 1 1 1 1
COLON COLON COLON MCF7A-repro BREAST
1 1 1 1 1
MCF7D-repro BREAST NSCLC NSCLC NSCLC
1 1 1 1 1
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 109

NSCLC

20
NSCLC COLON
OVARIAN OVARIAN
RENAL COLON
RENAL RENAL
RENAL
OVARIAN
PROSTATE
OVARIAN
NSCLC
OVARIAN
● ● COLON
●COLON MCF7A−repro
COLON
MCF7D−repro
BREAST ●
UNKNOWN BREAST
RENAL
NSCLC
RENAL NSCLC
BREAST
●
OVARIAN
COLON ● ● ● ● ●
PROSTATE
RENAL
BREAST ● ● COLON
NSCLC
CNS
CNS
RENAL
NSCLC ● ●
NSCLC NSCLC ●
CNS CNS
BREAST CNS MELANOMA
● ●
0
K562B−repro
K562A−repro
LEUKEMIA
pca$x[, 2]

LEUKEMIALEUKEMIA
LEUKEMIA LEUKEMIA
LEUKEMIA
RENAL
−20

MELANOMA
MELANOMA
MELANOMA
MELANOMA
−40

MELANOMA
MELANOMA MELANOMA
BREAST

BREAST

−40 −20 0 20 40 60

pca$x[, 1]

Figure 4.1: k-means clustering (with k = 4) overlaid on a scatterplot of the first two principal
components of the nci microarray data

MELANOMA BREAST BREAST MELANOMA MELANOMA

3 3 3 3 3
MELANOMA MELANOMA MELANOMA MELANOMA
3 3 3 3
Really it would be nice to be able to visualise the results of the clustering in some way.
Obviously this is not completely trivial, due to the very high dimension of the observations.
However, we saw in Chapter 2, Figure 2.4, that projecting down to the first two principal
components provides for a useful visualisation of the relationship between the samples, so
we can examine the results of our clustering in this context using the following commands,
pca=prcomp(t(nci))
plot(pca$x[,1],pca$x[,2],pch=km$cluster,col=km$cluster)
text(pca$x[,1],pca$x[,2],colnames(nci),cex=0.3,pos=3)
leading to the plot shown in Figure 4.1.

Choice of k
So far we haven’t discussed the choice of k. In some applications it will be clear from the
context exactly what k is most appropriate, and in many others, it will be clear “roughly”
what values of k should be considered. But in many cases a particular choice of k will
not be given a priori, and some indication from the data of what value might be most
appropriate can be very useful.
In practice, the choice of k is handled by running the k-means algorithm for several
values of k in the range of interest, and examining the way in which SSW decreases as k
increases (it should be clear that the minimum SSW should decrease monotonically with
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 110

260000
220000
SS_W

180000

1 2 3 4 5 6 7 8

Figure 4.2: SSW against k for k-means clustering of the nci microarray data

increasing k). Ideally we look for some kind of “kink”, or at least, levelling off of the sum
of squares as k increases, indicating diminishing returns from increasing k.

Example: choosing k for the microarray data

We can examine the behaviour of the k-means algorithm for varying values of k using the
following fragment of R code,
kmax=8
wss=numeric(kmax)
for (i in 1:kmax) {
km=kmeans(t(nci),i,iter.max=50,nstart=20)
wss[i]=km$tot.withinss
}
plot(1:kmax,wss,type="l",xlab="k",ylab="SS_W",lwd=2,col=2)
leading to the plot shown in Figure 4.2. Unfortunately, but quite typically, there does not
appear to be a very distinctive kink or flattening off of the curve at any point, but our
original choice of k = 4 seems reasonable, as there are diminishing returns for further
increasing k by this point.

Example: handwritten digit images

We can also cluster the zip.train data. Obviously we have a label which classifies the
image, but we can strip this off and run a k-means with 10 clusters to see what happens.
Ideally, in doing this, we would like to see each cluster representing a different digit, as this
would correspond to the different digits occupying distinct regions of R256 . The command
km=kmeans(zip.train[,-1],10,iter.max=30,nstart=10)
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 111

is quite slow to run due to the size of the data set. When it is finished we can examine the
content of the clusters as illustrated in the following R session.
> table(zip.train[,1][km$cluster==1])

0 1 2 3 4 5 6 7 8 9
7 3 46 73 5 71 2 4 433 7
> table(zip.train[,1][km$cluster==2])

1 2 3 4 6 7 8 9
1001 3 1 42 9 3 11 5
> table(Cluster=km$cluster,Digit=zip.train[,1])
Digit
Cluster 0 1 2 3 4 5 6 7 8 9
1 7 3 46 73 5 71 2 4 433 7
2 0 1001 3 1 42 0 9 3 11 5
3 524 0 3 2 0 15 42 0 2 0
4 0 0 13 2 5 3 0 445 1 61
5 512 0 12 1 1 7 29 0 4 0
6 15 0 34 546 0 296 1 0 30 2
7 7 0 565 12 14 9 36 2 10 0
8 10 0 40 8 440 40 10 23 10 125
9 119 0 6 2 6 110 533 0 3 0
10 0 1 9 11 139 5 2 168 38 444
From the results of this analysis, we see that the cluster labelled 1 represents mainly
the digit 8, together will a few other digits, especially 3 and 5, which are fairly similar to
a 8. The cluster labelled 2 mainly represents the digit 1. Cluster 3 consists mainly of
the digit 0, but also contains a number of 6s. Note that the cluster labels are completely
arbitrary, and re-running the analysis will (hopefully) lead to similar clusters, but typically
with a different permutation of the labels. On the basis of this very superficial analysis, it
does appear that the digits do largely occupy different regions of R256 , and so this gives
some encouragement that one might be able to automatically classify digits given just the
images (we will look at classification in more detail in the next chapter).
The following code fragment shows how to look at SSW as a function of k.
kmax=15
wss=numeric(kmax)
for (i in 1:kmax) {
print(i)
km=kmeans(zip.train[,-1],i,iter.max=30,nstart=10)
wss[i]=km$tot.withinss
}
plot(1:kmax,wss,type="l",xlab="k",ylab="SS_W",lwd=2,col=2)
The resulting plot (not shown), perhaps surprisingly, fails to show any especially notice-
able kink or flattening off at k = 10. In fact, this is just symptomatic of the general difficulty
of fixing k purely on the basis of the data. More sophisticated model-based clustering pro-
cedures provide principled methods for choosing k, but in most real data scenarios the
evidence for one value of k over another is often very weak. See section 14.3.6 (p.509)
of [HTF] for further details.
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 112

Pre-transformation, Mahalanobis distance and the QR factorisation

Note that k-means is strongly dependent on the assumption that Euclidean distance is
an appropriate metric for measuring dissimilarity of observations. If that is not the case,
some kind of transformation of the data prior to clustering can be used in order to try
and ensure that Euclidean distance for the transformed data matrix is appropriate. For
example, if Mahalanobis distance is felt to be a more appropriate metric (as it properly
accounts for the correlation structure in the data), then the data can be standardised using
a Mahalanobis transformation prior to clustering. We have seen previously that both the
symmetric square root and Cholesky factor of the sample variance matrix can be used
to compute Mahalanobis distance, and so it will not matter if we standardise using the
symmetric square root or Cholesky factor. Further, we have shown that standardisation
using the Cholesky factor is essentially equivalent to forming the QR factorisation of the
centered data matrix. Consequently, we can cluster based on Mahalanobis distance by
running k-means on the Q matrix of a QR factorisation. For example, we can do this for
the zip.train data as follows.
xbar=colMeans(zip.train[,-1])
W=sweep(zip.train[,-1],2,xbar)
QR=qr(W)
km=kmeans(qr.Q(QR),10,iter.max=30,nstart=10)
Whether or not the resulting clustering will be more appropriate will be very problem de-
pendent. We will come back to this issue later when we look at discrimination and classi-
fication.

4.2.2 Hierarchical clustering

k-means is fast and efficient, and scales well to large data sets. However, it only makes
sense for observations in Euclidean space, and requires specification of the number of
clusters, k. We will now examine an approach to clustering which does not require pre-
specification of the desired number of clusters, and does not require a metric distance
matrix. The idea behind hierarchical clustering, often also known as agglomerative clus-
tering, is to start with each observation in a separate cluster, and then join the two nearest
observations into a cluster containing just two observations. Then the algorithm proceeds
at each stage by combining the clusters that are “closest” together into a new cluster.
Clearly for this to make sense, we need a way of defining the distance between two
clusters, given that we know the distances between all pairs of observations. There are
several different ways that this can be done, and each method leads to a clustering algo-
rithm with different properties. However, provided that we have some way to do this, we
can describe a generic hierarchical clustering algorithm as follows.

1. Start with n clusters, C1 , C2 , . . . , Cn , with Ci containing just the single observation xi .

The distance matrix for the n clusters is just taken to be the distance matrix for the
n observations.

2. Find the minimum distance, dij (if there is more than one minimum, pick one at
random).
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 113

3. Combine clusters Ci and Cj into a single cluster, remove the ith and jth rows and
columns from the distance matrix, and add a new row and column corresponding
to the new cluster by calculating the distances between the new cluster and the
remaining clusters.

4. Note that the number of clusters has decreased by one.

5. If more than one cluster remains, return to step 2.

To put this algorithm into practice, we need ways of computing the distance between
clusters. One of the most commonly used methods, and arguably the most intuitive, is to
define the distance between two clusters to be the minimum distance between observa-
tions in the clusters, that is, for clusters A and B we define

dAB = min dij .

i∈A,j∈B

This definition leads to the so-called single-linkage clustering method, often known as
nearest-neighbour clustering. Alternatively, we can define the distance between clusters
to be the maximum distance between observations in the clusters, that is

dAB = max dij .

i∈A,j∈B

This definition leads to the complete-linkage clustering method, often known as furthest-
neighbour clustering. Both of these methods have the property that they are invariant to
any monotone transformation of the distance matrix. A further possibility (which does not
have this monotonicity property), is to use group average clustering, where
1 XX
dAB = dij .
nA nB i∈A j∈B

Example: toy problem

Let us begin by working through single-linkage and complete-linkage clustering by hand
for 5 observations, A, B, C, D, E having distance matrix
 
0
7 0 
 
D = 4 1 0
 .

6 4 6 0 
8 9 3 2 0

Let us begin by working through the single-linkage algorithm.

It is helpful to write out the distance matrix in the form of a

CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 114

table, as
A 0
B 7 0
C 4 1 0
D 6 4 6 0
E 8 9 3 2 0
A B C D E
The minimum (off-diagonal) distance is clearly d(B, C) = 1,
and so we eliminate the rows and columns for B and C, then
add a new final row for the new cluster {B, C} as follows:
A 0
D 6 0
E 8 2 0
{B, C} 4 4 3 0
A D E {B, C}
Here the new distances have been calculated using the single-
linkage minimum distance rule. Looking at the new table, we
see that the smallest off-diagonal distance is d(D, E) = 2, and
so we remove D and E and add the new row {D, E} as fol-
lows:
A 0
{B, C} 4 0
{D, E} 6 3 0
A {B, C} {D, E}
Inspecting this new table, we see that the minimum distance
is given by d({B, C}, {D, E}) = 3, and this leads to the final
table
A 0
{B, C, D, E} 4 0
A {B, C, D, E}
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 115

The final step combines the remaining two clusters into a sin-
gle cluster consisting of all observations.
We can display the results of a clustering process in a tree diagram known as a den-
drogram by successively joining each pair of clusters in the order they were merged in the
clustering algorithm, and using the vertical axis to mark the distance between the clusters
at the stage they were merged. The result of this for the single-linkage clustering is shown
in Figure 4.3.
We now re-do the clustering procedure using the complete-linkage method, based on
maximum distance.

We start with the distance table as previously.

A 0
B 7 0
C 4 1 0
D 6 4 6 0
E 8 9 3 2 0
A B C D E
The minimum (off-diagonal) distance is clearly d(B, C) = 1,
just as before, and so we eliminate the rows and columns for
B and C, then add a new final row for the new cluster {B, C}
as follows:
A 0
D 6 0
E 8 2 0
{B, C} 7 6 9 0
A D E {B, C}
Here the new distances have been calculated using the complete-
linkage maximum distance rule, and hence differ from those
obtained previously. The minimum distance in this new table
is d(D, E) = 2, and so we merge D and E to get the new table
A 0
{B, C} 7 0
{D, E} 8 9 0
A {B, C} {D, E}
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 116

In this new table, the minimum distance is d(A, {B, C}) = 7,

and so we merge those to get
{D, E} 0
{A, B, C} 9 0
{D, E} {A, B, C}
Again, the final step is to merge the final two clusters.
We can plot the results of this clustering in the dendrogram shown in Figure 4.4. Note
that this dendrogram is different to that for single-linkage clustering. We can now repeat
this analysis using R. We begin by constructing a distance matrix, as follows.
> dm=matrix(c(0,7,4,6,8,7,0,1,4,9,4,1,0,6,3,6,4,6,0,2,8,9,3,2,0),ncol
=5)
> colnames(dm)=c("A","B","C","D","E")
> d=as.dist(dm)
> d
A B C D
B 7
C 4 1
D 6 4 6
E 8 9 3 2
We can then use the hclust() function to carry out hierarchical clustering and plot the
result as a dendrogram using
hc=hclust(d,method="single")
plot(hc,xlab="single-linkage clustering")
leading to the plot shown in Figure 4.3. Note the use of method="single" to force the
use of single-linkage clustering. See ?hclust for other algorithms which can be used.
We can re-do the analysis using complete-linkage clustering with
hc=hclust(d)
plot(hc,xlab="complete-linkage clustering")
giving the plot shown in Figure 4.4, since the method used by hclust() defaults to
complete.
To obtain an actual clustering, we need to decide at some point to stop merging clus-
ters. In principle we could do this by terminating the clustering algorithm at some point. In
practice however, the algorithm is usually run to completion, and the number of clusters
is decided on post-hoc, usually after inspecting the dendrogram. Choosing a number of
clusters corresponds to “cutting” the dendrogram “tree” at some height, and the R com-
mand which does this is cutree(). The tree can be cut by specifying a desired number
of clusters, as in
> ct=cutree(hc,3)
> ct
A B C D E
1 2 2 3 3
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 117

Cluster Dendrogram
4.0

A
3.0
Height

2.0

E
1.0

C
single−linkage clustering
hclust (*, "single")

Figure 4.3: Dendrogram for a single-linkage hierarchical clustering algorithm

Cluster Dendrogram
8
6

A
Height

4
2

E
0

complete−linkage clustering
hclust (*, "complete")

Figure 4.4: Dendrogram for a complete-linkage hierarchical clustering algorithm

CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 118

Cluster Dendrogram

90

LEUKEMIA
NSCLC
70

LEUKEMIA
LEUKEMIA
NSCLC
BREAST
BREAST
RENAL
NSCLC

BREAST
CNS
CNS
OVARIAN
NSCLC
OVARIAN
RENAL
MELANOMA
Height

OVARIAN
COLON

COLON
NSCLC
MELANOMA

BREAST

COLON
MELANOMA

CNS
OVARIAN

NSCLC
OVARIAN
PROSTATE
NSCLC
MELANOMA
MELANOMA

RENAL
PROSTATE
COLON

RENAL
RENAL
COLON
NSCLC
NSCLC
LEUKEMIA
LEUKEMIA
50

RENAL
COLON
COLON

RENAL
RENAL
RENAL
MELANOMA
MELANOMA
MELANOMA

CNS
CNS
LEUKEMIA
K562B−repro
K562A−repro

UNKNOWN
OVARIAN
30

BREAST
BREAST

MCF7A−repro
BREAST
MCF7D−repro
dist(t(nci))
hclust (*, "single")

Figure 4.5: Dendrogram for a single-linkage hierarchical clustering algorithm applied to the nci
microarray data

to obtain 3 clusters. Alternatively, the clustering can be obtained by specifying the height
at which the tree is to be cut, for example
ct=cutree(hc,h=2.5)
Just as for k-means, ad hoc methods are used to justify the number of clusters. Often
people try to look for a “break” in the tree, where there is a big jump in distance before the
next merging, but very often no such obvious gap is present in practice.

