Understanding Machine Learning

Machine learning is one of the fastest growing areas of computer science,

with far-reaching applications. The aim of this textbook is to introduce
machine learning, and the algorithmic paradigms it offers, in a principled way. The book provides an extensive theoretical account of the
fundamental ideas underlying machine learning and the mathematical
derivations that transform these principles into practical algorithms. Following a presentation of the basics of the field, the book covers a wide
array of central topics that have not been addressed by previous textbooks. These include a discussion of the computational complexity of
learning and the concepts of convexity and stability; important algorithmic paradigms including stochastic gradient descent, neural networks,
and structured output learning; and emerging theoretical concepts such as
the PAC-Bayes approach and compression-based bounds. Designed for
an advanced undergraduate or beginning graduate course, the text makes
the fundamentals and algorithms of machine learning accessible to students and nonexpert readers in statistics, computer science, mathematics,
and engineering.
Shai Shalev-Shwartz is an Associate Professor at the School of Computer
Science and Engineering at The Hebrew University, Israel.
Shai Ben-David is a Professor in the School of Computer Science at the
University of Waterloo, Canada.

UNDERSTANDING
MACHINE LEARNING
From Theory to
Algorithms

Shai Shalev-Shwartz
The Hebrew University, Jerusalem

Shai Ben-David
University of Waterloo, Canada

32 Avenue of the Americas, New York, NY 10013-2473, USA

Cambridge University Press is part of the University of Cambridge.
It furthers the Universitys mission by disseminating knowledge in the pursuit of
education, learning and research at the highest international levels of excellence.
www.cambridge.org
Information on this title: www.cambridge.org/9781107057135
c Shai Shalev-Shwartz and Shai Ben-David 2014


This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2014
Printed in the United States of America
A catalog record for this publication is available from the British Library
Library of Congress Cataloging in Publication Data
ISBN 978-1-107-05713-5 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party Internet Web sites referred to in this publication,
and does not guarantee that any content on such Web sites is, or will remain,
accurate or appropriate.

Triple-S dedicates the book to triple-M

Contents

Preface

Introduction
1.1 What Is Learning?
1.2 When Do We Need Machine Learning?
1.3 Types of Learning
1.4 Relations to Other Fields
1.5 How to Read This Book
1.6 Notation

Part 1
2

page xv

Foundations

A Gentle Start
2.1
2.2
2.3
2.4

A Formal Learning Model

3.1
3.2
3.3
3.4
3.5

A Formal Model The Statistical Learning Framework

Empirical Risk Minimization
Empirical Risk Minimization with Inductive Bias
Exercises

PAC Learning
A More General Learning Model
Summary
Bibliographic Remarks
Exercises

Learning via Uniform Convergence

4.1
4.2
4.3
4.4
4.5

Uniform Convergence Is Sufficient for Learnability

Finite Classes Are Agnostic PAC Learnable
Summary
Bibliographic Remarks
Exercises

1
1
3
4
6
7
8
11
13
13
15
16
20
22
22
23
28
28
28
31
31
32
34
35
35
vii

viii

Contents

The Bias-Complexity Tradeoff

5.1 The No-Free-Lunch Theorem
5.2 Error Decomposition
5.3 Summary
5.4 Bibliographic Remarks
5.5 Exercises

36

The VC-Dimension

43

6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8

43
44
46
48
49
53
53
54

Nonuniform Learnability
7.1 Nonuniform Learnability
7.2 Structural Risk Minimization
7.3 Minimum Description Length and Occams Razor
7.4 Other Notions of Learnability Consistency
7.5 Discussing the Different Notions of Learnability
7.6 Summary
7.7 Bibliographic Remarks
7.8 Exercises

58

The Runtime of Learning

73

8.1
8.2
8.3
8.4
8.5
8.6
8.7

Part 2
9

Infinite-Size Classes Can Be Learnable

The VC-Dimension
Examples
The Fundamental Theorem of PAC learning
Proof of Theorem 6.7
Summary
Bibliographic remarks
Exercises

37
40
41
41
41

Computational Complexity of Learning

Implementing the ERM Rule
Efficiently Learnable, but Not by a Proper ERM
Hardness of Learning*
Summary
Bibliographic Remarks
Exercises

From Theory to Algorithms

Linear
9.1
9.2
9.3
9.4
9.5
9.6

Predictors
Halfspaces
Linear Regression
Logistic Regression
Summary
Bibliographic Remarks
Exercises

58
60
63
66
67
70
70
71

74
76
80
81
82
82
83
87
89
90
94
97
99
99
99

Contents

10

11

12

13

14

Boosting
10.1 Weak Learnability
10.2 AdaBoost
10.3 Linear Combinations of Base Hypotheses
10.4 AdaBoost for Face Recognition
10.5 Summary
10.6 Bibliographic Remarks
10.7 Exercises

101

Model
11.1
11.2
11.3
11.4
11.5

114
115
116
120
123
123

Convex Learning Problems

12.1 Convexity, Lipschitzness, and Smoothness
12.2 Convex Learning Problems
12.3 Surrogate Loss Functions
12.4 Summary
12.5 Bibliographic Remarks
12.6 Exercises

124

Regularization and Stability

13.1 Regularized Loss Minimization
13.2 Stable Rules Do Not Overfit
13.3 Tikhonov Regularization as a Stabilizer
13.4 Controlling the Fitting-Stability Tradeoff
13.5 Summary
13.6 Bibliographic Remarks
13.7 Exercises

137

Stochastic Gradient Descent

150

14.1
14.2
14.3
14.4
14.5
14.6
14.7
14.8

15

Selection and Validation

Model Selection Using SRM
Validation
What to Do If Learning Fails
Summary
Exercises

102
105
108
110
111
111
112

Gradient Descent
Subgradients
Stochastic Gradient Descent (SGD)
Variants
Learning with SGD
Summary
Bibliographic Remarks
Exercises

Support Vector Machines

15.1 Margin and Hard-SVM
15.2 Soft-SVM and Norm Regularization
15.3 Optimality Conditions and Support Vectors*

124
130
134
135
136
136

137
139
140
144
146
146
147

151
154
156
159
162
165
166
166
167
167
171
175

ix

Contents

15.4
15.5
15.6
15.7
15.8

16

17

Kernel
16.1
16.2
16.3
16.4
16.5
16.6

19

Methods
Embeddings into Feature Spaces
The Kernel Trick
Implementing Soft-SVM with Kernels
Summary
Bibliographic Remarks
Exercises

179

One-versus-All and All-Pairs

Linear Multiclass Predictors
Structured Output Prediction
Ranking
Bipartite Ranking and Multivariate Performance Measures
Summary
Bibliographic Remarks
Exercises

179
181
186
187
188
188
190
190
193
198
201
206
209
210
210

Decision Trees
18.1 Sample Complexity
18.2 Decision Tree Algorithms
18.3 Random Forests
18.4 Summary
18.5 Bibliographic Remarks
18.6 Exercises

212

Nearest Neighbor

219

19.1
19.2
19.3
19.4
19.5
19.6

20

175
176
177
177
178

Multiclass, Ranking, and Complex Prediction Problems

17.1
17.2
17.3
17.4
17.5
17.6
17.7
17.8

18

Duality*
Implementing Soft-SVM Using SGD
Summary
Bibliographic Remarks
Exercises

Neural
20.1
20.2
20.3
20.4
20.5
20.6

213
214
217
217
218
218

k Nearest Neighbors
Analysis
Efficient Implementation*
Summary
Bibliographic Remarks
Exercises

219
220
225
225
225
225

Networks
Feedforward Neural Networks
Learning Neural Networks
The Expressive Power of Neural Networks
The Sample Complexity of Neural Networks
The Runtime of Learning Neural Networks
SGD and Backpropagation

228
229
230
231
234
235
236

Contents

20.7 Summary
20.8 Bibliographic Remarks
20.9 Exercises

Part 3 Additional Learning Models

21 Online Learning
21.1
21.2
21.3
21.4
21.5
21.6
21.7

22

Clustering
22.1
22.2
22.3
22.4
22.5
22.6
22.7
22.8

23

24

25

Online Classification in the Realizable Case

Online Classification in the Unrealizable Case
Online Convex Optimization
The Online Perceptron Algorithm
Summary
Bibliographic Remarks
Exercises

Linkage-Based Clustering Algorithms

k-Means and Other Cost Minimization Clusterings
Spectral Clustering
Information Bottleneck*
A High Level View of Clustering
Summary
Bibliographic Remarks
Exercises

240
240
240
243
245
246
251
257
258
261
261
262
264
266
268
271
273
274
276
276
276

Dimensionality Reduction
23.1 Principal Component Analysis (PCA)
23.2 Random Projections
23.3 Compressed Sensing
23.4 PCA or Compressed Sensing?
23.5 Summary
23.6 Bibliographic Remarks
23.7 Exercises

278

Generative Models
24.1 Maximum Likelihood Estimator
24.2 Naive Bayes
24.3 Linear Discriminant Analysis
24.4 Latent Variables and the EM Algorithm
24.5 Bayesian Reasoning
24.6 Summary
24.7 Bibliographic Remarks
24.8 Exercises

295

Feature Selection and Generation

25.1 Feature Selection
25.2 Feature Manipulation and Normalization
25.3 Feature Learning

309

279
283
285
292
292
292
293

295
299
300
301
305
307
307
308

310
316
319

xi

xii

Contents

25.4 Summary
25.5 Bibliographic Remarks
25.6 Exercises

Part 4
26

27

28

Advanced Theory

Rademacher Complexities
26.1 The Rademacher Complexity
26.2 Rademacher Complexity of Linear Classes
26.3 Generalization Bounds for SVM
26.4 Generalization Bounds for Predictors with Low
1 Norm
26.5 Bibliographic Remarks

30

31

323
325
325
332
333
335
336

Covering Numbers
27.1 Covering
27.2 From Covering to Rademacher Complexity via Chaining
27.3 Bibliographic Remarks

337

Proof of the Fundamental Theorem of Learning Theory

341

28.1 The Upper Bound for the Agnostic Case

28.2 The Lower Bound for the Agnostic Case
28.3 The Upper Bound for the Realizable Case

29

321
321
322

337
338
340

341
342
347

Multiclass Learnability
29.1 The Natarajan Dimension
29.2 The Multiclass Fundamental Theorem
29.3 Calculating the Natarajan Dimension
29.4 On Good and Bad ERMs
29.5 Bibliographic Remarks
29.6 Exercises

351

Compression Bounds
30.1 Compression Bounds
30.2 Examples
30.3 Bibliographic Remarks

359

PAC-Bayes

364

31.1 PAC-Bayes Bounds

31.2 Bibliographic Remarks
31.3 Exercises

351
352
353
355
357
357

359
361
363

364
366
366

Appendix A Technical Lemmas

369

Appendix B
B.1
B.2
B.3
B.4

372

Measure Concentration
Markovs Inequality
Chebyshevs Inequality
Chernoffs Bounds
Hoeffdings Inequality

372
373
373
375

Contents

B.5
B.6
B.7

Appendix C
C.1
C.2
C.3
C.4
References
Index

Bennets and Bernsteins Inequalities

Sluds Inequality
Concentration of 2 Variables

376
378
378

Linear Algebra
Basic Definitions
Eigenvalues and Eigenvectors
Positive definite matrices
Singular Value Decomposition (SVD)

380
380
381
381
381
385
395

xiii

Preface

The term machine learning refers to the automated detection of meaningful patterns
in data. In the past couple of decades it has become a common tool in almost any
task that requires information extraction from large data sets. We are surrounded
by a machine learning based technology: Search engines learn how to bring us the
best results (while placing profitable ads), antispam software learns to filter our email messages, and credit card transactions are secured by a software that learns
how to detect frauds. Digital cameras learn to detect faces and intelligent personal
assistance applications on smart-phones learn to recognize voice commands. Cars
are equipped with accident prevention systems that are built using machine learning
algorithms. Machine learning is also widely used in scientific applications such as
bioinformatics, medicine, and astronomy.
One common feature of all of these applications is that, in contrast to more traditional uses of computers, in these cases, due to the complexity of the patterns that
need to be detected, a human programmer cannot provide an explicit, fine-detailed
specification of how such tasks should be executed. Taking example from intelligent
beings, many of our skills are acquired or refined through learning from our experience (rather than following explicit instructions given to us). Machine learning tools
are concerned with endowing programs with the ability to learn and adapt.
The first goal of this book is to provide a rigorous, yet easy to follow, introduction
to the main concepts underlying machine learning: What is learning? How can a
machine learn? How do we quantify the resources needed to learn a given concept?
Is learning always possible? Can we know whether the learning process succeeded or
failed?
The second goal of this book is to present several key machine learning algorithms. We chose to present algorithms that on one hand are successfully used in
practice and on the other hand give a wide spectrum of different learning techniques. Additionally, we pay specific attention to algorithms appropriate for large
scale learning (a.k.a. Big Data), since in recent years, our world has become
increasingly digitized and the amount of data available for learning is dramatically increasing. As a result, in many applications data is plentiful and computation

xv

xvi

Preface

time is the main bottleneck. We therefore explicitly quantify both the amount of
data and the amount of computation time needed to learn a given concept.
The book is divided into four parts. The first part aims at giving an initial rigorous answer to the fundamental questions of learning. We describe a generalization
of Valiants Probably Approximately Correct (PAC) learning model, which is a first
solid answer to the question What is learning? We describe the Empirical Risk
Minimization (ERM), Structural Risk Minimization (SRM), and Minimum Description Length (MDL) learning rules, which show how a machine can learn. We
quantify the amount of data needed for learning using the ERM, SRM, and MDL
rules and show how learning might fail by deriving a no-free-lunch theorem. We
also discuss how much computation time is required for learning. In the second part
of the book we describe various learning algorithms. For some of the algorithms,
we first present a more general learning principle, and then show how the algorithm
follows the principle. While the first two parts of the book focus on the PAC model,
the third part extends the scope by presenting a wider variety of learning models.
Finally, the last part of the book is devoted to advanced theory.
We made an attempt to keep the book as self-contained as possible. However,
the reader is assumed to be comfortable with basic notions of probability, linear
algebra, analysis, and algorithms. The first three parts of the book are intended
for first year graduate students in computer science, engineering, mathematics, or
statistics. It can also be accessible to undergraduate students with the adequate
background. The more advanced chapters can be used by researchers intending to
gather a deeper theoretical understanding.

ACKNOWLEDGMENTS
The book is based on Introduction to Machine Learning courses taught by Shai
Shalev-Shwartz at the Hebrew University and by Shai Ben-David at the University
of Waterloo. The first draft of the book grew out of the lecture notes for the course
that was taught at the Hebrew University by Shai Shalev-Shwartz during 20102013.
We greatly appreciate the help of Ohad Shamir, who served as a TA for the course
in 2010, and of Alon Gonen, who served as a TA for the course in 20112013. Ohad
and Alon prepared a few lecture notes and many of the exercises. Alon, to whom
we are indebted for his help throughout the entire making of the book, has also
prepared a solution manual.
We are deeply grateful for the most valuable work of Dana Rubinstein. Dana
has scientifically proofread and edited the manuscript, transforming it from lecturebased chapters into fluent and coherent text.
Special thanks to Amit Daniely, who helped us with a careful read of the
advanced part of the book and wrote the advanced chapter on multiclass learnability. We are also grateful for the members of a book reading club in Jerusalem who
have carefully read and constructively criticized every line of the manuscript. The
members of the reading club are Maya Alroy, Yossi Arjevani, Aharon Birnbaum,
Alon Cohen, Alon Gonen, Roi Livni, Ofer Meshi, Dan Rosenbaum, Dana Rubinstein, Shahar Somin, Alon Vinnikov, and Yoav Wald. We would also like to thank
Gal Elidan, Amir Globerson, Nika Haghtalab, Shie Mannor, Amnon Shashua, Nati
Srebro, and Ruth Urner for helpful discussions.

1
Introduction

The subject of this book is automated learning, or, as we will more often call it,
Machine Learning (ML). That is, we wish to program computers so that they can
learn from input available to them. Roughly speaking, learning is the process of
converting experience into expertise or knowledge. The input to a learning algorithm is training data, representing experience, and the output is some expertise,
which usually takes the form of another computer program that can perform some
task. Seeking a formal-mathematical understanding of this concept, well have to
be more explicit about what we mean by each of the involved terms: What is the
training data our programs will access? How can the process of learning be automated? How can we evaluate the success of such a process (namely, the quality of
the output of a learning program)?

1.1 WHAT IS LEARNING?

Let us begin by considering a couple of examples from naturally occurring animal
learning. Some of the most fundamental issues in ML arise already in that context,
which we are all familiar with.
Bait Shyness Rats Learning to Avoid Poisonous Baits: When rats encounter
food items with novel look or smell, they will first eat very small amounts, and subsequent feeding will depend on the flavor of the food and its physiological effect.
If the food produces an ill effect, the novel food will often be associated with the
illness, and subsequently, the rats will not eat it. Clearly, there is a learning mechanism in play here the animal used past experience with some food to acquire
expertise in detecting the safety of this food. If past experience with the food was
negatively labeled, the animal predicts that it will also have a negative effect when
encountered in the future.
Inspired by the preceding example of successful learning, let us demonstrate
a typical machine learning task. Suppose we would like to program a machine that
learns how to filter spam e-mails. A naive solution would be seemingly similar to the
way rats learn how to avoid poisonous baits. The machine will simply memorize all
previous e-mails that had been labeled as spam e-mails by the human user. When a
1

Introduction

new e-mail arrives, the machine will search for it in the set of previous spam e-mails.
If it matches one of them, it will be trashed. Otherwise, it will be moved to the users
inbox folder.
While the preceding learning by memorization approach is sometimes useful,
it lacks an important aspect of learning systems the ability to label unseen e-mail
messages. A successful learner should be able to progress from individual examples
to broader generalization. This is also referred to as inductive reasoning or inductive
inference. In the bait shyness example presented previously, after the rats encounter
an example of a certain type of food, they apply their attitude toward it on new,
unseen examples of food of similar smell and taste. To achieve generalization in the
spam filtering task, the learner can scan the previously seen e-mails, and extract a set
of words whose appearance in an e-mail message is indicative of spam. Then, when
a new e-mail arrives, the machine can check whether one of the suspicious words
appears in it, and predict its label accordingly. Such a system would potentially be
able correctly to predict the label of unseen e-mails.
However, inductive reasoning might lead us to false conclusions. To illustrate
this, let us consider again an example from animal learning.
Pigeon Superstition: In an experiment performed by the psychologist
B. F. Skinner, he placed a bunch of hungry pigeons in a cage. An automatic mechanism had been attached to the cage, delivering food to the pigeons at regular
intervals with no reference whatsoever to the birds behavior. The hungry pigeons
went around the cage, and when food was first delivered, it found each pigeon
engaged in some activity (pecking, turning the head, etc.). The arrival of food reinforced each birds specific action, and consequently, each bird tended to spend some
more time doing that very same action. That, in turn, increased the chance that the
next random food delivery would find each bird engaged in that activity again. What
results is a chain of events that reinforces the pigeons association of the delivery of
the food with whatever chance actions they had been performing when it was first
delivered. They subsequently continue to perform these same actions diligently.1
What distinguishes learning mechanisms that result in superstition from useful
learning? This question is crucial to the development of automated learners. While
human learners can rely on common sense to filter out random meaningless learning
conclusions, once we export the task of learning to a machine, we must provide
well defined crisp principles that will protect the program from reaching senseless
or useless conclusions. The development of such principles is a central goal of the
theory of machine learning.
What, then, made the rats learning more successful than that of the pigeons?
As a first step toward answering this question, let us have a closer look at the bait
shyness phenomenon in rats.
Bait Shyness revisited rats fail to acquire conditioning between food and electric
shock or between sound and nausea: The bait shyness mechanism in rats turns out to
be more complex than what one may expect. In experiments carried out by Garcia
(Garcia & Koelling 1996), it was demonstrated that if the unpleasant stimulus that
follows food consumption is replaced by, say, electrical shock (rather than nausea),
then no conditioning occurs. Even after repeated trials in which the consumption
1

See: http://psychclassics.yorku.ca/Skinner/Pigeon

1.2 When Do We Need Machine Learning?

of some food is followed by the administration of unpleasant electrical shock, the

rats do not tend to avoid that food. Similar failure of conditioning occurs when the
characteristic of the food that implies nausea (such as taste or smell) is replaced
by a vocal signal. The rats seem to have some built in prior knowledge telling
them that, while temporal correlation between food and nausea can be causal, it is
unlikely that there would be a causal relationship between food consumption and
electrical shocks or between sounds and nausea.
We conclude that one distinguishing feature between the bait shyness learning and the pigeon superstition is the incorporation of prior knowledge that biases
the learning mechanism. This is also referred to as inductive bias. The pigeons in
the experiment are willing to adopt any explanation for the occurrence of food.
However, the rats know that food cannot cause an electric shock and that the
co-occurrence of noise with some food is not likely to affect the nutritional value
of that food. The rats learning process is biased toward detecting some kind of
patterns while ignoring other temporal correlations between events.
It turns out that the incorporation of prior knowledge, biasing the learning process, is inevitable for the success of learning algorithms (this is formally stated and
proved as the No-Free-Lunch theorem in Chapter 5). The development of tools
for expressing domain expertise, translating it into a learning bias, and quantifying
the effect of such a bias on the success of learning is a central theme of the theory
of machine learning. Roughly speaking, the stronger the prior knowledge (or prior
assumptions) that one starts the learning process with, the easier it is to learn from
further examples. However, the stronger these prior assumptions are, the less flexible the learning is it is bound, a priori, by the commitment to these assumptions.
We shall discuss these issues explicitly in Chapter 5.

1.2 WHEN DO WE NEED MACHINE LEARNING?

When do we need machine learning rather than directly program our computers to
carry out the task at hand? Two aspects of a given problem may call for the use of
programs that learn and improve on the basis of their experience: the problems
complexity and the need for adaptivity.
Tasks That Are Too Complex to Program.

Tasks Performed by Animals/Humans: There are numerous tasks that we
human beings perform routinely, yet our introspection concerning how
we do them is not sufficiently elaborate to extract a well defined program. Examples of such tasks include driving, speech recognition, and
image understanding. In all of these tasks, state of the art machine learning programs, programs that learn from their experience, achieve quite
satisfactory results, once exposed to sufficiently many training examples.

Tasks beyond Human Capabilities: Another wide family of tasks that benefit from machine learning techniques are related to the analysis of very
large and complex data sets: astronomical data, turning medical archives
into medical knowledge, weather prediction, analysis of genomic data, Web
search engines, and electronic commerce. With more and more available

Introduction

digitally recorded data, it becomes obvious that there are treasures of meaningful information buried in data archives that are way too large and too
complex for humans to make sense of. Learning to detect meaningful patterns in large and complex data sets is a promising domain in which the
combination of programs that learn with the almost unlimited memory
capacity and ever increasing processing speed of computers opens up new
horizons.
Adaptivity. One limiting feature of programmed tools is their rigidity once the
program has been written down and installed, it stays unchanged. However,
many tasks change over time or from one user to another. Machine learning
tools programs whose behavior adapts to their input data offer a solution to
such issues; they are, by nature, adaptive to changes in the environment they
interact with. Typical successful applications of machine learning to such problems include programs that decode handwritten text, where a fixed program can
adapt to variations between the handwriting of different users; spam detection
programs, adapting automatically to changes in the nature of spam e-mails; and
speech recognition programs.

1.3 TYPES OF LEARNING

Learning is, of course, a very wide domain. Consequently, the field of machine
learning has branched into several subfields dealing with different types of learning
tasks. We give a rough taxonomy of learning paradigms, aiming to provide some
perspective of where the content of this book sits within the wide field of machine
learning.
We describe four parameters along which learning paradigms can be classified.
Supervised versus Unsupervised Since learning involves an interaction between the
learner and the environment, one can divide learning tasks according to the
nature of that interaction. The first distinction to note is the difference between
supervised and unsupervised learning. As an illustrative example, consider the
task of learning to detect spam e-mail versus the task of anomaly detection.
For the spam detection task, we consider a setting in which the learner receives
training e-mails for which the label spam/not-spam is provided. On the basis of
such training the learner should figure out a rule for labeling a newly arriving
e-mail message. In contrast, for the task of anomaly detection, all the learner
gets as training is a large body of e-mail messages (with no labels) and the
learners task is to detect unusual messages.
More abstractly, viewing learning as a process of using experience to gain
expertise, supervised learning describes a scenario in which the experience,
a training example, contains significant information (say, the spam/not-spam
labels) that is missing in the unseen test examples to which the learned expertise is to be applied. In this setting, the acquired expertise is aimed to predict
that missing information for the test data. In such cases, we can think of the
environment as a teacher that supervises the learner by providing the extra
information (labels). In unsupervised learning, however, there is no distinction
between training and test data. The learner processes input data with the goal

1.3 Types of Learning

of coming up with some summary, or compressed version of that data. Clustering a data set into subsets of similar objets is a typical example of such a
task.
There is also an intermediate learning setting in which, while the training examples contain more information than the test examples, the learner is
required to predict even more information for the test examples. For example, one may try to learn a value function that describes for each setting of a
chess board the degree by which Whites position is better than the Blacks.
Yet, the only information available to the learner at training time is positions
that occurred throughout actual chess games, labeled by who eventually won
that game. Such learning frameworks are mainly investigated under the title of
reinforcement learning.
Active versus Passive Learners Learning paradigms can vary by the role played
by the learner. We distinguish between active and passive learners. An
active learner interacts with the environment at training time, say, by posing
queries or performing experiments, while a passive learner only observes the
information provided by the environment (or the teacher) without influencing or directing it. Note that the learner of a spam filter is usually passive
waiting for users to mark the e-mails coming to them. In an active setting, one could imagine asking users to label specific e-mails chosen by the
learner, or even composed by the learner, to enhance its understanding of what
spam is.
Helpfulness of the Teacher When one thinks about human learning, of a baby at
home or a student at school, the process often involves a helpful teacher, who
is trying to feed the learner with the information most useful for achieving
the learning goal. In contrast, when a scientist learns about nature, the environment, playing the role of the teacher, can be best thought of as passive
apples drop, stars shine, and the rain falls without regard to the needs of the
learner. We model such learning scenarios by postulating that the training data
(or the learners experience) is generated by some random process. This is the
basic building block in the branch of statistical learning. Finally, learning also
occurs when the learners input is generated by an adversarial teacher. This
may be the case in the spam filtering example (if the spammer makes an effort
to mislead the spam filtering designer) or in learning to detect fraud. One also
uses an adversarial teacher model as a worst-case scenario, when no milder
setup can be safely assumed. If you can learn against an adversarial teacher,
you are guaranteed to succeed interacting any odd teacher.
Online versus Batch Learning Protocol The last parameter we mention is the distinction between situations in which the learner has to respond online, throughout the learning process, and settings in which the learner has to engage the
acquired expertise only after having a chance to process large amounts of data.
For example, a stockbroker has to make daily decisions, based on the experience collected so far. He may become an expert over time, but might have
made costly mistakes in the process. In contrast, in many data mining settings,
the learner the data miner has large amounts of training data to play with
before having to output conclusions.

Introduction

In this book we shall discuss only a subset of the possible learning paradigms.
Our main focus is on supervised statistical batch learning with a passive learner
(for example, trying to learn how to generate patients prognoses, based on large
archives of records of patients that were independently collected and are already
labeled by the fate of the recorded patients). We shall also briefly discuss online
learning and batch unsupervised learning (in particular, clustering).

1.4 RELATIONS TO OTHER FIELDS

As an interdisciplinary field, machine learning shares common threads with the
mathematical fields of statistics, information theory, game theory, and optimization.
It is naturally a subfield of computer science, as our goal is to program machines so
that they will learn. In a sense, machine learning can be viewed as a branch of AI
(Artificial Intelligence), since, after all, the ability to turn experience into expertise or to detect meaningful patterns in complex sensory data is a cornerstone of
human (and animal) intelligence. However, one should note that, in contrast with
traditional AI, machine learning is not trying to build automated imitation of intelligent behavior, but rather to use the strengths and special abilities of computers
to complement human intelligence, often performing tasks that fall way beyond
human capabilities. For example, the ability to scan and process huge databases
allows machine learning programs to detect patterns that are outside the scope of
human perception.
The component of experience, or training, in machine learning often refers to
data that is randomly generated. The task of the learner is to process such randomly
generated examples toward drawing conclusions that hold for the environment from
which these examples are picked. This description of machine learning highlights its
close relationship with statistics. Indeed there is a lot in common between the two
disciplines, in terms of both the goals and techniques used. There are, however, a
few significant differences of emphasis; if a doctor comes up with the hypothesis
that there is a correlation between smoking and heart disease, it is the statisticians
role to view samples of patients and check the validity of that hypothesis (this is the
common statistical task of hypothesis testing). In contrast, machine learning aims
to use the data gathered from samples of patients to come up with a description of
the causes of heart disease. The hope is that automated techniques may be able to
figure out meaningful patterns (or hypotheses) that may have been missed by the
human observer.
In contrast with traditional statistics, in machine learning in general, and in this
book in particular, algorithmic considerations play a major role. Machine learning
is about the execution of learning by computers; hence algorithmic issues are pivotal. We develop algorithms to perform the learning tasks and are concerned with
their computational efficiency. Another difference is that while statistics is often
interested in asymptotic behavior (like the convergence of sample-based statistical estimates as the sample sizes grow to infinity), the theory of machine learning
focuses on finite sample bounds. Namely, given the size of available samples,
machine learning theory aims to figure out the degree of accuracy that a learner
can expect on the basis of such samples.

1.5 How to Read This Book

There are further differences between these two disciplines, of which we shall
mention only one more here. While in statistics it is common to work under the
assumption of certain presubscribed data models (such as assuming the normality of data-generating distributions, or the linearity of functional dependencies), in
machine learning the emphasis is on working under a distribution-free setting,
where the learner assumes as little as possible about the nature of the data distribution and allows the learning algorithm to figure out which models best approximate
the data-generating process. A precise discussion of this issue requires some technical preliminaries, and we will come back to it later in the book, and in particular in
Chapter 5.

1.5 HOW TO READ THIS BOOK

The first part of the book provides the basic theoretical principles that underlie
machine learning (ML). In a sense, this is the foundation upon which the rest of
the book is built. This part could serve as a basis for a minicourse on the theoretical
foundations of ML.
The second part of the book introduces the most commonly used algorithmic
approaches to supervised machine learning. A subset of these chapters may also be
used for introducing machine learning in a general AI course to computer science,
Math, or engineering students.
The third part of the book extends the scope of discussion from statistical classification to other learning models. It covers online learning, unsupervised learning,
dimensionality reduction, generative models, and feature learning.
The fourth part of the book, Advanced Theory, is geared toward readers who
have interest in research and provides the more technical mathematical techniques
that serve to analyze and drive forward the field of theoretical machine learning.
The Appendixes provide some technical tools used in the book. In particular, we
list basic results from measure concentration and linear algebra.
A few sections are marked by an asterisk, which means they are addressed
to more advanced students. Each chapter is concluded with a list of exercises. A
solution manual is provided in the course Web site.

1.5.1 Possible Course Plans Based on This Book

A 14 Week Introduction Course for Graduate Students:
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.

Chapters 24.
Chapter 9 (without the VC calculation).
Chapters 56 (without proofs).
Chapter 10.
Chapters 7, 11 (without proofs).
Chapters 12, 13 (with some of the easier proofs).
Chapter 14 (with some of the easier proofs).
Chapter 15.
Chapter 16.
Chapter 18.

Introduction

11.
12.
13.
14.

Chapter 22.
Chapter 23 (without proofs for compressed sensing).
Chapter 24.
Chapter 25.

A 14 Week Advanced Course for Graduate Students:

1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.

Chapters 26, 27.

(continued)
Chapters 6, 28.
Chapter 7.
Chapter 31.
Chapter 30.
Chapters 12, 13.
Chapter 14.
Chapter 8.
Chapter 17.
Chapter 29.
Chapter 19.
Chapter 20.
Chapter 21.

1.6 NOTATION
Most of the notation we use throughout the book is either standard or defined on
the spot. In this section we describe our main conventions and provide a table summarizing our notation (Table 1.1). The reader is encouraged to skip this section and
return to it if during the reading of the book some notation is unclear.
We denote scalars and abstract objects with lowercase letters (e.g. x and ).
Often, we would like to emphasize that some object is a vector and then we use
boldface letters (e.g. x and ). The i th element of a vector x is denoted by x i . We use
uppercase letters to denote matrices, sets, and sequences. The meaning should be
clear from the context. As we will see momentarily, the input of a learning algorithm
is a sequence of training examples. We denote by z an abstract example and by
S = z 1 , . . . , z m a sequence of m examples. Historically, S is often referred to as a
training set; however, we will always assume that S is a sequence rather than a set.
A sequence of m vectors is denoted by x1 , . . . , xm . The i th element of xt is denoted
by x t,i .
Throughout the book, we make use of basic notions from probability. We denote
by D a distribution over some set,2 for example, Z . We use the notation z D to
denote that z is sampled according to D. Given a random variable f : Z R, its
expected value is denoted by EzD [ f (z)]. We sometimes use the shorthand E [ f ]
when the dependence on z is clear from the context. For f : Z {true, false} we
also use PzD [ f (z)] to denote D({z : f (z) = true}). In the next chapter we will also
2

To be mathematically precise, D should be defined over some -algebra of subsets of Z . The user who
is not familiar with measure theory can skip the few footnotes and remarks regarding more formal
measurability definitions and assumptions.

1.6 Notation

Table 1.1. Summary of notation

symbol

meaning

R
Rd
R+
N

O, o, , , , O
1[Boolean expression]
[a]+
[n]
x, v, w
xi , vi , wi
x, v
x2 or x
x1
x
x0
A Rd,k
A

A i, j
x x

x1 , . . . , xm
xi, j
w(1) , . . ., w(T )
(t)
wi
X
Y
Z
H

: H Z R+
D
D(A)
z D
S = z 1 , . . ., z m
S Dm
P, E
PzD [ f (z)]
EzD [ f (z)]
N(, C)
f  (x)
f  (x)

the set of real numbers

the set of d-dimensional vectors over R
the set of non-negative real numbers
the set of natural numbers
asymptotic notation (see text)
indicator function (equals 1 if expression is true and 0 o.w.)
= max{0, a}
the set {1, . . . , n} (for n N)
(column) vectors
the ith element of a vector


= di=1 xi vi (inner product)
= x, x (the
2 norm of x)


= di=1 |xi | (the
1 norm of x)
= maxi |xi | (the
norm of x)
the number of nonzero elements of x
a d k matrix over R
the transpose of A
the (i, j ) element of A
the d d matrix A s.t. A i, j = xi x j (where x Rd )
a sequence of m vectors
the j th element of the ith vector in the sequence
the values of a vector w during an iterative algorithm
the ith element of the vector w(t)
instances domain (a set)
labels domain (a set)
examples domain (a set)
hypothesis class (a set)
loss function
a distribution over some set (usually over Z or over X )
the probability of a set A Z according to D
sampling z according to D
a sequence of m examples
sampling S = z 1 , . . . , z m i.i.d. according to D
probability and expectation of a random variable
= D({z : f (z) = true}) for f : Z {true, false}
expectation of the random variable f : Z R
Gaussian distribution with expectation and covariance C
the derivative of a function f : R R at x
the second derivative of a function f : R R at x
the partial derivative of a function f : Rd R at w w.r.t. wi
the gradient of a function f : Rd R at w
the differential set of a function f : Rd R at w
= min{ f (x) : x C} (minimal value of f over C)
= max{ f (x) : x C} (maximal value of f over C)
the set {x C : f (x) = minzC f (z)}
the set {x C : f (x) = maxzC f (z)}
the natural logarithm

f (w)
wi

f (w)
f (w)
min xC f (x)
max xC f (x)
argmin xC f (x)
argmax xC f (x)
log

10

Introduction

introduce the notation Dm to denote the probability over Z m induced by sampling

(z 1 , . . . , z m ) where each point z i is sampled from D independently of the other points.
In general, we have made an effort to avoid asymptotic notation. However, we
occasionally use it to clarify the main results. In particular, given f : R R+ and
g : R R+ we write f = O(g) if there exist x 0 , R+ such that for all x > x 0 we
have f (x) g(x). We write f = o(g) if for every > 0 there exists x 0 such that for
all x > x 0 we have f (x) g(x). We write f = (g) if there exist x 0 , R+ such that
for all x > x 0 we have f (x) g(x). The notation f = (g) is defined analogously.
The notation f = (g) means that f = O(g) and g = O( f ). Finally, the notation

f = O(g)
means that there exists k N such that f (x) = O(g(x) logk (g(x))).
The inner product between vectors x and w is denoted by x, w. Whenever we
do not specify the vector

space we assume that it is the d-dimensional Euclidean
space and then x, w = di=1 x i wi . The Euclidean (or
2 ) norm of a vector w is

w2 = w, w. We omit the subscript from the
2 norm when it is clear from the
1/ p


context. We also use other
p norms, w p =
|wi | p
, and in particular w1 =
i


|w
|
and
w
=
max
|w
|.
i

i
i
i
We use the notation minxC f (x) to denote the minimum value of the set
{ f (x) : x C}. To be mathematically more precise, we should use infxC f (x) whenever the minimum is not achievable. However, in the context of this book the
distinction between infimum and minimum is often of little interest. Hence, to simplify the presentation, we sometimes use the min notation even when inf is more
adequate. An analogous remark applies to max versus sup.

PART 1

Foundations

2
A Gentle Start

Let us begin our mathematical analysis by showing how successful learning can be
achieved in a relatively simplified setting. Imagine you have just arrived in some
small Pacific island. You soon find out that papayas are a significant ingredient in the
local diet. However, you have never before tasted papayas. You have to learn how
to predict whether a papaya you see in the market is tasty or not. First, you need
to decide which features of a papaya your prediction should be based on. On the
basis of your previous experience with other fruits, you decide to use two features:
the papayas color, ranging from dark green, through orange and red to dark brown,
and the papayas softness, ranging from rock hard to mushy. Your input for figuring
out your prediction rule is a sample of papayas that you have examined for color
and softness and then tasted and found out whether they were tasty or not. Let
us analyze this task as a demonstration of the considerations involved in learning
problems.
Our first step is to describe a formal model aimed to capture such learning tasks.

2.1 A FORMAL MODEL THE STATISTICAL LEARNING FRAMEWORK

The learners input: In the basic statistical learning setting, the learner has access
to the following:
Domain set: An arbitrary set, X . This is the set of objects that we may wish
to label. For example, in the papaya learning problem mentioned before,
the domain set will be the set of all papayas. Usually, these domain
points will be represented by a vector of features (like the papayas color
and softness). We also refer to domain points as instances and to X as
instance space.
Label set: For our current discussion, we will restrict the label set to be a
two-element set, usually {0, 1} or {1, +1}. Let Y denote our set of possible labels. For our papayas example, let Y be {0, 1}, where 1 represents
being tasty and 0 stands for being not-tasty.
Training data: S = ((x 1 , y1 ) . . . (x m , ym )) is a finite sequence of pairs in X Y:
that is, a sequence of labeled domain points. This is the input that the
13

14

A Gentle Start

learner has access to (like a set of papayas that have been tasted and their
color, softness, and tastiness). Such labeled examples are often called
training examples. We sometimes also refer to S as a training set.1
The learners output: The learner is requested to output a prediction rule,
h : X Y. This function is also called a predictor, a hypothesis, or a classifier.
The predictor can be used to predict the label of new domain points. In our
papayas example, it is a rule that our learner will employ to predict whether
future papayas he examines in the farmers market are going to be tasty or not.
We use the notation A(S) to denote the hypothesis that a learning algorithm,
A, returns upon receiving the training sequence S.
A simple data-generation model We now explain how the training data is generated. First, we assume that the instances (the papayas we encounter) are
generated by some probability distribution (in this case, representing the
environment). Let us denote that probability distribution over X by D. It is
important to note that we do not assume that the learner knows anything about
this distribution. For the type of learning tasks we discuss, this could be any
arbitrary probability distribution. As to the labels, in the current discussion
we assume that there is some correct labeling function, f : X Y, and that
yi = f (x i ) for all i . This assumption will be relaxed in the next chapter. The
labeling function is unknown to the learner. In fact, this is just what the learner
is trying to figure out. In summary, each pair in the training data S is generated
by first sampling a point x i according to D and then labeling it by f .
Measures of success: We define the error of a classifier to be the probability that
it does not predict the correct label on a random data point generated by the
aforementioned underlying distribution. That is, the error of h is the probability to draw a random instance x, according to the distribution D, such that
h(x) does not equal f (x).
Formally, given a domain subset,2 A X , the probability distribution, D,
assigns a number, D(A), which determines how likely it is to observe a point
x A. In many cases, we refer to A as an event and express it using a function
: X {0, 1}, namely, A = {x X : (x) = 1}. In that case, we also use the
notation PxD [(x)] to express D(A).
We define the error of a prediction rule, h : X Y, to be
def

L D, f (h) =

def

P [h(x) = f (x)] = D({x : h(x) = f (x)}).

xD

(2.1)

That is, the error of such h is the probability of randomly choosing an example
x for which h(x) = f (x). The subscript (D, f ) indicates that the error is measured with respect to the probability distribution D and the correct labeling
function f . We omit this subscript when it is clear from the context. L (D, f ) (h)
has several synonymous names such as the generalization error, the risk, or
the true error of h, and we will use these names interchangeably throughout
1
2

Despite the set notation, S is a sequence. In particular, the same example may appear twice in S and
some algorithms can take into account the order of examples in S.
Strictly speaking, we should be more careful and require that A is a member of some -algebra of
subsets of X , over which D is defined. We will formally define our measurability assumptions in the
next chapter.

2.2 Empirical Risk Minimization

the book. We use the letter L for the error, since we view this error as the loss
of the learner. We will later also discuss other possible formulations of such
loss.
A note about the information available to the learner The learner is blind to the
underlying distribution D over the world and to the labeling function f. In our
papayas example, we have just arrived in a new island and we have no clue
as to how papayas are distributed and how to predict their tastiness. The only
way the learner can interact with the environment is through observing the
training set.
In the next section we describe a simple learning paradigm for the preceding
setup and analyze its performance.

2.2 EMPIRICAL RISK MINIMIZATION

As mentioned earlier, a learning algorithm receives as input a training set S, sampled from an unknown distribution D and labeled by some target function f , and
should output a predictor h S : X Y (the subscript S emphasizes the fact that
the output predictor depends on S). The goal of the algorithm is to find h S that
minimizes the error with respect to the unknown D and f .
Since the learner does not know what D and f are, the true error is not directly
available to the learner. A useful notion of error that can be calculated by the
learner is the training error the error the classifier incurs over the training sample:
def

L S (h) =

|{i [m] : h(x i ) = yi }|

,
m

(2.2)

where [m] = {1, . . . , m}.

The terms empirical error and empirical risk are often used interchangeably for
this error.
Since the training sample is the snapshot of the world that is available to the
learner, it makes sense to search for a solution that works well on that data. This
learning paradigm coming up with a predictor h that minimizes L S (h) is called
Empirical Risk Minimization or ERM for short.

2.2.1 Something May Go Wrong Overfitting

Although the ERM rule seems very natural, without being careful, this approach
may fail miserably.
To demonstrate such a failure, let us go back to the problem of learning to predict the taste of a papaya on the basis of its softness and color. Consider a sample as
depicted in the following:

15

16

A Gentle Start

Assume that the probability distribution D is such that instances are distributed
uniformly within the gray square and the labeling function, f , determines the label
to be 1 if the instance is within the inner square, and 0 otherwise. The area of the
gray square in the picture is 2 and the area of the inner square is 1. Consider the
following predictor:

yi if i [m] s. t. x i = x
(2.3)
h S (x) =
0 otherwise.
While this predictor mig
might
ht seem rather artificia
artificial,
l, in Exercise 2.1 we show a natural representation of it using polynomials. Clearly, no matter what the sample is,
L S (h S ) = 0, and therefore this predictor may be chosen by an ERM algorithm (it is
one of the empirical-minimum-cost hypotheses; no classifier can have smaller error).
On the other hand, the true error of any classifier that predicts the label 1 only on a
finite number of instances is, in this case, 1/2. Thus, L D (h S ) = 1/2. We have found
a predictor whose performance on the training set is excellent, yet its performance
on the true world is very poor. This phenomenon is called overfitting. Intuitively,
overfitting occurs when our hypothesis fits the training data too well (perhaps like
the everyday experience that a person who provides a perfect detailed explanation
for each of his single actions may raise suspicion).

2.3 EMPIRICAL RISK MINIMIZATION WITH INDUCTIVE BIAS

We have just demonstrated that the ERM rule might lead to overfitting. Rather
than giving up on the ERM paradigm, we will look for ways to rectify it. We will
search for conditions under which there is a guarantee that ERM does not overfit,
namely, conditions under which when the ERM predictor has good performance
with respect to the training data, it is also highly likely to perform well over the
underlying data distribution.
A common solution is to apply the ERM learning rule over a restricted search
space. Formally, the learner should choose in advance (before seeing the data) a set
of predictors. This set is called a hypothesis class and is denoted by H. Each h H
is a function mapping from X to Y. For a given class H, and a training sample, S,
the ERMH learner uses the ERM rule to choose a predictor h H, with the lowest
possible error over S. Formally,
ERMH (S) argmin L S (h),
hH

where argmin stands for the set of hypotheses in H that achieve the minimum value
of L S (h) over H. By restricting the learner to choosing a predictor from H, we bias it
toward a particular set of predictors. Such restrictions are often called an inductive
bias. Since the choice of such a restriction is determined before the learner sees the
training data, it should ideally be based on some prior knowledge about the problem
to be learned. For example, for the papaya taste prediction problem we may choose
the class H to be the set of predictors that are determined by axis aligned rectangles
(in the space determined by the color and softness coordinates). We will later show
that ERMH over this class is guaranteed not to overfit. On the other hand, the
example of overfitting that we have seen previously, demonstrates that choosing H

2.3 Empirical Risk Minimization with Inductive Bias

to be a class of predictors that includes all functions that assign the value 1 to a finite
set of domain points does not suffice to guarantee that ERMH will not overfit.
A fundamental question in learning theory is, over which hypothesis classes
ERMH learning will not result in overfitting. We will study this question later in
the book.
Intuitively, choosing a more restricted hypothesis class better protects us against
overfitting but at the same time might cause us a stronger inductive bias. We will get
back to this fundamental tradeoff later.

2.3.1 Finite Hypothesis Classes

The simplest type of restriction on a class is imposing an upper bound on its size
(that is, the number of predictors h in H). In this section, we show that if H is a
finite class then ERMH will not overfit, provided it is based on a sufficiently large
training sample (this size requirement will depend on the size of H).
Limiting the learner to prediction rules within some finite hypothesis class may
be considered as a reasonably mild restriction. For example, H can be the set of all
predictors that can be implemented by a C++ program written in at most 109 bits
of code. In our papayas example, we mentioned previously the class of axis aligned
rectangles. While this is an infinite class, if we discretize the representation of real
numbers, say, by using a 64 bits floating-point representation, the hypothesis class
becomes a finite class.
Let us now analyze the performance of the ERMH learning rule assuming that
H is a finite class. For a training sample, S, labeled according to some f : X Y, let
h S denote a result of applying ERM H to S, namely,
h S argmin L S (h).
hH

(2.4)

In this chapter, we make the following simplifying assumption (which will be

relaxed in the next chapter).
Definition 2.1 (The Realizability Assumption). There exists h  H s.t.
L (D, f ) (h  ) = 0. Note that this assumption implies that with probability 1 over random samples, S, where the instances of S are sampled according to D and are labeled
by f , we have L S (h  ) = 0.
The realizability assumption implies that for every ERM hypothesis we have
that3 L S (h S ) = 0. However, we are interested in the true risk of h S , L (D, f ) (h S ),
rather than its empirical risk.
Clearly, any guarantee on the error with respect to the underlying distribution,
D, for an algorithm that has access only to a sample S should depend on the relationship between D and S. The common assumption in statistical machine learning
is that the training sample S is generated by sampling points from the distribution D
independently of each other. Formally,
3

Mathematically speaking, this holds with probability 1. To simplify the presentation, we sometimes
omit the with probability 1 specifier.

17

18

A Gentle Start

The i.i.d. assumption: The examples in the training set are independently and
identically distributed (i.i.d.) according to the distribution D. That is, every
x i in S is freshly sampled according to D and then labeled according to the
labeling function, f . We denote this assumption by S Dm where m is the
size of S, and Dm denotes the probability over m-tuples induced by applying D
to pick each element of the tuple independently of the other members of the
tuple.
Intuitively, the training set S is a window through which the learner gets
partial information about the distribution D over the world and the labeling
function, f . The larger the sample gets, the more likely it is to reflect more
accurately the distribution and labeling used to generate it.
Since L (D, f ) (h S ) depends on the training set, S, and that training set is picked by
a random process, there is randomness in the choice of the predictor h S and, consequently, in the risk L (D, f ) (h S ). Formally, we say that it is a random variable. It is not
realistic to expect that with full certainty S will suffice to direct the learner toward
a good classifier (from the point of view of D), as there is always some probability
that the sampled training data happens to be very nonrepresentative of the underlying D. If we go back to the papaya tasting example, there is always some (small)
chance that all the papayas we have happened to taste were not tasty, in spite of the
fact that, say, 70% of the papayas in our island are tasty. In such a case, ERMH (S)
may be the constant function that labels every papaya as not tasty (and has 70%
error on the true distribution of papapyas in the island). We will therefore address
the probability to sample a training set for which L (D, f ) (h S ) is not too large. Usually, we denote the probability of getting a nonrepresentative sample by , and call
(1 ) the confidence parameter of our prediction.
On top of that, since we cannot guarantee perfect label prediction, we introduce
another parameter for the quality of prediction, the accuracy parameter, commonly
denoted by . We interpret the event L (D, f ) (h S ) > as a failure of the learner, while
if L (D, f ) (h S ) we view the output of the algorithm as an approximately correct
predictor. Therefore (fixing some labeling function f : X Y), we are interested
in upper bounding the probability to sample m-tuple of instances that will lead to
failure of the learner. Formally, let S|x = (x 1 , . . . , x m ) be the instances of the training
set. We would like to upper bound
Dm ({S|x : L (D, f ) (h S ) > }).
Let H B be the set of bad hypotheses, that is,
H B = {h H : L (D, f ) (h) > }.
In addition, let
M = {S|x : h H B , L S (h) = 0}
be the set of misleading samples: Namely, for every S|x M, there is a bad hypothesis, h H B , that looks like a good hypothesis on S|x . Now, recall that we would
like to bound the probability of the event L (D, f ) (h S ) > . But, since the realizability assumption implies that L S (h S ) = 0, it follows that the event L (D, f ) (h S ) > can
only happen if for some h H B we have L S (h) = 0. In other words, this event will

2.3 Empirical Risk Minimization with Inductive Bias

only happen if our sample is in the set of misleading samples, M. Formally, we have
shown that
{S|x : L (D, f ) (h S ) > } M.
Note that we can rewrite M as
M=

{S|x : L S (h) = 0}.

(2.5)

hH B

Hence,
Dm ({S|x : L (D, f ) (h S ) > }) Dm (M) = Dm ( hH B {S|x : L S (h) = 0}).

(2.6)

Next, we upper bound the right-hand side of the preceding equation using the
union bound a basic property of probabilities.
Lemma 2.2 (Union Bound).

For any two sets A, B and a distribution D we have

D(A B) D(A) + D(B).

Applying the union bound to the right-hand side of Equation (2.6) yields

Dm ({S|x : L S (h) = 0}).
(2.7)
Dm ({S|x : L (D, f ) (h S ) > })
hH B

Next, let us bound each summand of the right-hand side of the preceding inequality.
Fix some bad hypothesis h H B . The event L S (h) = 0 is equivalent to the event
i , h(x i ) = f (x i ). Since the examples in the training set are sampled i.i.d. we get that
Dm ({S|x : L S (h) = 0}) = Dm ({S|x : i , h(x i ) = f (x i )})
=

m


D({x i : h(x i ) = f (x i )}).

(2.8)

i=1

For each individual sampling of an element of the training set we have

D({x i : h(x i ) = yi }) = 1 L (D, f ) (h) 1 ,
where the last inequality follows from the fact that h H B . Combining the previous
equation with Equation (2.8) and using the inequality 1 e we obtain that for
every h H B ,
Dm ({S|x : L S (h) = 0}) (1 )m e m .

(2.9)

Combining this equation with Equation (2.7) we conclude that

Dm ({S|x : L (D, f ) (h S ) > }) |H B | e m |H| e m .
A graphical illustration which explains how we used the union bound is given in
Figure 2.1.
Corollary 2.3. Let H be a finite hypothesis class. Let (0, 1) and > 0 and let m be
an integer that satisfies
log (|H|/)
.
m

19

20

A Gentle Start

Figure 2.1. Each point in the large circle represents a possible m-tuple of instances. Each
colored oval represents the set of misleading m-tuple of instances for some bad predictor h H B . The ERM can potentially overfit whenever it gets a misleading training set
S. That is, for some h H B we have L S (h) = 0. Equation (2.9) guarantees that for each
individual bad hypothesis, h H B , at most (1 )m -fraction of the training sets would be
misleading. In particular, the larger m is, the smaller each of these colored ovals becomes.
The union bound formalizes the fact that the area representing the training sets that are
misleading with respect to some h H B (that is, the training sets in M) is at most the
sum of the areas of the colored ovals. Therefore, it is bounded by |H B | times the maximum
size of a colored oval. Any sample S outside the colored ovals cannot cause the ERM rule
to overfit.

Then, for any labeling function, f , and for any distribution, D, for which the realizability assumption holds (that is, for some h H, L (D, f ) (h) = 0), with probability of
at least 1 over the choice of an i.i.d. sample S of size m, we have that for every
ERM hypothesis, h S , it holds that
L (D, f ) (h S ) .
The preceeding corollary tells us that for a sufficiently large m, the ERMH rule
over a finite hypothesis class will be probably (with confidence 1 ) approximately
(up to an error of ) correct. In the next chapter we formally define the model of
Probably Approximately Correct (PAC) learning.

2.4 EXERCISES
2.1 Overfitting of polynomial matching: We have shown that the predictor defined in
Equation (2.3) leads to overfitting. While this predictor seems to be very unnatural,
the goal of this exercise is to show that it can be described as a thresholded polyd
m
nomial. That is, show that given a training set S = {(xi , f (xi ))}m
i=1 (R {0, 1}) ,
there exists a polynomial p S such that h S (x) = 1 if and only if p S (x) 0, where h S
is as defined in Equation (2.3). It follows that learning the class of all thresholded
polynomials using the ERM rule may lead to overfitting.
2.2 Let H be a class of binary classifiers over a domain X . Let D be an unknown distribution over X , and let f be the target hypothesis in H. Fix some h H. Show that
the expected value of L S (h) over the choice of S| x equals L (D, f ) (h), namely,
E

S|x D m

[L S (h)] = L (D, f ) (h).

2.3 Axis aligned rectangles: An axis aligned rectangle classifier in the plane is a classifier that assigns the value 1 to a point if and only if it is inside a certain rectangle.

2.4 Exercises
R*

+
R(S)

+
+

+
R1

Figure 2.2. Axis aligned rectangles.

Formally, given real numbers a1 b1 , a2 b2 , define the classifier h (a1 ,b1 ,a2 ,b2 ) by

1 if a1 x1 b1 and a2 x2 b2
h (a1 ,b1 ,a2 ,b2 ) (x1 , x2 ) =
.
(2.10)
0 otherwise
The class of all axis aligned rectangles in the plane is defined as
H2rec = {h (a1 ,b1 ,a2 ,b2 ) : a1 b1 , and a2 b2 }.

Note that this is an infinite size hypothesis class. Throughout this exercise we rely
on the realizability assumption.
1. Let A be the algorithm that returns the smallest rectangle enclosing all positive
examples in the training set. Show that A is an ERM.
2. Show that if A receives a training set of size 4log (4/) then, with probability of
at least 1 it returns a hypothesis with error of at most .
Hint: Fix some distribution D over X , let R = R(a1 , b1 , a2 , b2 ) be the rectangle that generates the labels, and let f be the corresponding hypothesis. Let
a1 a1 be a number such that the probability mass (with respect to D) of the
rectangle R1 = R(a1 , a1 , a2 , b2 ) is exactly /4. Similarly, let b1 , a2 , b2 be numbers
such that the probability masses of the rectangles R2 = R(b1 , b1 , a2 , b2 ), R3 =
R(a1 , b1 , a2 , a2 ), R4 = R(a1 , b1 , b2 , b2 ) are all exactly /4. Let R(S) be the
rectangle returned by A. See illustration in Figure 2.2.

Show that R(S) R .

Show that if S contains (positive) examples in all of the rectangles
R1 , R2 , R3 , R4 , then the hypothesis returned by A has error of at most .

For each i {1, . . . , 4}, upper bound the probability that S does not contain
an example from Ri .

Use the union bound to conclude the argument.
3. Repeat the previous question for the class of axis aligned rectangles in Rd .
4. Show that the runtime of applying the algorithm A mentioned earlier is polynomial in d, 1/ , and in log (1/).

21

3
A Formal Learning Model

In this chapter we define our main formal learning model the PAC learning model
and its extensions. We will consider other notions of learnability in Chapter 7.

3.1 PAC LEARNING

In the previous chapter we have shown that for a finite hypothesis class, if the ERM
rule with respect to that class is applied on a sufficiently large training sample (whose
size is independent of the underlying distribution or labeling function) then the output hypothesis will be probably approximately correct. More generally, we now
defineProbably Approximately Correct (PAC) learning.
Definition 3.1 (PAC Learnability). A hypothesis class H is PAC learnable if there
exist a function m H : (0, 1)2 N and a learning algorithm with the following property: For every , (0, 1), for every distribution D over X , and for every labeling
function f : X {0, 1}, if the realizable assumption holds with respect to H, D, f ,
then when running the learning algorithm on m m H ( , ) i.i.d. examples generated by D and labeled by f , the algorithm returns a hypothesis h such that, with
probability of at least 1 (over the choice of the examples), L (D, f ) (h) .
The definition of Probably Approximately Correct learnability contains two
approximation parameters. The accuracy parameter determines how far the output classifier can be from the optimal one (this corresponds to the approximately
correct), and a confidence parameter indicating how likely the classifier is to meet
that accuracy requirement (corresponds to the probably part of PAC). Under
the data access model that we are investigating, these approximations are inevitable.
Since the training set is randomly generated, there may always be a small chance that
it will happen to be noninformative (for example, there is always some chance that
the training set will contain only one domain point, sampled over and over again).
Furthermore, even when we are lucky enough to get a training sample that does
faithfully represent D, because it is just a finite sample, there may always be some
fine details of D that it fails to reflect. Our accuracy parameter, , allows forgiving
the learners classifier for making minor errors.
22

3.2 A More General Learning Model

Sample Complexity
The function m H : (0, 1)2 N determines the sample complexity of learning H: that
is, how many examples are required to guarantee a probably approximately correct
solution. The sample complexity is a function of the accuracy ( ) and confidence ()
parameters. It also depends on properties of the hypothesis class H for example,
for a finite class we showed that the sample complexity depends on log the size of H.
Note that if H is PAC learnable, there are many functions m H that satisfy the
requirements given in the definition of PAC learnability. Therefore, to be precise,
we will define the sample complexity of learning H to be the minimal function,
in the sense that for any , , m H ( , ) is the minimal integer that satisfies the
requirements of PAC learning with accuracy and confidence .
Let us now recall the conclusion of the analysis of finite hypothesis classes from
the previous chapter. It can be rephrased as stating:
Corollary 3.2. Every finite hypothesis class is PAC learnable with sample complexity


log (|H|/)
m H ( , )
.

There are infinite classes that are learnable as well (see, for example, Exercise
3.3). Later on we will show that what determines the PAC learnability of a class is
not its finiteness but rather a combinatorial measure called the VC dimension.

3.2 A MORE GENERAL LEARNING MODEL

The model we have just described can be readily generalized, so that it can be
made relevant to a wider scope of learning tasks. We consider generalizations in
two aspects:
Removing the Realizability Assumption
We have required that the learning algorithm succeeds on a pair of data distribution D and labeling function f provided that the realizability assumption is met. For
practical learning tasks, this assumption may be too strong (can we really guarantee that there is a rectangle in the color-hardness space that fully determines which
papayas are tasty?). In the next subsection, we will describe the agnostic PAC model
in which this realizability assumption is waived.
Learning Problems beyond Binary Classification
The learning task that we have been discussing so far has to do with predicting a
binary label to a given example (like being tasty or not). However, many learning
tasks take a different form. For example, one may wish to predict a real valued
number (say, the temperature at 9:00 p.m. tomorrow) or a label picked from a finite
set of labels (like the topic of the main story in tomorrows paper). It turns out
that our analysis of learning can be readily extended to such and many other scenarios by allowing a variety of loss functions. We shall discuss that in Section 3.2.2
later.

23

24

A Formal Learning Model

3.2.1 Releasing the Realizability Assumption Agnostic

PAC Learning
A More Realistic Model for the Data-Generating Distribution
Recall that the realizability assumption requires that there exists h  H such that
PxD [h  (x) = f (x)] = 1. In many practical problems this assumption does not hold.
Furthermore, it is maybe more realistic not to assume that the labels are fully determined by the features we measure on input elements (in the case of the papayas,
it is plausible that two papayas of the same color and softness will have different taste). In the following, we relax the realizability assumption by replacing the
target labeling function with a more flexible notion, a data-labels generating
distribution.
Formally, from now on, let D be a probability distribution over X Y, where,
as before, X is our domain set and Y is a set of labels (usually we will consider
Y = {0, 1}). That is, D is a joint distribution over domain points and labels. One can
view such a distribution as being composed of two parts: a distribution Dx over unlabeled domain points (sometimes called the marginal distribution) and a conditional
probability over labels for each domain point, D((x, y)|x). In the papaya example,
Dx determines the probability of encountering a papaya whose color and hardness
fall in some color-hardness values domain, and the conditional probability is the
probability that a papaya with color and hardness represented by x is tasty. Indeed,
such modeling allows for two papayas that share the same color and hardness to
belong to different taste categories.

The empirical and the True Error Revised

For a probability distribution, D, over X Y, one can measure how likely h is to
make an error when labeled points are randomly drawn according to D. We redefine
the true error (or risk) of a prediction rule h to be
def

L D (h) =

def

(x,y)D

[h(x) = y] = D({(x, y) : h(x) = y}).

(3.1)

We would like to find a predictor, h, for which that error will be minimized.
However, the learner does not know the data generating D. What the learner does
have access to is the training data, S. The definition of the empirical risk remains
the same as before, namely,
def

L S (h) =

|{i [m] : h(x i ) = yi }|

.
m

Given S, a learner can compute L S (h) for any function h : X {0, 1}. Note that
L S (h) = L D(uniform over S) (h).

The Goal
We wish to find some hypothesis, h : X Y, that (probably approximately)
minimizes the true risk, L D (h).

3.2 A More General Learning Model

The Bayes Optimal Predictor.

Given any probability distribution D over X {0, 1}, the best label predicting
function from X to {0, 1} will be

1 if P [ y = 1|x] 1/2
f D (x) =
0 otherwise
It is easy to verify (see Exercise 3.7) that for every probability distribution D, the
Bayes optimal predictor f D is optimal, in the sense that no other classifier, g : X
{0, 1}, has a lower error. That is, for every classifier g, L D ( f D ) L D (g).
Unfortunately, since we do not know D, we cannot utilize this optimal predictor
f D . What the learner does have access to is the training sample. We can now present
the formal definition of agnostic PAC learnability, which is a natural extension of
the definition of PAC learnability to the more realistic, nonrealizable, learning setup
we have just discussed.
Clearly, we cannot hope that the learning algorithm will find a hypothesis whose
error is smaller than the minimal possible error, that of the Bayes predictor. Furthermore, as we shall prove later, once we make no prior assumptions about the
data-generating distribution, no algorithm can be guaranteed to find a predictor that
is as good as the Bayes optimal one. Instead, we require that the learning algorithm
will find a predictor whose error is not much larger than the best possible error of a
predictor in some given benchmark hypothesis class. Of course, the strength of such
a requirement depends on the choice of that hypothesis class.
Definition 3.3 (Agnostic PAC Learnability). A hypothesis class H is agnostic PAC
learnable if there exist a function m H : (0, 1)2 N and a learning algorithm with the
following property: For every , (0, 1) and for every distribution D over X Y,
when running the learning algorithm on m m H ( , ) i.i.d. examples generated by
D, the algorithm returns a hypothesis h such that, with probability of at least 1
(over the choice of the m training examples),
L D (h  ) + .
L D (h) min

h H

Clearly, if the realizability assumption holds, agnostic PAC learning provides the
same guarantee as PAC learning. In that sense, agnostic PAC learning generalizes
the definition of PAC learning. When the realizability assumption does not hold, no
learner can guarantee an arbitrarily small error. Nevertheless, under the definition
of agnostic PAC learning, a learner can still declare success if its error is not much
larger than the best error achievable by a predictor from the class H. This is in
contrast to PAC learning, in which the learner is required to achieve a small error in
absolute terms and not relative to the best error achievable by the hypothesis class.

3.2.2 The Scope of Learning Problems Modeled

We next extend our model so that it can be applied to a wide variety of learning
tasks. Let us consider some examples of different learning tasks.

Multiclass Classification Our classification does not have to be binary. Take, for
example, the task of document classification: We wish to design a program that

25

26

A Formal Learning Model

will be able to classify given documents according to topics (e.g., news, sports,
biology, medicine). A learning algorithm for such a task will have access to
examples of correctly classified documents and, on the basis of these examples,
should output a program that can take as input a new document and output
a topic classification for that document. Here, the domain set is the set of all
potential documents. Once again, we would usually represent documents by a
set of features that could include counts of different key words in the document,
as well as other possibly relevant features like the size of the document or its origin. The label set in this task will be the set of possible document topics (so Y will
be some large finite set). Once we determine our domain and label sets, the other
components of our framework look exactly the same as in the papaya tasting
example; Our training sample will be a finite sequence of (feature vector, label)
pairs, the learners output will be a function from the domain set to the label
set, and, finally, for our measure of success, we can use the probability, over
(document, topic) pairs, of the event that our predictor suggests a wrong label.

Regression In this task, one wishes to find some simple pattern in the data a
functional relationship between the X and Y components of the data. For example, one wishes to find a linear function that best predicts a babys birth weight
on the basis of ultrasound measures of his head circumference, abdominal circumference, and femur length. Here, our domain set X is some subset of R3 (the
three ultrasound measurements), and the set of labels, Y, is the the set of real
numbers (the weight in grams). In this context, it is more adequate to call Y the
target set. Our training data as well as the learners output are as before (a finite
sequence of (x, y) pairs, and a function from X to Y respectively). However,
our measure of success is different. We may evaluate the quality of a hypothesis
function, h : X Y, by the expected square difference between the true labels
and their predicted values, namely,
def

L D (h) =

(x,y)D

(h(x) y)2 .

(3.2)

To accommodate a wide range of learning tasks we generalize our formalism of

the measure of success as follows:
Generalized Loss Functions
Given any set H (that plays the role of our hypotheses, or models) and some domain
Z let
be any function from H Z to the set of nonnegative real numbers,
:
H Z R+ . We call such functions loss functions.
Note that for prediction problems, we have that Z = X Y. However, our notion
of the loss function is generalized beyond prediction tasks, and therefore it allows
Z to be any domain of examples (for instance, in unsupervised learning tasks such
as the one described in Chapter 22, Z is not a product of an instance domain and a
label domain).
We now define the risk function to be the expected loss of a classifier, h H, with
respect to a probability distribution D over Z , namely,
def

L D (h) =

E [
(h, z)].

zD

(3.3)

3.2 A More General Learning Model

That is, we consider the expectation of the loss of h over objects z picked randomly according to D. Similarly, we define the empirical risk to be the expected loss
over a given sample S = (z 1 , . . . , z m ) Z m , namely,
1

(h, z i ).
m
m

def

L S (h) =

(3.4)

i=1

The loss functions used in the preceding examples of classification and regression
tasks are as follows:

01 loss: Here, our random variable z ranges over the set of pairs X Y and the
loss function is

0 if h(x) = y
def

01 (h, (x, y)) =
1 if h(x) = y

This loss function is used in binary or multiclass classification problems.

One should note that, for a random variable, , taking the values {0, 1},
ED [] = PD [ = 1]. Consequently, for this loss function, the definitions
of L D (h) given in Equation (3.3) and Equation (3.1) coincide.

Square Loss: Here, our random variable z ranges over the set of pairs X Y and
the loss function is
def

sq (h, (x, y)) = (h(x) y)2 .
This loss function is used in regression problems.
We will later see more examples of useful instantiations of loss functions.
To summarize, we formally define agnostic PAC learnability for general loss
functions.
Definition 3.4 (Agnostic PAC Learnability for General Loss Functions). A hypothesis class H is agnostic PAC learnable with respect to a set Z and a loss function

: H Z R+ , if there exist a function m H : (0, 1)2 N and a learning algorithm
with the following property: For every , (0, 1) and for every distribution D over
Z , when running the learning algorithm on m m H ( , ) i.i.d. examples generated
by D, the algorithm returns h H such that, with probability of at least 1 (over
the choice of the m training examples),
L D (h  ) + ,
L D (h) min

h H

where L D (h) = EzD [
(h, z)].

Remark 3.1 (A Note About Measurability*). In the aforementioned definition, for
every h H, we view the function
(h, ) : Z R+ as a random variable and define
L D (h) to be the expected value of this random variable. For that, we need to require
that the function
(h, ) is measurable. Formally, we assume that there is a -algebra
of subsets of Z , over which the probability D is defined, and that the preimage
of every initial segment in R+ is in this -algebra. In the specific case of binary
classification with the 01 loss, the -algebra is over X {0, 1} and our assumption
on
is equivalent to the assumption that for every h, the set {(x, h(x)) : x X } is in
the -algebra.

27

28

A Formal Learning Model

Remark 3.2 (Proper vs. Representation-Independent Learning*). In the preceding definition, we required that the algorithm will return a hypothesis from H. In
some situations, H is a subset of a set H , and the loss function can be naturally
extended to be a function from H Z to the reals. In this case, we may allow
the algorithm to return a hypothesis h  H , as long as it satisfies the requirement
L D (h  ) minhH L D (h) + . Allowing the algorithm to output a hypothesis from
H is called representation independent learning, while proper learning occurs when
the algorithm must output a hypothesis from H. Representation independent learning is sometimes called improper learning, although there is nothing improper in
representation independent learning.

3.3 SUMMARY
In this chapter we defined our main formal learning model PAC learning. The
basic model relies on the realizability assumption, while the agnostic variant does
not impose any restrictions on the underlying distribution over the examples. We
also generalized the PAC model to arbitrary loss functions. We will sometimes refer
to the most general model simply as PAC learning, omitting the agnostic prefix
and letting the reader infer what the underlying loss function is from the context.
When we would like to emphasize that we are dealing with the original PAC setting
we mention that the realizability assumption holds. In Chapter 7 we will discuss
other notions of learnability.

3.4 BIBLIOGRAPHIC REMARKS

Our most general definition of agnostic PAC learning with general loss functions
follows the works of Vladimir Vapnik and Alexey Chervonenkis (Vapnik and
Chervonenkis 1971). In particular, we follow Vapniks general setting of learning
(Vapnik 1982, Vapnik 1992, Vapnik 1995, Vapnik 1998).
The term PAC learning was introduced by Valiant (1984). Valiant was named
the winner of the 2010 Turing Award for the introduction of the PAC model.
Valiants definition requires that the sample complexity will be polynomial in 1/
and in 1/, as well as in the representation size of hypotheses in the class (see also
Kearns and Vazirani (1994)). As we will see in Chapter 6, if a problem is at all PAC
learnable then the sample complexity depends polynomially on 1/ and log (1/).
Valiants definition also requires that the runtime of the learning algorithm will be
polynomial in these quantities. In contrast, we chose to distinguish between the
statistical aspect of learning and the computational aspect of learning. We will elaborate on the computational aspect later on in Chapter 8. Finally, the formalization
of agnostic PAC learning is due to Haussler (1992).

3.5 EXERCISES
3.1 Monotonicity of Sample Complexity: Let H be a hypothesis class for a binary classification task. Suppose that H is PAC learnable and its sample complexity is given
by m H (, ). Show that m H is monotonically nonincreasing in each of its parameters. That is, show that given (0, 1), and given 0 < 1 2 < 1, we have that

3.5 Exercises

m H ( 1 , ) m H ( 2 , ). Similarly, show that given (0, 1), and given 0 < 1 2 < 1,
we have that m H ( , 1 ) m H ( , 2 ).
3.2 Let X be a discrete domain, and let HSingleton = {h z : z X } {h }, where for each
z X , h z is the function defined by h z (x) = 1 if x = z and h z (x) = 0 if x = z. h
is simply the all-negative hypothesis, namely, x X, h (x) = 0. The realizability
assumption here implies that the true hypothesis f labels negatively all examples in
the domain, perhaps except one.
1. Describe an algorithm that implements the ERM rule for learning HSingleton in
the realizable setup.
2. Show that HSingleton is PAC learnable. Provide an upper bound on the sample
complexity.
3.3 Let X = R2 , Y = {0, 1}, and let H be the class of concentric circles in the plane, that
is, H = {h r : r R+ }, where h r (x) = 1[xr ] . Prove that H is PAC learnable (assume
realizability), and its sample complexity is bounded by


log (1/)
m H ( , )
.

3.4 In this question, we study the hypothesis class of Boolean conjunctions defined as
follows. The instance space is X = {0, 1}d and the label set is Y = {0, 1}. A literal over
the variables x1 , . . ., xd is a simple Boolean function that takes the form f (x) = xi , for
some i [d], or f (x) = 1 xi for some i [d]. We use the notation x i as a shorthand
for 1 xi . A conjunction is any product of literals. In Boolean logic, the product is
denoted using the sign. For example, the function h(x) = x1 (1 x2 ) is written as
x1 x 2 .
We consider the hypothesis class of all conjunctions of literals over the d variables. The empty conjunction is interpreted as the all-positive hypothesis (namely,
the function that returns h(x) = 1 for all x). The conjunction x1 x 1 (and similarly
any conjunction involving a literal and its negation) is allowed and interpreted as
the all-negative hypothesis (namely, the conjunction that returns h(x) = 0 for all x).
We assume realizability: Namely, we assume that there exists a Boolean conjunction
that generates the labels. Thus, each example (x, y) X Y consists of an assignment to the d Boolean variables x1 , . . ., xd , and its truth value (0 for false and 1 for
true).
For instance, let d = 3 and suppose that the true conjunction is x1 x 2 . Then, the
training set S might contain the following instances:
((1, 1, 1), 0), ((1, 0, 1), 1), ((0, 1, 0), 0)((1, 0, 0), 1).
Prove that the hypothesis class of all conjunctions over d variables is PAC learnable and bound its sample complexity. Propose an algorithm that implements the
ERM rule, whose runtime is polynomial in d m.
3.5 Let X be a domain and let D1 , D2 , . . ., Dm be a sequence of distributions over X . Let
H be a finite class of binary classifiers over X and let f H. Suppose we are getting
a sample S of m examples, such that the instances are independent but are not identically distributed; the i th instance is sampled from Di and then yi is set to be f (xi ).
m denote the average, that is, D m = (D1 + + Dm )/m.
Let D
Fix an accuracy parameter (0, 1). Show that


P h H s.t. L (D m , f ) (h) > and L (S, f ) (h) = 0 |H|e m .

Hint: Use the geometric-arithmetic mean inequality.

29

30

A Formal Learning Model

3.6 Let H be a hypothesis class of binary classifiers. Show that if H is agnostic PAC
learnable, then H is PAC learnable as well. Furthermore, if A is a successful agnostic
PAC learner for H, then A is also a successful PAC learner for H.
3.7 (*) The Bayes optimal predictor: Show that for every probability distribution D, the
Bayes optimal predictor fD is optimal, in the sense that for every classifier g from
X to {0, 1}, L D ( f D ) L D (g).
3.8 (*) We say that a learning algorithm A is better than B with respect to some
probability distribution, D, if
L D ( A(S)) L D (B(S))
for all samples S (X {0, 1})m . We say that a learning algorithm A is better than B,
if it is better than B with respect to all probability distributions D over X {0, 1}.
1. A probabilistic label predictor is a function that assigns to every domain point
x a probability value, h(x) [0, 1], that determines the probability of predicting
the label 1. That is, given such an h and an input, x, the label for x is predicted by
tossing a coin with bias h(x) toward Heads and predicting 1 iff the coin comes up
Heads. Formally, we define a probabilistic label predictor as a function, h : X
[0, 1]. The loss of such h on an example (x, y) is defined to be |h(x) y|, which is
exactly the probability that the prediction of h will not be equal to y. Note that
if h is deterministic, that is, returns values in {0, 1}, then |h(x) y| = 1[h(x)= y] .
Prove that for every data-generating distribution D over X {0, 1}, the Bayes
optimal predictor has the smallest risk (w.r.t. the loss function
(h, (x, y)) =
|h(x) y|, among all possible label predictors, including probabilistic ones).
2. Let X be a domain and {0, 1} be a set of labels. Prove that for every distribution
D over X {0, 1}, there exist a learning algorithm AD that is better than any
other learning algorithm with respect to D.
3. Prove that for every learning algorithm A there exist a probability distribution,
D, and a learning algorithm B such that A is not better than B w.r.t. D.
3.9 Consider a variant of the PAC model in which there are two example oracles: one
that generates positive examples and one that generates negative examples, both
according to the underlying distribution D on X . Formally, given a target function
f : X {0, 1}, let D+ be the distribution over X + = {x X : f (x) = 1} defined by
D+ ( A) = D( A)/D(X + ), for every A X + . Similarly, D is the distribution over X
induced by D.
The definition of PAC learnability in the two-oracle model is the same as the
standard definition of PAC learnability except that here the learner has access to
+

m+
H ( , ) i.i.d. examples from D and m ( , ) i.i.d. examples from D . The learners
goal is to output h s.t. with probability at least 1 (over the choice of the two
training sets, and possibly over the nondeterministic decisions made by the learning
algorithm), both L (D+ , f ) (h) and L (D, f ) (h) .
1. (*) Show that if H is PAC learnable (in the standard one-oracle model), then H
is PAC learnable in the two-oracle model.
2. (**) Define h + to be the always-plus hypothesis and h to be the always-minus
hypothesis. Assume that h + , h H. Show that if H is PAC learnable in the
two-oracle model, then H is PAC learnable in the standard one-oracle model.

4
Learning via Uniform Convergence

The first formal learning model that we have discussed was the PAC model. In
Chapter 2 we have shown that under the realizability assumption, any finite hypothesis class is PAC learnable. In this chapter we will develop a general tool, uniform
convergence, and apply it to show that any finite class is learnable in the agnostic PAC model with general loss functions, as long as the range loss function is
bounded.

4.1 UNIFORM CONVERGENCE IS SUFFICIENT FOR LEARNABILITY

The idea behind the learning condition discussed in this chapter is very simple.
Recall that, given a hypothesis class, H, the ERM learning paradigm works as follows: Upon receiving a training sample, S, the learner evaluates the risk (or error)
of each h in H on the given sample and outputs a member of H that minimizes this
empirical risk. The hope is that an h that minimizes the empirical risk with respect to
S is a risk minimizer (or has risk close to the minimum) with respect to the true data
probability distribution as well. For that, it suffices to ensure that the empirical risks
of all members of H are good approximations of their true risk. Put another way, we
need that uniformly over all hypotheses in the hypothesis class, the empirical risk
will be close to the true risk, as formalized in the following.
Definition 4.1 ( -representative sample). A training set S is called -representative
(w.r.t. domain Z , hypothesis class H, loss function
, and distribution D) if
h H, |L S (h) L D (h)| .
The next simple lemma states that whenever the sample is ( /2)-representative,
the ERM learning rule is guaranteed to return a good hypothesis.
Lemma 4.2. Assume that a training set S is 2 -representative (w.r.t. domain Z ,
hypothesis class H, loss function
, and distribution D). Then, any output of
ERMH (S), namely, any h S argminhH L S (h), satisfies
L D (h S ) min L D (h) + .
hH

31

32

Learning via Uniform Convergence

Proof. For every h H,

L D (h S ) L S (h S ) + 2 L S (h) + 2 L D (h) + 2 + 2 = L D (h) + ,
where the first and third inequalities are due to the assumption that S is

2 -representative (Definition 4.1) and the second inequality holds since h S is an ERM
predictor.
The preceding lemma implies that to ensure that the ERM rule is an agnostic
PAC learner, it suffices to show that with probability of at least 1 over the random choice of a training set, it will be an -representative training set. The uniform
convergence condition formalizes this requirement.
Definition 4.3 (Uniform Convergence). We say that a hypothesis class H has the
uniform convergence property (w.r.t. a domain Z and a loss function
) if there exists
2
a function m UC
H : (0, 1) N such that for every , (0, 1) and for every probability distribution D over Z , if S is a sample of m m UC
H ( , ) examples drawn i.i.d.
according to D, then, with probability of at least 1 , S is -representative.
Similar to the definition of sample complexity for PAC learning, the function
m H measures the (minimal) sample complexity of obtaining the uniform convergence property, namely, how many examples we need to ensure that with
probability of at least 1 the sample would be -representative.
The term uniform here refers to having a fixed sample size that works for all
members of H and over all possible probability distributions over the domain.
The following corollary follows directly from Lemma 4.2 and the definition of
uniform convergence.
UC

Corollary 4.4. If a class H has the uniform convergence property with a function m UC
H
then the class is agnostically PAC learnable with the sample complexity m H ( , )
m UC
H ( /2, ). Furthermore, in that case, the ERMH paradigm is a successful agnostic
PAC learner for H.

4.2 FINITE CLASSES ARE AGNOSTIC PAC LEARNABLE

In view of Corollary 4.4, the claim that every finite hypothesis class is agnostic PAC
learnable will follow once we establish that uniform convergence holds for a finite
hypothesis class.
To show that uniform convergence holds we follow a two step argument, similar
to the derivation in Chapter 2. The first step applies the union bound while the
second step employs a measure concentration inequality. We now explain these two
steps in detail.
Fix some , . We need to find a sample size m that guarantees that for any D,
with probability of at least 1 of the choice of S = (z 1 , . . . , z m ) sampled i.i.d. from
D we have that for all h H, |L S (h) L D (h)| . That is,
Dm ({S : h H, |L S (h) L D (h)| }) 1 .
Equivalently, we need to show that
Dm ({S : h H, |L S (h) L D (h)| > }) < .

4.2 Finite Classes Are Agnostic PAC Learnable

Writing
{S : h H, |L S (h) L D (h)| > } = hH {S : |L S (h) L D (h)| > },
and applying the union bound (Lemma 2.2) we obtain

Dm ({S : h H, |L S (h) L D (h)| > })
Dm ({S : |L S (h) L D (h)| > }). (4.1)
hH

Our second step will be to argue that each summand of the right-hand side of
this inequality is small enough (for a sufficiently large m). That is, we will show that
for any fixed hypothesis, h, (which is chosen in advance prior to the sampling of the
training set), the gap between the true and empirical risks, |L S (h) L D (h)|, is likely
to be small.


Recall that L D (h) = EzD [
(h, z)] and that L S (h) = m1 m
i=1
(h, z i ). Since each z i
is sampled i.i.d. from D, the expected value of the random variable
(h, z i ) is L D (h).
By the linearity of expectation, it follows that L D (h) is also the expected value of
L S (h). Hence, the quantity |L D (h) L S (h)| is the deviation of the random variable
L S (h) from its expectation. We therefore need to show that the measure of L S (h) is
concentrated around its expected value.
A basic statistical fact, the law of large numbers, states that when m goes to
infinity, empirical averages converge to their true expectation. This is true for L S (h),
since it is the empirical average of m i.i.d random variables. However, since the law
of large numbers is only an asymptotic result, it provides no information about the
gap between the empirically estimated error and its true value for any given, finite,
sample size.
Instead, we will use a measure concentration inequality due to Hoeffding, which
quantifies the gap between empirical averages and their expected value.
Lemma 4.5 (Hoeffdings Inequality). Let 1 , . . . , m be a sequence of i.i.d. random
variables and assume that for all i , E [i ] = and P [a i b] = 1. Then, for any
>0

 m






1
i > 2 exp 2 m 2 /(b a)2 .
P m


i=1

The proof can be found in Appendix B.

Getting back to our problem, let i be the random variable
(h, z i ). Since h is
fixed and z 1 , . . . , z m are sampled i.i.d.,

it follows that 1 , . . . , m are also i.i.d. random
variables. Furthermore, L S (h) = m1 m
i=1 i and L D (h) = . Let us further assume
that the range of
is [0, 1] and therefore i [0, 1]. We therefore obtain that

 m





1

m
D ({S : |L S (h) L D (h)| > }) = P m
i > 2 exp 2 m 2 . (4.2)


i=1

Combining this with Equation (4.1) yields

Dm ({S : h H, |L S (h) L D (h)| > })


hH



2 exp 2 m 2



= 2 |H| exp 2 m 2 .

33

34

Learning via Uniform Convergence

Finally, if we choose
m

log (2|H|/)
2 2

then
Dm ({S : h H, |L S (h) L D (h)| > }) .
Corollary 4.6. Let H be a finite hypothesis class, let Z be a domain, and let
: H
Z [0, 1] be a loss function. Then, H enjoys the uniform convergence property with
sample complexity


log (2|H|/)
m UC
( ,
)

.
H
2 2
Furthermore, the class is agnostically PAC learnable using the ERM algorithm with
sample complexity


2 log (2|H|/)
m H ( , ) m UC
( /2,
)

.
H
2
Remark 4.1 (The Discretization Trick). While the preceding corollary only
applies to finite hypothesis classes, there is a simple trick that allows us to get a
very good estimate of the practical sample complexity of infinite hypothesis classes.
Consider a hypothesis class that is parameterized by d parameters. For example,
let X = R, Y = {1}, and the hypothesis class, H, be all functions of the form
h (x) = sign(x ). That is, each hypothesis is parameterized by one parameter,
R, and the hypothesis outputs 1 for all instances larger than and outputs 1
for instances smaller than . This is a hypothesis class of an infinite size. However,
if we are going to learn this hypothesis class in practice, using a computer, we will
probably maintain real numbers using floating point representation, say, of 64 bits.
It follows that in practice, our hypothesis class is parameterized by the set of scalars
that can be represented using a 64 bits floating point number. There are at most 264
such numbers; hence the actual size of our hypothesis class is at most 264 . More generally, if our hypothesis class is parameterized by d numbers, in practice we learn
a hypothesis class of size at most 264d . Applying Corollary 4.6 we obtain that the
. This upper bound
sample complexity of such classes is bounded by 128d+2 log(2/)
2
on the sample complexity has the deficiency of being dependent on the specific representation of real numbers used by our machine. In Chapter 6 we will introduce
a rigorous way to analyze the sample complexity of infinite size hypothesis classes.
Nevertheless, the discretization trick can be used to get a rough estimate of the
sample complexity in many practical situations.

4.3 SUMMARY
If the uniform convergence property holds for a hypothesis class H then in most
cases the empirical risks of hypotheses in H will faithfully represent their true
risks. Uniform convergence suffices for agnostic PAC learnability using the ERM
rule. We have shown that finite hypothesis classes enjoy the uniform convergence
property and are hence agnostic PAC learnable.

4.5 Exercises

4.4 BIBLIOGRAPHIC REMARKS

Classes of functions for which the uniform convergence property holds are also
called Glivenko-Cantelli classes, named after Valery Ivanovich Glivenko and
Francesco Paolo Cantelli, who proved the first uniform convergence result in the
1930s. See (Dudley, Gine & Zinn 1991). The relation between uniform convergence
and learnability was thoroughly studied by Vapnik see (Vapnik 1992, Vapnik 1995,
Vapnik 1998). In fact, as we will see later in Chapter 6, the fundamental theorem of
learning theory states that in binary classification problems, uniform convergence is
not only a sufficient condition for learnability but is also a necessary condition. This
is not the case for more general learning problems (see (Shalev-Shwartz, Shamir,
Srebro & Sridharan 2010)).

4.5 EXERCISES
4.1 In this exercise, we show that the ( , ) requirement on the convergence of errors in
our definitions of PAC learning, is, in fact, quite close to a simpler looking requirement about averages (or expectations). Prove that the following two statements are
equivalent (for any learning algorithm A, any probability distribution D, and any
loss function whose range is [0, 1]):
1. For every , > 0, there exists m( , ) such that m m( , )
P [L D ( A(S)) > ] <

SD m

2.
lim

E [L D ( A(S))] = 0

m SD m

(where E SDm denotes the expectation over samples S of size m).

4.2 Bounded loss functions: In Corollary 4.6 we assumed that the range of the loss function is [0, 1]. Prove that if the range of the loss function is [a, b] then the sample
complexity satisfies


2 log (2|H|/)(b a)2
UC
m H ( , ) m H ( /2, )
.
2

35

5
The Bias-Complexity Tradeoff

In Chapter 2 we saw that unless one is careful, the training data can mislead the
learner, and result in overfitting. To overcome this problem, we restricted the search
space to some hypothesis class H. Such a hypothesis class can be viewed as reflecting
some prior knowledge that the learner has about the task a belief that one of
the members of the class H is a low-error model for the task. For example, in our
papayas taste problem, on the basis of our previous experience with other fruits,
we may assume that some rectangle in the color-hardness plane predicts (at least
approximately) the papayas tastiness.
Is such prior knowledge really necessary for the success of learning? Maybe
there exists some kind of universal learner, that is, a learner who has no prior knowledge about a certain task and is ready to be challenged by any task? Let us elaborate
on this point. A specific learning task is defined by an unknown distribution D over
X Y, where the goal of the learner is to find a predictor h : X Y, whose risk,
L D (h), is small enough. The question is therefore whether there exist a learning
algorithm A and a training set size m, such that for every distribution D, if A receives
m i.i.d. examples from D, there is a high chance it outputs a predictor h that has a
low risk.
The first part of this chapter addresses this question formally. The No-FreeLunch theorem states that no such universal learner exists. To be more precise, the
theorem states that for binary classification prediction tasks, for every learner there
exists a distribution on which it fails. We say that the learner fails if, upon receiving
i.i.d. examples from that distribution, its output hypothesis is likely to have a large
risk, say, 0. 3, whereas for the same distribution, there exists another learner that
will output a hypothesis with a small risk. In other words, the theorem states that no
learner can succeed on all learnable tasks every learner has tasks on which it fails
while other learners succeed.
Therefore, when approaching a particular learning problem, defined by some
distribution D, we should have some prior knowledge on D. One type of such prior
knowledge is that D comes from some specific parametric family of distributions.
We will study learning under such assumptions later on in Chapter 24. Another type
of prior knowledge on D, which we assumed when defining the PAC learning model,
36

5.1 The No-Free-Lunch Theorem

is that there exists h in some predefined hypothesis class H, such that L D (h) = 0. A
softer type of prior knowledge on D is assuming that minhH L D (h) is small. In a
sense, this weaker assumption on D is a prerequisite for using the agnostic PAC
model, in which we require that the risk of the output hypothesis will not be much
larger than minhH L D (h).
In the second part of this chapter we study the benefits and pitfalls of using a
hypothesis class as a means of formalizing prior knowledge. We decompose the
error of an ERM algorithm over a class H into two components. The first component reflects the quality of our prior knowledge, measured by the minimal risk of a
hypothesis in our hypothesis class, minhH L D (h). This component is also called the
approximation error, or the bias of the algorithm toward choosing a hypothesis from
H. The second component is the error due to overfitting, which depends on the size
or the complexity of the class H and is called the estimation error. These two terms
imply a tradeoff between choosing a more complex H (which can decrease the bias
but increases the risk of overfitting) or a less complex H (which might increase the
bias but decreases the potential overfitting).

5.1 THE NO-FREE-LUNCH THEOREM

In this part we prove that there is no universal learner. We do this by showing that
no learner can succeed on all learning tasks, as formalized in the following theorem:
Theorem 5.1. (No-Free-Lunch) Let A be any learning algorithm for the task of
binary classification with respect to the 01 loss over a domain X . Let m be any number smaller than |X |/2, representing a training set size. Then, there exists a distribution
D over X {0, 1} such that:
1. There exists a function f : X {0, 1} with L D ( f ) = 0.
2. With probability of at least 1/7 over the choice of S Dm we have that
L D (A(S)) 1/8.
This theorem states that for every learner, there exists a task on which it fails,
even though that task can be successfully learned by another learner. Indeed, a
trivial successful learner in this case would be an ERM learner with the hypothesis class H = { f }, or more generally, ERM with respect to any finite hypothesis
class that contains f and whose size satisfies the equation m 8 log (7|H|/6) (see
Corollary 2.3).
Proof. Let C be a subset of X of size 2m. The intuition of the proof is that any
learning algorithm that observes only half of the instances in C has no information
on what should be the labels of the rest of the instances in C. Therefore, there exists
a reality, that is, some target function f , that would contradict the labels that A(S)
predicts on the unobserved instances in C.
Note that there are T = 22m possible functions from C to {0, 1}. Denote these
functions by f 1 , . . . , f T . For each such function, let Di be a distribution over C {0, 1}
defined by

1/|C| if y = f i (x)
Di ({(x, y)}) =
0
otherwise.

37

38

The Bias-Complexity Tradeoff

That is, the probability to choose a pair (x, y) is 1/|C| if the label y is indeed the true
label according to f i , and the probability is 0 if y = f i (x). Clearly, L Di ( fi ) = 0.
We will show that for every algorithm, A, that receives a training set of m
examples from C {0, 1} and returns a function A( S) : C {0, 1}, it holds that
max

E [ L Di ( A( S))] 1/4.

i[T ] SDim

(5.1)

Clearly, this means that for every algorithm, A , that receives a training set of m
examples from X {0, 1} there exist a function f : X {0, 1} and a distribution D
over X {0, 1}, such that L D ( f ) = 0 and
E [ L D ( A ( S))] 1/4.

SD m

(5.2)

It is easy to verify that the preceding suffices for showing that P [L D (A (S)) 1/8]
1/7, which is what we need to prove (see Exercise 5.1).
We now turn to proving that Equation (5.1) holds. There are k = (2m)m possible
sequences of m examples from C. Denote these sequences by S1 , . . . , Sk . Also, if
S j = (x 1 , . . . , x m ) we denote by S ij the sequence containing the instances in S j labeled
by the function f i , namely, S ij = ((x 1 , f i (x 1 )), . . . , (x m , fi (x m ))). If the distribution is
Di then the possible training sets A can receive are S1i , . . . , Ski , and all these training
sets have the same probability of being sampled. Therefore,

E m [L Di (A(S))] =

SDi

k
1
L Di (A(S ij )).
k

(5.3)

j =1

Using the facts that maximum is larger than average and that average is larger
than minimum, we have

max
i[T ]

k
T
k
1
1 1
L Di (A(S ij ))
L Di (A(S ij ))
k
T
k
j =1

i=1

j =1

k
T
1 1 
L Di (A(S ij ))
k
T
j =1

min

j [k]

i=1

T
1
L Di (A(S ij )).
T

(5.4)

i=1

Next, fix some j [k]. Denote S j = (x 1 , . . . , x m ) and let v1 , . . . , v p be the examples in

C that do not appear in S j . Clearly, p m. Therefore, for every function h : C

5.1 The No-Free-Lunch Theorem

{0, 1} and every i we have

L Di (h) =

1 
1[h(x)= f i (x)]
2m
xC

1 

1[h(vr )= f i (vr )]

2m
p

r=1

p
1 

1[h(vr )= f i (vr )] .

2p

(5.5)

r=1

Hence,
p
T
T
1
1 1 
i
L Di (A(S j ))
1[A(S i )(vr )= f i (vr )]
j
T
T
2p
i=1

i=1

1
2p

p

r=1

r=1

1
T

T


1[A(S i )(vr )= f i (vr )]

j

i=1

T
1
1
min
1[A(S i )(vr )= f i (vr )] .
j
2 r[ p] T

(5.6)

i=1

Next, fix some r [ p]. We can partition all the functions in f 1 , . . . , f T into T /2 disjoint pairs, where for a pair ( f i , fi  ) we have that for every c C, f i (c) = f i  (c) if and

only if c = vr . Since for such a pair we must have S ij = S ij , it follows that
1[A(S i )(vr )= f i (vr )] + 1[A(S i  )(v
j

r )= f i  (vr )]

= 1,

which yields
T
1 
1
1[A(S i )(vr )= f i (vr )] = .
j
T
2
i=1

Combining this with Equation (5.6), Equation (5.4), and Equation (5.3), we obtain
that Equation (5.1) holds, which concludes our proof.

5.1.1 No-Free-Lunch and Prior Knowledge

How does the No-Free-Lunch result relate to the need for prior knowledge? Let us
consider an ERM predictor over the hypothesis class H of all the functions f from
X to {0, 1}. This class represents lack of prior knowledge: Every possible function
from the domain to the label set is considered a good candidate. According to the
No-Free-Lunch theorem, any algorithm that chooses its output from hypotheses in
H, and in particular the ERM predictor, will fail on some learning task. Therefore,
this class is not PAC learnable, as formalized in the following corollary:
Corollary 5.2. Let X be an infinite domain set and let H be the set of all functions
from X to {0, 1}. Then, H is not PAC learnable.
Proof. Assume, by way of contradiction, that the class is learnable. Choose some
< 1/8 and < 1/7. By the definition of PAC learnability, there must be some

39

40

The Bias-Complexity Tradeoff

learning algorithm A and an integer m = m( , ), such that for any data-generating

distribution over X {0, 1}, if for some function f : X {0, 1}, L D ( f ) = 0, then with
probability greater than 1 when A is applied to samples S of size m, generated
i.i.d. by D, L D (A(S)) . However, applying the No-Free-Lunch theorem, since
|X | > 2m, for every learning algorithm (and in particular for the algorithm A), there
exists a distribution D such that with probability greater than 1/7 > , L D (A(S)) >
1/8 > , which leads to the desired contradiction.
How can we prevent such failures? We can escape the hazards foreseen by the
No-Free-Lunch theorem by using our prior knowledge about a specific learning
task, to avoid the distributions that will cause us to fail when learning that task.
Such prior knowledge can be expressed by restricting our hypothesis class.
But how should we choose a good hypothesis class? On the one hand, we want
to believe that this class includes the hypothesis that has no error at all (in the PAC
setting), or at least that the smallest error achievable by a hypothesis from this class
is indeed rather small (in the agnostic setting). On the other hand, we have just seen
that we cannot simply choose the richest class the class of all functions over the
given domain. This tradeoff is discussed in the following section.

5.2 ERROR DECOMPOSITION

To answer this question we decompose the error of an ERMH predictor into two
components as follows. Let h S be an ERMH hypothesis. Then, we can write
L D (h S ) = app + est where : app = min L D (h),
hH

est = L D (h S ) app .

(5.7)

The Approximation Error the minimum risk achievable by a predictor in

the hypothesis class. This term measures how much risk we have because we
restrict ourselves to a specific class, namely, how much inductive bias we have.
The approximation error does not depend on the sample size and is determined
by the hypothesis class chosen. Enlarging the hypothesis class can decrease the
approximation error.
Under the realizability assumption, the approximation error is zero. In the
agnostic case, however, the approximation error can be large.1

The Estimation Error the difference between the approximation error and the
error achieved by the ERM predictor. The estimation error results because the
empirical risk (i.e., training error) is only an estimate of the true risk, and so
the predictor minimizing the empirical risk is only an estimate of the predictor
minimizing the true risk.
The quality of this estimation depends on the training set size and on the size,
or complexity, of the hypothesis class. As we have shown, for a finite hypothesis class, est increases (logarithmically) with |H| and decreases with m. We can
1

In fact, it always includes the error of the Bayes optimal predictor (see Chapter 3), the minimal yet
inevitable error, because of the possible nondeterminism of the world in this model. Sometimes in the
literature the term approximation error refers not to minhH L D (h), but rather to the excess error over
that of the Bayes optimal predictor, namely, minhH L D (h) Bayes .

5.5 Exercises

think of the size of H as a measure of its complexity. In future chapters we will

define other complexity measures of hypothesis classes.
Since our goal is to minimize the total risk, we face a tradeoff, called the biascomplexity tradeoff. On one hand, choosing H to be a very rich class decreases the
approximation error but at the same time might increase the estimation error, as a
rich H might lead to overfitting. On the other hand, choosing H to be a very small
set reduces the estimation error but might increase the approximation error or, in
other words, might lead to underfitting. Of course, a great choice for H is the class
that contains only one classifier the Bayes optimal classifier. But the Bayes optimal
classifier depends on the underlying distribution D, which we do not know (indeed,
learning would have been unnecessary had we known D).
Learning theory studies how rich we can make H while still maintaining reasonable estimation error. In many cases, empirical research focuses on designing good
hypothesis classes for a certain domain. Here, good means classes for which the
approximation error would not be excessively high. The idea is that although we are
not experts and do not know how to construct the optimal classifier, we still have
some prior knowledge of the specific problem at hand, which enables us to design
hypothesis classes for which both the approximation error and the estimation error
are not too large. Getting back to our papayas example, we do not know how exactly
the color and hardness of a papaya predict its taste, but we do know that papaya is
rectangle in the color-hardness space may be a good predictor.

5.3 SUMMARY
The No-Free-Lunch theorem states that there is no universal learner. Every learner
has to be specified to some task, and use some prior knowledge about that task, in
order to succeed. So far we have modeled our prior knowledge by restricting our
output hypothesis to be a member of a chosen hypothesis class. When choosing
this hypothesis class, we face a tradeoff, between a larger, or more complex, class
that is more likely to have a small approximation error, and a more restricted class
that would guarantee that the estimation error will be small. In the next chapter we
will study in more detail the behavior of the estimation error. In Chapter 7 we will
discuss alternative ways to express prior knowledge.

5.4 BIBLIOGRAPHIC REMARKS

(Wolpert & Macready 1997) proved several no-free-lunch theorems for optimization, but these are rather different from the theorem we prove here. The theorem
we prove here is closely related to lower bounds in VC theory, as we will study in
the next chapter.

5.5 EXERCISES
5.1 Prove that Equation (5.2) suffices for showing that P [L D ( A(S)) 1/8] 1/7.
Hint: Let be a random variable that receives values in [0, 1] and whose expectation
satisfies E [ ] 1/4. Use Lemma B.1 to show that P [ 1/8] 1/7.

