lng.

Thornas PARISINI
Universita di G';T'Y, " i,::)"o","r;',o tIi Informatica
$.:,.:' .. ;
Via all'Cpt";, 13 . 161 .. 5 G..:'DO"JS" !talia
Tel. + jl) lO 353 27.99 - fax + 3::1 10 353 29.41
thomas @ rust. unige.it
Harald Niederreiter
Austrian Academy of Sciences
n N er
r tion an
nte Carlo
t s
SOCIETY FOR INDUSTRIAL AND APPLIED MATHEMATICS
PHILADELPHIA, PENt--ISYLVANIA 1992
CBMSNSF REGIONAL CONFERENCE SERIES
IN APPLIED MATHEI\J1ATICS
A series of lectures on topics of current research interest in applied mathematics under the
direction of the Conference Board of the Mathematical Sciences, supported by the National
Science Foundation and published by SIAM.
GARRETI' BIRKHOFF, The Numerical Solution of Elliptic Equations
D. V. LINDLEY, Bayesian Statistics, ~ Review
R. S. VARGA, Functional Analysis and Approximation Theory in Numerical Analysis
R. R. BAHADUR, Some Limit Theorems in Statistics
PATRICK BILLINGSLEY, Weak Convergence of Measures: Applications in Probability
J. L. LIONS, Some Aspects of the Optimal Control of Distributed Parameter Systems
ROOER PENROSE, Techniques of Differential Topology in Relativity
HERMAN CHERNOFF, Sequential Analysis and Optimal Design
1. DURBIN, Distribution Theory for Tests Based on the Sample Distribution Function
SOL 1. RUBINOW, Mathematical Problems in the Biological Sciences
P. D. LAX, Hyperbolic Systems of Conservation Laws and the Mathematical Theory of Shock
Waves
1. J. SCHOENBERG, Cardinal Spline interpolation
IVAN SINGER, The Theory of Best Approximation mid Functional Analysis
WERNER C. RHElNBOLDT, Methods of Solving Systems of Nonlinear Equations
HANS F. WEINBERGER, Variational Methods for Eigenvalue Approximation
R. TYRRELL ROCKAFELLAR, Conjugate Duality and Optimization
SIR JAMES LIGHTHILL, Mathematical Biofluiddynamics
GERARD SALTON, Theory of indexing
CATHLEEN S. MORAWETZ, Notes on Time Decay and Scattering for Some Hyperbolic Problems
F. HOPPENSTEADT, Mathematical Theories of Populations: Demographics, Genetics and
Epidemics
RICHARD ASKEY, Orthogonal Polynomials and Special Functions
L. E. PAYNE, improperly Posed Problems in Partial Differential Equations
S. ROSEN, Lectures on the Measurement and Evaluation of the Peiformance of Computing
Systems
HERBERT B. KELLER, Numerical Solution afTwo Point Boundary Value Problems
J. P. LASALLE, The Stability of Dynamical Systems - Z. ARTSTEIN, Appendix A: Limiting
Equations and Stability of Nonautonomous Ordinary Differential Equations
D. GOTTLIEB AND S. A. ORSZAG, Numerical Analysis of Spectral Methods: Theory and
Applications
PETER 1. HUBER, Robust Statistical Procedures
HERBERT SOLOMON, Geometric Probability
FRED S. ROBERTS, Graph Theory and its Applications to Problems of Society
(continued on inside back cover)
Contents
PREFACE v
CHAPTER 1. Monte Carlo Methods and Quasi-Monte Carlo
Methods
L 1 Introduction. . . . . . . .
1.2 Monte Carlo methods. . .
1.3 Quasi-Monte Carlo methods.
Notes ............ .
CHAPTER 2. Quasi-Monte Carlo Methods for Numerical Integra-
1
1
3
9
11
t ~ n 13
2.1 Discrepancy.. 13
2.2 Error bounds. 18
Notes .... 21
CHAPTER 3. Low-Discrepancy Point Sets and Sequences 23
3.1 Classical constructions.. . . . 23
3.2 General discrepancy bounds. 33
Notes ............. 44
CHAPTER 4. Nets and (t,s)-Sequences 47
4.1 Definitions and discrepancy bounds. 47
4.2 Combinatorial connections. .. 60
4.3 General construction principles. . . . 63
4.4 A special construction of nets. . . . . 74
4.5 A special construction of (t,s)-sequences. 90
Notes .................... 98
CHAPTER 5. Lattice Rules for Numerical Integration 101
5.1 The method of good lattice points. . . . . . 102
5.2 Existence theorems for good lattice points. 109
5.3 General lattice rules and their classification. 125
lV CONTENTS
5.4 Existence theorems for efficient lattice rules. .
Notes ..................... .
138
145
CHAPTER 6. Quasi-Monte Carlo Methods for Optimization 147
6.1 General theory of quasirandom search methods. . 147
6.2 Low-dispersion point sets and sequences. . 152
Notes ........................ 158
CHAPTER 7. Random Numbers and Pseu.dorandom Numbers 161
7.1 Random number generation. . 161
7.2 Pseudorandom numbers. 164
7.3 Classical generators. 168
Notes .......... 175
CHAPTER 8. Nonlinear Congruential Pseudorandom Numbers 177
8.1 The general nonlinear congruential method. 177
8.2 The inversive congruential method. 182
Notes ..................... 189
CHAPTER 9. Shift-Register Pseudorandom Numbers 191
9.1 The digital multistep method. . . . . . . . . . . . . . . 191
9.2 The generalized feedback shift-register (GFSR) method. 198
Notes ........................ 204
CHAPTER 10. Pseudorandom Vector Generation 205
10.1 The matrix method. 205
10.2 Nonlinear methods. . 213
Notes ........ 215
APPENDIX A. Finite Fields and Linear Recurring Sequences 217
APPENDIX B. Continued Fractions 219
BIBLIOGRAPHY 223
Preface
The NSF -CBMS Regional Research Conference on Random Number Generation
and Quasi-Monte Carlo Methods was held at the University of Alaska at Fair-
banks from August 13-17, 1990. The present lecture notes are an expanded
written record of a series of ten talks presented by the author as the principal
speaker at that conference. It was the aim of this series of lectures to familiar-
ize a selected group of researchers with important recent developments in the
related areas of quasi-Monte Carlo methods and uniform pseudorandom number
generation. Accordingly, the exposition concentrates on recent work in these
areas and stresses the interplay between uniform pseudorandom numbers and
quasi-Monte Carlo methods. To make these lecture notes more accessible to
nonspecialists, some background material was added.
Quasi-Monte Carlo methods can be succinctly described as deterministic ver-
sions of Monte Carlo methods. Determinism enters in two ways, namely, by work-
ing with deterministic points rather than random samples and by the availabil-
ity of deterministic error bounds instead of the probabilistic Monte Carlo error
bounds. It could be argued that most practical implementations of Monte Carlo
methods are, in fact, quasi-Monte Carlo methods since the purportedly random
samples that are used in a Monte Carlo calculation are often generated in the
computer by a deterministic algorithm. This is one good reason for a serious
study of quasi-Monte Carlo methods, and another reason is provided by the fact
that a quasi-Monte Carlo method with judiciously chosen deterministic points
usually leads to a faster rate of convergence than a corresponding Monte Carlo
method. The connections between quasi-Monte Carlo methods and uniform
pseudorandom numbers arise in the theoretical analysis of various methods for
the generation of uniform pseudorandom numbers.
The last five years have seen tremendous progress in the subject areas of
these lecture notes. The field of quasi-Monte Carlo methods was enriched by
the systematic development of the theory of lattice rules and of the theory of
nets and (t, s )-sequences, and new and better low-discrepancy point sets and
sequences were constructed. Important advances in the area of uniform pseudo-
random number generation include the introduction and the analysis of nonlin-
ear congruential methods and the much deeper understanding we have gained
v
VI PREFACE
of shift-register pseudorandom numbers. Furthermore, a systematic study of
methods for pseudorandom vector generation was initiated.
The main aim of these lecture notes is to give an account of the recent develop-
ments mentioned above. The material in Chapter 4 on nets and (t,s)-sequences
and in Chapter 5 on lattice rules is of central importance, since many of the
new advances in quasi-Monte Carlo methods, and even in uniform pseudoran-
dom number generation, revolve around the concepts and results in these two
chapters. Indeed, nets and lattices appear in so many different contexts that
they must be viewed as basic structures. Another fundamental notion is that of
discrepancy, for which the essential facts are presented in Chapters 2 and 3. The
concept of dispersion plays a role in Chapter 6 on quasi-Monte Carlo methods
for optimization. In our discussion of uniform pseudorandom number generation
in Chapters 7, 8, and 9, we emphasize those algorithms for which a detailed
theoretical analysis has been performed.
For some results, especially for classical ones, the proof has been omitted, but
a reference is always provided. The notes to each chapter contain supplemen-
tary information and historical comments. The bibliography is not meant to be
comprehensive, but lists only those references that are cited in the text. Since
a detailed bibliography up to 1978 is available in [225], the present bibliography
concentrates on the more recent literature.
The conference would not have occurred without the initiative and the en-
thusiasm of Professor John P. Lambert of the University of Alaska at Fairbanks,
who originated the idea, did all the necessary paper work, and organized the
conference in a perfect manner. Many thanks, Pat, for the generous hospitality
you extended to all participants and for making sure that everybody survived the
white-water rafting on the Nenana River. I also want to express my gratitude to
NSF -CBMS for the financial support of the conference. The actual production of
these lecture notes relied heavily on the word-processing skills of Rainer Gottfert
and Irene Hosch. Special words of appreciation go to Rainer Gottfert for his dedi-
cated help in proofreading the manuscript and to Dipl.-Ing. Leonid Dimitrov and
Dipl.-Ing. Reinhard Thaller for expert technical advice. The cooperation of the
SIAM staff is also noted gratefully. Last, but by no means least, I wish to thank
Gerlinde for putting up with a husband who was somewhere between the real
world and mathematical elysium during the last 15 months.
Vienna, August 1991 H. NIEDERREITER
CHAPTER
Monte Carlo Methods and
Quasi-Monte Carlo ethods


In this chapter we set for the more detailed discu.ssi,ot of quasi-
Carlo in c4 the o(quaai-
Vente Car19 metpods,is, without an rudi,wentary undemtan<i-
. ' , ',' . - .' , . ." \ ,;!.j . . - i - . '
ing of Monte Carlo methods. For tJris reason, and also to motiwte the introduc-
tion of Carlo JD.ethodB, we inciude a brief exposition of ,the statistical
Mep.te Carlo method. ,A numerical problem that lends itself to a straightforward
,illustrative comparison of ,classical, Monte Carlo, and quasi-Monte Carlo
methods is that of numerical integration. We will use this, problem to describe
the basic ideM behind Monte Carlo and quasi-Monte Carlo methods in 1.2 and
1.3, respectively.
1.1. Introduction.
We consider the problem of numerical integration in dimension 8. For 8 = 1 there
are classical integration rules such as the trapezoidal rule and Simpson's rule, to
mention only two of the most well-kno'li'ffi ones. For concreteness, let us look at
the trapezoidal rule for the unit interval [0, 1]. In this case, the trapezoidal rule
yields the approximation
(1.1)
where m is a positive integer and the weights Wn are given by Wo = Wm = 1/(2m)
and Wn = 11m for 1 :$ n :$ m - 1. The error involved in approximation is
O(m-
2
), provided that f has a continuous second derivative,ob [0,1]_
In the multidimensional case s 2 with an interval as integration do-
main, the classical numerical integration methods use Cartesian products of
one-dimensional integration rilles. In such multidimensional integration rules,
the node set is a Cartesian product of one-dimensional node sets, and the weights
are appropriate products of weights from the one-dimensional rules. These multi-
dimensional integration rules are obtained by viewing the a-dimensional integral
as an iteration of one-dimensional integrals and by applying a one-dimensional
integration rule in each iteration. To illustrate this procedure, we state the s-fold
1
2 CHAPTER 1
Cartesian product of the trapezoidal rule (1.1), which is
(1.2) { I(u) du f f W
n1
wn./(n
1
, , n
8
),
JIs .. m m
. n1=0 n,,=O
where Is - [0, 1]8 is the closed 8-dimensional unit cube and the Wn are as
in (1.1). The total number of nodes in (1.2) is N = (m + 1)8. From the error
bound for (1.1), it follows tha.t the error in (1.2) is O(m-
2
), provided that
fj2 1 is continuous on 1
8
for 1 i $ 8. To see that the error in (1.2)
need not, in be smaller than the one-dimensional error, it suffices to
apply (1.2) with a function Ion 1
8
that depends only on one variable, in which
case (1.2) reduces to (1.1). In terms of the number N of nodes, the error in (1.2)
is O(N-
2
/S). With increasing dimension 8, the usefulness of the error bound
O(N-2/s) declines drastically. Specifically, to guarantee a prescribed level of
accuracy, say an error that is in absolute value $ 10-
2
, we must use roughly 10
8
nodes; hence the reqUired number of nodes increases exponentially with s. This
phenomenon is often called the "curse of dimensionality. II
The phenomenon described above manifests itself in an analogous way for
the Cartesian product of anyone-dimensional integration rule. For an s-fold
Cartesian product, the order of magrutude of the error hound, in terms of the
total number of nodes, is the 8th root of the order of magnitude of the one-
dimensionalmtegration rule.
A decisive step in overcoming the curse of dimensionality the develop-
ment of the Monte Carlo method in the 19408. The Monte Carlo method is
actually a very general tool, and its applications are by no means restricted to
(for this reason, the plural.uMonte Carlo methods" is also
used frequently). In simple and general terms, the Monte Carlo method may
be described as a numerical method based on rap.<iom sampling. It is there-
f.ore a method with a strong statistical and probabilistic flavor. We will discuss
the Monte Carlo method for numerical integration in 1.2and the Monte Carlo
method fDr global optimizatiDn in 6.1. An important {act about the Monte CarlD
method fDr numerical integration is that it promises integratiDn errors for which
the ord" of magnitude, in terms of the number of nodes, is independent of the
dimensi,p. . is obviously a big step forward compared to classical methDds
based o:p.:, ian products of one-dimensi.onal integration rules. However, the
stochastic' nature of the Monte Carlo method causes some unpleasant side ef-
fects. These drawbacks of the Monte Carlo method will be discussed more fully
in 1.2. Let us mention now that the Monte Carlo method yields no guaran-
teed error bound (this is why we used the phrase "it promises integration errors
... " earlier 011)' ..
A crucial.task in t1;le of any Monte Carlo method is the generat'ion
success of a Monte Garlo calculatiDn often
stands or falls the of the random samples that are used, where,
by "quality," we mean how well the random samples reflect true randomness.
This intuitive notion will attain a more precise meaning in 1.2 and Chapter 7,
where the requirements on random samples will be made concrete.
MONTE CARLO METHODS AND QUASI-MONTE CARLO METHODS 3
Because of the decisive role played by random samples, the subject of random
number generation has an important spin-off the study Monte Carlo
With the present trend toward parallelized algorithms there is a surge
of interest in random vector generation as well. In the computational practice
of Monte Carlo methods" the required random numbers and random vectors are
actually generated by the' computet in a subroutine. this esse of
deterministic generation, we speak of pseudorandom numbers and plJeooorandom
vectors- The development of the theory of pseudorandom numbers and pseudo-
random vectors has accompanied the development of the Monte Carlo method.
Specific algorithms for the generation of pseudorandom numbers wereakeady
in the late 1940s. We refer to Chapters 7-10 for a detailed treatment
of pseudorandom number generation and pseudorandom vector. genera.tion.
,:" AB mentioned above, the Monte Carlo method has ooD;).e disadvantages, which
in the stochastic character of method . 'v'o.r, let us
again return tp our standard example of numerical In this case,
the Monte Carlo methOd yields only a probabilistic botmd'on the integration
error. On the other hand, a detailed analysis (see Chapter'2) reveals that, in
Monte Carlo integra,tion, it is not 00' much the true randomness of the sam-
ples that is relevant, but ra.ther that the' samples should be spread in a uniform
riianner over the integration domain (in a sense that can be made precise). More-
the analysis shows that a deterministic error bound can be established if
deterministic nodes are used. This leads to the idea of selecting deterministic
nodes in BUch a way that the error bound is as small as possible. This idea suc-
cinctly expresses the fundamental principle of a quasi-Monte Carlo method. A
quasi-Monte Carlo method can be described in simple terms as the deterministic
version of a Monte Carlo method, in the sense that the random samples in the
Monte Carlo method are replaced by well-chosen deterministic points. The main
aim is to select deterministic points for which the deterministic eITor bound is
smaller than the probabilistic Monte Carlo error bound. In the case of numerical
integration
1
this a.im can be achieved in a convincing manner. There are, in fact,
various types of quasi-Monte Carlo methods, and so this term is also often used
in the plural.
It turns out that the technical tools developed in the theory of quasi-
Monte Carlo methods are very useful in the theoretical analysis of several impor-
tant methods for pseudorandom number generation and pseudorandom vector
generation. Consequently, the material in these lecture notes is arranged in such
a way that the chapters on pseudorandom numbers and pseudorandom vectors
are preceded by the discussion of quasi-Monte Carlo methods.
1.2. Monte Carlo methods.
a Monte Carlo method, the quantity to be calculated is interpreted a
stochastic model and subsequently estimated by random sampling. In this sec-
tion, we illustrate this method in the context of numerical integration. A further
application to global optimization will be described in 6.1.
Consider the problem of approximately calculating the integral IB feu) du
4 CHAPTER 1
with an integration domain B ~ R.s satisfying 0 < As(B) < 00, where As hence-
forth denotes the s-dimensional Lebesgue measure. VIe tum B into a probability
space with probability measure dp, = du/>.-s(B). Then, for f E Ll(p.), we have
(1.3)
L feu) du = All (B) L f dp. = Aa(B)E(J),
where E(J) is the expected value of the random variable f. The problem of
numerical integration is thus reduced to the problem of the approximate calcu-
lation of an expected value. For the latter purpose, we use a basic idea from
statistics, namely, to estimate an expected value by a sample means. This idea
can be developed in the general framework of an arbitrary probability space.
Let f be a random variable on an arbitrary probability space (A,A, ),). Then
the Monte Carlo estimate for the expected value E(J) is obtained by taking N
'"
independent A-distributed random samples aI, ... ,aN E A and letting
(1.4)
1 N
E(f) ~ N L f(an).
n=l
The strong law of large numbers guarantees that this procedure converges almost
surely in the sense that
1 N
N ~ O O N L J(an) = E(f) ),oo-a.e.,
n=l
where A 00 is the product measure of denumerably many copies of"\. For the
probabilistic etfur'analysis of the estimate (1.4), we need the variance
. ~ . ,'") ; j ',I ;.- ,
, ~ i' :
which is finite whenever f E L2(..\).
THEOREM 1.1. If f E L2(>..), then, for any N ;?: 1, we have
Proof Write 9 = f - E(f)i then fA gd..\ = 0 and
1 N . 1 N
'N Lf(an ) -E(f) = N Lg(an).
n=l n=l
MONTE CARLO METHODS AND QUASI-MONTE CARLO METHODS 5
Thus
= L'"
, . 1
Theorem 1.1 may be interpreted to mean that the aboolutev,alue of the error
in (1.4) is, on the a v e r ~ j u(f)N-1/2, where a(f) = (u
2
(fV
2
the standard
deviation of I, Further probabilistic information about the error is obtained
from the central limit theorem, which states that, if 0 < u(f) < 00, then
for any constants Cl < C2 j where is the A 00 - measure of the set of all
sequences aI, a'}., . of elements the property indicated between
the parentheses.
