Inverse Problems: Basics, Theory and

Applications in Geophysics 2nd Edition

Mathias Richter
Visit to download the full and correct content document:
https://textbookfull.com/product/inverse-problems-basics-theory-and-applications-in-g
eophysics-2nd-edition-mathias-richter/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Variational Source Conditions Quadratic Inverse

Problems Sparsity Promoting Regularization New Results
in Modern Theory of Inverse Problems and an Application
in Laser Optics Jens Flemming
https://textbookfull.com/product/variational-source-conditions-
quadratic-inverse-problems-sparsity-promoting-regularization-new-
results-in-modern-theory-of-inverse-problems-and-an-application-
in-laser-optics-jens-flemming/

An Introduction to the Mathematical Theory of Inverse

Problems 3rd Edition Andreas Kirsch

https://textbookfull.com/product/an-introduction-to-the-
mathematical-theory-of-inverse-problems-3rd-edition-andreas-
kirsch/

Biota Grow 2C gather 2C cook Loucas

https://textbookfull.com/product/biota-grow-2c-gather-2c-cook-
loucas/

Geometry of the generalized geodesic flow and inverse

spectral problems Second Edition Petkov

https://textbookfull.com/product/geometry-of-the-generalized-
geodesic-flow-and-inverse-spectral-problems-second-edition-
petkov/
Nanomaterials Theory Problems and Solutions 2nd Edition
Upendranath Nandi Debnarayan Jana

https://textbookfull.com/product/nanomaterials-theory-problems-
and-solutions-2nd-edition-upendranath-nandi-debnarayan-jana/

Square Matrices of Order 2 Theory Applications and

Problems Vasile Pop

https://textbookfull.com/product/square-matrices-of-
order-2-theory-applications-and-problems-vasile-pop/

Inverse Heat Transfer Fundamentals and Applications

First Edition M Necat Ozisik

https://textbookfull.com/product/inverse-heat-transfer-
fundamentals-and-applications-first-edition-m-necat-ozisik/

Hydraulics, Hydrology and Environmental Engineering 2nd

Edition Simon A. Mathias

https://textbookfull.com/product/hydraulics-hydrology-and-
environmental-engineering-2nd-edition-simon-a-mathias/

The Basics of Psychotherapy An Introduction to Theory

and Practice 2nd Ed 2nd Edition Bruce E. Wampold

https://textbookfull.com/product/the-basics-of-psychotherapy-an-
introduction-to-theory-and-practice-2nd-ed-2nd-edition-bruce-e-
wampold/
 Lecture Notes in Geosystems Mathematics
and Computing

Mathias Richter

Inverse
Problems
Basics, Theory and Applications in
Geophysics

Second Edition
Lecture Notes in Geosystems Mathematics
and Computing

Series Editors
Willi Freeden, Technische Universität Kaiserslautern, Kaiserslautern, Germany
M. Zuhair Nashed, University of Central Florida, Orlando, FL, USA
Otmar Scherzer, University of Vienna, Vienna, Austria

Editorial Board
Hans-Peter Bunge, Munich University, Munich, Germany
Yalchin Efendiev, Texas A&M University, College Station, TX, USA
Bulent Karasozen, Middle East Technical University, Ankara, Turkey
Volker Michel, University of Siegen, Siegen, Germany
Tarje Nissen-Meyer, University of Oxford, Oxford, UK
Nils Olsen, Technical University of Denmark, Kongens Lyngby, Denmark
Helmut Schaeben, Technical University Bergakademie Freiberg, Freiberg, Germany
Frederik J. Simons, Princeton University, Princeton, NJ, USA
Thomas Sonar, Technical University of Braunschweig, Braunschweig, Germany
Peter J.G. Teunissen, Delft University of Technology, Delft, The Netherlands
Johannes Wicht, Max Planck Institute for Solar System Research,
Göttingen, Germany

More information about this series at http://www.springer.com/series/15481

Mathias Richter

Inverse Problems
Basics, Theory and Applications
in Geophysics

Second Edition
Mathias Richter
Faculty for Electrical Engineering
and Information Technology
Bundeswehr University Munich
Neubiberg, Germany

ISSN 2730-5996 ISSN 2512-3211 (electronic)

Lecture Notes in Geosystems Mathematics and Computing
ISBN 978-3-030-59316-2 ISBN 978-3-030-59317-9 (eBook)
https://doi.org/10.1007/978-3-030-59317-9

Mathematics Subject Classification: 65F22, 65M32, 65R32, 86A22

© The Editor(s) (if applicable) and The Author(s), under exclusive license to
Springer Nature Switzerland AG 2016, 2020
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors, and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This book is published under the imprint Birkhäuser, www.birkhauser-science.com by the registered
company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface of the First Edition

The term “inverse problem” has no acknowledged mathematical definition, and its
meaning relies on notions from physics. One assumes that there is a known mapping

T : U → W,

which models a physical law or a physical device. Here, U is a set of “causes” and
W is a set of “effects.” The computation of an effect T (u) for a given cause u is
called a direct problem. Finding a cause u ∈ U that entails a given effect w ∈ W
is called an inverse problem. Solving an inverse problem thus means to ask for the
solution of an equation T (u) = w.
Maybe a certain effect w ∈ W is desirable and one is looking for some u ∈
U to produce it. An inverse problem of this kind is called a control problem. In
the following, it will be assumed that an effect is actually observed and that its
cause has to be found. An inverse problem of this kind is called an identification
problem. It arises when an interesting physical quantity is not directly amenable
to measurements but can only be derived from observations of its effects. There
are numerous examples of identification problems in science and engineering such
as:
• Inverse gravimetry: Given its mass density distribution (cause), the gravitational
force (effect) exercised by a body can be computed (direct problem). The inverse
problem is to derive the mass density distribution from measured gravitational
forces. An application is the detection of oil or gas reservoirs in geological
prospecting.
• Transmission tomography: Given the density distribution of tissue (cause),
the intensity loss (effect) of an X-ray traversing it can be computed (direct
problem). Inversely, one tries to find the tissue’s density from measured intensity
losses of X-rays traversing it. Applications exist in diagnostic radiology and
nondestructive testing.
• Elastography: From known material properties (cause), the displacement field
(effect) of an elastic body under external forces can be computed (direct

v
vi Preface of the First Edition

problem). The inverse problem is to derive material properties from observed

displacements. An application is the medical imaging of soft tissue.
• Seismic tomography: Knowing the spatial distribution of mechanical character-
istics (cause) of the earth’s crust and mantle, seismic energy propagation (effect)
from controlled sound sources or earthquakes can be computed (direct problem).
Inversely, one seeks to find mechanical characteristics from records of seismic
data. A geophysical application is to map the earth’s interior.
Very often, the solution u∗ of an inverse problem T (u) = w depends in an extremely
sensitive manner on w, two very similar effects having very different causes. This
would not be a principal problem, if w was known exactly. However, in practice,
one never knows w perfectly well but can only construct some approximation w̃ of
it from a limited number of generally inexact measurements. Then the true cause
u∗ ∈ U defined by T (u) = w cannot even be found to good approximation, no
matter the available computing power. This is so because w is not known, and,
by the aforementioned sensitivity, solving T (u) = w̃ would produce a completely
wrong answer ũ, far away from u∗ . There is no way to miraculously overcome this
difficulty, unless we are lucky and have some additional information of the kind “the
true cause u∗ has property P .” If this is the case, then we might replace the problem
of solving T (u) = w̃ by the different problem of finding a cause ũ within the set
of all candidates from U also having property P , such that some distance between
T (ũ) and w̃ is minimized. If the solution ũ of this replacement problem depends
less sensitively on the effect w̃ than the solution u∗ of the original problem depends
on w but converges to u∗ if w̃ converges to w, then one speaks of a regularization
of the original inverse problem.

Scope

Many excellent books on inverse problems have been written, we only mention
[Kir96, EHN96], and [Isa06]. These books rely on functional analysis as an
adequate mathematical tool for a unified approach to analysis and regularized
solution of inverse problems. Functional analysis, however, is not easily accessible
to non-mathematicians. It is the first goal of the present book to provide an access to
inverse problems without requiring more mathematical knowledge than is taught in
undergraduate math courses for scientists and engineers. From abstract analysis, we
will only need the concept of functions as vectors. Function spaces are introduced
informally in the course of the text, when needed. Additionally, a more detailed,
but still condensed, introduction is given in Appendix B. We will not deal with
regularization theory for operators. Instead, inverse problems will first be discretized
and described approximately by systems of algebraic equations and only then
will be regularized by setting up a replacement problem, which will always be
a minimization problem. A second goal is to elaborate on the single steps to be
taken when solving an inverse problem: discretization, regularization, and practical
Preface of the First Edition vii

solution of the regularized optimization problem. Rather than being as general as

possible, we want to work out these steps for model problems from the fields
of inverse gravimetry and seismic tomography. We will not delve into details of
numerical algorithms, though, when high-quality software is readily available on
which we can rely for computations. For the numerical examples in this book, the
programming environment Matlab [Mat14]1 was used as well as C programs from
[PTVF92].

