Ebook Quasi Interpolation Martin Buhmann Online PDF All Chapter

Quasi Interpolation Martin Buhmann
Visit to download the full and correct content document:

https://ebookmeta.com/product/quasi-interpolation-martin-buhmann/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
The Statistical Analysis of Quasi Experiments

Christopher H. Achen
https://ebookmeta.com/product/the-statistical-analysis-of-quasi-
experiments-christopher-h-achen/
Diplomacy of quasi alliance in the Middle East Degang

Sun (Author)
https://ebookmeta.com/product/diplomacy-of-quasi-alliance-in-the-
middle-east-degang-sun-author/
Banach Space Valued Neural Network: Ordinary and

Fractional Approximation and Interpolation 1st Edition
George A. Anastassiou
https://ebookmeta.com/product/banach-space-valued-neural-network-
ordinary-and-fractional-approximation-and-interpolation-1st-
edition-george-a-anastassiou/
Banach Space Valued Neural Network Ordinary and

Fractional Approximation and Interpolation 1st Edition
George A Anastassiou
https://ebookmeta.com/product/banach-space-valued-neural-network-
ordinary-and-fractional-approximation-and-interpolation-1st-
edition-george-a-anastassiou-2/
Lens Design: Automatic and Quasi-Autonomous
Computational Methods and Techniques (IPH001) Donald
Dilworth
https://ebookmeta.com/product/lens-design-automatic-and-quasi-
autonomous-computational-methods-and-techniques-iph001-donald-
dilworth/
Vibration Protection Systems Negative and Quasi Zero

Stiffness 1st Edition Chang-Myung Lee
https://ebookmeta.com/product/vibration-protection-systems-
negative-and-quasi-zero-stiffness-1st-edition-chang-myung-lee/
Spectral Invariants with Bulk Quasi Morphisms and

Lagrangian Floer Theory 1st Edition Kenji Fukaya
https://ebookmeta.com/product/spectral-invariants-with-bulk-
quasi-morphisms-and-lagrangian-floer-theory-1st-edition-kenji-
fukaya/
Lens Design Automatic and Quasi Autonomous

Computational Methods and Techniques 2nd Edition Donald
C. Dilworth
https://ebookmeta.com/product/lens-design-automatic-and-quasi-
autonomous-computational-methods-and-techniques-2nd-edition-
donald-c-dilworth/
High-Order Harmonics Generation in Plasmas: Resonance

Processes, Quasi-Phase-Matching, and Nanostructures
Rashid A. Ganeev
https://ebookmeta.com/product/high-order-harmonics-generation-in-
plasmas-resonance-processes-quasi-phase-matching-and-
nanostructures-rashid-a-ganeev/
CAMBRIDGE MONOGRAPHS ON
A P P L I E D A N D C O M P U TAT I O NA L
MATHEMATICS
Series Editors
M . A B L O W I T Z , S . D AV I S , J . H I N C H ,
A . I S E R L E S , J . O C K E N D O N , P. O LV E R
37 Quasi-Interpolation
The Cambridge Monographs on Applied and Computational Mathematics series reflects the crucial
role of mathematical and computational techniques in contemporary science The series publishes
expositions on all aspects of applicable and numerical mathematics, with an emphasis on new
developments in this fast-moving area of research.
State-of-the-art methods and algorithms as well as modern mathematical descriptions of physical
and mechanical ideas are presented in a manner suited to graduate research students and professionals
alike. Sound pedagogical presentation is a prerequisite. It is intended that books in the series will serve
to inform a new generation of researchers.
A complete list of books in the series can be found at

www.cambridge.org/mathematics.
Recent titles include the following:
23. Curve and surface reconstruction, Tamal K. Dey

24. Learning theory, Felipe Cucker & Ding Xuan Zhou
25. Algebraic geometry and statistical learning theory, Sumio Watanabe
26. A practical guide to the invariant calculus, Elizabeth Louise Mansfield
27. Difference equations by differential equation methods, Peter E. Hydon
28. Multiscale methods for Fredholm integral equations, Zhongying Chen, Charles A. Micchelli &
Yuesheng Xu
29. Partial differential equation methods for image inpainting, Carola-Bibiane Schönlieb
30. Volterra integral equations, Hermann Brunner
31. Symmetry, phase modulation and nonlinear waves, Thomas J. Bridges
32. Multivariate approximation, V. Temlyakov
33. Mathematical modelling of the human cardiovascular system, Alfio Quarteroni, Luca Dede’,
Andrea Manzoni & Christian Vergara
34. Numerical bifurcation analysis of maps, Yuri A. Kuznetsov & Hil G.E. Meijer
35 Operator-adapted wavelets, fast solvers, and numerical homogenization, Houman Owhadi &
Clint Scovel
36. Spaces of measures and their applications to structured population models, Christian Düll,
Piotr Gwiazda, Anna Marciniak-Czochra & Jakub Skrzeczkowski
37. Quasi-interpolation, Martin Buhmann & Janin Jäger
38. The Christoffel–Darboux kernel for data analysis, Jean Bernard Lasserre, Edouard Pauwels &
Mihai Putinar
Quasi-Interpolation
M A RT I N B U H M A N N
Justus Liebig Universität, Giessen, Germany
J A N I N J Ä G E R
Justus-Liebig-Universität, Giessen, Germany
University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314 321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi 110025, India
103 Penang Road, #05–06/07, Visioncrest Commercial, Singapore 238467
Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence
www.cambridge.org
Information on this title: www.cambridge.org/9781107072633
DOI: 10.1017/9781139680523
© Martin Buhmann and Janin Jager 2022
This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2022
A catalogue record for this publication is available from the British Library.
Library of Congress Cataloging-in-Publication Data
Names: Buhmann, M D (Martin Dietrich), 1963 author | Jager, Janin, 1988 author
Title: Quasi-interpolation / Martin D. Buhmann, Justus-Liebig-Universitat Giessen, Germany,
Janin Jager, Justus-Liebig-Universitat Giessen, Germany.
Description: Cambridge, United Kingdom ; New York, NY : Cambridge University Press, 2021. |
Series: Cambridge monographs on applied and computational mathematics ; 37 |
Includes bibliographical references and index.
Identifiers: LCCN 2021044729 | ISBN 9781107072633 (hardback)
Subjects: LCSH: Interpolation.
Classification: LCC QA281 .B94 2021 | DDC 511/.422–dc23/eng/20211006
LC record available at https://lccn.loc.gov/2021044729
ISBN 978-1-107-07263-3 Hardback
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
Dedicated to M. J. D. Powell (1936–2015) in memory of his outstanding contributions to
approximation theory and for his guidance to us.
Contents
Preface page xi
1 Introduction 1
2 Generalities on Quasi-Interpolation 6
2.1 Approximation Properties 9
2.2 Main Examples of Quasi-Interpolants 13
2.3 Some Applications Connected with Different Forms of
Quasi Interpolants 24
3 Univariate RBF Quasi-Interpolants 26

3.1 Univariate Quasi-Interpolants 27
3.2 Examples of Convergence Results 29
3.3 Further Examples 33
3.4 Notes 37
4 Spline Quasi-Interpolants 38
4.1 Spaces of Piecewise Polynomials 38
4.2 General Form of Spline Quasi-Interpolants 41
4.3 Marsden’s Identity and de Boor–Fix Differential
4.4 Approximation Order and Error Estimates 50
4.5 Discrete Quasi Interpolants 56
4.6 Integral Quasi-Interpolants 63
4.7 Local Spline Projectors 66
4.8 Near-Minimally Normed Quasi-Interpolants 67
4.9 Schoenberg’s Quasi Interpolation and Discrete Spline
4.10 Notes 74
viii Contents
5 Quasi-Interpolants for Periodic Functions 76

5.1 Trigonometric Spline Quasi-interpolants 77
5.2 Multiquadric Trigonometric Spline Quasi-Interpolants 90
5.3 De la Vallée Poussin Quasi-Interpolants 93
5.4 Notes 104
6 Multivariate Spline Quasi-Interpolants 106
6.1 Quasi-Interpolants from Tensor Product Splines 107
6.2 Spline Spaces Constructed from Shifts of One Spline 118
6.3 Notes 141
7 Multivariate Quasi-Interpolants: Construction in n Dimensions 143
7.1 Introduction and Further Notation 143
7.2 Polynomial Reproduction 146
7.3 Examples 154
7.4 Convergence Theorems 161
7.5 Quasi Interpolation via Approximate Approximation 165
7.6 Notes 169
8 Quasi-Interpolation on the Sphere 171
8.1 Spherical Quasi Interpolation using Tensor Products 172
8.2 Quasi Interpolation on the Sphere Using Approximate
Fourier Coefficients 176
8.3 Quasi-Interpolants using Shepard’s Method for the Sphere 184
8.4 Notes 186
9 Other Quasi-Interpolants and Wavelets 188
9.1 Prewavelets with Radial Basis Functions and Quasi-Lagrange
Bases 188
9.2 Least-Squares Approximations as Quasi-Interpolants 203
9.3 Weierstrass Operators 215
9.4 Notes 225
10 Special Cases and Applications 226
10.1 Even and Odd Dimensional Spaces and Equally Spaced
Centres 226
10.2 A Look at the Inverse Multiquadric Radial Function in One
Dimension 240
10.3 Applications: Fredholm Integral Equations 248
10.4 Applications: Numerical Solution of Partial Differential
Equations 252
Contents ix
10.5 Compression in the Space of Continuous Functions 259

10.6 Notes 265
References 266
Index 274
Preface
In the range of current applications of numerical mathematics and approximation

theory, it appears that there is a particular demand for techniques admitting a
relatively easy formulation of approximants to data and functions with noise
or that require smoothing. This is especially true when functions and data in
high dimensions are to be dealt with. A conceptually easy way to formulate
approximations is by interpolation, that is, data and approximant should match
precisely at certain points, but this is in contrast to the demand of smoothing (for
instance noisy data, which we may not want to interpolate at all), and of course
there is often absolutely no guarantee that such interpolants exist, notably in more
than one dimension. On the other hand, mere smoothing algorithms may not meet
the accuracy and asymptotic convergence properties we wish to have.
The concept of quasi interpolation stands in the intersection of these two demands
and turns out to be exceptionally useful in a large field of applications, including
the important fields of solving partial differential equations and learning, and it
also has a rich and interesting mathematical theory in its own right. The idea is to
mimic the concept of interpolation by replacing Lagrange functions (which may
not even exist, but quasi-interpolation is almost always possible, in any dimension)
with quasi-Lagrange functions that can be computed in advance. The latter need no
longer match data precisely, but are local and exact on some subspaces (typically
polynomials) by design, in contrast to the classical Lagrange functions, whose
existence has to be ensured first, and only then can one look at whether they
are local, for instance, or give any convergence, etc. Using those quasi-Lagrange
functions one can then proceed to forming approximants in a similar way to standard
interpolation, and study their properties.
It turns out that this opens the door to a rich mathematical theory on function
spaces, approximation methods, their properties, and efficient and stable algorithms.
At the same time, it not only justifies the use of quasi interpolation in applications
xii Preface
but the theory itself then provides further insight into how to develop new methods
with improved properties.
In view of all this, we have tried to summarise a theory (while keeping in mind
relevant applications we ourselves are keenly interested in) of quasi-interpolation
in many different contexts: univariate and multivariate quasi-interpolation, periodic
and non-periodic polynomials, on spheres, piecewise or not, with kernels (often
radial basis function kernels), with and without shift-invariance. We have given
particular emphasis to the approximation (function) spaces that are behind the
quasi-interpolants (since it is crucial to choose a good space before we form the
approximants themselves) and to convergence questions, but readers may of course
have other preferences. For readability we summarise the references to the original
literature, the history of the derivation of the results as well as references to further
results in the ‘Notes’ section at the end of each chapter.
We hope the reader will find this collection of aspects of quasi-interpolation
useful, noting several new ideas of our own, and perhaps take our contribution
as a basis for further developments as well. Of course, we also hope that readers
with specific scientific and engineering applications in mind can take advantage of
our work on quasi-interpolation, for instance when employing some of the explicit
expressions we have for quasi Lagrange functions. No prerequisites are needed for
reading this book other than mathematics related undergraduate studies, so any
graduate students (or later) of mathematics, sciences or engineering should be able
to benefit from this book. We have placed an emphasis on keeping our work self
contained. In our ‘Notes’ sections we include information about further reading.
We could only have written this book on the basis of the enormous contributions of
our colleagues, past and present, for which we are immensely grateful. In particular,
our colleague Paul Sablonnière helped substantially when we began this project.
In fact, not only our book but also our earlier mathematical work, for both of us, in
approximation theory depends in very many respects specifically on the concepts and
theorems about quasi-interpolation that were created by other people from whom we
have learned. Instead of giving a long list of names of those mathematicians, we will
just mention Iso Schoenberg as one of the outstanding fathers of the approximation
theory of quasi-interpolation.
We have discovered many new aspects of quasi-interpolation and approximation
while writing this book; this has given us much pleasure and we hope that some
of this will shine through the text. As ever, it was a pleasure to work together
with our colleagues at Cambridge University Press when writing this book, and we
specifically thank David Tranah for his friendship and advice. We would also like to
express our gratitude to the University of Giessen for providing such an enjoyable
working environment.
Preface xiii
Due to the sheer volume of contributions from all sides, both mathematical and
from practice, we cannot, of course, cover the entire theory and applications of quasi
interpolation. But we hope the reader will decide that our choices of the various
parts are reasonable and provide a contribution that demonstrates the remarkable
usefulness, versatility and interesting theory of quasi interpolation.
1
Introduction
In this chapter we will give the reader a general overview of our themes and
motivations. We will give a brief outline of what functional approximations are for,
especially multivariate ones, what kind of approximation methods are typical, and
which ones we shall study in this book and why.
In many mathematical and scientific applications, approximation of functions of
possibly very many variables (unknowns) is often needed because the theoretically
known function is in fact too complicated to evaluate (especially when it would
have to be evaluated very many times and/or the number of unknowns is very
high) or in fact not known at all in the application except at a few points. Those
could be predefined or not initially available. On top of this, the given data may be
inaccurate or noisy, to a level often estimated in advance. Therefore the question
arises how to approximate univariable or multivariable functions efficiently in such
circumstances.
Interpolation and quasi-interpolation are both highly useful means of approxim-
ating functions and data in multivariable dimensions, say, for the notation in this
book, in Rn . They are useful methods that approximate from spaces spanned for
example by polynomials, piecewise polynomials, trigonometric polynomials and
exponentials or radial basis functions.
Interpolation and generally approximation using radial basis functions are very
good examples since they have become a well-known and appreciated tool for
approximating multivariate functions, especially when the dimensions n are really
large, for which polynomial interpolation and approximation from polynomial
spaces become difficult and very much depend on the geometry of the data. By
contrast, the success of radial basis function approximation is linked in particular
to the available variety of approximations of different kinds from the vector spaces
spanned by the translates of the radial basis functions.
We point out that the most important choice at the beginning of the approximation
procedure is the selection of the space of approximants.
2 Introduction
Important features are its approximation power (highly useful convergence