IT we apply the statistical estimate (1.4) to original problem of approxi-
mately calculating the integral fB I(u)du, then, in view of (1.3), we obtain the
Monte Carlo estimate
(1.5)
where Jell' " ,XN are N independent ,urdistributed random samples from B. By
the above analysis, the absolute value of the error (1.5) is then, on the average,
As(B)Cf(f)N-1/2. On the basis of this fact, or on basis of the central limit
theorem, we can state that the Monte Carlo method for numerical integration
yields a probabilistic error bound of the form O(N-
I
/'}.) in terms of the number
N of nodes, The remarkable feature here is that order of magnitude does
not depend on the dimension s. This should be compared with the error bound
O(N-2/s) for the classical s-dimensional integration rule discussed in 1.1. The
Monte Carlo method for numerical integration is definitely preferable to the
classical integration rule for dimensions s ~ 5.
There is a case in the statistical estimate (1.5) is not useful com-
putational practice, namely, when the integration is so complicated
6 CHAPTER 1
that ).s(B) is difficult to calculate. In this case, we can take recourse to the
following alternative Monte Carlo estimate. By applying, if necessary, a suitable
change of variables, it suffices to consider the situation where B is contained in
the unit cube i8. Then we can write
( f(u)du = J{U)CB(U) du,
1B jI8
where C B is the characteristic function of B. If the last integral is estimated
according to (1.5), then we arrive at the Monte Carlo estimate
(1.6)
r 1 N
jf f(u)du N L !(:xn),
B n=l
x"eB
where Xl, .. , ,XN are N independent random samples from the uniform distribu-
tion on i". Since the estimate (1.0) lli aenvea trom \.La}, .n:; lOllOW8wal; lor (1.0)
we also have a probabilistic error bound of the form O(N-
1
/2). '
The preceding description of the Monte Carlo method for numerical integra.-
tion gives some idea as to what the basic elements of Monte Carlo methods are,
in general. The first step invariably consists of the statistical modeling of the
given numerical problem, usually in terms of suitable random variables. This
step is obvious in the case of numerical integration, since an integral can readily
be interpreted in, terms of the expected value of a random variable (compare
with (1.3. Subsequently, the random variables occurring in the model must be
analyzed with regard to their statistical pmperties, such as law of distribution,
statistical dependence or independence, and so on. Another important step is
the generation- of random samples that retiect these statistical properties. It is
advisable not to be satisfied with just one run of a Monte Carlo calculation,
but to use as many :runs as is time with renewed sampling-to
increase the, reliability, of the result from, the statistical point of view . These pro-
cedural steps in a Monte Carlo method are typical of the larger area of stochastic
simulation to which Monte Carlo methods belong. Stochastic simulation deals
with the simulation of stochastic phenomena by special, but randomly selected,
instances. The methods used in stochastic siniula.tion are often called simulation
methods. Monte Carlo methods are thus examples of simulation methods.
Monte Carlo methods are used in a wide range of applications. In addition
to the applications to munerical integration and global optimization discussed
in these lecture notes, there are applications of Monte Carlo methods to many
other basic problems of numerical analysis, such as the solution of large systems
of linear equations, the solution of partial. differential equations and of integral
equations., Natural applications arise in problems with a background,
for instance, in computational statistics, stochastic optimization, and queueing
theory. There is,,,also a wealth of applications to real-world problems in areas
such as computational. physics, structural mechanics, reliability theory, l'iystems
analysis, and so on.
AB we have explained above, the Monte Carlo method for numerical integra.-
tion offers a way of overcoming the curse of dimensionality. However, it must be
MONTE CARLO METHODS AND QUASI.MONTE CARLO METHODS 1
pointed out that the Monte Carlo method is not a panacea. On the oontrary, it
has 2everal deficiencies that may hf;!,Illper its usefulness. First, the Monte Carlo
method for numerical integration provides only a probabilistic error bound. In
other words, there is never any guarantee that'the expected accuracy is aclUeved
in a concrete calculation. This is certainly not a desirable State of affairs, and,
for some applications---Q.g., ,in .sensitiwareas where highly" reliable :results are
needed-it is actually an untenable situation.
We also draw attention to the following fact: the probabilistic error bound
O(N-
1
/
2
) in the Monte Carlo method for numerical integration holds under
a very weak regularity condition, namely, for any square-integrable integrand,
but no benefit is derived from any additional regularity of the integrand. The
numerical practice teaches . a different lesson: more regular functio!lll are "easier"
to integrate numerically, . in the BenBe that they lead to a faster convergence
rate when the of nodes in integration
rules. The fact that the Monte Carlo error bound does not reflect any additional
regularity of the integrand must therefore be seen as another dra.wback of the
A fundamental difficulty of the Monte Carlo method for numerical integration
stems from the requirement that the nodes be independent random samples. This
mimediately raises the question of how to generate independent'r\:hldom sampieS
concretely. The practitioners avoid this question and use pseudorandom numbers
instead of truly random samples. Since "tandom sample"is a statistical. concept
that only makes sense in the context of a whole ensemble of samples, there is
actually no way to define this concept rigorously for an individual sample.
To summarize, the following points must be viewed' as deficiencies of the
Monte Carlo method for numerical integration:
(i) There are only probabilistic error bounds;
(li) The regularity of the integrand is not reflected;
(iii) Generating random samples is difficult.
AP, we have seen in Theorem 1.1, the mean-square error of the Monte Carlo
estimate is equal. to 0'2 (f) / N. To improve the efficiency of the Monte Carlo
method for numerical integration, a number of techniques for variance reduction,
i.e., for decreasing the variance factor 0'2(1), have been developed. One such
technique is called stratified sampling and proceeds as follows. Let f again be
a random variable on the probability space (A, A, .:\). Partition A into the sets
AI, . .. ,Ak E A with ':\(Aj) > 0 for 1 ::; j ::; k. For each j with 1 ::; j ::; k,
choose N
j
independent j.tj-distributed random samples ap), ... ,aW E A
j
, where
J
j.tj = A(Aj)-l A is the probability measure on Aj induced by A. Then use the
estimate
The_mean-square error of this estimate can be calculated by the method in the
8 CHAPTER 1
proof of Theorem 1.1, and this yields, instead of (1'2(f)jN, the expression
The numbers N
j
must be chosen suita.bly so as to achieve a variance reduction.
Let N = E;=1 Nj be the total number of sample points. Then the following
standard choice yields a. variance reduction. -
PROPOSITION 1.2. If the numbers Nj = A(Aj)N, 1 $ j $ kJ are integers,
then
t A(Aj) f (I __ 1_ f 1 dA)2 d>' $ (12(1).
j=1 Nj JAj A(Aj) JAj N
p.roof, 'Nit h the sTIBc1al fo:rm. the Ni it suffices to show that
t { (I ____ I 1, 1 dA)2 d)' $ f (f _E(J2 dA.
j=l JAj >'(Aj) Aj iA
By expanding the square on both sides, it is seen that this is equivalent to proving
2 k 1 (1 \2
E(f) $ L >'(A .) I I d>.) ,
j=1 J \Aj
and the latter is obtained from the Cauchy-Schwarz inequality by
noting that
Another for variance reduction is the method of antithetic variates.
. -
We explain this method in the simple case of an integral-
E(f) = 11 I(u)du.
We introduce the auxiliary function
(1.7)
y(u) = ! (f(u) + - u for 0 $ u $ 1
and use the estimate
: 1 N 1 N
E(f) = E(g) N L 9(Xn} = 2N L (f(Xn) + 1(1 - xn
\ n=l n=l
MONTE CARLO METHODS AND QUASI. MONTE CARLO METHODS 9
with N independent and uniformly distributed random samples Xi, E
[0,1]. According to Theorem 1.1, the error of this estimate is
rr (g) IN. Since there are 2N function values of / involved in this estimate, the
mean-eq1W'e er1'Of u2(g)/N should beoompared. with the quantity (72(J)/(2N).
The fOllowing result shows a cue in which a variance reduction is achieved.
PROPOSITION 1.3. II/is a continuous monotone function 00 [0, 1] and g
is tkfinetlby (L7) I
Proof. We have
o"(g) = J.' (g(,,) - E(g))' 11" = J.' G/(tt) + u) - E(f), du
= 11 /2(u) du + 11 I(u)!(l - u) du - E(f)2.
The de&red inequality iEl therefore eqluvalent to
.11 I(u)f(l- u)du '5 E(f)2.
By replacing, if necessary, f by -I, we can assume that f is nondecreasing.
Then the function
F(u) = l
u
J(1 - t) dt - E(f)u
on [0,1] has the nonincroosing derivative F'(u) = J(1-u) -E(J). Since F(O) =
F(l) = 0, it follows that F(u) 2:: 0 for all u E [0,1]. This implies that
11 F{u) dJ(u) 2:: o.
Integration by parts yields
11 f(u) dF(u) = 11 J(u)F'(u) $ O.
By inserting the formula for F'(u), we arrive at (1.8). 0
1.S. Quasi-Monte Carlo methods.
We :recall that in Monte Carlo integration with N random nodes the absolute
value of the error has the average order of magnitude N-l/2. Clearly, there exist
sets of N nodes for which the absolute value of the error is not larger than the
average. IT we could construct such sets of nodes explicitly, this would already be
a useful methodological advance. The quasi-Monte Carlo method for numerical
integration aims much higher! as it seeks to construct sets of nodes that perform
significantly better than average. The integration rules for the quasi-Monte Carlo
method are taken from the appropriate Monte Carlo estimate. For instance, for
10 CHAPTER 1
the normalized integration domain ill, we have the quasi-Monte Carlo approxi-
mation
(1.9)
f 1 N
}f8 feu) du N L f(Xn),
1 n=l
which formally looks like the Monte Carlo estimate but is now used with de-
terministic nodes Xll" . ,XN E is. These nodes should be chosen judiciously
so as to guarantee a small error in (1.9). In analogy with the more general
Monte Carlo estimate (1.6), we have the quasi-Monte Carlo approximation
(1.10)
f 1 N
}1 feu) au N L f(x.n),
B n=l
xnEB
where B is a subset of /s and Xl, , XN E Is are deterministic points.
The error analysis for the approximations (1.9) and (1.10) will be performed
in 2.2 and for a special situation in Chapter 5; Explicit constructions of sets of
deterministic nodes guaranteeing small errors will be the subject of ChapterS 3,
4, and 5. We do not wish to preview these results at length, but the fonowing
very concise summary may be given at this early stage. If we assume (1.9) to
be the standard case, then the benchmark is the probabilistic Monte Carlo error
bound O(N-l/2). The quasi-Monte Carlo method yields a much better result,
giving us the deterministic error bound D(N-'-l(logN)S-l) for suitably chosen
sets of nodes and for integrands with a relatively low degree of regularity. In the
special circumstan<;es considered in Chapter 5, even smaller error bounds can .
be achieved for sufficiently regular integrands. The sets of nodes producing this
high accuracy in (1.9) are obtained by effective constructions.
This brings us to a discussion of the advantages of the quasi-Monte Carlo
methl>d for numerical integration, which should be compared with the list of
deficiencies of the Monte Carlo method in 1.2. The very nature of the quasi-
Monte Carlo method, with its completely deterministic procedures, implies that
we get deterministic-and tl1us guanmteed-:-er.ror bounds. In principle, it is
therefore always possible to determine in advance an integration r-ule that yields
a prescribed level of accuracy. Moreover, with thesame computational effort, Le.,
with the same number of function evaluations (which a.rethe costly operations
in numerical integration), the quasi-Monte Carlo method achieves a significantly
higher accuracy than the Monte Carlo method. Thus, on two crucial accounts--
determinism and precision-the Carlo method is superior to the
Monte Carlo method. For a special type of quasi-Monte Carlo method, the lat-
tice rules to be discussed iuChapter 5, we have the desirable pr9perty that a
of regularity of the leads to precision in the inte-
gration rule., The one problem with the Monte Carlo method that attains almost
philosophiFal dim,ensions, namely, the difficulty of generating truly random sam-
ples, when we consider a quagi-Monte Carlo method, since here we
just implement ready-roMe constructions to obtain the required nodes.
,
MONTE CARLO METHODS AND QUASI-MONTE CARLO METHODS 11
There are quasi-Monte Carlo methods not only for numerical integra.tion,
but also for various other numerical problems. In fact, for many Monte Carlo
methods, it is possible to develop corresponding quasi-Monte Carlo methods as
their. deterministic versions. Invariably, the basic idea is to replace the random
samples in the Monte Carlo method by deterministic points that are well sUited
for the problem at hand. A further illustra.tion of this principle will be provided
Chapter 6, where we discuss quasi-Monte Carlo methods for global optimization.
Quasi-Monte Carlo methods have first been proposed in the 19508, and their
theory has since developed vigorously. However, for a ra.ther long time, these
methods remained the province of specialists. The wider acceptance of quasi-
Monte Carlo methods W8B pioneered in computational physics where large-scale
Monte Carlo calculations are very common' and where it was recognized that
quasi-Monte Carlo methods can lead to significant gains in' efficiency. The re-
finements of quasi-Monte Carlo methods and the expanding scope of their ap-
plications have recently made these methods known to 'larger segments of the
scientific computing community.
oh";ioUJl methods Dfoblere.B
of numerical analysis that can be reduced to numerical integration. An exam-
ple is the numerical solution of integral equations for which quaai-Monte Carlo
methods are discussed in the books of Korobov [160, Chap. 4] and Hua and
Wang [145, Chap. 10] and in the more recent work of Sarkar and Prasad [301]
and Tichy [349]. For applications to initial value and boundary value problems,
we refer to HUB. and \Vang [145, Chap. 10] and Taschner [34OJ. Important applica-
tions of quasi-Monte Carlo methods to the numerical solution of difficult integro-
differential equations such as the Boltzmann equation have been developed by
Babovsky et al. [12}, Lecot [181], [182], [184J, and Motta et al. [214]. Applica.-
tions to approximation theory can already be found in the books of Korobov [160,
Chap. 4] and Hua and Wang [145, Chap. 9]. Quasi-Monte Carlo methods for
the numerical solution of systems of equations are studied in Hlawk..a [140] and
Taschner [341 J. A survey of quasi-Monte Carlo methods in computational statis-
tics is presented in Shaw [309]. Deal< [52] reviews applications to stochastic
optimization, and Adlakha [1J describes an application to stochastic algorithms.
Drmota [69] and Drmota and Tichy [70] have developed applications of quasi-
Monte Carlo methods to statistical measurement techniques. Quasi-Monte Carlo
methods for the approximate calculation of vector-valued integrals are sketched
in Trendafilov [354]. Some further applications of quasi-Monte Carlo methods
and corresponding references are listed in Niederreiter [225, 2].
Notes.
The "official" history of the Monte Carlo method began in 1949 with the publi-
cation of a paper by Metropolis and Ula.m [211 J, but at that time the method had
already been used for several years in secret projects ofthe U. S. Defense Depart-
ment. This rather fascinating chapter in the early history of the Monte Carlo
method is outlined in Eckhardt [74J and Metropolis [210]. Several precursors are
mentIoned in Hammersley and Handscomb [130, Chap. 1]. The latter book is
12 CHAPTER 1
the da.ssic on Monte Carlo methods. More recent books on Monte Carlo meth-
ods and on the general. area of simulation are Bratley, Fox, and Schrage [34],
Deale [53], Kalos and Whitlock [150], Morgan [213], Ripley [295], and Rubin-
stein [291]. For a general background on numerical integration, we refer to Davis
and Rabinowitz [51).
Expository accounts of quasi-Monte Carlo methods are given in the books of
Korobov [160], Hlawka, Fimeis, and Zinterhof [141], and Hua and Wang [145].
A detailed survey of quasi-Monte Carlo methods with an extensive bibliography
is presented in the article of Niederreiter [225], and a recent update of this
survey can be found in Niederreiter [251]. A brief review of quasi-Monte Carlo
methods is contained in Lambert [168]. The first article in which the expression
"quasi-Monte Carlo method" appeared seems to have been a technical report of
Richtmyer [293] from 1951.
CHAPTER 2
Q,uasi-Mo Carlo M od
Numerical Integration
According to 1.3, the basic idea. of a quasi-Monte .;method is to repla.ce
samples in a Monte C&."10 :method by well-chosen de,terministic points.
The criterion for the choice of . points
ul:'oblem at 'Pm imuortant urohlem nnmflcicPl.] hY1;llZrstion, the se-
lection criterIon is easy to and leads to the concepts of uni/ormi.y distributed
sequence and discrepancy. The discrepancy can be viewed as a quantitative
measure for the deviation from uniform distribution. Various types of discrep.-
ancies and their basic properties will be discussed in 2.1. The important role of
the discrepancy in quasi-Monte Carlo integration is documented in 2.2, where
deterministic bounds for the integration error in terms of the discrepancy are
presented.
2.1. Discrepancy.
For the following discussion, we normalize the integration domain to be Is :=
[0, 1] 8, the dosed s-mmensional unit cube. For an integrand f) we use the
quasi-Monte Carlo approximation
(2.1)
[ 1 N
JIB f(U) du N L f(Xn)
1 11,=1
with Xl j ,XN E 1
8
In an idealized model, which we adopt for the moment, we
replace the set of nodes Xl, ,XN by an infinite sequence Xl, X2,' .. of points ill
[a. A basic requirement for this sequence is that we obtain a convergent method
from (2.1). Thus we want a sequence XI, for which
(2.2)
1 N J,
lim - L f(xn } = f(u) du
N--.oo N 1-8
n=l
holds for a reasonable class of integrands, say, for all continuous f on Is. The
resulting condition means that the sequence Xl, X2, .. should be uniformly dis-
tributed in [s, according to one of the standard definitions of this notion (com-
pare::.with Kuipers and Niederreiter [163]). An equivalent definition states that
13
14 CHAPTER 2
(2.3)
for all subintervals J of i
8
! where CJ is the characteristic function of J and >"8
denotes the s-dimensiona.l Lebesgue measure. It is well known that, ifxl,X2, .. '
is uniformly distributed in is, then (2.2) holds even for all Riemann-integrable
functions f on js (see [163]).
The above discussion suggests that the desirable nodes in (2.1) are those for
which the empirical distribution is dose to the uniform distribution on is. In an
intuitive language, the nodes Xl, ... ,XN should be "evenly distributed" over is.
The various notions of discrepancy that we will consider are quantitative mea-
sures for the deviation from uniform distribution, or, in other words, for the irreg-
',' "--' - ~
integration will become eyen more pronounced in 2.2, where bounds for the in-
tegration error in terms of the discrepancy are shown.
Let P be a point set consisting of Xl,'" ,XN E js. \Ve will always interpret
"point set" in the sense of the combinatorial notion of "multiset," Le., a set in
which the multiplicity of elements matters. In the theory of uniform distribution
of sequences, the term Ufinite sequence" is often used instead of "point set,"
but we will restrict the usage of "sequence" to its proper meaning in analysis,
namely, that of an infinite sequence. For an arbitrary subset B of is, we define
N
A(B; P) = L CB(Xn),
n=1
where CB is the characteristic function of B. Thus A(B; P) is the counting
function that indicates the number of n with 1 $; n $; N for which Xn E B.