Content

We start in Chap. 1 by presenting four typical examples of inverse problems, already

illustrating the sensitivity issue. We then formalize inverse problems as equations in
vector spaces and also formalize their sensitivity as “ill-posedness.” In the chapter’s
last two sections we have a closer look on problems from inverse gravimetry and
seismic tomography. We define specific model problems that we will tackle and
solve in later chapters. Special attention will be paid to the question whether sought-
after parameters (causes) can be uniquely identified from observed effects, at least
in the hypothetical case of perfect observations.
All model problems introduced in Chap. 1 are posed in function spaces: effects
and causes are functions. Chapter 2 is devoted to the discretization of such problems,
that is to the question of how equations in function spaces can approximately be
turned into equations in finite-dimensional spaces. Discretization is a prerequisite
for solving inverse problems on a computer. A first section on spline approximation
describes how general functions can be represented approximately by a finite set
of parameters. We then focus on linear inverse problems. For these, the least
squares method, investigated in Sect. 2.2, is a generally applicable discretization
method. Two alternatives, the collocation method and the method of Backus–
Gilbert, are presented in Sects. 2.3 and 2.4 for special, but important, classes of
inverse problems, known as Fredholm equations of the first kind. Two of our model
problems belong to this class. Even more specific are convolutional equations, which
play a significant role not only in inverse gravimetry but also in other fields like
communications engineering. Convolutional equations can be discretized in the
Fourier domain, which leads to very efficient solution algorithms. This is discussed
in Sect. 2.5. In the nonlinear case, discretizations are tailored to the specific problem
to be solved. We present discretizations of two nonlinear model problems from
inverse gravimetry and seismic tomography in Sect. 2.6.
The last two chapters investigate possibilities for a regularized solution of inverse
problems in finite-dimensional spaces (i.e., after discretization). In Chap. 3 we treat
the linear case, where the discretized problem takes the form of a linear system of
equations Ax = b. More generally, one considers the linear least squares problem

1 For the second edition, we used Release 2019b of Matlab.

viii Preface of the First Edition

of minimizing b − Ax2 , which is equivalent to solving Ax = b, if the latter does

have a solution, and is still solvable, if the system Ax = b no longer is. In the
first two sections, the linear least squares problem is analyzed, with an emphasis on
quantifying the sensitivity of its solution with respect to the problem data, consisting
of the vector b and of the matrix A. A measure of sensitivity can be given, known
as condition number in numerical analysis, which allows to estimate the impact of
data perturbations on the solution of a linear least squares problem. If the impact is
too large, a meaningful result cannot be obtained, since data perturbations can never
be completely avoided. The least squares problem then is not solvable practically.
Regularization tries to set up a new problem, the solution of which is close to
the actually desired one and which can be computed reliably. The general ideas
behind regularization are outlined in Sect. 3.3. The most popular regularization
technique, Tikhonov regularization, is explained and analyzed in Sect. 3.4, including
a discussion of how a regularized solution can be computed efficiently. Tikhonov
regularization requires the choice of a so-called regularization parameter. The proper
choice of this parameter is much debated. In Sect. 3.5 we present only a single
method to make it, the discrepancy principle, which is intuitively appealing and
often successful in practice. For alternative choices, we refer only to the literature.
Tikhonov regularization can also be used to solve problems derived by the Backus–
Gilbert method or for convolutional equations transformed to the Fourier domain.
These topics are treated in Sects. 3.7 and 3.8. Very interesting alternatives to
Tikhonov regularization are iterative regularization methods, which are attractive
for their computational efficiency. Two of these methods, the Landweber iteration
and the conjugate gradient method, are described in Sects. 3.9 and 3.10. It will be
found that the Landweber iteration can be improved by two measures, which relates
it back to Tikhonov regularization. One of these measures, a coordinate transform,
can also be taken to improve the conjugate gradient method. Technical details about
this coordinate transform are given in Sect. 3.6, which describes a transformation of
Tikhonov regularization to some standard form. This is also of independent interest,
since Tikhonov regularization in standard form is the easiest to analyze.
Regularization of nonlinear problems is studied in Chap. 4. In Sect. 4.1, nonlinear
Tikhonov regularization is treated abstractly, whereas in Sect. 4.2 this method is
applied to a model problem from nonlinear inverse gravimetry. Nonlinear Tikhonov
regularization leads to an optimization problem, namely a nonlinear least squares
problem. Section 4.3 discusses various possibilities for solving nonlinear least
squares problem numerically. All of these methods require the computation of
gradients, which can mean a prohibitive computational effort, if not done carefully.
The “adjoint method,” presented in Sect. 4.4, can sometimes drastically reduce the
numerical effort to obtain gradients. The adjoint method is presented in the context
of a nonlinear problem from seismic tomography, which is then solved in Sect. 4.5.
A final section presents one example of a nonlinear iterative regularization method,
the “inexact Newton-CG method.” The usefulness of this method is illustrated for a
problem of inverse gravimetry.
Appendix A lists required material from linear algebra and derives the singular
value decomposition of a matrix. Appendix B gives a condensed introduction
Preface of the First Edition ix

into functions spaces, i.e., into abstract vector spaces containing functions as
elements. Appendix C gives an introduction into the subject of (multidimensional)
Fourier transforms and their numerical implementation, including the case of
non-equidistant sample points. Finally, Appendix D contains a technical proof
outsourced from Chap. 3, which shows the regularizing property of the conjugate
gradient method applied to linear least squares problems.

Acknowledgments

I am much indebted to my editor Dr. Clemens Heine from Birkhäuser-Springer

publishers for his promotion of this book. I am greatly thankful to my friend
and colleague Professor Stefan Schäffler for his critical advice when reading my
manuscript and for having supported my work for well over twenty years.

Munich, Germany Mathias Richter

Preface of the Second Edition

Obvious goals of any second edition are the correction of known errors and the
improvement of the textual and graphical presentation of the material. This led to
minor changes throughout the monograph. Beyond that, we reworked our nonlinear
inverse model problems of inverse gravimetry and seismic tomography. After
discretization, in both cases a nonlinear least squares problem is obtained, the global
minimum point of which has to be found numerically. In the absence of convexity of
the objective function, this is challenging. On the one hand, all efficient solvers for
nonlinear least squares problems like the Gauß–Newton method, the Levenberg–
Marquardt method, or the reflective Newton method of Coleman and Li are not
fit to find a global minimum point. They iteratively improve an initial guess of a
minimum point by proceeding into directions of decreasing function values—this
way they can at best detect a local minimum point, where the iteration terminates.
These solvers must be provided a “good” starting point, close enough to the global
minimum point, which might be very hard to find. On the other hand, stochastic
algorithms for finding global minima do exist, but we do not know of any such
method that works efficiently when it has to face an optimization problem in many
dimensions and with many local minima at the same time, as it is the case for seismic
tomography and, to a lesser extent, for inverse gravimetry. While we have no all-
purpose method for solving any nonlinear least squares problem, we offer solutions
for our model problems.
For inverse gravimetry, multiscale optimization, as developed by Chavent and
Kunisch (see, e.g., [Cha09]), is quite effective. It relies on multiscale discretiza-
tion, which is presented in Sect. 2.6. In this section, we also give a reason for the
success of multiscale optimization, when applied to inverse gravimetry.
We completely revised our treatment of seismic tomography. First of all, we
now prefer to speak of full-waveform seismic inversion (FWI). The latter name
is commonly used in the literature for the problem we want to solve, whereas
“seismic tomography” could be confounded with “travel-time tomography,” which
in fact means a different kind of inverse problem considered in seismology. In our
previous treatment of FWI, we relied on the so-called travel-time transformation
and on the method of characteristics to derive a rather well-behaved nonlinear

xi
xii Preface of the Second Edition

least squares problem. However, this approach is limited to FWI in one space
dimension. We still restrict ourselves to the one-dimensional case only but now
base our inversion method on a standard finite difference discretization of the
underlying initial/boundary-value problem. This can be generalized to FWI in two
or three space dimensions. The new discretization is exposed in Sect. 2.7. It leads
to an optimization problem more challenging to solve than the one from the first
edition, due to the presence of multiple local minima of the objective function.
Multiscale optimization is no remedy in this case, since the objective function is
oscillatory even at large scales, as explained in the text. Instead, we resort to a
reconstruction method that gets us the sought-after subsurface parameters layer by
layer for increasing depths.
The structure of Chap. 4 was modified with a detailed discussion of our two
model problems now pushed to the last two sections. We also added a new appendix
on existence and uniqueness theorems for FWI.
I thank Dr. Rainer von Chossy, who carefully read parts of the first edition and
pointed out several errors to me. I also wish to thank my editor Christopher Tominich
from Birkhäuser-Springer publishers for his promotion of this book and the very
pleasant cooperation.

Munich, Germany Mathias Richter

February 2020
Contents

1 Characterization of Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Examples of Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Ill-Posed Problems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Model Problems for Inverse Gravimetry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Model Problems for Full-Waveform Seismic Inversion . . . . . . . . . . . . . 19
2 Discretization of Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.1 Approximation of Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.2 Discretization of Linear Problems by Least Squares Methods . . . . . . 39
2.3 Discretization of Fredholm Equations by Collocation Methods. . . . . 48
2.4 The Backus–Gilbert Method and the Approximative Inverse . . . . . . . 54
2.5 Discrete Fourier Inversion of Convolutional Equations. . . . . . . . . . . . . . 61
2.6 Discretization of Nonlinear Inverse Gravimetry . . . . . . . . . . . . . . . . . . . . . 68
2.7 Discretization of Nonlinear Waveform Inversion . . . . . . . . . . . . . . . . . . . . 75
3 Regularization of Linear Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.1 Linear Least Squares Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
3.2 Sensitivity Analysis of Linear Least Squares Problems . . . . . . . . . . . . . 88
3.3 The Concept of Regularization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
3.4 Tikhonov Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
3.5 Discrepancy Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
3.6 Reduction of Least Squares Regularization to Standard Form . . . . . . 128
3.7 Regularization of the Backus–Gilbert Method . . . . . . . . . . . . . . . . . . . . . . . 135
3.8 Regularization of Fourier Inversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
3.9 Landweber Iteration and the Curve of Steepest Descent . . . . . . . . . . . . 142
3.10 The Conjugate Gradient Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
4 Regularization of Nonlinear Inverse Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
4.1 Tikhonov Regularization of Nonlinear Problems . . . . . . . . . . . . . . . . . . . . 165
4.2 Nonlinear Least Squares Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
4.3 Computation of Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
4.4 Iterative Regularization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182

xiii
xiv Contents

4.5 Solution of Model Problem for Nonlinear Inverse Gravimetry. . . . . . 185

4.6 Solution of Model Problem for Nonlinear Waveform Inversion . . . . 201

A Results from Linear Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213

A.1 The Singular Value Decomposition (SVD) . . . . . . . . . . . . . . . . . . . . . . . . . . . 215

B Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221

B.1 Linear Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
B.2 Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
B.3 Normed Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
B.4 Inner Product Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
B.5 Convexity, Best Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233

C The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

C.1 One-Dimensional Discrete Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . 240
C.2 Discrete Fourier Transform for Non-equidistant Samples . . . . . . . . . . . 242
C.3 Error Estimates for Fourier Inversion in Sect. 2.5 . . . . . . . . . . . . . . . . . . . . 245

D Regularization Property of CGNE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253

E Existence and Uniqueness Theorems for Waveform Inversion . . . . . . . . . 259

E.1 Wave Equation with Constant Coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
E.2 Identifiability of Acoustic Impedance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
Chapter 1
Characterization of Inverse Problems

In Sect. 1.1 we present four typical examples of inverse problems from physics and
engineering and give an idea of why these problems can be difficult to solve in
practice. In all examples causes as well as effects are functions. Therefore, any
useful notion of inverse problems as equations relating causes to effects must be
general enough to include equations in function spaces, like differential or integral
equations. In Sect. 1.2 we do indeed describe inverse problems as equations in vector
spaces and define “ill-posedness,” which is the primary concern when dealing with
inverse problems. In Sects. 1.3 and 1.4, we formulate four model problems from
inverse gravimetry and full-waveform seismic inversion, which will be taken up in
later chapters to illustrate the methods presented in this book.