41

42

The Bias-Complexity Tradeoff

5.2 Assume you are asked to design a learning algorithm to predict whether patients
are going to suffer a heart attack. Relevant patient features the algorithm may have
access to include blood pressure (BP), body-mass index (BMI), age (A), level of
physical activity (P), and income (I).
You have to choose between two algorithms; the first picks an axis aligned rectangle in the two dimensional space spanned by the features BP and BMI and the
other picks an axis aligned rectangle in the five dimensional space spanned by all
the preceding features.
1. Explain the pros and cons of each choice.
2. Explain how the number of available labeled training samples will affect your
choice.
5.3 Prove that if |X | km for a positive integer k 2, then we can replace the lower
1
1
bound of 1/4 in the No-Free-Lunch theorem with k1
2k = 2 2k . Namely, let A be a
learning algorithm for the task of binary classification. Let m be any number smaller
than |X |/k, representing a training set size. Then, there exists a distribution D over
X {0, 1} such that:

There exists a function f : X {0, 1} with L D ( f ) = 0.

E SDm [L D ( A(S))] 12 2k1 .

6
The VC-Dimension

In the previous chapter, we decomposed the error of the ERMH rule into approximation error and estimation error. The approximation error depends on the fit
of our prior knowledge (as reflected by the choice of the hypothesis class H) to
the underlying unknown distribution. In contrast, the definition of PAC learnability requires that the estimation error would be bounded uniformly over all
distributions.
Our current goal is to figure out which classes H are PAC learnable, and to
characterize exactly the sample complexity of learning a given hypothesis class. So
far we have seen that finite classes are learnable, but that the class of all functions
(over an infinite size domain) is not. What makes one class learnable and the other
unlearnable? Can infinite-size classes be learnable, and, if so, what determines their
sample complexity?
We begin the chapter by showing that infinite classes can indeed be learnable, and thus, finiteness of the hypothesis class is not a necessary condition for
learnability. We then present a remarkably crisp characterization of the family of
learnable classes in the setup of binary valued classification with the zero-one loss.
This characterization was first discovered by Vladimir Vapnik and Alexey Chervonenkis in 1970 and relies on a combinatorial notion called the Vapnik-Chervonenkis
dimension (VC-dimension). We formally define the VC-dimension, provide several
examples, and then state the fundamental theorem of statistical learning theory,
which integrates the concepts of learnability, VC-dimension, the ERM rule, and
uniform convergence.

6.1 INFINITE-SIZE CLASSES CAN BE LEARNABLE

In Chapter 4 we saw that finite classes are learnable, and in fact the sample complexity of a hypothesis class is upper bounded by the log of its size. To show that the size
of the hypothesis class is not the right characterization of its sample complexity, we
first present a simple example of an infinite-size hypothesis class that is learnable.

43

44

The VC-Dimension

Example 6.1. Let H be the set of threshold functions over the real line, namely,
H = {h a : a R}, where h a : R {0, 1} is a function such that h a (x) = 1[x<a] . To
remind the reader, 1[x<a] is 1 if x < a and 0 otherwise. Clearly, H is of infinite size.
Nevertheless, the following lemma shows that H is learnable in the PAC model
using the ERM algorithm.
Lemma 6.1. Let H be the class of thresholds as defined earlier. Then, H is PAC
learnable, using the ERM rule, with sample complexity of m H ( , ) log (2/)/ .
Proof. Let a  be a threshold such that the hypothesis h  (x) = 1[x<a  ] achieves
L D (h  ) = 0. Let Dx be the marginal distribution over the domain X and let a0 <
a  < a1 be such that
P [x (a0 , a  )] = P [x (a  , a1 )] = .

xDx

xDx

mass

mass

a0

a1

a*

(If Dx ( , a  ) we set a0 = and similarly for a1 ). Given a training set S,

let b0 = max{x : (x, 1) S} and b1 = min{x : (x, 0) S} (if no example in S is positive
we set b0 = and if no example in S is negative we set b1 = ). Let b S be a
threshold corresponding to an ERM hypothesis, h S , which implies that b S (b0 , b1 ).
Therefore, a sufficient condition for L D (h S ) is that both b0 a0 and b1 a1 . In
other words,
P [L D (h S ) > ] P m [b0 < a0 b1 > a1 ],

SD m

SD

and using the union bound we can bound the preceding by

P [L D (h S ) > ] P m [b0 < a0 ] + P m [b1 > a1 ].

SD m

SD

SD

(6.1)

The event b0 < a0 happens if and only if all examples in S are not in the interval
(a0 , a ), whose probability mass is defined to be , namely,
P [b0 < a0 ] = P m [(x, y) S, x  (a0 , a  )] = (1 )m e m .

SD m

SD

Since we assume m > log (2/)/ it follows that the equation is at most /2. In the
same way it is easy to see that P SDm [b1 > a1 ] /2. Combining with Equation (6.1)
we conclude our proof.

6.2 THE VC-DIMENSION

We see, therefore, that while finiteness of H is a sufficient condition for learnability,
it is not a necessary condition. As we will show, a property called the VC-dimension
of a hypothesis class gives the correct characterization of its learnability. To motivate the definition of the VC-dimension, let us recall the No-Free-Lunch theorem
(Theorem 5.1) and its proof. There, we have shown that without restricting the
hypothesis class, for any learning algorithm, an adversary can construct a distribution for which the learning algorithm will perform poorly, while there is another

6.2 The VC-Dimension

learning algorithm that will succeed on the same distribution. To do so, the adversary used a finite set C X and considered a family of distributions that are
concentrated on elements of C. Each distribution was derived from a true target function from C to {0, 1}. To make any algorithm fail, the adversary used the
power of choosing a target function from the set of all possible functions from C to
{0, 1}.
When considering PAC learnability of a hypothesis class H, the adversary is
restricted to constructing distributions for which some hypothesis h H achieves a
zero risk. Since we are considering distributions that are concentrated on elements
of C, we should study how H behaves on C, which leads to the following definition.
Definition 6.2 (Restriction of H to C). Let H be a class of functions from X to {0, 1}
and let C = {c1 , . . . , cm } X . The restriction of H to C is the set of functions from C
to {0, 1} that can be derived from H. That is,
HC = {(h(c1 ), . . . , h(cm )) : h H},
where we represent each function from C to {0, 1} as a vector in {0, 1}|C| .
If the restriction of H to C is the set of all functions from C to {0, 1}, then we say
that H shatters the set C. Formally:
Definition 6.3 (Shattering). A hypothesis class H shatters a finite set C X if the
restriction of H to C is the set of all functions from C to {0, 1}. That is, |HC | = 2|C| .
Example 6.2. Let H be the class of threshold functions over R. Take a set C = {c1 }.
Now, if we take a = c1 + 1, then we have h a (c1 ) = 1, and if we take a = c1 1, then
we have h a (c1 ) = 0. Therefore, HC is the set of all functions from C to {0, 1}, and H
shatters C. Now take a set C = {c1 , c2 }, where c1 c2 . No h H can account for the
labeling (0, 1), because any threshold that assigns the label 0 to c1 must assign the
label 0 to c2 as well. Therefore not all functions from C to {0, 1} are included in HC ;
hence C is not shattered by H.
Getting back to the construction of an adversarial distribution as in the proof
of the No-Free-Lunch theorem (Theorem 5.1), we see that whenever some set C is
shattered by H, the adversary is not restricted by H, as they can construct a distribution over C based on any target function from C to {0, 1}, while still maintaining
the realizability assumption. This immediately yields:
Corollary 6.4. Let H be a hypothesis class of functions from X to {0, 1}. Let m be a
training set size. Assume that there exists a set C X of size 2m that is shattered by
H. Then, for any learning algorithm, A, there exist a distribution D over X {0, 1}
and a predictor h H such that L D (h) = 0 but with probability of at least 1/7 over the
choice of S Dm we have that L D (A(S)) 1/8.
Corollary 6.4 tells us that if H shatters some set C of size 2m then we cannot learn
H using m examples. Intuitively, if a set C is shattered by H, and we receive a sample
containing half the instances of C, the labels of these instances give us no information about the labels of the rest of the instances in C every possible labeling of the
rest of the instances can be explained by some hypothesis in H. Philosophically,

45

46

The VC-Dimension

If someone can explain every phenomenon, his explanations are worthless.

This leads us directly to the definition of the VC dimension.
Definition 6.5 (VC-dimension). The VC-dimension of a hypothesis class H,
denoted VCdim(H), is the maximal size of a set C X that can be shattered by H. If
H can shatter sets of arbitrarily large size we say that H has infinite VC-dimension.
A direct consequence of Corollary 6.4 is therefore:
Theorem 6.6. Let H be a class of infinite VC-dimension. Then, H is not PAC
learnable.
Proof. Since H has an infinite VC-dimension, for any training set size m, there exists
a shattered set of size 2m, and the claim follows by Corollary 6.4.
We shall see later in this chapter that the converse is also true: A finite VCdimension guarantees learnability. Hence, the VC-dimension characterizes PAC
learnability. But before delving into more theory, we first show several examples.

6.3 EXAMPLES
In this section we calculate the VC-dimension of several hypothesis classes. To show
that VCdim(H) = d we need to show that
1. There exists a set C of size d that is shattered by H.
2. Every set C of size d + 1 is not shattered by H.

6.3.1 Threshold Functions

Let H be the class of threshold functions over R. Recall Example 6.2, where we have
shown that for an arbitrary set C = {c1 }, H shatters C; therefore VCdim(H) 1. We
have also shown that for an arbitrary set C = {c1 , c2 } where c1 c2 , H does not
shatter C. We therefore conclude that VCdim(H) = 1.

6.3.2 Intervals
Let H be the class of intervals over R, namely, H = {h a,b : a, b R, a < b}, where h a,b :
R {0, 1} is a function such that h a,b (x) = 1[x(a,b)] . Take the set C = {1, 2}. Then, H
shatters C (make sure you understand why) and therefore VCdim(H) 2. Now take
an arbitrary set C = {c1 , c2 , c3 } and assume without loss of generality that c1 c2 c3 .
Then, the labeling (1, 0, 1) cannot be obtained by an interval and therefore H does
not shatter C. We therefore conclude that VCdim(H) = 2.

6.3.3 Axis Aligned Rectangles

Let H be the class of axis aligned rectangles, formally:
H = {h (a1 ,a2 ,b1 ,b2 ) : a1 a2 and b1 b2 }

6.3 Examples
c1
c4

c5

c2

c3

Figure 6.1. Left: 4 points that are shattered by axis aligned rectangles. Right: Any axis
aligned rectangle cannot label c5 by 0 and the rest of the points by 1.

where


h (a1 ,a2 ,b1 ,b2 ) (x 1 , x 2 ) =

1 if a1 x 1 a2
0 otherwise

and b1 x 2 b2

(6.2)

We shall show in the following that VCdim(H) = 4. To prove this we need to find
a set of 4 points that are shattered by H, and show that no set of 5 points can be
shattered by H. Finding a set of 4 points that are shattered is easy (see Figure 6.1).
Now, consider any set C R2 of 5 points. In C, take a leftmost point (whose first
coordinate is the smallest in C), a rightmost point (first coordinate is the largest), a
lowest point (second coordinate is the smallest), and a highest point (second coordinate is the largest). Without loss of generality, denote C = {c1 , . . . , c5 } and let c5
be the point that was not selected. Now, define the labeling (1, 1, 1, 1, 0). It is impossible to obtain this labeling by an axis aligned rectangle. Indeed, such a rectangle
must contain c1 , . . . , c4 ; but in this case the rectangle contains c5 as well, because
its coordinates are within the intervals defined by the selected points. So, C is not
shattered by H , and therefore VCdim( H ) = 4.

6.3.4 Finite Classes

Let H be a finite class. Then, clearly, for any set C we have |HC | |H| and thus
C cannot be shattered if |H| < 2|C| . This implies that VCdim(H) log2 (|H|). This
shows that the PAC learnability of finite classes follows from the more general statement of PAC learnability of classes with finite VC-dimension, which we shall see in
the next section. Note, however, that the VC-dimension of a finite class H can be
significantly smaller than log2 (|H|). For example, let X = {1, . . . , k}, for some integer k, and consider the class of threshold functions (as defined in Example 6.2).
Then, |H| = k but VCdim(H) = 1. Since k can be arbitrarily large, the gap between
log2 (|H|) and VCdim(H) can be arbitrarily large.

6.3.5 VC-Dimension and the Number of Parameters

In the previous examples, the VC-dimension happened to equal the number of
parameters defining the hypothesis class. While this is often the case, it is not
always true. Consider, for example, the domain X = R, and the hypothesis class
H = {h : R} where h : X {0, 1} is defined by h (x) = 0. 5 sin ( x). It is possible to prove that VCdim(H) = , namely, for every d, one can find d points that
are shattered by H (see Exercise 6.8).

47

48

The VC-Dimension

6.4 THE FUNDAMENTAL THEOREM OF PAC LEARNING

We have already shown that a class of infinite VC-dimension is not learnable. The
converse statement is also true, leading to the fundamental theorem of statistical
learning theory:
Theorem 6.7 (The Fundamental Theorem of Statistical Learning). Let H be a
hypothesis class of functions from a domain X to {0, 1} and let the loss function be the
01 loss. Then, the following are equivalent:
1.
2.
3.
4.
5.
6.

H has the uniform convergence property.

Any ERM rule is a successful agnostic PAC learner for H.
H is agnostic PAC learnable.
H is PAC learnable.
Any ERM rule is a successful PAC learner for H.
H has a finite VC-dimension.

The proof of the theorem is given in the next section.

Not only does the VC-dimension characterize PAC learnability; it even determines the sample complexity.
Theorem 6.8 (The Fundamental Theorem of Statistical Learning Quantitative
Version). Let H be a hypothesis class of functions from a domain X to {0, 1} and let
the loss function be the 01 loss. Assume that VCdim(H) = d < . Then, there are
absolute constants C1 , C2 such that
1. H has the uniform convergence property with sample complexity
C1

d + log (1/)
d + log (1/)
m UC
H ( , ) C2
2
2

2. H is agnostic PAC learnable with sample complexity

C1

d + log (1/)
d + log (1/)
m H ( , ) C2
2
2

3. H is PAC learnable with sample complexity

C1

d + log (1/)
d log (1/ ) + log (1/)
m H ( , ) C2

The proof of this theorem is given in Chapter 28.

Remark 6.3. We stated the fundamental theorem for binary classification tasks. A
similar result holds for some other learning problems such as regression with the
absolute loss or the squared loss. However, the theorem does not hold for all learning tasks. In particular, learnability is sometimes possible even though the uniform
convergence property does not hold (we will see an example in Chapter 13, Exercise
6.2). Furthermore, in some situations, the ERM rule fails but learnability is possible
with other learning rules.

6.5 Proof of Theorem 6.7

6.5 PROOF OF THEOREM 6.7

We have already seen that 1 2 in Chapter 4. The implications 2 3 and 3 4
are trivial and so is 2 5. The implications 4 6 and 5 6 follow from the NoFree-Lunch theorem. The difficult part is to show that 6 1. The proof is based on
two main claims:

If VCdim(H) = d, then even though H might be infinite, when restricting it to
a finite set C X , its effective size, |HC |, is only O(|C|d ). That is, the size of
HC grows polynomially rather than exponentially with |C|. This claim is often
referred to as Sauers lemma, but it has also been stated and proved independently by Shelah and by Perles. The formal statement is given in Section 6.5.1
later.

In Section 4 we have shown that finite hypothesis classes enjoy the uniform convergence property. In Section 6.5.2 later we generalize this result and show
that uniform convergence holds whenever the hypothesis class has a small
effective size. By small effective size we mean classes for which |HC | grows
polynomially with |C|.

6.5.1 Sauers Lemma and the Growth Function

We defined the notion of shattering, by considering the restriction of H to a finite
set of instances. The growth function measures the maximal effective size of H on
a set of m examples. Formally:
Definition 6.9 (Growth Function). Let H be a hypothesis class. Then the growth
function of H, denoted H : N N, is defined as
H (m) =

max

CX :|C|=m

|HC |.

In words, H (m) is the number of different functions from a set C of size m to {0, 1}
that can be obtained by restricting H to C.
Obviously, if VCdim(H) = d then for any m d we have H (m) = 2m . In such
cases, H induces all possible functions from C to {0, 1}. The following beautiful
lemma, proposed independently by Sauer, Shelah, and Perles, shows that when m
becomes larger than the VC-dimension, the growth function increases polynomially
rather than exponentially with m.
Lemma 6.10 (Sauer-Shelah-Perles).
 H be a hypothesis class with VCdim(H)

Let
d < . Then, for all m, H (m) di=0 mi . In particular, if m > d + 1 then H (m)
(em/d)d .

Proof of Sauers Lemma*

To prove the lemma it suffices to prove the following stronger claim: For any C =
{c1 , . . . , cm } we have
H, |HC | |{B C : H shatters B}|.

(6.3)

49

50

The VC-Dimension

The reason why Equation (6.3) is sufficient to prove the lemma is that if
VCdim(H) d then no set whose size is larger than d is shattered by H and therefore
d  

m
|{B C : H shatters B}|
.
i
i=0

When m > d + 1 the right-hand side of the preceding is at most (em/d)d (see
Lemma A.5 in Appendix A).
We are left with proving Equation (6.3) and we do it using an inductive argument. For m = 1, no matter what H is, either both sides of Equation (6.3) equal
1 or both sides equal 2 (the empty set is always considered to be shattered by H).
Assume Equation (6.3) holds for sets of size k < m and let us prove it for sets of size
m. Fix H and C = {c1 , . . . , cm }. Denote C  = {c2 , . . . , cm } and in addition, define the
following two sets:
Y0 = {(y2 , . . . , ym ) : (0, y2 , . . . , ym ) HC (1, y2 , . . . , ym ) HC },
and
Y1 = {(y2 , . . . , ym ) : (0, y2 , . . . , ym ) HC (1, y2 , . . . , ym ) HC }.
It is easy to verify that |HC | = |Y0 | + |Y1 |. Additionally, since Y0 = HC  , using the
induction assumption (applied on H and C  ) we have that
|Y0 | = |HC  | |{B C  : H shatters B}| = |{B C : c1  B H shatters B}|.
Next, define H H to be
H = {h H : h  H s.t. (1 h  (c1 ), h  (c2 ), . . . , h  (cm ))
= (h(c1 ), h(c2 ), . . . , h(cm )},
namely, H contains pairs of hypotheses that agree on C  and differ on c1 . Using
this definition, it is clear that if H shatters a set B C  then it also shatters the set
B {c1 } and vice versa. Combining this with the fact that Y1 = HC  and using the
inductive assumption (now applied on H and C  ) we obtain that
|Y1 | = |HC  | |{B C  : H shatters B}| = |{B C  : H shatters B {c1 }}|
= |{B C : c1 B H shatters B}| |{B C : c1 B H shatters B}|.
Overall, we have shown that
|HC | = |Y0 | + |Y1 |
|{B C : c1  B H shatters B}| + |{B C : c1 B H shatters B}|
= |{B C : H shatters B}|,
which concludes our proof.

6.5.2 Uniform Convergence for Classes of Small Effective Size

In this section we prove that if H has small effective size then it enjoys the uniform
convergence property. Formally,

6.5 Proof of Theorem 6.7

Theorem 6.11. Let H be a class and let H be its growth function. Then, for every D
and every (0, 1), with probability of at least 1 over the choice of S Dm we
have

4 + log (H (2m))

.
|L D (h) L S (h)|
2m
Before proving the theorem, let us first conclude the proof of Theorem 6.7.
Proof of Theorem 6.7. It suffices to prove that if the VC-dimension is finite then the
uniform convergence property holds. We will prove that


16d
16d
16 d log (2e/d)
UC
log
.
+
m H ( , ) 4
( )2
( )2
( )2
From Sauers lemma we have that for m > d, H (2m) (2em/d)d . Combining this
with Theorem 6.11 we obtain that with probability of at least 1 ,

4 + d log (2em/d)

|L S (h) L D (h)|
.
2m

For simplicity assume that d log (2em/d) 4; hence,

1 2d log (2em/d)
.
|L S (h) L D (h)|

m
To ensure that the preceding is at most we need that
m

2d log (m) 2 d log (2e/d)

+
.
( )2
( )2

Standard algebraic manipulations (see Lemma A.2 in Appendix A) show that a

sufficient condition for the preceding to hold is that


2d
2d
4 d log (2e/d)
log
.
m4
+
( )2
( )2
( )2

Remark 6.4. The upper bound on m UC

H we derived in the proof Theorem 6.7 is not
the tightest possible. A tighter analysis that yields the bounds given in Theorem 6.8
can be found in Chapter 28.

Proof of Theorem 6.11*

We will start by showing that



4 + log (H (2m))

E m sup |L D (h) L S (h)|

.
SD
2m
hH

(6.4)

Since the random variable suphH |L D (h) L S (h)| is nonnegative, the proof of
the theorem follows directly from the preceding using Markovs inequality (see
Section B.1).

51

52

The VC-Dimension

To bound the left-hand side of Equation (6.4) we first note that for every h H,
 is an additional i.i.d.
we can rewrite L D (h) = E S  Dm [L S  (h)], where S  = z 1 , . . . , z m
sample. Therefore,



E

sup |L D (h) L S (h)| = E m

SD m hH

SD





sup  E m L S  (h) L S (h) .
S D

hH

A generalization of the triangle inequality yields

E [L S  (h) L S (h)] E |L S  (h) L S (h)|,
S  D m
S  D m
and the fact that supermum of expectation is smaller than expectation of supremum
yields
sup

E |L S  (h) L S (h)|


m
hH S D

sup |L S  (h) L S (h)|.

S  D m hH

Formally, the previous two inequalities follow from Jensens inequality. Combining
all we obtain




E m sup |L D (h) L S (h)| E m sup |L S  (h) L S (h)|
SD

S,S D

hH

hH

S,S D m

m


1 

(
(h, z i )
(h, z i )) .
sup


hH m

(6.5)

i=1

The expectation on the right-hand side is over a choice of two i.i.d. samples S =

z 1 , . . . , z m and S  = z 1 , . . . , z m
. Since all of these 2m vectors are chosen i.i.d., nothing
will change if we replace the name of the random vector z i with the name of the
random vector z i . If we do it, instead of the term (
(h, z i )
(h, z i )) in Equation (6.5)
we will have the term (
(h, z i )
(h, z i )). It follows that for every {1}m we
have that Equation (6.5) equals

E

S,S D m



m

1 


i (
(h, z i )
(h, z i ))
sup

hH m
i=1

Since this holds for every {1}m , it also holds if we sample each component of
uniformly at random from the uniform distribution over {1}, denoted U . Hence,
Equation (6.5) also equals
m


1 

i (
(h, z i )
(h, z i )) ,
sup


hH m

m S,S  D m
U

i=1

and by the linearity of expectation it also equals


E

m
S,S D m U

m


1 


i (
(h, z i )
(h, z i )) .
sup


m
hH
i=1

6.7 Bibliographic Remarks

Next, fix S and S  , and let C be the instances appearing in S and S  . Then, we can
take the supremum only over h HC . Therefore,
m



1 


E m sup
i (
(h, z i )
(h, z i ))

U hH m
i=1



m

1 

i (
(h, z i )
(h, z i )) .
= E m max

U hHC m
i=1

m


Fix some h HC and denote h =
i=1 i (
(h, z i )
(h, z i )). Since E [h ] = 0 and
h is an average of independent variables, each of which takes values in [ 1, 1], we
have by Hoeffdings inequality that for every > 0,


P [|h | > ] 2 exp 2 m 2 .
1
m

Applying the union bound over h HC , we obtain that for any > 0,




P max |h | > 2 |HC | exp 2 m 2 .
hHC

Finally, Lemma A.4 in Appendix A tells us that the preceding implies




4 + log (|HC |)

E max |h |
.
hHC
2m
Combining all with the definition of H , we have shown that



4 + log (H (2m))

.
E m sup |L D (h) L S (h)|
SD
2m
hH

6.6 SUMMARY
The fundamental theorem of learning theory characterizes PAC learnability of
classes of binary classifiers using VC-dimension. The VC-dimension of a class is a
combinatorial property that denotes the maximal sample size that can be shattered
by the class. The fundamental theorem states that a class is PAC learnable if and
only if its VC-dimension is finite and specifies the sample complexity required for
PAC learning. The theorem also shows that if a problem is at all learnable, then
uniform convergence holds and therefore the problem is learnable using the ERM
rule.

6.7 BIBLIOGRAPHIC REMARKS

The definition of VC-dimension and its relation to learnability and to uniform convergence is due to the seminal work of Vapnik and Chervonenkis (1971). The
relation to the definition of PAC learnability is due to Blumer, Ehrenfeucht,
Haussler, and Warmuth (1989).
Several generalizations of the VC-dimension have been proposed. For example,
the fat-shattering dimension characterizes learnability of some regression problems (Kearns, Schapire & Sellie 1994; Alon, Ben-David, Cesa-Bianchi & Haussler

53

54

The VC-Dimension

1997; Bartlett, Long & Williamson 1994; Anthony & Bartlet 1999), and the
Natarajan dimension characterizes learnability of some multiclass learning problems (Natarajan 1989). However, in general, there is no equivalence between
learnability and uniform convergence. See (Shalev-Shwartz, Shamir, Srebro &
Sridharan 2010; Daniely, Sabato, Ben-David & Shalev-Shwartz 2011).
Sauers lemma has been proved by Sauer in response to a problem of Erdos
(Sauer 1972). Shelah (with Perles) proved it as a useful lemma for Shelahs theory
of stable models (Shelah 1972). Gil Kalai tells1 us that at some later time, Benjy
Weiss asked Perles about such a result in the context of ergodic theory, and Perles,
who forgot that he had proved it once, proved it again. Vapnik and Chervonenkis
proved the lemma in the context of statistical learning theory.

6.8 EXERCISES
6.1 Show the following monotonicity property of VC-dimension: For every two hypothesis classes if H H then VCdim(H ) VCdim(H).
6.2 Given some finite domain set, X , and a number k |X |, figure out the VC-dimension
of each of the following classes (and prove your claims):
X
1. HX
=k = {h {0, 1} : |{x : h(x) = 1}| = k}: that is, the set of all functions that assign
the value 1 to exactly k elements of X .
2. Hatmostk = {h {0, 1}X : |{x : h(x) = 1}| k or |{x : h(x) = 0}| k}.
6.3 Let X be the Boolean hypercube {0, 1}n . For a set I {1, 2, . . . , n} we define a parity
function h I as follows. On a binary vector x = (x1 , x2 , . . ., xn ) {0, 1}n ,



xi mod 2 .
h I (x) =
iI

(That is, h I computes parity of bits in I .) What is the VC-dimension of the class of
all such parity functions, Hn-parity = {h I : I {1, 2, . . . , n}}?
6.4 We proved Sauers lemma by proving that for every class H of finite VC-dimension
d, and every subset A of the domain,
|H A | |{B A : H shatters B}|


d 

| A|
i=0

Show that there are cases in which the previous two inequalities are strict (namely,
the can be replaced by <) and cases in which they can be replaced by equalities.
Demonstrate all four combinations of = and <.
6.5 VC-dimension of axis aligned rectangles in Rd : Let Hdrec be the class of axis aligned
rectangles in Rd . We have already seen that VCdim(H2rec ) = 4. Prove that in general,
VCdim(Hdrec ) = 2d.
6.6 VC-dimension of Boolean conjunctions: Let Hdcon be the class of Boolean conjunctions over the variables x1 , . . ., xd (d 2). We already know that this class is finite
and thus (agnostic) PAC learnable. In this question we calculate VCdim(Hdcon ).
1. Show that |Hdcon | 3d + 1.
2. Conclude that VCdim(H) d log 3.
3. Show that Hdcon shatters the set of unit vectors {ei : i d}.
1

http://gilkalai.wordpress.com/2008/09/28/extremal-combinatorics-iii-some-basic-theorems

6.8 Exercises

4. (**) Show that VCdim(Hdcon ) d.

Hint: Assume by contradiction that there exists a set C = {c1 , . . ., cd+1 } that is
shattered by Hdcon . Let h 1 , . . ., h d+1 be hypotheses in Hdcon that satisfy

0 i=j
i , j [d + 1], h i (c j ) =
1 otherwise
For each i [d + 1], h i (or more accurately, the conjunction that corresponds to
h i ) contains some literal
i which is false on ci and true on c j for each j = i . Use
the Pigeonhole principle to show that there must be a pair i < j d + 1 such
that
i and
j use the same xk and use that fact to derive a contradiction to the
requirements from the conjunctions h i , h j .
5. Consider the class Hdmcon of monotone Boolean conjunctions over {0, 1}d . Monotonicity here means that the conjunctions do not contain negations. As in Hdcon ,
the empty conjunction is interpreted as the all-positive hypothesis. We augment
Hdmcon with the all-negative hypothesis h . Show that VCdim(Hdmcon ) = d.
6.7 We have shown that for a finite hypothesis class H, VCdim(H) log (|H|). However, this is just an upper bound. The VC-dimension of a class can be much lower
than that:
1. Find an example of a class H of functions over the real interval X = [0, 1] such
that H is infinite while VCdim(H) = 1.
2. Give an example of a finite hypothesis class H over the domain X = [0, 1], where
VCdim(H) = log2 (|H|).
6.8 (*) It is often the case that the VC-dimension of a hypothesis class equals (or can
be bounded above by) the number of parameters one needs to set in order to define
each hypothesis in the class. For instance, if H is the class of axis aligned rectangles in
Rd , then VCdim(H) = 2d, which is equal to the number of parameters used to define
a rectangle in Rd . Here is an example that shows that this is not always the case.
We will see that a hypothesis class might be very complex and even not learnable,
although it has a small number of parameters.
Consider the domain X = R, and the hypothesis class
H = {x  sin ( x) : R}

(here, we take 1 = 0). Prove that VCdim(H) = .

Hint: There is more than one way to prove the required result. One option is by
applying the following lemma: If 0. x1 x2 x3 . . ., is the binary expansion of x (0, 1),
then for any natural number m, sin (2m x) = (1 xm ), provided that k m s.t.
xk = 1.
6.9 Let H be the class of signed intervals, that is,
H = {h a,b,s : a b, s {1, 1}} where

s
if x [a, b]
h a,b,s (x) =
s if x
/ [a, b]
Calculate VCdim(H).
6.10 Let H be a class of functions from X to {0, 1}.
1. Prove that if VCdim(H) d, for any d, then for some probability distribution D
over X {0, 1}, for every sample size, m,
E [L D ( A(S))] min L D (h) +

SD m

hH

d m
2d

55

56

The VC-Dimension

Hint: Use Exercise 6.3 in Chapter 5.

2. Prove that for every H that is PAC learnable, VCdim(H) < . (Note that this is
the implication 3 6 in Theorem 6.7.)
6.11 VC of union: Let H1 , . . . , Hr be hypothesis classes over some fixed domain set X .
Let d = maxi VCdim(Hi ) and assume for simplicity that d 3.
1. Prove that


VCdim ri=1 Hi 4d log (2d) + 2 log (r ).
Hint: Take a set of k examples and assume that they are shattered by the union
class. Therefore, the union class can produce all 2k possible labelings on these
examples. Use Sauers lemma to show that the union class cannot produce more
than r k d labelings. Therefore, 2k r k d . Now use Lemma A.2.
2. (*) Prove that for r = 2 it holds that
VCdim (H1 H2 ) 2d + 1.
6.12 Dudley classes: In this question we discuss an algebraic framework for defining concept classes over Rn and show a connection between the VC dimension of such
classes and their algebraic properties. Given a function f : Rn R we define the
corresponding function, POS( f )(x) = 1[ f (x)>0] . For a class F of real valued functions we define a corresponding class of functions POS(F ) = {POS( f ) : f F }. We
say that a family, F , of real valued functions is linearly closed if for all f , g F and
r R, ( f +r g) F (where addition and scalar multiplication of functions are defined
point wise, namely, for all x Rn , ( f + r g)(x) = f (x) + r g(x)). Note that if a family
of functions is linearly closed then we can view it as a vector space over the reals.
def

For a function g : Rn R and a family of functions F , let F + g = { f + g : f F }.

Hypothesis classes that have a representation as POS(F + g) for some vector space
of functions F and some function g are called Dudley classes.
1. Show that for every g : Rn R and every vector space of functions F as defined
earlier, VCdim(POS(F + g)) = VCdim(POS(F )).
2. (**) For every linearly closed family of real valued functions F , the VCdimension of the corresponding class POS(F ) equals the linear dimension of
F (as a vector space). Hint: Let f1 , . . ., f d be a basis for the vector space F . Consider the mapping x  ( f 1 (x), . . ., f d (x)) (from Rn to Rd ). Note that this mapping
induces a matching between functions over Rn of the form POS( f ) and homogeneous linear halfspaces in Rd (the VC-dimension of the class of homogeneous
linear halfspaces is analyzed in Chapter 9).
3. Show that each of the following classes can be represented as a Dudley class:
1. The class H Sn of halfspaces over Rn (see Chapter 9).
2. The class H H Sn of all homogeneous halfspaces over Rn (see Chapter 9).
3. The class Bd of all functions defined by (open) balls in Rd . Use the Dudley
representation to figure out the VC-dimension of this class.
4. Let Pnd denote the class of functions defined by polynomial inequalities of
degree d, namely,
Pnd = {h p : p is a polynomial of degree d in the variables x1 , . . ., xn },
where for x = (x1 . . . ., xn ), h p (x) = 1[ p(x)0] (the degree of a multivariable polynomial is the maximal sum of variable exponents over all of its terms. For
example, the degree of p(x) = 3x13 x22 + 4x3 x72 is 5).

6.8 Exercises

1. Use the Dudley representation to figure out the VC-dimension of the class
P1d the class of all d-degree polynomials over R.
2. Prove that the class of all polynomial classifiers over R has infinite VCdimension.
3. Use the Dudley representation to figure out the VC-dimension of the class
Pnd (as a function of d and n).

57

7
Nonuniform Learnability

The notions of PAC learnability discussed so far in the book allow the sample
sizes to depend on the accuracy and confidence parameters, but they are uniform
with respect to the labeling rule and the underlying data distribution. Consequently, classes that are learnable in that respect are limited (they must have a
finite VC-dimension, as stated by Theorem 6.7). In this chapter we consider more
relaxed, weaker notions of learnability. We discuss the usefulness of such notions
and provide characterization of the concept classes that are learnable using these
definitions.
We begin this discussion by defining a notion of nonuniform learnability that
allows the sample size to depend on the hypothesis to which the learner is compared. We then provide a characterization of nonuniform learnability and show that
nonuniform learnability is a strict relaxation of agnostic PAC learnability. We also
show that a sufficient condition for nonuniform learnability is that H is a countable union of hypothesis classes, each of which enjoys the uniform convergence
property. These results will be proved in Section 7.2 by introducing a new learning
paradigm, which is called Structural Risk Minimization (SRM). In Section 7.3 we
specify the SRM paradigm for countable hypothesis classes, which yields the Minimum Description Length (MDL) paradigm. The MDL paradigm gives a formal
justification to a philosophical principle of induction called Occams razor. Next,
in Section 7.4 we introduce consistency as an even weaker notion of learnability. Finally, we discuss the significance and usefulness of the different notions of
learnability.

7.1 NONUNIFORM LEARNABILITY

Nonuniform learnability allows the sample size to be nonuniform with respect to
the different hypotheses with which the learner is competing. We say that a hypothesis h is ( , )-competitive with another hypothesis h  if, with probability higher than
(1 ),
L D (h) L D (h  ) + .
58

7.1 Nonuniform Learnability

In PAC learnability, this notion of competitiveness is not very useful, as we

are looking for a hypothesis with an absolute low risk (in the realizable case) or
with a low risk compared to the minimal risk achieved by hypotheses in our class
(in the agnostic case). Therefore, the sample size depends only on the accuracy and
confidence parameters. In nonuniform learnability, however, we allow the sample
size to be of the form m H ( , , h); namely, it depends also on the h with which we
are competing. Formally,
Definition 7.1. A hypothesis class H is nonuniformly learnable if there exist a
: (0, 1)2 H N such that, for every
learning algorithm, A, and a function m NUL
H
NUL
, (0, 1) and for every h H, if m m H ( , , h) then for every distribution D,
with probability of at least 1 over the choice of S Dm , it holds that
L D (A(S)) L D (h) + .
At this point it might be useful to recall the definition of agnostic PAC learnability (Definition 3.3):
A hypothesis class H is agnostically PAC learnable if there exist a learning algorithm,
A, and a function m H : (0, 1)2 N such that, for every , (0, 1) and for every distribution D, if m m H ( , ), then with probability of at least 1 over the choice of
S Dm it holds that
L D (A(S)) min
L D (h  ) + .

h H

Note that this implies that for every h H

L D (A(S)) L D (h) + .
In both types of learnability, we require that the output hypothesis will be ( , )competitive with every other hypothesis in the class. But the difference between
these two notions of learnability is the question of whether the sample size m may
depend on the hypothesis h to which the error of A(S) is compared. Note that that
nonuniform learnability is a relaxation of agnostic PAC learnability. That is, if a
class is agnostic PAC learnable then it is also nonuniformly learnable.

7.1.1 Characterizing Nonuniform Learnability

Our goal now is to characterize nonuniform learnability. In the previous chapter
we have found a crisp characterization of PAC learnable classes, by showing that a
class of binary classifiers is agnostic PAC learnable if and only if its VC-dimension is
finite. In the following theorem we find a different characterization for nonuniform
learnable classes for the task of binary classification.
Theorem 7.2. A hypothesis class H of binary classifiers is nonuniformly learnable if
and only if it is a countable union of agnostic PAC learnable hypothesis classes.
The proof of Theorem 7.2 relies on the following result of independent interest:
Theorem 7.3. Let H be a hypothesis class that can be written as a countable union

of hypothesis classes, H = nN Hn , where each Hn enjoys the uniform convergence
property. Then, H is nonuniformly learnable.

59

60

Nonuniform Learnability

Recall that in Chapter 4 we have shown that uniform convergence is sufficient for
agnostic PAC learnability. Theorem 7.3 generalizes this result to nonuniform learnability. The proof of this theorem will be given in the next section by introducing a
new learning paradigm. We now turn to proving Theorem 7.2.

Proof of Theorem 7.2. First assume that H = nN Hn where each Hn is agnostic
PAC learnable. Using the fundamental theorem of statistical learning, it follows
that each Hn has the uniform convergence property. Therefore, using Theorem 7.3
we obtain that H is nonuniform learnable.
For the other direction, assume that H is nonuniform learnable using some
algorithm A. For every n N, let Hn = {h H : m NUL
H (1/8, 1/7, h) n}. Clearly,
H = nN Hn . In addition, using the definition of m NUL
H we know that for any distribution D that satisfies the realizability assumption with respect to Hn , with probability
of at least 6/7 over S Dn we have that L D (A(S)) 1/8. Using the fundamental
theorem of statistical learning, this implies that the VC-dimension of Hn must be
finite, and therefore Hn is agnostic PAC learnable.
The following example shows that nonuniform learnability is a strict relaxation of agnostic PAC learnability; namely, there are hypothesis classes that are
nonuniform learnable but are not agnostic PAC learnable.
Example 7.1. Consider a binary classification problem with the instance domain
being X = R. For every n N let Hn be the class of polynomial classifiers of degree
n; namely, Hn is the set of all classifiers of theform h(x) = sign( p(x)) where p :
R R is a polynomial of degree n. Let H = nN Hn . Therefore, H is the class
of all polynomial classifiers over R. It is easy to verify that VCdim(H) = while
VCdim(Hn ) = n + 1 (see Exercise 7.12). Hence, H is not PAC learnable, while on
the basis of Theorem 7.3, H is nonuniformly learnable.

7.2 STRUCTURAL RISK MINIMIZATION

So far, we have encoded our prior knowledge by specifying a hypothesis class H,
which we believe includes a good predictor for the learning task at hand. Yet
another way to express our prior knowledge is by specifying preferences over
hypotheses within H. In the Structural Risk Minimization
 (SRM) paradigm, we
do so by first assuming that H can be written as H = nN Hn and then specifying a weight function, w : N [0, 1], which assigns a weight to each hypothesis
class, Hn , such that a higher weight reflects a stronger preference for the hypothesis
class. In this section we discuss how to learn with such prior knowledge. In the next
section we describe a couple of important weighting schemes, including Minimum
Description Length.

Concretely, let H be a hypothesis class that can be written as H = nN Hn . For
example, H may be the class of all polynomial classifiers where each Hn is the class
of polynomial classifiers of degree n (see Example 7.1). Assume that for each n, the
class Hn enjoys the uniform convergence property (see Definition 4.3 in Chapter 4)
with a sample complexity function m UC
Hn ( , ). Let us also define the function n :
N (0, 1) (0, 1) by
n (m, ) = min{ (0, 1) : m UC
Hn ( , ) m}.

(7.1)

7.2 Structural Risk Minimization

In words, we have a fixed sample size m, and we are interested in the lowest possible
upper bound on the gap between empirical and true risks achievable by using a
sample of m examples.
From the definitions of uniform convergence and n , it follows that for every m
and , with probability of at least 1 over the choice of S Dm we have that
(7.2)
h Hn , |L D (h) L S (h)| n (m, ).


Let w : N [0, 1] be a function such that n=1 w(n) 1. We refer to w as a
weight function over the hypothesis classes H1 , H2 , . . .. Such a weight function can
reflect the importance that the learner attributes to each hypothesis class, or some
measure of the complexity of different hypothesis classes. If H is a finite union of N
hypothesis classes, one can simply assign the same weight of 1/N to all hypothesis
classes. This equal weighting corresponds to no a priori preference to any hypothesis
class. Of course, if one believes (as prior knowledge) that a certain hypothesis class is
more likely to contain the correct target function, then it should be assigned a larger
weight, reflecting this prior knowledge. When H is a (countable) infinite union of
hypothesis classes, a uniform weighting is not possible but many other weighting
schemes may work. For example, one can choose w(n) = 26n2 or w(n) = 2n . Later
in this chapter we will provide another convenient way to define weighting functions
using description languages.
The SRM rule follows a bound minimization approach. This means that the
goal of the paradigm is to find a hypothesis that minimizes a certain upper bound
on the true risk. The bound that the SRM rule wishes to minimize is given in the
following theorem.


Theorem 7.4. Let w : N [0, 1] be a function
 such that n=1 w(n) 1. Let H be a
hypothesis class that can be written as H = nN Hn , where for each n, Hn satisfies
the uniform convergence property with a sample complexity function m UC
Hn . Let n
be as defined in Equation (7.1). Then, for every (0, 1) and distribution D, with
probability of at least 1 over the choice of S Dm , the following bound holds
(simultaneously) for every n N and h Hn .
|L D (h) L S (h)| n (m, w(n) ).
Therefore, for every (0, 1) and distribution D, with probability of at least 1 it
holds that
(7.3)
h H, L D (h) L S (h) + min n (m, w(n) ).
n:hHn

Proof. For each n define n = w(n). Applying the assumption that uniform convergence holds for all n with the rate given in Equation (7.2), we obtain that if we fix n
in advance, then with probability of at least 1 n over the choice of S Dm,
h Hn , |L D (h) L S (h)| n (m, n ).
Applying
the union bound
over n = 1, 2, . . ., we obtain that with probability of at


least 1 n n = 1 n w(n) 1 , the preceding holds for all n, which concludes
our proof.
Denote
n(h) = min{n : h Hn },

(7.4)

61

62

Nonuniform Learnability

and then Equation (7.3) implies that

L D (h) L S (h) + n(h) (m, w(n(h)) ).
The SRM paradigm searches for h that minimizes this bound, as formalized in
the following pseudocode:
Structural Risk Minimization (SRM)
prior knowledge:

H = n Hn where Hn has uniform convergence with m UC
Hn


w : N [0, 1] where n w(n) 1
define: n as in Equation (7.1); n(h) as in Equation (7.4)
m
input: training set S D
 , confidence

output: h argminhH L S (h) + n(h) (m, w(n(h)) )
Unlike the ERM paradigm discussed in previous chapters, we no longer just care
about the empirical risk, L S (h), but we are willing to trade some of our bias toward
low empirical risk with a bias toward classes for which n(h) (m, w(n(h)) ) is smaller,
for the sake of a smaller estimation error.
Next we show that the SRM paradigm can be used for nonuniform learning of
every class, which is a countable union of uniformly converging hypothesis classes.

Theorem 7.5. Let H be a hypothesis class such that H = nN Hn , where each Hn has
the uniform convergence property with sample complexity m UC
Hn . Let w : N [0, 1] be
6
such that w(n) = n2 2 . Then, H is nonuniformly learnable using the SRM rule with
rate


UC
6
m NUL
/2
,
.
( ,
,
h)

m
H
Hn(h)
(n(h))2
Proof. Let A be the SRM algorithm with respect to the weighting function w. For


every h H, , and , let m m UC
Hn(h) ( , w(n(h))). Using the fact that
n w(n) = 1,
we can apply Theorem 7.4 to get that, with probability of at least 1 over the
choice of S Dm , we have that for every h  H,
L D (h  ) L S (h  ) + n(h  ) (m, w(n(h  ))).
The preceding holds in particular for the hypothesis A(S) returned by the SRM rule.
By the definition of SRM we obtain that


L S (h  ) + n(h  ) (m, w(n(h  ))) L S (h) + n(h) (m, w(n(h))).
L D (A(S)) min

h

Finally, if m m Hn(h) ( /2, w(n(h))) then clearly n(h) (m, w(n(h))) /2. In addition, from the uniform convergence property of each Hn we have that with
probability of more than 1 ,
UC

L S (h) L D (h) + /2.

Combining all the preceding we obtain that L D (A(S)) L D (h) + , which concludes
our proof.
Note that the previous theorem also proves Theorem 7.3.

7.3 Minimum Description Length and Occams Razor

Remark 7.2 (No-Free-Lunch for Nonuniform Learnability). We have shown that

any countable union of classes of finite VC-dimension is nonuniformly learnable. It
turns out that, for any infinite domain set, X , the class of all binary valued functions
over X is not a countable union of classes of finite VC-dimension. We leave the
proof of this claim as a (nontrivial) exercise (see Exercise 7.5). It follows that, in
some sense, the No-Free-Lunch theorem holds for nonuniform learning as well:
namely, whenever the domain is not finite, there exists no nonuniform learner with
respect to the class of all deterministic binary classifiers (although for each such
classifier there exists a trivial algorithm that learns it ERM with respect to the
hypothesis class that contains only this classifier).
It is interesting to compare the nonuniform learnability result given in Theorem 7.5 to the task of agnostic PAC learning any specific Hn separately. The prior
knowledge, or bias, of a nonuniform learner for H is weaker it is searching for a
model throughout the entire class H, rather than being focused on one specific Hn .
The cost of this weakening of prior knowledge is the increase in sample complexity needed to compete with any specific h Hn . For a concrete evaluation of this
gap, consider the task of binary classification with the zero-one loss. Assume that
n+log (1/)
(where C is the contant
for all n, VCdim(Hn ) = n. Since m UC
Hn ( , ) = C
2
appearing in Theorem 6.8), a straightforward calculation shows that
UC
m NUL
H ( , , h) m Hn ( /2, ) 4C

2 log (2n)
.
2

That is, the cost of relaxing the learners prior knowledge from a specific Hn that
contains the target h to a countable union of classes depends on the log of the index
of the first class in which h resides. That cost increases with the index of the class,
which can be interpreted as reflecting the value of knowing a good priority order on
the hypotheses in H.

7.3 MINIMUM DESCRIPTION LENGTH AND OCCAMS RAZOR

Let H be a countable hypothesis
class. Then, we can write H as a countable union
of singleton classes, namely, H = nN {h n }. By Hoeffdings inequality (Lemma 4.5),
each singleton class has the uniform convergence property with rate m UC ( , ) =
log (2/)
. Therefore, the function n given in Equation (7.1) becomes n (m, ) =
2 2
log (2/)
and the SRM rule becomes
2m



log (w(n)) + log (2/)
argmin L S (h) +
.
2m
h n H
Equivalently, we can think of w as a function from H to [0, 1], and then the SRM
rule becomes



log (w(h)) + log (2/)
.
argmin L S (h) +
2m
hH
It follows that in this case, the prior knowledge is solely determined by the weight we
assign to each hypothesis. We assign higher weights to hypotheses that we believe

63

64

Nonuniform Learnability

are more likely to be the correct one, and in the learning algorithm we prefer
hypotheses that have higher weights.
In this section we discuss a particular convenient way to define a weight function
over H, which is derived from the length of descriptions given to hypotheses. Having a hypothesis class, one can wonder about how we describe, or represent, each
hypothesis in the class. We naturally fix some description language. This can be
English, or a programming language, or some set of mathematical formulas. In any
of these languages, a description consists of finite strings of symbols (or characters)
drawn from some fixed alphabet. We shall now formalize these notions.
Let H be the hypothesis class we wish to describe. Fix some finite set  of symbols (or characters), which we call the alphabet. For concreteness, we let  =
{0, 1}. A string is a finite sequence of symbols from ; for example, = (0, 1, 1, 1, 0)
is a string of length 5. We denote by | | the length of a string. The set of all finite
length strings is denoted  . A description language for H is a function d : H  ,
mapping each member h of H to a string d(h). d(h) is called the description of h,
and its length is denoted by |h|.
We shall require that description languages be prefix-free; namely, for every distinct h, h  , d(h) is not a prefix of d(h  ). That is, we do not allow that any string d(h)
is exactly the first |h| symbols of any longer string d(h  ). Prefix-free collections of
strings enjoy the following combinatorial property:
Lemma 7.6 (Kraft Inequality). If S {0, 1} is a prefix-free set of strings, then
 1
1.
2| |
S
Proof. Define a probability distribution over the members of S as follows: Repeatedly toss an unbiased coin, with faces labeled 0 and 1, until the sequence of outcomes
is a member of S; at that point, stop. For each S, let P( ) be the probability that
this process generates the string . Note that since S is prefix-free, for every S, if
the coin toss outcomes follow the bits of then we will stop only once the sequence
of outcomes equals . We therefore get that, for every S, P( ) = 2|1 | . Since
probabilities add up to at most 1, our proof is concluded.
In light of Krafts inequality, any prefix-free description language of a hypothesis
class, H, gives rise to a weighting function w over that hypothesis class we will
simply set w(h) = 21|h| . This observation immediately yields the following:
Theorem 7.7. Let H be a hypothesis class and let d : H {0, 1} be a prefix-free
description language for H. Then, for every sample size, m, every confidence parameter, > 0, and every probability distribution, D, with probability greater than 1
over the choice of S Dm we have that,

|h| + ln (2/)
,
h H, L D (h) L S (h) +
2m
where |h| is the length of d(h).
Proof. Choose w(h) = 1/2|h| , apply Theorem 7.4 with n (m, ) =
that ln (2|h| ) = |h| ln (2) < |h|.

ln (2/)
2m ,

and note

7.3 Minimum Description Length and Occams Razor

As was the case with Theorem 7.4, this result suggests a learning paradigm for
H given
a training set, S, search for a hypothesis h H that minimizes the bound,

L S (h) + |h|+ln(2/)
. In particular, it suggests trading off empirical risk for saving
2m
description length. This yields the Minimum Description Length learning paradigm.
Minimum Description Length (MDL)
prior knowledge:
H is a countable hypothesis class
H is described by a prefix-free language over {0, 1}
For every h H, |h| is the length of the representation of h
input: A training set S 
Dm , confidence 

(2/)
output: h argminhH L S (h) + |h|+ln
2m

Example 7.3. Let H be the class of all predictors that can be implemented using
some programming language, say, C++. Let us represent each program using the
binary string obtained by running the gzip command on the program (this yields
a prefix-free description language over the alphabet {0, 1}). Then, |h| is simply
the length (in bits) of the output of gzip when running on the C++ program
corresponding to h.

7.3.1 Occams Razor

Theorem 7.7 suggests that, having two hypotheses sharing the same empirical risk,
the true risk of the one that has shorter description can be bounded by a lower value.
Thus, this result can be viewed as conveying a philosophical message:
A short explanation (that is, a hypothesis that has a short length) tends to be more
valid than a long explanation.
This is a well known principle, called Occams razor, after William of Ockham, a
14th-century English logician, who is believed to have been the first to phrase it
explicitly. Here, we provide one possible justification to this principle. The inequality of Theorem 7.7 shows that the more complex a hypothesis h is (in the sense of
having a longer description), the larger the sample size it has to fit to guarantee that
it has a small true risk, L D (h).
At a second glance, our Occam razor claim might seem somewhat problematic.
In the context in which the Occam razor principle is usually invoked in science, the
language according to which complexity is measured is a natural language, whereas
here we may consider any arbitrary abstract description language. Assume that we
have two hypotheses such that |h  | is much smaller than |h|. By the preceding result,
if both have the same error on a given training set, S, then the true error of h may
be much higher than the true error of h  , so one should prefer h  over h. However,
we could have chosen a different description language, say, one that assigns a string
of length 3 to h and a string of length 100000 to h  . Suddenly it looks as if one should

65

66

Nonuniform Learnability

prefer h over h  . But these are the same h and h  for which we argued two sentences
ago that h  should be preferable. Where is the catch here?
Indeed, there is no inherent generalizability difference between hypotheses.
The crucial aspect here is the dependency order between the initial choice of language (or, preference over hypotheses) and the training set. As we know from
the basic Hoeffdings bound (Equation (4.2)), if we commit to any hypothesis
before seeing the 
data, then we are guaranteed a rather small estimation error term
(2/)
. Choosing a description language (or, equivalently, some
L D (h) L S (h) + ln 2m
weighting of hypotheses) is a weak form of committing to a hypothesis. Rather than
committing to a single hypothesis, we spread out our commitment among many. As
long as it is done independently of the training sample, our generalization bound
holds. Just as the choice of a single hypothesis to be evaluated by a sample can be
arbitrary, so is the choice of description language.

7.4 OTHER NOTIONS OF LEARNABILITY CONSISTENCY

The notion of learnability can be further relaxed by allowing the needed sample
sizes to depend not only on , , and h but also on the underlying data-generating
probability distribution D (that is used to generate the training sample and to determine the risk). This type of performance guarantee is captured by the notion of
consistency1 of a learning rule.
Definition 7.8 (Consistency). Let Z be a domain set, let P be a set of probability
distributions over Z , and let H be a hypothesis class. A learning rule A is consistent
: (0, 1)2 H P N such
with respect to H and P if there exists a function m CON
H
that, for every , (0, 1), every h H, and every D P, if m m NUL
H ( , , h, D) then
with probability of at least 1 over the choice of S Dm it holds that
L D (A(S)) L D (h) + .
If P is the set of all distributions,2 we say that A is universally consistent with respect
to H.
The notion of consistency is, of course, a relaxation of our previous notion of
nonuniform learnability. Clearly if an algorithm nonuniformly learns a class H it is
also universally consistent for that class. The relaxation is strict in the sense that
there are consistent learning rules that are not successful nonuniform learners. For
example, the algorithm Memorize defined in Example 7.4 later is universally consistent for the class of all binary classifiers over N. However, as we have argued before,
this class is not nonuniformly learnable.
Example 7.4. Consider the classification prediction algorithm Memorize defined as
follows. The algorithm memorizes the training examples, and, given a test point x, it
1

In the literature, consistency is often defined using the notion of either convergence in probability (corresponding to weak consistency) or almost sure convergence (corresponding to strong
consistency).
Formally, we assume that Z is endowed with some sigma algebra of subsets , and by all distributions
we mean all probability distributions that have  contained in their associated family of measurable
subsets.

7.5 Discussing the Different Notions of Learnability

predicts the majority label among all labeled instances of x that exist in the training
sample (and some fixed default label if no instance of x appears in the training set).
It is possible to show (see Exercise 7.6) that the Memorize algorithm is universally
consistent for every countable domain X and a finite label set Y (w.r.t. the zero-one
loss).
Intuitively, it is not obvious that the Memorize algorithm should be viewed as a
learner, since it lacks the aspect of generalization, namely, of using observed data to
predict the labels of unseen examples. The fact that Memorize is a consistent algorithm for the class of all functions over any countable domain set therefore raises
doubt about the usefulness of consistency guarantees. Furthermore, the sharp-eyed
reader may notice that the bad learner we introduced in Chapter 2, which led
to overfitting, is in fact the Memorize algorithm. In the next section we discuss the
significance of the different notions of learnability and revisit the No-Free-Lunch
theorem in light of the different definitions of learnability.

7.5 DISCUSSING THE DIFFERENT NOTIONS OF LEARNABILITY

We have given three definitions of learnability and we now discuss their usefulness.
As is usually the case, the usefulness of a mathematical definition depends on what
we need it for. We therefore list several possible goals that we aim to achieve by
defining learnability and discuss the usefulness of the different definitions in light of
these goals.
What Is the Risk of the Learned Hypothesis?
The first possible goal of deriving performance guarantees on a learning algorithm is
bounding the risk of the output predictor. Here, both PAC learning and nonuniform
learning give us an upper bound on the true risk of the learned hypothesis based on
its empirical risk. Consistency guarantees do not provide such a bound. However, it
is always possible to estimate the risk of the output predictor using a validation set
(as will be described in Chapter 11).
How Many Examples Are Required to Be as Good as the Best Hypothesis in H?
When approaching a learning problem, a natural question is how many examples we
need to collect in order to learn it. Here, PAC learning gives a crisp answer. However, for both nonuniform learning and consistency, we do not know in advance
how many examples are required to learn H. In nonuniform learning this number depends on the best hypothesis in H, and in consistency it also depends on the
underlying distribution. In this sense, PAC learning is the only useful definition of
learnability. On the flip side, one should keep in mind that even if the estimation
error of the predictor we learn is small, its risk may still be large if H has a large
approximation error. So, for the question How many examples are required to be
as good as the Bayes optimal predictor? even PAC guarantees do not provide us
with a crisp answer. This reflects the fact that the usefulness of PAC learning relies
on the quality of our prior knowledge.
PAC guarantees also help us to understand what we should do next if our learning algorithm returns a hypothesis with a large risk, since we can bound the part

67

68

Nonuniform Learnability

of the error that stems from estimation error and therefore know how much of the
error is attributed to approximation error. If the approximation error is large, we
know that we should use a different hypothesis class. Similarly, if a nonuniform
algorithm fails, we can consider a different weighting function over (subsets of)
hypotheses. However, when a consistent algorithm fails, we have no idea whether
this is because of the estimation error or the approximation error. Furthermore,
even if we are sure we have a problem with the estimation error term, we do not
know how many more examples are needed to make the estimation error small.

How to Learn? How to Express Prior Knowledge?

Maybe the most useful aspect of the theory of learning is in providing an answer to
the question of how to learn. The definition of PAC learning yields the limitation
of learning (via the No-Free-Lunch theorem) and the necessity of prior knowledge.
It gives us a crisp way to encode prior knowledge by choosing a hypothesis class,
and once this choice is made, we have a generic learning rule ERM. The definition
of nonuniform learnability also yields a crisp way to encode prior knowledge by
specifying weights over (subsets of) hypotheses of H. Once this choice is made, we
again have a generic learning rule SRM. The SRM rule is also advantageous in
model selection tasks, where prior knowledge is partial. We elaborate on model
selection in Chapter 11 and here we give a brief example.
Consider the problem of fitting a one dimensional polynomial to data; namely,
our goal is to learn a function, h : R R, and as prior knowledge we consider the
hypothesis class of polynomials. However, we might be uncertain regarding which
degree d would give the best results for our data set: A small degree might not fit
the data well (i.e., it will have a large approximation error), whereas a high degree
might lead to overfitting (i.e., it will have a large estimation error). In the following we depict the result of fitting a polynomial of degrees 2, 3, and 10 to the same
training set.
Degree 2

Degree 3

Degree 10

It is easy to see that the empirical risk decreases as we enlarge the degree. Therefore, if we choose H to be the class of all polynomials up to degree 10 then the
ERM rule with respect to this class would output a 10 degree polynomial and would
overfit. On the other hand, if we choose too small a hypothesis class, say, polynomials up to degree 2, then the ERM would suffer from underfitting (i.e., a large
approximation error). In contrast, we can use the SRM rule on the set of all polynomials, while ordering subsets of H according to their degree, and this will yield a 3rd
degree polynomial since the combination of its empirical risk and the bound on its
estimation error is the smallest. In other words, the SRM rule enables us to select
the right model on the basis of the data itself. The price we pay for this flexibility
(besides a slight increase of the estimation error relative to PAC learning w.r.t. the

7.5 Discussing the Different Notions of Learnability

optimal degree) is that we do not know in advance how many examples are needed
to compete with the best hypothesis in H.
Unlike the notions of PAC learnability and nonuniform learnability, the definition of consistency does not yield a natural learning paradigm or a way to encode
prior knowledge. In fact, in many cases there is no need for prior knowledge at all.
For example, we saw that even the Memorize algorithm, which intuitively should not
be called a learning algorithm, is a consistent algorithm for any class defined over
a countable domain and a finite label set. This hints that consistency is a very weak
requirement.
Which Learning Algorithm Should We Prefer?
One may argue that even though consistency is a weak requirement, it is desirable
that a learning algorithm will be consistent with respect to the set of all functions
from X to Y, which gives us a guarantee that for enough training examples, we will
always be as good as the Bayes optimal predictor. Therefore, if we have two algorithms, where one is consistent and the other one is not consistent, we should prefer
the consistent algorithm. However, this argument is problematic for two reasons.
First, maybe it is the case that for most natural distributions we will observe in
practice that the sample complexity of the consistent algorithm will be so large so
that in every practical situation we will not obtain enough examples to enjoy this
guarantee. Second, it is not very hard to make any PAC or nonuniform learner consistent with respect to the class of all functions from X to Y. Concretely, consider
a countable domain, X , a finite label set Y, and a hypothesis class, H, of functions
from X to Y. We can make any nonuniform learner for H be consistent with respect
to the class of all classifiers from X to Y using the following simple trick: Upon
receiving a training set, we will first run the nonuniform learner over the training
set, and then we will obtain a bound on the true risk of the learned predictor. If this
bound is small enough we are done. Otherwise, we revert to the Memorize algorithm.
This simple modification makes the algorithm consistent with respect to all functions
from X to Y. Since it is easy to make any algorithm consistent, it may not be wise to
prefer one algorithm over the other just because of consistency considerations.

7.5.1 The No-Free-Lunch Theorem Revisited

Recall that the No-Free-Lunch theorem (Theorem 5.1 from Chapter 5) implies that
no algorithm can learn the class of all classifiers over an infinite domain. In contrast,
in this chapter we saw that the Memorize algorithm is consistent with respect to the
class of all classifiers over a countable infinite domain. To understand why these two
statements do not contradict each other, let us first recall the formal statement of
the No-Free-Lunch theorem.
Let X be a countable infinite domain and let Y = {1}. The No-Free-Lunch
theorem implies the following: For any algorithm, A, and a training set size, m,
there exist a distribution over X and a function h  : X Y, such that if A will get
a sample of m i.i.d. training examples, labeled by h  , then A is likely to return a
classifier with a larger error.
The consistency of Memorize implies the following: For every distribution over
X and a labeling function h  : X Y, there exists a training set size m (that depends

69

70

Nonuniform Learnability

on the distribution and on h  ) such that if Memorize receives at least m examples it

is likely to return a classifier with a small error.
We see that in the No-Free-Lunch theorem, we first fix the training set size, and
then find a distribution and a labeling function that are bad for this training set size.
In contrast, in consistency guarantees, we first fix the distribution and the labeling
function, and only then do we find a training set size that suffices for learning this
particular distribution and labeling function.

7.6 SUMMARY
We introduced nonuniform learnability as a relaxation of PAC learnability and consistency as a relaxation of nonuniform learnability. This means that even classes of
infinite VC-dimension can be learnable, in some weaker sense of learnability. We
discussed the usefulness of the different definitions of learnability.
For hypothesis classes that are countable, we can apply the Minimum Description Length scheme, where hypotheses with shorter descriptions are preferred,
following the principle of Occams razor. An interesting example is the hypothesis class of all predictors we can implement in C++ (or any other programming
language), which we can learn (nonuniformly) using the MDL scheme.
Arguably, the class of all predictors we can implement in C++ is a powerful class
of functions and probably contains all that we can hope to learn in practice. The ability to learn this class is impressive, and, seemingly, this chapter should have been the
last chapter of this book. This is not the case, because of the computational aspect
of learning: that is, the runtime needed to apply the learning rule. For example, to
implement the MDL paradigm with respect to all C++ programs, we need to perform an exhaustive search over all C++ programs, which will take forever. Even the
implementation of the ERM paradigm with respect to all C++ programs of description length at most 1000 bits requires an exhaustive search over 21000 hypotheses.
, the runtime is
While the sample complexity of learning this class is just 1000+log(2/)
2
1000
2
. This is a huge number much larger than the number of atoms in the visible
universe. In the next chapter we formally define the computational complexity of
learning. In the second part of this book we will study hypothesis classes for which
the ERM or SRM schemes can be implemented efficiently.

7.7 BIBLIOGRAPHIC REMARKS

Our definition of nonuniform learnability is related to the definition of an Occamalgorithm in Blumer, Ehrenfeucht, Haussler and Warmuth (1987). The concept of
SRM is due to (Vapnik & Chervonenkis 1974, Vapnik 1995). The concept of MDL
is due to (Rissanen 1978, Rissanen 1983). The relation between SRM and MDL
is discussed in Vapnik (1995). These notions are also closely related to the notion
of regularization (e.g., Tikhonov 1943). We will elaborate on regularization in the
second part of this book.
The notion of consistency of estimators dates back to Fisher (1922). Our presentation of consistency follows Steinwart and Christmann (2008), who also derived
several no-free-lunch theorems.

7.8 Exercises

7.8 EXERCISES
7.1 Prove that for any finite class H, and any description language d : H {0, 1} , the
VC-dimension of H is at most 2 sup{|d(h)| : h H} the maximum description length
of a predictor in H. Furthermore, if d is a prefix-free description then VCdim(H )
sup{|d(h)| : h H}.
7.2 Let H = {h n : n N} be an infinite countable hypothesis class for binary classification.
Show that it is impossible to assign weights to the hypotheses in H such that

H could be learnted nonuniformly using these weights.
That is, the weighting


function w : H [0, 1] should satisfy the condition hH w(h) 1.

The weights would be monotonically nondecreasing. That is, if i < j , then
w(h i ) w(h j ).

7.3
Consider a hypothesis class H =
for every n N, Hn is finite.
n=1 Hn , where

Find a weighting function w : H [0, 1] such that hH w(h) 1 and so that for
all h H, w(h) is determined by n(h) = min{n : h Hn } and by |Hn(h)|.

(*) Define such a function w when for all n Hn is countable (possibly infinite).
7.4 Let H be some hypothesis class. For any h H, let |h| denote the description length
of h, according to some fixed description language. Consider the MDL learning
paradigm in which the algorithm returns:



|h| + ln(2/)
h S arg min L S (h) +
,
hH
2m
where S is a sample of size m. For any B > 0, let H B = {h H : |h| B}, and define
h B = arg min L D (h).
hH B

Prove a bound on L D (h S ) L D (h B ) in terms of B, the confidence parameter , and

the size of the training set m.

Note: Such bounds are known as oracle inequalities in the literature: We wish to
estimate how good we are compared to a reference classifier (or oracle) h B .
7.5 In this question we wish to show a No-Free-Lunch result for nonuniform learnability: namely, that, over any infinite domain, the class of all functions is not learnable
even under the relaxed nonuniform variation of learning.
Recall that an algorithm, A, nonuniformly learns a hypothesis class H if there
2
exists a function m NUL
H : (0, 1) H N such that, for every , (0, 1) and for every
NUL
h H, if m m H ( , , h) then for every distribution D, with probability of at least
1 over the choice of S D m , it holds that
L D ( A(S)) L D (h) + .
If such an algorithm exists then we say that H is nonuniformly learnable.
1. Let A be a nonuniform learner for a class H. For each n N define HnA = {h H :
m NUL (0. 1, 0. 1, h) n}. Prove that each such class Hn has a finite VC-dimension.
2. Prove
that if a class H is nonuniformly learnable then there are classes Hn so that
H = nN Hn and, for every n N, VCdim(Hn ) is finite.
3. Let H be a class that shatters an infinite
set. Then, for every sequence of

classes (Hn : n N) such that H = nN Hn , there exists some n for which
VCdim(Hn ) = .
Hint: Given a class H that shatters some infinite set K , and a sequence of classes
(Hn : n N), each having a finite VC-dimension, start by defining subsets K n K
such that, for all n, |K n | > VCdim(Hn ) and for any n = m, K n K m = . Now, pick

71

72

Nonuniform Learnability

for each such K n a function fn : K n {0, 1} so that no h Hn agrees with f n on

the domain
n . Finally,
 K
 define f : X {0, 1} by combining these f n s and prove
that f H \ nN Hn .
5. Construct a class H1 of functions from the unit interval [0, 1] to {0, 1} that is
nonuniformly learnable but not PAC learnable.
6. Construct a class H2 of functions from the unit interval [0, 1] to {0, 1} that is not
nonuniformly learnable.
7.6 In this question we wish to show that the algorithm Memorize is a consistent learner
for every class of (binary-valued) functions over any countable domain. Let X be a
countable domain and let D be a probability distribution over X .
1. Let {xi : i N} be an enumeration of the elements of X so that for all i j ,
D({xi }) D({x j }). Prove that

D({xi }) = 0.
lim
n

in

2. Given any > 0 prove that there exists D > 0 such that
D({x X : D({x}) < D }) < .

3. Prove that for every > 0, if n is such that D({xi }) < for all i > n, then for every
m N,
/ S)] nem .
P m [xi : (D({xi }) > and xi
SD

4. Conclude that if X is countable then for every probability distribution D over

X there exists a function m D : (0, 1) (0, 1) N such that for every , > 0 if
m > m D ( , ) then
/ S}) > ] < .
P m [D({x : x
SD

5. Prove that Memorize is a consistent learner for every class of (binary-valued)

functions over any countable domain.

8
The Runtime of Learning

So far in the book we have studied the statistical perspective of learning, namely,
how many samples are needed for learning. In other words, we focused on the
amount of information learning requires. However, when considering automated
learning, computational resources also play a major role in determining the complexity of a task: that is, how much computation is involved in carrying out a learning
task. Once a sufficient training sample is available to the learner, there is some computation to be done to extract a hypothesis or figure out the label of a given test
instance. These computational resources are crucial in any practical application of
machine learning. We refer to these two types of resources as the sample complexity and the computational complexity. In this chapter, we turn our attention to the
computational complexity of learning.
The computational complexity of learning should be viewed in the wider context
of the computational complexity of general algorithmic tasks. This area has been
extensively investigated; see, for example, (Sipser 2006). The introductory comments that follow summarize the basic ideas of that general theory that are most
relevant to our discussion.
The actual runtime (in seconds) of an algorithm depends on the specific machine
the algorithm is being implemented on (e.g., what the clock rate of the machines
CPU is). To avoid dependence on the specific machine, it is common to analyze
the runtime of algorithms in an asymptotic sense. For example, we say that the
computational complexity of the merge-sort algorithm, which sorts a list of n items,
is O(n log (n)). This implies that we can implement the algorithm on any machine
that satisfies the requirements of some accepted abstract model of computation,
and the actual runtime in seconds will satisfy the following: there exist constants c
and n 0 , which can depend on the actual machine, such that, for any value of n > n 0 ,
the runtime in seconds of sorting any n items will be at most c n log (n). It is common
to use the term feasible or efficiently computable for tasks that can be performed
by an algorithm whose running time is O( p(n)) for some polynomial function p.
One should note that this type of analysis depends on defining what is the input
size n of any instance to which the algorithm is expected to be applied. For purely
algorithmic tasks, as discussed in the common computational complexity literature,
73

74

The Runtime of Learning

this input size is clearly defined; the algorithm gets an input instance, say, a list to
be sorted, or an arithmetic operation to be calculated, which has a well defined
size (say, the number of bits in its representation). For machine learning tasks, the
notion of an input size is not so clear. An algorithm aims to detect some pattern in
a data set and can only access random samples of that data.
We start the chapter by discussing this issue and define the computational
complexity of learning. For advanced students, we also provide a detailed formal
definition. We then move on to consider the computational complexity of implementing the ERM rule. We first give several examples of hypothesis classes where
the ERM rule can be efficiently implemented, and then consider some cases where,
although the class is indeed efficiently learnable, ERM implementation is computationally hard. It follows that hardness of implementing ERM does not imply
hardness of learning. Finally, we briefly discuss how one can show hardness of a
given learning task, namely, that no learning algorithm can solve it efficiently.

8.1 COMPUTATIONAL COMPLEXITY OF LEARNING

Recall that a learning algorithm has access to a domain of examples, Z , a hypothesis
class, H, a loss function,
, and a training set of examples from Z that are sampled
i.i.d. according to an unknown distribution D. Given parameters , , the algorithm
should output a hypothesis h such that with probability of at least 1 ,
L D (h  ) + .
L D (h) min

h H

As mentioned before, the actual runtime of an algorithm in seconds depends

on the specific machine. To allow machine independent analysis, we use the standard approach in computational complexity theory. First, we rely on a notion of
an abstract machine, such as a Turing machine (or a Turing machine over the reals
[Blum, Shub & Smale 1989]). Second, we analyze the runtime in an asymptotic
sense, while ignoring constant factors; thus the specific machine is not important as
long as it implements the abstract machine. Usually, the asymptote is with respect
to the size of the input to the algorithm. For example, for the merge-sort algorithm
mentioned before, we analyze the runtime as a function of the number of items that
need to be sorted.
In the context of learning algorithms, there is no clear notion of input size. One
might define the input size to be the size of the training set the algorithm receives,
but that would be rather pointless. If we give the algorithm a very large number
of examples, much larger than the sample complexity of the learning problem, the
algorithm can simply ignore the extra examples. Therefore, a larger training set
does not make the learning problem more difficult, and, consequently, the runtime
available for a learning algorithm should not increase as we increase the size of the
training set. Just the same, we can still analyze the runtime as a function of natural
parameters of the problem such as the target accuracy, the confidence of achieving that accuracy, the dimensionality of the domain set, or some measures of the
complexity of the hypothesis class with which the algorithms output is compared.
To illustrate this, consider a learning algorithm for the task of learning axis
aligned rectangles. A specific problem of learning axis aligned rectangles is derived

8.1 Computational Complexity of Learning

by specifying , , and the dimension of the instance space. We can define a

sequence of problems of the type rectangles learning by fixing , and varying the
dimension to be d = 2, 3, 4, . . .. We can also define another sequence of rectangles
learning problems by fixing d, and varying the target accuracy to be = 12 , 13 , . . ..
One can of course choose other sequences of such problems. Once a sequence of the
problems is fixed, one can analyze the asymptotic runtime as a function of variables
of that sequence.
Before we introduce the formal definition, there is one more subtlety we need to
tackle. On the basis of the preceding, a learning algorithm can cheat, by transferring the computational burden to the output hypothesis. For example, the algorithm
can simply define the output hypothesis to be the function that stores the training set
in its memory, and whenever it gets a test example x it calculates the ERM hypothesis on the training set and applies it on x. Note that in this case, our algorithm has a
fixed output (namely, the function that we have just described) and can run in constant time. However, learning is still hard the hardness is now in implementing the
output classifier to obtain a label prediction. To prevent this cheating, we shall
require that the output of a learning algorithm must be applied to predict the label
of a new example in time that does not exceed the runtime of training (that is, computing the output classifier from the input training sample). In the next subsection
the advanced reader may find a formal definition of the computational complexity
of learning.

8.1.1 Formal Definition*

The definition that follows relies on a notion of an underlying abstract machine,
which is usually either a Turing machine or a Turing machine over the reals. We
will measure the computational complexity of an algorithm using the number of
operations it needs to perform, where we assume that for any machine that implements the underlying abstract machine there exists a constant c such that any such
operation can be performed on the machine using c seconds.
Definition 8.1 (The Computational Complexity of a Learning Algorithm). We
define the complexity of learning in two steps. First we consider the computational
complexity of a fixed learning problem (determined by a triplet (Z , H,
) a domain
set, a benchmark hypothesis class, and a loss function). Then, in the second step we
consider the rate of change of that complexity along a sequence of such tasks.
1. Given a function f : (0, 1)2 N, a learning task (Z , H,
), and a learning
algorithm, A, we say that A solves the learning task in time O( f ) if there
exists some constant number c, such that for every probability distribution D
over Z , and input , (0, 1), when A has access to samples generated i.i.d.
by D,

A terminates after performing at most c f ( , ) operations

The output of A, denoted h A , can be applied to predict the label of a new
example while performing at most c f ( , ) operations

75

76

The Runtime of Learning

The output of A is probably approximately correct; namely, with probability of at least 1 (over the random samples A receives), L D (h A )
minh  H L D (h  ) +
2. Consider a sequence of learning problems, (Z n , Hn ,
n )
n=1 , where problem
n is defined by a domain Z n , a hypothesis class Hn , and a loss function

n . Let A be a learning algorithm designed for solving learning problems
of this form. Given a function g : N (0, 1)2 N, we say that the runtime
of A with respect to the preceding sequence is O(g), if for all n, A solves
the problem (Z n , Hn ,
n ) in time O( fn ), where f n : (0, 1)2 N is defined by
fn ( , ) = g(n, , ).

We say that A is an efficient algorithm with respect to a sequence (Z n , Hn ,
n ) if

its runtime is O( p(n, 1/ , 1/)) for some polynomial p.
From this definition we see that the question whether a general learning problem can be solved efficiently depends on how it can be broken into a sequence
of specific learning problems. For example, consider the problem of learning a
finite hypothesis class. As we showed in previous chapters, the ERM rule over
H is guaranteed to ( , )-learn H if the number of training examples is order of
m H ( , ) = log (|H|/)/ 2 . Assuming that the evaluation of a hypothesis on an
example takes a constant time, it is possible to implement the ERM rule in time
O(|H| m H ( , )) by performing an exhaustive search over H with a training set of
size m H ( , ). For any fixed finite H, the exhaustive search algorithm runs in polynomial time. Furthermore, if we define a sequence of problems in which |Hn | = n,
then the exhaustive search is still considered to be efficient. However, if we define a
sequence of problems for which |Hn | = 2n , then the sample complexity is still polynomial in n but the computational complexity of the exhaustive search algorithm
grows exponentially with n (thus, rendered inefficient).

8.2 IMPLEMENTING THE ERM RULE

Given a hypothesis class H, the ERMH rule is maybe the most natural learning
paradigm. Furthermore, for binary classification problems we saw that if learning
is at all possible, it is possible with the ERM rule. In this section we discuss the
computational complexity of implementing the ERM rule for several hypothesis
classes.
Given a hypothesis class, H, a domain set Z , and a loss function
, the
corresponding ERMH rule can be defined as follows:
On a finite input sample S Z m output some h H that minimizes the empirical
1

loss, L S (h) = |S|
zS
(h, z).
This section studies the runtime of implementing the ERM rule for several
examples of learning tasks.

8.2 Implementing the ERM Rule

8.2.1 Finite Classes

Limiting the hypothesis class to be a finite class may be considered as a reasonably
mild restriction. For example, H can be the set of all predictors that can be implemented by a C++ program written in at most 10000 bits of code. Other examples
of useful finite classes are any hypothesis class that can be parameterized by a finite
number of parameters, where we are satisfied with a representation of each of the
parameters using a finite number of bits, for example, the class of axis aligned rectangles in the Euclidean space, Rd , when the parameters defining any given rectangle
are specified up to some limited precision.
As we have shown in previous chapters, the sample complexity of learning a
finite class is upper bounded by m H ( , ) = c log (c|H|/)/ c , where c = 1 in the realizable case and c = 2 in the nonrealizable case. Therefore, the sample complexity
has a mild dependence on the size of H. In the example of C++ programs mentioned before, the number of hypotheses is 210,000 but the sample complexity is only
c(10, 000 + log(c/))/ c .
A straightforward approach for implementing the ERM rule over a finite
hypothesis class is to perform an exhaustive search. That is, for each h H we calculate the empirical risk, L S (h), and return a hypothesis that minimizes the empirical
risk. Assuming that the evaluation of
(h, z) on a single example takes a constant
amount of time, k, the runtime of this exhaustive search becomes k|H|m, where
m is the size of the training set. If we let m to be the upper bound on the sample
complexity mentioned, then the runtime becomes k|H|c log (c|H|/)/ c .
The linear dependence of the runtime on the size of H makes this approach
inefficient (and unrealistic) for large classes. Formally, if we define a sequence
of problems (Z n , Hn ,
n )
n=1 such that log (|Hn |) = n, then the exhaustive search
approach yields an exponential runtime. In the example of C++ programs, if Hn
is the set of functions that can be implemented by a C++ program written in at
most n bits of code, then the runtime grows exponentially with n, implying that the
exhaustive search approach is unrealistic for practical use. In fact, this problem is
one of the reasons we are dealing with other hypothesis classes, like classes of linear
predictors, which we will encounter in the next chapter, and not just focusing on
finite classes.
It is important to realize that the inefficiency of one algorithmic approach (such
as the exhaustive search) does not yet imply that no efficient ERM implementation
exists. Indeed, we will show examples in which the ERM rule can be implemented
efficiently.

8.2.2 Axis Aligned Rectangles

Let Hn be the class of axis aligned rectangles in Rn , namely,
Hn = {h (a1 ,...,an ,b1 ,...,bn ) : i , ai bi }
where


h (a1 ,...,an ,b1 ,...,bn ) (x, y) =

1 if i , x i [ai , bi ]
0 otherwise

(8.1)

77

78

The Runtime of Learning

Efficiently Learnable in the Realizable Case

Consider implementing the ERM rule in the realizable case. That is, we are given
a training set S = (x1 , y1 ), . . . , (xm , ym ) of examples, such that there exists an axis
aligned rectangle, h Hn , for which h(xi ) = yi for all i . Our goal is to find such an axis
aligned rectangle with a zero training error, namely, a rectangle that is consistent
with all the labels in S.
We show later that this can be done in time O(nm). Indeed, for each i [n],
set ai = min{x i : (x, 1) S} and bi = max{x i : (x, 1) S}. In words, we take ai to be
the minimal value of the i th coordinate of a positive example in S and bi to be the
maximal value of the i th coordinate of a positive example in S. It is easy to verify
that the resulting rectangle has zero training error and that the runtime of finding
each ai and bi is O(m). Hence, the total runtime of this procedure is O(nm).

Not Efficiently Learnable in the Agnostic Case

In the agnostic case, we do not assume that some hypothesis h perfectly predicts
the labels of all the examples in the training set. Our goal is therefore to find h
that minimizes the number of examples for which yi = h(xi ). It turns out that for
many common hypothesis classes, including the classes of axis aligned rectangles we
consider here, solving the ERM problem in the agnostic setting is NP-hard (and,
in most cases, it is even NP-hard to find some h H whose error is no more than
some constant c > 1 times that of the empirical risk minimizer in H). That is, unless
P = NP, there is no algorithm whose running time is polynomial in m and n that
is guaranteed to find an ERM hypothesis for these problems (Ben-David, Eiron &
Long 2003).
On the other hand, it is worthwhile noticing that, if we fix one specific hypothesis
class, say, axis aligned rectangles in some fixed dimension, n, then there exist efficient learning algorithms for this class. In other words, there are successful agnostic
PAC learners that run in time polynomial in 1/ and 1/ (but their dependence on
the dimension n is not polynomial).
To see this, recall the implementation of the ERM rule we presented for the
realizable case, from which it follows that an axis aligned rectangle is determined by
at most 2n examples. Therefore, given a training set of size m, we can perform an
exhaustive search over all subsets of the training set of size at most 2n examples and
construct a rectangle from each such subset. Then, we can pick the rectangle with
the minimal training error. This procedure is guaranteed to find an ERM hypothesis, and the runtime of the procedure is m O(n) . It follows that if n is fixed, the runtime
is polynomial in the sample size. This does not contradict the aforementioned hardness result, since there we argued that unless P=NP one cannot have an algorithm
whose dependence on the dimension n is polynomial as well.

8.2.3 Boolean Conjunctions

A Boolean conjunction is a mapping from X = {0, 1}n to Y = {0, 1} that can be
expressed as a proposition formula of the form x i1 . . . x ik x j1 . . . x jr , for
some indices i 1 , . . . ,i k , j1 , . . . , jr [n]. The function that such a proposition formula

8.2 Implementing the ERM Rule

defines is


h(x) =

1 if x i1 = = x ik = 1 and x j1 = = x jr = 0
0 otherwise

Let HCn be the class of all Boolean conjunctions over {0, 1}n . The size of HCn is
at most 3n + 1 (since in a conjunction formula, each element of x either appears,
or appears with a negation sign, or does not appear at all, and we also have the all
negative formula). Hence, the sample complexity of learning HCn using the ERM
rule is at most d log (3/)/ .
Efficiently Learnable in the Realizable Case
Next, we show that it is possible to solve the ERM problem for HCn in time polynomial in n and m. The idea is to define an ERM conjunction by including in the
hypothesis conjunction all the literals that do not contradict any positively labeled
example. Let v1 , . . . , vm + be all the positively labeled instances in the input sample S.
We define, by induction on i m + , a sequence of hypotheses (or conjunctions). Let
h 0 be the conjunction of all possible literals. That is, h 0 = x 1 x 1 x 2 . . . x n x n .
Note that h 0 assigns the label 0 to all the elements of X . We obtain h i+1 by deleting
from the conjunction h i all the literals that are not satisfied by vi+1 . The algorithm
outputs the hypothesis h m + . Note that h m + labels positively all the positively labeled
examples in S. Furthermore, for every i m + , h i is the most restrictive conjunction
that labels v1 , . . . , vi positively. Now, since we consider learning in the realizable
setup, there exists a conjunction hypothesis, f HCn , that is consistent with all the
examples in S. Since h m + is the most restrictive conjunction that labels positively all
the positively labeled members of S, any instance labeled 0 by f is also labeled 0
by h m + . It follows that h m + has zero training error (w.r.t. S) and is therefore a legal
ERM hypothesis. Note that the running time of this algorithm is O(mn).
Not Efficiently Learnable in the Agnostic Case
As in the case of axis aligned rectangles, unless P = NP, there is no algorithm whose
running time is polynomial in m and n that guaranteed to find an ERM hypothesis
for the class of Boolean conjunctions in the unrealizable case.

8.2.4 Learning 3-Term DNF

We next show that a slight generalization of the class of Boolean conjunctions leads
to intractability of solving the ERM problem even in the realizable case. Consider
the class of 3-term disjunctive normal form formulae (3-term DNF). The instance
space is X = {0, 1}n and each hypothesis is represented by the Boolean formula of
the form h(x) = A1 (x) A2 (x) A3 (x), where each Ai (x) is a Boolean conjunction (as
defined in the previous section). The output of h(x) is 1 if either A1 (x) or A2 (x) or
A3 (x) outputs the label 1. If all three conjunctions output the label 0 then h(x) = 0.
n
be the hypothesis class of all such 3-term DNF formulae. The size
Let H3DNF
n
n
of H3DNF is at most 33n . Hence, the sample complexity of learning H3DNF
using the
ERM rule is at most 3n log (3/)/ .
However, from the computational perspective, this learning problem is hard.
It has been shown (see (Pitt & Valiant 1988, Kearns, Schapire & Sellie 1994))

79

80

The Runtime of Learning

that unless RP = NP, there is no polynomial time algorithm that properly learns
a sequence of 3-term DNF learning problems in which the dimension of the nth
problem is n. By properly we mean that the algorithm should output a hypothesis
that is a 3-term DNF formula. In particular, since ERMHn3 D N F outputs a 3-term DNF
formula it is a proper learner and therefore it is hard to implement it. The proof
uses a reduction of the graph 3-coloring problem to the problem of PAC learning
3-term DNF. The detailed technique is given in Exercise 8.4. See also (Kearns and
Vazirani 1994, section 1.4).

8.3 EFFICIENTLY LEARNABLE, BUT NOT BY A PROPER ERM

In the previous section we saw that it is impossible to implement the ERM rule
n
efficiently for the class H3DNF
of 3-DNF formulae. In this section we show that it is
possible to learn this class efficiently, but using ERM with respect to a larger class.

Representation Independent Learning Is Not Hard

Next we show that it is possible to learn 3-term DNF formulae efficiently. There is
no contradiction to the hardness result mentioned in the previous section as we now
allow representation independent learning. That is, we allow the learning algorithm to output a hypothesis that is not a 3-term DNF formula. The basic idea is to
replace the original hypothesis class of 3-term DNF formula with a larger hypothesis class so that the new class is easily learnable. The learning algorithm might
return a hypothesis that does not belong to the original hypothesis class; hence the
name representation independent learning. We emphasize that in most situations,
returning a hypothesis with good predictive ability is what we are really interested
in doing.
We start by noting that because distributes over , each 3-term DNF formula
can be rewritten as


A1 A2 A3 =

(u v w)

u A1 ,v A2 ,w A3
3

Next, let us define: : {0, 1}n {0, 1}(2n) such that for each triplet of literals u, v, w
there is a variable in the range of indicating if u v w is true or false. So, for each
3
3-DNF formula over {0, 1}n there is a conjunction over {0, 1}(2n) , with the same truth
table. Since we assume that the data is realizable, we can solve the ERM problem
3
with respect to the class of conjunctions over {0, 1}(2n) . Furthermore, the sample
complexity of learning the class of conjunctions in the higher dimensional space
is at most n 3 log (1/)/ . Thus, the overall runtime of this approach is polynomial
in n.
Intuitively, the idea is as follows. We started with a hypothesis class for which
learning is hard. We switched to another representation where the hypothesis class
is larger than the original class but has more structure, which allows for a more
efficient ERM search. In the new representation, solving the ERM problem is easy.

8.4 Hardness of Learning

,1}

(2n)

on

uncti

Conj

r {0
s ove

3-term-DNF formulae over {0,1}n

8.4 HARDNESS OF LEARNING*

We have just demonstrated that the computational hardness of implementing
ERMH does not imply that such a class H is not learnable. How can we prove that
a learning problem is computationally hard?
One approach is to rely on cryptographic assumptions. In some sense, cryptography is the opposite of learning. In learning we try to uncover some rule underlying
the examples we see, whereas in cryptography, the goal is to make sure that nobody
will be able to discover some secret, in spite of having access to some partial information about it. On that high level intuitive sense, results about the cryptographic
security of some system translate into results about the unlearnability of some corresponding task. Regrettably, currently one has no way of proving that a cryptographic
protocol is not breakable. Even the common assumption of P = NP does not suffice
for that (although it can be shown to be necessary for most common cryptographic
scenarios). The common approach for proving that cryptographic protocols are
secure is to start with some cryptographic assumptions. The more these are used
as a basis for cryptography, the stronger is our belief that they really hold (or, at
least, that algorithms that will refute them are hard to come by).
We now briefly describe the basic idea of how to deduce hardness of learnability
from cryptographic assumptions. Many cryptographic systems rely on the assumption that there exists a one way function. Roughly speaking, a one way function is
a function f : {0, 1}n {0, 1}n (more formally, it is a sequence of functions, one for
each dimension n) that is easy to compute but is hard to invert. More formally, f
can be computed in time poly(n) but for any randomized polynomial time algorithm
A, and for every polynomial p( ),
P [ f (A( f (x))) = f (x)] <

1
p(n) ,

where the probability is taken over a random choice of x according to the uniform
distribution over {0, 1}n and the randomness of A.
A one way function, f , is called trapdoor one way function if, for some polynomial function p, for every n there exists a bit-string sn (called a secret key) of length
p(n), such that there is a polynomial time algorithm that, for every n and every
x {0, 1}n , on input ( f (x), sn ) outputs x. In other words, although f is hard to invert,
once one has access to its secret key, inverting f becomes feasible. Such functions
are parameterized by their secret key.

81

82

The Runtime of Learning

Now, let Fn be a family of trapdoor functions over {0, 1}n that can be calculated
by some polynomial time algorithm. That is, we fix an algorithm that given a secret
key (representing one function in Fn ) and an input vector, it calculates the value
of the function corresponding to the secret key on the input vector in polynomial
time. Consider the task of learning the class of the corresponding inverses, H Fn =
{ f 1 : f Fn }. Since each function in this class can be inverted by some secret key
sn of size polynomial in n, the class H Fn can be parameterized by these keys and its
size is at most 2 p(n) . Its sample complexity is therefore polynomial in n. We claim
that there can be no efficient learner for this class. If there were such a learner, L,
then by sampling uniformly at random a polynomial number of strings in {0, 1}n ,
and computing f over them, we could generate a labeled training sample of pairs
( f (x), x), which should suffice for our learner to figure out an ( , ) approximation
of f 1 (w.r.t. the uniform distribution over the range of f ), which would violate the
one way property of f .
A more detailed treatment, as well as a concrete example, can be found in
(Kearns and Vazirani 1994, chapter 6). Using reductions, they also show that the
class of functions that can be calculated by small Boolean circuits is not efficiently
learnable, even in the realizable case.

8.5 SUMMARY
The runtime of learning algorithms is asymptotically analyzed as a function of different parameters of the learning problem, such as the size of the hypothesis class,
our measure of accuracy, our measure of confidence, or the size of the domain
set. We have demonstrated cases in which the ERM rule can be implemented
efficiently. For example, we derived efficient algorithms for solving the ERM problem for the class of Boolean conjunctions and the class of axis aligned rectangles,
under the realizability assumption. However, implementing ERM for these classes
in the agnostic case is NP-hard. Recall that from the statistical perspective, there
is no difference between the realizable and agnostic cases (i.e., a class is learnable in both cases if and only if it has a finite VC-dimension). In contrast, as we
saw, from the computational perspective the difference is immense. We have also
shown another example, the class of 3-term DNF, where implementing ERM is
hard even in the realizable case, yet the class is efficiently learnable by another
algorithm.
Hardness of implementing the ERM rule for several natural hypothesis classes
has motivated the development of alternative learning methods, which we will
discuss in the next part of this book.

8.6 BIBLIOGRAPHIC REMARKS

Valiant (1984) introduced the efficient PAC learning model in which the runtime of
the algorithm is required to be polynomial in 1/ , 1/, and the representation size
of hypotheses in the class. A detailed discussion and thorough bibliographic notes
are given in Kearns and Vazirani (1994).

8.7 Exercises

8.7 EXERCISES
8.1 Let H be the class of intervals on the line (formally equivalent to axis aligned rectangles in dimension n = 1). Propose an implementation of the ERMH learning rule
(in the agnostic case) that given a training set of size m, runs in time O(m 2 ).
Hint: Use dynamic programming.
8.2 Let H1 , H2 , . . . be a sequence of hypothesis classes for binary classification. Assume
that there is a learning algorithm that implements the ERM rule in the realizable
case such that the output hypothesis of the algorithm for each class Hn only depends
on O(n) examples out of the training set. Furthermore, assume that such a hypothesis can be calculated given these O(n) examples in time O(n), and that the empirical
risk of each such hypothesis can be evaluated in time O(mn). For example, if Hn is
the class of axis aligned rectangles in Rn , we saw that it is possible to find an ERM
hypothesis in the realizable case that is defined by at most 2n examples. Prove that
in such cases, it is possible to find an ERM hypothesis for Hn in the unrealizable case
in time O(mn m O(n)).
8.3 In this exercise, we present several classes for which finding an ERM classifier is
computationally hard. First, we introduce the class of n-dimensional halfspaces,
H Sn , for a domain X = Rn . This is the class of all functions of the form h w,b (x) =
sign(w, x+b) where w, x Rn , w, x is their inner product, and b R. See a detailed
description in Chapter 9.
1. Show that ERMH over the class H = H Sn of linear predictors is computationally
hard. More precisely, we consider the sequence of problems in which the dimension n grows linearly and the number of examples m is set to be some constant
times n.
Hint: You can prove the hardness by a reduction from the following problem:
Max FS: Given a system of linear inequalities, Ax > b with A R mn and
b Rm (that is, a system of m linear inequalities in n variables, x = (x1 , . . ., xn )),
find a subsystem containing as many inequalities as possible that has a
solution (such a subsystem is called feasible).
It has been shown (Sankaran 1993) that the problem Max FS is NP-hard.
Show that any algorithm that finds an ERM H Sn hypothesis for any training sample S (Rn {+1, 1})m can be used to solve the Max FS problem of size m, n.
Hint: Define a mapping that transforms linear inequalities in n variables into
labeled points in Rn , and a mapping that transforms vectors in Rn to halfspaces,
such that a vector w satisfies an inequality q if and only if the labeled point
that corresponds to q is classified correctly by the halfspace corresponding to
w. Conclude that the problem of empirical risk minimization for halfspaces in
also NP-hard (that is, if it can be solved in time polynomial in the sample size,
m, and the Euclidean dimension, n, then every problem in the class NP can be
solved in polynomial time).
2. Let X = Rn and let Hnk be the class of all intersections of k-many linear halfspaces
in Rn . In this exercise, we wish to show that ERMHnk is computationally hard for
every k 3. Precisely, we consider a sequence of problems where k 3 is a
constant and n grows linearly. The training set size, m, also grows linearly with n.
Toward this goal, consider the k-coloring problem for graphs, defined as
follows:
Given a graph G = (V , E), and a number k, determine whether there exists a
function f : V {1 . . . k} so that for every (u, v ) E, f (u) = f (v ).

83

84

The Runtime of Learning

The k-coloring problem is known to be NP-hard for every k 3 (Karp 1972).

We wish to reduce the k-coloring problem to E RMHnk : that is, to prove that if
there is an algorithm that solves the E RMHnk problem in time polynomial in k, n,
and the sample size m, then there is a polynomial time algorithm for the graph
k-coloring problem.
Given a graph G = (V , E), let {v1 . . . vn } be the vertices in V . Construct a sample
S(G) (Rn {1})m , where m = |V | + |E|, as follows:

For every vi V , construct an instance ei with a negative label.

For every edge (vi , v j ) E, construct an instance (ei + e j )/2 with a positive
label.
1. Prove that if there exists some h Hnk that has zero error over S(G) then G is
k-colorable.
Hint: Let h = kj=1 h j be an ERM classifier in Hnk over S. Define a
coloring of V by setting f (vi ) to be the minimal j such that h j (ei ) =
1. Use the fact that halfspaces are convex sets to show that it cannot
be true that two vertices that are connected by an edge have the same
color.
2. Prove that if G is k-colorable then there exists some h Hkn that has zero error
over S(G).
Hint: Given a coloring f of the vertices of G, we should come up
with k hyperplanes, h 1 . . .h k whose intersection is a perfect classifier for
S(G). Let b = 0. 6 for all of these hyperplanes and, for t k let the
i th weight of the tth hyperplane, wt,i , be 1 if f (vi ) = t and 0
otherwise.
3. On the basis of the preceding, prove that for any k 3, the ERMHnk problem
is NP-hard.
8.4 In this exercise we show that hardness of solving the ERM problem is equivalent to
hardness of proper PAC learning. Recall that by properness of the algorithm we
mean that it must output a hypothesis from the hypothesis class. To formalize this
statement, we first need the following definition.
Definition 8.2. The complexity class Randomized Polynomial (RP) time is the class
of all decision problems (that is, problems in which on any instance one has to find
out whether the answer is YES or NO) for which there exists a probabilistic algorithm (namely, the algorithm is allowed to flip random coins while it is running) with
these properties:

On any input instance the algorithm runs in polynomial time in the input size.

If the correct answer is NO, the algorithm must return NO.

If the correct answer is YES, the algorithm returns YES with probability a 1/2
and returns NO with probability 1 a.1
Clearly the class RP contains the class P. It is also known that RP is contained in the
class NP. It is not known whether any equality holds among these three complexity
classes, but it is widely believed that NP is strictly larger than RP. In particular, it is
believed that NP-hard problems cannot be solved by a randomized polynomial time
algorithm.

Show that if a class H is properly PAC learnable by a polynomial time algorithm,
then the ERMH problem is in the class RP. In particular, this implies that whenever the ERMH problem is NP-hard (for example, the class of intersections of
1

The constant 1/2 in the definition can be replaced by any constant in (0, 1).

8.7 Exercises

halfspaces discussed in the previous exercise), then, unless NP = RP, there exists
no polynomial time proper PAC learning algorithm for H.
Hint: Assume you have an algorithm A that properly PAC learns a class H in
time polynomial in some class parameter n as well as in 1/ and 1/. Your
goal is to use that algorithm as a subroutine to contract an algorithm B for
solving the ERMH problem in random polynomial time. Given a training set,
S (X {1}m ), and some h H whose error on S is zero, apply the PAC
learning algorithm to the uniform distribution over S and run it so that with
probability 0. 3 it finds a function h H that has error less than = 1/|S| (with
respect to that uniform distribution). Show that the algorithm just described
satisfies the requirements for being a RP solver for ERMH .

85

PART 2

From Theory to Algorithms

9
Linear Predictors

In this chapter we will study the family of linear predictors, one of the most useful
families of hypothesis classes. Many learning algorithms that are being widely used
in practice rely on linear predictors, first and foremost because of the ability to learn
them efficiently in many cases. In addition, linear predictors are intuitive, are easy
to interpret, and fit the data reasonably well in many natural learning problems.
We will introduce several hypothesis classes belonging to this family
halfspaces, linear regression predictors, and logistic regression predictors and
present relevant learning algorithms: linear programming and the Perceptron
algorithm for the class of halfspaces and the Least Squares algorithm for linear
regression. This chapter is focused on learning linear predictors using the ERM
approach; however, in later chapters we will see alternative paradigms for learning
these hypothesis classes.
First, we define the class of affine functions as
L d = {h w,b : w Rd , b R},
where


h w,b (x) = w, x + b =

d



wi x i

+ b.

i=1

It will be convenient also to use the notation

L d = {x  w, x + b : w Rd , b R},
which reads as follows: L d is a set of functions, where each function is parameterized
by w Rd and b R, and each such function takes as input a vector x and returns as
output the scalar w, x + b.
The different hypothesis classes of linear predictors are compositions of a function : R Y on L d . For example, in binary classification, we can choose to be
the sign function, and for regression problems, where Y = R, is simply the identity
function.
It may be more convenient to incorporate b, called the bias, into w as an extra
coordinate and add an extra coordinate with a value of 1 to all x X ; namely,
89

90

Linear Predictors

let w = (b, w1 , w2 , . . . wd ) Rd+1 and let x = (1, x 1 , x 2 , . . . , x d ) Rd+1 . Therefore,

h w,b (x) = w, x + b = w , x .
It follows that each affine function in Rd can be rewritten as a homogenous linear
function in Rd+1 applied over the transformation that appends the constant 1 to
each input vector. Therefore, whenever it simplifies the presentation, we will omit
the bias term and refer to L d as the class of homogenous linear functions of the form
h w (x) = w, x.
Throughout the book we often use the general term linear functions for both
affine functions and (homogenous) linear functions.

