Visit https://ebookfinal.

com to download the full version and

explore more ebooks

Introduction to Convex Optimization 3-4 Edition

Elad Hazan

_____ Click the link below to download _____

https://ebookfinal.com/download/introduction-to-convex-
optimization-3-4-edition-elad-hazan/

Explore and download more ebooks at ebookfinal.com

Here are some suggested products you might be interested in.
Click the link to download

Convex optimization and Euclidean distance geometry

Dattorro J.

https://ebookfinal.com/download/convex-optimization-and-euclidean-
distance-geometry-dattorro-j/

An Introduction to Nonlinear Optimization Theory Durea

https://ebookfinal.com/download/an-introduction-to-nonlinear-
optimization-theory-durea/

Introduction to 3 1 Numerical Relativity Miguel Alcubierre

https://ebookfinal.com/download/introduction-to-3-1-numerical-
relativity-miguel-alcubierre/

Introduction to Clinical Neurology 4 Edition Edition

Douglas James Gelb

https://ebookfinal.com/download/introduction-to-clinical-
neurology-4-edition-edition-douglas-james-gelb/
Black Box Optimization with Exact Subsolvers A Radial
Basis Function Algorithm for Problems with Convex
Constraints 1st Edition Christine Edman
https://ebookfinal.com/download/black-box-optimization-with-exact-
subsolvers-a-radial-basis-function-algorithm-for-problems-with-convex-
constraints-1st-edition-christine-edman/

Introduction to Spectroscopy 4 ed 4th ed Edition Pavia

Donald L.

https://ebookfinal.com/download/introduction-to-spectroscopy-4-ed-4th-
ed-edition-pavia-donald-l/

Introduction to the Theory of Optimization in Euclidean

Space 1st Edition Samia Challal (Author)

https://ebookfinal.com/download/introduction-to-the-theory-of-
optimization-in-euclidean-space-1st-edition-samia-challal-author/

VCE Accounting Units 3 4 5th Edition Neville Box

https://ebookfinal.com/download/vce-accounting-units-3-4-5th-edition-
neville-box/

Biologically Inspired Optimization Methods An Introduction

1st Edition Mattias Wahde

https://ebookfinal.com/download/biologically-inspired-optimization-
methods-an-introduction-1st-edition-mattias-wahde/
Introduction to Convex Optimization 3-4 Edition Elad
Hazan Digital Instant Download
Author(s): Elad Hazan
ISBN(s): 9781680831702, 1680831704
Edition: 3-4
File Details: PDF, 4.22 MB
Year: 2015
Language: english
Introduction to Online
Convex Optimization
Introduction to Online
Convex Optimization

Elad Hazan
Princeton University
[email protected]

Boston — Delft
Foundations and Trends R in Optimization

Published, sold and distributed by:

now Publishers Inc.
PO Box 1024
Hanover, MA 02339
United States
Tel. +1-781-985-4510
www.nowpublishers.com
[email protected]
Outside North America:
now Publishers Inc.
PO Box 179
2600 AD Delft
The Netherlands
Tel. +31-6-51115274
The preferred citation for this publication is
E. Hazan . Introduction to Online Convex Optimization. Foundations and Trends R
in Optimization, vol. 2, no. 3-4, pp. 157–325, 2015.
This Foundations and Trends R issue was typeset in LATEX using a class file designed
by Neal Parikh. Printed on acid-free paper.
ISBN: 978-1-68083-170-2
c 2016 E. Hazan

All rights reserved. No part of this publication may be reproduced, stored in a retrieval
system, or transmitted in any form or by any means, mechanical, photocopying, recording
or otherwise, without prior written permission of the publishers.
Photocopying. In the USA: This journal is registered at the Copyright Clearance Cen-
ter, Inc., 222 Rosewood Drive, Danvers, MA 01923. Authorization to photocopy items for
internal or personal use, or the internal or personal use of specific clients, is granted by
now Publishers Inc for users registered with the Copyright Clearance Center (CCC). The
‘services’ for users can be found on the internet at: www.copyright.com
For those organizations that have been granted a photocopy license, a separate system
of payment has been arranged. Authorization does not extend to other kinds of copy-
ing, such as that for general distribution, for advertising or promotional purposes, for
creating new collective works, or for resale. In the rest of the world: Permission to pho-
tocopy must be obtained from the copyright owner. Please apply to now Publishers Inc.,
PO Box 1024, Hanover, MA 02339, USA; Tel. +1 781 871 0245; www.nowpublishers.com;
[email protected]
now Publishers Inc. has an exclusive license to publish this material worldwide. Permission
to use this content must be obtained from the copyright license holder. Please apply to
now Publishers, PO Box 179, 2600 AD Delft, The Netherlands, www.nowpublishers.com;
e-mail: [email protected]
 Foundations and Trends R in Optimization
Volume 2, Issue 3-4, 2015
Editorial Board

Editors-in-Chief

Stephen Boyd Yinyu Ye

Stanford University Stanford University
United States United States

Editors

Dimitris Bertsimas Georgia Institute of Technology

Massachusetts Institute of Technology Yurii Nesterov
Dimitri P. Bertsekas UC Louvain
Massachusetts Institute of Technology Jorge Nocedal
John R. Birge Northwestern University
University of Chicago Pablo A. Parrilo
Robert E. Bixby Massachusetts Institute of Technology
Rice University Boris T. Polyak
Emmanuel Candès Institute for Control Science, Moscow
Stanford University Tamás Terlaky
David Donoho Lehigh University
Stanford University Michael J. Todd
Laurent El Ghaoui Cornell University
University of California, Berkeley Kim-Chuan Toh
Donald Goldfarb National University of Singapore
Columbia University John N. Tsitsiklis
Michael I. Jordan Massachusetts Institute of Technology
University of California, Berkeley Lieven Vandenberghe
Zhi-Quan (Tom) Luo University of California, Los Angeles
University of Minnesota, Twin Cites Robert J. Vanderbei
George L. Nemhauser Princeton University
Georgia Institute of Technology Stephen J. Wright
Arkadi Nemirovski University of Wisconsin
 Editorial Scope

Topics

Foundations and Trends R in Optimization publishes survey and tuto-

rial articles on methods for and applications of mathematical optimiza-
tion, including the following topics:

• Algorithm design, analysis, and implementation (especially on modern

computing platforms)

• Models and modeling systems

• New optimization formulations for practical problems

• Applications of optimization in:

– Machine learning
– Statistics
– Data analysis
– Signal and image processing
– Computational economics and finance
– Engineering design
– Scheduling and resource allocation
– and other areas

Information for Librarians

Foundations and Trends R in Optimization, 2015, Volume 2, 4 issues. ISSN

paper version 2167-3888. ISSN online version 2167-3918. Also available as a
combined paper and online subscription.
Foundations and Trends R in Optimization
Vol. 2, No. 3-4 (2015) 157–325
c 2016 E. Hazan
DOI: 10.1561/2400000013

Introduction to Online Convex Optimization

Elad Hazan
Princeton University
[email protected]
viii

to Dana
 Preface

This book serves as an introduction to the expanding theory of online

convex optimization. It was written as an advanced text to serve as
a basis for a graduate course, and/or as a reference to the researcher
diving into this fascinating world at the intersection of optimization
and machine learning.
Such a course was given at the Technion in the years 2010–2014
with slight variations from year to year, and later at Princeton Univer-
sity in the years 2015-2016. The core material in these courses is fully
covered in this book, along with exercises that allow the students to
complete parts of proofs, or that were found illuminating and thought-
provoking. Most of the material is given with examples of applications,
which are interlaced throughout different topics. These include pre-
diction from expert advice, portfolio selection, matrix completion and
recommendation systems, SVM training and more.
Our hope is that this compendium of material and exercises will be
useful to you; the educator and the researcher.

Placing this book in the machine learning library

The broad field of machine learning broadly speaking, as in the sub-

disciplines of online learning, boosting, regret minimization in games,
universal prediction and other related topics, have seen a plethora of

ix
x Preface

introductory texts in recent years. With this note we can hardly do

justice to all, but perhaps point to the location of this book in the
readers’ virtual library.
The neighboring book, which served as an inspiration to the current
manuscript, and indeed an inspiration to the entire field of learning
in games, is the wonderful text of Cesa-Bianchi and Lugosi (29). On
the other side, there are the numerous introductory essays to convex
optimization and convex analysis, to name a few (23; 78; 76; 77; 21; 92).
The more broad texts on machine learning are too numerous to state
hereby, but are definitely not far.
The primary purpose of this manuscript is to serve as an educa-
tional textbook for a dedicated course on online convex optimization
and the convex optimization approach to machine learning. Online con-
vex optimization has already had enough impact to appear in several
surveys and introductory texts, such as (53; 97; 85; 87). We hope this
compilation of material and exercises will further enrich the literature.

Book’s structure

This book is intended to serve as a reference for a self-contained course

for the educated graduate student in computer science/electrical engi-
neering/operations research/statistics and related fields. As such, its or-
ganization follows the structure of the course “decision analysis” taught
at the Technion.
Each chapter should take one or two weeks of classes, depending
on the depth and breadth of the intended course. The first chapter is
designed to be a “teaser” for the field, and thus less rigorous than the
rest of the book.
Roughly speaking, the book can be thought of as two units. The
first, from chapter 2 through 5, contains the basic definitions, frame-
work and core algorithms for online convex optimization. The rest of
the book deals with more advanced algorithms, more difficult settings
and relationships to well-known machine learning paradigms.
 Contents

Preface ix

1 Introduction 3
1.1 The online convex optimization model . . . . . . . . . . . 4
1.2 Examples of problems that can be modeled via OCO . . . 6
1.3 A gentle start: learning from expert advice . . . . . . . . . 10
1.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 Bibliographic remarks . . . . . . . . . . . . . . . . . . . . 19

2 Basic concepts in convex optimization 21

2.1 Basic definitions and setup . . . . . . . . . . . . . . . . . 21
2.2 Gradient/subgradient descent . . . . . . . . . . . . . . . . 25
2.3 Reductions to non-smooth and non-strongly convex functions 30
2.4 Example: support vector machine (SVM) training . . . . . 35
2.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.6 Bibliographic remarks . . . . . . . . . . . . . . . . . . . . 40

3 First order algorithms for online convex optimization 41

3.1 Online gradient descent . . . . . . . . . . . . . . . . . . . 42
3.2 Lower bounds . . . . . . . . . . . . . . . . . . . . . . . . 45
3.3 Logarithmic regret . . . . . . . . . . . . . . . . . . . . . . 46
3.4 Application: stochastic gradient descent . . . . . . . . . . 48

xi
xii Preface

3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6 Bibliographic remarks . . . . . . . . . . . . . . . . . . . . 53

4 Second order methods 55

4.1 Motivation: universal portfolio selection . . . . . . . . . . 55
4.2 Exp-concave functions . . . . . . . . . . . . . . . . . . . . 59
4.3 The online Newton step algorithm . . . . . . . . . . . . . 60
4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
4.5 Bibliographic Remarks . . . . . . . . . . . . . . . . . . . . 68

5 Regularization 69
5.1 Regularization functions . . . . . . . . . . . . . . . . . . . 70
5.2 The RFTL algorithm and its analysis . . . . . . . . . . . . 72
5.3 Online Mirrored Descent . . . . . . . . . . . . . . . . . . 76
5.4 Application and special cases . . . . . . . . . . . . . . . . 78
5.5 Randomized regularization . . . . . . . . . . . . . . . . . 81
5.6 * Optimal regularization . . . . . . . . . . . . . . . . . . 89
5.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.8 Bibliographic Remarks . . . . . . . . . . . . . . . . . . . . 97