1.1 Examples of Inverse Problems

Example 1.1 (Determination of Growth Rates) The initial value problem

dw(t)
w  (t) = = u(t) · w(t), w(t0 ) = w0 > 0, t0 ≤ t ≤ t1 ,
t0 < t1 ,
dt
(1.1)
is a simple and well-known mathematical model for a growth process. Here, w(t)
might represent the population size of a colony of bacteria at time t. Then, w0
represents the known initial population size and u(t) represents a variable bacterial
growth rate. For a given, continuous function u : [t0 , t1 ] → R, (1.1) has a unique,
continuously differentiable solution w : [t0 , t1 ] →]0, ∞[, to wit:

t
w(t) = w0 · e U (t)
, U (t) = u(s) ds, t0 ≤ t ≤ t1 . (1.2)
t0

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2020 1

M. Richter, Inverse Problems, Lecture Notes in Geosystems
Mathematics and Computing, https://doi.org/10.1007/978-3-030-59317-9_1
2 1 Characterization of Inverse Problems

Thus, a cause u (the growth rate) entails an effect w (the population size). Formally,
this can be described as a mapping T : u → w, parameterized by t0 , t1 , and w0
(a mathematically correct definition of this mapping from one function space into
another is deferred to Sect. 1.2). Inversely, for a given function w, one asks for a
function u such that T (u) = w. This inverse problem also has a unique solution,
explicitly given by

w  (t) d
u(t) = = ln(w(t)), t0 ≤ t ≤ t1 . (1.3)
w(t) dt

Function u is the input of the direct problem (“the data”) and function w is its
result. For the inverse problem, function w is the input and function u is the result.
In practice, we never know the respective inputs exactly, since this would mean to
exactly know infinitely many values u(t) (for the direct problem) or w(t) (for the
inverse problem). Rather, we only have a finite number of measurements, subject to
measurement errors. From these measurements, we can only approximate the true
input function.
For the direct problem, such unavoidable errors do not have serious conse-
quences. To see this, assume that ũ is an approximation of the true input u with

max{|u(t) − ũ(t)|; t0 ≤ t ≤ t1 } ≤ ε for some ε > 0.

Then, for the corresponding results w̃ and w we have

max{|w(t) − w̃(t)|; t0 ≤ t ≤ t1 } ≤ εC

with some constant C.1 This means that deviations of results can be kept under
control if one cares to keep deviations of inputs small enough. This quality of
the mapping from inputs to outputs is called stability or robustness of the direct
problem. The inverse problem behaves very differently. For example, take
   
w : [t0 , t1 ] → R u : [t0 , t1 ] → R
input and result
t → esin(t) t → cos(t)

and consider the specific perturbed inputs

 
1
wn : [t0 , t1 ] → R, t → w(t) · 1 + √ cos(nt) , n ∈ N, n ≥ 2, (1.4)
n

1 From Theorem 12.V in [Wal98] we get

w0 (μ+ε)(t1 −t0 ) μ(t1 −t0 )
C= e (e − 1) where μ := max{|u(t)|; t0 ≤ t ≤ t1 }.
μ
1.1 Examples of Inverse Problems 3

with max{|wn (t) − w(t)|; t0 ≤ t ≤ t1 } → 0 for n → ∞. Inputs wn lead to results

√
n sin(nt)
un : [t0 , t1 ] → R, t → u(t) − , n ∈ N, n ≥ 2,
1+ √1
n
cos(nt)

with

max{|un (t) − u(t)|; t0 ≤ t ≤ t1 } → ∞ for n → ∞.

The better wn (t) approximates w(t), the larger becomes the deviation of un (t)
from u(t)! The reason for this is the differentiation operation in (1.3). Whereas
integration as in (1.2) is a smoothing operation (which, for example, will level out
sharp function peaks), the inverse differentiation operation necessarily must roughen
all function details which are smoothed by integration. But this will eventually also
blow up perturbations of the true input function. As a consequence, the explicit
formula (1.3), albeit mathematically correct, is practically useless. Figure 1.1
illustrates this example for n = 40. The graphs of input w and result u are shown as
red lines on the left and on the right, respectively. The wildly oscillating black lines
are the graphs of wn and un . ♦
The next example probably is the best known inverse problem.

8 8

6 6

4 4

2 2

0 0

-2 -2

-4 -4

-6 -6

-8 -8

0 2 4 6 8 0 2 4 6 8

Fig. 1.1 Effects w, wn and causes u, un in Example 1.1, n = 40

4 1 Characterization of Inverse Problems

Example 1.2 (Computerized Tomography, CT) A plane section of a body is

characterized by a density function f : R2 → R, x → f (x), where we assume
that

f (x) = 0 for x ∈ D := {x ∈ R2 ; x2 < 1} ⊆ R2 ,

which can always be achieved by proper scaling. The value f (x) models the tissue’s
density at position x. Knowing f (x) at every position x ∈ D means to know the
body’s interior. If we knew f , then we also could compute the intensity loss of an
X-ray sent through the body section along a straight line L: by Beer’s law, the ray’s
intensity loss at position x is proportional to f (x) and so the total loss corresponds
to the line integral

f (x) ds,
L

see Fig. 1.2, left. According to this figure, f is assumed to be a piecewise constant
function, equal function values being indicated by the same shading intensity with
darker shades meaning denser tissue.
We will rewrite the line integral. For a given angle ϕ let us define vectors
   
cos ϕ ⊥ − sin ϕ
θ := and θ := .
sin ϕ cos ϕ

Fig. 1.2 Principle of CT

1.1 Examples of Inverse Problems 5

Since every straight line L through D can be parameterized in the form

L = {sθ + tθ ⊥ ; t ∈ R} for ϕ ∈ [0, π [ and s ∈] − 1, 1[,

the line integral becomes

∞
Rf (ϕ, s) := Rϕ f (s) := f (sθ + tθ ⊥ ) dt. (1.5)
−∞

Since f (x) = 0 for x ∈ D, this is only seemingly an improper integral. Figure 1.2
in its right part illustrates Rϕ f for a fixed angle ϕ as a function of s. Note that local
maxima appear where the total intensity loss reaches a local maximum (depending
on the density of the traversed tissue). The function

Rf : [0, π [×] − 1, 1[→ R, (ϕ, s) → Rf (ϕ, s),

is called Radon transform of f . The role of the physical law modelled by a

mapping T is now played by the transform R, function f is the “cause,” and Rf
is the “effect.” The inverse problem consists in solving the equation Rf = g for an
unknown f when g is known. An explicit inversion formula was found in 1917 by
the Austrian mathematician Johann Radon:

π 1 d
1 ds g(ϕ, s)
f (x) = dsdϕ, x ∈ D. (1.6)
2π 2 x·θ −s
0 −1

The appearance of derivatives in (1.6) is a hint that f depends sensitively on g, as it

was the case in Example 1.1. ♦
The next example has the form of an integral equation. Consider a relation
T (u) = w between cause and effect u, w : [a, b] → R, which is given in the form

b
k(s, t, u(t)) dt = w(s), a ≤ s ≤ b, (1.7)
a

with k : [a, b]2 × R → R a known, so-called kernel function. Equation (1.7) is

called Fredholm integral equation of the first kind. A special case is the linear
Fredholm integral equation of the first kind having the form

b
k(s, t)u(t) dt = w(s), a ≤ s ≤ b, (1.8)
a
6 1 Characterization of Inverse Problems

where again k : [a, b]2 → R is given. If this kernel function has the special property

k(s, t) = 0 for t > s,

then (1.8) can be written in the form

s
k(s, t)u(t) dt = w(s), a ≤ s ≤ b, (1.9)
a

and this equation is called Volterra integral equation of the first kind. Another
special case are kernel functions with the property

k(s, t) = k(s − t).

