Sie sind auf Seite 1von 15

IMA Journal of Numerical Analysis (1989) 9, 405-419

An Introduction to Lattice Roles and their Generator Matrices

J. N. LYNESS
Mathematics and Computer Science Division, Argonne National Laboratory,
9700 South Cass Avenue, Argonne, IL 60439-4844, USA

[Received 31 May 1988 and in revised form 30 October 1988]

For the one-dimensional quadrature of a naturally periodic function over its


period, the trapezoidal rule is an excellent choice, its efficiency being predicted

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


theoretically and confirmed in practice. However, for 5-dimensional quadrature
over a hypercube, the s-dimensional product trapezoidal rule is not generally cost
effective even for naturally periodic functions. The search for more effective rules
has led first to number theoretic rules and then more recently to lattice rules. This
survey outlines the motivation for and present results of this theory. It is
particularly designed to introduce the reader to lattice rules.

0. Introduction

LATTICE rules are generalizations of the number theoretic quadrature rules of


Korobov (1959) for integration over an s-dimensional hypercube. They were first
introduced by Sloan & Kachoyan (1984, 1987). The rules use an abscissa set of
points that lie in the intersection of the hypercube and an integration lattice. The
theory presented previously (some of which is described in more detail in Section
7) was based firmly on the geometric character of the lattice. Up to now, little use
has been made of the algebraic properties of the lattice generator matrices.
In this paper we develop a new approach, based on the definition of a lattice
through its generator matrix. This has led to two new characterizations of an
integration lattice. One (Definition 3.3) demands that it be a sublattice of a scaled
version of the unit lattice. The other is simply that its reciprocal lattice should
have a generator matrix composed of integers (Theorem 5.2). These definitions
have led in turn to a self-contained and relatively uncomplicated theory; this
involves a particularly simple proof of the Poisson summation formula for lattice
rules (Corollary 6.1).
In Section 1 we describe the one-dimensional lattice rule (the trapezoidal rule)
and so cover completely the one-dimensional specialization of all the theory in
this paper. In Section 2 we introduce the problems that arise when one attempts
to generalize the one-dimensional theory. Sections 3 to 5 then deal with the basic
theory of lattice rules; we treat lattices, integration lattices, lattice rules, the
reciprocal lattice, and integer lattices in that order. This material is mainly a
question of interrelated definition. Then in Section 6 we return to the analysis of
the quadrature rule error and show that many terms in the Poisson summation
formula drop out naturally when it is applied to a lattice rule.

© Oxford Univtnity Prea 1989


406 J. N. LYNESS

Section 7 is devoted to a brief outline of previous and current work on lattice


rules.

1. The one-dimenskmal background


The statement that the trapezoidal rule is good for periodic functions is often
made. In this section we discuss this statement. This provides a suitable
background for the subsequent introduction to lattice rules.
We shall treat an integrand function f(x) that is ( ^ [ 0 , 1 ] with p s= 0. Let / / be
its exact integral over the integration interval [0,1], and let/(x) be the periodic
continuation of f(x) defined as the sum of the Fourier series of f(x) with respect
to this interval. We shall denote the Fourier coefficients by

M for all integer r. (1.1)


Jo

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


Let
Qf=i^f(xj) with i > y = i (i.2)
y-i y-i

be a general quadrature rule, and

"'"' -; M>+%t (£) + H - ip(£


be the m-panel trapezoidal rule. These are both approximations to / / = ao(f), the
exact integral. By substituting the Fourier series of f(x) into (1.2) we find a
general form of the Poisson summation formula, namely,

Q? = If+ £ «,a)G(e2-te) (1.4)


r*0

where the multiplier of the rth Fourier coefficient is

dr(Q) = Q(e2"i") = £ w, exp Qnirx,), (1.5)


7-1

i.e. the result of applying the rule Q to the indicated exponential function.
The Poisson summation formula (1.4) expressed the quadrature error Qf'-If
as a series. We seek to construct Q so that the most significant terms in this series
drop out. These are usually the earliest terms. To do this, we seek Q that for a
fixed value of v annihilates as many as is feasible of the coefficients dr(Q)- This
turns out to be a well-defined problem. The solution is a one-parameter system of
rules that comprise /? (m) together with all its offset (displaced) versions, with
m = v or v — 1. We find
dr(Rim)) = l when r/m = integer, and 0 otherwise. (1.6)
Consequently, by applying (1.4) to the trapezoidal rule, we have found that