Example: microarray data

We can easily apply our clustering methods to the nci microarray data. We begin with
single-linkage
hc=hclust(dist(t(nci)),method="single")
plot(hc,cex=0.4)
leading to the plot shown in Figure 4.5. Alternatively, we can use complete-linkage clus-
tering with
hc=hclust(dist(t(nci)))
plot(hc,cex=0.4)
to get the plot shown in Figure 4.6, or average-linkage clustering with
hc=hclust(dist(t(nci)),method="average")
plot(hc,cex=0.4)
to get the plot shown in Figure 4.7.
Comparing the figures, we see that single-linkage methods are prone to chaining to-
gether nearby observations one-by-one (the algorithm is essentially the algorithm for find-
ing a minimal spanning tree), and can lead to long, thin clusters. The complete-linkage
 Height Height

microarray data
30 50 70 90 20 60 100 140

nci microarray data

LEUKEMIA BREAST
LEUKEMIA MCF7A−repro
LEUKEMIA BREAST
LEUKEMIA MCF7D−repro
K562B−repro COLON
K562A−repro COLON
LEUKEMIA COLON
LEUKEMIA COLON
MELANOMA COLON
BREAST COLON
BREAST COLON
MELANOMA LEUKEMIA
MELANOMA LEUKEMIA
MELANOMA LEUKEMIA
MELANOMA LEUKEMIA
MELANOMA K562B−repro
MELANOMA K562A−repro
RENAL LEUKEMIA
BREAST LEUKEMIA
NSCLC NSCLC
OVARIAN RENAL
OVARIAN BREAST
UNKNOWN NSCLC
OVARIAN NSCLC
NSCLC MELANOMA
MELANOMA MELANOMA
RENAL MELANOMA
RENAL MELANOMA
RENAL MELANOMA
RENAL BREAST
RENAL BREAST
RENAL MELANOMA
RENAL MELANOMA
NSCLC RENAL
OVARIAN UNKNOWN
OVARIAN OVARIAN
NSCLC BREAST

dist(t(nci))
dist(t(nci))
NSCLC NSCLC
NSCLC CNS
PROSTATE CNS
OVARIAN CNS
PROSTATE CNS

hclust (*, "average")

RENAL BREAST
hclust (*, "complete")

CNS OVARIAN
CNS OVARIAN

Cluster Dendrogram
Cluster Dendrogram

CNS RENAL
CNS RENAL
CNS RENAL
BREAST RENAL
NSCLC RENAL
NSCLC RENAL
BREAST RENAL
MCF7A−repro OVARIAN
BREAST OVARIAN
MCF7D−repro NSCLC
COLON NSCLC
COLON NSCLC
COLON NSCLC
COLON MELANOMA
COLON CNS
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING

COLON NSCLC
COLON PROSTATE
BREAST OVARIAN
NSCLC PROSTATE

Figure 4.7: Dendrogram for a average-linkage hierarchical clustering algorithm applied to the nci
Figure 4.6: Dendrogram for a complete-linkage hierarchical clustering algorithm applied to the
119
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 120

NSCLC

20
NSCLC COLON
OVARIAN OVARIAN
RENAL COLON
RENAL RENAL OVARIAN OVARIAN COLON
● ● PROSTATE ● COLON MCF7A−repro
COLON
MCF7D−repro
● ●
RENAL OVARIAN
NSCLC
● RENAL
NSCLC
RENAL
UNKNOWN
NSCLC ● ● ●
OVARIAN BREAST
BREAST

RENAL ● ●
● ●
BREAST ● ● ● ●
BREAST PROSTATE
●
COLON
COLON

●
NSCLC
CNS ●
RENAL
●
CNS
CNS ●CNS ●MELANOMANSCLC
● CNS
NSCLC NSCLC
BREAST ●● ● ●
● ●
●●
0
K562B−repro
● LEUKEMIA
K562A−repro
pca$x[, 2]

LEUKEMIALEUKEMIA
LEUKEMIA LEUKEMIA
LEUKEMIA
RENAL
−20

MELANOMA
MELANOMA
MELANOMA
● ●
MELANOMA
●
−40

MELANOMA
●
MELANOMA MELANOMA
BREAST
●● ●
● BREAST

−40 −20 0 20 40 60

pca$x[, 1]

Figure 4.8: Four clusters obtained from a complete-linkage hierarchical clustering algorithm over-
laid on the first two principal components of the nci microarray data

method tends to give fairly spherical clusters, and is generally to be preferred, which is
why it is the default method in R. Other methods tend to try and interpolate between these
two extremes, including the average-linkage method.
To obtain an actual cluster allocation, we can cut the tree at 4 clusters and overlay the
results on the first two principal components, just as we did for the k-means procedure.
hc=hclust(dist(t(nci)))
ct=cutree(hc,4)
pca=prcomp(t(nci))
plot(pca$x[,1],pca$x[,2],pch=ct,col=ct)
text(pca$x[,1],pca$x[,2],colnames(nci),cex=0.3,pos=3)
This gives rise to the plot shown in Figure 4.8. The clustering perhaps looks unexpected,
but is difficult to interpret, since the clustering was carried out on (a distance matrix de-
rived from) the full observation vectors, but we are looking at the clusters on a simple 2d
projection.
Hierarchical clustering can be used to order a set of observations in such a way that
nearby observations are typically more closely related than observations that are far away.
This is exploited by the heatmap() function that we examined briefly in Chapter 1. Con-
sider now the following command
heatmap(nci,Rowv=NA,labRow=NA,col=grey((15:0)/15),cexCol=0.3)
which produces the plot shown in Figure 4.9. The use of the option Rowv=NA suppresses
the clustering of the rows (which is very time consuming and not particularly useful),
but since we have not explicitly suppressed clustering of the columns, this default action
is executed, and ensures that the columns are ordered meaningfully. Inspection of the
classification labels at the bottom of the columns suggests that the clustering agrees well
with the manually obtained diagnoses.
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 121

Figure 4.9: A heatmap of the nci microarray data with columns ordered according to a hierarchi-
cal clustering

Example: handwritten digits

In principle, we can use hierarchical clustering to cluster the handwritten digits into 10
clusters using the following R commands.
hc=hclust(dist(zip.train[,-1]))
ct=cutree(hc,10)
However, hierarchical clustering does not scale well to large n scenarios such as this, and
the above commands will take a very long time to complete on all but the most powerful
computers. This should be contrasted with the microarray data (large p, but small n),
which clustered fine. Nevertheless, we can run the above commands and eventually we
will get a clustering, which we can examine similarly to previously as illustrated below.
> table(Cluster=ct,Digit=zip.train[,1])
Digit
Cluster 0 1 2 3 4 5 6 7 8 9
1 21 2 33 78 104 387 451 108 23 193
2 56 1001 110 11 176 24 84 6 166 23
3 2 0 25 28 286 71 0 506 264 411
4 256 0 93 303 0 24 8 0 19 0
5 3 2 285 229 7 13 0 1 52 0
6 361 0 6 1 4 10 116 0 5 0
7 1 0 175 0 29 0 0 1 3 1
8 80 0 0 4 0 25 4 0 10 0
9 413 0 1 0 0 2 0 0 0 0
10 1 0 3 4 46 0 1 23 0 16
Here cluster 1 consists mainly of the digits 5 and 6, cluster 2 represents the digit 1, cluster
3 is a mixture of digits 4, 7, 8 and 9, and cluster 4 is a mix of 0 and 3. Again, there is not
CHAPTER 4. CLUSTER ANALYSIS AND UNSUPERVISED LEARNING 122

a strong separation of digits in the clusters.

See section 14.3.12 (p.520) of [HTF] for further details of hierarchical clustering meth-
ods.

4.2.3 Model-based clustering

The clustering methods we have discussed so far (k-means and hierarchical clustering)
are both just heuristic data analysis techniques. However useful they are for data ex-
ploration, it is difficult to use them to make strong statements about clustering in the
population from which they were sampled without making some modelling assumptions.
The standard modelling assumption is that each cluster has a multivariate normal distri-
bution, but that the different clusters have different MVN distributions. It is then possible to
use maximum likelihood or Bayesian methods in order to fit these models to the available
data. Such techniques are beyond the scope of this course, but are becoming increas-
ingly widely used in modern data analysis. Such methods can be found in the CRAN
package mclust, which can be installed using install.packages("mclust") if it is not
already installed. The package can be loaded with require(mclust) and help can be
obtained with ?Mclust. Note that the methods in this package typically do not scale well
to very large data sets.
In summary, we have looked in some detail at k-means clustering and at hierarchical
clustering. The advantage of k-means clustering is that it exploits the geometry of Eu-
clidean space (when this makes sense), and that it scales well to very large data sets
(large n). The advantages of hierarchical clustering include that it does not require a
“real” metric distance matrix, and that it provides very rich information about the cluster-
ing structure in the form of a dendrogram. We have not looked in detail at model-based
clustering methods, but in general these work very well (when the modelling assumptions
are reasonable), and allow much stronger inferences to be made about the population
from which the data is sampled, but typically do not scale well to large data scenarios
(large n or large p).
Chapter 5

Discrimination and classification

5.1 Introduction
In the previous chapter we considered the rather ambitious unsupervised learning task of
automatically grouping observations together into clusters of “similar” observations. Often
however, we have some a priori idea of what kinds of groups we expect to see (or at
least, some examples of observations that have already been classified into pre-specified
groups), and we want to use our current understanding of the group structure in order
to automatically classify new observations which have not yet been assigned to a group.
This is the statistical classification problem, and relies on partitioning the sample space
so as to be able to discriminate between the different groups. Data that has already been
classified into groups is often known as a training set, and methods which make use of
such data are often known as supervised learning algorithms.

5.2 Heuristic classifiers

5.2.1 Closest group mean classifier
Let us start by assuming that our p-dimensional observations belong to one of k groups,
and that there is a mean vector associated with each group, so the group means are
µ1 , µ2 , . . . , µk . Now suppose that we are given an observation x ∈ Rp . Which group
should we allocate the observation to? One obvious method would be to allocate the
observation to the group with mean that is closest to x. That is, we could allocate x to
group i if
kx − µi k < kx − µj k, ∀j 6= i.
This is a very simple example of a classification rule. Note that it has the effect of parti-
tioning Rp into k regions, R1 , R2 , . . . , Rk , where Ri ⊆ Rp , i = 1, 2, . . . , k, Ri ∩Rj = ∅, ∀i 6= j
and ki=1 Ri = Rp in such a way that x is assigned to group i if x ∈ Ri . Different classifica-
S
tion rules lead to different partitions, and clearly some methods of choosing the partition
will be more effective than others for ensuring that most observations will be assigned to
the correct group. In this case, the partition is known as the Voronoi tessellation of Rp gen-
erated by the “seeds” µ1 , µ2 , . . . , µk . The boundaries between the classes are piecewise
linear, and to see why, we will begin by considering the case k = 2.

123
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 124

Binary classification
The case k = 2 is of particular interest, as it represents a very commonly encountered
example in practice, and is known as binary classification. In this case one is typically
interested in deciding whether a particular binary variable of interest is “true” or “false”.
Examples include deciding whether or not a patient has a particular disease, or deciding
if a manufactured item should be flagged as potentially faulty.
In this binary case we will have 2 group means, which we can label µ1 and µ2 . Under
our simple classification rule we allocate x to group 1 provided that kx − µ1 k < kx − µ2 k,
otherwise we allocate to group 2. But then

kx − µ1k < kx − µ2k

⇔ kx − µ1k2 < kx − µ2k2
⇔ (x − µ1)T(x − µ1) < (x − µ2)T(x − µ2)
⇔ xTx − 2µ1Tx + µ1Tµ1 < xTx − 2µ2Tx + µ2Tµ2
⇔ 2(µ2 − µ1)Tx < µ2Tµ2 − µ1Tµ1
1
⇔ (µ2 − µ2)Tx < (µ2 − µ1)T(µ2 + µ1)
  2
1
⇔ (µ2 − µ1)T x − (µ1 + µ2) < 0
2
 
1
⇔ (µ1 − µ2)T x − (µ1 + µ2) > 0.
2
Note carefully how the quadratic term xT x cancels out to leave a discrimination rule
that is linear in x. If we think about the boundary between the two classes, this is clearly
given by solutions to the equation
 
T 1
(µ1 − µ2 ) x − (µ1 + µ2 ) = 0.
2
The first thing to note is that this boundary passes through the midpoint of the group
means, 21 (µ1 + µ2 ), and the next thing to note is that it represents a (shifted) hyperplane∗
orthogonal to the vector µ1 −µ2 . That is, the boundary is the separating hyperplane which
perpendicularly bisects µ1 and µ2 at their mid-point.
For k > 2, it is clear that x will be allocated to group i if
 
T 1
(µi − µj ) x − (µi + µj ) > 0, ∀j 6= i.
2
∗
Because of the shift, the boundary is technically an affine set and not a hyperplane, which should pass through the
origin. However, most people ignore this distinction and refer to the boundary as a hyperplane.
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 125

It is then clear that the region Ri will be an intersection of half-spaces, and hence a convex
polytope.

Example: Toy problem

To illustrate the basic idea, we will illustrate it using an example with k = 2 and p =
2. Suppose that the group means are µ1 = (−1, 2)T and µ2 = (2, 1)T . What is the
classification rule?

We first compute the mid-point

 
1 0.5
(µ + µ2) = ,
2 1 1.5
and the difference between the means
 
−3
µ1 − µ2 = .
1
Then we allocate to group 1 if
   
x1 0.5
(−3, 1) − >0
x2 1.5
⇒ −3x1 + x2 + 1.5 − 1.5 > 0
⇒ x2 > 3x1.
Therefore the boundary is given by the line x2 = 3x1 . Observations lying above the
line will be allocated to group 1, and those falling below will be allocated to group 2.

5.2.2 Linear discriminant analysis (LDA)

The closest group mean classifier is a simple and natural way to discriminate between
groups, and is clearly very closely related to the method of k-means clustering. However,
like (basic) k-means clustering, it ignores the covariance structure in the data, including
the fact that some variables are more variable than others. The more variable (and more
highly correlated) variables are likely to dominate the Euclidean distance, and hence will
have a disproportionate effect on the classification rule. We could correct for this by first
applying a standardisation transformation to the data and the group means, and then carry
out closest group mean classification, or we can directly adapt our rule to take covariance
structure into account. The simplest case of this arises when the k groups have different
means µ1 , µ2 , . . . , µk , but common variance matrix Σ. In this case we can standardise
using (say) the symmetric square root Σ−1/2 , and allocate x to group i if

kΣ−1/2 x − Σ−1/2 µi k < kΣ−1/2 x − Σ−1/2 µj k, ∀j 6= i.

CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 126

Note that we are implicitly relying on linearity of expectation here to appropriately trans-
form the group means. We can write this rule differently by noting that

kΣ−1/2 x − Σ−1/2 µi k < kΣ−1/2 x − Σ−1/2 µj k, ∀j 6= i

−1/2 2 −1/2 2
⇔ kΣ (x − µi )k < kΣ (x − µj )k , ∀j 6= i
T −1 T −1
⇔ (x − µi ) Σ (x − µi ) < (x − µj ) Σ (x − µj ), ∀j 6= i.

That is, we will allocate to the group whose mean is closest when measured according to
Mahalanobis distance. Again, this rule simplifies considerably as

(x − µi)TΣ−1(x − µi) < (x − µj )TΣ−1(x − µj )

⇔ xTΣ−1x − 2µiTΣ−1x + µiTΣ−1µi < xTΣ−1x − 2µj TΣ−1x + µj TΣ−1µ
⇔ 2(µj − µi)TΣ−1x < (µj − µi)TΣ−1(µj + µi)
1
⇔ (µj − µi)TΣ−1x − (µj − µi)TΣ−1(µj + µi) < 0
 2 
1
⇔ (µj − µi)TΣ−1 x − (µi + µj ) < 0
2
 
1
⇔ (µi − µj )TΣ−1 x − (µi + µj ) > 0.
2
Again note how the quadratic term, here xT Σ−1 x, cancels on both sides to give a linear
discrimination rule. Again region Ri is a convex polytope corresponding to an intersection
of half-spaces. In the binary case, the boundary between groups 1 and 2 is given by
 
T −1 1
(µ1 − µ2 ) Σ x − (µ1 + µ2 ) = 0,
2

which again passes through the mid-point 21 (µ1 + µ2 ), but is now the hyperplane orthog-
onal to Σ−1 (µ1 − µ2 ). This classification rule is known as Fisher’s linear discriminant, and
is one of the most fundamental results in classification theory.

Example: Toy problem

Let us reconsider the toy example introduced previously, but now assume a common
variance of  
1 1
Σ= .
1 4
What is the linear discriminant in this case?
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 127

The mid-point is as before, but the normal is now given by

    
1 4 −1 −3 1 −13
Σ−1(µ1 − µ2) = = .
3 −1 1 1 3 4
Therefore we allocate to group 1 if
   
1 x1 0.5
(−13, 4) − >0
3 x2 1.5
⇒ −13x1 + 4x2 + 0.5 > 0
13 1
⇒ x2 > x1 − .
4 8
So the boundary is now given by the line x2 = 13x1 /4 − 1/8, which is very similar,
but different, to the boundary we computed by ignoring the variance structure. The 2
boundaries, together with some simulated data, are shown in Figure 5.1.

Computational considerations
Computationally, it is often most efficient to Mahalanobis transform the data and then carry
out nearest group mean classification. This can be accomplished using the Cholesky
factorisation of Σ.
Note that in the high-dimensional case, p >> k, there is an additional simplification
which can arise, due to the fact that the k group means must all lie in a (k −1)-dimensional
subspace of Rp . Since the rule is to allocate to the closest group mean, all data may be
projected down to this (k −1)-dimensional space, and the classification may be carried out
in Rk−1 , since any components orthogonal to this subspace contribute the same distance
to each of the group means, and therefore cancel out of the classification rule. When
k << p this dimension reduction can result in significant computational savings.
The natural basis vectors for this (k − 1)-dimensional space are known as discrimina-
tion coordinates, or crimcoords, but their derivation and use is beyond the scope of this
course.

5.2.3 Quadratic discrimination

When the variance matrices in the groups are unequal, so that the group means are
µ1 , . . . , µk and the variance matrices are Σ1 , . . . , Σk , an obvious first attempt at adapting
the previous classification method would be to compare Mahalanobis distances for each
group using a distance appropriate to each group. That is, allocate observation x to group
i if
(x − µi )T Σi −1 (x − µi ) < (x − µj )T Σj −1 (x − µj ), ∀j 6= i.
Note that now the quadratic terms will not cancel, and so we end up with quadratic bound-
aries. In particular, in the binary case, we will allocate x to group 1 if
(x − µ1 )T Σ1 −1 (x − µ1 ) < (x − µ2 )T Σ2 −1 (x − µ2 ),
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 128

and so the boundary between the two groups will correspond to a contour of a quadratic
form.
Note that although the above approach to constructing a quadratic discrimination rule
is attractive, it does not penalise groups with large variances sufficiently. We will see later
how to improve on this quadratic rule by attempting to take a more principled model-based
approach.

5.2.4 Discrimination functions

We have now looked at several different methods for discrimination and classification,
each leading to a different rule. However, all of the rules we have considered so far can
be described in within a common framework by introducing the concept of a discriminant
function. For each class i = 1, 2, . . . , k, we define a corresponding function
Qi (·) : Rp −→ R,
known as a discriminant function which determines a partition of Rp , R1 , R2 , . . . , Rk , by
assigning an observation x to Ri if
Qi (x) > Qj (x), ∀j 6= i.
Note the use of a > inequality rather than a < inequality, so instead of a measure of dis-
tance or dissimilarity, the discriminant functions represent the likelihood or propensity of
an observation to belong to a particular group. This turns out to be more natural and con-
venient, especially in the context of model-based methods for classification. Our distance-
based methods can be easily viewed within this framework by introducing a minus sign.
Thus, our closest group mean classifier corresponds to the discriminant functions
Qi (x) = −kx − µi k2 , i = 1, 2, . . . , k,
and the LDA method corresponds to
Qi (x) = −(x − µi )T Σ−1 (x − µi ), i = 1, 2, . . . , k.
Our first attempt at a quadratic classification rule corresponds to the discriminant func-
tions,
Qi (x) = −(x − µi )T Σi −1 (x − µi ), i = 1, 2, . . . , k,
but as previously mentioned, we will improve on this in due course.
It turns out that most classification methods of practical interest can be cast into the
framework of discriminant functions, and having a common framework can be useful in
order to develop generic techniques for the computation, interpretation and analysis of
classification methods and procedures that are not specific to a particular discrimination
technique.
There are many different approaches that can be taken to classification. Many, such
as those we have been considering so far, are heuristically derived. However, given
a particular set of discriminant functions, we can study the properties of the resulting
classifier, such as misclassification rates, either empirically, or theoretically, to try and
establish whether the method is good or bad.
Other approaches to classification start from modelling assumptions, and derive clas-
sification rules which are “optimal” in some narrowly defined sense. We will next examine
one such approach, based on the assumption of multivariate normality and the principle
of maximum likelihood.
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 129

5.3 Maximum likelihood discrimination

Model-based approaches to classification assume a probability model for each group, of
the form fi (·) : Rp −→ R, i = 1, 2, . . . , k. So for observations x ∈ Rp , each model fi (x)
represents a probability density function (or likelihood) for the random variable X from
group i. The maximum likelihood discriminant rule (MLDR) is to classify observations
by assigning them to the group with the maximum likelihood. In other words, MLDR
corresponds to using the discriminant functions

Qi (x) = fi (x), i = 1, 2, . . . , k.

5.3.1 LDA
The simplest case of maximum likelihood discrimination arises in the case where it is as-
sumed that observations from all groups are multivariate normal, and all share a common
variance matrix, Σ. That is observations from group i are assumed to be iid N (µi , Σ)
random variables. In other words,
 
−p/2 −1/2 1 T −1
Qi (x) = fi (x) = (2π) |Σ| exp − (x − µi ) Σ (x − µi ) .
2

We can simplify things by noting that

Qi (x) > Qj (x)

 
−p/2 −1/2 1
⇔ (2π) |Σ| exp − (x − µi ) Σ (x − µi ) > (2π)−p/2 |Σ|−1/2
T −1
2
 
1 T −1
× exp − (x − µj ) Σ (x − µj )
2
   
1 T −1 1 T −1
⇔ exp − (x − µi ) Σ (x − µi ) > exp − (x − µj ) Σ (x − µj )
2 2
⇔ −(x − µi ) Σ (x − µi ) > −(x − µj )T Σ−1 (x − µj ),
T −1

and so the MLDR is in fact exactly equivalent to using

Qi (x) = −(x − µi )T Σ−1 (x − µi ).

But this is just classification by minimising Mahalanobis distance, and hence corresponds
to Fisher’s linear discriminant, which allocates to group i if
 
T −1 1
(µi − µj ) Σ x − (µi + µj ) > 0, ∀j 6= i.
2

So in the case of equal variance matrices, the MLDR corresponds exactly to Fisher’s
linear discriminant.
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 130

5.3.2 Quadratic discriminant analysis (QDA)

It is natural to next consider how the MLDR changes when we allow the variance matrices
associated with each group to be unequal. That is, we assume that observations from
group i are iid N (µi , Σi ). In this case we have
 
−p/2 −1/2 1 T −1
Qi (x) = fi (x) = (2π) |Σi | exp − (x − µi ) Σi (x − µi ) .
2

We can simplify this a little by noting that

Qi (x) > Qj (x)

 
−p/2 −1/2 1
⇔ (2π) |Σi | exp − (x − µi ) Σi (x − µi ) > (2π)−p/2 |Σj |−1/2
T −1
2
 
1 T −1
× exp − (x − µj ) Σj (x − µj )
2
   
−1/2 1 T −1 −1/2 1 T −1
⇔ |Σi | exp − (x − µi ) Σi (x − µi ) > |Σj | exp − (x − µj ) Σj (x − µj )
2 2
⇔ − log |Σi | − (x − µi ) Σi (x − µi ) > − log |Σj | − (x − µj )T Σj −1 (x − µj ),
T −1

and so the MLDR is exactly equivalent to using the discriminant function

Qi (x) = − log |Σi | − (x − µi )T Σi −1 (x − µi ).

Note that this is a quadratic form in x, but is different to the quadratic discriminant func-
tion we derived heuristically, due to the presence of the log |Σi | term, which has the effect
of appropriately penalising the distances associated with groups having large variances.
This penalty term, which corresponds to the normalisation constant of the density, gen-
erally improves the performance of the classifier, and hence is the form typically used in
QDA.

The binary case

In the k = 2 case, we assign to group 1 if Q1 (x) > Q2 (x), that is

− log |Σ1 | − (x − µ1 )T Σ1 −1 (x − µ1 ) > − log |Σ2 | − (x − µ2 )T Σ2 −1 (x − µ2 )

⇔ log |Σ1 | + (x − µ1 )T Σ1 −1 (x − µ1 ) < log |Σ2 | + (x − µ2 )T Σ2 −1 (x − µ2 ).

Multiplying out and gathering terms gives

|Σ1 |
xT (Σ1 −1 − Σ2 −1 )x + 2(µ2 T Σ2 −1 − µ1 T Σ1 −1 )x + µ1 T Σ1 −1 µ1 − µ2 T Σ2 −1 µ2 + log < 0.
|Σ2 |

Here we can see explicitly that the quadratic term does not cancel out, and that the bound-
ary between the two classes corresponds to the contour of a quadratic form.
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 131

5.3.3 Estimation from data

Typically the group mean vectors and variance matrices are not known in advance, but
can be estimated from data that has already been classified, typically referred to as a
training set. For the MLDR, maximum likelihood parameter estimates should be plugged
into the discriminant functions. So, one can start with the group sample means x̄i and
plug these in for the µi , for both LDA and QDA. The variance matrices are slightly more
complicated. For QDA, the group sample variance matrices, Si can be plugged in for Σi .
Technically, a divisor of ni rather than ni −1 should be used, since it is the MLE of variance
that is required, but it will obviously not make much difference if ni is large. However, if
Σ̂i = (ni − 1)Si /ni is used, the discriminant function becomes

Qi (x) = − log |Σ̂i | − (x − x̄i )T Σ̂i −1 (x − x̄i ).

For LDA, a single pooled estimate of variance is required. The unbiased pooled estimate
of Σ is
k
1 X
SW = (ni − 1)Si ,
n − k i=1
and this can be plugged in for Σ. The corresponding MLE is

n−k
Σ̂ = SW ,
n
but note that here it doesn’t matter what divisor is used, since it will cancel out of Fisher’s
linear discriminant. Also note that it is not appropriate to use the overall sample variance
matrix, S, since the means within the groups are different. The discriminant function
becomes
Qi (x) = −(x − x̄i )T SW −1 (x − x̄i ).
In this case, Fisher’s linear discriminant for k = 2 takes the form, allocate to group 1 if
 
T −1 1
(x̄1 − x̄2 ) SW x − (x̄1 + x̄2 ) > 0.
2

Example: Toy problem

Now that we know how to use data in order to construct discriminant rules, we can see
how to use R to do so. We will first consider simulated data based on the simple toy model
considered previously. We will simulate 20 observations from each group, using the group
means and common variance previously given, and store the results in a 40 × 2 matrix, X,
and a group vector Class.
mu1=c(-1,2)
mu2=c(2,1)
diff=mu1-mu2
ave=0.5*(mu1+mu2)
Sigma=matrix(c(1,1,1,4),ncol=2)
Gt=chol(Sigma)
Z=matrix(rnorm(80),ncol=2)
W=Z %*% Gt
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 132

6
●
● ●
●

4 ●
●
● ● ●●
● ●
●
2
X[,2]

● ● ●
●
● ●
0
−2

−2 −1 0 1 2 3

X[,1]

Figure 5.1: Simulated data with two different classification boundaries overlaid

shift1=matrix(rep(mu1,each=20),ncol=2)
shift2=matrix(rep(mu2,each=20),ncol=2)
shift=rbind(shift1,shift2)
X=W+shift
Class=rep(1:2,each=20)
We can plot the data, and overlay the decision boundaries we computed earlier as follows.
plot(X,pch=Class,col=Class)
abline(0,3,col=2)
abline(-1/8,13/4,col=3)
This gives the plot shown in Figure 5.1.
We can compute the sample means and pooled variance estimate, and then construct
the normal for Fisher’s discriminant rule using:
> xbar1=colMeans(X[Class==1,])
> xbar2=colMeans(X[Class==2,])
> V1=var(X[Class==1,])
> V2=var(X[Class==2,])
> V=(19*V1+19*V2)/(40-2)
> normal=solve(V,xbar1-xbar2)
> normal
[1] -6.41715 1.76622
We can then also use the lda() function from the MASS package to automate the process,
as
> require(MASS)
> X.lda=lda(X,Class,prior=c(0.5,0.5))
> X.lda
Call:
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 133

lda(X, grouping = Class, prior = c(0.5, 0.5))

Prior probabilities of groups:

1 2
0.5 0.5

Group means:
1 2
1 -0.6744606 2.554725
2 2.1881006 1.346963

Coefficients of linear discriminants:

LD1
[1,] 1.4540364
[2,] -0.4002007
Setting the prior option to be uniform over the two classes ensures that we obtain the
MLDR. However, leaving this option out allows the lda() function to estimate the group
frequencies and use this in the classification rule. This turns out to be very helpful in
general — see the discussion about Bayesian classification at the end of the chapter.
By comparing the ratios of elements, it should be clear that the normal computed by
the lda() function is in exactly the same direction as that which we computed explicitly.
The object X.lda can be passed into the predict() function in order to make actual
classifications on new data. For example, to classify the single observation x = (−3, 1),
we can use predict(X.lda,c(-3,1))$class, but we can also pass in a 2-column matrix
and get back a vector of classifications. For example, if we want to classify the data we
used in order to build the discriminant function (in order to see how well the classifier
works), then we can do so using
> predict(X.lda,X)$class
[1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2
[34] 2 2 2 2 2 2 2
Levels: 1 2
We can also use R to carry out QDA for us, using the qda() function (also from the
MASS package).
X.qda=qda(X,Class)
X.qda=qda(X,Class,prior=c(0.5,0.5))
X.qda
The object returned from the qda() function can also be passed into the predict()
function in order to classify new data.

