Professional Documents
Culture Documents
Ebook Quasi Interpolation Martin Buhmann Online PDF All Chapter
Ebook Quasi Interpolation Martin Buhmann Online PDF All Chapter
https://ebookmeta.com/product/the-statistical-analysis-of-quasi-
experiments-christopher-h-achen/
https://ebookmeta.com/product/diplomacy-of-quasi-alliance-in-the-
middle-east-degang-sun-author/
https://ebookmeta.com/product/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/
https://ebookmeta.com/product/vibration-protection-systems-
negative-and-quasi-zero-stiffness-1st-edition-chang-myung-lee/
https://ebookmeta.com/product/spectral-invariants-with-bulk-
quasi-morphisms-and-lagrangian-floer-theory-1st-edition-kenji-
fukaya/
https://ebookmeta.com/product/lens-design-automatic-and-quasi-
autonomous-computational-methods-and-techniques-2nd-edition-
donald-c-dilworth/
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.
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
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
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
Quasi Interpolants 43
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
Quasi Interpolants 73
4.10 Notes 74
viii Contents
References 266
Index 274
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
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
{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,
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
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.
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 9
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
2.1 Approximation Properties 11
1 ij x ˆ
e f ( j), x ∈ Rn .
(2π) n j ∈Zn
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
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=−∞
12 Generalities on Quasi-Interpolation
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=−∞
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 .
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.
⎧
⎪
⎪ O(h m+ ), when 0 < < 1,
⎪
⎪
Qh f f ∞ = ⎨
⎪ O(h m+1 log(1/h)), when = 1,
⎪
⎪
⎪O(h m+1 ),
⎩ when > 1,
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
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.
2.2 Main Examples of Quasi-Interpolants 15
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
(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
16 Generalities on Quasi-Interpolation
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,
2.2 Main Examples of Quasi-Interpolants 17
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
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
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
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
2.2 Main Examples of Quasi-Interpolants 21
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
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
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
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)!
24 Generalities on Quasi-Interpolation
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].
2.3.1 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
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.
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
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=−∞
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.
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,
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
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
(1 + h 1 c)ω f (h),
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.