R™f = If+ £ aml(f). (1.7)


(— — oo
LATTICE RULES AND THEIR GENERATOR MATRICES 407

This rule then is clearly suitable for functions fix) whose Fourier coefficients
ar(f) decay rapidly with r. It is relatively well known that the decay rate is
connected with the degree of continuity of j(x). An immediate demonstration of
this statement follows by applying integration by parts to the right-hand side of
(1.1) to obtain the Fourier coefficient asymptotic expansion

valid when / e C* p) [0,1]. It is, of course, also valid when / e C^^O, 1) in which
case/^*" 1 ^) and/ ( f l - 1 ) (l) should be replaced by the appropriate right-hand and
left-hand derivatives respectively. Since J(x) is periodic we have
(1.9)

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


which, of course, is zero if the (q — l)th derivative of fix) is continuous.
THEOREM 1.1 (See e.g. Copson, 1965; page 21.) When fix) and its first p-\
derivatives are continuous,
aril) = (f(p\0+) - / ^ ( 0 - ) ) / ( 2 j r i r y + 1 + o(r^p+1)). (1.10)
When the derivatives of f(x) of all orders are continuous, this implies that ar(J)
decays more rapidly than any inverse power of r. One can show that when fix) is
analytic in the strip |Imx| < L, there exists M such that
\ar(f)\<Me~Mr. (1.11)
These results about the rate of decay of the Fourier coefficients imply that, when
f{x) is not C , this rate depends on the degree of continuity of fix), and when
fix) is analytic, this rate depends on a measure of the magnitude of the region of
analyticity.
The statement that the trapezoidal rule is good for naturally periodic functions
is an abbreviation for the following set of statements.
(a) The trapezoidal rule is better for functions fix) whose Fourier coefficients
decay more rapidly.
(b) These are functions whose periodic continuations fix) have higher degrees
of continuity.
(c) The appearance of fix) is more naturally periodic according as /(x) has
higher degree of continuity.
This one-dimensional theory contains three principal features which we shall
want to generalize to s dimensions:
(i) The Poisson summation formula (1.4) which expressed the error Qf — If
as a sum of terms ar(J)driQ)- This is easy to generalize.
(ii) An implied ordering of the Fourier coefficients. That is, we made the
readily justifiable assumption that ar(J) is likely to be more significant
than a,(J) when \r\ < \t\. There is no self-evident generalization in s
dimensions. In Section 2 we simply report three somewhat arbitrary but
reasonable possibilities, one of which has been widely studied.
408 J. N. LYNESS

(iii) The derivation of the rule Q which 'zeroed out' the coefficients dr(Q)
associated with the more significant Fourier coefficients ar(J). In one
dimension, the result that /? (m) is the solution has the status of a theorem.
In s dimensions, the corresponding nonlinear equations are more compli-
cated, and no general solution is known. The reason they are more
complicated stems from the more involved ordering in (ii) above. Sections
2-6 are devoted to describing the nature of ad hoc attempts to find
apparently reasonable rules. These include number theoretic rules and
lattice rules.

2. Problems arising in the s-dhnenskHial generalization


The J-dimensional generalization of features (i) and (ii) at the end of Section 1
has been treated extensively in the theory of number theoretic rules, and we

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


have little to add. In the context of lattice rules, this theory is covered in Lyness
(1988) using the same notation as is used here. Briefly, the one-dimensional
formulas (1.1), (12), (1.4), and (1.5) have obvious ^-dimensional generalizations.
We treat the unit hypercube H; [0,1) J ; the Fourier coefficients are

~"du, du2 • • • du, (2.1)


JH

and the Poisson summation formula takes the form

Qf = lf+2atf)dr(Q) (2.2)
re/V)

where as expected

dr(Q) = Q(e2nir-X) = S w, exp (2nir • x,). (2.3)

In (2.2) we have used the terminology re A^to indicate that all components of r
are integers. (Later we shall introduce the unit lattice A, to denote precisely this
set of points.) The details about the convergence of the Poisson summation
formula are complicated but play no significant role in this theory, /(x) is the sum
of the ^-dimensional Fourier series of/(x). This is a periodic continuation of /(x)
and may be defined by