5.4 Misclassification
Obviously, whatever discriminant functions we use, we will not characterise the group
of interest perfectly, and so some future observations will be classified incorrectly. An
obvious way to characterise a classification scheme is by some measure of the degree of
misclassification associated with the scheme. For a fully specified model, we can define
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 134

a k × k matrix P of classification probabilities with elements

pij = P(allocate to group i|observation from group j) .

Ideally we would like this to be Ik , but in practice the best we can hope for is that the
diagonal elements are close to one, and that the off-diagonal elements are close to zero.†
One way of characterising the overall quality of a classification scheme would be using
Tr (P) /k, with larger values of this quality score representing better schemes, and a max-
imum value of 1 indicating a perfect classification scheme.
For simple fully specified models and classification schemes, it may be possible to
compute the pij directly, analytically (see homework assignment). For models where the
parameters have been estimated from data, it may be possible to estimate p̂ij , the prob-
ability obtained from the model by plugging in MLEs for the model parameters. Even
where this is possible, it is likely to lead to over-estimates of the diagonal elements and
under-estimates of the off-diagonal elements, due to over-fitting, and ignoring sampling
variability in the parameter estimates.
Where there is no underlying model, or the model or classification scheme is complex,
it is quite likely that it will not be possible to compute P directly. In this case we can obtain
empirical estimates of the pij directly from data. The plug-in method estimates the pij by
re-using the data originally used to derive the classification rules. So, if we define nij to
be the number of class j observations allocated to class i (so that nj = ki=1 nij ), we can
P
estimate pij by
nij
p̂ij = .
nj
That is, we estimate the probability that a class j observation is classified as class i by
the observed proportion of class j observations that are classified as class i. Although
this method is very simple, and very widely used, it also leads to over-optimistic estimates
of the quality of the classification scheme by testing the classifier on the data used in its
construction (over-fitting).

Example: Toy problem

We can create classification tables in R for the simple examples we considered earlier by
using the table() command.
> tab=table(Predicted=predict(X.lda)$class,Actual=Class)
> tab
Actual
Predicted 1 2
1 20 0
2 0 20
>
> tab=table(Predicted=predict(X.qda)$class,Actual=Class)
> tab
Actual
Predicted 1 2
†
Note that P is a left stochastic matrix, having similar properties (such as columns summing to one) as the transition
matrix of a Markov chain.
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 135

1 20 0
2 0 20
A more reliable way of estimating the performance of a classifier in the case of very
large n is to split the data set (randomly) into two, and then use one half of the data
set to train the classifier and the other half to test the performance of the classifier on
independent data.

Example: handwritten digits

We can use the data in the file zip.train in order to build a linear classifier for the
handwritten digits, and then test this classifier on the data zip.test.
> zip.lda=lda(zip.train[,-1],zip.train[,1],prior=rep(0.1,10))
> tab=table(Predicted=predict(zip.lda,zip.test[,-1])$class,Actual=zip.
test[,1])
> tab
Actual
Predicted 0 1 2 3 4 5 6 7 8 9
0 342 0 7 3 1 4 0 0 5 0
1 0 251 2 0 4 0 0 1 0 0
2 0 0 155 3 6 0 3 0 2 0
3 4 2 4 142 0 17 0 2 11 0
4 3 5 12 3 174 3 3 7 7 4
5 1 0 2 9 0 125 3 0 4 0
6 4 3 1 0 2 0 158 0 0 0
7 0 0 1 1 2 0 0 129 0 5
8 4 1 14 4 1 6 3 1 135 3
9 1 2 0 1 10 5 0 7 2 165
> tab2=t(t(tab)/colSums(tab))
> round(tab2,digits=2)
Actual
Predicted 0 1 2 3 4 5 6 7 8 9
0 0.95 0.00 0.04 0.02 0.00 0.02 0.00 0.00 0.03 0.00
1 0.00 0.95 0.01 0.00 0.02 0.00 0.00 0.01 0.00 0.00
2 0.00 0.00 0.78 0.02 0.03 0.00 0.02 0.00 0.01 0.00
3 0.01 0.01 0.02 0.86 0.00 0.11 0.00 0.01 0.07 0.00
4 0.01 0.02 0.06 0.02 0.87 0.02 0.02 0.05 0.04 0.02
5 0.00 0.00 0.01 0.05 0.00 0.78 0.02 0.00 0.02 0.00
6 0.01 0.01 0.01 0.00 0.01 0.00 0.93 0.00 0.00 0.00
7 0.00 0.00 0.01 0.01 0.01 0.00 0.00 0.88 0.00 0.03
8 0.01 0.00 0.07 0.02 0.00 0.04 0.02 0.01 0.81 0.02
9 0.00 0.01 0.00 0.01 0.05 0.03 0.00 0.05 0.01 0.93
> sum(diag(tab2))/10
[1] 0.8745323
We see that the classification is generally good, but ’2’s are sometimes misclassified as
’4’s or ’8’s, and similarly, ’5’s are sometimes misclassified as ’3’s or ’8’s, etc.
We can repeat the analysis without requiring a uniform prior as follows.
> zip.lda=lda(zip.train[,-1],zip.train[,1])
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 136

> tab=table(Predicted=predict(zip.lda,zip.test[,-1])$class,Actual=zip.
test[,1])
> tab
Actual
Predicted 0 1 2 3 4 5 6 7 8 9
0 342 0 7 3 1 6 1 0 5 0
1 0 251 2 0 4 0 0 1 0 0
2 0 0 157 3 6 0 3 0 2 0
3 4 2 4 142 0 16 0 2 11 0
4 3 5 12 3 174 3 3 7 7 4
5 1 0 2 9 0 125 3 0 4 0
6 5 3 1 0 2 0 157 0 0 0
7 0 0 1 1 2 0 0 129 0 5
8 3 1 12 4 1 5 3 1 135 3
9 1 2 0 1 10 5 0 7 2 165
> tab2=t(t(tab)/colSums(tab))
> round(tab2,digits=2)
Actual
Predicted 0 1 2 3 4 5 6 7 8 9
0 0.95 0.00 0.04 0.02 0.00 0.04 0.01 0.00 0.03 0.00
1 0.00 0.95 0.01 0.00 0.02 0.00 0.00 0.01 0.00 0.00
2 0.00 0.00 0.79 0.02 0.03 0.00 0.02 0.00 0.01 0.00
3 0.01 0.01 0.02 0.86 0.00 0.10 0.00 0.01 0.07 0.00
4 0.01 0.02 0.06 0.02 0.87 0.02 0.02 0.05 0.04 0.02
5 0.00 0.00 0.01 0.05 0.00 0.78 0.02 0.00 0.02 0.00
6 0.01 0.01 0.01 0.00 0.01 0.00 0.92 0.00 0.00 0.00
7 0.00 0.00 0.01 0.01 0.01 0.00 0.00 0.88 0.00 0.03
8 0.01 0.00 0.06 0.02 0.00 0.03 0.02 0.01 0.81 0.02
9 0.00 0.01 0.00 0.01 0.05 0.03 0.00 0.05 0.01 0.93
> sum(diag(tab2))/10
[1] 0.8749542
Here the predictive performance improves only slightly (two additional ’2’s were correctly
classified), due to the fact that the groups are approximately equally likely. However, in
cases where groups are very unbalanced, estimation of prior probabilities can make a
large difference to classification performance.
Using either method we see that we are able to automatically classify around 90% of
digit images correctly using a simple LDA classifier. This is probably not good enough
to be usable in practice, but nevertheless illustrates that simple linear algebraic methods
can be used to solve very high-dimensional classification problems well.

5.5 Bayesian classification

The MLDR is often simple to implement, but only really works well when the k groups
are all a priori equally likely. To see this, consider the binary classification case being
applied to testing for a very rare disease. The MLDR blindly classifies observations to
maximise the probability of the test result given disease status, when what we really want
to maximise is the probability of disease status given the observation. We need to use
Bayes theorem to do this, and this requires knowing the a priori group probabilities, which
CHAPTER 5. DISCRIMINATION AND CLASSIFICATION 137

can also be estimated from data. This has the effect of introducing additional terms into
the discriminant functions. Bayesian classifies typically perform much better than heuristic
methods, or methods based on MLDR, but are beyond the scope of this course.
See section 4.3 (p.106) of [HTF] for further information about techniques for classifi-
cation.
Chapter 6

Graphical modelling

6.1 Introduction
Variables of multivariate distributions and data sets are often correlated in complex ways.
It would be convenient if many variables were independent, leading to a sparse indepen-
dence structure, as this would lead to many simplifications of much multivariate theory
and methodology. Unfortunately variables tend not to be independent in practice. It is
more often the case that variables are conditionally independent. Although not quite as
strong a property as independence, this too turns out to be enough to lead to consider-
able simplification of a general dependence structure, and is useful for both conceptual
and theoretical understanding, and also for computational efficiency. Sparse conditional
independence structures can be represented using graphs, and this leads to the theory
of graphical models. But before we explore graphical models, we must first ensure that
we understand the notion of conditional independence.

6.2 Independence, conditional independence and factorisation

Let us begin by reminding ourselves of the essential notion of independence of events
and random variables. For events A and B, we say that A and B are independent, and
write A ⊥ B, if
P(A ∩ B) = P(A) P(B) .
Note that for P(B) > 0 we get
P(A ∩ B)
P(A|B) = = P(A) ,
P(B)
which captures the intuitive notion of independence, that learning the outcome of B should
not affect the probability of A.
The notion extends straightforwardly to (continuous) random quantities, where now
we say that variables X and Y are independent and write X ⊥ Y if their joint density
factorises as
fX,Y (x, y) = fX (x)fY (y).
Again, for fY (y) > 0 we have
fX,Y (x, y)
fX|Y (x|y) = = fX (x),
fY (y)

138
CHAPTER 6. GRAPHICAL MODELLING 139

capturing the intuitive notion that the conditional distribution of X given Y = y should not
depend on y. That is, the conditional and marginal distributions should be the same.
It turns out that independence is a special case of the more general notion of con-
ditional independence. For events A, B and C, we say that A and B are conditionally
independent given C, and write A⊥⊥B|C, if

P(A ∩ B|C) = P(A|C) P(B|C) .

Now, when P(B|C) > 0 we have

P(A ∩ B|C)
P(A|B ∩ C) = = P(A|C) ,
P(B|C)

and so conditional on C, learning the outcome of B does not change the probability of A.
This notion again generalises easily to (continuous) random quantities, so that for
random variables X, Y and Z, we have that X and Y are conditionally independent given
Z, written X⊥⊥Y |Z if the conditional density of X and Y given Z = z factorises as

fX,Y |Z (x, y|z) = fX|Z (x|z)fY |Z (y|z).

So then if fY |Z > 0 we have

fX,Y |Z (x, y|z)

fX|Y,Z (x|y, z) = = fX|Z (x|z).
fY |Z (y|z)

There are numerous other important properties of this factorisation for the joint density of
X, Y and Z, fX,Y,Z (x, y, z).

Proposition 31 If X⊥⊥Y |Z then the joint density factorises as

fX,Y,Z (x, y, z) = fZ (z)fX|Z (x|z)fY |Z (y|z).

This factorisation property turns out to be key to understanding directed acyclic graph
(DAG) models.
Proof

fX,Y,Z (x, y, z) = fZ (z)fX,Y |Z (x, y|z) = fZ (z)fX|Z (x|z)fY |Z (y|z).



Proposition 32 If X⊥⊥Y |Z and fZ (z) > 0, then the joint density factorises as

fX,Z (x, z)fY,Z (y, z)

fX,Y,Z (x, y, z) = .
fZ (z)
CHAPTER 6. GRAPHICAL MODELLING 140

This factorisation property turns out to be important for understanding undirected graphi-
cal models.
Proof

fX,Y,Z = fZ (z)fX|Z (x|z)fY |Z (y|z)

fX,Z (x, z) fY,Z (y, z)
= fZ (z)
fZ (z) fZ (z)
fX,Z (x, z)fY,Z (y, z)
= .
fZ (z)


Proposition 33 Assuming that fZ (z) > 0, we have that X⊥⊥Y |Z if and only if the joint
density factorises in the form

fX,Y,Z (x, y, z) = h(x, z)k(y, z),

for some bivariate functions h(·, ·) and k(·, ·).

Proof
First note that the forward implication follows from Proposition 32. That is, if X⊥⊥Y |Z,
then there exist functions h(·, ·) and k(·, ·) such that

fX,Y,Z (x, y, z) = h(x, z)k(y, z).

For example, we could choose h(x, z) = fX,Z (x, z) and k(y, z) = fY,Z (y, z)/fZ (z), but there
are obviously many other possibilities.
The reverse direction is less clear. We assume that we have the factorisation

fX,Y,Z (x, y, z) = h(x, z)k(y, z),

and want to prove that

fX,Y |Z (x, y|z) = fX|Z (x|z)fY |Z (y|z).
This can be demonstrated with straightforward algebraic manipulation, but is left as an
exercise. 
Note that although we may have been thinking about univariate random quantities X,
Y and Z, nothing we have discussed is specific to the univariate case, and all applies
similarly to the case of multivariate random quantities X, Y and Z.
The above factorisation results are very powerful, and can be used to prove a number
of general algebraic properties associated with independence and conditional indepen-
dence of random quantities.
CHAPTER 6. GRAPHICAL MODELLING 141

Proposition 34

1. X⊥⊥Y |Z ⇔ Y ⊥⊥X|Z

2. X⊥⊥Y |Z ⇒ h(X)⊥⊥Y |Z

3. X⊥⊥Y |Z ⇒ X⊥⊥Y |(Z, h(X))

4. X⊥⊥Y |Z and X⊥⊥W |(Y, Z) ⇔ X⊥⊥(W, Y )|Z

5. If all densities are positive, then

X⊥⊥Y |Z and X⊥⊥Z|Y ⇒ X ⊥ (Y, Z).

Proof
The proofs of these results are mainly straightforward, and left as an exercise. 

6.3 Undirected graphs

6.3.1 Graph theory
It turns out that when we have a large number of variables with many associated condi-
tional independence statements, it is useful to represent known statements using a graph.
The graphs used can be directed or undirected, and can be given several different inter-
pretations. We begin with a reminder of some basic properties of undirected graphs, and
then later examine how such graphs may be associated with conditional independence
properties.
A graph G is a tuple G = (V, E), where V = {v1 , v2 , . . . , vn } is a finite set of vertices (or
nodes), and E is a finite set of undirected edges, with
n

each edge e ∈ E being of the form
e = {vi , vj }, vi , vj ∈ V , i 6= j. There are clearly 2 = n(n − 1)/2 possible edges of this
form, and we write V2 for the set of all such edges, so that E ⊆ V2 . A graph is said to

be complete if E = V2 .