9.1 HALFSPACES
The first hypothesis class we consider is the class of halfspaces, designed for binary
classification problems, namely, X = Rd and Y = {1, +1}. The class of halfspaces is
defined as follows:
H Sd = sign L d = {x  sign(h w,b (x)) : h w,b L d }.
In other words, each halfspace hypothesis in H Sd is parameterized by w Rd and
b R and upon receiving a vector x the hypothesis returns the label sign(w, x + b).
To illustrate this hypothesis class geometrically, it is instructive to consider the
case d = 2. Each hypothesis forms a hyperplane that is perpendicular to the vector
w and intersects the vertical axis at the point (0, b/w2 ). The instances that are
above the hyperplane, that is, share an acute angle with w, are labeled positively.
Instances that are below the hyperplane, that is, share an obtuse angle with w, are
labeled negatively.
w

In Section 9.1.3 we will show that VCdim(H Sd ) = d + 1. It follows that we

can
using the ERM paradigm, as long as the sample size is

 learn halfspaces
d+log (1/)
. Therefore, we now discuss how to implement an ERM procedure


for halfspaces.
We introduce in the following two solutions to finding an ERM halfspace in the
realizable case. In the context of halfspaces, the realizable case is often referred to
as the separable case, since it is possible to separate with a hyperplane all the
positive examples from all the negative examples. Implementing the ERM rule in
the nonseparable case (i.e., the agnostic case) is known to be computationally hard
(Ben-David and Simon, 2001). There are several approaches to learning nonseparable data. The most popular one is to use surrogate loss functions, namely, to learn
a halfspace that does not necessarily minimize the empirical risk with the 01 loss,
but rather with respect to a diffferent loss function. For example, in Section 9.3 we

9.1 Halfspaces

will describe the logistic regression approach, which can be implemented efficiently
even in the nonseparable case. We will study surrogate loss functions in more detail
later on in Chapter 12.

9.1.1 Linear Programming for the Class of Halfspaces

Linear programs (LP) are problems that can be expressed as maximizing a linear
function subject to linear inequalities. That is,
u, w

max

wRd

subject to

Aw v

where w Rd is the vector of variables we wish to determine, A is an m d matrix,

and v Rm , u Rd are vectors. Linear programs can be solved efficiently,1 and
furthermore, there are publicly available implementations of LP solvers.
We will show that the ERM problem for halfspaces in the realizable case can be
expressed as a linear program. For simplicity, we assume the homogenous case. Let
S = {(xi , yi )}m
i=1 be a training set of size m. Since we assume the realizable case, an
ERM predictor should have zero errors on the training set. That is, we are looking
for some vector w Rd for which
sign(w, xi ) = yi ,

i = 1, . . . , m.

Equivalently, we are looking for some vector w for which

yi w, xi  > 0,

i = 1, . . . , m.

Let w be a vector that satisfies this condition (it must exist since we assume real
= w . Therefore, for all i we
izability). Define = mini (yi w , xi ) and let w
have
1
x i  = yi w , xi  1.
yi w,

We have thus shown that there exists a vector that satisfies

yi w, xi  1,

i = 1, . . . , m.

(9.1)

And clearly, such a vector is an ERM predictor.

To find a vector that satisfies Equation (9.1) we can rely on an LP solver as
follows. Set A to be the m d matrix whose rows are the instances multiplied by yi .
That is, Ai, j = yi x i, j , where x i, j is the j th element of the vector xi . Let v be the
vector (1, . . . , 1) Rm . Then, Equation (9.1) can be rewritten as
Aw v.
The LP form requires a maximization objective, yet all the w that satisfy the
constraints are equal candidates as output hypotheses. Thus, we set a dummy
objective, u = (0, . . . , 0) Rd .
1

Namely, in time polynomial in m, d, and in the representation size of real numbers.

91

92

Linear Predictors

9.1.2 Perceptron for Halfspaces

A different implementation of the ERM rule is the Perceptron algorithm of Rosenblatt (Rosenblatt 1958). The Perceptron is an iterative algorithm that constructs a
sequence of vectors w(1) , w(2) , . . .. Initially, w(1) is set to be the all-zeros vector. At
iteration t, the Perceptron finds an example i that is mislabeled by w(t) , namely, an
example for which sign(w(t) , xi ) = yi . Then, the Perceptron updates w(t) by adding
to it the instance xi scaled by the label yi . That is, w(t+1) = w(t) + yi xi . Recall that
our goal is to have yi w, xi  > 0 for all i and note that
yi w(t+1) , xi  = yi w(t) + yi xi , xi  = yi w(t) , xi  + xi 2 .
Hence, the update of the Perceptron guides the solution to be more correct on
the i th example.
Batch Perceptron
input: A training set (x1 , y1 ), . . . , (xm , ym )
initialize: w(1) = (0, . . . , 0)
for t = 1, 2, . . .
if ( i s.t. yi w(t) , xi  0) then
w(t+1) = w(t) + yi xi
else
output w(t)
The following theorem guarantees that in the realizable case, the algorithm stops
with all sample points correctly classified.
Theorem 9.1. Assume that (x1 , y1 ), . . . , (xm , ym ) is separable, let B = min{w : i
[m], yi w, xi  1}, and let R = maxi xi . Then, the Perceptron algorithm stops after
at most (R B)2 iterations, and when it stops it holds that i [m], yi w(t) , xi  > 0.
Proof. By the definition of the stopping condition, if the Perceptron stops it must
have separated all the examples. We will show that if the Perceptron runs for T
iterations, then we must have T (R B)2 , which implies the Perceptron must stop
after at most (R B)2 iterations.
Let w be a vector that achieves the minimum in the definition of B. That is,
yi w , x i  1 for all i , and among all vectors that satisfy these constraints, w is of
minimal norm.
The idea of the proof is to show that after
performing T iterations, the cosine of

T

(T
+1)
is at least R B . That is, we will show that
the angle between w and w

w , w(T +1) 
T
.
(9.2)


(T
+1)
RB
w  w

By the Cauchy-Schwartz inequality, the left-hand side of Equation (9.2) is at most
1. Therefore, Equation (9.2) would imply that

T
1
T (R B)2 ,
RB
which will conclude our proof.

9.1 Halfspaces

To show that Equation (9.2) holds, we first show that w , w(T +1)  T . Indeed,
at the first iteration, w(1) = (0, . . . , 0) and therefore w , w(1)  = 0, while on iteration
t, if we update using example (xi , yi ) we have that
w , w(t+1)  w , w(t)  = w , w(t+1) w(t) 
= w , yi xi  = yi w , xi 
1.
Therefore, after performing T iterations, we get
w , w(T +1)  =

T 


w , w(t+1)  w , w(t)  T ,

(9.3)

t=1

as required.
Next, we upper bound w(T +1) . For each iteration t we have that
w(t+1) 2 = w(t) + yi xi 2
= w(t) 2 + 2yi w(t) , xi  + yi2 xi 2
w(t) 2 + R 2

(9.4)

where the last inequality is due to the fact that example i is necessarily such that
yi w(t) , xi  0, and the norm of xi is at most R. Now, since w(1) 2 = 0, if we use
Equation (9.4) recursively for T iterations, we obtain that

w(T +1) 2 TR2 w(T +1)  T R.

(9.5)
Combining Equation (9.3) with Equation (9.5), and using the fact that w  = B, we
obtain that

w(T +1) , w 
T
T
.
=


(T
+1)
BR
w  w
 B TR
We have thus shown that Equation (9.2) holds, and this concludes our proof.
Remark 9.1. The Perceptron is simple to implement and is guaranteed to converge.
However, the convergence rate depends on the parameter B, which in some situations might be exponentially large in d. In such cases, it would be better to
implement the ERM problem by solving a linear program, as described in the previous section. Nevertheless, for many natural data sets, the size of B is not too large,
and the Perceptron converges quite fast.

9.1.3 The VC Dimension of Halfspaces

To compute the VC dimension of halfspaces, we start with the homogenous case.
Theorem 9.2. The VC dimension of the class of homogenous halfspaces in Rd is d.
Proof. First, consider the set of vectors e1 , . . . , ed , where for every i the vector ei is
the all zeros vector except 1 in the i th coordinate. This set is shattered by the class

93

94

Linear Predictors

of homogenous halfspaces. Indeed, for every labeling y1 , . . . , yd , set w = (y1 , . . . , yd ),

and then w, ei  = yi for all i .
Next, let x1 , . . . , xd+1 be a set of d + 1 vectors in Rd . Then,

there must exist
real numbers a1 , . . . , ad+1 , not all of them are zero, such that d+1
i=1 ai xi = 0. Let
I = {i : ai > 0} and J = { j : a j < 0}. Either I or J is nonempty. Let us first assume
that both of them are nonempty. Then,


ai xi =
|a j |x j .
iI

jJ

Now, suppose that x1 , . . . , xd+1 are shattered by the class of homogenous classes.
Then, there must exist a vector w such that w, xi  > 0 for all i I while w, x j  < 0
for every j J . It follows that
!
" !
"




ai xi , w =
ai xi , w =
|a j |x j , w =
|a j |x j , w < 0,
0<
iI

iI

jJ

jJ

which leads to a contradiction. Finally, if J (respectively, I ) is empty then the rightmost (respectively, left-most) inequality should be replaced by an equality, which
still leads to a contradiction.
Theorem 9.3. The VC dimension of the class of nonhomogenous halfspaces in Rd is
d + 1.
Proof. First, as in the proof of Theorem 9.2, it is easy to verify that the set of vectors
0, e1 , . . . , ed is shattered by the class of nonhomogenous halfspaces. Second, suppose
that the vectors x1 , . . . , xd+2 are shattered by the class of nonhomogenous halfspaces.
But, using the reduction we have shown in the beginning of this chapter, it follows
that there are d + 2 vectors in Rd+1 that are shattered by the class of homogenous
halfspaces. But this contradicts Theorem 9.2.

9.2 LINEAR REGRESSION

Linear regression is a common statistical tool for modeling the relationship between
some explanatory variables and some real valued outcome. Cast as a learning
problem, the domain set X is a subset of Rd , for some d, and the label set Y is the
set of real numbers. We would like to learn a linear function h : Rd R that best
approximates the relationship between our variables (say, for example, predicting
the weight of a baby as a function of her age and weight at birth). Figure 9.1 shows
an example of a linear regression predictor for d = 1.
The hypothesis class of linear regression predictors is simply the set of linear
functions,
Hreg = L d = {x  w, x + b : w Rd , b R}.
Next we need to define a loss function for regression. While in classification
the definition of the loss is straightforward, as
(h, (x, y)) simply indicates whether
h(x) correctly predicts y or not, in regression, if the babys weight is 3 kg, both the
predictions 3.00001 kg and 4 kg are wrong, but we would clearly prefer the former
over the latter. We therefore need to define how much we shall be penalized for

9.2 Linear Regression

Figure 9.1. Linear regression for d = 1. For instance, the x-axis may denote the age of the
baby, and the y-axis her weight.

the discrepancy between h(x) and y. One common way is to use the squared-loss
function, namely,

(h, (x, y)) = (h(x) y)2 .
For this loss function, the empirical risk function is called the Mean Squared Error,
namely,
m
1
(h(xi ) yi )2 .
L S (h) =
m
i=1

In the next subsection, we will see how to implement the ERM rule for linear regression with respect to the squared loss. Of course, there are a variety of
other loss functions that one can use, for example, the absolute value loss function,

(h, (x, y)) = |h(x) y|. The ERM rule for the absolute value loss function can be
implemented using linear programming (see Exercise 9.1).
Note that since linear regression is not a binary prediction task, we cannot analyze its sample complexity using the VC-dimension. One possible analysis of the
sample complexity of linear regression is by relying on the discretization trick
(see Remark 4.1 in Chapter 4); namely, if we are happy with a representation of
each element of the vector w and the bias b using a finite number of bits (say a 64
bits floating point representation), then the hypothesis class becomes finite and its
size is at most 264(d+1). We can now rely on sample complexity bounds for finite
hypothesis classes as described in Chapter 4. Note, however, that to apply the sample complexity bounds from Chapter 4 we also need that the loss function will be
bounded. Later in the book we will describe more rigorous means to analyze the
sample complexity of regression problems.

9.2.1 Least Squares

Least squares is the algorithm that solves the ERM problem for the hypothesis class
of linear regression predictors with respect to the squared loss. The ERM problem

95

96

Linear Predictors

with respect to this class, given a training set S, and using the homogenous version
of L d , is to find
1
(w, xi  yi )2 .
m
m

argmin L S (h w ) = argmin
w

i=1

To solve the problem we calculate the gradient of the objective function and
compare it to zero. That is, we need to solve
2
(w, xi  yi )xi = 0.
m
m

i=1

We can rewrite the problem as the problem Aw = b where


 m
m


xi x

yi x i .
and b =
A=
i
i=1

(9.6)

i=1

Or, in matrix form:

..
.
A=
x1
..
.

...

..
.
b=
x1
..
.

...

..
.
. ..

xm
x1
..
..
.
.

..
y1
.
..

xm .
..
ym
.

...

..
.
xm
,
..
.

(9.7)

(9.8)

If A is invertible then the solution to the ERM problem is

w = A1 b.
The case in which A is not invertible requires a few standard tools from linear algebra, which are available in Appendix C. It can be easily shown that if the training
instances do not span the entire space of Rd then A is not invertible. Nevertheless,
we can always find a solution to the system Aw = b because b is in the range of A.
Indeed, since A is symmetric we can write it using its eigenvalue decomposition as
A = V DV
, where D is a diagonal matrix and V is an orthonormal matrix (that is,
V
V is the identity d d matrix). Define D + to be the diagonal matrix such that
+
+
= 0 if Di,i = 0 and otherwise Di,i
= 1/Di,i . Now, define
Di,i
A+ = V D + V

= A+ b.
and w

Let vi denote the i th column of V . Then, we have

= A A+ b = V DV
V D + V
b = V D D + V
b =
Aw

vi v

i b.

i:Di,i =0

is the projection of b onto the span of those vectors vi for which Di,i = 0.
That is, Aw
Since the linear span of x1 , . . . , xm is the same as the linear span of those vi , and b is
= b, which concludes our argument.
in the linear span of the xi , we obtain that Aw

9.3 Logistic Regression

9.2.2 Linear Regression for Polynomial Regression Tasks

Some learning tasks call for nonlinear predictors, such as polynomial predictors.
Take, for instance, a one dimensional polynomial function of degree n, that is,
p(x) = a0 + a1 x + a2 x 2 + + an x n
where (a0 , . . . , an ) is a vector of coefficients of size n + 1. In the following we depict
a training set that is better fitted using a 3rd degree polynomial predictor than using
a linear predictor.

We will focus here on the class of one dimensional, n-degree, polynomial

regression predictors, namely,
Hnpoly = {x  p(x)},
where p is a one dimensional polynomial of degree n, parameterized by a vector of
coefficients (a0 , . . . , an ). Note that X = R, since this is a one dimensional polynomial,
and Y = R, as this is a regression problem.
One way to learn this class is by reduction to the problem of linear regression,
which we have already shown how to solve. To translate a polynomial regression
problem to a linear regression problem, we define the mapping : R Rn+1 such
that (x) = (1, x, x 2 , . . . , x n ). Then we have that
p((x)) = a0 + a1 x + a2 x 2 + + an x n = a, (x)
and we can find the optimal vector of coefficients a by using the Least Squares
algorithm as shown earlier.

9.3 LOGISTIC REGRESSION

In logistic regression we learn a family of functions h from Rd to the interval [0, 1].
However, logistic regression is used for classification tasks: We can interpret h(x) as
the probability that the label of x is 1. The hypothesis class associated with logistic
regression is the composition of a sigmoid function sig : R [0, 1] over the class of
linear functions L d . In particular, the sigmoid function used in logistic regression is
the logistic function, defined as
sig (z) =

1
.
1 + exp( z)

(9.9)

97

98

Linear Predictors

The name sigmoid means S-shaped, referring to the plot of this function, shown
in the figure:

The hypothesis class is therefore (where for simplicity we are using homogenous
linear functions):
Hsig = sig L d = {x  sig (w, x) : w Rd }.
Note that when w, x is very large then sig (w, x) is close to 1, whereas if w, x
is very small then sig (w, x) is close to 0. Recall that the prediction of the halfspace corresponding to a vector w is sign(w, x). Therefore, the predictions of the
halfspace hypothesis and the logistic hypothesis are very similar whenever |w, x| is
large. However, when |w, x| is close to 0 we have that sig (w, x) 12 . Intuitively,
the logistic hypothesis is not sure about the value of the label so it guesses that the
label is sign(w, x) with probability slightly larger than 50%. In contrast, the halfspace hypothesis always outputs a deterministic prediction of either 1 or 1, even if
|w, x| is very close to 0.
Next, we need to specify a loss function. That is, we should define how bad it is
to predict some h w (x) [0, 1] given that the true label is y {1}. Clearly, we would
like that h w (x) would be large if y = 1 and that 1 h w (x) (i.e., the probability of
predicting 1) would be large if y = 1. Note that
1 h w (x) = 1

exp ( w, x)

1
1
=
=
.
1 + exp( w, x) 1 + exp( w, x) 1 + exp(w, x)

Therefore, any reasonable loss function would increase monotonically with

1
, or equivalently, would increase monotonically with 1 + exp( yw, x).
1+exp(yw,x)
The logistic loss function used in logistic regression penalizes h w based on the log of
1 + exp( yw, x) (recall that log is a monotonic function). That is,



(h w , (x, y)) = log 1 + exp( yw, x) .
Therefore, given a training set S = (x1 , y1 ), . . . , (xm , ym ), the ERM problem associated with logistic regression is
argmin
wRd

m


1
log 1 + exp( yi w, xi ) .
m

(9.10)

i=1

The advantage of the logistic loss function is that it is a convex function with respect
to w; hence the ERM problem can be solved efficiently using standard methods.
We will study how to learn with convex functions, and in particular specify a simple
algorithm for minimizing convex functions, in later chapters.
The ERM problem associated with logistic regression (Equation (9.10)) is identical to the problem of finding a Maximum Likelihood Estimator, a well-known

9.6 Exercises

statistical approach for finding the parameters that maximize the joint probability of
a given data set assuming a specific parametric probability function. We will study
the Maximum Likelihood approach in Chapter 24.

9.4 SUMMARY
The family of linear predictors is one of the most useful families of hypothesis
classes, and many learning algorithms that are being widely used in practice rely
on linear predictors. We have shown efficient algorithms for learning linear predictors with respect to the zero-one loss in the separable case and with respect to the
squared and logistic losses in the unrealizable case. In later chapters we will present
the properties of the loss function that enable efficient learning.
Naturally, linear predictors are effective whenever we assume, as prior knowledge, that some linear predictor attains low risk with respect to the underlying
distribution. In the next chapter we show how to construct nonlinear predictors by
composing linear predictors on top of simple classes. This will enable us to employ
linear predictors for a variety of prior knowledge assumptions.

9.5 BIBLIOGRAPHIC REMARKS

The Perceptron algorithm dates back to Rosenblatt (1958). The proof of its convergence rate is due to (Agmon 1954, Novikoff 1962). Least Squares regression goes
back to Gauss (1795), Legendre (1805), and Adrain (1808).

9.6 EXERCISES
9.1 Show how to cast the ERM problem of linear regression with respect to the absolute
value loss function,
(h, (x, y)) = |h(x) y|, as a linear program; namely, show how
to write the problem
m

|w, xi  yi |
min
w

i=1

as a linear program.
Hint: Start with proving that for any c R,
|c| = min a s.t. c a and c a.
a0

9.2 Show that the matrix A defined in Equation (9.6) is invertible if and only if x1 , . . ., xm
span Rd .
9.3 Show that Theorem 9.1 is tight in the following sense: For any positive integer m,
there exist a vector w Rd (for some appropriate d) and a sequence of examples
{(x1 , y1 ), . . ., (xm , ym )} such that the following hold:

R = maxi xi  1.

w 2 = m, and for all i m, yi xi , w  1. Note that, using the notation in
Theorem 9.1, we therefore get

B = min{w : i [m], yi w, xi  1} m.

Thus, (B R)2 m.

99

100

Linear Predictors

When running the Perceptron on this sequence of examples it makes m updates

before converging.
Hint: Choose d = m and for every i choose xi = ei .
9.4 (*) Given any number m, find an example of a sequence of labeled examples
((x1 , y1 ), . . ., (xm , ym )) (R3 {1, +1})m on which the upper bound of Theorem 9.1
equals m and the perceptron algorithm is bound to make m mistakes.
Hint: Set each xi to be a third dimensional vector of the form (a, b, yi ), where
a 2 + b2 = R 2 1. Let w be the vector (0, 0, 1). Now, go over the proof of the Perceptrons upper bound (Theorem 9.1), see where we used inequalities () rather
than equalities (=), and figure out scenarios where the inequality actually holds with
equality.
9.5 Suppose we modify the Perceptron algorithm as follows: In the update step, instead
of performing w(t+1) = w(t) + yi xi whenever we make a mistake, we perform w(t+1) =
w(t) + yi xi for some > 0. Prove that the modified Perceptron will perform the
same number of iterations as the vanilla Perceptron and will converge to a vector
that points to the same direction as the output of the vanilla Perceptron.
9.6 In this problem, we will get bounds on the VC-dimension of the class of (closed)
balls in Rd , that is,
Bd = {Bv,r : v Rd ,r > 0},

where

)
Bv,r (x) =

1
0

if x v r
.
otherwise

1. Consider the mapping : Rd Rd+1 defined by (x) = (x, x2 ). Show that if
x1 , . . ., xm are shattered by Bd then (x1 ), . . ., (xm ) are shattered by the class of
halfspaces in Rd+1 (in this question we assume that sign(0) = 1). What does this
tell us about VCdim(Bd )?
2. (*) Find a set of d + 1 points in Rd that is shattered by Bd . Conclude that
d + 1 VCdim(Bd ) d + 2.

10
Boosting

Boosting is an algorithmic paradigm that grew out of a theoretical question and

became a very practical machine learning tool. The boosting approach uses a generalization of linear predictors to address two major issues that have been raised
earlier in the book. The first is the bias-complexity tradeoff. We have seen (in
Chapter 5) that the error of an ERM learner can be decomposed into a sum of
approximation error and estimation error. The more expressive the hypothesis class
the learner is searching over, the smaller the approximation error is, but the larger
the estimation error becomes. A learner is thus faced with the problem of picking a
good tradeoff between these two considerations. The boosting paradigm allows the
learner to have smooth control over this tradeoff. The learning starts with a basic
class (that might have a large approximation error), and as it progresses the class
that the predictor may belong to grows richer.
The second issue that boosting addresses is the computational complexity of
learning. As seen in Chapter 8, for many interesting concept classes the task of
finding an ERM hypothesis may be computationally infeasible. A boosting algorithm amplifies the accuracy of weak learners. Intuitively, one can think of a weak
learner as an algorithm that uses a simple rule of thumb to output a hypothesis
that comes from an easy-to-learn hypothesis class and performs just slightly better
than a random guess. When a weak learner can be implemented efficiently, boosting provides a tool for aggregating such weak hypotheses to approximate gradually
good predictors for larger, and harder to learn, classes.
In this chapter we will describe and analyze a practically useful boosting algorithm, AdaBoost (a shorthand for Adaptive Boosting). The AdaBoost algorithm
outputs a hypothesis that is a linear combination of simple hypotheses. In other
words, AdaBoost relies on the family of hypothesis classes obtained by composing
a linear predictor on top of simple classes. We will show that AdaBoost enables us
to control the tradeoff between the approximation and estimation errors by varying
a single parameter.
AdaBoost demonstrates a general theme, that will recur later in the book, of
expanding the expressiveness of linear predictors by composing them on top of
other functions. This will be elaborated in Section 10.3.
101

102

Boosting

AdaBoost stemmed from the theoretical question of whether an efficient weak

learner can be boosted into an efficient strong learner. This question was raised by
Kearns and Valiant in 1988 and solved in 1990 by Robert Schapire, then a graduate
student at MIT. However, the proposed mechanism was not very practical. In 1995,
Robert Schapire and Yoav Freund proposed the AdaBoost algorithm, which was
the first truly practical implementation of boosting. This simple and elegant algorithm became hugely popular, and Freund and Schapires work has been recognized
by numerous awards.
Furthermore, boosting is a great example for the practical impact of learning theory. While boosting originated as a purely theoretical problem, it has led to popular
and widely used algorithms. Indeed, as we shall demonstrate later in this chapter,
AdaBoost has been successfully used for learning to detect faces in images.

10.1 WEAK LEARNABILITY

Recall the definition of PAC learning given in Chapter 3: A hypothesis class, H,
is PAC learnable if there exist m H : (0, 1)2 N and a learning algorithm with the
following property: For every , (0, 1), for every distribution D over X , and for
every labeling function f : X {1}, if the realizable assumption holds with respect
to H, D, f , then when running the learning algorithm on m m H ( , ) i.i.d. examples
generated by D and labeled by f , the algorithm returns a hypothesis h such that,
with probability of at least 1 , L (D, f ) (h) .
Furthermore, the fundamental theorem of learning theory (Theorem 6.8 in
Chapter 6) characterizes the family of learnable classes and states that every PAC
learnable class can be learned using any ERM algorithm. However, the definition of
PAC learning and the fundamental theorem of learning theory ignores the computational aspect of learning. Indeed, as we have shown in Chapter 8, there are cases in
which implementing the ERM rule is computationally hard (even in the realizable
case).
However, perhaps we can trade computational hardness with the requirement
for accuracy. Given a distribution D and a target labeling function f , maybe there
exists an efficiently computable learning algorithm whose error is just slightly better
than a random guess? This motivates the following definition.
Definition 10.1 ( -Weak-Learnability).

A learning algorithm, A, is a -weak-learner for a class H if there exists a function m H : (0, 1) N such that for every (0, 1), for every distribution D over
X , and for every labeling function f : X {1}, if the realizable assumption
holds with respect to H, D, f , then when running the learning algorithm on m
m H () i.i.d. examples generated by D and labeled by f , the algorithm returns a
hypothesis h such that, with probability of at least 1 , L (D, f ) (h) 1/2 .

A hypothesis class H is -weak-learnable if there exists a -weak-learner for
that class.

This definition is almost identical to the definition of PAC learning, which here
we will call strong learning, with one crucial difference: Strong learnability implies
the ability to find an arbitrarily good classifier (with error rate at most for an

10.1 Weak Learnability

arbitrarily small > 0). In weak learnability, however, we only need to output a
hypothesis whose error rate is at most 1/2 , namely, whose error rate is slightly
better than what a random labeling would give us. The hope is that it may be easier
to come up with efficient weak learners than with efficient (full) PAC learners.
The fundamental theorem of learning (Theorem 6.8) states that if a hypothesis
class H has a VC dimension d, then the sample complexity of PAC learning H satisfies m H ( , ) C1 d+log (1/) , where C1 is a constant. Applying this with = 1/2 we
immediately obtain that if d = then H is not -weak-learnable. This implies that
from the statistical perspective (i.e., if we ignore computational complexity), weak
learnability is also characterized by the VC dimension of H and therefore is just as
hard as PAC (strong) learning. However, when we do consider computational complexity, the potential advantage of weak learning is that maybe there is an algorithm
that satisfies the requirements of weak learning and can be implemented efficiently.
One possible approach is to take a simple hypothesis class, denoted B, and to
apply ERM with respect to B as the weak learning algorithm. For this to work, we
need that B will satisfy two requirements:

ERM B is efficiently implementable.

For every sample that is labeled by some hypothesis from H, any ERM B
hypothesis will have an error of at most 1/2 .

Then, the immediate question is whether we can boost an efficient weak learner
into an efficient strong learner. In the next section we will show that this is indeed
possible, but before that, let us show an example in which efficient weak learnability
of a class H is possible using a base hypothesis class B.
Example 10.1 (Weak Learning of 3-Piece Classifiers Using Decision Stumps). Let
X = R and let H be the class of 3-piece classifiers, namely, H = {h 1 ,2 ,b : 1 , 2
R, 1 < 2 , b {1}}, where for every x,

+b if x < 1 or x > 2
h 1 ,2 ,b (x) =
b if 1 x 2
An example hypothesis (for b = 1) is illustrated as follows:
+

+
2

Let B be the class of Decision Stumps, that is, B = {x  sign(x )b : R, b

{1}}. In the following we show that ERM B is a -weak learner for H, for = 1/12.
To see that, we first show that for every distribution that is consistent with H,
there exists a decision stump with L D (h) 1/3. Indeed, just note that every classifier
in H consists of three regions (two unbounded rays and a center interval) with alternate labels. For any pair of such regions, there exists a decision stump that agrees
with the labeling of these two components. Note that for every distribution D over
R and every partitioning of the line into three such regions, one of these regions
must have D-weight of at most 1/3. Let h H be a zero error hypothesis. A decision
stump that disagrees with h only on such a region has an error of at most 1/3.

103

104

Boosting

Finally, since the VC-dimension of decision stumps is 2, if the sample size is

greater than ( log (1/)/ 2 ), then with probability of at least 1 , the ERM B rule
returns a hypothesis with an error of at most 1/3 + . Setting = 1/12 we obtain that
the error of ERM B is at most 1/3 + 1/12 = 1/2 1/12.
We see that ERM B is a -weak learner for H. We next show how to implement
the ERM rule efficiently for decision stumps.

10.1.1 Efficient Implementation of ERM for Decision Stumps

Let X = Rd and consider the base hypothesis class of decision stumps over Rd ,
namely,
HDS = {x  sign( x i ) b : R,i [d], b {1}}.
For simplicity, assume that b = 1; that is, we focus on all the hypotheses in HDS of the
form sign( x i ). Let S = ((x1 , y1 ), . . . , (xm , ym )) be a training set. We will show how
to implement an ERM rule, namely, how to find a decision stump that minimizes
L S (h). Furthermore, since in the next section we will show that AdaBoost requires
finding a hypothesis with a small risk relative to some distribution over S, we will
show here how to minimize such risk functions. Concretely, let D be a probability


vector in Rm (that is, all elements of D are nonnegative and i Di = 1). The weak
learner we describe later receives D and S and outputs a decision stump h : X Y
that minimizes the risk w.r.t. D,
L D (h) =

m


Di 1[h(xi )= yi ] .

i=1

Note that if D = (1/m, . . . , 1/m) then L D (h) = L S (h).

Recall that each decision stump is parameterized by an index j [d] and a
threshold . Therefore, minimizing L D (h) amounts to solving the problem



min min
(10.1)
Di 1[xi, j >] +
Di 1[xi, j ] .
j [d] R

i:yi =1

i:yi =1

Fix j [d] and let us sort the examples so that x 1, j x 2, j . . . x m, j . Define

x +x
 j = { i, j 2 i+1, j : i [m 1]} {(x 1, j 1), (x m, j + 1)}. Note that for any R there
exists   j that yields the same predictions for the sample S as the threshold .
Therefore, instead of minimizing over R we can minimize over  j .
This already gives us an efficient procedure: Choose j [d] and  j that
minimize the objective value of Equation (10.1). For every j and  j we have to
calculate a sum over m examples; therefore the runtime of this approach would be
O(dm 2 ). We next show a simple trick that enables us to minimize the objective in
time O(dm).
The observation is as follows. Suppose we have calculated the objective for
(x i1, j , x i, j ). Let F( ) be the value of the objective. Then, when we consider 
(x i, j , x i+1, j ) we have that
F(  ) = F( ) Di 1[yi =1] + Di 1[yi =1] = F( ) yi Di .

10.2 AdaBoost

Therefore, we can calculate the objective at  in a constant time, given the objective
at the previous threshold, . It follows that after a preprocessing step in which we
sort the examples with respect to each coordinate, the minimization problem can be
performed in time O(dm). This yields the following pseudocode.
ERM for Decision Stumps
input:
training set S = (x1 , y1 ), . . . , (xm , ym )
distribution vector D
goal: Find j  ,  that solve Equation (10.1)
initialize: F  =
for j = 1, . . . , d
sort S using the j th coordinate, and denote
def

x 1,

j x 2, j x m, j x m+1, j = x m, j + 1
F = i:yi =1 Di
if F < F 
F  = F,  = x 1, j 1, j  = j
for i = 1, . . . , m
F = F y i Di
if F < F  and x i, j = x i+1, j
F  = F,  = 12 (x i, j + x i+1, j ), j  = j
output j  , 

10.2 ADABOOST
AdaBoost (short for Adaptive Boosting) is an algorithm that has access to a weak
learner and finds a hypothesis with a low empirical risk. The AdaBoost algorithm
receives as input a training set of examples S = (x1 , y1 ), . . . , (xm , ym ), where for
each i , yi = f (xi ) for some labeling function f . The boosting process proceeds in
a sequence of consecutive rounds. At round t, the booster first defines a distribution

m
(t)
over the examples in S, denoted D(t) . That is, D(t) Rm
+ and
i=1 Di = 1. Then,
(t)
the booster passes the distribution D and the sample S to the weak learner. (That
way, the weak learner can construct i.i.d. examples according to D(t) and f .) The
weak learner is assumed to return a weak hypothesis, h t , whose error,
def

def

t = L D(t) (h t ) =

m


(t)

Di 1[h t (xi )= yi ] ,

i=1

is at most

1
2

(of course, there is a probability of at most that theweak learner

fails). Then, AdaBoost assigns a weight for h t as follows: wt = 12 log 1t 1 . That

is, the weight of h t is inversely proportional to the error of h t . At the end of the
round, AdaBoost updates the distribution so that examples on which h t errs will
get a higher probability mass while examples on which h t is correct will get a lower
probability mass. Intuitively, this will force the weak learner to focus on the problematic examples in the next round. The output of the AdaBoost algorithm is a

105

106

Boosting

strong classifier that is based on a weighted sum of all the weak hypotheses. The
pseudocode of AdaBoost is presented in the following.
AdaBoost
input:
training set S = (x1 , y1 ), . . . , (xm , ym )
weak learner WL
number of rounds T
initialize D(1) = ( m1 , . . . , m1 ).
for t = 1, . . . , T :
invoke weak learner h t = WL(D(t) , S)


(t)
compute t = m
i=1 D
i 1[yi =h t (xi )]
let wt = 12 log

(t)

Di exp (wt yi h t (xi ))

for all i = 1, . . . , m

m
(t)
j h t (x j ))
j =1 D j exp (wt y


T
hypothesis h s (x) = sign
w
h
(x)
.
t
t
t=1
(t+1)

update Di
output the

1
t

The following theorem shows that the training error of the output hypothesis
decreases exponentially fast with the number of boosting rounds.
Theorem 10.2. Let S be a training set and assume that at each iteration of AdaBoost,
the weak learner returns a hypothesis for which t 1/2 . Then, the training error
of the output hypothesis of AdaBoost is at most
1
1[h s (xi )= yi ] exp( 2 2 T ) .
m
m

L S (h s ) =

i=1

Proof. For each t, denote f t =

f T . In addition, denote

pt

w p h p . Therefore, the output of AdaBoost is

1  yi f t (xi )
e
.
m
m

Zt =

i=1

Note that for any hypothesis we have that 1[h(x)= y] eyh(x) . Therefore, L S ( f T )
Z T , so it suffices to show that Z T e2
ZT =

2T

. To upper bound Z T we rewrite it as

ZT
ZT
Z T 1
Z2 Z1
=

,
Z0
Z T 1 Z T 2
Z1 Z0

(10.2)

where we used the fact that Z 0 = 1 because f 0 0. Therefore, it suffices to show that
for every round t,
Z t+1
2
e2 .
(10.3)
Zt
To do so, we first note that using a simple inductive argument, for all t and i ,
(t+1)

Di

eyi f t (xi )
=
m y f (x ) .
j t j
j =1 e

10.2 AdaBoost

Hence,
Z t+1
=
Zt

m

yi f t+1 ( xi )
i=1 e
m

y f t (x )
j
e j

j =1

m

i=1 e

yi f t ( xi ) e yi wt+1 h t+1 ( xi )
m

e y j f t (x j )

j =1

m


(t+1) yi wt+1 h t+1 ( xi )

Di

i=1

= ewt+1

(t+1)

Di

i: yi h t+1 ( xi )=1

=e

wt+1

(1 t+1 ) + e

+ ewt+1

(t+1)

Di

i: yi h t+1 ( xi )=1
wt+1

t+1

(1 t+1 ) + 1/ t+1 1 t+1

1
=
1/ t+1 1
*

t+1
1 t+1
=
(1 t+1 ) +
t+1
1 t+1
t+1

= 2 t+1 (1 t+1 ).

By our assumption, t+1 12 . Since the function g(a) = a(1 a) is monotonically

increasing in [0, 1/2], we obtain that
*

 

1
1
2 t+1 (1 t+1 ) 2

+ = 1 4 2.
2
2

2
2
Finally, using the inequality 1 a ea we have that 1 4 2 e4 /2 = e2 .
This shows that Equation (10.3) holds and thus concludes our proof.
Each iteration of AdaBoost involves O(m) operations as well as a single call to
the weak learner. Therefore, if the weak learner can be implemented efficiently (as
happens in the case of ERM with respect to decision stumps) then the total training
process will be efficient.
Remark 10.2. Theorem 10.2 assumes that at each iteration of AdaBoost, the weak
learner returns a hypothesis with weighted sample error of at most 1/2 . According to the definition of a weak learner, it can fail with probability . Using the union
bound, the probability that the weak learner will not fail at all of the iterations is at
least 1 T . As we show in Exercise 10.1, the dependence of the sample complexity on can always be logarithmic in 1/, and therefore invoking the weak learner
with a very small is not problematic. We can therefore assume that T is also
small. Furthermore, since the weak learner is only applied with distributions over
the training set, in many cases we can implement the weak learner so that it will have
a zero probability of failure (i.e., = 0). This is the case, for example, in the weak

107

108

Boosting

learner that finds the minimum value of L D (h) for decision stumps, as described in
the previous section.
Theorem 10.2 tells us that the empirical risk of the hypothesis constructed by
AdaBoost goes to zero as T grows. However, what we really care about is the true
risk of the output hypothesis. To argue about the true risk, we note that the output
of AdaBoost is in fact a composition of a halfspace over the predictions of the T
weak hypotheses constructed by the weak learner. In the next section we show that
if the weak hypotheses come from a base hypothesis class of low VC-dimension,
then the estimation error of AdaBoost will be small; namely, the true risk of the
output of AdaBoost would not be very far from its empirical risk.

10.3 LINEAR COMBINATIONS OF BASE HYPOTHESES

As mentioned previously, a popular approach for constructing a weak learner is to
apply the ERM rule with respect to a base hypothesis class (e.g., ERM over decision
stumps). We have also seen that boosting outputs a composition of a halfspace over
the predictions of the weak hypotheses. Therefore, given a base hypothesis class B
(e.g., decision stumps), the output of AdaBoost will be a member of the following
class:

+

 T

T
wt h t (x) : w R , t, h t B .
(10.4)
L(B, T ) = x  sign
t=1

That is, each h L(B, T ) is parameterized by T base hypotheses from B and by

a vector w RT . The prediction of such an h on an instance x is obtained by first
applying the T base hypotheses to construct the vector (x) = (h 1 (x), . . . , h T (x))
RT , and then applying the (homogenous) halfspace defined by w on (x).
In this section we analyze the estimation error of L(B, T ) by bounding the VCdimension of L(B, T ) in terms of the VC-dimension of B and T. We will show that,
up to logarithmic factors, the VC-dimension of L(B, T ) is bounded by T times the
VC-dimension of B. It follows that the estimation error of AdaBoost grows linearly
with T . On the other hand, the empirical risk of AdaBoost decreases with T . In
fact, as we demonstrate later, T can be used to decrease the approximation error
of L(B, T ). Therefore, the parameter T of AdaBoost enables us to control the biascomplexity tradeoff.
To demonstrate how the expressive power of L(B, T ) increases with T , consider
the simple example, in which X = R and the base class is Decision Stumps,
HDS1 = {x  sign(x ) b : R, b {1}}.
Note that in this one dimensional case, HDS1 is in fact equivalent to (nonhomogenous) halfspaces on R.
Now, let H be the rather complex class (compared to halfspaces on the line) of
piece-wise constant functions. Let gr be a piece-wise constant function with at most
r pieces; that is, there exist thresholds = 0 < 1 < 2 < < r = such that
gr (x) =

r

i=1

i 1[x(i1 ,i ]] i , i {1}.

10.3 Linear Combinations of Base Hypotheses

Denote by Gr the class of all such piece-wise constant classifiers with at most r
pieces.
In the following we show that GT L(HDS1 , T ); namely, the class of halfspaces
over T decision stumps yields all the piece-wise constant classifiers with at most T
pieces.
Indeed, without loss of generality consider any g GT with t = ( 1)t . This
implies that if x is in the interval (t1 , t ], then g(x) = ( 1)t . For example:

Now, the function


h(x) = sign

T



wt sign(x t1) ,

(10.5)

t=1

where w1 = 0. 5 and for t > 1, wt = ( 1)t , is in L(HDS1 , T ) and is equal to g (see

Exercise 10.2).
From this example we obtain that L(HDS1 , T ) can shatter any set of T + 1
instances in R; hence the VC-dimension of L(HDS1 , T ) is at least T + 1. Therefore,
T is a parameter that can control the bias-complexity tradeoff: Enlarging T yields
a more expressive hypothesis class but on the other hand might increase the estimation error. In the next subsection we formally upper bound the VC-dimension of
L(B, T ) for any base class B.

10.3.1 The VC-Dimension of L(B, T )

The following lemma tells us that the VC-dimension of L(B, T ) is upper bounded

by O(VCdim(B)
T ) (the O notation ignores constants and logarithmic factors).
Lemma 10.3. Let B be a base class and let L(B, T ) be as defined in Equation (10.4).
Assume that both T and VCdim(B) are at least 3. Then,
VCdim(L(B, T )) T (VCdim(B) + 1) (3 log (T (VCdim(B) + 1)) + 2).
Proof. Denote d = VCdim(B). Let C = {x 1 , . . . , x m } be a set that is shattered by
L(B, T ). Each labeling of C by h L(B, T ) is obtained by first choosing h 1 , . . . , h T
B and then applying a halfspace hypothesis over the vector (h 1 (x), . . . , h T (x)). By
Sauers lemma, there are at most (em/d)d different dichotomies (i.e., labelings)
induced by B over C. Therefore, we need to choose T hypotheses, out of at most
(em/d)d different hypotheses. There are at most (em/d)d T ways to do it. Next,
for each such choice, we apply a linear predictor, which yields at most (em/T )T
dichotomies. Therefore, the overall number of dichotomies we can construct is

109

110

Boosting

upper bounded by
(em/d)d T (em/T )T m (d+1)T ,
where we used the assumption that both d and T are at least 3. Since we assume that
C is shattered, we must have that the preceding is at least 2m , which yields
2m m (d+1)T .
Therefore,
m log (m)

(d + 1)T
.
log (2)

Lemma A.1 in Appendix A tells us that a necessary condition for the preceding to
hold is that
m2

(d + 1)T
(d + 1)T
log
(d + 1)T (3 log ((d + 1)T ) + 2),
log (2)
log (2)

which concludes our proof.

In Exercise 10.4 we show that for some base classes, B, it also holds that
VCdim( L( B, T )) (VCdim( B) T ).

10.4 ADABOOST FOR FACE RECOGNITION

We now turn to a base hypothesis that has been proposed by Viola and Jones for
the task of face recognition. In this task, the instance space is images, represented
as matrices of gray level values of pixels. To be concrete, let us take images of size
24 24 pixels, and therefore our instance space is the set of real valued matrices of
size 24 24. The goal is to learn a classifier, h : X {1}, that given an image as
input, should output whether the image is of a human face or not.
Each hypothesis in the base class is of the form h(x) = f (g(x)), where f is a
decision stump hypothesis and g : R24,24 R is a function that maps an image to a
scalar. Each function g is parameterized by

An axis aligned rectangle R. Since each image is of size 24 24, there are at most
244 axis aligned rectangles.

A type, t {A, B, C, D}. Each type corresponds to a mask, as depicted in
Figure 10.1.

To calculate g we stretch the mask t to fit the rectangle R and then calculate the
sum of the pixels (that is, sum of their gray level values) that lie within the outer
rectangles and subtract it from the sum of pixels in the inner rectangles.
Since the number of such functions g is at most 244 4, we can implement a weak
learner for the base hypothesis class by first calculating all the possible outputs of
g on each image, and then apply the weak learner of decision stumps described in
the previous subsection. It is possible to perform the first step very efficiently by
a preprocessing step in which we calculate the integral image of each image in the
training set. See Exercise 10.5 for details.
In Figure 10.2 we depict the first two features selected by AdaBoost when
running it with the base features proposed by Viola and Jones.

10.6 Bibliographic Remarks

A

Figure 10.1. The four types of functions, g, used by the base hypotheses for face recognition. The value of g for type A or B is the difference between the sum of the pixels within
two rectangular regions. These regions have the same size and shape and are horizontally
or vertically adjacent. For type C, the value of g is the sum within two outside rectangles
subtracted from the sum in a center rectangle. For type D, we compute the difference
between diagonal pairs of rectangles.

Figure 10.2. The first and second features selected by AdaBoost, as implemented by Viola
and Jones. The two features are shown in the top row and then overlaid on a typical training face in the bottom row. The first feature measures the difference in intensity between
the region of the eyes and a region across the upper cheeks. The feature capitalizes on
the observation that the eye region is often darker than the cheeks. The second feature
compares the intensities in the eye regions to the intensity across the bridge of the nose.

10.5 SUMMARY
Boosting is a method for amplifying the accuracy of weak learners. In this chapter
we described the AdaBoost algorithm. We have shown that after T iterations of
AdaBoost, it returns a hypothesis from the class L(B, T ), obtained by composing a
linear classifier on T hypotheses from a base class B. We have demonstrated how the
parameter T controls the tradeoff between approximation and estimation errors. In
the next chapter we will study how to tune parameters such as T, on the basis of the
data.

10.6 BIBLIOGRAPHIC REMARKS

As mentioned before, boosting stemmed from the theoretical question of whether
an efficient weak learner can be boosted into an efficient strong learner (Kearns
& Valiant 1988) and solved by Schapire (1990). The AdaBoost algorithm has been
proposed in Freund and Schapire (1995).

111

112

Boosting

Boosting can be viewed from many perspectives. In the purely theoretical context, AdaBoost can be interpreted as a negative result: If strong learning of a
hypothesis class is computationally hard, so is weak learning of this class. This negative result can be useful for showing hardness of agnostic PAC learning of a class
B based on hardness of PAC learning of some other class H, as long as H is weakly
learnable using B. For example, Klivans and Sherstov (2006) have shown that PAC
learning of the class of intersection of halfspaces is hard (even in the realizable case).
This hardness result can be used to show that agnostic PAC learning of a single halfspace is also computationally hard (Shalev-Shwartz, Shamir & Sridharan 2010). The
idea is to show that an agnostic PAC learner for a single halfspace can yield a weak
learner for the class of intersection of halfspaces, and since such a weak learner can
be boosted, we will obtain a strong learner for the class of intersection of halfspaces.
AdaBoost also shows an equivalence between the existence of a weak learner
and separability of the data using a linear classifier over the predictions of base
hypotheses. This result is closely related to von Neumanns minimax theorem (von
Neumann 1928), a fundamental result in game theory.
AdaBoost is also related to the concept of margin, which we will study later
on in Chapter 15. It can also be viewed as a forward greedy selection algorithm, a
topic that will be presented in Chapter 25. A recent book by Schapire and Freund
(2012) covers boosting from all points of view and gives easy access to the wealth of
research that this field has produced.

10.7 EXERCISES
10.1 Boosting the Confidence: Let A be an algorithm that guarantees the following:
There exist some constant 0 (0, 1) and a function m H : (0, 1) N such that
for every (0, 1), if m m H ( ) then for every distribution D it holds that with
probability of at least 1 0 , L D ( A(S)) minhH L D (h) + .
Suggest a procedure that relies on A and learns H in the usual agnostic PAC
learning model and has a sample complexity of


2 log (4k/)
m H ( , ) k m H ( ) +
,
2
where
k = log ()/ log (0 ).
Hint: Divide the data into k + 1 chunks, where each of the first k chunks is of size
m H ( ) examples. Train the first k chunks using A. Argue that the probability that
for all of these chunks we have L D ( A(S)) > minhH L D (h) + is at most 0k /2.
Finally, use the last chunk to choose from the k hypotheses that A generated from
the k chunks (by relying on Corollary 4.6).
10.2 Prove that the function h given in Equation (10.5) equals the piece-wise constant
function defined according to the same thresholds as h.
10.3 We have informally argued that the AdaBoost algorithm uses the weighting mechanism to force the weak learner to focus on the problematic examples in the next
iteration. In this question we will find some rigorous justification for this argument.

10.7 Exercises

Show that the error of h t w.r.t. the distribution D(t+1) is exactly 1/2. That is, show
that for every t [T ]
m

(t+1)
Di
1[yi =h t (xi )] = 1/2.
i=1

10.4 In this exercise we discuss the VC-dimension of classes of the form L(B, T ). We
proved an upper bound of O(dT log (dT )), where d = VCdim(B). Here we wish to
prove an almost matching lower bound. However, that will not be the case for all
classes B.
1. Note that for every class B and every number T 1, VCdim(B)
VCdim(L(B, T )). Find a class B for which VCdim(B) = VCdim(L(B, T )) for
every T 1.
Hint: Take X to be a finite set.
2. Let Bd be the class of decision stumps over Rd . Prove that log (d)
VCdim(Bd ) 5 + 2 log (d).
Hints:

For the upper bound, rely on Exercise 10.11.

For the lower bound, assume d = 2k . Let A be a k d matrix whose columns
are all the d binary vectors in {1}k . The rows of A form a set of k vectors
in R d . Show that this set is shattered by decision stumps over Rd .
3. Let T 1 be any integer. Prove that VCdim(L(Bd , T )) 0. 5 T log (d).
Hint: Construct a set of T2 k instances by taking the rows of the matrix A from
the previous question, and the rows of the matrices 2A, 3A, 4A, . . ., T2 A. Show
that the resulting set is shattered by L(Bd , T ).
10.5 Efficiently Calculating the Viola and Jones Features Using an Integral Image: Let
A be a 24 24 matrix representing an

image. The integral image of A, denoted by
I ( A), is the matrix B such that Bi, j = i  i, j  j Ai, j .

Show that I ( A) can be calculated from A in time linear in the size of A.

Show how every Viola and Jones feature can be calculated from I ( A) in a constant amount of time (that is, the runtime does not depend on the size of the
rectangle defining the feature).

113

11
Model Selection and Validation

In the previous chapter we have described the AdaBoost algorithm and have shown
how the parameter T of AdaBoost controls the bias-complexity tradeoff. But how
do we set T in practice? More generally, when approaching some practical problem,
we usually can think of several algorithms that may yield a good solution, each of
which might have several parameters. How can we choose the best algorithm for the
particular problem at hand? And how do we set the algorithms parameters? This
task is often called model selection.
To illustrate the model selection task, consider the problem of learning a one
dimensional regression function, h : R R. Suppose that we obtain a training set as
depicted in the figure.

We can consider fitting a polynomial to the data, as described in Chapter 9. However, we might be uncertain regarding which degree d would give the best results
for our data set: A small degree may not fit the data well (i.e., it will have a large
approximation error), whereas a high degree may lead to overfitting (i.e., it will have
a large estimation error). In the following we depict the result of fitting a polynomial of degrees 2, 3, and 10. It is easy to see that the empirical risk decreases as we
enlarge the degree. However, looking at the graphs, our intuition tells us that setting
the degree to 3 may be better than setting it to 10. It follows that the empirical risk
alone is not enough for model selection.

114

11.1 Model Selection Using SRM

Degree 2

Degree 3

Degree 10

In this chapter we will present two approaches for model selection. The first
approach is based on the Structural Risk Minimization (SRM) paradigm we have
described and analyzed in Chapter 7.2. SRM is particularly useful when a learning
algorithm depends on a parameter that controls the bias-complexity tradeoff (such
as the degree of the fitted polynomial in the preceding example or the parameter T
in AdaBoost). The second approach relies on the concept of validation. The basic
idea is to partition the training set into two sets. One is used for training each of the
candidate models, and the second is used for deciding which of them yields the best
results.
In model selection tasks, we try to find the right balance between approximation and estimation errors. More generally, if our learning algorithm fails to find
a predictor with a small risk, it is important to understand whether we suffer from
overfitting or underfitting. In Section 11.3 we discuss how this can be achieved.

11.1 MODEL SELECTION USING SRM

The SRM paradigm has been described and analyzed in Section 7.2. Here we show
how SRM can be used for tuning the tradeoff between bias and complexity without
deciding on a specific hypothesis class in advance. Consider a countable sequence
of hypothesis classes H1 , H2 , H3 , . . .. For example, in the problem of polynomial
regression mentioned, we can take Hd to be the set of polynomials of degree at
most d. Another example is taking Hd to be the class L(B, d) used by AdaBoost, as
described in the previous chapter.
We assume that for every d, the class Hd enjoys the uniform convergence property (see Definition 4.3 in Chapter 4) with a sample complexity function of the
form
g(d) log (1/)
,
(11.1)
m UC
Hd ( , )
2
where g : N R is some monotonically increasing function. For example, in the case
of binary classification problems, we can take g(d) to be the VC-dimension of the
class Hd multiplied by a universal constant (the one appearing in the fundamental
theorem of learning; see Theorem 6.8). For the classes L(B, d) used by AdaBoost,
the function g will simply grow with d.
Recall that the SRM rule follows a bound minimization approach, where in
our case the bound is as follows: With probability of at least 1 , for every d N
and h Hd ,
*
g(d)( log (1/) + 2 log (d) + log ( 2 /6))
.
(11.2)
L D (h) L S (h) +
m

115

116

Model Selection and Validation

This bound, which follows directly from Theorem 7.4, shows that for every d and
every h Hd , the true risk is bounded by two terms the empirical risk, L S (h), and
a complexity term that depends on d. The SRM rule will search for d and h Hd
that minimize the right-hand side of Equation (11.2).
Getting back to the example of polynomial regression described earlier, even
though the empirical risk of the 10th degree polynomial is smaller than that of the
3rd degree polynomial, we would still prefer the 3rd degree polynomial since its
complexity (as reflected by the value of the function g(d)) is much smaller.
While the SRM approach can be useful in some situations, in many practical
cases the upper bound given in Equation (11.2) is pessimistic. In the next section we
present a more practical approach.

11.2 VALIDATION
We would often like to get a better estimation of the true risk of the output predictor
of a learning algorithm. So far we have derived bounds on the estimation error of
a hypothesis class, which tell us that for all hypotheses in the class, the true risk
is not very far from the empirical risk. However, these bounds might be loose and
pessimistic, as they hold for all hypotheses and all possible data distributions. A
more accurate estimation of the true risk can be obtained by using some of the
training data as a validation set, over which one can evalutate the success of the
algorithms output predictor. This procedure is called validation.
Naturally, a better estimation of the true risk is useful for model selection, as we
will describe in Section 11.2.2.

11.2.1 Hold Out Set

The simplest way to estimate the true error of a predictor h is by sampling an additional set of examples, independent of the training set, and using the empirical error
on this validation set as our estimator. Formally, let V = (x1 , y1 ), . . . , (xm v , ym v ) be a
set of fresh m v examples that are sampled according to D (independently of the m
examples of the training set S). Using Hoeffdings inequality ( Lemma 4.5) we have
the following:
Theorem 11.1. Let h be some predictor and assume that the loss function is in [0, 1].
Then, for every (0, 1), with probability of at least 1 over the choice of a
validation set V of size m v we have
*


log (2/)
L V (h) L D (h)
.
2 mv
The bound in Theorem 11.1 does not depend on the algorithm or the training
set used to construct h and is tighter than the usual bounds that we have seen so far.
The reason for the tightness of this bound is that it is in terms of an estimate on a
fresh validation set that is independent of the way h was generated. To illustrate this
point, suppose that h was obtained by applying an ERM predictor with respect to a
hypothesis class of VC-dimension d, over a training set of m examples. Then, from

11.2 Validation

the fundamental theorem of learning (Theorem 6.8) we obtain the bound


d + log (1/)
,
L D (h) L S (h) + C
m
where C is the constant appearing in Theorem 6.8. In contrast, from Theorem 11.1
we obtain the bound
*
log (2/)
L D (h) L V (h) +
.
2m v
Therefore, taking m v to be order of m, we obtain an estimate that is more accurate
by a factor that depends on the VC-dimension. On the other hand, the price we
pay for using such an estimate is that it requires an additional sample on top of the
sample used for training the learner.
Sampling a training set and then sampling an independent validation set is equivalent to randomly partitioning our random set of examples into two parts, using one
part for training and the other one for validation. For this reason, the validation set
is often referred to as a hold out set.

11.2.2 Validation for Model Selection

Validation can be naturally used for model selection as follows. We first train different algorithms (or the same algorithm with different parameters) on the given
training set. Let H = {h 1 , . . . , h r } be the set of all output predictors of the different
algorithms. For example, in the case of training polynomial regressors, we would
have each h r be the output of polynomial regression of degree r . Now, to choose a
single predictor from H we sample a fresh validation set and choose the predictor
that minimizes the error over the validation set. In other words, we apply ERMH
over the validation set.
This process is very similar to learning a finite hypothesis class. The only difference is that H is not fixed ahead of time but rather depends on the training set.
However, since the validation set is independent of the training set we get that
it is also independent of H and therefore the same technique we used to derive
bounds for finite hypothesis classes holds here as well. In particular, combining
Theorem 11.1 with the union bound we obtain:
Theorem 11.2. Let H = {h 1 , . . . , h r } be an arbitrary set of predictors and assume that
the loss function is in [0, 1]. Assume that a validation set V of size m v is sampled
independent of H. Then, with probability of at least 1 over the choice of V we
have
*


log (2|H|/)
.
h H, L D (h) L V (h)
2 mv
This theorem tells us that the error on the validation set approximates the true
error as long as H is not too large. However, if we try too many methods (resulting
in |H| that is large relative to the size of the validation set) then were in danger of
overfitting.

117

Model Selection and Validation

To illustrate how validation is useful for model selection, consider again the
example of fitting a one dimensional polynomial as described in the beginning of
this chapter. In the following we depict the same training set, with ERM polynomials of degree 2, 3, and 10, but this time we also depict an additional validation set
(marked as red, unfilled circles). The polynomial of degree 10 has minimal training
error, yet the polynomial of degree 3 has the minimal validation error, and hence it
will be chosen as the best model.

11.2.3 The Model-Selection Curve

The model selection curve shows the training error and validation error as a function
of the complexity of the model considered. For example, for the polynomial fitting
problem mentioned previously, the curve will look like:
0.4

Train
Validation

0.3
Error

118

0.2

0.1

0
2

10

As can be shown, the training error is monotonically decreasing as we increase the

polynomial degree (which is the complexity of the model in our case). On the other
hand, the validation error first decreases but then starts to increase, which indicates
that we are starting to suffer from overfitting.
Plotting such curves can help us understand whether we are searching the correct
regime of our parameter space. Often, there may be more than a single parameter
to tune, and the possible number of values each parameter can take might be quite
large. For example, in Chapter 13 we describe the concept of regularization, in which
the parameter of the learning algorithm is a real number. In such cases, we start

11.2 Validation

with a rough grid of values for the parameter(s) and plot the corresponding modelselection curve. On the basis of the curve we will zoom in to the correct regime
and employ a finer grid to search over. It is important to verify that we are in the
relevant regime. For example, in the polynomial fitting problem described, if we
start searching degrees from the set of values {1, 10, 20} and do not employ a finer
grid based on the resulting curve, we will end up with a rather poor model.

11.2.4 k-Fold Cross Validation

The validation procedure described so far assumes that data is plentiful and that we
have the ability to sample a fresh validation set. But in some applications, data is
scarce and we do not want to waste data on validation. The k-fold cross validation
technique is designed to give an accurate estimate of the true error without wasting
too much data.
In k-fold cross validation the original training set is partitioned into k subsets
(folds) of size m/k (for simplicity, assume that m/k is an integer). For each fold, the
algorithm is trained on the union of the other folds and then the error of its output
is estimated using the fold. Finally, the average of all these errors is the estimate of
the true error. The special case k = m, where m is the number of examples, is called
leave-one-out (LOO).
k-Fold cross validation is often used for model selection (or parameter tuning),
and once the best parameter is chosen, the algorithm is retrained using this parameter on the entire training set. A pseudocode of k-fold cross validation for model
selection is given in the following. The procedure receives as input a training set,
S, a set of possible parameter values, , an integer, k, representing the number of
folds, and a learning algorithm, A, which receives as input a training set as well as a
parameter . It outputs the best parameter as well as the hypothesis trained by
this parameter on the entire training set.
k-Fold Cross Validation for Model Selection
input:
training set S = (x1 , y1 ), . . . , (xm , ym )
set of parameter values 
learning algorithm A
integer k
partition S into S1 , S2 , . . . , Sk
foreach 
for i = 1 . . . k
h i, = A(S \ Si ; )


error( ) = k1 ki=1 L Si (h i, )
output
 = argmin [error( )]
h  = A(S;  )
The cross validation method often works very well in practice. However, it might
sometime fail, as the artificial example given in Exercise 11.1 shows. Rigorously

119

120

Model Selection and Validation

understanding the exact behavior of cross validation is still an open problem. Rogers
and Wagner (Rogers & Wagner 1978) have shown that for k local rules (e.g., k
Nearest Neighbor; see Chapter 19) the cross validation procedure gives a very good
estimate of the true error. Other papers show that cross validation works for stable
algorithms (we will study stability and its relation to learnability in Chapter 13).

11.2.5 Train-Validation-Test Split

In most practical applications, we split the available examples into three sets. The
first set is used for training our algorithm and the second is used as a validation set
for model selection. After we select the best model, we test the performance of the
output predictor on the third set, which is often called the test set. The number
obtained is used as an estimator of the true error of the learned predictor.

11.3 WHAT TO DO IF LEARNING FAILS

Consider the following scenario: You were given a learning task and have
approached it with a choice of a hypothesis class, a learning algorithm, and parameters. You used a validation set to tune the parameters and tested the learned
predictor on a test set. The test results, unfortunately, turn out to be unsatisfactory.
What went wrong then, and what should you do next?
There are many elements that can be fixed. The main approaches are listed in
the following:

Get a larger sample

Change the hypothesis class by
Enlarging it
Reducing it
Completely changing it
Changing the parameters you consider

Change the feature representation of the data

Change the optimization algorithm used to apply your learning rule

In order to find the best remedy, it is essential first to understand the cause of
the bad performance. Recall that in Chapter 5 we decomposed the true error of
the learned predictor into approximation error and estimation error. The approximation error is defined to be L D (h  ) for some h  argminhH L D (h), while the
estimation error is defined to be L D (h S ) L D (h  ), where h S is the learned predictor
(which is based on the training set S).
The approximation error of the class does not depend on the sample size or on
the algorithm being used. It only depends on the distribution D and on the hypothesis class H. Therefore, if the approximation error is large, it will not help us to
enlarge the training set size, and it also does not make sense to reduce the hypothesis class. What can be beneficial in this case is to enlarge the hypothesis class or
completely change it (if we have some alternative prior knowledge in the form of a
different hypothesis class). We can also consider applying the same hypothesis class
but on a different feature representation of the data (see Chapter 25).

11.3 What to Do If Learning Fails

The estimation error of the class does depend on the sample size. Therefore,
if we have a large estimation error we can make an effort to obtain more training
examples. We can also consider reducing the hypothesis class. However, it doesnt
make sense to enlarge the hypothesis class in that case.
Error Decomposition Using Validation
We see that understanding whether our problem is due to approximation error or
estimation error is very useful for finding the best remedy. In the previous section we
saw how to estimate L D (h S ) using the empirical risk on a validation set. However,
it is more difficult to estimate the approximation error of the class. Instead, we
give a different error decomposition, one that can be estimated from the train and
validation sets.
L D (h S ) = (L D (h S ) L V (h S )) + (L V (h S ) L S (h S )) + L S (h S ).
The first term, (L D (h S ) L V (h S )), can be bounded quite tightly using Theorem 11.1.
Intuitively, when the second term, (L V (h S ) L S (h S )), is large we say that our algorithm suffers from overfitting while when the empirical risk term, L S (h S ), is large
we say that our algorithm suffers from underfitting. Note that these two terms
are not necessarily good estimates of the estimation and approximation errors. To
illustrate this, consider the case in which H is a class of VC-dimension d, and D is
a distribution such that the approximation error of H with respect to D is 1/4. As
long as the size of our training set is smaller than d we will have L S (h S ) = 0 for every
ERM hypothesis. Therefore, the training risk, L S (h S ), and the approximation error,
L D (h  ), can be significantly different. Nevertheless, as we show later, the values of
L S (h S ) and (L V (h S ) L S (h S )) still provide us useful information.
Consider first the case in which L S (h S ) is large. We can write
L S (h S ) = (L S (h S ) L S (h  )) + (L S (h  ) L D (h  )) + L D (h  ).
When h S is an ERMH hypothesis we have that L S (h S ) L S (h  ) 0. In addition,
since h  does not depend on S, the term (L S (h  ) L D (h  )) can be bounded quite
tightly (as in Theorem 11.1). The last term is the approximation error. It follows that
if L S (h S ) is large then so is the approximation error, and the remedy to the failure
of our algorithm should be tailored accordingly (as discussed previously).
Remark 11.1. It is possible that the approximation error of our class is small, yet
the value of L S (h S ) is large. For example, maybe we had a bug in our ERM implementation, and the algorithm returns a hypothesis h S that is not an ERM. It may
also be the case that finding an ERM hypothesis is computationally hard, and our
algorithm applies some heuristic trying to find an approximate ERM. In some cases,
it is hard to know how good h S is relative to an ERM hypothesis. But, sometimes it
is possible at least to know whether there are better hypotheses. For example, in the
next chapter we will study convex learning problems in which there are optimality
conditions that can be checked to verify whether our optimization algorithm converged to an ERM solution. In other cases, the solution may depend on randomness
in initializing the algorithm, so we can try different randomly selected initial points
to see whether better solutions pop out.

121

122

Model Selection and Validation

Error

Error
Validation error

Val

idat

ion

erro

Train error

Train error
m

Figure 11.1. Examples of learning curves. Left: This learning curve corresponds to the
scenario in which the number of examples is always smaller than the VC dimension of the
class. Right: This learning curve corresponds to the scenario in which the approximation
error is zero and the number of examples is larger than the VC dimension of the class.

Next consider the case in which L S (h S ) is small. As we argued before, this does
not necessarily imply that the approximation error is small. Indeed, consider two
scenarios, in both of which we are trying to learn a hypothesis class of VC-dimension
d using the ERM learning rule. In the first scenario, we have a training set of m < d
examples and the approximation error of the class is high. In the second scenario,
we have a training set of m > 2d examples and the approximation error of the class
is zero. In both cases L S (h S ) = 0. How can we distinguish between the two cases?
Learning Curves
One possible way to distinguish between the two cases is by plotting learning curves.
To produce a learning curve we train the algorithm on prefixes of the data of increasing sizes. For example, we can first train the algorithm on the first 10% of the
examples, then on 20% of them, and so on. For each prefix we calculate the training
error (on the prefix the algorithm is being trained on) and the validation error (on
a predefined validation set). Such learning curves can help us distinguish between
the two aforementioned scenarios. In the first scenario we expect the validation
error to be approximately 1/2 for all prefixes, as we didnt really learn anything.
In the second scenario the validation error will start as a constant but then should
start decreasing (it must start decreasing once the training set size is larger than the
VC-dimension). An illustration of the two cases is given in Figure 11.1.
In general, as long as the approximation error is greater than zero we expect the
training error to grow with the sample size, as a larger amount of data points makes
it harder to provide an explanation for all of them. On the other hand, the validation
error tends to decrease with the increase in sample size. If the VC-dimension is
finite, when the sample size goes to infinity, the validation and train errors converge
to the approximation error. Therefore, by extrapolating the training and validation
curves we can try to guess the value of the approximation error, or at least to get a
rough estimate on an interval in which the approximation error resides.
Getting back to the problem of finding the best remedy for the failure of our
algorithm, if we observe that L S (h S ) is small while the validation error is large, then
in any case we know that the size of our training set is not sufficient for learning
the class H. We can then plot a learning curve. If we see that the validation error is
starting to decrease then the best solution is to increase the number of examples (if
we can afford to enlarge the data). Another reasonable solution is to decrease the

11.5 Exercises

complexity of the hypothesis class. On the other hand, if we see that the validation
error is kept around 1/2 then we have no evidence that the approximation error of
H is good. It may be the case that increasing the training set size will not help us at
all. Obtaining more data can still help us, as at some point we can see whether the
validation error starts to decrease or whether the training error starts to increase.
But, if more data is expensive, it may be better first to try to reduce the complexity
of the hypothesis class.
To summarize the discussion, the following steps should be applied:
1. If learning involves parameter tuning, plot the model-selection curve to make
sure that you tuned the parameters appropriately (see Section 11.2.3).
2. If the training error is excessively large consider enlarging the hypothesis
class, completely change it, or change the feature representation of the data.
3. If the training error is small, plot learning curves and try to deduce from them
whether the problem is estimation error or approximation error.
4. If the approximation error seems to be small enough, try to obtain more data.
If this is not possible, consider reducing the complexity of the hypothesis class.
5. If the approximation error seems to be large as well, try to change the
hypothesis class or the feature representation of the data completely.

11.4 SUMMARY
Model selection is the task of selecting an appropriate model for the learning task
based on the data itself. We have shown how this can be done using the SRM learning paradigm or using the more practical approach of validation. If our learning
algorithm fails, a decomposition of the algorithms error should be performed using
learning curves, so as to find the best remedy.

11.5 EXERCISES
11.1 Failure of k-fold cross validation Consider a case in that the label is chosen at random according to P [y = 1] = P [y = 0] = 1/2. Consider a learning algorithm that
outputs the constant predictor h(x) = 1 if the parity of the labels on the training set
is 1 and otherwise the algorithm outputs the constant predictor h(x) = 0. Prove that
the difference between the leave-one-out estimate and the true error in such a case
is always 1/2.
11.2 Let H1 , . . ., Hk be k hypothesis classes. Suppose you are given m i.i.d. training examples and you would like to learn the class H = ki=1 Hi . Consider two alternative
approaches:

Learn H on the m examples using the ERM rule

Divide the m examples into a training set of size (1 )m and a validation
set of size m, for some (0, 1). Then, apply the approach of model selection using validation. That is, first train each class Hi on the (1 )m training
examples using the ERM rule with respect to Hi , and let h 1 , . . ., h k be the resulting hypotheses. Second, apply the ERM rule with respect to the finite class
{h 1 , . . . , h k } on the m validation examples.
Describe scenarios in which the first method is better than the second and vice
versa.

123

12
Convex Learning Problems

In this chapter we introduce convex learning problems. Convex learning comprises

an important family of learning problems, mainly because most of what we can
learn efficiently falls into it. We have already encountered linear regression with
the squared loss and logistic regression, which are convex problems, and indeed
they can be learned efficiently. We have also seen nonconvex problems, such as
halfspaces with the 0-1 loss, which is known to be computationally hard to learn in
the unrealizable case.
In general, a convex learning problem is a problem whose hypothesis class is a
convex set, and whose loss function is a convex function for each example. We begin
the chapter with some required definitions of convexity. Besides convexity, we will
define Lipschitzness and smoothness, which are additional properties of the loss
function that facilitate successful learning. We next turn to defining convex learning
problems and demonstrate the necessity for further constraints such as Boundedness and Lipschitzness or Smoothness. We define these more restricted families of
learning problems and claim that Convex-Smooth/Lipschitz-Bounded problems are
learnable. These claims will be proven in the next two chapters, in which we will
present two learning paradigms that successfully learn all problems that are either
convex-Lipschitz-bounded or convex-smooth-bounded.
Finally, in Section 12.3, we show how one can handle some nonconvex problems
by minimizing surrogate loss functions that are convex (instead of the original
nonconvex loss function). Surrogate convex loss functions give rise to efficient
solutions but might increase the risk of the learned predictor.

12.1 CONVEXITY, LIPSCHITZNESS, AND SMOOTHNESS

12.1.1 Convexity
Definition 12.1 (Convex Set). A set C in a vector space is convex if for any two
vectors u, v in C, the line segment between u and v is contained in C. That is, for any
[0, 1] we have that u + (1 )v C.
124

12.1 Convexity, Lipschitzness, and Smoothness

Examples of convex and nonconvex sets in R2 are given in the following. For the
nonconvex sets, we depict two points in the set such that the line between the two
points is not contained in the set.
Nonconvex

Convex

Given [0, 1], the combination, u + (1 )v of the points u, v is called a

convex combination.
Definition 12.2 (Convex Function). Let C be a convex set. A function f : C R is
convex if for every u, v C and [0, 1],
f (u + (1 )v) f (u) + (1 ) f (v).
In words, f is convex if for any u, v, the graph of f between u and v lies below the
line segment joining f (u) and f (v). An illustration of a convex function, f : R R,
is depicted in the following.

f (v)

f (u) + (1 ) f (v)
f (u)

f ( u + (1 )v)
u

u + (1 )v

The epigraph of a function f is the set

epigraph(f) = {(x, ) : f (x) }.

(12.1)

It is easy to verify that a function f is convex if and only if its epigraph is a convex
set. An illustration of a nonconvex function f : R R, along with its epigraph, is
given in the following.

125

Convex Learning Problems

f (x)

An important property of convex functions is that every local minimum of the

function is also a global minimum. Formally, let B(u,r ) = {v : v u r } be a
ball of radius r centered around u. We say that f (u) is a local minimum of f at
u if there exists some r > 0 such that for all v B(u,r ) we have f (v) f (u). It
follows that for any v (not necessarily in B), there is a small enough > 0 such that
u + (v u) B(u,r ) and therefore
f (u) f (u + (v u)).

(12.2)

If f is convex, we also have that

f (u + (v u)) = f (v + (1 )u) (1 ) f (u) + f (v).

(12.3)

Combining these two equations and rearranging terms, we conclude that f (u)
f (v). Since this holds for every v, it follows that f (u) is also a global minimum of f .
Another important property of convex functions is that for every w we can
construct a tangent to f at w that lies below f everywhere. If f is differentiable, this tangent is the linear function l(u) = f (w) +  f (w), u w, where
f (w) is the gradient of f at w, namely, the vector of partial derivatives of f ,
f (w) = fw(w) , . . . , fw(w)
. That is, for convex differentiable functions,
d
1

u, f (u) f (w) +  f (w), u w.

(12.4)

,
f(
w

In Chapter 14 we will generalize this inequality to nondifferentiable functions. An

illustration of Equation (12.4) is given in the following.

)+

f (u)

f(
w

126

f (w)
w

If f is a scalar differentiable function, there is an easy way to check whether it is

convex.
Lemma 12.3. Let f : R R be a scalar twice differential function, and let f  , f  be
its first and second derivatives, respectively. Then, the following are equivalent:

12.1 Convexity, Lipschitzness, and Smoothness

1. f is convex
2. f  is monotonically nondecreasing
3. f  is nonnegative
Example 12.1.

The scalar function f (x) = x 2 is convex. To see this, note that f  (x) = 2x and
f  (x) = 2 > 0.

The scalar function f (x) = log (1 + exp(x)) is convex. To see this, observe that
exp(x)
1
f  (x) = 1+exp(x)
= exp (x)+1
. This is a monotonically increasing function since
the exponent function is a monotonically increasing function.

The following claim shows that the composition of a convex scalar function with
a linear function yields a convex vector-valued function.
Claim 12.4. Assume that f : Rd R can be written as f (w) = g(w, x + y), for some
x Rd , y R, and g : R R. Then, convexity of g implies the convexity of f .
Proof. Let w1 , w2 Rd and [0, 1]. We have
f (w1 + (1 )w2 ) = g(w1 + (1 )w2 , x + y)
= g(w1 , x + (1 )w2 , x + y)
= g((w1 , x + y) + (1 )(w2 , x + y))
g(w1 , x + y) + (1 )g(w2 , x + y),
where the last inequality follows from the convexity of g.
Example 12.2.

Given some x Rd and y R, let f : Rd R be defined as f (w) = (w, x y)2 .
Then, f is a composition of the function g(a) = a 2 onto a linear function, and
hence f is a convex function.

Given some x Rd and y {1}, let f : Rd R be defined as f (w) = log (1 +
exp ( yw, x)). Then, f is a composition of the function g(a) = log (1 + exp(a))
onto a linear function, and hence f is a convex function.

Finally, the following lemma shows that the maximum of convex functions is
convex and that a weighted sum of convex functions, with nonnegative weights, is
also convex.
Claim 12.5. For i = 1, . . . ,r , let fi : Rd R be a convex function. The following
functions from Rd to R are also convex.

g(x) = maxi[r] fi (x)



g(x) = ri=1 wi f i (x), where for all i , wi 0.

127

128

Convex Learning Problems

Proof. The first claim follows by

g(u + (1 )v) = max f i (u + (1 )v)
i

max [ fi (u) + (1 ) fi (v)]

i

max fi (u) + (1 ) max f i (v)

i

= g(u) + (1 )g(v).
For the second claim
g(u + (1 )v) =

wi f i (u + (1 )v)

wi [ fi (u) + (1 ) fi (v)]

wi fi (u) + (1 )

wi f i (v)

= g(u) + (1 )g(v).

Example 12.3. The function g(x) = |x| is convex. To see this, note that g(x) =
max{x, x} and that both the function f 1 (x) = x and f 2 (x) = x are convex.

12.1.2 Lipschitzness
The definition of Lipschitzness that follows is with respect to the Euclidean norm
over Rd . However, it is possible to define Lipschitzness with respect to any norm.
Definition 12.6 (Lipschitzness). Let C Rd . A function f : Rd Rk is -Lipschitz
over C if for every w1 , w2 C we have that  f (w1 ) f (w2 ) w1 w2 .
Intuitively, a Lipschitz function cannot change too fast. Note that if f : R R is
differentiable, then by the mean value theorem we have
f (w1 ) f (w2 ) = f  (u)(w1 w2 ),
where u is some point between w1 and w2 . It follows that if the derivative of f is
everywhere bounded (in absolute value) by , then the function is -Lipschitz.
Example 12.4.

The function f (x) = |x| is 1-Lipschitz over R. This follows from the triangle
inequality: For every x 1 , x 2 ,

|x 1 | |x 2 | = |x 1 x 2 + x 2 | |x 2 | |x 1 x 2 | + |x 2 | |x 2| = |x 1 x 2 |.
Since this holds for both x 1 , x 2 and x 2 , x 1 , we obtain that ||x 1| |x 2 || |x 1 x 2 |.

The function f (x) = log (1 + exp(x)) is 1-Lipschitz over R. To see this, observe
that




exp (x)
1
=
1.
| f  (x)| =


1 + exp(x)
exp( x) + 1

12.1 Convexity, Lipschitzness, and Smoothness

The function f (x) = x 2 is not -Lipschitz over R for any . To see this, take
x 1 = 0 and x 2 = 1 + , then

f (x 2 ) f (x 1 ) = (1 + )2 > (1 + ) = |x 2 x 1 |.
However, this function is -Lipschitz over the set C = {x : |x| /2}. Indeed, for
any x 1 , x 2 C we have
|x 12 x 22 | = |x 1 + x 2 | |x 1 x 2 | 2(/2) |x 1 x 2 | = |x 1 x 2 |.

The linear function f : Rd R defined by f (w) = v, w + b where v Rd is
v-Lipschitz. Indeed, using Cauchy-Schwartz inequality,

| f (w1 ) f (w2 )| = |v, w1 w2 | v w1 w2 .

The following claim shows that composition of Lipschitz functions preserves
Lipschitzness.
Claim 12.7. Let f (x) = g1 (g2 (x)), where g1 is 1 -Lipschitz and g2 is 2 -Lipschitz.
Then, f is (1 2 )-Lipschitz. In particular, if g2 is the linear function, g2 (x) = v, x+b,
for some v Rd , b R, then f is (1 v)-Lipschitz.
Proof.
| f (w1 ) f (w2 )| = |g1 (g2 (w1 )) g1 (g2 (w2 ))|
1 g2 (w1 ) g2 (w2 )
1 2 w1 w2 .

12.1.3 Smoothness
The definition of a smooth function relies on the notion of gradient. Recall that the
gradient of a differentiable function f : Rd R at w, denoted f (w), is the vector
.
of partial derivatives of f , namely, f (w) = fw(w) , . . . , fw(w)
d
1

Definition 12.8 (Smoothness). A differentiable function f : Rd R is -smooth if

its gradient is -Lipschitz; namely, for all v, w we have  f (v) f (w) vw.
It is possible to show that smoothness implies that for all v, w we have
f (v) f (w) +  f (w), v w +

v w2 .
2

(12.5)

Recall that convexity of f implies that f (v) f (w) +  f (w), v w. Therefore,
when a function is both convex and smooth, we have both upper and lower bounds
on the difference between the function and its first order approximation.
Setting v = w 1 f (w) in the right-hand side of Equation (12.5) and rearranging terms, we obtain
1
 f (w)2 f (w) f (v).
2

129

130

Convex Learning Problems

If we further assume that f (v) 0 for all v we conclude that smoothness implies the
following:
 f (w)2 2 f (w).

(12.6)

A function that satisfies this property is also called a self-bounded function.

Example 12.5.

The function f (x) = x 2 is 2-smooth. This follows directly from the fact
that f  (x) = 2x. Note that for this particular function Equation (12.5) and
Equation (12.6) hold with equality.

The function f (x) = log (1 + exp (x)) is (1/4)-smooth. Indeed, since f  (x) =
1
we have that
1+exp(x)

| f  (x)| =

exp ( x)
1
1/4.
=
(1 + exp( x))2 (1 + exp( x))(1 + exp(x))

Hence, f  is (1/4)-Lipschitz. Since this function is nonnegative, Equation (12.6)

holds as well.
The following claim shows that a composition of a smooth scalar function over a
linear function preserves smoothness.
Claim 12.9. Let f (w) = g(w, x + b), where g : R R is a -smooth function, x Rd ,
and b R. Then, f is ( x2 )-smooth.
Proof. By the chain rule we have that f (w) = g  (w, x + b)x, where g  is the
derivative of g. Using the smoothness of g and the Cauchy-Schwartz inequality we
therefore obtain
f (v) = g(v, x + b)

(v w, x)2
2

g(w, x + b) + g  (w, x + b)v w, x + (v w x)2

2
x2
v w2 .
= f (w) +  f (w), v w +
2
g(w, x + b) + g  (w, x + b)v w, x +

Example 12.6.

For any x Rd and y R, let f (w) = (w, x y)2 . Then, f is (2x2 )-smooth.

For any x Rd and y {1}, let f (w) = log (1 + exp( yw, x)). Then, f is
(x2 /4)-smooth.

12.2 CONVEX LEARNING PROBLEMS

Recall that in our general definition of learning (Definition 3.4 in Chapter 3), we
have a hypothesis class H, a set of examples Z , and a loss function
: H Z R+ .
So far in the book we have mainly thought of Z as being the product of an instance

12.2 Convex Learning Problems

space and a target space, Z = X Y, and H being a set of functions from X to

Y. However, H can be an arbitrary set. Indeed, throughout this chapter, we consider hypothesis classes H that are subsets of the Euclidean space Rd . That is, every
hypothesis is some real-valued vector. We shall, therefore, denote a hypothesis in H
by w. Now we can finally define convex learning problems:
Definition 12.10 (Convex Learning Problem). A learning problem, (H, Z ,
), is
called convex if the hypothesis class H is a convex set and for all z Z , the loss
function,
(, z), is a convex function (where, for any z,
(, z) denotes the function
f : H R defined by f (w) =
(w, z)).
Example 12.7 (Linear Regression with the Squared Loss). Recall that linear regression is a tool for modeling the relationship between some explanatory variables
and some real valued outcome (see Chapter 9). The domain set X is a subset of
Rd , for some d, and the label set Y is the set of real numbers. We would like to
learn a linear function h : Rd R that best approximates the relationship between
our variables. In Chapter 9 we defined the hypothesis class as the set of homogenous linear functions, H = {x  w, x : w Rd }, and used the squared loss function,

(h, (x, y)) = (h(x) y)2 . However, we can equivalently model the learning problem
as a convex learning problem as follows. Each linear function is parameterized by a
vector w Rd . Hence, we can define H to be the set of all such parameters, namely,
H = Rd . The set of examples is Z = X Y = Rd R = Rd+1 , and the loss function
is
(w, (x, y)) = (w, x y)2 . Clearly, the set H is a convex set. The loss function is
also convex with respect to its first argument (see Example 12.2).
Lemma 12.11. If
is a convex loss function and the class H is convex, then the
ERMH problem, of minimizing the empirical loss over H, is a convex optimization
problem (that is, a problem of minimizing a convex function over a convex set).
Proof. Recall that the ERMH problem is defined by
ERMH (S) = argmin L S (w).
wH

Since, for a sample S = z 1 , . . . , z m , for every w, L S (w) = m1 m
i=1
(w, z i ), Claim 12.5
implies that L S (w) is a convex function. Therefore, the ERM rule is a problem of
minimizing a convex function subject to the constraint that the solution should be in
a convex set.
Under mild conditions, such problems can be solved efficiently using generic
optimization algorithms. In particular, in Chapter 14 we will present a very simple
algorithm for minimizing convex functions.

12.2.1 Learnability of Convex Learning Problems

We have argued that for many cases, implementing the ERM rule for convex learning problems can be done efficiently. But is convexity a sufficient condition for the
learnability of a problem?
To make the quesion more specific: In VC theory, we saw that halfspaces in ddimension are learnable (perhaps inefficiently). We also argued in Chapter 9 using

131

132

Convex Learning Problems

the discretization trick that if the problem is of d parameters, it is learnable with

a sample complexity being a function of d. That is, for a constant d, the problem
should be learnable. So, maybe all convex learning problems over Rd , are learnable?
Example 12.8 later shows that the answer is negative, even when d is low. Not
all convex learning problems over Rd are learnable. There is no contradiction to
VC theory since VC theory only deals with binary classification while here we consider a wide family of problems. There is also no contradiction to the discretization
trick as there we assumed that the loss function is bounded and also assumed that
a representation of each parameter using a finite number of bits suffices. As we will
show later, under some additional restricting conditions that hold in many practical
scenarios, convex problems are learnable.
Example 12.8 (Nonlearnability of Linear Regression Even If d = 1). Let H = R, and
the loss be the squared loss:
(w, (x, y)) = (wx y)2 (were referring to the homogenous case). Let A be any deterministic algorithm.1 Assume, by way of contradiction,
that A is a successful PAC learner for this problem. That is, there exists a function
m(, ), such that for every distribution D and for every , if A receives a training set
of size m m( , ), it should output, with probability of at least 1 , a hypothesis
minw L D (w) .
w = A(S), such that L D (w)
. We will define
Choose = 1/100, = 1/2, let m m( , ), and set = log (100/99)
2m
two distributions, and will show that A is likely to fail on at least one of them. The
first distribution, D1 , is supported on two examples, z 1 = (1, 0) and z 2 = (, 1),
where the probability mass of the first example is while the probability mass of the
second example is 1 . The second distribution, D2 , is supported entirely on z 2 .
Observe that for both distributions, the probability that all examples of the training set will be of the second type is at least 99%. This is trivially true for D2 , whereas
for D1 , the probability of this event is
(1 )m e2m = 0. 99.
Since we assume that A is a deterministic algorithm, upon receiving a training
set of m examples, each of which is (, 1), the algorithm will output some w.
Now,
if w < 1/(2), we will set the distribution to be D1 . Hence,
(w)
2 1/(4).
L D1 (w)
However,
min L D1 (w) L D1 (0) = (1 ).
w

It follows that
min L D1 (w)
L D1 (w)
w

1
(1 ) > .
4

Therefore, such algorithm A fails on D1 . On the other hand, if w 1/(2)

1/4 while
then well set the distribution to be D2 . Then we have that L D2 (w)
minw L D2 (w) = 0, so A fails on D2 . In summary, we have shown that for every A
there exists a distribution on which A fails, which implies that the problem is not
PAC learnable.
1

Namely, given S the output of A is determined. This requirement is for the sake of simplicity. A slightly
more involved argument will show that nondeterministic algorithms will also fail to learn the problem.

12.2 Convex Learning Problems

A possible solution to this problem is to add another constraint on the hypothesis

class. In addition to the convexity requirement, we require that H will be bounded;
namely, we assume that for some predefined scalar B, every hypothesis w H
satisfies w B.
Boundedness and convexity alone are still not sufficient for ensuring that the
problem is learnable, as the following example demonstrates.
Example 12.9. As in Example 12.8, consider a regression problem with the squared
loss. However, this time let H = {w : |w| 1} R be a bounded hypothesis class. It is
easy to verify that H is convex. The argument will be the same as in Example 12.8,
except that now the two distributions, D1 , D2 will be supported on z 1 = (1/, 0) and
z 2 = (1, 1). If the algorithm A returns w < 1/2 upon receiving m examples of the
second type, then we will set the distribution to be D1 and have that
2
L D1 (w)
min L D1 (w) (w/)

L D1 (0) 1/(4) (1 ) > .

w

Similarly, if w 1/2 we will set the distribution to be D2 and have that

min L D2 (w) ( 1/2 + 1)2 0 > .
L D2 (w)
w

This example shows that we need additional assumptions on the learning problem, and this time the solution is in Lipschitzness or smoothness of the loss function.
This motivates a definition of two families of learning problems, convex-Lipschitzbounded and convex-smooth-bounded, which are defined later.

12.2.2 Convex-Lipschitz/Smooth-Bounded Learning Problems

Definition 12.12 (Convex-Lipschitz-Bounded Learning Problem). A learning problem, (H, Z ,
), is called Convex-Lipschitz-Bounded, with parameters , B if the
following holds:

The hypothesis class H is a convex set and for all w H we have w B.

For all z Z , the loss function,
(, z), is a convex and -Lipschitz function.

Example 12.10. Let X = {x Rd : x } and Y = R. Let H = {w Rd : w B}

and let the loss function be
(w, (x, y)) = |w, x y|. This corresponds to a regression
problem with the absolute-value loss, where we assume that the instances are in a
ball of radius and we restrict the hypotheses to be homogenous linear functions
defined by a vector w whose norm is bounded by B. Then, the resulting problem is
Convex-Lipschitz-Bounded with parameters , B.
Definition 12.13 (Convex-Smooth-Bounded Learning Problem). A learning problem, (H, Z ,
), is called Convex-Smooth-Bounded, with parameters , B if the
following holds:

The hypothesis class H is a convex set and for all w H we have w B.

For all z Z , the loss function,
(, z), is a convex, nonnegative, and -smooth
function.

133

134

Convex Learning Problems

Note that we also required that the loss function is nonnegative. This is needed
to ensure that the loss function is self-bounded, as described in the previous section.
Example 12.11. Let X = {x Rd : x /2} and Y = R. Let H = {w Rd : w B}
and let the loss function be
(w, (x, y)) = (w, x y)2 . This corresponds to a regression problem with the squared loss, where we assume that the instances are in a
ball of radius /2 and we restrict the hypotheses to be homogenous linear functions
defined by a vector w whose norm is bounded by B. Then, the resulting problem is
Convex-Smooth-Bounded with parameters , B.
We claim that these two families of learning problems are learnable. That is, the
properties of convexity, boundedness, and Lipschitzness or smoothness of the loss
function are sufficient for learnability. We will prove this claim in the next chapters
by introducing algorithms that learn these problems successfully.

12.3 SURROGATE LOSS FUNCTIONS

As mentioned, and as we will see in the next chapters, convex problems can be
learned effficiently. However, in many cases, the natural loss function is not convex
and, in particular, implementing the ERM rule is hard.
As an example, consider the problem of learning the hypothesis class of
halfspaces with respect to the 01 loss. That is,

01 (w, (x, y)) = 1[y=sign(w,x)] = 1[yw,x0] .
This loss function is not convex with respect to w and indeed, when trying to minimize the empirical risk with respect to this loss function we might encounter local
minima (see Exercise 12.1). Furthermore, as discussed in Chapter 8, solving the
ERM problem with respect to the 01 loss in the unrealizable case is known to be
NP-hard.
To circumvent the hardness result, one popular approach is to upper bound the
nonconvex loss function by a convex surrogate loss function. As its name indicates,
the requirements from a convex surrogate loss are as follows:
1. It should be convex.
2. It should upper bound the original loss.
For example, in the context of learning halfspaces, we can define the so-called hinge
loss as a convex surrogate for the 01 loss, as follows:
def

hinge (w, (x, y)) = max{0, 1 yw, x}.

Clearly, for all w and all (x, y),
01 (w, (x, y))
hinge (w, (x, y)). In addition, the
convexity of the hinge loss follows directly from Claim 12.5. Hence, the hinge loss
satisfies the requirements of a convex surrogate loss function for the zero-one loss.
An illustration of the functions
01 and
hinge is given in the following.

12.4 Summary

hinge

0 1

y w, x

Once we have defined the surrogate convex loss, we can learn the problem with
respect to it. The generalization requirement from a hinge loss learner will have the
form
hinge

LD

hinge

(A(S)) min L D
wH

(w) + ,

hinge

where L D (w) = E(x,y)D [
hinge (w, (x, y))]. Using the surrogate property, we can
lower bound the left-hand side by L 01
D (A(S)), which yields
hinge

L 01
D (A(S)) min L D
wH

(w) + .

We can further rewrite the upper bound as follows:



hinge
01
01
L 01
(A(S))

min
L
(w)
+
min
L
(w)

min
L
(w)
+ .
D
D
D
D
wH

wH

wH

That is, the 01 error of the learned predictor is upper bounded by three terms:

Approximation error: This is the term minwH L 01
D (w), which measures how
well the hypothesis class performs on the distribution. We already elaborated
on this error term in Chapter 5.

Estimation error: This is the error that results from the fact that we only receive
a training set and do not observe the distribution D. We already elaborated on
this error term in Chapter 5.


hinge

Optimization error: This is the term minwH L D (w) minwH L 01

D (w) that
measures the difference between the approximation error with respect to the
surrogate loss and the approximation error with respect to the original loss.
The optimization error is a result of our inability to minimize the training loss
with respect to the original loss. The size of this error depends on the specific
distribution of the data and on the specific surrogate loss we are using.

12.4 SUMMARY
We introduced two families of learning problems: convex-Lipschitz-bounded and
convex-smooth-bounded. In the next two chapters we will describe two generic

135

136

Convex Learning Problems

learning algorithms for these families. We also introduced the notion of convex
surrogate loss function, which enables us also to utilize the convex machinery for
nonconvex problems.

12.5 BIBLIOGRAPHIC REMARKS

There are several excellent books on convex analysis and optimization (Boyd &
Vandenberghe 2004, Borwein & Lewis 2006, Bertsekas 1999, Hiriart-Urruty &
Lemarchal 1993). Regarding learning problems, the family of convex-Lipschitzbounded problems was first studied by Zinkevich (2003) in the context of online
learning and by Shalev-Shwartz, Shamir, Sridharan, and Srebro ((2009)) in the
context of PAC learning.

12.6 EXERCISES
12.1 Construct an example showing that the 01 loss function may suffer from local
minima; namely, construct a training sample S (X {1})m (say, for X = R2 ), for
which there exist a vector w and some > 0 such that
1. For any w such that w w  we have L S (w) L S (w ) (where the loss here
is the 01 loss). This means that w is a local minimum of L S .
2. There exists some w such that L S (w ) < L S (w). This means that w is not a
global minimum of L S .
12.2 Consider the learning problem of logistic regression: Let H = X = {x Rd : x B},
for some scalar B > 0, let Y = {1}, and let the loss function
be defined as

(w, (x, y)) = log (1 + exp ( yw, x)). Show that the resulting learning problem is both convex-Lipschitz-bounded and convex-smooth-bounded. Specify the
parameters of Lipschitzness and smoothness.
12.3 Consider the problem of learning halfspaces with the hinge loss. We limit our
domain to the Euclidean ball with radius R. That is, X = {x : x2 R}. The label set
is Y = {1} and the loss function
is defined by
(w, (x, y)) = max{0, 1 yw, x}.
We already know that the loss function is convex. Show that it is R-Lipschitz.
12.4 (*) Convex-Lipschitz-Boundedness Is Not Sufficient for Computational Efficiency:
In the next chapter we show that from the statistical perspective, all convexLipschitz-bounded problems are learnable (in the agnostic PAC model). However,
our main motivation to learn such problems resulted from the computational perspective convex optimization is often efficiently solvable. Yet the goal of this
exercise is to show that convexity alone is not sufficient for efficiency. We show
that even for the case d = 1, there is a convex-Lipschitz-bounded problem which
cannot be learned by any computable learner.
Let the hypothesis class be H = [0, 1] and let the example domain, Z , be the set of
all Turing machines. Define the loss function as follows. For every Turing machine
T Z , let
(0, T ) = 1 if T halts on the input 0 and
(0, T ) = 0 if T doesnt halt on the
input 0. Similarly, let
(1, T ) = 0 if T halts on the input 0 and
(1, T ) = 1 if T doesnt
halt on the input 0. Finally, for h (0, 1), let
(h, T ) = h
(0, T ) + (1 h)
(1, T ).
1. Show that the resulting learning problem is convex-Lipschitz-bounded.
2. Show that no computable algorithm can learn the problem.

13
Regularization and Stability

In the previous chapter we introduced the families of convex-Lipschitz-bounded

and convex-smooth-bounded learning problems. In this section we show that all
learning problems in these two families are learnable. For some learning problems
of this type it is possible to show that uniform convergence holds; hence they are
learnable using the ERM rule. However, this is not true for all learning problems of
this type. Yet, we will introduce another learning rule and will show that it learns all
convex-Lipschitz-bounded and convex-smooth-bounded learning problems.
The new learning paradigm we introduce in this chapter is called Regularized
Loss Minimization, or RLM for short. In RLM we minimize the sum of the empirical
risk and a regularization function. Intuitively, the regularization function measures
the complexity of hypotheses. Indeed, one interpretation of the regularization function is the structural risk minimization paradigm we discussed in Chapter 7. Another
view of regularization is as a stabilizer of the learning algorithm. An algorithm is
considered stable if a slight change of its input does not change its output much. We
will formally define the notion of stability (what we mean by slight change of input
and by does not change much the output) and prove its close relation to learnability. Finally, we will show that using the squared
2 norm as a regularization function
stabilizes all convex-Lipschitz or convex-smooth learning problems. Hence, RLM
can be used as a general learning rule for these families of learning problems.

13.1 REGULARIZED LOSS MINIMIZATION

Regularized Loss Minimization (RLM) is a learning rule in which we jointly minimize the empirical risk and a regularization function. Formally, a regularization
function is a mapping R : Rd R, and the regularized loss minimization rule outputs
a hypothesis in


argmin L S (w) + R(w) .
(13.1)
w

Regularized loss minimization shares similarities with minimum description length

algorithms and structural risk minimization (see Chapter 7). Intuitively, the complexity of hypotheses is measured by the value of the regularization function, and
137

138

Regularization and Stability

the algorithm balances between low empirical risk and simpler, or less complex,
hypotheses.
There are many possible regularization functions one can use, reflecting some
prior belief about the problem (similarly to the description language in Minimum
Description Length). Throughout this section we will focus on one of the most simple regularization functions:
R(w) = w2 , where > 0 is a scalar and the norm is


d
2
the
2 norm, w =
i=1 wi . This yields the learning rule:


(13.2)
A(S) = argmin L S (w) + w2 .
w

This type of regularization function is often called Tikhonov regularization.

As mentioned before, one interpretation of Equation (13.2) is using structural
risk minimization, where the norm of w is a measure of its complexity. Recall that
in the previous chapter we introduced the notion of bounded hypothesis classes.
Therefore, we can define a sequence of hypothesis classes, H1 H2 H3 . . ., where
Hi = {w : w2 i }. If the sample complexity of each Hi depends on i then the RLM
rule is similar to the SRM rule for this sequence of nested classes.
A different interpretation of regularization is as a stabilizer. In the next section
we define the notion of stability and prove that stable learning rules do not overfit.
But first, let us demonstrate the RLM rule for linear regression with the squared
loss.

13.1.1 Ridge Regression

Applying the RLM rule with Tikhonov regularization to linear regression with the
squared loss, we obtain the following learning rule:


m
1 1
2
2
(w, xi  yi ) .
argmin w2 +
(13.3)
m
2
wRd
i=1

Performing linear regression using Equation (13.3) is called ridge regression.

To solve Equation (13.3) we compare the gradient of the objective to zero and
obtain the set of linear equations
(2m I + A)w = b,
where I is the identity matrix and A, b are as defined in Equation (9.6), namely,
 m

m



A=
xi xi
yi x i .
(13.4)
and b =
i=1

i=1

Since A is a positive semidefinite matrix, the matrix 2m I + A has all its eigenvalues
bounded below by 2m. Hence, this matrix is invertible and the solution to ridge
regression becomes
(13.5)
w = (2m I + A)1 b.
In the next section we formally show how regularization stabilizes the algorithm
and prevents overfitting. In particular, the analysis presented in the next sections
(particularly, Corollary 13.11) will yield:

13.2 Stable Rules Do Not Overfit

Theorem 13.1. Let D be a distribution over X [ 1, 1], where X = {x Rd : x 1}.

Let H = {w Rd : w B}. For any (0, 1), let m 150 B 2/ 2 . Then, applying the
ridge regression algorithm with parameter = /(3 B 2 ) satisfies
E [ L D ( A( S))] min L D (w) + .
wH

SD m

Remark 13.1. The preceding theorem tells us how many examples are needed to
guarantee that the expected value of the risk of the learned predictor will be bounded
by the approximation error of the class plus . In the usual definition of agnostic
PAC learning we require that the risk of the learned predictor will be bounded
with probability of at least 1 . In Exercise 13.1 we show how an algorithm with a
bounded expected risk can be used to construct an agnostic PAC learner.