/(*)=i 2 2 • • • 2 /(*./,. y^-.y.j)


C
(2.4)
7i-0 ;j-0 j,-0

where
[{**} xk * integer,
^*J 1/ xk = integer.
Here we have used braces to denote the fractional part. This somewhat
cumbersome definition of/(x) disguises an intrinsically simple concept, and the
difference between /(x) and /(x) also plays no significant role in this theory.
LATTICE RULES AND THEIR GENERATOR MATRICES 409

We now turn to step (ii) of the list at the end of Section 1, that of deciding
which of the s-dimensional Fourier coefficients ar(J) are the more significant. We
need this to establish a rule construction criterion, in this case a criterion for
deciding which coefficients dr(Q) should be 'zeroed out'. This problem is
discussed in some detail in Lyness (1988). Let us define three measures.
P(r) = fxr2 • • • f, where f, = max (1, \r,\)
S(r) = |r,| + \r2\ + • • • + \rs\ (2.5)
M(r) = max (|r,|).

(P, 5, and M are the product, the sum, and the maximum respectively of the
non-zero components of r.) The traditional criterion in this area (Korobov, 1959;
Zaremba, 1972) incorporates the view that

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


ar(f) is more significant than a,(f) when P(r) < P(t).
For functions that have low-order derivative discontinuities, this statement is
realistic. For naturally periodic functions, e.g. /(x) analytic in all components,
examination of Fourier coefficients suggests strongly that one should use S in
place of P. The use of M, which leads to the product trapezoidal rule, has little to
recommend it. Further information is available in Lyness (1988). Corresponding
to each measure, there are indices p and 5 to indicate to what extent a rule Q has
'zeroed out' dr(Q)- Thus,
Q has Zaremba index p when dr{Q) = 0 for 1 s£ P(r) < p except r = 0
Q has overall degree 6 when dr(Q) = 0 for 0 < S(r) =e 6 (2.6)
Q has product degree 6 when dr{Q) = 0 for 0 < M{r) =s 6.
Note that in all cases d9(Q) = 1.
In one dimension all these three measures become the same, reducing to the
one used in Section 1.
Step (iii) of the list at the end of Section 1 involves determining cost effective
rules that satisfy the construction criterion we have chosen.
If we choose the measure M(r) above, there is an exact s-parameter solution;
this is the .s-dimensional product of s displaced trapezoidal rules, each having m
panels. The parameters are the displacements which may all be different. Any
such rule which uses m' function values of/is of product degree m — 1.
In contrast, the measures P(r) and S{r) are difficult to handle. An analytic
solution requires solving large, somewhat ad hoc sets of nonlinear equations. As
is usual in this sort of situation, progress continues only on a trial and error basis.
One makes an educated guess about the form of what one thinks might be a
suitable rule. Then one tries to choose the incidental parameters to maximize the
degree 6, or the Zaremba index p. And, of course, progress in this search will be
quicker, at least in the early stages, if one chooses the rule form so that 6 or p is
relatively easy to measure - or to bound.
410 J. N. LYNESS

In general, at first sight, for every candidate rule Q, one has to calculate

= £ w^i (2.7)

for various values of r = (rlv.., r,), searching for a rule for which many of these are
zero. The first 'simplification' we can make is to consider only rules whose
one-dimensional projections are /V-panel trapezoidal rules. Then at least some of
the dr(Q) are zero, i.e. those having 5 — 1 zero components and one nonzero
component less than N. Such a rule would use some, but not necessarily all, of
the abscissas on the lattice (l/N)Ao, i.e. abscissas for which all components are
integer multiples of 1//V. We start by considering rules using TV points which have
this property.
An obvious candidate is a number theoretic rule

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


Qf = {VN)'2KjzlN), (2.8)
i-i

where z = (zx,..., z,) has integer components. Let us assume for this illustration
that TV is a prime number, and 0 < zt < N for all i. It is easy to verify that all
one-dimensional projections are TV-panel trapezoidal rules. And with this
structure we find many zero coefficients. Specifically,

If we consider only rules of the form (2.8) defined by parameters TV and z, we can
determine which dr(Q) are zero simply by seeing whether r • z/N is an integer, a
process that is much faster than carrying out the sum in (2.7).
There is a vast literature devoted to number theoretic rules. Major contributors
include Korobov (1959, 1960), Hawka (1962), Zaremba (1966, 1972), Niederrei-
ter (1978), Hua & Wang (1981), and Haber (1983). A reference list with over
seventy entries may be found in Niederreiter (1988). We now return to the
comparatively simple general theory of lattice rules.