6 Bandit Convex Optimization 99

6.1 The Bandit Convex Optimization model . . . . . . . . . . 100
6.2 The Multi Armed Bandit (MAB) problem . . . . . . . . . 100
6.3 A reduction from limited information to full information . . 105
6.4 Online gradient descent without a gradient . . . . . . . . . 110
6.5 * Optimal regret algorithms for BLO . . . . . . . . . . . . 112
6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.7 Bibliographic Remarks . . . . . . . . . . . . . . . . . . . . 119

7 Projection-free Algorithms 121

7.1 Review: relevant concepts from linear algebra . . . . . . . 122
7.2 Motivation: matrix completion and recommendation systems123
7.3 The conditional gradient method . . . . . . . . . . . . . . 124
7.4 Projections vs. linear optimization . . . . . . . . . . . . . 129
7.5 The online conditional gradient algorithm . . . . . . . . . 131
7.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Preface xiii

7.7 Bibliographic Remarks . . . . . . . . . . . . . . . . . . . . 137

8 Games, Duality and Regret 139

8.1 Linear programming and duality . . . . . . . . . . . . . . 140
8.2 Zero-sum games and equilibria . . . . . . . . . . . . . . . 141
8.3 Proof of von Neumann Theorem . . . . . . . . . . . . . . 145
8.4 Approximating Linear Programs . . . . . . . . . . . . . . . 147
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.6 Bibliographic Remarks . . . . . . . . . . . . . . . . . . . . 150

9 Learning Theory, Generalization and OCO 151

9.1 The setting of statistical learning theory . . . . . . . . . . 151
9.2 Agnostic learning using OCO . . . . . . . . . . . . . . . . 157
9.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
9.4 Bibliographic Remarks . . . . . . . . . . . . . . . . . . . . 166

Acknowledgements 167

References 169
Abstract

This manuscript portrays optimization as a process. In many practical

applications the environment is so complex that it is infeasible to lay
out a comprehensive theoretical model and use classical algorithmic
theory and mathematical optimization. It is necessary as well as ben-
eficial to take a robust approach, by applying an optimization method
that learns as one goes along, learning from experience as more aspects
of the problem are observed. This view of optimization as a process
has become prominent in varied fields and has led to some spectacular
success in modeling and systems that are now part of our daily lives.

E. Hazan . Introduction to Online Convex Optimization. Foundations and Trends R

in Optimization, vol. 2, no. 3-4, pp. 157–325, 2015.
DOI: 10.1561/2400000013.
 1
Introduction

This manuscript concerns the view of optimization as a process. In

many practical applications the environment is so complex that it is
infeasible to lay out a comprehensive theoretical model and use classical
algorithmic theory and mathematical optimization. It is necessary as
well as beneficial to take a robust approach, by applying an optimiza-
tion method that learns as one goes along, learning from experience as
more aspects of the problem are observed. This view of optimization as
a process has become prominent in various fields and led to spectacular
successes in modeling and systems that are now part of our daily lives.
The growing literature of machine learning, statistics, decision sci-
ence and mathematical optimization blur the classical distinctions be-
tween deterministic modeling, stochastic modeling and optimization
methodology. We continue this trend in this book, studying a promi-
nent optimization framework whose precise location in the mathemat-
ical sciences is unclear: the framework of online convex optimization,
which was first defined in the machine learning literature (see bibliogra-
phy at the end of this chapter). The metric of success is borrowed from
game theory, and the framework is closely tied to statistical learning
theory and convex optimization.

3
4 Introduction

We embrace these fruitful connections and, on purpose, do not try

to fit any particular jargon. Rather, this book will start with actual
problems that can be modeled and solved via online convex optimiza-
tion. We will proceed to present rigorous definitions, background, and
algorithms. Throughout, we provide connections to the literature in
other fields. It is our hope that you, the reader, will contribute to our
understanding of these connections from your domain of expertise, and
expand the growing literature on this fascinating subject.

1.1 The online convex optimization model

In online convex optimization, an online player iteratively makes deci-

sions. At the time of each decision, the outcomes associated with the
choices are unknown to the player.
After committing to a decision, the decision maker suffers a loss:
every possible decision incurs a (possibly different) loss. These losses
are unknown to the decision maker beforehand. The losses can be ad-
versarially chosen, and even depend on the action taken by the decision
maker.
Already at this point, several restrictions are necessary for this
framework to make any sense at all:

• The losses determined by an adversary should not be allowed to

be unbounded. Otherwise the adversary could keep decreasing
the scale of the loss at each step, and never allow the algorithm
to recover from the loss of the first step. Thus we assume the
losses lie in some bounded region.

• The decision set must be somehow bounded and/or structured,

though not necessarily finite.
To see why this is necessary, consider decision making with an
infinite set of possible decisions. An adversary can assign high
loss to all the strategies chosen by the player indefinitely, while
setting apart some strategies with zero loss. This precludes any
meaningful performance metric.
1.1. The online convex optimization model 5

Surprisingly, interesting statements and algorithms can be derived

with not much more than these two restrictions. The Online Convex
Optimization (OCO) framework models the decision set as a convex set
in Euclidean space denoted K ⊆ Rn . The costs are modeled as bounded
convex functions over K.
The OCO framework can be seen as a structured repeated game.
The protocol of this learning framework is as follows:
At iteration t, the online player chooses xt ∈ K . After the player
has committed to this choice, a convex cost function ft ∈ F : K 7→ R
is revealed. Here F is the bounded family of cost functions available to
the adversary. The cost incurred by the online player is ft (xt ), the value
of the cost function for the choice xt . Let T denote the total number
of game iterations.
What would make an algorithm a good OCO algorithm? As the
framework is game-theoretic and adversarial in nature, the appropriate
performance metric also comes from game theory: define the regret of
the decision maker to be the difference between the total cost she has
incurred and that of the best fixed decision in hindsight. In OCO we
are usually interested in an upper bound on the worst case regret of an
algorithm.
Let A be an algorithm for OCO, which maps a certain game history
to a decision in the decision set. We formally define the regret of A after
T iterations as:
( T T
)
X X
regretT (A) = sup ft (xt ) − min ft (x) (1.1)
{f1 ,...,fT }⊆F x∈K
t=1 t=1

Intuitively, an algorithm performs well if its regret is sublinear as

a function of T , i.e. regretT (A) = o(T ), since this implies that on the
average the algorithm performs as well as the best fixed strategy in
hindsight.
The running time of an algorithm for OCO is defined to be the
worst-case expected time to produce xt , for an iteration t ∈ [T ]1 in a T -
iteration repeated game. Typically, the running time will depend on n
(the dimensionality of the decision set K), T (the total number of game
1
Here and henceforth we denote by [n] the set of integers {1, ..., n}.
6 Introduction

iterations), and the parameters of the cost functions and underlying

convex set.

1.2 Examples of problems that can be modeled via OCO

Perhaps the main reason that OCO has become a leading online learn-
ing framework in recent years is its powerful modeling capability: prob-
lems from diverse domains such as online routing, ad selection for search
engines and spam filtering can all be modeled as special cases. In this
section, we briefly survey a few special cases and how they fit into the
OCO framework.

Prediction from expert advice

Perhaps the most well known problem in prediction theory is the so-
called “experts problem”. The decision maker has to choose among the
advice of n given experts. After making her choice, a loss between
zero and one is incurred. This scenario is repeated iteratively, and at
each iteration the costs of the various experts are arbitrary (possibly
even adversarial, trying to mislead the decision maker). The goal of the
decision maker is to do as well as the best expert in hindsight.
The online convex optimization problem captures this problem as
a special case: the set of decisions is the set of all distributions over
n elements (experts), i.e., the n-dimensional simplex K = ∆n = {x ∈
Rn , i xi = 1 , xi ≥ 0}. Let the cost of the i’th expert at iteration
P

t be gt (i), and let gt be the cost vector of all n experts. Then the
cost function is the expected cost of choosing an expert according to
distribution x, and is given by the linear function ft (x) = gt> x.
Thus, prediction from expert advice is a special case of OCO in
which the decision set is the simplex and the cost functions are linear
and bounded, in the `∞ norm, to be at most one. The bound on the
cost functions is derived from the bound on the elements of the cost
vector gt .
The fundamental importance of the experts problem in machine
learning warrants special attention, and we shall return to it and ana-
lyze it in detail at the end of this chapter.
1.2. Examples of problems that can be modeled via OCO 7

Online spam filtering

Consider an online spam-filtering system. Repeatedly, emails arrive into

the system and are classified as spam/valid. Obviously such a system
has to cope with adversarially generated data and dynamically change
with the varying input—a hallmark of the OCO model.
The linear variant of this model is captured by representing the
emails as vectors according to the “bag-of-words” representation. Each
email is represented as a vector x ∈ Rd , where d is the number of words
in the dictionary. The entries of this vector are all zero, except for those
coordinates that correspond to words appearing in the email, which are
assigned the value one.
To predict whether an email is spam, we learn a filter, for example
a vector x ∈ Rd . Usually a bound on the Euclidean norm of this vector
is decided upon a priori, and is a parameter of great importance in
practice.
Classification of an email a ∈ Rd by a filter x ∈ Rd is given by the
sign of the inner product between these two vectors, i.e., ŷ = signhx, ai
(with, for example, +1 meaning valid and −1 meaning spam).
In the OCO model of online spam filtering, the decision set is taken
to be the set of all such norm-bounded linear filters, i.e., the Euclidean
ball of a certain radius. The cost functions are determined according to
a stream of incoming emails arriving into the system, and their labels
(which may be known by the system, partially known, or not known
at all). Let (a, y) be an email/label pair. Then the corresponding cost
function over filters is given by f (x) = `(ŷ, y). Here ŷ is the classifi-
cation given by the filter x, y is the true label, and ` is a convex loss
function, for example, the square loss `(ŷ, y) = (ŷ − y)2 .

Online shortest paths

In the online shortest path problem, the decision maker is given a

directed graph G = (V, E) and a source-sink pair u, v ∈ V . At each
iteration t ∈ [T ], the decision maker chooses a path pt ∈ Pu,v , where
Pu,v ⊆ E |V | is the set of all u-v-paths in the graph. The adversary
independently chooses weights (lengths) on the edges of the graph,
8 Introduction

given by a function from the edges to the real numbers wt : E 7→ R,

which can be represented as a vector wt ∈ Rm , where m = |E|. The
decision maker suffers and observes a loss, which is the weighted length
P
of the chosen path e∈pt wt (e).
The discrete description of this problem as an experts problem,
where we have an expert for each path, presents an efficiency challenge.
There are potentially exponentially many paths in terms of the graph
representation size.
Alternatively, the online shortest path problem can be cast in the
online convex optimization framework as follows. Recall the standard
description of the set of all distributions over paths (flows) in a graph as
a convex set in Rm , with O(m + |V |) constraints (Figure 1.1). Denote
this flow polytope by K. The expected cost of a given flow x ∈ K
(distribution over paths) is then a linear function, given by ft (x) =
wt> x, where, as defined above, wt (e) is the length of the edge e ∈ E.
This inherently succinct formulation leads to computationally efficient
algorithms.