Example
Consider the graph G = (V, E) where V = {A, B, C} and E = {{A, C}, {B, C}}. We can
draw a pictorial representation of the graph as given in Figure 6.1. Note that this graph is
not complete, since the edge {A, B} is missing. Also note that we can plot this graph in
R using the ggm package, using the following commands
require(ggm)
drawGraph(UG( ∼ A*C+B*C))
Note that the drawGraph() function allows moving nodes around using a very simple
point-and-click interface. The command plotGraph() can also be used, and has a more
sophisticated interface, but may not be available on all systems.
CHAPTER 6. GRAPHICAL MODELLING 142

C●

A● B●

Figure 6.1: A graph with 3 vertices and 2 edges

Vertices vi , vj ∈ V are said to be adjacent, written vi ∼ vj , if {vi , vj } ∈ E. Any subset

of vertices U ⊆ V induces a subgraph GU = (U, F ), where

F = {{u, v} ∈ E|u, v ∈ U }.

Example
For the graph considered in the previous example, nodes A and C are adjacent (A ∼ C),
but nodes A and B are not. Similarly, the subgraph G{A,C} is complete, but the subgraph
G{A,B} is not.

A subset U ⊆ V is a (maximal) clique if it is maximally complete. That is, the associ-

ated subgraph GU is complete, and for all strict supersets of U , W (so that U ⊂ W ⊆ V ),
the induced subgraph GW is not complete.

Example
Again, considering again the previous example, the graph G has two cliques, {A, C} and
{B, C}.

Example
The previous example is somewhat uninteresting in that the cliques correspond with
edges. So consider now the graph G = (V, E) where V = {D, E, F, G} and

E = {{D, E}, {D, F }, {E, F }, {E, G}}.

Draw the graph and write down the cliques. How many cliques are there?
CHAPTER 6. GRAPHICAL MODELLING 143

W● Z●

V●

X● Y●

Figure 6.2: A graph with 5 vertices and 3 cliques

The cliques are {D, E, F } and {E, G}. There are two cliques.
Note that the set of cliques defines the graph. That is, complete knowledge of all
cliques of the graph allows construction of the graph.

Example
Consider the graph G which has three cliques: {V, W, X}, {W, X, Y } and {W, X, Z}. We
can construct and plot this graph in R using
drawGraph(UG( ∼ V*W*X+W*X*Y+W*Y*Z))
leading to the plot shown in Figure 6.2.

The neighbours of a node v ∈ V , often known as the boundary of v, written bd(v),

is {u ∈ V |u ∼ v}. The closure of v, written cl(v), is given by cl(v) = v ∪ bd(v). A
path is a sequence x0 , x1 , . . . , xm ∈ V such that xi−1 ∼ xi , i = 1, 2, . . . , m. A graph
is connected if there exists a path between every pair of vertices. A component is a
maximal connected subgraph. A cycle is a (non-trivial) path starting and ending at the
same vertex. A connected graph is a tree if it does not contain any cycles. A leaf is any
node of a tree connected to just one edge. A forest is a graph with components which are
all trees.
For any undirected graph G = (V, E) with vertex subsets A, B, C ⊆ V , we say that C
separates A and B if all paths from a node in A to a node in B pass through a node in C.

6.3.2 Graphical models

Now, given a collection of conditional independence statements relating to a collection of
random variables, we can define a graph by associating nodes with variables, and using
CHAPTER 6. GRAPHICAL MODELLING 144

the edge structure to encode the conditional independence structure. This is useful for
both visualisation and computational analysis. It turns out that there are several different
ways that we can do this, but they all turn out to be equivalent in practice in most cases.
Given a set of random variables X1 , X2 , . . . , Xn , we can define an associated graph
G = (V, E), where V = {X1 , X2 , . . . , Xn }, and E is a set of edges.

Definition 10 We say that G has the factorisation property (F) if the joint density of the
random variables factorises in the form
Y
fX (x) = φc (xc ),
c∈C

for some functions φc (·), where C denotes the set of cliques associated with the graph G.

Definition 11 We say that G has the global Markov property (G) if for any disjoint vertex
subsets A, B and C such that C separates A and B in G we have A⊥⊥B|C.

Definition 12 We say that G has the local Markov property (L) if for all v ∈ V we have

v⊥⊥V \ cl(v)|bd(v).

Definition 13 We say that G has the pairwise Markov property (P) if for all u, v ∈ V such
that u 6∼ v we have
u⊥⊥v|V \{u, v}.

We have just presented four different ways that one can associate a graph with a con-
ditional independence structure of a set of random variables. If these different interpreta-
tions all turn out to be fundamentally different, this is potentially confusing. Fortunately, it
turns out that these different interpretations are all closely linked.

Proposition 35 (F) ⇒ (G) ⇒ (L) ⇒ (P)

This important result tells us that if a graph satisfies the factorisation property (F), then
the other properties all follow.
Proof
Let’s start with the simplest result, (G) ⇒ (L): This is trivial, since by definition, bd(v)
separates v from V \ cl(v).
Now let us consider (L) ⇒ (P):

Suppose (L), so that we know

v⊥⊥V \ cl(v)|bd(v).
Now using property 3 of Proposition 34, we can “copy” ver-
tices from V \ cl(v) to the conditioning set as desired in order
to deduce that
v⊥⊥V \ cl(v)|V \{u, v}.
CHAPTER 6. GRAPHICAL MODELLING 145

Finally, since u and v are not adjacent, we know that u ∈

V \ cl(v), and so using property 2 of Proposition 34, we con-
clude that
v⊥⊥u|V \{u, v},
and (P) is satisfied.
Finally, let us consider (F) ⇒ (G): Assume (F), and suppose that A, B and C are
disjoint subsets of V with C separating A and B. The graph GV \C can’t be connected, so
let à be the components of GV \C containing vertices in A, so that A ⊆ à and B ∩ à = ∅
due to the separation assumption. Now put B̃ = V \(Ã ∪ C) so that B ⊆ B̃, and V is the
disjoint union of Ã, B̃ and C. Again, due to the separation assumption, every c ∈ C must
be such that c ⊆ à ∪ C or c ⊆ B̃ ∪ C. Call the set of all cliques in à ∪ C, CA , and the rest
CB . Using (F), we have
Y Y Y
f (x) = φc (x) = φc (x) φc (x) = h(xÃ∪C )k(xB̃∪C ),
c∈C c∈CA c∈CB

and so Ã⊥⊥B̃|C. Now using property 2 of Proposition 34 we conclude first that A⊥⊥B̃|C,
since A ⊆ Ã and then that A⊥⊥B|C, since B ⊆ B̃. That is, (G) is satisfied. 

Proposition 36 (Hammersley-Clifford) If all densities are positive, then (P) ⇒ (F).

This important (and difficult) result is stated without proof. However, the implication is that
in most practical cases, all four interpretations of a graph of conditional independence
structures are equivalent, and may be considered interchangeably.

Corollary 1 If all densities are positive, then (F) ⇔ (G) ⇔ (L) ⇔ (P).

6.4 Gaussian graphical models (GGMs)

We now turn attention to the simplest parametrised class of graphical models — Gaus-
sian graphical models (GGMs). These are simply multivariate normal distributions where
the variables are subject to conditional independence properties. It turns out to be very
straightforward to characterise GGMs in terms of the pairwise Markov property (P), but to
understand why, we must first understand a little more about the MVN, and its parametri-
sation in terms of the precision matrix. If X is a random vector with variance matrix
Var(X), then Q = Var(X) −1 is the precision matrix for X. So, if X ∼ N (µ, Σ), we can
put Q = Σ−1 , and then X ∼ N (µ, Q−1 ). Note that we can express the density of the MVN
in terms of µ and Q as
 
−p/2 1/2 1 T
f (x) = (2π) |Q| exp − (x − µ) Q(x − µ) .
2

Writing the density in this way allows us to partition x, µ and Q, and derive the conditional
distribution of the first part of x given the rest.
CHAPTER 6. GRAPHICAL MODELLING 146

Proposition 37 For X ∼ N (µ, Q−1 ), partition

       
X1 x1 µ1 Q11 Q12
X= , x= , µ= , Q= .
X2 x2 µ2 Q21 Q22

Then
(X 1 |X 2 = x2 ) ∼ N (µ1|2 , Q11 −1 ),
where
µ1|2 = µ1 − Q11 −1 Q12 (x2 − µ2 ).

Proof

f (x1|x2) ∝ f (x)
 
1
∝ exp − (x − µ)TQ(x − µ)
2

1
∝ exp − (x1 − µ1)TQ11(x1 − µ1)+
2
+ (x2 − µ2)TQ22(x2 − µ2)
T
o
+2(x1 − µ1) Q12(x2 − µ2)
 
1 T
∝ exp − x1 Q11x1 − 2x1TQ11µ1 + 2x1TQ12(x2 − µ2)

2
 
1 T
∝ exp − x1 Q11x1 − 2x1TQ11µ1|2

2
 
1
∝ exp − (x1 − µ1|2)TQ11(x1 − µ1|2) .

2


Using this result, it is straightforward to see why zeroes of Q correspond to pairwise con-
ditional independence statements.
CHAPTER 6. GRAPHICAL MODELLING 147

First consider the case q12 = 0 (= q21 ). This is wlog, since we can always re-order the
variables to ensure that the zero of interest is in this position.∗ Partition X as
 
X3
 X4 
   
X1 X1  
X= , where X 1 = and X 2 =  ..  .
X2 X2  . 
Xp
Then X 1 |X2 = x2 has variance Q11 −1 , which must be diagonal, since Q11 is. Now this
means that X1 and X2 are conditionally uncorrelated, but for the MVN, uncorrelated and
independent are equivalent. Consequently X1 ⊥⊥X2 |(X3 , . . . , Xp ), and so q12 = 0 leads
directly to this CI statement. In general, we have the following result.
Proposition 38 If X ∼ N (µ, Q−1 ), then for i 6= j we have
qij = 0 (= qji ) ⇔ Xi ⊥⊥Xj |X\{Xi , Xj }.
Therefore, there is a direct correspondence between the zero structure of Q and an undi-
rected graph with the pairwise Markov property (P), where zeroes in Q correspond to
missing edges in the graph. Further, provided that Q (or equivalently, the variance matrix,
Σ) is strictly positive definite, the densities are all positive, and we can conclude that all of
our properties (F), (G), (L) and (P) are satisfied by the associated graph.

6.4.1 Partial covariance and correlation

We have shown that the MVN has the important and interesting property that Var(X 1 |X 2 = x2 )
does not depend on the observed value of x2 . This allows us to consider the conditional
variance matrix without reference to the observed values of the variables we are condition-
ing on. We define the partial (co)variance matrix for X 1 (given X 2 ) to be the conditional
covariance matrix Q11 −1 . In particular, in the case X 1 = (X1 , X2 )T , we have
 −1  
−1 q11 q12 1 q22 −q12
Q11 = = ,
q12 q22 2
q11 q22 − q12 −q12 q11
2
and so we define the partial covariance between X1 and X2 to be −q12 /(q11 q22 − q12 ). The
√
corresponding partial correlation is therefore given by −q12 / q11 q22 . In general, the partial
correlation between Xi and Xj , for i 6= j is given by
−qij
√ .
qii qjj
Consequently, we can define the partial correlation matrix to be the matrix of partial cor-
relations, and this is clearly given by
PCorr(X) = 2 Ip −D−1/2 QD−1/2 , where D = diag {q11 , . . . , qpp } .
Note that in the typical case of positive definite Q, the zeroes of Q match those of PCorr(X).
Consequently, we can understand the CI structure of a GGM by studying the zero struc-
ture of the partial correlation matrix.
∗
Note that if P is a permutation matrix, then P is orthogonal. Then Var(PX) = P Var(X) PT , and so
Var(PX) −1 = P Var(X) −1 PT , as can be verified by direct multiplication. In other words, the precision matrix
for the re-ordered variables is just the precision matrix for the variables in their original order, but with rows and
columns re-ordered to match.
CHAPTER 6. GRAPHICAL MODELLING 148

Example: toy problem

Suppose that X = (X1 , X2 , X3 )T ∼ N (0, Σ), where
 
3 −2 1
Σ = −2 4 −2 .
1 −2 3
Are there any conditional independence relations between the variables?
We can use R to answer this question. Note that as there are no zeroes in the variance
matrix, Σ, there are no marginal independence relations between the variables. To look for
conditional independence relations, we can compute the corresponding partial correlation
matrix. We could do this by hand, by first using Gaussian elimination to compute the
precision matrix, but it will be easier to use R.
> Sigma=matrix(c(3,-2,1,-2,4,-2,1,-2,3),ncol=3)
> Q=solve(Sigma)
> Q
[,1] [,2] [,3]
[1,] 0.50 0.25 0.00
[2,] 0.25 0.50 0.25
[3,] 0.00 0.25 0.50
> 2*diag(3)-diag(diag(Q)ˆ(-0.5))%*%Q%*%diag(diag(Q)ˆ(-0.5))
[,1] [,2] [,3]
[1,] 1.0 -0.5 0.0
[2,] -0.5 1.0 -0.5
[3,] 0.0 -0.5 1.0
Note that the ggm package includes the command parcor() for computing the partial
correlation matrix directly from a variance matrix:
> parcor(Sigma)
[,1] [,2] [,3]
[1,] 1.0 -0.5 0.0
[2,] -0.5 1.0 -0.5
[3,] 0.0 -0.5 1.0
which is much easier. We now see that the partial correlation in position (1,3) and (3,1)
is zero, and this implies the pairwise Markov property, X1 ⊥⊥X3 |X2 . Thus, the associated
undirected graph will be a chain with X2 separating X1 and X3 .
This also gives us a potential strategy for estimating a GGM from data: compute
the sample partial correlation matrix for the data set, and then use the zero structure
of this estimated matrix in order to identify missing edges in the GGM. Now, following this
procedure directly will typically lead to very dense graphs, since sample estimates of zero
partial correlations will often be small, but not exactly zero. In this case we can apply a
threshold to the sample partial correlation matrix, or attempt to apply statistical tests for
whether or not particular elements are significantly different from zero.