3. Lattices and integration lattices

A lattice rule is a quadrature rule that uses as abscissas a finite selection of


points of a particular kind of lattice. In this section we introduce lattices.
An 5-dimensional lattice is an infinite array of points in j-dimensional space
having certain regularity and periodicity properties. The simplest example is the
unit lattice, denoted by AQ which comprises all points having integer components.
A scaled unit lattice denoted by 6A^, comprises all points all of whose
components are integer multiples of «5. It is convenient to define the unit vectors
e) (J = 1,..., s). The ;th component of et is 1 and the other components are all zero.
Thus the scaled unit lattice 6AQ comprises all points p of the form p = E/_i
with ft = (/J,,..,/i,)6 4
LATTICE RULES AND THEIR GENERATOR MATRICES 411

A widely used definition of a lattice follows.


DEFINITION 3.1 A lattice A comprises a set of points such that
p,qeA => p + qe A andp-qe A (3.1)
and there exists no limit point, i.e. there exists positive e(A) such that
\p-q\** e(A) unless p = q. (3.2)
Corresponding to every s-dimensional lattice A is a non-singular 5 X 5 generator
matrix A . The rows of A , namely, ar = (a r>1 , ar2,-,ar^), are elements of A
chosen in such a way that
peA O p = 2^r (3.3)

for some A = (A^..., A,) e AQ. The rows of a generator matrix are sometimes

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


termed a 'set of generators of A'. The right-hand relation of (3.3) may be written
(3.4)
LEMMA 3.1. Any nonsingular s xs matrix A generates an s-dimensional lattice A.
Proof. Qearly the set of points (3.3) satisfy (3.1). When A is nonsingular we
have
\iA\2^o\M2^o VA#0,
where a is the smallest singular value of A. This establishes (3.2).
These results about generator matrices are introduced and discussed in greater
detail in most texts on the geometry of numbers (see, e.g. Cassells, 1959:
Chap. 1).
There are many different generator matrices corresponding to the same lattice
A. To see this we note that the set of points defined by (3.4) is unaffected when
one replaces
fl by a
' "' (3.5)
ar by ar + \ia, (r*t, /z = integer).
So one may apply elementary row operations (using integer coefficients) to a
generator matrix without altering the lattice A. It can readily be shown that any
generator matrix of A may be obtained from any other generator matrix of A
using elementary row operations with integer coefficients. It follows that:
(i) if all elements of one generator matrix of A are integers, then all elements
of any generator matrix of A are integers;
(ii) the transformations do not alter |detA| which is a constant. Thus,
N = N(A) = |det A\~l depends on the lattice.
The interpretation of this constant follows from treating the matrix A as a space
transformation p T = ATAT from A-space, in which is embedded the AQ lattice, to
p-space, in which is embedded the A lattice. These lattices are in one-to-one
correspondence with each other. Thus |det A\ is the ratio of the point densities in
412 J. N. LYNESS

these two lattices. Since the point density of A) is 1, |det A p 1 is the point density
of A. We shall state a particular case of this fundamental result in Theorem 3.1.
In some of the proofs below, we denote a scaled up or down version of a lattice A
by kA where A: is a nonzero scalar. (The corresponding generator matrices are A
and kA.)
We shall now define a special class of lattice.
DEFINITION 3.2 An integration lattice A is a lattice that contains the unit
lattice AQ.
The rest of this paper, and the theory of lattice rules, is exclusively concerned
with integration lattices.
We shall see below that, when A 3 AQ, the structure of the lattice is periodic
with period 1, and IdetAp 1 is precisely the number of points N(A) in the unit
hypercube. We restate this critical result as a theorem.