X X
xe = 1 = xe flow value is one
e=(u,w),w∈V e=(w,v),w∈V
X X
∀w ∈ V \ {u, v} xe = xe flow conservation
e=(v,x)∈E e=(x,v)∈E

∀e ∈ E 0 ≤ xe ≤ 1 capacity constraints

Figure 1.1: Linear equalities and inequalities that define the flow polytope, which
is the convex hull of all u-v paths.

Portfolio selection

In this section we consider a portfolio selection model that does not

make any statistical assumptions about the stock market (as opposed
to the standard geometric Brownian motion model for stock prices),
and is called the “universal portfolio selection” model.
1.2. Examples of problems that can be modeled via OCO 9

At each iteration t ∈ [T ], the decision maker chooses a distribution

of her wealth over n assets xt ∈ ∆n . The adversary independently
chooses market returns for the assets, i.e., a vector rt ∈ Rn with strictly
positive entries such that each coordinate rt (i) is the price ratio for the
i’th asset between the iterations t and t + 1. The ratio between the
wealth of the investor at iterations t + 1 and t is r> t xt , and hence
the gain in this setting is defined to be the logarithm of this change
ratio in wealth log(r>
t xt ). Notice that since xt is the distribution of the
investor’s wealth, even if xt+1 = xt , the investor may still need to trade
to adjust for price changes.
The goal of regret minimization, which in this case corresponds to
minimizing the difference maxx? ∈∆n Tt=1 log(r> ? PT >
t x )−
P
t=1 log(rt xt ),
has an intuitive interpretation. The first term is the logarithm of the
wealth accumulated by the best possible in-hindsight distribution x? .
Since this distribution is fixed, it corresponds to a strategy of rebal-
ancing the position after every trading period, and hence, is called a
constant rebalanced portfolio. The second term is the logarithm of the
wealth accumulated by the online decision maker. Hence regret mini-
mization corresponds to maximizing the ratio of the investor’s wealth
to the wealth of the best benchmark from a pool of investing strategies.
A universal portfolio selection algorithm is defined to be one that,
in this setting, attains regret converging to zero. Such an algorithm,
albeit requiring exponential time, was first described by Cover (see
bibliographic notes at the end of this chapter). The online convex op-
timization framework has given rise to much more efficient algorithms
based on Newton’s method. We shall return to study these in detail in
Chapter 4.

Matrix completion and recommendation systems

The prevalence of large-scale media delivery systems such as the Netflix

online video library, Spotify music service and many others, give rise
to very large scale recommendation systems. One of the most popular
and successful models for automated recommendation is the matrix
completion model.
10 Introduction

In this mathematical model, recommendations are thought of as

composing a matrix. The customers are represented by the rows, the
different media are the columns, and at the entry corresponding to a
particular user/media pair we have a value scoring the preference of
the user for that particular media.
For example, for the case of binary recommendations for music,
we have a matrix X ∈ {0, 1}n×m where n is the number of persons
considered, m is the number of songs in our library, and 0/1 signifies
dislike/like respectively:

 0, person i dislikes song j

Xij = .


1, person i likes song j

In the online setting, for each iteration the decision maker outputs
a preference matrix Xt ∈ K, where K ⊆ {0, 1}n×m is a subset of all
possible zero/one matrices. An adversary then chooses a user/song pair
(it , jt ) along with a “real” preference for this pair yt ∈ {0, 1}. Thus, the
loss experienced by the decision maker can be described by the convex
loss function,
ft (X) = (Xit ,jt − yt )2 .
The natural comparator in this scenario is a low-rank matrix, which
corresponds to the intuitive assumption that preference is determined
by few unknown factors. Regret with respect to this comparator means
performing, on the average, as few preference-prediction errors as the
best low-rank matrix.
We return to this problem and explore efficient algorithms for it in
Chapter 7.

1.3 A gentle start: learning from expert advice

Consider the following fundamental iterative decision making problem:

At each time step t = 1, 2, . . . , T , the decision maker faces a choice
between two actions A or B (i.e., buy or sell a certain stock). The
decision maker has assistance in the form of N “experts” that offer
their advice. After a choice between the two actions has been made,
1.3. A gentle start: learning from expert advice 11

the decision maker receives feedback in the form of a loss associated

with each decision. For simplicity one of the actions receives a loss of
zero (i.e., the “correct” decision) and the other a loss of one.
We make the following elementary observations:

1. A decision maker that chooses an action uniformly at random

each iteration, trivially attains a loss of T2 and is “correct” 50%
of the time.

2. In terms of the number of mistakes, no algorithm can do better

in the worst case! In a later exercise, we will devise a random-
ized setting in which the expected number of mistakes of any
algorithm is at least T2 .

We are thus motivated to consider a relative performance metric:

can the decision maker make as few mistakes as the best expert in
hindsight? The next theorem shows that the answer in the worst case
is negative for a deterministic decision maker.

Theorem 1.1. Let L ≤ T2 denote the number of mistakes made by

the best expert in hindsight. Then there does not exist a deterministic
algorithm that can guarantee less than 2L mistakes.

Proof. Assume that there are only two experts and one always chooses
option A while the other always chooses option B. Consider the setting
in which an adversary always chooses the opposite of our prediction (she
can do so, since our algorithm is deterministic). Then, the total number
of mistakes the algorithm makes is T . However, the best expert makes
no more than T2 mistakes (at every iteration exactly one of the two
experts is mistaken). Therefore, there is no algorithm that can always
guarantee less than 2L mistakes.

This observation motivates the design of random decision making

algorithms, and indeed, the OCO framework gracefully models deci-
12 Introduction

sions on a continuous probability space. Henceforth we prove Lemmas

1.3 and 1.4 that show the following:

Theorem 1.2. Let ε ∈ (0, 12 ). Suppose the best expert makes L mis-
takes. Then:

1. There is an efficient deterministic algorithm that can guarantee

N
less than 2(1 + ε)L + 2 log
ε mistakes;

2. There is an efficient randomized algorithm for which the expected

number of mistakes is at most (1 + ε)L + logε N .

1.3.1 The weighted majority algorithm

Simple observations: The weighted majority (WM) algorithm is in-
tuitive to describe: each expert i is assigned a weight Wt (i) at every
iteration t. Initially, we set W1 (i) = 1 for all experts i ∈ [N ]. For all
t ∈ [T ] let St (A), St (B) ⊆ [N ] be the set of experts that choose A (and
respectively B) at time t. Define,
X X
Wt (A) = Wt (i) Wt (B) = Wt (i)
i∈St (A) i∈St (B)

and predict according to


A if Wt (A) ≥ Wt (B)
at =
B otherwise.

Next, update the weights Wt (i) as follows:


W (i) if expert i was correct
t
Wt+1 (i) = ,
Wt (i)(1 − ε) if expert i was wrong

where ε is a parameter of the algorithm that will affect its performance.

This concludes the description of the WM algorithm. We proceed to
bound the number of mistakes it makes.
1.3. A gentle start: learning from expert advice 13

Lemma 1.3. Denote by Mt the number of mistakes the algorithm

makes until time t, and by Mt (i) the number of mistakes made by
expert i until time t. Then, for any expert i ∈ [N ] we have
2 log N
MT ≤ 2(1 + ε)MT (i) + .
ε
We can optimize ε to minimize the above bound. The expression on the
right hand side is of the form f (x) = ax+b/x, that reaches its minimum
at x = b/a. Therefore the bound is minimized at ε? = log N/MT (i).
p p

Using this optimal value of ε, we get that for the best expert i?
q 
?
MT ≤ 2MT (i ) + O MT (i? ) log N .

Of course, this value of ε? cannot be used in advance since we do not

know which expert is the best one ahead of time (and therefore we
do not know the value of MT (i? )). However, we shall see later on that
the same asymptotic bound can be obtained even without this prior
knowledge.
Let us now prove Lemma 1.3.

Proof. Let Φt = N i=1 Wt (i) for all t ∈ [T ], and note that Φ1 = N .

P

Notice that Φt+1 ≤ Φt . However, on iterations in which the WM

algorithm erred, we have
ε
Φt+1 ≤ Φt (1 − ),
2
the reason being that experts with at least half of total weight were
wrong (else WM would not have erred), and therefore
1 1 ε
Φt+1 ≤ Φt (1 − ε) + Φt = Φt (1 − ).
2 2 2
From both observations,
ε ε
Φt ≤ Φ1 (1 − )Mt = N (1 − )Mt .
2 2
On the other hand, by definition we have for any expert i that

WT (i) = (1 − ε)MT (i) .

14 Introduction

Since the value of WT (i) is always less than the sum of all weights ΦT ,
we conclude that
ε
(1 − ε)MT (i) = WT (i) ≤ ΦT ≤ N (1 − )MT .
2
Taking the logarithm of both sides we get
ε
MT (i) log(1 − ε) ≤ log N + MT log (1 − ).
2
Next, we use the approximations
1
−x − x2 ≤ log (1 − x) ≤ −x 0<x< ,
2
which follow from the Taylor series of the logarithm function, to obtain
that
ε
−MT (i)(ε + ε2 ) ≤ log N − MT ,
2
and the lemma follows.

1.3.2 Randomized weighted majority

In the randomized version of the WM algorithm, denoted RWM, we

choose expert i w.p. pt (i) = Wt (i)/ N
P
j=1 Wt (j) at time t.

Lemma 1.4. Let Mt denote the number of mistakes made by RWM

until iteration t. Then, for any expert i ∈ [N ] we have

log N
E[MT ] ≤ (1 + ε)MT (i) + .
ε
The proof of this lemma is very similar to the previous one, where the
factor of two is saved by the use of randomness:

Proof. As before, let Φt = N i=1 Wt (i) for all t ∈ [T ], and note that
P

Φ1 = N . Let m̃t = Mt − Mt−1 be the indicator variable that equals

one if the RWM algorithm makes a mistake on iteration t. Let mt (i)
equal one if the i’th expert makes a mistake on iteration t and zero
1.3. A gentle start: learning from expert advice 15

otherwise. Inspecting the sum of the weights:

X
Φt+1 = Wt (i)(1 − εmt (i))
i

pt (i) = PWW
t (i)
X
= Φt (1 − ε pt (i)mt (i)) (j)
j t
i
= Φt (1 − ε E[m̃t ])
≤ Φ e−ε E[m̃t ] .
t 1 + x ≤ ex

On the other hand, by definition we have for any expert i that

WT (i) = (1 − ε)MT (i)

Since the value of WT (i) is always less than the sum of all weights ΦT ,
we conclude that

(1 − ε)MT (i) = WT (i) ≤ ΦT ≤ N e−ε E[MT ] .

Taking the logarithm of both sides we get

MT (i) log(1 − ε) ≤ log N − ε E[MT ]

Next, we use the approximation

1
−x − x2 ≤ log (1 − x) ≤ −x , 0<x<
2
to obtain
−MT (i)(ε + ε2 ) ≤ log N − ε E[MT ],
and the lemma follows.

1.3.3 Hedge
The RWM algorithm is in fact more general: instead of considering
a discrete number of mistakes, we can consider measuring the perfor-
mance of an expert by a non-negative real number `t (i), which we refer
to as the loss of the expert i at iteration t. The randomized weighted
majority algorithm guarantees that a decision maker following its ad-
vice will incur an average expected loss approaching that of the best
expert in hindsight.
16 Introduction

Algorithm 1 Hedge
1: Initialize: ∀i ∈ [N ], W1 (i) = 1
2: for t = 1 to T do
3: Pick it ∼R Wt , i.e., it = i with probability xt (i) = PWW
t (i)
(j) j t

4: Incur loss `t (it ).