13.2 STABLE RULES DO NOT OVERFIT

Intuitively, a learning algorithm is stable if a small change of the input to the algorithm does not change the output of the algorithm much. Of course, there are many
ways to define what we mean by a small change of the input and what we mean
by does not change the output much. In this section we define a specific notion of
stability and prove that under this definition, stable rules do not overfit.
Let A be a learning algorithm, let S = (z 1 , . . . , z m ) be a training set of m examples,
and let A(S) denote the output of A. The algorithm A suffers from overfitting if the
difference between the true risk of its output, L D (A(S)), and the empirical risk of its
output, L S (A(S)), is large. As mentioned in Remark 13.1, throughout this chapter
we focus on the expectation (with respect to the choice of S) of this quantity, namely,
E S [L D (A(S)) L S (A(S))].
We next define the notion of stability. Given the training set S and an additional
example z  , let S (i) be the training set obtained by replacing the i th example of S
with z  ; namely, S (i) = (z 1 , . . . , z i1 , z  , z i+1 , . . . , z m ). In our definition of stability, a
small change of the input means that we feed A with S (i) instead of with S. That is,
we only replace one training example. We measure the effect of this small change
of the input on the output of A, by comparing the loss of the hypothesis A(S) on
z i to the loss of the hypothesis A(S (i) ) on z i . Intuitively, a good learning algorithm
will have
(A(S (i) ), z i )
(A(S), z i ) 0, since in the first term the learning algorithm
does not observe the example z i while in the second term z i is indeed observed. If
the preceding difference is very large we suspect that the learning algorithm might
overfit. This is because the learning algorithm drastically changes its prediction on
z i if it observes it in the training set. This is formalized in the following theorem.
Theorem 13.2. Let D be a distribution. Let S = (z 1 , . . . , z m ) be an i.i.d. sequence of
examples and let z  be another i.i.d. example. Let U (m) be the uniform distribution
over [m]. Then, for any learning algorithm,
E [L D (A(S)) L S (A(S))] =

SD m

(S,z  )D m+1 ,iU (m)

[
(A(S (i) , z i ))
(A(S), z i )].

(13.6)

Proof. Since S and z  are both drawn i.i.d. from D, we have that for every i ,
E [L D (A(S))] = E  [
(A(S), z  )] = E  [
(A(S (i) ), z i )].
S

S,z

S,z

139

140

Regularization and Stability

On the other hand, we can write

E [L S (A(S))] = E [
(A(S), z i )].
S

S,i

Combining the two equations we conclude our proof.

When the right-hand side of Equation (13.6) is small, we say that A is a stable algorithm changing a single example in the training set does not lead to a
significant change. Formally,
Definition 13.3 (On-Average-Replace-One-Stable). Let : N R be a monotonically decreasing function. We say that a learning algorithm A is on-average-replaceone-stable with rate (m) if for every distribution D
E

(S,z  )D m+1 ,iU (m)

[
(A(S (i) , z i ))
(A(S), z i )] (m).

Theorem 13.2 tells us that a learning algorithm does not overfit if and only if
it is on-average-replace-one-stable. Of course, a learning algorithm that does not
overfit is not necessarily a good learning algorithm take, for example, an algorithm A that always outputs the same hypothesis. A useful algorithm should find
a hypothesis that on one hand fits the training set (i.e., has a low empirical risk)
and on the other hand does not overfit. Or, in light of Theorem 13.2, the algorithm
should both fit the training set and at the same time be stable. As we shall see, the
parameter of the RLM rule balances between fitting the training set and being
stable.

13.3 TIKHONOV REGULARIZATION AS A STABILIZER

In the previous section we saw that stable rules do not overfit. In this section we
show that applying the RLM rule with Tikhonov regularization, w2 , leads to a
stable algorithm. We will assume that the loss function is convex and that it is either
Lipschitz or smooth.
The main property of the Tikhonov regularization that we rely on is that it makes
the objective of RLM strongly convex, as defined in the following.
Definition 13.4 (Strongly Convex Functions). A function f is -strongly convex if
for all w, u, and (0, 1) we have

f (w + (1 )u) f (w) + (1 ) f (u) (1 )w u2 .

2
Clearly, every convex function is 0-strongly convex. An illustration of strong
convexity is given in the following figure.

13.3 Tikhonov Regularization as a Stabilizer

f (u)

f (w)
w
w + (1) u

(1 ) || u w ||2

The following lemma implies that the objective of RLM is (2)-strongly convex.
In addition, it underscores an important property of strong convexity.
Lemma 13.5.
1. The function f (w) = w2 is 2-strongly convex.
2. If f is -strongly convex and g is convex, then f + g is -strongly convex.
3. If f is -strongly convex and u is a minimizer of f , then, for any w,

f (w) f (u) w u2 .

2
Proof. The first two points follow directly from the definition. To prove the last
point, we divide the definition of strong convexity by and rearrange terms to get
that

f (u + (w u)) f (u)
f (w) f (u) (1 )w u2 .

2
Taking the limit 0 we obtain that the right-hand side converges to f (w) f (u)

2
2 w u . On the other hand, the left-hand side becomes the derivative of the
function g() = f (u + (w u)) at = 0. Since u is a minimizer of f , it follows that
= 0 is a minimizer of g, and therefore the left-hand side of the preceding goes to
zero in the limit 0, which concludes our proof.
We now turn to prove that RLM is stable. Let S = (z 1 , . . . , z m ) be a training set,
let z  be an additional example, and let S (i) = (z 1 , . . . , z i1 , z  , z i+1 , . . . , z m ). Let A be
the RLM rule, namely,


A(S) = argmin L S (w) + w2 .
w

Denote f S (w) = L S (w) + w2 , and on the basis of Lemma 13.5 we know that f S is
(2)-strongly convex. Relying on part 3 of the lemma, it follows that for any v,
f S (v) f S (A(S)) v A(S)2 .

(13.7)

141

142

Regularization and Stability

On the other hand, for any v and u, and for all i , we have
f S (v) f S (u) = L S (v) + v2 (L S (u) + u2 )

(13.8)

= L S (i) (v) + v (L S (i) (u) + u )

2

(v, z i )
(u, z i )
(u, z  )
(v, z  )

+
.
m
m

In particular, choosing v = A(S (i) ), u = A(S), and using the fact that v minimizes
L S (i) (w) + w2 , we obtain that
f S (A(S (i) )) f S (A(S))

(A(S (i) ), z i )
(A(S), z i )
(A(S), z  )
(A(S (i) ), z  )

+
.
m
m
(13.9)

Combining this with Equation (13.7) we obtain that

A(S (i) ) A(S)2

(A(S (i) ), z i )
(A(S), z i )
(A(S), z  )
(A(S (i) ), z  )

+
.
m
m
(13.10)

The two subsections that follow continue the stability analysis for either Lipschitz or smooth loss functions. For both families of loss functions we show that
RLM is stable and therefore it does not overfit.

13.3.1 Lipschitz Loss

If the loss function,
(, z i ), is -Lipschitz, then by the definition of Lipschitzness,

(A(S (i) ), z i )
(A(S), z i ) A(S (i) ) A(S).

(13.11)

Similarly,

(A(S), z  )
(A(S (i) ), z  ) A(S (i) ) A(S).
Plugging these inequalities into Equation (13.10) we obtain
A(S (i) ) A(S)2

2 A(S (i) ) A(S)

,
m

which yields
A(S (i) ) A(S)

2
.
m

Plugging the preceding back into Equation (13.11) we conclude that

(A(S (i) ), z i )
(A(S), z i )

2 2
.
m

Since this holds for any S, z  , i we immediately obtain:

Corollary 13.6. Assume that the loss function is convex and -Lipschitz. Then, the
2
RLM rule with the regularizer w2 is on-average-replace-one-stable with rate 2m .

13.3 Tikhonov Regularization as a Stabilizer

It follows (using Theorem 13.2) that

E m [L D (A(S)) L S (A(S))]

SD

2 2
.
m

13.3.2 Smooth and Nonnegative Loss

If the loss is -smooth and nonnegative then it is also self-bounded (see
Section 12.1):
 f (w)2 2 f (w).

(13.12)

We further assume that 2

m , or, in other words, that m/2. By the smoothness
assumption we have that

(A(S (i) ), z i )
(A(S), z i ) 
(A(S), z i ), A(S (i) ) A(S) +

A(S (i) ) A(S)2 .

2
(13.13)

Using the Cauchy-Schwartz inequality and Equation (12.6) we further obtain that

(A(S (i) ), z i )
(A(S), z i )

(A(S), z i ) A(S (i) ) A(S) +

A(S (i) ) A(S)2

2

2
(A(S), z i ) A(S (i) ) A(S) + A(S (i) ) A(S)2 .

2

(13.14)

By a symmetric argument it holds that

(A(S), z  )
(A(S (i) ), z  )


2
(A(S (i) ), z  ) A(S (i) ) A(S) + A(S (i) ) A(S)2 .

2
Plugging these inequalities into Equation (13.10) and rearranging terms we obtain
that




2
(i)


(A(S), z i ) +
(A(S ), z ) .
A(S ) A(S)
( m )
(i)

Combining the preceding with the assumption m/2 yields

A(S (i) ) A(S)




8 

(A(S), z i ) +
(A(S (i) ), z  ) .
m

143

144

Regularization and Stability

Combining the preceding with Equation (13.14) and again using the assumption
m/2 yield

(A(S (i) ), z i )
(A(S), z i )


2
(A(S), z i ) A(S (i) ) A(S) + A(S (i) ) A(S)2

2


2


8 2
4
(i) ), z  )
+

(A(S),
z
)
+

(A(S
i
m (m)2
2


8 
(i)


(A(S), z i ) +
(A(S ), z )
m

24 

(A(S), z i ) +
(A(S (i) ), z  ) ,

m
where in the last step we used the inequality (a + b)2 3(a 2 + b2 ). Taking expectation with respect to S, z  , i and noting that E [
(A(S), z i )] = E [
(A(S (i) ), z  )] =
E [L S (A(S))], we conclude that:
Corollary 13.7. Assume that the loss function is -smooth and nonnegative. Then, the
RLM rule with the regularizer w2 , where 2
m , satisfies

 48
E
(A(S (i) ), z i )
(A(S), z i )
E [L S (A(S))].
m
Note that if for all z we have
(0, z) C, for some scalar C > 0, then for every S,
L S (A(S)) L S (A(S)) + A(S)2 L S (0) + 02 = L S (0) C.
Hence, Corollary 13.7 also implies that
 48 C

.
E
(A(S (i) ), z i )
(A(S), z i )
m

13.4 CONTROLLING THE FITTING-STABILITY TRADEOFF

We can rewrite the expected risk of a learning algorithm as
E [L D (A(S))] = E [L S (A(S))] + E [L D (A(S)) L S (A(S))].
S

(13.15)

The first term reflects how well A(S) fits the training set while the second term
reflects the difference between the true and empirical risks of A(S). As we have
shown in Theorem 13.2, the second term is equivalent to the stability of A. Since
our goal is to minimize the risk of the algorithm, we need that the sum of both terms
will be small.
In the previous section we have bounded the stability term. We have shown
that the stability term decreases as the regularization parameter, , increases. On
the other hand, the empirical risk increases with . We therefore face a tradeoff
between fitting and overfitting. This tradeoff is quite similar to the bias-complexity
tradeoff we discussed previously in the book.

13.4 Controlling the Fitting-Stability Tradeoff

We now derive bounds on the empirical risk

rule. Recall that
 term for the RLM

the RLM rule is defined as A( S) = argminw L S (w) + w2 . Fix some arbitrary
vector w . We have
L S ( A( S)) L S ( A( S)) +  A( S)2 L S (w ) + w 2 .
Taking expectation of both sides with respect to S and noting that E S [ L S (w )] =
L D (w ), we obtain that
E [ L S ( A( S))] L D (w ) + w 2 .

(13.16)

Plugging this into Equation (13.15) we obtain

E [ L D ( A( S))] L D (w ) + w 2 + E [ L D ( A( S)) L S ( A( S))].
S

Combining the preceding with Corollary 13.6 we conclude:

Corollary 13.8. Assume that the loss function is convex and -Lipschitz. Then, the
RLM rule with the regularization function w2 satisfies
w , E [ L D ( A( S))] L D (w ) + w 2 +
S

2 2
.
m

This bound is often called an oracle inequality if we think of w as a hypothesis

with low risk, the bound tells us how many examples are needed so that A( S) will
be almost as good as w , had we known the norm of w . In practice, however, we
usually do not know the norm of w . We therefore usually tune on the basis of a
validation set, as described in Chapter 11.
We can also easily derive a PAC-like guarantee1 from Corollary 13.8 for convexLipschitz-bounded learning problems:
Corollary 13.9. Let (H, Z ,
) be a convex-Lipschitz-bounded
learning problem with


parameters , B. For any training set size m, let =

the regularization function w2 satisfies

2 2
.
B2 m

E [ L D ( A( S))] min L D (w) + B

S

wH

In particular, for every > 0, if m

E S [ L D ( A( S))] minwH L D (w) + .

8 2 B 2
2

Then, the RLM rule with

8
.
m

then for every distribution D,

The preceding corollary holds for Lipschitz loss functions. If instead the loss
function is smooth and nonnegative, then we can combine Equation (13.16) with
Corollary 13.7 to get:
Corollary 13.10. Assume that the loss function is convex, -smooth, and nonnegative.
Then, the RLM rule with the regularization function w2 , for 2
m , satisfies the
1

Again, the bound below is on the expected risk, but using Exercise 13.1 it can be used to derive an
agnostic PAC learning guarantee.

145

146

Regularization and Stability

following for all w :






48
48 
L D (w ) + w 2 .
E [L D (A(S))] 1 +
E [L S (A(S))] 1 +
S
m S
m
For example, if we choose = 48
m we obtain from the preceding that the
expected true risk of A(S) is at most twice the expected empirical risk of A(S).
Furthermore, for this value of , the expected empirical risk of A(S) is at most
2
L D (w ) + 48
m w  .
We can also derive a learnability guarantee for convex-smooth-bounded learning problems based on Corollary 13.10.
Corollary 13.11. Let (H, Z ,
) be a convex-smooth-bounded learning problem with
parameters , B. Assume in addition that
(0, z) 1 for all z Z . For any (0, 1)
B2
let m 150
and set = /(3B 2 ). Then, for every distribution D,
2
E [L D (A(S))] min L D (w) + .
S

wH

13.5 SUMMARY
We introduced stability and showed that if an algorithm is stable then it does not
overfit. Furthermore, for convex-Lipschitz-bounded or convex-smooth-bounded
problems, the RLM rule with Tikhonov regularization leads to a stable learning algorithm. We discussed how the regularization parameter, , controls the
tradeoff between fitting and overfitting. Finally, we have shown that all learning
problems that are from the families of convex-Lipschitz-bounded and convexsmooth-bounded problems are learnable using the RLM rule. The RLM paradigm
is the basis for many popular learning algorithms, including ridge regression (which
we discussed in this chapter) and support vector machines (which will be discussed
in Chapter 15).
In the next chapter we will present Stochastic Gradient Descent, which gives
us a very practical alternative way to learn convex-Lipschitz-bounded and convexsmooth-bounded problems and can also be used for efficiently implementing the
RLM rule.

13.6 BIBLIOGRAPHIC REMARKS

Stability is widely used in many mathematical contexts. For example, the necessity
of stability for so-called inverse problems to be well posed was first recognized by
Hadamard (1902). The idea of regularization and its relation to stability became
widely known through the works of Tikhonov (1943) and Phillips (1962). In the
context of modern learning theory, the use of stability can be traced back at least to
the work of Rogers and Wager (1978), which noted that the sensitivity of a learning
algorithm with regard to small changes in the sample controls the variance of the
leave-one-out estimate. The authors used this observation to obtain generalization
bounds for the k-nearest neighbor algorithm (see Chapter 19). These results were
later extended to other local learning algorithms (see Devroye, Gyrfi & Lugosi
(1996) and references therein). In addition, practical methods have been developed

13.7 Exercises

to introduce stability into learning algorithms, in particular the Bagging technique

introduced by (Breiman 1996).
Over the last decade, stability was studied as a generic condition for learnability.
See (Kearns & Ron 1999, Bousquet & Elisseeff 2002, Kutin & Niyogi 2002, Rakhlin,
Mukherjee & Poggio 2005, Mukherjee, Niyogi, Poggio & Rifkin 2006). Our presentation follows the work of Shalev-Shwartz, Shamir, Srebo, and Sridharan (2010),
who showed that stability is sufficient and necessary for learning. They have also
shown that all convex-Lipschitz-bounded learning problems are learnable using
RLM, even though for some convex-Lipschitz-bounded learning problems uniform
convergence does not hold in a strong sense.

13.7 EXERCISES
13.1 From Bounded Expected Risk to Agnostic PAC Learning: Let A be an algorithm
that guarantees the following: If m m H ( ) then for every distribution D it holds
that
E m [L D ( A(S))] min L D (h) + .
SD

hH

Show that for every (0, 1), if m m H ( ) then with probability of at least
1 it holds that L D ( A(S)) minhH L D (h) + .
Hint: Observe that the random variable L D ( A(S)) minhH L D (h) is nonnegative and rely on Markovs inequality.

For every (0, 1) let


log (4/) + log (log2 (1/))
m H ( , ) = m H ( /2)log2 (1/) +
.
2
Suggest a procedure that agnostic PAC learns the problem with sample complexity of m H ( , ), assuming that the loss function is bounded by 1.
Hint: Let k = log2 (1/). Divide the data into k + 1 chunks, where each of the
first k chunks is of size m H ( /2) examples. Train the first k chunks using A. On
the basis of the previous question argue that the probability that for all of these
chunks we have L D ( A(S)) > minhH L D (h) + is at most 2k /2. Finally, use
the last chunk as a validation set.
13.2 Learnability without Uniform Convergence: Let B be the unit ball of Rd , let H = B,
let Z = B {0, 1}d , and let
: Z H R be defined as follows:

(w, (x, )) =

d


i (xi wi )2 .

i=1

This problem corresponds to an unsupervised learning task, meaning that we do

not try to predict the label of x. Instead, what we try to do is to find the center of
mass of the distribution over B. However, there is a twist, modeled by the vectors
. Each example is a pair (x, ), where x is the instance x and indicates which
features of x are active and which are turned off. A hypothesis is a vector
w representing the center of mass of the distribution, and the loss function is the
squared Euclidean distance between x and w, but only with respect to the active
elements of x.

Show that this problem is learnable using the RLM rule with a sample
complexity that does not depend on d.

147

148

Regularization and Stability

Consider a distribution D over Z as follows: x is fixed to be some x0 , and each

element of is sampled to be either 1 or 0 with equal probability. Show that
the rate of uniform convergence of this problem grows with d.
Hint: Let m be a training set size. Show that if d % 2m , then there is a high
probability of sampling a set of examples such that there exists some j [d] for
which j = 1 for all the examples in the training set. Show that such a sample
cannot be -representative. Conclude that the sample complexity of uniform
convergence must grow with log (d).

Conclude that if we take d to infinity we obtain a problem that is learnable but
for which the uniform convergence property does not hold. Compare to the
fundamental theorem of statistical learning.
13.3 Stability and Asymptotic ERM Are Sufficient for Learnability:
We say that a learning rule A is an AERM (Asymptotic Empirical Risk Minimizer)
with rate (m) if for every distribution D it holds that


E m L S ( A(S)) min L S (h) (m).
SD

hH

We say that a learning rule A learns a class H with rate (m) if for every distribution

D it holds that



E m L D ( A(S)) min L D (h) (m).

SD

hH

Prove the following:

Theorem 13.12. If a learning algorithm A is on-average-replace-one-stable with rate
1 (m) and is an AERM with rate 2 (m), then it learns H with rate 1 (m) + 2 (m).
13.4 Strong Convexity with Respect to General Norms:
Throughout the section we used the
2 norm. In this exercise we generalize some
of the results to general norms. Let   be some arbitrary norm, and let f be a
strongly convex function with respect to this norm (see Definition 13.4).
1. Show that items 23 of Lemma 13.5 hold for every norm.
2. (*) Give an example of a norm for which item 1 of Lemma 13.5 does not hold.
3. Let R(w) be a function that is (2)-strongly convex with respect to some norm
 . Let A be an RLM rule with respect to R, namely,


A(S) = argmin L S (w) + R(w) .
w

Assume that for every z, the loss function
(, z) is -Lipschitz with respect to
the same norm, namely,
z, w, v,
(w, z)
(v, z) w v.
Prove that A is on-average-replace-one-stable with rate
4. (*) Let q (1, 2) and consider the
q -norm

wq =

d

i=1

1/q
| wi |

2 2
m .

13.7 Exercises

It can be shown (see, for example, Shalev-Shwartz (2007)) that the function
R(w) =

1
w2q
2(q 1)

(d)
is 1-strongly convex with respect to wq . Show that if q = loglog(d)1
then R(w) is


1
d
3log (d) -strongly convex with respect to the
1 norm over R .

149

14
Stochastic Gradient Descent

Recall that the goal of learning is to minimize the risk function, L D (h) =
EzD [
(h, z)]. We cannot directly minimize the risk function since it depends on
the unknown distribution D. So far in the book, we have discussed learning methods that depend on the empirical risk. That is, we first sample a training set S and
define the empirical risk function L S (h). Then, the learner picks a hypothesis based
on the value of L S (h). For example, the ERM rule tells us to pick the hypothesis
that minimizes L S (h) over the hypothesis class, H. Or, in the previous chapter, we
discussed regularized risk minimization, in which we pick a hypothesis that jointly
minimizes L S (h) and a regularization function over h.
In this chapter we describe and analyze a rather different learning approach,
which is called Stochastic Gradient Descent (SGD). As in Chapter 12 we will focus
on the important family of convex learning problems, and following the notation
in that chapter, we will refer to hypotheses as vectors w that come from a convex
hypothesis class, H. In SGD, we try to minimize the risk function L D (w) directly
using a gradient descent procedure. Gradient descent is an iterative optimization
procedure in which at each step we improve the solution by taking a step along the
negative of the gradient of the function to be minimized at the current point. Of
course, in our case, we are minimizing the risk function, and since we do not know
D we also do not know the gradient of L D (w). SGD circumvents this problem by
allowing the optimization procedure to take a step along a random direction, as
long as the expected value of the direction is the negative of the gradient. And, as
we shall see, finding a random direction whose expected value corresponds to the
gradient is rather simple even though we do not know the underlying distribution D.
The advantage of SGD, in the context of convex learning problems, over the
regularized risk minimization learning rule is that SGD is an efficient algorithm that
can be implemented in a few lines of code, yet still enjoys the same sample complexity as the regularized risk minimization rule. The simplicity of SGD also allows us
to use it in situations when it is not possible to apply methods that are based on the
empirical risk, but this is beyond the scope of this book.
We start this chapter with the basic gradient descent algorithm and analyze its
convergence rate for convex-Lipschitz functions. Next, we introduce the notion of
150

14.1 Gradient Descent

subgradient and show that gradient descent can be applied for nondifferentiable
functions as well. The core of this chapter is Section 14.3, in which we describe
the Stochastic Gradient Descent algorithm, along with several useful variants. We
show that SGD enjoys an expected convergence rate similar to the rate of gradient
descent. Finally, we turn to the applicability of SGD to learning problems.

14.1 GRADIENT DESCENT

Before we describe the stochastic gradient descent method, we would like to
describe the standard gradient descent approach for minimizing a differentiable
convex function f (w).
The gradient of a differentiable function f : Rd R at w, denoted f (w), is
f (w)
. Gradient
the vector of partial derivatives of f , namely, f (w) = w
, . . . , wf (w)
[1]
[d]
descent is an iterative algorithm. We start with an initial value of w (say, w(1) = 0).
Then, at each iteration, we take a step in the direction of the negative of the gradient
at the current point. That is, the update step is
w(t+1) = w(t) f (w(t) ),

(14.1)

where > 0 is a parameter to be discussed later. Intuitively, since the gradient

points in the direction of the greatest rate of increase of f around w(t) , the algorithm makes a small step in the opposite direction, thus decreasing the value of the
function. Eventually, after T iterations, the algorithm outputs the averaged vector,

T
= T1 t=1
w
w(t) . The output could also be the last vector, w(T ) , or the best performing vector, argmint[T ] f (w(t) ), but taking the average turns out to be rather useful,
especially when we generalize gradient descent to nondifferentiable functions and
to the stochastic case.
Another way to motivate gradient descent is by relying on Taylor approximation. The gradient of f at w yields the first order Taylor approximation of f around
w by f (u) f (w) + u w, f (w). When f is convex, this approximation lower
bounds f , that is,
f (u) f (w) + u w, f (w).
Therefore, for w close to w(t) we have that f (w) f (w(t) ) + w w(t) , f (w(t) ).
Hence we can minimize the approximation of f (w). However, the approximation
might become loose for w, which is far away from w(t) . Therefore, we would like to
minimize jointly the distance between w and w(t) and the approximation of f around
w(t) . If the parameter controls the tradeoff between the two terms, we obtain the
update rule


1
w(t+1) = argmin w w(t) 2 + f (w(t) ) + w w(t) , f (w(t) ) .
2
w
Solving the preceding by taking the derivative with respect to w and comparing it to
zero yields the same update rule as in Equation (14.1).

151

152

Stochastic Gradient Descent

Figure 14.1. An illustration of the gradient descent algorithm. The function to be

minimized is 1. 25(x1 + 6)2 + (x2 8)2 .

14.1.1 Analysis of GD for Convex-Lipschitz Functions

To analyze the convergence rate of the GD algorithm, we limit ourselves to the case
of convex-Lipschitz functions (as we have seen, many problems lend themselves
easily to this setting). Let w be any vector and let B be an upper bound on w . It
is convenient to think of w as the minimizer of f (w), but the analysis that follows
holds for every w .
We would like to obtain an upper bound on the suboptimality of our solu
T
f (w ), where w
= T1 t=1
tion with respect to w , namely, f (w)
w(t) . From the
and using Jensens inequality, we have that
definition of w,
 T


f (w ) = f T1
w(t) f (w )
f (w)
t=1

T

1 
f (w(t) ) f (w )
T
t=1

T

1 
=
f (w(t) ) f (w ) .
T

(14.2)

t=1

For every t, because of the convexity of f , we have that

f (w(t) ) f (w ) w(t) w , f (w(t) ).

(14.3)

Combining the preceding we obtain

T
1  (t)
f (w )
f (w)
w w , f (w(t) ).
T


t=1

To bound the right-hand side we rely on the following lemma:

Lemma 14.1. Let v1 , . . . , vT be an arbitrary sequence of vectors. Any algorithm with
an initialization w(1) = 0 and an update rule of the form
w(t+1) = w(t) vt

(14.4)

14.1 Gradient Descent

satisfies
T


w 2 
vt 2 .
+
2
2
T

w(t) w , vt 

t=1

(14.5)

t=1

In particular, for every B, > 0, if for all t we have that vt  and if we set =

B2
, then for every w with w  B we have
2 T
T
B
1  (t)
w w , vt  .
T
T
t=1

Proof. Using algebraic manipulations (completing the square), we obtain:

1
w(t) w , vt  = w(t) w , vt 

1
= ( w(t) w vt 2 + w(t) w 2 + 2 vt 2 )
2
=

( w(t+1) w 2 + w(t) w 2 ) + vt 2 ,

2
2

where the last equality follows from the definition of the update rule. Summing the
equality over t, we have
T

t=1

w

(t)

 
1 
w(t+1) w 2 + w(t) w 2 +
w , vt  =
vt 2 . (14.6)
2
2


t=1

t=1

The first sum on the right-hand side is a telescopic sum that collapses to
w(1) w 2 w(T +1) w 2 .
Plugging this in Equation (14.6), we have
T


w(t) w , vt  =

t=1

1

(w(1) w 2 w(T +1) w 2 ) +
vt 2
2
2
T

t=1

1

w(1) w 2 +
vt 2
2
2
T

t=1

w 2 +
2
2

T


vt 2 ,

t=1

where the last equality is due to the definition w(1) = 0. This proves the first part of
the lemma (Equation (14.5)). The second part follows by upper bounding w  by
B, vt  by , dividing by T , and plugging in the value of .
Lemma 14.1 applies to the GD algorithm with vt = f (w(t) ). As we will show
later in Lemma 14.7, if f is -Lipschitz, then  f (w(t) ) . We therefore satisfy

153

154

Stochastic Gradient Descent

the lemmas conditions and achieve the following corollary:

and let w
2
argmin{w:wB} f (w). If we run the GD algorithm on f for T steps with = B2 T ,
satisfies
then the output vector w

Corollary 14.2. Let f be a convex,

-Lipschitz function,

B
f (w ) .
f (w)
T
f (w ) , it suffices to run the GD
Furthermore, for every > 0, to achieve f (w)
algorithm for a number of iterations that satisfies
T

B 2 2
.
2

14.2 SUBGRADIENTS
The GD algorithm requires that the function f be differentiable. We now generalize
the discussion beyond differentiable functions. We will show that the GD algorithm
can be applied to nondifferentiable functions by using a so-called subgradient of
f (w) at w(t) , instead of the gradient.
To motivate the definition of subgradients, recall that for a convex function f ,
the gradient at w defines the slope of a tangent that lies below f , that is,
u,

f (u) f (w) + u w, f (w).

(14.7)

An illustration is given on the left-hand side of Figure 14.2.

The existence of a tangent that lies below f is an important property of convex
functions, which is in fact an alternative characterization of convexity.
Lemma 14.3. Let S be an open convex set. A function f : S R is convex iff for
every w S there exists v such that
u S,

f (u) f (w) + u w, v.

(14.8)

The proof of this lemma can be found in many convex analysis textbooks (e.g.,
(Borwein & Lewis 2006)). The preceding inequality leads us to the definition of
subgradients.
Definition 14.4. (Subgradients). A vector v that satisfies Equation (14.8) is called a
subgradient of f at w. The set of subgradients of f at w is called the differential set
and denoted f (w).
An illustration of subgradients is given on the right-hand side of Figure 14.2. For
scalar functions, a subgradient of a convex function f at w is a slope of a line that
touches f at w and is not above f elsewhere.

14.2.1 Calculating Subgradients

How do we construct subgradients of a given convex function? If a function is differentiable at a point w, then the differential set is trivial, as the following claim
shows.

,
f(
w
)

14.2 Subgradients

)+

f (u)

f(

f (w)
w

Figure 14.2. Left: The right-hand side of Equation (14.7) is the tangent of f at w. For a
convex function, the tangent lower bounds f. Right: Illustration of several subgradients
of a nondifferentiable convex function.

Claim 14.5. If f is differentiable at w then f (w) contains a single element the

gradient of f at w, f (w).
Example 14.1 (The Differential Set of the Absolute Function). Consider the absolute value function f (x) = |x|. Using Claim 14.5, we can easily construct the
differential set for the differentiable parts of f , and the only point that requires
special attention is x 0 = 0. At that point, it is easy to verify that the subdifferential is
the set of all numbers between 1 and 1. Hence:

if x > 0
{1}
f (x) = {1}
if x < 0

[ 1, 1] if x = 0
For many practical uses, we do not need to calculate the whole set of subgradients at a given point, as one member of this set would suffice. The following claim
shows how to construct a sub-gradient for pointwise maximum functions.
Claim 14.6. Let g(w) = maxi[r] gi (w) for r convex differentiable functions g1 , . . . , gr .
Given some w, let j argmaxi gi (w). Then g j (w) g(w).
Proof. Since g j is convex we have that for all u
g j (u) g j (w) + u w, g j (w).
Since g(w) = g j (w) and g(u) g j (u) we obtain that
g(u) g(w) + u w, g j (w),
which concludes our proof.
Example 14.2 (A Subgradient of the Hinge Loss). Recall the hinge loss function
from Section 12.3, f (w) = max{0, 1 yw, x} for some vector x and scalar y. To
calculate a subgradient of the hinge loss at some w we rely on the preceding claim
and obtain that the vector v defined in the following is a subgradient of the hinge
loss at w:

0
if 1 yw, x 0
v=
yx if 1 yw, x > 0

155

156

Stochastic Gradient Descent

14.2.2 Subgradients of Lipschitz Functions

Recall that a function f : A R is -Lipschitz if for all u, v A
| f (u) f (v)| u v.
The following lemma gives an equivalent definition using norms of subgradients.
Lemma 14.7. Let A be a convex open set and let f : A R be a convex function.
Then, f is -Lipschitz over A iff for all w A and v f (w) we have that v .
Proof. Assume that for all v f (w) we have that v . Since v f (w) we have
f (w) f (u) v, w u.
Bounding the right-hand side using Cauchy-Schwartz inequality we obtain
f (w) f (u) v, w u v w u w u.
An analogous argument can show that f (u) f (w) w u. Hence f is Lipschitz.
Now assume that f is -Lipschitz. Choose some w A, v f (w). Since A is
open, there exists > 0 such that u = w+ v/v belongs to A. Therefore, uw, v =
v and u w = . From the definition of the subgradient,
f (u) f (w) v, u w = v.
On the other hand, from the Lipschitzness of f we have
= u w f (u) f (w).
Combining the two inequalities we conclude that v .

14.2.3 Subgradient Descent

The gradient descent algorithm can be generalized to nondifferentiable functions
by using a subgradient of f (w) at w(t) , instead of the gradient. The analysis of the
convergence rate remains unchanged: Simply note that Equation (14.3) is true for
subgradients as well.

14.3 STOCHASTIC GRADIENT DESCENT (SGD)

In stochastic gradient descent we do not require the update direction to be based
exactly on the gradient. Instead, we allow the direction to be a random vector and
only require that its expected value at each iteration will equal the gradient direction.
Or, more generally, we require that the expected value of the random vector will be
a subgradient of the function at the current vector.

14.3 Stochastic Gradient Descent (SGD)

Figure 14.3. An illustration of the gradient descent algorithm (left) and the stochastic
gradient descent algorithm (right). The function to be minimized is 1. 25(x + 6)2 + (y 8)2 .
For the stochastic case, the solid line depicts the averaged value of w.

Stochastic Gradient Descent (SGD) for minimizing f (w)

parameters: Scalar > 0, integer T > 0
initialize: w(1) = 0
for t = 1, 2, . . . , T
choose vt at random from a distribution such that E [vt | w(t) ] f (w(t) )
update w(t+1) = w(t) vt

T
= T1 t=1
output w
w(t)

An illustration of stochastic gradient descent versus gradient descent is given

in Figure 14.3. As we will see in Section 14.5, in the context of learning problems,
it is easy to find a random vector whose expectation is a subgradient of the risk
function.

14.3.1 Analysis of SGD for Convex-Lipschitz-Bounded Functions

Recall the bound we achieved for the GD algorithm in Corollary 14.2. For the
stochastic case, in which only the expectation of vt is in f (w(t) ), we cannot directly
apply Equation (14.3). However, since the expected value of vt is a subgradient of
f at w(t) , we can still derive a similar bound on the expected output of stochastic
gradient descent. This is formalized in the following theorem.
Theorem 14.8. Let B, > 0. Let f be a convex function and let w
B2
argminw:wB f (w). Assume that SGD is run for T iterations with =
.
2 T
Assume also that for all t, vt  with probability 1. Then,
B
f (w ) .
E [ f (w)]
T

157

158

Stochastic Gradient Descent

f (w ) , it suffices to run the SGD

Therefore, for any > 0, to achieve E [ f (w)]
algorithm for a number of iterations that satisfies
T

B 2 2
.
2

Proof. Let us introduce the notation v1:t to denote the sequence v1 , . . . , vt . Taking
expectation of Equation (14.2), we obtain

 T

f (w )] E T1
( f (w(t) ) f (w )) .
E [ f (w)
v1:T

v1:T

t=1

Since Lemma 14.1 holds for any sequence v1 , v2 , . . . vT , it applies to SGD as well. By
taking expectation of the bound in the lemma we have


T
B
1  (t)

(14.9)
w w , vt  .
E
v1:T T
T
t=1

It is left to show that



 T
 T


(t)

(t)

1
1
E T
( f (w ) f (w )) E T
w w , vt  ,
v1:T

v1:T

t=1

(14.10)

t=1

which we will hereby prove.

Using the linearity of the expectation we have


T
T
1
1  (t)

w w , vt  =
E [w(t) w , vt ].
E
v1:T T
v1:T
T
t=1

t=1

Next, we recall the law of total expectation: For every two random variables , ,
and a function g, E [g()] = E E [g()|]. Setting = v1:t and = v1:t1 we get
that
E [w(t) w , vt ] = E [w(t) w , vt ]

v1:T

v1: t

= E E [w(t) w , vt  | v1: t1 ].
v1: t1 v1: t

Once we know v1:t1 , the value of w(t) is not random any more and therefore
E E [w(t) w , vt  | v1: t1 ] = E w(t) w , E [vt | v1: t1].

v1: t1 v1: t

v1: t1

vt

Since w(t) only depends on v1:t1 and SGD requires that Evt [vt | w(t) ] f (w(t) ) we
obtain that Evt [vt | v1:t1 ] f (w(t) ). Thus,
E w(t) w , E [vt | v1: t1 ]

v1: t1

vt

E [ f (w(t) ) f (w )].

v1: t1

Overall, we have shown that

E [w(t) w , vt ] E [ f (w(t) ) f (w )]

v1:T

v1: t1

= E [ f (w(t) ) f (w )].
v1:T

14.4 Variants

Summing over t, dividing by T , and using the linearity of expectation, we get that
Equation (14.10) holds, which concludes our proof.

14.4 VARIANTS
In this section we describe several variants of Stochastic Gradient Descent.

14.4.1 Adding a Projection Step

In the previous analyses of the GD and SGD algorithms, we required that the norm
of w will be at most B, which is equivalent to requiring that w is in the set H =
{w : w B}. In terms of learning, this means restricting ourselves to a B-bounded
hypothesis class. Yet any step we take in the opposite direction of the gradient (or
its expected direction) might result in stepping out of this bound, and there is even
satisfies it. We show in the following how to overcome this
no guarantee that w
problem while maintaining the same convergence rate.
The basic idea is to add a projection step; namely, we will now have a two-step
update rule, where we first subtract a subgradient from the current value of w and
then project the resulting vector onto H. Formally,
1

1. w(t+ 2 ) = w(t) vt
1
2. w(t+1) = argminwH w w(t+ 2 ) 
The projection step replaces the current value of w by the vector in H closest to it.
Clearly, the projection step guarantees that w(t) H for all t. Since H is convex
H as required. We next show that the analysis of SGD with
this also implies that w
projections remains the same. This is based on the following lemma.
Lemma 14.9 (Projection Lemma). Let H be a closed convex set and let v be the
projection of w onto H, namely,
v = argmin x w2 .
xH

Then, for every u H,

w u2 v u2 0.
Proof. By the convexity of H, for every (0, 1) we have that v + (u v) H.
Therefore, from the optimality of v we obtain
v w2 v + (u v) w2
= v w2 + 2v w, u v + 2 u v2 .
Rearranging, we obtain
2v w, u v u v2 .
Taking the limit 0 we get that
v w, u v 0.

159

160

Stochastic Gradient Descent

Therefore,
w u2 = w v + v u2
= w v2 + v u2 + 2v w, u v
v u2 .

Equipped with the preceding lemma, we can easily adapt the analysis of SGD to
the case in which we add projection steps on a closed and convex set. Simply note
that for every t,
w(t+1) w 2 w(t) w 2
1

= w(t+1) w 2 w(t+ 2 ) w 2 + w(t+ 2 ) w 2 w(t) w 2

1

w(t+ 2 ) w 2 w(t) w 2 .
Therefore, Lemma 14.1 holds when we add projection steps and hence the rest of
the analysis follows directly.

14.4.2 Variable Step Size

Another variant of SGD is decreasing the step size as a function of t. That is, rather
than updating with a constant , we use t . For instance, we can set t = Bt and
achieve a bound similar to Theorem 14.8. The idea is that when we are closer to the
minimum of the function, we take our steps more carefully, so as not to overshoot
the minimum.

14.4.3 Other Averaging Techniques

T
= T1 t=1
We have set the output vector to be w
w(t) . There are alternative
(t)
approaches such as outputting w for some random t [t], or outputting the average of w(t) over the last T iterations, for some (0, 1). One can also take
a weighted average of the last few iterates. These more sophisticated averaging
schemes can improve the convergence speed in some situations, such as in the case
of strongly convex functions defined in the following.

14.4.4 Strongly Convex Functions*

In this section we show a variant of SGD that enjoys a faster convergence rate for
problems in which the objective function is strongly convex (see Definition 13.4 of
strong convexity in the previous chapter). We rely on the following claim, which
generalizes Lemma 13.5.
Claim 14.10. If f is -strongly convex then for every w, u and v f (w) we have
w u, v f (w) f (u) + 2 w u2 .
The proof is similar to the proof of Lemma 13.5 and is left as an exercise.

14.4 Variants

SGD for minimizing a -strongly convex function

Goal: Solve minwH f (w)
parameter: T
initialize: w(1) = 0
for t = 1, . . . , T
Choose a random vector vt s.t. E [vt |w(t) ] f (w(t) )
Set t = 1/( t)
1

Set w(t+ 2 ) = w(t) t vt

1

Set w(t+1) = arg minwH w w(t+ 2 ) 2

T
= T1 t=1
output: w
w(t)
Theorem 14.11. Assume that f is -strongly convex and that E [vt 2 ] 2 . Let w
argminwH f (w) be an optimal solution. Then,
f (w )
E [ f (w)]

2
(1 + log (T )).
2T

Proof. Let (t) = E [vt |w(t) ]. Since f is strongly convex and (t) is in the subgradient
set of f at w(t) we have that
w(t) w , (t)  f (w(t) ) f (w ) + 2 w(t) w 2 .

(14.11)

Next, we show that

w(t) w , (t) 

E [w(t) w 2 w(t+1) w 2 ] t 2
+ .
2 t
2

(14.12)

Since w(t+1) is the projection of w(t+ 2 ) onto H, and w H we have that w(t+ 2 )
w 2 w(t+1) w 2 . Therefore,
1

w(t) w 2 w(t+1) w 2 w(t) w 2 w(t+ 2 ) w 2

= 2t w(t) w , vt  t2 vt 2 .
Taking expectation of both sides, rearranging, and using the assumption E [vt 2 ]
2 yield Equation (14.12). Comparing Equation (14.11) and Equation (14.12) and
summing over t we obtain
T


( E [ f (w(t) )] f (w ))

t=1

 T 
 w(t) w 2 w(t+1) w 2
t=1

2 t


2 w(t) w 2

2 
t .
2
T

t=1

Next, we use the definition t = 1/( t) and note that the first sum on the right-hand
side of the equation collapses to T w(T +1) w 2 0. Thus,
T

t=1

( E [ f (w(t) )] f (w ))

2  1 2

(1 + log (T )).
2
t
2
T

t=1

161

162

Stochastic Gradient Descent

The theorem follows from the preceding by dividing by T and using Jensens
inequality.
Remark 14.3. Rakhlin, Shamir, and Sridharan ((2012)) derived a convergence rate
in which the log (T ) term is eliminated for a variant
in which we

T of the algorithm
(t) . Shamir and Zhang
= T2 t=T
output the average of the last T /2 iterates, w
w
/2+1
= w( T ) .
(2013) have shown that Theorem 14.11 holds even if we output w

14.5 LEARNING WITH SGD

We have so far introduced and analyzed the SGD algorithm for general convex
functions. Now we shall consider its applicability to learning tasks.

14.5.1 SGD for Risk Minimization

Recall that in learning we face the problem of minimizing the risk function
L D (w) = E [
(w, z)].
zD

We have seen the method of empirical risk minimization, where we minimize the
empirical risk, L S (w), as an estimate to minimizing L D (w). SGD allows us to take
a different approach and minimize L D (w) directly. Since we do not know D, we
cannot simply calculate L D (w(t) ) and minimize it with the GD method. With SGD,
however, all we need is to find an unbiased estimate of the gradient of L D (w), that
is, a random vector whose conditional expected value is L D (w(t) ). We shall now
see how such an estimate can be easily constructed.
For simplicity, let us first consider the case of differentiable loss functions. Hence
the risk function L D is also differentiable. The construction of the random vector vt
will be as follows: First, sample z D. Then, define vt to be the gradient of the
function
(w, z) with respect to w, at the point w(t) . Then, by the linearity of the
gradient we have
E [vt |w(t) ] = E [
(w(t) , z)] = E [
(w(t) , z)] = L D (w(t) ).
zD

zD

(14.13)

The gradient of the loss function
(w, z) at w(t) is therefore an unbiased estimate of
the gradient of the risk function L D (w(t) ) and is easily constructed by sampling a
single fresh example z D at each iteration t.
The same argument holds for nondifferentiable loss functions. We simply let vt
be a subgradient of
(w, z) at w(t) . Then, for every u we have

(u, z)
(w(t) , z) u w(t) , vt .
Taking expectation on both sides with respect to z D and conditioned on the value
of w(t) we obtain
L D (u) L D (w(t) ) = E [
(u, z)
(w(t) , z)|w(t) ]
E [u w(t) , vt |w(t) ]
= u w(t) , E [vt |w(t) ].

14.5 Learning with SGD

It follows that E [vt |w(t) ] is a subgradient of L D (w) at w(t) .

To summarize, the stochastic gradient descent framework for minimizing the
risk is as follows.
Stochastic Gradient Descent (SGD) for minimizing L D (w)
parameters: Scalar > 0, integer T > 0
initialize: w(1) = 0
for t = 1, 2, . . . , T
sample z D
pick vt
(w(t) , z)
update w(t+1) = w(t) vt

T
= T1 t=1
output w
w(t)
We shall now use our analysis of SGD to obtain a sample complexity analysis for
learning convex-Lipschitz-bounded problems. Theorem 14.8 yields the following:
Corollary 14.12. Consider a convex-Lipschitz-bounded learning problem with
parameters , B. Then, for every > 0, if we run the SGD method for minimizing
L D (w) with a number of iterations (i.e., number of examples)
T
and with =

B2
2 T

B 2 2
2

, then the output of SGD satisfies

min L D (w) + .
E [L D (w)]
wH

It is interesting to note that the required sample complexity is of the same order
of magnitude as the sample complexity guarantee we derived for regularized loss
minimization. In fact, the sample complexity of SGD is even better than what we
have derived for regularized loss minimization by a factor of 8.

14.5.2 Analyzing SGD for Convex-Smooth Learning Problems

In the previous chapter we saw that the regularized loss minimization rule also
learns the class of convex-smooth-bounded learning problems. We now show that
the SGD algorithm can be also used for such problems.
Theorem 14.13. Assume that for all z, the loss function
(, z) is convex, -smooth,
and nonnegative. Then, if we run the SGD algorithm for minimizing L D (w) we have
that for every w ,


1
w 2


E [L D (w)]
L D (w ) +
.
1
2 T
Proof. Recall that if a function is -smooth and nonnegative then it is self-bounded:
 f (w)2 2 f (w).

163

164

Stochastic Gradient Descent

To analyze SGD for convex-smooth problems, let us define z 1 , . . . , z T the random

samples of the SGD algorithm, let f t ( ) =
(, z t ), and note that vt = f t (w(t) ).
For all t, f t is a convex function and therefore f t (w(t) ) f t (w ) vt , w(t) w .
Summing over t and using Lemma 14.1 we obtain
T


( f t (w(t) ) f t (w ))

t=1

T


vt , w(t) w 

t=1

w 2 
+
vt 2 .
2
2
T

t=1

Combining the preceding with the self-boundedness of ft yields

T


( ft (w(t) ) f t (w ))

t=1


w 2
+
f t (w(t) ).
2
T

t=1

Dividing by T and rearranging, we obtain



T
T
1
1
w 2
1
(t)

f t (w )
ft (w ) +
.
T
1 T
2 T
t=1

t=1

Next, we take expectation of the two sides of the preceding equation with respect to
z 1 , . . . , z T . Clearly, E [ f t (w )] = L D (w ). In addition, using the same argument as in
the proof of Theorem 14.8 we have that




T
T

1
1

E
f t (w(t) ) = E
L D (w(t) ) E [L D (w)].
T
T
t=1

t=1

Combining all we conclude our proof.

As a direct corollary we obtain:
Corollary 14.14. Consider a convex-smooth-bounded learning problem with parameters , B. Assume in addition that
(0, z) 1 for all z Z . For every > 0, set
1
= (1+3/ )
. Then, running SGD with T 12B 2/ 2 yields
min L D (w) + .
E [L D (w)]
wH

14.5.3 SGD for Regularized Loss Minimization

We have shown that SGD enjoys the same worst-case sample complexity bound
as regularized loss minimization. However, on some distributions, regularized loss
minimization may yield a better solution. Therefore, in some cases we may want
to solve the optimization problem associated with regularized loss minimization,
namely,1



w2 + L S (w) .
min
(14.14)
w
2
Since we are dealing with convex learning problems in which the loss function is
convex, the preceding problem is also a convex optimization problem that can be
solved using SGD as well, as we shall see in this section.
1

We divided by 2 for convenience.

14.6 Summary

Define f (w) = 2 w2 + L S (w). Note that f is a -strongly convex function;

therefore, we can apply the SGD variant given in Section 14.4.4 (with H = Rd ).
To apply this algorithm, we only need to find a way to construct an unbiased estimate of a subgradient of f at w(t) . This is easily done by noting that if we pick z
uniformly at random from S, and choose vt
(w(t) , z) then the expected value of
w(t) + vt is a subgradient of f at w(t) .
To analyze the resulting algorithm, we first rewrite the update rule (assuming
that H = Rd and therefore the projection step does not matter) as follows

1  (t)
w + vt
w(t+1) = w(t)
t


1
1
= 1
w(t) vt
t
t
t 1 (t)
1
w vt
t
t


t 1 t 2 (t1)
1
1
w
vt1 vt
=

t
t 1
(t 1)
t
=

t
1 
vi .
t

(14.15)

i=1

If we assume that the loss function is -Lipschitz, it follows that for all t we have
vt  and therefore w(t)  , which yields
w(t) + vt  2.
Theorem 14.11 therefore tells us that after performing T iterations we have that
f (w )
E [ f (w)]

4 2
(1 + log (T )).
T

14.6 SUMMARY
We have introduced the Gradient Descent and Stochastic Gradient Descent algorithms, along with several of their variants. We have analyzed their convergence rate
and calculated the number of iterations that would guarantee an expected objective
of at most plus the optimal objective. Most importantly, we have shown that by
using SGD we can directly minimize the risk function. We do so by sampling a
point i.i.d from D and using a subgradient of the loss of the current hypothesis w(t)
at this point as an unbiased estimate of the gradient (or a subgradient) of the risk
function. This implies that a bound on the number of iterations also yields a sample complexity bound. Finally, we have also shown how to apply the SGD method
to the problem of regularized risk minimization. In future chapters we show how
this yields extremely simple solvers to some optimization problems associated with
regularized risk minimization.

165

166

Stochastic Gradient Descent

14.7 BIBLIOGRAPHIC REMARKS

SGD dates back to Robbins and Monro (1951). It is especially effective in large scale
machine learning problems. See, for example, (Murata 1998, Le Cun 2004, Zhang
2004, Bottou & Bousquet 2008, Shalev-Shwartz, Singer & Srebro 2007, ShalevShwartz & Srebro 2008). In the optimization community it was studied in the context
of stochastic optimization. See, for example, (Nemirovski & Yudin 1978, Nesterov &
Nesterov 2004, Nesterov 2005, Nemirovski, Juditsky, Lan & Shapiro 2009, Shapiro,

Dentcheva & Ruszczynski

2009).
The bound we have derived for strongly convex function is due to Hazan,
Agarwal, and Kale (2007). As mentioned previously, improved bounds have been
obtained in Rakhlin, Shamir & Sridharan (2012).

14.8 EXERCISES
14.1 Prove Claim 14.10. Hint: Extend the proof of Lemma 13.5.
14.2 Prove Corollary 14.14.
14.3 Perceptron as a subgradient descent algorithm: Let S = ((x1 , y1 ), . . ., (xm , ym ))
(Rd {1})m . Assume that there exists w Rd such that for every i [m] we have
yi w, xi  1, and let w be a vector that has the minimal norm among all vectors
that satisfy the preceding requirement. Let R = maxi xi . Define a function
f (w) = max (1 yi w, xi ) .
i[m]

Show that minw:ww  f (w) = 0 and show that any w for which f (w) < 1
separates the examples in S.

Show how to calculate a subgradient of f .

Describe and analyze the subgradient descent algorithm for this case. Compare the algorithm and the analysis to the Batch Perceptron algorithm given in
Section 9.1.2.
14.4 Variable step size (*): Prove an analog of Theorem 14.8 for SGD with a variable
step size, t = Bt .

15
Support Vector Machines

In this chapter and the next we discuss a very useful machine learning tool: the
support vector machine paradigm (SVM) for learning linear predictors in high
dimensional feature spaces. The high dimensionality of the feature space raises both
sample complexity and computational complexity challenges.
The SVM algorithmic paradigm tackles the sample complexity challenge by
searching for large margin separators. Roughly speaking, a halfspace separates
a training set with a large margin if all the examples are not only on the correct
side of the separating hyperplane but also far away from it. Restricting the algorithm to output a large margin separator can yield a small sample complexity even
if the dimensionality of the feature space is high (and even infinite). We introduce
the concept of margin and relate it to the regularized loss minimization paradigm as
well as to the convergence rate of the Perceptron algorithm.
In the next chapter we will tackle the computational complexity challenge using
the idea of kernels.

15.1 MARGIN AND HARD-SVM

Let S = (x1 , y1 ), . . . , (xm , ym ) be a training set of examples, where each xi Rd and
yi {1}. We say that this training set is linearly separable, if there exists a halfspace,
(w, b), such that yi = sign(w, xi  + b) for all i . Alternatively, this condition can be
rewritten as
i [m], yi (w, xi  + b) > 0.
All halfspaces (w, b) that satisfy this condition are ERM hypotheses (their 0-1
error is zero, which is the minimum possible error). For any separable training
sample, there are many ERM halfspaces. Which one of them should the learner
pick?
Consider, for example, the training set described in the picture that
follows.
167

168

Support Vector Machines

While both the dashed and solid hyperplanes separate the four examples, our intuition would probably lead us to prefer the dashed hyperplane over the solid one.
One way to formalize this intuition is using the concept of margin.
The margin of a hyperplane with respect to a training set is defined to be the
minimal distance between a point in the training set and the hyperplane. If a hyperplane has a large margin, then it will still separate the training set even if we slightly
perturb each instance.
We will see later on that the true error of a halfspace can be bounded in terms
of the margin it has over the training sample (the larger the margin, the smaller the
error), regardless of the Euclidean dimension in which this halfspace resides.
Hard-SVM is the learning rule in which we return an ERM hyperplane that
separates the training set with the largest possible margin. To define Hard-SVM
formally, we first express the distance between a point x to a hyperplane using the
parameters defining the halfspace.
Claim 15.1. The distance between a point x and the hyperplane defined by (w, b)
where w = 1 is |w, x + b|.
Proof. The distance between a point x and the hyperplane is defined as
min{x v : w, v + b = 0}.
Taking v = x (w, x + b)w we have that
w, v + b = w, x (w, x + b)w2 + b = 0,
and
x v = |w, x + b| w = |w, x + b|.
Hence, the distance is at most |w, x + b|. Next, take any other point u on the
hyperplane, thus w, u + b = 0. We have
x u2 = x v + v u2
= x v2 + v u2 + 2x v, v u
x v2 + 2x v, v u
= x v2 + 2(w, x + b)w, v u
= x v2 ,
where the last equality is because w, v = w, u = b. Hence, the distance between
x and u is at least the distance between x and v, which concludes our proof.

15.1 Margin and Hard-SVM

On the basis of the preceding claim, the closest point in the training set to the
separating hyperplane is mini[m] |w, xi  + b|. Therefore, the Hard-SVM rule is
argmax min |w, xi  + b| s.t. i , yi (w, xi  + b) > 0.

(w,b):w=1 i[m]

Whenever there is a solution to the preceding problem (i.e., we are in the separable
case), we can write an equivalent problem as follows (see Exercise 15.1):
argmax min yi (w, xi  + b).

(15.1)

(w,b):w=1 i[m]

Next, we give another equivalent formulation of the Hard-SVM rule as a quadratic

optimization problem.1
Hard-SVM
input: (x1 , y1 ), . . . , (xm , ym )
solve:
(w0 , b0 ) = argmin w2 s.t. i , yi (w, xi  + b) 1

(15.2)

(w,b)

=
output: w

w0
w0  ,

b =

b0
w0 

The lemma that follows shows that the output of hard-SVM is indeed the separating hyperplane with the largest margin. Intuitively, hard-SVM searches for
w of minimal norm among all the vectors that separate the data and for which
|w, xi  + b| 1 for all i . In other words, we enforce the margin to be 1, but now
the units in which we measure the margin scale with the norm of w. Therefore, finding the largest margin halfspace boils down to finding w whose norm is minimal.
Formally:
Lemma 15.2. The output of Hard-SVM is a solution of Equation (15.1).
Proof. Let (w , b  ) be a solution of Equation (15.1) and define the margin achieved
by (w , b  ) to be  = mini[m] yi (w , xi  + b ). Therefore, for all i we have
yi (w , xi  + b ) 
or equivalently

yi ( w  , xi  + b  ) 1.


b
Hence, the pair ( w
 ,  ) satisfies the conditions of the quadratic optimization prob

1
lem given in Equation (15.2). Therefore, w0   w
  =  . It follows that for
all i ,
= 1 yi (w0 , xi  + b0 ) 1  .
xi  + b)
yi (w,
w0 
w0 

is an optimal solution of Equation (15.1).

= 1 we obtain that (w,
b)
Since w
1

A quadratic optimization problem is an optimization problem in which the objective is a convex

quadratic function and the constraints are linear inequalities.

169

170

Support Vector Machines

15.1.1 The Homogenous Case

It is often more convenient to consider homogenous halfspaces, namely, halfspaces
that pass through the origin and are thus defined by sign(w, x), where the bias term
b is set to be zero. Hard-SVM for homogenous halfspaces amounts to solving
min w2 s.t. i , yi w, xi  1.
w

(15.3)

As we discussed in Chapter 9, we can reduce the problem of learning

nonhomogenous halfspaces to the problem of learning homogenous halfspaces by
adding one more feature to each instance of xi , thus increasing the dimension to
d + 1.
Note, however, that the optimization problem given in Equation (15.2) does not
regularize the bias term b, while if we learn a homogenous halfspace in Rd+1 using
Equation (15.3) then we regularize the bias term (i.e., the d + 1 component of the
weight vector) as well. However, regularizing b usually does not make a significant
difference to the sample complexity.

15.1.2 The Sample Complexity of Hard-SVM

Recall that the VC-dimension of halfspaces in Rd is d + 1. It follows that the sample
complexity of learning halfspaces grows with the dimensionality of the problem.
Furthermore, the fundamental theorem of learning tells us that if the number of
examples is significantly smaller than d/ then no algorithm can learn an -accurate
halfspace. This is problematic when d is very large.
To overcome this problem, we will make an additional assumption on the underlying data distribution. In particular, we will define a separability with margin
assumption and will show that if the data is separable with margin then the sample complexity is bounded from above by a function of 1/ 2 . It follows that even if
the dimensionality is very large (or even infinite), as long as the data adheres to the
separability with margin assumption we can still have a small sample complexity.
There is no contradiction to the lower bound given in the fundamental theorem of
learning because we are now making an additional assumption on the underlying
data distribution.
Before we formally define the separability with margin assumption, there is a
scaling issue we need to resolve. Suppose that a training set S = (x1 , y1 ), . . . , (xm , ym )
is separable with a margin ; namely, the maximal objective value of Equation (15.1)
is at least . Then, for any positive scalar > 0, the training set S  =
(x1 , y1 ), . . . , (xm , ym ) is separable with a margin of . That is, a simple scaling
of the data can make it separable with an arbitrarily large margin. It follows that in
order to give a meaningful definition of margin we must take into account the scale
of the examples as well. One way to formalize this is using the definition that follows.
Definition 15.3. Let D be a distribution over Rd {1}. We say that D is separable
with a ( , )-margin if there exists (w , b  ) such that w  = 1 and such that with
probability 1 over the choice of (x, y) D we have that y(w , x + b ) and x
. Similarly, we say that D is separable with a ( , )-margin using a homogenous
halfspace if the preceding holds with a halfspace of the form (w , 0).

15.2 Soft-SVM and Norm Regularization

In the advanced part of the book (Chapter 26), we will prove that the sample
complexity of Hard-SVM depends on (/ )2 and is independent of the dimension
d. In particular, Theorem 26.13 in Section 26.3 states the following:
Theorem 15.4. Let D be a distribution over Rd {1} that satisfies the ( , )separability with margin assumption using a homogenous halfspace. Then, with
probability of at least 1 over the choice of a training set of size m, the 0-1 error of
the output of Hard-SVM is at most
*
4 (/ )2
+
m

2 log (2/)
.
m

Remark 15.1 (Margin and the Perceptron). In Section 9.1.2 we have described and
analyzed the Perceptron algorithm for finding an ERM hypothesis with respect to
the class of halfspaces. In particular, in Theorem 9.1 we upper bounded the number of updates the Perceptron might make on a given training set. It can be shown
(see Exercise 15.2) that the upper bound is exactly (/ )2 , where is the radius of
examples and is the margin.

15.2 SOFT-SVM AND NORM REGULARIZATION

The Hard-SVM formulation assumes that the training set is linearly separable,
which is a rather strong assumption. Soft-SVM can be viewed as a relaxation of the
Hard-SVM rule that can be applied even if the training set is not linearly separable.
The optimization problem in Equation (15.2) enforces the hard constraints
yi (w, xi  + b) 1 for all i . A natural relaxation is to allow the constraint to be
violated for some of the examples in the training set. This can be modeled by
introducing nonnegative slack variables, 1 , . . . , m , and replacing each constraint
yi (w, xi  + b) 1 by the constraint yi (w, xi  + b) 1 i . That is, i measures
by how much the constraint yi (w, xi  + b) 1 is being violated. Soft-SVM jointly
minimizes the norm of w (corresponding to the margin) and the average of i (corresponding to the violations of the constraints). The tradeoff between the two terms
is controlled by a parameter . This leads to the Soft-SVM optimization problem:
Soft-SVM
input: (x1 , y1 ), . . . , (xm , ym )
parameter: > 0
solve:


m

1
min w2 +
i
w,b,
m
i=1

s.t. i , yi (w, xi  + b) 1 i and i 0

output: w, b

(15.4)

171

172

Support Vector Machines

We can rewrite Equation (15.4) as a regularized loss minimization problem.

Recall the definition of the hinge loss:

hinge ((w, b), (x, y)) = max{0, 1 y(w, x + b)}.
Given (w, b) and a training set S, the averaged hinge loss on S is denoted by
hinge
L S ((w, b)). Now, consider the regularized loss minimization problem:


hinge
(15.5)
min w2 + L S ((w, b)) .
w,b

Claim 15.5. Equation (15.4) and Equation (15.5) are equivalent.

Proof. Fix some w, b and consider the minimization over in Equation (15.4).
Fix some i . Since i must be nonnegative, the best assignment to i would be 0
if yi (w, xi  + b) 1 and would be 1 yi (w, xi  + b) otherwise. In other words,
i =
hinge ((w, b), (xi , yi )) for all i , and the claim follows.
We therefore see that Soft-SVM falls into the paradigm of regularized loss minimization that we studied in the previous chapter. A Soft-SVM algorithm, that is, a
solution for Equation (15.5), has a bias toward low norm separators. The objective
function that we aim to minimize in Equation (15.5) penalizes not only for training
errors but also for large norm.
It is often more convenient to consider Soft-SVM for learning a homogenous
halfspace, where the bias term b is set to be zero, which yields the following
optimization problem:


hinge
(15.6)
min w2 + L S (w) ,
w

where

1
max{0, 1 yw, xi }.
m
m

hinge

LS

(w) =

i=1

15.2.1 The Sample Complexity of Soft-SVM

We now analyze the sample complexity of Soft-SVM for the case of homogenous
halfspaces (namely, the output of Equation (15.6)). In Corollary 13.8 we derived
a generalization bound for the regularized loss minimization framework assuming
that the loss function is convex and Lipschitz. We have already shown that the hinge
loss is convex so it is only left to analyze the Lipschitzness of the hinge loss.
Claim 15.6. Let f (w) = max{0, 1 yw, x}. Then, f is x-Lipschitz.
Proof. It is easy to verify that any subgradient of f at w is of the form x where
|| 1. The claim now follows from Lemma 14.7.
Corollary 13.8 therefore yields the following:
Corollary 15.7. Let D be a distribution over X {0, 1}, where X = {x : x }.
Consider running Soft-SVM (Equation (15.6)) on a training set S Dm and let A(S)
be the solution of Soft-SVM. Then, for every u,
hinge

E m [L D

SD

hinge

(A(S))] L D

(u) + u2 +

2 2
.
m

15.2 Soft-SVM and Norm Regularization

Furthermore, since the hinge loss upper bounds the 01 loss we also have
hinge

E m [L 01
D (A(S))] L D

SD

Last, for every B > 0, if we set =

2 2
B2m

(u) + u2 +

2 2
.
m

then
*

01

E m [L D (A(S))]

SD

hinge

E m [L D

SD

(A(S))]

hinge

min L D

w:wB

(w) +

8 2 B 2
.
m

We therefore see that we can control the sample complexity of learning a halfspace as a function of the norm of that halfspace, independently of the Euclidean
dimension of the space over which the halfspace is defined. This becomes highly
significant when we learn via embeddings into high dimensional feature spaces, as
we will consider in the next chapter.
Remark 15.2. The condition that X will contain vectors with a bounded norm follows from the requirement that the loss function will be Lipschitz. This is not just
a technicality. As we discussed before, separation with large margin is meaningless
without imposing a restriction on the scale of the instances. Indeed, without a constraint on the scale, we can always enlarge the margin by multiplying all instances
by a large scalar.

15.2.2 Margin and Norm-Based Bounds versus Dimension

The bounds we have derived for Hard-SVM and Soft-SVM do not depend on the
dimension of the instance space. Instead, the bounds depend on the norm of the
examples, , the norm of the halfspace B (or equivalently the margin parameter
) and, in the nonseparable case, the bounds also depend on the minimum hinge
loss of all halfspaces of norm B. In contrast, the VC-dimension of the class of
homogenous halfspaces is d, which implies that the error of an ERM hypothesis

decreases as d/m does. We now give an example in which 2 B 2 & d; hence the
bound given in Corollary 15.7 is much better than the VC bound.
Consider the problem of learning to classify a short text document according
to its topic, say, whether the document is about sports or not. We first need to
represent documents as vectors. One simple yet effective way is to use a bag-ofwords representation. That is, we define a dictionary of words and set the dimension
d to be the number of words in the dictionary. Given a document, we represent it
as a vector x {0, 1}d , where x i = 1 if the i th word in the dictionary appears in the
document and x i = 0 otherwise. Therefore, for this problem, the value of 2 will be
the maximal number of distinct words in a given document.
A halfspace for this problem assigns weights to words. It is natural to assume
that by assigning positive and negative weights to a few dozen words we will be
able to determine whether a given document is about sports or not with reasonable
accuracy. Therefore, for this problem, the value of B 2 can be set to be less than 100.
Overall, it is reasonable to say that the value of B 2 2 is smaller than 10,000.
On the other hand, a typical size of a dictionary is much larger than 10,000. For
example, there are more than 100,000 distinct words in English. We have therefore

173

174

Support Vector Machines

shown a problem in which there can be an order of magnitude difference between

learning a halfspace with the SVM rule and learning a halfspace using the vanilla
ERM rule.
Of course, it is possible to construct problems in which the SVM bound will be
worse than the VC bound. When we use SVM, we in fact introduce another form
of inductive bias we prefer large margin halfspaces. While this inductive bias can
significantly decrease our estimation error, it can also enlarge the approximation
error.

15.2.3 The Ramp Loss*

The margin-based bounds we have derived in Corollary 15.7 rely on the fact that
we minimize
the hinge loss. As we have shown in the previous subsection, the

term 2 B 2 /m can be much smaller than the corresponding term in the VC bound,

d/m. However, the approximation error in Corollary 15.7 is measured with respect
to the hinge loss while the approximation error in VC bounds is measured with
respect to the 01 loss. Since the hinge loss upper bounds the 01 loss, the approximation error with respect to the 01 loss will never exceed that of the hinge
loss.
 It is not possible to derive bounds that involve the estimation error term
2 B 2 /m for the 01 loss. This follows from the fact that the 01 loss is scale
insensitive, and therefore there is no meaning to the norm of w or its margin when
we measure error with the 01 loss. However, it is possible to define a loss
function
that on one hand it is scale sensitive and thus enjoys the estimation error 2 B 2 /m
while on the other hand it is more similar to the 01 loss. One option is the ramp
loss, defined as

ramp (w, (x, y)) = min{1,
hinge (w, (x, y))} = min{1 , max{0, 1 yw, x}}.
The ramp loss penalizes mistakes in the same way as the 01 loss and does not
penalize examples that are separated with margin. The difference between the ramp
loss and the 01 loss is only with respect to examples that are correctly classified but
not with a significant margin. Generalization bounds for the ramp loss are given in
the advanced part of this book.

hinge

0 1

ramp

y w, x

15.4 Duality

The reason SVM relies on the hinge loss and not on the ramp loss is that the
hinge loss is convex and, therefore, from the computational point of view, minimizing the hinge loss can be performed efficiently. In contrast, the problem of
minimizing the ramp loss is computationally intractable.

15.3 OPTIMALITY CONDITIONS AND SUPPORT VECTORS*

The name Support Vector Machine stems from the fact that the solution of hardSVM, w0 , is supported by (i.e., is in the linear span of) the examples that are exactly
at distance 1/w0  from the separating hyperplane. These vectors are therefore
called support vectors. To see this, we rely on Fritz John optimality conditions.
Theorem 15.8. Let w0 be as defined in Equation (15.3) and let I = {i : |w0 , xi | = 1}.
Then, there exist coefficients 1 , . . . , m such that

i xi .
w0 =
iI

The examples {xi : i I } are called support vectors.

The proof of this theorem follows by applying the following lemma to
Equation (15.3).
Lemma 15.9 (Fritz John). Suppose that
w argmin f (w) s.t. i [m], gi (w) 0,
w

where f , g1 , . . . , gm are differentiable. Then, there exists Rm such that f (w ) +



iI i gi (w ) = 0, where I = {i : gi (w ) = 0}.

15.4 DUALITY*
Historically, many of the properties of SVM have been obtained by considering
the dual of Equation (15.3). Our presentation of SVM does not rely on duality. For
completeness, we present in the following how to derive the dual of Equation (15.3).
We start by rewriting the problem in an equivalent form as follows. Consider the
function

m

0
if i , yi w, xi  1
.
i (1 yi w, xi ) =
g(w) = max
Rm :0
otherwise
i=1
We can therefore rewrite Equation (15.3) as


min w2 + g(w) .
w

(15.7)

Rearranging the preceding we obtain that Equation (15.3) can be rewritten as the
problem


m

1
2
w +
min max
i (1 yi w, xi ) .
(15.8)
w Rm :0 2
i=1

175

176

Support Vector Machines

Now suppose that we flip the order of min and max in the equation. This can only
decrease the objective value (see Exercise 15.4), and we have


m

1
2
w +
i (1 yi w, xi )
min max
w Rm :0 2
i=1


m

1
max
w2 +
min
i (1 yi w, xi ) .
Rm :0 w
2
i=1

The preceding inequality is called weak duality. It turns out that in our case, strong
duality also holds; namely, the inequality holds with equality. Therefore, the dual
problem is


m

1
2
w +
i (1 yi w, xi ) .
(15.9)
max min
Rm :0 w
2
i=1

We can simplify the dual problem by noting that once is fixed, the optimization
problem with respect to w is unconstrained and the objective is differentiable; thus,
at the optimum, the gradient equals zero:
w

m


i yi xi = 0 w =

i=1

m


i yi xi .

i=1

This shows us that the solution must be in the linear span of the examples, a
fact we will use later to derive SVM with kernels. Plugging the preceding into
Equation (15.9) we obtain that the dual problem can be rewritten as
0

02
!
"
m
m
0
0


1
0
0
max 0
i yi xi 0 +
i 1 yi
j y j x j , x i .
(15.10)
0
Rm :0
20
i=1

i=1

Rearranging yields the dual problem

m
m 
m


1
max
i
i j yi y j x j , xi  .
Rm :0
2
i=1

(15.11)

i=1 j =1

Note that the dual problem only involves inner products between instances and
does not require direct access to specific elements within an instance. This property is important when implementing SVM with kernels, as we will discuss in the
next chapter.

15.5 IMPLEMENTING SOFT-SVM USING SGD

In this section we describe a very simple algorithm for solving the optimization
problem of Soft-SVM, namely,


m
1

2
min
w +
max{0, 1 yw, xi } .
(15.12)
w
2
m
i=1

We rely on the SGD framework for solving regularized loss minimization problems,
as described in Section 14.5.3.

15.7 Bibliographic Remarks

Recall that, on the basis of Equation (14.15), we can rewrite the update rule of
SGD as
t
1 
vj,
w(t+1) =
t
j =1

where v j is a subgradient of the loss function at w( j ) on the random example chosen

at iteration j . For the hinge loss, given an example (x, y), we can choose v j to be 0 if


yw( j ) , x 1 and v j = y x otherwise (see Example 14.2). Denoting (t) = j <t v j
we obtain the following procedure.
SGD for Solving Soft-SVM
goal: Solve Equation (15.12)
parameter: T
initialize: (1) = 0
for t = 1, . . . , T
Let w(t) = 1t (t)
Choose i uniformly at random from [m]
If (yi w(t) , xi  < 1)
Set (t+1) = (t) + yi xi
Else
Set (t+1) = (t)

T
= T1 t=1
output: w
w(t)

15.6 SUMMARY
SVM is an algorithm for learning halfspaces with a certain type of prior knowledge,
namely, preference for large margin. Hard-SVM seeks the halfspace that separates
the data perfectly with the largest margin, whereas soft-SVM does not assume separability of the data and allows the constraints to be violated to some extent. The
sample complexity for both types of SVM is different from the sample complexity
of straightforward halfspace learning, as it does not depend on the dimension of the
domain but rather on parameters such as the maximal norms of x and w.
The importance of dimension-independent sample complexity will be realized
in the next chapter, where we will discuss the embedding of the given domain into
some high dimensional feature space as means for enriching our hypothesis class.
Such a procedure raises computational and sample complexity problems. The latter is solved by using SVM, whereas the former can be solved by using SVM with
kernels, as we will see in the next chapter.

15.7 BIBLIOGRAPHIC REMARKS

SVMs have been introduced in (Cortes and Vapnik 1992, Boser, Guyor and Vapnik
1992). There are many good books on the theoretical and practical aspects of SVMs.
For example, (Vapnik 1995, Cristianini & Shawe-Taylor 2000, Schlkopf & Smola
2002, Hsu et al. 2003, Steinwart and Christmann 2008). Using SGD for solving softSVM has been proposed in Shalev-Shwartz et al. (2007).

177

178

Support Vector Machines

15.8 EXERCISES
15.1 Show that the hard-SVM rule, namely,
argmax

min |w, xi  + b| s.t. i , yi (w, xi  + b) > 0,

(w,b):w=1 i[m]

is equivalent to the following formulation:

argmax

min yi (w, xi  + b).

(w,b):w=1 i[m]

(15.13)

Hint: Define G = {(w, b) : i , yi (w, xi  + b) > 0}.

1. Show that
argmax min yi (w, xi  + b) G
(w,b):w=1 i[m]

2. Show that (w, b) G ,

min yi (w, xi  + b) = min |w, xi  + b|

i[m]

i[m]

15.2 Margin and the Perceptron Consider a training set that is linearly separable with a
margin and such that all the instances are within a ball of radius . Prove that the
maximal number of updates the Batch Perceptron algorithm given in Section 9.1.2
will make when running on this training set is (/ )2.
15.3 Hard versus soft SVM: Prove or refute the following claim:
There exists > 0 such that for every sample S of m > 1 examples, which is separable by the class of homogenous halfspaces, the hard-SVM and the soft-SVM (with
parameter ) learning rules return exactly the same weight vector.
15.4 Weak duality: Prove that for any function f of two vector variables x X , y Y , it
holds that
min max f (x, y) max min f (x, y).
xX yY

yY xX

16
Kernel Methods

In the previous chapter we described the SVM paradigm for learning halfspaces in
high dimensional feature spaces. This enables us to enrich the expressive power of
halfspaces by first mapping the data into a high dimensional feature space, and then
learning a linear predictor in that space. This is similar to the AdaBoost algorithm,
which learns a composition of a halfspace over base hypotheses. While this approach
greatly extends the expressiveness of halfspace predictors, it raises both sample
complexity and computational complexity challenges. In the previous chapter we
tackled the sample complexity issue using the concept of margin. In this chapter we
tackle the computational complexity challenge using the method of kernels.
We start the chapter by describing the idea of embedding the data into a high
dimensional feature space. We then introduce the idea of kernels. A kernel is a
type of a similarity measure between instances. The special property of kernel similarities is that they can be viewed as inner products in some Hilbert space (or
Euclidean space of some high dimension) to which the instance space is virtually
embedded. We introduce the kernel trick that enables computationally efficient
implementation of learning, without explicitly handling the high dimensional representation of the domain instances. Kernel based learning algorithms, and in
particular kernel-SVM, are very useful and popular machine learning tools. Their
success may be attributed both to being flexible for accommodating domain specific prior knowledge and to having a well developed set of efficient implementation
algorithms.

16.1 EMBEDDINGS INTO FEATURE SPACES

The expressive power of halfspaces is rather restricted for example, the following
training set is not separable by a halfspace.
Let the domain be the real line; consider the domain points {10, 9, 8, . . ., 0,
1, . . . , 9, 10} where the labels are +1 for all x such that |x| > 2 and 1 otherwise.
To make the class of halfspaces more expressive, we can first map the original
instance space into another space (possibly of a higher dimension) and then learn a
halfspace in that space. For example, consider the example mentioned previously.
179

180

Kernel Methods

Instead of learning a halfspace in the original representation let us first define a

mapping : R R2 as follows:
(x) = (x, x 2 ).
We use the term feature space to denote the range of . After applying the
data can be easily explained using the halfspace h(x) = sign(w, (x) b), where
w = (0, 1) and b = 5.
The basic paradigm is as follows:
1. Given some domain set X and a learning task, choose a mapping : X F ,
for some feature space F , that will usually be Rn for some n (however, the
range of such a mapping can be any Hilbert space, including such spaces of
infinite dimension, as we will show later).
2. Given a sequence of labeled examples, S = (x1 , y1 ), . . . , (xm , ym ), create the
image sequence S = ((x1 ), y1 ), . . . , ((xm ), ym ).

3. Train a linear predictor h over S.

4. Predict the label of a test point, x, to be h((x)).
Note that, for every probability distribution D over X Y, we can readily
define its image probability distribution D over F Y by setting, for every subset
A F Y, D (A) = D( 1 (A)).1 It follows that for every predictor h over the
feature space, L D (h) = L D (h ), where h is the composition of h onto .
The success of this learning paradigm depends on choosing a good for a given
learning task: that is, a that will make the image of the data distribution (close to
being) linearly separable in the feature space, thus making the resulting algorithm a
good learner for a given task. Picking such an embedding requires prior knowledge
about that task. However, often some generic mappings that enable us to enrich
the class of halfspaces and extend its expressiveness are used. One notable example
is polynomial mappings, which are a generalization of the we have seen in the
previous example.
Recall that the prediction of a standard halfspace classifier on an instance x is
based on the linear mapping x  w, x. We can generalize linear mappings to a
polynomial mapping, x  p(x), where p is a multivariate polynomial of degree
k. For simplicity, consider first the case in which x is 1 dimensional. In that case,


p(x) = kj =0 w j x j , where w Rk+1 is the vector of coefficients of the polynomial we
need to learn. We can rewrite p(x) = w, (x) where : R Rk+1 is the mapping
x  (1, x, x 2 , x 3 , . . . , x k ). It follows that learning a k degree polynomial over R can
be done by learning a linear mapping in the (k + 1) dimensional feature space.
More generally, a degree k multivariate polynomial from Rn to R can be
written as
r


p(x) =
wJ
x Ji .
(16.1)
J [n]r :rk

i=1

As before, we can rewrite p(x) = w, (x) where now : Rn Rd is such that
for every J [n]r , r k, the coordinate of (x) associated with J is the monomial
1r
i=1 x Ji .
1

This is defined for every A such that 1 ( A) is measurable with respect to D.

16.2 The Kernel Trick

Naturally, polynomial-based classifiers yield much richer hypothesis classes than

halfspaces. We have seen at the beginning of this chapter an example in which the
training set, in its original domain (X = R), cannot be separable by a halfspace, but
after the embedding x  (x, x 2 ) it is perfectly separable. So, while the classifier
is always linear in the feature space, it can have highly nonlinear behavior on the
original space from which instances were sampled.
In general, we can choose any feature mapping that maps the original
instances into some Hilbert space.2 The Euclidean space Rd is a Hilbert space for
any finite d. But there are also infinite dimensional Hilbert spaces (as we shall see
later on in this chapter).
The bottom line of this discussion is that we can enrich the class of halfspaces by
first applying a nonlinear mapping, , that maps the instance space into some feature space, and then learning a halfspace in that feature space. However, if the range
of is a high dimensional space we face two problems. First, the VC-dimension of
halfspaces in Rn is n + 1, and therefore, if the range of is very large, we need
many more samples in order to learn a halfspace in the range of . Second, from
the computational point of view, performing calculations in the high dimensional
space might be too costly. In fact, even the representation of the vector w in the
feature space can be unrealistic. The first issue can be tackled using the paradigm
of large margin (or low norm predictors), as we already discussed in the previous
chapter in the context of the SVM algorithm. In the following section we address
the computational issue.

16.2 THE KERNEL TRICK

We have seen that embedding the input space into some high dimensional feature
space makes halfspace learning more expressive. However, the computational complexity of such learning may still pose a serious hurdle computing linear separators
over very high dimensional data may be computationally expensive. The common
solution to this concern is kernel based learning. The term kernels is used in
this context to describe inner products in the feature space. Given an embedding
of some domain space X into some Hilbert space, we define the kernel function K (x, x ) = (x), (x ). One can think of K as specifying similarity between
instances and of the embedding as mapping the domain set X into a space where
these similarities are realized as inner products. It turns out that many learning algorithms for halfspaces can be carried out just on the basis of the values of the kernel
function over pairs of domain points. The main advantage of such algorithms is that
they implement linear separators in high dimensional feature spaces without having to specify points in that space or expressing the embedding explicitly. The
remainder of this section is devoted to constructing such algorithms.
2

A Hilbert space is a vector space with an inner product, which is also complete. A space is complete if
all
Cauchy sequences in the space converge. In our case, the norm w is defined by the inner product
w, w. The reason we require the range of to be in a Hilbert space is that projections in a Hilbert
space are well defined. In particular, if M is a linear subspace of a Hilbert space, then every x in the
Hilbert space can be written as a sum x = u + v where u M and v, w = 0 for all w M. We use this
fact in the proof of the representer theorem given in the next section.

181

182

Kernel Methods

In the previous chapter we saw that regularizing the norm of w yields a small
sample complexity even if the dimensionality of the feature space is high. Interestingly, as we show later, regularizing the norm of w is also helpful in overcoming the
computational problem. To do so, first note that all versions of the SVM optimization problem we have derived in the previous chapter are instances of the following
general problem:
3
2
3

 2
(16.2)
min f w, (x1 ) , . . . , w, (xm ) + R(w) ,
w

where f : R R is an arbitrary function and R : R+ R is a monotonically nondecreasing function. For example, Soft-SVM for homogenous halfspaces
(Equation (15.6)) can be derived from Equation (16.2) by letting R(a) = a 2 and


f (a1 , . . . , am ) = m1 i max{0, 1 yi ai }. Similarly, Hard-SVM for nonhomogenous
halfspaces (Equation (15.2)) can be derived from Equation (16.2) by letting
R(a) = a 2 and letting f (a1 , . . . , am ) be 0 if there exists b such that yi (ai + b) 1
for all i , and f (a1 , . . . , am ) = otherwise.
The following theorem shows that there exists an optimal solution of
Equation (16.2) that lies in the span of {(x1 ), . . . , (xm )}.
m

Theorem 16.1 (Representer Theorem). Assume that is a mapping

from X to a


Hilbert space. Then, there exists a vector Rm such that w = m

i=1 i (xi ) is an
optimal solution of Equation (16.2).
Proof. Let w be an optimal solution of Equation (16.2). Because w is an element
of a Hilbert space, we can rewrite w as
w =

m


i (xi ) + u,

i=1

where u, (xi ) = 0 for all i . Set w = w u. Clearly, w 2 = w2 + u2 ,

thus w w . Since R is nondecreasing we obtain that R(w) R(w ).
Additionally, for all i we have that
yi w, (xi ) = yi w u, (xi ) = yi w , (xi ),
hence
3
3
 2
3
3
2
2
 2
f y1 w, (x1 ) , . . . , ym w, (xm ) = f y1 w , (x1 ) , . . . , ym w , (xm ) .
We have shown that the objective of Equation (16.2) at w cannot be larger than the


objective at w and therefore w is also an optimal solution. Since w = m
i=1 i (xi )
we conclude our proof.
On the basis of the representer theorem we can optimize Equation (16.2)
with

respect to the coefficients instead of the coefficients w as follows. Writing
w = mj=1 j (x j ) we have that for all i
!
"
m


3
2
j (x j ), (xi ) =
j (x j ), (xi ).
w, (xi ) =
j

j =1

16.2 The Kernel Trick

Similarly,
!
w2 =

j (x j ),

"
j (x j )

m


i j (xi ), (x j ).

i, j =1

Let K (x, x ) = (x), (x ) be a function that implements the kernel function with
respect to the embedding . Instead of solving Equation (16.2) we can solve the
equivalent problem

m
m


j K (x j , x1 ), . . . ,
j K (x j , xm )
minm f
R

j =1

j =1

4
5
5 m
i j K (x j , xi ).
+ R 6

(16.3)

i, j =1

To solve the optimization problem given in Equation (16.3), we do not need any
direct access to elements in the feature space. The only thing we should know is
how to calculate inner products in the feature space, or equivalently, to calculate
the kernel function. In fact, to solve Equation (16.3) we solely need to know the
value of the m m matrix G s.t. G i, j = K (xi , x j ), which is often called the Gram
matrix.
In particular, specifying the preceding to the Soft-SVM problem given in
Equation (15.6), we can rewrite the problem as


m

1
min T G +
max {0, 1 yi (G)i } ,
(16.4)
Rm
m
i=1

where (G)i is the i th element of the vector obtained by multiplying the Gram
matrix G by the vector . Note that Equation (16.4) can be written as quadratic
programming and hence can be solved efficiently. In the next section we describe an
even simpler algorithm for solving Soft-SVM with kernels.
Once we learn the coefficients we can calculate the prediction on a new
instance by
m
m


w, (x) =
j (x j ), (x) =
j K (x j , x).
j =1

j =1

The advantage of working with kernels rather than directly optimizing w in

the feature space is that in some situations the dimension of the feature space
is extremely large while implementing the kernel function is very simple. A few
examples are given in the following.
Example 16.1 (Polynomial Kernels). The k degree polynomial kernel is defined
to be
K (x, x ) = (1 + x, x )k .
Now we will show that this is indeed a kernel function. That is, we will show that
there exists a mapping from the original space to some higher dimensional space

183

184

Kernel Methods

2
3
for which K (x, x ) = (x), (x ) . For simplicity, denote x 0 = x 0 = 1. Then, we have
K (x, x ) = (1 + x, x )k = (1 + x, x ) (1 + x, x )

n
n


=
x j x j
x j x j
j =0

j =0

k


x Ji x Ji

J {0,1,...,n}k i=1

k


x Ji

J {0,1,...,n}k i=1

k


x Ji .

i=1

Now, if we define : Rn R(n+1) such that for J {0, 1, . . . , n}k there is an element
1
of (x) that equals ki=1 x Ji , we obtain that
K (x, x ) = (x), (x ).
Since contains all the monomials up to degree k, a halfspace over the range
of corresponds to a polynomial predictor of degree k over the original space.
Hence, learning a halfspace with a k degree polynomial kernel enables us to learn
polynomial predictors of degree k over the original space.
Note that here the complexity of implementing K is O(n) while the dimension
of the feature space is on the order of n k .
Example 16.2 (Gaussian Kernel). Let the original instance space be R and consider the mapping where for each nonnegative integer n 0 there exists an
x2

element (x)n that equals 1 e 2 x n . Then,

n!



 2

1 x2 n
1 (x )

 n
2
2
e
e
x
(x )
(x), (x ) =
n!
n!
n=0


x 2 +(x  )2 
(x x  )n

2
=e
n!
n=0

xx 2
= e 2 .

Here the feature space is of infinite dimension while evaluating the kernel is very
simple. More generally, given a scalar > 0, the Gaussian kernel is defined to be
K (x, x ) = e

xx 2
2 .

Intuitively, the Gaussian kernel sets the inner product in the feature space
between x, x to be close to zero if the instances are far away from each other (in
the original domain) and close to 1 if they are close. is a parameter that controls
the scale determining what we mean by close. It is easy to verify that K implements an inner product in a space in which for any n and any monomial of order k

16.2 The Kernel Trick

x2 1
n
there exists an element of (x) that equals 1 e 2
i=1 x Ji . Hence, we can learn
n!
any polynomial predictor over the original space by using a Gaussian kernel.
Recall that the VC-dimension of the class of all polynomial predictors is infinite (see Exercise 16.12). There is no contradiction, because the sample complexity
required to learn with Gaussian kernels depends on the margin in the feature space,
which will be large if we are lucky, but can in general be arbitrarily small.
The Gaussian kernel is also called the RBF kernel, for Radial Basis Functions.

16.2.1 Kernels as a Way to Express Prior Knowledge

As we discussed previously, a feature mapping, , may be viewed as expanding the
class of linear classifiers to a richer class (corresponding to linear classifiers over
the feature space). However, as discussed in the book so far, the suitability of any
hypothesis class to a given learning task depends on the nature of that task. One can
therefore think of an embedding as a way to express and utilize prior knowledge
about the problem at hand. For example, if we believe that positive examples can be
distinguished by some ellipse, we can define to be all the monomials up to order
2, or use a degree 2 polynomial kernel.
As a more realistic example, consider the task of learning to find a sequence of
characters (signature) in a file that indicates whether it contains a virus or not.
Formally, let X be the set of all finite strings over some alphabet set , and let
Xd be the set of all such strings of length at most d. The hypothesis class that one
wishes to learn is H = {h v : v Xd }, where, for a string x X , h v (x) is 1 iff v is a
substring of x (and h v (x) = 1 otherwise). Let us show how using an appropriate
embedding this class can be realized by linear classifiers over the resulting feature
space. Consider a mapping to a space Rs where s = |Xd |, so that each coordinate
of (x) corresponds to some string v and indicates whether v is a substring of x
(that is, for every x X , (x) is a vector in {0, 1}|Xd | ). Note that the dimension of
this feature space is exponential in d. It is not hard to see that every member of the
class H can be realized by composing a linear classifier over (x), and, moreover, by
such a halfspace whose norm is 1 and that attains
a margin of 1 (see Exercise 16.1).
Furthermore, for every x X , (x) = O( d). So, overall, it is learnable using
SVM with a sample complexity that is polynomial in d. However, the dimension of
the feature space is exponential in d so a direct implementation of SVM over the
feature space is problematic. Luckily, it is easy to calculate the inner product in the
feature space (i.e., the kernel function) without explicitly mapping instances into
the feature space. Indeed, K (x, x  ) is simply the number of common substrings of x
and x  , which can be easily calculated in time polynomial in d.
This example also demonstrates how feature mapping enables us to use
halfspaces for nonvectorial domains.

16.2.2 Characterizing Kernel Functions*

As we have discussed in the previous section, we can think of the specification of
the kernel matrix as a way to express prior knowledge. Consider a given similarity
function of the form K : X X R. Is it a valid kernel function? That is, does it

185

186

Kernel Methods

represent an inner product between (x) and (x ) for some feature mapping ?
The following lemma gives a sufficient and necessary condition.
Lemma 16.2. A symmetric function K : X X R implements an inner product in
some Hilbert space if and only if it is positive semidefinite; namely, for all x1 , . . . , xm ,
the Gram matrix, G i, j = K (xi , x j ), is a positive semidefinite matrix.
Proof. It is trivial to see that if K implements an inner product in some Hilbert
space then the Gram matrix is positive semidefinite. For the other direction, define
the space of functions over X as RX = { f : X R}. For each x X let (x) be
the function x  K (, x). Define a vector space by taking all linear combinations of
elements of the form K (, x). Define an inner product on this vector space to be
"
!




i K (, xi ),
j K (, x j ) =
i j K (xi , xj ).
i

i, j

This is a valid inner product since it is symmetric (because K is symmetric), it is

linear (immediate), and it is positive definite (it is easy to see that K (x, x) 0 with
equality only for (x) being the zero function). Clearly,
2
3 2
3
(x), (x ) = K (, x), K (, x) = K (x, x ),
which concludes our proof.

16.3 IMPLEMENTING SOFT-SVM WITH KERNELS

Next, we turn to solving Soft-SVM with kernels. While we could have designed
an algorithm for solving Equation (16.4), there is an even simpler approach that
directly tackles the Soft-SVM optimization problem in the feature space,


m
1

2
min
w +
max{0, 1 yw, (xi )} ,
(16.5)
w
2
m
i=1

while only using kernel evaluations. The basic observation is that the vector w(t)
maintained by the SGD procedure we have described in Section 15.5 is always in
the linear span of {(x1 ), . . . , (xm )}. Therefore, rather than maintaining w(t) we
can maintain the corresponding coefficients .
Formally, let K be the kernel function, namely, for all x, x , K (x, x ) =
(x), (x ). We shall maintain two vectors in Rm , corresponding to two vectors
(t) and w(t) defined in the SGD procedure of Section 15.5. That is, (t) will be a
vector such that
m

(t)
(t) =
j (x j )
(16.6)
j =1

and (t) be such that

w(t) =

m


(t)

j (x j ).

j =1

The vectors and are updated according to the following procedure.

(16.7)

16.4 Summary

SGD for Solving Soft-SVM with Kernels

Goal: Solve Equation (16.5)
parameter: T
Initialize: (1) = 0
for t = 1, . . . , T
Let (t) = 1t (t)
Choose i uniformly at random from [m]
(t+1)
(t)
For all j = i set j
= j

m
(t)
If (yi j =1 j K (x j , xi ) < 1)
(t+1)

(t)

= i + yi
Set i
Else
(t+1)
(t)
Set i
=

m i
= j =1 j (x j ) where =
Output: w

1
T

T

t=1

(t)

The following lemma shows that the preceding implementation is equivalent to

running the SGD procedure described in Section 15.5 on the feature space.
be the output of the SGD procedure described in Section 15.5,
Lemma 16.3. Let w


= mj=1 j (x j ) be the output of
when applied on the feature space, and let w
= w.

applying SGD with kernels. Then w

Proof. We will show that for every t Equation (16.6) holds, where (t) is the result
of running the SGD procedure described in Section 15.5 in the feature space. By the
definition of (t) = 1t (t) and w(t) = 1t (t) , this claim implies that Equation (16.7)
also holds, and the proof of our lemma will follow. To prove that Equation (16.6)
holds we use a simple inductive argument. For t = 1 the claim trivially holds. Assume
it holds for t 1. Then,
"
!
m
7
8
 (t)

(t)
(t)
yi w , (xi ) = yi
j (x j ), (xi ) = yi
j K (x j , xi ).
j =1

Hence, the condition in the two algorithms is equivalent and if we update we have
(t+1) = (t) + yi (xi ) =

m

j =1

(t)

j (x j ) + yi (xi ) =

m


(t+1)

(x j ),

j =1

which concludes our proof.

16.4 SUMMARY
Mappings from the given domain to some higher dimensional space, on which a
halfspace predictor is used, can be highly powerful. We benefit from a rich and
complex hypothesis class, yet need to solve the problems of high sample and computational complexities. In Chapter 10, we discussed the AdaBoost algorithm, which
faces these challenges by using a weak learner: Even though were in a very high
dimensional space, we have an oracle that bestows on us a single good coordinate
to work with on each iteration. In this chapter we introduced a different approach,

187

188

Kernel Methods

the kernel trick. The idea is that in order to find a halfspace predictor in the high
dimensional space, we do not need to know the representation of instances in that
space, but rather the values of inner products between the mapped instances. Calculating inner products between instances in the high dimensional space without
using their representation in that space is done using kernel functions. We have also
shown how the SGD algorithm can be implemented using kernels.
The ideas of feature mapping and the kernel trick allow us to use the framework
of halfspaces and linear predictors for nonvectorial data. We demonstrated how
kernels can be used to learn predictors over the domain of strings.
We presented the applicability of the kernel trick in SVM. However, the kernel trick can be applied in many other algorithms. A few examples are given as
exercises.
This chapter ends the series of chapters on linear predictors and convex problems. The next two chapters deal with completely different types of hypothesis
classes.

16.5 BIBLIOGRAPHIC REMARKS

In the context of SVM, the kernel-trick has been introduced in Boser et al. (1992).
See also Aizerman et al. (1964). The observation that the kernel-trick can be applied
whenever an algorithm only relies on inner products was first stated by Schlkopf
et al. (1998). The proof of the representer theorem is given in (Schlkopf et al.
2000, Schlkopf et al. 2001). The conditions stated in Lemma 16.2 are simplification
of conditions due to Mercer. Many useful kernel functions have been introduced in
the literature for various applications. We refer the reader to Schlkopf & Smola
(2002).

16.6 EXERCISES
16.1 Consider the task of finding a sequence of characters in a file, as described in
Section 16.2.1. Show that every member of the class H can be realized by composing
a linear classifier over (x), whose norm is 1 and that attains a margin of 1.
16.2 Kernelized Perceptron: Show how to run the Perceptron algorithm while only
accessing the instances via the kernel function. Hint: The derivation is similar to
the derivation of implementing SGD with kernels.
16.3 Kernel Ridge Regression: The ridge regression problem, with a feature mapping
, is the problem of finding a vector w that minimizes the function
1 
(w, (xi ) yi )2 ,
2m
m

f (w) = w2 +

(16.8)

i=1

and then returning the predictor

h(x) = w, x.
Show how to implement the ridge regression algorithm with kernels.
Hint: The representer theorem tells us that there exists a vector Rm such that


m
i=1 i (xi ) is a minimizer of Equation (16.8).

16.6 Exercises

1. Let G be the Gram matrix with regard to S and K . That is, G i j = K (xi , x j ).
Define g : Rm R by
m
1 
(, G ,i  yi )2 ,
2m

g() = T G +

(16.9)

i=1

where G ,i

is the i th column of G. Show that if minimizes Equation (16.9)

then w = m
i=1 i (xi ) is a minimizer of f .
2. Find a closed form expression for .
16.4 Let N be any positive integer. For every x, x  {1, . . . , N } define
K (x, x  ) = min{x, x  }.
Prove that K is a valid kernel; namely, find a mapping : {1, . . . , N } H where H
is some Hilbert space, such that
x, x  {1, . . . , N }, K (x, x  ) = (x), (x  ).
16.5 A supermarket manager would like to learn which of his customers have babies on
the basis of their shopping carts. Specifically, he sampled i.i.d. customers, where
for customer i , let xi {1, . . . , d} denote the subset of items the customer bought,
and let yi {1} be the label indicating whether this customer has a baby. As prior
knowledge, the manager knows that there are k items such that the label is determined to be 1 iff the customer bought at least one of these k items. Of course, the
identity of these k items is not known (otherwise, there was nothing to learn). In
addition, according to the store regulation, each customer can buy at most s items.
Help the manager to design a learning algorithm such that both its time complexity
and its sample complexity are polynomial in s, k, and 1/ .
16.6 Let X be an instance set and let be a feature mapping of X into some Hilbert
feature space V . Let K : X X R be a kernel function that implements inner
products in the feature space V .
Consider the binary classification algorithm that predicts the label of an unseen
instance according to the class with the closest average. Formally, given a training
sequence S = (x1 , y1 ), . . ., (xm , ym ), for every y {1} we define
cy =

1 
(xi ).
m y i:y =y
i

where m y = |{i : yi = y}|. We assume that m + and m are nonzero. Then, the
algorithm outputs the following decision rule:

1 (x) c+  (x) c 
h(x) =
0 otherwise.
1. Let w = c+ c and let b = 12 (c 2 c+ 2 ). Show that
h(x) = sign(w, (x) + b).
2. Show how to express h(x) on the basis of the kernel function, and without
accessing individual entries of (x) or w.

189

17
Multiclass, Ranking, and Complex
Prediction Problems

Multiclass categorization is the problem of classifying instances into one of several

possible target classes. That is, we are aiming at learning a predictor h : X Y,
where Y is a finite set of categories. Applications include, for example, categorizing
documents according to topic (X is the set of documents and Y is the set of possible
topics) or determining which object appears in a given image (X is the set of images
and Y is the set of possible objects).
The centrality of the multiclass learning problem has spurred the development of
various approaches for tackling the task. Perhaps the most straightforward approach
is a reduction from multiclass classification to binary classification. In Section 17.1
we discuss the most common two reductions as well as the main drawback of the
reduction approach.
We then turn to describe a family of linear predictors for multiclass problems.
Relying on the RLM and SGD frameworks from previous chapters, we describe
several practical algorithms for multiclass prediction.
In Section 17.3 we show how to use the multiclass machinery for complex prediction problems in which Y can be extremely large but has some structure on it.
This task is often called structured output learning. In particular, we demonstrate
this approach for the task of recognizing handwritten words, in which Y is the set of
all possible strings of some bounded length (hence, the size of Y is exponential in
the maximal length of a word).
Finally, in Section 17.4 and Section 17.5 we discuss ranking problems in which
the learner should order a set of instances according to their relevance. A typical
application is ordering results of a search engine according to their relevance to the
query. We describe several performance measures that are adequate for assessing
the performance of ranking predictors and describe how to learn linear predictors
for ranking problems efficiently.

17.1 ONE-VERSUS-ALL AND ALL-PAIRS

The simplest approach to tackle multiclass prediction problems is by reduction to binary classification. Recall that in multiclass prediction we would like
190

17.1 One-versus-All and All-Pairs

to learn a function h : X Y. Without loss of generality let us denote

Y = {1, . . . , k}.
In the One-versus-All method (a.k.a. One-versus-Rest) we train k binary classifiers, each of which discriminates between one class and the rest of the classes. That
is, given a training set S = (x1 , y1 ), . . . , (xm , ym ), where every yi is in Y, we construct
1
k binary training sets, S1 , . . . , Sk , where Si = (x1 , ( 1) [y1 =i] ), . . . , (xm , ( 1)1[ym =i] ). In
words, Si is the set of instances labeled 1 if their label in S was i , and 1 otherwise.
For every i [k] we train a binary predictor h i : X {1} based on Si , hoping that
h i (x) should equal 1 if and only if x belongs to class i . Then, given h 1 , . . . , h k , we
construct a multiclass predictor using the rule
h(x) argmax h i (x).