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


THEOREM 3.1 When A is a generator matrix of an integration lattice A, i.e.
A 3 AQ, then the number of points of A in the hypercube [ 0 , 1 / is precisely

Definition 3.2 is a little cumbersome in some cases because, to verify that a set of
points forms a lattice, one has to verify the criterion (3.2) about the absence of
limit points. It is convenient to replace this criterion as follows.
We define
djx = min {\pf\ : Pj # 0, p e A}.
i.e. djx is the smallest nonzero xt component of any lattice point. For a general
lattice, dj need not exist. But since A => AQ, it follows that dj exists and is an
integer. We may take d = lcmy (dt) in the following definition.
DEFINITION 3.3 An integration lattice A comprises a set of points such that
p,q GA => p + q e A and p — q e A
and there exists an integer d such that
d-'Ao^A^Ao. (3.6)
This means that every point of A lies on the scaled unit lattice, and that every
point of AQ is also a member of A. The advantage of this definition is that it is
self-contained; there is no reference to limit points and it does not require the
prior definition of a lattice (except the scaled unit lattice).
Yet another definition of the same concept follows.
DEFINITION 3.4 An 5-dimensional integration lattice A may be denned by any set
{n,,zi :i-l,...,t},
where t and n( are positive integers and zt e AQ. A comprises all points of the form

P = 2 W « i + X WJ (3-7)
LATTICE RULES AND THEIR GENERATOR MATRICES 413

where A, and /iy may be any integers. Note that the same point p may be defined
using different selections of A, and \Xj. Nevertheless, it is trivial to see that, if we
take d = n^n2 • • • n,, all the postulates of Definition 3.3 are satisfied.
In Definition 3.4, we employed t + s parameters A, and \i-t to describe an
j-dimensional system. Naturally, when the generator matrix definition (3.4) is
used, only s parameters are required. The question arises as to what condition on
a generator matrix A is required to ensure that A is an integration matrix. The
answer is that A~l should have only integer elements. This is one of the most
crucial facts of the theory and is proved in Section 5.
We have mentioned that the integration lattice A having a generator matrix A
has precisely N = |det A\~l points within the unit hypercube H = [0,1) 1 . Let us
denote these points by
A n / / = {x, ,...,*„}. (3.8)
Since the lattice contains the unit lattice AQ, it follows that every point in the

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


lattice has the form
/>=** + 2 V ; with A: e [1, N]. (3.9)
y-i

The lattice is, of course, infinite in extent. But it is periodic with period 1 in each
direction. If one imagines all space to be subdivided into unit hypercubes, one of
which is Ily-i [A;, A, + 1) for integer Ay, the pattern of points of A within each
hypercube is identical to the pattern in the unit hypercube [ 0 , 1 / . It is this basic
set of points that is used in the lattice rule.

4. The lattice rale Q(A)


We are now in a position to define a lattice rule Q(A). Given an integration
lattice A, we define the abscissa set A(Q) as the set in (3.8), i.e.
nH. (4.1)
Then the lattice rule is

l ^ 2 f(p) (4-2)
with/(x) being the periodic continuation of f(x) defined in (2.4). Here v(Q) = N
is the order of the set A(Q). This is the number of function values of/required to
evaluate Qf. (The number of function values of / required exceeds \{Q) by at
least T — 1 since all lattice rules require /(0) which is the average of f(x) taken
over all ¥ vertices of H.)
A completely self-contained definition of the abscissa set for a lattice rule is the
following. We define the 'fractional part' of a vector in a standard way, i.e.
y = {x} when y e H and y — x e AQ. The abscissa set comprises N points xh with
xteH and x, E /V"'AI
having the property that
*,-,*, 6i4(G) => {or, + xt) eA(Q) and {x, -x,} e A(Q). (4.3)
414 J. N. LYNESS

The abscissa set A(Q) forms a group of order N under the operation 'plus' where
x 'plus' y = {x + y}. This aspect of the abscissa set is treated in considerable detail
in Sloan and Lyness (1989).
A graphic description of the construction of a lattice rule is the following. One
chooses an integer mesh ratio d and chooses t points Xi,...,x, on the lattice
d - 1 / V Then one constructs the lattice A by including 0 and all points «,,
j = l,...,s, and all points obtainable from any existing pair using ordinary vector
addition and subtraction. The abscissa set A(Q) is the set of N points of A lying
in H = [0 , iy, i.e. those whose components all lie in [0,1). And the rule is given
by (4.2) above. Incidentally N is a factor of or a multiple of d. In this process, if
we had chosen / = 1 we would have found a number theoretic rule (2.8) above.
However, with O l w e might have found a rule not expressible in that form.
If one returns to Definition 3.4 of an integration lattice, one might define