5: Update weights Wt+1 (i) = Wt (i)e−ε`t (i)
6: end for

Historically, this was observed by a different and closely related

algorithm called Hedge, whose total loss bound will be of interest to us
later on in the book.
Henceforth, denote in vector notation the expected loss of the al-
gorithm by
N
xt (i)`t (i) = xt> `t
X
E[`t (it )] =
i=1

`2t
Theorem 1.5. Let denote the N -dimensional vector of square losses,
i.e., `2t (i) = `t (i)2 , let ε > 0, and assume all losses to be non-negative.
The Hedge algorithm satisfies for any expert i? ∈ [N ]:
T T T
log N
xt> `t ≤ xt> `2t +
X X X
`t (i? ) + ε
t=1 t=1 t=1
ε
PN
Proof. As before, let Φt = i=1 Wt (i) for all t ∈ [T ], and note that
Φ1 = N .
Inspecting the sum of weights:
P −ε`t (i)
Φt+1 = i Wt (i)e
−ε`t (i) xt (i) = PWW
t (i)
P
= Φt i xt (i)e (j)
j t

≤ Φt − ε`t (i) + ε2 ` (i)2 )) for x ≥ 0,

P
i xt (i)(1 t
e−x ≤ 1 − x + x2
= Φt (1 − εxt> `t + ε2 xt> `2t )
> ` +ε2 x> `2
≤ Φt e−εxt t t t . 1 + x ≤ ex
On the other hand, by definition, for expert i? we have that
PT ?)
WT (i? ) = e−ε ` (i
t=1 t
1.3. A gentle start: learning from expert advice 17

Since the value of WT (i? ) is always less than the sum of all weights Φt ,
we conclude that
P P
xt> `t +ε2 xt> `2t
WT (i? ) ≤ ΦT ≤ N e−ε t t .

Taking the logarithm of both sides we get

T T T
xt> `t + ε2 xt> `2t
X X X
−ε `t (i? ) ≤ log N − ε
t=1 t=1 t=1

and the theorem follows by simplifying.

18 Introduction

1.4 Exercises

1. (Attributed to Claude Shannon)

Construct market returns over two stocks for which the wealth ac-
cumulated over any single stock decreases exponentially, whereas
the best constant rebalanced portfolio increases wealth exponen-
tially. More precisely, construct two sequences of numbers in the
range (0, ∞), that represent returns, such that:

(a) Investing in any of the individual stocks results in expo-

nential decrease in wealth. This means that the product of
the prefix of numbers in each of these sequences decreases
exponentially.
(b) Investing evenly on the two assets and rebalancing after ev-
ery iteration increases wealth exponentially.

2. (a) Consider the experts problem in which the payoffs are be-
tween zero and a positive real number G > 0. Give an algo-
rithm that attains expected payoff lower bounded by:
T
X T
X
`t (i? ) − c T log N
p
E[`t (it )] ≥ max
?
i ∈[N ]
t=1 t=1

for the best constant c you can (the constant c should be

independent of the number of game iterations T , and the
number of experts n. Assume that T is known in advance).
(b) Suppose the upper bound G is not known in advance. Give
an algorithm whose performance is asymptotically as good
as your algorithm in part (a), up to an additive and/or mul-
tiplicative constant which is independent of T, n, G. Prove
your claim.

3. Consider the experts problem in which the payoffs can be negative

and are real numbers in the range [−1, 1]. Give an algorithm with
√
regret guarantee of O( T log n) and prove your claim.
1.5. Bibliographic remarks 19

1.5 Bibliographic remarks

The OCO model was first defined by Zinkevich (110) and has since
become widely influential in the learning community and significantly
extended since (see thesis and surveys (52; 53; 97)).
The problem of prediction from expert advice and the Weighted
Majority algorithm were devised in (71; 73). This seminal work was
one of the first uses of the multiplicative updates method—a ubiquitous
meta-algorithm in computation and learning, see the survey (11) for
more details. The Hedge algorithm was introduced in (44).
The Universal Portfolios model was put forth in (32), and is one
of the first examples of a worst-case online learning model. Cover gave
an optimal-regret algorithm for universal portfolio selection that runs
in exponential time. A polynomial time algorithm was given in (62),
which was further sped up in (7; 54). Numerous extensions to the model
also appeared in the literature, including addition of transaction costs
(20) and relation to the Geometric Brownian Motion model for stock
prices (56).
In their influential paper, Awerbuch and Kleinberg (14) put forth
the application of OCO to online routing. A great deal of work has been
devoted since then to improve the initial bounds, and generalize it into
a complete framework for decision making with limited feedback. This
framework is an extension of OCO, called Bandit Convex Optimization
(BCO). We defer further bibliographic remarks to chapter 6 which is
devoted to the BCO framework.
 2
Basic concepts in convex optimization

In this chapter we give a gentle introduction to convex optimization

and present some basic algorithms for solving convex mathematical
programs. Although offline convex optimization is not our main topic,
it is useful to recall the basic definitions and results before we move on
to OCO. This will help in assessing the advantages and limitations of
OCO. Furthermore, we describe some tools that will be our bread-and-
butter later on.
The material in this chapter is far from being new. A broad and
significantly more detailed literature exists, and the reader is deferred
to the bibliography at the end of this chapter for references. We give
here only the most elementary analysis, and focus on the techniques
that will be of use to us later on.

2.1 Basic definitions and setup

The goal in this chapter is to minimize a continuous and convex func-

tion over a convex subset of Euclidean space. Henceforth, let K ⊆ Rd
be a bounded convex and compact set in Euclidean space. We denote

21
22 Basic concepts in convex optimization

by D an upper bound on the diameter of K:

∀x, y ∈ K, kx − yk ≤ D.

A set K is convex if for any x, y ∈ K, all the points on the line segment
connecting x and y also belong to K, i.e.,

∀α ∈ [0, 1], αx + (1 − α)y ∈ K.

A function f : K 7→ R is convex if for any x, y ∈ K

∀α ∈ [0, 1], f ((1 − α)x + αy) ≤ (1 − α)f (x) + αf (y).

Equivalently, if f is differentiable, that is, its gradient ∇f (x) exists for

all x ∈ K, then it is convex if and only if ∀x, y ∈ K

f (y) ≥ f (x) + ∇f (x)> (y − x).

For convex and non-differentiable functions f , the subgradient at x is

defined to be any member of the set of vectors {∇f (x)} that satisfies
the above for all y ∈ K.
We denote by G > 0 an upper bound on the norm of the subgradi-
ents of f over K, i.e., k∇f (x)k ≤ G for all x ∈ K. Such an upper bound
implies that the function f is Lipschitz continuous with parameter G,
that is, for all x, y ∈ K

|f (x) − f (y)| ≤ Gkx − yk.

The optimization and machine learning literature studies special

types of convex functions that admit useful properties, which in turn
allow for more efficient optimization. Notably, we say that a function
is α-strongly convex if
α
f (y) ≥ f (x) + ∇f (x)> (y − x) + ky − xk2 .
2
A function is β-smooth if
β
f (y) ≤ f (x) + ∇f (x)> (y − x) +ky − xk2 .
2
The latter condition is equivalent to a Lipschitz condition over the
gradients, i.e.,
k∇f (x) − ∇f (y)k ≤ βkx − yk.
2.1. Basic definitions and setup 23

If the function is twice differentiable and admits a second deriva-

tive, known as a Hessian for a function of several variables, the above
conditions are equivalent to the following condition on the Hessian,
denoted ∇2 f (x):
αI 4 ∇2 f (x) 4 βI,
where A 4 B if the matrix B − A is positive semidefinite.
When the function f is both α-strongly convex and β-smooth, we
say that it is γ-well-conditioned where γ is the ratio between strong
convexity and smoothness, also called the condition number of f
α
γ= ≤1
β

2.1.1 Projections onto convex sets

In the following algorithms we shall make use of a projection operation
onto a convex set, which is defined as the closest point inside the convex
set to a given point. Formally,

Π(y) , arg min kx − yk.

K x∈K

When clear from the context, we shall remove the K subscript. It is

left as an exercise to the reader to prove that the projection of a given
point over a compact convex set exists and is unique.
The computational complexity of projections is a subtle issue that
depends much on the characterization of K itself. Most generally, K can
be represented by a membership oracle—an efficient procedure that is
capable of deciding whether a given x belongs to K or not. In this
case, projections can be computed in polynomial time. In certain spe-
cial cases, projections can be computed very efficiently in near-linear
time. The computational cost of projections, as well as optimization
algorithms that avoid them altogether, is discussed in chapter 7.
A crucial property of projections that we shall make extensive use
of is the Pythagorean theorem, which we state here for completeness:
Theorem 2.1 (Pythagoras, circa 500 BC). Let K ⊆ Rd be a convex set,
y ∈ Rd and x = ΠK (y). Then for any z ∈ K we have
ky − zk ≥ kx − zk.
24 Basic concepts in convex optimization

Figure 2.1: Pythagorean theorem.

We note that there exists a more general version of the Pythagorean

theorem. The above theorem and the definition of projections are true
and valid not only for Euclidean norms, but for any norm. In addition,
projections according to other distances that are not norms can be
defined, in particular, with respect to Bregman divergences (see chapter
5), and an analogue of the Pythagorean theorem remains valid.

2.1.2 Introduction to optimality conditions

The standard curriculum of high school mathematics contains the basic

facts concerning when a function (usually in one dimension) attains a
local optimum or saddle point. The generalization of these conditions
to more than one dimension is called the KKT (Karush-Kuhn-Tucker)
conditions, and the reader is referred to the bibliographic material at
the end of this chapter for an in-depth rigorous discussion of optimality
conditions in general mathematical programming.
For our purposes, we describe only briefly and intuitively the main
facts that we will require henceforth. Naturally, we restrict ourselves to
2.2. Gradient/subgradient descent 25

convex programming, and thus a local minimum of a convex function

is also a global minimum (see exercises at the end of this chapter).
The generalization of the fact that a minimum of a convex differen-
tiable function on R is a point in which its derivative is equal to zero,
is given by the multi-dimensional analogue that its gradient is zero:
∇f (x) = 0 ⇐⇒ x ∈ arg min f (x).
x∈Rn
We will require a slightly more general, but equally intuitive, fact for
constrained optimization: at a minimum point of a constrained convex
function, the inner product between the negative gradient and direc-
tion towards the interior of K is non-positive. This is depicted in Figure
2.2, which shows that −∇f (x? ) defines a supporting hyperplane to K.
The intuition is that if the inner product were positive, one could im-
prove the objective by moving in the direction of the projected negative
gradient. This fact is stated formally in the following theorem.
Theorem 2.2 (Karush-Kuhn-Tucker). Let K ⊆ Rd be a convex set, x? ∈
arg minx∈K f (x). Then for any y ∈ K we have
∇f (x? )> (y − x? ) ≥ 0.

2.2 Gradient/subgradient descent

Gradient descent (GD) is the simplest and oldest of optimization meth-

ods. It is an iterative method—the optimization procedure proceeds
in iterations, each improving the objective value. The basic method
amounts to iteratively moving the current point in the direction of the
gradient, which is a linear time operation if the gradient is given explic-
itly (indeed, for many functions computing the gradient at a certain
point is a simple linear-time operation).
The following table summarises the convergence rates of GD vari-
ants for convex functions with different convexity parameters. The rates
described omit the (usually small) constants in the bounds—we focus
on asymptotic rates.
In this section we address only the first row of Table 2.1. For accel-
erated methods and their analysis see references at the bibliographic
section.
26 Basic concepts in convex optimization

Figure 2.2: Optimality conditions: negative (sub)gradient pointing outwards.