properties and error estimates are available, as will be seen with a special emphasis
in this book), its applicability in basically any space dimension any number
of variables and its simplicity in formulating and stating the approximation (or
interpolation) problem.
Despite wishing to avoid pointwise collocation and its potential disadvantages,
and therefore using quasi-interpolation in the end, much of the theoretical analysis
began using interpolants from radial basis function spaces. That is, we wish to
meet the approximand (the function f that is being approximated) at a number
(sometimes infinite, sometimes gridded, in practice of course finite) of given points
ξ ∈ Rn .
Returning to the general concepts, during the past few decades quasi-interpolation
has become a particularly popular approach in approximation theory, especially for
smoothing purposes (e.g. when given data are noisy) and in contrast to interpolation
schemes. The approach we take in this book uses a quasi-Lagrange function in
linear combinations with coefficients that are derived from the approximand by
linear operators, most often point evaluation, in such a way that certain low-degree
polynomials are not just approximated but recovered exactly. Often this type of
reproduction is only required for the leading monomial term of the approximation.
On top of this, we do not always require point evaluation of approximands to
form the quasi interpolants, but also admit more general operators applied to the
approximands (e.g. local integrals or derivative evaluations) before they are used to
formulate the quasi-interpolating approximation.
These fundamental ideas can replace pointwise interpolation, and the approach
should lead to approximation orders when sufficiently smooth general functions
are approximated under suitable conditions (such as localness of the said quasi
Lagrange functions).
But now let us return for a moment specifically to the approach using radial basis
functions.
We will speak about more general approaches to quasi-interpolation later on;
this, however, is a good way to start explaining these ideas. This is because we
note another highly important feature of radial basis functions, namely that for large
well-known classes of radial basis functions the interpolation problem is uniquely
solvable, sometimes with no or at most some very easily verifiable conditions on
the mentioned points and their geometry, and thus completely different from the
far more complicated case of multivariable polynomial interpolation, for example,
where the geometry of the interpolation points is of central importance to the
unique solvability of the interpolation problem. This is true, for example, for the
multiquadric or inverse multiquadric radial basis function specified below, which in
many applications and parts of the literature is the prime example of approximation
Introduction 3
with radial basis functions, others being inverse multiquadrics, Gaussian and Poisson
kernels and thin plate splines. We note already at this point that other than in
straight polynomial interpolation, for example, the linear spaces used here to form
the approximants depend on the initial points, also called ‘centres’ as well, because
we take the radial basis functions and shift them by these centres (thus they are
radially symmetric about these points, which explains their name).
Initial examples of radial basis function quasi-interpolants
√ that will turn out to be
useful are the aforementioned multiquadrics φ(r) = r + c for a real parameter c
2 2
that is explicitly allowed to be zero, or thin-plate splines φ(r) = r 2 log r. The

former case is an example (of which there are many) where unique interpolants are
guaranteed when the points are distinct and there are at least two of them. This is
the simplest conceivable condition. There are some easy linear side conditions, and
conditions requiring that the data points do not lie all on a straight line, that give
the same property to thin-plate splines.
So for multiquadrics and any other radial basis function that immediately gives
non-singularity to the interpolation problem with finitely many distinct centres
ξ ∈ Ξ (such as inverse multiquadrics, Gaussian or Poisson kernels, or multiquadrics
with c = 0 ‘linears’; see Buhmann, 2003), we formulate the search for an interpolant
of the type

s(x) = λ ξ φ( x − ξ ), x ∈ Rn,
ξ ∈Ξ
by solving the linear system of equations

λ ξ φ(ζ ξ ) = f (ζ ), ζ ∈ Ξ,
ξ ∈Ξ
with the interpolation matrix
{φ(ζ ξ )}ζ,ξ ∈Ξ
In the case of extra conditions such as with thin-plate splines, this would be the
search for an interpolant

s(x) = λ ξ φ( x ξ ) + p(x), x ∈ Rn,
ξ ∈Ξ
with a polynomial of degree one in the thin-plate spline case and extra conditions

λ ξ q(ξ) = 0
ξ ∈Ξ
for all at most linear polynomials q. Such extra conditions are not unknown for
univariate splines either; even straightforward natural cubic splines
4 Introduction

s(x) = λ ξ φ(|x ξ |) + cx + d, x ∈ R,
ξ ∈Ξ
with φ(r) = r 3 enabling us to express the spline as a radial basis function, and
with real constants c and d, demand side conditions for uniqueness, namely second
derivatives of s vanishing at the first and last knot.
While interpolation is sometimes explicitly preferred or demanded, as it
reproduces the data values at the interpolation points exactly, in many circumstances
quasi-interpolation is to be preferred because it has other reproducing properties,
is easier to compute, and possesses a tendency to smooth data, which is frequently
desired. Quasi-interpolation forms the approximant Q f from the (data) function
(approximand) f by building sums of kernel functions, which we call ‘quasi-
Lagrange functions’ ψ, shifted as ψ(· ξ) or in other ways depending on the data
points ψ ξ ∈ Rn multiplied by the function values f (ξ), their derivatives, integrals
or other linear operators λ ξ f applied to them.
The interpolation property is now replaced by other conditions, typically of ψ or
ψ ξ being from the space spanned by φ( · ξ ) but restricted to locally supported or
decaying functions, and that the quasi-interpolation operator be exact (reproducing)
on certain function spaces, so that approximants are linear combinations of such ψ
functions. For these, polynomials of a fixed maximal total degree in n unknowns
is a typical example. The latter opens the door to powerful convergence results for
sufficiently smooth approximands f by Taylor series or related arguments.
In summary, the difference between interpolation and quasi-interpolation is that
an approximation from a finite- or infinite-dimensional linear space S spanned
by basis functions φ ξ , ξ ∈ Ξ, need not satisfy pointwise interpolation conditions
such as
s(ζ ) = f (ζ ) for all ζ ∈ Ξ,
but the approximants are essentially from the same spaces.

We now have new means to approximate certain approximands well that are
smooth enough, from certain Sobolev spaces, etc. The basis functions are no
longer Lagrange functions but ‘bell-shaped’ quasi-Lagrange functions. Therefore,
in particular, approximations can be computed without solving linear (interpolation
equations) systems of equations. Moreover, it will turn out that convergence
estimates are simpler to achieve with quasi-interpolation when they are employed in
place of interpolation, as in particular there are no complicated estimates of operator
norms (Lebesgue constants).
We have used the elements ξ ∈ Ξ, for the time being, as indices for the basis
functions, but in due course it will be seen that they are inherently related to
the linear spaces spanned by radial basis functions. However, giving up the explicit
Introduction 5
interpolation conditions raises the need for other conditions that fix the approximant
and avoid trivial choices. In particular, we note explicitly that formulating quasi
interpolants via linear combinations

f (ξ)φ( x ξ )
ξ ∈Ξ
is normally useless because the radial basis functions are not at all local with respect
to space (except for the special cases of Gaussian or Poisson kernels or compactly
supported radial basis functions). The above approximation would provide no
locality, and normally approximations to smooth functions that are locally described
by Taylor expansions, for instance, will not be any good. Spline approximations,
however, do give compact support (localness) and some polynomial recovery, but
are much harder to formulate in more than one or two dimensions, partly due to
their piecewise structure. We shall see that, nonetheless, at least in one dimension
quasi-interpolation is related to approximation by piecewise polynomials.
In contrast to the standard approach of interpolation, incidentally, approximations
created by quasi interpolants are normally not unique, which usually has little impact
on practical applications.
It is also often much more difficult to estimate the desired Lebesgue constant
(operator norms) for interpolation, whereas for quasi-interpolants this is usually
straightforward.
We should point out that also other approximation methods from spaces
spanned by radial basis functions are possible, that is, neither interpolation nor
quasi-interpolation but wavelets (Buhmann, 2003), prewavelets, compression or
smoothing splines with generalised cross-validation (Wahba and Wendelberger,
1980), for example.
We address some of these ideas at the end of the book.
Having given the reasons why we consider the approximations by quasi-
interpolation to be useful and interesting, we shall proceed to some general remarks,
univariate approximations by quasi-interpolation, and then come to multivariable
methods.
2
Generalities on Quasi Interpolation
After we have defined quasi-interpolation and clarified the differences from ordinary
interpolation (e.g. polynomial or piecewise polynomial, trigonometric or piecewise
exponential, i.e. spline-based), we wish to study this new method of choice as
an approximation method. This specifically includes searching for conditions on
the distributions of data points, examples of a selection of spline or radial basis
functions and computing approximants. It also includes showing the differences
between using vector spaces created by shifts of ordinary radial basis functions and
those spanned by the aforementioned quasi Lagrange functions.
A first general form of quasi-interpolants is the following, which we will use to
describe the different important parts and properties of a quasi interpolant. This
will emphasise the choices we have in the construction of quasi interpolants.
We let F denote some normed linear space of functions defined on a domain
Ω in Rn (bounded or not). For the purpose of good approximation properties, it
is assumed that F contains a subspace P, which we aim to reconstruct precisely
via our approximation scheme. These subspaces are often linear spaces of algebraic
polynomials (or trigonometric polynomials for periodic functions) that at least
contain the constant functions.
For f belonging to F , we define the quasi-interpolant Q as a linear operator of
the type

Q f (x) λ α ( f )ϕα (x), x ∈ Ω.
α∈Λ
Here Λ is a (finite or infinite) set of indices usually closely connected to the

information about the function that is available for the approximation. This typically
means we use a set of data sites Ξ ⊂ Ω, where f and ϕ are evaluated in some way,
and then it is convenient to use Ξ itself as the index set. In other words we will use
indexing λ ξ ( f )ϕ ξ over ξ ∈ Ξ ⊂ Ω, rather than λ ξα ( f )ϕ ξα over α ∈ Λ, and so on.
Generalities on Quasi-Interpolation 7
The distribution of these data sets whether they are scattered or gridded will
play an important rôle in the further discussion.
The λ α are continuous linear forms (coefficient functionals) defined on F . In
choosing the functionals we decide which information about the approximated
function ( f the approximand) is to be provided, for example evaluations of averages
of the approximand at a small number of values or simply evaluations of the
approximand. The most usual way this information is given is as point evaluations
at the aforementioned data points. In this case the functionals take the form
λ ξ ( f ) = f (ξ) for all ξ ∈ Ξ.
But there are also other operators available that might be applied to the function
(values): divided differences (particularly easy to use when the data points and
centres are equally spaced and gridded), derivatives (usually of some fixed total
order), or local integrals, for instance averages over balls of small radius placed at
the given centres. The divided difference approach is often applied when we are
not using predetermined linear combinations of the (radial or spline or other) basis
functions but use the plain vanilla shifts themselves. In the case where supp(ϕα ) is
bounded, the λ α are often finite linear combinations of values of f at points located
in a neighbourhood of this support. This property makes Q a local operator in the
sense that the value of Q f only depends on values of f in a neighbourhood of its
argument.
But now, having decided on the information available for approximation, the next
decision is the choice of the function space from which these approximants (here and
now: quasi interpolants) should stem. This is determined by the ϕα , which are given
functions whose properties are rather general; often they have compact support or
fast decay at infinity. One can regard them as basic functions or quasi Lagrange
functions of the family of quasi-interpolants. In this work we choose spaces spanned
by either polynomial or spline or radial basis functions in different ways. If we are
given a set of data sites as mentioned above, the data sites are often the same as
the knots in the spline case and are usually also to be employed as centres for the
radial basis function spaces. In both cases it is often possible to define the function
ϕ ξ (x) = ϕ( x − ξ ) as a shift of just one radial or polynomial spline basis function.
Now that we have discussed the form of the quasi-interpolant, we will take a
closer look at the properties a quasi-interpolant should satisfy in order to give good
approximation results.
Properties of quasi-interpolants with splines, radial basis functions, etc., can be
studied in L p Banach or Hilbert spaces, especially for p = 1 and p = 2. Here,
L p (Ω), where the domain Ω ⊂ Rn may be unbounded, is for 1 ≤ p < ∞ the
space of all real-valued Lebesgue-measurable functions with finite L p -norm defined
by its pth power
8 Generalities on Quasi-Interpolation