H B is a nonempty family of Lebesgue-measurable subsets of is, then a general
notion of discrepancy of the point set P is given by
(2.4)
(
I
A(B;P) ()I
DN B; P) = ~ ~ ~ N - >"8 B .
Note that 0 $; DN(B; P) $; 1 always. By suitable specializations of the family B,
we obtain the two most important concepts of discrepancy. \file put r' = [0, l)s.
DEFINITION 2.1. The star discrepancy DN(P) = D'N(-xlJ'" j XN) of the
point set P is defined by Dj.l(P) = D1V(3*; P)j where 3>' is the family of all
subintervals of IS of the form n:=l [0, Ui).
DEFINITION 2.2. The (extreme) discrep13ncy D1V(P) = DN(Xl, ... j XN) of
the point set P is defined by DN(P) = DN(:!; P)j where :J is the family of all
subintervals of J8 of the form n:=l [Uij V'i)'
- -
REMARK 2.3. If the points of P are in j which happens in most cases of
practical interest, then DN(P) = DN(.:J
c
*; P) and DN(P) = DN(.:Jc; P), where
\
QUASI-MONTE CARLO METHODS FOR NUMERICAL INTEGRATION 15
.1: is the family of all subintervals of I" of the form n:=l [ 0, Ui 1 and .1c is the
family of all closed subintervals of IS.
PaqPOSITION 2.4. For any P oomi.sting of points in Is} we have
Proof. The first inequality is trivial.. For s = 1, the second inequality is
immediately obtained. from
A(( u,v)j P} = A([O, v}; P) - A([ 0, u); P), Al([ ti, v)) = Al([O, v)) - Al([O, u)),
and for 8 ;::: 2 it is obtained from analogous identities (compare with [163, p. 93]
for the case where 8 = 2). 0
In the case where 8 = 1, simple explicit formulas for the discrepancies D'N(P)
and DN(P) can be given. The following lemma is useful for the proofs.
LEMMA 2.5. If XI, ,XN,Yl, ,YN E [0,1] satisfy IXn - Ynl ~ c for
Proof. Denote by P the point set consisting of Xl,'" ,XN, and by Q the
point set consisting of Yl, . .. , Y N . Consider any interval J = [0, u) ~ [0, 1).
Whenever Yn E J, then Xn E J1 := [O,u_r-e) n [0,1]; hence
A(J;Q)
N
Whenever Xn E J'}, := [O,u - e), then Yn E J; hence
Thus Div(Q) ~ Div(P) + e. By interchanging the roles of P and Q, we obtain
D'N(P) ~ D'N(Q) + e, and so IDN(P) - D'NCQ)I ~ c. The second part of the
lemma is shown similarly. 0
For s = 1, we may arrange the points Xl, .. ,XN of a given point set in
nondecreasing order. The formula in Theorem 2.6 is due to Niederreiter [217J,
and the one in Theorem 2.7 is a simplified version of a formula of de Clerck [55].
THEOREM 2.6. If 0 :::; Xl ~ X2 ~ .. ~ XN ~ 1, then
Proof. Since D'N is a continuous function of Xl,'" ,XN (see Lemma 2.5), we
can ~ s u m e that < Xl < X2 < ... < XN < 1. Put Xo = 0 and XN+l = 1. P
16 CHAPTER 2
is the point set consisting of Xl, . .. ,X N, then
Div(P) = max sup I A([ 0, U)i P) - ul
O'5.n'5.N Xv> <U'5.Xn +l N
= max: sup - ul
O'5.n'5.N z" <u'5.x,,+l N
= - xnj, I; -Xn+ll)
= - Xnl, I n;; 1 - Xnl)
1 I 2n -11
= 2N + Xn - 2N .
o
THEOREM 2.7. If 0 Xl X2 .. :::; XN 1, then
DN(Xll .. , ,::EN) = N
l
+ max ('N
n
- xn)' - min (' N
n
- xn )'
l'5.n'5.N l'5.n'5.N
Proof. As in the proof of Theorem 2.6, we can assume that Xo := 0 < Xl <
X2 < .. , < XN < 1 =: XN+1' If P is the point set consisting of Xl!'" ,XN, then
D (P)
I
A([u,v);P) ( >1
N = max sup -v-u
O'5.i'5.i'5.N Z.<U'5.Zi+l N
Zj<tI'5.Zj+l ' -
u<tI
Put'r
n
= n/N - Xn for 0:::; n:::; N + 1; then
DN(P) = max max(lr-+l - 1'- - !..llr, - r'+1 + !..I)
O<i< -<N ' J S N' J S N
__ 3_
= max I!.. + ri - 1'-1
O<i<N N 3
1 '5.3 '5.N +1
If the last maximum is restricted to 1 i,j N, then its value is clearly given
by the expression for DN(P) in the theorem. Using
. 1
nun r <1'1<-
1'5.n'5.N n - - N'
it is Bee!l that in the last expression for DN(P) the terms in the maximum
corresponding to either i = 0 or j = N + 1 are dominated by the expression for
DN(P) in the theorem. 0
\
QUASIMONTE CARLO METHODS FOR NUMERICAL INTEGRATION 17
For a. sequence S of elements of is, we write DN(S) for the discrepancy and
DN (S) for the star discrepancy of the first N terms of S. According to a classical
result in the theory of uniform distribution of sequences (see [183]); the following
properties are equivalent:
., .. Jo,., ,', ""'
(i) Sis uniformly distributed in is;
(li) DN(S) = OJ
(ill) limN-+co D'N(S) = o ..
In this sense, the discrepancy and the star discrepancy can be viewed as quantifi-
cations of the definition of a uniformly distributed sequence in is given in (2.3).
For integration domains that are more general than is, we need various other
types of discrepancies. The following notion is the appropriate one for general
i
L
;, qpnvex with Theorem 2.14).,
... )fIPE.rINITION 2.8. The ISOtroPIC alScrepa!lfY IN(P). IN{Xl 1 ''' ,XN) of the
definedby IN(P) = P), where C isthe family of all convex
We always 5 IN(P) :5 4sDN(P)1/iJ, where the inequality is
tHVial, and the second one follows from a result of Niederreiter and Wills [278].
By combining thm with a criterion mentioned above, we see that a sequence S
is uniformly distributed in is, if and only if limN-+co IN(S) = 0, where IN(S) is
the isotropic discrepancy of the first N te:rms of S.
There is a classification of all Jordan-measurable subsets of i
8
(Le., of all
uOOets of i
8
for which the characteristic function is Riemann integrable) in
te:rms of the complexity of their boundary. For B is and e > 0, define
Be = {x E Is : d(x,y) < e for some y E B},
= {x E Is : d(x,y) e for all y E 1
8
" B},
where d is the standard Euclidean metric in JR8. Let b = b( E:) be a positive
nondecreasing function defined for all e > and satisfying lime-+D+ bee) = o.
Then we let Mb/be the family of all Lebesgue-measurable B 1
8
for which
A8(Be "B) :5 bee) and As(B" B-e) :5 bee) for all e > O.
Every B E Mb is actually Jordan measurable, and, conversely, every Jordan-
measurable subset of i
8
belongs to Mb for a suitable function b (see Nieder-
reiter [219, pp. 168-169]).
For a family Mb, we now consider the discrepancy DN(Mbi P) defined ac-
cording to (2.4). Niederreiter and Wills [278] have given a bound for this dis-
crepancy in terms of DN(P). If the function b satisfies beE:) 2: E for all c > 0,
then the bound can be simplified to yield
DN(Mbi P) :5 4b{2.jSDN(P)1/B).
In many cases of interest, the function b Mll have the form b(e) = C for some
constant C > 0, and then the bound in [278] reduces to
18 CHAPTER 2
Since every convex subset of i
8
belongs to Mba with b
o
being the function
bo(e) = 2se, we have
IN(P) $; DN(Mb;P)
whenever the function b satisfies b( e) ~ 2Se for all e > 0. IT the sequence S is uni-
formly distributed in ja, then, for any function b, we have limN __ CD DN(Mbi S) =
0, where DN(Mbi S) is the appropriate discrepancy of the first N terms of S.
Under suitable conditions on the function h, e.g., if b(e) ~ 2se for all e > 0, we
a.lso have the converse result; namely, limN-+CD DN(Mb; S) = 0 implies that S is
uniformly distributed in i
a

2.2. Error bowids.
We discuss the most important error bounds for the quasi-Monte Carlo approx-
imation (2.1) and analogous bounds for more general integration domains. All
these bounds involve a suitable notion of discrepancy. We start with the one-
dimensional case in which the proofs are quite easy. A classical result is the
following inequality of Koksma [155].
THEOREM 2.9. If f has bounded variation V (f) on [0, 1], then, for any
Xl! .. ,XN E [0,1 j, we have
Proof. We can assume that Xl ::5 X2 ::5 . . . ::5 x N. Put Xo = 0 and x N +1 = 1.
Using summation by parts and integration by parts, we obtain
1 N 11 . N . 11
N L f(Xn) - f(u) du = - L ; (f(Xn+d - f(xn}) + udf(u)
n=1 ' () n=O ()
N /Zn+l ( n)
= ~ l z : " u- N df(u).
For fixed n with ::5 11, ::5 N, we have
lu- ;1 ::5l)N(Xl!'" ,XN) for xn::5 u::5 Xn+l
by Theorem 2.6, and the desired inequality follows immediately. 0
In Theorem 2.12, below, we will prove a result that implies that Koksma's
inequality is, in general, the best possible, even for Coo functions. We recall
that, for a continuous function f on [0,1], its mOOulm of contin/uity is defined
by
w(fj t) = sup If(u) - f(v)1 for t ~ O.
u,vE[ 0,1]
lu-vl:9
The following error bound for continuous integrands was established by Nieder-
reiter [218].
QUASI-MONTE CARLO METHODS FOR NUMERICAL INTEGRATION 19
2.10. II I is continuow on [0,1], then, for any Xli'" jXN E
1], we have
Proof. We can again assume that Xl X2 ::; X N. By the mean-value
theorem for integrals, we have
r1 N (n.IN 1 N
Jo I(u) du = L J(, f(u) du = N L J(tn }
o n=l (n-l)/N n=l
with (n -l)IN < tn < niN. Therefore
Now Ix", - tnl :::; D:N(Xl,'" ) XN) for 1 n ::; N by Theorem
follows. 0
the result
To extend Koksma's inequa.1..ity to the multidimeooional. COOiEi, an appropriate
concept of total variation for functions of several'!2\Tiables ill needed. For a
function J on i
8
and a subinterval J of 11\ let l},,(f; J) be an alternating sum of
the values of I at the vertices of J (Le., function -v'alues at adjacent vertices have
opposite signs). The variation off on is in the se'f!.JJe is defined by
V(S)(f) = sup L J)I,
'P JE'P
where the supremum is extended over all partitions P of is into subintervals.
The more convenient formula
(2.5)
1
1 11 oaJ
V(8) (f) = ... I ----ldu1du
s
o 0 871,1'" 8us
holds whenever the indicated partial derivative is continuous on is. For 1 ::; k ::; s
and 1 ::; il < i2 < ... < ik 3, let V(k)(f; i
l
), ,. ,ik) be the variation in the
sense of Vitali of the restriction of f to the k-dimensional face {( 71,1, . ,Us) E
is: Uj = 1 for j i= ill'" ,ik}. Then
8
V(f) = L
V
(k)Cf . . )
; 21" .. ,'I,k
is called the va:riation of f on is in the sense oj Hardy and Krouse, and f is of
bounded variation in this sense if V (f) is finite. With this notion of variation, we
have the following inequality of Hlawka [134]' which is often called the Koksma-
fIlawlfa inequality. We again put IS = [0,1)8,
20 CHAPTER 2
THEOREM 2.11. ]f f has bounded variation V(f) on [8 in the sense of Hardy
and Krouse, then, lor any X:t, .. ,XN E 1
8
, we have
I t. J(x,.) -hJ(n) dul ,XN).
THEOREM 2.12. For any Xl, ... j'X.N E]II and any E: > 0, there exists a
function f E Coo (ill) with V (f) = 1 and
E f(Xn) - hs feu) dul> DN(Xl 1 '" ,XN) - e.
I '11=1 1
Proof. By the definition of DN(P) = DN(xt, ... , XN), there exists an inter-
val J = n:=l[O, Vi} with 0 < Vi :$ 1 for 1:$ i:5 8 and
! A(J: P) " I e
r--'- N_ A<J j I ./' "-/ N \.<) 2'
Furthermore, there exists an interval K = n:=l [ 0, t,,] with 0 :5 ti < Vi and
Vi - ti < e/(2s) for 1 :5 i :5 s such that J ...... K does not contain any point
len. For any given 0 :$ t <v :$ 1, we can construct a nonincreasing function
Jt,v E COO{[O, I]} with ft.v(u) = 1 for 0 :$ u :$ t and ft,v(u) = 0 for v :$ u :$ 1.
Then
II
feu) = f(uI, ... ,us) = IIft.,tI.(Ui)
i=l
is a function in COO(III) with 0 :$ feu) :$ 1 for all u E 1
s
, feu) = 1 for U E K,
and feu) = 0 for u rt J. From (2.5) we obtain V(lI)(f) = 1. Since feu) =
o whenever one of the eoorclinates of u is 1, we have V(k)(f; h, ... ,ik) = 0
whenever 1 :5 k < 8. Altogether, we obtain VU) = 1. It is clear that
and that
N
!.. J(Xn) = A(J;
NL.J N
'11=1
A8(K):$ feu) du :$ As(J)
}ls
Since As(J) - As(K) < 2::=1 (Vi tli) < c:/2, it follows that
1
1 ;" (. I IA(J;P) . I E: *
N f(x,..) - }t,,-/(U) du > - As(J) - 2 > DN(P) - e. 0
There is also a multidimensional version of Theorem 2.10. For a continuous
function J on is, we define its modul'US of continu.ity by
(2.6) W(ji t) = sup IJ(u) - f(v)l, for t 0,
u,vEi
s

where lIuli = Iu.i I for 11 = (11,'" ,Ua ) E ]R", The following bound is
due to Proinov [288].
\
QUASI-MONTE CARLO METHODS FOR NUMERICAL INTEGRATION 21
,THEOREM 2.13. III is continuous on js, then,
MtJe
I t./(x..) - .'(U) dul !> 4w(f; DN(x" ... ,xN )'/,).
" The error bounds given already in this section all lead to the same conclusion,
namely, that point sets with small star discrepancy guarantee small errors in
qu.asi-Monte Carlo integration over is. Constructions of such point sets will be
discussed in detail in the following chapters.
We now consider more general bounded integration domains. By applying,
if necessary, a tra,nslation and a contraction, we am aasume that such an inte-
gration domain B is contained in il). We have already seen in (1.10) what the
quasi .. MonteCarlo approximation looks like in this case. For convex integration
domains, the following error bound in terms of the isotropic discrepancy was
p1-oved by Zaremba [364].
,'. THEOREM 2.14. If B jB is convex and f hiM bounded varfutioo V(f) on
, "

,XN E ]8, we have
"
,(
. N
I I(x..) - 1. /(U) dul !> (V(f) + 1/(1, . .. ,1 )i)JN(P).
x",EB
The most general situation that we consider is where B belongs to a. family
Mb of Jordan-measurable sets introduced in 2.1. The following theorem is an
improvement by de Clerck [54] on a result of Niederreiter [219].
THEOREM 2.15. If BE Mb and f has bounded variation V(f) on is in the
sense Hardy and Krause, then, for any point set P consisting of Xl, I XN E
we ha've
I t. I(x..) - 1. f( u) dul !> (V (f) + 1/(1, ... ,1 )I)DN(M,; Pl
x"EB
The term 11(1, ... ,1)1 is needed in the inequalities in Theorems 2.14 and 2.15
because without this term the bounds would fail to hold even for constant func-
tions f.
Notes.
The notion of discrepancy was first studied in its own right by Bergstrom [24J,
but the applications to numerical analysis became apparent only after the later
work of Koksma [155]. The use of the isotropic discrepancy was suggested
by Hlawka [136]. The discrepancies DN(Mb; P) were introduced by Nieder-
reiter [219]. If, instead of the maximum deviation between the empirical dis-
tribution and the uniform distribution, we considel' the mean-square deviation,
then -1;he L2 discrepancy is obtained, and, more generally, an IJ> discrepancy
22 CHAPTER 2
may be defined for 1 ~ P < 00. Discrepancies for point sets with nonequal
weights and for nonuniform distributions have also been studied. See Nieder-
reiter [225] for a survey of various notions of discrepancy and Niederreiter and
Tichy [274], Niederreiter, Tichy, and Turnwald [277], Proinov [283], [285], [289],
and Tichy [348] for more recent literature.
Apart from variOilS types of discrepancy, there are other quantities measuring
the irregularity of distribution of point sets, for instance, the non uniformity due
to Sobol' [322] (see also Niederreiter [225,2J) and the diaphony introduced by
Zinterhof [369] (see also Zinterhof and Stegbuchner [371]).
Multidimensional analogues of Theorems 2.6 and 2.7 are not easy to obtain.
Partial results are given in Niederreiter [217] and de Clerck [55], [56]. For two-
dimensional point sets with no repeated coordinates, de Clerck [56] established
an explicit formula for the star discrepancy. The problem of calculating the
isotropic discrepancy was considered by Niooerreiter [217] and Zaremba [3671.
example of Zaremba (362] shows that the exponent l/s the inequaJity
IN(P) ~ 4sD
N
(P)1/11 is the best possible.
A proof of Theorem 2.11 diH'erent from that of Hlawka [134] can be found
in Kuipers and Niederreiter [163, Chap. 2]. Theorem 2.12 is due to Nieder-
reiter [234}. Versions of Theorem 2.13 with_ 4 replaced by larger constants were
proved by IDawka [137] and Shi [310], [311]; according to Proinov [288], the
constant 4 cannot, in general, be replaced by a constant < 1. The paper of
IDawka [137] also contains an ~ r r o r bound for arbitrary Riemann-integrable func-
tions; for the case where 8 = 1, see a100 Proinov [284]. A survey of further error
bounds can be found in Niederreiter [225, 2]. A number of more recent papers
have dealt with error bounds for integration rules with general weights in terms
of an appropriate discrepancy; see Niederreiter and Tichy [214}, Proinov [283),
[284], [286], [281], (288], [289], and Totkov [353]. A variant of Koksma's inequality
was shown by Horbowicz [142]. Integration rules involving values of derivatives
of the integrand have also been considered, and error bounds in terms of suit-
able discrepancies were established; see Proinov [286], [281], [289], Proinovand
Kirov [290], and StegbuChner [334].
Since it is often difficult to obtain the exact value of the variation V(f),
a discretization scheme for the approximate calculation of V(f) described in
Niederreiter [231] can be useful in practice. Upper bounds for V(f) are presented
in Bliimlinger and Tichy [29], Hua and Wang [145, Chap. 5], and Tichy [349].