2.2.1 Basic gradient descent—linear convergence

Algorithmic box 2 describes a template for the basic gradient descent

method for mathematical optimization. It is a template since the se-
quence of step sizes {ηt } is left as an input parameter, and the several
variants of the algorithm differ on its choice.

Algorithm 2 Basic gradient descent

1: Input: f , T , initial point x1 ∈ K, sequence of step sizes {ηt }
2: for t = 1 to T do
3: Let yt+1 = xt − ηt ∇f (xt ), xt+1 = ΠK (yt+1 )
4: end for
5: return xT +1

Although the choice of ηt can make a difference in practice, in theory

the convergence of the vanilla GD algorithm is well understood and
given in the following theorem. Below, let ht = f (xt ) − f (x? ).
Exploring the Variety of Random
Documents with Different Content
Parslewe was a strange, even an eccentric man, who would do
things in his own fashion, and I was sufficiently learned in the ways
of the world, young as I was, to know that such men run into
danger. Where was he going that night, and to do—what? Evidently
on some mission which might need police interference. And
supposing that interference came along too late? What was I to do
then? When all was said and done, and in spite of what he had said
about myself and Madrasia in his easy-going fashion, I was almost a
stranger to him and to her, and I foresaw complications if anything
serious happened to him. I did not particularly love Parslewe that
evening; I thought he might have given me more of his confidence.
But there it was! and there was Madrasia. And Madrasia seemed to
have been restored to a more serene state of mind by this rejoining
of her guardian; evidently she possessed a sound belief in Parslewe’s
powers.
“Did he tell you anything this afternoon when I was out shopping?”
she inquired suddenly as we rode up Grey Street. “I’m sure you
must have talked.”
“Well,” I answered. “A little. He thinks the copper box has a false
bottom, a narrow cavity, in it, and that something has been
concealed there, and stolen from it. Possibly by our friend Weech.”
“Weech!” she exclaimed. “When?”
“When we left him alone with the box while we went up to the
library,” I replied. “Of course, that’s possible—if Weech knew the
secret.”
This seemed to fill her with new ideas.
“I wonder!” she said, musingly. “And is—is that what he’s after, here,
in Newcastle?”
“Something of the sort,” I assented. “At least, I gather so. You know
what he is—better than I do.”
She sat for a time in silence—in fact, till the cab drew up at the
theatre. Then she spoke.
“There’s one thing about Jimmie,” she remarked, reassuringly,
“nobody will get the better of him! So let him work things out.”
I did not tell her that there was no question of choice on my part.
We saw the play which Parslewe had commended so highly; we
went back to the hotel; we had supper together; then Madrasia went
off to her room. And it was then twenty minutes to twelve, and I sat
out every one of them, waiting, watching the door, listening for a
step in the corridor without. But Parslewe did not come.
And at one minute past twelve I seized my overcoat and cap and left
the room.
 IX
The Whitesmith’s Parlour
THERE were few people about in the big hall of the hotel, but
amongst them was the principal hall-porter, who, as I came there,
appeared to be handing over his duties to his deputy for the night.
An idea occurred to me, and I went up to him, drawing him aside.
“You know Mr. Parslewe?” I asked.
“Mr. Parslewe, sir—yes, sir!” he answered.
“Have you seen him go out this evening?”
“Yes, sir. Mr. Parslewe went out just about eleven, sir—not many
minutes before you came in with the young lady.”
“You haven’t seen him come in again?”
“Not yet, sir—not been in since then.”
I nodded, and went out into the street. So Parslewe’s business,
whatever it was, had been fixed for a late hour, after eleven. He
might have gone with us to the theatre, then! Unless, indeed, he
had been doing other business at the hotel. But it was no use
speculating on these things, my job was to do his bidding. And it
was not cowardice on my part that I heartily disliked the doing of it.
I had no idea as to the whereabouts of the police station. Parslewe
had said it was close by, but I did not know in which direction. I
might have inquired in the hotel, but I did not wish the hotel people
to know that we were or were about to be mixed up with the police;
it might have got to Madrasia’s ears before I got back. There were
still people in the streets, I could ask my way. And just then, as I
might have expected, a policeman came round the corner, and at my
question directed me. Parslewe had been quite correct, the place
was close at hand.
I went in, wonderingly, having never been in such a building before,
and not knowing what to expect, I had no more idea of what a
police headquarters was like than of the interior of an Eastern
palace, perhaps less. It was all very ordinary, when I got inside;
there was a well-lighted office, with a counter, and tables, and
desks; three or four policemen stood or sat about, examining papers
or writing in books. One of them, seeing me approach the counter
and probably noticing my diffident and greenhorn air, got off his
stool, put his pen behind his ear, and came across with an almost
fatherly solicitude on his fresh-coloured face.
“Can I see some responsible official?” I asked.
He half turned, indicating a man who wore braid on his closely
buttoned tunic, and sat at a desk in the corner.
“Inspector, sir,” he said. “Speak to him.”
He lifted a hinged door in the counter, and I went across to the man
in question. He looked up as I drew near, and gave me a swift
glance from top to toe. I had a vague sense of thankfulness that I
was well dressed.
“Yes!” he said.
I got close to him. Possibly I looked mysterious—anyway, I felt so.
“You know Mr. James Parslewe of Kelpieshaw, near Wooler?” I
suggested.
“Yes!”
“Mr. Parslewe is staying at the North Eastern Station Hotel. I am
staying there with him. He asked me, in case of a certain eventuality
——”
He interrupted me with an almost imperceptible smile—amused, I
think, at my precision of language.
“What eventuality?”
“In case he was not back at his hotel by midnight, I was to bring
and give you this note from him,” I answered, and laid the letter on
his desk. “He was not back—so I came straight to you.”
He picked up and opened the letter and began to read it; from
where I stood I could see that it covered three sides of a sheet of
the hotel notepaper. There was not a sign of anything—surprise,
perplexity, wonder—on the man’s face as he read—and he only read
the thing over once. Then he folded the letter, put it in his desk, and
turned to me.
“Mr. Alvery Craye, I think?” he asked.
“Yes,” said I.
“Do you know what’s in this letter, Mr. Craye?” he went on. “Did Mr.
Parslewe tell you its contents?”
“No,” I replied. “But he said I could answer any questions you asked,
and, if you go anywhere in consequence of the letter, I could go with
you.”
“I’ve only one question,” he remarked. “Do you know what time Mr.
Parslewe left the hotel?”
“Yes,” I said. “I found that out. He left just about eleven—a few
minutes before, I gathered from the hall-porter.”
He nodded, turned a key in his desk, put the key in his pocket, rose,
and asking me to sit down a moment, went across the room and
through a door at its farther extremity. Within a couple of minutes
he was back again, in company with a man in plain clothes; he
himself had put on a uniform overcoat and peaked cap. He made
some whispered communication to a sergeant who was busily
writing at a table in the centre of the room; then he beckoned to
me, and the three of us went out into the night.
At that moment I had not the slightest idea as to our destination.
There was a vague notion, utterly cloudy, in my mind that we might
be going to some dark and unsavoury quarter of the city; I had been
in Newcastle two or three times previously, and in my wanderings
had realised that it harboured some slums which were quite as
disreputable as anything you can find in Liverpool or Cardiff. But my
companions turned up town, towards the best parts of the place. It
was quiet in those spacious and stately streets, and the echo of our
footsteps sounded eerie in the silence. Nobody spoke until we had
walked some distance; then the inspector turned to me.
“Did Mr. Parslewe speak to you of any possible danger, Mr. Craye, as
regards what he was after?” he asked.
“I’ll tell you precisely what he did say,” I answered. “He said, ‘Not so
much danger as difficulty, though I won’t deny that there may be
danger.’ His exact words!”
“Just so,” he remarked. “And he didn’t tell you much more?”
“He told me nothing, except that he was hoping to get hold of a
possible something,” I replied. “If you know Mr. Parslewe, you know
that on occasion, when it suits him, he can be both vague and
ambiguous.”
“I know Mr. Parslewe—well enough!” he answered, with a sly
chuckle. “Highly eccentric gentleman, Mr. Parslewe, and
uncommonly fond of having his own way, and going his own way,
and taking his own line about everything. There isn’t one of his
brother magistrates in all Northumberland who isn’t aware of that,
Mr. Craye! Then, you have no idea of where we are going just now?”
“No idea whatever!” I answered.
“Well, as he said you could go with us, I may as well tell you,” he
remarked, with another laugh. “We’re going to the house and shop
of one Bickerdale, a whitesmith and coppersmith, in a side street
just up here. That’s where Mr. Parslewe’s gone.”
Of course, I might have known it! I felt myself an ass for not having
thought of it before. But I started, involuntarily.
“The name seems familiar to you,” suggested the inspector.
“Yes, I know it!” I asserted. “I’ve been in that shop. Oh! so he’s
there, is he?”
“That’s where we’re to look for him, anyway,” he replied. “But
whether we do find him there, or, if we do, under what conditions
it’ll be, that I don’t know. However, we’re carrying out his
instructions, and here’s the corner of the street.”
I knew that corner well enough, and the street, too. It was there
that I had shadowed White Whiskers and the Newcastle solicitor, and
thence that I had retreated after my passage-at-arms with
Bickerdale. Presently we stood before the side door of Bickerdale’s
shop, the door which presumably led to his house at the rear. There
was no light visible through the transom over the door, none in the
shop window, none in the windows over the shop. And when the
plain-clothes man, in response to the inspector’s order, rang the bell
and knocked in addition, no reply came.
It was not until we had knocked and rung three times, each more
loudly and urgently, that we heard sounds inside the door. They
were the sounds of somebody cautiously drawing back a bolt and
turning a key. But no light showed through keyhole or letter-box, or
the glass in the transom, and the inspector gave his man a
whispered instruction.
“Turn your lamp full on whoever opens the door!” he said. “And get
a foot over the threshold.”
I held my breath as the door was opened. It moved back; the plain-
clothes man’s light from a bull’s-eye lantern flashed on a frightened,
inquiring face looking round the edge of the door.
Weech!
I could have laughed aloud as Weech turned and fled, for he let out
a squeal at the sight of us, and bolted for all the world like a
frightened rabbit. And, of course, he left the door wide open, and we
were at once on his heels, and after him down the passage. He
swept aside a curtain, flung open a door behind it, and burst into a
well-lighted parlour or living-room with a sharp cry of warning.
“Police!”
I got a full view of the men in that room in one quick glance from
between the two policemen as they walked in. There was a table in
its centre, an oblong table; at our end of it, with his back to us, sat
Parslewe, calmly smoking a cigar; at the other, morose, perplexed,
defiant, sat Bickerdale. And behind Bickerdale, leaning against a
dresser or sideboard, stood Pawley!
These three all looked towards us as we entered, each with a
different expression. Bickerdale’s face became angry, almost savage;
Pawley appeared, after his first glance of surprise, to be intensely
annoyed. But Parslewe, half turning, motioned to the inspector and
whispered a few words to him; the inspector, his plain-clothes man,
and myself remained after that in the doorway by which we had
entered, and Parslewe gave his attention to Bickerdale, to whose
side, near the fireplace, Weech, still nervous and upset, had made
his way round the table.
“Now then, Bickerdale!” he said. “Without any more to do about it,