In this case, a linear Fredholm integral equation is called convolutional equation

and can be written as

b
k(s − t)u(t) dt = w(s), a ≤ s ≤ b. (1.10)
a

One speaks of Fredholm or Volterra integral equations of the second kind, if the
function u also appears outside the integral. An example is

s
u(s) + λ k(s, t)u(t) dt = w(s), a ≤ s ≤ b, λ ∈ R.
a

Linear Fredholm equations of the first and second kind do have quite different
properties. Technical details are given in [Eng97], Corollary 2.40. We only
provide an informal explanation, as follows. If the kernel function k is “smooth”
(continuous, for example), then the mapping u → w defined by integral equations
of the first kind also has a smoothing property. Thus, the solution of an integral
equation, which means an inversion of integration, necessarily has to roughen the
right-hand side, thereby also amplifying errors of approximations w̃ to w. In fact
the computation of derivatives, which was found to be problematic in Example 1.1,
can be interpreted as solving a Volterra integral equation:

t

u(t) = w (t), w(t0 ) = 0 ⇐⇒ w(t) = u(s) ds .
t0

In contrast, integral equations of the second kind also contain an unsmoothed copy
of u. The solution of such equations thus does not necessarily have to roughen a
1.1 Examples of Inverse Problems 7

given right-hand side. Fredholm equations are also defined for functions having
multiple arguments. In this case, the integration interval [a, b] generalizes to a
compact subset of Rs , s ∈ N, see the following example.
Example 1.3 (Inverse Gravimetry) Let D ⊂ R3 be a bounded domain, i.e., an open,
connected set, and let S := D be its closure (consisting of all interior and boundary
points of D). Let a body B occupy the compact region S ⊂ R3 in space and have
a mass density given by a function ρ : S → R, x → ρ(x) ≥ 0. The gravitational
potential of B at any point x ∈ R3 \ S is defined by the volume integral

ρ(y)
V (x) = −G dy. (1.11)
x − y2
S

Here, x − y2 is the Euclidean distance between x and y, G is the gravitational

constant, and S and ρ are assumed to be regular enough to make (1.11) a
properly defined (Lebesgue) integral. Equation (1.11) is a convolutional equation
for functions of three independent variables. The gravitational force exercised by
B on a unit mass located in x ∈ R3 \ S is given by the negative gradient of the
potential: F (x) = −∇V (x).
It is unrealistic to assume that F can be measured everywhere outside B. We
consider the following situation. Let S ⊂ Ω ⊂ R3 , where Ω is a convex domain
with sufficiently smooth boundary (a ball or a half-space would do), let Γ ⊂ ∂Ω
contain an interior (with respect to ∂Ω) point, and assume that the magnitude of the
gravitational force induced by B can be measured on Γ , as illustrated by Fig. 1.3.
We thus assume that

g : Γ → R, x → ∇V (x)2 (1.12)

can be observed. Then, (1.11) and (1.12) define a mapping of the form

T :ρ→g (1.13)

(a precise definition of this mapping from one function space into another will only
be given in Sect. 1.3). The density function ρ is interpreted as being the cause of the
observed effect g. The inverse problem of gravimetry consists in finding ρ such that
T (ρ) = g, where g is given.
It can be shown that under appropriate conditions on S, Ω, Γ , and ρ, two
potentials V1 and V2 necessarily coincide everywhere on R3 \ S, if ∇V1 (x)2 =
∇V2 (x)2 for all x ∈ Γ (see Lemma 2.1.1 in [Isa90]). In this sense it is sufficient
to observe ∇V (x)2 for x ∈ Γ . It is also known, however, that even complete
knowledge of V on R3 \ S is not enough to uniquely determine ρ. To achieve
uniqueness, restrictions beyond Eq. (1.11) have to be imposed on ρ. For example,
within the set
8 1 Characterization of Inverse Problems

Fig. 1.3 Gravitational force

exercised by a body,
measured on a surface

⎧ ⎫
⎨  ⎬
L := ρ : S → [0, ∞); ρ solves (1.11) and ρ(x)2 dx < ∞
⎩ ⎭
S

(it is to be understood that all functions ρ ∈ L are assumed to be regular enough for
all involved integrals to make sense), there exists a unique mass density distribution
ρh , such that
 
ρh (x) dx ≤ ρ(x)2 dx for all ρ ∈ L .
2

S S

This function ρh is harmonic, i.e., the Laplace equation Δρh (x) = 0 holds for all x
in the interior of S, see, e.g., the survey article [MF08]. By the maximum principle
for harmonic functions, this means that ρh takes its maximal values at the surface
of S, which might be undesirable. For example, the mass density distribution of the
earth’s interior does not have maximal values at the surface. There may be situations,
however, where we are only interested in knowing any ρ satisfying (1.11)—for
example if the real goal is to construct V (x) for all x ∈ S from observed data (1.12).
This could in principle be done by finding ρh and then using (1.11) to compute V .
Another situation where uniqueness holds, and which is of evident interest in
geological prospecting, is the one illustrated by Fig. 1.4, where a body of constant
mass density is included in another body, also having constant mass density.2 One

2 Actually, Fig. 1.4 shows a two-dimensional plane section of a three-dimensional scenery. Thus the

three-dimensional body B and its inclusion appear as two-dimensional regions. The corresponding
section of the boundary part Γ , where gravitational forces are measured, is shown as a fat line.
1.1 Examples of Inverse Problems 9

Fig. 1.4 Sections of equipotential surfaces of a body with inhomogeneous mass density

might think of the task to detect a subterranean ore deposit. In certain cases it is
possible to determine the exact shape and location of the hidden body. This is the
case if both mass densities are known and if the inclusion is convex “along one
dimension.” A set S ⊂ R3 is xj -convex for some j ∈ {1, 2, 3} and thus “convex
along one dimension,” if its intersection with any straight line parallel to the xj -
axis is an interval (convex sets are x1 -convex, x2 -convex, and x3 -convex at the same
time). This uniqueness result goes back to the Russian mathematician Novikov and
the year 1938. Technical details are described in [Isa90], see for example Theorem
3.1.4. We will consider the problem of finding a hidden body when formulating
a model problem for inverse gravimetry in Sect. 1.3. Of course, the assumption of
constant mass densities can only be approximately true in reality. ♦
Example 1.4 (Full-Waveform Seismic Inversion) The goal of full-waveform seis-
mic inversion (FWI)3 is to obtain information about the earth’s subsurface material
parameters. To this end, seismic waves are generated at or near the surface and
the pressure or the velocity field of the reflected waves is measured, usually on a
part of the earth’s surface or in the ocean. Figure 1.5 schematically illustrates a
marine seismic experiment. An acoustic source, towed behind a survey ship, creates

Sections of equipotential surfaces, where the potential V is constant, are also shown. The gradient
∇V (x) at a point x is orthogonal to the equipotential surface through x.
3 Also: waveform inversion, for short.
10 1 Characterization of Inverse Problems

Fig. 1.5 Marine seismic

experiment hydrophone streamer
acoustic source
sea

layer 1
layer 2
layer 3

layer N

waves traveling into the subsurface. In the simplified situation shown in Fig. 1.5, the
medium is composed of plane homogeneous layers. Within each layer as well as in
sea water, waves propagate at a constant speed. At the interface between two layers,
a traveling wave is transmitted in part and reflected in part. Hydrophones record the
pressure field of upcoming reflected waves below the sea surface. Note that despite
the simplicity of the model, multiple reflections are considered.
Mathematically, the propagation of seismic waves in a domain Ω ⊂ R3 is
modelled by the elastic wave equation, see, e.g., [Sch07] or [Sym09]. We will
exclusively use a simpler mathematical model, the acoustic wave equation, refer-
ring to [Sym09] for a discussion of the conditions under which this simplification
is adequate. Acoustic waves consist of small localized material displacements,
which propagate through the medium and lead to deviations from time-independent
equilibrium pressure. The displacement at location x ∈ Ω and at time t ≥ 0 is
a vector u(x, t) ∈ R3 , defining a vector field u : Ω × [0, ∞[→ R3 . The sound
pressure, which is the deviation from equilibrium pressure, at time t and location
x ∈ Ω is a scalar p(x, t) ∈ R, defining a scalar function p : Ω × [0, ∞[→ R. Both
are related by Hooke’s law

p = −κ∇ · u + S, (1.14)

where κ = κ(x) is a material parameter, called elasticity coefficient or bulk

modulus and where
∂u ∂u ∂u
∇ · u := + + = div(u)
∂x1 ∂x2 ∂x3

is a convenient notation for the divergence of u taken only with respect to x. Further,
S = S(x, t) is a time and space dependent “source term.” This could be a source
1.1 Examples of Inverse Problems 11

injecting material into the medium and thus exciting the wave, like an air gun
in the marine seismic experiment. A second relation between sound pressure and
displacement is given by Newton’s law

∂ 2u
∇p = −ρ , (1.15)
∂t 2
where the gradient ∇p again is to be taken only with respect to the space variable x
and where ρ = ρ(x) is another material parameter, the mass density. In the acoustic
model, knowing Ω, ρ, and κ means to completely know the medium. From (1.14)
and (1.15) one can immediately derive the (scalar) acoustic wave equation for p,
reading
 
1 ∂ 2p 1 1 ∂ 2S
−∇ · ∇p = , (1.16)
ρc2 ∂t 2 ρ κ ∂t 2

where we have introduced the wave velocity

κ(x)
c = c(x) = . (1.17)
ρ(x)

The value c(x) is the acoustic wave’s propagation speed at x ∈ Ω. The differential
equation (1.16) is complemented with initial conditions

∂p
p(x, 0) = 0, (x, 0) = 0, x ∈ Ω, (1.18)
∂t
and boundary conditions

p(x, t) = 0, x ∈ ∂Ω, t > 0. (1.19)

Under appropriate assumptions concerning ρ, κ, Ω, and S, one can prove that a

unique solution p of (1.16), (1.18), and (1.19) exists. This means that there is a
mapping F : (ρ, κ) → p. Hydrophones measure the sound pressure p(x, t) on a
subset M ⊂ Ω × [0, ∞[, so we have another mapping

T : (ρ, κ) → p|M

with p|M the restriction of p to the subset M. A precise definition of a variant of

this mapping will be given in Sect. 1.4. The inverse problem takes the usual form of
finding the cause (ρ, κ) responsible for an observed effect p|M . This is an example
of a parameter identification problem for a partial differential equation.
Similar to the situation in inverse gravimetry, it is known that T in general is
not one-to-one, i.e., ρ and κ cannot be determined from p|M . In Sect. 1.4 we will
12 1 Characterization of Inverse Problems

consider a much simplified situation, which is to determine κ alone, assuming ρ to

be known. ♦

1.2 Ill-Posed Problems

We present Hadamard’s definition of ill-posedness, formulating it specifically

for identification problems, i.e., in the context of equation solving. “Solving an
equation” has to be understood in a broad sense, including equations in function
spaces, like differential or integral equations. Function spaces and mappings from
one function space into another are studied in functional analysis. In the following,
we do not need deep results from functional analysis. We only make use of some
of its formalism to precisely state the essential difficulties coming up with inverse
problems. A condensed introduction into function spaces is given in Appendix B.