(17.1)

i[k]

When more than one binary hypothesis predicts 1 we should somehow decide
which class to predict (e.g., we can arbitrarily decide to break ties by taking the
minimal index in argmaxi h i (x)). A better approach can be applied whenever each
h i hides additional information, which can be interpreted as the confidence in the
prediction y = i . For example, this is the case in halfspaces, where the actual prediction is sign(w, x), but we can interpret w, x as the confidence in the prediction.
In such cases, we can apply the multiclass rule given in Equation (17.1) on the real
valued predictions. A pseudocode of the One-versus-All approach is given in the
following.

One-versus-All
input:
training set S = (x1 , y1 ), . . . , (xm , ym )
algorithm for binary classification A
foreach i Y
1
let Si = (x1 , ( 1) [y1 =i] ), . . . , (xm , ( 1)1[ym =i] )
let h i = A(Si )
output:
the multiclass hypothesis defined by h(x) argmaxiY h i (x)

Another popular reduction is the All-Pairs approach, in which all pairs of classes
are compared to each other. Formally, given a training set S = (x1 , y1 ), . . . , (xm , ym ),
where every yi is in [k], for every 1 i < j k we construct a binary training
sequence, Si, j , containing all examples from S whose label is either i or j . For each
such an example, we set the binary label in Si, j to be +1 if the multiclass label in
S is i and 1 if the multiclass label in S is j . Next, we train a binary classification
algorithm based on every Si, j to get h i, j . Finally, we construct a multiclass classifier
by predicting the class that had the highest number of wins. A pseudocode of the
All-Pairs approach is given in the following.

191

192

Multiclass and Ranking

All-Pairs
input:
training set S = (x1 , y1 ), . . . , (xm , ym )
algorithm for binary classification A
foreach i , j Y s.t. i < j
initialize Si, j to be the empty sequence
for t = 1, . . . , m
If yt = i add (xt , 1) to Si, j
If yt = j add (xt , 1) to Si, j
let h i, j = A(Si, j )
output:
the multiclass hypothesis
defined by



h(x) argmaxiY
sign(
j

i
)
h
(x)
i, j
j Y
Although reduction methods such as the One-versus-All and All-Pairs are simple and easy to construct from existing algorithms, their simplicity has a price. The
binary learner is not aware of the fact that we are going to use its output hypotheses
for constructing a multiclass predictor, and this might lead to suboptimal results, as
illustrated in the following example.
Example 17.1. Consider a multiclass categorization problem in which the instance
space is X = R2 and the label set is Y = {1, 2, 3}. Suppose that instances of the
different classes are located in nonintersecting balls as depicted in the following.

Suppose that the probability masses of classes 1, 2, 3 are 40%, 20%, and 40%,
respectively. Consider the application of One-versus-All to this problem, and
assume that the binary classification algorithm used by One-versus-All is ERM with
respect to the hypothesis class of halfspaces. Observe that for the problem of discriminating between class 2 and the rest of the classes, the optimal halfspace would
be the all negative classifier. Therefore, the multiclass predictor constructed by Oneversus-All might err on all the examples from class 2 (this will be the case if the tie in
the definition of h(x) is broken by the numerical
value
 of the class label). Incontrast,


1 1

, w2 = (0, 1), and w3 = 1 , 1 ,

if we choose h i (x) = wi , x, where w1 = ,
2

then the classifier defined by h(x) = argmaxi h i (x) perfectly predicts all the examples. We see that even though the approximation error of the class of predictors of
the form h(x) = argmaxi wi , x is zero, the One-versus-All approach might fail to
find a good predictor from this class.

17.2 Linear Multiclass Predictors

17.2 LINEAR MULTICLASS PREDICTORS

In light of the inadequacy of reduction methods, in this section we study a more
direct approach for learning multiclass predictors. We describe the family of linear
multiclass predictors. To motivate the construction of this family, recall that a linear
predictor for binary classification (i.e., a halfspace) takes the form
h(x) = sign(w, x).
An equivalent way to express the prediction is as follows:
h(x) = argmax w, yx,
y{1}

where yx is the vector obtained by multiplying each element of x by y.

This representation leads to a natural generalization of halfspaces to multiclass
problems as follows. Let  : X Y Rd be a class-sensitive feature mapping. That
is,  takes as input a pair (x, y) and maps it into a d dimensional feature vector.
Intuitively, we can think of the elements of (x, y) as score functions that assess
how well the label y fits the instance x. We will elaborate on  later on. Given 
and a vector w Rd , we can define a multiclass predictor, h : X Y, as follows:
h(x) = argmax w, (x, y).
yY

That is, the prediction of h for the input x is the label that achieves the highest
weighted score, where weighting is according to the vector w.
Let W be some set of vectors in Rd , for example, W = {w Rd : w B},
for some scalar B > 0. Each pair (, W ) defines a hypothesis class of multiclass
predictors:
H,W = {x  argmax w, (x, y) : w W }.
yY

Of course, the immediate question, which we discuss in the sequel, is how to construct a good . Note that if Y = {1} and we set (x, y) = yx and W = Rd , then
H,W becomes the hypothesis class of homogeneous halfspace predictors for binary
classification.

17.2.1 How to Construct 

As mentioned before, we can think of the elements of (x, y) as score functions
that assess how well the label y fits the instance x. Naturally, designing a good 
is similar to the problem of designing a good feature mapping (as we discussed in
Chapter 16 and as we will discuss in more detail in Chapter 25). Two examples of
useful constructions are given in the following.
The Multivector Construction:
Let Y = {1, . . . , k} and let X = Rn . We define  : X Y Rd , where d = nk, as
follows
(x, y) = [ 0, . . . , 0 , x 1 , . . . , x n , 0, . . . , 0 ].
9 :; < 9 :; < 9 :; <
R(y1)n

Rn

R(ky)n

(17.2)

193

194

Multiclass and Ranking

That is, (x, y) is composed of k vectors, each of which is of dimension n, where

we set all the vectors to be the all zeros vector except the yth vector, which is set
to be x. It follows that we can think of w Rnk as being composed of k weight
vectors in Rn , that is, w = [w1 ; . . . ; wk ], hence the name multivector construction.
By the construction we have that w, (x, y) = w y , x, and therefore the multiclass
prediction becomes
h(x) = argmax w y , x.
yY

A geometric illustration of the multiclass prediction over X = R2 is given in the

following.
w2
w1

w3

w4

TF-IDF:
The previous definition of (x, y) does not incorporate any prior knowledge about
the problem. We next describe an example of a feature function  that does incorporate prior knowledge. Let X be a set of text documents and Y be a set of possible
topics. Let d be a size of a dictionary of words. For each word in the dictionary,
whose corresponding index is j , let T F( j , x) be the number of times the word corresponding to j appears in the document x. This quantity is called Term-Frequency.
Additionally, let D F( j , y) be the number of times the word corresponding to j
appears in documents in our training set that are not about topic y. This quantity
is called Document-Frequency and measures whether word j is frequent in other
topics. Now, define  : X Y Rd to be such that


 j (x, y) = T F( j , x) log DFm( j ,y) ,
where m is the total number of documents in our training set. The preceding quantity
is called term-frequency-inverse-document-frequency or TF-IDF for short. Intuitively,  j (x, y) should be large if the word corresponding to j appears a lot in the
document x but does not appear at all in documents that are not on topic y. If this
is the case, we tend to believe that the document x is on topic y. Note that unlike
the multivector construction described previously, in the current construction the
dimension of  does not depend on the number of topics (i.e., the size of Y).

17.2.2 Cost-Sensitive Classification

So far we used the zero-one loss as our performance measure of the quality of h(x).
That is, the loss of a hypothesis h on an example (x, y) is 1 if h(x) = y and 0 otherwise. In some situations it makes more sense to penalize different levels of loss for
different mistakes. For example, in object recognition tasks, it is less severe to predict that an image of a tiger contains a cat than predicting that the image contains a

17.2 Linear Multiclass Predictors

whale. This can be modeled by specifying a loss function,  : Y Y R+ , where for

every pair of labels, y  , y, the loss of predicting the label y  when the correct label is
y is defined to be ( y  , y). We assume that ( y, y) = 0. Note that the zero-one loss
can be easily modeled by setting ( y  , y) = 1[ y  = y] .

17.2.3 ERM
We have defined the hypothesis class H,W and specified a loss function . To learn
the class with respect to the loss function, we can apply the ERM rule with respect
to this class. That is, we search for a multiclass hypothesis h H,W , parameterized
by a vector w, that minimizes the empirical risk with respect to ,
1
(h(xi ), yi ).
m
m

L S (h) =

i=1

We now show that when W = Rd and we are in the realizable case, then it is
possible to solve the ERM problem efficiently using linear programming. Indeed, in
the realizable case, we need to find a vector w Rd that satisfies
i [m], yi = argmaxw, (xi , y).
yY

Equivalently, we need that w will satisfy the following set of linear inequalities
i [m], y Y \ {yi }, w, (xi , yi ) > w, (xi , y).
Finding w that satisfies the preceding set of linear equations amounts to solving a
linear program.
As in the case of binary classification, it is also possible to use a generalization
of the Perceptron algorithm for solving the ERM problem. See Exercise 17.2.
In the nonrealizable case, solving the ERM problem is in general computationally hard. We tackle this difficulty using the method of convex surrogate loss
functions (see Section 12.3). In particular, we generalize the hinge loss to multiclass
problems.

17.2.4 Generalized Hinge Loss

Recall that in binary classification, the hinge loss is defined to be max{0, 1 yw, x}.
We now generalize the hinge loss to multiclass predictors of the form
h w (x) = argmax w, (x, y ).
y  Y

Recall that a surrogate convex loss should upper bound the original nonconvex loss,
which in our case is (h w (x), y). To derive an upper bound on (h w (x), y) we first
note that the definition of h w (x) implies that
w, (x, y) w, (x, h w (x)).
Therefore,
(h w (x), y) (h w (x), y) + w, (x, h w (x)) (x, y).

195

Multiclass and Ranking

Since h w (x) Y we can upper bound the right-hand side of the preceding by

 def
max (y  , y) + w, (x, y  ) (x, y) =
(w, (x, y)).
(17.3)
y  Y

We use the term generalized hinge loss to denote the preceding expression. As we
have shown,
(w, (x, y)) (h w (x), y). Furthermore, equality holds whenever the
score of the correct label is larger than the score of any other label, y  , by at least
(y  , y), namely,
y  Y \ {y}, w, (x, y) w, (x, y ) + (y  , y).
It is also immediate to see that
(w, (x, y)) is a convex function with respect to w
since it is a maximum over linear functions of w (see Claim 12.5 in Chapter 12), and
that
(w, (x, y)) is -Lipschitz with = max y  Y (x, y  ) (x, y).
Remark 17.2. We use the name generalized hinge loss since in the binary case,
when Y = {1}, if we set (x, y) = yx
2 , then the generalized hinge loss becomes the
vanilla hinge loss for binary classification,

(w, (x, y)) = max{0, 1 yw, x}.
Geometric Intuition:
The feature function  : X Y Rd maps each x into |Y| vectors in Rd . The value of

(w, (x, y)) will be zero if there exists a direction w such that when projecting the |Y|
vectors onto this direction we obtain that each vector is represented by the scalar
w, (x, y), and we can rank the different points on the basis of these scalars so
that

The point corresponding to the correct y 2is top-ranked
3
2
3

For each y  = y, the difference between w, (x, y) and w, (x,
y  ) is larger
2
3
than
the loss
of predicting y  instead of y. The difference w, (x, y)
3
2
w, (x, y  ) is also referred to as the margin (see Section 15.1).

This is illustrated in the figure following:

w

(x, y)
(x, y")

")
,y

y
(

196

(y

(y

,y
')

,y

')

(x, y')

17.2.5 Multiclass SVM and SGD

Once we have defined the generalized hinge loss, we obtain a convex-Lipschitz
learning problem and we can apply our general techniques for solving such

17.2 Linear Multiclass Predictors

problems. In particular, the RLM technique we have studied in Chapter 13 yields

the multiclass SVM rule:

Multiclass SVM
input: (x1 , y1 ), . . . , (xm , ym )
parameters:
regularization parameter > 0
loss function  : Y Y R+
class-sensitive feature mapping  : X Y Rd
solve:


m



1
min w2 +
max (y  , yi ) + w, (xi , y  ) (xi , yi )
m
y  Y
wRd
i=1

output the predictor h w (x) = argmax yY w, (x, y)

We can solve the optimization problem associated with multiclass SVM

using generic convex optimization algorithms (or using the method described in
Section 15.5). Let us analyze the risk of the resulting hypothesis. The analysis
seamlessly follows from our general analysis for convex-Lipschitz problems given
in Chapter 13. In particular, applying Corollary 13.8 and using the fact that the generalized hinge loss upper bounds the  loss, we immediately obtain an analog of
Corollary 15.7:
Corollary 17.1. Let D be a distribution over X Y, let  : X Y Rd , and assume
that for all x X and y 
Y we have (x, y) /2. Let B > 0. Consider running
Multiclass SVM with =
Multiclass SVM. Then,

2 2
B 2m

on a training set S Dm and let h w be the output of

*


E m [L D (h w )]

SD

ghinge

E m [L D

SD

(w)]

ghinge

min L D

u:uB

(u) +

8 2 B 2
,
m

where L 
(w) = E(x,y)D [
(w, (x, y))] with

D (h) = E(x,y)D [(h(x), y)] and L D
being the generalized hinge-loss as defined in Equation (17.3).
ghinge

ghinge

We can also apply the SGD learning framework for minimizing L D

(w)
as described in Chapter 14. Recall Claim 14.6, which dealt with subgradients
of max functions. In light of this claim, in order to find a subgradient of the
generalized hinge loss all we need to do is to find y Y that achieves the maximum in the definition of the generalized hinge loss. This yields the following

197

198

Multiclass and Ranking

algorithm:
SGD for Multiclass Learning
parameters:
Scalar > 0, integer T > 0
loss function  : Y Y R+
class-sensitive feature mapping  : X Y Rd
initialize: w(1) = 0 Rd
for t = 1, 2, . . . , T
sample (x, y) D
find y argmaxy  Y ((y  , y) + w(t) , (x, y  ) (x, y))
set vt = (x, y ) (x, y)
update w(t+1) = w(t) vt

T
= T1 t=1
output w
w(t)
Our general analysis of SGD given in Corollary 14.12 immediately implies:
Corollary 17.2. Let D be a distribution over X Y, let  : X Y Rd , and assume
that for all x X and y Y we have (x, y) /2. Let B > 0. Then, for every
> 0, if we run SGD for multiclass learning with a number of iterations (i.e., number
of examples)
B 2 2
T 2


B2
and with = 2 T , then the output of SGD satisfies
E [L 
)]
D (h w

SD m

ghinge

E [L D

SD m

(w)]

ghinge

min L D

u:uB

(u) + .

Remark 17.3. It is interesting to note that the risk bounds given in Corollary 17.1 and
Corollary 17.2 do not depend explicitly on the size of the label set Y, a fact we will
rely on in the next section. However, the bounds may depend implicitly on the size
of Y via the norm of (x, y) and the fact that the bounds are meaningful only when
ghinge
(u) is not excessively large.
there exists some vector u, u B, for which L D

17.3 STRUCTURED OUTPUT PREDICTION

Structured output prediction problems are multiclass problems in which Y is very
large but is endowed with a predefined structure. The structure plays a key role in
constructing efficient algorithms. To motivate structured learning problems, consider the problem of optical character recognition (OCR). Suppose we receive an
image of some handwritten word and would like to predict which word is written in
the image. To simplify the setting, suppose we know how to segment the image into
a sequence of images, each of which contains a patch of the image corresponding
to a single letter. Therefore, X is the set of sequences of images and Y is the set
of sequences of letters. Note that the size of Y grows exponentially with the maximal length of a word. An example of an image x corresponding to the label y =
workable is given in the following.

17.3 Structured Output Prediction

To tackle structure prediction we can rely on the family of linear predictors

described in the previous section. In particular, we need to define a reasonable loss
function for the problem, , as well as a good class-sensitive feature mapping, .
By good we mean a feature mapping that will lead to a low approximation error
for the class of linear predictors with respect to  and . Once we do this, we can
rely, for example, on the SGD learning algorithm defined in the previous section.
However, the huge size of Y poses several challenges:
1. To apply the multiclass prediction we need to solve a maximization problem
over Y. How can we predict efficiently when Y is so large?
2. How do we train w efficiently? In particular, to apply the SGD rule we again
need to solve a maximization problem over Y.
3. How can we avoid overfitting?
In the previous section we have already shown that the sample complexity of
learning a linear multiclass predictor does not depend explicitly on the number of
classes. We just need to make sure that the norm of the range of  is not too large.
This will take care of the overfitting problem. To tackle the computational challenges we rely on the structure of the problem, and define the functions  and  so
that calculating the maximization problems in the definition of h w and in the SGD
algorithm can be performed efficiently. In the following we demonstrate one way to
achieve these goals for the OCR task mentioned previously.
To simplify the presentation, let us assume that all the words in Y are of length
r and that the number of different letters in our alphabet is q. Let y and y be two
 , y) to be the
words (i.e., sequences of letters) in Y. We define the function (y
1
r

average number of letters that are different in y and y, namely, r i=1 1[yi = yi ] .
Next, let us define a class-sensitive feature mapping (x, y). It will be convenient
to think about x as a matrix of size n r , where n is the number of pixels in each
image, and r is the number of images in the sequence. The j th column of x corresponds to the j th image in the sequence (encoded as a vector of gray level values
of pixels). The dimension of the range of  is set to be d = n q + q 2 .
The first nq feature functions are type 1 features and take the form:
1
x i,t 1[yt = j] .
r
r

i, j ,1 (x, y) =

t=1

That is, we sum the value of the i th pixel only over the images for which y assigns
the letter j . The triple index (i , j , 1) indicates that we are dealing with feature (i , j )
of type 1. Intuitively, such features can capture pixels in the image whose gray level

199

200

Multiclass and Ranking

values are indicative of a certain letter. The second type of features take the form
1
1[yt =i] 1[yt1 = j] .
r
r

i, j ,2 (x, y) =

t=2

That is, we sum the number of times the letter i follows the letter j . Intuitively,
these features can capture rules like It is likely to see the pair qu in a word or It
is unlikely to see the pair rz in a word. Of course, some of these features will not
be very useful, so the goal of the learning process is to assign weights to features by
learning the vector w, so that the weighted score will give us a good prediction via
h w (x) = argmax w, (x, y).
yY

It is left to show how to solve the optimization problem in the definition of h w (x)
efficiently, as well as how to solve the optimization problem in the definition of y in
the SGD algorithm. We can do this by applying a dynamic programming procedure.
We describe the procedure for solving the maximization in the definition of h w and
leave as an exercise the maximization problem in the definition of y in the SGD
algorithm.
To derive the dynamic programming procedure, let us first observe that we can
write
r

(x, yt , yt1 ),
(x, y) =
t=1

for an appropriate : X [q] [q] {0} Rd , and for simplicity we assume that y0
is always equal to 0. Indeed, each feature function i, j ,1 can be written in terms of
i, j ,1 (x, yt , yt1 ) = x i,t 1[yt = j] ,
while the feature function i, j ,2 can be written in terms of
i, j ,2 (x, yt , yt1 ) = 1[yt =i] 1[yt1 = j] .
Therefore, the prediction can be written as
h w (x) = argmax
yY

r


w, (x, yt , yt1 ).

(17.4)

t=1

In the following we derive a dynamic programming procedure that solves every

problem of the form given in Equation (17.4). The procedure will maintain a matrix
M Rq,r such that
Ms, =

max

(y1 ,...,y ):y =s

w, (x, yt , yt1 ).

t=1

Clearly, the maximum of w, (x, y) equals maxs Ms,r . Furthermore, we can
calculate M in a recursive manner:

2
3

 , 1 + w, (x, s, s )
Ms, = max
M
.
(17.5)
s

s

17.4 Ranking

This yields the following procedure:

Dynamic Programming for Calculating h w (x) as Given in
Equation (17.4)
input: a matrix x Rn,r and a vector w
initialize:
foreach s [q]
Ms,1 = w, (x, s, 1)
for = 2, . . . ,r
foreach s [q]
set Ms, as in Equation (17.5)
set Is, to be the s  that maximizes Equation (17.5)
set yt = argmaxs Ms,r
for = r , r 1, . . . , 2
set y 1 = I y ,
output: y = (y1 , . . . , yr )

17.4 RANKING
Ranking is the problem of ordering a set of instances according to their relevance. A typical application is ordering results of a search engine according to their
relevance to the query. Another example is a system that monitors electronic transactions and should alert for possible fraudulent transactions. Such a system should
order transactions according to how suspicious they are.

n
Formally, let X =
n=1 X be the set of all sequences of instances from X of
arbitrary length. A ranking hypothesis, h, is a function that receives a sequence of
instances x = (x1 , . . . , xr ) X , and returns a permutation of [r ]. It is more convenient to let the output of h be a vector y Rr , where by sorting the elements of y
we obtain the permutation over [r ]. We denote by (y) the permutation over [r ]
induced by y. For example, for r = 5, the vector y = (2, 1, 6, 1, 0. 5) induces the
permutation (y) = (4, 3, 5, 1, 2). That is, if we sort y in an ascending order, then
we obtain the vector ( 1, 0. 5, 1, 2, 6). Now, (y)i is the position of yi in the sorted
vector ( 1, 0. 5, 1, 2, 6). This notation reflects that the top-ranked instances are
those that achieve the highest values in (y).
In the notation of our PAC learning model, the examples domain is Z =

r
r
r=1 (X R ), and the hypothesis class, H, is some set of ranking hypotheses. We
next turn to describe loss functions for ranking. There are many possible ways to
define such loss functions, and here we list a few examples. In all the examples we

define
(h, (x, y)) = (h(x), y), for some function  : r=1
(Rr Rr ) R+ .

01 Ranking loss: (y , y) is zero if y and y induce exactly the same ranking
and (y , y) = 1 otherwise. That is, (y , y) = 1[(y )=(y)] . Such a loss function is
almost never used in practice as it does not distinguish between the case in which
(y ) is almost equal to (y) and the case in which (y ) is completely different
from (y).

201

202

Multiclass and Ranking

Kendall-Tau Loss: We count the number of pairs (i , j ) that are in different order
in the two permutations. This can be written as

(y , y) =

r1 
r

2
1[sign(yi y j )=sign(yi y j )] .
r (r 1)
i=1 j =i+1

This loss function is more useful than the 01 loss as it reflects the level of
similarity between the two rankings.

Normalized Discounted Cumulative Gain (NDCG): This measure emphasizes
the correctness at the top of the list by using a monotonically nondecreasing
discount function D : N R+ . We first define a discounted cumulative gain
measure:
r

D((y )i ) yi .
G(y , y) =
i=1

In words, if we interpret yi as a score of the true relevance of item i , then

we take a weighted sum of the relevance of the elements, while the weight of
yi is determined on the basis of the position of i in (y ). Assuming that all
elements of y are nonnegative, it is easy to verify that 0 G(y , y) G(y, y).
We can therefore define a normalized discounted cumulative gain by the ratio
G(y , y)/G(y, y), and the corresponding loss function would be
(y , y) = 1

r



1
G(y , y)
=
D((y)i ) D((y )i ) yi .
G(y, y)
G(y, y)
i=1

We can easily see that (y , y) [0, 1] and that (y , y) = 0 whenever
(y ) = (y).
A typical way to define the discount function is by

1
if i {r k + 1, . . . ,r }
D(i ) = log2 (ri+2)
0
otherwise
where k < r is a parameter. This means that we care more about elements that
are ranked higher, and we completely ignore elements that are not at the top-k
ranked elements. The NDCG measure is often used to evaluate the performance
of search engines since in such applications it makes sense completely to ignore
elements that are not at the top of the ranking.
Once we have a hypothesis class and a ranking loss function, we can learn a
ranking function using the ERM rule. However, from the computational point of
view, the resulting optimization problem might be hard to solve. We next discuss
how to learn linear predictors for ranking.

17.4.1 Linear Predictors for Ranking

A natural way to define a ranking function is by projecting the instances onto some
vector w and then outputting the resulting scalars as our representation of the ranking function. That is, assuming that X Rd , for every w Rd we define a ranking

17.4 Ranking

function
h w ((x1 , . . . , xr )) = (w, x1 , . . . , w, xr ).

(17.6)

As we discussed in Chapter 16, we can also apply a feature mapping that maps
instances into some feature space and then takes the inner products with w in the
feature space. For simplicity, we focus on the simpler form as in Equation (17.6).
Given some W Rd , we can now define the hypothesis class HW = {h w : w W }.
Once we have defined this hypothesis class, and have chosen a ranking loss function,
we can apply the ERM rule as follows: Given a training set, S = (x1 , y1 ), . . . , (xm , ym ),
R)ri , for some ri N, we should search w W that
where each (xi , yi ) is in (X

xi ), yi ). As in the case of binary classificaminimizes the empirical loss, m
i=1 (h w (
tion, for many loss functions this problem is computationally hard, and we therefore
turn to describe convex surrogate loss functions. We describe the surrogates for the
Kendall tau loss and for the NDCG loss.

A Hinge Loss for the Kendall Tau Loss Function:

We can think of the Kendall tau loss as an average of 01 losses for each pair. In
particular, for every (i , j ) we can rewrite
1[sign( y  y  )=sign( yi y j )] = 1[sign( yi y j )( y  y  )0] .
i

In our case, yi y j = w, xi x j . It follows that we can use the hinge loss upper
bound as follows:
2
3>
=

1[sign( yi y j )( y  y  )0] max 0, 1 sign yi y j w, xi x j .
i

Taking the average over the pairs we obtain the following surrogate convex loss for
the Kendall tau loss function:
(h w (x), y)

r
r1 

3>
=
2
2
max 0, 1 sign( yi y j ) w, xi x j .
r (r 1)
i=1 j =i+1

The right-hand side is convex with respect to w and upper bounds the Kendall tau
loss. It is also a -Lipschitz function with parameter maxi, j xi x j .

A Hinge Loss for the NDCG Loss Function:

The NDCG loss function depends on the predicted ranking vector y Rr via the
permutation it induces. To derive a surrogate loss function we first make the following observation. Let V be the set of all permutations of [r ] encoded as vectors;
namely, each v V is a vector in [r ]r such that for all i = j we have vi = v j . Then
(see Exercise 17.4),
(y ) = argmax
vV

r

i=1

vi yi .

(17.7)

203

204

Multiclass and Ranking

Let us denote (x, v) =

r

i=1 vi xi ;

it follows that

(h w (x)) = argmax
vV

r


vi w, xi 

i=1

= argmax w,
vV

r


"
vi xi

i=1

= argmaxw, (x, v).

vV

On the basis of this observation, we can use the generalized hinge loss for costsensitive multiclass classification as a surrogate loss function for the NDCG loss as
follows:
(h w (x), y) (h w (x), y) + w, (x, (h w (x))) w, (x , (y))

2
3 2
3
max (v, y) + w, (x, v) w, (x, (y))
vV 

r

(vi (y)i ) w, xi  .
= max (v, y) +
vV

(17.8)

i=1

The right-hand side is a convex function with respect to w.

We can now solve the learning problem using SGD as described in
Section 17.2.5. The main computational bottleneck is calculating a subgradient of
the loss function, which is equivalent to finding v that achieves the maximum in
Equation (17.8) (see Claim 14.6). Using the definition of the NDCG loss, this is
equivalent to solving the problem

argmin
vV

r


(i vi + i D(vi )),

i=1

where i = w, xi  and i = yi /G(y, y). We can think of this problem a little bit
differently by defining a matrix A Rr,r where
Ai, j = j i + D( j ) i .
Now, let us think about each j as a worker, each i as a task, and Ai, j as the cost
of assigning task i to worker j . With this view, the problem of finding v becomes
the problem of finding an assignment of the tasks to workers of minimal cost. This
problem is called the assignment problem and can be solved efficiently. One particular algorithm is the Hungarian method (Kuhn 1955). Another way to solve
the assignment problem is using linear programming. To do so, let us first write the

17.4 Ranking

assignment problem as
argmin

r


BRr,r
+ i, j =1

(17.9)

Ai, j Bi, j

s.t. i [r ],

r


Bi, j = 1

j =1

j [r ],

r


Bi, j = 1

i=1

i , j , Bi, j {0, 1}
A matrix B that satisfies the constraints in the preceding optimization problem is
called a permutation matrix. This is because the constraints guarantee that there is
at most a single entry of each row that equals 1 and a single entry of each column
that equals 1. Therefore, the matrix B corresponds to the permutation v V defined
by vi = j for the single index j that satisfies Bi, j = 1.
The preceding optimization is still not a linear program because of the combinatorial constraint Bi, j {0, 1}. However, as it turns out, this constraint is redundant
if we solve the optimization problem while simply omitting the combinatorial
constraint, then we are still guaranteed that there is an optimal solution that will
satisfy this constraint.

This is formalized later.
Denote A, B = i, j Ai, j Bi, j . Then, Equation (17.9) is the problem of minimizing A, B such that B is a permutation matrix.
A matrix B Rr,r is called doubly stochastic if all elements of B are nonnegative,
the sum of each row of B is 1, and the sum of each column of B is 1. Therefore,
solving Equation (17.9) without the constraints Bi, j {0, 1} is the problem
argminA, B s.t. B is a doubly stochastic matrix.

(17.10)

BRr,r

The following claim states that every doubly stochastic matrix is a convex
combination of permutation matrices.
Claim 17.3 (Birkhoff 1946, Von Neumann 1953). The set of doubly stochastic
matrices in Rr,r is the convex hull of the set of permutation matrices in Rr,r .
On the basis of the claim, we easily obtain the following:
Lemma 17.4. There exists an optimal solution of Equation (17.10) that is also an
optimal solution of Equation (17.9).
Proof.
can write

Let B be a solution of Equation (17.10). Then, by Claim 17.3, we

B = i i Ci , where each Ci is a permutation matrix, each i > 0, and i i = 1.
Since all the Ci are also doubly stochastic, we clearly have that A, B A, Ci  for
every i . We claim that there is some i for which A, B = A, Ci . This must be true
since otherwise, if for every i A, B < A, Ci , we would have that
"
!



i C i =
i A, Ci  >
i A, B = A, B,
A, B = A,
i

205

206

Multiclass and Ranking

which cannot hold. We have thus shown that some permutation matrix, Ci , satisfies A, B = A, Ci . But, since for every other permutation matrix C we have
A, B A, C we conclude that Ci is an optimal solution of both Equation (17.9)
and Equation (17.10).

17.5 BIPARTITE RANKING AND MULTIVARIATE

PERFORMANCE MEASURES
In the previous section we described the problem of ranking. We used a vector
y Rr for representing an order over the elements x1 , . . . , xr . If all elements in y
are different from each other, then y specifies a full order over [r ]. However, if two
elements of y attain the same value, yi = y j for i = j , then y can only specify a partial
order over [r ]. In such a case, we say that xi and x j are of equal relevance according
to y. In the extreme case, y {1}r , which means that each xi is either relevant
or nonrelevant. This setting is often called bipartite ranking. For example, in the
fraud detection application mentioned in the previous section, each transaction is
labeled as either fraudulent (yi = 1) or benign (yi = 1).
Seemingly, we can solve the bipartite ranking problem by learning a binary classifier, applying it on each instance, and putting the positive ones at the top of the
ranked list. However, this may lead to poor results as the goal of a binary learner
is usually to minimize the zero-one loss (or some surrogate of it), while the goal of
a ranker might be significantly different. To illustrate this, consider again the problem of fraud detection. Usually, most of the transactions are benign (say 99. 9%).
Therefore, a binary classifier that predicts benign on all transactions will have a
zero-one error of 0. 1%. While this is a very small number, the resulting predictor is
meaningless for the fraud detection application. The crux of the problem stems from
the inadequacy of the zero-one loss for what we are really interested in. A more adequate performance measure should take into account the predictions over the entire
set of instances. For example, in the previous section we have defined the NDCG
loss, which emphasizes the correctness of the top-ranked items. In this section we
describe additional loss functions that are specifically adequate for bipartite ranking
problems.
As in the previous section, we are given a sequence of instances, x = (x1 , . . . , xr ),
and we predict a ranking vector y Rr . The feedback vector is y {1}r . We define
a loss that depends on y and y and depends on a threshold R. This threshold
transforms the vector y Rr into the vector (sign(yi ), . . . , sign(yr )) {1}r .
Usually, the value of is set to be 0. However, as we will see, we sometimes set
while taking into account additional constraints on the problem.
The loss functions we define in the following depend on the following 4 numbers:
True positives: a = |{i : yi = +1 sign(yi ) = +1}|
False positives: b = |{i : yi = 1 sign(yi ) = +1}|
False negatives: c = |{i : yi = +1 sign(yi ) = 1}|

(17.11)

True negatives: d = |{i : yi = 1 sign(yi ) = 1}|

y

The recall (a.k.a. sensitivity) of a prediction vector is the fraction of true positives
a
catches, namely, a+c
. The precision is the fraction of correct predictions among

17.5 Bipartite Ranking and Multivariate Performance Measures

a
the positive labels we predict, namely, a+b
. The specificity is the fraction of true
d
negatives that our predictor catches, namely, d+b
.
Note that as we decrease the recall increases (attaining the value 1 when =
). On the other hand, the precision and the specificity usually decrease as we
decrease . Therefore, there is a tradeoff between precision and recall, and we can
control it by changing . The loss functions defined in the following use various
techniques for combining both the precision and recall.

Averaging sensitivity and specificity:

Thismeasure is the average of the sensitiv

a
d
+ d+b
. This is also the accuracy on positive
ity and specificity, namely, 12 a+c
examples averaged with the accuracy on negative examples.
Here,

 we set = 0
1
a
d

and the corresponding loss function is (y , y) = 1 2 a+c + d+b .

F1 -score: The F1 score is the harmonic mean of the precision and recall:
2
. Its maximal value (of 1) is obtained when both precision and recall
1
1
Precision + Recall

are 1, and its minimal value (of 0) is obtained whenever one of them is 0 (even
if the other one is 1). The F1 score can be written using the numbers a, b, c
2a
. Again, we set = 0, and the loss function becomes
as follows; F1 = 2a+b+c

(y , y) = 1 F1 .

F -score: It is like F1 score, but we attach 2 times more importance to
recall than to precision, that is,
(1+ 2 )a
(1+ 2 )a+b+ 2 c

1+ 2

1
2 1
Precision + Recall

. It can also be written as F =

. Again, we set = 0, and the loss function becomes (y , y) =

1 F .

Recall at k: We measure the recall while the prediction must contain at most k
positive labels. That is, we should set so that a + b k. This is convenient, for
example, in the application of a fraud detection system, where a bank employee
can only handle a small number of suspicious transactions.

Precision at k: We measure the precision while the prediction must contain at
least k positive labels. That is, we should set so that a + b k.
The measures defined previously are often referred to as multivariate performance measures. Note that these measures are highly different from the average
b+d
. In the aforemenzero-one loss, which in the preceding notation equals a+b+c+d
tioned example of fraud detection, when 99. 9% of the examples are negatively
labeled, the zero-one loss of predicting that all the examples are negatives is 0. 1%.
In contrast, the recall of such prediction is 0 and hence the F1 score is also 0, which
means that the corresponding loss will be 1.

17.5.1 Linear Predictors for Bipartite Ranking

We next describe how to train linear predictors for bipartite ranking. As in the
previous section, a linear predictor for ranking is defined to be
h w (x) = (w, x1 , . . . , w, xr ).
The corresponding loss function is one of the multivariate performance measures
described before. The loss function depends on y = h w (x) via the binary vector it

207

208

Multiclass and Ranking

induces, which we denote by

b(y ) = (sign( y1 ), . . . , sign( yr )) {1}r .

(17.12)

As in the previous section, to facilitate an efficient algorithm we derive a convex surrogate loss function on . The derivation is similar to the derivation of the
generalized hinge loss for the NDCG ranking loss, as described in the previous
section.
Our first observation is that for all the values of defined before, there is some
V {1}r such that b(y ) can be rewritten as
b(y ) = argmax
v V

r


vi yi .

(17.13)

i=1

This is clearly true for the case = 0 if we choose V = {1}r . The two measures for
which is not taken to be 0 are precision at k and recall at k. For precision at k we
can take V to be the set Vk , containing all vectors in {1}r whose number of ones
is at least k. For recall at k, we can take V to be Vk , which is defined analogously.
See Exercise 17.5.
Once we have defined b as in Equation (17.13), we can easily derive a convex
surrogate loss as follows. Assuming that y V , we have that
(h w (x), y) = (b(h w (x)), y)
(b(h w (x)), y) +

max
vV

(v, y) +

r

i=1

r


(bi (h w (x)) yi )w, xi 



(vi yi ) w, xi  .

(17.14)

i=1

The right-hand side is a convex function with respect to w.

We can now solve the learning problem using SGD as described in
Section 17.2.5. The main computational bottleneck is calculating a subgradient of
the loss function, which is equivalent to finding v that achieves the maximum in
Equation (17.14) (see Claim 14.6).
In the following we describe how to find this maximizer efficiently for any performance measure that can be written as a function of the numbers a, b, c, d given
in Equation (17.11), and for which the set V contains all elements in {1}r for which
the values of a, b satisfy some constraints. For example, for recall at k the set V is
all vectors for which a + b k.
The idea is as follows. For any a, b [r ], let
Y a,b = {v : |{i : vi = 1 yi = 1}| = a |{i : vi = 1 yi = 1}| = b }.
Any vector v V falls into Y a,b for some a, b [r ]. Furthermore, if Y a,b V is not
empty for some a, b [r ] then Y a,b V = Ya,b . Therefore, we can search within each
Y a,b that has a nonempty intersection with V separately, and then take the optimal
value. The key observation is that once we are searching only within Y a,b , the value

17.6 Summary

of  is fixed so we only need to maximize the expression

max

vY a,b

r


vi w, xi .

i=1

Suppose the examples are sorted so that w, x1  w, xr . Then, it is easy to
verify that we would like to set vi to be positive for the smallest indices i . Doing
this, with the constraint on a, b, amounts to setting vi = 1 for the a top ranked positive examples and for the b top-ranked negative examples. This yields the following
procedure.
Solving Equation (17.14)
input:
(x1 , . . . , xr ), (y1 , . . . , yr ), w, V , 
assumptions:
 is a function of a, b, c, d
V contains all vectors for which f (a, b) = 1 for some function f
initialize:
P = |{i : yi = 1}|, N = |{i : yi = 1}|
= (w, x1 , . . . , w, xr ),  =
sort examples so that 1 2 r
let i 1 , . . . ,i P be the (sorted) indices of the positive examples
let j1 , . . . , j N be the (sorted) indices of the negative examples
for a = 0, 1, . . . , P
c = P a
for b = 0, 1, . . . , N such that f (a, b) = 1
d = N b
calculate  using a, b, c, d
set v1 , . . . , vr s.t. vi1 = = via = v j1 = = v jb = 1
and the rest

of the elements of v equal 1
set =  + ri=1 vi i
if 
 = , v = v
output v

17.6 SUMMARY
Many real world supervised learning problems can be cast as learning a multiclass
predictor. We started the chapter by introducing reductions of multiclass learning
to binary learning. We then described and analyzed the family of linear predictors
for multiclass learning. We have shown how this family can be used even if the
number of classes is extremely large, as long as we have an adequate structure on
the problem. Finally, we have described ranking problems. In Chapter 29 we study
the sample complexity of multiclass learning in more detail.

209

210

Multiclass and Ranking

17.7 BIBLIOGRAPHIC REMARKS

The One-versus-All and All-Pairs approach reductions have been unified under the
framework of Error Correction Output Codes (ECOC) (Dietterich & Bakiri 1995,
Allwein, Schapire & Singer 2000). There are also other types of reductions such
as tree-based classifiers (see, for example, Beygelzimer, Langford & Ravikumar
(2007)). The limitations of reduction techniques have been studied in (Daniely et
al. 2011, Daniely et al. 2012). See also Chapter 29, in which we analyze the sample
complexity of multiclass learning.
Direct approaches to multiclass learning with linear predictors have been studied
in (Vapnik 1998, Weston & Watkins 1999, Crammer & Singer 2001). In particular,
the multivector construction is due to Crammer and Singer (2001).
Collins (2000) has shown how to apply the Perceptron algorithm for structured
output problems. See also Collins (2002). A related approach is discriminative
learning of conditional random fields; see Lafferty et al. (2001). Structured output SVM has been studied in (Collins 2002, Taskar et al. 2003, Tsochantaridis et al.
2004).
The dynamic procedure we have presented for calculating the prediction h w (x)
in the structured output section is similar to the forward-backward variables
calculated by the Viterbi procedure in HMMs (see, for instance, (Rabiner &
Juang 1986)). More generally, solving the maximization problem in structured output is closely related to the problem of inference in graphical models (see, for
example, Koller & Friedman (2009a)).
Chapelle, Le, and Smola (2007) proposed to learn a ranking function with
respect to the NDCG loss using ideas from structured output learning. They also
observed that the maximization problem in the definition of the generalized hinge
loss is equivalent to the assignment problem.
Agarwal and Roth (2005) analyzed the sample complexity of bipartite ranking. Joachims (2005) studied the applicability of structured output SVM to bipartite
ranking with multivariate performance measures.

17.8 EXERCISES
17.1 Consider a set S of examples in Rn [k] for which there exist vectors 1 , . . ., k
such that every example (x, y) S falls within a ball centered at y whose radius
is r 1. Assume also that for every i = j , i j  4r . Consider concatenating
each instance by the constant 1 and then applying the multivector construction,
namely,
(x, y) = [ 0, . . . , 0 , x1 , . . ., xn , 1, 0, . . ., 0 ].
9 :; < 9
:;
< 9 :; <
R(y1)(n+1)

Rn+1

R(ky)(n+1)

Show that there exists a vector w Rk(n+1) such that
(w, (x, y)) = 0 for every
(x, y) S.
Hint: Observe that for every example (x, y) S we can write x = y + v for some
v r . Now, take w = [w1 , . . ., wk ], where wi = [i , i 2 /2].

17.8 Exercises

17.2 Multiclass Perceptron: Consider the following algorithm:

Multiclass Batch Perceptron

Input:
A training set (x1 , y1 ), . . ., (xm , ym )
A class-sensitive feature mapping  : X Y Rd
Initialize: w(1) = (0, . . ., 0) Rd
For t = 1, 2, . . .
If ( i and y = yi s.t. w(t) , (xi , yi ) w(t) , (xi , y)) then
w(t+1) = w(t) + (xi , yi ) (xi , y)
else
output w(t)
Prove the following:
Theorem 17.5. Assume that there exists w such that for all i and for all y = yi it
holds that w , (xi , yi ) w , (xi , y) + 1. Let R = maxi,y (xi , yi ) (xi , y).
Then, the multiclass Perceptron algorithm stops after at most (Rw )2 iterations,
and when it stops it holds that i [m], yi = argmax y w(t) , (xi , y).
17.3 Generalize the dynamic programming procedure given in Section 17.3 for solving the maximization problem given in the definition of h

in the SGD procedure
for multiclass prediction. You can assume that (y , y) = rt=1 (yt , yt ) for some
arbitrary function .
17.4 Prove that Equation (17.7) holds.
17.5 Show that the two definitions of as defined in Equation (17.12) and
Equation (17.13) are indeed equivalent for all the multivariate performance
measures.

211

18
Decision Trees

A decision tree is a predictor, h : X Y, that predicts the label associated with

an instance x by traveling from a root node of a tree to a leaf. For simplicity we
focus on the binary classification setting, namely, Y = {0, 1}, but decision trees can
be applied for other prediction problems as well. At each node on the root-to-leaf
path, the successor child is chosen on the basis of a splitting of the input space.
Usually, the splitting is based on one of the features of x or on a predefined set of
splitting rules. A leaf contains a specific label. An example of a decision tree for the
papayas example (described in Chapter 2) is given in the following:
Color?
Pale green to pale yellow
Other

Not-tasty

Softness?
Other

Not-tasty

Gives slightly to palm pressure

Tasty

To check if a given papaya is tasty or not, the decision tree first examines the
color of the Papaya. If this color is not in the range pale green to pale yellow, then
the tree immediately predicts that the papaya is not tasty without additional tests.
Otherwise, the tree turns to examine the softness of the papaya. If the softness level
of the papaya is such that it gives slightly to palm pressure, the decision tree predicts
that the papaya is tasty. Otherwise, the prediction is not-tasty. The preceding
example underscores one of the main advantages of decision trees the resulting
classifier is very simple to understand and interpret.

212

18.1 Sample Complexity

18.1 SAMPLE COMPLEXITY

A popular splitting rule at internal nodes of the tree is based on thresholding
the value of a single feature. That is, we move to the right or left child of the
node on the basis of 1[xi <] , where i [d] is the index of the relevant feature
and R is the threshold. In such cases, we can think of a decision tree as a
splitting of the instance space, X = Rd , into cells, where each leaf of the tree
corresponds to one cell. It follows that a tree with k leaves can shatter a set
of k instances. Hence, if we allow decision trees of arbitrary size, we obtain a
hypothesis class of infinite VC dimension. Such an approach can easily lead to
overfitting.
To avoid overfitting, we can rely on the minimum description length (MDL)
principle described in Chapter 7, and aim at learning a decision tree that on one
hand fits the data well while on the other hand is not too large.
For simplicity, we will assume that X = {0, 1}d . In other words, each instance is a
vector of d bits. In that case, thresholding the value of a single feature corresponds
to a splitting rule of the form 1[xi =1] for some i = [d]. For instance, we can model
the papaya decision tree earlier by assuming that a papaya is parameterized by a
two-dimensional bit vector x {0, 1}2 , where the bit x 1 represents whether the color
is pale green to pale yellow or not, and the bit x 2 represents whether the softness is
gives slightly to palm pressure or not. With this representation, the node Color? can
be replaced with 1[x1 =1] , and the node Softness? can be replaced with 1[x2 =1] . While
this is a big simplification, the algorithms and analysis we provide in the following
can be extended to more general cases.
With the aforementioned simplifying assumption, the hypothesis class becomes
finite, but is still very large. In particular, any classifier from {0, 1}d to {0, 1} can be
represented by a decision tree with 2d leaves and depth of d + 1 (see Exercise 18.1).
Therefore, the VC dimension of the class is 2d , which means that the number of
examples we need to PAC learn the hypothesis class grows with 2d . Unless d is very
small, this is a huge number of examples.
To overcome this obstacle, we rely on the MDL scheme described in Chapter 7.
The underlying prior knowledge is that we should prefer smaller trees over larger
trees. To formalize this intuition, we first need to define a description language for
decision trees, which is prefix free and requires fewer bits for smaller decision trees.
Here is one possible way: A tree with n nodes will be described in n + 1 blocks, each
of size log2 (d + 3) bits. The first n blocks encode the nodes of the tree, in a depthfirst order (preorder), and the last block marks the end of the code. Each block
indicates whether the current node is:

An internal node of the form 1[xi =1] for some i [d]

A leaf whose value is 1
A leaf whose value is 0
End of the code

Overall, there are d + 3 options, hence we need log2 (d + 3) bits to describe each
block.

213

214

Decision Trees

Assuming each internal node has two children,1 it is not hard to show that this
is a prefix-free encoding of the tree, and that the description length of a tree with n
nodes is (n + 1) log2 (d + 3).
By Theorem 7.7 we have that with probability of at least 1 over a sample of
size m, for every n and every decision tree h H with n nodes it holds that

(n + 1) log2 (d + 3) + log (2/)
L D (h) L S (h) +
.
(18.1)
2m
This bound performs a tradeoff: on the one hand, we expect larger, more complex
decision trees to have a smaller training risk, L S (h), but the respective value of n
will be larger. On the other hand, smaller decision trees will have a smaller value of
n, but L S (h) might be larger. Our hope (or prior knowledge) is that we can find a
decision tree with both low empirical risk, L S (h), and a number of nodes n not too
high. Our bound indicates that such a tree will have low true risk, L D (h).

18.2 DECISION TREE ALGORITHMS

The bound on L D (h) given in Equation (18.1) suggests a learning rule for decision trees search for a tree that minimizes the right-hand side of Equation (18.1).
Unfortunately, it turns out that solving this problem is computationally hard.2 Consequently, practical decision tree learning algorithms are based on heuristics such
as a greedy approach, where the tree is constructed gradually, and locally optimal
decisions are made at the construction of each node. Such algorithms cannot guarantee to return the globally optimal decision tree but tend to work reasonably well
in practice.
A general framework for growing a decision tree is as follows. We start with a
tree with a single leaf (the root) and assign this leaf a label according to a majority
vote among all labels over the training set. We now perform a series of iterations.
On each iteration, we examine the effect of splitting a single leaf. We define some
gain measure that quantifies the improvement due to this split. Then, among all
possible splits, we either choose the one that maximizes the gain and perform it, or
choose not to split the leaf at all.
In the following we provide a possible implementation. It is based on a popular decision tree algorithm known as ID3 (short for Iterative Dichotomizer 3).
We describe the algorithm for the case of binary features, namely, X = {0, 1}d , and
therefore all splitting rules are of the form 1[xi =1] for some feature i [d]. We discuss
the case of real valued features in Section 18.2.3.
The algorithm works by recursive calls, with the initial call being ID3(S, [d]), and
returns a decision tree. In the pseudocode that follows, we use a call to a procedure
Gain(S,i ), which receives a training set S and an index i and evaluates the gain of a
split of the tree according to the i th feature. We describe several gain measures in
Section 18.2.1.
1
2

We may assume this without loss of generality, because if a decision node has only one child, we can
replace the node by its child without affecting the predictions of the decision tree.
More precisely, if NP=P then no algorithm can solve Equation (18.1) in time polynomial in n, d, and m.

18.2 Decision Tree Algorithms

ID3(S, A)
INPUT: training set S, feature subset A [d]
if all examples in S are labeled by 1, return a leaf 1
if all examples in S are labeled by 0, return a leaf 0
if A = , return a leaf whose value = majority of labels in S
else :
Let j = argmaxi A Gain(S,i )
if all examples in S have the same label
Return a leaf whose value = majority of labels in S
else
Let T1 be the tree returned by ID3({(x, y) S : x j = 1}, A \ { j }).
Let T2 be the tree returned by ID3({(x, y) S : x j = 0}, A \ { j }).
Return the tree:
xj = 1?

T2

T1

18.2.1 Implementations of the Gain Measure

Different algorithms use different implementations of Gain(S,i ). Here we present
three. We use the notation P S [F] to denote the probability that an event holds with
respect to the uniform distribution over S.
Train Error: The simplest definition of gain is the decrease in training error.
Formally, let C(a) = min{a, 1 a}. Note that the training error before splitting on
feature i is C( P S [y = 1]), since we took a majority vote among labels. Similarly, the
error after splitting on feature i is
P [x i = 1] C( P [y = 1|x i = 1]) + P [x i = 0]C( P [y = 1|x i = 0]).
S

Therefore, we can define Gain to be the difference between the two, namely,
Gain(S,i ) := C( P [y = 1])
S


P [x i = 1] C( P [y = 1|x i = 1]) + P [x i = 0]C( P [y = 1|x i = 0]) .
S

Information Gain: Another popular gain measure that is used in the ID3 and
C4.5 algorithms of Quinlan (1993) is the information gain. The information gain
is the difference between the entropy of the label before and after the split, and
is achieved by replacing the function C in the previous expression by the entropy
function,
C(a) = a log (a) (1 a) log (1 a).

215

216

Decision Trees

Gini Index: Yet another definition of a gain, which is used by the CART
algorithm of Breiman, Friedman, Olshen, and Stone (1984), is the Gini index,
C(a) = 2a(1 a).
Both the information gain and the Gini index are smooth and concave upper bounds
of the train error. These properties can be advantageous in some situations (see,
for example, Kearns & Mansour (1996)).

18.2.2 Pruning
The ID3 algorithm described previously still suffers from a big problem: The
returned tree will usually be very large. Such trees may have low empirical risk,
but their true risk will tend to be high both according to our theoretical analysis,
and in practice. One solution is to limit the number of iterations of ID3, leading
to a tree with a bounded number of nodes. Another common solution is to prune
the tree after it is built, hoping to reduce it to a much smaller tree, but still with a
similar empirical error. Theoretically, according to the bound in Equation (18.1), if
we can make n much smaller without increasing L S (h) by much, we are likely to get
a decision tree with a smaller true risk.
Usually, the pruning is performed by a bottom-up walk on the tree. Each node
might be replaced with one of its subtrees or with a leaf, based on some bound or
estimate of L D (h) (for example, the bound in Equation (18.1)). A pseudocode of a
common template is given in the following.
Generic Tree Pruning Procedure
input:
function f (T , m) (bound/estimate for the generalization error
of a decision tree T , based on a sample of size m),
tree T .
foreach node j in a bottom-up walk on T (from leaves to root):
find T  which minimizes f (T  , m), where T  is any of the following:
the current tree after replacing node j with a leaf 1.
the current tree after replacing node j with a leaf 0.
the current tree after replacing node j with its left subtree.
the current tree after replacing node j with its right subtree.
the current tree.
let T := T  .

18.2.3 Threshold-Based Splitting Rules for Real-Valued Features

In the previous section we have described an algorithm for growing a decision tree
assuming that the features are binary and the splitting rules are of the form 1[xi =1] .
We now extend this result to the case of real-valued features and threshold-based
splitting rules, namely, 1[xi <] . Such splitting rules yield decision stumps, and we
have studied them in Chapter 10.

18.4 Summary

The basic idea is to reduce the problem to the case of binary features as follows.
Let x1 , . . . , xm be the instances of the training set. For each real-valued feature i ,
sort the instances so that x 1,i x m,i . Define a set of thresholds 0,i , . . . , m+1,i
such that j ,i (x j ,i , x j +1,i ) (where we use the convention x 0,i = and x m+1,i =
). Finally, for each i and j we define the binary feature 1[xi < j ,i ] . Once we have
constructed these binary features, we can run the ID3 procedure described in the
previous section. It is easy to verify that for any decision tree with threshold-based
splitting rules over the original real-valued features there exists a decision tree over
the constructed binary features with the same training error and the same number
of nodes.
If the original number of real-valued features is d and the number of examples
is m, then the number of constructed binary features becomes dm. Calculating the
Gain of each feature might therefore take O(dm 2 ) operations. However, using a
more clever implementation, the runtime can be reduced to O(dm log (m)). The
idea is similar to the implementation of ERM for decision stumps as described in
Section 10.1.1.

18.3 RANDOM FORESTS

As mentioned before, the class of decision trees of arbitrary size has infinite VC
dimension. We therefore restricted the size of the decision tree. Another way to
reduce the danger of overfitting is by constructing an ensemble of trees. In particular, in the following we describe the method of random forests, introduced by
Breiman (2001).
A random forest is a classifier consisting of a collection of decision trees, where
each tree is constructed by applying an algorithm A on the training set S and an
additional random vector, , where is sampled i.i.d. from some distribution. The
prediction of the random forest is obtained by a majority vote over the predictions
of the individual trees.
To specify a particular random forest, we need to define the algorithm A and
the distribution over . There are many ways to do this and here we describe one
particular option. We generate as follows. First, we take a random subsample
from S with replacements; namely, we sample a new training set S  of size m  using
the uniform distribution over S. Second, we construct a sequence I1 , I2 , . . ., where
each It is a subset of [d] of size k, which is generated by sampling uniformly at
random elements from [d]. All these random variables form the vector . Then,
the algorithm A grows a decision tree (e.g., using the ID3 algorithm) based on the
sample S  , where at each splitting stage of the algorithm, the algorithm is restricted
to choosing a feature that maximizes Gain from the set It . Intuitively, if k is small,
this restriction may prevent overfitting.

18.4 SUMMARY
Decision trees are very intuitive predictors. Typically, if a human programmer
creates a predictor it will look like a decision tree. We have shown that the VC
dimension of decision trees with k leaves is k and proposed the MDL paradigm for

217

218

Decision Trees

learning decision trees. The main problem with decision trees is that they are computationally hard to learn; therefore we described several heuristic procedures for
training them.

18.5 BIBLIOGRAPHIC REMARKS

Many algorithms for learning decision trees (such as ID3 and C4.5) have been
derived by Quinlan (1986). The CART algorithm is due to Breiman, Friedman,
Olshen & Stone (1984). Random forests were introduced by Breiman (2001). For
additional reading we refer the reader to (Hastie, Tibshirani & Friedman 2001,
Rokach 2007).
The proof of the hardness of training decision trees is given in Hyafil and Rivest
(1976).

18.6 EXERCISES
18.1 1. Show that any binary classifier h : {0, 1}d  {0, 1} can be implemented as a decision tree of height at most d + 1, with internal nodes of the form (xi = 0?) for
some i {1, . . ., d}.
2. Conclude that the VC dimension of the class of decision trees over the domain
{0, 1}d is 2d .
18.2 (Suboptimality of ID3)
Consider the following training set, where X = {0, 1}3 and Y = {0, 1}:
((1, 1, 1), 1)
((1, 0, 0), 1)
((1, 1, 0), 0)
((0, 0, 1), 0)
Suppose we wish to use this training set in order to build a decision tree of depth
2 (i.e., for each input we are allowed to ask two questions of the form (xi = 0?)
before deciding on the label).
1. Suppose we run the ID3 algorithm up to depth 2 (namely, we pick the root
node and its children according to the algorithm, but instead of keeping on
with the recursion, we stop and pick leaves according to the majority label in
each subtree). Assume that the subroutine used to measure the quality of each
feature is based on the entropy function (so we measure the information gain),
and that if two features get the same score, one of them is picked arbitrarily.
Show that the training error of the resulting decision tree is at least 1/4.
2. Find a decision tree of depth 2 that attains zero training error.

19
Nearest Neighbor

Nearest Neighbor algorithms are among the simplest of all machine learning algorithms. The idea is to memorize the training set and then to predict the label of
any new instance on the basis of the labels of its closest neighbors in the training
set. The rationale behind such a method is based on the assumption that the features that are used to describe the domain points are relevant to their labelings in a
way that makes close-by points likely to have the same label. Furthermore, in some
situations, even when the training set is immense, finding a nearest neighbor can
be done extremely fast (for example, when the training set is the entire Web and
distances are based on links).
Note that, in contrast with the algorithmic paradigms that we have discussed
so far, like ERM, SRM, MDL, or RLM, that are determined by some hypothesis
class, H, the Nearest Neighbor method figures out a label on any test point without
searching for a predictor within some predefined class of functions.
In this chapter we describe Nearest Neighbor methods for classification and
regression problems. We analyze their performance for the simple case of binary
classification and discuss the efficiency of implementing these methods.

19.1 k NEAREST NEIGHBORS

Throughout the entire chapter we assume that our instance domain, X , is endowed
with a metric function . That is, : X X R is a function that returns the distance
between any two elements of X
For example, if X = Rd then can be the Euclidean
.

d
 2
distance, (x, x ) = x x  =
i=1 (x i x i ) .
Let S = (x1 , y1 ), . . . , (xm , ym ) be a sequence of training examples. For each x X ,
let 1 (x), . . . , m (x) be a reordering of {1, . . . , m} according to their distance to x,
(x, xi ). That is, for all i < m,
(x, xi (x) ) (x, xi+1 (x) ).

219

220

Nearest Neighbor

Figure 19.1. An illustration of the decision boundaries of the 1-NN rule. The points
depicted are the sample points, and the predicted label of any new point will be the
label of the sample point in the center of the cell it belongs to. These cells are called a
Voronoi Tessellation of the space.

For a number k, the k-NN rule for binary classification is defined as follows:
k-NN
input: a training sample S = (x1 , y1 ), . . . , (xm , ym )
output: for every point x X ,
return the majority label among {yi (x) : i k}
When k = 1, we have the 1-NN rule:
h S (x) = y1 (x) .
A geometric illustration of the 1-NN rule is given in Figure 19.1.
For regression problems, namely, Y = R, one can define the

prediction to be
the average target of the k nearest neighbors. That is, h S (x) = 1k ki=1 yi (x) . More
generally, for some function : (X Y)k Y, the k-NN rule with respect to is:


(19.1)
h S (x) = (x1 (x) , y1 (x) ), . . . , (xk (x) , yk (x) ) .
It is easy to verify that we can cast the prediction by majority of labels (for classification) or by the averaged target (for regression) as in Equation (19.1) by an
appropriate choice of . The generality can lead to other rules; for example, if Y = R,
we can take a weighted average of the targets according to the distance from x:
h S (x) =

k

i=1

(x, xi (x) )
yi (x) .

k
j =1 (x, x j (x) )

19.2 ANALYSIS
Since the NN rules are such natural learning methods, their generalization properties have been extensively studied. Most previous results are asymptotic consistency
results, analyzing the performance of NN rules when the sample size, m, goes to
infinity, and the rate of convergence depends on the underlying distribution. As we
have argued in Section 7.4, this type of analysis is not satisfactory. One would like to
learn from finite training samples and to understand the generalization performance
as a function of the size of such finite training sets and clear prior assumptions on
the data distribution. We therefore provide a finite-sample analysis of the 1-NN rule,

19.2 Analysis

showing how the error decreases as a function of m and how it depends on properties of the distribution. We will also explain how the analysis can be generalized to
k-NN rules for arbitrary values of k. In particular, the analysis specifies the number
of examples required to achieve a true error of 2L D (h  ) + , where h  is the Bayes
optimal hypothesis, assuming that the labeling rule is well behaved" (in a sense we
will define later).

19.2.1 A Generalization Bound for the 1-NN Rule

We now analyze the true error of the 1-NN rule for binary classification with the 0-1
loss, namely, Y = {0, 1} and
(h, (x, y)) = 1[h(x)= y] . We also assume throughout the
analysis that X = [0, 1]d and is the Euclidean distance.
We start by introducing some notation. Let D be a distribution over X Y. Let
DX denote the induced marginal distribution over X and let : Rd R be the
conditional probability1 over the labels, that is,
(x) = P [y = 1|x].
Recall that the Bayes optimal rule (that is, the hypothesis that minimizes L D (h) over
all functions) is
h  (x) = 1[(x)>1/2] .
We assume that the conditional probability function is c-Lipschitz for some
c > 0: Namely, for all x, x X , |(x) (x )| c x x . In other words, this
assumption means that if two vectors are close to each other then their labels are
likely to be the same.
The following lemma applies the Lipschitzness of the conditional probability
function to upper bound the true error of the 1-NN rule as a function of the expected
distance between each test instance and its nearest neighbor in the training set.
Lemma 19.1. Let X = [0, 1]d , Y = {0, 1}, and D be a distribution over X Y
for which the conditional probability function, , is a c-Lipschitz function. Let
S = (x1 , y1 ), . . . , (xm , ym ) be an i.i.d. sample and let h S be its corresponding 1-NN
hypothesis. Let h  be the Bayes optimal rule for . Then,
E [L D (h S )] 2 L D (h  ) + c

SD m

SD m ,xD

[x x1 (x) ].

Proof. Since L D (h S ) = E(x,y)D [1[h S (x)= y] ], we obtain that E S [L D (h S )] is the probability to sample a training set S and an additional example (x, y), such that the
label of 1 (x) is different from y. In other words, we can first sample m unlabeled
examples, Sx = (x1 , . . . , xm ), according to DX , and an additional unlabeled example,
x DX , then find 1 (x) to be the nearest neighbor of x in Sx , and finally sample

Formally, P [y = 1|x] = lim0

around x.

D({(x ,1):x B(x,)})

,
D({(x ,y):x B(x,),yY})

where B(x, ) is a ball of radius centered

221

222

Nearest Neighbor

y (x) and y1 (x) (1 (x)). It follows that

E [L D (h S )] =
S

m ,xD ,y(x),y  ( (x))

S x D X
X
1

[1[y= y  ] ]

E
m

Sx DX ,xDX

y(x),y  (1 (x))


[y = y  ] .

(19.2)

We next upper bound P y(x),y  (x ) [y = y  ] for any two domain points x, x :

P

y(x),y  (x )

[y = y  ] = (x )(1 (x)) + (1 (x))(x)

= ((x) (x) + (x ))(1 (x))
+ (1 (x) + (x) (x))(x)
= 2(x)(1 (x)) + ((x) (x))(2(x) 1).

Using |2(x) 1| 1 and the assumption that is c-Lipschitz, we obtain that the
probability is at most:
P

y(x),y  (x )

[y = y  ] 2(x)(1 (x)) + c x x.

Plugging this into Equation (19.2) we conclude that

E [L D (h S )] E [2(x)(1 (x))] + c E [x x1 (x) ].
x

S,x

Finally, the error of the Bayes optimal classifier is

L D (h  ) = E [ min{(x), 1 (x)}] E [(x)(1 (x))].
x

Combining the preceding two inequalities concludes our proof.

The next step is to bound the expected distance between a random x and its
closest element in S. We first need the following general probability lemma. The
lemma bounds the probability weight of subsets that are not hit by a random sample,
as a function of the size of that sample.
Lemma 19.2. Let C1 , . . . , Cr be a collection of subsets of some domain set, X . Let S
be a sequence of m points sampled i.i.d. according to some probability distribution, D
over X . Then,

E m

SD

r
.
me

P [Ci ]

i:Ci S=

Proof. From the linearity of expectation, we can rewrite:

r




E
P [Ci ] =
P [Ci ] E 1[Ci S=] .
S

i:Ci S=

i=1

Next, for each i we have



E 1[Ci S=] = P [Ci S = ] = (1 P[Ci ])m e P [Ci ] m .
S

19.2 Analysis

Combining the preceding two equations we get

r


P [Ci ]
P [Ci ] e P [Ci ] m r max P [Ci ] e P [Ci ] m .
E
S

i:Ci S=

i=1

Finally, by a standard calculus, maxa aema

1
me

and this concludes the proof.

Equipped with the preceding lemmas we are now ready to state and prove the
main result of this section an upper bound on the expected error of the 1-NN
learning rule.
Theorem 19.3. Let X = [0, 1]d , Y = {0, 1}, and D be a distribution over X Y for
which the conditional probability function, , is a c-Lipschitz function. Let h S denote
the result of applying the 1-NN rule to a sample S Dm . Then,
1
E m [L D (h S )] 2 L D (h  ) + 4 c d m d+1 .

SD

Proof. Fix some = 1/T , for some integer T , let r = T d and let C1 , . . . , Cr be the
cover of the set X using boxes of length : Namely, for every (1 , . . . , d ) [T ]d ,
there exists a set Ci of the form {x : j , x j [( j 1)/T , j /T ]}. An illustration for
d = 2, T = 5 and the set corresponding to = (2, 4) is given in the following.
1

For each x, x in the same box we have x x  d . Otherwise, x x  d.

Therefore,




Ci d + P
Ci d ,
E [x x1 (x) ] E P
x,S

i:Ci S=

i:Ci S=

and by combining Lemma 19.2 with the trivial bound P [

i:Ci S= Ci ] 1 we get that

r

E [x x1(x) ] d me
+ .

x,S

Since the number of boxes is r = (1/ )d we get that


 d d
E [x x1(x) ] d 2 m e + .
S,x

Combining the preceding with Lemma 19.1 we obtain that


 d d
E [L D (h S )] 2 L D (h  ) + c d 2 m e + .
S

223

224

Nearest Neighbor

Finally, setting = 2 m 1/(d+1) and noting that

2d d
2d 2d m d/(d+1)
+ =
+ 2 m 1/(d+1)
me
me
= m 1/(d+1)(1/e + 2) 4m 1/(d+1)
we conclude our proof.
The theorem implies that if we first fix the data-generating distribution and
then let m go to infinity, then the error of the 1-NN rule converges to twice the
Bayes error. The analysis can be generalized to larger
values of k, showing that
the expected error of the k-NN rule converges to (1 + 8/k) times the error of the
Bayes classifier. This is formalized in Theorem 19.5, whose proof is left as a guided
exercise.

19.2.2 The Curse of Dimensionality

The upper bound given in Theorem 19.3 grows with c (the Lipschitz coefficient of )
and with d, the Euclidean dimension of the domain set X . In fact, it is easy to see that
a necessary
condition for the last term in Theorem 19.3 to be smaller than is that
m (4 c d/ )d+1 . That is, the size of the training set should increase exponentially
with the dimension. The following theorem tells us that this is not just an artifact
of our upper bound, but, for some distributions, this amount of examples is indeed
necessary for learning with the NN rule.
Theorem 19.4. For any c > 1, and every learning rule, L, there exists a distribution
over [0, 1]d {0, 1}, such that (x) is c-Lipschitz, the Bayes error of the distribution is
0, but for sample sizes m (c + 1)d /2, the true error of the rule L is greater than 1/4.
Proof. Fix any values of c and d. Let G dc be the grid on [0, 1]d with distance of
1/c between points on the grid. That is, each point on the grid is of the form
(a1 /c, . . . , ad /c) where ai is in {0, . . . , c 1, c}. Note that, since any two distinct points
on this grid are at least 1/c apart, any function : G CD [0, 1] is a c-Lipschitz
function. It follows that the set of all c-Lipschitz functions over G dc contains the
set of all binary valued functions over that domain. We can therefore invoke the
No-Free-Lunch result (Theorem 5.1) to obtain a lower bound on the needed sample sizes for learning that class. The number of points on the grid is (c + 1)d ; hence,
if m < (c + 1)d /2, Theorem 5.1 implies the lower bound we are after.
The exponential dependence on the dimension is known as the curse of dimensionality. As we saw, the 1-NN rule might fail if the number of examples is smaller
than ((c + 1)d ). Therefore, while the 1-NN rule does not restrict itself to a predefined set of hypotheses, it still relies on some prior knowledge its success depends
on the assumption that the dimension and the Lipschitz constant of the underlying
distribution, , are not too high.

19.6 Exercises

19.3 EFFICIENT IMPLEMENTATION*

Nearest Neighbor is a learning-by-memorization type of rule. It requires the entire
training data set to be stored, and at test time, we need to scan the entire data set in
order to find the neighbors. The time of applying the NN rule is therefore (d m).
This leads to expensive computation at test time.
When d is small, several results from the field of computational geometry have
proposed data structures that enable to apply the NN rule in time o(d O(1) log (m)).
However, the space required by these data structures is roughly m O(d) , which makes
these methods impractical for larger values of d.
To overcome this problem, it was suggested to improve the search method by
allowing an approximate search. Formally, an r -approximate search procedure is
guaranteed to retrieve a point within distance of at most r times the distance to the
nearest neighbor. Three popular approximate algorithms for NN are the kd-tree,
balltrees, and locality-sensitive hashing (LSH). We refer the reader, for example, to
(Shakhnarovich, Darrell & Indyk 2006).

19.4 SUMMARY
The k-NN rule is a very simple learning algorithm that relies on the assumption
that things that look alike must be alike. We formalized this intuition using the
Lipschitzness of the conditional probability. We have shown that with a sufficiently
large training set, the risk of the 1-NN is upper bounded by twice the risk of the
Bayes optimal rule. We have also derived a lower bound that shows the curse of
dimensionality the required sample size might increase exponentially with the
dimension. As a result, NN is usually performed in practice after a dimensionality
reduction preprocessing step. We discuss dimensionality reduction techniques later
on in Chapter 23.

19.5 BIBLIOGRAPHIC REMARKS

Cover and Hart (1967) gave the first analysis of 1-NN, showing that its risk converges to twice the Bayes optimal error under mild conditions. Following a lemma
due to Stone (1977), Devroye and Gyrfi (1985) have shown that the k-NN rule is
consistent (with respect to the hypothesis class of all functions from Rd to {0, 1}).
A good presentation of the analysis is given in the book by Devroye et al. (1996).
Here, we give a finite sample guarantee that explicitly underscores the prior assumption on the distribution. See Section 7.4 for a discussion on consistency results.
Finally, Gottlieb, Kontorovich, and Krauthgamer (2010) derived another finite
sample bound for NN that is more similar to VC bounds.

19.6 EXERCISES
In this exercise we will prove the following theorem for the k-NN rule.
Theorem 19.5. Let X = [0, 1]d , Y = {0, 1}, and D be a distribution over X Y for
which the conditional probability function, , is a c-Lipschitz function. Let h S denote

225

226

Nearest Neighbor

the result of applying the k-NN rule to a sample S Dm , where k 10. Let h  be the
Bayes optimal hypothesis. Then,

 


8
E [L D (h S )] 1 +
L D (h  ) + 6 c d + k m 1/(d+1).
S
k
19.1 Prove the following lemma.
Lemma 19.6. Let C1 , . . ., Cr be a collection of subsets of some domain set, X . Let S
be a sequence of m points sampled i.i.d. according to some probability distribution,
D over X . Then, for every k 2,


2r k
P [Ci ]
E
.
SD m
m
i:|C S|<k
i

Hints:

Show that

E
S

P [Ci ] =

i:|Ci S|<k

r


P [Ci ] P [|Ci S| < k] .

i=1

Fix some i and suppose that k < P [Ci ] m/2. Use Chernoffs bound to show that
P [|Ci S| < k] P [|Ci S| < P [Ci ]m/2] e P [Ci ]m/8 .
S

Use the inequality maxa

aema

1
me

to show that for such i we have

P [Ci ] P [|Ci S| < k] P [Ci ]e P [Ci ]m/8

S

8
.
me

Conclude the proof by using the fact that for the case k P [Ci ] m/2 we clearly
have:
2k
P [Ci ] P [|Ci S| < k] P [Ci ] .
S
m
19.2 We use the notation y p as a shorthand for y is a Bernoulli random variable
with expected value p. Prove the following lemma:
Z k be independent
Lemma 19.7. Let k 10 and let Z 1 , . . .,


Bernoulli random
variables with P [Z i = 1] = pi . Denote p = 1k i pi and p = 1k ki=1 Z i . Show that

 
8
P [y = 1[ p >1/2] ] 1 +
P [y = 1[ p>1/2] ].
E
Z 1 ,...,Z k y p
k y p
Hints:
W.l.o.g. assume that p 1/2. Then, P y p [y = 1[ p>1/2] ] = p. Let y  = 1[ p >1/2] .

Show that
E

P [y = y  ] p =

Z 1 ,...,Z k y p

Z 1 ,...,Z k

[ p > 1/2](1 2 p).

Use Chernoffs bound (Lemma B.3) to show that

P [ p > 1/2] e



k p h 21p 1
,

where
h(a) = (1 + a) log (1 + a) a.

19.6 Exercises

To conclude the proof of the lemma, you can rely on the following inequality
(without proving it): For every p [0, 1/2] and k 10:

8
k p + k2 (log (2 p)+1)

(1 2 p) e
p.
k
19.3 Fix some p, p [0, 1] and y  {0, 1}. Show that
P [ y = y  ] P  [ y = y  ] + | p p |.

y p

y p

19.4 Conclude the proof of the theorem according to the following steps:

As in the proof of Theorem 19.3, six some > 0 and let C1 , . . ., Cr be the cover

of the set
X using boxes of length . For
each x, x in the same box we have


x x  d . Otherwise, x x  2 d. Show that


P [Ci ]
E [ L D (h S )] E
S

+ max
i

i:|Ci S|<k


P h S (x) = y | j [k], x x j (x)  d .

S,(x, y)

(19.3)

Bound the first summand using Lemma 19.6.

To bound the second summand, let us
fix S| x and x such that all the k neighbors
x. W.l.o.g assume that the k NN
of x in S| x are at distance of at most d from

are x1 , . . ., xk . Denote pi = (xi ) and let p = 1k i pi . Use Exercise 19.3 to show
that
P [h S (x) = y] E P [h S (x) = y] + | p (x)|.
E
y1 ,...,y j y(x)

y1 ,...,y j y p

W.l.o.g. assume that p 1/2. Now use Lemma 19.7 to show that

 
8
P
P [h S (x) = y] 1 +
P [1[ p>1/2] = y].
y1 ,...,y j y p
k y p

Show that
P [1[ p>1/2] = y] = p = min{ p, 1 p} min{(x), 1 (x)} + | p (x)|.

y p

Combine all the preceding to obtain that the second summand in

Equation (19.3) is bounded by

 

8
1+
L D (h  ) + 3 c d.
k

Use r = (2/ )d to obtain that:

 

8
2(2/ )d k
E [L D (h S )] 1 +
L D (h  ) + 3 c d +
.
S
k
m
Set = 2m 1/(d+1) and use



2k
6 c m 1/(d+1) d + m 1/(d+1) 6c d + k m 1/(d+1)
e
to conclude the proof.

227

20
Neural Networks

An artificial neural network is a model of computation inspired by the structure of

neural networks in the brain. In simplified models of the brain, it consists of a large
number of basic computing devices (neurons) that are connected to each other in
a complex communication network, through which the brain is able to carry out
highly complex computations. Artificial neural networks are formal computation
constructs that are modeled after this computation paradigm.
Learning with neural networks was proposed in the mid-20th century. It yields
an effective learning paradigm and has recently been shown to achieve cutting-edge
performance on several learning tasks.
A neural network can be described as a directed graph whose nodes correspond
to neurons and edges correspond to links between them. Each neuron receives as
input a weighted sum of the outputs of the neurons connected to its incoming edges.
We focus on feedforward networks in which the underlying graph does not contain
cycles.
In the context of learning, we can define a hypothesis class consisting of neural
network predictors, where all the hypotheses share the underlying graph structure
of the network and differ in the weights over edges. As we will show in Section 20.3,
every predictor over n variables that can be implemented in time T (n) can also be
expressed as a neural network predictor of size O(T (n)2 ), where the size of the network is the number of nodes in it. It follows that the family of hypothesis classes
of neural networks of polynomial size can suffice for all practical learning tasks, in
which our goal is to learn predictors which can be implemented efficiently. Furthermore, in Section 20.4 we will show that the sample complexity of learning such
hypothesis classes is also bounded in terms of the size of the network. Hence, it
seems that this is the ultimate learning paradigm we would want to adapt, in the
sense that it both has a polynomial sample complexity and has the minimal approximation error among all hypothesis classes consisting of efficiently implementable
predictors.
The caveat is that the problem of training such hypothesis classes of neural network predictors is computationally hard. This will be formalized in Section 20.5.

228

20.1 Feedforward Neural Networks

A widely used heuristic for training neural networks relies on the SGD framework we studied in Chapter 14. There, we have shown that SGD is a successful
learner if the loss function is convex. In neural networks, the loss function is highly
nonconvex. Nevertheless, we can still implement the SGD algorithm and hope it will
find a reasonable solution (as happens to be the case in several practical tasks). In
Section 20.6 we describe how to implement SGD for neural networks. In particular,
the most complicated operation is the calculation of the gradient of the loss function with respect to the parameters of the network. We present the backpropagation
algorithm that efficiently calculates the gradient.

20.1 FEEDFORWARD NEURAL NETWORKS

The idea behind neural networks is that many neurons can be joined together by
communication links to carry out complex computations. It is common to describe
the structure of a neural network as a graph whose nodes are the neurons and each
(directed) edge in the graph links the output of some neuron to the input of another
neuron. We will restrict our attention to feedforward network structures in which
the underlying graph does not contain cycles.
A feedforward neural network is described by a directed acyclic graph, G =
(V, E), and a weight function over the edges, w : E R. Nodes of the graph correspond to neurons. Each single neuron is modeled as a simple scalar function,
: R R. We will focus on three possible functions for : the sign function, (a) =
sign(a), the threshold function, (a) = 1[a>0] , and the sigmoid function, (a) =
1/(1 + exp( a)), which is a smooth approximation to the threshold function. We
call the activation function of the neuron. Each edge in the graph links the
output of some neuron to the input of another neuron. The input of a neuron is
obtained by taking a weighted sum of the outputs of all the neurons connected to it,
where the weighting is according to w.
To simplify the description of the calculation performed by the network, we
further assume that the network is organized in layers. That is, the set of nodes can
T V , such that
be decomposed into a union of (nonempty) disjoint subsets, V = t=0
t
every edge in E connects some node in Vt1 to some node in Vt , for some t [T ].
The bottom layer, V0 , is called the input layer. It contains n + 1 neurons, where n is
the dimensionality of the input space. For every i [n], the output of neuron i in V0
is simply x i . The last neuron in V0 is the constant neuron, which always outputs 1.
We denote by vt,i the i th neuron of the tth layer and by ot,i (x) the output of vt,i when
the network is fed with the input vector x. Therefore, for i [n] we have o0,i (x) = x i
and for i = n + 1 we have o0,i (x) = 1. We now proceed with the calculation in a layer
by layer manner. Suppose we have calculated the outputs of the neurons at layer t.
Then, we can calculate the outputs of the neurons at layer t + 1 as follows. Fix some
vt+1, j Vt+1 . Let at+1, j (x) denote the input to vt+1, j when the network is fed with
the input vector x. Then,

w((vt,r , vt+1, j )) ot,r (x),
at+1, j (x) =
r: (vt,r ,vt+1, j )E

and



ot+1, j (x) = at+1, j (x) .

229

230

Neural Networks

That is, the input to vt+1, j is a weighted sum of the outputs of the neurons in Vt that
are connected to vt+1, j , where weighting is according to w, and the output of vt+1, j
is simply the application of the activation function on its input.
Layers V1 , . . . , VT 1 are often called hidden layers. The top layer, VT , is called
the output layer. In simple prediction problems the output layer contains a single
neuron whose output is the output of the network.
We refer to T as the number of layers in the network (excluding V0 ), or the
depth of the network. The size of the network is |V |. The width of the network
is maxt |Vt |. An illustration of a layered feedforward neural network of depth 2, size
10, and width 5, is given in the following. Note that there is a neuron in the hidden
layer that has no incoming edges. This neuron will output the constant (0).
Input

Hidden

Output

Layer

Layer

Layer

(V0 )

(V1 )

(V2 )

v1,1
x1

v0,1

x2

v0,2

x3

v0,3

Constant

v0,4

v1,2
v1,3

v2,1

Output

v1,4
v1,5

20.2 LEARNING NEURAL NETWORKS

Once we have specified a neural network by (V , E, , w), we obtain a function
h V ,E,,w : R|V0 |1 R|VT | . Any set of such functions can serve as a hypothesis class
for learning. Usually, we define a hypothesis class of neural network predictors by
fixing the graph (V , E) as well as the activation function and letting the hypothesis
class be all functions of the form h V ,E,,w for some w : E R. The triplet (V , E, )
is often called the architecture of the network. We denote the hypothesis class by
HV ,E, = {h V ,E,,w : w is a mapping from E to R}.

(20.1)

That is, the parameters specifying a hypothesis in the hypothesis class are the
weights over the edges of the network.
We can now study the approximation error, estimation error, and optimization
error of such hypothesis classes. In Section 20.3 we study the approximation error
of HV ,E, by studying what type of functions hypotheses in HV ,E, can implement,
in terms of the size of the underlying graph. In Section 20.4 we study the estimation
error of HV ,E, , for the case of binary classification (i.e., VT = 1 and is the sign
function), by analyzing its VC dimension. Finally, in Section 20.5 we show that it
is computationally hard to learn the class HV ,E, , even if the underlying graph is
small, and in Section 20.6 we present the most commonly used heuristic for training
HV ,E, .

20.3 The Expressive Power of Neural Networks

20.3 THE EXPRESSIVE POWER OF NEURAL NETWORKS

In this section we study the expressive power of neural networks, namely, what type
of functions can be implemented using a neural network. More concretely, we will
fix some architecture, V , E, , and will study what functions hypotheses in HV ,E,
can implement, as a function of the size of V .
We start the discussion with studying which type of Boolean functions (i.e.,
functions from {1}n to {1}) can be implemented by HV ,E,sign . Observe that for
every computer in which real numbers are stored using b bits, whenever we calculate a function f : Rn R on such a computer we in fact calculate a function
g : {1}nb {1}b . Therefore, studying which Boolean functions can be implemented by HV ,E,sign can tell us which functions can be implemented on a computer
that stores real numbers using b bits.
We begin with a simple claim, showing that without restricting the size of the
network, every Boolean function can be implemented using a neural network of
depth 2.
Claim 20.1. For every n, there exists a graph (V, E) of depth 2, such that HV ,E,sign
contains all functions from {1}n to {1}.
Proof. We construct a graph with |V0 | = n +1, |V1 | = 2n +1, and |V2 | = 1. Let E be all
possible edges between adjacent layers. Now, let f : {1}n {1} be some Boolean
function. We need to show that we can adjust the weights so that the network will
implement f . Let u1 , . . . , uk be all vectors in {1}n on which f outputs 1. Observe
that for every i and every x {1}n , if x = ui then x, ui  n 2 and if x = ui then
x, u i  = n. It follows that the function gi (x) = sign(x, ui  n + 1) equals 1 if and
only if x = ui . It follows that we can adapt the weights between V0 and V1 so that for
every i [k], the neuron v1,i implements the function gi (x). Next, we observe that
f (x) is the disjunction of the functions gi (x), and therefore can be written as

 k

gi (x) + k 1 ,
f (x) = sign
i=1

which concludes our proof.

The preceding claim shows that neural networks can implement any Boolean
function. However, this is a very weak property, as the size of the resulting network
might be exponentially large. In the construction given at the proof of Claim 20.1,
the number of nodes in the hidden layer is exponentially large. This is not an artifact
of our proof, as stated in the following theorem.
Theorem 20.2. For every n, let s(n) be the minimal integer such that there exists a
graph (V , E) with |V | = s(n) such that the hypothesis class HV ,E,sign contains all the
functions from {0, 1}n to {0, 1}. Then, s(n) is exponential in n. Similar results hold for
HV ,E, where is the sigmoid function.
Proof. Suppose that for some (V , E) we have that HV ,E,sign contains all functions
from {0, 1}n to {0, 1}. It follows that it can shatter the set of m = 2n vectors in {0, 1}n
and hence the VC dimension of HV ,E,sign is 2n . On the other hand, the VC dimension of HV ,E,sign is bounded by O(|E| log (|E|)) O(|V |3 ), as we will show in the

231

232

Neural Networks

next section. This implies that |V | (2n/3 ), which concludes our proof for the
case of networks with the sign activation function. The proof for the sigmoid case is
analogous.
Remark 20.1. It is possible to derive a similar theorem for HV ,E, for any , as long
as we restrict the weights so that it is possible to express every weight using a number
of bits which is bounded by a universal constant. We can even consider hypothesis
classes where different neurons can employ different activation functions, as long as
the number of allowed activation functions is also finite.
Which functions can we express using a network of polynomial size? The preceding claim tells us that it is impossible to express all Boolean functions using a
network of polynomial size. On the positive side, in the following we show that all
Boolean functions that can be calculated in time O(T (n)) can also be expressed by
a network of size O(T (n)2 ).
Theorem 20.3. Let T : N N and for every n, let Fn be the set of functions that can
be implemented using a Turing machine using runtime of at most T (n). Then, there
exist constants b, c R+ such that for every n, there is a graph (Vn , E n ) of size at most
c T (n)2 + b such that HVn ,En ,sign contains Fn .
The proof of this theorem relies on the relation between the time complexity
of programs and their circuit complexity (see, for example, Sipser (2006)). In a
nutshell, a Boolean circuit is a type of network in which the individual neurons
implement conjunctions, disjunctions, and negation of their inputs. Circuit complexity measures the size of Boolean circuits required to calculate functions. The
relation between time complexity and circuit complexity can be seen intuitively as
follows. We can model each step of the execution of a computer program as a simple
operation on its memory state. Therefore, the neurons at each layer of the network
will reflect the memory state of the computer at the corresponding time, and the
translation to the next layer of the network involves a simple calculation that can
be carried out by the network. To relate Boolean circuits to networks with the sign
activation function, we need to show that we can implement the operations of conjunction, disjunction, and negation, using the sign activation function. Clearly, we
can implement the negation operator using the sign activation function. The following lemma shows that the sign activation function can also implement conjunctions
and disjunctions of its inputs.
Lemma 20.4. Suppose that a neuron v, that implements the sign activation function,
has k incoming edges, connecting it to neurons whose outputs are in {1}. Then, by
adding one more edge, linking a constant neuron to v, and by adjusting the weights
on the edges to v, the output of v can implement the conjunction or the disjunction of
its inputs.
function, f (x) =
Proof. Simply observe that if f : {1}k {1} is the conjunction



i x i , then it can be written as f (x) = sign 1 k + ki=1 x i . Similarly, the disjunc



tion function, f (x) = i x i , can be written as f (x) = sign k 1 + ki=1 x i .

20.3 The Expressive Power of Neural Networks

So far we have discussed Boolean functions. In Exercise 20.1 we show that neural
networks are universal approximators. That is, for every fixed precision parameter,
> 0, and every Lipschitz function f : [ 1, 1]n [ 1, 1], it is possible to construct
a network such that for every input x [ 1, 1]n , the network outputs a number
between f (x) and f (x) + . However, as in the case of Boolean functions, the
size of the network here again cannot be polynomial in n. This is formalized in the
following theorem, whose proof is a direct corollary of Theorem 20.2 and is left as
an exercise.
Theorem 20.5. Fix some (0, 1). For every n, let s(n) be the minimal integer such
that there exists a graph (V , E) with |V | = s(n) such that the hypothesis class HV ,E, ,
with being the sigmoid function, can approximate, to within precision of , every
1-Lipschitz function f : [ 1, 1]n [ 1, 1]. Then s(n) is exponential in n.

20.3.1 Geometric Intuition

We next provide several geometric illustrations of functions f : R2 {1} and show
how to express them using a neural network with the sign activation function.
Let us start with a depth 2 network, namely, a network with a single hidden layer.
Each neuron in the hidden layer implements a halfspace predictor. Then, the single
neuron at the output layer applies a halfspace on top of the binary outputs of the
neurons in the hidden layer. As we have shown before, a halfspace can implement
the conjunction function. Therefore, such networks contain all hypotheses which are
an intersection of k 1 halfspaces, where k is the number of neurons in the hidden
layer; namely, they can express all convex polytopes with k 1 faces. An example
of an intersection of 5 halfspaces is given in the following.

We have shown that a neuron in layer V2 can implement a function that indicates
whether x is in some convex polytope. By adding one more layer, and letting the
neuron in the output layer implement the disjunction of its inputs, we get a network
that computes the union of polytopes. An illustration of such a function is given in
the following.

233

234

Neural Networks

20.4 THE SAMPLE COMPLEXITY OF NEURAL NETWORKS

Next we discuss the sample complexity of learning the class HV , E, . Recall that the
fundamental theorem of learning tells us that the sample complexity of learning a
hypothesis class of binary classifiers depends on its VC dimension. Therefore, we
focus on calculating the VC dimension of hypothesis classes of the form HV , E, ,
where the output layer of the graph contains a single neuron.
We start with the sign activation function, namely, with HV , E,sign . What is the VC
dimension of this class? Intuitively, since we learn | E| parameters, the VC dimension should be order of | E|. This is indeed the case, as formalized by the following
theorem.
Theorem 20.6. The VC dimension of HV , E,sign is O(| E| log (| E|)).
Proof. To simplify the notation throughout the proof, let us denote the hypothesis
class by H. Recall the definition of the growth function, H (m), from Section 6.5.1.
This function measures maxCX :|C|=m |HC |, where HC is the restriction of H to functions from C to {0, 1}. We can naturally extend the definition for a set of functions
from X to some finite set Y, by letting HC be the restriction of H to functions from
C to Y, and keeping the definition of H (m) intact.
Our neural network is defined by a layered graph. Let V0 , . . . , VT be the layers
of the graph. Fix some t [T ]. By assigning different weights on the edges between
Vt1 and Vt , we obtain different functions from R|Vt1 | {1}|Vt | . Let H(t) be the
class of all possible such mappings from R|Vt1 | {1}|Vt | . Then, H can be written
as a composition, H = H(T ) . . . H(1) . In Exercise 20.4 we show that the growth
function of a composition of hypothesis classes is bounded by the products of the
growth functions of the individual classes. Therefore,
T


H (m)

H(t) (m).

t=1

In addition, each H(t) can be written as a product of function classes, H(t) = H(t,1)
H(t,|Vt |) , where each H(t, j ) is all functions from layer t 1 to {1} that the j th
neuron of layer t can implement. In Exercise 20.3 we bound product classes, and
this yields
H(t) (m)

|Vt |


H(t,i) (m).

i=1

Let dt,i be the number of edges that are headed to the i th neuron of layer t.
Since the neuron is a homogenous halfspace hypothesis and the VC dimension of
homogenous halfspaces is the dimension of their input, we have by Sauers lemma
that
 dt,i
H(t,i) (m) dem
(em)dt,i .
t,i
Overall, we obtained that

H (m) (em)

t,i dt,i

= (em)|E| .

20.5 The Runtime of Learning Neural Networks

Now, assume that there are m shattered points. Then, we must have H (m) = 2m ,
from which we obtain
2m (em)| E| m |E| log (em)/ log (2).
The claim follows by Lemma A.2.
Next, we consider HV , E, , where is the sigmoid function. Surprisingly, it turns
out that the VC dimension of HV , E, is lower bounded by (| E|2 ) (see Exercise
20.5.) That is, the VC dimension is the number of tunable parameters squared. It
is also possible to upper bound the VC dimension by O(|V |2 | E|2 ), but the proof
is beyond the scope of this book. In any case, since in practice we only consider
networks in which the weights have a short representation as floating point numbers
with O(1) bits, by using the discretization trick we easily obtain that such networks
have a VC dimension of O(| E|), even if we use the sigmoid activation function.

20.5 THE RUNTIME OF LEARNING NEURAL NETWORKS

In the previous sections we have shown that the class of neural networks with an
underlying graph of polynomial size can express all functions that can be implemented efficiently, and that the sample complexity has a favorable dependence on
the size of the network. In this section we turn to the analysis of the time complexity
of training neural networks.
We first show that it is NP hard to implement the ERM rule with respect to
HV , E,sign even for networks with a single hidden layer that contain just 4 neurons in
the hidden layer.
Theorem 20.7. Let k 3. For every n, let (V , E) be a layered graph with n input
nodes, k + 1 nodes at the (single) hidden layer, where one of them is the constant
neuron, and a single output node. Then, it is NP hard to implement the ERM rule
with respect to HV ,E,sign .
The proof relies on a reduction from the k-coloring problem and is left as
Exercise 20.6.
One way around the preceding hardness result could be that for the purpose of
learning, it may suffice to find a predictor h H with low empirical error, not necessarily an exact ERM. However, it turns out that even the task of finding weights that
result in close-to-minimal empirical error is computationally infeasible (see (Bartlett
& Ben-David 2002)).
One may also wonder whether it may be possible to change the architecture
of the network so as to circumvent the hardness result. That is, maybe ERM
with respect to the original network structure is computationally hard but ERM
with respect to some other, larger, network may be implemented efficiently (see
Chapter 8 for examples of such cases). Another possibility is to use other activation
functions (such as sigmoids, or any other type of efficiently computable activation
functions). There is a strong indication that all of such approaches are doomed to
fail. Indeed, under some cryptographic assumption, the problem of learning intersections of halfspaces is known to be hard even in the representation independent
model of learning (see Klivans & Sherstov (2006)). This implies that, under the

235

236

Neural Networks

same cryptographic assumption, any hypothesis class which contains intersections

of halfspaces cannot be learned efficiently.
A widely used heuristic for training neural networks relies on the SGD framework we studied in Chapter 14. There, we have shown that SGD is a successful
learner if the loss function is convex. In neural networks, the loss function is highly
nonconvex. Nevertheless, we can still implement the SGD algorithm and hope
it will find a reasonable solution (as happens to be the case in several practical
tasks).

20.6 SGD AND BACKPROPAGATION

The problem of finding a hypothesis in HV ,E, with a low risk amounts to the problem of tuning the weights over the edges. In this section we show how to apply a
heuristic search for good weights using the SGD algorithm. Throughout this section
we assume that is the sigmoid function, (a) = 1/(1 + ea ), but the derivation
holds for any differentiable scalar function.
Since E is a finite set, we can think of the weight function as a vector w R|E| .
Suppose the network has n input neurons and k output neurons, and denote by
h w : Rn Rk the function calculated by the network if the weight function is defined
by w. Let us denote by (h w (x), y) the loss of predicting h w (x) when the target
is y Y. For concreteness, we will take  to be the squared loss, (h w (x), y) =
1
2
2 h w (x) y ; however, similar derivation can be obtained for every differentiable
function. Finally, given a distribution D over the examples domain, Rn Rk , let
L D (w) be the risk of the network, namely,
L D (w) =

(x,y)D

[(h w (x), y)] .

Recall the SGD algorithm for minimizing the risk function L D (w). We repeat
the pseudocode from Chapter 14 with a few modifications, which are relevant to the
neural network application because of the nonconvexity of the objective function.
First, while in Chapter 14 we initialized w to be the zero vector, here we initialize w
to be a randomly chosen vector with values close to zero. This is because an initialization with the zero vector will lead all hidden neurons to have the same weights
(if the network is a full layered network). In addition, the hope is that if we repeat
the SGD procedure several times, where each time we initialize the process with
a new random vector, one of the runs will lead to a good local minimum. Second,
while a fixed step size, , is guaranteed to be good enough for convex problems,
here we utilize a variable step size, t , as defined in Section 14.4.2. Because of the
nonconvexity of the loss function, the choice of the sequence t is more significant,
and it is tuned in practice by a trial and error manner. Third, we output the best
performing vector on a validation set. In addition, it is sometimes helpful to add regularization on the weights, with parameter . That is, we try to minimize L D (w) +

2
2 w . Finally, the gradient does not have a closed form solution. Instead, it is
implemented using the backpropagation algorithm, which will be described in the
sequel.

20.6 SGD and Backpropagation

SGD for Neural Networks

parameters:
number of iterations
step size sequence 1 , 2 , . . . ,
regularization parameter > 0
input:
layered graph (V , E)
differentiable activation function : R R
initialize:
choose w(1) R|E| at random
(from a distribution s.t. w(1) is close enough to 0)
for i = 1, 2, . . . ,
sample (x, y) D
calculate gradient vi = backpropagation(x, y, w, (V , E), )
update w(i+1) = w(i) i (vi + w(i) )
output:
is the best performing w(i) on a validation set
w
Backpropagation
input:
example (x, y), weight vector w, layered graph (V , E),
activation function : R R
initialize:
denote layers of the graph V0 , . . . , VT where Vt = {vt,1 , . . . , vt,kt }
define Wt,i, j as the weight of (vt, j , vt+1,i )
/ E)
(where we set Wt,i, j = 0 if (vt, j , vt+1,i )
forward:
set o0 = x
for t = 1, . . . , T
for i = 1, . . . , kt

k
set at,i = jt1
=1 Wt1,i, j ot1, j
set ot,i = (at,i )
backward:
set T = oT y
for t = T 1, T 2, . . . , 1
for i = 1, . . . , kt

k

t,i = jt+1
=1 Wt, j ,i t+1, j (at+1, j )
output:
foreach edge (vt1, j , vt,i ) E
set the partial derivative to t,i  (at,i ) ot1, j
Explaining How Backpropagation Calculates the Gradient:
We next explain how the backpropagation algorithm calculates the gradient of the
loss function on an example (x, y) with respect to the vector w. Let us first recall

237

238

Neural Networks

a few definitions from vector calculus. Each element of the gradient is the partial
derivative with respect to the variable in w corresponding to one of the edges of the
network. Recall the definition of a partial derivative. Given a function f : Rn R,
the partial derivative with respect to the i th variable at w is obtained by fixing the
values of w1 , . . . , wi1 , wi+1 , wn , which yields the scalar function g : R R defined
by g(a) = f ((w1 , . . . , wi1 , wi + a, wi+1 , . . . , wn )), and then taking the derivative of g
at 0. For a function with multiple outputs, f : Rn Rm , the Jacobian of f at w Rn ,
denoted Jw (f), is the m n matrix whose i , j element is the partial derivative of f i :
Rn R w.r.t. its j th variable at w. Note that if m = 1 then the Jacobian matrix is the
gradient of the function (represented as a row vector). Two examples of Jacobian
calculations, which we will later use, are as follows.

Let f(w) = Aw for A Rm,n . Then Jw (f) = A.

For every n, we use the notation to denote the function from Rn to Rn which
applies the sigmoid function element-wise. That is, = ( ) means that for
1
. It is easy to verify that J ( ) is a diagoevery i we have i = (i ) = 1+exp(
i)

nal matrix whose (i ,i ) entry is (i ), where  is the derivative function of the
1
. We also use the
(scalar) sigmoid function, namely,  (i ) = (1+exp( ))(1+exp(
i
i ))

notation diag( ( )) to denote this matrix.

The chain rule for taking the derivative of a composition of functions can be
written in terms of the Jacobian as follows. Given two functions f : Rn Rm and
g : Rk Rn , we have that the Jacobian of the composition function, (f g) : Rk Rm ,
at w, is
Jw (f g) = Jg(w) (f)Jw (g).
For example, for g(w) = Aw, where A Rn,k , we have that
Jw ( g) = diag(  (Aw)) A.
To describe the backpropagation algorithm, let us first decompose V into the
T V . For every t, let us write V = {v , . . . , v
layers of the graph, V = t=0
t
t
t,kt }, where
t,1
kt = |Vt |. In addition, for every t denote Wt Rkt+1 ,kt a matrix which gives a weight to
every potential edge between Vt and Vt+1 . If the edge exists in E then we set Wt,i, j to
be the weight, according to w, of the edge (vt, j , vt+1,i ). Otherwise, we add a phantom edge and set its weight to be zero, Wt,i, j = 0. Since when calculating the partial
derivative with respect to the weight of some edge we fix all other weights, these
additional phantom edges have no effect on the partial derivative with respect
to existing edges. It follows that we can assume, without loss of generality, that all
edges exist, that is, E = t (Vt Vt+1 ).
Next, we discuss how to calculate the partial derivatives with respect to the edges
from Vt1 to Vt , namely, with respect to the elements in Wt1 . Since we fix all other
weights of the network, it follows that the outputs of all the neurons in Vt1 are fixed
numbers which do not depend on the weights in Wt1 . Denote the corresponding
vector by ot1 . In addition, let us denote by
t : Rkt R the loss function of the
subnetwork defined by layers Vt , . . . , VT as a function of the outputs of the neurons
in Vt . The input to the neurons of Vt can be written as at = Wt1 ot1 and the output
of the neurons of Vt is ot = (at ). That is, for every j we have ot, j = (at, j ). We

20.6 SGD and Backpropagation

obtain that the loss, as a function of Wt1 , can be written as

gt (Wt1 ) =
t (ot ) =
t ( (at )) =
t ( (Wt1 ot1 )).
It would be convenient to rewrite this as follows. Let wt1 Rkt1 kt be the column
vector obtained by concatenating the rows of Wt1 and then taking the transpose of
the resulting long vector. Define by Ot1 the kt (kt1 kt ) matrix

ot1
0

0
0
o

t1

.
(20.2)
Ot1 = .
..
..
..
..
.
.
.

t1

Then, Wt1 ot1 = Ot1 wt1 , so we can also write

gt (wt1 ) =
t ( (Ot1 wt1 )).
Therefore, applying the chain rule, we obtain that
Jwt1 (gt ) = J (Ot1 wt1 ) (
t ) diag(  (Ot1 wt1 )) Ot1 .
Using our notation we have ot = (Ot1 wt1 ) and at = Ot1 wt1 , which yields
Jwt1 (gt ) = Jot (
t ) diag(  (at )) Ot1 .
Let us also denote t = Jot (
t ). Then, we can further rewrite the preceding as




Jwt1 (gt ) = t,1  (at,1 ) o

(20.3)
t1 , . . . , t,kt (at,kt ) ot1 .
It is left to calculate the vector t = Jot (
t ) for every t. This is the gradient of
t
at ot . We calculate this in a recursive manner. First observe that for the last layer
we have that
T (u) = (u, y), where  is the loss function. Since we assume that
(u, y) = 12 u y2 we obtain that Ju (
T ) = (u y). In particular, T = JoT (
T ) =
(oT y). Next, note that

t (u) =
t+1 ( (Wt u)).
Therefore, by the chain rule,
Ju (
t ) = J (Wt u) (
t+1 )diag(  (Wt u))Wt .
In particular,
t = Jot (
t ) = J (Wt ot ) (
t+1 )diag(  (Wt ot ))Wt
= Jot+1 (
t+1 )diag(  (at+1 ))Wt
= t+1 diag(  (at+1 ))Wt .
In summary, we can first calculate the vectors {at , ot } from the bottom of the
network to its top. Then, we calculate the vectors { t } from the top of the network
back to its bottom. Once we have all of these vectors, the partial derivatives are
easily obtained using Equation (20.3). We have thus shown that the pseudocode of
backpropagation indeed calculates the gradient.