you’ll give me that document—you, or Weech, or both of you! Do
you hear—hand it over!”
“No!” exclaimed Pawley. “I object. If there’s any handing over, Mr.
Parslewe, it’ll be to me. And as you’ve brought police here, I’d better
say at once that——”
Parslewe suddenly rose from his chair. He held up his left hand—
towards Pawley. There was something in the gesture that made
Pawley break off short in his words and remain silent. As for
Parslewe’s right hand, it went into his pocket and brought out his
cigar case. Silently he handed it to the inspector, motioning him to
help himself and to pass it to his man. Then he turned to Pawley
again.
“Mr. Pawley!” he said in his most matter-of-fact tones. “It’s very
evident to me that you and I had better have a little conversation in
strict privacy. Bickerdale!—where have you a spare room?”
Bickerdale turned to Weech, growling something that sounded more
like a curse than an intimation. But Weech opened a door in the rear
of the room, and revealed a lighted kitchen place, and Parslewe,
motioning Pawley to follow him, went within. The door closed on
them.
They were in that kitchen a good half-hour. As for those of us—five
men—who were left in the sitting-room, we kept to our respective
camps. Bickerdale and Weech, at their end of the place, hung
together, eyeing us furtively, and occasionally whispering. Weech in
particular looked venomous, and by that time I had come to the
conclusion that he had bluffed Madrasia and myself very cleverly on
the occasion of his visit to Kelpieshaw. As for the inspector and his
man and myself, we sat in a line, on three very stiff-seated, straight-
backed chairs, smoking Parslewe’s cigars, for lack of anything better
to do, and watched and waited. Only once during that period of
suspense did any of us speak; that was when the inspector,
happening to catch my eye, gave me a quiet whisper.
“Queer business, Mr. Craye!” he said. “Odd!”
“Very,” said I.
He smiled and looked round at his man. But the man was one of
those stolid-faced individuals who seem as if nothing could move
them; also, he appeared to be relishing Parslewe’s cigar. With this in
one corner of his lips he sat immovable, watching the door through
which Parslewe and Pawley had vanished. I think he never took his
eyes off it; anyway, my recollection of him is as of a man who could
sit down and watch a thing or a person for hours and hours and
hours, without as much as flickering an eyelid—an uncanny,
uncomforting man—doubtless fitted, by some freak of nature, to his
trade of sleuth-hound.
As for me, I was wondering all round the affair. What was Parslewe
after? What was Pawley doing there? Who was Pawley, anyway?
Why, at that juncture, when his reinforcements had come up, did
Parslewe want to parley with Pawley? For a parley it was that was
going on in that kitchen, without a doubt. We heard nothing; there
were no raised voices, no evidence of any angry or wordy
discussion. All we heard was an occasional whisper from Weech, a
muttered growl from Bickerdale in response. But on the hearth a
kettle was singing, and at a corner of the rug, near Bickerdale’s
slippered feet, a cat sat and purred and purred....
The door of the kitchen suddenly opened; just as suddenly I saw
that whatever had been said, or whatever had taken place in that
kitchen, a marvellous transformation had developed in Pawley’s
manner. He held open the door for Parslewe, and stood aside
deferentially as Parslewe passed him. When he followed Parslewe
into the parlour it was with the air of a man who has either met his
master or been made subject to some revelation. And it was he who
spoke first—in answer to a nod from Parslewe. He turned to
Bickerdale.
“Mr. Bickerdale!” he said in suave, placatory tones. “I think you’d
better do what Mr. Parslewe asks! I—I’ve had some conversation
with Mr. Parslewe, and—and I think that’s what you’d better do, Mr.
Bickerdale—just so!”
Bickerdale turned on him with a sudden glare which denoted nothing
but sheer surprise. I could see that the man was fairly astonished—
amazed.
“Why—why!” he exclaimed. “It was you—you!—that told me just
now to do nothing of the sort!”
Pawley smiled in a queer, sickly, deprecating sort of fashion.
“Circumstances alter cases, Mr. Bickerdale,” he said. “I—I didn’t
know then what I know now. My advice is, now—do as Mr. Parslewe
wants.”
Weech sprang to his feet—an epitome of anger and chagrin.
“But us!” he vociferated. “Us—me and him! What are we going to
get out of it? Where shall we profit?” He turned almost savagely on
Bickerdale. “Don’t!” he went on. “Don’t you do it! Never mind those
fellows over there! there’s no police business in this that I know of,
and——”
“I’ll give you in charge of the police in two minutes, my lad!” said
Parslewe suddenly. “Just to show you——”
“Mr. Bickerdale,” said Pawley. “Take my advice! I—I understand—
from Mr. Parslewe—you’ll not be a loser.”
Bickerdale gave him a searching look. Then, suddenly, he thrust a
hand into his inner breast pocket and drew out a small square
envelope, which, with equal quickness, he handed across the table
to Parslewe. In its passage, the light from the lamp gleamed upon
this envelope; it seemed to me that I saw a crest on the flap.
“Rid of it now, anyway!” growled Bickerdale, sullenly. “Done!”
We were all watching Parslewe. He drew back to a corner of the
room, where a second lamp stood on a wall bracket. Beneath this he
turned the envelope over, examining it back and front; I saw then
that it had been slit open by Bickerdale or Weech, or somebody
through whose hands it had passed. And out of it Parslewe drew
what seemed to be an ordinary sheet of notepaper. Whatever was
written on it, he had read through in a minute. There were six pairs
of eyes watching him, but you might as well have hoped to get news
out of a stone wall as gain any information from his face; it was
more inscrutable and impassive than I had ever seen it. He showed
nothing—and suddenly he thrust paper and envelope into his pocket,
sat down at the table, pulled out a cheque-book and a fountain-pen,
and began to write. A moment later, he threw a cheque across to
Bickerdale; then, without a word to him, or to Weech, or to Pawley,
he strode out, motioning us to follow.
We made a little procession down town. Parslewe and the inspector
walked first; I heard them talking about county business—the
levying of a new rate, or some triviality of that sort. The plain-
clothes man and I brought up the rear; we talked about the weather,
and he told me that he had an allotment garden somewhere on the
outskirts and wanted rain for what he had just planted. Presently we
all parted, and Parslewe and I went to the hotel and up to the
private sitting-room. There was whisky and soda on the sideboard,
and he mixed a couple of glasses, handed me one, and drank his
own off at a draught. Then, when I had finished mine, he gave me a
questioning look.
“Bed, my lad?” he suggested. “Just so! Come on, then—your room’s
next to mine; we’ll go together.” We walked along the corridor
outside. “Do you want an idea—not an original one—to go to bed
with, Craye?” he asked abruptly, as we reached our doors. “I’ll give
you one. There are some damned queer things in this world!”
Then, with one of his loud, sardonic peals of laughter, he shook my
hand and shot into his room.
 X
Known at the Crown
THE various doings of that evening had not been of a nature that
conduced to sleep, and I lay awake for a long time wondering about
them. Naturally, my speculations chiefly ran on what I had seen in
Bickerdale’s back parlour. How came Pawley there? What did
Parslewe say to Pawley in that kitchen that made Pawley suddenly
transformed into a state of almost servile acquiescence in Parslewe’s
further doings? What was the document that Bickerdale handed over
to Parslewe? Had Bickerdale found it during his repairing of the
copper box, or had Augustus Weech abstracted it when Madrasia
and I left him alone in the parlour at Kelpieshaw? All these questions
ran helter-skelter through my brain as I lay there, anything but
sleepy, and, needless to say, I hadn’t satisfactorily answered one of
them when at last I dropped off to a more or less uneasy slumber.
The truth was that in spite of various developments the whole thing,
at two o’clock that morning, was to me a bigger and more
exasperating mystery than ever.
And it was still there when I woke, sharply, at six o’clock; so much
so, indeed, that I felt as if I should like to march into Parslewe’s
room next door, shake him out of his no doubt sound sleep, and tell
him that he’d got to make a clean breast of things, there and then.
As I knew no man in the world less likely to be forced into
confession until such time as he chose to speak, I took another
course to calm my perplexed state of mind. I rose, shaved, dressed,
and going downstairs, went out into the big station. Railway
stations, at any hour of the day, but especially early in the morning,
have a fascination for me—the goings and comings of the first trains,
the gradually increasing signs of waking up, the arrival of the
newspapers and opening of the bookstalls, even the unloading and
carrying away of the milk cans, are sources of mighty attraction. I
lounged about for some time, watching and observing; finally, as the
hands of the clock pointed to seven, and knowing that neither
Parslewe nor Madrasia would be ready for breakfast much before
nine, I turned into the refreshment room for a cup of coffee. And
there at the counter, a suit-case at his feet and a rug over his arm,
stood Pawley.
Pawley gave me a smile which was half bland and half sickly, and
wholly mysterious. And suddenly feeling that I had as good a right
as another to indulge an entirely natural sense of inquisitiveness, I
went up to him and bade him good morning. He responded civilly
enough, and it struck me that he was rather glad to see me and not
indisposed to talk. He was eating bread and butter and sipping tea; I
got some coffee and biscuits, and for a moment or two we stood
side by side, silent. But I had an idea that Pawley wanted me to
speak.
“Leaving?” I asked, with a glance at his belongings.
“That’s it, Mr. Craye,” he replied, almost eagerly. “By the seven-forty,
sir. I’m through with my job after last night.”
I noticed a difference in his tone and manner. He was no longer the
amateur antiquary, affecting a knowledge and a jargon carefully
acquired; he talked like what he probably was, an inquiry agent of
some sort. And in consonance with my previous feeling of intuition, I
thought that however much he might keep back, he was not against
communicating some of his knowledge.
“Last night’s proceedings,” I remarked, “were somewhat mysterious,
Mr. Pawley.”
“Mysterious!” he exclaimed. “I believe you! I’ve been concerned in
some queer things in my time, Mr. Craye, but in none queerer than
this! Beyond me! But no doubt you know more than I do.”
“I know nothing,” I answered. “Nothing, that is, beyond what I’ve
seen. And what I’ve seen I haven’t understood. For instance, I didn’t
understand how you came to be at Bickerdale’s last night.”
“Oh, that’s easy!” he said. “I was left here to keep an eye on
Bickerdale and to get in touch with him. And, incidentally, to find
out, if I could, whether Bickerdale had discovered anything in that
copper box when he had it.”
“Did anybody suspect that something might have been concealed in
the copper box?” I asked.
“To be sure, Mr. Craye! Sir Charles Sperrigoe suspected—does
suspect. That’s what he sent me up here for—to take a preliminary
look round. Then—came himself. Gone back now—but kept me here
for a day or two. To watch Bickerdale—as I said. And last night, just
as I was hoping to worm something out of Bickerdale and his son-in-
law, that ratty little chap—Weech—in walks Mr. Parslewe!”
“And did—what?” I asked.
He smiled, enigmatically.
“Mr. Parslewe, sir, is an odd gentleman!” he answered. “You’ll know
that by now, I think, Mr. Craye, though I understand you’re almost a
stranger to him. Well—Mr. Parslewe, he lost no time. He told
Bickerdale that he knew he’d found a document in a secret place in
that copper box—he’d the copper box with him, and he showed us
the trick of opening it.”
“Oh, he did, did he?” I interrupted, in surprise. “Found it out, eh?”
“He knew it, anyway,” replied Pawley. “It’s done by unscrewing the
feet—the little knobs—that the thing stands on. And he demanded of
Bickerdale that he should hand that document over there and then!
Sharp!”
“And—what then?” I inquired.
Pawley poured out more tea, and stirred it thoughtfully.
“That was where I came in,” he said. “I objected—as representing