Definition and Practical Significance of Ill-Posedness

The following definition refers to general normed spaces (X,  • X ), as for-

mally defined in Appendix B. For example, X = Rn is the Euclidean space,
which can be equipped with the norm  • X =  • 2 defined by x2 =

x12 + . . . + xn2 for every x ∈ Rn . As another example, X = C[a, b] =
{f : [a, b] → R; f continuous} is the space of all continuous mappings from
[a, b] ⊂ R into R, which can be equipped with the maximum norm  • X =
 • C[a,b] defined by f C[a,b] = max{|f (t); t ∈ [a, b]} for every f ∈ C[a, b].
Definition 1.5 (Ill-Posed Inverse Problem) Let (X,  • X ) and let (Y,  • Y ) be
normed linear spaces and let

T :U⊆X→W⊆Y

be a given function. Consider the inverse problem of solving the equation

T (u) = w, u ∈ U, w ∈ W, (1.20)

for u, when w is given. This problem is called well-posed (in the sense of
Hadamard), or equivalently is called properly posed, if
(1) for every w ∈ W a solution u ∈ U does exist (condition of existence),
(2) the solution is unique (condition of uniqueness), and
(3) the inverse function T −1 : W → U is continuous (condition of stability).
Otherwise, problem (1.20) is called ill-posed, or equivalently is called improperly
posed.
1.2 Ill-Posed Problems 13

Properties (1) and (2) are formally equivalent to the existence of an inverse function
T −1 : W → U. This means that for every effect w ∈ W a unique cause u = T −1 (w)
can be identified. Continuity of T −1 means

lim wn − wY = 0 ⇒ lim T −1 (wn ) − T −1 (w)X = 0. (1.21)

n→∞ n→∞

The significance of this stability condition lies in the following: the solution u∗ =
T −1 (w) of T (u) = w can be approximated arbitrarily well by the solution ũ =
T −1 (w̃) of T (u) = w̃, if w̃ ∈ W approximates w arbitrarily well. If the stability
condition is violated, we cannot hope that ũ tells us anything about u∗ , no matter
how much effort we undertake to get a good approximation w̃ of w.
Example 1.6 Computing the inverse Fourier transform of a known square integrable
function is a well-posed problem. Plancherel’s Theorem (C.6) shows that the
condition of stability is met. ♦
Example 1.7 (Determination of Growth Rates, Part 2) In Example 1.1 the vector
space X = U = C[t0 , t1 ] of continuous functions had been equipped with the norm
 • X :=  • C[t0 ,t1 ] , defined by vC[t0 ,t1 ] := max{|v(t)|; t ∈ [t0 , t1 ]}. The space
Y = C 1 [t0 , t1 ] of continuously differentiable functions had been equipped with
the same norm  • Y =  • C[t0 ,t1 ] . We also had used W := {w ∈ Y ; w(t) >
0 for t0 ≤ t ≤ t1 }. Direct and inverse problem are determined by the mapping

t
T : U → W, u → w, w(t) = w0 e U (t)
, U (t) = u(s) ds.
t0

Formula (1.3) shows that a unique solution of the inverse problem exists for every
w ∈ W. But the inverse operator4 T −1 is not continuous, as was already observed in
Example 1.1: for the sequence (wn )n∈N defined in (1.4) and for w(t) = exp(sin(t))
we had seen

lim wn − wY = 0, but lim T −1 (wn ) − T −1 (w)X = ∞.

n→∞ n→∞

Consequently, the determination of growth rates is an ill-posed problem, since the

condition of stability is violated. ♦
In practice, already the existence condition can be problematic, because one usually
only disposes of an approximation w̃ of the true effect w and there is no guarantee
that w̃ ∈ T (U). In the presence of modelling errors, it would even be possible
that w ∈ T (U). This will happen if T is a simplified mathematical description of
an existing physical law, such that not even the exact effect w can be described
in the form T (u) = w. However, we will not consider this case. Rather, we will

4 The term “operator” is used synonymously to “function,” see Appendix B.

14 1 Characterization of Inverse Problems

always assume w ∈ T (U), tacitly adding an eventual modelling error to the data
error w̃ − w. Since we investigate identification problems, we cannot concede the
uniqueness condition. If it was violated, then it would not be possible to identify a
specific cause u responsible for the effect w. In such a case we would either have
to observe more or other kinds of effects or to somehow restrict the set of possible
causes in order to restore uniqueness. The latter was done in Example 1.3 for inverse
gravimetry and will also be done for waveform inversion. Taking the uniqueness
condition for granted, one could approximately solve an inverse problem by first
finding a “best” approximation ŵ ∈ T (U) of w̃, as defined, for example, by the
projection theorem (see Theorem B.12), if T (U) is closed and convex and if Y is
a Hilbert space, and by then finding the unique element û ∈ U with T (û) = ŵ.
Whether this approach produces a meaningful approximation û of the sought-after
solution depends on the stability condition. Violation of the stability condition is the
remaining problem and in fact is a major concern when dealing with identification
problems.

Amendments to Definition 1.5

Definition 1.5 is academic and depends on a meaningful choice of U, W,  • X ,

and  • Y to be useful. These choices must reflect the characteristics of the actual
problem to be solved.
First of all, stability means continuity of the operator T −1 . Continuity, as defined
in (1.21), depends on the chosen norms  • X and  • Y . It is well possible that an
operator is continuous with respect to one norm, but discontinuous with respect to
another. If  • X and | • |X are two norms defined on a linear space X and if there
is a constant C > 0 with

xX ≤ C|x|X for all x ∈ X,

then | • |X is called stronger than  • X and  • X is called weaker than | • |X ,

since convergence of a sequence (xn )n∈N with respect to | • |X entails convergence
with respect to  • X , but not vice versa. Consequently, if one replaces the norm
 • Y on Y by a stronger norm | • |Y , then there are less sequences (wn )n∈N
converging to w and thus there are less sequences for which the implication in
(1.21) must hold. The condition of stability imposed on T −1 is thus weakened
and a discontinuous operator T −1 may artificially be turned into a continuous
one. Likewise, if the norm  • X on X is replaced by a weaker norm, then the
implication in (1.21) is easier to fulfill. Again, this weakens the stability condition
imposed on T −1 .
Example 1.8 (Determination of Growth Rates, Part 3) On the linear space Y =
C 1 [t0 , t1 ], we can replace the norm  • C[t0 ,t1 ] by the stronger one  • C 1 [t0 ,t1 ] ,
defined by f C 1 [t0 ,t1 ] = f C[t0 ,t1 ] + f  C[t0 ,t1 ] . The sequence (wn )n∈N from
1.2 Ill-Posed Problems 15

(1.4) is no longer convergent with respect to  • C 1 [t0 ,t1 ] and can no longer furnish
an example for the discontinuity of T −1 . Quite to the contrary, T −1 is continuous
with respect to the norms  • C 1 [t0 ,t1 ] on Y and  • C[t0 ,t1 ] on X, compare
Proposition B.9. We have turned an ill-posed inverse problem into a well-posed
one! ♦
Enforcing stability by changing norms is a mathematical trick, which is not helpful
in practice. In Example 1.1 we may observe w(t), but we cannot observe w (t), the
determination of which is the essence of the inverse problem. If we could somehow
observe w  directly, the inverse problem would be trivial. Its difficulty only comes up
since w  cannot be observed and thus we cannot hope to find some approximation w̃
being close to w with respect to the norm  • C 1 [t0 ,t1 ] . Instead, (noisy) observations
might get us a function w̃ which approximates well the function values of w, but at
the same time w̃  might approximate badly the values of w  .
A second warning concerns the dependence of Definition 1.5 on the choice
of U and W. These sets must be chosen appropriately for the application under
consideration. For example, arguing that one is only interested in solving the
problem T (u) = w0 for a single, specific effect w0 , one could be tempted to define
W = {w0 }, a set containing only this single element. Continuity of T −1 then would
become trivially true. But such an artificial choice of W does not reflect the fact that
w0 never is exactly known in practice. W will have to contain some neighborhood
of w0 , if the mathematical model of the inverse problem is to be meaningful. If one
was in fact only interested in the solution for a specific right-hand side w0 ∈ W,
then the definition of well-posedness should be changed. One would still require
existence and uniqueness of an element u0 ∈ U such that T (u0 ) = w0 , but the
stability condition would have to be modified. One should demand the existence of
r > 0 such that for any sequence (un )n∈N ⊆ U with u0 − un X < r for all n ∈ N,
the following implication holds:
n→∞ n→∞
T (un ) − T (u0 )Y −→ 0 ⇒ un − u0 X −→ 0.

In this case, the inverse problem is called locally well-posed in w0 .

A third objection to Definition 1.5 again concerns the stability condition.
Continuity of the inverse T −1 just tells us that we can compute u to arbitrary
precision, if w is known to arbitrary precision. In practice we only know w to some
finite precision, i.e., we have at hand some w̃ and hopefully know some finite ε
such that w̃ − wY ≤ ε. Then we would like to estimate how close T −1 (w̃) is
to T −1 (w), i.e., we would like to give a bound for T −1 (w̃) − T −1 (w)X . Such
bounds will be given in Sect. 3.2, where we introduce so-called condition numbers
as a quantitative measure of ill-posedness for linear least squares problems in finite
dimensions. The same could be done for general operators in vector spaces by means
of “linearization,” but this would require us to introduce operator derivatives, so-
called Fréchet derivatives.
Another random document with
no related content on Scribd:
sweet and lovely in English womanhood would soon be his to have and to
hold for ever.

Her eyes, large, brown, and true, were fixed steadfastly on him, and
found no less pleasure in what they saw than his did. In his evening dress
Ralph looked taller than the six feet that he actually measured; fair hair
curled crisply over a sun-tanned face, in every line of which frank candour
was written, and his athletic figure was graceful in every involuntary pose.
Gwendolen had reason to be proud of her lover as he thus stood silhouetted
against the moonlit sky, and she made no secret of it to herself that she
found pleasure in his unconscious show of great strength in restraint. He
could kill her with so little effort of those well-shaped, nervous hands, and
yet one look from her could make his whole frame tremble.