p
f p = | f (x)| p dx.
Ω
This is well-defined as long as this integral is finite, and for p = ∞ we take the
supremum of | f (x)| as the norm, except possibly on a subset of Ω of measure zero.
The image Q f of f ∈ F , where F = L p (Ω) would be a common choice,
belongs to some subspace of F containing the function space we want to reproduce.
And also, in many cases, we wish to reproduce non-integrable functions such as
polynomials, so in particular we wish the quasi-interpolant to be exact on P 1,
that is,
Qp = p for all p ∈ P.
If we aim to reconstruct polynomials up to a certain degree, we naturally see that

this is possible if we use the space of polynomials of at least the given order as a
basis function space, or a spline space (i.e. a space of piecewise polynomials), also
at least of the desired order.
For non polynomial functions such as radial basis functions, results will show
that under certain conditions on the Fourier transform of the radial basis function,
a linear combination of such radial basis functions can be used to construct quasi
interpolants with polynomial-reproducing properties.
If we further assume that Q is a bounded operator, meaning that
Q f F
Q = sup < ∞,
f ∈F f F
it is well known and easy to verify, by adding and subtracting an arbitrary element
of the space P in the next equation, that for all f ∈ F
f Q f F ≤ (1 + Q)d( f , P),
where
d( f , P) = inf{ f p | p ∈ P}.
In the next section we will discuss in more detail the implications of these
properties for the approximation order of the quasi-interpolant.
Quasi-interpolation is such a method devoted to multiple dimensions, as both
splines and radial basis functions should, of course, be constructed for multiple
dimensions. Nonetheless, the general idea is made more accessible by looking at
one dimension first, because several tools in the desired construction are much
the same in one as in several dimensions. And naturally, the radial basis functions
themselves are, as such, one-dimensional, although taking their Fourier transforms
later on, for instance, we generalise them to arbitrary space dimensions.
2.1 Approximation Properties

We required that quasi-interpolation linear operators reproduce at least all constant
functions, i.e. P0 . The majority are at least exact on linear polynomials P1 too,
whether in one or higher dimensions. An exception to this rule is the so-called
approximate approximations of Maz’ya and Schmidt (2007), where no asymptotic
convergence orders (to zero) are established but errors go down only to a low
threshold that is chosen a priori.
The next step is to consider quasi-interpolation operators that are exact on
polynomials (algebraic or trigonometric) of higher degree. There are remarkable
results on polynomial or rational quasi-interpolants as well as on quasi-interpolants
using radial basis functions. Increasing the order of polynomial reproduction is
one possible option to get better error estimates, as stated in the previous section,
but these will generally be more complex and demand more information about the
function.
The second option is to keep the order of reproduction fixed and increase the
amount of information available about a certain bounded domain.
To give an introductory example, we assume we have decided on a quasi-
interpolant that uses data on a uniform grid in one dimension. The grid centres
are therefore hZ and we further assume that we have a Lagrange type form of the
quasi-interpolant:

Qh f = λ h j ( f )ϕ(x h j). (2.1)
j ∈Z
With decreasing h, the amount of available information on f increases as the number

of basis functions used (locally) increases, so we expect a good quasi-interpolant
to satisfy lower error estimates in this case. We say that a quasi-interpolant has
approximation order k if at least
f Qh f ≤ Chk , 0 < h < 1,
i.e. this is O(hk ) for h → 0, with an h-independent constant C.
For non-gridded data sets the mesh distance is used to derive similar results.
For basis functions such as splines, which are usually locally supported, such
approximation order error estimates are mostly proved using applicable forms of
the Taylor expansion and using polynomial reproduction in a local form. Hence
the resulting approximation order is generally equal to the order of the polynomial
space to be reconstructed.
As we have already mentioned, for radial basis functions there is a bit more
work to be done in order to derive quasi-interpolants that reproduce polynomials.
An important result in this context states conditions on the function ϕ for which
the shift invariant space spanned by ϕ(x h j) contains polynomials of order k.
This is of course necessary if the quasi-interpolation operator Qh is to reproduce

these polynomials.
2.1.1 Polynomial Reproduction and Strang–Fix Conditions

We now want to turn our attention to discrete quasi interpolants of the Schoenberg
type. These were among the first quasi interpolants introduced, and are discrete
quasi-interpolants in Lagrange form

Q f (x) = f (ξ)ϕ ξ (x), x ∈ R,
ξ ∈Ξ
for any real x. The fundamental contributions of Schoenberg are essential for almost
all the work in the area of quasi-interpolation and cardinal interpolation, not only
by splines but also by radial basis functions, as described in Chapter 3 and later on.
Schoenberg was the first to describe this kind of approximation for function
values in Z (see Schoenberg 1946a, 1946b). These two papers are the foundation
for most of the quasi-interpolants described in this book. Schoenberg used integer
shifts of just one basis function, so his quasi-interpolant takes the form

∞
Q f (x) = f ( j)ϕ(x j), x ∈ R.
j=−∞
The key goal was to find conditions on the function ϕ that is local (e.g. decaying
quickly, perhaps exponentially, or locally supported), which enforce Q f to reproduce
polynomials and possibly to interpolate any function f on Z.
In the original literature, Schoenberg referred to this kind of approximation as
smoothing interpolation, but it is exactly what we call quasi interpolation today. We
will see many applications of the next theorem and its enhancements, but we are
going to state the proof of the simplest versions so that the essence of the proof
becomes clear.
The basic tool of Schoenberg’s approximation is that of Fourier transforms in
one and more dimensions,

f (ξ) =
ˆ f (x) e−iξ ·x dx, ξ ∈ Rn,
Rn
and the Poisson summation formula. The Poisson summation formula taken from
Stein and Weiss (1971) states that for any Schwartz test function f ∈ S with Fourier
transform fˆ, if we consider its periodisation

F (x) = f (x + 2π ), x ∈ Rn,
∈Z n
then F is a 2π-periodic function with Fourier expansion
1 ij x ˆ
e f ( j), x ∈ Rn .
(2π) n j ∈Zn
We use the Schwartz space of infinitely smooth functions γ̃ in n variables that

decay like (1 + x) −ν for all ν, to define the generalised Fourier transforms of non-
integrable functions that have at most polynomial growth. We also use D, the Dirac
space of infinitely smooth functions γ̃ on Rn that are real valued and compactly
supported.
Generally the Poisson summation formula refers to the implication of the above
for x = 0 only; to be precise,
1 ˆ
f ( j) = f (2π ).
(2π) n j ∈Zn ∈Z n
The Poisson summation formula holds in any dimension with suitable changes of
constants for the Fourier transform and indices. Continuing our univariate example,
we now give conditions that ensure the polynomial reproduction of the quasi
interpolant.
Theorem 2.1 (Schoenberg, 1946b) Let the basis function ϕ satisfy |ϕ(x)| ≤ A e−B |x |
for positive B. Then the quasi-interpolant

Q f (x) = f ( j)ϕ(x j), x ∈ R,
j ∈Z
reproduces polynomials, so Qq = q, for all q ∈ P1m−1 , if
1. ϕ̂(u) 1 has a zero of order m at u = 0 and
2. ϕ̂(u) has zeros of order m for all u = 2π j ( j 0 is an integer).
The theorem is proved as follows. We know from the decay of the function ϕ that
it is representable as an inverse Fourier transform
∞
1 1
ϕ(x y) = ϕ̂(t) ei(x−y)t dt = ( ϕ̂(t) eixt )(y), x, y ∈ R.
2π −∞ 2π
Applying the Poisson summation formula yields

∞
∞
ϕ̂(u + 2πn) ei(u+2πn)x = ϕ(x n) einu . (2.2)
n=−∞ n=−∞
We now regard x as a fixed number and expand the exponentials on both sides of
the last equation in terms of powers of u. Then we will compare the coefficients on
both sides. The right hand side is

∞
i u
∞
n ϕ(x n), x ∈ R,
=0
! n=−∞
and we see that the second sum in this expression is an approximation by quasi
interpolation Q(x ) to the polynomial x .
On the left-hand side of equation (2.2) we now compute the Taylor expansion of
ϕ̂(u + 2πn) at u = 0. Then only the term with n = 0 contributes to the sum (because
of Conditions 1 and 2, up to order m, if we ignore the higher-order terms), and
therefore, for a suitable g, whose properties will be stated below, we have

∞
i u
∞
n ϕ(x n) = eixu g(u).
=0
! n=−∞
The first condition in our statement of the theorem implies g(u) = 1 + O(um ), and
with the expansion of eixu we get

∞
i u
∞
∞
i u
n ϕ(x n) = x + O(um ).
=0
! n=−∞ =0
!
Comparing the first m 1 coefficients of the two expansions, we see that, for all x,

∞
n ϕ(x n) = x ,
n=−∞
which implies the polynomial reproduction.

We will refer to this theorem as an example of the basic idea of the proofs of
all the theorems of this kind. While Schoenberg assumed that |ϕ(x)| ≤ A e−B |x |
(B > 0), which applies to spline basis functions or exponentially decaying kernels,
this condition was dropped in later enhancements of the theorem in Light and
Cheney (1992). This made Schoenberg’s approximation accessible to a much wider
range of basis functions.
The statement was generalised in several ways by Strang and Fix for multi
dimensional uniform grids a fundamental contribution and then extended
from compactly supported functions to functions with certain decay properties.
We cite a version matching our current setting from Rabut (1992, Theorem 3.1); the
original version is described in Strang and Fix (1973). Throughout this book we will
use the following standard notation for the index quantities in n dimensions. For
α = (α1, , α n ) ∈ Z+n and β ∈ Z+n we define
α! = α1 ! · · · α n !,
α β = (α1 β1, . . . , α n βn ),
xα = x 1α1 · · · x αn n ,
β ≤ α if and only if α β ∈ Z+n and
β < α if and only if α β, α β ∈ Z+n and
D α = D1α1 · · · Dnαn .
Theorem 2.2 Let m be a positive integer and let ϕ : Rn → R be a function such

that
1. there exists a non negative real valued such that, when x → ∞, |ϕ(x)| =
O( x −n−m− ),
2. D α ϕ̂(0) = 0 for all α ∈ Z+n , 1 ≤ |α| ≤ m, and ϕ̂(0) = 1,
3. D α ϕ̂(2π j) = 0 for all j ∈ Zn \ {0} and all α ∈ Z+n with |α| ≤ m.
Then the quasi interpolant

Qh f (x) = f (h j)ϕ(x/h j), x ∈ Rn
j ∈Z n
is well-defined and exact on the space of polynomials of degree m, and the

approximation error can be estimated by
⎧
⎪
⎪ O(h m+ ), when 0 < < 1,
⎪
⎪
Qh f f ∞ = ⎨
⎪ O(h m+1 log(1/h)), when = 1,
⎪
⎪
⎪O(h m+1 ),
⎩ when > 1,
for h → 0 and a bounded function f ∈ C m+1 (Rn ) with bounded derivatives.
Conditions 2 and 3 are generally referred to as Strang–Fix conditions, and

basis functions possessing these properties are used in linear as well as nonlinear
approximation operators.
2.2 Main Examples of Quasi-Interpolants

Now that our first encouraging error estimates for general quasi-interpolants have
been established, we will briefly give an introduction on how such quasi-interpolants
are formed and derived for different choices of basis functions. This section consists
of a brief overview and initial examples for each method, and the details are then
given in the corresponding chapters of this volume.
2.2.1 Quasi-Interpolants Based on Radial Basis Functions

In quasi interpolants based on radial basis functions, the aforementioned translates
to span the linear space are
φ( · ξ 2 ), ξ ∈ Ξ.
Other norms (especially p -norms) can and have been applied, with greatly √ varying
success, but in particular isotropy can easily be given up by using x H x, for
instance, with positive (semi-) definite matrices H of various eigenstructures. We
will normally leave out the subscript 2 in the norm notation.
Many of the standard (and indeed practically useful) examples of radial basis
functions (identity, multiquadrics, inverse and generalised multiquadrics, thin-plate
splines, shifted thin-plate splines, all to be specified below in this book) are not at all
locally supported or even decaying, other than exponentials, for example (Gaussian
or Poisson kernel or quartic exponentials), or the compactly supported ones due
to the first author of this book or Wendland, and thus direct quasi-interpolation
using shifts of these functions as given above is not useful. The same is true for
polynomials or trigonometric functions (unless we work in an automatically periodic
setting such as on spheres).
Instead, a priori linear combinations such as (divided) differences of the radial
basis functions are formed to provide ψ or ψ ξ , which are from the same linear
function space but are now able to be local: they are at least absolutely integrable, or
indeed quickly decaying like a negative power, but not usually compactly supported
(excluding trivial cases), which in practice is often less important than initially
expected.
Therefore the quasi-interpolant is not formed directly from the basis function φ
but using the quasi-Lagrange functions ψ ξ :

Q f (x) f (ξ)ψ ξ (x), x ∈ Rn,
ξ ∈Ξ
where the fundamental functions ψ ξ are a (usually finite) linear combination of

translates of the basic radial basis function φ( x):