Example: Mathematics students exam marks

The ggm package includes a very famous (old!) data set on exam marks of mathematics
students in 5 subject areas. It can be loaded and plotted with
CHAPTER 6. GRAPHICAL MODELLING 149

20 40 60 80 10 30 50 70

40 80
● ● ●● ●
● ●
●
● ●● ● ●●
●●● ● ●● ● ● ● ● ● ●● ●●●● ●
●●●●●
●● ● ●●●
●●
●
● ● ● ● ●●●
●●
●
● ● ●●●
●●●
● ●●●●
● ●
●● ●
●
●●●●●
●● ● ●
●●●●●●●
●
● ●●●● ● ●
● ● ●● ● ●●●●
●●●●● ●●
●● ● ● ●●● ●●●●●●●●●● ●●●
● ● ● ●● ● ●● ●●●●●● ●
● ●●●
● ●●● ●
mechanics ●●●●
●
●
●
●
●
●●
● ● ● ●●●● ●
●●
●
●
● ●
●● ●●
●
●
●
●●●
●
●●
●
●
●
●
● ● ●● ●●
●●
●
●●
●
●
●
●
●●●
● ●
●
●
● ●● ●
● ●●●●
● ●
● ● ● ●●● ●● ●
● ●
●
● ●●
●● ●●● ● ●●●● ●
●●● ● ● ●● ●● ●● ●●● ●● ●
●● ●
●● ●
● ● ● ●● ●
● ●
●● ●●●● ●● ●●●● ● ●● ● ● ● ● ● ●
● ● ● ● ●●● ●

0
● ● ● ● ●● ●
●
● ●● ● ●●● ●
● ●●●●●●●●●● ●●● ● ●● ●
● ● ● ●● ●
●● ● ●
20 60
●●●● ●
● ●●
●● ● ●●● ●
●●● ● ● ● ●●● ● ● ●●● ●●● ●●●●
● ● ●●
●● ●
●●●● ●●●●
● ●●
●
●
●●
●● ●
●●
●●
●●
●
● ● ● ●●●●●●●●●
● ●
●
●●● ●● ●●●●●●
● ●
●●●●● ● ●●● ● ● ●● ●
●●● ●● ●●● ●● ●● ● ●● ● ● ●●
●●● ●
● ●
● ● ● ●
●
●
●●
● ● ● ●●●●
●
●●
● ●
●●
●
●●
●●
●●
●
● vectors ● ●
●●●●
● ●
●●●●●
● ●●
● ●
●
●●
●●
●
●
●
●
●
●
● ●●
●●●●
●●
● ● ●● ●●
● ●
● ●
●
●
●●
●●●
●●
●● ●
●
●
●
●
●
● ●
● ●
●●
●
●●
●
●
●
●●
●●
●●
●●
● ●
● ●
●● ● ●
●
●●
●● ●
● ● ● ● ● ● ● ●
● ● ● ●

● ● ● ●
●● ●● ● ●● ● ●● ● ●
●●● ● ●●●●● ●●● ●● ●
●● ●
●● ● ● ● ●● ● ●● ●

60
●● ● ● ●● ●
●●●● ● ●●●● ●● ● ● ● ● ● ●● ●●●●
●● ●●●●●
●●
● ●● ●●● ●●●●●
● ●●● ●●
● ●●
●●● ●
●●●●
●● ●
●● ●
●
●●
●
●●
●
●●
●
●
●
● ● ●●●●
●● ●● ●
●
●● ●
● ●●
● ●
● ●● ●
●●
●
●●
●●
●
●●
●●
●
●
●● ●
●
●● ● ●● ●
●
●
●
●
●●
●●
●
●
●
●
●
●
●● ●
● algebra ●●●●●
●
●●
●●
●●
●●●
●●●
●●●
● ●●
● ●●
●
●
●●●
●●●
●
●
●
●
●● ●
●●
●
●
● ●
●
●●●
●●
●
●
●
● ● ● ●
●
●
● ●●●●●●● ●
●
●
●●
●●●
●●
●
●
●
●
●● ●
●● ●●● ●
● ●

20
● ● ● ●
● ● ● ●

● ● ● ●●●● ●● ● ● ●● ● ●
● ● ●●●●● ●●●●● ●●
● ● ●●●●●●
● ●●●
●●● ●● ●●●●
●●●●
● ●● ● ●
●
● ●●●
● ●● ●●
●● ●●●●●●●●●●●●● ●●
●● ●●
● ●●●●●●
● ●● ●●
● ●●● ●●●● ● ●●● ● ●
● ● ●●
●●●● ● ●●● ●● ●
50

●●● ●● ●
●● ●● ● ●● ●●
●● ●●
●●●●●●
● ●● ●
●● ● ●●● ●●●● ● ●● ●
● ● ●●●● ●● ●●●●
●●●● ●
●
●●● ●●●● ● ●●●●●●●●● ● ● ●●● ●
●● ●● ● ●
● ●● ● ●●●●
●●● ● ●●●●
●
●● ●
●●
●●
●●● ●
●● ●●●
●
●
● ● ●●
●●●●
●
●●
●
●
●●
●
●
●● ●
analysis ●●●
●
●
●
●● ●
● ●● ●●●
●
● ●
●
●
●● ● ● ●● ● ●● ●
● ● ●●●
●
●
●●●
● ●●
●● ● ●
● ● ● ● ● ●
10

● ● ● ●

● ● ●● ● ●●● ● ●● ● ● ● ●●
●
● ●●● ● ●●●●●● ● ●● ● ●●●● ● ● ●●
● ● ● ●
● ● ● ● ●●● ● ● ●●● ● ●
●●● ● ●● ●●●●●●
● ●
● ●●● ● ● ●●● ●

50
● ● ● ●● ● ● ●● ●● ●●● ●●
● ●● ●
●●● ●●
●●
●●
● ●●●
●●
●
●
●●●●●●
●
●●
●
●
● ●
●●
●●
●●
●
●●
●●
●●● ● ●● ●●
●●
●●● ●
●●
●●●
●
●●●
●●
●●
●● ●● ●
● ●●●
●●●
●
●●●
● ●
●●● ● ●●●
●●●
● ●
●
●
●●
●
●
● ●
●●●●
●
●
●●●
●
●
●
● ●
●●
●
● ● ●● ●
●●●
●
● ● ●●●●●●●
●
●
●●
●●
●
●
●●
●●
●●● ●
●
●●●
● statistics
●●● ●
●
● ● ●●● ● ● ●●●●● ●
● ● ●●● ●●
● ●
● ● ●● ●●●●● ●● ● ●● ●● ● ●●

10
● ● ● ● ●

0 20 60 20 40 60 80 10 30 50 70

Figure 6.3: Pairwise scatter-plots of exam marks

data(marks)
pairs(marks)
giving the plot shown in Figure 6.3. The plot shows a general positive correlation between
marks in different subject areas (unsurprisingly), and this can be confirmed by inspecting
the sample variance matrix for the data.
> Sigma=var(marks)
> round(Sigma,digits=2)
mechanics vectors algebra analysis statistics
mechanics 305.69 127.04 101.47 106.32 117.49
vectors 127.04 172.84 85.16 94.67 99.01
algebra 101.47 85.16 112.89 112.11 121.87
analysis 106.32 94.67 112.11 220.38 155.54
statistics 117.49 99.01 121.87 155.54 297.76
There is clearly no indication that any of these variables are marginally independent. To
investigate conditional independence structure, we inspect the partial correlation matrix.
> PCorr=parcor(Sigma)
> round(PCorr,digits=2)
mechanics vectors algebra analysis statistics
mechanics 1.00 0.33 0.23 0.00 0.03
vectors 0.33 1.00 0.28 0.08 0.02
algebra 0.23 0.28 1.00 0.43 0.36
analysis 0.00 0.08 0.43 1.00 0.25
statistics 0.03 0.02 0.36 0.25 1.00
We see immediately that, to 2dps, the partial correlation between mechanics and analysis
is zero. However, it is also clear that the sample partial correlations between mechanics
and statistics, and between vectors and statistics, are also very small, and probably not
CHAPTER 6. GRAPHICAL MODELLING 150

mech stat●
●

alg
●

vec ana
● ●

Figure 6.4: Possible conditional independence graph describing the exam marks data

significantly different from zero. In fact, even the partial correlation between vectors and
analysis is smaller than 0.1. If we arbitrarily threshold the partial correlations at a value of
0.1, then we immediately see the clique structure in the adjacency matrix:
> adj=1*(abs(PCorr)>0.1)
> adj
mechanics vectors algebra analysis statistics
mechanics 1 1 1 0 0
vectors 1 1 1 0 0
algebra 1 1 1 1 1
analysis 0 0 1 1 1
statistics 0 0 1 1 1
and so we can plot the corresponding conditional independence graph with
drawGraph(adj)
giving the plot shown in Figure 6.4. What we see is that algebra separates mechanics and
vectors from analysis and statistics. It is not that (say) ability in mechanics and statistics
is uncorrelated, but that they are uncorrelated given ability in algebra. That is, although
knowing how good someone is at mechanics is useful for predicting how good someone
is at statistics, if we already knew how good they were at algebra, learning how good
they are at mechanics will give us no additional information about how good they are at
statistics.
Obviously, there exist a range of statistical methods for estimating the variance and
precision matrix of a GGM conditional on a given CI structure, and also methods for
statistical testing of partial correlations, and methods for searching model space in a
principled way. Unfortunately we do not have time to explore these methods in detail in
this course, but note that the ggm package includes functions such as fitConGraph()
and pcor.test() which can be useful in this context.
CHAPTER 6. GRAPHICAL MODELLING 151

angle velocity
● ●

radial.position
●

north.south east.west
●
●

Figure 6.5: Possible conditional independence graph describing the galaxy data

Example: galaxy data

Before moving on, it is worth looking briefly at the conditional independence structure
of the variables in the galaxy data from the ElemStatLearn package. We can follow
the procedure used above for the exam marks, as follows, leading to the plot shown in
Figure 6.5.
> require(ElemStatLearn)
> Sigma=var(galaxy)
> PCorr=parcor(Sigma)
> round(PCorr,digits=2)
east.west north.south angle radial.position velocity
east.west 1.00 0.37 -0.04 0.80 0.14
north.south 0.37 1.00 0.15 -0.03 -0.88
angle -0.04 0.15 1.00 -0.05 0.16
radial.position 0.80 -0.03 -0.05 1.00 0.30
velocity 0.14 -0.88 0.16 0.30 1.00
> adj=1*(abs(PCorr)>0.1)
> drawGraph(adj)

See section 17.3 (p.630) of [HTF] for further details of estimating undirected GGMs
from data.
CHAPTER 6. GRAPHICAL MODELLING 152

6.5 Directed acyclic graph (DAG) models

6.5.1 Introduction
Given a conditional independence statement such as X⊥⊥Y |Z, we know that the joint
density for the variables X, Y and Z must factorise in a special way, but it does not tell
us what exactly the “natural” factorisation is. For example, given the above statement, the
following three factorisations are all consistent with the conditional independence state-
ment:

f (x, y, z) = f (x)f (z|x)f (y|z)

f (x, y, z) = f (y)f (z|y)f (x|z)
f (x, y, z) = f (z)f (x|z)f (y|z).

But these factorisations are different, and in the context of conditionally specified statisti-
cal models, one of these choices may be much more natural or convenient than another.
In many applications, a particular choice of factorisation will be needed, and so it is useful
to have a method of encoding a particular factorisation and associated conditional inde-
pendence statements in a formal way. This turns out to be very convenient using directed
acyclic graphs (DAGs).

6.5.2 Directed graphs

A directed graph, or digraph, G, is a tuple (V, E) where V is a finite set of vertices and E a
set of directed edges, with each edge e ∈ E being an ordered pair of the form e = (vi , vj ),
vi , vj ∈ V , i 6= j, representing an edge from vi to vj , often written vi → vj . Clearly
E ⊆ V × V = V 2.
A directed path is a sequence x0 , x1 , . . . , xm ∈ V such that xi−1 → xi , i = 1, 2, . . . , m.
A directed cycle is a (non-trivial) directed path starting and ending at the same vertex.
A directed acyclic graph (DAG) is a digraph which does not contain any directed cycles.
DAGs turn out to be a very important concept in many areas of discrete mathematics and
computing science.
x is a parent of y if x → y. The set of all parents of a node y is denoted pa(y). x is a
child of y if y → x. The set of all children of a node y is denoted ch(y).
If there exists a directed path from x to y, then x is said to be an ancestor of y, and
y is said to be a descendant of x. The set of all ancestors of a node y is written an(y),
and the set of all descendants of a node x is written de(x). The non-descendants of x
are defined by nd(x) ≡ V \{x ∪ de(x)}. Note that for a DAG, we must have an(x) ⊆ nd(x)
(easy exercise).
A list ordering of a DAG is a sequential ordering of the nodes, x1 , x2 , . . . , xp in such a
way that xi ∈ an(xj ) ⇒ i < j. A list ordering can always be constructed for a DAG, but
they are usually not unique.

Example: toy DAG

Consider the graph G = (V, E), where V = {X, Y, Z} and E = {(X, Z), (Z, Y )}. We can
plot this in R using
CHAPTER 6. GRAPHICAL MODELLING 153

X● Z●

Y●

Figure 6.6: Simple DAG with 3 nodes and 2 edges

drawGraph(DAG(Z ∼ X,Y ∼ Z))

giving the graph shown in Figure 6.6. Note, for example, that Z is a child of X and a
parent of Y . Similarly, Y is a descendant of X and X is an ancestor of Y . Also note that
here, an(Y ) = nd(Y ) = {X, Z}. X, Z, Y is a directed path, but Y, Z, X is not. Similarly,
X, Z, Y is the only valid list ordering for this DAG.

From a DAG G, it is possible to construct an associated undirected graph, G m , known

as the moral graph of G, by joining (marrying) all pairs of parents of each node, and then
dropping arrows from all edges. This graph has the property that for each node v and
its parents, pa(v) in the DAG, the corresponding subgraph in the moral graph induced by
v ∪ pa(v) is complete. The marrying of parents ensures that this is the case, and is a key
property of the moral graph, as we shall see.

Example: toy DAG

For the simple toy DAG previously considered, each node has just one parent, so no
parents need marrying, and the moral graph is obtained simply by dropping arrows from
the DAG.

6.5.3 DAG models

As with undirected graphs, we now wish to associate factorisation and conditional inde-
pendence properties of random variables with graphs, and as for the undirected case,
there are several ways to do this. We consider the two most fundamental here.
CHAPTER 6. GRAPHICAL MODELLING 154

Definition 14 We say that the DAG G has the recursive factorisation or directed factori-
sation (DF) property if the joint density of the random variables factorises in the form
Y
f (x) = f (v| pa(v)).
v∈V

Models constructed in this way are sometimes referred to as Bayesian networks, due to
their occurrence in many applications of Bayesian inference.

Example: toy DAG

For the DAG we have been considering, the implied factorisation is

f (x, y, z) = f (x)f (z|x)f (y|z).

Definition 15 We say that the DAG G has the directed local (DL) Markov property if for
all v ∈ V we have
v⊥⊥ nd(v)|pa(v).

Example: toy DAG

For the DAG we have been considering, the only non-trivial implied CI statement is

Y ⊥⊥(X, Z)|Z,

or equivalently,
Y ⊥⊥X|Z.

Lemma 1 (DF) ⇒ (DL).

Proof
We assume (DF), and pick an arbitrary u ∈ V . We wish to show that u⊥⊥ nd(u)|pa(u).
First partition V = u ∪ nd(u) ∪ de(u), and write our factorisation in the form
Y Y
f (x) = f (u| pa(u)) f (v| pa(v)) f (v| pa(v)).
v∈nd(u) v∈de(u)

We are not interested in the descendants of u, so we can marginalise those away to leave
Y
f (u ∪ nd(u)) = f (u| pa(u)) f (v| pa(v)).
v∈nd(u)

But now the only term involving u is the first, giving

f (u| nd(u)) = f (u| pa(u)),

or in other words, u⊥⊥ nd(u)|pa(u). 

CHAPTER 6. GRAPHICAL MODELLING 155

Lemma 2 (DL) ⇒ (DF).

Proof

Start with a list ordering of the nodes x1, x2, . . . , xp, then fac-
torise the joint density in the form
f (x) = f (x1)f (x2|x1) · · · f (xp|x1, . . . , xp−1)
Yp
= f (xi|x1, . . . , xi−1).
i=1

Considering the ith term in this product, we know that since we

have a list ordering we must have {x1, . . . , xi−1} ⊆ nd(xi), and
since we are assuming that xi⊥⊥ nd(xi)|pa(xi), we have that
xi⊥⊥x1, . . . , xi−1|pa(xi), and so f (xi|x1, . . . , xi−1) = f (xi| pa(xi)),
giving the factorisation
p
Y
f (x) = f (xi| pa(xi)),
i=1

and so (DF) is satisfied.