So in silence they communed together, as is the way with lovers who

know that no words can express a tithe of their deep emotion. And, indeed,
while lovers have eyes to see, they do not need tongues to speak. Silence is
best when two hearts are in accord.

The silence was broken by Sir Geoffrey's voice talking to Mrs. Austen
as they came over the velvety turf. Sir Geoffrey helped his companion on to
the houseboat and followed her up the stairway.

"Forgive us for being so long, Gwendolen," he said in his cheery,

bantering fashion. "I hope my nephew has been doing his best to entertain
you."

"He has been behaving very nicely," Gwendolen replied, "and I think
you have brought him up very well."

"I told Martin to bring us some coffee and liqueurs," Sir Geoffrey went
on, "and I'm going to smoke, if you ladies will allow me, and look at the
reflections in the water, and fancy I'm young again."

Mrs. Austen protested.

"You are young, until you feel old," she said, "and you don't feel that to-
night."
 "No, I don't," said Sir Geoffrey stoutly. "This is an ideal ending to one
of the happiest days of my life, and if a man is only as old as he feels, I
shall come of age on Ralph's wedding-day." He lighted a cigar and flung the
match into the river. "Have a cigar, Ralph? I'm sure you have earned it."
The old fellow was pleased that his nephew could not chime in with his
trivial chatter, and pulling up a chair by Gwendolen's side, he patted her
hand. "Happy, Gwen?" he asked, and as the answering smile dawned in the
girl's dark eyes, he wiped his own, which suddenly grew misty. "That's
right, that's right," he said quickly. "Ah! here is Martin with the tray."

Allured by the material pleasures of tobacco and liqueurs, Ralph

descended to earth again, and soon the little party were laughing and
chatting merrily enough. Soft strains of music from another houseboat were
carried down to them, and presently a young fellow poled a racing punt
swiftly down the stream; two swans floated out from underneath the trees,
rocking gracefully on the water ruffled by the punt; and from the tender
came suggestively domestic sounds as the old butler put away the cups and
saucers and decanted whiskey for the men.

Then presently they strolled back to the manor house and lingered for a
little in the hall; and while Ralph took his time to bid Gwendolen good-
night, Sir Geoffrey found opportunity to say a few more words to Mrs.
Austen.

"I wish I could tell you how happy I am," he said. "I have hoped for this
all my life, and now it has come to pass. They both are worthy of each
other, and to see such happiness as theirs is almost as good as having it
oneself."

Mrs. Austen cordially agreed, but she wondered if Sir Geoffrey's hearty
words were at all belied by the sigh that accompanied them. Yet she stifled
the suspicion as it was born, for no woman lives long enough to give her
child in marriage without learning the truth that underlies the words:

"Our sincerest laughter

With some pain is fraught;
Our sweetest songs are those that
tell of saddest thought."
Then, with the curiosity of her sex, she wondered again, as she had so often
wondered before, why Sir Geoffrey Holt himself had never married.

CHAPTER III.

FRAUD.

Rather more than a week elapsed, during which Melville saw practically
nothing of the outer world. His chambers were at the top of the house in
Jermyn Street, the suite consisting of a sitting, bed, and bath rooms, which
he rented furnished for seventy pounds a year. His food and attendance
were all supplied to him by the general manager of the house, and his credit
for these bare necessaries of life was still good. So Melville gave orders to
the hall porter to reply uniformly to all enquirers that he was not at home,
and remained in his chambers steeped in dull melancholy. One evening he
stole out and pawned his violin, but that very night he lost nearly all the
proceeds of the transaction in some utterly foolish wager, and the next
morning he woke up face to face with the fact that he only possessed ten
shillings in the world. It was pouring with rain and the wind was howling
round the balustrade outside his windows. Melville shivered; he felt cold
and ill, and recollected that he had eaten no dinner the night before. He rang
the bell and told the valet, whose services he shared with the other tenants
on his floor, to bring him up some breakfast and some shaving water.

"What is the time?" he asked curtly, as the man came from his bedroom
to say the shaving water was ready.

"About twelve, sir. I will bring up your breakfast in a quarter of an

hour."

Melville turned to the window again. If only the rain would stop! And
how he missed his violin! No human being could realise what his
instrument had been to him, or what a wrench it had been to part with it. He
felt utterly destitute.

"What am I to do?" he muttered vainly. "Sir Geoffrey—no, it's worse

than useless to apply to him—last time was the last time, unless some
marvellous inspiration helps me to pitch some plausible yarn."

While he was still harping on the one perpetual theme, the valet
returned with his breakfast, and Melville drank some tea and disposed of
some excellent kidneys.

"I was getting quite faint," he said to the man who was attending to him.
"Don't bother about things this morning. I shall go out presently, and you
can do whatever you've got to do then."

"It's a very wet day, sir," the other answered.

"Wet?" said Melville disgustedly. "I should think it is wet. The weather
certainly means business." He drank some more tea and lighted a cigarette.
"By the way, put out my dress clothes early this evening. I probably shall
not be dining at home."

The valet hesitated.

"Have you any more linen in any other portmanteau, sir?" he asked.

"I'm sure I don't know," Melville replied testily. "You'd better look and
see. Anyhow, find some."

The valet looked still more uncomfortable.

"I sent all I could find to the wash, sir," he stammered; "and the laundry
people have refused to leave any clean linen until your account is settled."

Melville grew scarlet with anger.

"What do I owe them?" he asked.

"It's a little over four pounds, sir. Will you write a cheque?"
 "No, I won't," said Melville shortly. "Go to the Burlington Arcade and
tell my hosiers to send me over three dozen, and put them down in my
account."

"Yes, sir," said the valet civilly, and left the room.

Melville laughed when the door closed behind the servant. When the
devil laughs it is time for good folks to beware, and Melville felt like a
fiend at that moment. It was grotesquely funny that he could get three dozen
shirts on credit, but had not the money to pay his washerwoman. But the
fact was a staggering reminder of his real position. He got up preparatory to
going out, when he remembered that he had still to shave; he went,
therefore, into his bedroom, and, having stropped his razor, took off his
collar and tie and began to make a lather for his face.

And then suddenly the idea came to him with the force of a conviction
that the way out of his trouble lay plain before him. It was the cowardly
way which it yet requires a measure of courage to take. Death was the
solution of the problem. He did not know how to live, but it was very
simple to die. He sat down in a chair and, almost closing his eyes, peered at
his reflection in the mirror. Very little paler—only with eyes quite closed—
he would not look very different presently if he did this thing. And, unless
his courage failed him in the act, it would not hurt. Then what would
happen? The scene here, in this room, with the dead body stretched upon
the floor, was easy to imagine. It might not be very appalling. Had he ever
contemplated such a deed before he would have provided himself with
some poison, which, while it was as fatal as the razor blade, would not
disfigure him; for to the living man the idea of being disfigured after death
is always repugnant. But he had no poison, and here was the razor ready to
his hand. He would be found quite soon—but it must not be too soon—and
he rose and stealthily locked the outer door.

Again he sat before the impassive mirror. There was no one who would
care. In all the world, so far as Melville knew, there was no one who would
care if he were dead, only a few who would resent the manner of his dying.

He had nothing left to lose. Penniless and friendless in the present,

bankrupt of hope for the future, he had nothing material left to lose, at any
rate, and he stood to gain emancipation from an environment to which he
had ceased to be adapted.

He would have to draw the blade across his throat—so! He must do it

very strongly, very swiftly, or he would fail.

The man leant forward on the dressing-table and gazed closely at

himself in the glass; he saw exactly where he must make the gash, and
without any hesitation or nervousness he felt the edge of the razor with the
thumb of his left hand. As he did so he cut the skin, and some blood fell
upon the snow-white cover of the table. In the extraordinary mental state in
which he was, the horrible incongruity in his reasoning did not strike him,
but, in actual fact, the bloodstain on the cloth gave him offence, and he
paused and looked around him. This—would make such a mess! And there
was a revolver in his bag. How stupid of him not to have remembered that!
It had another advantage, too, for people might think the pistol had gone off
by accident while he was cleaning it, whereas there could be no doubt about
the intention in the other case. It mattered a great deal what people would
think.

He laid down the razor on a chest of drawers and removed the soiled
toilet cover from the dressing table. Then he went to his bag to take out his
revolver. The valet had disarranged the contents of the bag, and Melville
turned over a lot of things and could not find the little pistol case. Instead,
his hand fell upon a heap of letters, and on the top of them was the one that
had come to him from Ralph asking for a loan of a hundred pounds.

A sudden revulsion of thought made Melville sick and giddy. It was as

if a gambler who had lost all but his last five-franc piece had, after
hesitation, staked en plein and followed with a run of wins on single
numbers. One cannot follow up the gambler's line of thought, but many a
one whom that fortune befell would be almost sick to think how narrowly
he missed his chance. Melville was a gambler pure and simple. An instant
before he had been upon the very point of death because he did not know
whence money could be got, and without money he did not want to live. Yet
here in his bag was a letter which might mean at least a hundred pounds. Of
course, he might lose his stake, but to kill himself without having made the
venture was intolerable.
 The physical endurance of the strongest man has its limitations, and
Melville staggered into the sitting-room and threw himself into a great
armchair. Here presently he was discovered by his valet, who was
frightened by his master's complete collapse. Some hours passed by before
he regained anything like his usual self-control, and then, resolutely putting
out of his mind all thought of how close he had been to death, he began to
consider the best time and manner of making his final venture to raise
money.

A train left Waterloo at six-forty, which would land him at Fairbridge

Manor at eight o'clock. If he went by that train he would in all probability
find Sir Geoffrey Holt in a good humour after dinner. He even took the
precaution of changing his clothes again, substituting a somewhat shabby
lounge suit for his elaborate frock coat. "May as well look the part," he said
sardonically to himself. "The Prodigal Son was a bit baggy at the knees, I
imagine, and that is the scene I'm on in now. I shall have to draw on my
imagination about the husks all the same."