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


*k=zklnk (k = l,...,t),
and generating the lattice then leads to the following form for a lattice rule.
THEOREM 4.1 (Sloan & Lyness, 1989) The form {t-cycle)

Qf=

where z, e AQ and nt and t are positive integers, is a lattice rule, and all lattice rules
can be written in this form. This is called a t-cycle form if n^,..., n, are each greater
than 1.
Note that, since f(x) =/({*}), the effect of the generators ejt j = l,...,s, is
automatically taken into account. It is readily verified that all the points in A n H
are included in this finite sum. The only nontrivial problem concerns the
circumstance that points may be counted more than once. However, one may
show that if this set includes 0 more than once, say k times, then it includes every
other point precisely k times, and vice versa. In this case we term form (4.4) a
^-repetitive form of Q. While it would be ridiculous to program directly the
^-repetitive form, it can be used with confidence to derive theoretical results. We
shall make use of this form in Section 6.
To obtain a form (4.4) directly from the /4-matrix, one sets
Zj/n, = aj, (4.5)

where Z/ e A>- This is possible since we recall each element of A is a rational


number. Using these values of zt and nt in (4.4) gives an s-cycle form that is likely
to be repetitive. This is easy to prove, but not completely self evident. One has to
convince oneself that the integers ns provided by (4.5) are suitable as upper limits
on the summations in (4.4).

5. The reciprocal lattice AL


Corresponding to every lattice A is its reciprocal (polar, dual) lattice,
sometimes denoted by A x . It may be defined as follows.
LATTICE RULES AND THEIR GENERATOR MATRICES 415

DEFINITION 5.1 r e A x if and only if


r 'p = integer for all p e A. (5.1)
One has to show that, given A, points r defined above exist, and then that they do
indeed form a lattice. This is essentially established in Theorem 5.1. After this, it
is straightforward to derive a host of relations of which the following are only a
selection.
(A^ = A; (dAoy = d-1Ao
( }
'
Naturally, the generator matrices of a lattice and its reciprocal lattice are closely
related.
THEOREM 5.1 When A has generator matrix A, then A x has generator matrix
T

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


B = (A y\
Proof. Denote by A x the reciprocal lattice of A and by A(B) the lattice whose
generator matrix is B. We have to show A x = A(B).
(a) When r e A(B) and p e A we have
p • r = kABTfiv = kfiy = integer
which, according to (5.1), is the condition that r e A x .
(b) When r e A x , since aj e A we have aj • r = /^ = integer and so

giving

This, according to (3.3), is the condition that r e A(B). O


We now treat the case when A is an integration lattice; then B has a special
property described below.
THEOREM 5.2 A nonsingular s xs matrix A is the generator matrix of an
integration lattice A if and only if B = (A T ) - 1 has only integer elements.
Proof. Nearly all the proof follows from applying the second and third members
of (5.2) to A and AQ giving
A2Ao O A-'-cAo. (5.3)
x
(a) If A(A) is an integration lattice, it satisfies the left-hand relation. A whose
generator matrix is B then satisfies the right relation. Thus all elements of
A x have integer components. Since the rows of B are elements of A±, it
follows that the elements of B are all integers.
(b) Since B is nonsingular, Lemma 3.1 assures us that A(B) exists and since it
has integer elements, Ao3 A(B). Since A = (B7)'1, A(A) is the reciprocal
lattice of A(B). Thus A(A) 3 AQ and so is an integration lattice. D

The reciprocal lattice of an integration lattice, like the integration lattice, is


periodic in structure. We note that, when N = |det A\~l is the number of points of
416 J. N. LYNESS

A lying in [0,1) J , we find that AL has period N. That is, A x has Ns~x points ry
lying in [0fNY and this pattern of points is repeated in every hypercube
II/=i [kjN , (A; + 1)N) such that XEAQ. Much of the burden of constructing lattice
rules lies in examining this reciprocal lattice to find points near the origin. The
large number N*~* of points that may be involved in a cursory search adds to the
computational difficulty of this task.