239

240

Neural Networks

20.7 SUMMARY
Neural networks over graphs of size s(n) can be used to describe
 hypothesis classes
of all predictors that can be implemented in runtime of O( s(n)). We have also
shown that their sample complexity depends polynomially on s(n) (specifically,
it depends on the number of edges in the network). Therefore, classes of neural network hypotheses seem to be an excellent choice. Regrettably, the problem
of training the network on the basis of training data is computationally hard. We
have presented the SGD framework as a heuristic approach for training neural networks and described the backpropagation algorithm which efficiently calculates the
gradient of the loss function with respect to the weights over the edges.

20.8 BIBLIOGRAPHIC REMARKS

Neural networks were extensively studied in the 1980s and early 1990s, but with
mixed empirical success. In recent years, a combination of algorithmic advancements, as well as increasing computational power and data size, has led to a
breakthrough in the effectiveness of neural networks. In particular, deep networks (i.e., networks of more than 2 layers) have shown very impressive practical
performance on a variety of domains. A few examples include convolutional networks (LeCun & Bengio 1995), restricted Boltzmann machines (Hinton, Osindero
& Teh 2006), auto-encoders (Ranzato et al. 2007, Bengio & LeCun 2007, Collobert
& Weston 2008, Lee et al. 2009, Le et al. 2012), and sum-product networks (Livni,
Shalev-Shwartz & Shamir 2013, Poon & Domingos 2011). See also (Bengio 2009)
and the references therein.
The expressive power of neural networks and the relation to circuit complexity
have been extensively studied in (Parberry 1994). For the analysis of the sample
complexity of neural networks we refer the reader to (Anthony & Bartlet 1999).
Our proof technique of Theorem 20.6 is due to Kakade and Tewari lecture notes.
Klivans and Sherstov (2006) have shown that for any c > 0, intersections of
n c halfspaces over {1}n are not efficiently PAC learnable, even if we allow representation independent learning. This hardness result relies on the cryptographic
assumption that there is no polynomial time solution to the unique-shortest-vector
problem. As we have argued, this implies that there cannot be an efficient algorithm
for training neural networks, even if we allow larger networks or other activation
functions that can be implemented efficiently.
The backpropagation algorithm has been introduced in Rumelhart, Hinton, and
Williams (1986).

20.9 EXERCISES
20.1 Neural Networks are universal approximators: Let f : [ 1, 1]n [ 1, 1] be a
-Lipschitz function. Fix some > 0. Construct a neural network N : [ 1, 1]n
[ 1, 1], with the sigmoid activation function, such that for every x [ 1, 1]n it
holds that | f (x) N (x)| .
Hint: Similarly to the proof of Theorem 19.3, partition [ 1, 1]n into small boxes.
Use the Lipschitzness of f to show that it is approximately constant at each box.

20.9 Exercises

20.2

20.3

20.4

20.5

Finally, show that a neural network can first decide which box the input vector
belongs to, and then predict the averaged value of f at that box.
Prove Theorem 20.5.
Hint: For every f : {1, 1}n {1, 1} construct a 1-Lipschitz function g :
[ 1, 1]n [ 1, 1] such that if you can approximate g then you can express f .
Growth function of product: For i = 1, 2, let Fi be a set of functions from X to Yi .
Define H = F1 F2 to be the Cartesian product class. That is, for every f 1 F1
and f2 F2 , there exists h H such that h(x) = ( f 1 (x), f 2 (x)). Prove that H (m)
F1 (m) F2 (m).
Growth function of composition: Let F1 be a set of functions from X to Z and let
F2 be a set of functions from Z to Y . Let H = F2 F1 be the composition class. That
is, for every f 1 F1 and f2 F2 , there exists h H such that h(x) = f 2 ( f 1 (x)). Prove
that H (m) F2 (m)F1 (m).
VC of sigmoidal networks: In this exercise we show that there is a graph (V , E)
such that the VC dimension of the class of neural networks over these graphs with
the sigmoid activation function is (|E|2 ). Note that for every > 0, the sigmoid
activation function can approximate the threshold activation function, 1[
i xi ] , up
to accuracy . To simplify the presentation, throughout the exercise we assume
that we can exactly implement the activation function 1[
i xi >0] using a sigmoid
activation function.
Fix some n.
1. Construct a network, N1 , with O(n) weights, which implements a function from
R to {0, 1}n and satisfies the following property. For every x {0, 1}n , if we feed
the network with the real number 0. x1 x2 . . . xn , then the output of the network
will be x.
Hint: Denote = 0. x1 x2 . . . xn and observe that 10k 0. 5 is at least 0. 5 if xk = 1
and is at most 0. 3 if xk = 1.
2. Construct a network, N2 , with O(n) weights, which implements a function from
[n] to {0, 1}n such that N2 (i ) = ei for all i . That is, upon receiving the input i , the
network outputs the vector of all zeros except 1 at the i th neuron.
(i) (i)
(i)
3. Let 1 , . . ., n be n real numbers such that every i is of the form 0. a1 a2 . . .an ,
(i)

with a j {0, 1}. Construct a network, N3 , with O(n) weights, which implements
a function from [n] to R, and satisfies N2 (i ) = i for every i [n].
4. Combine N1 , N3 to obtain a network that receives i [n] and output a(i) .
(i)
5. Construct a network N4 that receives (i , j ) [n] [n] and outputs a j .
Hint: Observe that the AND function over {0, 1}2 can be calculated using O(1)
weights.
6. Conclude that there is a graph with O(n) weights such that the VC dimension
of the resulting hypothesis class is n 2 .
20.6 Prove Theorem 20.7.
Hint: The proof is similar to the hardness of learning intersections of halfspaces
see Exercise 32 in Chapter 8.

241

PART 3

Additional Learning Models

21
Online Learning

In this chapter we describe a different model of learning, which is called online

learning. Previously, we studied the PAC learning model, in which the learner first
receives a batch of training examples, uses the training set to learn a hypothesis,
and only when learning is completed uses the learned hypothesis for predicting
the label of new examples. In our papayas learning problem, this means that we
should first buy a bunch of papayas and taste them all. Then, we use all of this
information to learn a prediction rule that determines the taste of new papayas. In
contrast, in online learning there is no separation between a training phase and a
prediction phase. Instead, each time we buy a papaya, it is first considered a test
example since we should predict whether it is going to taste good. Then, after taking
a bite from the papaya, we know the true label, and the same papaya can be used
as a training example that can help us improve our prediction mechanism for future
papayas.
Concretely, online learning takes place in a sequence of consecutive rounds.
On each online round, the learner first receives an instance (the learner buys a
papaya and knows its shape and color, which form the instance). Then, the learner
is required to predict a label (is the papaya tasty?). At the end of the round, the
learner obtains the correct label (he tastes the papaya and then knows whether
it is tasty or not). Finally, the learner uses this information to improve his future
predictions.
To analyze online learning, we follow a similar route to our study of PAC
learning. We start with online binary classification problems. We consider both the
realizable case, in which we assume, as prior knowledge, that all the labels are generated by some hypothesis from a given hypothesis class, and the unrealizable case,
which corresponds to the agnostic PAC learning model. In particular, we present
an important algorithm called Weighted-Majority. Next, we study online learning
problems in which the loss function is convex. Finally, we present the Perceptron
algorithm as an example of the use of surrogate convex loss functions in the online
learning model.

245

246

Online Learning

21.1 ONLINE CLASSIFICATION IN THE REALIZABLE CASE

Online learning is performed in a sequence of consecutive rounds, where at round
t the learner is given an instance, xt , taken from an instance domain X , and is
required to provide its label. We denote the predicted label by pt . After predicting
the label, the correct label, yt {0, 1}, is revealed to the learner. The learners goal
is to make as few prediction mistakes as possible during this process. The learner
tries to deduce information from previous rounds so as to improve its predictions
on future rounds.
Clearly, learning is hopeless if there is no correlation between past and present
rounds. Previously in the book, we studied the PAC model in which we assume that
past and present examples are sampled i.i.d. from the same distribution source. In
the online learning model we make no statistical assumptions regarding the origin
of the sequence of examples. The sequence is allowed to be deterministic, stochastic, or even adversarially adaptive to the learners own behavior (as in the case of
spam e-mail filtering). Naturally, an adversary can make the number of prediction mistakes of our online learning algorithm arbitrarily large. For example, the
adversary can present the same instance on each online round, wait for the learners
prediction, and provide the opposite label as the correct label.
To make nontrivial statements we must further restrict the problem. The realizability assumption is one possible natural restriction. In the realizable case, we
assume that all the labels are generated by some hypothesis, h  : X Y. Furthermore, h  is taken from a hypothesis class H, which is known to the learner. This is
analogous to the PAC learning model we studied in Chapter 3. With this restriction
on the sequence, the learner should make as few mistakes as possible, assuming
that both h  and the sequence of instances can be chosen by an adversary. For an
online learning algorithm, A, we denote by M A (H) the maximal number of mistakes
A might make on a sequence of examples which is labeled by some h  H. We
emphasize again that both h  and the sequence of instances can be chosen by an
adversary. A bound on M A (H) is called a mistake-bound and we will study how to
design algorithms for which M A (H) is minimal. Formally:
Definition 21.1 (Mistake Bounds, Online Learnability). Let H be a hypothesis class and let A be an online learning algorithm. Given any sequence
S = (x 1 , h  (y1 )), . . . , (x T , h  (yT )), where T is any integer and h  H, let M A (S) be
the number of mistakes A makes on the sequence S. We denote by M A (H) the
supremum of M A (S) over all sequences of the preceding form. A bound of the form
M A (H) B < is called a mistake bound. We say that a hypothesis class H is online
learnable if there exists an algorithm A for which M A (H) B < .
Our goal is to study which hypothesis classes are learnable in the online model,
and in particular to find good learning algorithms for a given hypothesis class.
Remark 21.1. Throughout this section and the next, we ignore the computational
aspect of learning, and do not restrict the algorithms to be efficient. In Section 21.3
and Section 21.4 we study efficient online learning algorithms.
To simplify the presentation, we start with the case of a finite hypothesis class,
namely, |H| < .

21.1 Online Classification in the Realizable Case

In PAC learning, we identified ERM as a good learning algorithm, in the sense

that if H is learnable then it is learnable by the rule ERMH . A natural learning rule
for online learning is to use (at any online round) any ERM hypothesis, namely, any
hypothesis which is consistent with all past examples.
Consistent
input: A finite hypothesis class H
initialize: V1 = H
for t = 1, 2, . . .
receive xt
choose any h Vt
predict pt = h(xt )
receive true label yt = h  (xt )
update Vt+1 = {h Vt : h(xt ) = yt }
The Consistent algorithm maintains a set, Vt , of all the hypotheses which are
consistent with (x1 , y1 ), . . . , (xt1 , yt1 ). This set is often called the version space. It
then picks any hypothesis from Vt and predicts according to this hypothesis.
Obviously, whenever Consistent makes a prediction mistake, at least one hypothesis is removed from Vt . Therefore, after making M mistakes we have |Vt | |H| M.
Since Vt is always nonempty (by the realizability assumption it contains h  ) we have
1 |Vt | |H| M. Rearranging, we obtain the following:
Corollary 21.2. Let H be a finite hypothesis class. The Consistent algorithm enjoys the
mistake bound MConsistent (H) |H| 1.
It is rather easy to construct a hypothesis class and a sequence of examples on
which Consistent will indeed make |H| 1 mistakes (see Exercise 21.1.) Therefore,
we present a better algorithm in which we choose h Vt in a smarter way. We shall
see that this algorithm is guaranteed to make exponentially fewer mistakes.
Halving
input: A finite hypothesis class H
initialize: V1 = H
for t = 1, 2, . . .
receive xt
predict pt = argmaxr{0,1} |{h Vt : h(xt ) = r }|
(in case of a tie predict pt = 1)
receive true label yt = h  (xt )
update Vt+1 = {h Vt : h(xt ) = yt }

Theorem 21.3. Let H be a finite hypothesis class. The Halving algorithm enjoys the
mistake bound MHalving (H) log2 (|H|).

247

248

Online Learning

Proof. We simply note that whenever the algorithm errs we have |Vt+1 | |Vt |/2,
(hence the name Halving). Therefore, if M is the total number of mistakes, we have
1 |VT +1 | |H| 2M .
Rearranging this inequality we conclude our proof.
Of course, Halvings mistake bound is much better than Consistents mistake
bound. We already see that online learning is different from PAC learningwhile
in PAC, any ERM hypothesis is good, in online learning choosing an arbitrary ERM
hypothesis is far from being optimal.

21.1.1 Online Learnability

We next take a more general approach, and aim at characterizing online learnability.
In particular, we target the following question: What is the optimal online learning
algorithm for a given hypothesis class H?
We present a dimension of hypothesis classes that characterizes the best achievable mistake bound. This measure was proposed by Nick Littlestone and we
therefore refer to it as Ldim(H).
To motivate the definition of Ldim it is convenient to view the online learning
process as a game between two players: the learner versus the environment. On
round t of the game, the environment picks an instance xt , the learner predicts a
label pt {0, 1}, and finally the environment outputs the true label, yt {0, 1}. Suppose that the environment wants to make the learner err on the first T rounds of the
game. Then, it must output yt = 1 pt , and the only question is how it should choose
the instances xt in such a way that ensures that for some h  H we have yt = h  (xt )
for all t [T ].
A strategy for an adversarial environment can be formally described as a binary
tree, as follows. Each node of the tree is associated with an instance from X . Initially,
the environment presents to the learner the instance associated with the root of the
tree. Then, if the learner predicts pt = 1 the environment will declare that this is a
wrong prediction (i.e., yt = 0) and will traverse to the right child of the current node.
If the learner predicts pt = 0 then the environment will set yt = 1 and will traverse
to the left child. This process will continue and at each round, the environment will
present the instance associated with the current node.
Formally, consider a complete binary tree of depth T (we define the depth of
the tree as the number of edges in a path from the root to a leaf). We have 2T +1 1
nodes in such a tree, and we attach an instance to each node. Let v1 , . . . , v2T +1 1 be
these instances. We start from the root of the tree, and set x1 = v1 . At round t, we
set xt = vit where i t is the current node. At the end of round t, we go to the left child
of i t if yt = 0 or to the right child if yt = 1. That is, i t+1 = 2i t + yt . Unraveling the


t1 j .
recursion we obtain i t = 2t1 + t1
j =1 y j 2
The preceding strategy for the environment succeeds only if for every
(y1 , . . . , yT ) there exists h H such that yt = h(xt ) for all t [T ]. This leads to the
following definition.
Definition 21.4 (H Shattered Tree). A shattered tree of depth d is a sequence
of instances v1 , . . . , v2d 1 in X such that for every labeling (y1 , . . . , yd ) {0, 1}d

21.1 Online Classification in the Realizable Case

v1
v2

v1
v2
v3

v3

h1

h2

h3

h4

0
0

0
1

Figure 21.1. An illustration of a shattered tree of depth 2. The dashed path corresponds
to the sequence of examples ((v1 , 1), (v3 , 0)). The tree is shattered by H = {h 1 , h 2 , h 3 , h 4 },
where the predictions of each hypothesis in H on the instances v1 , v2 , v3 is given in the
table (the * mark means that h j (vi ) can be either 1 or 0).

there exists h H such that for all t [d] we have h(vit ) = yt where i t = 2t1 +

t1
t1 j .
j =1 y j 2
An illustration of a shattered tree of depth 2 is given in Figure 21.1.
Definition 21.5 (Littlestones Dimension (Ldim)). Ldim(H) is the maximal integer
T such that there exists a shattered tree of depth T , which is shattered by H.
The definition of Ldim and the previous discussion immediately imply the
following:
Lemma 21.6. No algorithm can have a mistake bound strictly smaller than
Ldim(H); namely, for every algorithm, A, we have M A (H) Ldim(H).
Proof. Let T = Ldim(H) and let v1 , . . . , v2T 1 be a sequence that satisfies the
requirements in the definition of Ldim. If the environment sets xt = vit and yt =
1 pt for all t [T ], then the learner makes T mistakes while the definition of Ldim
implies that there exists a hypothesis h H such that yt = h(xt ) for all t.
Let us now give several examples.
Example 21.2. Let H be a finite hypothesis class. Clearly, any tree that is shattered
by H has depth of at most log2 (|H|). Therefore, Ldim(H) log2 (|H|). Another way
to conclude this inequality is by combining Lemma 21.6 with Theorem 21.3.
Example 21.3. Let X = {1, . . . , d} and H = {h 1 , . . . , h d } where h j (x) = 1 iff x = j .
Then, it is easy to show that Ldim(H) = 1 while |H| = d can be arbitrarily large.
Therefore, this example shows that Ldim(H) can be significantly smaller than
log2 (|H|).
Example 21.4. Let X = [0, 1] and H = {x  1[x<a] : a [0, 1]}; namely, H is the
class of thresholds on the interval [0, 1]. Then, Ldim(H) = . To see this, consider
the tree
1/2
1/4
1/8

3/4
3/8

5/8

7/8

249

250

Online Learning

This tree is shattered by H. And, because of the density of the reals, this tree can be
made arbitrarily deep.
Lemma 21.6 states that Ldim(H) lower bounds the mistake bound of any algorithm. Interestingly, there is a standard algorithm whose mistake bound matches this
lower bound. The algorithm is similar to the Halving algorithm. Recall that the prediction of Halving is made according to a majority vote of the hypotheses which are
consistent with previous examples. We denoted this set by Vt . Put another way, Halving partitions Vt into two sets: Vt+ = {h Vt : h(xt ) = 1} and Vt = {h Vt : h(xt ) = 0}.
It then predicts according to the larger of the two groups. The rationale behind this
prediction is that whenever Halving makes a mistake it ends up with |Vt+1 | 0. 5 |Vt |.
The optimal algorithm we present in the following uses the same idea, but
instead of predicting according to the larger class, it predicts according to the class
with larger Ldim.
Standard Optimal Algorithm (SOA)
input: A hypothesis class H
initialize: V1 = H
for t = 1, 2, . . .
receive xt
(r)
for r {0, 1} let Vt = {h Vt : h(xt ) = r }
(r)
predict pt = argmaxr{0,1} Ldim(Vt )
(in case of a tie predict pt = 1)
receive true label yt
update Vt+1 = {h Vt : h(xt ) = yt }
The following lemma formally establishes the optimality of the preceding
algorithm.
Lemma 21.7. SOA enjoys the mistake bound MSOA (H) Ldim(H).
Proof. It suffices to prove that whenever the algorithm makes a prediction mistake
we have Ldim(Vt+1 ) Ldim(Vt ) 1. We prove this claim by assuming the contrary,
that is, Ldim(Vt+1 ) = Ldim(Vt ). If this holds true, then the definition of pt implies
(r)
that Ldim(Vt ) = Ldim(Vt ) for both r = 1 and r = 0. But, then we can construct
a shaterred tree of depth Ldim(Vt ) + 1 for the class Vt , which leads to the desired
contradiction.
Combining Lemma 21.7 and Lemma 21.6 we obtain:
Corollary 21.8. Let H be any hypothesis class. Then, the standard optimal algorithm enjoys the mistake bound MSOA (H) = Ldim(H) and no other algorithm can
have M A (H) < Ldim(H).
Comparison to VC Dimension
In the PAC learning model, learnability is characterized by the VC dimension of
the class H. Recall that the VC dimension of a class H is the maximal number d
such that there are instances x1 , . . . , xd that are shattered by H. That is, for any

21.2 Online Classification in the Unrealizable Case

sequence of labels (y1 , . . . , yd ) {0, 1}d there exists a hypothesis h H that gives
exactly this sequence of labels. The following theorem relates the VC dimension to
the Littlestone dimension.
Theorem 21.9. For any class H, VCdim(H) Ldim(H), and there are classes for
which strict inequality holds. Furthermore, the gap can be arbitrarily larger.
Proof. We first prove that VCdim(H) Ldim(H). Suppose VCdim(H) = d and
let x1 , . . . , xd be a shattered set. We now construct a complete binary tree of
instances v1 , . . . , v2d 1 , where all nodes at depth i are set to be xi see the following
illustration:
x1
x2
x3

x2
x3

x3

x3

Now, the definition of a shattered set clearly implies that we got a valid shattered
tree of depth d, and we conclude that VCdim(H) Ldim(H). To show that the gap
can be arbitrarily large simply note that the class given in Example 21.4 has VC
dimension of 1 whereas its Littlestone dimension is infinite.

21.2 ONLINE CLASSIFICATION IN THE UNREALIZABLE CASE

In the previous section we studied online learnability in the realizable case. We now
consider the unrealizable case. Similarly to the agnostic PAC model, we no longer
assume that all labels are generated by some h  H, but we require the learner to
be competitive with the best fixed predictor from H. This is captured by the regret
of the algorithm, which measures how sorry the learner is, in retrospect, not to
have followed the predictions of some hypothesis h H. Formally, the regret of an
algorithm A relative to h when running on a sequence of T examples is defined as

 T
T


Regret A (h, T ) =
sup
| p t yt |
|h(x t ) yt | ,
(21.1)
(x1 ,y1 ),...,(x T ,yT )

t=1

t=1

and the regret of the algorithm relative to a hypothesis class H is

Regret A (H, T ) = sup Regret A (h, T ).
hH

(21.2)

We restate the learners goal as having the lowest possible regret relative to H. An
interesting question is whether we can derive an algorithm with low regret, meaning
that Regret A (H, T ) grows sublinearly with the number of rounds, T , which implies
that the difference between the error rate of the learner and the best hypothesis in
H tends to zero as T goes to infinity.
We first show that this is an impossible missionno algorithm can obtain a
sublinear regret bound even if |H| = 2. Indeed, consider H = {h 0 , h 1 }, where h 0
is the function that always returns 0 and h 1 is the function that always returns 1. An

251

252

Online Learning

adversary can make the number of mistakes of any online algorithm be equal to T ,
by simply waiting for the learners prediction and then providing the opposite label
as the true label. In contrast, for any sequence of true labels, y1 , . . . , yT , let b be
the majority of labels in y1 , . . . , yT , then the number of mistakes of h b is at most T /2.
Therefore, the regret of any online algorithm might be at least T T /2 = T /2, which
is not sublinear in T . This impossibility result is attributed to Cover (Cover 1965).
To sidestep Covers impossibility result, we must further restrict the power of the
adversarial environment. We do so by allowing the learner to randomize his predictions. Of course, this by itself does not circumvent Covers impossibility result, since
in deriving this result we assumed nothing about the learners strategy. To make the
randomization meaningful, we force the adversarial environment to decide on yt
without knowing the random coins flipped by the learner on round t. The adversary
can still know the learners forecasting strategy and even the random coin flips of
previous rounds, but it does not know the actual value of the random coin flips used
by the learner on round t. With this (mild) change of game, we analyze the expected
number of mistakes of the algorithm, where the expectation is with respect to the
learners own randomization. That is, if the learner outputs yt where P [ yt = 1] = pt ,
then the expected loss he pays on round t is
P [ yt = yt ] = | pt yt |.
Put another way, instead of having the predictions of the learner being in {0, 1} we
allow them to be in [0, 1], and interpret pt [0, 1] as the probability to predict the
label 1 on round t.
With this assumption it is possible to derive a low regret algorithm. In particular,
we will prove the following theorem.
Theorem 21.10. For every hypothesis class H, there exists an algorithm for online
classification, whose predictions come from [0, 1], that enjoys the regret bound
h H,

T

t=1

| p t yt |

T


|h(xt ) yt |


2 min{log (|H|) , Ldim(H) log (eT )} T .

t=1

Furthermore,
no algorithm can achieve an expected regret bound smaller than

Ldim(H) T .

We will provide a constructive proof of the upper bound part of the preceding
theorem. The proof of the lower bound part can be found in (Ben-David, Pal, &
Shalev-Shwartz 2009).
The proof of Theorem 21.10 relies on the Weighted-Majority algorithm for learning with expert advice. This algorithm is important by itself and we dedicate the next
subsection to it.

21.2.1 Weighted-Majority
Weighted-majority is an algorithm for the problem of prediction with expert advice.
In this online learning problem, on round t the learner has to choose the advice
of d given experts. We also allow the learner to randomize his choice by defining a distribution over the d experts, that is, picking a vector w(t) [0, 1]d , with

21.2 Online Classification in the Unrealizable Case

(t)

(t)

wi = 1, and choosing the i th expert with probability wi . After the learner

chooses an expert, it receives a vector of costs, vt [0, 1]d , where vt,i is the cost of
following the advice of the i th expert. If the learners predictions are randomized,

(t)
then its loss is defined to be the averaged cost, namely, i wi vt,i = w(t) , vt . The
algorithm assumes that the number of rounds T is given. In Exercise 21.4 we show
how to get rid of this dependence using the doubling trick.
i

Weighted-Majority
input: number of
experts, d ; number of rounds, T
parameter: = 2 log (d)/T
(1) = (1, . . . , 1)
initialize: w
for t = 1, 2, . . .

(t)
(t) /Z t where Z t = i w i
set w(t) = w
(t)
choose expert i at random according to P [i ] = wi
d
receive costs of all experts vt [0, 1]
pay cost w(t) , vt 
(t+1)
(t)
= w i evt,i
update rule i , w i
The following theorem is key for analyzing the regret bound of WeightedMajority.
Theorem 21.11. Assuming that T > 2 log (d), the Weighted-Majority algorithm enjoys
the bound
T
T



w(t) , vt  min
vt,i 2 log (d) T .
i[d]

t=1

t=1

Proof. We have:
 w
 (t)
Z t+1
i
= log
evt,i = log
wi evt,i .
Zt
Zt
(t)

log

Using the inequality ea 1 a + a 2 /2, which holds for all a (0, 1), and the fact

(t)
that i wi = 1, we obtain
log


 (t) 
Z t+1
2
log
wi 1 vt,i + 2 vt,i
/2
Zt
i

 (t) 
2
wi vt,i 2 vt,i
/2 ).
= log(1
9

:;

<

def

=b

Next, note that b (0, 1). Therefore, taking log of the two sides of the inequality
1 b eb we obtain the inequality log (1 b) b, which holds for all b 1,

253

254

Online Learning

and obtain
log


 (t) 
Z t+1
2

wi vt,i 2 vt,i
/2
Zt
i
 (t)
2
= w(t) , vt  + 2
wi vt,i
/2
i

w , vt  + /2.
(t)

Summing this inequality over t we get

log (Z T +1 ) log (Z 1 ) =

T



Z t+1
T 2
.

w(t) , vt  +
Zt
2
T

log

t=1

(21.3)

t=1

(T +1)

Next, we lower bound Z T +1 . For each i , we can rewrite w i

= e t vt,i and we
get that










t vt,i
log Z T +1 = log
e
vt,i .
log max e t vt,i = min
i

Combining the preceding with Equation (21.3) and using the fact that log (Z 1 ) =
log (d) we get that
min
i

vt,i log (d)

T


w(t) , vt  +

t=1

T 2
,
2

which can be rearranged as follows:

T


w(t) , vt  min
i

t=1


t

vt,i

log (d) T
+
.

Plugging the value of into the equation concludes our proof.

Proof of Theorem 21.10

Equipped with the Weighted-Majority algorithm and Theorem 21.11, we are ready to
prove Theorem 21.10. We start with the simpler case, in which H is a finite class,
and let us write H = {h 1 , . . . , h d }. In this case, we can refer to each hypothesis, h i ,
as an expert, whose advice is to predict h i (xt ), and whose cost is vt,i = |h i (xt ) yt |.

(t)
The prediction of the algorithm will therefore be pt = i wi h i (xt ) [0, 1], and the
loss is



d
d






(t)
(t)
| p t yt | =
wi h i (xt ) yt =
wi (h i (xt ) yt ) .



i=1

i=1

Now, if yt = 1, then for all i , h i (xt ) yt 0. Therefore, the above equals to

(t)
i wi |h i (xt ) yt |. If yt = 0 then for all i , h i (xt ) yt 0, and the above also equals

21.2 Online Classification in the Unrealizable Case

(t)

wi |h i (xt ) yt |. All in all, we have shown that

| p t yt | =

d


(t)

wi |h i (xt ) yt | = w(t) , vt .

i=1

Furthermore, for each i , t vt,i is exactly the number of mistakes hypothesis h i
makes. Applying Theorem 21.11 we obtain
Corollary 21.12. Let H be a finite hypothesis class. There exists an algorithm for
online classification, whose predictions come from [0, 1], that enjoys the regret bound
T

t=1

| pt yt | min
hH

T


|h(xt ) yt |


2 log (|H|) T .

t=1

Next, we consider the case of a general hypothesis class. Previously, we constructed an expert for each individual hypothesis. However, if H is infinite this leads
to a vacuous bound. The main idea is to construct a set of experts in a more sophisticated way. The challenge is how to define a set of experts that, on one hand, is
not excessively large and, on the other hand, contains experts that give accurate
predictions.
We construct the set of experts so that for each hypothesis h H and every
sequence of instances, x1 , x2 , . . . , xT , there exists at least one expert in the set which
behaves exactly as h on these instances. For each L Ldim(H) and each sequence
1 i 1 < i 2 < < i L T we define an expert. The expert simulates the game between
SOA (presented in the previous section) and the environment on the sequence
of instances x1 , x2 , . . . , xT assuming that SOA makes a mistake precisely in rounds
i 1 ,i 2 , . . . ,i L . The expert is defined by the following algorithm.
Expert (i 1 ,i 2 , . . . ,i L )
input A hypothesis class H ; Indices i 1 < i 2 < < i L
initialize: V1 = H
for t = 1, 2, . . . , T
receive xt
(r)
for r {0, 1} let Vt = {h  Vt : h(x
 t) = r}
(r)

define yt = argmaxr Ldim Vt

(in case of a tie set yt = 0)

if t {i 1 ,i 2 , . . . ,i L }
predict yt = 1 y t
else
predict yt = yt
( y )
update Vt+1 = Vt t
Note that each such expert can give us predictions at every round t while only
observing the instances x1 , . . . , xt . Our generic online learning algorithm is now an
application of the Weighted-Majority algorithm with these experts.

255

256

Online Learning

To analyze the algorithm we first note that the number of experts is

Ldim(H) 

d=


L=0


T
.
L

(21.4)

It can be shown that when T Ldim(H) + 2, the right-hand side of the equation is

Ldim(H)
(the proof can be found in Lemma A.5).
bounded by eT /Ldim(H)
Theorem 21.11 tells us that the expected number of mistakes
 of WeightedMajority is at most the number of mistakes of the best expert plus 2 log (d) T . We
will next show that the number of mistakes of the best expert is at most the number
of mistakes of the best hypothesis in H. The following key lemma shows that, on
any sequence of instances, for each hypothesis h H there exists an expert with the
same behavior.
Lemma 21.13. Let H be any hypothesis class with Ldim(H) < . Let x1 , x2 , . . . , xT
be any sequence of instances. For any h H, there exists L Ldim(H) and indices
1 i 1 < i 2 < < i L T such that when running Expert (i 1 ,i 2 , . . . ,i L ) on the sequence
x1 , x2 , . . . , xT , the expert predicts h(xt ) on each online round t = 1, 2, . . . , T .
Proof. Fix h H and the sequence x1 , x2 , . . . , xT . We must construct L and the
indices i 1 ,i 2 , . . . ,i L . Consider running SOA on the input (x1 , h(x1 )), (x2 , h(x2 )), . . .,
(xT , h(xT )). SOA makes at most Ldim(H) mistakes on such input. We define L to
be the number of mistakes made by SOA and we define {i 1 ,i 2 , . . . ,i L } to be the set
of rounds in which SOA made the mistakes.
Now, consider the Expert (i 1 ,i 2 , . . . ,i L ) running on the sequence x1 , x2 , . . . , xT .
By construction, the set Vt maintained by Expert (i 1 ,i 2 , . . . ,i L ) equals the set Vt
maintained by SOA when running on the sequence (x1 , h(x1 )), . . . , (xT , h(xT )). The
predictions of SOA differ from the predictions of h if and only if the round is
in {i 1 ,i 2 , . . . ,i L }. Since Expert (i 1 ,i 2 , . . . ,i L ) predicts exactly like SOA if t is not
in {i 1 ,i 2 , . . . ,i L } and the opposite of SOAs predictions if t is in {i 1 ,i 2 , . . . ,i L }, we
conclude that the predictions of the expert are always the same as the predictions of h.
The previous lemma holds in particular for the hypothesis in H that makes the
least number of mistakes on the sequence of examples, and we therefore obtain the
following:
Corollary 21.14. Let (x1 , y1 ), (x2 , y2 ), . . . , (xT , yT ) be a sequence of examples and let
H be a hypothesis class with Ldim(H) < . There exists L Ldim(H) and indices
1 i 1 < i 2 < < i L T , such that Expert (i 1 ,i 2 , . . . ,i L ) makes at most as many
mistakes as the best h H does, namely,
min
hH

T


|h(xt ) yt |

t=1

mistakes on the sequence of examples.

Together with Theorem 21.11, the upper bound part of Theorem 21.10 is proven.

21.3 Online Convex Optimization

21.3 ONLINE CONVEX OPTIMIZATION

In Chapter 12 we studied convex learning problems and showed learnability results
for these problems in the agnostic PAC learning framework. In this section we
show that similar learnability results hold for convex problems in the online learning
framework. In particular, we consider the following problem.
Online Convex Optimization
definitions:
hypothesis class H ; domain Z ; loss function
: H Z R
assumptions:
H is convex
z Z ,
(, z) is a convex function
for t = 1, 2, . . . , T
learner predicts a vector w(t) H
environment responds with z t Z
learner suffers loss
(w(t) , z t )
As in the online classification problem, we analyze the regret of the algorithm.
Recall that the regret of an online algorithm with respect to a competing hypothesis,
which here will be some vector w H, is defined as


Regret A (w , T ) =

T


(w , z t )
(t)

t=1

T


(w , z t ).

(21.5)

t=1

As before, the regret of the algorithm relative to a set of competing vectors, H, is

defined as
Regret A (H, T ) = sup Regret A (w , T ).
w H

In Chapter 14 we have shown that Stochastic Gradient Descent solves convex

learning problems in the agnostic PAC model. We now show that a very similar
algorithm, Online Gradient Descent, solves online convex learning problems.
Online Gradient Descent
parameter: > 0
initialize: w(1) = 0
for t = 1, 2, . . . , T
predict w(t)
receive z t and let f t ( ) =
(, z t )
choose vt f t (w(t) )
update:
1

1. w(t+ 2 ) = w(t) vt
1

2. w(t+1) = argminwH w w(t+ 2 ) 

257

258

Online Learning

Theorem 21.15. The Online Gradient Descent algorithm enjoys the following regret
bound for every w H,
Regret A (w , T )

w 2 
+
vt 2 .
2
2
T

t=1

If we further assume that ft is -Lipschitz for all t, then setting = 1/ T yields

1
Regret A (w , T ) (w 2 + 2 ) T .
2
If we further assume that H is B-bounded and we set =
Regret A (H, T ) B

then

T.

Proof. The analysis is similar to the analysis of Stochastic Gradient Descent with
1
projections. Using the projection lemma, the definition of w(t+ 2 ) , and the definition
of subgradients, we have that for every t,
w(t+1) w 2 w(t) w 2
1

= w(t+1) w 2 w(t+ 2 ) w 2 + w(t+ 2 ) w 2 w(t) w 2

1

w(t+ 2 ) w 2 w(t) w 2
= w(t) vt w 2 w(t) w 2
= 2w(t) w , vt  + 2 vt 2
2( f t (w(t) ) f t (w )) + 2 vt 2 .
Summing over t and observing that the left-hand side is a telescopic sum we
obtain that
w(T +1) w 2 w(1) w 2 2

T


( ft (w(t) ) f t (w )) + 2

T


t=1

vt 2 .

t=1

Rearranging the inequality and using the fact that w(1) = 0, we get that
T


( ft (w(t) ) f t (w ))

t=1

w(1) w 2 w(T +1) w 2 

+
vt 2
2
2
T

t=1

w 2
+
2
2

T


vt 2 .

t=1

This proves the first bound in the theorem. The second bound follows from the
assumption that f t is -Lipschitz, which implies that vt  .

21.4 THE ONLINE PERCEPTRON ALGORITHM

The Perceptron is a classic online learning algorithm for binary classification with
the hypothesis class of homogenous halfspaces, namely, H = {x  sign(w, x) :

21.4 The Online Perceptron Algorithm

w Rd }. In Section 9.1.2 we have presented the batch version of the Perceptron,

which aims to solve the ERM problem with respect to H. We now present an online
version of the Perceptron algorithm.
Let X = Rd , Y = {1, 1}. On round t, the learner receives a vector xt Rd . The
learner maintains a weight vector w(t) Rd and predicts pt = sign(w(t) , xt ). Then,
it receives yt Y and pays 1 if pt = yt and 0 otherwise.
The goal of the learner is to make as few prediction mistakes as possible. In
Section 21.1 we characterized the optimal algorithm and showed that the best
achievable mistake bound depends on the Littlestone dimension of the class. We
show later that if d 2 then Ldim(H) = , which implies that we have no hope
of making few prediction mistakes. Indeed, consider the tree for which v1 =
( 12 , 1, 0, . . . , 0), v2 = ( 14 , 1, 0, . . . , 0), v3 = ( 34 , 1, 0, . . . , 0), etc. Because of the density
of the reals, this tree is shattered by the subset of H which contains all hypotheses that are parametrized by w of the form w = ( 1, a, 0, . . . , 0), for a [0, 1]. We
conclude that indeed Ldim(H) = .
To sidestep this impossibility result, the Perceptron algorithm relies on the technique of surrogate convex losses (see Section 12.3). This is also closely related to the
notion of margin we studied in Chapter 15.
A weight vector w makes a mistake on an example (x, y) whenever the sign of
w, x does not equal y. Therefore, we can write the 01 loss function as follows

(w, (x, y)) = 1[yw,x0] .
On rounds on which the algorithm makes a prediction mistake, we shall use the
hinge-loss as a surrogate convex loss function
f t (w) = max{0, 1 yt w, xt }.
The hinge-loss satisfies the two conditions:

ft is a convex function

For all w, f t (w)
(w, (xt , yt )). In particular, this holds for w(t) .

On rounds on which the algorithm is correct, we shall define ft (w) = 0. Clearly, f t

is convex in this case as well. Furthermore, f t (w(t) ) =
(w(t) , (xt , yt )) = 0.
Remark 21.5. In Section 12.3 we used the same surrogate loss function for all the
examples. In the online model, we allow the surrogate to depend on the specific
round. It can even depend on w(t) . Our ability to use a round specific surrogate
stems from the worst-case type of analysis we employ in online learning.
Let us now run the Online Gradient Descent algorithm on the sequence of functions, f1 , . . . , f T , with the hypothesis class being all vectors in Rd (hence, the
projection step is vacuous). Recall that the algorithm initializes w(1) = 0 and its
update rule is
w(t+1) = w(t) vt
for some vt f t (w(t) ). In our case, if yt w(t) , xt  > 0 then f t is the zero function and
we can take vt = 0. Otherwise, it is easy to verify that vt = yt xt is in f t (w(t) ). We

259

260

Online Learning

therefore obtain the update rule


w(t)
w(t+1) =
w(t) + yt xt

if yt w(t) , xt  > 0
otherwise

Denote by M the set of rounds in which sign(w(t) , xt ) = yt . Note that on round t,

the prediction of the Perceptron can be rewritten as



(t)
pt = sign(w , xt ) = sign
yi xi , xt  .
iM:i<t

This form implies that the predictions of the Perceptron algorithm and the set M
do not depend on the actual value of as long as > 0. We have therefore obtained
the Perceptron algorithm:
Perceptron
initialize: w1 = 0
for t = 1, 2, . . . , T
receive xt
predict pt = sign(w(t) , xt )
if yt w(t) , xt  0
w(t+1) = w(t) + yt xt
else
w(t+1) = w(t)
To analyze the Perceptron, we rely on the analysis of Online Gradient Descent
given in the previous section. In our case, the subgradient of f t we use in the
Perceptron is vt = 1[yt w(t) ,xt 0] yt xt . Indeed, the Perceptrons update is w(t+1) =

w(t) vt , and as discussed before this is equivalent to w(t+1) = w(t) vt for every
> 0. Therefore, Theorem 21.15 tells us that
T


f t (w(t) )

t=1

T


f t (w )

t=1

1

w 22 +
vt 22 .
2
2
T

t=1

Since f t (w(t) ) is a surrogate for the 01 loss we know that

Denote R = maxt xt ; then we obtain
|M|

T

t=1

Setting =

w 

R | M|

f t (w )

T

t=1 f t (w

(t) )

|M|.

w 22 + |M| R 2
2
2

and rearranging, we obtain

T


|M| Rw  |M|
f t (w ) 0.
t=1

This inequality implies

(21.6)

21.6 Bibliographic Remarks

Theorem 21.16. Suppose that the Perceptron algorithm runs on a sequence

(x1 , y1 ), . . . , (xT , yT ) and let R = maxt xt . Let M be the rounds on which the
Perceptron errs and let f t (w) = 1[tM] [1 yt w, xt ]+ . Then, for every w
*

|M|
f t (w ) + R w 
ft (w ) + R 2 w 2 .
t

In particular, if there exists w such that yt w , xt  1 for all t then

|M| R 2 w 2 .
Proof. The theorem follows from Equation (21.6) and the following claim: Given

x, b, c R+ , the inequality x b x c 0 implies that x c + b2 + b c. The last

claim can be easily derived by analyzing the roots of the convex parabola Q(y) =
y 2 by c.
The last assumption of Theorem 21.16 is called separability with large margin
(see Chapter 15). That is, there exists w that not only satisfies that the point xt lies
on the correct side of the halfspace, it also guarantees that xt is not too close to the
decision boundary. More specifically, the distance from xt to the decision boundary
is at least = 1/w  and the bound becomes (R/ )2 .
When the separability assumption does not hold, the bound involves the term
[1 yt w , xt ]+ which measures how much the separability with margin requirement is violated.
As a last remark we note that there can be cases in which there exists some w
that makes zero errors on the sequence but the Perceptron will make many errors.
Indeed, this is a direct consequence of the fact that Ldim(H) = . The way we
sidestep this impossibility result is by assuming more on the sequence of examples
the bound in Theorem 21.16 will be meaningful only if the cumulative surrogate


loss, t f t (w ) is not excessively large.

21.5 SUMMARY
In this chapter we have studied the online learning model. Many of the results we
derived for the PAC learning model have an analog in the online model. First, we
have shown that a combinatorial dimension, the Littlestone dimension, characterizes online learnability. To show this, we introduced the SOA algorithm (for the
realizable case) and the Weighted-Majority algorithm (for the unrealizable case).
We have also studied online convex optimization and have shown that online gradient descent is a successful online learner whenever the loss function is convex and
Lipschitz. Finally, we presented the online Perceptron algorithm as a combination
of online gradient descent and the concept of surrogate convex loss functions.

21.6 BIBLIOGRAPHIC REMARKS

The Standard Optimal Algorithm was derived by the seminal work of Littlestone
(1988). A generalization to the nonrealizable case, as well as other variants like
margin-based Littlestones dimension, were derived in (Ben-David et al. 2009).
Characterizations of online learnability beyond classification have been obtained

261

262

Online Learning

in (Abernethy, Bartlett, Rakhlin & Tewari 2008, Rakhlin, Sridharan & Tewari
2010, Daniely et al. 2011). The Weighted-Majority algorithm is due to (Littlestone
& Warmuth 1994) and (Vovk 1990).
The term online convex programming was introduced by Zinkevich (2003)
but this setting was introduced some years earlier by Gordon (1999). The Perceptron dates back to Rosenblatt (Rosenblatt 1958). An analysis for the realizable
case (with margin assumptions) appears in (Agmon 1954, Minsky & Papert 1969).
Freund and Schapire (Freund & Schapire 1999) presented an analysis for the
unrealizable case with a squared-hinge-loss based on a reduction to the realizable
case. A direct analysis for the unrealizable case with the hinge-loss was given by
Gentile (Gentile 2003).
For additional information we refer the reader to Cesa-Bianchi and Lugosi
(2006) and Shalev-Shwartz (2011).

21.7 EXERCISES
21.1 Find a hypothesis class H and a sequence of examples on which Consistent makes
|H| 1 mistakes.
21.2 Find a hypothesis class H and a sequence of examples on which the mistake bound
of the Halving algorithm is tight.
21.3 Let d 2, X = {1, . . . , d} and let H = {h j : j [d]}, where h j (x) = 1[x= j] . Calculate
MHalving (H) (i.e., derive lower and upper bounds on MHalving (H), and prove that
they are equal).
21.4 The Doubling Trick:
In Theorem 21.15, the parameter depends on the time horizon T . In this exercise
we show how to get rid of this dependence by a simple trick.

Consider an algorithm that enjoys a regret bound of the form T , but its
parameters require the knowledge of T . The doubling trick, described in the following, enables us to convert such an algorithm into an algorithm that does not need
to know the time horizon. The idea is to divide the time into periods of increasing
size and run the original algorithm on each period.

The Doubling Trick

input: algorithm A whose parameters depend on the time horizon
for m = 0, 1, 2, . . .
run A on the 2m rounds t = 2m , . . ., 2m+1 1

Show that if the regret of A on each period of 2m rounds is at most 2m , then the
total regret is at most

T.
21
21.5 Online-to-batch Conversions: In this exercise we demonstrate how a successful
online learning algorithm can be used to derive a successful PAC learner as well.
Consider a PAC learning problem for binary classification parameterized by an
instance domain, X , and a hypothesis class, H. Suppose that there exists an online
learning algorithm, A, which enjoys a mistake bound M A (H) < . Consider running this algorithm on a sequence of T examples which are sampled i.i.d. from a
distribution D over the instance space X , and are labeled by some h  H. Suppose

21.7 Exercises

that for every round t, the prediction of the algorithm is based on a hypothesis
h t : X {0, 1}. Show that
M A (H)
,
E [L D (h r )]
T
where the expectation is over the random choice of the instances as well as a random choice of r according to the uniform distribution over [T ].
Hint: Use similar arguments to the ones appearing in the proof of Theorem 14.8.

263

22
Clustering

Clustering is one of the most widely used techniques for exploratory data analysis.
Across all disciplines, from social sciences to biology to computer science, people
try to get a first intuition about their data by identifying meaningful groups among
the data points. For example, computational biologists cluster genes on the basis of
similarities in their expression in different experiments; retailers cluster customers,
on the basis of their customer profiles, for the purpose of targeted marketing; and
astronomers cluster stars on the basis of their spacial proximity.
The first point that one should clarify is, naturally, what is clustering? Intuitively,
clustering is the task of grouping a set of objects such that similar objects end up in
the same group and dissimilar objects are separated into different groups. Clearly,
this description is quite imprecise and possibly ambiguous. Quite surprisingly, it is
not at all clear how to come up with a more rigorous definition.
There are several sources for this difficulty. One basic problem is that the two
objectives mentioned in the earlier statement may in many cases contradict each
other. Mathematically speaking, similarity (or proximity) is not a transitive relation,
while cluster sharing is an equivalence relation and, in particular, it is a transitive
relation. More concretely, it may be the case that there is a long sequence of objects,
x 1 , . . . , x m such that each x i is very similar to its two neighbors, x i1 and x i+1 , but x 1
and x m are very dissimilar. If we wish to make sure that whenever two elements
are similar they share the same cluster, then we must put all of the elements of
the sequence in the same cluster. However, in that case, we end up with dissimilar
elements (x 1 and x m ) sharing a cluster, thus violating the second requirement.
To illustrate this point further, suppose that we would like to cluster the points
in the following picture into two clusters.

A clustering algorithm that emphasizes not separating close-by points (e.g., the
Single Linkage algorithm that will be described in Section 22.1) will cluster this input
264

22.0 Clustering

by separating it horizontally according to the two lines:

In contrast, a clustering method that emphasizes not having far-away points share
the same cluster (e.g., the 2-means algorithm that will be described in Section 22.1)
will cluster the same input by dividing it vertically into the right-hand half and the
left-hand half:

Another basic problem is the lack of ground truth for clustering, which is a
common problem in unsupervised learning. So far in the book, we have mainly dealt
with supervised learning (e.g., the problem of learning a classifier from labeled training data). The goal of supervised learning is clear we wish to learn a classifier
which will predict the labels of future examples as accurately as possible. Furthermore, a supervised learner can estimate the success, or the risk, of its hypotheses
using the labeled training data by computing the empirical loss. In contrast, clustering is an unsupervised learning problem; namely, there are no labels that we
try to predict. Instead, we wish to organize the data in some meaningful way.
As a result, there is no clear success evaluation procedure for clustering. In fact,
even on the basis of full knowledge of the underlying data distribution, it is not
clear what is the correct clustering for that data or how to evaluate a proposed
clustering.
Consider, for example, the following set of points in R2 :

265

266

Clustering

and suppose we are required to cluster them into two clusters. We have two highly
justifiable solutions:

This phenomenon is not just artificial but occurs in real applications. A given set
of objects can be clustered in various different meaningful ways. This may be due
to having different implicit notions of distance (or similarity) between objects, for
example, clustering recordings of speech by the accent of the speaker versus clustering them by content, clustering movie reviews by movie topic versus clustering
them by the review sentiment, clustering paintings by topic versus clustering them
by style, and so on.
To summarize, there may be several very different conceivable clustering solutions for a given data set. As a result, there is a wide variety of clustering algorithms
that, on some input data, will output very different clusterings.
A Clustering Model:
Clustering tasks can vary in terms of both the type of input they have and the type
of outcome they are expected to compute. For concreteness, we shall focus on the
following common setup:
Input a set of elements, X , and a distance function over it. That is, a function
d : X X R+ that is symmetric, satisfies d(x, x) = 0 for all x X , and often
also satisfies the triangle inequality. Alternatively, the function could be a similarity function s : X X [0, 1] that is symmetric and satisfies s(x, x) = 1
for all x X . Additionally, some clustering algorithms also require an input
parameter k (determining the number of required clusters).
Output a partition of the domain set X into subsets. That is, C = (C1 , . . . Ck )

where ki=1 Ci = X and for all i = j , Ci C j = . In some situations the
clustering is soft, namely, the partition of X into the different clusters is
probabilistic where the output is a function assigning to each domain point,
x X , a vector ( p1 (x), . . . , pk (x)), where pi (x) = P [x Ci ] is the probability
that x belongs to cluster Ci . Another possible output is a clustering dendrogram (from Greek dendron = tree, gramma = drawing), which is a hierarchical
tree of domain subsets, having the singleton sets in its leaves, and the full
domain as its root. We shall discuss this formulation in more detail in the
following.
In the following we survey some of the most popular clustering methods. In the
last section of this chapter we return to the high level discussion of what is clustering.

22.1 LINKAGE-BASED CLUSTERING ALGORITHMS

Linkage-based clustering is probably the simplest and most straightforward
paradigm of clustering. These algorithms proceed in a sequence of rounds. They

22.1 Linkage-Based Clustering Algorithms

start from the trivial clustering that has each data point as a single-point cluster.
Then, repeatedly, these algorithms merge the closest clusters of the previous clustering. Consequently, the number of clusters decreases with each such round. If kept
going, such algorithms would eventually result in the trivial clustering in which all of
the domain points share one large cluster. Two parameters, then, need to be determined to define such an algorithm clearly. First, we have to decide how to measure
(or define) the distance between clusters, and, second, we have to determine when
to stop merging. Recall that the input to a clustering algorithm is a between-points
distance function, d. There are many ways of extending d to a measure of distance
between domain subsets (or clusters). The most common ways are
1. Single Linkage clustering, in which the between-clusters distance is defined
by the minimum distance between members of the two clusters, namely,
def

D(A, B) = min{d(x, y) : x A, y B}
2. Average Linkage clustering, in which the distance between two clusters is
defined to be the average distance between a point in one of the clusters and
a point in the other, namely,
def

D(A, B) =

1
|A||B|

d(x, y)

x A, yB

3. Max Linkage clustering, in which the distance between two clusters is defined
as the maximum distance between their elements, namely,
def

D(A, B) = max{d(x, y) : x A, y B}.

The linkage-based clustering algorithms are agglomerative in the sense that they
start from data that is completely fragmented and keep building larger and larger
clusters as they proceed. Without employing a stopping rule, the outcome of such
an algorithm can be described by a clustering dendrogram: that is, a tree of domain
subsets, having the singleton sets in its leaves, and the full domain as its root. For
example, if the input is the elements X = {a, b, c, d, e} R2 with the Euclidean distance as depicted on the left, then the resulting dendrogram is the one depicted on
the right:
{a, b, c, d, e}
{b, c, d, e}

a
e

{b, c}

d
c
b

{a}

{b}

{c}

{d, e}
{d}

{e}

The single linkage algorithm is closely related to Kruskals algorithm for finding
a minimal spanning tree on a weighted graph. Indeed, consider the full graph whose
vertices are elements of X and the weight of an edge (x, y) is the distance d(x, y).
Each merge of two clusters performed by the single linkage algorithm corresponds
to a choice of an edge in the aforementioned graph. It is also possible to show that

267

268

Clustering

the set of edges the single linkage algorithm chooses along its run forms a minimal
spanning tree.
If one wishes to turn a dendrogram into a partition of the space (a clustering),
one needs to employ a stopping criterion. Common stopping criteria include

Fixed number of clusters fix some parameter, k, and stop merging clusters as
soon as the number of clusters is k.

Distance upper bound fix some r R+ . Stop merging as soon as all
the between-clusters distances are larger than r . We can also set r to be
max{d(x, y) : x, y X } for some < 1. In that case the stopping criterion is
called scaled distance upper bound.

22.2 k-MEANS AND OTHER COST MINIMIZATION CLUSTERINGS

Another popular approach to clustering starts by defining a cost function over a
parameterized set of possible clusterings and the goal of the clustering algorithm is
to find a partitioning (clustering) of minimal cost. Under this paradigm, the clustering task is turned into an optimization problem. The objective function is a function
from pairs of an input, (X , d), and a proposed clustering solution C = (C1 , . . . , Ck ),
to positive real numbers. Given such an objective function, which we denote by G,
the goal of a clustering algorithm is defined as finding, for a given input (X , d), a
clustering C so that G((X , d), C) is minimized. In order to reach that goal, one has
to apply some appropriate search algorithm.
As it turns out, most of the resulting optimization problems are NP-hard, and
some are even NP-hard to approximate. Consequently, when people talk about,
say, k-means clustering, they often refer to some particular common approximation
algorithm rather than the cost function or the corresponding exact solution of the
minimization problem.
Many common objective functions require the number of clusters, k, as a parameter. In practice, it is often up to the user of the clustering algorithm to choose the
parameter k that is most suitable for the given clustering problem.
In the following we describe some of the most common objective functions.
The k-means objective function is one of the most popular clustering objectives.
In k-means the data is partitioned into disjoint sets C1 , . . . , Ck where each Ci is
represented by a centroid i . It is assumed that the input set X is embedded in
some larger metric space (X  , d) (so that X X  ) and centroids are members
of X  . The k-means objective function measures the squared distance between
each point in X to the centroid of its cluster. The centroid of Ci is defined to be

d(x, )2 .
i (Ci ) = argmin
X 

xCi

Then, the k-means objective is

G kmeans ((X , d), (C1 , . . . , Ck )) =

k 

i=1 xCi

d(x, i (Ci ))2 .

22.2 k-Means and Other Cost Minimization Clusterings

This can also be rewritten as

G kmeans ((X , d), (C1 , . . . , Ck )) =

min

k 


1 ,...k X 

d(x, i )2 .

(22.1)

i=1 xCi

The k-means objective function is relevant, for example, in digital communication tasks, where the members of X may be viewed as a collection of signals
that have to be transmitted. While X may be a very large set of real valued vectors, digital transmission allows transmitting of only a finite number of bits for
each signal. One way to achieve good transmission under such constraints is to
represent each member of X by a close member of some finite set 1 , . . . k ,
and replace the transmission of any x X by transmitting the index of the
closest i . The k-means objective can be viewed as a measure of the distortion
created by such a transmission representation scheme.
The k-medoids objective function is similar to the k-means objective, except that
it requires the cluster centroids to be members of the input set. The objective
function is defined by
G Kmedoid ((X , d), (C1 , . . . , Ck )) =

k 


min

1 ,...k X

d(x, i )2 .

i=1 xCi

The k-median objective function is quite similar to the k-medoids objective,

except that the distortion between a data point and the centroid of its cluster
is measured by distance, rather than by the square of the distance:
G Kmedian ((X , d), (C1 , . . . , Ck )) =

min

k 


1 ,...k X

d(x, i ).

i=1 xCi

An example where such an objective makes sense is the facility location problem. Consider the task of locating k fire stations in a city. One can model
houses as data points and aim to place the stations so as to minimize the
average distance between a house and its closest fire station.
The previous examples can all be viewed as center-based objectives. The solution to such a clustering problem is determined by a set of cluster centers, and the
clustering assigns each instance to the center closest to it. More generally, centerbased objective is determined by choosing some monotonic function f : R+ R+
and then defining
G f ((X , d), (C1 , . . . Ck )) =

min

k 


1 ,...k X 

f (d(x, i )),

i=1 xCi

where X  is either X or some superset of X .

Some objective functions are not center based. For example, the sum of in-cluster
distances (SOD)
G SOD ((X , d), (C1 , . . . Ck )) =

k


i=1 x,yCi

d(x, y)

269

270

Clustering

and the MinCut objective that we shall discuss in Section 22.3 are not center-based
objectives.

22.2.1 The k-Means Algorithm

The k-means objective function is quite popular in practical applications of clustering. However, it turns out that finding the optimal k-means solution is often
computationally infeasible (the problem is NP-hard, and even NP-hard to approximate to within some constant). As an alternative, the following simple iterative
algorithm is often used, so often that, in many cases, the term k-means Clustering
refers to the outcome of this algorithm rather than to the clustering that minimizes
the k-means objective cost. We describe the algorithm with respect to the Euclidean
distance function d(x, y) = x y.
k-Means
input: X Rn ; Number of clusters k
initialize: Randomly choose initial centroids 1 , . . . , k
repeat until convergence
i [k] set Ci = {x X : i = argmin j x j }
(break ties in some arbitrary manner)


i [k] update i = |C1i | xCi x
Lemma 22.1. Each iteration of the k-means algorithm does not increase the k-means
objective function (as given in Equation (22.1)).
Proof. To simplify the notation, let us use the shorthand G(C1 , . . . , Ck ) for the
k-means objective, namely,
G(C1 , . . . , Ck ) =

min

k 


1 ,...,k Rn

x i 2 .

(22.2)

i=1 xCi

It is convenient to define (Ci ) = |C1i | xCi x and note that (Ci ) =


argminRn xCi x 2 . Therefore, we can rewrite the k-means objective as
G(C1 , . . . , Ck ) =

k 


x (Ci )2 .

(22.3)

i=1 xCi
(t1)

Consider the update at iteration t of the k-means algorithm. Let C1

(t1)

(t1)

(t)

k 


(t)

(t1)

, . . . , Ck

(t)

be the previous partition, let i

= (Ci
), and let C1 , . . . , Ck be the new
partition assigned at iteration t. Using the definition of the objective as given in
Equation (22.2) we clearly have that
(t)

G(C1 , . . . , Ck )

(t1) 2

x i

 .

(22.4)

i=1 xC (t)
i

(t)

(t)

In addition, the definition of the new partition (C1 , . . . , Ck ) implies that it

k

(t1) 2
 over all possible partitions
minimizes the expression
xCi x i
i=1

22.3 Spectral Clustering

(C1 , . . . , Ck ). Hence,
k 


(t1) 2

x i

i=1 xC (t)

k



(t1) 2

x i

 .

(22.5)

i=1 xC (t1)

Using Equation (22.3) we have that the right-hand side of Equation (22.5) equals
(t1)
(t1)
, . . . , Ck
). Combining this with Equation (22.4) and Equation (22.5), we
G(C1
(t)

(t)

(t1)

obtain that G(C1 , . . . , Ck ) G(C1

(t1)

, . . . , Ck

), which concludes our proof.

While the preceding lemma tells us that the k-means objective is monotonically
nonincreasing, there is no guarantee on the number of iterations the k-means algorithm needs in order to reach convergence. Furthermore, there is no nontrivial lower
bound on the gap between the value of the k-means objective of the algorithms
output and the minimum possible value of that objective function. In fact, k-means
might converge to a point which is not even a local minimum (see Exercise 22.2).
To improve the results of k-means it is often recommended to repeat the procedure
several times with different randomly chosen initial centroids (e.g., we can choose
the initial centroids to be random points from the data).

22.3 SPECTRAL CLUSTERING

Often, a convenient way to represent the relationships between points in a data set
X = {x 1 , . . . , x m } is by a similarity graph; each vertex represents a data point x i , and
every two vertices are connected by an edge whose weight is their similarity, Wi, j =
s(x i , x j ), where W Rm,m . For example, we can set Wi, j = exp( d(x i , x j )2 / 2 ),
where d(, ) is a distance function and is a parameter. The clustering problem can
now be formulated as follows: We want to find a partition of the graph such that the
edges between different groups have low weights and the edges within a group have
high weights.
In the clustering objectives described previously, the focus was on one side of
our intuitive definition of clustering making sure that points in the same cluster
are similar. We now present objectives that focus on the other requirement points
separated into different clusters should be nonsimilar.

22.3.1 Graph Cut

Given a graph represented by a similarity matrix W , the simplest and most direct
way to construct a partition of the graph is to solve the mincut problem, which
chooses a partition C1 , . . . , Ck that minimizes the objective
cut(C1 , . . . , Ck ) =

k


Wr,s .

/ i
i=1 rCi ,s C

For k = 2, the mincut problem can be solved efficiently. However, in practice it

often does not lead to satisfactory partitions. The problem is that in many cases, the
solution of mincut simply separates one individual vertex from the rest of the graph.

271

272

Clustering

Of course, this is not what we want to achieve in clustering, as clusters should be

reasonably large groups of points.
Several solutions to this problem have been suggested. The simplest solution is
to normalize the cut and define the normalized mincut objective as follows:
RatioCut(C1 , . . . , Ck ) =

k

1
|Ci |
i=1

Wr,s .

rCi ,s
/ Ci

The preceding objective assumes smaller values if the clusters are not too small.
Unfortunately, introducing this balancing makes the problem computationally hard
to solve. Spectral clustering is a way to relax the problem of minimizing RatioCut.

22.3.2 Graph Laplacian and Relaxed Graph Cuts

The main mathematical object for spectral clustering is the graph Laplacian matrix.
There are several different definitions of graph Laplacian in the literature, and in
the following we describe one particular definition.
Definition 22.2 (Unnormalized Graph Laplacian). The unnormalized graph Laplacian is
the m m matrix L = D W where D is a diagonal matrix with
Di,i = mj=1 Wi, j . The matrix D is called the degree matrix.
The following lemma underscores the relation between RatioCut and the
Laplacian matrix.
Lemma 22.3. Let C1 , . . . , Ck be a clustering and let H Rm,k be the matrix such that
Hi, j = 1

1
.
|C j | [iC j ]

Then, the columns of H are orthonormal to each other and

RatioCut(C1 , . . . , Ck ) = trace(H
L H ).
Proof. Let h1 , . . . , hk be the columns of H . The fact that these vectors are orthonormal is immediate from the definition.

Next, by standard algebraic manipulations, it
can be shown that trace(H
L H ) = ki=1 h

i Lhi and that for any vector v we have





1 
1

2
2
v Lv =
Dr,r vr 2
vr vs Wr,s +
Ds,s vs =
Wr,s (vr vs )2 .
2
2
r
r,s
s
r,s
Applying this with v = hi and noting that (h i,r h i,s )2 is nonzero only if r Ci , s
/ Ci
or the other way around, we obtain that
h

i Lhi =

1
|Ci |

Wr,s .

rCi ,s C
/ i

Therefore, to minimize RatioCut we can search for amatrix H whose columns

are orthonormal and such that each Hi, j is either 0 or 1/ |C j |. Unfortunately, this
is an integer programming problem which we cannot solve efficiently. Instead, we

22.4 Information Bottleneck

relax the latter requirement and simply search an orthonormal matrix H Rm,k
that minimizes trace(H
L H ). As we will see in the next chapter about PCA (particularly, the proof of Theorem 23.2), the solution to this problem is to set U to
be the matrix whose columns are the eigenvectors corresponding to the k minimal eigenvalues of L. The resulting algorithm is called Unnormalized Spectral
Clustering.

22.3.3 Unnormalized Spectral Clustering

Unnormalized Spectral Clustering
Input: W Rm,m ; Number of clusters k
Initialize: Compute the unnormalized graph Laplacian L
Let U Rm,k be the matrix whose columns are the eigenvectors of L
corresponding to the k smallest eigenvalues
Let v1 , . . . , vm be the rows of U
Cluster the points v1 , . . . , vm using k-means
Output: Clusters C1 , . . . , C K of the k-means algorithm
The spectral clustering algorithm starts with finding the matrix H of the k eigenvectors corresponding to the smallest eigenvalues of the graph Laplacian matrix. It
then represents points according to the rows of H . It is due to the properties of the
graph Laplacians that this change of representation is useful. In many situations,
this change of representation enables the simple k-means algorithm to detect the
clusters seamlessly. Intuitively, if H is as defined in Lemma 22.3 then each point in
the new representation is an indicator vector whose value is nonzero only on the
element corresponding to the cluster it belongs to.

22.4 INFORMATION BOTTLENECK*

The information bottleneck method is a clustering technique introduced by Tishby,
Pereira, and Bialek. It relies on notions from information theory. To illustrate the
method, consider the problem of clustering text documents where each document
is represented as a bag-of-words; namely, each document is a vector x = {0, 1}n ,
where n is the size of the dictionary and x i = 1 iff the word corresponding to index
i appears in the document. Given a set of m documents, we can interpret the bagof-words representation of the m documents as a joint probability over a random
variable x, indicating the identity of a document (thus taking values in [m]), and a
random variable y, indicating the identity of a word in the dictionary (thus taking
values in [n]).
With this interpretation, the information bottleneck refers to the identity of a
clustering as another random variable, denoted C, that takes values in [k] (where
k will be set by the method as well). Once we have formulated x, y, C as random
variables, we can use tools from information theory to express a clustering objective.
In particular, the information bottleneck objective is
min I (x; C) I (C; y) ,

p(C|x)

273

274

Clustering

where I (; ) is the mutual information between two random variables,1 is a parameter, and the minimization is over all possible probabilistic assignments of points to
clusters. Intuitively, we would like to achieve two contradictory goals. On one hand,
we would like the mutual information between the identity of the document and
the identity of the cluster to be as small as possible. This reflects the fact that we
would like a strong compression of the original data. On the other hand, we would
like high mutual information between the clustering variable and the identity of the
words, which reflects the goal that the relevant information about the document
(as reflected by the words that appear in the document) is retained. This generalizes
the classical notion of minimal sufficient statistics2 used in parametric statistics to
arbitrary distributions.
Solving the optimization problem associated with the information bottleneck
principle is hard in the general case. Some of the proposed methods are similar
to the EM principle, which we will discuss in Chapter 24.

22.5 A HIGH LEVEL VIEW OF CLUSTERING

So far, we have mainly listed various useful clustering tools. However, some fundamental questions remain unaddressed. First and foremost, what is clustering? What
is it that distinguishes a clustering algorithm from any arbitrary function that takes
an input space and outputs a partition of that space? Are there any basic properties
of clustering that are independent of any specific algorithm or task?
One method for addressing such questions is via an axiomatic approach. There
have been several attempts to provide an axiomatic definition of clustering. Let us
demonstrate this approach by presenting the attempt made by Kleinberg (2003).
Consider a clustering function, F, that takes as input any finite domain X with a
dissimilarity function d over its pairs and returns a partition of X .
Consider the following three properties of such a function:
Scale Invariance (SI) For any domain set X , dissimilarity function d, and any
def

> 0, the following should hold: F(X , d) = F(X , d) (where (d)(x, y) =

d(x, y)).
Richness (Ri) For any finite X and every partition C = (C1 , . . . Ck ) of X (into
nonempty subsets) there exists some dissimilarity function d over X such that
F(X , d) = C.
Consistency (Co) If d and d  are dissimilarity functions over X , such that for every
x, y X , if x, y belong to the same cluster in F(X , d) then d  (x, y) d(x, y)
and if x, y belong to different clusters in F(X , d) then d  (x, y) d(x, y), then
F(X , d) = F(X , d  ).
1






p(a,b)
,
That is, given a probability function, p over the pairs (x, C), I (x; C) = a b p(a, b) log p(a)
p(b)
where the sum is over all values x can take and all values C can take.
A sufficient statistic is a function of the data which has the property of sufficiency with respect to a
statistical model and its associated unknown parameter, meaning that no other statistic which can be
calculated from the same sample provides any additional information as to the value of the parameter.
For example, if we assume that a variable is distributed normally with a unit variance and an unknown
expectation, then the average function is a sufficient statistic.

22.5 A High Level View of Clustering

A moment of reflection reveals that the Scale Invariance is a very natural

requirement it would be odd to have the result of a clustering function depend
on the units used to measure between-point distances. The Richness requirement
basically states that the outcome of the clustering function is fully controlled by
the function d, which is also a very intuitive feature. The third requirement, Consistency, is the only requirement that refers to the basic (informal) definition of
clustering we wish that similar points will be clustered together and that dissimilar
points will be separated to different clusters, and therefore, if points that already
share a cluster become more similar, and points that are already separated become
even less similar to each other, the clustering function should have even stronger
support of its previous clustering decisions.
However, Kleinberg (2003) has shown the following impossibility result:
Theorem 22.4. There exists no function, F, that satisfies all the three properties: Scale
Invariance, Richness, and Consistency.
Proof. Assume, by way of contradiction, that some F does satisfy all three properties. Pick some domain set X with at least three points. By Richness, there must be
some d1 such that F(X , d1 ) = {{x} : x X } and there also exists some d2 such that
F(X , d2 ) = F(X , d1 ).
Let R+ be such that for every x, y X , d2 (x, y) d1 (x, y). Let d3 =
d2 . Consider F(X , d3 ). By the Scale Invariance property of F, we should have
F(X , d3 ) = F(X , d2 ). On the other hand, since all distinct x, y X reside in different clusters w.r.t. F(X , d1 ), and d3 (x, y) d1 (x, y), the Consistency of F implies
that F(X , d3 ) = F(X , d1 ). This is a contradiction, since we chose d1 , d2 so that
F(X , d2 ) = F(X , d1 ).
It is important to note that there is no single bad axiom there is no single
bad property among the three properties. For every pair of the three axioms,
there exist natural clustering functions that satisfy the two properties in that pair
(one can even construct such examples just by varying the stopping criteria for the
Single Linkage clustering function). On the other hand, Kleinberg shows that any
clustering algorithm that minimizes any center-based objective function inevitably
fails the consistency property (yet, the k-sum-of-in-cluster-distances minimization
clustering does satisfy Consistency).
The Kleinberg impossibility result can be easily circumvented by varying the
properties. For example, if one wishes to discuss clustering functions that have
a fixed number-of-clusters parameter, then it is natural to replace Richness by kRichness (namely, the requirement that every partition of the domain into k subsets
is attainable by the clustering function). k-Richness, Scale Invariance and Consistency all hold for the k-means clustering and are therefore consistent. Alternatively,
one can relax the Consistency property. For example, say that two clusterings
C = (C1 , . . . Ck ) and C  = (C1 , . . . Cl ) are compatible if for every clusters Ci C and
C j C  , either Ci C j or C j Ci or Ci C j = (it is worthwhile noting that for
every dendrogram, every two clusterings that are obtained by trimming that dendrogram are compatible). Refinement Consistency is the requirement that, under
the assumptions of the Consistency property, the new clustering F(X , d  ) is compatible with the old clustering F(X , d). Many common clustering functions satisfy this

275

276

Clustering

requirement as well as Scale Invariance and Richness. Furthermore, one can come
up with many other, different, properties of clustering functions that sound intuitive
and desirable and are satisfied by some common clustering functions.
Furthermore, one can come up with many other, different, properties of clustering functions that sound intuitive and desirable and are satisfied by some common
clustering functions.
There are many ways to interpret these results. We suggest to view it as indicating that there is no ideal clustering function. Every clustering function will
inevitably have some undesirable properties. The choice of a clustering function
for any given task must therefore take into account the specific properties of that
task. There is no generic clustering solution, just as there is no classification algorithm that will learn every learnable task (as the No-Free-Lunch theorem shows).
Clustering, just like classification prediction, must take into account some prior
knowledge about the specific task at hand.

22.6 SUMMARY
Clustering is an unsupervised learning problem, in which we wish to partition a set
of points into meaningful subsets. We presented several clustering approaches
including linkage-based algorithms, the k-means family, spectral clustering, and
the information bottleneck. We discussed the difficulty of formalizing the intuitive
meaning of clustering.

22.7 BIBLIOGRAPHIC REMARKS

The k-means algorithm is sometimes named Lloyds algorithm, after Stuart Lloyd,
who proposed the method in 1957. For a more complete overview of spectral clustering we refer the reader to the excellent tutorial by Von Luxburg (2007). The
information bottleneck method was introduced by Tishby, Pereira, and Bialek
(1999). For an additional discussion on the axiomatic approach see Ackerman and
Ben-David (2008).

22.8 EXERCISES
22.1 Suboptimality of k-Means: For every parameter t > 1, show that there exists an
instance of the k-means problem for which the k-means algorithm (might) find a
solution whose k-means objective is at least t OPT, where OPT is the minimum
k-means objective.
22.2 k-Means Might Not Necessarily Converge to a Local Minimum: Show that the kmeans algorithm might converge to a point which is not a local minimum. Hint:
Suppose that k = 2 and the sample points are {1, 2, 3, 4} R suppose we initialize
the k-means with the centers {2, 4}; and suppose we break ties in the definition of
Ci by assigning i to be the smallest value in argmin j x j .
22.3 Given a metric space (X , d), where |X | < , and k N, we would like to find a
partition of X into C1 , . . ., Ck which minimizes the expression
G kdiam ((X , d), (C1 , . . . , Ck )) = max diam(C j ),
j[d]

22.8 Exercises

where diam(C j ) = max x,x  C j d(x, x  ) (we use the convention diam(C j ) = 0 if
|C j | < 2).
Similarly to the k-means objective, it is NP-hard to minimize the k-diam objective. Fortunately, we have a very simple approximation algorithm: Initially, we pick
some x X and set 1 = x. Then, the algorithm iteratively sets
j {2, . . ., k}, j = argmax min d(x, i ).
xX

i[ j1]

Finally, we set
i [k], Ci = {x X : i = argmin d(x, j )}.
j[k]

Prove that the algorithm described is a 2-approximation algorithm. That is, if we

denote its output by C 1 , . . . , C k , and denote the optimal solution by C1 , . . ., Ck , then,
G kdiam ((X , d), (C 1 , . . ., C k )) 2 G kdiam ((X , d), (C1 , . . ., Ck )).
Hint: Consider the point k+1 (in other words, the next center we would have chosen, if we wanted k + 1 clusters). Let r = min j[k] d( j , k+1 ). Prove the following
inequalities
G kdiam ((X , d), (C 1 , . . ., C k )) 2r
G kdiam ((X, d), (C1 , . . ., Ck )) r .
22.4 Recall that a clustering function, F, is called Center-Based Clustering if, for
some monotonic function f : R+ R+ , on every given input (X , d), F(X , d) is
a clustering that minimizes the objective
G f ((X , d), (C1 , . . . Ck )) =

min

1 ,...k X 

k 


f (d(x, i )),

i=1 xCi

where X  is either X or some superset of X .

Prove that for every k > 1 the k-diam clustering function defined in the previous
exercise is not a center-based clustering function.
Hint: Given a clustering input (X , d), with |X | > 2, consider the effect of adding
many close-by points to some (but not all) of the members of X , on either the
k-diam clustering or any given center-based clustering.
22.5 Recall that we discussed three clustering properties: Scale Invariance, Richness,
and Consistency. Consider the Single Linkage clustering algorithm.
1. Find which of the three properties is satisfied by Single Linkage with the Fixed
Number of Clusters (any fixed nonzero number) stopping rule.
2. Find which of the three properties is satisfied by Single Linkage with the
Distance Upper Bound (any fixed nonzero upper bound) stopping rule.
3. Show that for any pair of these properties there exists a stopping criterion for
Single Linkage clustering, under which these two axioms are satisfied.
22.6 Given some number k, let k-Richness be the following requirement:
For any finite X and every partition C = (C1 , . . .Ck ) of X (into nonempty subsets)
there exists some dissimilarity function d over X such that F(X , d) = C.
Prove that, for every number k, there exists a clustering function that satisfies the
three properties: Scale Invariance, k-Richness, and Consistency.

277

23
Dimensionality Reduction

Dimensionality reduction is the process of taking data in a high dimensional space

and mapping it into a new space whose dimensionality is much smaller. This process
is closely related to the concept of (lossy) compression in information theory. There
are several reasons to reduce the dimensionality of the data. First, high dimensional
data impose computational challenges. Moreover, in some situations high dimensionality might lead to poor generalization abilities of the learning algorithm (for
example, in Nearest Neighbor classifiers the sample complexity increases exponentially with the dimensionsee Chapter 19). Finally, dimensionality reduction can
be used for interpretability of the data, for finding meaningful structure of the data,
and for illustration purposes.
In this chapter we describe popular methods for dimensionality reduction. In
those methods, the reduction is performed by applying a linear transformation to
the original data. That is, if the original data is in Rd and we want to embed it into
Rn (n < d) then we would like to find a matrix W Rn,d that induces the mapping
x  W x. A natural criterion for choosing W is in a way that will enable a reasonable
recovery of the original x. It is not hard to show that in general, exact recovery of x
from W x is impossible (see Exercise 23.1).
The first method we describe is called Principal Component Analysis (PCA).
In PCA, both the compression and the recovery are performed by linear transformations and the method finds the linear transformations for which the differences
between the recovered vectors and the original vectors are minimal in the least
squared sense.
Next, we describe dimensionality reduction using random matrices W . We
derive an important lemma, often called the Johnson-Lindenstrauss lemma,
which analyzes the distortion caused by such a random dimensionality reduction
technique.
Last, we show how one can reduce the dimension of all sparse vectors using
again a random matrix. This process is known as Compressed Sensing. In this case,
the recovery process is nonlinear but can still be implemented efficiently using linear
programming.

278

23.1 Principal Component Analysis (PCA)

We conclude by underscoring the underlying prior assumptions behind PCA

and compressed sensing, which can help us understand the merits and pitfalls of the
two methods.

23.1 PRINCIPAL COMPONENT ANALYSIS (PCA)

Let x1 , . . . , xm be m vectors in Rd . We would like to reduce the dimensionality of
these vectors using a linear transformation. A matrix W Rn,d , where n < d, induces
a mapping x  W x, where W x Rn is the lower dimensionality representation of x.
Then, a second matrix U Rd,n can be used to (approximately) recover each original
vector x from its compressed version. That is, for a compressed vector y = W x,
where y is in the low dimensional space Rn , we can construct x = U y, so that x is the
recovered version of x and resides in the original high dimensional space Rd .
In PCA, we find the compression matrix W and the recovering matrix U so that
the total squared distance between the original and recovered vectors is minimal;
namely, we aim at solving the problem
argmin
W Rn,d ,U Rd,n

m


xi U W xi 22 .

(23.1)

i=1

To solve this problem we first show that the optimal solution takes a specific
form.
Lemma 23.1. Let (U , W ) be a solution to Equation (23.1). Then the columns of U
are orthonormal (namely, U
U is the identity matrix of Rn ) and W = U
.
Proof. Fix any U, W and consider the mapping x  U W x. The range of this mapping, R = {U W x : x Rd }, is an n dimensional linear subspace of Rd . Let V Rd,n
be a matrix whose columns form an orthonormal basis of this subspace, namely, the
range of V is R and V
V = I . Therefore, each vector in R can be written as V y
where y Rn . For every x Rd and y Rn we have
x V y22 = x2 + y
V
V y 2y
V
x = x2 + y2 2y
(V
x),
where we used the fact that V
V is the identity matrix of Rn . Minimizing the preceding expression with respect to y by comparing the gradient with respect to y to
zero gives that y = V
x. Therefore, for each x we have that
V V
x = argmin x x 22 .
x R

In particular this holds for x1 , . . . , xm and therefore we can replace U , W by V , V

and by that do not increase the objective

m

i=1

xi U W xi 22

m


xi V V
xi 22 .

i=1

Since this holds for every U , W the proof of the lemma follows.

279

280

Dimensionality Reduction

On the basis of the preceding lemma, we can rewrite the optimization problem
given in Equation (23.1) as follows:
argmin

m


U Rd,n :U
U =I i=1

xi UU
xi 22 .

(23.2)

We further simplify the optimization problem by using the following elementary

algebraic manipulations. For every x Rd and a matrix U Rd,n such that U
U = I
we have
x UU
x2 = x2 2x
UU
x + x
UU
UU
x
= x2 x
UU
x
= x2 trace(U
xx
U ),

(23.3)

where the trace of a matrix is the sum of its diagonal entries. Since the trace is a
linear operator, this allows us to rewrite Equation (23.2) as follows:


m

argmax trace U

xi x

(23.4)
i U .
U Rd,n :U
U =I

i=1

m

Let A = i=1 xi x

i . The matrix A is symmetric and therefore it can be written

using its spectral decomposition as A = VDV
, where D is diagonal and V
V =
VV
= I . Here, the elements on the diagonal of D are the eigenvalues of A and
the columns of V are the corresponding eigenvectors. We assume without loss of
generality that D1,1 D2,2 Dd,d . Since A is positive semidefinite it also holds
that Dd,d 0. We claim that the solution to Equation (23.4) is the matrix U whose
columns are the n eigenvectors of A corresponding to the largest n eigenvalues.

Theorem 23.2. Let x1 , . . . , xm be arbitrary vectors in Rd , let A = m

i=1 xi xi , and let
u1 , . . . , un be n eigenvectors of the matrix A corresponding to the largest n eigenvalues
of A. Then, the solution to the PCA optimization problem given in Equation (23.1) is
to set U to be the matrix whose columns are u1 , . . . , un and to set W = U
.
Proof. Let VDV
be the spectral decomposition of A. Fix some matrix U Rd,n
with orthonormal columns and let B = V
U . Then, VB = VV
U = U . It follows
that
U
AU = B
V
VDV
VB = B
D B,
and therefore
trace(U
AU) =

d

j =1

D j, j

n


B 2j ,i .

i=1

Note that B
B = U
VV
U = U
U = I . Therefore, the columns of B are also



orthonormal, which implies that dj =1 ni=1 B 2j ,i = n. In addition, let B Rd,d be
a matrix such that its first n columns
are the columns of B and in addition
B
B = I .

n

d
2
2

Then, for every j we have i=1 B j ,i = 1, which implies that i=1 B j ,i 1. It

23.1 Principal Component Analysis (PCA)

follows that

trace(U AU)

max

[0,1]d : 1 n

d


D j, j j .

j =1

It is not hard to verify (see 23.2) that the right-hand side equals nj =1 D j , j . We have
therefore shown that for every matrix U Rd,n with orthonormal columns it holds


that trace(U
AU) nj =1 D j , j . On the other hand, if we set U to be the matrix
whose
columns are the n leading eigenvectors of A we obtain that trace(U
AU) =

n
j =1 D j , j , and this concludes our proof.
Remark 23.1. The proof

of Theorem 23.2 also tells us that the value of the objective
of Equation (23.4) is ni=1 Di,i . Combining this with Equation (23.3) and noting




2
A) = di=1 Di,i we obtain that the optimal objective value
that m
i=1 xi  = trace(

d
of Equation (23.1) is i=n+1 Di,i .
Remark 23.2. It is a common practice
to center the examples before applying
PCA. That is, we first calculate = m1 m
i=1 xi and then apply PCA on the vectors
(x1 ), . . . , (xm ). This is also related to the interpretation of PCA as variance
maximization (see Exercise 23.4).

23.1.1 A More Efficient Solution for the Case d % m

In some situations the original dimensionality of the data is much larger than the
number of examples m. The computational complexity of calculating the PCA solution as described previously is O(d 3 ) (for calculating eigenvalues of A) plus O(md 2 )
(for constructing the matrix A). We now show a simple trick that enables us to
calculate the PCA solution more efficiently when

d %
m.
Recall that the matrix A is defined to be m
i=1 xi xi . It is convenient to rewrite
A = X
X where X Rm,d is a matrix whose i th row is x

i . Consider the matrix

B = X X
. That is, B Rm,m is the matrix whose i , j element equals xi , x j . Suppose
that u is an eigenvector of B: That is, Bu = u for some R. Multiplying the
equality by X
and using the definition of B we obtain X
X X
u = X
u. But, using

the definition of A, we get that A(X

u) = (X
u). Thus, XX
uu is an eigenvector of
A with eigenvalue of .
We can therefore calculate the PCA solution by calculating the eigenvalues of
B instead of A. The complexity is O(m 3 ) (for calculating eigenvalues of B) and m 2 d
(for constructing the matrix B).
Remark 23.3. The previous discussion also implies that to calculate the PCA solution we only need to know how to calculate inner products between vectors. This
enables us to calculate PCA implicitly even when d is very large (or even infinite)
using kernels, which yields the kernel PCA algorithm.

23.1.2 Implementation and Demonstration

A pseudocode of PCA is given in the following.

281

282

Dimensionality Reduction

PCA
input
A matrix of m examples X Rm,d
number of components n
if (m > d)
A = X
X
Let u1 , . . . , un be the eigenvectors of A with largest eigenvalues
else
B = X X

Let v1 , . . . , vn be the eigenvectors of B with largest eigenvalues

for i = 1, . . . , n set ui = X
1 v  X
vi
i
output: u1 , . . . , un
To illustrate how PCA works, let us generate vectors in R2 that approximately
reside on a line, namely, on a one dimensional subspace of R2 . For example, suppose
that each example is of the form (x, x + y) where x is chosen uniformly at random
from [ 1, 1] and y is sampled from a Gaussian distribution with mean 0 and standard deviation of 0. 1. Suppose we apply PCA on this data. Then, the eigenvector

corresponding to the largest eigenvalue will be close to the vector (1/ 2, 1/ 2).
When projecting a point (x, x + y) on this principal component we will obtain the
. The reconstruction of the original vector will be ((x + y/2), (x + y/2)).
scalar 2x+y
2
In Figure 23.1 we depict the original versus reconstructed data.

1.5

0.5

0.5

1.5
1.5

0.5

0.5

1.5

Figure 23.1. A set of vectors in R2 (xs) and their reconstruction after dimensionality
reduction to R1 using PCA (circles).

23.2 Random Projections

o oo oo o

x
+

+++ +++

x xx x x
x
*

* *
*

**

Figure 23.2. Images of faces extracted from the Yale data set. Top-left: the original
images in R50x50 . Top-right: the images after dimensionality reduction to R10 and reconstruction. Middle row: an enlarged version of one of the images before and after PCA.
Bottom: the images after dimensionality reduction to R2 . The different marks indicate
different individuals.

Next, we demonstrate the effectiveness of PCA on a data set of faces. We

extracted images of faces from the Yale data set (Georghiades, Belhumeur &
Kriegman 2001). Each image contains 50 50 = 2500 pixels; therefore the original
dimensionality is very high.
Some images of faces are depicted on the top-left side of Figure 23.2. Using PCA,
we reduced the dimensionality to R10 and reconstructed back to the original dimension, which is 502 . The resulting reconstructed images are depicted on the top-right
side of Figure 23.2. Finally, on the bottom of Figure 23.2 we depict a 2 dimensional representation of the images. As can be seen, even from a 2 dimensional
representation of the images we can still roughly separate different individuals.

23.2 RANDOM PROJECTIONS

In this section we show that reducing the dimension by using a random linear transformation leads to a simple compression scheme with a surprisingly low distortion.
The transformation x  W x, when W is a random matrix, is often referred to
as a random projection. In particular, we provide a variant of a famous lemma

283

284

Dimensionality Reduction

due to Johnson and Lindenstrauss, showing that random projections do not distort
Euclidean distances too much.
Let x1 , x2 be two vectors in Rd . A matrix W does not distort too much the
distance between x1 and x2 if the ratio
W x1 W x2 
x1 x2 
is close to 1. In other words, the distances between x1 and x2 before and after the
transformation are almost the same. To show that W x1 W x2  is not too far away
from x1 x2  it suffices to show that W does not distort the norm of the difference
x
vector x = x1 x2 . Therefore, from now on we focus on the ratio W
x .
We start with analyzing the distortion caused by applying a random projection
to a single vector.
Lemma 23.3. Fix some x Rd . Let W Rn,d be a random matrix such that each Wi, j
is an independent normal random variable. Then, for every (0, 3) we have


 0

0(1/ n)W x0
02
2


1 > 2 e n/6 .
P
2


x
Proof. Without loss of generality we can assume that x2 = 1. Therefore, an
equivalent inequality is


2
P (1 )n W x2 (1 + )n 1 2e n/6 .
Let wi be the i th row of W . The random variable wi , x is a weighted sum of
d independent normal random variables and therefore it is normally distributed

2
2
with zero mean and variance
j x j = x = 1. Therefore, the random variable


n
W x2 = i=1 (wi , x)2 has a n2 distribution. The claim now follows directly from
a measure concentration property of 2 random variables stated in Lemma B.12
given in Section B.7.
The Johnson-Lindenstrauss lemma follows from this using a simple union bound
argument.
Lemma 23.4 (Johnson-Lindenstrauss Lemma). Let Q be a finite set of vectors in
Rd . Let (0, 1) and n be an integer such that

6 log (2|Q|/)
=
3.
n
Then, with probability of at least 1 over a choice of a random matrix W Rn,d
such that each element of W is distributed normally with zero mean and variance of
1/n we have


W x2



sup
1 < .
2


x
xQ

23.3 Compressed Sensing

Proof. Combining Lemma 23.3 and the union bound we have that for every
(0, 3):




W x2

2


1 > 2 |Q| e n/6 .
P sup
2


x
xQ
Let denote the right-hand side of the inequality; thus we obtain that

6 log (2|Q|/)
.
=
n

Interestingly, the bound given in Lemma 23.4 does not depend on the original
dimension of x. In fact, the bound holds even if x is in an infinite dimensional Hilbert
space.

23.3 COMPRESSED SENSING

Compressed sensing is a dimensionality reduction technique which utilizes a prior
assumption that the original vector is sparse in some basis. To motivate compressed
sensing, consider a vector x Rd that has at most s nonzero elements. That is,
def

x0 = |{i : x i = 0}| s.

Clearly, we can compress x by representing it using s (index,value) pairs. Furthermore, this compression is lossless we can reconstruct x exactly from the s
(index,value) pairs. Now, lets take one step forward and assume that x = U , where
is a sparse vector, 0 s, and U is a fixed orthonormal matrix. That is, x has a
sparse representation in another basis. It turns out that many natural vectors are (at
least approximately) sparse in some representation. In fact, this assumption underlies many modern compression schemes. For example, the JPEG-2000 format for
image compression relies on the fact that natural images are approximately sparse
in a wavelet basis.
Can we still compress x into roughly s numbers? Well, one simple way to do this
is to multiply x by U
, which yields the sparse vector , and then represent by its s
(index,value) pairs. However, this requires us first to sense x, to store it, and then
to multiply it by U
. This raises a very natural question: Why go to so much effort
to acquire all the data when most of what we get will be thrown away? Cannot we
just directly measure the part that will not end up being thrown away?
Compressed sensing is a technique that simultaneously acquires and compresses
the data. The key result is that a random linear transformation can compress x without losing information. The number of measurements needed is order of s log (d).
That is, we roughly acquire only the important information about the signal. As we
will see later, the price we pay is a slower reconstruction phase. In some situations,
it makes sense to save time in compression even at the price of a slower reconstruction. For example, a security camera should sense and compress a large amount of
images while most of the time we do not need to decode the compressed data at
all. Furthermore, in many practical applications, compression by a linear transformation is advantageous because it can be performed efficiently in hardware. For

285

286

Dimensionality Reduction

example, a team led by Baraniuk and Kelly has proposed a camera architecture
that employs a digital micromirror array to perform optical calculations of a linear
transformation of an image. In this case, obtaining each compressed measurement
is as easy as obtaining a single raw measurement. Another important application
of compressed sensing is medical imaging, in which requiring fewer measurements
translates to less radiation for the patient.
Informally, the main premise of compressed sensing is the following three
surprising results:
1. It is possible to reconstruct any sparse signal fully if it was compressed by
x  W x, where W is a matrix which satisfies a condition called the Restricted
Isoperimetric Property (RIP). A matrix that satisfies this property is guaranteed to have a low distortion of the norm of any sparse representable
vector.
2. The reconstruction can be calculated in polynomial time by solving a linear
program.
3. A random n d matrix is likely to satisfy the RIP condition provided that n is
greater than an order of s log (d).
Formally,
Definition 23.5 (RIP). A matrix W Rn,d is ( , s)-RIP if for all x = 0 s.t. x0 s we
have


W x2



2

1

.
x22

The first theorem establishes that RIP matrices yield a lossless compression
scheme for sparse vectors. It also provides a (nonefficient) reconstruction scheme.
Theorem 23.6. Let < 1 and let W be a ( , 2s)-RIP matrix. Let x be a vector s.t.
x0 s, let y = W x be the compression of x, and let
x argmin v0
v:W v=y

be a reconstructed vector. Then, x = x.

Proof. We assume, by way of contradiction, that x = x. Since x satisfies the constraints in the optimization problem that defines x we clearly have that x0 x0
s. Therefore, x x 0 2s and we can apply the RIP inequality on the vector x x .
But, since W (x x ) = 0 we get that |0 1| , which leads to a contradiction.
The reconstruction scheme given in Theorem 23.6 seems to be nonefficient
because we need to minimize a combinatorial objective (the sparsity of v). Quite
surprisingly, it turns out that we can replace the combinatorial objective, v0 , with
a convex objective, v1 , which leads to a linear programming problem that can be
solved efficiently. This is stated formally in the following theorem.
Theorem 23.7. Assume that the conditions of Theorem 23.6 holds and that <
Then,
x = argmin v0 = argmin v1 .
v:W v=y

v:W v=y

1
.
1+ 2

23.3 Compressed Sensing

In fact, we will prove a stronger result, which holds even if x is not a sparse
vector.
Theorem 23.8. Let <
vector and denote

1
1+ 2

and let W be a ( , 2s)-RIP matrix. Let x be an arbitrary

xs argmin x v1 .
v:v0 s

That is, xs is the vector which equals x on the s largest elements of x and equals 0
elsewhere. Let y = W x be the compression of x and let
x argmin v1
v:W v=y

be the reconstructed vector. Then,

x x2 2
where =

2 /(1 ).

1 + 1/2
s
x xs 1 ,
1

Note that in the special case that x = xs we get an exact recovery, x = x, so

Theorem 23.7 is a special case of Theorem 23.8. The proof of Theorem 23.8 is given
in Section 23.3.1.
Finally, the third result tells us that random matrices with n (s log (d)) are
likely to be RIP. In fact, the theorem shows that multiplying a random matrix by an
orthonormal matrix also provides an RIP matrix. This is important for compressing
signals of the form x = U where x is not sparse but is sparse. In that case, if W is a
random matrix and we compress using y = W x then this is the same as compressing
by y = (W U ) and since W U is also RIP we can reconstruct (and thus also x)
from y.
Theorem 23.9. Let U be an arbitrary fixed d d orthonormal matrix, let , be scalars
in (0, 1), let s be an integer in [d], and let n be an integer that satisfies
n 100

s log (40d/( ))
.
2

Let W Rn,d be a matrix s.t. each element of W is distributed normally with zero
mean and variance of 1/n. Then, with proabability of at least 1 over the choice of
W , the matrix W U is ( , s)-RIP.

23.3.1 Proofs*
Proof of Theorem 23.8
We follow a proof due to Cands (2008).
Let h = x x. Given a vector v and a set of indices I we denote by v I the vector
whose i th element is vi if i I and 0 otherwise.
The first trick we use is to partition the set of indices [d] = {1, . . . , d} into disjoint
sets of size s. That is, we will write [d] = T0 T1 T2 . . . Td/s1 where for all i , |Ti | =
s, and we assume for simplicity that d/s is an integer. We define the partition as
follows. In T0 we put the s indices corresponding to the s largest elements in absolute

287

288

Dimensionality Reduction

values of x (ties are broken arbitrarily). Let T0c = [d]\T0 . Next, T1 will be the s indices
corresponding to the s largest elements in absolute value of hT0c . Let T0,1 = T0 T1
c
and T0,1
= [d] \ T0,1 . Next, T2 will correspond to the s largest elements in absolute
c . And, we will construct T3 , T4 , . . . in the same way.
value of hT0,1
To prove the theorem we first need the following lemma, which shows that RIP
also implies approximate orthogonality.
Lemma 23.10 Let W be an ( , 2s)-RIP matrix. Then, for any two disjoint sets I , J ,
both of size at most s, and for any vector u we have that W u I , W u J  u I 2 u J 2 .
Proof. W.l.o.g. assume u I 2 = u J 2 = 1.
W u I , W u J  =

W u I + W u J 22 W u I W u J 22
.
4

But, since |J I | 2s we get from the RIP condition that W u I + W u J 22 (1 +

)(u I 22 + u J 22 ) = 2(1 + ) and that W u I W u J 22 (1 )(u I 22 + u J 22 ) =
2(1 ), which concludes our proof.
We are now ready to prove the theorem. Clearly,
c 2 hT 2 + hT c 2 .
h2 = hT0,1 + hT0,1
0,1
0,1

(23.5)

To prove the theorem we will show the following two claims:

c 2 hT 2 + 2s 1/2 x xs 1 .
Claim 1: hT0,1
0

Claim 2: hT0,1 2

2 1/2
x xs 1 .
1 s

Combining these two claims with Equation (23.5) we get that

1/2
c 2 2hT 2 + 2s
h2 hT0,1 2 + hT0,1
x xs 1
0,1


2
2 1
+ 1 s 1/2 x xs 1

=2

1 + 1/2
x xs 1 ,
s
1

and this will conclude our proof.

Proving Claim 1:
To prove this claim we do not use the RIP condition at all but only use the fact that
x minimizes the
1 norm. Take j > 1. For each i T j and i  T j 1 we have that
|h i | |h i  |. Therefore, hT j  hT j 1 1 /s. Thus,
hT j 2 s 1/2 hT j  s 1/2 hT j 1 1 .
Summing this over j = 2, 3, . . . and using the triangle inequality we obtain that

c 2
hT0,1
hT j 2 s 1/2 hT0c 1
(23.6)
j 2

23.3 Compressed Sensing

Next, we show that hT0c 1 cannot be large. Indeed, from the definition of x we
have that x1 x 1 = x + h1 . Thus, using the triangle inequality we obtain that


x1 x + h1 =
|x i + h i | +
|x i + h i | xT0 1 hT0 1 + hT0c 1 xT0c 1
iT0c

iT0

(23.7)
and since xT0c 1 = x xs 1 = x1 xT0 1 we get that
hT0c 1 hT0 1 + 2xT0c 1 .

(23.8)

Combining this with Equation (23.6) we get that



1/2
c 2 s
hT0 1 + 2xT0c 1 hT0 2 + 2s 1/2 xT0c 1 ,
hT0,1
which concludes the proof of claim 1.
Proving Claim 2:
For the second claim we use the RIP condition to get that
(1 )hT0,1 22 W hT0,1 22 .




Since W hT0,1 = W h j 2 W hT j = j 2 W hT j we have that


W hT0,1 22 =
W hT0,1 , W hT j  =
W hT0 + W hT1 , W hT j .
j 2

(23.9)

j 2

From the RIP condition on inner products we obtain that for all i {1, 2} and j 2
we have
|W hTi , W hT j | hTi 2 hT j 2 .

Since hT0 2 + hT1 2 2hT0,1 2 we therefore get that



W hT0,1 22 2 hT0,1 2

hT j 2 .
j 2

Combining this with Equation (23.6) and Equation (23.9) we obtain

(1 )hT0,1 22 2 hT0,1 2 s 1/2 hT0c 1 .

Rearranging the inequality gives

2 1/2
s
hT0c 1 .
hT0,1 2
1
Finally, using Equation (23.8) we get that
hT0,1 2 s 1/2 (hT0 1 + 2xT0c 1 ) hT0 2 + 2s 1/2 xT0c 1 ,
but since hT0 2 hT0,1 2 this implies
hT0,1 2

2 1/2
s
xT0c 1 ,
1

which concludes the proof of the second claim.