Sir Charles Sperrigoe. I said he was the proper person to have any
document, and I was his representative. We were disputing that,
and Bickerdale was getting more obstinate about handing anything
over to anybody till he knew what he was getting out of it, when you
and those police chaps arrived. And the rest you know, Mr. Craye.”
“On the contrary,” said I, “I don’t! I don’t know anything. What
happened between you and Mr. Parslewe in that kitchen?”
But at that he shook his head, and I saw that there were things he
wouldn’t tell.
“As to that, Mr. Craye,” he answered, “Mr. Parslewe’s closed my lips!
But he’s a gentleman of his word, and after what he said to me, I’d
no choice, sir, no choice at all, but to fall in with his suggestion that
the document should be handed over to him. I couldn’t do anything
else—after what he told me. But as to what he told me—mum is the
word, Mr. Craye!—mum! At present.”
“Have you any idea what that document is?” I asked, going at last
straight to a principal point.
“None!” he replied quickly. “But—I’ve a very good idea!”
“What, then?” I put it to him. “I’d give a good deal to know.”
He glanced round, as if he feared to be overheard, though there was
no one near us.
“Well,” he said. “Have you heard of an old gentleman named
Palkeney, Mr. Matthew Palkeney, of Palkeney Manor, away there in
the Midlands, who died some little time ago, leaving money and a
fine place and no relatives, and from whose library that copper box
and those books were undoubtedly stolen?”
“I’ve heard of him—and of the rest,” I replied.
“Just so,” he said. “Well, Mr. Craye, between you and me, it’s my
belief that the document Mr. Parslewe got from Bickerdale last night
—or, rather, early this morning—was neither more nor less than Mr.
Matthew Palkeney’s will, which the old fellow—a queer old stick!—
had hidden in that copper box! That’s what I think!”
“Is it known that he made any will?” I asked.
“In the ordinary way, no,” he answered. “But things come out. This
would have come out before, but for the slowness of country folk to
tell anything. Sperrigoes, as the old gentleman’s solicitors, have
never been able to find any will, or trace of any. But recently—quite
recently—they’ve come across this—a couple of men on the estate,
one a woodman, the other a gamekeeper, have come forward to say
that some time ago they set their names to a paper which they saw
their old master sign his name to. What’s that but a will, Mr. Craye?
Come!”
“It sounds like it,” I agreed. “And you think that it was that that Mr.
Parslewe recovered last night?”
“I do,” he answered. “For I’ve heard—Sir Charles told me himself—
that when the old man was struck down and lay dying, all he spoke
of—as far as they could make out—was the copper box, coupled
with the family name. Mr. Craye, I think he hid that will in the copper
box, and that Mr. Parslewe now has it in his pocket!”
It seemed a probable suggestion, and I nodded my assent.
“I suppose we shall hear,” I said.
Pawley picked up his suit-case.
“I must go to my train,” he said. “Hear? Yes—and see, too, Mr.
Craye! I think you’ll hear and see some queer things within this next
day or two, if you’re remaining in Mr. Parslewe’s company. But, I’ll
say this—Mr. Parslewe, though unmistakably a queer, a very queer,
eccentric gentleman, is a straight ’un, and whatever he got from
Bickerdale, it’s safe with him. Otherwise I shouldn’t be going south.
And, as I say, if you’re stopping with Mr. Parslewe, I think you’ll have
some entertainment. Better than a tale, I call it!”
He said good morning at that, and went off to his train, and after
buying a morning newspaper, I turned into the hotel and went up to
the private sitting-room. And then, presently, came Madrasia.
I was not going to say anything to Madrasia—I mean, as regarded
the events of the night. Fortunately, she asked no questions—about
the past, at any rate; her sole concern seemed to be about the
immediate future. There was a waiter in the room, laying the table
for breakfast, when she came in; she and I withdrew into the
embrasure of the window, looking out on streets that had now
grown busy.
“Have you seen him this morning?” she asked significantly. “No? Did
you see him last night?”
“For a few minutes,” I answered.
“Did he say what we’re going to do to-day?” she inquired.
“Not a word!” said I. “Said nothing!”
“Not even whether we’re going home or not—or anything?” she
demanded. “No? But what are we here for?”
“I don’t know,” I replied. “Ask him!”
“Might as well ask the man on that monument!” she retorted,
pointing out of the window. “I feel like a marionette!—with Jimmie
pulling the strings just as he pleases.”
“Do you mind?” I asked.
“Well, it does seem as if one hadn’t a mind or a will of one’s own,”
she said. “Look here!—if he comes to breakfast with some new
scheme, or plan, or mad notion, what are you going to do—
yourself?”
I gave her a purposely steady look.
“Fall in with it,” I answered.
“You are!” she exclaimed. “Why?”
“Nothing else to do,” I replied.
She regarded me steadily for a while.
“I hope he hasn’t hypnotised us,” she said. “Seems to me he’s only
got to lift a finger and we walk after him like lambs!”
“Rather amusing, though, after all,” I observed. “Adds variety to life.
He may have something quite exciting in store for us to-day.”
“Oh, well,” she said. “If you like to be led about like a performing
bear—however, here he is!”
Parslewe and breakfast came together. He had evidently been down
to the bookstall to buy a financial paper, and he sat, grim and fixed
of expression, as he ate and drank, and read figures and statistics. It
was not until he had made an end of his food that he betrayed any
particular consciousness of our presence. Then, laying down his
paper, he bent across the table to his ward, favouring her with one
of his charming smiles.
“Well, my dear,” he said, “did you buy all that you wanted yesterday
afternoon?”
“Yes,” replied Madrasia, promptly, “I did.”
“Enough to last you till you get home?” he continued.
“Quite!” said Madrasia.
“You didn’t happen to buy a bag, or a case, or something to carry
your impedimenta in?” he asked with a grin.
“Yes, I did,” retorted Madrasia. “Do you think I was going to carry a
brown paper parcel back to Wooler? I bought a very nice bag.”
“Oh!” he said sweetly. “All right! Then you can just go and pack it,
my girl, and be ready in three quarters of an hour, for we’re going.”
“Going where?” demanded Madrasia. “Home?”
“Not just yet,” he said. “We’re going south—by the express. A good
way, too, and we must get seats in the luncheon car. So you be
ready.”
Madrasia pointed a slim finger at me.
“You’ve never asked him!” she exclaimed.
“That’s all right,” replied Parslewe, calmly. “He’s going too. We’re all
going. You go and pack your duds.” He turned to me as she went
out of the room and his smile was sweeter than ever. “You may as
well see it through, Craye,” he said. “I think we’ll have about settled
things up by noon to-morrow. And I’ll show you something that’ll
appeal to your artistic eye. Eh?”
“In for a penny, in for a pound, Mr. Parslewe,” I replied. “I’m game!
—but hanged if I know what it’s all about!”
“I’m not so certain that I know that myself, my lad!” he answered.
“But I think we’re getting near it. Well, be ready!”
He went off then, and I saw him no more until we all met to walk
across to the express. There was considerable satisfaction in
travelling with Parslewe, for he reduced the whole thing to the
perfection of comfort. He had already booked three places and a
table in the first-class dining-car, armed himself with a heap of
magazines and newspapers, and, as he said, we had nothing to do
but drop on our padded seats, put our toes under the table, and go
away in luxurious idleness. We went—but when that train moved out
of Newcastle and started on its long journey southward, neither
Madrasia nor myself had the faintest idea as to where we were
going. I think it was much to our credit—especially to hers—that we
made no inquiry, and allowed Parslewe to do just as he liked with us.
Certainly I had some vague, shadowy idea as to our destination;
probably it was Palkeney Manor, or to some place where somebody
—Sperrigoe, perhaps—lived who had some connection with it. But
then—I did not know where Palkeney Manor was; Pawley, to be
sure, had referred to it as being in the Midlands, but the Midlands
are wide-stretching.
Anyway, we remained in that express until, after we had had lunch,
it ran into Peterborough. There Parslewe, without notice, bundled us
out. He treated us to one of his sardonic grins when he had
shepherded us on to the platform.
“This is where we begin to travel,” he remarked, drily. “That was a
preliminary—no bother about that—all straight running. But now
——”
He broke off abruptly, and left us; presently we saw him in close
conversation with an official. Madrasia turned to me.
“Of course, you haven’t the least idea where we’re going,” she
suggested.
“If you mean, do I know exactly and precisely, no!” I answered. “If
you mean, do I know to within fifty or a hundred miles, yes!”
“Well?” she demanded.
“I think we’re going into one of the Midland counties,” I said,
resignedly. “There are several of them. I remember sufficient
geography to repeat their names if you want to hear them.”
“I don’t!” she answered. “But I wish we were there, wherever it is.
Where are we now? I mean, in relation to where we’re going?”
“A long way off,” I replied, consolingly. “That’s what he means when
he says we’re about to travel. The fact is, we’ve so far been on a
straight line; now, I suppose, we’ve got to cut across country. We’re
in the east, and we’ve got to go west.”
There was a clock over our heads, and Madrasia looked at it. We
were half-way through the afternoon.
“I suppose we shall land somewhere about midnight,” she said. “But
it’s just what I expected.”
She was wrong. We travelled a long way, to be sure, after leaving
Peterborough, and I knew, by passing such places as Rugby and
Warwick, that we were making into the heart of mid-England. But at
eight o’clock, and at a small station, Parslewe had us out of our
carriage and into a cab; within a few minutes we were in the quaint
old streets of what looked like a mediæval town. And even then we
did not know its name; all we knew was that he had ordered our
driver to carry us to the Crown. Presently we were there, and saw
an old-world hostelry, out of which came a very modern hall-porter,
who, at sight of Parslewe, smiled widely and touched his forehead.
“Glad to see you again, sir!” said this functionary. “Rooms, sir?”
Parslewe looked at the man with a quizzical, inquiring glance.
“So you remember me, do you?” he asked. “Eh?”
“Never forget a face, sir,” replied the hall-porter. “This way, sir!”
 XI
Back to Elizabeth
WE found ourselves in a wonderful old house, a place of nooks and
corners, old oak, old everything; it would not have surprised me if,
as I made my way along its queer passages and stairways, I had
met men in ruffs and women in farthingales. But it was modern
enough in its present administration, and they served us with a
capital dinner, late as it was in the evening. That and the charm of
our surroundings put Madrasia and myself into a better humour than
we had been in during the last tiresome stages of our long journey. I
think that by that time we were both inclined to enter into the spirit
of the thing; it was, after all, an adventure that had possibilities. And
in the middle of dinner Madrasia, with a whimsical laugh and a
roguish sparkle in her eyes, bent towards Parslewe, who, as usual,
was unconcerned, phlegmatic, and apparently quite at home.
“If it’s not stripping away too many of the wrappings of your
precious mystery, Jimmie,” she said, “may one ask a question?”
“Aye, why not?” he answered. “As long as it’s sensible.”
“Sensible enough,” she retorted. “Where are we?”
“Just so!” said I. “The same question has suggested itself to me,
limited as my intelligence is. For I don’t know!”
Parslewe regarded us with calm eyes and manner.
“We’re in the ancient and eminently picturesque and very
comfortable Crown Hotel at Medminster,” he answered. “Which
stands in the Market Place, and commands a remarkable view of one
of the last bits of mediæval England, as you can read in the guide
books, and, better still, see for yourselves in the morning.”
“Medminster,” exclaimed Madrasia. “But that’s where old Sperrigoe
comes from! At least, that was the address given in The Times
advertisement that Weech showed us.”
“Precisely!” agreed Parslewe, drily. “It was here in this very room, at
that table, that I had the honour of meeting Sperrigoe.”
“When?” demanded Madrasia.