The notion of discrepancy plays a central role in the theOry of quasi-
Monte Carlo methods and is also important for the theoretical analysis of pseudo-
random numbers, as will become apparent in later chapters. Various types of dis-
crepancy have also found useful applications in a wide range of other areas, such
as the combinatorial theory of irregularities of partitions (Beck [18], SOO [332]),
geometry (Laczkovich [164], Linhart [194]), statistics (Gyires [120], [121]),ap-
proximation theory (Proinov [286},[287]), stochastic algorithms (Lapeyre, Pages,
and Sab [170]), and cryptology (Niederreiter and Schnorr [210]).
CHAPTER
Discre
Seq e ces
Point Sets n
The error analysis for quasi-Monte Carlo integration in 2.2 has demonstrated
that small. errors are guaranteed if point sets with small star or extreme discrep-
ancy are used.. This is true not only for the normalized integration domain Is,
f'nr 1nbcgI'Htion domains contained [8
1
occur-
ring in the inequalities in Theorems 2.14 and 2.15 can be bounded. terms of
the extreme discrepancy by results in 2.L A point set P consisting of N ele-
m.ents of Is is informally called. a low-discrepancy point set if Div(P) or DN(P)
is sm.all. In the computational practice of quasi-Monte Carlo integ-ration, it is
often convenient to be able to change the value of N 'Without losing the previ-
ously calculated. function values. For this purpose, it is recommended. to work
with a sequence of nodes and then to take its first N terms whenever a value of
N has been selected.. In this way, N can be increased while all data from the
earlier computation can still be used. The desirable notion in this connection is
that of a low-discrepancy sequence, which is informally defined as a sequence S
of elements of Is for which Div(S) or DN(S) is small for all N :2: L
this chapter, we initiate the study of low-discrepancy point sets and se-
quences (some authors speak of quasirandom points and quasirandom sequences,
respectively). Further information on this topic can also be found in later chap-
ters. In 3.1 we review some classical constructions of low-discrepancy point sets
and sequences. Several general principles for obtaining upper and lower bounds
for the discrepancy are presented in 3.2.
3.1. Classical constructions.
In the one-dimensional case, it is easy to determine the minimum of the star
discrepancy Div (Xl, ... , X N) and of the discrepancy D N (Xl, ... ,X N) if N is
fixed. In fact, it follows from Theorem 2.6 that we always have
D'N(xl, ... , XN) 2:: 2 ~ '
and equality holds if Xn = (2n-l)j(2N) for 1 ::; n ::; N. The quasi-Monte Carlo
approximation
(I f(u)du ~ ~ t f (2n-l)
Jo lV n=l \ 2N
23
24 CHAPTER 3
with these nodes is a classical integration rule, namely, the N-panel midpoint
rule for the interval [0,1]. Similarly, Theorem 2.7 shows that we always have
1
DN(Xl, ... ,XN) ~ N
and that equality holds if Xn = (2n - 1)/(2N) for 1 ~ n :::; N.
While D N (P) = O( N-l) is possible for N -element point sets P from [0, 1 ],
there is no sequence S of elements of [0,1] for which DN(S) = O(N-l) for all
N ~ 1. On the contrary, we have the phenomenon of irregularities of distribu-
tion, according to which the discrepancy DN(S) is infinitely often of a larger
order of magnitude. Concretely, we have a result of Schmidt [303], which says
that there exists an a.bsolute constant c > 0 such that, for any sequence S of
elements of [0, 1 J, we have
'1 '
.uJg
Currently, the best value of c is c = 0.12 (see Bejian [22]). In view of Proposi-
tion 2.4, we obtain that, for any sequence S of elements of [0, 1], we have
D'N(S) ~ (0.06)N-
1
log N for infinitely many N.
Thus, for low-discrepancy sequences in the one-dimensional case, we cannot
expect anything better than D'N(S) = O(N-1logN) for all N ~ 2. This or-
der of ma.gnitude can indeed be achieved by several constructions. Although
quasi-Monte Carlo methods are not so important in the one-dimensional case
(classical integration rules are to be preferred for dimension 8 = 1), we discuss
two constructions of one-dimensional low-discrepancy sequences, since these con-
structions form the basis for the extension to the multidimensional case.
For an integer b ~ 2, we put Zb = {O, 1, ... ,b-l}; i.e., Zb is the least residue
system mod b. Every integer n 2: 0 has a unique digit expansion
00
(3.1) n = Laj(n)o1
j=O
in base b, where aj (n) E Zb for all j ~ 0 and aj (n) = 0 for all sufficiently large
j; Le., the sum in (3.1) is actually finite.
DEFINITION 3.1. For an integer b ~ 2, the radical-inverse function b in base
b is defined by
00
<Pb(n) = Laj(n)b-
j
-
1
for all integers n ~ 0,
j=O
where n is given by its digit expansion (3.1) in base b.
Thus b (n) is obtained from n by a symmetric reflection of the expansion (3.1)
in the "decimal point." Note that b(n) E I := [0,1) for all n ~ O. In the
construction of\ many low-discrepancy sequences, we find it convenient to index
LOW-DISCREPANCY POINT SETS AND SEQUENCES 25
the terms by n = 0, 1, .... In particular, in our constructions of one-dimensional
low-discrepancy sequences 8, we denote the terms XO, Xl, .. Accordingly,
we then write
DN,(8) = DN(XO,Xl, . ,XN-l)
for the discrepancy of the first N terms of S, and similarly for the star discrep-
ancy.
DEFINITION 3.2. For an integer b ~ 2, the van der Corput seq'uence in base
b is the sequence Xo, Xl, ... with Xn = <Pb(n) for all n ~ O.
" If 8b is the van der Corput sequence in base b, then Div(Sb) = O(N-1log N)
for all N ~ 2 with an implied constant depending only on b. This follows from
a more general result shown in Theorem 3.6 below. Faure [94] established the
following more precise asymptotic result:
f
b2 for even b,
_ ND'" (8) _ ND. (8) .. ' 4(b+. l)logb
lim N b=lim N b=
N ~ o o log N N ~ o o log N b _ l "',
l410gb uuu u.
Consequently, 8
a
is asymptotically the best van: der Corput sequence. The case
where b = 2 is of historical interest, since 8
2
is the sequence tha.t was originally
introduced by van der Corput [356]. In this case, Bejian and Fa.ure [23] proved
that
and
lim (NDN(8
2
) _ lOgN) = ~ + log 3 .
N->oo log 8 9 log 8
Further improvements in terms of the asymptotic behavior of the discrepancy
are obtained by considering a generalized van der Corput sequence in base b,
which is defined by
co
Xn = L u (aj(n)) b-
j
-
1
for all n ~ 0,
j=O
where n is given by (3.1) and (f is a permutation of Zb. This sequence can also
be generated recursively by
u(O)
Xo = b -1 j
1 .
Xbn+r = b(xn + a(r)) for n ;:: 0 and 0:::; r :::; b - 1.
Currently, the best choice of parameters is that found by Faure [100], who used
b = 36 and a specific permutation u of Z36 and showed that the resulting gener-
alized van der Corput sequence S satisfies
lim NDN(S) = 23 = 0.366" ..
N --00 log N 35 log 6
26 CHAPTER 3
At this time, this is the sequence that yields the smallest value of the limit
superior on the left-hand side for any known sequence of elements of [0,1]. For
the star discrepancy, the current "record holder" is a generalized van der Corput
sequence S* in base 12 constructed by Faure [94], which satisfies
-. ND'N(S*) 1919
J ~ o o log N = 3454 log 12 = 0.223 ....
Another class of one-dimensional low-discrepancy sequences is obtained by
considering the multiples of suitable irrational numbers modulo 1. For u E R let
{u} = u - l u J be the fractional part of u. Note that {u} E I always. For an
irrational number z, let S(z) be the sequence XQ, Xl.' . with
Xn = {nz} for all n ~ O
.Let
z = [aO;al,a2,"']
be the continued fraction expansion of z, with partial quotients aD, ar, ... E Z
and ai ~ 1 for i ~ 1 (compare with Appendix B for a background on continued
fractions). The denominators of the convergents to z are denoted by qo, qI, ....
THEOREM 3.3. Let z be irrational and N ~ 1. Then N can be represented
in the form
I(N)
N = L Ciqi,
;=0
where l(N) is the unique nonnegative integer with ql(N) :::; N < q'(N)+l and
where the Ci are integers with 0:::; Ci :::; ai+l for 0 :::; i :::; leN). Furthermore, we
have
1 leN) 1 l(N)+l
DN(S(Z)) < N L: (Ci + 1) :::; N L ai
i=O i=l
Ci;CO
Proof. Since 1 = qo :::; ql < q2 < ' , . , the existence and uniqueness of leN) is
guaranteed. Now we can write N = CI(N)qZ(N) + r with integers Cl(N) ~ 1 and
o :::; r < q'(N)' If we had C!(N) > aZ(N)+l, then
a contradiction. Thus Cl(N) :::; aZ(N)+l' If r > 0, then we apply this procedure
to r instead of. N, and, continuing in this manner, we arrive at the desired
representation for N.
Given this representation, we decompose the point set SN(Z) consisting of
the fiI'.st N terms of S(z) into blocks of consecutive terms, namely, Ci blocks of
length qi for 0 :::; i :::; l(N). Consider such a block of length qi for fixed. ij it is a
point set Pi consisting of the fractional parts {nz}, n = ni, ni + 1, . .. , ni + qi - 1,
\
LOW-DISCREPANCY POINT SETS AND SEQUENCES 27
for some integer ni ;?:: O. Let Pi/qi, be the ith convergent to z. Then, from the
theory of continued fractions,
Pi 8i
z = - + -- with 18i I < l.
qi qiqi+1
Thus, if n = ni + j, j E Zq" as above, then we obtain
{nz} = -' +niz+ . {
jp. j8 }
qi qiqi+1
Since gcd(Pi, qi) = 1, the fractional parts {jpi/qi + niz}, j E Zqp form a point
set Qi of qi equidistant points in I with distance 1/q;., and these points have
discrepancy 1/qi by Theorem 2.7. Because
the point set Pi is obtained by displacing modulo 1 the elements of Qi in one
'In 15; ) h'T < ','
",/ '" .. ,.
1 1
Dq.(Pi ) < - +-.
qi qi+1
From the triangle inequality for discrepancies (see [163, p. 115]) and the way in
which we decomposed SN(Z), it follows that
I(N) l(N) l(N)
NDN(S(Z)) < L Ci (1 + L (Ci + I)Ci + 1),
;'=0 i=O q,+l i=O
c.#O Ci#O
which is the first bound for DN(S(Z) in the theorem. The second bound is
deduced from two properties of the coefficients Ci obtained by the algorithm at
the beginning of the proof, namely, that Co < ql = al and that Ci = aHl implies
Ci-l = O. To prove the second property, we note that, if qi N < qHl and
Ci = aH1! then N - Giqi = N - aHlqi < qHl - aHlqi = qi-l 0
COROLLARY 3.4. If the irrational z is such that ai = O(m), then
DN(S(Z)) = O(N-llogN) for all N ;?:: 2.
Proof. Theorem 3.3 and the hypothesis imply DN(S(Z)) = O(N-l(l(N) +
1)). Induction shows that qi ;?:: for all i 0, where a = (1 + -15)/2. Thus
N ;?:: ql(N) ;?:: al(N)-l, and the desired bound follows. 0
Theorem 3.3 and Coronary 3.4 show that S(z) is a low-discrepancy sequence
if the partial quotients ai, i 1, in the continued fraction expansion of z are
small. The following is a particularly interesting special case.
COROLLARY 3.5. If the irrational z is such that there exists a positive integer
K with ai :::; K fOT i 1, then
DN(S(Z) < G(K)N-llog(N + 1) for all N 1,
where G(K) = 2/ log 2 for K = 1,2,3 and G(K) = (K + 1)/ log(K + 1) for

28 CHAPTER 3
Proof. In view of Theorem 3.3, it suffices to show that
l(N)
(3.2) seN) := L (Ci + 1) ~ G(K) log(N + 1) for all N ~ 1.
i=O
c.o
Here seN) is well defined if we use the coefficients Ci produced by the algorithm
at the beginning of the proof of Theorem 3.3. We establish (3.2) by induction
on the value of leN). If qo < ql, then the least possible value of leN) is zero,
and a corresponding N satisfies 1 :::; N < ql :$ K. If qo = ql = 1, then the least
possible value of leN) is 1, and a corresponding N satisfies 1 :::; N:::; q2 -1 :$ K.
Since seN) = N + 1 for these N, it suffices to show, for the first step in the
induction, that
(3.3) N + 1 ~ G(K) lOf!;(N + 1) fry!" 1 < <
This follows readily from the definition of G(K), however. Now consider an
arbitrary 1 with qz > 1 and a cor.responding N with leN) = 1, hence with
qz :::; N < ql+l We write N = czql+r with 0:::; r < qt. Then s(N) = cl+l+s(r),
and the induction hypothesis yields
seN) :::; Cl + 1 + G(K) log(r + 1),
which also holds for r = O. Now N + 1 = czqz + r + 1 ;:::: (Cl + l)(r + 1) and
1 ~ Cl :$ aHl :::; K. Thus, by (3.3),
seN) :::; G(K) log(CI + 1) + G(K) log(r + 1) :$ G(K) log(N + 1),
and (3.2) is shown. 0
The sequences 8(z) have an obvious analogue in the multidimensional case
s ~ 2. We first define the fractional part of u. = (Ul' ... jUs) E JR,s by
Now let z = (z}, . .. ,zs) E lR
8
be such that 1, Zl, . ,Zs are linearly independent
over the rationals, and let 8(z) be the sequence Xo, Xl!'" with
Xn = {nz} for all 11, ~ O.
Then, by a. classical result [163, p. 48], the sequence 8('.1.) is uniformly dis-
tributed in Is. For the discrepancy of 8(z), there is a probabilistic result
of Schmidt [302] to the effect that, for every e > 0, we have DN(S(Z)) =
O(N-l(l + logN)8+
H
e) for A.-almost all Z E :lR
8
. For s ~ 2, no individual
z E lR.!I with DN(S(Z = O(N-l(l + logN)8+1) is known. If the coordinates
of Z E lR
s
are algebraic numbers satisfying the linear independence condition
stated ab<!ve, then DN(8(z)) = O(N-1+e) for every e > 0 by a result of Nieder-
reiter [218]. Point sets that can be viewed as discrete ve:rsiOI1B of the sequences
8(z) will be discussed in 5.1.
\
LOW-DISCREPANCY POINT SETS AND SEQUENCES 29
A more satisfactory way of producing low-discrepancy sequences in any di-
mension is obtained by extending the notion of Ph van der Corput sequence in-
troduced in Definition 3.2. This leads to the family of Halton sequences. For a
given dimension s ~ 1, let b
l
,. .. ,b
s
be integers ~ 2. Then, using the radical.-
inverse functions <Ph in Definition 3.1, we define the Halton sequence in the bases
bI,'" ,ba as the sequence Xo,Xl,'" with
For s = 1 this definition reduces to that of Ph van der Corput sequence.
THEOREM 3.6. II S is the Halton sequence in the pairwise relatively prime
bases hI, . .. ,b
s
, then
D* (8) ~ ~ n
S
(b
i
- 1 1 N b
i
+ 1)
N < N + N 21 b. og + 2
i=l og ,
-,-,)
, ~ ; ---.IV) ~
'i\T .......,. --.j
.J.1
nl
for aU N ~ 1.
(
\ .
J ~ 1
8
, where 8
N
is the point set consisting of the first N terJr1l of the Halton
sequence. For 1 SiS ,8 and an integer e ~ 0, let Ei (e) be the family of all
intervals [0, ab;e) with a E Z, 0 < a S bf, and let Fi (e) be the family of all
intervals [cb;f, (c + l)b;f) with c, fEZ satisfying 0 ::; I ::; e and 0 ::; C < h{.
For integers ell ... , e
8
~ 0, let ' (ell .. , , e
a
) be the family of all intervals E =
I1:=1 Ei with B;, E i(ei) U Fi(ei) for 1 S i ::; s. !e claim that
(3.4)
ID(E)I:S IT n(b, -l)e,+ 1)
a
for all E = II Ei E &(eI, ... ,ea ),
i=l
E.(/.:Fi(e;)
where an empty product is meant to be 1. We prove (3.4) by induction on the
number k of indices i for which Ei . .1i(ei).
First, let k = 0, i.e., E = rr:=tl Cih;f" (Ci + l)b;li) with Ci, Ii E Z, 0::; Ii ::;
e 0 < c < b!' for 1 < i < s. Note that
~ , - ~ ~ --
if and only if, for 1 S i :::; s, the first Ii digits (in base bi) of <Phi (n) after the
"decimal point" have certain prescribed values. Equivalently, for 1 :::; i ::; s, the
Ii least significant digits of n in base b
i
must have certain prescribed values, or
in other words, n must lie in a prescribed residue dass mod b{i. Since b
I
, ... ,b
8
are pairwise relatively prime, it follows from the Chinese remainder theorem
that the last condition is equivalent to n lying in a prescribed residue class
mod m = b{l ... b!s. Consequently, among any m consecutive terms of S, exactly
one of them lies in E. This implies that ID(E)I :::; 1,
Now suppose that for some k ~ 1 the claim has been established for k - 1,
and.consider E = rr:=l Ei E &(el! ... , es ), where we assume without loss of
30 CHAPTER 3
generality that Ei Fi(ei) for lSi 5 k and Ei E Fi(ei) for k + lSi 5 s.
Then El = [0,ab
l
e1
) for some a E Z, 0 < a < Let
el
ab
l
el
= L dj b
1
j
j=l
be the digit expansion in base h
l
. Then E1 can be written as the disjoint union
of d
1
intervals in .1i(e1) of length b
l
\ of d
2
intervals in ri(el) of length b
l
2
,
and so on. Thus
r=l
with pairwise disjoint Fr E Fl(el) for 1 5 r 5 d := dj , and so E is the
disjoint union
It follows that
d
17'_ ! 117'
, - V l
r=l
d Ie (1 )
(3.5) ID(E)/5 x E2 x ... x Es)15 d!! 2(bi -l)ei + 1 ,
where in the second inequality we applied the induction hypothesis. With G =
[ab
l
e1
, 1), we have El = [0,1) " G; thus
ID(E)15 ID((O, 1) x E2 X . x Es)1 + ID(G x E2 x ... x Es)1
g ( - 1) .. + 1) + ID(G X E2 x ... x E.)I.
Now G can be written as the disjoint union of (b1 - l)el - d + 1 intervals in
FI(el); hence
ID(E) I b1 - 1)el - d + 2) f! -1)e, + 1).
Adding this inequality to (3.5) and dividing by 2, we get (3.4) for the value k.
Now let J = 1 [ 0, Ui) IS be arbitrary. For lSi 5 8, let ei be the
least integer with b:
i
N and let ai be the least integer with aib;e, Ui. Put
E = n:=l[O,aib;e,). Since the ith coordinates of all points of SN are rationals
with denominator b:', we have A(JjSN) = A(E;SN). Furthermore, we have
E E E(el; ... ,e
s
), and so
ID(J)I 5 N (AaCE) - Aa(J + ID(E)/
8
5 NLb;e
i
+ ID(E) I 5 s + ID(E)I
i=l
LOW-DISCREPANCY POINT SETS AND SEQUENCES 31
by (3.4). To finish the proof, note that ei < 1 + (log N) / (log hi) for 1 ::;;
i::;; s. 0
For a dimension 8 ~ 2 and for integers N ~ 1 and bI, ... ,bs - l ~ 2, the
N -element Hammersley point set in the bases hI, ... ,b,-l is given by
x.= (;,qI,,(n), ... , >'._, (n)) eI
8
for n=O,l, ... ,N-l.