ψ ξ (x) = μ j,ξ φ( x ξ j ), x ∈ Rn,
j=0
where the shift values ξ j can vary with ξ. If we use gridded data, the basis function
will be formed as a shift of just one basis function via ψ ξ = ψ(· ξ). Therefore
the coefficients, μ j do not depend on the points at which the corresponding ψ ξ are
associated (‘centred’) due to the periodicity of the data.
These cases are the so-called shift-invariant spaces, studied in great detail by
de Boor et al. (1994). Intermediate cases have been studied too (see e.g. Buhmann
and Micchelli, 1992b), where there is a finite number of scattered data that are then
periodised.
That being said, the most important necessary step is to ensure that quasi
interpolants (linear combinations of the said ψ, which we call quasi interpolating
basis functions) are capable of exactly reproducing all polynomials of a fixed total
degree, at least constants.
For the reproduction of constants, the reproducing condition can immediately be
reformulated as

ψ ξ (x) ≡ 1, x ∈ Rn .
ξ ∈Ξ
For gridded data, higher-order polynomial reproduction is achieved when the

coefficients μ j of the quasi-Lagrange functions are such that the resulting function
ψ satisfies the Strang–Fix conditions stated in Theorem 2.2.
Depending on the choice of radial basis function, choosing applicable coefficients
is not always possible and sometimes depends on the parity of the dimension.
In particular, it is usually not possible for those radial basis functions which are
sufficiently local in theory, for example exponentials (Gaussian or Poisson kernels).
Given the large choice of different radial basis functions, this in fact poses no
problem in reality.
The list of applicable basis functions is now enhanced, starting with instances of
multiquadric quasi Lagrange functions decaying in odd dimensional cases such as
(1 + x) −3n−2
for large x. Note that in particular the decay depends on the spatial dimension n
and even increases with larger n, although in principle the work becomes even more
difficult for large dimensions! This is in great contrast to the well-known curse of
dimensionality.
The decay of the radial basis function will be inherited by any finite linear
combination, thereby ensuring that such a combination will satisfy Property 1 of
Theorem 2.2.
Since the radial basis functions normally employed are non-polynomial and not
of compact support, neither φ nor ψ are usually compactly supported. This might
be a disadvantage from time to time, but the mathematical approximation estimates
are nonetheless powerful, and in practice hardly any accuracy problems occur.
On the other hand it is important that the quasi-interpolating basis functions are
easy to formulate and, in particular, made from finite sums of translates of the
original basis kernel, and this creates the aforementioned restrictions on the parity
of the space dimension. It has been shown that if this requirement is also given up,
even the restrictions on n’s parity are no longer needed. Regarding the number of
variables, it is interesting to notice at this early stage that the total degree of possible
polynomial reproduction usually depends on n and increases with n.
The faster decay admits inserting polynomials into the sums, and it is interesting
to notice cases such as exponentials, Gaussian kernels and Poisson kernels which
are initially exponentially decaying, so no quasi-interpolating basis functions need
be formed, but there are no polynomials in their span and therefore no polynomials
to be reproduced by the quasi-interpolation operator approach.
Beginning with these new bases for the radial basis function spaces, many
new, linear and nonlinear approximation operators may be formed: wavelet or
prewavelet approaches of different forms enabling nested decompositions of Hilbert
spaces, compression operators using them or other approaches, least-squares
approximations and other minimising approximations, best L p -approximations, e.g.
p = 1 or p = ∞, to name but a few. We will revisit some of these methods in Chapter
9 but focus on quasi-interpolation for the main part.
It is interesting to compare this with other approximation methods using
polynomials, for example, where we have similar choices.
Take subdivision algorithms as approximation means, for instance. We can choose
between interpolating and non interpolating approaches, where the initial control
points are kept on the limit curve and always on its intermediate approximations
during the subdivision process in the first case, and where this is no longer required
in the second case. Using the subdivision iteration we can in fact compute even
inner products of the limit functions with its own shift instead of points, or for
example inner products of function with its derivative shifted by using the so-
called refinement equation and solving certain eigenvalue problems with eigenvalues
depending on the degree of the aforementioned derivative.
All polynomial-reproducing features of our quasi-interpolants, which in a way
replace the interpolation conditions, and the decay features of ψ, are expressed by
properties of the generalised Fourier transforms of the radial basis functions, which
are themselves radially symmetric. The generalised (distributional) form of the oth-
erwise standard Fourier transform is to be used, as the radial functions are generally
not integrable. In fact these Fourier transforms are singular at the origin, precisely
because the radial functions are not absolutely integrable, and their singularity
orders increase with dimension and order of growth of the original function.
We use the space of infinitely smooth functions γ : Rn → R that decay with
arbitrarily large negative powers of (1 + x) as we let x become arbitrarily large.
Then the generalised Fourier transform
σ̂ : Rn \ {0} → R,
of an at most polynomially growing test function σ : Rn → R, satisfies the Parseval–

Plancherel type formula

σ(x) γ̂(x) dx = σ̂(x)γ(x) dx
Rn Rn
for all γ satisfying

γ(x)q(x) dx = 0, q ∈ Pk−1
n ,
Rn
where k is an integer, and the smallest such integer possible for σ is its singularity
order.
The essential features of the approximation with quasi interpolation and their
approximation orders (approximation powers) stem from suitable conditions on the
coefficients μ ·, · used to derive ψ ξ from φ. Those conditions and the obtainable
approximation powers depend essentially on the underlying space dimensions,
the radial basis function and the singularity of its generalised Fourier transform
(recalling that the radial basis functions that interest us are not absolutely integrable
and therefore their generalised Fourier transforms are singular at the origin).
As radial basis functions, the choices of multiquadrics, inverse and generalised
multiquadrics and inverse quadratics, odd powers of r and even powers of r times
a logarithm are some of the most interesting choices. Since the expansions of these
radial functions (series about zero) and their generalised Fourier transforms are
well understood, we can use classical tools such as the Poisson summation formula
to consider the Fourier transforms of the linear combinations of the shifts, ψ̂, in
order to determine coefficients μ which ensure that the quasi-Lagrange functions
satisfy the Strang–Fix conditions. By resolving the singularity of the multivariate
generalised Fourier transform of the radial basis function using a trigonometric
polynomial with a sufficiently high order zero at the origin, with coefficients μ ·, · ,
the result becomes an integrable function.
We will demonstrate the method using a simple example, so we assume that the
space is constructed from a shift of one basis function ψ ξ = ψ(· ξ).
The generalised Fourier transform of the linear combination of basis functions
ψ is

ψ̂(x) = μ j eiξ j ·x φ̂(x),
j=0
which allows us to choose appropriate values of ξ j and μ j that ensure higher-order

Strang–Fix conditions as in Theorem 2.2.
These quasi-interpolants have been studied by Buhmann (1990a) and Jackson
(1988), for instance.
We take the following example from Jackson (1988, Chapter 4), when φ(r) =
r 2 log(r) is the Duchon or thin plate spline. The Fourier transform of this in two
dimensions is a constant multiple of x −4 . Now, from the resulting form of the
Fourier transform, the above conditions on the coefficients are deduced by replacing

eix ·ξ j with a polynomial series expansion ∞ s=0 (ix · ξ j ) /s!. Condition 2 of Theorem
s
2.2 is now satisfied for m = 0 if

∞ ∞
(iξ j · x) s
8π(x 41 + 2x 21 x 22 + x 42 ) −1 μj = 1 + O( x)
j=0 s=0
s!
near zero. This results in the following list of linear equations for the coefficients μ j
and the points ξ j = (a j , b j ):

12
12
12
12
12
12
μj = aj μj = bj μ j = a2j μ j = a j bj μ j = b2j μ j = 0,
j=0 j=0 j=0 j=0 j=0 j=0

12
12
12
12
12
12
a3j μ j = a2j b j μ j = a j b2j μ j = b3j μ j = a3j b j μ j = a j b3j μ j = 0,
j=0 j=0 j=0 j=0 j=0 j=0

12
12
12
a4j μ j = b4j μ j = 3/π, a2j b2j μ j = 1/π (2.3)
j=0 j=0 j=0
Condition 3 of Theorem 2.2 follows immediately from (2.3) since we choose the
data sites on an integer grid. When choosing the set of points
ξ0 = (0, 0), ξ1 = (1, 0), ξ2 = (0, 1), ξ3 = ( 1, 0), ξ4 = (0, 1),

ξ5 = (1, 1), ξ6 = ( 1, 1), ξ7 = ( 1, 1), ξ8 = (1, 1),
ξ9 = (2, 0), ξ10 = (0, 2), ξ11 = ( 2, 0), ξ12 = (0, 2),
we get the following coefficients:

5 1
, μ0 = μ1 = μ2 = μ3 = μ4 = ,
2π π
1 1
μ5 = μ6 = μ7 = μ8 = , μ9 = μ10 = μ11 = μ12 =
4π 8π
These coefficients are proportional to those of the discrete bi Laplacian scheme
based on the same points. The associated quasi interpolant is of Schoenberg type,

Q f (x) = f ( j)ψ(x j), x ∈ R2,
j ∈Z2
and it reproduces constants.

Once we have formed the modified quasi interpolating basis functions, we

arrive at decaying and specifically at absolutely integrable functions. The seeming
disadvantages of radial basis functions of classical choices are therefore eliminated.
We should now note that not all function spaces admit our constructions of
quasi-interpolants, such as the following or in the following situations. If the radial
functions are absolutely integrable or even compactly supported or exponentially
decaying, their Fourier transforms exist in the classical sense and are certainly
continuous about the origin and sometimes even analytic. When the Fourier
transforms have no singularity, and the original radial functions have a non-zero
integral, this Fourier transform has a well-defined non-zero value at the origin, there
is no singularity to resolve, and this admits no linear combinations of translates
that would then be used to form the quasi-interpolants. Examples include all of
Wu’s, Wendland’s and Buhmann’s compactly supported radial basis functions (see
e.g. Porcu et al., 2013), and the Gaussian and Poisson kernels. We have already
mentioned other examples, but there are also the so-called Dagum functions of the
general form with certain parameter conditions:
r β γ
φ(r) = 1 .
rβ + 1
The parameters that render these radial basis functions useful by making the
interpolating matrices employing them strictly positive definite have been studied
by Hofmann (2013), for example.
2.2.2 Polynomial Quasi-Interpolants

We now consider easy examples that are one dimensional quasi-interpolants and
use simple finite sums to form the approximation. Of those, the simplest model in
one dimension is the classical Bernstein operator (Bernstein, 1912)
m i
Q m f (x) = f ψi(m) (x), x ∈ [0, 1].
i=0
m
We can describe this classical operator as a quasi-interpolant of the general form

introduced at the beginning of this chapter. Therefore we set F = C[0, 1], Λ =
{0, 1, , m}, and we take ϕi (x) ψi(m) (x) as the Bernstein basis of the polynomials
of degree m. The space we reconstruct is the space of linear polynomials and the
linear functionals are evaluation functionals, λ i ( f ) = f i f (i/m).
It is also possible to construct polynomial quasi-interpolants with higher
approximation order, that use the Bernstein-operator. One can be defined by
1 x(1 x) 2
m f (x) = Q m f (x) +
Q[2] D Q m f (x), x ∈ [0, 1]
2 m 1
It can also be written as

m
(m)
m f (x) =
Q[2] λ [2]
i ( f )ψi (x), x ∈ [0, 1],
i=0
where the coefficient functional is accordingly

i(m i) i(m i) i(m i)
λ [2]
i ( f ) f i−1 + 1 + fi f i+1
2(m 1) m 1 2(m 1)
The operators Q m and Q[2]
m are uniformly bounded with respect to m: one obeys the
bound Q m ∞ = 1 and the other one Q[2]
m ∞ ≤ 3. As for approximation properties,
if f is smooth enough, we have
|Q m f (x) − f (x)| = O(1/m), |Q[2]

m f (x) − f (x)| = O(1/m ),
2
since Q m is exact on linear polynomials, while Q[2]

m is also exact on quadratic
polynomials.
More generally, we can define a sequence of quasi-interpolants that have even
higher approximation orders. Setting {Q[k] [0]
m | 0 ≤ k ≤ m} so that Q m = Q m and
[m]
Q m = L m , which are Lagrange interpolants on equally spaced points with spacing
1/m, we therefore have

m
(m)
m f (x) =
Q[k] λ [k]
i ( f )ψi (x), x ∈ [0, 1],
i=0
where the coefficient functional λ [m]

i ( f ) is a linear combination of finitely many
function values f j , where the j are restricted to neighbouring values i in a
precise definition. Details of the construction of this quasi-interpolant are given
in Sablonnière (1989, 1992).
These results also apply to Bernstein-type quasi-interpolants on simplices or
hypercubes (by tensor product techniques) and to Bernstein–Durrmeyer-type or
Bernstein–Kantorovich-type quasi-interpolants

m
m
M m f (x) = μi ( f )ψi(m) (x), Km f (x) = κ i ( f )ψi(m) (x),
i=0 i=0
where we define the functionals by integration (averages, in other words)

1 (i+1)/(m+1)
μi ( f ) (m + 1) ψi(m) (t) f (t) dt, κ i ( f ) f (t) dt
0 i/(m+1)
There are also other types of quasi-interpolants of Bernstein type, namely discrete
Bernstein operators, or for instance the q-operators of Phillips, or Lewanowicz
Woźny generalisations. The general method of construction introduced for these
types of quasi-interpolants in Sablonnière (1999) will be featured and demons-

trated as an example of quasi-interpolants derived from the Weierstrass operator
in Chapter 9.
2.2.3 Polynomial Spline Quasi-Interpolants