“Oh! Some time ago,” he answered, indifferently.
“Then, if you’d been at home when he called at Kelpieshaw the
other morning, he’d have known you?” said Madrasia, giving me a
kick under the table. “Of course he would, as you’d met before.”
“He might have known my face,” answered Parslewe. “But he
wouldn’t have known my name; at least, I mean, he never knew my
name when I met him here. We chanced to meet here, as strangers,
happening to dine together at the same table; we smoked a cigar
together afterwards, and chatted about the old town. I heard his
name, but he never heard mine; to him I was a mere bird-of-
passage. I guess,” he went on, with one of his cynical laughs, “old
Sperrigoe would have been vastly astonished if he’d found me in at
Kelpieshaw and had recognised in me the stranger of the Crown!”
“Have you come here to see old Sperrigoe?” demanded Madrasia.
Parslewe was the best hand I ever came across at the fine art of
disregarding a direct question. His face became utterly blank and his
lips set, and remained so until a new whim came over him, and he
began to tell us something of the history of the old house to which
he had brought us. Of that he would talk, but we both saw that it
was no use questioning him on any other subject, and we left him
alone. But I had already learned something—Parslewe had been
there before; he had met White Whiskers there; White Whiskers
would know him: without a doubt he had come there to meet White
Whiskers. But why on earth did he elude White Whiskers at
Kelpieshaw?
Before the evening closed I learned something else. Madrasia retired
early; Parslewe began writing a letter in the smoking-room; left to
myself, I strolled out to the front door of the hotel to take a look at
my surroundings. The old Market Place was flooded in bright
moonlight, and I saw at once that Parslewe had been right when he
spoke of it as a bit of mediæval England. On all sides of me were
ancient half-timbered houses with high gables and quaintly
ornamented fronts; above them, at one end of the square, rose the
tall square tower of a church; fronting it, at the other end, was what
I took to be an old Moot Hall. But for the gas-lamps which twinkled
here and there, and for the signs and names above the shops, I
should have thought myself thrown back to Tudor times.
The hall-porter came out as I stood there, looked up at the sky, and
remarked that we should have a fine day to-morrow, and that good
weather was desirable, for people were beginning to go about.
“You get tourists here, I suppose?” said I.
“No end of ’em, sir,” he answered. “Deal of Americans come here—
they like that sort of thing”—waving his hand towards the old houses
opposite. “Nothing of that sort in their country, I understand, sir. Oh,
yes, full of tourists all the summer months, sir.”
“But you don’t remember all their faces, do you?” I suggested.
He laughed at my reference to his remark on our arrival.
“Why, no, sir, not chance comers like that,” he admitted. “Though I
wouldn’t be too sure on that point—one gets into a habit of noticing,
you know, sir. But in the case of anybody who stops here a day or
two—never forget, sir. I knew your friend at once—noticeable
gentleman, of course.”
“Easily recognised,” I suggested.
“Just so, sir. Though it’ll be—let me see—yes, three years or so since
he was here,” he answered. “Oh, I remember him well enough.
Stopped here two or three nights, one autumn. A collector of
curiosities, I think, sir. I remember I bought him a couple of packing-
cases to carry away odds and ends that he’d bought in the town.
Considerable trade done in that way here, sir—half those shops you
see on the other side are curio shops.”
“How do they keep up their stock?” I asked.
“Ah! that’s a question that a lot of people have put to me, sir,” he
replied. “You’d almost think they manufacture things! But the fact is,
sir, this is an old part of England, with a lot of old houses about, old
country seats, and the like. And families die out, and the stuff
they’ve been accumulating for generations comes to auction, and a
lot of it gets into these curio shops—that’s how it’s done, sir. Plenty
of antiques in those shops, sir, but nothing to what there is in the old
houses in the neighbourhood.”
“Do you know a house near here called Palkeney Manor?” I asked,
thinking that as this was an intelligent and communicative man I
might as well improve my own knowledge. “There is such a place, I
think?”
“Palkeney Manor, sir!” he answered readily. “To be sure, sir! Three
miles out—fine old house that is—sort of show-place; you can look
round it by paying a shilling—all our American customers go there,
and the shillings go to the local charities. Oh, yes! that was old Mr.
Matthew Palkeney’s. Dead now, he is, and they do say that the
lawyers don’t know who the property belongs to—haven’t found out
yet, anyway. Fine property it is, too. Queer old gentleman, old Mr.
Palkeney!—and that reminds me that I think your friend knew him,
sir. Leastways, the last day your friend was here I remember that old
Mr. Palkeney drove up in his carriage and gave me a parcel for him—
I helped him to pack that parcel in one of the cases I’d bought for
him.”
“You’ve an excellent memory,” I remarked.
“Oh, well, one thinks of things, sir,” he answered. “Faces, now, sir,
they stir your memory up, don’t you think? And I’ve seen some
remarkable faces in my time—faces that you’d remember twenty
years after. Some faces, of course, is that ordinary that you never
notice ’em. But others——”
At that moment Parslewe put his face through the swing door behind
us, and seeing the hall-porter on the steps came out. He had a letter
in his hand. Coming to the hall-porter he waved the letter towards
the west end of the Market Place.
“Isn’t Sir Charles Sperrigoe’s office just round that corner?” he
asked. “Aye? Well, just go and put this note into his letter-box, will
you? Then he’ll get it first thing in the morning. Go now, there’s a
good fellow!” He turned to me when the man had gone on his
errand. “Well, master!” he asked, in his half-cynical, half-humorous
fashion. “How does this appeal to your artistic sense?”
“A fine setting for a mystery, Mr. Parslewe,” I answered.
“I dare say you’re right,” he said with a laugh. “But I think we shall
have done with mysteries to-morrow, my lad. And what mystery
there is has been none of my making! Well, I’m off to my bed. Good
night.”
With that, and a pleasant nod, he went unconcernedly off, and
presently I followed his example, more mystified than ever by his
last remark. For if he had not made all this mystery, who had?
Whether the mystery was going to be done with next day or not, its
atmosphere was still thick upon us next morning. At ten o’clock,
Parslewe, who invariably made all his arrangements without
consulting anybody who was affected by them, marshalled us into a
carriage and pair at the door of the Crown and gave some
instructions, aside, to the coachman. We drove off into a singularly
picturesque and well-wooded country. Madrasia, fresh from the
almost treeless slopes of the Cheviots, was immediately in raptures
with it. Already the trees were in leaf, the wide-spreading meadows
were covered with fresh green, and in the vistas of woodland
through which we passed daffodils and wood anemones made
splashes of colour against the bursting verdure. New to her, too,
were the quaint thatched cottages, many of them half-timbered, and
all ancient, by the roadside.
“It’s like the old England that one sees in pictures!” she exclaimed.
“It’s as if we’d gone back!”
“We have gone back,” said Parslewe, with one of his queer, grim
looks. “Back to Elizabeth! There’s not much that’s altered hereabouts
since Shakespeare’s time—neither houses nor men. And if you’re
going to develop a taste for mediævalism, my dear, you’ll soon be
satisfied—we’re presently going to set foot in a house that’s as old
as they make ’em.”
But before this came about the carriage stopped at a wayside
cottage, and Parslewe, without a word to us, got out, knocked at the
door, and went in. He remained inside for several minutes; when he
emerged again it was in the company of a tall, weather-beaten old
man whom, because of his velveteen coat and general appearance, I
took to be a gamekeeper. Motioning to the coachman to follow him
up the road, Parslewe walked on ahead with his companion, and
presently turned into a wayside wood. Coming abreast of the bridle-
gate by which they had entered, we saw them in conversation with a
third man, also elderly, who was felling trees; for some minutes the
three stood talking together.
“What is he after now?” asked Madrasia.
I shook my head—nothing was going to draw me into speculations
about Parslewe’s proceedings.
“The best thing at the present juncture,” said I, as oracularly as
possible, “is just to let things occur. I don’t know what he’s after! Let
him pursue it! We shall find things out as we go on.”
“Where are we going?” she asked.
“I imagine—but I may be wrong—that we’re on our way to that
Palkeney Manor of which Murthwaite told us,” said I. “I think we’ve
been on our way to it ever since we left Newcastle yesterday
morning. That’s probably the house that’s as old as they make ’em.
You’d better be prepared—for anything. I am!”
“What do you mean?” she demanded. “You’re getting cryptic too!”
“I think my brains are addled!” I answered. “Never mind! I know
something!”
“What?” she asked.
I bent forward to her, endeavouring to look as mysterious as I
possibly could, and spoke in a hushed voice.
“This!” I said, thrillingly. “Parslewe has the copper box in his pocket!”
She drew back, staring at me as if she wondered whether my mind
had given way. I nodded solemnly.
“Fact!” said I. “I saw him put it there! It’s in the right-hand outer
pocket of his coat. And in the copper box lies the explanation of—
everything! Hush!—not a word! He’s coming back.”
Parslewe came back, leaving the two old men talking together, and I
noticed that they stared after his retreating figure with vast interest.
But instead of getting into the carriage again, he motioned to us, in
his usual imperious fashion, to get out. Then he turned to the
coachman.
“Isn’t there an inn along the road there, near the village?” he asked.
“Just so; then you go and put up your horses at it, and wait there till
I send for you. We’ll do the rest on foot,” he went on, turning to us.
“There’s a path through the woods a little farther on.”
He led us up the road for another hundred yards, then turned into a
bridle-track that wound through a mass of venerable old trees for a
good half-mile. We made slow progress, for Madrasia insisted on
gathering a bunch of primroses. She was putting the last finishing
touches to this when Parslewe, who had got a little ahead, called to
us.
“Now then, here you are!” he said. “Here’s the place!”
We went on, and found him at the edge of the wood, leaning over a
gate. He pointed before him with his stick.
“Palkeney Manor,” he remarked, drily.
Madrasia let out a sudden, whole-souled exclamation of delighted
wonder. I was not surprised; the scene before us was one of that
peculiar charm and quiet beauty which no other country than our
own can show. We were looking on an undulating park, vividly
green, studded with old trees beneath which antlered deer were
browsing; there was a tree-shaded stretch of water in one of the
miniature valleys, and cattle standing knee-deep in it, and above
this, on a rising ground, backed by tall elms and giant chestnuts,
stood a beautiful old house, mellowed by centuries of age.
We were all intent for a time, staring. Then Madrasia spoke, softly.
“What a picture of a place!” she said. “Jimmie! even you must think
it is!”
But Parslewe gave us one of his queer looks.
“Um!” he answered. “To tell you the truth, my girl, I was wondering
if the drains are all right! Picturesqueness is all very well—but,
however, we’ll go a bit nearer.”
We went slowly across the park, admiring its sylvan beauties, past
the shining water, past the shy deer, and up to the front of the
house, Madrasia’s ecstasies of admiration increasing with every step
we took. And as for myself, I was beginning to have a great wonder
and an itching curiosity—especially the itching curiosity. What were
we doing here?
But Parslewe seemed to know. He led us straight to the front door,
which stood open.
“This is a show-place on certain days in the week,” he said. “This is
one of those days, so we can go in.”
We went in. An elderly woman appeared. She wore a black silk
apron, and a thin gold chain round her neck; I took these to be
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

ebookfinal.com