A bound for the star discrepancy of this point set is obtained from the following
general principle.
LEMMA 3.7. Fof'S 2: 2, let 8 be an arbitrary sequence of points XO,XI,'"
in ]B-1. For N ~ 1, let P be the point set consisting of (n/N,x.n) E IS for
n = 0, 1, ... ,N - 1. Then
ND'N(P) < max MD'M(S) + L
- I$M$N
Fbr an arbitrary J = l 0, ud ~ ,we have (nll'v, An) E J it
and only if Xn E JI := n:=2[O,Ui) and n < NUl' If J.\1 is the largest integer
< NUl + 1, then A(J; P} = A(J'; 8M), where 8M is the point set consisting of
the first M terms of 8. Therefore
IA(J; P) - NAs(J) I ::;; IA(J'j 8M ) - MAs-l (J')I + IMAs-I(J') - NA,(J)I
::;; MD'M(S) + IMAs-I(J') - NAs(J)I
Now NUl::;; M < NUl + 1; thus
s s
0::;; MAs-l (J') - NAs(J) ::;; (NUl + 1) II Ui - NIl Ui ::;; 1,
i=2 i=l
and the desired result follows. 0
THEOREM 3.8. If P is the N -element Hammersley point set in the pairwise
relatively prime bases hI,. >. ,b
s
- 1 , then
a-;( )
s 1 - b - 1 b + 1
D'N(P) < N + N 2 io b. log N + T .
i=l g z
Proof. For the proof, use Theorem 3.6 and Lemma 3.7. 0
It follows from Theorem 3.6 that, if S is the Halton sequence in the pairwise
relatively prime bases b
1
, ,hs, then
where the coefficient of the leading term is given by
32 CHAPTER 3
The minimum value of this coefficient is obtained by letting b
l
, ... ,b
s
be the
first s primes PI = 2, 1>2 = 3, ... ,Ps' In this case,
where As = A(pI,'" ,Ps)' In a similar vein, to minimize the bound in The<r
rem 3.8, we choose bl , ... ,bs- 1 to be the first 8 - 1 primes PI, .. ,Ps-I, and
then we obtain
for the corresponding N-element Hammersley point set P with N 2. The dis-
crepancy bounds for Hammersley point sets and Halton sequences serve as bench-
marks for further constructions of low-discrepancy point sets 8....11d sequences. In
fact, it has become customary to speak of low-discrepancy point sets and se-
n,,.... "A ""':':;:;38 - .,.,\ ... 0(1\7---1
1'""" """v )"-' " i J
dimension 8, with N being the number of points considered.
We emphasize that, in view of Theorem 2.11 and the discrepancy bounds
above, the use of Hammersley point sets and Halton sequences in quasi-Monte
Carlo integration leads to a dramatic improvement on the Monte Carlo error
bound O(N-
1
/
2
), at least as far as the order of magnitUde is concerned.. Con-
cretely, if the integrand is of bounded variation on Is in the sense of Hardy
and Krause, then the quasi-Monte Carlo error bound is O(N-I(1ogN)8-1) for
an N -element Hammersley point set in pairwise relatively prime bases, and
O(N-l(logN)B) for the first N terms of a Halton sequence in pairwise rel-
atively prime bases, where N 2. Further advances in the construction of
low-discrepancy point sets and sequences (see Chapter 4) yield corresponding
improvements in quasi-Monte Carlo methods for numerical integration.
It is widely believed. that, in the a-dimensional case, the star discrepancy of
any N -element point set P satisfies
(3.8)
where the constant Bs > 0 depends only on s. This would mean that the star
discrepancy of an N -element Hammersley point set in pairwise relatively prime
bases attains the least possible order of magnitude. As mentioned previously
in this section, (3.8) is obvious in the case where s = 1. For s = 2, (3.8) was
established by Schmidt [303], but (3.8) is still open for s 3. The best general
result in this direction is due to Roth [296], who showed. that, for any N-element
point set P in dimension s, we have
For s = 3, a slight improvement was recently obtained by Beck [19], namely, that
Div(P) BsN-
1
(log N) (log log N)C for N 2: 3,
where c > 0 is an absolute constant.
LOW-DISCREPANCY POINT SETS AND SEQUENCES 33
If (3.8) holds, then, with the help of Lemma 3.7, we would easily obtain that
any s-dimemlional sequence S satisfies
(3.9) DN(S) for infinitely many N,
where the oomltant > 0 depends only on 8. This would mean that the
star discrepancy of a Halton sequence in pairwise relatively prime bases attains
the least possible order of magnitude. By a result of Schmidt [303], mentioned
previously in this section, (3.9) holds for 8 = 1, but this is the only case in
which (3.9) is known. The results of Roth [296] and Beck [19], stated above,
have analogues for sequences. For arbitrary 8 and any s-dimensional sequence
S, we have
DN(S) for infinitely many N,
while any two-dimensional sequence S satisfies
>
many N,
where d > 0 is an absolute constant. These results on irregularities of distribu-
tion are quite difficult to prove. Expository accounts of this topic are given in
the books of Beck and Chen [20] and Kuipers and Niederreiter [163J.
3.2. Gen.eral discrepancy bounds.
Because of the finite-precision arithmetic of computers, all point sets and se-
quences of practical interest consist only of points with rational coordinates. For
this case, there are several important principles for obtaining upper and lower
discrepancy bounds.
For an integer M 2, let G(M) = (-M/2,M/2] nz and G*(M) = G(M) "-
{O}. Furthermore, let Gs(M) be the Cartesian product of s copies of GeM) and
G:(M) = Gs(M) "- {o}. Put
r(h, M) = {Ml sin for h E G*(M),
for h = O.
8
r(h, M) = II r(hi' M).
i=l
We write e(u) = e
2
'1rv'-T
u
for u E J.R and X Y for the standard inner product of
x,yEJ.RB.
LEMMA 3.9. Let ti, Ui E [0,1] for 1 :::; i :::; s and let v E [0,1 J be such that
Iti - uil :::; v for 1 :::; i :::; s. Then .
I
B ti - 8 Ui I :::; 1 - (1 - v) 8.
i=l i=l
34 CHAPTER 3
Proof. We proceed by induction on s, with the case 8 = 1 being trivial. If
the inequality is shown for some 8 2:: 1 and we assume without loss of generality
that ts+! 2:: Ua+!, then
lIt ti -It Uil ~ (t'+1- U'+1) g ti +u,+lg ti - i ~ ' Uil
:::; ts+l - Us+l + us+! (1 - (1 - v)S)
= ts+! (1 - (1 - v)S) + (ts+l - us+d (1- v)S
:::; 1- (1- v)S + v(l - v)S = 1 - (1- v)s+l.
o
THEOREM 3.10. For an integer M 2:: 2 and Yo, ... ,YN-l E ZS, let P be the
point set consisting of the fractional parts {M-lyO}"" ,{M-1YN_d. Then
(
1 ) s 1 1 N-l (1 )
DN(P)::; 1- 1- lvI + L reh M) N 2: e' M
h
' Yn .
hEC:(M)' n=O
Proof For k = (k
1
, ... ,ks) E ZS, let A(k) be the number of n with 0 :::;
n :::; N - 1 and Y n = k mod M, where a congruence between vectors is meant
componentwise. Then
since the inner sum has the value MS if Yn = k modM and the value zero
otherwise. Therefore
N 1 (1) N-l (1 )
(3.10) A(k) - Ms = Ms L e - M
h
. k L e M
h
. Yn .
hEC;(M) n=O
Now let J = rr:=l [ Ui, Vi) be a subinterval of IS. For each i, 1 :::; i :::; 8, we choose
the largest closed subinterval of [Ui' Vi) of the form [ai/M, bi/M] with integers
ai:::; bi, which we again denote by [ai/M,bi/M]. The case where for some ina
such subinterval of [Ui' Vi) exists can be easily dealt with, since we then have
A(J; P) = 0 and Vi - Ui < ljM; hence
(3.11)
A(J; P) _ A (J) = A (J) < ~ < 1 - l 1 - ~
I )8
N s 8 M- ! M
\
In the remaining Fase, the integers ai, b
i
, 1 :::; i :::; 8, are well defined, and we
LOW-DISCREPANCY POINT SETS AND SEQUENCES 35
obtain, by (3.10),
_ A.(J) (At) - + -a; + 1) - A.(J)

Now
I
b
i-
a
i+
1
I 1
M - (Vi - Ui) < M
for 1 :::; i :::; s,
and so an application of Lemma 3.9 yields
, A {P\ . I (1 \ 8
t
3
.
12
) - AS(J)I < 1- \1- M)
L I L
hEe(M) k . n=O
a,$k.$bi
}lor fixed h = (hI!", , h
s
) E C;(M), we have
If hi = 0, then
whereas, for hi E C*(M), we have
Therefore
We use this in (3.12) and obtain a bound for the left-hand side, which is valid
for all_ J because of (3.11). 0
36 CHAPTER 3
COROLLARY 3.11. Let P be as in Theorem 3,10 and suppose that the real
number B is such that
for all hE C;(M).
Then
DN(P)"; 1- (1- r + ! (,,; logM +1.72 r
Proof. From (3.12) we obtain
From the proof of Theorem 3,10, we obtain
L I L e( < L
heC;(M) k \ hECs(M)i=l k.=O

= fI( L /bI:' i)
i=l heC{M) k.=O M ' I
fI(M + - ai + l)/M) I).
i=l h=l sm(1rh/M)
By an inequality of Cochrane [42], we have
sin ( 1rhj / M) I 4
(3.13) sin(1rh/M) < 11"2 Mlog M + (0.41)M + 0.61
for any integer j, and this yields
I
A(J;P) I ( 1)8 B (4 0.61)8
N - AB(J) < 1 - . 1 - !vI + N 1r2 log M + 1.41 + M .
In view of (3.11), this holds for all J, so that the desired bound for DN(P)
follows. 0
We now cO:rulider point sets for which all coordinates of all points have a finite
digit expansion in a fixed base b 2. Let
(3.14)
- ( (1) (8)) E 1
8
wn - 'W
n
, .. ,W
n
for n = 0,1, ... ,N - 1,
where, for an integer Til 1, we have
m
(i) _ '" (i) b-
j
Wn - L.; Wn,j
for 0 :; n N - 1, 1:; i 8,
j=l
LOW-DISCREPANCY POINT SETS AND SEQUENCES 37
with E Zb for 0 ::::; n ::::; N -1, 1 ::; i ::::; 8, 1 j ::;: m. For simplicity, we write
instea.d of and a. similar convention applies in the rest of this section.
For (hI, ... ,hm) E Cm(b), define
(3.15)
d(h
1
, ... ,h
m
} = {olargest d with hd "# 0 if (hI!'" ,h
m
) =I: 0,
if (hI!'" ,hm ) = O.
For b = 2, put
(3.16)
Q (h h )
- 2-d(h1 , . ,hm )
b 1, .. , m - .
For b> 2, put
\ "1 J"> - 1 _ ,/ 'J\
) 'l/-"j - J '-"" i j f.; \, I ......
Let C(b)8Xm be the set of all s x m matrices with entries in C(b). For H =
(h.j) E C(b)BXm, we define
8
(3.18) Wb(H) = II Qb(hib . ,him).
i=l
THEOREM 3.12. If P is the point set (3.14), then
where the outer sum is over all nonzero matrices H = (h
ij
) E C (b) a x m.
Proof. Let J = n:=1 (0, ti J IS. Then
N-l
(3.19) A(Jj P) = L C(ll! . .. c(ta,
n=O
where c(t, .) is the characteristic function of the interval [0, t]. For 1 ::::; i ::::; 5, let
00
ti = L tir b-
r
r=l
be the digit expansion of ti in base b, where all tir E Zb! and, for each i, we have
tir < b - 1 for infinitely many r. Put
(3.20)
Uir =
{
tir for 1 ::::; r < m,
tir -+ 1 for r = m.
38 CHAPTER 3
For v, wE Zb let6(v,w) = 1 if v = wand 6(v, w) = 0 if v =I- w. Then
m Uir-1
C(til = L 8(Ui} , ... 6(Ui,r-l, L 6(v,
r=l v=O
where we henceforth use the standard convention that empty sums have the
value zero and empty products have the value 1. Now
6(v,w) = L -V)),
hEC(b)
and so
with
T(h,L) e(
for L 2:: O.
Substitution in (3.19) yields
N-l m
A(J; P) = I:: L b-
r1
- .. -
r
"
In the innermost sum, we. split off' the term corresponding to the choice h
ij
= 0
for 1 j ri, 1 i s. Also using T(O, L) = L and putting
m
Ui = LUirb-r for 1 i 8,
r=l
LOW-DISCREPANCY POINT SETS AND SEQUENCES 39
we obtain
8 m
A(JjP) = NIl u, + L b-
r1
--
r
"
i=l 1'1 ..... 1'.=1
Lemma. 3.9 yields
with
m
V(hil Ui) = 2:: b-rT(hir , Uir).
r=a(hi )
To complete the proof, it thus suffices to show, in view of Remark 2.3 and (3.18),
that
(3.21)
If hi = 0, then
jV(hi' Ui)! = It b-rT(O, Uir)! = It b-rUir! = Ui :::; 1,
1'=1 11'=1
and (3.21) is shown. If hi =I- 0 and b = 2, then put d = a(hi ) = d(hil, . .. ,him)
and note that T(l, L) attains only the values 0 or -l. Together with (3.15), this
yields
m
V(hi,Ui) = 2-
d
T(1,uid) + L 2-
rU
ir,
r=d+l
40 CHAPTER 3
and so !V(h.i, ui)1 2-
d
, which proves (3.21) in view of (3.16). If h.; =1= 0 and
b> 2, then, again putting d = a(hi ) = d(hil , ... j him), we obtain
IV(h.;, Ui)! = jb-dT{hid, Uid) +- t b-rUirl
r=d+l
b-
d
IT(hid, uid)1 +- b-du(d, m).
For h E C*(b), we have
I
e(hL/b) - 11 2 1rlhl
IT(h, L)I = e(h/b) _ 1 leeh/b) _ 11 = esc -b-;
thus (3.21) follows because of (3.17). 0
LEMMA 3.13. Let s 2:: 1 and m 2:: 1 be integers. Then for b> 2 we have
where the first sum is extended over all mat-rices H E C{b).!ixm. In the case
where b = 2, we have
(;+
Proof. From (3.18) we obtain, for any b,
8
LWb(H) =
H H=(h'j) i=l
= ( Qb(hl, ... ,
11. 1 , ,h",C(b)
To evaluate the last sum, we split it up according to the value of d(hb ... ,h
m
)
for (h
ll
... , h
m
) E Cm{b). For 1 d m, there are b
d
-
1
elements (hI, ... ,h
m
)
of Cm(b) with d(hl"" , hm) = d and a fi.."'{ed nonzero value of he!.. Thus by (3.17)
we obtain, for b > 2,
Q fh h) 1 0;';""< Ed-l,-d (00. o. 0 1rlhl (d))
L.-J b\ 1,, m = + [) 0 L.:{ .0 CSC-
b
- +u ,m
hl, ... ,hmEC{b) d-1 hEC ,b)
m ?ilhl b-1
= 1 + b
CSC
T + --r;- u(d, m)
0=1
m 1rlhl m-1
- b csc-b-+m--
b
-
hEC*(b)
LOW-DISCREPANCY POINT SETS AND SEQUENCES 41
By [222, p. 574J, we have
(3.22)
and, by combining these results, we obtain the first part of the lemma. In the
case where b = 2, we use (3.16) to obtain
m
L Qb(hl, ... ,h
m
) = 1 + L2d-12-d = ; + 1,
hl\ ... ,hmEC(b) d=l
and the second part of the lemma follows. 0
We now consider some general principles of obtaining lower bounds for the
discrepancy. The following is an elementary lower bound for the discrepancy of
point sets comprising only points with rational coordinates.
point set with the property that aU CJxYY'dinates of all points are rational numbers
in [0, 1) with denominator JIll. Then
DN(P) ~ 1 - (
1
- ~ r
Proof. All points of P lie the interval J = [0,1- M-1]S ~ JB. Thus, by
Remark 2.3, we obtain
o
REMARK 3.15. For an integer M ~ 2, consider the point set P consisting of
the N = MS points (11,1/ M, ... ,ris / M) E JB, where 1/,1, ... , 11,8 run independently
through Z M. Then
DN (p)' > DN$ (P) > 1 _ (1 _ ~ . )\ 8
- \ - \ .M
\
by Theorem 3.14. On the other hand, if the points of are denoted by M-
1
y.n,)
:::; 11, :::; N - 1, then, for any h = (hI, .. , ,hs) E C;(JlIl), we have
42 CHAPTER 3
and so Theorem 3.10 implies that DN(P) ~ 1 - (1- M-
I
r. Therefore
This shows, in particular, that the term 1 - (1 - AI-I) s in Theorems 3.10 and
3.14 is, in general, the best possible. The expression 1 - (1 - M-Ir may be
viewed as the discretization error, since it arises from the specific discreteness
property of the point sets in Theorem 3.14. Note that in first approximation the
discretization error is s / M for large M.
It is often easier to obtain formulas or lower bounds for cosine or exponential
sums than for discrepancies. The following results can then be used to derive
lower bounds for discrepancies. For h = (hI, . 0 ,hs) E ZS, we write
s
r(h) = II max(l, Ihil).
i=l
Note that Theorem 3.16 and Corollary 3.17 can be viewed as special versions of
Theorem 2.11.
THEOREM 3.16. For arbitrary to, tlJ ... ,tN-l E lIt
s
let P be the point set
consisting of the fractional parts {to}, {tl}, ... ,{tN-I}. Then, for any nonzero
hE ZS and any real 6, we have
1
1 N-l I 2
N ~ cos 21r(b . tn - 6) ~ ;((11" + l)m - l)r(h)DN(P),
where m is the number of nonzero coordinates of h.
Proof We mst consider the case where h = (hI, ... ,ha) with Ihil = 1 for
1 ~ i ~ s. Then the function f(u!, ... ,us) = feu) = coo 211"(h u - 6) satisfies
f(ul,'" ,Uk-ll Uk + !, Uk+h 0 ,'Us) = - f(ul, ... ,Uk-I, Uk, Uk+lI. 0, Us)
for each k with 1 ~ k :5 s and all UI, 0 ,Uk-lJ Uk+l," 0 ,Us E [0,1], Uk E (0, ~ ].
Thus we can apply [234, Lemma. 5.3]. With suitable 6
i
E lR, this yields
I
N-I I
~ ~ coo 21r(h . tn - 6)
Now
. - r
l/2
r
l/
'},/ ( j ) I 1
Jo ... Jo cos 21r t; hiui - 6i dUl ... dUj = 2i-I1r'
\
LOW-DISCREPANCY POINT SETS AND SEQUENCES 43
and we arrive at the inequality of the theorem in the case under consideration.