Therefore the two properties we have considered are equivalent.
Proposition 39 (DF) ⇔ (DL).
We now link DAG models back to undirected graphical models via the moral graph.
Proposition 40 If the DAG G satisfies (DF), then the associated moral graph G m satisfies
(F).
Proof

We start with the factorisation

Yp
f (x) = f (xi| pa(xi))
i=1
Yp
= hi(xi, pa(xi)),
i=1
CHAPTER 6. GRAPHICAL MODELLING 156

but by construction of the moral graph, xi ∪ pa(xi) is complete

in G m, and hence will be contained in a clique of G m. By allo-
cating each term to a clique, we see that the density factorises
according to the clique structure of G m, and (F) is satisfied.

This is a key result, since using this, and the results we already have, we can immediately
conclude that

Corollary 2 The moral graph G m associated with a DAG G satisfying (DL) or (DF) satisfies
the Markov properties (G), (L) and (P).

Note that none of the above discussion relies on the positivity assumption necessary for
the Hammersley-Clifford theorem. In summary, one way to understand the conditional
independence assumptions associated with a DAG is to form the associated moral graph,
and then read of conditional assumptions from the moral graph according to any of the
interpretations (G), (L) or (P). Note that this process loses information, in that not all of
the (conditional) independence statements associated with the original DAG model are
present in the corresponding moral graph, but we will not pursue this issue in detail here.

Example: toy DAG

For our toy example, we have already seen that the moral graph is just the undirected ver-
sion of the DAG, encoding the CI statement X⊥⊥Y |Z, as we have already deduced. Note
however, that the three different factorisations consistent with this statement, considered
at the start of this section, all encapsulate the same CI statement and all have the same
moral graph. This is an important point to understand — there will typically be many
DAGs consistent with a particular moral graph, and with a particular set of conditional
independence statements.

Example: collider
At the beginning of this section we considered three different factorisations corresponding
to the conditional independence statement X⊥⊥Y |Z. Each of these factorisations corre-
sponds to a graph with 3 nodes and 2 directed edges, with Z in the middle. There is a
graph with two edges with Z in the middle we have not considered, known as a collider.
It is the graph with edges X → Z and Y → Z. We can draw it using R with
drawGraph(DAG(Z ∼ X+Y))
giving the plot shown in Figure 6.7.
Note that this DAG corresponds to the factorisation

f (x, y, z) = f (x)f (y)f (z|x, y).

There is no (non-trivial) CI statement associated with this DAG. In particular, since X and
Y are both parents of Z, they get “married” in forming the moral graph, and so the moral
CHAPTER 6. GRAPHICAL MODELLING 157

X● Y●

Z●

Figure 6.7: DAG for a “collider”

graph for this DAG is complete. However, it should be noted that this factorisation and the
corresponding DAG do encode the marginal independence of X and Y . So here X and Y
are marginally independent, but not conditionally independent given Z. It is not possible
to encode this information in an undirected graphical model. It is therefore possible to
encode information in DAG models that is not possible to encode in undirected graphs.

Example
Draw the DAG for the random variables X1 , X2 , X3 and X4 given the following factorisation
of the joint density,

f (x1 , x2 , x3 , x4 ) = f (x1 )f (x2 )f (x3 |x1 , x2 )f (x4 |x3 ).

Draw the associated moral graph. Is it true that X1 ⊥⊥X2 |X3 ? Is it true that X2 ⊥⊥X4 |X3 ?

Can draw DAG with drawGraph(DAG(x3∼x1+x2,x4∼x3)). Can draw

moral graph with drawGraph(UG(∼x1*x2*x3+x3*x4)).
CHAPTER 6. GRAPHICAL MODELLING 158

It is not true that X1⊥⊥X2|X3, since X1 and X2 are married in

the moral graph. It is true that X2⊥⊥X4|X3, since X3 separates
X2 and X4 in the moral graph.

Example
Write down the factorisation of the full joint density implied by the following DAG:

X1● X2● X3●

X4● X5●

X6●

f (x) = f (x1)f (x2)f (x3)f (x4|x1, x2)f (x5|x2, x3)f (x6|x4, x5)

6.5.4 Fitting to data

Given a particular DAG structure, we know the form of the factorisation of the joint density,
and a variety of methods may be used to estimate the individual factors. In the graphical
Gaussian case, this essentially involves linearly regressing each variable on its parents.
See the function fitDag() in the ggm package.
Estimating DAG structure from data is an altogether more delicate matter, which we
will not consider in detail here. Note, however, that in general it is not possible to infer
directionality from observational data. Consider first the bivariate case for variables X
and Y . We know that we can factor the joint density either as f (x, y) = f (x)f (y|x) or
f (x, y) = f (y)f (x|y). The first of these corresponds to X → Y and the second to X ← Y .
We cannot distinguish these without some additional information. If we have a particular
parametric form for the terms in the factorisation, that could in principle help, but in the
CHAPTER 6. GRAPHICAL MODELLING 159

graphical Gaussian case we assume that the variables are bivariate normal, and the two
factorisations are indistinguishable from data. In general, the best strategy is often to
estimate an undirected graph structure from data, and then explore the set of directed
graphs consistent with this undirected graph, informally.

6.6 Conclusion
Graphical models are one of the most important tools in modern multivariate statistical
modelling and analysis. In particular, they are central to Bayesian hierarchical modelling,
and many modern Bayesian computational methods. We do not have time to explore
these ideas here, but the basic properties of undirected and directed graphs, and graphi-
cal Gaussian models forms a foundation for further study in this area.
Note that the standard reference on the theory of graphical models is:

Lauritzen, S. L. (1996) Graphical Models, Oxford Science Publications.

Some of the material in this chapter was derived from the above text.
Chapter 7

Variable selection and multiple testing

7.1 Regularisation and variable selection

7.1.1 Introduction
Let us reconsider the problem of multiple linear regression, in the form

y = Xβ + ε,

where X is an n × p data matrix, and we choose β in order to minimise kεk2 . If p is large,

so that X contains many variables, there may be many choices of β that almost minimise
our loss function
L(β) = (y − Xβ)T (y − Xβ).
We know that the β minimising L(β) is given by the solution of the normal equations

XT Xβ̂ = XT y.

When n ≥ p and X has full column rank, there is a unique solution given by

β̂ = (XT X)−1 XT y.

However, if X does not have full column rank (for example, because some of the variables
are co-linear), or p > n, then XT X will not be invertible, and there will be many solutions
to the normal equations, corresponding to many different minimisers of the loss function
L(β). Even if X is full rank, if it is close to rank degenerate, then solution of the normal
equations may be numerically unstable, and the “optimal” β may turn out to have poor
predictive properties.
Regularisation and variable selection are two ways to deal with the above problem,
which turn out to be closely related. Regularisation is concerned with smoothing the
loss function, ensuring it has a unique minimum, and “shrinking” the solution towards a
sensible default, often zero. Regularisation methods are often termed shrinkage methods
for this reason. We begin by looking at the classical method of regularising regression
problems, often known as ridge regression, or Tikhonov regularisation.

160
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 161

7.1.2 Ridge regression

There are several ways to think about ridge regression. We start by viewing it as a con-
strained minimisation problem, where we now find a minimum of L(β) subject to the
constraint that kβk does not exceed some threshold size. This way we ensure that we
do not consider unrealistic choices of β. Now, we often tackle constrained optimisation
problems using a Lagrange multiplier approach. In this case we will minimise the loss
function
Lr (β) = L(β) + λkβk2 ,
and then later use λ in order to impose the required constraint. However, this gives us a
different, and perhaps more natural, way to think about ridge regression. We modify our
loss function in order to include a penalty term which penalises choices of β with large
norm. Small values of λ lead to small penalties, and hence solutions similar to those that
would be obtained through regular least squares approaches. However, larger values of
λ will lead to strong penalties for β with large norm, and hence will have the effect of
“shrinking” the optimal solution towards the origin. Let us now consider how to solve for
the optimal β for given fixed λ ≥ 0.
Proposition 41 The β minimising Lr (β) is given by the solution to the equations
(XT X + λ I)β = XT y.
For λ > 0, these equations are guaranteed to have a unique solution given by
β = (XT X + λ I)−1 XT y.
Proof

First let us expand and simplify the loss function

Lr (β) = L(β) + λkβk2
= (y − Xβ)T(y − Xβ) + λβ Tβ
= y Ty + β TXTXβ − 2β TXTy + λβ Tβ
= y Ty + β T(XTX + λ I)β − 2β TXTy.
Now differentiating wrt β and equating to zero gives
(XTX + λ I)β = XTy.
For λ > 0, the matrix λ I is strictly positive definite, and since
XTX is non-negative definite, XTX + λ I is strictly positive defi-
nite, and hence is invertible. This gives the final result.

The attraction of ridge regression is that irrespective of the column rank of X, or the relative
sizes of n and p, for any λ > 0 the ridge regression has a unique solution, and for large
values of λ, the numerical solution is computationally very stable. It is in this sense that
the problem has been regularised.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 162

Example: galaxy data

We can look at how to implement ridge regression using R for the galaxy data. We being
by thinking about regressing velocity on the other 4 variables, and start using lm().
> require(ElemStatLearn)
> y=galaxy[,5]
> X=as.matrix(galaxy[,-5])
> lm(y ∼ X)

Call:
lm(formula = y ∼ X)

Coefficients:
(Intercept) Xeast.west Xnorth.south Xangle
1589.4229 0.7741 -3.1918 0.1245
Xradial.position
0.9012
As usual, we can append a column of 1s to the front of X and then compute the least
squares solution either using lm() or by solving the normal equations directly.
> X0=cbind(rep(1,length(y)),X)
> lm(y ∼ 0+X0)

Call:
lm(formula = y ∼ 0 + X0)

Coefficients:
X0 X0east.west X0north.south
X0angle
1589.4229 0.7741 -3.1918
0.1245
X0radial.position
0.9012

> solve(t(X0)%*%X0,t(X0)%*%y)
[,1]
1589.4229455
east.west 0.7740997
north.south -3.1917877
angle 0.1245414
radial.position 0.9011797
> QR=qr(X0)
> solve(qr.R(QR),t(qr.Q(QR))%*%y)
[,1]
1589.4229455
east.west 0.7740997
north.south -3.1917877
angle 0.1245414
radial.position 0.9011797
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 163

We can avoid the complications associated with intercepts by first centering the output
and the predictor matrix.
> y=y-mean(y)
> W=sweep(X,2,colMeans(X))
> solve(t(W)%*%W,t(W)%*%y)
[,1]
east.west 0.7740997
north.south -3.1917877
angle 0.1245414
radial.position 0.9011797
>
> QR=qr(W)
> solve(qr.R(QR),t(qr.Q(QR))%*%y)
[,1]
east.west 0.7740997
north.south -3.1917877
angle 0.1245414
radial.position 0.9011797
We can now carry out ridge regression (with λ = 100) by direct solution of the regularised
normal equations
> solve(t(W)%*%W+100*diag(4),t(W)%*%y)
[,1]
east.west 0.7646416
north.south -3.1881190
angle 0.1244684
radial.position 0.9059122
We see that the predictor is very similar to the usual least squares solution, but the addi-
tion of a diagonal term will ensure that the numerical solution of the equations will be very
stable. Note that we could also compute the optimal predictor by numerically optimising
the loss function.
> loss<-function(beta)
+ {
+ eps=y-W%*%beta
+ return(sum(eps*eps)+lambda1*sum(abs(beta))+lambda2*sum(beta*beta))
+ }
> lambda1=0
> lambda2=0
> optim(rep(0,4),loss,control=list(maxit=10000,reltol=1e-12))
$par
[1] 0.7740986 -3.1917888 0.1245412 0.9011769

$value
[1] 288650.8

$counts
function gradient
537 NA
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 164

$convergence
[1] 0

$message
NULL
Note that in the case lambda1=0, the loss is exactly as we require for ridge regression.
The other term in this loss will be explained in the following section. Note that direct
numerical optimisation of multivariate functions is fraught with difficulty, and so whenever
a direct alternative exists, it is almost always to be preferred. In this case, direct solution
of the regularised normal equations is a much better way to compute the solution.

7.1.3 The LASSO and variable selection

Ridge regression is one way to regularise an ill-conditioned least squares problem, but
by no means the only one. In the case of p being large, another obvious approach would
be to select just a subset of the available variables. The intuition here is that if there are
many variables, it is likely that not all will be useful for predicting y, and that since vari-
ables are likely to be correlated anyway, it will not be necessary to include all variables in
the regression even if all happen to be marginally predictive, since the column span of the
subset of variables will be similar to that of a larger subset if the variables dropped are
highly correlated with some of those that remain. There are many possible approaches to
variable selection, including forwards and backwards selection. Perhaps initially variable
selection may appear to have little to do with penalised regularisation methods, but this is
not the case, and the connection between them becomes clear in the context of a regular-
isation method known as the LASSO (least absolute shrinkage and selection operator),
which is actually rather closely related to the method of ridge regression we have already
considered.
The Lasso uses the penalty

Ll (β) = L(β) + λkβk1 ,

where p
X
kβk1 = |βi |
i=1

is the l1 -norm of β. The switch from 2-norm to 1-norm seems at first to be unlikely to make
a significant difference to the shrinkage behaviour of the estimator, but in fact it does, and
importantly, encourages “sparsity” in the optimal solution. It also complicates analysis
somewhat, since it is no longer possible to compute a simple closed-form solution for the
optimal β, and since the l1 -norm is not everywhere differentiable (it is not differentiable
on the coordinate axes), methods of optimisation which assume smooth functions cannot
be used. In fact, it is precisely this lack of differentiability of the l1 -norm on the axes
which causes the method to often have minima including components of β which are
exactly zero. Zeroes in the optimal β correspond to variables which are dropped out
of the regression, and similarly, the non-zero elements of β correspond to the selected
variables. Thus, the simple switch from an l2 to an l1 regularisation penalty leads directly
to a method which simultaneously regularises and selects variables for analysis.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 165

Clearly for λ = 0 and for small values of λ > 0, the method will behave very like
ordinary least squares regression. However, as the value of λ > 0 increases, the effect of
the l1 penalty takes effect, and variables of weak predictive effect begin to drop out of the
optimal predictor.

Example: galaxy data

We can now think about using the Lasso in order to compute an optimal predictor for the
galaxy data. We can begin by trying to directly optimise the loss as follows.
> lambda1=100000
> lambda2=0
> optim(rep(0,4),loss,control=list(maxit=10000,reltol=1e-12))
$par
[1] -1.784425e-11 -2.837611e+00 7.050503e-03 1.103656e+00

$value
[1] 713727.1

$counts
function gradient
469 NA

$convergence
[1] 0

$message
NULL
Note that for lambda2=0, the loss is exactly what we need for the Lasso. Note that the first
coefficient is very close to zero, so the first variable has dropped out of the regression. As
already explained, direct numerical optimisation is problematic, especially for non-smooth
objectives such as this. Fortunately there is an R package called elasticnet which has
efficient procedures for optimising loss functions of this form. We can use it to solve this
problem as follows.
> require(elasticnet)
> predict(enet(X,y,lambda=lambda2,normalize=FALSE),s=lambda1,mode="
penalty",type="coefficients")$coefficients
east.west north.south angle radial.position
0.000000000 -2.836045617 0.007405521 1.104725175
Note how the value for the ridge loss term (here, 0) is passed into the call to enet(), but
the coefficient of the l1 penalty is only needed for the call to the generic function predict
(). This is due to the way that the algorithm is implemented, in that solutions for all values
of the l1 penalty are computed simultaneously. We can exploit this in order to understand
the effect of varying the l1 penalty over a range of values. The command
> plot(enet(X,y,lambda=lambda2,normalize=FALSE))
gives the plot in Figure 7.1. The left hand side of this plot shows the results of using a
very large shrinkage parameter which shrinks all of the regression coefficients to zero.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 166

Standardized Coefficients

angle
0
−1
−2

north.south
−3

0.0 0.2 0.4 0.6 0.8 1.0

|beta|/max|beta|

Figure 7.1: A graphical illustration of the effect of varying the LASSO shrinkage parameter on the
coefficients of the optimal predictor.