We now aim to construct the quasi-interpolant from univariate or multivariate spline
spaces. The functions used are piecewise polynomials either on intervals or cells of
a grid in multivariate space, and possess additional overall smoothness conditions.
For this introduction we focus on the univariate setting where Ω is an interval,
which is divided into subintervals by a knot sequence. The polynomials of order m
(i.e. degree m 1) are elements of the resulting spline space consisting of piecewise
polynomials of order m on each subinterval and overall smoothness m 2. This is the
most common choice of these properties. It follows immediately that, given a basis
of the spline space N1, , N j , which will usually be a B spline or box spline basis,
and since the polynomials are elements of the spline space, there exist coefficients
for the monomial basis of Pm−1 to satisfy

j
−1
x = λ k Nk (x), 1≤ ≤ m, x ∈ Ω.
k=0
Formulas of the above form play a crucial rôle in the construction of polynomial
spline, box spline and trigonometric spline quasi-interpolants and are usually
referred to as a Marsden identity. As the functionals and polynomial space are
linear, we can now construct polynomial-reproducing quasi-interpolants by using
any linear functional that satisfies
λ k (x −1 ) = λ k for all 1 ≤ ≤ m, k ∈ {0, . . . , j}.
For the basis functions a desirable property is having a support that is minimal for a
set of basis functions; this, for example, is the case for a B-spline basis. Further, one
aims to choose the functionals λ j to depend only on function values, or possibly
averages or derivatives of the function within or near the support of N j . Under these
assumptions the quasi-interpolant

j
Q f (x) = λ k ( f )Nk (x), x ∈ R,
k=0
is reproducing polynomials of order m, and under the additional locality conditions,

it will have approximation order m 1.
The simplest model is the Schoenberg operator

Sm f (x) = f (k)Mm (x − k), x ∈ R, (2.4)
k ∈Z
where Mm is the normalised B spline of order m with support [ 12 m, 12 m), where

the node set is Z if m is even and Z+ 12 if m is odd. This quasi-interpolant reproduces
constants since the normalised B spline basis forms a partition of unity.
In Chapters 4 and 6 on univariate and multivariate spline quasi interpolants, we
will present a great variety of functionals that ensure the reproduction of higher-
order polynomials and thereby give higher approximation orders.
Also in these chapters the concept of quasi-interpolants that are derived from local
spline or polynomial projectors are described. The method uses the local support
of the spline basis functions to derive a solution on Ω by using other techniques on
small subintervals, e.g. least-squares or L 2 -projections, and then recombining the
information to yield a quasi-interpolant on the whole domain.
Quasi-interpolants using splines can be generalised to a multivariate setting, for
example by tensor product methods or by box splines, as will be discussed in detail
in Chapter 6.
2.2.4 Trigonometric Quasi-Interpolants

When the function to be approximated is periodic, the function space we use for
approximation, as well as the function space we aim to reproduce, should consist
of periodic functions only. The periodic basis functions that replace polynomials in
this setting are usually trigonometric polynomials taking the form

n
a0 + ak cos(k x) + bk sin(k x) π ≤ x ≤ π, ak , bk ∈ R
k=1
Generally quasi-interpolation in this setting proceeds in the same way as described

before, especially if the piecewise trigonometric polynomials or trigonometric
splines are chosen as basis functions of the quasi interpolant.
The focus in this book is on trigonometric spline approximation, and since the
trigonometric versions of, for example, the Taylor expansion and Marsden identity
are of higher complexity and not as well known as these results for polynomials, the
focus will be on describing these results.
Therefore not all extensions and generalisations of these methods are discussed
in detail, but a subset is referred to in Section 5.4. Besides trigonometric splines,
we will also discuss trigonometric multiquadric-based quasi-interpolants, which
correspond to the adaptation to periodic functions of the radial basis function
method described earlier in this section.
The trigonometric approximation also has important applications for approxim

ation on the sphere, which we will discuss in Chapter 8.
2.2.5 Quasi-Interpolants Obtained from Other Types of Operators

A large number of other types of quasi-interpolation operators have been developed
during the past 100 years.
Examples include those by Dunham Jackson (1930):
π
Jm(r ) ( f , x) ϕm,r (x t) f (t) dt, m, r > 0, π ≤ x ≤ π,
−π
where
sin( t/2) 2r
ϕm,r (t) cm,r , [m/r] + 1,
sin(t/2)
and cm,r is defined by normalisation, that is,
π
ϕm,r (x) dx = 1
−π
The kernel has a removable singularity at the origin.

By approximating the integral by a quadrature formula based on the nodes t k ∈
[ π, π], with weights wk , we get a discrete quasi-interpolant of the type

N
N
J˜m(r ) ( f , x) wk ϕm,r (x t k ) f (t k ) = f (t k )ϕ(x t k ), x ∈ R,
k=1 k=1
which has the general form given above. Another example of non-discrete, continu
ous quasi interpolants is the convolution operator named after the mathematician
de la Vallée Poussin,
π
1
Vm ( f , x) ϕm (x t) f (t) dt, x ∈ R,
π π
whose kernel, i.e. the basis function the quasi-interpolants depend upon, is
1
m
1 m!
ϕm (t) (2 cos(t/2)) 2m = + ρm (k) cos(kt), π ≤ x ≤ π,
2 (2m)! 2 k=1
with
(m!) 2
ρm (k) = .
(m k)!(m + k)!
One can generalise the discretisation idea to the de la Vallée Poussin operator
and to the Weierstrass operator

1
e−(x−t) /m f (t) dt, x ∈ R,
2 2
Wm ( f , x) √
πm R
by using a quadrature formula on the reals rather than [ π, π].
2.3 Some Applications Connected with Different Forms of Quasi-Interpolants

The approximants derived using quasi-interpolation techniques can be used for the
solution of various other problems in applications. In particular, their usefulness for
the solution of integral equations and partial differential equations will be discussed
in detail in Chapter 10. The other more basic applications will only be touched
upon.
2.3.1 Integration
Using quasi-interpolants, as introduced in (2.1), for the approximation of integrals
is straightforward. The approximation is simply achieved by

f (x) dx ≈ Q f (x) dx = aα λ α ( f ), aα ϕα (x) dx
Ω Ω α∈Λ Ω
In the simpler quasi-Lagrange form with point evaluation, the above simplifies to

f (x) dx ≈ Q f (x) dx = w ξ f (ξ), w ξ ϕ ξ (x) dx
Ω Ω ξ ∈Ξ Ω
If, further, the approximation space is created by shifts of just one kernel
ϕ ξ = ϕ(· − ξ), the coefficients will also be independent of ξ if Ω = Rn .
2.3.2 Differentiation
If we aim to approximate a linear differential operator D, then we have the
approximation

D f (x) ≈ DQ f (x) = λ α ( f )Dϕα (x)
α∈Λ
to the differential operator. This allows the approximation of the derivatives of a

function by quasi-interpolation.
2.3 Some Applications Connected with Different Forms of Quasi-Interpolants 25
2.3.3 Integral Equations

Quasi interpolation methods have been successfully used to solve Fredholm
(integral) equations,
b
u(x) = f (x) + k (x, t)u(t) dt, x ∈ [a, b],
a
where f and k are given and u is to be approximated. Here is the basic idea of
solving this problem using quasi-interpolation (more of this below). Let Q be a
general quasi-interpolant.
The quasi-interpolant is used to approximate k (·, t) : x → k (x, t) and we get the
approximated kernel

Qk (·, t)(x) = λ α (k (·, t))ϕα (x) = k̃ α (t)ϕα (x).
α∈Λ α∈Λ
Replacing the kernel with the quasi-interpolant allows us to solve the integral
equation on a subspace depending on the basis function used; the details will be
described in Section 10.3.
2.3.4 Partial Differential Equations

Given that radial basis functions are useful for approximating multivariate functions
by quasi-interpolation, it is now usual to apply them to approximate solutions of
partial differential equations numerically. The goal is to use the quasi-interpolants
and quasi-Lagrange functions and their approximation power to derive useful
schemes for solving PDEs, much like the piecewise polynomials in finite element
algorithms (see e.g. Ciarlet, 2002) and other methods using splines.
Indeed, several approaches have been tried and tested in this direction, namely
collocation, i.e. interpolation techniques (of which we speak much less in this
book as we discuss quasi interpolation more than strict interpolation), variational
formulations and boundary element methods. For us, their main purpose is to solve
elliptic partial differential equations with boundary values given. There are several
useful approaches; they include collocation methods (see e.g. Kansa, 1990) and
Galerkin methods, and strong and weak formulations (Franke and Schaback, 1998).
For details we refer the reader to Chapter 10 on applications.
3
Univariate RBF Quasi Interpolants
We begin with a collection of simple cases in one dimension that illustrate many
of our techniques in some detail, which we will use again frequently below. This
includes in particular the work on radial basis functions and examples such as the
well-known multiquadrics.
The general formulation of quasi-interpolants for equally spaced knots in one
dimension is

∞
Qh f (x) = f ( j h)ϕ(h−1 x j), x ∈ R,
j=−∞
where the infinite sum on the right hand side is usually required to be absolutely
summable. This is a condition that is easily formulated for this series in one
dimension and it is usually met by letting f have growth of some maximal order
(bounded or polynomially increasing up to some degree) and requiring ϕ to decay
fast (or be compactly supported). Alternatively we could require ϕ simply to be
bounded, for instance, and let f decay. The latter situation occurs rarely; the former
conditions are standard. We may also replace function evaluations with mean
values or derivatives, and will need similar requirements of decay and absolute
summability.
Examples are exponentially decaying functions or compactly supported ϕ such
as B-splines.
Many known techniques are available for establishing the question that is always
asked first, namely how fast Qh f will converge to f (if at all) when h → 0. In this
chapter we will give a few examples using radial basis functions that increase but
allow an elegant formulation of a kernel ϕ created by taking divided differences of
the increasing radial basis function.
3.1 Univariate Quasi-Interpolants 27
3.1 Univariate Quasi-Interpolants

As we have seen, many radial basis functions allow an easy formulation of quasi
interpolants as sums of function values times shifts of the kernel function ϕ as long as
they decay fast enough, because if ϕ decays exponentially (think of Gaussian kernels
ϕ(r) = exp( c2r 2 ) or Poisson kernels ϕ(r) = exp( cr)) or is compactly supported
(e.g. B splines), there are no issues with the convergence of the corresponding
infinite series.
The start of the quasi-interpolation analysis for kernel functions in the one-
dimensional setting with radial basis functions (such as multiquadrics) that increase,
and then forming spaces generated by linear combinations of them, is due to Powell
(1990). He noticed that the linear space
√ spanned by second-order divided differences
of the multiquadric function φ(r) = r 2 + c2 contains P11 (as required, or suggested,
in the previous chapter; Pkn is the space of polynomials in n variables and of total
degree at most k) if there are infinitely many centres that diverge (go to infinity) to
both +∞ and ∞. They need not be equally spaced.
Here, divided differences of a function g at points ξk are defined recursively as
[ξk ]g g(ξk ) and
[ξk+1 ξk+2 . . . ξk+j ]g [ξk ξk+1 . . . ξk+j−1 ]g
[ξk ξk+1 . . . ξk+j ]g .
ξk+j ξk
This observation about the space containing some polynomials was preceded by
one in Buhmann (1988), who noticed that this happens when the centres form the
set of all the integers. This is also related to the polynomial quasi-interpolants of
Chapter 2.
Specifically, Powell (1990) proved the following.
Theorem 3.1 Let Ξ = {· · · < ξ−1 < ξ0 < ξ1 < ξ2 < · · · } be a sequence of centres
such that asymptotically ξ j → ±∞ as j → ±∞. If p is a linear polynomial and φ is
the multiquadric function for any real c, then the quasi-Lagrange functions
φ(|x ξ j−1 |) (ξ j+1 ξ j−1 )φ(|x ξ j |) φ(|x ξ j+1 |)
ψ ξ j (x) = − + , x ∈ R,
2(ξ j ξ j 1) 2(ξ j ξ j 1 )(ξ j+1 ξj ) 2(ξ j+1 ξ j )
so that

2
ψ ξ j (x) = [ξ j−1 ξ j ξ j+1 ]φ(x − ·) = μ φ(x − ξ j−−1 )
=0
(coefficients as above) and

∞
Qp = p(ξ j )ψ ξ j
j=−∞
yield Qp ≡ p.
28 Univariate RBF Quasi-Interpolants
This theorem is a well known fact for c = 0, i.e. piecewise linear splines, where
the approximant is actually an interpolant; see Chapter 4 on univariate spline quasi
interpolants. Its relevance here lies in the close relationship between polynomial
recovery properties of quasi interpolants, which are studied in this book, and their
order of accuracy when the centres become dense in R, as we shall see below.
Its further relevance lies in its demonstration of the links between classical spline
theory and its use in quasi-interpolation; see also Chapter 6 on multivariate spline
quasi-interpolants, in particular using multiquadrics, and the radial basis function
approach to quasi-interpolation, using so-called ‘quasi-Lagrange functions’, which
is central to this book.
The conclusion that Qp = p for linear polynomials p is based on the observation
that
1
ψ ξ j (x) = N j2 ∗ φ (x), x ∈ R,
2
where N j2 is the linear B-spline that occurs as a radial basis function when c vanishes,
and that

1 ∞ 2
∞ ∞
p(ξ j )ψ ξ j (x) = N j (t)p(ξ j )φ (x t) dt
j=−∞
2 −∞ j=−∞
∞
1
= p(t)φ (x − t) dt
2 −∞
≡ p(x),
if and only if p is a linear polynomial. The ∗ denotes convolution. The expression
is also instrumental to some slightly more recent results about multiquadric
prewavelets in Buhmann (2003), which will be discussed below.
It is important to note that, in stark contrast to the result above, if no quasi-
Lagrange functions of the suggested form are used, there are neither linear
polynomials nor even non-zero constant functions, in the univariate space defined
by all absolutely convergent series (this is a condition we shall always require)

∞
λ j φ(|x ξ j |), x ∈ R,
j=−∞
where φ is still the multiquadric function, that is, we must make use of nothing more
than divided differences or other suitable linear combinations of translates of the
radial function. This observation is contained in Powell’s paper, but we remark that
it is a trivial observation for vanishing parameter c, as

∞
λ j |x ξ j |, x ∈ R,
j=−∞
has derivative discontinuities at the centres unless it is identically zero.