Next, we treat the case where h = (hI,.,. ,hs) with all hi i= O. Let tn =
... ,t}:)) and put Wn = ... ,lh.'llt!: for 0 ::; n ::; N - 1. Then
b tn = I W n , where all coordinates of I have absolute value 1. According to
we have already established,
1
1 N-l I 11 N-l I
N cos 211"(b. tn - 0) = N cos 211"(1 Wn - 0)
:::; + 1)8 -l)DN({wo}, ... ,{WN-l})'
11"
To bound the last discrepancy, we note that, for a positive integer h and t E R,
we have {ht} E [u, v) [0, 1) if and only if {t} belongs to one of the h intervals
[(q + u)jh, (q + v)jh), q = 0,1, ... ,h - 1. Using this property, we obtain
this again yields the inequality of the theorem,
Finally, we consider an arbitrary h = (hI, ... ,hs) i- We can assume
loss of generality that hi =f. 0 for 1 i :::; m and hi = 0 m + 1 :::; i 8.
T. b
i
- Ih h . ) d 'f t - (t(l) t(8) "", - ft(l)
. - \ 1, .. " m, an 1 n - n, . , n ., we put [,n - \ n )... , n
o S n :::; N - 1. According to what we have already established, we obtain
1
1 N-l I I 1 N-l !
N L c08211"(b tn - 0) = I N L cos 2'li(h' . - 8)1
n=O n=O I
:::; + 1)7n - l)r(h')D
N
( ... ) }).
11" -
Now r(h') = r(h), and, from the definition of the discrepancy, we see that
Thus we have proved the theorem in the general case. 0
COROLLARY 3.17. For arbitmry to, , ... , tN-l E R8
7
let P be the point set
consisting of the fractional parts {to}, {td, ... ,{tN-d. Then, for any nonzero
h E Z
8
, we have
whe'i'e m is the number of nonzero coordinates of h.
Pmoj. We write
N-l I
N
-
1
I
2: e(h . t
n
) = e(8) e(h . t
n
)
n=O n=O
44 CHAPTER 3
with some real 0, a representation that is possible for any complex number. Then
and, taking real parts, we obtain
IN-l I N-l
e(h.tn)1 = cos27r(htn -0).
The desired bound now follows from Theorem 3.16. 0
Notes,
For the van Corput sequence 8
2
in bnse '2,a proof of DN(S,),) =
O(N-llog(N + 1) can be found in [163, p. 127]. Haber [122] showed that
NDj.,(S2) (logN)j(log8) + 0(1) and that the constant Ij(log8) is the best
possible. The van del' Corput sequence in base b can be obtained by iteration
of an ergodic transformation on [0,1), as was demonstrated by Lambert [165];
see also Lapeyre and Pages [169]. This fact was used by Lambert [166] for anal-
ogous constructions in dimensions 8 =2,3,4. Further work on the (star) dis-
crepancy of generalized van del' Corput sequences was carried out by Bejian [21],
Faure [94], [96], [97], [99], [100], and Thomas [347]. The study of the discrep-
ancy of the sequences S(z) with irrational z is a classical area of number theory;
see [163, p. 128] for work prior to 1974. More recent papers on this topic are
Dupain [72], Dupain and 80s [73], Ramshaw [291J, SchoiBengeier [305], [306],
[307], and 80s [331]. In particular, SchoiBengeier [305] has shown the converse of
Corollary 3.4; i.e., if DN(8(z)) = O(N-llogN) for all N ;:::: 2, then necessarily
E::l ai = Oem).
Distribution properties of the multidimensional sequences 8(z) were recently
studied by Larcher [172], [176], [178] and Liardet [191]; see [163, p. 129] for
references to earlier work. The sequences 8(z) play an important role in a quasi-
Monte Carlo method for the numerical integration of periodic functions, which
is based on the theory of diophantine approximations (see [225, 5]).
Halton sequences were introduced in Halton [125], and Hammersley point
sets in Hammersley [129]. Discrepancy bounds of the orders of magnitude in
Theorems 3.6 and 3.8, respectively, were established in Halton [125], and the im-
plied constants were first improved by Meijer [209] and then by Faure [93]. The
bounds in Theorems 3.6 and 3.8 are essentially those of the latter paper, but we
have used a somewhat different method of proof. The principle in Lemma 3.7 is
attributed to Roth [296]; for a converse of this principle, see [163, p. 106, Ex. 2.2].
The paper of Roth [296] already contains the construction of two-dimensional
Hammersley point sets in the base 2. For this case, and for the number N of
being a power of 2, an exact formula for the star discrepancy was es-
tablished by Halton and Zaremba [128J. More generally, an exact formula for
the star discrepancy of two-dimensional N -element Hammersley point sets in an
\
LOW-DISCREPANCY POINT SETS AND SEQUENCES 45
arbitrary base b, with N being a power of b, was given by de Clerck [55], [56].
Computer implementations of Halton sequences and Hammersley point sets are
described in Fox [107], Halton and Smith [127], and Ucot [183]. Generalized Hal-
ton sequences, which are the obvious analogues of generalized van der Corput
sequences, were first considered by Braaten and Weller [32], and further results
on these sequences can be found in Hellekalek [132]. For generalized Hammer-
sley point sets in the two-dimensional case, see Faure [98]. Other variants of
Hammersley point sets are discussed in [225, pp. 977-978].
Theorem 3.10 is due to Niederreiter [224]. Theorem 3.12 is a special case
of a result of Niederreiter [241] in which the points (3.14) may be such that in
each coordinate the digit expansions can have a different length and a different
base. A somewhat weaker form of Theorem 3.12 was already stated in Nieder-
reiter [229], and the one-dimensional case was proved in Niederreiter [232]. A
discrepancy formula that is more general than that in Remark 3.15 was estab-
lished in Niederreiter [241]; note that the discrepancy formula in Remark 3.15
can also be derived by elementary counting arguments. Theorem 3.16 was shown
dimensional case, the problem of obtaining a result like Corollary 3.17 with opti-
mal constants was studied by Horbowicz and Niederreiter [143] and Niederreiter
and Horbowicz [265].
Low-discrepancy point sets and sequences and irregularities of distribution
have also been considered for domains other than intervals. A useful survey
can be found in Beck and Chen [20]. The case of spheres has received special
attention; see Hlawka [138] and Lubotzky, Phillips, and Sarnak [195], [196] for
important work on thiB case, and Tichy [351] for applications. A detailed study
of various other special domains was recently carried out by Hlawka [139].
CHAPTER 4
N'ets and ces
We have seen in 3.1 that, for an s-dimensional Halton sequence in pairwise
, relatively prime bases, we have Div(S) = O(N"71(logN)S) for all N 2. By
,optimizing the choice of bases, we arrived at the discrepancy bound (3.6). Let
mi' now take a closer look at tHe coefficient 'As of the leading term in thiS bound.
"Ve have As = A(Pl,'" ,Ps), where Pb'" JP& afe the mst s By the
for A(pl, ... ,Ps) and by prime r.nllnber theorem, we obtain
lim Iog'A.'! = L
s-+oo slogs
Thus As increases superexponentially as' s '00. Similarly, for Hammersley
point sets with an optimal choice of bases, we have the discrepancy bound (3.7),
where the coefficient of the leading term again increases superexponentially as
s 00. This fast growth of ,As (compare also with Table 4.4) makes the
bounds (3.6) and (3.7) practically useless for all but very small dimensions s.
For most applications, we need point sets and sequences satisfying discrepancy
bounds with' much smaller implied constants. Constructions of such point sets
and sequences will be described in this chapter.
In 4.1 we define point sets and sequences with a very regular distribution
behavior. These will be called (t, m, s)-nets and (t, s)-sequences, respectively.
On the basis of the strong properties enjoyed by these nets and (t, s)-sequences,
we can derive very good discrepancy bounds. The definitions of (t, m, s }-nets
and (t, s)-sequences have a certain combinatorial flavor, and some concrete con-
nections with classical combinatorial problems are explored in 4.2. General
principles for the construction of nets and (t, s}-sequences are presented in 4.3.
These principles are used in 4.4 and 4.5 for the construction of special fami-
lies of nets and (t, s)-sequences, respectively. By optimizing the parameters in
the construction of (t, s )-sequences in 4.5, we obtain sequences for any dimen-
sion s 2, which asymptotically have the smallest discrepancy that is currently
known.
4.1. Definitions and discrepancy bounds.
To motivate the following definitions, we focus on the following special property
of the van der Corput sequence XO,Xl, . in an arbitrary base b 2 (this
47
48 CHAPTER 4
sequence was introduced in Definition 3.2). For fixed integers k 2:: 0 and m 2:: 1,
consider the b
m
points Xn with kb
m
~ n < (k + l)b
m
We claim that every
b-adic interval [ab-
m
, (a + l)b-
m
), where a E Z and 0 ~ a < b
m
, contains
exactly one point Xn with kb
m
~ n < (k + 1 )b
m
. To prove this, note that, for
kb
m
~ n < (k + l)b
m
; the m least significant digits in the digit expansion of
n in base b can range freely, whereas the remaining leading digits are fixed; for
Xn = b(n) this means that its m leading digits after the "decimal point" can
range freely, whereas the remaining digits are fixed; so each b-adic interval of
length b-
m
contains exactly one of these X
n
.
For the subsequent definitions, we fix the dimension s 2:: 1 and an integer
b 2:: 2. A subinterval E of IS of the form
s
E = TI[ ai b-
d
., (ai + l)b-
d
,)
; ~ c l
with ai, ~ E Z, di 2:: 0, 0 ::; ai < b
d
; for 1 ~ i ::; s is called an elementary interval
in base b.
DEFINITION 4.1. Let 0 ::; t ::; m be integers. A (t, m, s )- net in base b is
a point set P of b
m
points in IS such that A(E; P) = b
t
for every elementary
interval E in base b with As (E) = b
t
-
m
.
DEFINITION 4.2. Let t 2:: Obe an integer. A sequence Xo, Xl, ... of points in
1
8
is a (t,s)-sequence in base b if, for all integers k 2:: 0 and m > t, the point set
consisting of the:xn with kb
m
::; n< (k + l)b
m
is a (t, m, s)-net in base b.
In t:Qis.Janguage, the van der Corput sequence in base b is a (0, I)-sequence
in base 'b. Definitions 4.1 and 4.2 :Were introduced by Sobol' [323] in the case
where b = 2jthe general definitions were first given by NiedeIT.eiter [244].
REMARK 4.3. If P is a (t,m,s)-net in base b, then A(EjP) = bmAs(E) for
eVery elementary interval E iii base b with As (E) = b
t
-
m
. More generally, it
follows that, if E is an elementary interval in base b with As (E) ~ b
t
-
m
, or a
disjoint Union of such intervals, then A(E; P) = b
m
As(E) (observe that such sets
E can be written as disjoint unions of elementary intervals in base b with As-
measure b
t
-
m
). This implies that any (t, m, s)-net in base b is also a(u, m, s)-net
in base b for integers t ~ u ::; m and that any (t, B)-sequence in base b is also
a (u, 13 )-sequence in base b for integers u 2:: t. Therefore it is clear that smaller
values of t mean stronger regularity properties.
We now establish upper bounds for the star discrepancy of nets. These
results are due to Niederreiter[244]. We use a standard convention for binomial
coefficients, namely, that (D = 0 for i > r or i < O.
LEMMA 4.4. Let P be a (1, m, s)-net in base b, let E be an elementary interval
in base -b with As(E) = b-
u
, where 0 $ u S m ~ t,and let T be an affine
traf1formation from E onto IS. Then the points of P that belong to E are
transformed by T into a (t, m -- u, s)-net 'in base b.
\
NETS AND (t,S)-SEQUENCES 49
Proof. It follows from Remark 4.3 that A(E; P) = b
m
-
u
. If the points of P
belonging to E are transformed by T, we obtain a point set pi of b
m
-
u
points in
[s. To prove that pI is a (t, m -U, 8 )-net in base h, we take an. elementary interval
E' in base b with As(E') = b
t
-
m
+
u
and note that for x E E we have T(x) E E' if
and only if x E T-l(E
'
). Now T-l(E') is an elementary interval in base b with
AB(T-l(E' = bt-mj hence A(T-
1
(E')jP) = b
t
, and so AtE';P') = b
t
0
THEOREM 4.5. The star discrepancy of a (t, m, a)-net P in base b ~ 3 satis-
fies
(4.1)
Proof. Write DCJ; P) = A(Jj P) -NA
B
(}) for an interval J ~ IS, where N is
the number of points of P, and let Ab(t,m, 8) denote the right-hand side of (4.1).
We fix t ~ 0 and proceed by double induction on 8 ~ 1 and m ~ t. First, let
8 '= 1 and consider an arbitrary m ~ t. If an. interval J = [0, ti), 0 < U S 1;
is given) we split it up into the disjoint intervals Jh = [hfl-
m
, + l)b
t
-
m
))
k ...... 0, 1, ... ,k - 1, and JIe = [kb
t
-
m
, u), where k = L ub
m
-
t
J. For a (t, m, I)-net
P'm base b, we have D(Jhi P) = 0 for 0 S h < kj hence D(J;P) = D(Jki P).
Now 0 S A(J:; P} < b
t
and 0 S b
m
Al(J:) S bt; thus ID(J; P)I S b
t
. It follows
that ND'N(P) S b
t
= Ab(t, m, 1), and so (4.1) is shown for 8 = 1.
Now let s ~ 2 be given' and suppose that (4.1) has been shown for the
dimension s -1 and all m ~ t. We verify (4.1) for the dimension 8 by induction
on m 2:: t. Ifm = t1 then it is trivial that ND'N(P) ~ N = b
t
= Llb(t,t,s).
Suppose that (4.1) has been establiShed for some m 2:: t and consider a
(t,m + l,s)-net P in base b. We must show
(4.2) ID(JjP)1 ~ Ab(t,m+1,s)
for evef'J interval J = n:=l [0, Ui) ~ IS. We distinguish cases depending on the
value of Us. If Us = 1, then we apply to P the projection T : [S -+ [s-1 defined
by
This transforms P into a (t, m + 1, s - I)-net PI in base b. Also D(J; P) =
D(T(J); PI), and so by induction hypothesis
(4.4) ID(J; P)I ~ Ab(t, m + 1, s - 1).
Hence (4.2) follows from the obvious inequality
.6.
b
(t, m + 1, s - 1) S .6.b(t, m + 1,8).
If Us < 1, then we introduce the integer l = LbusJ satisfying 0 ~ 1 S b - l.
Next, we consider the case where 0 ~ l S L b /2 J. 'We split up J into the disjoint
50 CHAPTER 4
intervals
for h = 0, 1, ... ,l - 1,
Then
z
(4.5) D(J;P) = LD(JhiP),
h=O
Put El = [0,1)8-1 X [lib, (l + I)/b) and let 11 be an affine transformation from
E, onto J8. By Lemma 4.4, 11 transforms the points of P that belong to EE into
'" m, P., b . . /,>.180, DCJI ; P) = D(Tl(J&); P2), and so, 'oy inductIon
hypothesis,
(4.6)
For 0 h < l, the projection T : J8 - /8-1 from (4.3) transforms the points of
P that belong to Eh = [0,1)8-1 X [h/b, (h + 1}/b) into a (t, m, 8 - I)-net pJh)
in base b. Also, D(Jh; P) = D(T(Jh); pJh}), and so, by induction hypothesis,
(4.7) jD(Jhi P)I Ab(t, m, S - 1) for 0 h < l.
Combining (4.5), (4.6), and (4.7), we obtain
(4.8) jD(J; P)! Ab(t, m, s) + lb/2JAb(t, m, 8 -1).
Therefore
JD(J;P)J b't. (8 1) (m; t) ur +b' (: = D (:"--n
(" 1) (m; t) (8 1) (7 --n
b' (8 1) (m +i
1
- t) = + 1,s),
and 80 (4.2) is established.
Finally, we consider the case where Lb/2J + 1 I b - 1. We view J as
the set-theoretic difference of the intervals L = n::; [ 0, Ui) X [0, 1) and M =
n::; [ 0, Ui) X [us, 1), and we split up M into the disjoint intervals
for h = l + 1, 1 + 2, . . . ,b - 1.
NETS AND (t,S)-SEQUENCES 51
Then we have
0-1
(4.9)
.
h='
By (4.4) we obtain
ID(Lj P)I :::; m + 1, s - 1).
The interval M, is treated like the interval J, in (4.6); hence
and, for l < h :::; b - 1, we obtain, as in (4.7),
ID(Mh; P)l :::; 6.b(t, Tn, oS -1).
Altogether, we obtain, from (4.9),
(4.10)
ID(Jj P)I :::; + 1, s -1) + m, s) + (b -2 m, s - 1).
Therefore
ID(J;P)I $ (" /) (m+/ -t) 1) (m;t) lH
+ WJ 2) (m ;t) lH
(8 2) (7_-1t) H' (: (7_-1t)
(8 1) (m; t)
= b
t
(8 1) (m +;1 - t) Llb(t, m + 1,s),
and so we again have (4.2). 0
THEOREM 4.6. The star discrepancy of a (t,m,s)-net P in an even base b
satisfies
(4.11)
NDj,(P) $ (m;t) G), + G (m-tt+ I) Gr
52 CHAPTER 4
Proof. Denote the right-hand side of (4,11) by b
i
E(m, 8), where we suppress
the dependence of E( m, s) on b and t for simplicity. If we follow the proof of
Theorem 4.5, then it is clear that it suffices to verify the analogue of (4.2). In
view of (4.8) and (4.10), this amounts to proving the inequalities
(4.12)
(4.13)
b
E(m, s) + "2E(m, s - 1) ::;; E(m + 1, s),
E{m+ 1,s-1) +E(m,s) + 2)E(m,S-1)::; E(m+ 1,8)
for B 2 and m 2 t. Now
E(m,8) + (m; ') + G)'
+ G -I) (m-
t
7i+ 1) G)' + G -1) 1) GY
= (m+jl- t) GY + G-l) (m -t7 i+ 2) GY = E(m+1,8),
and 00 (4.12) is established. Note tha.t the case corresponding to (4.13) in the
proof of Theorem 4.5, namely, Lb/2J + 1 ::; l ::; b - 1, does not occur for b = 2;
henceforth we can assume b 4. We write
(4.14) E(m,s) = F(m,s) +
with
$-1 (m _ t) (b)i
F(m, s) = i "2'
Now
NETS AND (t,S)-SEQUENCES 53
Furthermore,
G(m+ l,s -1) + G(m,s) + 2)G(m,s -1)
= G(m+ 1,8) _ (m .-2 + -1) B-2
+ G)' -
= G(m+ 1,8) - 2) Gr' - (m-it+i) G)'.
Because of (4.14), this yields
E(m + i, s -1) +E(m, s) + - 2 )E(m, 8 -1) = E(m + 1, s)
+ - (m;t)) W' - G -1)
_ (m -.t+i) i.
2 i=O 1. 2/
To prove (4.13), it remains to show that
For 8 = 2 and 8 = 3, this is a simple verification. For 8 ;?: 4, we prove that,
for 0 i 8 - 2, the coefficient of (b/2)i on the left-hand side does not exceed
the corresponding coefficient on the right-hand side of (4.15). This is trivial for
i = O. For 1 i 8 - 3, we have
(
m-t) _ (m-t) < (m-t) < fm-t+i-1) < (m-t+i)
i-I i - i-I - \. i-I - i

For i = /3 - 2, we must prove that
(
m -t) _ (m -t) (m -t + s - 2\) .. _ 1).