6. The Poisson summation formula of a lattice rale


The reader cannot have failed to notice that, after Section 2, no results
pertaining to the accuracy or feasibility of lattice rules have been presented.
Sections 3,4, and 5 dealt with the concepts of lattice, lattice rules, and reciprocal
lattice, without mentioning at all why such concepts are relevant. In this section
we return to the analysis.

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


We recall that our error analysis is based on the Poisson summation formula
(2.2), namely

Of -V = 2 arif)dr(Q),

and our concern is to reduce this error by choosing Q in such a way that the
coefficients dr(Q) in the significant terms vanish. The basic result of this section is
Theorem 6.1 which links these coefficients to the reciprocal lattice A x . Surpris-
ingly enough, the most straightforward proof of this theorem stems from the
redundant but graphic Definition 3.4 of the lattice and the corresponding f-cycle
form (4.4) of the lattice rule. We recall that dr(Q) is the result of applying Q to
the function exp(2jtir -x). Replacing/(r) by this function in (4.4) gives
</,((?) = G(exp2ra>-x)

THEOREM 6.1 Let Q be a lattice rule having lattice A. Then


whenreAx, (6.2a)
whenr$Ax. (6.2b)
Comment. The search for efficient lattice rules is based on constructing rules for
which dr(Q) = 0 for r satisfying properties outlined in Section 2. Thus the whole
thrust of the construction is shifted to examining properties of the reciprocal
lattice. A second comment is that (6.2a) is much easier to establish than (6.2b).
Proof. Expanding (6.1) we find

dr(Q) = Qie2""-*) = II ^ 2 exp (2**7* ^ ) . (6.3)


The kth term of this product is 1 or is 0 according as zk • r/nk is an integer or is
not an integer. Consequently, there are only two possible values for dr(Q). These
LATTICE RULES AND THEIR GENERATOR MATRICES 417

are
when zk • r/nk = integer for all k = 1,..., /,
(6-4)
otherwise.
We now apply (6.4) to the reciprocal lattice Ax.
(a) When r = A±, then by Definition 5.1 r«x = integer for all x e A. Since
each of zk/nk is an element of A, it follows that zk • r/nk = integer,
it = 1,...,/, and in view of (6.4) dr{Q) = 1.
(b) Conversely, when dr(Q) = 1, each term in the product (6.3) is 1. Thus
zk • r/nk = integer, k = 1,..., f. It follows that, for a general abscissa x of
(4.4), x • r = E*=i/*z* • r/nk is also an integer. Since x • r is an integer for
all x e A, it follows that reAx.
Parts (a) and (b) above establish that reAx is a necessary and sufficient
condition for dr(Q) = 1- In view of (6.4) the only other value which dr(Q) can

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


take is zero. This establishes (6.2). •
COROLLARY 6.1 For a lattice rule Q(A)f, the Poisson summation formula is

where Ax is the reciprocal lattice of A.


We may now reconsider the situation described in Section 2 where we
introduced the Zaremba p-index and the trigonometric degrees of a rule Q. If we
are dealing with lattice rules Q(A), these degrees are simple properties of the
reciprocal lattice A±. The search for rules that are cost effective using any of
these definitions may be accomplished by searching reciprocal lattices. Each point
on such a lattice corresponds to an error term. We look for lattices whose lattice
points are not close to the origin; here the term close may be defined in terms of
P(r), S(r), or M(r) of Section 2.

7. Other work on lattice rules

The author is developing further theory based on the approach of this article,
that is, exploiting various properties of the generator matrices A and B of the
lattice A and its reciprocal. This work seems at present to be almost orthogonal to
two other areas in which significant progress has recently been made.
The first, chronologically, is work by Sloan & Kachoyan (1984, 1987). A
convenient review of that work is given in Sloan (1985). This generalizes to lattice
rules some of the deeper number theoretical results previously established only in
the context of number theoretical rules (sets of good lattice points). An example
is a generalization of a result due to Hua & Wang (1981); in the notation of (2.5)
and (2.6), this is to the effect that
Qf-If^cd(s, *)p~' r (l + lnp) tr ~ 1
when f(x) has Fourier coefficients satisfying
418 J. N. LYNESS