3.2 Examples of Convergence Results 29
This is why the sum of the coefficients λ j must cancel, as explained below in

Chapter 7: the series λ j cos(ξ j x) must have a zero at the origin. We write this
condition as λ ⊥ 1. Here 1 denotes the vector whose entries are all one. For other
radial basis functions the same approach works but we need higher-order conditions
on the coefficients: λ ⊥ P1m | Ξ for bigger m. It will also not be enough to take simple
divided differences; results of this type can be found in Chapter 7, and we have
another example at the end of this section.
Powell also noted that the space spanned by the ψ ξ j cannot contain all quadratic
polynomials, although the ψ ξ j decay fast enough to admit them. This does not
exclude the possibility, incidentally, that certain other polynomials are in that linear
space.
3.2 Examples of Convergence Results

Beatson and Powell (1992b) made use of Theorem 3.1 by identifying approximation
orders of quasi-interpolants. Before we begin this discussion, however, we briefly
mention the results of Buhmann (1988) because they will provide a useful
background to those of Beatson and Powell. Among other things, that paper showed
that the following two assertions are true. We remark immediately that it is of utmost
importance to pay attention to the different cases of stationary and non stationary
approximation, where in the former case scalings by step size h are also introduced
into the parameters of the radial basis function (so in the next theorem one would
take for instance c = c(h) = c̃h or other powers of the spacing), and in the latter
case parameters c, say, in multiquadrics, inverse multiquadrics, Gaussian or Poisson
kernels remain, even when the spacing tends to zero, fixed and independent.
Theorem 3.2 Let Q and ψ ξ j be as above and let Ξ = hZ be equally spaced centres
with positive spacing. Let f ∈ C 3 (R) have bounded derivatives. Then, if both c and
h from (0, 1) tend to zero,
Q f − f ∞ = O(h2 + c2 min[log(1/c), log(1/h)]).
The following will work and be an interesting result even in the simple case
Ξ = Z.
Theorem 3.3 Let Q, ψ ξ j and Ξ be as in the previous theorem. Suppose that c stays
bounded away from zero as h → 0. Then there exists an exponentially decaying
infinitely smooth f that has exponentially decaying derivatives such that Q f f ∞
is uniformly bounded away from zero.
This work on multiquadrics has been carried much further by Beatson and Powell
(1992b) by studying three different forms of quasi-interpolants when finitely many
30 Univariate RBF Quasi-Interpolants
non equally spaced centres a < ξ2 < · · · < ξ m−1 < b are given. We will quote these
results as they are also required later on.
Concretely, the following quasi interpolants are studied, where one assumes that
the set of centres {ξ j } m ∞
j=1 is augmented so that we have a bi-infinite set Ξ = {ξ j } j=−∞
satisfying the requirements of Theorem 3.1,

0
m
∞
Q f (x) = f (a)ψ ξ j (x) + f (ξ j )ψ ξ j (x) + f (b)ψ ξ j (x) on [a, b],
j=−∞ j=1 j=m+1
where ψ ξ j is still the quasi-Lagrange function.

Beatson and Powell (1992b) pointed out that the above is well-defined and
independent of the choice of the centres, in [a, b], as long as they are strictly
monotone ascending, for j < 1 and j > m.
Secondly, they considered a quasi-interpolant s(x) defined by second-order
divided differences:

m−1
f (a)
f (ξ j )ψ ξ j (x) + ([a ξ2 ]φ(x) + [a b]φ(x))
j=2
2
f (b)
+ ([b ξ m 1 ]φ(x) + [a b]φ(x)).
2
Thirdly, for f continuously differentiable in the interval [a, b], they analysed

0
m
Q f (x) = ( f (a) + (ξ j a) f (a))ψ ξ j (x) + f (ξ j )ψ ξ j (x)
j=−∞ j=1
∞
+ ( f (b) + (ξ j b) f (b))ψ ξ j (x), x ∈ [a, b].
j=m+1
By virtue of the above theorem and the non-singularity results of Micchelli (1986)
on radial basis functions, mentioned in Chapter 1, the quasi-interpolant used in this
chapter is in the (m + 1)-dimensional space span{φ(· ξ j )} mj=1 ∪ P1 .
0
Our other quasi-interpolant is in the m-dimensional linear space spanned just by

the m translates of φ. The first and final quasi-interpolants are in the same space
enlarged by linear polynomials, so the dimension of the vector space is m + 2. Upper
bounds for the accuracy obtained for the three quasi-interpolants are
(1 + h 1 c)ω f (h),
where ω f is the modulus of continuity of f , h is the maximal spacing between

centres, and
c{| f (a)| + | f (b)|}(b a) −1 + (1 + h−1 c)ω f (h)
Another random document with
no related content on Scribd:
CHAPTER XVI.
A DEADLY STRIFE.
“Your arrival is well-timed,” said Jewan, turning to Zeemit.