8-3 8-2 8-4 2
54 CHAPTER 4
This is trivial for m = t, and, for m t + 1, we obtain
(
m -t) _ (m -t) = (m -t - 1) _ (m -t - 1) :::; (m -t -1)
8-3 8-2 \ s-4 8-2 s-4
:::;
so that (4.15) is shown in all cases. 0
If the bounds in (4.1) and (4.11) are expressed in terms of powers of m - t
with m > t, then they can be written in the form
where the implied constant depends only on b and 8, In terIDB of the number
N = bID of points in a (t, m, s)-net in base b, this says that for m > 0 we have
ND* (P) :::; b
t
(lb/2J )S-l (logN)S-l + O(bt(logN)S-2)
N (s -I)! 10gb
with an implied constant depending only on band 8.
In the cases where s = 2,3, 4, there is a different method of proof, which yields
results that are sometimes better than Theorems 4.5 and 4.6. We illustrate this
method in the case where s = 2, and we only state the resulting discrepancy
bounds for s = 3,4 (see Niederreiter [244] for the complete proofs in the latter
cases).
THEOREM 4.7. For s = 2, the star discrepancy of a (t,m,s)-net P in base b
satisfies
ND'N(P):::;
Proof. Write D(J;P) = A(JiP) -'- b
m
).2(J) for an interval J 1
2
, Now let
J = [O,u) x [O,v) [2. Write u = Ej:l ujb-
j
'!}rith Uj E Zb for all j 1,
and put r = m - t and d = 2:;=1 Uj. Then [0,2:;=1 ujb-
j
) can be represented
as the disjoint union of Ul elementary intervals in base b of length b-
1
, of U2
elementary intervals in base b of length b-
2
, and so on, hence of done-dimensional
elementary intervals El, ... ,Ed in base b. With F = [I:;=1 ujb-j,u), we then
obtain
d
(4.16) ID(J;P)I:; L ID(Eh x [O,v)jP)1 + ID(F x [O,v)jP)I
h=l
Now let K be an arbitrary two-dimensional interval that is contained in some
tW<rdimensional elementary interval in base b of area b-
r
Then 0 ::; A(K; P) :::;
b
t
and 0 :::; b
m
).2(K) :::; b
t
; thus
( 4.17)
NETS AND (t,S)-SEQUENCES 55
Next, we claim that, for anyone-dimensional elementary interval E in base b,
we have
(4.18) ID(E x [O,v)i P)I5: bt.
If 'xl (E) b-", then (4.18) follows from (4.17). Otherwise, we have 'xl (E) = b-
a
for some a E Z with 0 5: a < r. In this we split up [0, v) into one-
dimensional elementary intervals in base b of length ba.-r and a remaining interval
Fl of length < b
a
-
r
. From Remark 4.3, it follows that DC E x [0, v); P) =
D(E x FI ; P). However, Ex FI is contained in a two-dimensional elementary
interval in base b of area b-", and so (4.18) is implied by (4.17).
Note that F x [0, v) is contained in a two-dimensional elementary interval in
base b of area b-". Thus it follows from (4.16), (4.17), and (4.18) that
(4.19) ID(J; P)I :S (d+ l)b
t
,
J Now consider L -:- [u,l) x [O,v). By, treating the interval [u,l) in the same
way as [0, u) at the beginning of the proof, we see .that [u, 1) can be represented
as the disjoint union of (b-l)r - d one-dimensional elementary intervals in base
b and an interval that is contained in a elementary interval in,
base b of length b-
r
. Hence, by the argument leading to (4.19), we obtain
/D(L; P)I :S (b - 1)r - d + l)b
t
,
FUrthermore,
D(J; P) = D([O, 1) x [0, v); P) - D(L; P),
and, together with (4.18), this yields
(4,20) ID(J; P)I :S ((b - 1)r - d + 2)b
t
.
It follows that /D(Jj P)I is bounded by the minimum of the right-hand sides
of (4.19) and (4.20). Maximizing over 0 :S d:S (b - 1)r, we obtain the result of
the theorem. 0
THEOREM 4.8. For s = 3, the star discrepancy of a (t, m, s) -net P in base b
satisfies
* b-l 2 b-l 9 t
l(
)
2 J
NDN(P):S -2- (m-t) +-2-(m-t)+4 b.
THEOREM 4.9. For s = 4, the staT discrepancy of a (t, m, s) -net P in base b
satisfies
ND'N(P) < l(b -1)3 (m - t)3 + - t)2 + -l)(m - t) + 15J b
t
,
- 2 8 8 4
We combine these discrepancy bounds in such a way that, in the asymptotic
form, the coefficient of the leading term is as small as possible. This yields the
following result.
56 CHAPTER 4
THEOREM 4.10. The star discrepancy of a (t, m, s)-net P in base b 'l.J.Jith
m > 0 satisfies
NDiv(P) B(s,b)b
t
(logN)8-1 + O(b
t
(logN)8-2) ,
where the implied constant depends only on band s. Here
(
b -1 )8-1
B(s, b) = 210gb
if either s = 2 or b = 2, s = 3,4; otherwise
1 (lb/2J)B-I
B(s, b) = (s _ 1)1 log b
The fact that 1;hesB uppal' tn.t) 3a,{ diSClep&-:iCY of Iiets are lncrews-
ing functions of t is in accordance with an earlier observation on the values of t in
Remark 4.3. We now consider the problem of bounding the star discrepancy of
(t, s}-sequences. In the following general lemma, we let Llb(t, m, s) be a number
for which NDiv(P) Llb(t,m,s) holds for any (t,m,s)-net P in base b.
LEMMA 4.11. For the star discrepancy Div(S) of the first N terms of a (t, s)-
sequence S in base b, we have
k
ND'N(S) b; 1 L .6.b(t,m,s) + 1,s) +
m=t
for N b
t
, where k is the largest integer 'l.J.Jith ble N, where b
r
is the largest
power of b dividing N, and where we set Llb(t, T, s) = 0 if T < t.
Proof. Let Xo, Xl, ... be a (t,s)-sequence in base b. For N b
t
let
Ie
N= Lamb
m
m=O
be the digit expansion of N in base b, so that all am E Zb, ale =I- 0, and k t. We
split up the point set Xo, Xl, ... , XN-I mto the point sets Pm, 0 m k, defined
as follows:, Let PI!: consist of the Xn with 0 n < akbk, and, for 0 m k - 1,
let Pm consist of the Xn with ahbh n < ahb
h
. The Pm with
am =1= 0 are nonempty, and, by the definition of a (t, s)-sequence in base b, they
can be split up into am (t, m, s)-nets in base b provided. that m t. Therefore
Ie t-l
( 4.21) ND'N(S) L amLlb(t,m,S) + L amb
m
.
m=t
Now we apply the same method to the point set consisting of the Xn with N
n < This point set has
Ie
b
le
+
1
- N = L emb
m
m=O
NETS AND (t,S)-SEQUENCES 57
points, where the right-hand side is the digit expansion in base b. Also, using
the fact that the point set consisting of the len with 0 :5 n < b
k
+
1
forms a.
k + 1,8 )-net in base b, we obtain
k t-l
(4.22) ND'N(S) :5 Ab(t, k + 8) + L emA,,(t, m, s) + L emb
m

, 'm='t m=O
Adding (4.21) and (4.22) and dividing by 2, see that
1 k 1
ND'N(S) :5 2" L(am + em)Ab(t,m, 8) + '2Ab(t, k + 1,8)
m=t
(4.23)
1 t-l
+ 2" L(arn + em)b
ffl
,
m=O
From the definition of r, it follows that am = em = 0 for 0 :5 m < r, ar + c,. = b,
, and am + em = b - 1 for't < m :5 k. If r -:; t - 1, (4.23) yields
b-l k 1
ND;"(S) :5 -2- L Ab(t, m, 8) + 2" Ab(t, k + 1, s)
m=t
+ (br+l + t-l (b _ l)b
m
)
m=r+l
b-l k lIt
= -2- L Ab(t, m, 8) + 2"Ab(t, k + 1, s) + 2"b ,
m=t
which is the result of the lemma. If r t, then (4.23) yields
... b b-l Ie 1
ND
N
(S):52"Ab(t,r,s)+-2- L Ab(t,m,s)+'2 Ab(t,k+l,s)
m=r+l
b-l k 1 1
S -2- L Ab(t, m, s) + 2"Ab(t, r, s) + 2Ab(t, k + 1,8),
m=t
and the result of the lemma follows. 0
THEOREM 4.12. The star discrepancy D'N(S) of the first N terms of a (t, 8)-
seq'uence S in base b 3 satisfies
NDN(S) <;, b; 1
b
,t, G = D (H -t)
+ 1) -t) + lH
for N b
t
J where k is the largest integer with b
k
S N.
58 CHAPTER 4
Proof. By Theorem 4.5, we can use Lemma. 4.11 with
t 8-1 m-t b
8-1 ( ) ( ) l J'
Ab(t, m, s) = b ~ i i 2'
For N ~ b
t
, we have k ~ t, and it is clear that the integer r in Lemma 4.11
satisfies r ~ k. We then obtain
An easy induction on M yields
and this implies the desired result. 0
THEOREM 4.13. The star discrepancy Div(S) of the first N terms of a (t, s)-
sequence S in an even base b satisfies
for N ~ b
t
, where k is the largest integer with b
k
~ N.
Proof. By Theorem 4.6, we can use Lemma. 4.11 with
Proceeding as in the proof of Theorem 4.12, we obtain
NETS AND (t,S)-SEQUENCES
for N? b
t
. By (4.24) we have
and
k-t (m + i + 1) = k+i+l-t (m) k+i+l-t (m) = (k + i + 2 -t)
L, < "'. '+1'
'l. 'l. L- 'l. 'l.
m=O m=t+l m=O
and this implies the desired result. 0
59
H the bounds in Theorems 4.12 and 4.13 are expressed in terms of powers of
k - t with k > t, then they can be written in the form
:,b
tb
; -t)1I + O(bt(k - t)8-1),
<where the implied constant depends only on b and 8. Since k (log N) / log b,
this yields
.N D* (8) b - 1 (L b/2 J ) 1/ (log N)B + o (b
t
(log N)8-1) for N? 2,
.. N 8! 2 Lb/2J 10gb
with an implied constant depending only on b and 8.
In the cases where s = 2, 3, 4, we obtain the following additional bounds
by using Theorems 4.7, 4.8, and 4.9, together with Lemma 4.11 (see Nieder-
reiter [244] for details).
THEOREM 4.14. For s = 2, the star discrepancy Div(S) of the first N terms
of a (t, s) -sequence 8 in base b satisfies
i(b - 1)2b
t
(k - t)2 + l(b - l)(b + 9)b
t
(k - t) + !(b + l)b
t
for N ? b
t
, where k is the largest integer with b
k
N.
THEOREM 4.15. For s = 3, the star discrepancy Div(8) of the first N terms
of a (t, s)-sequence S in base b satisfies
NDiv(S) (b -1)3b
t
(k - t)3 + (b - 1)2(b+ 5)b
t
(k - t)2
+ }s (b - 1)(b
2
+ 16b + 61)b
t
(k - t) + l(b
2
+ 4b + 13)b
t
for N ? b
t
, where k is the largest integer with b
k
N.
THEOREM 4.16. For s = 4, the star discrepancy D'N(8) of the first N terms
of a (t, s)-sequence 8 in base b satisfies
N Div(S) (b - 1)4b
t
(k - t)4 + i2 (b - 1)3(b + 5)b
t
(k - t)3
+ - 1)2(b
2
+ 16b + 13)b
t
(k - t)2
+ i2(b - 1)(7b
2
+ b + 64)b
t
(k - t) + 116 (b
3
+ 8b + 51)b
t
for N ? b
t
, where k is the largest integer with b
k
N.
60 CHAPTER 4
We combine these discrepancy bounds for (t, s )-sequences in such a way that,
in the asymptotic form, the coefficient of the leading term is as small as possible.
This leads to the following theorem.
THEOREM 4.17. The star discrepancy Dj.,(S) of the first N terms of a (t, 8)-
sequence S in base b satisfies
where the implied constant depends only on band 8. Here
1(b-1)B
C (s, b) = S 210gb
if either s = 2 or b = 2, s = 3,4; otherwise
1 b-l (Lh/2J) 8
C(s,b) = s! . 2lb/2J 10gb .
4.2. Combinatorial connections.
The definitions and results in 4.1 show that the (t, m, s)-nets and (t, s)-sequences
with the most regular distribution behavior are those with t = O. The main
results of this section demonstrate that, except in trivial cases, (0, m, s )-nets in
a given base b can only exist if s does not exceed a certain function of h, and
a similar condition on SIDust hold so that a (0, s)-sequence in base b can exist.
The principal tool is a connection between nets and the combinatorial theory of
orthogonal latin squares.
In the cases where m = 0,1, it is trivial to construct a (0, m, s)-net in base b
for any s 1 and b 2. For m = 0 take one arbitrary point from 1
B
, and for
m = 1 take the point set consisting of (n/b, ... ,nib) E JB, n = 0,1, ... ,b - 1.
Thus we can hen.ceforth assiune that m 2.
We recall the following standard concepts from combinatorics (compare
with Hall [124,. Chap. 13]). an integer b 2, two b
2
-tuples e =
(e(0),e(I), ... ,e(b
2
-1)) and f = (f(0),f(1), ... ,f(b
2
-1)) with entries from
the same set of cardinality b are caIled orthogonal if the b
2
ordered pairs
(e(h),f(h)), h = Oj1, ... ,b
2
-1, are all distinct. The b
2
-tuples ell ... ,e
B
with
entries from the same set of cardinality b are called mutuaUy orthogonal if ei and
ej are orthogonal for all 1 i < j s. Equivalently, the entries of a b
2
-tuple
may be arranged in a prescribed manner in a square array with b rows and b
columns, and, with such an identification of b
2
-tuples and b x b arrays, we speak
of orthogonal squares of order b and mutually orthogonal squares of order b. A
b x b array is called a latin square of order b if each row and each column is a per-
mutation of the same set of cardinality b. If M (b) is the maximum cardinality of
a set of mutually orthogonal latin squares of order b, then we have M(b) b-l
for all 2 (see [59, p. 158], [299, BO]).
THEOREM 4.1B. There exists a (0,2, s )-net in base b if and only if there exist
s - 2 mutually squares of order b.
i.
NETS AND (t,S)-SEQUENCES 61
Proof. For 8 = 1, both conditions are trivially satisfied, so we assume that
8 2. Let Xn = ... ,x!: E ]'1, n = 0,1, ... ,b
2
- 1, be the points
of a (0,2, a)-net in base b. For 1 SiS 8, we define a b
2
-tuple ei =
(e,(O),ei(l), ... -1 with entries in Zb by setting
We will prove that 181; . ,e
s
are mutually orthogonal. If 1 S i < j S s are
given, then and ej will be shown to be orthogonal once we prove that, for
any c,d E Zb, we have (ei(h),ej(h = (c,d) for some h. Now E = n:=1 E.,.
with Ei = [c/b,(c+ 1)/b), Ej = [d/b,(d+ 1)/b), and E.,. = [0,1) for r =f: i,j
is an elementary interval in base b with As(E) = b-
2
; hence it contains a point
Jell. From Xh E E, it follows that ei(h) = c and ej(h) = d. By [124, pp. 222-
223] there exist 8 mutually orthogonal b
2
-tuples if and only if there exist 8 - 2
w,utually orthogonal latin squares of order b.
,1 Conversely, if there exist s - 2 mutually orthogonal latin squares of order h,
then there exist s mutually orthogonal b
2
-tuples ei = (ei(O), ei(l), . .. j ei(b
2
-1)),
1 :5 i S 8, with entries in Zb. Ii we rix i, then from the orthogonality of ei with
one of the other Sj, j =1= i, it follows that each v E Zb occurs exactly b times as an
. entry of eli. Thus there exists a map 1/Ji : Zb2 -+ Zb with the property that, for
each v E Z", the restriction of tPi to the set {n E Zb2 : ei (n) = v} is a bijection.
For n E Zb2, we define
We will show that the point set P consistingQf the Xn, n = 0,1, ... ,b
2
- 1,
is a (0,2, s)-net in base b. If E = n:=l E.,. is an elementary interval in base
b with As(E) = b-
2
, then we have Er = [0,1) for only one or for only two
subscripts r. IT Er =f: [0,1) for only say for r = i, then Ei =
(e + 1)/b
2
) for some e E Zb2 and E.,. = [0,1) if r =1= i. Thus Xn E E if and
only if ei(n)b + 'lfJi(n) = e. If we write e = vb + w with v, w E Zb, then Xn E E
if and only if ei ( n) = v and 'l/Ji (n) = w. From the definition of 'l/Ji it follows that
there is exactly one n E ZlfA satisfying these conditions, hence A(E; P) = 1. If
Br =J. [0, 1) for two subscripts, say for r = i and r = j with 1 ::::; i < j S s, then
with suitable c, dE Zb we have Ei = [c/b, (c + 1)/b), Ej = [d/b, (d + 1)/b), and
E.,. = [0,1) if r =1= i,j. Thus Xn E E if and only if ei(n) = c and ej(n) = d. Since
ei and ej are orthogonal, we have (ei (n), ej (n)) = (c, d) for exactly one n E Zb2,
hence A( Ej P) = 1. 0
REMARK 4.19. Since the e:ldstence of b-1 mutually orthogonal latin squares
of order b is equivalent to the existence of a finite projective plane of order b
(see [124, pp. 209-210]), it follows from Theorem 4.18 that there exists a finite
projective plane of order b if and only if there exists a (0,2, b + 1}-net in base b.
Thus the most celebrated problem in finite geometry is equivalent to a problem
on the existence of certain nets.
62 CHAPTER 4
THEOREM 4.20. For m 2:: 2, a (0, m, B)-net in base b can only exist if B :::;
M(b) + 2.
Proof. If there exists a (0, m, s )-:net in base b for some m 2:: 2, then, by
Lemma 4.4, there exists a (0,2, s )-net in base b. From Theorem 4.18, it follows
that there exist s - 2 mutually orthogonal latin squares of order bj therefore
S - 2 $ M(b). 0
COROLLARY 4.21. For m 2:: 2, a (0, m, B)-net in base b can only exist if
s:::;b+1.
Proof. This follows from Theorem 4.20 and the inequality M (b) :::; b - 1. 0
Analogous necessary conditions hold for the existence of (0, s )-sequences in
base b. The link with the above results is established by the following lemma.
LEMMA 4.22. If there exists a (t, s)-sequence in base h, then, for evenJ m 2:: t,
the'l'E exists a (t, m, S + 1 )-net in baBe b.