The second area, currently under investigation, exploits the group character
(see (4.3)) of the abscissa set A(Q). This leads to a canonical form for a lattice
rule expressed in the form (4.4). Specifically, every rule has a set of positive
integer invariants nx,...,ns. Each divides the previous one, and Qf, expressed in
form (4.4) using these values of nh is nonrepetitive. The rule is termed to be of
rank t when n, > 1 and nt = 1 for / > t. The number theoretic rules of Korobov are
all rank 1. Other rules, including the product trapezoidal rules, are of higher
rank. These ideas are developed in a sequence of articles (Sloan & Lyness, 1989)
and summarized in Sloan & Walsh (1988).

Acknowledgment

It is a pleasure to acknowledge the cooperation of Drs Tor S0revik and Pat


Keast in preparing this article. This work was supported by the Applied

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011


Mathematical Sciences subprogram of the Office of Energy Research, US
Department of Energy, under contract W-31-109-Eng-38.

REFERENCES
CASSELS, J. W. S. 1959 Introduction to the Geometry of Numbers. Berlin: Springer-Verlag.
COPSON, E. T. 1965 Asymptotic Expansion. Cambridge: Cambridge University Press.
HABER, S. 1983 Parameters for integrating periodic functions of several variables. Math
Comp. 41, 115-129.
HLAWKA, E. 1962 Zur angenaherten Berechnung mehrfacher Integrale. Monatsh. Math.
66, 140-151.
HUA Loo KENG, & WANG YUAN 1981 Applications of Number Theory to Numerical
Analysis. Berlin: Springer-Verlag.
KOROBOV, N. M. 1959 The approximate computation of multiple integrals (Russian).
Dokl. Akad. Nauk. SSSR 12A, 1207-1210.
KOROBOV, N. M. 1960 Properties and calculation of optimal coefficients (Russian). Dokl
Akad. Nauk SSSR 132, 1009-1012; Soviet Math. Dokl. 1, 696-700.
LYNESS, J. N. 1988 Some comments on quadrature rule construction criteria. In:
Numerical Integration III (G. Hammerlin & H. Brass, Eds), ISNM vol. 85. Basel:
Birkhauser Verlag, 117-129.
NIEDERREITER, H. 1978 Quasi-Monte Carlo method and pseudo-random numbers. Bull.
Amer. Math. Soc. 84, 957-1041.
NIEDERREITER, H. 1988 Quasi-Monte Carlo methods for multidimensional numerical
integration. In: Numerical Integration III (G. Hammerlin and H. Brass, Eds), ISNM
vol. 85. BirkhSuser Verlag, 157-171.
SLOAN, I. H. 1985 Lattice methods for multiple integration. J. Comp. Appl. Math. 12,13,
131-143.
SLOAN, I. H., & KACHOYAN, P. 1984 Lattice for multiple integration. Proc. Cent. Math.
Anal. Aust. Natl. Univ. 6, 147-165.
SLOAN, I. H., & KACHOYAN, P. J. 1987 Lattice methods for multiple integration: theory,
error analysis and examples. SIAM J. Numer. Anal. 24, 116-128.
SLOAN, I. H., & LYNESS, J. N. 1989 The representation of lattice quadrature rules as
multiple sums. Math. Comp. 52, 81-94.
SLOAN, I. H., & WALSH, L. 1988 Lattice rules - classification and searches. In: Numerical
Integration III (G. Hammerlin and H. Brass, Eds), ISNM vol. 85. Birkhauser Verlag,
251-260.
LATTICE RULES AND THEIR GENERATOR MATRICES 419

ZAREMBA, S. K. 1966 Good lattice points, discrepancy and numerical integration. Ann.
Mat. Pura Appl. 73, 293-317.
ZAREMBA, S. K. 1972 La me'thode des "bons treillis" pour le calcul des int6grales
multiples. In: Application of Number Theory to Numerical Analysis (S. K. Zaremba,
Ed.) London: Academic Press, 39-119.

Downloaded from imajna.oxfordjournals.org by guest on January 27, 2011

Das könnte Ihnen auch gefallen