There was something almost heroic—in a wicked fashion—in his effort

to pull himself together, for his recent temptation to commit suicide had
really shaken him. He drank freely of the spirits in his tantalus as he was
dressing, and all the while tried to anticipate every difficulty in the
interview before him.

"If only Ralph is out of the way I may pull it off. His letter will serve to
account for one hundred of the last two-fifty, and I can gas about some
forgotten bills to explain how most of the rest has gone. It's a fighting
chance anyhow, and if I fail there is still the pistol."

From one thought sprang another.

"There is still the pistol!"

With a curiously furtive action Melville took the revolver from his
portmanteau and slipped it into his pocket. Then he crept downstairs, and,
hailing a hansom, drove to Waterloo.
 But when the train steamed into Fairbridge Station, Melville was not in
it. He was so restless that he could not endure the swaying of the carriage,
and getting out two stations short of his destination he resolved to walk the
rest of the way.

Leaving the high road he made his way down to the river and followed
the towing path. It was getting dark, but the rain had ceased; the silence was
intense, and the occasional splash of a water-rat startled him so much that
he was angry with himself for being in so highly strung and nervous a
condition.

When at last he reached the gardens of the Manor House he was feeling
very shaky; he walked quickly towards the house, wondering, now the
moment was at hand, how he should begin.

"Ralph may be cornered for money," he muttered, "but I notice he hasn't

got rid of his houseboat. I wonder whether he is here to-night. Everything
depends upon that."

He crept cautiously up to the dining-room windows and tried to peep

through the blinds. As he did so he heard the front door open, and crouching
down hid himself in some shrubbery. He recognised Sir Geoffrey's firm,
quick step, and peering over the laurels saw his uncle walking with Ralph
down the drive. He watched them shake hands, and saw Ralph walk briskly
away; then he drew back among the laurels as Sir Geoffrey returned to the
house and quietly closed the door.

"So Ralph is here to-night," said Melville under his breath; "my luck
again!"

He felt horribly uncertain what to do. His first impulse was to follow
Ralph, who might be going up to town, but he refrained, and walked softly
down to the towing path again, turning round at every other step to see if
Sir Geoffrey were coming. The evening grew colder, and Melville turned up
the collar of his coat and stood back among the shadows, steadying himself
against a tree.
 "Perhaps that is Sir Geoffrey," he thought, as the sound of footsteps fell
upon his ear. "No! it's someone going the other way. 'Pon my word, I'm
beginning to feel quite guilty. Still—I'm not going back without seeing him.
Perhaps I'd better go up to the house and get it over. Why can't he come
down here as usual?"

He retraced his steps, and as he reached the garden gate came face to
face with Sir Geoffrey, who, apparently, was not at all surprised to see him.

"How do you do, uncle?" Melville said. "I thought I would look you
up."

"Very good of you, I'm sure," said Sir Geoffrey drily. "Have you just
come from the station?"

"Yes, just this instant," Melville answered, without thinking.

"H'm!" said Sir Geoffrey; "I suppose they must have put on a new down
train. Did you meet Ralph?"

"No," said Melville shortly.

"H'm!" said Sir Geoffrey again; "I thought not."

"Not a very promising beginning!" said Melville to himself; then he

added aloud, "Is Ralph staying with you?"

"He has been," said Sir Geoffrey, "and he's coming back to-morrow, so I
am sorry I cannot offer you his room."

Melville was annoyed.

"I am not aware that I have asked you to give me his room, and I am
aware that you prefer his company."

"That being so," said Sir Geoffrey, "it seems to me that you have chosen
a somewhat unconventional hour for your visit."
 "I've only just returned to England," Melville replied; "otherwise I
should have called earlier."

"May I ask the object of your visit now you have called?" enquired Sir
Geoffrey. "What is it you want?" and he looked keenly at his nephew.

"Well," Melville stammered, "the fact is I wanted to ask you to give me

some more money. I—I——"

"But it's not two months since I gave you two hundred and fifty
pounds," cried Sir Geoffrey. "What on earth have you done with that?"

Melville was at a loss how to begin the explanation he had invented.

"I've been away," he said lamely, "and ill, and—and it's gone."

"I can quite believe it's gone," said his uncle bitterly. "Money melts
before you like pyramids of snow. I wonder you have the face to ask me
again."

Melville flushed. He knew that Sir Geoffrey had detected him in one lie,
and that in his present state of excitement he would only make matters
worse if he faltered in his suddenly improvised story.

"Well, what am I to do?" he asked.

"Do what every other man does," Sir Geoffrey said. "Work, instead of
idling about in the club and playing the fiddle—and the fool."

"But I can't get any work," Melville objected.

"What have you tried to do?"

"Oh, it's no good going into all that."

"I should think not," said Sir Geoffrey with a bitter laugh, "but, anyhow,
I won't help you any more; men of your type never will work while they've
got any relations on whom they can sponge. You give up the fiddle, as a
start."
 "I have," said Melville, "to a pawn-broker."

"Best place for it," grunted Sir Geoffrey unsympathetically. "I'll pay the
interest for you next year, if you'll agree to leave it there."

Melville clenched his fists and walked on in silence for a few yards.

"You don't mind helping Ralph," he said, with a sneer; "he's so different,
isn't he?"

"That is my own affair," Sir Geoffrey said, "but I don't mind saying I've
never had to refuse him, because he has never asked me. He's a thoroughly
fine fellow."

"He's a humbug," said Melville. "He's not above borrowing from me at

all events. As you insist upon knowing what I did with the last two hundred
and fifty I had from you, I will tell you that I gave Ralph a hundred of it."

"I don't believe you," said Sir Geoffrey. "You're a liar, Melville, and I've
proved it."

"Read that," said Melville shortly.

He took Ralph's letter out of his pocket and gave it to his uncle, who
read it in the fading light. A spasm of pain crossed the old man's face, but
he drew himself up with dignity.

"I detected you in one lie, sir," he said, "but I may have made a mistake
about this. If so, I apologise. You did what your brother asked you? Sent
him this hundred pounds?"

Melville met his keen eyes steadily.

"I did. I sent it to him at once."

"How? By cheque?"

"No," said Melville; "in notes—twenty fivers." His wonted effrontery

returned to him. "I can tell you the numbers if you like."
 "Thank you, no," replied Sir Geoffrey. "I'm not proposing to try to trace
the notes now, and Ralph can give me his own explanation of his temporary
embarrassment later. Come to the house and I will repay you for him now."

Melville's heart beat rapidly with excitement. He felt absolutely no

shame at his fraud, no fear of the subsequent inevitable exposure. He had to
get money somehow, and with incredible swiftness it was already almost in
his grasp. They walked in silence to the Manor House. As they passed the
drawing-room windows Melville caught sight of Gwendolen Austen's
figure and involuntarily paused, but Sir Geoffrey noticed the action and
harshly interrupted him.

"Mrs. and Miss Austen are staying here as my guests. As this is purely a
business visit on your part we will, if you please, go to my library," and he
strode along the terrace.

Melville followed him, and turning to the right came to the west front of
the house, on which side lay Sir Geoffrey's private set of rooms. To
Melville, overwrought with excitement as he was, the library with its great
armchairs and well-filled bookcases looked very homelike and comfortable,
but he did not venture to sit down unasked, and Sir Geoffrey pointedly
refrained from everything approaching hospitality. He unlocked a drawer in
his writing table and, taking out his cheque book, filled in a form payable to
Melville for one hundred pounds. Before signing the cheque, he laid down
his pen and looked scrutinisingly at his nephew.

"There are a few things I wish to say to you, Melville," he said very
slowly, "before we finally part, and I beg you to remember them, as they
may prevent any future misunderstanding. For more than thirty years I have
treated you as my son, in spite of endless disappointments at your total
failure to give me any return in consideration or affection. You have always
been utterly selfish, and, as I think, utterly bad. Now I am a rich man, and
you may perhaps argue that I am only anticipating the provision I have
doubtless made for you in my will. Please understand that that is not the
case. Over and above the just expenses of your life up to now you have
already had from me many thousand pounds, which have been squandered
by you in wanton vice. I do not intend you to have any more. I hold that my
money was given to me for some other purpose than that. In point of fact, I
have not made my will, but when I choose to do so, you will not be a
legatee. You understand perfectly?"

Melville bowed.

"Very good. Now I am giving you this cheque because for once you
have done an unselfish action and have lent your brother two-fifths of what
you had reason to suppose was the last money you would ever receive from
me. I am very, very sorry Ralph asked you for it, but very glad you sent it to
him. I repay you on his behalf, and will see that he in turn repays me."

Sir Geoffrey signed the cheque and gave it to Melville.

"I have left it open so that you may obtain the money in the morning.
This, too, is your property," and he gave him Ralph's letter, which Melville
had forgotten.

Then Sir Geoffrey rose.

"This is a final parting, Melville," he said solemnly, "and I wish to

heaven it were not so. If in these last few weeks I had any reason to hope
you had been trying to be a better man I might have been more harsh to-
night, but not so relentless. But the money I gave you the other day, apart
from this hundred pounds, has gone in gambling as all the rest has gone,
and as everything else I might give you would go. And I declare now, upon
my word of honour as a gentleman, that I hold myself free of you at last.
From whatever you may do in the future to bring shame upon your family I,
in their name, declare we are absolved, and you must look for no more help
or countenance from us. And now I will ask you to go. You can walk to the
station, and will not have long to wait for a train to town."

And opening the French windows on to the lawn, Sir Geoffrey stood
with set lips and stern eyes until his nephew disappeared among the
shrubbery that fringed the drive.

Outside, Melville drew a deep breath.