289

290

Dimensionality Reduction

Proof of Theorem 23.9

To prove the theorem we follow an approach due to (Baraniuk, Davenport, DeVore
& Wakin 2008). The idea is to combine the Johnson-Lindenstrauss (JL) lemma with
a simple covering argument.
We start with a covering property of the unit ball.
 d
Lemma 23.11 Let (0, 1). There exists a finite set Q Rd of size |Q| 3 such
that
sup min x v .
x:x1 vQ

Proof. Let k be an integer and let

Q  = {x Rd : j [d], i {k, k + 1, . . . , k} s. t. x j = ki }.
Clearly, |Q  | = (2k + 1)d . We shall set Q = Q  B2 (1), where B2 (1) is the unit
2 ball
of Rd . Since the points in Q  are distributed evenly on the unit
ball, the size of Q
is the size of Q  times the ratio between the volumes of the unit
2 and
balls. The
volume of the
ball is 2d and the volume of B2 (1) is
d/2
.
(1 + d/2)
For simplicity, assume that d is even and therefore
 d/2
(1 + d/2) = (d/2)! d/2
,
e
where in the last inequality we used Stirlings approximation. Overall we obtained
that
|Q| (2k + 1)d (/e)d/2 (d/2)d/2 2d .

(23.10)

Now lets us specify k. For each x B2 (1) let v Q be the vector whose i th element
is sign(x i ) |x i | k/k. Then, for each element we have that |x i vi | 1/k and thus

d
.
x v
k

To ensure that the right-hand side will be at most we shall set k =  d/ . Plugging
this value into Equation (23.10) we conclude that
  d  d

|Q| (3 d/(2 ))d (/e)d/2 (d/2)d/2 = 3 2e

3 .
Let x be a vector that can be written as x = U with U being some orthonormal
matrix and 0 s. Combining the earlier covering property and the JL lemma
(Lemma 23.4) enables us to show that a random W will not distort any such x.
Lemma 23.12 Let U be an orthonormal d d matrix and let I [d] be a set of
indices of size |I | = s. Let S be the span of {Ui : i I }, where Ui is the i th column of
U . Let (0, 1), (0, 1), and n N such that
n 24

log (2/) + s log (12/ )

.
2

23.3 Compressed Sensing

Then, with probability of at least 1 over a choice of a random matrix W Rn,d

such that each element of W is independently distributed according to N(0, 1/n), we
have


W x


sup
1 < .
x
xS
Proof. It suffices to prove the lemma for all x S with x = 1. We can write x = U I
where Rs , 2 = 1, and U I is the matrix whose columns are {Ui : i I }. Using
Lemma 23.11 we know that there exists a set Q of size |Q| (12/ )s such that
sup min  v ( /4).

:=1 vQ

But since U is orthogonal we also have that

sup min U I U I v ( /4).

:=1 vQ

Applying Lemma 23.4 on the set {U I v : v Q} we obtain that for n satisfying the
condition given in the lemma, the following holds with probability of at least 1 :



W U v2
I


1 /2,
sup
2


U
v
I
vQ
This also implies that

W U I v

1 /2.
sup
U I v
vQ

Let a be the smallest number such that

x S,

W x
1 + a.
x

Clearly a < . Our goal is to show that a . This follows from the fact that for any
x S of unit norm there exists v Q such that x U I v /4 and therefore
W x W U I v + W (x U I v) 1 + /2 + (1 + a) /4.
Thus,
x S,



W x
1 + /2 + (1 + a) /4 .
x

But the definition of a implies that

a /2 + (1 + a) /4 a
This proves that for all x S we have
as well since

W x
x

/2 + /4
.
1 /4

1 . The other side follows from this

W x W U I v W (x U I v) 1 /2 (1 + ) /4 1 .

The preceding lemma tells us that for x S of unit norm we have

(1 ) W x (1 + ),

291

292

Dimensionality Reduction

which implies that

(1 2 ) W x2 (1 + 3 ).
The proof of Theorem 23.9 follows from this by a union bound over all choices of I .

23.4 PCA OR COMPRESSED SENSING?

Suppose we would like to apply a dimensionality reduction technique to a given set
of examples. Which method should we use, PCA or compressed sensing? In this
section we tackle this question, by underscoring the underlying assumptions behind
the two methods.
It is helpful first to understand when each of the methods can guarantee perfect recovery. PCA guarantees perfect recovery whenever the set of examples is
contained in an n dimensional subspace of Rd . Compressed sensing guarantees
perfect recovery whenever the set of examples is sparse (in some basis). On the
basis of these observations, we can describe cases in which PCA will be better than
compressed sensing and vice versa.
As a first example, suppose that the examples are the vectors of the standard
basis of Rd , namely, e1 , . . . , ed , where each ei is the all zeros vector except 1 in the i th
coordinate. In this case, the examples are 1-sparse. Hence, compressed sensing will
yield a perfect recovery whenever n ( log (d)). On the other hand, PCA will lead
to poor performance, since the data is far from being in an n dimensional subspace,
as long as n < d. Indeed, it is easy ro verify that in such a case, the averaged recovery
error of PCA (i.e., the objective of Equation (23.1) divided by m) will be (d n)/d,
which is larger than 1/2 whenever n d/2.
We next show a case where PCA is better than compressed sensing. Consider
m examples that are exactly on an n dimensional subspace. Clearly, in such a case,
PCA will lead to perfect recovery. As to compressed sensing, note that the examples are n-sparse in any orthonormal basis whose first n vectors span the subspace.
Therefore, compressed sensing would also work if we will reduce the dimension
to (n log (d)). However, with exactly n dimensions, compressed sensing might
fail. PCA has also better resilience to certain types of noise. See (Chang, Weiss
& Freeman 2009) for a discussion.

23.5 SUMMARY
We introduced two methods for dimensionality reduction using linear transformations: PCA and random projections. We have shown that PCA is optimal in the
sense of averaged squared reconstruction error, if we restrict the reconstruction procedure to be linear as well. However, if we allow nonlinear reconstruction, PCA is
not necessarily the optimal procedure. In particular, for sparse data, random projections can significantly outperform PCA. This fact is at the heart of the compressed
sensing method.

23.6 BIBLIOGRAPHIC REMARKS

PCA is equivalent to best subspace approximation using singular value decomposition (SVD). The SVD method is described in Appendix C. SVD dates back to

23.7 Exercises

Eugenio Beltrami (1873) and Camille Jordan (1874). It has been rediscovered many
times. In the statistical literature, it was introduced by Pearson (1901). Besides PCA
and SVD, there are additional names that refer to the same idea and are being
used in different scientific communities. A few examples are the Eckart-Young theorem (after Carl Eckart and Gale Young who analyzed the method in 1936), the
Schmidt-Mirsky theorem, factor analysis, and the Hotelling transform.
Compressed sensing was introduced in Donoho (2006) and in (Candes & Tao
2005). See also Candes (2006).

23.7 EXERCISES
23.1 In this exercise we show that in the general case, exact recovery of a linear
compression scheme is impossible.
1. let A Rn,d be an arbitrary compression matrix where n d 1. Show that there
exists u, v Rn , u = v such that Au = Av.
2. Conclude that exact recovery of a linear compression scheme is impossible.
23.2 Let Rd such that 1 2 d 0. Show that
max

[0,1]d :1 n

d


jj =

j=1

n


j.

j=1

Hint: Take every vector [0, 1]d such that 1 n. Let i be the minimal index
for which i < 1. If i = n + 1 we are done. Otherwise, show that we can increase i ,
while possibly decreasing j for some j > i , and obtain a better solution. This will
imply that the optimal solution is to set i = 1 for i n and i = 0 for i > n.
23.3 Kernel PCA: In this exercise we show how PCA can be used for constructing nonlinear dimensionality reduction on the basis of the kernel trick (see
Chapter 16).
Let X be some instance space and let S = {x1 , . . . , xm } be a set of points in X .
Consider a feature mapping : X V , where V is some Hilbert space (possibly of infinite dimension). Let K : X X be a kernel function, that is, k(x, x ) =
(x), (x ). Kernel PCA is the process of mapping the elements in S into V
using , and then applying PCA over {(x1 ), . . ., (xm )} into Rn . The output of
this process is the set of reduced elements.
Show how this process can be done in polynomial time in terms of m and n,
assuming that each evaluation of K (, ) can be calculated in a constant time. In
particular, if your implementation requires multiplication of two matrices A and
B, verify that their product can be computed. Similarly, if an eigenvalue decomposition of some matrix C is required, verify that this decomposition can be
computed.
23.4 An Interpretation of PCA as Variance Maximization:
Let x1 , . . ., xm be m vectors in Rd , and let x be a random vector distributed according
to the uniform distribution over x1 , . . ., xm . Assume that E [x] = 0.
1. Consider the problem of finding a unit vector, w Rd , such that the random variable w, x has maximal variance. That is, we would like to solve the
problem
argmax Var[w, x] = argmax
w:w=1

w:w=1

m
1
(w, xi )2 .
m
i=1

293

294

Dimensionality Reduction

Show that the solution of the problem is to set w to be the first principle vector
of x1 , . . ., xm .
2. Let w1 be the first principal component as in the previous question. Now, suppose we would like to find a second unit vector, w2 Rd , that maximizes the
variance of w2 , x, but is also uncorrelated to w1 , x. That is, we would like to
solve
Var[w, x].
argmax
w:w=1, E [(w1 ,x)(w,x)]=0

Show that the solution to this problem is to set w to be the second principal
component of x1 , . . ., xm .
Hint: Note that

E [(w1 , x)(w, x)] = w

1 E [xx ]w = mw1 Aw,

where A = i xi x

i . Since w is an eigenvector of A we have that the constraint

E [(w1 , x)(w, x)] = 0 is equivalent to the constraint
w1 , w = 0.
23.5 The Relation between SVD and PCA: Use the SVD theorem (Corollary C.6) for
providing an alternative proof of Theorem 23.2.
23.6 Random Projections Preserve Inner Products: The Johnson-Lindenstrauss lemma
tells us that a random projection preserves distances between a finite set of vectors.
In this exercise you need to prove that if the set of vectors are within the unit ball,
then not only are the distances between any two vectors preserved, but the inner
product is also preserved.
Let Q be a finite set of vectors in Rd and assume that for every x Q we have
x 1.
1. Let (0, 1) and n be an integer such that
*
6 log (|Q|2 /)
=
3.
n
Prove that with probability of at least 1 over a choice of a random matrix
W Rn,d , where each element of W is independently distributed according to
N (0, 1/n), we have
|W u, W v u, v|
for every u, v Q.
(u+v)
(uv)
Hint: Use JL to bound both Wu+v
and Wuv
.
2. (*) Let x1 , . . ., xm be a set of vectors in Rd of norm at most 1, and assume that
these vectors are linearly separable with margin of . Assume that d % 1/ 2 .
Show that there exists a constant c > 0 such that if we randomly project these
vectors into Rn , for n = c/ 2 , then with probability of at least 99% it holds that
the projected vectors are linearly separable with margin /2.

24
Generative Models

We started this book with a distribution free learning framework; namely, we did not
impose any assumptions on the underlying distribution over the data. Furthermore,
we followed a discriminative approach in which our goal is not to learn the underlying distribution but rather to learn an accurate predictor. In this chapter we describe
a generative approach, in which it is assumed that the underlying distribution over
the data has a specific parametric form and our goal is to estimate the parameters of
the model. This task is called parametric density estimation.
The discriminative approach has the advantage of directly optimizing the quantity of interest (the prediction accuracy) instead of learning the underlying distribution. This was phrased as follows by Vladimir Vapnik in his principle for solving
problems using a restricted amount of information:
When solving a given problem, try to avoid a more general problem as an intermediate
step.
Of course, if we succeed in learning the underlying distribution accurately, we
are considered to be experts in the sense that we can predict by using the Bayes
optimal classifier. The problem is that it is usually more difficult to learn the underlying
distribution than to learn an accurate predictor. However, in some situations, it is
reasonable to adopt the generative learning approach. For example, sometimes it
is easier (computationally) to estimate the parameters of the model than to learn a
discriminative predictor. Additionally, in some cases we do not have a specific task at
hand but rather would like to model the data either for making predictions at a later
time without having to retrain a predictor or for the sake of interpretability of the data.
We start with a popular statistical method for estimating the parameters of the
data, which is called the maximum likelihood principle. Next, we describe two generative assumptions which greatly simplify the learning process. We also describe
the EM algorithm for calculating the maximum likelihood in the presence of latent
variables. We conclude with a brief description of Bayesian reasoning.

24.1 MAXIMUM LIKELIHOOD ESTIMATOR

Let us start with a simple example. A drug company developed a new drug to treat
some deadly disease. We would like to estimate the probability of survival when
295

296

Generative Models

using the drug. To do so, the drug company sampled a training set of m people and
gave them the drug. Let S = (x 1 , . . . , x m ) denote the training set, where for each i ,
x i = 1 if the i th person survived and x i = 0 otherwise. We can model the underlying
distribution using a single parameter, [0, 1], indicating the probability of survival.
We now would like to estimate the parameter on the basis of the training set S.
A natural idea is to use the average number of 1s in S as an estimator. That is,
1
xi .
=
m
m

(24.1)

i=1

= . That is, is an unbiased estimator of . Furthermore, since is

Clearly, E S []
the average of m i.i.d. binary random variables we can use Hoeffdings inequality to
get that with probability of at least 1 over the choice of S we have that

log (2/)

| |
.
(24.2)
2m
Another interpretation of is as the Maximum Likelihood Estimator, as we
formally explain now. We first write the probability of generating the sample S:
P [S = (x 1 , . . . , x m )] =

m


xi (1 )1xi =

i xi

(1 )

(1xi )

i=1

We define the log likelihood of S, given the parameter , as the log of the preceding
expression:




x i + log (1 )
(1 x i ).
L(S; ) = log P [S = (x 1 , . . . , x m )] = log ( )
i

The maximum likelihood estimator is the parameter that maximizes the likelihood
argmax L(S; ).

(24.3)

Next, we show that in our case, Equation (24.1) is a maximum likelihood estimator.
To see this, we take the derivative of L(S; ) with respect to and equate it to zero:




(1 x i )
i xi
i
= 0.

1
Solving the equation for we obtain the estimator given in Equation (24.1).

24.1.1 Maximum Likelihood Estimation for Continuous

Random Variables
Let X be a continuous random variable. Then, for most x R we have P [X = x] = 0
and therefore the definition of likelihood as given before is trivialized. To overcome
this technical problem we define the likelihood as log of the density of the probability of X at x. That is, given an i.i.d. training set S = (x 1 , . . . , x m ) sampled according
to a density distribution P we define the likelihood of S given as

m
m


P (x i ) =
log (P (x i )).
L(S; ) = log
i=1

i=1

24.1 Maximum Likelihood Estimator

As before, the maximum likelihood estimator is a maximizer of L( S; ) with respect

to .
As an example, consider a Gaussian random variable, for which the density
function of X is parameterized by = (, ) and is defined as follows:


1
(x )2
P (x) = exp
.
2 2
2
We can rewrite the likelihood as

1 
(x i )2 m log ( 2 ).
L( S; ) = 2
2
m

i=1

To find a parameter = (, ) that optimizes this we take the derivative of the

likelihood w.r.t. and w.r.t. and compare it to 0. We obtain the following two
equations:
1 
d
L( S; ) = 2
(x i ) = 0
d

i=1

1 
d
m
L( S; ) = 3
(x i )2 = 0
d

i=1

Solving the preceding equations we obtain the maximum likelihood estimates:

4
5
m
m
51 
1

=
xi
and = 6
(x i )
2
m
m
i=1

i=1

Note that the maximum likelihood estimate is not always an unbiased estimator.
For example, while
is unbiased, it is possible to show that the estimate of the
variance is biased (Exercise 24.1).
Simplifying Notation
To simplify our notation, we use P[X = x] in this chapter to describe both the probability that X = x (for discrete random variables) and the density of the distribution
at x (for continuous variables).

24.1.2 Maximum Likelihood and Empirical Risk Minimization

The maximum likelihood estimator shares some similarity with the Empirical Risk
Minimization (ERM) principle, which we studied extensively in previous chapters.
Recall that in the ERM principle we have a hypothesis class H and we use the
training set for choosing a hypothesis h H that minimizes the empirical risk. We
now show that the maximum likelihood estimator is an ERM for a particular loss
function.
Given a parameter and an observation x, we define the loss of on x as

(, x) = log (P [x]).

(24.4)

297

298

Generative Models

That is,
(, x) is the negation of the log-likelihood of the observation x, assuming the data is distributed according to P . This loss function is often referred to as
the log-loss. On the basis of this definition it is immediate that the maximum likelihood principle is equivalent to minimizing the empirical risk with respect to the loss
function given in Equation (24.4). That is,

argmin

m


( log (P [x i ])) = argmax

i=1

m


log (P [x i ]).

i=1

Assuming that the data is distributed according to a distribution P (not necessarily

of the parametric form we employ), the true risk of a parameter becomes
E [
(, x)] =
x

P[x] log (P [x])


9


 

P[x]
1
P[x] log
+
P[x] log
,
P [x]
P[x]
x
:;
< 9
:;
<
DRE [P ||P ]

(24.5)

H (P )

where DRE is called the relative entropy, and H is called the entropy function. The
relative entropy is a divergence measure between two probabilities. For discrete
variables, it is always nonnegative and is equal to 0 only if the two distributions are
the same. It follows that the true risk is minimal when P = P.
The expression given in Equation (24.5) underscores how our generative
assumption affects our density estimation, even in the limit of infinite data. It shows
that if the underlying distribution is indeed of a parametric form, then by choosing the correct parameter we can make the risk be the entropy of the distribution.
However, if the distribution is not of the assumed parametric form, even the best
parameter leads to an inferior model and the suboptimality is measured by the
relative entropy divergence.

24.1.3 Generalization Analysis

How good is the maximum likelihood estimator when we learn from a finite
training set?
To answer this question we need to define how we assess the quality of an approximated solution of the density estimation problem. Unlike discriminative learning,
where there is a clear notion of loss, in generative learning there are various ways
to define the loss of a model. On the basis of the previous subsection, one natural
candidate is the expected log-loss as given in Equation (24.5).
In some situations, it is easy to prove that the maximum likelihood principle
guarantees low true risk as well. For example, consider the problem of estimating the mean of a Gaussian variable of unit variance.

We saw previously that the
1
maximum likelihood estimator is the average:
= m i x i . Let  be the optimal

24.2 Naive Bayes

parameter. Then,



P [x]
[
(,
x)
( , x)] =
E
log
E
P [x]
x N( ,1)
x N( ,1)


1
1
 2
2

=
E
(x ) + (x )
2
2
x N( ,1)


2 ( )2

+ ( )

E
[x]
2
2
x N( ,1)

2 ( )2

+ ( )

2
2
1
 )2 .
= (
2
=

(24.6)

Next, we note that

is the average of m Gaussian variables and therefore it is also
distributed normally with mean  and variance  /m. From this fact we can derive
bounds of the form with probability of at least 1 we have that |
 | where

depends on /m and on .
In some situations, the maximum likelihood estimator clearly overfits. For example, consider a Bernoulli random variable X and let P[ X = 1] =  . As we saw
previously, using Hoeffdings inequality we can easily derive a guarantee on |  |
that holds with high probability (see Equation (24.2)). However, if our goal is to
obtain a small value of the expected log-loss function as defined in Equation (24.5)
we might fail. For example, assume that  is nonzero but very small. Then, the
probability that no element of a sample of size m will be 1 is (1  )m , which is

(2)
greater than e2 m . It follows that whenever m log
2  , the probability that the
sample is all zeros is at least 50%, and in that case, the maximum likelihood rule will
set = 0. But the true risk of the estimate = 0 is
x)] = 
( , 1) + (1  )
( , 0)
E [
(,

=  log (1/ ) + (1  ) log (1/(1 ))

=  log (1/0) = .
This simple example shows that we should be careful in applying the maximum
likelihood principle.
To overcome overfitting, we can use the variety of tools we encountered previously in the book. A simple regularization technique is outlined in Exercise 24.2.

24.2 NAIVE BAYES

The Naive Bayes classifier is a classical demonstration of how generative assumptions and parameter estimations simplify the learning process. Consider the problem
of predicting a label y {0, 1} on the basis of a vector of features x = (x 1 , . . . , x d ),
where we assume that each x i is in {0, 1}. Recall that the Bayes optimal classifier is
h Bayes (x) = argmax P[Y = y|X = x].
y{0,1}

299

300

Generative Models

To describe the probability function P[Y = y|X = x] we need 2d parameters, each

of which corresponds to P[Y = 1|X = x] for a certain value of x {0, 1}d . This
implies that the number of examples we need grows exponentially with the number
of features.
In the Naive Bayes approach we make the (rather naive) generative assumption
that given the label, the features are independent of each other. That is,
P[X = x|Y = y] =

d


P[X i = x i |Y = y].

i=1

With this assumption and using the Bayes rule, the Bayes optimal classifier can be
further simplified:
h Bayes (x) = argmax P[Y = y|X = x]
y{0,1}

= argmax P[Y = y]P[X = x|Y = y]

y{0,1}

= argmax P[Y = y]
y{0,1}

d


P[X i = x i |Y = y].

(24.7)

i=1

That is, now the number of parameters we need to estimate is only 2d + 1. Here, the
generative assumption we made reduced significantly the number of parameters we
need to learn.
When we also estimate the parameters using the maximum likelihood principle,
the resulting classifier is called the Naive Bayes classifier.

24.3 LINEAR DISCRIMINANT ANALYSIS

Linear discriminant analysis (LDA) is another demonstration of how generative
assumptions simplify the learning process. As in the Naive Bayes classifier we consider again the problem of predicting a label y {0, 1} on the basis of a vector of
features x = (x 1 , . . . , x d ). But now the generative assumption is as follows. First,
we assume that P[Y = 1] = P[Y = 0] = 1/2. Second, we assume that the conditional probability of X given Y is a Gaussian distribution. Finally, the covariance
matrix of the Gaussian distribution is the same for both values of the label. Formally,
let 0 , 1 Rd and let  be a covariance matrix. Then, the density distribution is
given by


1
1
T 1
P[X = x|Y = y] =
exp (x y )  (x y ) .
(2)d/2 ||1/2
2
As we have shown in the previous section, using the Bayes rule we can write
h Bayes (x) = argmax P[Y = y]P[X = x|Y = y].
y{0,1}

This means that we will predict h Bayes (x) = 1 iff



P[Y = 1]P[X = x|Y = 1]
log
> 0.
P[Y = 0]P[X = x|Y = 0]

24.4 Latent Variables and the EM Algorithm

This ratio is often called the log-likelihood ratio.

In our case, the log-likelihood ratio becomes
T 1
T 1
1
1
2 (x 0 )  (x 0 ) 2 (x 1 )  (x 1 )

We can rewrite this as w, x + b where

w = (1 0 )T  1 and b =

1
2



0T  1 0 1T  1 1 .

(24.8)

As a result of the preceding derivation we obtain that under the aforementioned

generative assumptions, the Bayes optimal classifier is a linear classifier. Additionally, one may train the classifier by estimating the parameter 0 , 1 and  from the
data, using, for example, the maximum likelihood estimator. With those estimators
at hand, the values of w and b can be calculated as in Equation (24.8).

24.4 LATENT VARIABLES AND THE EM ALGORITHM

In generative models we assume that the data is generated by sampling from a specific parametric distribution over our instance space X . Sometimes, it is convenient
to express this distribution using latent random variables. A natural example is a
mixture of k Gaussian distributions. That is, X = Rd and we assume that each x is
generated as follows. First, we choose a random number in {1, . . . , k}. Let Y be a
random variable corresponding to this choice, and denote P[Y = y] = c y . Second,
we choose x on the basis of the value of Y according to a Gaussian distribution
P[X = x|Y = y] =



1
1
T 1
(x

exp

)

(x

)
.
y
y
y
(2)d/2 | y |1/2
2

(24.9)

Therefore, the density of X can be written as:

P[X = x] =

k


P[Y = y]P[X = x|Y = y]

y=1

k

y=1



1
1
T 1
cy
exp (x y )  y (x y ) .
(2)d/2 | y |1/2
2

Note that Y is a hidden variable that we do not observe in our data. Nevertheless,
we introduce Y since it helps us describe a simple parametric form of the probability
of X.
More generally, let be the parameters of the joint distribution of X and Y (e.g.,
in the preceding example, consists of c y , y , and  y , for all y = 1, . . . , k). Then, the
log-likelihood of an observation x can be written as

k



log P [X = x] = log
P [X = x, Y = y] .
y=1

301

302

Generative Models

Given an i.i.d. sample, S = (x1 , . . . , xm ), we would like to find that maximizes

the log-likelihood of S,
L( ) = log

m


P [X = xi ]

i=1

m


log P [X = xi ]

i=1

m


log

i=1

k


P [X = xi , Y = y] .

y=1

The maximum-likelihood estimator is therefore the solution of the maximization

problem

m
k


log
P [X = xi , Y = y] .
argmax L( ) = argmax

i=1

y=1

In many situations, the summation inside the log makes the preceding optimization problem computationally hard. The Expectation-Maximization (EM) algorithm, due to Dempster, Laird, and Rubin, is an iterative procedure for searching a
(local) maximum of L( ). While EM is not guaranteed to find the global maximum,
it often works reasonably well in practice.
EM is designed for those cases in which, had we known the values of the latent
variables Y , then the maximum likelihood optimization problem would have been
tractable. More precisely, define the following function over m k matrices and the
set of parameters :
F(Q, ) =

m 
k




Q i,y log P [X = xi , Y = y] .

i=1 y=1

If each row of Q defines a probability over the i th latent variable given X = xi ,

then we can interpret F(Q, ) as the expected log-likelihood of a training set
(x1 , y1 ), . . . , (xm , ym ), where the expectation is with respect to the choice of each yi
on the basis of the i th row of Q. In the definition of F, the summation is outside
the log, and we assume that this makes the optimization problem with respect to
tractable:
Assumption 24.1. For any matrix Q [0, 1]m,k , such that each row of Q sums to 1,
the optimization problem
argmax F(Q, )

is tractable.
The intuitive idea of EM is that we have a chicken and egg problem. On one
hand, had we known Q, then by our assumption, the optimization problem of finding
the best is tractable. On the other hand, had we known the parameters we could
have set Q i,y to be the probability of Y = y given that X = xi . The EM algorithm

24.4 Latent Variables and the EM Algorithm

therefore alternates between finding given Q and finding Q given . Formally,

EM finds a sequence of solutions (Q (1) , (1) ), (Q (2) , (2) ), . . . where at iteration t, we
construct (Q (t+1) , (t+1) ) by performing two steps.

Expectation Step: Set
(t+1)

Q i,y

= P (t) [Y = y|X = xi ].

(24.10)

This step is called the Expectation step, because it yields a new probability over
the latent variables, which defines a new expected log-likelihood function over .

Maximization Step: Set (t+1) to be the maximizer of the expected loglikelihood, where the expectation is according to Q (t+1) :
(t+1) = argmax F(Q (t+1) , ).

(24.11)

By our assumption, it is possible to solve this optimization problem efficiently.

The initial values of (1) and Q (1) are usually chosen at random and the
procedure terminates after the improvement in the likelihood value stops being
significant.

24.4.1 EM as an Alternate Maximization Algorithm

To analyze the EM algorithm, we first view it as an alternate maximization
algorithm. Define the following objective function
G(Q, ) = F(Q, )

m 
k


Q i,y log (Q i,y ).

i=1 y=1

The second term is the sum of the entropies of the rows of Q. Let


Q = Q [0, 1]m,k : i ,
Q i,y = 1

y=1

be the set of matrices whose rows define probabilities over [k]. The following lemma
shows that EM performs alternate maximization iterations for maximizing G.
Lemma 24.2. The EM procedure can be rewritten as
Q (t+1) = argmax G(Q, (t) )
QQ

(t+1) = argmax G(Q (t+1) , ).

Furthermore, G(Q (t+1) , (t) ) = L( (t) ).

Proof. Given Q (t+1) we clearly have that
argmax G(Q (t+1) , ) = argmax F(Q (t+1) , ).

303

304

Generative Models

Therefore, we only need to show that for any , the solution of argmax QQ G(Q, )
is to set Q i,y = P [Y = y|X = xi ]. Indeed, by Jensens inequality, for any Q Q we
have that



m
k


P [X = xi , Y = y]

G(Q, ) =
Q i,y log
Q i,y
i=1
y=1

m
k


P
[X
=
x
,
Y
=
y]

i
log

Q i,y

Q i,y
i=1
y=1

m
k


log
P [X = xi , Y = y]
=
i=1

m


y=1



log P [X = xi ] = L( ),

i=1

while for Q i,y = P [Y = y|X = xi ] we have



m
k


[X
=
x
,
Y
=
y]
P

G(Q, ) =
P [Y = y|X = xi ] log
P [Y = y|X = xi ]
i=1

y=1

m 
k




P [Y = y|X = xi ] log P [X = xi ]

i=1 y=1

=
=

m


k


log P [X = xi ]
P [Y = y|X = xi ]

i=1

y=1

m




log P [X = xi ] = L( ).

i=1

This shows that setting Q i,y = P [Y = y|X = xi ] maximizes G(Q, ) over Q Q and
shows that G(Q (t+1) , (t) ) = L( (t) ).
The preceding lemma immediately implies:
Theorem 24.3. The EM procedure never decreases the log-likelihood; namely,
for all t,
L( (t+1) ) L( (t) ).
Proof. By the lemma we have
L( (t+1) ) = G(Q (t+2) , (t+1) ) G(Q (t+1) , (t) ) = L( (t) ).

24.4.2 EM for Mixture of Gaussians (Soft k-Means)

Consider the case of a mixture of k Gaussians in which is a triplet
(c, {1 , . . . , k }, {1 , . . . , k }) where P [Y = y] = c y and P [X = x|Y = y] is as given

24.5 Bayesian Reasoning

in Equation (24.9). For simplicity, we assume that 1 = 2 = = k = I , where

I is the identity matrix. Specifying the EM algorithm for this case we obtain the
following:

Expectation step: For each i [m] and y [k] we have that

1
P (t) [Y = y] P (t) [ X = xi |Y = y]
Zi


1 (t)
1
=
c y exp xi (yt) 2 ,
(24.12)
Zi
2


where Z i is a normalization factor which ensures that y P (t) [Y = y| X = xi ]
sums to 1.

Maximization step: We need to set t+1 to be a maximizer of Equation (24.11),
which in our case amounts to maximizing the following expression w.r.t. c and :
P (t) [Y = y| X = xi ] =

k
m 

i=1 y=1



1
P (t) [Y = y| X = xi ] log (c y ) xi y 2 .
2

(24.13)

Comparing the derivative of Equation (24.13) w.r.t. y to zero and rearranging

terms we obtain:
m

y =
P (t) [Y = y| X = xi ] xi .
i=1

That is, y is a weighted average of the xi where the weights are according to the
probabilities calculated in the E step. To find the optimal c we need to be more
careful since we must ensure that c is a probability vector. In Exercise 24.3 we
show that the solution is

m
P (t) [Y = y|X = xi ]
c y =
k i=1
.
(24.14)

m

y  =1
i=1 P (t) [Y = y |X = xi ]
It is interesting to compare the preceding algorithm to the k-means algorithm
described in Chapter 22. In the k-means algorithm, we first assign each example to a
cluster according to the distance xi y . Then, we update each center y according to the average of the examples assigned to this cluster. In the EM approach,
however, we determine the probability that each example belongs to each cluster.
Then, we update the centers on the basis of a weighted sum over the entire sample.
For this reason, the EM approach for k-means is sometimes called soft k-means.

24.5 BAYESIAN REASONING

The maximum likelihood estimator follows a frequentist approach. This means that
we refer to the parameter as a fixed parameter and the only problem is that we do
not know its value. A different approach to parameter estimation is called Bayesian
reasoning. In the Bayesian approach, our uncertainty about is also modeled using
probability theory. That is, we think of as a random variable as well and refer to the
distribution P[ ] as a prior distribution. As its name indicates, the prior distribution
should be defined by the learner prior to observing the data.

305

306

Generative Models

As an example, let us consider again the drug company which developed a new
drug. On the basis of past experience, the statisticians at the drug company believe
that whenever a drug has reached the level of clinic experiments on people, it is
likely to be effective. They model this prior belief by defining a density distribution
on such that

0. 8 if > 0. 5
P[ ] =
(24.15)
0. 2 if 0. 5
As before, given a specific value of , it is assumed that the conditional probability,
P[X = x| ], is known. In the drug company example, X takes values in {0, 1} and
P[X = x| ] = x (1 )1x .
Once the prior distribution over and the conditional distribution over X given
are defined, we again have complete knowledge of the distribution over X. This is
because we can write the probability over X as a marginal probability


P[X = x] =
P[X = x, ] =
P[ ]P[X = x| ],

where the last equality follows from the definition of conditional probability. If
is continuous we replace P[ ] with the density function and the sum becomes an
integral:
F
P[X = x] = P[ ]P[X = x| ] d .

Seemingly, once we know P[X = x], a training set S = (x 1 , . . . , x m ) tells us nothing

as we are already experts who know the distribution over a new point X. However,
the Bayesian view introduces dependency between S and X. This is because we
now refer to as a random variable. A new point X and the previous points in S are
independent only conditioned on . This is different from the frequentist philosophy
in which is a parameter that we might not know, but since it is just a parameter of
the distribution, a new point X and previous points S are always independent.
In the Bayesian framework, since X and S are not independent anymore, what
we would like to calculate is the probability of X given S, which by the chain rule
can be written as follows:


P[X = x|, S] P[ |S] =
P[X = x| ] P[ |S].
P[X = x|S] =

The second inequality follows from the assumption that X and S are independent
when we condition on . Using the Bayes rule we have
P[ |S] =

P[S| ] P[ ]
,
P[S]

and together with the assumption that points are independent conditioned on , we
can write
m
P[S| ] P[ ]
1 
P[ |S] =
=
P[X = x i | ] P[ ].
P[S]
P[S]
i=1

24.7 Bibliographic Remarks

We therefore obtain the following expression for Bayesian prediction:


1 
P[X = x| ]
P[X = x i | ] P[ ].
P[S]
m

P[X = x|S] =

(24.16)

i=1

Getting back to our drug company example, we can rewrite P[X = x|S] as
F




1
P[X = x|S] =
x+ i xi (1 )1x+ i (1xi ) P[ ] d .
P[S]
It is interesting to note that when P[ ] is uniform we obtain that
F




P[X = x|S] x+ i xi (1 )1x+ i (1xi ) d .
Solving the preceding integral (using integration by parts) we obtain


( i xi ) + 1
.
P[X = 1|S] =
m +2
Recall that the prediction
according to the maximum likelihood principle in this


i xi

case is P[X = 1|] = m . The Bayesian prediction with uniform prior is rather
similar to the maximum likelihood prediction, except it adds pseudoexamples to
the training set, thus biasing the prediction toward the uniform prior.
Maximum A Posteriori
In many situations, it is difficult to find a closed form solution to the integral given
in Equation (24.16). Several numerical methods can be used to approximate this
integral. Another popular solution is to find a single which maximizes P[ |S].
The value of which maximizes P[ |S] is called the Maximum A Posteriori estimator. Once this value is found, we can calculate the probability that X = x given the
maximum a posteriori estimator and independently on S.

24.6 SUMMARY
In the generative approach to machine learning we aim at modeling the distribution
over the data. In particular, in parametric density estimation we further assume that
the underlying distribution over the data has a specific parametric form and our goal
is to estimate the parameters of the model. We have described several principles
for parameter estimation, including maximum likelihood, Bayesian estimation, and
maximum a posteriori. We have also described several specific algorithms for implementing the maximum likelihood under different assumptions on the underlying
data distribution, in particular, Naive Bayes, LDA, and EM.

24.7 BIBLIOGRAPHIC REMARKS

The maximum likelihood principle was studied by Ronald Fisher in the beginning
of the 20th century. Bayesian statistics follow the Bayes rule, which is named after
the 18th century English mathematician Thomas Bayes.

307

308

Generative Models

There are many excellent books on the generative and Bayesian approaches
to machine learning. See, for example, (Bishop 2006, Koller & Friedman 2009b,
MacKay 2003, Murphy 2012, Barber 2012).

24.8 EXERCISES
24.1 Prove that the maximum likelihood estimator of the variance of a Gaussian variable
is biased.
24.2 Regularization for Maximum Likelihood: Consider the following regularized loss
minimization:
m

1
1
log (1/P [xi ]) +
log (1/ ) + log(1/(1 )) .
m
m
i=1

Show that the preceding objective is equivalent to the usual empirical error
had we added two pseudoexamples to the training set. Conclude that the
regularized maximum likelihood estimator would be


m

1
xi .
=
1+
m +2
i=1

Derive a high probability bound on |  |. Hint: Rewrite this as | E [ ] +

E [ ]  | and then use the triangle inequality and Hoeffding inequality.
1

Use this to bound the true risk. Hint: Use the fact that now m+2
to relate


| | to the relative entropy.

24.3 Consider a general optimization problem of the form
max
c

k

y=1

y log (c y ) s.t. c y > 0,

c y = 1,

where Rk+ is a vector of nonnegative weights.

Verify that the M step of soft k-means involves solving such an optimization
problem.

Let c =
1 y . Show that c is a probability vector.
y

Show that the optimization problem is equivalent to the problem


c y = 1.
min DRE (c ||c) s.t. c y > 0,
c

Using properties of the relative entropy, conclude that c is the solution to the
optimization problem.

25
Feature Selection and Generation

In the beginning of the book, we discussed the abstract model of learning, in which
the prior knowledge utilized by the learner is fully encoded by the choice of the
hypothesis class. However, there is another modeling choice, which we have so far
ignored: How do we represent the instance space X ? For example, in the papayas
learning problem, we proposed the hypothesis class of rectangles in the smoothnesscolor two dimensional plane. That is, our first modeling choice was to represent a
papaya as a two dimensional point corresponding to its smoothness and color. Only
after that did we choose the hypothesis class of rectangles as a class of mappings
from the plane into the label set. The transformation from the real world object
papaya into the scalar representing its smoothness or its color is called a feature
function or a feature for short; namely, any measurement of the real world object
can be regarded as a feature. If X is a subset of a vector space, each x X is sometimes referred to as a feature vector. It is important to understand that the way we
encode real world objects as an instance space X is by itself prior knowledge about
the problem.
Furthermore, even when we already have an instance space X which is represented as a subset of a vector space, we might still want to change it into a different
representation and apply a hypothesis class on top of it. That is, we may define a
hypothesis class on X by composing some class H on top of a feature function which
maps X into some other vector space X  . We have already encountered examples
of such compositions in Chapter 15 we saw that kernel-based SVM learns a composition of the class of halfspaces over a feature mapping that maps each original
instance in X into some Hilbert space. And, indeed, the choice of is another form
of prior knowledge we impose on the problem.
In this chapter we study several methods for constructing a good feature set. We
start with the problem of feature selection, in which we have a large pool of features and our goal is to select a small number of features that will be used by our
predictor. Next, we discuss feature manipulations and normalization. These include
simple transformations that we apply on our original features. Such transformations may decrease the sample complexity of our learning algorithm, its bias, or its

309

310

Feature Selection and Generation

computational complexity. Last, we discuss several approaches for feature learning.

In these methods, we try to automate the process of feature construction.
We emphasize that while there are some common techniques for feature learning one may want to try, the No-Free-Lunch theorem implies that there is no
ultimate feature learner. Any feature learning algorithm might fail on some problem. In other words, the success of each feature learner relies (sometimes implicitly)
on some form of prior assumption on the data distribution. Furthermore, the relative quality of features highly depends on the learning algorithm we are later going
to apply using these features. This is illustrated in the following example.
Example 25.1. Consider a regression problem in which X = R2 , Y = R, and the
loss function is the squared loss. Suppose that the underlying distribution is such
that an example (x, y) is generated as follows: First, we sample x 1 from the uniform
distribution over [ 1, 1]. Then, we deterministically set y = x 1 2 . Finally, the second
feature is set to be x 2 = y + z, where z is sampled from the uniform distribution over
[ 0. 01, 0. 01]. Suppose we would like to choose a single feature. Intuitively, the first
feature should be preferred over the second feature as the target can be perfectly
predicted based on the first feature alone, while it cannot be perfectly predicted
based on the second feature. Indeed, choosing the first feature would be the right
choice if we are later going to apply polynomial regression of degree at least 2. However, if the learner is going to be a linear regressor, then we should prefer the second
feature over the first one, since the optimal linear predictor based on the first feature
will have a larger risk than the optimal linear predictor based on the second feature.

25.1 FEATURE SELECTION

Throughout this section we assume that X = Rd . That is, each instance is represented as a vector of d features. Our goal is to learn a predictor that only relies
on k & d features. Predictors that use only a small subset of features require a
smaller memory footprint and can be applied faster. Furthermore, in applications
such as medical diagnostics, obtaining each possible feature (e.g., test result) can
be costly; therefore, a predictor that uses only a small number of features is desirable
even at the cost of a small degradation in performance, relative to a predictor that
uses more features. Finally, constraining the hypothesis class to use a small subset
of features can reduce its estimation error and thus prevent overfitting.
Ideally, we could have tried all subsets of k out of d features and choose the
subset which leads to the best performing predictor. However, such an exhaustive
search is usually computationally intractable. In the following we describe three
computationally feasible approaches for feature selection. While these methods
cannot guarantee finding the optimal subset, they often work reasonably well in
practice. Some of the methods come with formal guarantees on the quality of the
selected subsets under certain assumptions. We do not discuss these guarantees
here.

25.1.1 Filters
Maybe the simplest approach for feature selection is the filter method, in which
we assess individual features, independently of other features, according to some

25.1 Feature Selection

quality measure. We can then select the k features that achieve the highest score
(alternatively, decide also on the number of features to select according to the value
of their scores).
Many quality measures for features have been proposed in the literature. Maybe
the most straightforward approach is to set the score of a feature according to the
error rate of a predictor that is trained solely by that feature.
To illustrate this, consider a linear regression problem with the squared loss. Let
v = (x 1, j , . . . , x m, j ) Rm be a vector designating the values of the j th feature on a
training set of m examples and let y = ( y1 , . . . , ym ) Rm be the values of the target on
the same m examples. The empirical squared loss of an ERM linear predictor that
uses only the j th feature would be
min

a,bR

1
av + b y2 ,
m

where the meaning of adding a scalar b to a vector v is adding b to all coordinates of

v. To solve this problem, let v = m1 m
i=1 vi be the averaged value of the feature and


let y = m1 m
y
be
the
averaged
value
of the target. Clearly (see Exercise 25.1),
i=1 i
1
1
av + b y2 = min a(v v)
+ b (y y )2 .
a,bR m
a,bR m
min

(25.1)

Taking the derivative of the right-hand side objective with respect to b and comparing it to zero we obtain that b = 0. Similarly, solving for a (once we know that
b = 0) yields a = v v,
y y /v v
2 . Plugging this value back into the objective
we obtain the value
y y 2

(v v,
y y )2
.
v v
2

Ranking the features according to the minimal loss they achieve is equivalent to
ranking them according to the absolute value of the following score (where now a
higher score yields a better feature):
1
v v,
y y 
v v,
y y 

=  m
.
v v
y y 
1
1
2
2
v

v

y


m
m

(25.2)

The preceding expression is known as Pearsons correlation coefficient. The numerator is the empirical estimate of the covariance of the j th feature and the target
value, E [(v Ev)(y E y)], while the denominator is the squared root of the empirical estimate for the variance of the j th feature, E [(v E v)2 ], times the variance of
the target. Pearsons coefficient ranges from 1 to 1, where if the Pearsons coefficient is either 1 or 1, there is a linear mapping from v to y with zero empirical
risk.
If Pearsons coefficient equals zero it means that the optimal linear function from
v to y is the all-zeros function, which means that v alone is useless for predicting y.
However, this does not mean that v is a bad feature, as it might be the case that
together with other features v can perfectly predict y. Indeed, consider a simple
example in which the target is generated by the function y = x 1 + 2x 2 . Assume also
that x 1 is generated from the uniform distribution over {1}, and x 2 = 12 x 1 + 12 z,

311

312

Feature Selection and Generation

where z is also generated i.i.d. from the uniform distribution over {1}. Then,
E [x 1 ] = E [x 2 ] = E [y] = 0, and we also have
E [yx 1] = E [x 12 ] + 2 E[x 2 x 1 ] = E [x 12 ] E[x 12 ] + E [zx 1 ] = 0.
Therefore, for a large enough training set, the first feature is likely to have a
Pearsons correlation coefficient that is close to zero, and hence it will most probably not be selected. However, no function can predict the target value well without
knowing the first feature.
There are many other score functions that can be used by a filter method.
Notable examples are estimators of the mutual information or the area under the
receiver operating characteristic (ROC) curve. All of these score functions suffer
from similar problems to the one illustrated previously. We refer the reader to
Guyon and Elisseeff (2003).

25.1.2 Greedy Selection Approaches

Greedy selection is another popular approach for feature selection. Unlike filter
methods, greedy selection approaches are coupled with the underlying learning
algorithm. The simplest instance of greedy selection is forward greedy selection.
We start with an empty set of features, and then we gradually add one feature at a
time to the set of selected features. Given that our current set of selected features is
I , we go over all i
/ I , and apply the learning algorithm on the set of features I {i }.
Each such application yields a different predictor, and we choose to add the feature
that yields the predictor with the smallest risk (on the training set or on a validation
set). This process continues until we either select k features, where k is a predefined
budget of allowed features, or achieve an accurate enough predictor.
Example 25.2 (Orthogonal Matching Pursuit). To illustrate the forward greedy
selection approach, we specify it to the problem of linear regression with the squared
loss. Let X Rm,d be a matrix whose rows are the m training instances. Let y Rm
be the vector of the m labels. For every i [d], let X i be the i th column of X. Given
a set I [d] we denote by X I the matrix whose columns are {X i : i I }.
The forward greedy selection method starts with I0 = . At iteration t, we look
for the feature index jt , which is in
argmin mint X It1 { j } w y2 .
j

wR

Then, we update It = It1 { jt }.

We now describe a more efficient implementation of the forward greedy selection approach for linear regression which is called Orthogonal Matching Pursuit
(OMP). The idea is to keep an orthogonal basis of the features aggregated so far.
Let Vt be a matrix whose columns form an orthonormal basis of the columns of X It .
Clearly,
min X It w y2 = mint Vt y2 .
w

We will maintain a vector t which minimizes the right-hand side of the equations.
Initially, we set I0 = , V0 = , and 1 to be the empty vector. At round t, for

X j is the projection of X j
every j , we decompose X j = v j + u j where v j = Vt1 Vt1

25.1 Feature Selection

onto the subspace spanned by Vt1 and u j is the part of X j orthogonal to Vt1 (see
Appendix C). Then,
min Vt1 + u j y2
,


= min Vt1 y2 + 2 u j 2 + 2u j , Vt1 y
,


= min Vt1 y2 + 2 u j 2 + 2u j , y
,


= min Vt1 y2 + min 2 u j 2 2u j , y


= Vt1 t1 y2 + min 2 u j 2 2u j , y

= Vt1 t1 y2

(u j , y)2
.
u j 2

It follows that we should select the feature

jt = argmax
j

(u j , y)2
.
u j 2

The rest of the update is to set


Vt = Vt1 ,


u jt
,
u jt 2



u jt , y
t = t1 ;
.
u jt 2

The OMP procedure maintains an orthonormal basis of the selected features,

where in the preceding description, the orthonormalization property is obtained by
a procedure similar to Gram-Schmidt orthonormalization. In practice, the GramSchmidt procedure is often numerically unstable. In the pseudocode that follows we
use SVD (see Section C.4) at the end of each round to obtain an orthonormal basis
in a numerically stable manner.
Orthogonal Matching Pursuit (OMP)
input:
data matrix X Rm,d , labels vector y Rm ,
budget of features T
initialize: I1 =
for t = 1, . . . , T
use SVD to find an orthonormal basis V Rm,t1 of X It
(for t = 1 set V to be the all zeros matrix)
foreach j [d] \ It let u j = X j V V
X j
let jt = argmax j / It :u j >0
update It+1 = It { jt }
output IT +1

(u j ,y)2
u j 2

313

314

Feature Selection and Generation

More Efficient Greedy Selection Criteria

Let R(w) be the empirical risk of a vector w. At each round of the forward greedy
selection method, and for every possible j , we should minimize R(w) over the
vectors w whose support is It1 { j }. This might be time consuming.
A simpler approach is to choose jt that minimizes
argmin min R(wt1 + e j ),
R

where e j is the all zeros vector except 1 in the j th element. That is, we keep the
weights of the previously chosen coordinates intact and only optimize over the new
variable. Therefore, for each j we need to solve an optimization problem over a
single variable, which is a much easier task than optimizing over t.
An even simpler approach is to upper bound R(w) using a simple function and
then choose the feature which leads to the largest decrease in this upper bound. For
example, if R is a -smooth function (see Equation (12.5) in Chapter 12), then
R(w + e j ) R(w) +

R(w)
+ 2 /2.
w j

1
Minimizing the right-hand side over yields = R(w)
w j and plugging this value
into the inequality yields


1 R(w) 2
.
R(w)
2
w j

This value is minimized if the partial derivative of R(w) with respect to w j is maximal. We can therefore choose jt to be the index of the largest coordinate of the
gradient of R(w) at w.
Remark 25.3 (AdaBoost as a Forward Greedy Selection Procedure). It is possible
to interpret the AdaBoost algorithm from Chapter 10 as a forward greedy selection
procedure with respect to the function

m
d


exp yi
w j h j (x j ) .
(25.3)
R(w) = log
i=1

j =1

See Exercise 25.3.

Backward Elimination
Another popular greedy selection approach is backward elimination. Here, we start
with the full set of features, and then we gradually remove one feature at a time
from the set of features. Given that our current set of selected features is I , we go
over all i I , and apply the learning algorithm on the set of features I \ {i }. Each
such application yields a different predictor, and we choose to remove the feature
i for which the predictor obtained from I \ {i } has the smallest risk (on the training
set or on a validation set).
Naturally, there are many possible variants of the backward elimination idea. It
is also possible to combine forward and backward greedy steps.

25.1 Feature Selection

25.1.3 Sparsity-Inducing Norms

The problem of minimizing the empirical risk subject to a budget of k features can
be written as
min L S (w) s.t. w0 k,
w

where1
w0 = |{i : wi = 0}|.
In other words, we want w to be sparse, which implies that we only need to measure
the features corresponding to nonzero elements of w.
Solving this optimization problem is computationally hard (Natarajan 1995,
Davis, Mallat & Avellaneda 1997). A possible

relaxation is to replace the nonconvex
function w0 with the
1 norm, w1 = di=1 |wi |, and to solve the problem
min L S (w) s.t. w1 k1 ,
w

(25.4)

where k1 is a parameter. Since the
1 norm is a convex function, this problem can
be solved efficiently as long as the loss function is convex. A related problem is
minimizing the sum of L S (w) plus an
1 norm regularization term,


min L S (w) + w1 ,
(25.5)
w

where is a regularization parameter. Since for any k1 there exists a such that
Equation (25.4) and Equation (25.5) lead to the same solution, the two problems
are in some sense equivalent.
The
1 regularization often induces sparse solutions. To illustrate this, let us start
with the simple optimization problem


min 12 w2 xw + |w| .
(25.6)
w R

It is easy to verify (see Exercise 25.2) that the solution to this problem is the soft
thresholding operator
w = sign(x) [|x| ]+ ,
(25.7)
def

where [a]+ = max{a, 0}. That is, as long as the absolute value of x is smaller than ,
the optimal solution will be zero.
Next, consider a one dimensional regression problem with respect to the squared
loss:


m
1 
2
argmin
(x i w yi ) + |w| .
2m
w Rm
i=1

We can rewrite the problem as

 m
 




1 1 2
2
1
xi w m
x i yi w + |w| .
argmin
2 m
w Rm
i

i=1

The function  0 is often referred to as the
0 norm. Despite the use of the norm notation,  0
is not really a norm; for example, it does not satisfy the positive homogeneity property of norms,
aw0 = |a| w0 .

315

316

Feature Selection and Generation

For simplicity let us assume that m1 i x i2 = 1, and denote x, y = m
i=1 x i yi ; then the
optimal solution is
w = sign(x, y) [|x, y|/m ]+ .
That is, the solution will be zero unless the correlation between the feature x and
the labels vector y is larger than .
Remark 25.4. Unlike the
1 norm, the
2 norm does not induce sparse solutions.
Indeed, consider aforementioned problem with an
2 regularization, namely,


m
1 
2
2
argmin
(x i w yi ) + w .
2m
w Rm
i=1

Then, the optimal solution is

w=

x, y/m
.
x2 /m + 2

This solution will be nonzero even if the correlation between x and y is very small. In
contrast, as we have shown before, when using
1 regularization, w will be nonzero
only if the correlation between x and y is larger than the regularization parameter .
Adding
1 regularization to a linear regression problem with the squared loss
yields the LASSO algorithm, defined as


1
argmin
Xw y2 + w1 .
(25.8)
2m
w
Under some assumptions on the distribution and the regularization parameter ,
the LASSO will find sparse solutions (see, for example, (Zhao & Yu 2006) and
the references therein). Another advantage of the
1 norm is that a vector with
low
1 norm can be sparsified (see, for example, (Shalev-Shwartz, Zhang, and
Srebro 2010) and the references therein).

25.2 FEATURE MANIPULATION AND NORMALIZATION

Feature manipulations or normalization include simple transformations that we
apply on each of our original features. Such transformations may decrease the
approximation or estimation errors of our hypothesis class or can yield a faster algorithm. Similarly to the problem of feature selection, here again there are no absolute
good and bad transformations, but rather each transformation that we apply
should be related to the learning algorithm we are going to apply on the resulting
feature vector as well as to our prior assumptions on the problem.
To motivate normalization, consider a linear regression problem with the
squared loss. Let X Rm,d be a matrix whose rows are the instance vectors and let
y Rm be a vector of target values. Recall that ridge regression returns the vector


1
Xw y2 + w2 = (2m I + X
X)1 X
y.
argmin
m
w
Suppose that d = 2 and the underlying data distribution is as follows. First we sample
y uniformly at random from {1}. Then, we set x 1 to be y +0. 5, where is sampled
uniformly at random from {1}, and we set x 2 to be 0. 0001y. Note that the optimal

25.2 Feature Manipulation and Normalization

weight vector is w = [0; 10000], and L D (w ) = 0. However, the objective of ridge

regression at w is 108 . In contrast, the objective of ridge regression at w = [1; 0] is
8
, the
likely to be close to 0. 25 + . It follows that whenever > 100.25
8 1 0. 25 10
objective of ridge regression is smaller at the suboptimal solution w = [1; 0]. Since
typically should be at least 1/m (see the analysis in Chapter 13), it follows that in
the aforementioned example, if the number of examples is smaller than 108 then we
are likely to output a suboptimal solution.
The crux of the preceding example is that the two features have completely different scales. Feature normalization can overcome this problem. There are many
ways to perform feature normalization, and one of the simplest approaches is simply
to make sure that each feature receives values between 1 and 1. In the preceding
example, if we divide each feature by the maximal value it attains we will obtain that
and x 2 = y. Then, for 103 the solution of ridge regression is quite
x 1 = y+0.5
1.5

close to w .
Moreover, the generalization bounds we have derived in Chapter 13 for regularized loss minimization depend on the norm of the optimal vector w and on the
maximal norm of the instance vectors.2 Therefore, in the aforementioned example,
before we normalize the features we have that w 2 = 108 , while after we normalize the features we have that w 2 = 1. The maximal norm of the instance vector
remains roughly the same; hence the normalization greatly improves the estimation
error.
Feature normalization can also improve the runtime of the learning algorithm.
For example, in Section 14.5.3 we have shown how to use the Stochastic Gradient
Descent (SGD) optimization algorithm for solving the regularized loss minimization problem. The number of iterations required by SGD to converge also depends
on the norm of w and on the maximal norm of x. Therefore, as before, using
normalization can greatly decrease the runtime of SGD.
Next, we demonstrate in the following how a simple transformation on features,
such as clipping, can sometime decrease the approximation error of our hypothesis
class. Consider again linear regression with the squared loss. Let a > 1 be a large
number, suppose that the target y is chosen uniformly at random from {1}, and
then the single feature x is set to be y with probability (1 1/a) and set to be ay
with probability 1/a. That is, most of the time our feature is bounded but with a very
small probability it gets a very high value. Then, for any w, the expected squared
loss of w is
1
L D (w) = E (wx y)2
2


1 1
11
(wy y)2 +
(awy y)2 .
= 1
a 2
a2

More precisely, the bounds we derived in Chapter 13 for regularized loss minimization depend on
w 2 and on either the Lipschitzness or the smoothness of the loss function. For linear predictors
and loss functions of the form
(w, (x, y)) = (w, x, y), where is convex and either 1-Lipschitz or
1-smooth with respect to its first argument, we have that
is either x-Lipschitz or x2 -smooth. For
example, for the squared loss, (a, y) = 12 (a y)2 , and
(w, (x, y)) = 12 (w, x y)2 is x2 -smooth with
respect to its first argument.

317

318

Feature Selection and Generation

Solving for w we obtain that w = a 22a1

, which goes to zero as a goes to infinity.
+a1
Therefore, the objective at w goes to 0. 5 as a goes to infinity. For example, for
a = 100 we will obtain L D (w ) 0. 48. Next, suppose we apply a clipping transformation; that is, we use the transformation x  sign(x) min{1, |x|}. Then, following
this transformation, w becomes 1 and L D (w ) = 0. This simple example shows that
a simple transformation can have a significant influence on the approximation error.
Of course, it is not hard to think of examples in which the same feature transformation actually hurts performance and increases the approximation error. This
is not surprising, as we have already argued that feature transformations should rely
on our prior assumptions on the problem. In the aforementioned example, a prior
assumption that may lead us to use the clipping transformation is that features
that get values larger than a predefined threshold value give us no additional useful
information, and therefore we can clip them to the predefined threshold.

25.2.1 Examples of Feature Transformations

We now list several common techniques for feature transformations. Usually, it is
helpful to combine some of these transformations (e.g., centering + scaling). In the
following, we denote by f = ( f 1 , . . . , fm ) Rm the value of the feature f over the


m training examples. Also, we denote by f = m1 m
i=1 f i the empirical mean of the
feature over all examples.
Centering:
This transformation makes the feature have zero mean, by setting f i fi f.
Unit Range:
This transformation makes the range of each feature be [0, 1]. Formally, let fmax =
f i f min
maxi fi and f min = mini f i . Then, we set f i fmax
. Similarly, we can make
f
min

f i f min
the range of each feature be [ 1, 1] by the transformation f i 2 f max
f min 1. Of
course, it is easy to make the range [0, b] or [ b, b], where b is a user-specified
parameter.

Standardization:
This transformation

makes all2 features have a zero mean and unit variance. For
mally, let = m1 m
i=1 ( f i f ) be the empirical variance of the feature. Then, we
set fi

f
i f

Clipping:
This transformation clips high or low values of the feature. For example, f i
sign( fi ) max{b, | f i |}, where b is a user-specified parameter.
Sigmoidal Transformation:
As its name indicates, this transformation applies a sigmoid function on the fea1
ture. For example, f i 1+exp(b
f i ) , where b is a user-specified parameter. This
transformation can be thought of as a soft version of clipping: It has a small effect

25.3 Feature Learning

on values close to zero and behaves similarly to clipping on values far away from
zero.
Logarithmic Transformation:
The transformation is f i log (b + f i ), where b is a user-specified parameter. This is
widely used when the feature is a counting feature. For example, suppose that the
feature represents the number of appearances of a certain word in a text document.
Then, the difference between zero occurrences of the word and a single occurrence
is much more important than the difference between 1000 occurrences and 1001
occurrences.
Remark 25.5. In the aforementioned transformations, each feature is transformed
on the basis of the values it obtains on the training set, independently of other
features values. In some situations we would like to set the parameter of the
transformation on the basis of other features as well. A notable example is a transformation in which one applies a scaling to the features so that the empirical average
of some norm of the instances becomes 1.

25.3 FEATURE LEARNING

So far we have discussed feature selection and manipulations. In these cases, we
start with a predefined vector space Rd , representing our features. Then, we select a
subset of features (feature selection) or transform individual features (feature transformation). In this section we describe feature learning, in which we start with some
instance space, X , and would like to learn a function, : X Rd , which maps
instances in X into a representation as d-dimensional feature vectors.
The idea of feature learning is to automate the process of finding a good representation of the input space. As mentioned before, the No-Free-Lunch theorem
tells us that we must incorporate some prior knowledge on the data distribution in
order to build a good feature representation. In this section we present a few feature
learning approaches and demonstrate conditions on the underlying data distribution
in which these methods can be useful.
Throughout the book we have already seen several useful feature constructions.
For example, in the context of polynomial regression, we have mapped the original instances into the vector space of all their monomials (see Section 9.2.2 in
Chapter 9). After performing this mapping, we trained a linear predictor on top
of the constructed features. Automation of this process would be to learn a transformation : X Rd , such that the composition of the class of linear predictors on
top of yields a good hypothesis class for the task at hand.
In the following we describe a technique of feature construction called dictionary
learning.

25.3.1 Dictionary Learning Using Auto-Encoders

The motivation of dictionary learning stems from a commonly used representation
of documents as a bag-of-words: Given a dictionary of words D = {w1 , . . . , wk },

319

320

Feature Selection and Generation

where each wi is a string representing a word in the dictionary, and given a document, ( p1 , . . . , pd ), where each pi is a word in the document, we represent the
document as a vector x {0, 1}k , where x i is 1 if wi = p j for some j [d], and
x i = 0 otherwise. It was empirically observed in many text processing tasks that linear predictors are quite powerful when applied on this representation. Intuitively,
we can think of each word as a feature that measures some aspect of the document. Given labeled examples (e.g., topics of the documents), a learning algorithm
searches for a linear predictor that weights these features so that a right combination
of appearances of words is indicative of the label.
While in text processing there is a natural meaning to words and to the dictionary, in other applications we do not have such an intuitive representation of an
instance. For example, consider the computer vision application of object recognition. Here, the instance is an image and the goal is to recognize which object appears
in the image. Applying a linear predictor on the pixel-based representation of the
image does not yield a good classifier. What we would like to have is a mapping
that would take the pixel-based representation of the image and would output a
bag of visual words, representing the content of the image. For example, a visual
word can be there is an eye in the image. If we had such representation, we could
have applied a linear predictor on top of this representation to train a classifier for,
say, face recognition. Our question is, therefore, How can we learn a dictionary
of visual words such that a bag-of-words representation of an image would be
helpful for predicting which object appears in the image?
A first naive approach for dictionary learning relies on a clustering algorithm
(see Chapter 22). Suppose that we learn a function c : X {1, . . . , k}, where c(x)
is the cluster to which x belongs. Then, we can think of the clusters as words,
and of instances as documents, where a document x is mapped to the vector
(x) {0, 1}k , where (x)i is 1 if and only if x belongs to the i th cluster. Now, it
is straightforward to see that applying a linear predictor on (x) is equivalent to
assigning the same target value to all instances that belong to the same cluster. Furthermore, if the clustering is based on distances from a class center (e.g., k-means),
then a linear predictor on (x) yields a piece-wise constant predictor on x.
Both the k-means and PCA approaches can be regarded as special cases of a
more general approach for dictionary learning which is called auto-encoders. In an
auto-encoder we learn a pair of functions: an encoder function, : Rd Rk , and
a decoder function, : Rk Rd . The goal of the learning process is to find a pair


of functions such that the reconstruction error, i xi ((xi ))2 , is small. Of
course, we can trivially set k = d and both , to be the identity mapping, which
yields a perfect reconstruction. We therefore must restrict and in some way. In
PCA, we constrain k < d and further restrict and to be linear functions. In kmeans, k is not restricted to be smaller than d, but now and rely on k centroids,
1 , . . . , k , and (x) returns an indicator vector in {0, 1}k that indicates the closest
centroid to x, while takes as input an indicator vector and returns the centroid
representing this vector.
An important property of the k-means construction, which is key in allowing
k to be larger than d, is that maps instances into sparse vectors. In fact, in kmeans only a single coordinate of (x) is nonzero. An immediate extension of the
k-means construction is therefore to restrict the range of to be vectors with at

25.5 Bibliographic Remarks

most s nonzero elements, where s is a small integer. In particular, let and be

functions that depend on 1 , . . . , k . The function maps an instance vector x to a
vector (x) Rk , where

(x) should have at most s nonzero elements. The function
(v) is defined to be ki=1 vi i . As before, our goal is to have a small reconstruction
error, and therefore we can define
(x) = argmin x (v)2 s.t. v0 s,
v

where v0 = |{ j : v j = 0}|. Note that when s = 1 and we further restrict v1 = 1
then we obtain the k-means encoding function; that is, (x) is the indicator vector
of the centroid closest to x. For larger values of s, the optimization problem in the
preceding definition of becomes computationally difficult. Therefore, in practice,
we sometime use
1 regularization instead of the sparsity constraint and define
to be


(x) = argmin x (v)2 + v1 ,
v

where > 0 is a regularization parameter. Anyway, the dictionary learning problem

is now to find the vectors 1 , . . . , k such that the reconstruction error, m
i=1 xi
((x))2 , is as small as possible. Even if is defined using the
1 regularization,
this is still a computationally hard problem (similar to the k-means problem). However, several heuristic search algorithms may give reasonably good solutions. These
algorithms are beyond the scope of this book.

25.4 SUMMARY
Many machine learning algorithms take the feature representation of instances for
granted. Yet the choice of representation requires careful attention. We discussed
approaches for feature selection, introducing filters, greedy selection algorithms,
and sparsity-inducing norms. Next we presented several examples for feature transformations and demonstrated their usefulness. Last, we discussed feature learning, and in particular dictionary learning. We have shown that feature selection,
manipulation, and learning all depend on some prior knowledge on the data.

25.5 BIBLIOGRAPHIC REMARKS

Guyon and Elisseeff (2003) surveyed several feature selection procedures, including
many types of filters.
Forward greedy selection procedures for minimizing a convex objective subject to a polyhedron constraint date back to the Frank-Wolfe algorithm (Frank &
Wolfe 1956). The relation to boosting has been studied by several authors, including,
(Warmuth, Liao & Ratsch 2006, Warmuth, Glocer & Vishwanathan 2008, ShalevShwartz & Singer 2008). Matching pursuit has been studied in the signal processing
community (Mallat & Zhang 1993). Several papers analyzed greedy selection methods under various conditions. See, for example, Shalev-Shwartz, Zhang, and Srebro
(2010) and the references therein.
The use of the
1 -norm as a surrogate for sparsity has a long history
(e.g., Tibshirani (1996) and the references therein), and much work has been done

321

322

Feature Selection and Generation

on understanding the relationship between the
1 -norm and sparsity. It is also

closely related to compressed sensing (see Chapter 23). The ability to sparsify low

1 norm predictors dates back to Maurey (Pisier 19801981). In Section 26.4 we also
show that low
1 norm can be used to bound the estimation error of our predictor.
Feature learning and dictionary learning have been extensively studied recently
in the context of deep neural networks. See, for example, (LeCun & Bengio 1995,
Hinton et al. 2006, Ranzato et al. 2007, Collobert & Weston 2008, Lee et al. 2009, Le
et al. 2012, Bengio 2009) and the references therein.

25.6 EXERCISES
25.1 Prove the equality given in Equation (25.1). Hint: Let a , b be minimizers of the
left-hand side. Find a, b such that the objective value of the right-hand side is
smaller than that of the left-hand side. Do the same for the other direction.
25.2 Show that Equation (25.7) is the solution of Equation (25.6).
25.3 AdaBoost as a Forward Greedy Selection Algorithm: Recall the AdaBoost algorithm from Chapter 10. In this section we give another interpretation of AdaBoost
as a forward greedy selection algorithm.

Given a set of m instances x1 , . . ., xm , and a hypothesis class H of finite VC
dimension, show that there exist d and h 1 , . . ., h d such that for every h H
there exists i [d] with h i (x j ) = h(x j ) for every j [m].

Let R(w) be as defined in Equation (25.3). Given some w, define f w to be the
function
d

fw ( ) =
wi h i ( ).
i=1

Let D be the distribution over [m] defined by

Di =

exp ( yi f w (xi ))
,
Z

where Z is a normalization factor that ensures that D is a probability vector.

Show that
m

R(w)
=
Di yi h j (xi ).
wj

Furthermore, denoting j =

i=1

m

i=1 Di 1[h j (xi )= yi ] , show that

R(w)
wj

= 2 j 1.

Conclude that if j 1/2 then R(w)
w j /2.
(t+1) ) R(w(t) )

Show
 that the update of AdaBoost guarantees R(w
2
log ( 1 4 ). Hint: Use the proof of Theorem 10.2.

PART 4

Advanced Theory

26
Rademacher Complexities

In Chapter 4 we have shown that uniform convergence is a sufficient condition for

learnability. In this chapter we study the Rademacher complexity, which measures
the rate of uniform convergence. We will provide generalization bounds based on
this measure.

26.1 THE RADEMACHER COMPLEXITY

Recall the definition of an -representative sample from Chapter 4, repeated here
for convenience.
Definition 26.1 ( -Representative Sample). A training set S is called
-representative (w.r.t. domain Z , hypothesis class H, loss function
, and distribution D) if
sup |L D (h) L S (h)| .
hH

We have shown that if S is an /2 representative sample then the ERM rule is

-consistent, namely, L D (ERMH (S)) minhH L D (h) + .
To simplify our notation, let us denote
def

def

F =
H = {z 
(h, z) : h H},
and given f F , we define
1
f (z i ).
m
m

L D ( f ) = E [ f (z)],
zD

LS( f ) =

i=1

We define the representativeness of S with respect to F as the largest gap between

the true error of a function f and its empirical error, namely,
def

RepD (F , S) = sup (L D ( f ) L S ( f )).

f F

(26.1)

Now, suppose we would like to estimate the representativeness of S using the

sample S only. One simple idea is to split S into two disjoint sets, S = S1 S2 ;
325

326

Rademacher Complexities

refer to S1 as a validation set and to S2 as a training set. We can then estimate

the representativeness of S by
sup (L S1 ( f ) L S2 ( f )).

(26.2)

f F

This can be written more compactly by defining = (1 , . . . , m ) {1}m to be a

vector such that S1 = {z i : i = 1} and S2 = {z i : i = 1}. Then, if we further assume
that |S1 | = |S2 | then Equation (26.2) can be rewritten as
m

2
sup
i f (z i ).
m f F

(26.3)

i=1

The Rademacher complexity measure captures this idea by considering the expectation of the term appearing in Equation 26.3 with respect to a random choice of .
Formally, let F S be the set of all possible evaluations a function f F can achieve
on a sample S, namely,
F S = {( f (z 1 ), . . . , f (z m )) : f F }.
Let the variables in be distributed i.i.d. according to P [i = 1] = P [i = 1] = 12 .
Then, the Rademacher complexity of F with respect to S is defined as follows:


m

def 1
E
i f (z i ) .
(26.4)
sup
R(F S) =
m {1}m f F
i=1

Rm ,

we define

m

def 1
E sup
R(A) =
i ai .
m a A

More generally, given a set of vectors, A

(26.5)

i=1

The following lemma bounds the expected value of the representativeness of S

by twice the expected Rademacher complexity.
Lemma 26.2.
E [ RepD (F , S)] 2 E m R(F S).

SD m

SD

 } be another i.i.d. sample. Clearly, for all f F , L ( f ) =

Proof. Let S  = {z 1 , . . . , z m
D
E S  [L S  ( f )]. Therefore, for every f F we have

L D ( f ) L S ( f ) = E [L S  ( f )] L S ( f ) = E [L S  ( f ) L S ( f )].
S

Taking supremum over f F of both sides, and using the fact that the supremum
of expectation is smaller than expectation of the supremum we obtain
sup (L D ( f ) L S ( f )) = sup E [L S  ( f ) L S ( f )]
f F

f F S

E
S

sup (L S  ( f ) L S ( f )) .
f F

26.1 The Rademacher Complexity

Taking expectation over S on both sides we obtain




E
S

sup (L D ( f ) L S ( f )) E 

sup (L S  ( f ) L S ( f ))

S,S

f F

f F

1
E
=
m S,S 

sup

m


f F i=1


(

f (z i )

f (z i )) .

(26.6)

Next, we note that for each j , z j and z j are i.i.d. variables. Therefore, we can replace
them without affecting the expectation:


E  sup ( f (z j ) f (z j )) +
( f (z i ) f (z i ))
S,S

f F

i= j

= E  sup ( f (z j ) f (z j )) +
S,S

f F

( f (z i ) f (z i )) .

(26.7)

i= j

Let j be a random variable such that P [ j = 1] = P [ j = 1] = 1/2. From

Equation (26.7) we obtain that


E sup j ( f (z j ) f (z j )) +
( f (z i ) f (z i ))
S,S , j

f F

i= j

1
1
= (l.h.s. of Equation (26.7)) + (r.h.s. of Equation (26.7))
2
2


= E  sup ( f (z j ) f (z j )) +
( f (z i ) f (z i )) .

(26.8)

Repeating this for all j we obtain that





m
m




E  sup
( f (z i ) f (z i )) = E
i ( f (z i ) f (z i )) .
sup

(26.9)

S,S

S,S

f F

i= j

S,S ,

f F i=1

Finally,
sup

f F

i ( f (z i ) f (z i )) sup

f F i=1

f F

i f (z i ) + sup

f F

i f (z i )

and since the probability of is the same as the probability of , the right-hand
side of Equation (26.9) can be bounded by





E
i f (z i ) + sup
i f (z i )
sup
S,S ,

f F

f F

= m E [R(F S )] + m E [R(F S)] = 2m E [R(F S)].

S

The lemma immediately yields that, in expectation, the ERM rule finds a
hypothesis which is close to the optimal hypothesis in H.

327

328

Rademacher Complexities

Theorem 26.3. We have

E [L D (ERMH (S)) L S (ERMH (S))] 2 E m R(
H S).

SD m

SD

Furthermore, for any

E [L D (ERMH (S)) L D (h  )] 2 E m R(
H S).

SD m

SD

if h 

Furthermore,
= argminh L D (h) then for each (0, 1) with probability of at least
1 over the choice of S we have
2 E S  Dm R(
H S  )
.

Proof. The first inequality follows directly from Lemma 26.2. The second inequality
follows because for any fixed h  ,
L D (ERMH (S)) L D (h  )

L D (h  ) = E [L S (h  )] E [L S (ERMH (S))].
S

The third inequality follows from the previous inequality by relying on Markovs
inequality (note that the random variable L D (ERMH (S)) L D (h  ) is nonnegative).
Next, we derive bounds similar to the bounds in Theorem 26.3 with a better dependence on the confidence parameter . To do so, we first introduce the
following bounded differences concentration inequality.
Lemma 26.4 (McDiarmids Inequality). Let V be some set and let f : V m R
be a function of m variables such that for some c > 0, for all i [m] and for all
x 1 , . . . , x m , x i V we have
| f (x 1 , . . . , x m ) f (x 1 , . . . , x i1 , x i , x i+1 , . . . , x m )| c.
Let X 1 , . . . , X m be m independent random variables taking values in V . Then, with
probability of at least 1 we have
  
| f (X 1 , . . . , X m ) E [ f (X 1 , . . . , X m )]| c

ln

m/2.

On the basis of the McDiarmid inequality we can derive generalization bounds

with a better dependence on the confidence parameter.
Theorem 26.5. Assume that for all z and h H we have that |
(h, z)| c. Then,
1. With probability of at least 1 , for all h H,
L D (h) L S (h) 2

S  Dm

R(
H S ) + c

In particular, this holds for h = ERMH (S).

2. With probability of at least 1 , for all h H,
L D (h) L S (h) 2 R(
H S) + 4 c
In particular, this holds for h = ERMH (S).

2 ln (2/)
.
m

2 ln (4/)
.
m

26.1 The Rademacher Complexity

3. For any h  , with probability of at least 1 ,

2 ln (8/)
.
m


Proof. First note that the random variable RepD (F , S) = suphH L D (h) L S (h)
satisfies the bounded differences condition of Lemma 26.4 with a constant 2c/m.
Combining the bounds in Lemma 26.4 with Lemma 26.2 we obtain that with
probability of at least 1 ,


2 ln (2/)
2 ln (2/)

2 E R(
H S ) + c
.
RepD (F , S) E RepD (F , S) + c
S
m
m


L D (ERMH (S)) L D (h ) 2 R(
H S) + 5 c

The first inequality of the theorem follows from the definition of RepD (F , S). For
the second inequality we note that the random variable R(
H S) also satisfies
the bounded differences condition of Lemma 26.4 with a constant 2c/m. Therefore,
the second inequality follows from the first inequality, Lemma 26.4, and the union
bound. Finally, for the last inequality, denote h S = ERMH (S) and note that
L D (h S ) L D (h  )
= L D (h S ) L S (h S ) + L S (h S ) L S (h  ) + L S (h  ) L D (h  )

 

L D (h S ) L S (h S ) + L S (h  ) L D (h  ) .

(26.10)

The first summand on the right-hand side is bounded by the second inequality of
the theorem. For the second summand, we use the fact that h  does not depend on
S; hence by using Hoeffdings inequality we obtain that with probaility of at least
1 /2,

ln (4/)


L S (h ) L D (h ) c
.
(26.11)
2m
Combining this with the union bound we conclude our proof.
The preceding theorem tells us that if the quantity R(
H S) is small then it
is possible to learn the class H using the ERM rule. It is important to emphasize
that the last two bounds given in the theorem depend on the specific training set S.
That is, we use S both for learning a hypothesis from H as well as for estimating the
quality of it. This type of bound is called a data-dependent bound.

26.1.1 Rademacher Calculus

Let us now discuss some properties of the Rademacher complexity measure. These
properties will help us in deriving some simple bounds on R(
H S) for specific
cases of interest.
The following lemma is immediate from the definition.
Lemma 26.6. For any A Rm , scalar c R, and vector a0 Rm , we have
R({c a + a0 : a A}) |c| R(A).
The following lemma tells us that the convex hull of A has the same complexity
as A.

329

330

Rademacher Complexities

Lemma 26.7. Let A be a subset of Rm and let A = {

A, j 0, 1 = 1}. Then, R(A ) = R(A).

N

j =1 j a

( j)

: N N, j , a( j )

Proof. The main idea follows from the fact that for any vector v we have
N


sup

0:1 =1 j =1

j v j = max v j .
j

Therefore,
m R(A ) = E

sup

m


sup

0: =1 (1)
1
a ,...,a(N) i=1

=E

N


sup

0: =1
1
j =1

= E sup

m


a A
i=1

j sup
a( j )

m


N


( j)

j ai

j =1
( j)

i ai

i=1

i ai

= m R(A),
and we conclude our proof.
The next lemma, due to Massart, states that the Rademacher complexity of a
finite set grows logarithmically with the size of the set.
Lemma 26.8 (Massart Lemma). Let A = {a1 , . . . , a N } be a finite set of vectors in Rm .

N
ai . Then,
Define a = N1 i=1

2 log (N)
.
R(A) max a a 
a A
m
Proof. On the basis of Lemma 26.6, we can assume without loss of generality that
a = 0. Let > 0 and let A = {a1 , . . . , a N }. We upper bound the Rademacher
complexity as follows:


 


 ,a
m R(A ) = E max , a = E log max e

a A

E log


e ,a

a A

 
log E


= log


e

 ,a

a A

m


a A

i=1

a A

// Jensens inequality


E [ei ai ] ,
i

where the last equality occurs because the Rademacher variables are independent.
Next, using Lemma A.6 we have that for all ai R,
E e i a i =

exp (ai ) + exp( ai )

exp(ai2 /2),
2

26.1 The Rademacher Complexity

and therefore

m R(A ) log

m



exp

a A i=1

ai2
2




= log




2
exp a /2

a A





2
log |A | max exp a /2
= log (|A |) + max (a2 /2).
a A

a A

Since R(A) = 1 R(A ) we obtain from the equation that

R(A)
Setting =
proof.

log (|A|) + 2 maxa A (a2 /2)

.
m


2 log (|A|)/ maxa A a2 and rearranging terms we conclude our

The following lemma shows that composing A with a Lipschitz function does not
blow up the Rademacher complexity. The proof is due to Kakade and Tewari.
Lemma 26.9 (Contraction Lemma). For each i [m], let i : R R be a -Lipschitz
function; namely, for all , R we have |i () i ()| | |. For a Rm let
(a) denote the vector (1 (a1 ), . . . , m (ym )). Let A = {(a) : a A}. Then,
R( A) R(A).
Proof. For simplicity, we prove the lemma for the case = 1. The case =
1 will follow by defining  = 1 and then using Lemma 26.6. Let Ai =
{(a1 , . . . , ai1 , i (ai ), ai+1 , . . . , am ) : a A}. Clearly, it suffices to prove that for any
set A and all i we have R(Ai ) R(A). Without loss of generality we will prove the
latter claim for i = 1 and to simplify notation we omit the subscript from 1 . We have


m

m R(A1 ) = E sup
i ai

a A1 i=1

= E sup 1 (a1 ) +

a A

m



i ai

i=2






m
m


1
=
i ai + sup (a1 ) +
i ai
sup (a1 ) +
E
2 2 ,...,m a A
a A
i=2
i=2



m
m


1


E
=
i ai +
i ai
sup (a1 ) (a1 ) +
2 2 ,...,m a,a A
i=2
i=2



m
m


1
E

i ai +
i ai
sup |a1 a1 | +
,
(26.12)
2 2 ,...,m a,a A
i=2

i=2

where in the last inequality we used the assumption that is Lipschitz. Next, we
note that the absolute value on |a1 a1 | in the preceding expression can be omitted
since both a and a are from the same set A and the rest of the expression in the

331

332

Rademacher Complexities

supremum is not affected by replacing a and a . Therefore,

1
m R(A1 )
E
2 2 ,...,m


sup

a,a A

a1 a1 +

m


i ai +

i=2

m



i ai

(26.13)

i=2

But, using the same equalities as in Equation (26.12), it is easy to see that the righthand side of Equation (26.13) exactly equals m R(A), which concludes our proof.

26.2 RADEMACHER COMPLEXITY OF LINEAR CLASSES

In this section we analyze the Rademacher complexity of linear classes. To simplify
the derivation we first define the following two classes:
H1 = {x  w, x : w1 1},

H2 = {x  w, x : w2 1}.

(26.14)

The following lemma bounds the Rademacher complexity of H2 . We allow the

xi to be vectors in any Hilbert space (even infinite dimensional), and the bound
does not depend on the dimensionality of the Hilbert space. This property becomes
useful when analyzing kernel methods.
Lemma 26.10. Let S = (x1 , . . . , xm ) be vectors in a Hilbert space. Define: H2 S =
{(w, x1 , . . . , w, xm ) : w2 1}. Then,
maxi xi 2

.
m

R(H2 S)

Proof. Using Cauchy-Schwartz inequality we know that for any vectors w, v we

have w, v w v. Therefore,

m R(H2 S) = E

m


sup

aH2 S i=1


=E

sup


i w, xi 

w:w1 i=1

sup w,

=E

w:w1

E 

i ai

m


m


m



i xi 

i=1

i xi 2 .

(26.15)

i=1

Next, using Jensens inequality we have that

0
0 
02 1/2
02 1/2
0
0 m
m
m
0
0
0
0
0
0
0

0
0
0
0
0
E 0
i xi 0 = E 0
i xi 0 E 0
i xi 0 . (26.16)
0

0
0
0
0
0
i=1

i=1

i=1

26.3 Generalization Bounds for SVM

Finally, since the variables 1 , . . . , m are independent we have

 m



E 
i xi 22 = E
i j xi , x j 

i=1

i, j

xi , x j  E [i j ] +

i= j

m


m



xi , xi  E i2

i=1

xi 22 m max xi 22 .

i

i=1

Combining this with Equation (26.15) and Equation (26.16) we conclude our proof.
Next we bound the Rademacher complexity of H1 S.
Lemma 26.11. Let S = (x1 , . . . , xm ) be vectors in Rn . Then,

2 log (2n)
.
R(H1 S) max xi 
i
m
Proof. Using Holders inequality we know that for any vectors w, v we have w, v
w1 v . Therefore,


m

i ai
m R(H1 S) = E sup


=E

aH1 S i=1

sup


=E

w:w1 1 i=1

sup w,

w:w1 1

E 

m


m



i w, xi 

m



i xi 

i=1

i xi  .

(26.17)

i=1

For each j [n], let v j = (x 1, j , . . . , x m, j ) Rm . Note that v j 2 m maxi xi  . Let

V = {v1 , . . . , vn , v1 , . . . , vn }. The right-hand side of Equation (26.17) is m R(V ).
Using Massart lemma (Lemma 26.8) we have that

R(V ) max xi  2 log (2n)/m,
i

which concludes our proof.

26.3 GENERALIZATION BOUNDS FOR SVM

In this section we use Rademacher complexity to derive generalization bounds for
generalized linear predictors with Euclidean norm constraint. We will show how this
leads to generalization bounds for hard-SVM and soft-SVM.

333

334

Rademacher Complexities

We shall consider the following general constraint-based formulation. Let H =

{w : w2 B} be our hypothesis class, and let Z = X Y be the examples domain.
Assume that the loss function
: H Z R is of the form

(w, (x, y)) = (w, x, y),

(26.18)

where : R Y R is such that for all y Y, the scalar function a  (a, y) is Lipschitz. For example, the hinge-loss function,
(w, (x, y)) = max{0, 1 yw, x},
can be written as in Equation (26.18) using (a, y) = max{0, 1 ya}, and note
that is 1-Lipschitz for all y {1}. Another example is the absolute loss function,
(w, (x, y)) = |w, x y|, which can be written as in Equation (26.18) using
(a, y) = |a y|, which is also 1-Lipschitz for all y R.
The following theorem bounds the generalization error of all predictors in H
using their empirical error.
Theorem 26.12. Suppose that D is a distribution over X Y such that with probability
1 we have that x2 R. Let H = {w : w2 B} and let
: H Z R be a loss
function of the form given in Equation (26.18) such that for all y Y, a  (a, y)
is a -Lipschitz function and such that maxa[B R,B R] |(a, y)| c. Then, for any
(0, 1), with probability of at least 1 over the choice of an i.i.d. sample of size m,

2 ln (2/)
2 B R
+c
.
w H, L D (w) L S (w) +
m
m
Proof. Let F = {(x, y)  (w, x, y) : w H}. We will show that with probability 1,

R(F S) B R/ m and then the theorem will follow from Theorem 26.5. Indeed,
the set F S can be written as
F S = {((w, x1 , y1 ), . . . , (w, xm , ym )) : w H},
and the bound on R(F S) follows directly by combining Lemma 26.9, Lemma 26.10,
and the assumption that x2 R with probability 1.
We next derive a generalization bound for hard-SVM based on the previous
theorem. For simplicity, we do not allow a bias term and consider the hard-SVM
problem:
argmin w2 s.t. i , yi w, xi  1
(26.19)
w

Theorem 26.13. Consider a distribution D over X {1} such that there exists some
vector w with P(x,y)D [yw , x 1] = 1 and such that x2 R with probability 1.
Let w S be the output of Equation (26.19). Then, with probability of at least 1 over
the choice of S Dm , we have that

2 ln (2/)
2 R w 

+ (1 + R w )
.
P [y = sign(w S , x)]
m
m
(x,y)D
Proof. Throughout the proof, let the loss function be the ramp loss (see
Section 15.2.3). Note that the range of the ramp loss is [0, 1] and that it is a
1-Lipschitz function. Since the ramp loss upper bounds the zero-one loss, we
have that
P [y = sign(w S , x)] L D (w S ).
(x,y)D

26.4 Generalization Bounds for Predictors with Low
1 Norm

Let B = w 2 and consider the set H = {w : w2 B}. By the definition of hardSVM and our assumption on the distribution, we have that w S H with probability
1 and that L S (w S ) = 0. Therefore, using Theorem 26.12 we have that

2 ln (2/)
2B R
.
L D (w S ) L S (w S ) + +
m
m

Remark 26.1. Theorem 26.13 implies that the sample complexity of hard-SVM
2
 2
grows like R w
. Using a more delicate analysis and the separability assumption,
2
R 2 w 2
.


preceding theorem depends on w ,

it is possible to improve the bound to an order of

The bound in the

which is unknown. In the
following we derive a bound that depends on the norm of the output of SVM; hence
it can be calculated from the training set itself. The proof is similar to the derivation
of bounds for structure risk minimization (SRM).
Theorem 26.14. Assume that the conditions of Theorem 26.13 hold. Then, with
probability of at least 1 over the choice of S Dm , we have that
*
4 log2 (w S )
ln (
)
4Rw S 

+
.
P [y = sign(w S , x)]
m
m
(x,y)D
Proof. For any integer i , let Bi = 2i , Hi = {w : w Bi }, and let i = 2i2 . Fix i , then
using Theorem 26.12 we have that with probability of at least 1 i

2 ln (2/i )
2Bi R
w Hi , L D (w) L S (w) + +
m
m


Applying the union bound and using i=1 i we obtain that with probability of
at least 1 this holds for all i . Therefore, for all w, if we let i = log2 (w) then
w Hi , Bi 2w, and

2
i

(2i)2

(4 log2 (w))2
.

Therefore,

2 ln (2/i )
m

4wR
4( ln (4 log2 (w)) + ln (1/))
L S (w) +
+
.
m
m

2Bi R
L D (w) L S (w) + +
m

In particular, it holds for w S , which concludes our proof.

Remark 26.2. Note that all the bounds we have derived do not depend on the dimension of w. This property is utilized when learning SVM with kernels, where the
dimension of w can be extremely large.

26.4 GENERALIZATION BOUNDS FOR PREDICTORS WITH LOW
1 NORM

In the previous section we derived generalization bounds for linear predictors with
an
2 -norm constraint. In this section we consider the following general
1 -norm
constraint formulation. Let H = {w : w1 B} be our hypothesis class, and let

335

336

Rademacher Complexities

Z = X Y be the examples domain. Assume that the loss function,
: H Z R,

is of the same form as in Equation (26.18), with : R Y R being -Lipschitz
w.r.t. its first argument. The following theorem bounds the generalization error of
all predictors in H using their empirical error.
Theorem 26.15. Suppose that D is a distribution over X Y such that with probability
1 we have that x R. Let H = {w Rd : w1 B} and let
: H Z R be a loss
function of the form given in Equation (26.18) such that for all y Y, a  (a, y)
is an -Lipschitz function and such that maxa[B R,B R] |(a, y)| c. Then, for any
(0, 1), with probability of at least 1 over the choice of an i.i.d. sample of size m,


2 log (2d)
2 ln (2/)
+c
.
w H, L D (w) L S (w) + 2 B R
m
m
Proof. The proof is identical to the proof of Theorem 26.12, while relying on
Lemma 26.11 instead of relying on Lemma 26.10.
It is interesting to compare the two bounds given in Theorem 26.12 and Theorem 26.15. Apart from the extra log (d) factor that appears in Theorem 26.15, both
bounds look similar. However, the parameters B, R have different meanings in the
two bounds. In Theorem 26.12, the parameter B imposes an
2 constraint on w and
the parameter R captures a low
2 -norm assumption on the instances. In contrast,
in Theorem 26.15 the parameter B imposes an
1 constraint on w (which is stronger
than an
2 constraint) while the parameter R captures a low
-norm assumption on
the instance (which is weaker than a low
2 -norm assumption). Therefore, the choice
of the constraint should depend on our prior knowledge of the set of instances and
on prior assumptions on good predictors.

26.5 BIBLIOGRAPHIC REMARKS

The use of Rademacher complexity for bounding the uniform convergence is due to
(Koltchinskii & Panchenko 2000, Bartlett & Mendelson 2001, Bartlett & Mendelson
2002). For additional reading see, for example, (Bousquet 2002, Boucheron,
Bousquet & Lugosi 2005, Bartlett, Bousquet & Mendelson 2005). Our proof of
the concentration lemma is due to Kakade and Tewari lecture notes. Kakade,
Sridharan, and Tewari (2008) gave a unified framework for deriving bounds on the
Rademacher complexity of linear classes with respect to different assumptions on
the norms.

27
Covering Numbers

In this chapter we describe another way to measure the complexity of sets, which is
called covering numbers.

27.1 COVERING
Definition 27.1 (Covering). Let A Rm be a set of vectors. We say that A is r covered by a set A , with respect to the Euclidean metric, if for all a A there exists
a A with a a  r . We define by N(r , A) the cardinality of the smallest A that
r -covers A.
Example 27.1 (Subspace). Suppose that A Rm , let c = maxa A a,
andd assume
m
that A lies in a d-dimensional subspace of R . Then, N(r , A) (2c d/r ) . To see
this, let v1 , . . . , vd be an orthonormal basis of the subspace. Then, any a A can be


written as a = di=1 i vi with  2 = a2 c. Let R and consider the set

A =

d


+
i vi : i , i {c, c + , c + 2 , . . . , c} .

i=1

Given a A s.t. a =

d

i=1 i vi

a a 2 = 

with  c, there exists a A such that



(i i )vi 2 2

vi 2 2 d.

Choose = r / d; then a a  r and therefore A is an r -cover of A. Hence,



N(r , A) |A | =

2c

d
=

d
2c d
.
r

337

338

Covering Numbers

27.1.1 Properties
The following lemma is immediate from the definition.
Lemma 27.2. For any A Rm , scalar c > 0, and vector a0 Rm , we have
r > 0, N(r , {c a + a0 : a A}) N(cr , A).
Next, we derive a contraction principle.
Lemma 27.3. For each i [m], let i : R R be a -Lipschitz function; namely, for
all , R we have |i () i ()| | |. For a Rm let (a) denote the vector
(1 (a1 ), . . . , m (am )). Let A = {(a) : a A}. Then,
N( r , A) N(r , A).
Proof. Define B = A. Let A be an r -cover of A and define B  = A . Then, for
all a A there exists a A with a a  r . So,


(a) (a )2 =
(i (ai ) i (ai ))2 2
(ai ai )2 (r )2 .
i

Hence,

is an ( r )-cover of B.

27.2 FROM COVERING TO RADEMACHER COMPLEXITY VIA CHAINING

The following lemma bounds the Rademacher complexity of A based on the covering numbers N(r , A). This technique is called Chaining and is attributed to
Dudley.
Lemma 27.4. Let c = mina maxa A a a . Then, for any integer M > 0,
M

c 2M 6 c  k
R(A) +
2
log (N(c 2k , A)).
m
m
k=1

Proof. Let a be a minimizer of the objective function given in the definition of c.

On the basis of Lemma 26.6, we can analyze the Rademacher complexity assuming
that a = 0.
Consider the set B0 = {0} and note that it is a c-cover of A. Let B1 , . . . , B M
be sets such that each Bk corresponds to a minimal (c 2k )-cover of A. Let a =
argmaxa A  , a (where if there is more than one maximizer, choose one in an arbitrary way, and if a maximizer does not exist, choose a such that  , a  is close
enough to the supremum). Note that a is a function of . For each k, let b(k) be the
nearest neighbor of a in Bk (hence b(k) is also a function of ). Using the triangle
inequality,
b(k) b(k1)  b(k) a  + a b(k1)  c (2k + 2(k1) ) = 3 c 2k .
For each k define the set
B k = {(a a) : a Bk , a Bk1 , a a  3 c 2k }.

27.2 From Covering to Rademacher Complexity via Chaining

We can now write

1
E , a 
m


M

1
= E  , a b(M)  +
 , b(k) b(k1) 
m
k=1


M



1
1

(M)
E sup  , a .
E   a b  +
m
m
a B

R(A) =

k=1

Since   = m and a b(M)  c 2M , the first summand is at most cm 2M .

Additionally, by Massart lemma,


1
2 log (N(c 2k , A)2 )
log (N(c 2k , A))
k
k
E sup  , a 3 c 2
= 6c2
.
m a B
m
m
k

Therefore,
c 2M 6c  k
R(A) +
2
m
m
M


log (N(c2k , A)).

k=1

As a corollary we obtain the following:

Lemma 27.5. Assume that there are , > 0 such that for any k 1 we have

log (N(c2k , A)) + k.
Then,
R(A)

6c
( + 2).
m

Proof. The bound follows from Lemma 27.4 by taking M and noting that



k
k = 2.
= 1 and
k=1 2
k=1 k2
Example 27.2. Consider a set A which lies in a d dimensional subspace of Rm and
 d
such that c = maxa A a. We have shown that N(r , A) 2cr d . Therefore, for
any k,




k
log (N(c2 , A)) d log 2k+1 d


d log (2 d) + k d


d log (2 d) + d k.

Hence Lemma 27.5 yields

 




6c
c d log (d)
R(A)
d log (2 d) + 2 d = O
.
m
m

339

340

Covering Numbers

27.3 BIBLIOGRAPHIC REMARKS

The chaining technique is due to Dudley (1987). For an extensive study of covering
numbers as well as other complexity measures that can be used to bound the rate of
uniform convergence we refer the reader to (Anthony & Bartlet 1999).

28
Proof of the Fundamental Theorem
of Learning Theory

In this chapter we prove Theorem 6.8 from Chapter 6. We remind the reader the
conditions of the theorem, which will hold throughout this chapter: H is a hypothesis
class of functions from a domain X to {0, 1}, the loss function is the 0 1 loss, and
VCdim(H) = d < .
We shall prove the upper bound for both the realizable and agnostic cases and
shall prove the lower bound for the agnostic case. The lower bound for the realizable
case is left as an exercise.

28.1 THE UPPER BOUND FOR THE AGNOSTIC CASE

For the upper bound we need to prove that there exists C such that H is agnostic
PAC learnable with sample complexity
m H ( , ) C

d + ln (1/)
.
2

We will prove the slightly looser bound:

d log (d/ ) + ln (1/)
.
(28.1)
2
The tighter bound in the theorem statement requires a more involved proof, in
which a more careful analysis of the Rademacher complexity using a technique
called chaining should be used. This is beyond the scope of this book.
To prove Equation (28.1), it suffices to show that applying the ERM with a
sample size



64d
8 
32d
+ 2 8d log (e/d) + 2 log (4/)
m 4 2 log
2



m H ( , ) C

yields an , -learner for H. We prove this result on the basis of Theorem 26.5.
Let (x1 , y1 ), . . . , (xm , ym ) be a classification training set. Recall that the SauerShelah lemma tells us that if VCdim(H) = d then

d


{(h(x1 ), . . . , h(xm )) : h H} e m .
d
341

342

Proof of the Fundamental Theorem of Learning Theory

Denote A = {(1[h(x1 )= y1 ] , . . . , 1[h(xm )= ym ] ) : h H}. This clearly implies that
 e m d
|A|
.
d
Combining this with Lemma 26.8 we obtain the following bound on the Rademacher
complexity:

2d log (em/d)
R(A)
.
m
Using Theorem 26.5 we obtain that with probability of at least 1 , for every h H
we have that


8d log (em/d)
2 log (2/)
L D (h) L S (h)
+
.
m
m
Repeating the previous argument for minus the zero-one loss and applying the
union bound we obtain that with probability of at least 1 , for every h H it
holds that


8d log (em/d)
2 log (4/)
|L D (h) L S (h)|
+
m
m

8d log (em/d) + 2 log (4/)
.
2
m
To ensure that this is smaller than we need

4 
m 2 8d log (m) + 8d log (e/d) + 2 log (4/) .

Using Lemma A.2, a sufficient condition for the inequality to hold is that



32d
64d
8 
m 4 2 log
+ 2 8d log (e/d) + 2 log (4/) .
2

28.2 THE LOWER BOUND FOR THE AGNOSTIC CASE

Here, we prove that there exists C such that H is agnostic PAC learnable with
sample complexity
d + ln (1/)
m H ( , ) C
.
2
We will prove the lower bound in two parts. First, we will show that m( , )
0. 5 log (1/(4))/ 2 , and second we will show that for every 1/8 we have that
m( , ) 8d/ 2 . These two bounds will conclude the proof.

28.2.1 Showing That m(

, ) 0. 5 log (1/(4))/
2

We first show that for any < 1/ 2 and any (0, 1), we have that m( , )
0. 5 log (1/(4))/ 2 . To do so, we show that for m 0. 5 log (1/(4))/ 2 , H is not
learnable.
Choose one example that is shattered by H. That is, let c be an example such that
there are h + , h H for which h + (c) = 1 and h (c) = 1. Define two distributions,

28.2 The Lower Bound for the Agnostic Case

D+ and D , such that for b {1} we have


Db ({(x, y)}) =

1+yb
2

if x = c
otherwise.

That is, all the distribution mass is concentrated on two examples (c, 1) and (c, 1),
1b
where the probability of (c, b) is 1+b
2 and the probability of (c, b) is 2 .
Let A be an arbitrary algorithm. Any training set sampled from Db has the
form S = (c, y1 ), . . . , (c, ym ). Therefore, it is fully characterized by the vector y =
(y1 , . . . , ym ) {1}m . Upon receiving a training set S, the algorithm A returns a
hypothesis h : X {1}. Since the error of A w.r.t. Db only depends on h(c), we
can think of A as a mapping from {1}m into {1}. Therefore, we denote by A(y)
the value in {1} corresponding to the prediction of h(c), where h is the hypothesis
that A outputs upon receiving the training set S = (c, y1 ), . . . , (c, ym ).
Note that for any hypothesis h we have
L Db (h) =

1 h(c)b
.
2

In particular, the Bayes optimal hypothesis is h b and

1 A(y)b 1
L Db (A(y)) L Db (h b ) =

=
2
2

if A(y) = b

otherwise.

Fix A. For b {1}, let Y b = {y {0, 1}m : A(y) = b}. The distribution Db induces
a probability Pb over {1}m . Hence,
P [L Db (A(y)) L Db (h b ) = ] = Db (Y b ) =

Pb [y]1[A(y)=b] .

Denote N + = {y : |{i : yi = 1}| m/2} and N = {1}m \ N + . Note that for any y N +
we have P+ [y] P [y] and for any y N we have P [y] P+ [y]. Therefore,
max P[L Db (A(y)) L Db (h b ) = ]

b{1}

= max

b{1}

Pb [y]1[A(y)=b]

1
1
P+ [y]1[A(y)=+] +
P [y]1[A(y)=]
2 y
2 y

1 
(P+ [y]1[A(y)=+] + P [y]1[A(y)=] )
2
+
yN

1 
(P+ [y]1[A(y)=+] + P [y]1[A(y)=] )
2

yN

343

344

Proof of the Fundamental Theorem of Learning Theory

1 
(P [y]1[A(y)=+] + P [y]1[A(y)=] )
2
+
yN

1 
(P+ [y]1[A(y)=+] + P+ [y]1[A(y)=] )
2

yN

1 
1 
=
P [y] +
P+ [y].
2
2
+

yN

yN

Next note that yN + P [y] = yN P+ [y], and both values are the probability that
a Binomial (m, (1 )/2) random variable will have value greater than m/2. Using
Lemma B.11, this probability is lower bounded by






1
1
1 1 exp( m 2 /(1 2 ))
1 1 exp( 2m 2 ) ,
2
2
where we used the assumption that 2 1/2. It follows that if m 0. 5 log (1/(4))/ 2
then there exists b such that
P [L Db (A(y)) L Db (h b ) = ]




1 1 4 ,
2
where the last inequality follows by standard algebraic manipulations. This concludes our proof.