“I see that it is so,” she answered. “I soon discovered in Delhi that
you had left, and I determined to follow you, for poor old Zeemit is
alone in the world now. I was lucky in meeting with Wanna. Some
years ago I was in Cawnpore, and I knew her then. When she learnt
that I had followed you, she lost no time in conducting me here.”
“I am glad of it,” said Jewan. “My prize will be safely kept now. Guard
her well, Zeemit; and you, Wanna, if you value your life, look to her!
You understand? She has dared to defy me, and I swear to subdue
her!”
He crossed the room to where Flora still trembled, and crouched
upon the floor. He stooped over, and said, with bitterness—
“I leave you now. Business calls me hence, but I shall return to-night,
and then we will see who conquers.”
He passed out of the room, and Wanna locked the door after him. It
was an inexpressible relief to Flora when he had gone. But when she
raised her head, and her eyes fell upon Wanna’s face, she
shuddered. It was a face scarcely human in its expression of hate.
She turned to Zeemit—she had given her hope in Meerut—why had
she failed her now? She could read little or nothing in the dusky
features. Her heart sank, for the glimmering ray that had supported
her hitherto seemed to fade entirely.
“Come,” said Wanna, spurning the trembling girl with her foot, “here
is food for you; I suppose I must keep life in you until Jewan has
sucked your sweetness. What he can see in you I know not. It is a
mad infatuation, and he will get the better of it; but if I had my way I
would torture you. I would spoil your beauty—I would pluck your
eyes out—I would lop off a limb from your body every day—I would
burn you with hot irons. Ah, ah, ah! it would be sport! Eh, Zeemit,
what say you? We have been ground as corn in a mill by these
accursed Feringhees; and now that our day has come, have we not
a right to be glad?”
She hummed the air of an Indian ditty, and fairly danced about the
room with fiendish glee.
“Oh, woman!” moaned the unhappy Flora, “if you are not altogether
inhuman, have pity, and kill me.”
“Ugh, bah, pish! pity indeed,” cried Wanna, moving about backwards
and forwards in that restless and strange manner peculiar to caged,
wild animals. “Have we ever had pity from your countrymen? Have
you not crushed us into the earth?—subdued us with fire and sword?
And now that our power is coming back we know well how to
retaliate.”
As she spoke she spat upon the floor twice, and made a sort of
hissing sound with her lips.[5]
“Why do you not get up?” asked Zeemit, in a tone that contrasted
strangely with the savageness and cruelty of Wanna.
The ray brightened again for Flora. She caught comfort from that
voice; but when she looked into the face she saw nothing to justify
the inference she had drawn. The kindliness displayed in Zeemit’s
voice did not escape Wanna, who turned sharply upon her country-
woman and cried—
“How is this? You speak to the white-faced cat as if she were your
pet dove, instead of an enemy.”
“Scarcely an enemy, Wanna. Her only crime seems to be that she is
a Feringhee.”
“She is a beast.”
“She is a woman, and I feel as a woman should do for her.”
Zeemit’s words were to Flora like water to the parched earth. They
gave her hope, they gave her joy; she drank them in with avidity, and
gained strength. She rose up and would have clung around the neck
of her ayah, had not the attitude of Wanna appalled her.
The hag stood facing Zeemit. The bangles on her legs and arms
chinked as she shook with passion. She was clawing the air, and
almost foaming at the mouth. She struggled to speak, but her
passion well-nigh choked her. Words came at last.
“You sympathise with this Feringhee woman. I see through you—you
are an enemy to us, a friend to her. But, if you thought to liberate her,
you have set up a trap into which you yourself have blindly walked. I
go for Jewan.”
She made a movement towards the door. To let her go would
frustrate every plan. Zeemit knew that it was no time for reflection. It
was woman to woman—age to age; for on both the years pressed
heavily. With a lithe and agile spring she fastened upon Wanna, who,
with the sudden instinct of self-preservation and the ferocity of the
jungle cat, twisted her bony fingers round and dug her nails deep
into the flesh of the other’s arms.
It was a strange scene. From the wall the picture of the idol seemed
to grin hideously. Speechless with terror, poor Flora stood wringing
her hands. The two women, panting with the first shock of attack,
glared at each other, and over all there fell the weird, flickering light
of the swinging cocoa-lamp.
As in all Indian buildings of this kind, there was a long window in the
room opening on to a verandah. The jalousies were thrown back.
The stars in the heavens were shining, and from below came up the
sounds of the voices of the natives, who were beating their tom-toms
and making merry.
Miss Meredith moved to this verandah. She peered over. She could
see groups of people below. Her first impulse was to call for
assistance, but in an instant she was convinced of the madness of
such a proceeding. On the issue of the struggle her life depended.
She might go free if Zeemit conquered—die if the triumph was
Wanna’s.
“Give me the key of that door,” demanded Zeemit, when she had
recovered breath enough for speech.
“Never while my heart beats,” answered the other.
“Then I will take it from you when your heart has done beating,” said
Zeemit.
Mehal was slightly the taller of the two women, and her arms were
longer. In this respect she, perhaps, had an advantage.
The women struggled furiously. Now they were locked in a deadly
embrace, now parted, only to spring together again with increased
ferocity. Never did wild animals grip and tear, and hiss, and struggle
more savagely than did these two women. But the springs which
moved them both to action were of a totally different nature. A kindly
desire to render assistance to one in distress was Mehal’s motive—a
deadly hatred for the Englishwoman was the other’s.
They dragged each other round the room; they panted with the
extraordinary exertion which each made to gain the victory; their
muslin garments were encrimsoned with blood and rent to shreds.
Now they dashed against the stone walls, then reeled and tottered to
the floor, writhing in the agony of the terrible grip which each had of
the other. Rising again, covered with dust and blood, and their limbs
locked together like snakes—their faces contorted with pain and
passion, and their breath coming thick and fast.
It was an awful moment for Flora. She would have rendered
assistance to Mehal, but that was impracticable, as she found, for
Wanna twisted herself about so rapidly as to frustrate the attempts
which Flora made to grasp her.
It was truly a struggle for life; for, ere it ceased, one of the strugglers
must die. They knew that, and so they fought with the desperate
energy which nerves a human being when dear life is at stake.
The efforts of Wanna were growing gradually weaker. Mehal had
worked one of her hands up to the other’s throat, and she was
pressing her thumb and fingers together, until Wanna’s eyes started.
The hag knew now that only by a desperate effort could she free
herself, and save her life. But even if that were impossible, she was
determined that her antagonist should not live to enjoy her triumph.
She put forth what little strength remained in her withered frame. It
was an upleaping of the dying fire again, and for a moment the battle
raged fiercer than ever. They spun round, and reeled, and
staggered.
The end was coming. Wanna felt that. With an almost superhuman
effort, she managed to drag her foe to the verandah, and, with a
quick and sudden movement, drew the key from her girdle, and,
uttering a cry of ferocious joy, was about to hurl it over the railings.
But a counter-movement of Mehal’s broke the force of the jerk, and
the key fell on the extreme edge.
Flora darted forward, but she could not pass the combatants.
Wanna saw that her chance had gone. But nerving herself for one
final struggle, she dragged Mehal round. They lost their balance—
they fell to the floor—they rolled against the wooden railings, which,
old and rotten with age, broke down with a crash. Away went the key
into space. The two women were on the extreme edge of the
verandah!
Flora rushed forward once more. She made a frantic clutch at their
garments, with a view of dragging them back.
It was too late!
Death let fall his spear, and took the stakes. The fighters rolled over,
and Flora stood petrified with horror, still holding in her hands some
remnants of blood-stained garments.
The wind moaned amongst the ivy on the walls. In its wailing she
seemed to hear a prophetic voice that told her the struggle she had
been an unwilling witness to between the two women, but
represented the greater struggle between two races that had just
commenced; and, before it could end, the soil of India should be
drenched with blood.
The night wind moaned. It sounded in her ears like a requiem for her
slaughtered friends. It seemed like an agonised cry of pain, wrung
from hearts suffering almost more than mortal sorrow.
The night wind moaned—a dirge-like moan, that told that the Angel
of Peace had been beaten, broken-winged, into the dust; and
through the Orient land were stalking the grim demons, War and
Woe.
The night wind spoke. It told her that the catastrophe she had just
witnessed destroyed every hope of escape she might have had, for
with Zeemit her best friend had gone.
She heard Jewan Bukht’s voice in the wind—a voice malignant and
cruel.
“I will return to-night, and then we will see who conquers!”
Those were his parting words. As the wind repeated them to her, it
called her back to a sense of her awful danger. Her almost stilled
heart sprang into life again. It throbbed with the wildness of fear and
horror at what the consequences might be if he returned.
She could foil him yet; in her hands she held her own life. An effort of
will, and she could snap the “silver thread” and break the “golden
bowl.” Three paces forward, and a plunge down into the dark depth,
whence had rolled the bodies of Zeemit and Wanna.
Were it not better to die than to live to shame and misery?
When all hope has fled, when everything that can make life
endurable has gone, has not the time come to die? She thought this.
And the moaning wind answered her, and said “Yes.”
A plunge, a rapid descent, a terrific shock, and then the end.
She looked up to the silent stars. They seemed to look down
pityingly on her. Mentally her gaze wandered beyond the stars, to the
plains of peace, to the White Throne of Mercy and Justice, and she
put up a prayer for forgiveness.
Be still, wild heart! cease, oh, throbbing brain! death is merciful.
She took a step forward—she closed her eyes—she threw up her
arms; and, bending her body, she was about to take the fatal leap,
when a voice reached her.
Not of the wind this time, but a human voice, that cried for help, that
told of pain.
She went down on her knees. She peered over the broken verandah
into the darkness. She could see nothing. The voice had ceased,
and there was silence again, save that the “ivy rustled and the wind
moaned.”
FOOTNOTE:
[5] When the Hindoos wish to express a thorough loathing and
contempt for anything, they spit upon the ground, and make a
peculiar movement with the lips. During the mutiny, and for long
afterwards, it was common for the native servants in the European
houses, when ordered to do anything, to spit upon the ground when
they thought their masters were not looking. The language put into
the mouth of Wanna, and the ferocity depicted, are by no means an
exaggeration. In fact, words would almost fail to accurately express
the inhuman hatred for the English, which the natives—men and
women—took every opportunity of displaying during the revolt.
CHAPTER XVII.
FOR LIFE AND LOVE.
The cry that came up out of the darkness, and stayed Flora Meredith
in the very act of self murder, was uttered by one who had been
miraculously saved from an awful death.
For some minutes Flora continued to strain her eyes before she
could make anything out. Then she became conscious that the figure
of a woman was lying on a verandah about fifteen feet below, and
which projected considerably beyond the lines of the upper one on
which Flora stood. That it was one of the women who had rolled
over, Miss Meredith had no doubt; but which one was a question
difficult to answer. But presently the cry was repeated. Flora fancied
she detected Mehal’s voice, but could not be certain. Everything was
quiet below in the grounds, for the hour was late, and nobody was
about. She bent over the verandah as far as possible, and, in a low
tone, called—
“Mehal—Zeemit—Zeemit.”
She waited with palpitating heart for any reply, for on that reply it
might truly be said her life hung. But the reply did not come—only a
half-stifled moan telling of acute suffering.
Again she called—a little louder, this time; again she waited in
expectancy, to be disappointed once more. She rose to her feet, and
considered what was best to be done. There was little time to lose,
little time for thought.
Hope rose again. If she could manage to reach the lower balcony,
she might be saved. But how was that to be accomplished? Even if
she had been in possession of a rope, she doubted her ability either
to make it fast, or, having succeeded in that, to lower herself down;
for easy as such a thing seems to the uninitiated, it is practically a
task fraught with the utmost danger, and requiring an exertion of
physical strength severe for a man, and ten times more so for a
woman. But though she had possessed the acrobatic skill to have
performed the feat, the rope was not there, nor was there anything in
the room that would have answered as a substitute. What, then, was
to be done?
She stood irresolute, almost distracted by the painful tensity to which
her mental powers were stretched. But as she stood, hovering, as it
were, between life and death, the rustling creepers whispered to her
—
“Here is a way down.”
As the idea flashed upon her, she could have cried out with joy.
She moved to the end of the verandah. The great rope-like stems
were twined and twisted together, and spread out in all directions.
She looked at her hands, delicate and soft, and mentally asked
herself if she had strength of arm and wrist sufficient for the task.
Fear lends strength, as it gives wings, and even a woman, situated
as Flora was, will perform deeds that, under ordinary circumstances,
would seem impossible.
It was the sole chance, and she must avail herself of it. She
hesitated no longer; but mounting the railing of the verandah,
grasped firmly a thick stem of the ivy, and swung herself over.
It was an awful moment. The failure of the power of the arms, the
slightest giddiness, and a fall of fifty feet would close the book of life
for ever. But after the first nervous dread had passed, she found that
the descent was far easier than she had imagined.
The rough angles of the walls, and the thick ivy, gave her tolerable
foothold. But now and again her weight dragged the stems from their
hold of the wall, and she would slip down a little way with a jerk that
sent the blood back upon her heart with a rush.
It was hard work; it was a struggle for life—a life that, a few minutes
ago, she would have sacrificed, for then all hope seemed to have
gone. But since then the star had risen a little once more, by reason
of the pain-wrung cry of a human sufferer.
She struggled with desperate energy to save that life. Lower and
lower she went. It seemed as if she would never reach the goal.
The ivy ripped and gave way, painfully straining and jerking her
arms, and the rough stones lacerated and tore her hands. But there
was no giving up until she reached the wished-for point.
She clung desperately—she struggled bravely, and the reward came
at last—she was abreast of the lower verandah! She got a foothold,
then clutched the railing, and, in a few moments, stood on the floor,
breathless and exhausted, but safe so far.
The figure of the prostrate woman was a few feet off. She moved to
her, bent down, turned her over, and then uttered a silent prayer of
thankfulness, as she recognised the well-known features of her
faithful ayah.
But it was evident that Zeemit was wounded grievously. She was
unconscious, and lay in a pool of blood, which flowed from a deep
wound in the forehead. In her descent she had struck her head on
the railing of the verandah; but this probably saved her life, as it
caused her to roll inward, instead of outward.
Flora endeavoured to staunch the blood. She chafed the hands, and
raised the body to a sitting posture. Her efforts were at length
rewarded, for consciousness slowly returned to the old woman. It
was some time before she could realise her exact position. But, as
the truth dawned upon her, she grasped the hand of Flora, and cried
—
“Allah be praised, missy, you are still safe!”
“We both live,” answered Flora; “but we both stand in deadly peril.
How are we to save ourselves?”
“You must not think of me. You must endeavour to get free of this
place, and save your own life.”
“And leave you here!” cried Flora; “never!”
“You are a brave girl, and Zeemit thanks you; but you must go.
Wanna is, no doubt, dead. If she fell to the ground, which seems
probable, it would have been impossible to have survived such a fall.
Dead people tell no tales; therefore we have nothing to fear from her.
I feel that I cannot rise. For me to go with you would but impede your
flight. Leave me. I shall be discovered. I shall tell Jewan that Wanna
intended to set you free, tempted by a heavy bribe you offered. I
endeavoured to prevent her—we struggled, and fell over the
verandah—and then all is blank to me. This will give me an
opportunity of rendering you still further assistance, because,
however angry Jewan may be, he would scarcely dare to offer me
violence.”
“It is much against my will to have to leave you here, Zeemit, and I
can scarcely reconcile myself to such a course.”
“But it is the only chance there is for me to render you aid. Besides,
there is one below who waits anxiously for you.”
“Ah! tell me, tell me, where he is?” cried Flora, the opportunity
occurring for the first time to speak of him since Zeemit’s
appearance.
“He was safe when I left him,” answered the old woman. “Soon after
leaving Meerut we were attacked in a bungalow, where we had
sought shelter; but we managed to escape, and continue our journey
to Delhi. We gained entrance to the city, and I soon learned from
some of the Palace servants that Jewan had gone to Cawnpore. We
lost no time in following him, and we arrived here last night. In
yonder clump of trees,”—as the old woman spoke, she slightly raised
her head, and pointed with her finger across the compound—“is a
disused bullock-shed. There, on a heap of straw, you will find Mr.
Gordon. He was to remain secreted until I had learned tidings of you.
He was weary and footsore, and sleeping soundly when I came
away.”
“But how am I to reach there unobserved?” asked Flora, scarcely
able to restrain her impatience.
“I think that will be comparatively easy. Go through the room here till
you gain the landing, then down the stairs until you come to the
entrance-hall. The night is dark, and you may easily make your way
to the bullock-shed. Once there, you and Mr. Gordon must lose no
time in hurrying to the protection of the English quarters; but, if
possible, fly from Cawnpore without delay, for there is an awful time
coming for the place. The native troops are pledged to rise, and the
Nana Sahib is thirsting for revenge.”
“God help us all out of our tribulation,” murmured Flora. “I will
endeavour to carry out your directions, Zeemit, but be sure that you
join us. It is against my will to leave you here, but we must bow to
the circumstances that we cannot alter.”
“Go—go,” murmured Mehal; “I am old, and you are young. Join your
lover, and seek safety in flight. I have no doubt we shall meet again;
but be discreet. Jewan is wary, and the moment he discovers your
escape, he will use every endeavour to recapture you.”
“Farewell, Zeemit,” said Flora, as she stooped and kissed the old
woman, “we part in sorrow, but I trust when next we meet, it will be
under happier circumstances. You have been miraculously
preserved from death, and no doubt it is for some wise purpose.
When we reach our English friends, I shall lose no time in sending
for you.”
A hurried shake of the hands, a few final whispered words of parting,
and Zeemit Mehal was left wounded and sick, lying alone under the
stars; and Flora Meredith, like a timid hare, was descending the
stairs.
On the various landings the natives were lying about asleep, a
custom common to the servants in India, who coil themselves up
anywhere. With noiseless tread, and rapidly beating heart, the
fugitive picked her way amongst the sleepers, turning pale with
alarm, as one moved here, and another groaned there, almost
entirely holding her breath, lest even the act of breathing should
awaken those whom she had such cause to dread. But after nearly
half-an-hour of the most painful and intense anxiety, she stood at the
main entrance of the building.
Day was commencing to break; there was sufficient light in the sky to
enable her to see across the compound. Not a soul was in sight.
Without a moment’s delay, she sped towards the clump of trees. The
bullock-shed indicated by Zeemit was soon reached. It was a very
dilapidated structure, built of bamboo and mud. She entered through
the doorway, and advanced cautiously for some paces; then
listened, for there was scarcely sufficient light in the hut to
distinguish anything plainly. The sound of heavy breathing fell upon
her ears. It came from the extreme end, where she could make out a
heap of straw. She went a little farther, and stood again.
“Walter!” she called softly; “Walter!” she repeated, a little louder.
But there was no reply. The sleeper slept, and the heavy breathing
was her only answer. She went nearer. The rustling of her own dress
alarmed her, for her nerves were unstrung.
“Walter!” she whispered again, as she reached the straw. Still no
reply. “He is worn and weary, and he sleeps heavily,” she murmured
to herself.
The light had considerably increased, for the day breaks in India as
suddenly as the night closes in. She was close to the sleeping form.
She stooped down until she knelt on the straw. She stretched
forward to waken the sleeper, but instinctively drew back as she
noticed the muslin garments of a native. She rose to her feet again,
advanced a little, bent down and peered into the face, the dusky face
of, as she thought, a Hindoo. She had come expecting to find her
lover—in his place was a native. She uttered an involuntary cry of
alarm, and, turning round, sped quickly away.
The cry penetrated to the sleeper’s brain. He turned uneasily, then
assumed a sitting posture, and, as Walter Gordon rubbed his eyes,
he muttered—
“Bless my life, how soundly I have been sleeping. I could have
sworn, though, I heard a woman’s cry. It must have been fancy.”
He stretched himself out once more on the straw; for many weary
miles had he travelled, without being able to obtain a moment’s rest,
and nature was thoroughly exhausted.
“Poor Flo,” he thought, as sleep commenced to steal over him again,
“I hope she will come soon. Zeemit is a faithful creature, and I have
no doubt will succeed. God grant it.”
Walter Gordon slept once more, and she for whom he sighed was
speeding from him on the wings of terror, into the very jaws of death.
CHAPTER XVIII.
WITH A LOVE THAT PASSETH UNDERSTANDING.
The signs of dissatisfaction which had alarmed General Wheeler for