The RHS shows the effect of using a zero shrinkage parameter, leading to the usual co-
efficients for ordinary least squares. Moving from left to right, we see additional variables
being incorporated into the optimal predictor as the l1 penalty is gradually relaxed.

7.1.4 The elastic net

The lasso turns out to be good for variable selection, but less good than ridge regression
for regularisation and shrinkage, and also has limitations in the p > n scenario. It is
therefore natural to consider combining the two different penalties, and this is the idea
behind the elastic net. So the elastic net uses the loss function

Le (β) = L(β) + λ2 kβk2 + λ1 kβk1 .

Clearly in the special case λ1 = 0 we get ridge regression, and in the case λ2 = 0 we get
the lasso. So the elastic net is a generalisation of both ridge regression and the lasso,
and combines the best aspects of both.

Example: galaxy data

Again, we could attempt to optimise this loss function directly, as follows.
> lambda1=100000
> lambda2=100
> optim(rep(0,4),loss,control=list(maxit=10000,reltol=1e-12))$par
[1] 1.211979e-11 -2.834280e+00 7.376000e-03 1.104901e+00
We can also use the enet() function to compute this for us.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 167

> predict(enet(X,y,lambda=lambda2,normalize=FALSE),s=lambda1,mode="
penalty",type="coefficients")$coefficients/(1+lambda2)
east.west north.south angle radial.position
0.000000 -2.834275 0.007378 1.104903
Note the division by (1 + λ2 ) at the end. If we do not impose this correction, we get
> predict(enet(X,y,lambda=lambda2,normalize=FALSE),s=lambda1,mode="
penalty",type="coefficients")$coefficients
east.west north.south angle radial.position
0.000000 -286.261800 0.745178 111.595172
These coefficients are a correction to those of the naive elastic net we have presented,
which are thought to have better predictive performance in some scenarios.

7.1.5 p >> n
In some settings we have many more variables than observations, p >> n. We have seen
that in this case there is not a unique solution to the ordinary least squares problem. In
this case regularisation is used in order to make the problem well-defined. We have seen
that ridge regression, or Tikhonov regularisation, is very effective at making the problem
solvable, and shrinking the solution towards the origin. Some kind of shrinkage is vital in
the p >> n scenario, and loss functions containing an l2 penalty are very often used for
this purpose. Variable selection is very often desirable in high dimensional scenarios, but
the Lasso does not perform especially well for p >> n due to the lack of l2 regularisation.
In this case, the elastic net, which combines l1 and l2 regularisation, performs much better,
allowing for a combination of shrinkage and variable selection which can lead to sparse
predictors with good performance.

Example: microarray data

We can investigate the p >> n case using the nci microarray data. Suppose that we
want to regress the first gene on the rest. We can attempt to do ordinary least squares as
follows.
y=t(nci)[,1]
X=t(nci)[,-1]
lm(y ∼ X)$coefficients
but this fails. Trying to compute the solution directly doesn’t help.
X0=cbind(rep(1,length(y)),X)
solve(t(X0)%*%X0,t(X0)%*%y)
We get an error, due to the fact that the system we are attempting to solve is singular.
However, if we include an l2 penalty, we can easily solve the ridge regression problem as
follows.
solve(t(X0)%*%X0+100*diag(ncol(X0)),t(X0)%*%y)
This returns a vector of 6,830 regression coefficients, which we can use for prediction.
However, this representation is not sparse, and so it makes sense to use some kind of
variable selection method to obtain a more parsimonious predictor. We can use the elastic
net for this purpose.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 168

Figure 7.2: A graphical illustration of the effect of varying the l1 shrinkage parameter on the
coefficients of the optimal predictor for the nci microarray data.

lambda1=10
lambda2=100
nci.enet=enet(X,y,lambda=lambda2,normalize=FALSE)
predict(nci.enet,s=lambda1,mode="penalty",type="coefficients")$
coefficients/(1+lambda2)
plot(nci.enet)
This leads to the plot shown in Figure 7.2. We see many predictors being brought in as
the l1 penalty is relaxed. In practice, we will most likely decide on a sensible number of
predictors, and then choose λ1 appropriately. Alternatively, we could keep back some test
data and choose the parameters via cross-validation.

See section 3.3 (p.57) and 3.4 (p.61) of [HTF] for further details of regularisation and
variable selection and Chapter 18 (p.649) of [HTF] for further details of methods for p >>
n.

7.2 Multiple testing

7.2.1 Introduction
For this section, we will think in terms of an n × m data matrix X, rather than n × p, as
in this final section, p will be used to refer to p-values, in the usual statistical sense, and
so it will help to avoid confusion if we let m denote the number of variables of interest. In
the previous section we considered the problem of variable selection. Some methods of
variable selection involve statistical testing of the null hypothesis that adding in the variable
does not improve prediction. Indeed, there are many other contexts where it is natural to
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 169

apply a test to each variable separately in a multivariate problem. This is particularly

common in genomics, where statistical tests are often applied to each variable in turn,
in order to find genes which behave significantly differently in some appropriate sense.
In this context, we often end up with a p-value associated with each variable (each gene
in the genomic data context), and need to somehow threshold these p-values in order
to produce a list of significant variables. The p-value is the probability (under repeated
sampling) that a test statistic as extreme as that observed, will be observed, assuming
that the null hypothesis is true.

7.2.2 The multiple testing problem

We typically consider some threshold significance level, α (often α = 0.05 for a single
test), and then flag as “significant” tests with a p-value smaller than α. There are many
issues with this, but here we will focus on the problems which arise when many tests are
considered together (here, m). In order to get a simple understanding of the problem, we
will assume that all of the tests are independent, although in practice this is unlikely to be
the case.
If we carry out m tests, even if the null hypothesis is true in every case, so that we would
not wish any tests to be flagged as significant, we know that by construction, each of the
m tests will be significant with probability α. In this case the number of false positives will
be X, where X ∼ Bin(m, α). Then we have E(X) = mα, so for large m we expect to
get many false positive tests. For example, for m = 7, 000 and α = 0.05, we would expect
to get 350 false positives in the case where there are no true positives. Indeed, if we
anticipate the number of true positives to be small, will will a priori expect that most of the
positive tests will in fact be false positives, and we will have no way to decide on which or
how many of these we should consider. This is the multiple testing problem.

7.2.3 Bonferroni correction

The Bonferroni method is the classical solution to multiple testing. For a collection of (in-
dependent) tests, Bonferroni’s method controls the family-wise error rate (FWER) — this
is the probability of at least one false positive when there are no true positives. Therefore

FWER = P(At least one false positive)

= 1 − P(No false positives)
= 1 − (1 − α)m

Now, for small α, (1 − α)m ' 1 − mα (first 2 terms in binomial expansion), and so

FWER ' mα.

Note that although we have derived this as an approximation, and assuming independent
tests, it is possible to deduce directly using Boole’s inequality that

FWER ≤ mα,

without assuming independence or making any approximation. Either way, if a FWER of

α0 is required, choosing a significance level for the individual tests of α = α0 /m will achieve
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 170

this. For example, for m = 7, 000, if a FWER of α0 = 0.1 is required, then a significance
level of around α = 0.000014 is required. This incredibly stringent significance level is
required in order to control the FWER, but will clearly come at the expense of many more
false negatives. For large m, this leads to a procedure with poor power for identifying the
true positives of primary interest.

7.2.4 False discovery rate (FDR)

In the context of a large number of tests, m, as is typically the case in “discovery science”
such as genomics, Bonferroni’s method is considered to be too conservative. This is
because in that context, the FWER is really not of primary concern. Here it makes more
sense to consider the false discovery rate (FDR) — this is the proportion of false positives
among the collection of all rejected null hypotheses. For example, an FDR of 0.1 means
that 10% of the tests flagged as significant are false. The converse of this is that 90% of
the tests flagged as significant are true, and that will be fine in most discovery contexts —
people are generally willing to tolerate a small proportion of false positives, as long as the
majority of positives are true. So, we need to understand how to choose a significance
threshold in order to give a desired FDR.
Suppose that we have m tests, with p-values p1 , p2 , . . . , pm . It will be helpful later to
also have the p-values ordered in ascending order, p(1) , p(2) , . . . , p(m) , so that p(1) ≤ p(2) ≤
· · · ≤ p(m) , and p(1) is the most significant test, and p(m) is the least significant test. The
choice of significance level will then be equivalent to deciding how many p-values to take,
reading down the ordered list. Choosing a significance level α will lead to a particular
cutoff, which we will denote l(α), giving a list of significant p-values, p(1) , p(2) , . . . , p(l(α)) .
Clearly, decreasing α will shorten this list, and increasing α will lengthen the list. That is,
the function l(α) is monotonic increasing in α.
Assume now that the number of true positives is small, then the number of false posi-
tives is X, where X ∼ Bin(m, α), approximately. In that case the FDR is given by

X
FDR = .
l(α)

But then
E(X) mα
E(FDR) = = .
l(α) l(α)
If we want to have E(FDR) < α0 , we need to have

mα α0 l(α)
< α0 ⇒ α < .
l(α) m

That is, we need to choose α sufficiently small that this inequality is satisfied (but other-
wise as large as possible). However, we need to be a little bit careful, since both sides of
this inequality depend on α.
In order to see how to solve this, it is helpful to invert the relationship between l and α,
where we now regard l as given, and consider α to be the value of α giving rise to a list of
length l. Then we have
α0 l
α(l) < ,
m
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 171

but p(l) is the p-value giving rise to a list of length l, so

α0 l
p(l) <
m
is the inequality of interest. But then we can visualise the solution of this problem by
0
plotting p(l) and αml as functions of l, and take the crossing point as determining l and p(l) ,
the threshold significance level.

Example: microarray data

Suppose that for the nci microarray data, we want to find genes specifically associated
with melanoma. We could, for each gene, do a 2 sample t-test to see if the mean across
the melanoma samples is significantly different to the mean across the other samples. On
the basis of the computed t statistics, we can compute a p-value for each gene. We can
do this in R as follows, though the details are not especially important for this course.
nci.mel=nci[,colnames(nci)=="MELANOMA"]
nci.rest=nci[,colnames(nci)!="MELANOMA"]
mel.var=apply(nci.mel,1,var)
rest.var=apply(nci.rest,1,var)
pool.var=((8-1)*mel.var+(56-1)*rest.var)/(8+56-2)
pool.se=sqrt(pool.var)*sqrt(1/8+1/56)
tstat=(apply(nci.mel,1,mean)-apply(nci.rest,1,mean))/pool.se
pval=2*(1-pt(abs(tstat),8+7-2))
What we end up with is a vector pval of length 6,830, containing a p-value for each gene.
Under the assumption that the null hypothesis is true for every gene, we would expect
6, 830 × 0.05 = 341.5 false positives. The command length(pval[pval<0.05]) returns
1,377, and so it appears that there are likely to be many (around 1,000) true positives
among the 1,377 genes that are significant at the 5% level. We could employ a Bonferroni
correction, but the command pval[pval<0.05/6830] returns just six p-values meeting
this very stringent criteria. The FDR approach is a compromise between taking just the six
smallest p-values, which are very likely to be all true positives, but leads to vast numbers
of false negatives, and the 1,377 p-values below the 5% threshold, which we know will
contain large numbers of false positives.
We can sort the p-values and plot them as follows, leading to the plot shown in Fig-
ure 7.3.
pval.sort=sort(pval)
plot(1:6830,pval.sort,type="l",col=2)
abline(0.05,0,col=3)
abline(0.05/6830,0,col=5)
abline(0,0.05/6830,col=4)
The red line shows the ordered p-values. The green line represents the usual 0.05 cut-
off, crossed by the p-values at 1,377. The cyan line represents the Bonferroni-corrected
threshold, which just looks like it is at zero on the scale of this plot. The dark blue line
is the FDR threshold. From this plot, all we can really tell is that the p-values cross this
threshold sometime before 500. We can zoom in on the first 500 p-values as follows,
giving the plot shown in Figure 7.4.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 172

1.0
0.8
pval.sort

0.6
0.4
0.2
0.0

0 1000 2000 3000 4000 5000 6000 7000

1:6830

Figure 7.3: Ordered p-values for the nci microarray data.

plot(1:500,pval.sort[1:500],type="l",col=2)
abline(0.05/6830,0,col=5)
abline(0,0.05/6830,col=4)
We still can’t see exactly where the p-values cross the Bonferroni threshold, but we
know see that the p-values cross the FDR threshold at around 180 (in fact, it first exceeds
at 187), and so we will choose to look at the smallest 186 p-values (corresponding to a
significance threshold of around 0.0014), if we are only prepared to tolerate a FDR of 5%.
An alternative way to view the solution to the problem, which is also informative, is to
rewrite the inequality as
p(l) m
< α0 .
l
Then defining
p(l) m
f(l) = ,
l
we want to find the largest l such that
f(l) < α0 .
So if we plot f(l) against l, we look for the (last) crossing of the α0 threshold, from below.
Consequently, we can think informally of f(l) as representing the expected FDR associated
with the lth ordered p-value. This is closely related to (but not quite the same as) the
concept of a q-value, which is also a kind of FDR-corrected p-value. Further examination
of such concepts is beyond the scope of this course.

Example: microarray data

We can plot f(l) for 500 most significant microarray genes using the following commands,
leading to the plot shown in Figure 7.5.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 173

0.008
pval.sort[1:500]

0.004
0.000

0 100 200 300 400 500

1:500

Figure 7.4: First 500 ordered p-values for the nci microarray data.

fdr=6830*pval.sort/1:6830
plot(1:500,fdr[1:500],type="l",col=2)
abline(0.05,0,col=3)
Notice that this function is not monotonic in l, and this why it is not quite right to interpret
f(l) as an FDR-corrected p-value, but it is close enough for our purposes.
Before leaving this example, it is worth emphasising that when working with FDR,
people often work with thresholds above the 0.05 often used in classical statistical testing.
A threshold of 0.1 is very often used (tolerating 1 in 10 false positives), and thresholds of
0.15 are also used sometimes. We can see from Figure 7.5 that if we were to increase
our FDR threshold to 0.1, we would get a list containing around 400 genes, and most
scientists would consider that to be a more appropriate compromise.

See section 18.7 (p.683) of [HTF] for further details of multiple testing problems.
CHAPTER 7. VARIABLE SELECTION AND MULTIPLE TESTING 174

0.10
fdr[1:500]

0.06
0.02

0 100 200 300 400 500

1:500

Figure 7.5: First 500 f(l) statistics for the nci microarray data.