28.2.2 Showing That m( , 1/8) 8d/ 2

We shall now prove that for every < 1/(8 2) we have that m( , ) 8d
.
2

Let = 8 and note that (0, 1/ 2). We will construct a family of distributions
as follows. First, let C = {c1 , . . . , cd } be a set of d instances which are shattered by H.
Second, for each vector (b1 , . . . , bd ) {1}d , define a distribution Db such that

1 1+ybi
2
if i : x = ci
Db ({(x, y)}) = d
0
otherwise.
That is, to sample an example according to Db , we first sample an element ci C
uniformly at random, and then set the label to be bi with probability (1 + )/2 or
bi with probability (1 )/2.
It is easy to verify that the Bayes optimal predictor for Db is the hypothesis h H
such that h(ci ) = bi for all i [d], and its error is 1
2 . In addition, for any other
function f : X {1}, it is easy to verify that
L Db ( f ) =

1 + |{i [d] : f (ci ) = bi }| 1 |{i [d] : f (ci ) = bi }|

.
2
d
2
d

Therefore,
L Db ( f ) min L Db (h) =
hH

|{i [d] : f (ci ) = bi }|

.
d

(28.2)

28.2 The Lower Bound for the Agnostic Case

Next, fix some learning algorithm A. As in the proof of the No-Free-Lunch

theorem, we have that


max
E m L Db (A(S)) min L Db (h)
(28.3)
Db :b{1}d SDb

hH


E m L Db (A(S)) min L Db (h)


hH

Db :bU ({1}d ) SDb

(28.4)



|{i [d] : A(S)(ci ) = bi |
Em
d
Db :bU ({1}d ) SDb

(28.5)


E
E 1[A(S)(ci )=bi ] ,
d
Db :bU ({1}d ) SDbm

(28.6)

i=1

where the first equality follows from Equation (28.2). In addition, using the definition of Db , to sample S Db we can first sample ( j1 , . . . , jm ) U ([d])m , set xr = c ji ,
and finally sample yr such that P [yr = b ji ] = (1 + )/2. Let us simplify the notation
and use y b to denote sampling according to P[y = b] = (1 + )/2. Therefore, the
right-hand side of Equation (28.6) equals

E
E
E 1[A(S)(ci )=bi ] .
d
j U ([d])m bU ({1}d ) r,yr b jr
d

(28.7)

i=1

We now proceed in two steps. First, we show that among all learning algorithms,
A, the one which minimizes Equation (28.7) (and hence also Equation (28.4)) is the
Maximum-Likelihood learning rule, denoted A M L . Formally, for each i , A M L (S)(ci )
is the majority vote among the set {yr : r [m], xr = ci }. Second, we lower bound
Equation (28.7) for A M L .
Lemma 28.1. Among all algorithms, Equation (28.4) is minimized for A being the
Maximum-Likelihood algorithm, A M L , defined as



yr .
i , A M L (S)(ci ) = sign
r:xr =ci

Proof. Fix some j [d]m . Note that given j and y {1}m , the training set S is fully
determined. Therefore, we can write A( j , y) instead of A(S). Let us also fix i [d].
Denote bi the sequence (b1 , . . . , bi1 , bi+1 , . . . , bm ). Also, for any y {1}m , let y I
denote the elements of y corresponding to indices for which jr = i and let y I be
the rest of the elements of y. We have
E

bU ({1}d ) r,yr b jr

1 
2

1[A(S)(ci )=bi ]
E

bi U ({1}d1 ) y
bi {1}

bi U ({1}d1 ) I
y

P[y|bi , bi ]1[A( j ,y)(ci )=bi ]



1

P[y I |bi ]
P[y I |bi ]1[A( j ,y)(ci )=bi ] .
2 I
y

bi {1}

345

346

Proof of the Fundamental Theorem of Learning Theory

The sum within the parentheses is minimized when A( j , y)(ci ) is the maximizer of
P[y I |bi ] over bi {1}, which is exactly the Maximum-Likelihood rule. Repeating
the same argument for all i we conclude our proof.
Fix i . For every j , let n i ( j ) = {|t : jt = i |} be the number of instances in which the
instance is ci . For the Maximum-Likelihood rule, we have that the quantity
E

bU ({1}d ) r,yr b jr

1[A M L (S)(ci )=bi ]

is exactly the probability that a binomial (n i ( j ), (1 )/2) random variable will be

larger than n i ( j )/2. Using Lemma B.11, and the assumption 2 1/2, we have that



1
2
2n
(
j
)
i
P[B n i ( j )/2]
.
1 1e
2
We have thus shown that

E
E
E 1[A(S)(ci )=bi ]
d
j U ([d])m bU ({1}d ) r,yr b jr
d

i=1




d

2n ( j )
2
i

E
1 1e
2d
j U ([d])m
i=1



d


E
1 2 2 n i ( j ) ,
2d
j U ([d])m
i=1

where in the last inequality we used the inequality 1 ea a.

Since the square root function is concave, we can apply Jensens inequality to
obtain that the above is lower bounded by


d


2
E
ni ( j )

1 2
2d
j U ([d])m
i=1



d 

1 2 2 m/d
2d
i=1




2
=
1 2 m/d .
2
=

As long as m <

d
,
8 2

this term would be larger than /4.

In summary, we have shown that if m < 8d 2 then for any algorithm there exists a
distribution such that


E m L D (A(S)) min L D (h) /4.
SD

hH

28.3 The Upper Bound for the Realizable Case

Finally, Let  = 1 (L D (A(S)) minhH L D (h)) and note that  [0, 1] (see
Equation (28.5)). Therefore, using Lemma B.1, we get that




P [L D (A(S)) min L D (h) > ] = P  >
E []
hH

Choosing = 8 we conclude that if m <

will have L D (A(S)) minhH L D (h) .

8d
,
2

1
.
4

then with probability of at least 1/8 we

28.3 THE UPPER BOUND FOR THE REALIZABLE CASE

Here we prove that there exists C such that H is PAC learnable with sample
complexity
m H ( , ) C

d ln (1/ ) + ln (1/)
.

(1/)
We do so by showing that for m C d ln (1/ )+ln
, H is learnable using the ERM

rule. We prove this claim on the basis of the notion of -nets.

Definition 28.2 ( -net). Let X be a domain. S X is an -net for H 2X with

respect to a distribution D over X if
h H : D(h) h S = .
Theorem 28.3. Let H 2X with VCdim(H) = d. Fix (0, 1), (0, 1/4) and let



 
8
16e
2
m
2d log
+ log
.

Then, with probability of at least 1 over a choice of S Dm we have that S is an

-net for H.
Proof. Let
B = {S X : |S| = m, h H, D(h) , h S = }
be the set of sets which are not -nets. We need to bound P [S B]. Define
B  = {(S, T ) X : |S| = |T | = m, h H, D(h) , h S = , |T h| >
Claim 1
P [S B] 2 P [(S, T ) B  ].
Proof of Claim 1: Since S and T are chosen independently we can write







E
P[(S, T ) B ] =
E m 1[(S,T )B  ] .
1[(S,T )B  ] = E m
(S,T )D 2m

SD

T D

m
2 }.

347

348

Proof of the Fundamental Theorem of Learning Theory

Note that (S, T ) B  implies S B and therefore 1[(S,T )B  ] = 1[(S,T )B  ] 1[SB] , which
gives
P [(S, T ) B  ] = E m E m 1[(S,T )B  ] 1[SB]
SD T D

= E m 1[SB] E m 1[(S,T )B  ] .
T D

SD

Fix some S. Then, either 1[SB] = 0 or S B and then h S such that D(h S ) and
|h S S| = 0. It follows that a sufficient condition for (S, T ) B  is that |T h S | > m
2 .
Therefore, whenever S B we have
E 1[(S,T )B  ]

T D m

P [|T h S | >

T D m

m
2 ].

But, since we now assume S B we know that D(h S ) = . Therefore, |T h S |

is a binomial random variable with parameters (probability of success for a single
try) and m (number of tries). Chernoffs inequality implies
P [|T h S |

m
2 ]

2
m (mm/2)2

= em/2 em /2 ed log (1/)/2 = d/2 1/2.

Thus,
P [|T h S | >

m
2 ]

= 1 P[|T h S |

m
2 ]

1 P[|T h S |

m
2 ]

1/2.

Combining all the preceding we conclude the proof of Claim 1.

Claim 2 (Symmetrization):
P [(S, T ) B  ] e m/4 H (2m).
Proof of Claim 2: To simplify notation, let = m /2 and for a sequence A =
(x 1 , . . . , x 2m ) let A0 = (x 1 , . . . , x m ). Using the definition of B  we get that
P [ A B ] =

max 1[D(h) ] 1[|hA0 |=0] 1[|hA|]

max 1[|hA0 |=0] 1[|hA|] .

AD 2m hH
AD 2m hH

Now, let us define by H A the effective number of different hypotheses on A, namely,

H A = {h A : h H }. It follows that
P [ A B ]

max 1[|hA0 |=0] 1[|hA|]

AD 2m hH A

AD 2m

1[|hA0 |=0] 1[|hA|] .

hH A

Let J = {j [2m] : |j| = m}. For any j J and A = (x 1 , . . . , x 2m ) define Aj =

(x j1 , . . . , x jm ). Since the elements of A are chosen i.i.d., we have that for any j J and
any function f (A, A0 ) it holds that E AD2m [ f (A, A0 )] = E AD2m [ f (A, Aj )]. Since
this holds for any j it also holds for the expectation

of j chosen at random from J .
In particular, it holds for the function f (A, A0 ) = hH A 1[|hA0 |=0] 1[|hA|] . We

28.3 The Upper Bound for the Realizable Case

therefore obtain that

P [ A B ]
=

AD 2m j J

AD 2m

1[|hAj |=0] 1[|hA|]

hH A

1[|hA|] E 1[|hAj |=0] .

hH A

j J

Now, fix some A s.t. |h A| . Then, E j 1[|hAj |=0] is the probability that when
choosing m balls from a bag with at least red balls, we will never choose a red ball.
This probability is at most
(1 /(2m))m = (1 /4)m e m/4 .
We therefore get that
P [ A B ]

AD 2m

e m/4 e m/4

hH A

AD 2m

|H A |.

Using the definition of the growth function we conclude the proof of Claim 2.
Completing the Proof: By Sauers lemma we know that H (2m) (2em/d)d .
Combining this with the two claims we obtain that
P [S B] 2(2em/d)d e m/4 .
We would like the right-hand side of the inequality to be at most ; that is,
2(2em/d)d e m/4 .
Rearranging, we obtain the requirement
m

 4d
4
4
d log (2em/d) + log (2/) =
log (m) + (d log (2e/d) + log (2/).

Using Lemma A.2, a sufficient condition for the preceding to hold is that
 
8d
16d
8
log
m
+ (d log (2e/d) + log (2/).



A sufficient condition for this is that
 
8d
16d
16
log
m
+ (d log (2e/d) + 12 log (2/)



 

16d
8
8d 2e
=
log
+ log (2/)

d




 
8
16e
2
=
2d log
+ log
.

and this concludes our proof.

28.3.1 From -Nets to PAC Learnability

Theorem 28.4. Let H be a hypothesis class over X with VCdim(H) = d. Let D be a
distribution over X and let c H be a target hypothesis. Fix , (0, 1) and let m be
as defined in Theorem 28.3. Then, with probability of at least 1 over a choice of m

349

350

Proof of the Fundamental Theorem of Learning Theory

i.i.d. instances from X with labels according to c we have that any ERM hypothesis
has a true error of at most .




Proof. Define the class Hc = {c h : h H}, where c h = (h \ c) (c \ h). It is
easy to verify that if some A X is shattered by H then it is also shattered by Hc
and vice versa. Hence, VCdim(H) = VCdim(Hc ). Therefore, using Theorem 28.3
we know that with probability
of at least 1 , the sample S is an -net for Hc .


Note

that L D (h) = D(h c). Therefore, for any h H with L D (h) we have that
|(h c) S| > 0, which implies that h cannot be an ERM hypothesis, which concludes
our proof.

29
Multiclass Learnability

In Chapter 17 we have introduced the problem of multiclass categorization, in which

the goal is to learn a predictor h : X [k]. In this chapter we address PAC learnability of multiclass predictors with respect to the 0-1 loss. As in Chapter 6, the main
goal of this chapter is to:

Characterize which classes of multiclass hypotheses are learnable in the (multiclass) PAC model.

Quantify the sample complexity of such hypothesis classes.

In view of the fundamental theorem of learning theory (Theorem 6.8), it is natural to seek a generalization of the VC dimension to multiclass hypothesis classes.
In Section 29.1 we show such a generalization, called the Natarajan dimension, and
state a generalization of the fundamental theorem based on the Natarajan dimension. Then, we demonstrate how to calculate the Natarajan dimension of several
important hypothesis classes.
Recall that the main message of the fundamental theorem of learning theory is
that a hypothesis class of binary classifiers is learnable (with respect to the 0-1 loss)
if and only if it has the uniform convergence property, and then it is learnable by
any ERM learner. In Chapter 13, Exercise 29.2, we have shown that this equivalence breaks down for a certain convex learning problem. The last section of this
chapter is devoted to showing that the equivalence between learnability and uniform
convergence breaks down even in multiclass problems with the 0-1 loss, which are
very similar to binary classification. Indeed, we construct a hypothesis class which is
learnable by a specific ERM learner, but for which other ERM learners might fail
and the uniform convergence property does not hold.

29.1 THE NATARAJAN DIMENSION

In this section we define the Natarajan dimension, which is a generalization of the
VC dimension to classes of multiclass predictors. Throughout this section, let H
be a hypothesis class of multiclass predictors; namely, each h H is a function
from X to [k].
351

352

Multiclass Learnability

To define the Natarajan dimension, we first generalize the definition of

shattering.
Definition 29.1 (Shattering (Multiclass Version)). We say that a set C X is
shattered by H if there exist two functions f0 , f1 : C [k] such that

For every x C, f 0 (x) = f 1 (x).

For every B C, there exists a function h H such that

x B, h(x) = f 0 (x) and x C \ B, h(x) = f 1 (x).

Definition 29.2 (Natarajan Dimension). The Natarajan dimension of H, denoted
Ndim(H), is the maximal size of a shattered set C X .
It is not hard to see that in the case that there are exactly two classes,
Ndim(H) = VCdim(H). Therefore, the Natarajan dimension generalizes the VC
dimension. We next show that the Natarajan dimension allows us to generalize the
fundamental theorem of statistical learning from binary classification to multiclass
classification.

29.2 THE MULTICLASS FUNDAMENTAL THEOREM

Theorem 29.3 (The Multiclass Fundamental Theorem). There exist absolute constants C1 , C2 > 0 such that the following holds. For every hypothesis class H of
functions from X to [k], such that the Natarajan dimension of H is d, we have
1. H has the uniform convergence property with sample complexity
C1

d + log (1/)
d log (k) + log (1/)
m UC
.
H ( , ) C2
2
2

2. H is agnostic PAC learnable with sample complexity

C1

d + log (1/)
d log (k) + log (1/)
m H ( , ) C2
.
2
2

3. H is PAC learnable (assuming realizability) with sample complexity

 
d log kd
d + log (1/)
+ log (1/)
C1
m H ( , ) C2
.

29.2.1 On the Proof of Theorem 29.3

The lower bounds in Theorem 29.3 can be deduced by a reduction from the binary
fundamental theorem (see Exercise 29.5).
The upper bounds in Theorem 29.3 can be proved along the same lines of the
proof of the fundamental theorem for binary classification, given in Chapter 28
(see Exercise 29.4). The sole ingredient of that proof that should be modified in
a nonstraightforward manner is Sauers lemma. It applies only to binary classes and
therefore must be replaced. An appropriate substitute is Natarajans lemma:
Lemma 29.4 (Natarajan). |H| |X |Ndim(H) k 2Ndim(H) .

29.3 Calculating the Natarajan Dimension

The proof of Natarajans lemma shares the same spirit of the proof of Sauers
lemma and is left as an exercise (see Exercise 29.3).

29.3 CALCULATING THE NATARAJAN DIMENSION

In this section we show how to calculate (or estimate) the Natarajan dimension of
several popular classes, some of which were studied in Chapter 17. As these calculations indicate, the Natarajan dimension is often proportional to the number of
parameters required to define a hypothesis.

29.3.1 One-vs.-All Based Classes

In Chapter 17 we have seen two reductions of multiclass categorization to binary
classification: One-vs.-All and All-Pairs. In this section we calculate the Natarajan
dimension of the One-vs.-All method.
Recall that in One-vs.-All we train, for each label, a binary classifier that distinguishes between that label and the rest of the labels. This naturally suggests
considering multiclass hypothesis classes of the following form. Let Hbin {0, 1}X
:X
be a binary hypothesis class. For every h = (h 1 , . . . , h k ) (Hbin )k define T (h)
[k] by

T (h)(x)
= argmax h i (x).
i[k]

If there are two labels that maximize h i (x), we choose the smaller one. Also, let
OvA,k
: h (Hbin )k }.
Hbin
= {T (h)
OvA,k
? Intuitively, to specify a
What should be the Natarajan dimension of Hbin
hypothesis in Hbin we need d = VCdim(Hbin ) parameters. To specify a hypotheOvA,k
, we need to specify k hypotheses in Hbin . Therefore, kd parameters
sis in Hbin
should suffice. The following lemma establishes this intuition.

Lemma 29.5. If d = VCdim(Hbin ) then

OvA,k
Ndim(Hbin
) 3kd log (kd) .

Proof. Let C X be a shattered set. By the definition of shattering (for multiclass

hypotheses)


OvA,k
2|C| .
Hbin
C

OvA,k
Hbin

On the other hand, each hypothesis in

is determined by using k hypotheses
from Hbin . Therefore,


OvA,k
Hbin
| (Hbin )C |k .
C

By Sauers lemma, | (Hbin )C

| |C|d .

2|C|

We conclude that


OvA,k
Hbin
|C|dk .
C

The proof follows by taking the logarithm and applying Lemma A.1.

353

354

Multiclass Learnability

How tight is Lemma 29.5? It is not hard to see that for some classes,
OvA,k
) can be much smaller than dk (see Exercise 29.1). However, there
Ndim(Hbin
OvA,k
are several natural binary classes, Hbin (e.g., halfspaces), for which Ndim(Hbin
)=
(dk) (see Exercise 29.6).

29.3.2 General Multiclass-to-Binary Reductions

The same reasoning used to establish Lemma 29.5 can be used to upper bound the
Natarajan dimension of more general multiclass-to-binary reductions. These reductions train several binary classifiers on the data. Then, given a new instance, they
predict its label by using some rule that takes into account the labels predicted by
the binary classifiers. These reductions include One-vs.-All and All-Pairs.
Suppose that such a method trains l binary classifiers from a binary class Hbin ,
and r : {0, 1}l [k] is the rule that determines the (multiclass) label according to
the predictions of the binary classifiers. The hypothesis class corresponding to this
:
method can be defined as follows. For every h = (h 1 , . . . , h l ) (Hbin )l define R(h)
X [k] by
x) = r (h 1 (x), . . . , h l (x)).
R(h)(
Finally, let

: h (Hbin )l }.
Hrbin = {R(h)

Similarly to Lemma 29.5 it can be proven that:

Lemma 29.6. If d = VCdim(Hbin ) then
Ndim(Hrbin ) 3l d log (l d) .
The proof is left as Exercise 29.2.

29.3.3 Linear Multiclass Predictors

Next, we consider the class of linear multiclass predictors (see Section 17.2). Let
 : X [k] Rd be some class-sensitive feature mapping and let

+
H =

x  argmaxw, (x,i ) : w Rd

(29.1)

i[k]

Each hypothesis in H is determined by d parameters, namely, a vector w Rd .

Therefore, we would expect that the Natarajan dimension would be upper bounded
by d. Indeed:
Theorem 29.7. Ndim(H ) d .
Proof. Let C X be a shattered set, and let f 0 , f1 : C [k] be the two functions
that witness the shattering. We need to show that |C| d. For every x C let (x) =
def

(x, f 0 (x)) (x, f1 (x)). We claim that the set (C) = {(x) : x C} consists of
|C| elements (i.e., is one to one) and is shattered by the binary hypothesis class of
homogeneous linear separators on Rd ,
H = {x  sign(w, x) : w Rd }.

29.4 On Good and Bad ERMs

Since VCdim(H) = d, it will follow that |C| = |(C)| d, as required.

To establish our claim it is enough to show that |H(C) | = 2|C| . Indeed, given a
subset B C, by the definition of shattering, there exists h B H for which
x B, h B (x) = f0 (x) and x C \ B, h B (x) = f 1 (x).
Let w B Rd be a vector that defines h B . We have that, for every x B,
w, (x, f0 (x)) > w, (x, f1 (x)) w, (x) > 0.
Similarly, for every x C \ B,
w, (x) < 0.
It follows that the hypothesis g B H defined by the same w Rd label the points in
(B) by 1 and the points in (C \ B) by 0. Since this holds for every B C we obtain
that |C| = |(C)| and |H(C) | = 2|C| , which concludes our proof.
The theorem is tight in the sense that there are mappings  for which
Ndim(H ) = (d). For example, this is true for the multivector construction
(see Section 17.2 and the Bibliographic Remarks at the end of this chapter). We
therefore conclude:
Corollary 29.8. Let X = Rn and let  : X [k] Rnk be the class sensitive feature
mapping for the multi-vector construction:
(x, y) = [ 0, . . . , 0 , x 1 , . . . , x n , 0, . . . , 0 ].
9 :; < 9 :; < 9 :; <
R(y1)n

Rn

R(ky)n

Let H be as defined in Equation (29.1). Then, the Natarajan dimension of H

satisfies
(k 1)(n 1) Ndim(H ) kn.

29.4 ON GOOD AND BAD ERMS

In this section we present an example of a hypothesis class with the property that not
all ERMs for the class are equally successful. Furthermore, if we allow an infinite
number of labels, we will also obtain an example of a class that is learnable by some
ERM, but other ERMs will fail to learn it. Clearly, this also implies that the class is
learnable but it does not have the uniform convergence property. For simplicity, we
consider only the realizable case.
The class we consider is defined as follows. The instance space X will be any
finite or countable set. Let P f (X ) be the collection of all finite and cofinite subsets
of X (that is, for each A P f (X ), either A or X \ A must be finite). Instead of [k],
the label set is Y = P f (X ) {}, where is some special label. For every A P f (X )
define h A : X Y by

A xA
h A (x) =
x
/A
Finally, the hypothesis class we take is
H = {h A : A P f (X )}.

355

356

Multiclass Learnability

Let A be some ERM algorithm for H. Assume that A operates on a sample labeled
by h A H. Since h A is the only hypothesis in H that might return the label A, if
A observes the label A, it knows" that the learned hypothesis is h A , and, as an
ERM, must return it (note that in this case the error of the returned hypothesis
is 0). Therefore, to specify an ERM, we should only specify the hypothesis it returns
upon receiving a sample of the form
S = {(x 1 , ), . . . , (x m , )}.
We consider two ERMs: The first, Agood , is defined by
Agood (S) = h ;
that is, it outputs the hypothesis which predicts * for every x X . The second
ERM, Abad , is defined by
Abad (S) = h {x1 ,...xm }c .
The following claim shows that the sample complexity of Abad is about |X |-times
larger than the sample complexity of Agood . This establishes a gap between different
ERMs. If X is infinite, we even obtain a learnable class that is not learnable by
every ERM.
Claim 29.9.
1. Let , > 0, D a distribution
  over X and h A H. Let S be an i.i.d. sample
1
consisting of m log 1 examples, sampled according to D and labeled by
h A . Then, with probability of at least 1 , the hypothesis returned by Agood
will have an error of at most .
2. There exists a constant a > 0 such that for every 0 < < a there exists a distribution D over X and h A H such that the following holds. The hypothesis
|1
, sampled according
returned by Abad upon receiving a sample of size m |X6
1

to D and labeled by h A , will have error with probability e 6 .

Proof. Let D be a distribution over X and suppose that the correct labeling is h A .
For any sample, Agood returns either h or h A . If it returns h A then its true error is
zero. Thus, it returns a hypothesis with error only if all the m examples in the
sample are from X \ A while the error of h , L D (h ) = PD [ A], is . Assume m
1
1
m
m .
log ( ); then the probability of the latter event is no more than (1 ) e
This establishes item 1.
Next we prove item 2. We restrict the proof to the case that |X | = d < . The
proof for infinite X is similar. Suppose that X = {x 0 , . . . , x d1 }.
Let a > 0 be small enough such that 1 2 e4 for every < a and fix some
< a. Define a distribution on X by setting P [x 0 ] = 1 2 and for all 1 i d 1,
2
. Suppose that the correct hypothesis is h and let the sample size be m.
P [x i ] = d1
Clearly, the hypothesis returned by Abad will err on all the examples from X which
61
,
are not in the sample. By Chernoffs bound, if m d1
6 , then with probability e
the sample will include no more than d1
examples
from
X
.
Thus
the
returned
2
hypothesis will have error .

29.6 Exercises

The conclusion of the example presented is that in multiclass classification, the

sample complexity of different ERMs may differ. Are there good ERMs for every
hypothesis class? The following conjecture asserts that the answer is yes.
Conjecture 29.10. The realizable sample complexity of every hypothesis class H
[k]X is


Ndim(H)
m H ( , ) = O
.

We emphasize that the O notation may hide only poly-log factors of , , and
Ndim(H), but no factor of k.

29.5 BIBLIOGRAPHIC REMARKS

The Natarajan dimension is due to Natarajan (1989). That paper also established
the Natarajan lemma and the generalization of the fundamental theorem. Generalizations and sharper versions of the Natarajan lemma are studied in Haussler and
Long (1995). Ben-David, Cesa-Bianchi, Haussler, and Long (1995) defined a large
family of notions of dimensions, all of which generalize the VC dimension and may
be used to estimate the sample complexity of multiclass classification.
The calculation of the Natarajan dimension, presented here, together with calculation of other classes, can be found in Daniely et al. (2012). The example of good
and bad ERMs, as well as conjecture 29.10, are from Daniely et al. (2011).

29.6 EXERCISES
29.1 Let d, k > 0. Show that there exists a binary hypothesis Hbin of VC dimension d
) = d.
such that Ndim(HOvA,k
bin
29.2 Prove Lemma 29.6.
29.3 Prove Natarajans lemma.
Hint: Fix some x0 X . For i , j [k], denote by Hi j all the functions f : X \{x0 } [k]
that can be extended to a function in H
both by defining f (x0 ) = i and by defining
f (x0 ) = j . Show that |H| |HX \{x0 } | + i= j |Hi j | and use induction.
29.4 Adapt the proof of the binary fundamental theorem and Natarajans lemma to
prove that, for some universal constant C > 0 and for every hypothesis class of
Natarajan dimension d, the agnostic sample complexity of H is
 
d log kd
+ log (1/)
m H ( , ) C
.
2
29.5 Prove that, for some universal constant C > 0 and for every hypothesis class of
Natarajan dimension d, the agnostic sample complexity of H is
m H ( , ) C

d + log (1/)
.
2

Hint: Deduce it from the binary fundamental theorem.

29.6 Let H be the binary hypothesis class of (nonhomogenous) halfspaces in Rd . The
goal of this exercise is to prove that Ndim(HOvA,k ) (d 1) (k 1).

357

358

Multiclass Learnability

1. Let Hdiscrete be the class of all functions f : [k 1] [d 1] {0, 1} for which

there exists some i 0 such that, for every j [d 1]
i < i 0 , f (i , j ) = 1 while i > i 0 , f (i , j ) = 0.
Show that Ndim(HOvA,k
discrete ) = (d 1) (k 1).
2. Show that Hdiscrete can be realized by H. That is, show that there exists a
mapping : [k 1] [d 1] Rd such that
Hdiscrete {h : h H}.

Hint: You can take (i , j ) to be the vector whose j th coordinate is 1, whose

last coordinate is i , and the rest are zeros.
3. Conclude that Ndim(HOvA,k ) (d 1) (k 1).

30
Compression Bounds

Throughout the book, we have tried to characterize the notion of learnability using
different approaches. At first we have shown that the uniform convergence property of a hypothesis class guarantees successful learning. Later on we introduced
the notion of stability and have shown that stable algorithms are guaranteed to be
good learners. Yet there are other properties which may be sufficient for learning,
and in this chapter and its sequel we will introduce two approaches to this issue:
compression bounds and the PAC-Bayes approach.
In this chapter we study compression bounds. Roughly speaking, we shall see
that if a learning algorithm can express the output hypothesis using a small subset
of the training set, then the error of the hypothesis on the rest of the examples
estimates its true error. In other words, an algorithm that can compress its output
is a good learner.

30.1 COMPRESSION BOUNDS

To motivate the results, let us first consider the following learning protocol. First,
we sample a sequence of k examples denoted T . On the basis of these examples, we
construct a hypothesis denoted h T . Now we would like to estimate the performance
of h T so we sample a fresh sequence of m k examples, denoted V , and calculate the
error of h T on V . Since V and T are independent, we immediately get the following
from Bernsteins inequality (see Lemma B.10).
Lemma 30.1. Assume that the range of the loss function is [0, 1]. Then,

*
4
log
(1/)
2L
(h
)
log
(1/)
V T
.
P L D (h T ) L V (h T )
+
|V |
|V |
To derive this bound, all we needed was independence between T and V .
Therefore, we can redefine the protocol as follows. First, we agree on a sequence
of k indices I = (i 1 , . . . ,i k ) [m]k . Then, we sample a sequence of m examples
S = (z 1 , . . . , z m ). Now, define T = SI = (z i1 , . . . , z ik ) and define V to be the rest of
359

360

Compression Bounds

the examples in S. Note that this protocol is equivalent to the protocol we defined
before hence Lemma 30.1 still holds.
Applying a union bound over the choice of the sequence of indices we obtain
the following theorem.
Theorem 30.2. Let k be an integer and let B : Z k H be a mapping from sequences
of k examples to the hypothesis class. Let m 2k be a training set size and let A :
Z m H be a learning rule that receives a training sequence S of size m and returns
a hypothesis such that A(S) = B(z i1 , . . . , z ik ) for some (i 1 , . . . ,i k ) [m]k . Let V = {z j :
j
/ (i 1 , . . . ,i k )} be the set of examples which were not selected for defining A(S). Then,
with probability of at least 1 over the choice of S we have

4k log (m/) 8k log(m/)
+
.
L D (A(S)) L V (A(S)) + L V (A(S))
m
m
Proof. For any I [m]k let h I = B(z i1 , . . . , z ik ). Let n = m k. Combining
Lemma 30.1 with the union bound we have



2L V (h I ) log (1/) 4 log (1/)
k
+
P I [m] s. t. L D (h I ) L V (h I )
n
n




2L V (h I ) log (1/) 4 log (1/)
+
P L D (h I ) L V (h I )

n
n
k
I [m]

m k .
Denote  = m k . Using the assumption k m/2, which implies that n = m k m/2,
the above implies that with probability of at least 1  we have that

4k log (m/  ) 8k log (m/  )
+
,
L D (A(S)) L V (A(S)) + L V (A(S))
m
m
which concludes our proof.
As a direct corollary we obtain:
Corollary 30.3. Assuming the conditions of Theorem 30.2, and further assuming that
L V (A(S)) = 0, then, with probability of at least 1 over the choice of S we have
L D (A(S))

8k log (m/)
.
m

These results motivate the following definition:

Definition 30.4. (Compression Scheme) Let H be a hypothesis class of functions
from X to Y and let k be an integer. We say that H has a compression scheme of
size k if the following holds:
For all m there exists A : Z m [m]k and B : Z k H such that for all h H, if we
feed any training set of the form (x 1 , h(x 1 )), . . . , (x m , h(x m )) into A and then feed
(x i1 , h(x i1 )), . . . , (x ik , h(x ik )) into B, where (i 1 , . . . ,i k ) is the output of A, then the
output of B, denoted h  , satisfies L S (h  ) = 0.

30.2 Examples

It is possible to generalize the definition for unrealizable sequences as follows.

Definition 30.5. (Compression Scheme for Unrealizable Sequences) Let H be a
hypothesis class of functions from X to Y and let k be an integer. We say that H
has a compression scheme of size k if the following holds:
For all m there exists A : Z m [m]k and B : Z k H such that for all h H,
if we feed any training set of the form (x 1 , y1 ), . . . , (x m , ym ) into A and then feed
(x i1 , yi1 ), . . . , (x ik , yik ) into B, where (i 1 , . . . ,i k ) is the output of A, then the output of
B, denoted h  , satisfies L S (h  ) L S (h).
The following lemma shows that the existence of a compression scheme for
the realizable case also implies the existence of a compression scheme for the
unrealizable case.
Lemma 30.6. Let H be a hypothesis class for binary classification, and assume it
has a compression scheme of size k in the realizable case. Then, it has a compression
scheme of size k for the unrealizable case as well.
Proof. Consider the following scheme: First, find an ERM hypothesis and denote
it by h. Then, discard all the examples on which h errs. Now, apply the realizable
compression scheme on the examples that have not been removed. The output of
the realizable compression scheme, denoted h  , must be correct on the examples that
have not been removed. Since h errs on the removed examples it follows that the
error of h  cannot be larger than the error of h; hence h  is also an ERM hypothesis.

30.2 EXAMPLES
In the examples that follows, we present compression schemes for several hypothesis classes for binary classification. In light of Lemma 30.6 we focus on the realizable
case. Therefore, to show that a certain hypothesis class has a compression scheme,
it is necessary to show that there exist A, B, and k for which L S (h  ) = 0.

30.2.1 Axis Aligned Rectangles

Note that this is an uncountable infinite class. We show that there is a simple
compression scheme. Consider the algorithm A that works as follows: For each
dimension, choose the two positive examples with extremal values at this dimension.
Define B to be the function that returns the minimal enclosing rectangle. Then, for
k = 2d, we have that in the realizable case, L S (B(A(S))) = 0.

30.2.2 Halfspaces
Let X = Rd and consider the class of homogenous halfspaces, {x  sign(w, x) :
w Rd }.
A Compression Scheme:
W.l.o.g. assume all labels are positive (otherwise, replace xi by yi xi ). The compression scheme we propose is as follows. First, A finds the vector w which is in the

361

362

Compression Bounds

convex hull of {x1 , . . . , xm } and has minimal norm. Then, it represents it as a convex
combination of d points in the sample (it will be shown later that this is always possible). The output of A are these d points. The algorithm B receives these d points
and set w to be the point in their convex hull of minimal norm.
Next we prove that this indeed is a compression sceme. Since the data is linearly
separable, the convex hull of {x1 , . . . , xm } does not contain the origin. Consider the
point w in this convex hull closest to the origin. (This is a unique point which is the
Euclidean projection of the origin onto this convex hull.) We claim that w separates
the data.1 To see this, assume by contradiction that w, xi  0 for some i . Take
2

w = (1 )w + xi for = x w
2 +w2 (0, 1). Then w is also in the convex hull and
i

w 2 = (1 )2 w2 + 2 xi 2 + 2(1 )w, xi 

(1 )2 w2 + 2 xi 2
=

xi 4 w2 + xi 2 w4

(w2 + xi 2 )2

xi 2 w2
w2 + xi 2

= w2

1
w2 /xi 2 + 1

< w2 ,
which leads to a contradiction.
We have thus shown that w is also an ERM. Finally, since w is in the convex hull
of the examples, we can apply Caratheodorys theorem to obtain that w is also in the
convex hull of a subset of d + 1 points of the polygon. Furthermore, the minimality
of w implies that w must be on a face of the polygon and this implies it can be
represented as a convex combination of d points.
It remains to show that w is also the projection onto the polygon defined by the
d points. But this must be true: On one hand, the smaller polygon is a subset of the
larger one; hence the projection onto the smaller cannot be smaller in norm. On the
other hand, w itself is a valid solution. The uniqueness of projection concludes our
proof.

30.2.3 Separating Polynomials

Let X = Rd and consider the class x  sign( p(x)) where p is a degree r polynomial.
Note that p(x) can be rewritten as w, (x) where the elements of (x) are
all the monomials of x up to degree r . Therefore, the problem of constructing a
compression scheme for p(x) reduces to the problem of constructing a compression

scheme for halfspaces in Rd where d  = O(d r ).

It can be shown that w is the direction of the max-margin solution.

30.3 Bibliographic Remarks

30.2.4 Separation with Margin

Suppose that a training set is separated with margin . The Perceptron algorithm
guarantees to make at most 1/ 2 updates before converging to a solution that makes
no mistakes on the entire training set. Hence, we have a compression scheme of size
k 1/ 2 .

30.3 BIBLIOGRAPHIC REMARKS

Compression schemes and their relation to learning were introduced by Littlestone
and Warmuth (1986). As we have shown, if a class has a compression scheme then
it is learnable. For binary classification problems, it follows from the fundamental
theorem of learning that the class has a finite VC dimension. The other direction,
namely, whether every hypothesis class of finite VC dimension has a compression
scheme of finite size, is an open problem posed by Manfred Warmuth and is still
open (see also (Floyd 1989, Floyd & Warmuth 1995, Ben-David & Litman 1998,
Livni & Simon 2013).

363

31
PAC-Bayes

The Minimum Description Length (MDL) and Occams razor principles allow a
potentially very large hypothesis class but define a hierarchy over hypotheses and
prefer to choose hypotheses that appear higher in the hierarchy. In this chapter we
describe the PAC-Bayesian approach that further generalizes this idea. In the PACBayesian approach, one expresses the prior knowledge by defining prior distribution
over the hypothesis class.

31.1 PAC-BAYES BOUNDS

As in the MDL paradigm, we define a hierarchy over hypotheses in our class H.
Now, the hierarchy takes the form of a prior distribution over H. That is, we assign
a probability (or density if H is continuous) P(h) 0 for each h H and refer to
P(h) as the prior score of h. Following the Bayesian reasoning approach, the output
of the learning algorithm is not necessarily a single hypothesis. Instead, the learning
process defines a posterior probability over H, which we denote by Q. In the context
of a supervised learning problem, where H contains functions from X to Y, one can
think of Q as defining a randomized prediction rule as follows. Whenever we get a
new instance x, we randomly pick a hypothesis h H according to Q and predict
h(x). We define the loss of Q on an example z to be
def

(Q, z) = E [
(h, z)].

hQ

By the linearity of expectation, the generalization loss and training loss of Q can be
written as
def
def
L D (Q) = E [L D (h)] and L S (Q) = E [L S (h)].
hQ

hQ

The following theorem tells us that the difference between the generalization
loss and the empirical loss of a posterior Q is bounded by an expression that depends
on the Kullback-Leibler divergence between Q and the prior distribution P. The
Kullback-Leibler is a natural measure of the distance between two distributions.
The theorem suggests that if we would like to minimize the generalization loss of Q,
364

31.1 PAC-Bayes Bounds

we should jointly minimize both the empirical loss of Q and the Kullback-Leibler
distance between Q and the prior distribution. We will later show how in some cases
this idea leads to the regularized risk minimization principle.
Theorem 31.1. Let D be an arbitrary distribution over an example domain Z . Let H
be a hypothesis class and let
: H Z [0, 1] be a loss function. Let P be a prior
distribution over H and let (0, 1). Then, with probability of at least 1 over
the choice of an i.i.d. training set S = {z 1 , . . . , z m } sampled according to D, for all
distributions Q over H (even such that depend on S), we have
*

D(Q||P) + ln m/
,
2(m 1)

L D (Q) L S (Q) +
where
def

D(Q||P) = E [ ln (Q(h)/P(h))]
hQ

is the Kullback-Leibler divergence.

Proof. For any function f (S), using Markovs inequality:
P [ f (S) ] = P [e f (S) e ]
S

E S [e f (S) ]
.
e

(31.1)

Let (h) = L D (h) L S (h). We will apply Equation (31.1) with the function


2
f (S) = sup 2(m 1) E ((h)) D(Q||P) .
Q

hQ

We now turn to bound E S [e f (S)]. The main trick is to upper bound f (S) by using an
expression that does not depend on Q but rather depends on the prior probability
P. To do so, fix some S and note that from the definition of D(Q||P) we get that for
all Q,
2

2(m 1) E ((h))2 D(Q||P) = E [ ln (e2(m1)(h) P(h)/Q(h))]

hQ

hQ

ln E [e2(m1)(h) P(h)/Q(h)]
hQ

= ln E [e2(m1)(h) ],
hP

(31.2)

where the inequality follows from Jensens inequality and the concavity of the log
function. Therefore,
2

E [e f (S) ] E E [e2(m1)(h) ].
S

S hP

(31.3)

The advantage of the expression on the right-hand side stems from the fact that
we can switch the order of expectations (because P is a prior that does not depend

365

366

PAC-Bayes

on S), which yields

2

E [e f ( S)] E E [e2(m1)(h) ].
S

(31.4)

h P S

Next, we claim that for all h we have E S [e2(m1)(h) ] m. To do so, recall that
Hoeffdings inequality tells us that
P [(h) ] e2m .
2

This implies that E S [e2(m1)(h) ] m (see Exercise 31.1). Combining this with
Equation (31.4) and plugging into Equation (31.1) we get
P [ f (S) ]
S

m
.
e

(31.5)

Denote the right-hand side of the above , thus = ln (m/), and we therefore obtain
that with probability of at least 1 we have that for all Q
2(m 1) E ((h))2 D(Q||P) = ln (m/).
hQ

Rearranging the inequality and using Jensens inequality again (the function x 2 is
convex) we conclude that

2
ln (m/) + D(Q||P)
.
(31.6)
E (h) E ((h))2
hQ
hQ
2(m 1)

Remark 31.1 (Regularization). The PAC-Bayes bound leads to the following

learning rule:
Given a prior P, return a posterior Q that minimizes the function
*
D(Q||P) + ln m/
.
L S (Q) +
2(m 1)

(31.7)

This rule is similar to the regularized risk minimization principle. That is, we jointly
minimize the empirical loss of Q on the sample and the Kullback-Leibler distance
between Q and P.

31.2 BIBLIOGRAPHIC REMARKS

PAC-Bayes bounds were first introduced by McAllester (1998). See also
(McAllester 1999, McAllester 2003, Seeger 2003, Langford & Shawe-Taylor 2003,
Langford 2006).

31.3 EXERCISES
31.1 Let X be a random variable that satisfies P [X ] e2m . Prove that
2
E [e2(m1)X ] m.
2

31.3 Exercises

31.2
Suppose that H is a finite hypothesis class, set the prior to be uniform over H,
and set the posterior to be Q(h S ) = 1 for some h S and Q(h) = 0 for all other
h H. Show that
*
ln (|H|) + ln(m/)
L D (h S ) L S (h) +
.
2(m 1)
Compare to the bounds we derived using uniform convergence.

Derive a bound similar to the Occam bound given in Chapter 7 using the PACBayes bound

367

Appendix A
Technical Lemmas

Lemma A.1. Let a > 0. Then: x 2a log (a) x a log (x). It follows that a
necessary condition for the inequality x < a log (x) to hold is that x < 2a log (a).

Proof. First note that for a (0, e ] the inequality x a log (x) holds uncondition
ally and therefore the claim is trivial. From now on, assume that a > e. Consider
the function f (x) = x a log (x). The derivative is f  (x) = 1 a/x. Thus, for x > a
the derivative is positive and the function increases. In addition,
f (2a log (a)) = 2a log (a) a log (2a log (a))
= 2a log (a) a log (a) a log (2 log (a))
= a log (a) a log (2 log (a)).
Since a 2 log (a) > 0 for all a > 0, the proof follows.
Lemma A.2.

Let a 1 and b > 0. Then: x 4a log(2a) + 2b x a log (x) + b.

Proof. It suffices to prove that x 4a log (2a) + 2b implies that both x 2a log (x)
and x 2b. Since we assume a 1 we clearly have that x 2b. In addition, since
b > 0 we have that x 4a log (2a) which using Lemma A.1 implies that x 2a log (x).
This concludes our proof.
Lemma A.3. Let X be a random variable and x  R be a scalar and assume that
2 2
there exists a > 0 such that for all t 0 we have P [|X x  | > t] 2et /a . Then,
E [|X x  |] 4 a.
Proof. For all i = 0, 1, 2, . . . denote ti = a i . Since ti is monotonically increasing we



have that E [|X x  |] is at most
i=1 ti P [|X x | > ti1 ]. Combining this with the



(i1)2 . The proof
assumption in the lemma we get that E [|X x |] 2 a
i=1 i e
now follows from the inequalities


i=1

i e(i1)
2

5

i=1

i e(i1) +
2

xe(x1) d x < 1. 8 + 107 < 2 .

2

369

370

Technical Lemmas

Lemma A.4. Let X be a random variable and x  R be a scalar and assume that
2 2
there exists a > 0 and b e such
that for all t 0 we have P [|X x  | > t] 2b et /a .

Then, E [|X x  |] a(2 + log (b)).
Proof. For all i = 0, 1, 2, . . . denote ti = a (i +
increasing we have that


log (b)). Since ti is monotonically



ti P [|X x  | > ti1 ].
E [|X x |] a log (b) +


i=1

Using the assumption in the lemma we have

ti P [|X x  | > ti1 ] 2 a b

i=1

(i +


2
log (b))e(i1+ log (b))

i=1

2a b

1+

F
= 2a b
F

xe(x1) d x
2

log (b)

(y + 1)ey d y
2

log (b)

yey d y
2

4a b

log (b)

2
= 2 a b ey

log (b)

= 2 a b/b = 2 a.
Combining the preceding inequalities we conclude our proof.
Lemma A.5. Let m, d be two positive integers such that d m 2. Then,
d  

m

k=0

 e m d
d

Proof. We prove the claim by induction. For d = 1 the left-hand side equals 1 + m
while the right-hand side equals em; hence the claim is true. Assume that the claim
holds for d and let us prove it for d + 1. By the induction assumption we have
d+1  

m
k=0

 e m d
d


+



m
d +1



d d m(m 1)(m 2) (m d)
1+
=
d
em
(d + 1)d!


 d
 em d
d
(m d)

1+
.
d
e
(d + 1)d!
 e m d

Technical Lemmas

Using Stirlings approximation we further have that



 d
 e m d
d
(m d)

1+
d
e
(d + 1) 2d(d/e)d

 e m d 
m d
=
1+
d
2d(d + 1)
 e m d d + 1 + (m d)/2d
=

d
d +1
 e m d d + 1 + (m d)/2

d
d +1
 e m d d/2 + 1 + m/2

=
d
d +1
 e m d m

d
d +1
where in the last inequality we used the assumption that d m 2. On the other
hand,



d
d
e m d+1  e m d em

d +1
d
d +1
d +1
 e m d em
1

d
d + 1 (1 + 1/d)d
 e m d em 1

d
d +1 e
 e m d
m
=
,

d
d +1
which proves our inductive argument.
Lemma A.6.

For all a R we have

ea + ea
2
ea /2 .
2

Proof. Observe that

ea =


an
n=0

n!

Therefore,

ea + ea  a 2n
=
,
2
(2n)!
n=0

and
ea

2 /2


a 2n
.
2n n!
n=0

Observing that (2n)! 2 n! for every n 0 we conclude our proof.

n

371

Appendix B
Measure Concentration

Let Z 1 , . . . , Z m be an i.i.d. sequence of random variables and let be their mean.

The strong law of large numbers states that when m tends to infinity, the empirical


average, m1 m
i=1 Z i , converges to the expected value , with probability 1. Measure
concentration inequalities quantify the deviation of the empirical average from the
expectation when m is finite.

B.1 MARKOVS INEQUALITY

We start with an inequality which is called Markovs inequality. Let Z be a
nonnegative random variable. The expectation of Z can be written as follows:
F
P [Z x]d x.
(B.1)
E [Z ] =
x=0

Since P [Z x] is monotonically nonincreasing we obtain

F a
F a
P [Z x]d x
P [Z a]d x = a P [Z a].
a 0, E [Z ]
x=0

(B.2)

x=0

Rearranging the inequality yields Markovs inequality:

a 0, P [Z a]

E [Z ]
.
a

(B.3)

For random variables that take value in [0, 1], we can derive from Markovs
inequality the following.
Lemma B.1. Let Z be a random variable that takes values in [0, 1]. Assume that
E [Z ] = . Then, for any a (0, 1),
P [Z > 1 a]

(1 a)
.
a

This also implies that for every a (0, 1),

P [Z > a]
372

a
a.
1a

B.3 Chernoffs Bounds

Proof. Let Y = 1 Z . Then Y is a nonnegative random variable with

E [Y ] = 1 E[Z ] = 1 . Applying Markovs inequality on Y we obtain
P [Z 1 a] = P [1 Z a] = P [Y a]

E [Y ] 1
=
.
a
a

Therefore,
P [Z > 1 a] 1

1 a+1
=
.
a
a

B.2 CHEBYSHEVS INEQUALITY

Applying Markovs inequality on the random variable (Z E [Z ])2 we obtain
Chebyshevs inequality:
a > 0, P [|Z E [Z ]| a] = P [(Z E [Z ])2 a 2 ]

Var[Z ]
,
a2

(B.4)

where Var[Z ] = E [(Z E [Z ])2 ] is the variance of Z .

Consider the random variable m1 m
i=1 Z i . Since Z 1 , . . . , Z m are i.i.d. it is easy to
verify that


m
1
Var[Z 1 ]
.
Var
Zi =
m
m
i=1

Applying Chebyshevs inequality, we obtain the following:

Lemma B.2. Let Z 1 , . . . , Z m be a sequence of i.i.d. random variables and assume
that E [Z 1 ] = and Var[Z 1 ] 1. Then, for any (0, 1), with probability of at least
1 we have


m
1 

1


.
Z i

m

m
i=1

Proof. Applying Chebyshevs inequality we obtain that for all a > 0



m
1 

1
Var[Z 1 ]


P
Z i > a

.
2
m

ma
m a2
i=1

The proof follows by denoting the right-hand side and solving for a.
The deviation between the empirical average and the mean given previously
decreases polynomially with m. It is possible to obtain a significantly faster decrease.
In the sections that follow we derive bounds that decrease exponentially fast.

B.3 CHERNOFFS BOUNDS

Let Z 1 , . . . , Z m be independent

Bernoulli variables where

m for every i , P [Z i = 1] = pi
m
and P [Z i = 0] = 1 pi . Let p = i=1 pi and let Z = i=1 Z i . Using the monotonicity

373

374

Measure Concentration

of the exponent function and Markovs inequality, we have that for every t > 0
P [Z > (1 + ) p] = P [et Z > et(1+) p ]

E [et Z ]
.
e(1+)t p

(B.5)

Next,
E [e ] = E [e
tZ

i Zi ] = E


e

t Zi

E [et Z i ]



pi et + (1 pi )e0

by independence




=
1 + pi (et 1)
i

=e

e pi (e 1)
t

i


i

using 1 + x e x

pi (et 1)

= e(e 1) p .
t

Combining the equation with Equation (B.5) and choosing t = log (1 + ) we obtain
Lemma B.3. Let Z 1 , . . . , Z m be independent Bernoulli variables where for every i ,



m
P [Z i = 1] = pi and P [Z i = 0] = 1 pi . Let p = m
i=1 pi and let Z =
i=1 Z i . Then,
for any > 0,
P[Z > (1 + ) p] eh() p ,
where
h() = (1 + ) log (1 + ) .
Using the inequality h(a) a 2 /(2 + 2a/3) we obtain
Lemma B.4. Using the notation of Lemma B.3 we also have
P [Z > (1 + ) p] e

p 2+2/3

For the other direction, we apply similar calculations:

P [Z < (1 ) p] = P [ Z > (1 ) p] = P [et Z > et(1) p ]

E [et Z ]
,
e(1)t p

(B.6)

B.4 Hoeffdings Inequality

and
E [et Z ] = E [et
=

Zi

]=E


et Z i

E [e

t Z i

by independence



1 + pi (et 1)
i

e pi (e

t 1)

using 1 + x e x

= e(e

t 1) p

Setting t = log (1 ) yields

e p

P [Z < (1 ) p]

e(1) log (1) p

= e ph() .

It is easy to verify that h( ) h() and hence

Lemma B.5. Using the notation of Lemma B.3 we also have
P [Z < (1 ) p] e ph() e ph() e

p 2+2/3

B.4 HOEFFDINGS INEQUALITY

Lemma B.6 (Hoeffdings

Inequality). Let Z 1 , . . . , Z m be a sequence of i.i.d. random

variables and let Z = m1 m

i=1 Z i . Assume that E [ Z ] = and P [a Z i b] = 1 for
every i . Then, for any > 0


 m





1
Z i > 2 exp 2 m 2 /(b a)2 .
P m


i=1

Proof. Denote X i = Z i E [Z i ] and X = m1 i X i . Using the monotonicity of the
exponent function and Markovs inequality, we have that for every > 0 and > 0,

P [ X ] = P [e X e ] e E [e X ].
Using the independence assumption we also have




X
X i /m
E [e ] = E
e
E [eX i /m ].
=
i

By Hoeffdings lemma (Lemma B.7 later), for every i we have

E [eX i /m ] e
Therefore,
P [ X ] e


i

2 (ba)2
8m 2

2 (ba)2
8m 2

= e +

2 (ba)2
8m

375

376

Measure Concentration

Setting = 4m /(b a)2 we obtain

P[ X ] e

2
2m 2
(ba)

Applying the same arguments on the variable X we obtain that P [ X ]

e

2
2m 2
(ba)

. The theorem follows by applying the union bound on the two cases.

Lemma B.7 (Hoeffdings Lemma). Let X be a random variable that takes values in
the interval [a, b] and such that E [X] = 0. Then, for every > 0,
E [eX ] e

2 (ba)2
8

Proof. Since f (x) = ex is a convex function, we have that for every (0, 1), and
x [a, b],
f (x) f (a) + (1 ) f (b).
Setting =

bx
ba

[0, 1] yields
ex

b x a x a b
e +
e .
ba
ba

Taking the expectation, we obtain that

E [eX ]

b E [X] a E [x] a b
b a
a b
e +
e =
e
e ,
ba
ba
ba
ba

a
, and
where we used the fact that E [X] = 0. Denote h = (b a), p = ba
h
L(h) = hp + log (1 p + pe ). Then, the expression on the right-hand side of
the equation can be rewritten as e L(h) . Therefore, to conclude our proof it suf2
fices to show that L(h) h8 . This follows from Taylors theorem using the facts
L(0) = L  (0) = 0 and L  (h) 1/4 for all h.

B.5 BENNETS AND BERNSTEINS INEQUALITIES

Bennets and Bernseins inequalities are similar to Chernoffs bounds, but they
hold for any sequence of independent random variables. We state the inequalities
without proof, which can be found, for example, in Cesa-Bianchi and Lugosi (2006).
Lemma B.8 (Bennets Inequality). Let Z 1 , . . . , Z m be independent random variables
with zero mean, and assume that Z i 1 with probability 1. Let
1
E [Z i2 ].
m
m

i=1

Then for all > 0,

P

 m



Zi > e

m 2 h

m 2

i=1

where
h(a) = (1 + a) log (1 + a) a.

B.5 Bennets and Bernsteins Inequalities

By using the inequality h(a) a 2 /(2 + 2a/3) it is possible to derive the following:
Lemma B.9 (Bernsteins Inequality). Let Z 1 , . . . , Z m be i.i.d. random variables with
a zero mean. If for all i, P (|Z i | < M) = 1, then for all t > 0:


m


t 2 /2
Z i > t exp

P
E Z 2j + Mt/3
i=1


.

B.5.1 Application
Bernsteins inequality can be used to interpolate between the rate 1/ we derived
for PAC learning in the realizable case (in Chapter 2) and the rate 1/ 2 we derived
for the unrealizable case (in Chapter 4).
Lemma B.10. Let
: H Z [0, 1] be a loss function. Let D be an arbitrary
distribution over Z . Fix some h. Then, for any (0, 1) we have

2L D (h) log (1/) 2 log (1/)
P
+

L S (h) L D (h) +
SD m
3m
m



2L S (h) log (1/) 4 log (1/)
+
P
L D (h) L S (h) +

SD m
m
m


1.

2.

Proof. Define random variables 1 , . . . , m s.t. i =
(h, z i ) L D (h). Note that

E [i ] = 0 and that
E [i2 ] = E [
(h, z i )2 ] 2L D (h) E [
(h, z i )] + L D (h)2
= E [
(h, z i )2 ] L D (h)2
E [
(h, z i )2 ]
E [
(h, z i )] = L D (h),
where in the last inequality we used the fact that
(h, z i ) [0, 1] and thus
(h, z i )2

(h, z i ). Applying Bernseins inequality over the i s yields
 m



t 2 /2
i > t exp
2
P
E j + t/3
i=1


t 2 /2
def
exp
= .
m L D (h) + t/3

377

378

Measure Concentration

Solving for t yields

t 2 /2
= log (1/)
m L D (h) + t/3
log (1/)
t log (1/) m L D (h) = 0
3
*
log (1/)
log2 (1/)
+
t=
+ 2 log (1/) m L D (h)
3
32
log (1/) 
+ 2 log (1/) m L D (h)
2
3

t 2 /2

= L S (h) L D (h), it follows that with probability of at least 1 ,


2 log (1/) L D (h)
log (1/)
+
,
L S (h) L D (h) 2
3m
m
which proves the first inequality. The second part of the lemma follows in a
similar way.

Since

1
m

i i

B.6 SLUDS INEQUALITY

Let X be a (m, p) binomial variable. That is, X = m
i=1 Z i , where each Z i is 1 with
probability p and 0 with probability 1 p. Assume that p = (1 )/2. Sluds inequality (Slud 1977) tells us that P [X m/2] is lower bounded
by the probability that

a normal variable will be greater than or equal to m 2 /(1 2 ). The following
lemma follows by standard tail bounds for the normal distribution.
Lemma B.11. Let X be a (m, p) binomial variable and assume that p = (1 )/2.
Then,



1
2
2
P [X m/2]
1 1 exp( m /(1 )) .
2

B.7 CONCENTRATION OF 2 VARIABLES

Let X 1 , . . . , X k be k independent normally distributed random variables. That is,
for all i , X i N(0, 1). The distribution of the random variable X i2 is called 2 (chi
square) and the distribution of the random variable Z = X 12 + + X k2 is called k2 (chi
square with k degrees of freedom). Clearly, E [X i2 ] = 1 and E [Z ] = k. The following
lemma states that X k2 is concentrated around its mean.
Lemma B.12. Let Z k2 . Then, for all > 0 we have
P [Z (1 )k] e

2 k/6

and for all (0, 3) we have

P [Z (1 + )k] e

2 k/6

Finally, for all (0, 3),

P [(1 )k Z (1 + )k] 1 2e

2 k/6

B.7 Concentration of 2 Variables

Proof. Let us write Z =

k

2
i=1 X i

where X i N(0, 1). To prove both bounds we

2

use Chernoffs bounding method. For the first inequality, we first bound E [eX 1 ],
2
where > 0 will be specified later. Since ea 1 a + a2 for all a 0 we have that
E [eX 1 ] 1 E [X 12 ] +
2

2
E [X 14 ].
2

Using the well known equalities, E [X 12 ] = 1 and E [X 14 ] = 3, and the fact that 1 a
ea we obtain that
3 2
2
E [eX 1 ] 1 + 32 2 e+ 2 .
Now, applying Chernoffs bounding method we get that


P [ Z (1 )k] = P eZ e(1 )k


e(1 )k E eZ


k
2
= e(1 )k E eX 1
3 2
k

e(1 )k ek+ 2
3

= e k+ 2 k .
Choose = /3 we obtain the first inequality stated in the lemma.
For the second inequality, we use a known closed form expression for the
moment generating function of a k2 distributed random variable:

2
= (1 2)k/2 .
< 12 , E eZ
(B.7)
On the basis of the equation and using Chernoffs bounding method we have


P [Z (1 + )k)] = P eZ e(1+ )k
 
e(1+ )k E eZ
= e(1+ )k (1 2)k/2
e(1+ )k ek = e k ,
where the last inequality occurs because (1 a) ea . Setting = /6 (which is in
(0, 1/2) by our assumption) we obtain the second inequality stated in the lemma.
Finally, the last inequality follows from the first two inequalities and the union
bound.

379

Appendix C
Linear Algebra

C.1 BASIC DEFINITIONS

In this chapter we only deal with linear algebra over finite dimensional Euclidean
spaces. We refer to vectors as column vectors.
Given two d dimensional vectors u, v Rd , their inner product is
u, v =

d


u i vi .

i=1

The Euclidean norm (a.k.a. the
2 norm) is u = u, u. We also use the
1 norm,

d
u1 = i=1 |u i | and the
norm u = maxi |u i |.
A subspace of Rd is a subset of Rd which is closed under addition and scalar
multiplication. The span of a set of vectors u1 , . . . , uk is the subspace containing all
vectors of the form
k


i ui

i=1

where for all i , i R.

A set of vectors U = {u1 , . . . , uk } is independent if for every i , ui is not in the span
of u1 , . . . , ui1 , ui+1 , . . . , uk . We say that U spans a subspace V if V is the span of the
vectors in U . We say that U is a basis of V if it is both independent and spans V. The
dimension of V is the size of a basis of V (and it can be verified that all bases of V
have the same size). We say that U is an orthogonal set if for all i = j , ui , u j  = 0.
We say that U is an orthonormal set if it is orthogonal and if for every i , ui  = 1.
Given a matrix A Rn,d , the range of A is the span of its columns and the null
space of A is the subspace of all vectors that satisfy Au = 0. The rank of A is the
dimension of its range.
The transpose of a matrix A, denoted A
, is the matrix whose (i , j ) entry equals
the ( j ,i ) entry of A. We say that A is symmetric if A = A
.

380

C.4 Singular Value Decomposition (SVD)

C.2 EIGENVALUES AND EIGENVECTORS

Let A Rd,d be a matrix. A nonzero vector u is an eigenvector of A with a
corresponding eigenvalue if
Au = u.
Theorem C.1 (Spectral Decomposition). If A Rd,d is a symmetric matrix of rank
k, then there exists an orthonormal basis of Rd , u1 , . . . , ud , such that each ui is an


eigenvector of A. Furthermore, A can be written as A = di=1 i ui u

i , where each
i is the eigenvalue corresponding to the eigenvector ui . This can be written equivalently as A = U DU
, where the columns of U are the vectors u1 , . . . , ud , and D is
a diagonal matrix with Di,i = i and for i = j , Di, j = 0. Finally, the number of i
which are nonzero is the rank of the matrix, the eigenvectors which correspond to the
nonzero eigenvalues span the range of A, and the eigenvectors which correspond to
zero eigenvalues span the null space of A.

C.3 POSITIVE DEFINITE MATRICES

A symmetric matrix A Rd,d is positive definite if all its eigenvalues are positive. A
is positive semidefinite if all its eigenvalues are nonnegative.
Theorem C.2. Let A Rd,d be a symmetric matrix. Then, the following are equivalent
definitions of positive semidefiniteness of A:

All the eigenvalues of A are nonnegative.

For every vector u, u, Au 0.

There exists a matrix B such that A = B B
.

C.4 SINGULAR VALUE DECOMPOSITION (SVD)

Let A Rm,n be a matrix of rank r . When m = n, the eigenvalue decomposition given
in Theorem C.1 cannot be applied. We will describe another decomposition of A,
which is called Singular Value Decomposition, or SVD for short.
Unit vectors v Rn and u Rm are called right and left singular vectors of A with
corresponding singular value > 0 if
Av = u and A
u = v.
We first show that if we can find r orthonormal singular vectors with positive
singular values, then we can decompose A = UDV
, with the columns of U and V
containing the left and right singular vectors, and D being a diagonal r r matrix
with the singular values on its diagonal.
Lemma C.3. Let A Rm,n be a matrix of rank r . Assume that v1 , . . . , vr is an
orthonormal set of right singular vectors of A, u1 , . . . , ur is an orthonormal set of corresponding left singular vectors of A, and 1 , . . . , r are the corresponding singular

381

382

Linear Algebra

values. Then,
A=

r


i ui v

i .

i=1

It follows that if U is a matrix whose columns are the ui s, V is a matrix whose columns
are the vi s, and D is a diagonal matrix with Di,i = i , then
A = U DV
.
Proof. Any right singular vector of A must be in the range of A
(otherwise, the
singular value will have to be zero). Therefore, v1 , . . . , vr is an orthonormal basis
of the range of A. Let us complete it to an orthonormal basis of Rn by adding


the vectors vr+1 , . . . , vn . Define B = ri=1 i ui v

i . It suffices to prove that for all i ,

Avi = Bvi . Clearly, if i > r then Avi = 0 and Bvi = 0 as well. For i r
we have
r

Bvi =
j u j v
j vi = i ui = Avi ,
j =1

where the last equality follows from the definition.

The next lemma relates the singular values of A to the eigenvalues of A
A
and A A
.
Lemma C.4. v, u are right and left singular vectors of A with singular value iff
v is an eigenvector of A
A with corresponding eigenvalue 2 and u = 1 Av is an
eigenvector of A A
with corresponding eigenvalue 2 .
Proof. Suppose that is a singular value of A with v Rn being the corresponding
right singular vector. Then,
A
Av = A
u = 2 v.
Similarly,
A A
u = Av = 2 u.
For the other direction, if = 0 is an eigenvalue of A
A, with v being the corresponding
eigenvector, then > 0 because A
A is positive semidefinite. Let

= , u = 1 Av. Then,
Av
u = = Av,

and
A
u =

A Av = v = v.

Finally, we show that if A has rank r then it has r orthonormal singular vectors.

C.4 Singular Value Decomposition (SVD)

Lemma C.5. Let A Rm,n with rank r . Define the following vectors:
v1 = argmax Av
vRn :v=1

v2 = argmax Av
vRn :v=1
v,v1 =0

..
.
vr =

argmax
vRn :v=1
i<r, v,vi =0

Av

Then, v1 , . . . , vr is an orthonormal set of right singular vectors of A.

Proof. First note that since the rank of A is r , the range of A is a subspace of
dimension r , and therefore it is easy to verify that for all i = 1, . . . ,r , Avi  > 0.
Let W Rn,n be an orthonormal matrix obtained by the eigenvalue decomposition of A
A, namely, A
A = W DW
, with D being a diagonal matrix with
D1,1 D2,2 0. We will show that v1 , . . . , vr are eigenvectors of A
A that
correspond to nonzero eigenvalues, and, hence, using Lemma C.4 it follows that
these are also right singular vectors of A. The proof is by induction. For the basis of
the induction, note that any unit vector v can be written as v = W x, for x = W
v,
and note that x = 1. Therefore,
2

Av = AW x = W DW W x = W Dx = Dx =

n


2
Di,i
xi 2 .

i=1

Therefore,
max Av2 = max

v:v=1

x:x=1

n


2
Di,i
xi 2 .

i=1

The solution of the right-hand side is to set x = (1, 0, . . . , 0), which implies that v1 is
the first eigenvector of A
A. Since Av1  > 0 it follows that D1,1 > 0 as required.
For the induction step, assume that the claim holds for some 1 t r 1. Then,
any v which is orthogonal to v1 , . . . , vt can be written as v = W x with all the first t
elements of x being zero. It follows that
max

v:v=1,it,v
vi =0

Av2 = max

x:x=1

n


2
Di,i
xi 2 .

i=t+1

The solution of the right-hand side is the all zeros vector except x t+1 = 1. This
implies that vt+1 is the (t + 1)th column of W . Finally, since Avt+1  > 0 it follows
that Dt+1,t+1 > 0 as required. This concludes our proof.
Corollary C.6 (The SVD Theorem). Let A Rm,n with rank r . Then A = U DV

where D is an r r matrix with nonzero singular values of A and the columns of U , V

2 is
are orthonormal left and right singular vectors of A. Furthermore, for all i , Di,i
an eigenvalue of A
A, the i th column of V is the corresponding eigenvector of A
A
and the i th column of U is the corresponding eigenvector of A A
.

383

References

Abernethy, J., Bartlett, P. L., Rakhlin, A. & Tewari, A. (2008), Optimal strategies
and minimax lower bounds for online convex games, in Proceedings of the nineteenth
annual conference on computational learning theory.
Ackerman, M. & Ben-David, S. (2008), Measures of clustering quality: A working set
of axioms for clustering, in Proceedings of Neural Information Processing Systems
(NIPS), pp. 121128.
Agarwal, S. & Roth, D. (2005), Learnability of bipartite ranking functions, in
Proceedings of the 18th annual conference on learning theory, pp. 1631.
Agmon, S. (1954), The relaxation method for linear inequalities, Canadian Journal of
Mathematics 6(3), 382392.
Aizerman, M. A., Braverman, E. M. & Rozonoer, L. I. (1964), Theoretical foundations
of the potential function method in pattern recognition learning, Automation and
Remote Control 25, 821837.
Allwein, E. L., Schapire, R. & Singer, Y. (2000), Reducing multiclass to binary: A
unifying approach for margin classifiers, Journal of Machine Learning Research
1, 113141.
Alon, N., Ben-David, S., Cesa-Bianchi, N. & Haussler, D. (1997), Scale-sensitive
dimensions, uniform convergence, and learnability, Journal of the ACM 44(4),
615631.
Anthony, M. & Bartlet, P. (1999), Neural Network Learning: Theoretical Foundations,
Cambridge University Press.
Baraniuk, R., Davenport, M., DeVore, R. & Wakin, M. (2008), A simple proof of
the restricted isometry property for random matrices, Constructive Approximation
28(3), 253263.
Barber, D. (2012), Bayesian reasoning and machine learning, Cambridge University
Press.
Bartlett, P., Bousquet, O. & Mendelson, S. (2005), Local rademacher complexities,
Annals of Statistics 33(4), 14971537.
Bartlett, P. L. & Ben-David, S. (2002), Hardness results for neural network approximation problems, Theor. Comput. Sci. 284(1), 5366.
Bartlett, P. L., Long, P. M. & Williamson, R. C. (1994), Fat-shattering and the learnability of real-valued functions, in Proceedings of the seventh annual conference on
computational learning theory, (ACM), pp. 299310.

385

386

References

Bartlett, P. L. & Mendelson, S. (2001), Rademacher and Gaussian complexities: Risk

bounds and structural results, in 14th Annual Conference on Computational Learning
Theory (COLT) 2001, Vol. 2111, Springer, Berlin, pp. 224240.
Bartlett, P. L. & Mendelson, S. (2002), Rademacher and Gaussian complexities: Risk
bounds and structural results, Journal of Machine Learning Research 3, 463482.
Ben-David, S., Cesa-Bianchi, N., Haussler, D. & Long, P. (1995), Characterizations of
learnability for classes of {0, . . . , n}-valued functions, Journal of Computer and System
Sciences 50, 7486.
Ben-David, S., Eiron, N. & Long, P. (2003), On the difficulty of approximately
maximizing agreements, Journal of Computer and System Sciences 66(3), 496514.
Ben-David, S. & Litman, A. (1998), Combinatorial variability of vapnik-chervonenkis
classes with applications to sample compression schemes, Discrete Applied Mathematics 86(1), 325.
Ben-David, S., Pal, D., & Shalev-Shwartz, S. (2009), Agnostic online learning, in
Conference on Learning Theory (COLT).
Ben-David, S. & Simon, H. (2001), Efficient learning of linear perceptrons, Advances
in Neural Information Processing Systems, pp. 189195.
Bengio, Y. (2009), Learning deep architectures for AI, Foundations and Trends in
Machine Learning 2(1), 1127.
Bengio, Y. & LeCun, Y. (2007), Scaling learning algorithms towards AI, Large-Scale
Kernel Machines 34.
Bertsekas, D. (1999), Nonlinear programming, Athena Scientific.
Beygelzimer, A., Langford, J. & Ravikumar, P. (2007), Multiclass classification with
filter trees, Preprint, June .
Birkhoff, G. (1946), Three observations on linear algebra, Revi. Univ. Nac. Tucuman,
ser. A 5, 147151.
Bishop, C. M. (2006), Pattern recognition and machine learning, Vol. 1, Springer: New
York.
Blum, L., Shub, M. & Smale, S. (1989), On a theory of computation and complexity
over the real numbers: Np-completeness, recursive functions and universal machines,
Am. Math. Soc. 21(1), 146.
Blumer, A., Ehrenfeucht, A., Haussler, D. & Warmuth, M. K. (1987), Occams razor,
Information Processing Letters 24(6), 377380.
Blumer, A., Ehrenfeucht, A., Haussler, D. & Warmuth, M. K. (1989), Learnability
and the Vapnik-Chervonenkis dimension, Journal of the Association for Computing
Machinery 36(4), 929965.
Borwein, J. & Lewis, A. (2006), Convex analysis and nonlinear optimization, Springer.
Boser, B. E., Guyon, I. M. & Vapnik, V. N. (1992), A training algorithm for optimal
margin classifiers, in COLT, pp. 144152.
Bottou, L. & Bousquet, O. (2008), The tradeoffs of large scale learning, in NIPS,
pp. 161168.
Boucheron, S., Bousquet, O. & Lugosi, G. (2005), Theory of classification: A survey of
recent advances, ESAIM: Probability and Statistics 9, 323375.
Bousquet, O. (2002), Concentration Inequalities and Empirical Processes Theory
Applied to the Analysis of Learning Algorithms, PhD thesis, Ecole Polytechnique.
Bousquet, O. & Elisseeff, A. (2002), Stability and generalization, Journal of Machine
Learning Research 2, 499526.
Boyd, S. & Vandenberghe, L. (2004), Convex optimization, Cambridge University Press.
Breiman, L. (1996), Bias, variance, and arcing classifiers, Technical Report 460, Statistics
Department, University of California at Berkeley.
Breiman, L. (2001), Random forests, Machine Learning 45(1), 532.
Breiman, L., Friedman, J. H., Olshen, R. A. & Stone, C. J. (1984), Classification and
regression trees, Wadsworth & Brooks.

References

Cands, E. (2008), The restricted isometry property and its implications for compressed
sensing, Comptes Rendus Mathematique 346(9), 589592.
Candes, E. J. (2006), Compressive sampling, in Proc. of the int. congress of math.,
Madrid, Spain.
Candes, E. & Tao, T. (2005), Decoding by linear programming, IEEE Trans. on
Information Theory 51, 42034215.
Cesa-Bianchi, N. & Lugosi, G. (2006), Prediction, learning, and games, Cambridge
University Press.
Chang, H. S., Weiss, Y. & Freeman, W. T. (2009), Informative sensing, arXiv preprint
arXiv:0901.4275.
Chapelle, O., Le, Q. & Smola, A. (2007), Large margin optimization of ranking
measures, in NIPS workshop: Machine learning for Web search (Machine Learning).
Collins, M. (2000), Discriminative reranking for natural language parsing, in Machine
Learning.
Collins, M. (2002), Discriminative training methods for hidden Markov models: Theory
and experiments with perceptron algorithms, in Conference on Empirical Methods in
Natural Language Processing.
Collobert, R. & Weston, J. (2008), A unified architecture for natural language processing: deep neural networks with multitask learning, in International Conference on
Machine Learning (ICML).
Cortes, C. & Vapnik, V. (1995), Support-vector networks, Machine Learning
20(3), 273297.
Cover, T. (1965), Behavior of sequential predictors of binary sequences, Trans.
4th Prague conf. information theory statistical decision functions, random processes,
pp. 263272.
Cover, T. & Hart, P. (1967), Nearest neighbor pattern classification, Information
Theory, IEEE Transactions on 13(1), 2127.
Crammer, K. & Singer, Y. (2001), On the algorithmic implementation of multiclass kernel-based vector machines, Journal of Machine Learning Research 2,
265292.
Cristianini, N. & Shawe-Taylor, J. (2000), An introduction to support vector machines,
Cambridge University Press.
Daniely, A., Sabato, S., Ben-David, S. & Shalev-Shwartz, S. (2011), Multiclass
learnability and the erm principle, in COLT.
Daniely, A., Sabato, S. & Shwartz, S. S. (2012), Multiclass learning approaches: A
theoretical comparison with implications, in NIPS.
Davis, G., Mallat, S. & Avellaneda, M. (1997), Greedy adaptive approximation,
Journal of Constructive Approximation 13, 5798.
Devroye, L. & Gyrfi, L. (1985), Nonparametric density estimation: The L B1 S view,
Wiley.
Devroye, L., Gyrfi, L. & Lugosi, G. (1996), A probabilistic theory of pattern recognition,
Springer.
Dietterich, T. G. & Bakiri, G. (1995), Solving multiclass learning problems via errorcorrecting output codes, Journal of Artificial Intelligence Research 2, 263286.
Donoho, D. L. (2006), Compressed sensing, Information Theory, IEEE Transactions
52(4), 12891306.
Dudley, R., Gine, E. & Zinn, J. (1991), Uniform and universal glivenko-cantelli
classes, Journal of Theoretical Probability 4(3), 485510.
Dudley, R. M. (1987), Universal Donsker classes and metric entropy, Annals of
Probability 15(4), 13061326.
Fisher, R. A. (1922), On the mathematical foundations of theoretical statistics, Philosophical Transactions of the Royal Society of London. Series A, Containing Papers of
a Mathematical or Physical Character 222, 309368.

387

388

References

Floyd, S. (1989), Space-bounded learning and the Vapnik-Chervonenkis dimension,

in COLT, pp. 349364.
Floyd, S. & Warmuth, M. (1995), Sample compression, learnability, and the VapnikChervonenkis dimension, Machine Learning 21(3), 269304.
Frank, M. & Wolfe, P. (1956), An algorithm for quadratic programming, Naval Res.
Logist. Quart. 3, 95110.
Freund, Y. & Schapire, R. (1995), A decision-theoretic generalization of on-line learning and an application to boosting, in European Conference on Computational
Learning Theory (EuroCOLT), Springer-Verlag, pp. 2337.
Freund, Y. & Schapire, R. E. (1999), Large margin classification using the perceptron
algorithm, Machine Learning 37(3), 277296.
Garcia, J. & Koelling, R. (1996), Relation of cue to consequence in avoidance
learning, Foundations of animal behavior: classic papers with commentaries 4, 374.
Gentile, C. (2003), The robustness of the p-norm algorithms, Machine Learning
53(3), 265299.
Georghiades, A., Belhumeur, P. & Kriegman, D. (2001), From few to many: Illumination cone models for face recognition under variable lighting and pose, IEEE Trans.
Pattern Anal. Mach. Intelligence 23(6), 643660.
Gordon, G. (1999), Regret bounds for prediction problems, in Conference on
Learning Theory (COLT).
Gottlieb, L.-A., Kontorovich, L. & Krauthgamer, R. (2010), Efficient classification for
metric data, in 23rd conference on learning theory, pp. 433440.
Guyon, I. & Elisseeff, A. (2003), An introduction to variable and feature selection,
Journal of Machine Learning Research, Special Issue on Variable and Feature Selection
3, 11571182.
Hadamard, J. (1902), Sur les problmes aux drives partielles et leur signification
physique, Princeton University Bulletin 13, 4952.
Hastie, T., Tibshirani, R. & Friedman, J. (2001), The elements of statistical learning,
Springer.
Haussler, D. (1992), Decision theoretic generalizations of the PAC model for neural net and other learning applications, Information and Computation 100(1),
78150.
Haussler, D. & Long, P. M. (1995), A generalization of sauers lemma, Journal of
Combinatorial Theory, Series A 71(2), 219240.
Hazan, E., Agarwal, A. & Kale, S. (2007), Logarithmic regret algorithms for online
convex optimization, Machine Learning 69(23), 169192.
Hinton, G. E., Osindero, S. & Teh, Y.-W. (2006), A fast learning algorithm for deep
belief nets, Neural Computation 18(7), 15271554.
Hiriart-Urruty, J.-B. & Lemarchal, C. (1993), Convex analysis and minimization
algorithms, Springer.
Hsu, C.-W., Chang, C.-C., & Lin, C.-J. (2003), A practical guide to support vector
classification.
Hyafil, L. & Rivest, R. L. (1976), Constructing optimal binary decision trees is NPcomplete, Information Processing Letters 5(1), 1517.
Joachims, T. (2005), A support vector method for multivariate performance measures,
in Proceedings of the international conference on machine learning (ICML).
Kakade, S., Sridharan, K. & Tewari, A. (2008), On the complexity of linear prediction:
Risk bounds, margin bounds, and regularization, in NIPS.
Karp, R. M. (1972), Reducibility among combinatorial problems, Springer.
Kearns, M. & Mansour, Y. (1996), On the boosting ability of top-down decision tree
learning algorithms, in ACM Symposium on the Theory of Computing (STOC).
Kearns, M. & Ron, D. (1999), Algorithmic stability and sanity-check bounds for leaveone-out cross-validation, Neural Computation 11(6), 14271453.

References

Kearns, M. & Valiant, L. G. (1988), Learning Boolean formulae or finite automata

is as hard as factoring, Technical Report TR-14-88, Harvard University, Aiken
Computation Laboratory.
Kearns, M. & Vazirani, U. (1994), An Introduction to Computational Learning Theory,
MIT Press.
Kearns, M. J., Schapire, R. E. & Sellie, L. M. (1994), Toward efficient agnostic
learning, Machine Learning 17, 115141.
Kleinberg, J. (2003), An impossibility theorem for clustering, NIPS, pp. 463470.
Klivans, A. R. & Sherstov, A. A. (2006), Cryptographic hardness for learning intersections of halfspaces, in FOCS.
Koller, D. & Friedman, N. (2009), Probabilistic graphical models: Principles and
techniques, MIT Press.
Koltchinskii, V. & Panchenko, D. (2000), Rademacher processes and bounding the risk
of function learning, in High Dimensional Probability II, Springer, pp. 443457.
Kuhn, H. W. (1955), The hungarian method for the assignment problem, Naval
Research Logistics Quarterly 2(12), 8397.
Kutin, S. & Niyogi, P. (2002), Almost-everywhere algorithmic stability and generalization error, in Proceedings of the 18th conference in uncertainty in artificial intelligence,
pp. 275282.
Lafferty, J., McCallum, A. & Pereira, F. (2001), Conditional random fields: Probabilistic models for segmenting and labeling sequence data, in International conference on
machine learning, pp. 282289.
Langford, J. (2006), Tutorial on practical prediction theory for classification, Journal
of machine learning research 6(1), 273.
Langford, J. & Shawe-Taylor, J. (2003), PAC-Bayes & margins, in NIPS, pp. 423430.
Le, Q. V., Ranzato, M.-A., Monga, R., Devin, M., Corrado, G., Chen, K., Dean, J. & Ng,
A. Y. (2012), Building high-level features using large scale unsupervised learning,
in ICML.
Le Cun, L. (2004), Large scale online learning, in Advances in neural information
processing systems 16: Proceedings of the 2003 conference, Vol. 16, MIT Press, p. 217.
LeCun, Y. & Bengio, Y. (1995), Convolutional networks for images, speech, and time
series, in The handbook of brain theory and neural networks, The MIT Press.
Lee, H., Grosse, R., Ranganath, R. & Ng, A. (2009), Convolutional deep belief
networks for scalable unsupervised learning of hierarchical representations, in
ICML.
Littlestone, N. (1988), Learning quickly when irrelevant attributes abound: A new
linear-threshold algorithm, Machine Learning 2, 285318.
Littlestone, N. & Warmuth, M. (1986), Relating data compression and learnability.
Unpublished manuscript.
Littlestone, N. & Warmuth, M. K. (1994), The weighted majority algorithm, Information and Computation 108, 212261.
Livni, R., Shalev-Shwartz, S. & Shamir, O. (2013), A provably efficient algorithm for
training deep networks, arXiv preprint arXiv:1304.7045 .
Livni, R. & Simon, P. (2013), Honest compressions and their application to compression schemes, in COLT.
MacKay, D. J. (2003), Information theory, inference and learning algorithms, Cambridge
University Press.
Mallat, S. & Zhang, Z. (1993), Matching pursuits with time-frequency dictionaries,
IEEE Transactions on Signal Processing 41, 33973415.
McAllester, D. A. (1998), Some PAC-Bayesian theorems, in COLT.
McAllester, D. A. (1999), PAC-Bayesian model averaging, in COLT, pp. 164170.
McAllester, D. A. (2003), Simplified PAC-Bayesian margin bounds, in COLT,
pp. 203215.

389

390

References

Minsky, M. & Papert, S. (1969), Perceptrons: An introduction to computational geometry,

The MIT Press.
Mukherjee, S., Niyogi, P., Poggio, T. & Rifkin, R. (2006), Learning theory: stability is
sufficient for generalization and necessary and sufficient for consistency of empirical
risk minimization, Advances in Computational Mathematics 25(13), 161193.
Murata, N. (1998), A statistical study of on-line learning, Online Learning and Neural
Networks, Cambridge University Press.
Murphy, K. P. (2012), Machine learning: a probabilistic perspective, The MIT Press.
Natarajan, B. (1995), Sparse approximate solutions to linear systems, SIAM J.
Computing 25(2), 227234.
Natarajan, B. K. (1989), On learning sets and functions, Mach. Learn. 4, 6797.
Nemirovski, A., Juditsky, A., Lan, G. & Shapiro, A. (2009), Robust stochastic
approximation approach to stochastic programming, SIAM Journal on Optimization
19(4), 15741609.
Nemirovski, A. & Yudin, D. (1978), Problem complexity and method efficiency in
optimization, Nauka, Moscow.
Nesterov, Y. (2005), Primal-dual subgradient methods for convex problems, Technical report, Center for Operations Research and Econometrics (CORE), Catholic
University of Louvain (UCL).
Nesterov, Y. & Nesterov, I. (2004), Introductory lectures on convex optimization: A basic
course, Vol. 87, Springer, Netherlands.
Novikoff, A. B. J. (1962), On convergence proofs on perceptrons, in Proceedings of
the symposium on the mathematical theory of automata, Vol. XII, pp. 615622.
Parberry, I. (1994), Circuit complexity and neural networks, The MIT press.
Pearson, K. (1901), On lines and planes of closest fit to systems of points in space,
The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science
2(11), 559572.
Phillips, D. L. (1962), A technique for the numerical solution of certain integral
equations of the first kind, Journal of the ACM 9(1), 8497.
Pisier, G. (19801981), Remarques sur un rsultat non publi de B. maurey.
Pitt, L. & Valiant, L. (1988), Computational limitations on learning from examples,
Journal of the Association for Computing Machinery 35(4), 965984.
Poon, H. & Domingos, P. (2011), Sum-product networks: A new deep architecture, in
Conference on Uncertainty in Artificial Intelligence (UAI).
Quinlan, J. R. (1986), Induction of decision trees, Machine Learning 1, 81106.
Quinlan, J. R. (1993), C4.5: Programs for machine learning, Morgan Kaufmann.
Rabiner, L. & Juang, B. (1986), An introduction to hidden markov models, IEEE
ASSP Magazine 3(1), 416.
Rakhlin, A., Shamir, O. & Sridharan, K. (2012), Making gradient descent optimal for
strongly convex stochastic optimization, in ICML.
Rakhlin, A., Sridharan, K. & Tewari, A. (2010), Online learning: Random averages,
combinatorial parameters, and learnability, in NIPS.
Rakhlin, S., Mukherjee, S. & Poggio, T. (2005), Stability results in learning theory,
Analysis and Applications 3(4), 397419.
Ranzato, M., Huang, F., Boureau, Y. & Lecun, Y. (2007), Unsupervised learning
of invariant feature hierarchies with applications to object recognition, in Computer Vision and Pattern Recognition, 2007. CVPR07. IEEE Conference on, IEEE,
pp. 18.
Rissanen, J. (1978), Modeling by shortest data description, Automatica 14, 465471.
Rissanen, J. (1983), A universal prior for integers and estimation by minimum
description length, The Annals of Statistics 11(2), 416431.
Robbins, H. & Monro, S. (1951), A stochastic approximation method, The Annals of
Mathematical Statistics, pp. 400407.

References

Rogers, W. & Wagner, T. (1978), A finite sample distribution-free performance bound

for local discrimination rules, The Annals of Statistics 6(3), 506514.
Rokach, L. (2007), Data mining with decision trees: Theory and applications, Vol. 69,
World Scientific.
Rosenblatt, F. (1958), The perceptron: A probabilistic model for information storage and organization in the brain, Psychological Review 65, 386407. (Reprinted in
Neurocomputing, MIT Press, 1988).
Rumelhart, D. E., Hinton, G. E. & Williams, R. J. (1986), Learning internal representations by error propagation, in D. E. Rumelhart & J. L. McClelland, eds, Parallel
distributed processing explorations in the microstructure of cognition, MIT Press,
chapter 8, pp. 318362.
Sankaran, J. K. (1993), A note on resolving infeasibility in linear programs by constraint
relaxation, Operations Research Letters 13(1), 1920.
Sauer, N. (1972), On the density of families of sets, Journal of Combinatorial Theory
Series A 13, 145147.
Schapire, R. (1990), The strength of weak learnability, Machine Learning 5(2),
197227.
Schapire, R. E. & Freund, Y. (2012), Boosting: Foundations and algorithms, MIT Press.
Schlkopf, B. & Smola, A. J. (2002), Learning with kernels: Support vector machines,
regularization, optimization and beyond, MIT Press.
Seeger, M. (2003), Pac-bayesian generalisation error bounds for gaussian process
classification, The Journal of Machine Learning Research 3, 233269.
Shakhnarovich, G., Darrell, T. & Indyk, P. (2006), Nearest-neighbor methods in learning
and vision: Theory and practice, MIT Press.
Shalev-Shwartz, S. (2007), Online Learning: Theory, Algorithms, and Applications, PhD
thesis, The Hebrew University.
Shalev-Shwartz, S. (2011), Online learning and online convex optimization, FoundaR in Machine Learning 4(2), 107194.
tions and Trends

Shalev-Shwartz, S., Shamir, O., Srebro, N. & Sridharan, K. (2010), Learnability,
stability and uniform convergence, The Journal of Machine Learning Research
9999, 26352670.
Shalev-Shwartz, S., Shamir, O. & Sridharan, K. (2010), Learning kernel-based
halfspaces with the zero-one loss, in COLT.
Shalev-Shwartz, S., Shamir, O., Sridharan, K. & Srebro, N. (2009), Stochastic convex
optimization, in COLT.
Shalev-Shwartz, S. & Singer, Y. (2008), On the equivalence of weak learnability and
linear separability: New relaxations and efficient boosting algorithms, in Proceedings
of the nineteenth annual conference on computational learning theory.
Shalev-Shwartz, S., Singer, Y. & Srebro, N. (2007), Pegasos: Primal Estimated
sub-GrAdient SOlver for SVM, in International conference on machine learning,
pp. 807814.
Shalev-Shwartz, S. & Srebro, N. (2008), SVM optimization: Inverse dependence
on training set size, in International conference on machine learning ICML,
pp. 928935.
Shalev-Shwartz, S., Zhang, T. & Srebro, N. (2010), Trading accuracy for sparsity
in optimization problems with sparsity constraints, Siam Journal on Optimization
20, 28072832.
Shamir, O. & Zhang, T. (2013), Stochastic gradient descent for non-smooth optimization: Convergence results and optimal averaging schemes, in ICML.

Shapiro, A., Dentcheva, D. & Ruszczynski,

A. (2009), Lectures on stochastic programming: modeling and theory, Vol. 9, Society for Industrial and Applied Mathematics.
Shelah, S. (1972), A combinatorial problem; stability and order for models and theories
in infinitary languages, Pac. J. Math 4, 247261.

391

392

References

Sipser, M. (2006), Introduction to the Theory of Computation, Thomson Course

Technology.
Slud, E. V. (1977), Distribution inequalities for the binomial law, The Annals of
Probability 5(3), 404412.
Steinwart, I. & Christmann, A. (2008), Support vector machines, Springerverlag, New
York.
Stone, C. (1977), Consistent nonparametric regression, The Annals of Statistics
5(4), 595620.
Taskar, B., Guestrin, C. & Koller, D. (2003), Max-margin markov networks, in NIPS.
Tibshirani, R. (1996), Regression shrinkage and selection via the lasso, J. Royal.
Statist. Soc B. 58(1), 267288.
Tikhonov, A. N. (1943), On the stability of inverse problems, Dolk. Akad. Nauk SSSR
39(5), 195198.
Tishby, N., Pereira, F. & Bialek, W. (1999), The information bottleneck method, in
The 37th Allerton conference on communication, control, and computing.
Tsochantaridis, I., Hofmann, T., Joachims, T. & Altun, Y. (2004), Support vector
machine learning for interdependent and structured output spaces, in Proceedings
of the twenty-first international conference on machine learning.
Valiant, L. G. (1984), A theory of the learnable, Communications of the ACM
27(11), 11341142.
Vapnik, V. (1992), Principles of risk minimization for learning theory, in J. E. Moody,
S. J. Hanson & R. P. Lippmann, eds., Advances in Neural Information Processing
Systems 4, Morgan Kaufmann, pp. 831838.
Vapnik, V. (1995), The Nature of Statistical Learning Theory, Springer.
Vapnik, V. N. (1982), Estimation of Dependences Based on Empirical Data, SpringerVerlag.
Vapnik, V. N. (1998), Statistical Learning Theory, Wiley.
Vapnik, V. N. & Chervonenkis, A. Y. (1971), On the uniform convergence of relative frequencies of events to their probabilities, Theory of Probability and Its
Applications XVI(2), 264280.
Vapnik, V. N. & Chervonenkis, A. Y. (1974), Theory of pattern recognition, Nauka,
Moscow (In Russian).
Von Luxburg, U. (2007), A tutorial on spectral clustering, Statistics and Computing
17(4), 395416.
von Neumann, J. (1928), Zur theorie der gesellschaftsspiele (on the theory of parlor
games), Math. Ann. 100, 295320.
Von Neumann, J. (1953), A certain zero-sum two-person game equivalent to
the optimal assignment problem, Contributions to the Theory of Games 2,
512.
Vovk, V. G. (1990), Aggregating strategies, in COLT, pp. 371383.
Warmuth, M., Glocer, K. & Vishwanathan, S. (2008), Entropy regularized lpboost, in
Algorithmic Learning Theory (ALT).
Warmuth, M., Liao, J. & Ratsch, G. (2006), Totally corrective boosting algorithms that
maximize the margin, in Proceedings of the 23rd international conference on machine
learning.
Weston, J. & Watkins, C. (1999), Support vector machines for multi-class pattern
recognition, in Proceedings of the seventh european symposium on artificial neural
networks.
Wolpert, D. H. & Macready, W. G. (1997), No free lunch theorems for optimization,
Evolutionary Computation, IEEE Transactions on 1(1), 6782.

References

Zhang, T. (2004), Solving large scale linear prediction problems using stochastic gradient descent algorithms, in Proceedings of the twenty-first international conference on
machine learning.
Zhao, P. & Yu, B. (2006), On model selection consistency of Lasso, Journal of
Machine Learning Research 7, 25412567.
Zinkevich, M. (2003), Online convex programming and generalized infinitesimal
gradient ascent, in International conference on machine learning.

393

Index

3-term DNF, 79
F1 -score, 207

1 norm, 149, 286, 315, 335
accuracy, 18, 22
activation function, 229
AdaBoost, 101, 105, 314
all-pairs, 191, 353
approximation error, 37, 40
auto-encoders, 319
backpropagation, 237
backward elimination, 314
bag-of-words, 173
base hypothesis, 108
Bayes optimal, 24, 30, 221
Bayes rule, 306
Bayesian reasoning, 305
Bennets inequality, 376
Bernsteins inequality, 376
bias, 16, 37, 40
bias-complexity tradeoff, 41
Boolean conjunctions, 29, 54, 78
boosting, 101
boosting the confidence, 112
boundedness, 133
C4.5, 215
CART, 216
chaining, 338
Chebyshevs inequality, 373
Chernoff bounds, 373
class-sensitive feature mapping, 193
classifier, 14
clustering, 264
spectral, 271
compressed sensing, 285
compression bounds, 359
compression scheme, 360
computational complexity, 73

confidence, 18, 22
consistency, 66
Consistent, 247
contraction lemma, 331
convex, 124
function, 125
set, 124
strongly convex, 140, 160
convex-Lipschitz-bounded learning, 133
convex-smooth-bounded learning, 133
covering numbers, 337
curse of dimensionality, 224
decision stumps, 103, 104
decision trees, 212
dendrogram, 266, 267
dictionary learning, 319
differential set, 154
dimensionality reduction, 278
discretization trick, 34
discriminative, 295
distribution free, 295
domain, 13
domain of examples, 26
doubly stochastic matrix, 205
duality, 176
strong duality, 176
weak duality, 176
Dudley classes, 56
efficient computable, 73
EM, 301
Empirical Risk Minimization, see ERM
empirical error, 15
empirical risk, 15, 27
entropy, 298
relative entropy, 298
epigraph, 125
ERM, 15
error decomposition, 40, 135

395

396

Index
estimation error, 37, 40
Expectation-Maximization, see EM
face recognition, see Viola-Jones
feasible, 73
feature, 13
feature learning, 319
feature normalization, 316
feature selection, 309, 310
feature space, 179
feature transformations, 318
filters, 310
forward greedy selection, 312
frequentist, 305
gain, 215
GD, see gradient descent
generalization error, 14
generative models, 295
Gini index, 215
Glivenko-Cantelli, 35
gradient, 126
gradient descent, 151
Gram matrix, 183
growth function, 49
halfspace, 90
homogenous, 90, 170
nonseparable, 90
separable, 90
Halving, 247
hidden layers, 230
Hilbert space, 181
Hoeffdings inequality, 33, 375
holdout, 116
hypothesis, 14
hypothesis class, 16
i.i.d., 18
ID3, 214
improper, see representation independent
inductive bias, see bias
information bottleneck, 273
information gain, 215
instance, 13
instance space, 13
integral image, 113
Johnson-Lindenstrauss lemma, 284
k-means, 268, 270
soft k-means, 304
k-median, 269
k-medoids, 269
Kendall tau, 201
kernel PCA, 281
kernels, 179
Gaussian kernel, 184
kernel trick, 181
polynomial kernel, 183
RBF kernel, 184

label, 13
Lasso, 316, 335
generalization bounds, 335
latent variables, 301
LDA, 300
Ldim, 248, 249
learning curves, 122
least squares, 95
likelihood ratio, 301
linear discriminant analysis, see LDA
linear predictor, 89
homogenous, 90
linear programming, 91
linear regression, 94
linkage, 266
Lipschitzness, 128, 142, 157
subgradient, 155
Littlestone dimension, see Ldim
local minimum, 126
logistic regression, 97
loss, 15
loss function, 26
0-1 loss, 27, 134
absolute value loss, 95, 99, 133
convex loss, 131
generalized hinge loss, 195
hinge loss, 134
Lipschitz loss, 133
log-loss, 298
logistic loss, 98
ramp loss, 174
smooth loss, 133
square loss, 27
surrogate loss, 134, 259
margin, 168
Markovs inequality, 372
Massart lemma, 330
max linkage, 267
maximum a posteriori, 307
maximum likelihood, 295
McDiarmids inequality, 328
MDL, 63, 65, 213
measure concentration, 32, 372
Minimum Description Length, see MDL
mistake bound, 246
mixture of Gaussians, 301
model selection, 114, 117
multiclass, 25, 190, 351
cost-sensitive, 194
linear predictors, 193, 354
multivector, 193, 355
Perceptron, 211
reductions, 190, 354
SGD, 198
SVM, 197
multivariate performance measures, 206
Naive Bayes, 299
Natarajan dimension, 351
NDCG, 202

Index
Nearest Neighbor, 219
k-NN, 220
neural networks, 228
feedforward networks, 229
layered networks, 229
SGD, 236
No-Free-Lunch, 37
nonuniform learning, 59
Normalized Discounted Cumulative Gain, see
NDCG
Occams razor, 65
OMP, 312
one-vs.-all, 191, 353
one-vs.-rest, see one-vs.-all
online convex optimization, 257
online gradient descent, 257
online learning, 245
optimization error, 135
oracle inequality, 145
orthogonal matching pursuit, see OMP
overfitting, 15, 41, 121
PAC, 22
agnostic PAC, 23, 25
agnostic PAC for general loss, 27
PAC-Bayes, 364
parametric density estimation, 295
PCA, 279
Pearsons correlation coefficient, 311
Perceptron, 92
kernelized Perceptron, 188
multiclass, 211
online, 258
permutation matrix, 205
polynomial regression, 96
precision, 206
predictor, 14
prefix free language, 64
Principal Component Analysis, see PCA
prior knowledge, 39
Probably Approximately Correct, see PAC
projection, 159
projection lemma, 159
proper, 28
pruning, 216
Rademacher complexity, 325
random forests, 217
random projections, 283
ranking, 201
bipartite, 206
realizability, 17
recall, 206
regression, 26, 94, 138
regularization, 137
Tikhonov, 138, 140
regularized loss minimization, see RLM
representation independent, 28, 80
representative sample, 31, 325
representer theorem, 182

ridge regression, 138

kernel ridge regression, 188
RIP, 286
risk, 14, 24, 26
RLM, 137, 164
sample complexity, 22
Sauers lemma, 49
self-boundedness, 130
sensitivity, 206
SGD, 156
shattering, 45, 352
single linkage, 267
Singular Value Decomposition, see SVD
Sluds inequality, 378
smoothness, 129, 143, 163
SOA, 250
sparsity-inducing norms, 315
specificity, 206
spectral clustering, 271
SRM, 60, 115
stability, 139
Stochastic Gradient Descent, see SGD
strong learning, 102
Structural Risk Minimization, see SRM
structured output prediction, 198
subgradient, 154
Support Vector Machines, see SVM
SVD, 381
SVM, 167, 333
duality, 175
generalization bounds, 172, 333
hard-SVM, 168, 169
homogenous, 170
kernel trick, 181
soft-SVM, 171
support vectors, 175
target set, 26
term frequency, 194
TF-IDF, 194
training error, 15
training set, 13
true error, 14, 24
underfitting, 41, 121
uniform convergence, 31, 32
union bound, 19
unsupervised learning, 265
validation, 114, 116
cross validation, 119
train-validation-test split, 120
Vapnik-Chervonenkis dimension, see VC
dimension
VC dimension, 43, 46
version space, 247
Viola-Jones, 110
weak learning, 101, 102
Weighted-Majority, 252

397