 "The hysterical old idiot!" he said, half audibly; but his fingers trembled
as he placed the cheque in his inner pocket, and he was more nervous than
he thought himself capable of being. "Still, I've got a hundred pounds, and
as for the row which, I suppose, is bound to follow when the old man finds
out the truth—that can rip for the present. I'm glad he didn't cross the
cheque. There wouldn't have been much change out of it for me if I'd had to
pay it into my account, because I'm so overdrawn, and, what's more, it
might be stopped if Ralph turned up early to-morrow. Gad! I'll go to the
bank at nine."

He stumbled along until he reached the station. He had another stiff

glass of spirits at the refreshment bar, and found he had only a shilling left.

"Good thing I took a return ticket," he muttered, "and as for to-morrow I

can go to the bank in a cab, thank goodness, and go home in a balloon, if I
choose. And after that, I'll clear out of town for a bit and pull myself
together—and pull myself together."

He laughed stupidly as he found himself repeating his words, and then

huddled up in a corner of the carriage. How he got back to his chambers in
Jermyn Street he scarcely knew, but he had been there some time before his
attention was attracted by a letter which was lying on his table. It was
written in a hand that was not familiar to him. It bore date that morning, and
the paper was stamped with a monogram and the address, 5, The Vale,
South Kensington.

"Dear Mr. Melville Ashley," it ran, "there are many reasons—into none
of which do I deem it expedient to enter now—why I have hitherto
refrained from inviting you to my house. For the moment I will confine
myself to making the announcement, for which you may be wholly
unprepared, that I married Sir Geoffrey Holt many years ago, and am,
consequently, your aunt by marriage. I shall be obliged if you will call upon
me to-morrow at half-past four o'clock, and it is my desire that until I have
seen you, you shall not acquaint any third person with the contents of this
communication.—I am, yours faithfully, LAVINIA HOLT."
 At last the full significance of the note was borne in upon him.

"Married Sir Geoffrey many years ago!" Melville said slowly. "Strange!
that is very strange!"

He entered the address in his pocket book, and then carefully locked
away the letter, together with that from Ralph, in a despatch box.

"In spite of all you said, Sir Geoffrey, I fancy this letter, too, may mean
money in my pocket!" and the smile upon his face was very evil.

CHAPTER IV.

MEDIATION.

Breakfast is a period of probation for many people's temper. It is a

comparatively easy matter after dinner in the evening to assume light spirits
with one's evening dress, knowing that the work and worries of the day are
all behind one, but considerable philosophy is required to be entirely
amiable the first thing in the morning, when the same work and worries
have to be taken up anew.

So when, the morning following Melville's surprise visit to the Manor

House, Sir Geoffrey entered the dining-room, Gwendolen's loving eyes
perceived at once that something had occurred to ruffle his equanimity.
With her he was never irritable, but his greeting was absent-minded, and he
seemed to seek in vain for anything to interest him in the columns of the
Times.

Mrs. Austen usually breakfasted in bed, and as Ralph was not to return
until the middle of the day, Sir Geoffrey and Gwendolen were alone
together, and the meal passed almost in silence. At last Sir Geoffrey himself
appeared to become aware of the fact that he was discharging his duties as a
host with something less than his usual success.

"Forgive me, Gwen," he said pleasantly. "I'm an old bear this morning,
and poor company for my beautiful princess."

Gwendolen rose and put her arms round his neck.

"Then if the story books are to be believed, the beautiful Princess only
has to kiss the old bear, and he will be transformed into Prince Charming
again," and leaning over him she kissed him affectionately.

"You're a little witch," said Sir Geoffrey, smiling; "but tell me, aren't
you burning to know what has upset my temper to-day?"

"Not at all," Gwendolen answered quickly, "unless it is anything that I

have done."

"Of course it isn't," said Sir Geoffrey; "but it's the next thing to it. I've
got a bone to pick with Ralph."

Gwendolen's face clouded over.

"Oh! I am sorry," she said, but almost immediately her eyes shone
brightly again. "It can't be very serious, though, because he's sure to have
some perfectly satisfactory explanation for whatever he has done, and as
soon as you see him you'll find there's no bone to pick."

"You're a loyal little woman," said Sir Geoffrey, well pleased, "and I've
no doubt you're right. What time is the immaculate hero to honour us by his
reappearance?"

"About a quarter to one," Gwendolen replied.

"In time for luncheon," Sir Geoffrey remarked. "Whatever one may
think about his other meritorious qualities, there can be no doubt about the
excellence of Ralph's appetite."
 "You're trying to draw me," said Gwendolen cheerfully; "but I won't be
drawn. I like a man to have a good appetite, and, by the way, you're not a
bad trencherman yourself."

Sir Geoffrey laughed.

"I've got some work to do this morning," he said as he got up. "You
must kill the time somehow until Ralph returns, and after luncheon you will
be able to pick water lilies and gaze into each other's eyes to any extent. Are
you going to meet him at the station?"

"I thought of doing so," Gwendolen admitted.

"Did you really!" said Sir Geoffrey, with affected incredulity. "Well, I
don't want to interfere with your plans, but seriously, Gwen, as soon as
you've got over the shock—I mean the rapture—of seeing him again, will
you tell him to come to me in the library?"

"Of course I will," said Gwendolen, "and seriously, too, dear uncle, I'm
sure everything will be cleared up as soon as you see him."

"I daresay it will," Sir Geoffrey agreed, "but I have always believed in
getting to the bottom of things immediately. When you're married, Gwen,
avoid a misunderstanding with your husband as you would avoid the devil.
Quarrel if you must, but, at any rate, know what you're quarrelling about.
That's good advice."

"How can an old bachelor give any good advice about the married
state?" Gwendolen asked lightly, and she nodded gaily as she ran upstairs,
not noticing how the expression altered on Sir Geoffrey's face.

"Blows beneath the heart dealt by those one loves the most," he
muttered sadly. "Well, it's inevitable in this world, I suppose, and, after all,
there's compensation in the love itself. But Ralph ought not to have stooped
to borrow that money from Melville; and what on earth can he have wanted
it for that he was afraid to ask me? That's the sting," and the old gentleman
walked slowly to his library and shut himself in there alone.
 Both to Sir Geoffrey and to Gwendolen the morning seemed to drag, but
at last the train which brought Ralph from town arrived, and, heedless of the
bystanders, Gwendolen kissed her lover and walked down the hill with him
to the river.

"Had a happy morning, dear?" he asked.

"A very long one," Gwendolen replied. "Time is very inconsiderate to

people who are in love; it flies when they are together and halts when they
are alone, whereas, of course, it ought to do exactly the reverse."

"Of course it ought," Ralph assented, "but, anyhow, it's ripping to be

alive. By Jove, Gwen, I think I'm the happiest man in the whole world."

Gwen looked at him critically.

"I'm sure you are the nicest," she said enthusiastically, and did not
demur to his finding her approval an excuse for another kiss.

"Let's go on the houseboat," he said, "and after luncheon I will punt you
up to where the water lilies are."

"Pick water lilies and gaze in each other's eyes," said Gwen, laughing;
"that was the programme Sir Geoffrey mapped out for us. Oh! I forgot. He
asked me to send you to him directly you arrived. He's in the library."

"Can't it keep till after luncheon?" Ralph asked indifferently. "I want to
talk to you."

"No," Gwen replied; "you must go now. I promised that you would. He
said he had a bone to pick with you."

"Did he?" said Ralph. "I wonder what's the matter."

"I don't know," Gwen answered, "but he was very quiet at breakfast, and
I guessed there was something wrong; then he told me it was about you, and
I said you could explain anything you did or didn't do, and you've got to go
at once and do so."
 "A very lucid statement," Ralph said, smiling. "Well, it's a bore to have
to leave you at once, but if you've promised, there's no help for it."

"None," said Gwendolen gravely. "Come along, Ralph."

In her heart she was a little uneasy, for although she had absolute
confidence in Ralph's perfect integrity, she had never before seen Sir
Geoffrey look so troubled at anything in which his favourite nephew was
concerned. But she stifled her not unnatural curiosity, and, leaving Ralph at
the library door, ran off to the room where her mother was writing wholly
unnecessary letters.

Sir Geoffrey was so engrossed in a book that he did not hear Ralph
come into the room. Comfortably ensconced in a huge armchair, with
spectacles on his nose, and the sunlight streaming through the window upon
his silver hair, he embodied the general idea of a cultivated old English
gentleman. Ralph looked at him, and then spoke.

"Gwendolen tells me you want to see me, Uncle Geoffrey, so I've come
straight in."

Sir Geoffrey looked up.

"Yes," he said. "Melville was here last night."

Ralph was vexed, for he knew what was the usual reason for his
brother's visits to Fairbridge.

"Was he?" he said. "I didn't know he was back."

"Then you knew he was going abroad?"

"Oh, yes," said Ralph. "He made no secret about it to me."

Sir Geoffrey only grunted, and Ralph went on.

"In many ways I'm rather sorry for Melville, uncle. Of course, I know
he has been a lot of worry to you, but he's my brother after all, and it isn't
easy to get the sort of work that he could do."
 "He's had a good education," said Sir Geoffrey, "and he's got good
health and a pair of hands. What more does a man need to earn an honest
living?"

Ralph was very happy, and when one is happy it is difficult not to feel
generously disposed even to those one loves the least; so now he
championed his brother quite sincerely.

"I've got all that, too," he said, but Sir Geoffrey put up his hand in
deprecation of any comparison between the two brothers.

"You owe me a hundred pounds, Ralph," he said.

"My dear uncle," Ralph replied. "I owe you a great deal more than that.
I can never repay you a fraction of what I owe you."

Sir Geoffrey's face lighted up with pleasure at the young fellow's frank
expression of gratitude.

"One does not repay free gifts," he answered. "Let all that pass; but,
Ralph, why couldn't you tell me you were in need of ready money?"

"I don't quite understand," said Ralph, looking puzzled.

"A few weeks ago," said Sir Geoffrey, rather testily, "you borrowed it
from Melville, and I repaid him for you last night."

Ralph's face flushed with indignation.

"You paid Melville a hundred pounds for me?"

"Yes," said his uncle.

"But I don't owe him anything."

"You wrote to him at Monte Carlo, and asked him for a hundred pounds.
What did you want that for?"