the safety of his community gradually increased. The smothered fire
was gaining strength. It muttered and rumbled, and gave evidence
that a tremendous outbreak was imminent.
Sir Hugh was loath to believe in the infidelity of his troops, and
hesitated about taking steps for self-protection. But there were those
about him who had less of the optimist in their natures than he, and
who were loud in their condemnation of his supineness. They urged
him in every possible manner to take instant steps to place the
cantonments in a state of defence, until he could no longer turn a
deaf ear to their entreaties.
But though he had been slow to take this step, it must not be
assumed that Sir Hugh Wheeler was unmindful of the awful
responsibility that rested upon his shoulders. His was as brave a
heart as ever beat in human breast, but out of his very bravery arose
the danger to those under his charge.
He knew the character of the natives well. He knew that they writhed
under a sense of supposed wrong, and that the slightest touch will
cause an open wound to smart. He was, therefore, fearful of letting
them see that the English mistrusted them. He acted upon the old
principle that confidence begets confidence. Moreover, he had firm
faith in Nana Sahib. He knew that as a native the Rajah had infinitely
greater power over the native mind than an European could possibly
have had.
Sir Hugh’s confidence, too, seemed fully justified, for the Nana had
readily complied with the request made to him, and had posted two
hundred of his troops at the Newab-gung. This was a slightly
elevated position, and fully commanded the arsenal and treasury.
A couple of guns on the spot, served by determined and faithful
soldiers, could have kept a regiment at bay; but the fact of the
Nana’s assassins—for no other term is applicable to them—being
placed there was the very irony of fate. Into their hands had been
given a wealthy treasury, and a well-stocked arsenal. All they had to
do when the right moment came was to walk into these places, and
slay the English with their own weapons.
Listening at last—though reluctantly—to the entreaties of his people
General Wheeler looked about for the best means of securing his
position; and it occurred to him, in the emergency, that the only way
of defending the precious lives of the Christians was by throwing up
some defensive works, within which he might gather his people, so
that with their guns they could keep the enemy at bay.
He selected a spot for this purpose about six miles down the river to
the south-east, not far from the Sepoys’ huts, and about a mile from
the banks of the river. He was guided in this choice, to a great
extent, by the fact that on the spot were two long hospital barracks
that would make good quarters for the people. One of the buildings
was a substantial structure built wholly of masonry; but the other had
a heavy thatched roof.
Here, again, the cruel hand of Fate seemed to be, for a time, against
the English, for to the circumstance of the thatched roof some of the
most awful suffering endured by the besieged was due, as will be
hereafter shown. Both buildings were single-storied, and verandahs
ran all round them; they stood in an open and perfectly flat
compound. In the centre of the compound was a well, the only place
from which supplies of water could be drawn; and as will be
disclosed in the subsequent unfoldings of the story, this well was the
scene of almost unparalleled heroic deeds.
Having selected his place, Sir Hugh began to entrench it, and supply
it with a stock of provisions capable of feeding his people for several
weeks.
The so-called fortifications were paltry in the extreme, for the means
were not at hand to render them worthy the name. The earth-works
were only four feet high, and were not even proof against bullets at
the crest. The apertures for the artillery exposed both guns and
gunners; whilst, on all sides, adjacent buildings offered splendid
cover for the enemy. The excessive heat and dryness of the weather
had rendered the ground so hard that it could only be turned with the
greatest amount of difficulty, and by patient labour; and when it was
dug it was so friable that the cohesion necessary for solidity could
not be attained.
The month of May wore on; the expected mutiny did not occur. June
came in, and Sir Hugh then felt confident that all danger had passed;
and Lucknow being threatened, the General sent to the relief of the
neighbouring station a portion of his own little company of soldiers.
As these white troops crossed the bridge of boats, and set their
faces towards Lucknow, the natives fairly shook with suppressed
laughter as they thought what fools the English were. And at this
very time, Jewan Bukht and other agents of the Nana were visiting
the bazaars and the native lines, and fanning the smouldering fire to
flame.
Towards the latter end of May, there entered Cawnpore by the
pontoon bridge, two strangers. It was the close of a more than
usually sultry day, and the travellers, who were on foot, were dust-
stained and worn.
These travellers were Lieutenant Harper and Haidee. They had
come from Delhi—a long weary march; and along their line of route
they had experienced the greatest difficulty in procuring necessary
food and rest.
Nerved by the one all-powerful motive, Haidee had kept up, and
exhibited extraordinary powers of endurance. When her companion
sank exhausted from heat and thirst, this brave and beautiful woman
watched over him, encouraged him, and gave him hope. Her gentle
hand wiped his brow, her soft bosom pillowed his head. Her love for
him grew stronger each day. To lie at his feet, to pillow his head, to
watch him when he slept, was joy inexpressible to her. And yet
during this journey she never by a single word betrayed aught of the
strong passion which filled her heart; but every action, every deed
proclaimed it.
On his part he tried to think of her only as one who had befriended
him, and to whom it was his duty to offer such protection as lay in his
power. But on the road from Delhi he proved the weaker vessel of
the two, for the awful heat, aided by the want of proper rest and
sustenance, sorely tired him. She, on the other hand, inured from
birth to the heat, and strengthened by her great love for him, kept up
when he faltered, and exhibited, comparatively speaking, but little
weariness.
Hers was the devotion of a true woman; it was self-sacrificing, all-
absorbing, undying. Truly she had made him her star that gave her
only light. She had no selfish thought, except such selfishness as is
begotten by true love—for all love is selfish; it is its very nature to be
so. And yet this faithfulness made the man sad. He felt that he could
not return her love, however much he might admire her. However
much he might feel grateful, however great his worship for her
nobleness of nature might be, he must shut his eyes to her charms,
close his senses to her silent outpourings of love, for he was
another’s, and to that one he must be true, or feel that for evermore
the honour which was so very dear to him was sullied, and time
could never wipe out the stain again.
Often as he dragged his weary steps along, with the loving Haidee
by his side, he mentally asked himself if he was not pursuing a
phantom that was luring him to unknown danger. Had he done right
in setting his face towards Cawnpore, and could he justify the course
he had taken by any amount of logical reasoning? He was striving to
do his duty. If he failed, it would be through error of judgment, and
not through want of heart.
As the two travellers stood upon the Cawnpore bank of the river
Ganges, Harper gave vent to a sigh of relief. But Haidee seemed to
be pressed with a weight of sorrow.
“You do not seem well, Haidee,” Harper remarked casually, as he
observed the depressed look of his companion. “Your eyes are dull,
and your cheek is pale. What is the cause?”
She looked at him almost reproachfully, and her only answer was a
long-drawn sigh.
“What is the matter with you?” he asked again, with a good deal of
indifference in his tone; for, to confess the truth, his thoughts were
far away. He was racked with doubts and fears, and half-regretted
that he had yielded consent to come to Cawnpore, instead of
returning to his quarters at Meerut.
Her eyes glowed, and her face and neck crimsoned, as she
struggled to conceal the emotion which almost choked her, and
which his words had caused. Her sensitive nature was wounded by
his indifference, and she shrank away, as it were, like a startled
fawn.
“Why do you sting me?” she exclaimed, when she could speak.
“Sting you, Haidee! What do you mean?” as he turned upon her
quickly, and coming back again to a sense of his true position.
“Why do you ask me what is the matter, in a tone that betrays too
plainly that you take no interest in the question?”
“Nay, Haidee, there you wrong me.”
“Sooner would I wrong myself than you; but your words remain with
Haidee while your heart is far away.”
“My heart is divided, Haidee, and I give you all of it that I dare. You
are my friend. Every sacrifice I can make I will make for you, if it is
necessary. I will protect you with my life. I cannot do more.”
“Ah!” she sighed; “and yet you can ask me what it is that makes me
sad? There is sorrow at my heart; sorrow at the thought our journey
is ended, and you and I must probably part never to meet again.
That is what is the matter with me.”
“Forgive me, Haidee, if I have hurt you by my seeming
thoughtlessness. I assure you I had no intention of doing so. And
though our journey is for the present ended, do not say we shall part
for ever. You have grown precious to me as a noble, generous,
devoted woman; and I vow, by all that I hold sacred, that I will
endeavour never to lose sight of you as long as I live.”
She trembled with a nameless, pleasurable emotion; her nerves
vibrated like unto the strings of a harp that are swept with a strong
wind; for this man’s words were music to her. “I will endeavour not to
lose sight of you as long as I live.” Had he not spoken them? And
they sank to the deeper depths of her nature. They were like an elixir
of life, given to one whose strength was ebbing away. She yearned
for sympathy, and this man gave it to her. Her soul cried out for
kindredship, and it found it in him. What wonder then that she should
be taken captive?—that beat for beat her heart should answer his? It
is given to human beings to feel the burning rapture of love, but not
to solve its mystery; for it is a mystery as strange as the Sphinx of
old; as unsolvable as the cosmical problems which have puzzled
philosophers of all ages.
She loved him. Every look, every action, every tone betrayed that
she loved him with a true woman’s pure love. If it had sprung up
suddenly, it was none the less genuine or strong. She would have
been content to follow him, even if he, like the fabled “Wandering
Jew,” had been doomed to go on and on, restlessly and for
evermore. Still would she have followed, living in his shadow,
drawing her very life from his look and voice, sorrowing when he
sorrowed, laughing when he laughed. Nay, more; she would have
taken upon herself all the pains, however fearful, he might have had
to endure. She would have rendered that last and greatest sacrifice
that one human being can make for another—she would have laid
down her life to save his.
It was a grand love, this love of hers—not the sickly sentiment of a
wayward girl, but the strong, powerful, absorbing passion of a
woman; a love as heroic as any that Homer ever sang of, or that
moved the Roman women of old to follow the youths to the battle-
fields, and die when they died.

Ebook Quasi Interpolation Martin Buhmann Online PDF All Chapter

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ebook Quasi Interpolation Martin Buhmann Online PDF All Chapter

Uploaded by

Copyright:

Available Formats

Quasi Interpolation Martin Buhmann

Visit to download the full and correct content document:

The Statistical Analysis of Quasi Experiments

Diplomacy of quasi alliance in the Middle East Degang

Banach Space Valued Neural Network: Ordinary and

Banach Space Valued Neural Network Ordinary and

Vibration Protection Systems Negative and Quasi Zero

Spectral Invariants with Bulk Quasi Morphisms and

Lens Design Automatic and Quasi Autonomous

High-Order Harmonics Generation in Plasmas: Resonance

A complete list of books in the series can be found at

Recent titles include the following:

23. Curve and surface reconstruction, Tamal K. Dey

Cambridge University Press is part of the University of Cambridge.

3 Univariate RBF Quasi-Interpolants 26

5 Quasi-Interpolants for Periodic Functions 76

10.5 Compression in the Space of Continuous Functions 259

In the range of current applications of numerical mathematics and approximation

Important features are its approximation power (highly useful convergence

that is explicitly allowed to be zero, or thin-plate splines φ(r) = r 2 log r. The

by solving the linear system of equations

with the interpolation matrix

but the approximants are essentially from the same spaces.

Here Λ is a (ﬁnite or inﬁnite) set of indices usually closely connected to the

If we aim to reconstruct polynomials up to a certain degree, we naturally see that

2.1 Approximation Properties

With decreasing h, the amount of available information on f increases as the number

This is of course necessary if the quasi-interpolation operator Qh is to reproduce

2.1.1 Polynomial Reproduction and Strang–Fix Conditions

then F is a 2π-periodic function with Fourier expansion

We use the Schwartz space of inﬁnitely smooth functions γ̃ in n variables that

reproduces polynomials, so Qq = q, for all q ∈ P1m−1 , if

1. ϕ̂(u) 1 has a zero of order m at u = 0 and

2. ϕ̂(u) has zeros of order m for all u = 2π j ( j 0 is an integer).

which implies the polynomial reproduction.

Theorem 2.2 Let m be a positive integer and let ϕ : Rn → R be a function such

Then the quasi interpolant

is well-deﬁned and exact on the space of polynomials of degree m, and the

for h → 0 and a bounded function f ∈ C m+1 (Rn ) with bounded derivatives.

Conditions 2 and 3 are generally referred to as Strang–Fix conditions, and

2.2 Main Examples of Quasi-Interpolants

2.2.1 Quasi-Interpolants Based on Radial Basis Functions

where the fundamental functions ψ ξ are a (usually ﬁnite) linear combination of

For gridded data, higher-order polynomial reproduction is achieved when the

of an at most polynomially growing test function σ : Rn → R, satisﬁes the Parseval–

for all γ satisfying

which allows us to choose appropriate values of ξ j and μ j that ensure higher-order

2.2 is now satisﬁed for m = 0 if

ξ0 = (0, 0), ξ1 = (1, 0), ξ2 = (0, 1), ξ3 = ( 1, 0), ξ4 = (0, 1),

we get the following coeﬃcients:

and it reproduces constants.

Once we have formed the modiﬁed quasi interpolating basis functions, we

2.2.2 Polynomial Quasi-Interpolants

We can describe this classical operator as a quasi-interpolant of the general form

It can also be written as

where the coeﬃcient functional is accordingly

|Q m f (x) − f (x)| = O(1/m), |Q[2]

since Q m is exact on linear polynomials, while Q[2]

where the coeﬃcient functional λ [m]

where we deﬁne the functionals by integration (averages, in other words)

types of quasi-interpolants in Sablonnière (1999) will be featured and demons-

2.2.3 Polynomial Spline Quasi-Interpolants

λ k (x −1 ) = λ k for all 1 ≤ ≤ m, k ∈ {0, . . . , j}.

λ k (x −1 ) = λ k for all 1 ≤ ≤ m, k ∈ {0, . . . , j}.