Proof. Let Xo, Xl!'" be a (t, s)-sequence in base b. We will prove that, for
fixed m 2:: t, the point set P consisting of (nb-
m
, XllJ E Is+1, n = 0,1, ... ,bm-I,
is a (t, m, s + I)-net in base b. Let
8+1
E = II [ai b-
di
, (ai + l)b-
di
)
i=l
be an (s + I)-dimensional elementary interval in base b with A
8
+1(E) = bt-mj
hence d
i
= m - t. We have (nb-m,x
n
) E E if and only if a10m-d! s: n <
(al + 1)b
m
-
d1
and
8+1
Xn E E' := TI[ aib-di, (ai + l)b-
d
,).
i=2
Since m - d1 .2:: t and the Xn form a (t, s in base b, the point set
pI consisting of the Xn with al b
m
-
d1
:::; n < (al + 1 )b
m
-
d1
is a (t, m - d
1
,
net in base b. Now E' is an s-dimensional elementary interval in base b with
As(E') = b
t
-
m
+
d1
j hence A(E'jP') = b
t
, and so A{E;P) = bi. 0
THEOREM 4.23. A (0, s)-sequence in base b can only exist if s :::; M(b) + 1.
Proof. This foHows from Theorem 4.20 and Lemma 4.22. 0
COROLLARY 4.24. A (0, e)-sequence in base b can only exist if s ::; b.
Proof This follows from Theorem 4.23 and the inequality M(b) :::; b-l. 0
DEFINITION 4.25. For given integers b 2:: 2 and s 2:: I, let tb(S) be the least
value of t for which there exists a (t, s )-sequence in base b.
Theorem 4.23 shows that tb(e) 2:: 1 for 82: M(b) +2. Note that M(b) = b-l
for every prime power h, since b - 1 mutually orthogonal latin squares of order
b can rfe constructed by using the existence of a finite field with b elements (see
[192, Thm. 9.83]). For b = 6, we have M(b) = 1 by the negative solution of
Euler's officer problem (see [299, pp. 84-85]).
i.
NETS AND (t,S)-SEQUENCES
63
4.3. General construction principles.
We now present general principles for the construction of nets and (t, s )-sequences
in an arbitrary base. The special constructions of nets and (t, 8 )-sequences to be
discussed in 4.4 and 4.5, respectively, will be based on these principles.
First, we introduce a general principle for the construction of (t, m, s)-nets in
base b. Let the integers m 1, S 1, and b;::: 2 be given. As before, we write
Zb = {O, 1, ... ,b - I}. We choose the following:
(Nl) A commutative ring R with identity and card(R) = h;
(N2) Bijections 'l/Jr : Zb -l> R for 0 :::; r :::; m - 1;
(N3) Bijections "Iij : R -l> Zb for 1 :::; i :::; s and 1 :::; j :::; mj
(N4) Elements E R for 1 :::; i :::; s, 1 :::; j :::; m, and 0 :::; r :::; m - 1.
For n = 0,1; ... , b
m
- 1, let
n = 2: ar(n)b
r
with all ar(n) E Zb
1"=0
be the digit expansion of n in base b. We put
m
= for 0 ::s: n < b
m
and 1:::; i :::; s,
j=l
with
for 0 :::; n < b
m
, ;t :::; i :::; 5, 1:::; j :::; m,
and define the point set
(4.25)
THEOREM 4.26. Suppose that the integer t with 0 :::; t :::; m satisfies the
following property: For any integers d
11
.. ,d
s
;::: 0 with 2::=1 di = m - t and
any fY) E R, 1 :::; j :::; d
i1
1 :::; i :::; 8, the system of m - t linear equations
m-l
L Zr = fY) for 1 :::; j :::; di1 1 5: i :::; s,
r=O
in the unknowns zo, ... , Zm-l over R has exactly ht solutions. Then the point
set (4.25) is a (t, m, s )-net in base b.
Proof Let
s
E = II[aib-di,(ai + l)b-
di
)
i=l
64 CHAPTER 4
be an elementary interval in base b with >.-s(E) = bt-mj hence 2::=1 d
i
= m - t.
For 1 :::;; i :::;; 8, let
do
ai = L aijbd.,-j
j=l
be the digit expansion in base b, where all aij E Zb. For the points Xn of the
point set P in (4.25), we have Xn E E if and only if
for 1 :S i :::;; 8.
This is equivalent to
1<j<:::; 1 :S i S;
which is, in turn, equivalent to
m-l
L = TJi/(aij) for 1 :::;; j :::;; di , 1:::;; i :::;; 8,
1'=0
where TJi/ denotes the inverse map of 'fJij' By hypothesis, this system of linear
equations in the unknowns 'r(ar(n)), 0 :::;; r :S m - 1, over R has exactly b
t
solutions. Each solution corresponds to a unique m-tuple (ao(n), . .. ,am-l (n)),
and each such m-tuple corresponds to a unique integer n with 0 :::;; n < b
m
.
Therefore A(E; P) = b
t
. 0
We frequently consider the special case where b is a prime power. In this case,
we write q instead of b, and we can take for the ring R the finite field Fq with q
elements. -VVe collect the elements E Fq in (N4) into a system of vectors
(4.26)
(i) _ ( (i) (i)) F m
Cj - CjO "'" Cj,m_l E q
for 1 :::;; i :::;; s, 1:S j :S m,
and we introduce the following general notion.
DEFINITION 4.27. For a system C = {c)i) E V : 1 :::;; i :::;; s, 1 :S j :::;; m}
of vectors in a finite-dimensional vector space V, let p( G) be the largest integer
d such that any system {C)i) : 1:S j:::;; d
i
, 1:::;; i:::;; s} with 0:::;; d
i
:::;; m for
1 :::;; i :::;; s and E:=l d
i
= d is linearly independent in V (here the empty system
is viewed as linearly independent).
It is clear that we always have 0 :::;; p(G) :::;; dim(V). If we specialize G to
be as in (4.26), which is the only case of interest in this chapter, then V is the
vector space F;n over Fq; and so 0:::;; peG) :::;; m. The significance of the number
p(G) is apparent from the fonowing result.
THEOREM 4.28. If q is a prime power, R = Fq , and the system G of vectors
is given by (4.26), then the point set (4.25) is a (t,m,s)-net in base q with
t = m - p(G).
NETS AND (t,S)-SEQUENCES 65
Proof. It suffices to check the condition in Theorem 4.26 for t = m - peG).
Consider integers db'" ,ds 0 with 2::=1 di = peG), Then, by Definition 4.27,
the vectors cJ
i
), 1 :::; j :::; d
i
, 1 :::; i :::; s, are linearly independent over F
q
Thus
the coefficient matrix: of the system of p( G) linear equations in Theorem 4.26 has
rank peG), and so the system always has exactly qm-p(C) = q't solutions. 0
The principle in Theorem 4.28 can be extended to an arbitrary base b 2 by
using the following procedure. We can write b as a product of prime powers, say
b = n:=l qv' Let R be the ring direct product R = Fqt) of the finite fields
F
qv
' 1 :::; v :::; h. Then the ring R sa.tisfies the properties in (Nl). Note that all
operations in R are performed coordina.tewise. The elements in (N4) are of
the form
(4.27)
where E Fq". For fixed v, let Gv be the system of vectors
_ (c(i) Ji) ) r F.! m
Ljv - jOv' ,0j,m-l,v c: q"
for 1 i 5, 1 j
THEOREM 4.29. If b = n:=l qy is a product of prime powers qVI if R =
Fq", and if the E R are given by (4.27), then the point set (4.25) is a
(t,m,s)-net in base b with
where the system G
v
of vectors is given by (4.28).
Proof. We check the condition in Theorem 4.26 for t = m - minl<v<h p(Cv ).
For integers d
1
, .. ,d
8
0 with 2::=1 d
i
= m - t and elements - -
(i) (i) (i)
f
j
= (fjl , ... ,f
jh
) E R for 1 :::; j :::; di , 1:::; i :::; 8,
consider the system
m-l
(4.29) L Zr = fP) for 1 S:. j :::; d
i
, 1 S:. i :::; 8,
1'=0
in the unknowns Zo, ... ,Zm-l over R. Because of the direct product structure of
R, this is equivalent to considering, for 1 :::; v :s; h, the system of m - t equations
m-l
(4.30)
(i) _ (i)
L..t Cj1'v Zrv - Ijv E Fq"
1'=0
in the m unknowns ZOv,'" ,Zm-l,v over Fq". From p(Cv ) 2: m-t, we obtain that
the coefficient matrix of the system (4.30) has rank m - t for each v. Thus the
system (4.30) has exactly q; solutions for each v. Consequently, the system (4.29)
has exactly q! = b
t
solutions. 0
66 CHAPTER 4
The values of t obtained from Theorems 4.28 and 4.29 can be used in the
general discrepancy bound for (t, m, s )-nets in Theorem 4.10. For simplicity, we
state the resulting bound only for a prime power base q. In this case, the point
set P in (4.25) satisfies
(4.31)
where the implied constant depends only on q and s. We now show that there
is also a lower bound for Dj.,{P) in terms of peG).
THEOREM 4.30. If q is a prime power and R = F
q
, then, for the point set P
in (4.25), we have
> (! > ___ L 'n-P(C),
"" ,- 3 3q}"-
If, in addition, the bijections 'TJij are such that 'TJij(O) = 0 for 1 :::; i :::; sand
1 :::; j m, then
D'N-{P) - 21q) q-p(C).
Proof. By construction, we have
m-l
(4.32) L =
1'=0
for 0 n < qm, 1 i s, 1 j :::; m.
If s = 1 and p( G) = m, then the lower bounds for D'N (P) follow from The-
orem 2.6. Otherwise, the definition of p(C) implies the existence of integers
dl? ,dB with 0 di :::; m for 1 :::; i :::; s and L::=1 di = ,0(0) + 1 such that the
system {C)i) E F:: : 1 :::; j :::; di, 1 :::; i s} is linearly dependent over F
q
If w
with 1 :::; w :::; s is the largest index for which d
m
=1= 0, then there exist elements
h)i) E F
q
, 1 :::; j :::; d
i
, 1 i w, not all zero such that
w d.
LL:h)i)c)i) = 0 E F;:.
i=l j=l
We have ht) =1= 0 by the definition of peG). By comparing components, we
obtain
NETS AND (t,S)-SEQUENCES
Together with (4.32), this yields
w c4 w d. m-1
(4.33) L: L = L L hJi) L
i==1 ;=1 ;=1 j=1 r=O
m-l w d.
= L 1Pr(ar (n = 0
r=O i=l ;=1
Since ht) =F 0, there exists a unique C E Fq with
w-l d;, d",-l
(4.34) L L hJi} 1Jij
1
(0) + L + ht}c = O.
i=1 ;=1 ;=1
Put a = 7)wd,JC) E Zq. Define the intervals
Ji=[O,q-d
i
) for
J
w
= {. [(a + 1)q-d
w
, ql-d
w
) if a < (q - 1)/3,
, [O,aq-d
w
) if a;::: (q -1)/3.
The subinterval J of ]B is then defined by
w
J = 11 J
i
X J8-w.
i=l
67
We claim that A(J; P) = 0. Suppose, on the contrary, that Xn E J for some
n with 0 n < qm. Then E J
i
for 1 i ::; w. For 1 ::; i < w, it follows from
the definition of J
i
that = 0 for 1 ::; j < d
i
. For i = w, it follows from the
definition of J
w
that = 0 for 1 ::; j < dw and =F a. Thus (4.33) implies
that
w-l di d",-l
"" """ h(i) 'TI"':: 1 (0) + '" + h(w)1J-
1
(y(w) = O.
L...t 3 '1'3 3 WJ dw wdw ndw
i=1 ;=1 ;=1
In view of (4.34), this yields = Ci hence = 'f/wd
w
( c) = a, a
contradiction. Thus the claim is shown.
Now we consider two cases, as in the definition of J
w
' In the first case, let
a < (q - 1)/3. Define subintervals II and h of IS by
w-l
h = II J
i
x [O,ql-d
w
) X I
s
-
w
,
i=l
w-l
12 = IT J
i
X [0, (a + 1)q-d
w
) x I
S
-
w
.
i=l
68 CHAPTER 4
Then 11 is the disjoint union of J and 1
2
, Since A(Jj P) = 0, we obtain
Using 2:7=1 di = peG) + 1, we see that
1 1 q - a-I () q - 1 (0)
Div(P) ~ '2As(J) = '2(q - a - 1)q-d
1
--d
w
= 2q q-P 0 > --aqq-P .
In the second case, let a ~ (q -1)/3. Then, from A(JjP) = 0, we obtain
Thus in both cases we have the first inequality in the theorem.
If, in addition, we have l1ij(O) = 0 for 1 ::; i ::; s and 1 ::; j ::; m, then
from (4.34) we get c = OJ hence a = O. Thus,. from the first case above, we
obtain
D* (P) > q - a - 1
q
_p(0) = q -1
q
_ p(0}
N - 2q 2q'
which is the second inequality in the theorem. 0
REMARK 4.31. Since the proof of Theorem 4.30 is based on the construction
of an interval J containing none of the points Xn in (4.25), it follows that the
lower bounds in Theorem 4.30 also hold for any point set consisting of the Xn
with n running through an arbitrary nonempty subset of {O, 1, ... ,qm -1}.
If q is a prime, then Fq and Zq can be identified. There is also a canonical
way of identifying elements of Fq and C( q) . Let C = {C)i.) E F;: : 1 ::; i ::; s, 1 ::;
j ::; m} be the system of vectors in (4.26). Using the quantities Wq(H) defined
in (3.18), we set
(4.35)
Rq(C) = 2: Wq{H).
H
Here the sum runs over aU nonzero H = (h
ij
) E C(q)8Xm with
where the h
ij
are viewed as elements of F
q

LEMMA 4.32. If q is a prime, R = Fq; and eveTY "Iij is the identity map,
then t h e . ~ t a r discrepancy of the point set P in (4.25) satisfies
NETS AND (t,S)-SEQUENCES
Proof By the assumption on the "Iii' we have
m-l
(i) _ ""' (i)
Ynj - L...J Cj'1' 'r(ar(n
for 0:5 n < qm, 1:5 i :5 8, 1 :5 j :5 m.
'1'=0
A:n a.pplication of Theorem 3.12 yields
(4.36)
(
1)8 1 N-l (18m )
DN(P) $1- , 1- N + L Wq(H) N L e - '
\ HO n=O q i=l j=l
69
where the outer sum is over all nonzero H = (h
ij
) E C(q)8xm. For fixed H, we
have
= II 2:
e
- .
ffl-1(q-l (b 8 m ))
1"=0 b=O q i=l j=1
;; The last expression is equal to qm = N if
for 0 :5 r :5 m - 1,
i=1 ;=1
and equal to zero otherwise, where the h ij are viewed as elements of F
q
. The
lemma now follows from (4.36) and the definition of Rq(C) in (4.35). 0
On the basis of Lemma 4.32, we can show that the construction of the point
sets (4.25) yields on the average a low-discrepancy point set.
THEOREM 4.33. For a prime q and for integers m 1 and s 1, let
1
Mq(m, 8) = card(C) Rq(C)
be the mean value of Rq (C) extended over the set C of all choices for a system
C = {cji) E F;": 1:::; i:::; S, 1 $j :::;m}. Then, with N= qffl, we have
1 (logN )8 1
Mq (m, s) = N log 4 + 1 N if q = 2,
1 (m ""' nlhl m - 1) S 1
Mq(m,s) = N.- L..t csc-+m--
q hEC*(q) q q N
<..!.. + _7 ___ 1_) logN + !)S N
l
if q > 2.
N n 5 log q q log q q
70 CHAPTER 4
Proof. Inserting the definition of Rq(G) into the expression for Mq(m, s) and
interchanging the order of summation, we obtain
1
Mq(m, s) = card(C) ~ Wq(H) ~ 1,
where the outer sum is over all nonzero H = (hij ) E C(q)sxm and the inner sum
is over all C = (c)i E C for which
(4.37)
s m
'" ""' h (i) 0 F
m
L- L., ijC
j
= E q.
i=l j=l
For a fixed nonzero H E G(q)BXm, the inner sum in the last expression for
iI.J
q
(m,3) represents the number solutions (c)tJ) E C of the vector equa-
tion (4.37). Since at least one hij is a nonzero element of F
q
, we can choose
ms - 1 vectors c)i) E F:: arbitrarily, and the remaining vector is then uniquely
determined by (4.37). Therefore the number of solutions of (4.37) is q(ms-l)m.
Since card ( C) = qm
2
s, it follows that
Mq(m,8) = q-m
2
8 q(ms-l)m L Wq{H) = ~ L Wq(H).
H#O H,eO
H q = 2, then the second part of Lemma 3.13 yields
~ (m)S (lOgN)8
f::o Wq (H) = 2" + 1 - 1 = log 4 + 1 - 1,
and the formula for Mq (m, 8) follows. For q > 2, the desired result is obtained
from the first part of Lemma 3.13. 0 c,
If we combine Lemma 4.32 and Theorem 4.33, then we see that, if q is a
prime, R = Fq , and every rli'; is the identity map, and if m 2: f and s 2: 1 are
fixed, then the construction of the point sets (4.25) yields on the average a point
set P with Div(P) = O(N-l(logN)s). We now show that the quantities p(G) in
Definition 4.27 and R,.;( G) in (4.35) are connected by the following inequalities.
THEOREM 4.34. For s 2: 2 and any prime q, we have
q-p(C)-l :$ Rq(C) :$ (1- t)k(q)8 ((m + lr _ ( p ( C ~ + s) )q-P(C\
where k(q) = 1 if q = 2 and k(q) = csc(-n-jq) + 1 if q > 2.
Proof. If C = (c)i)) is given and a nonzero H = (h
ij
) E G(q)sxm is such that
B m
(4.38) L :E hijc;i) = 0 E F::,
i=l j=l
NETS AND (t,S)-SEQUENCES 71
then the system {C)i) : 1 $ j $ di(H), 1 $ i $ s} is linearly dependent over
Fq, where di(H) = d(h
i1
, ,him) for 1 $ i $ s with the notation introduced
in (3.15). From the definition of p(G), it follows that E:=l p(G) + 1.
We obtain' the lower bound in the theorem from the fact that there is an H
.sAtisfying (4.38) and E:=l di(H) = p(G) + HI:. "
, 'il To prove the upper bound, we put D(H)' = E:=1 di,(H), and note that
from (3.16), (3.17), and (3.18) we obtain
Wq(H) $ k(q)8q-D(H).
Therefore
(4.39) Rq(G) $ k(q)8 L q-D(H) =: k(q)8 Sq(G),
H
where the sum is extended over all nonzero H = (hij ) E G(q)8Xm satisfy-
ing (4.38). We have
(4.40)
Sq( G) = L A( d)q-d1--d
s
,
d
where we sum over all d = (d
1
, ,d
s
) E Z8 with 0 $ d
i
$ m for 1 $ i $ s and
E:=l di p(G) + 1, and where A(d) is the number of H = (hij ) E G(q)8Xm
'satisfying (4.38) and di(H) = d
i
for 1 $ i $ s. If d is fixed and H is counted by
A(d), then (4.38) attains the form
i=l j=1
Choose integers Ii with 0 $ Ii $ di for 1 $ i $ sand E:=l Ii = p(G). Then the
last vector identity can be written in the form
(4.41 )
Suppose that we allow arbitrary choices for the coefficients hij on the right-hand
side of (4.41), the only stipulation being that hid. =1= 0 whenever di > Ii. Since
the vectors C;i) on the left-hand side of (4.41) are linearly independent over Fq ,
there can be at most one choice for the coefficients hi'; on the left-hand side.
Therefore
A(d) $ ir (q - 1)qdi -
f
.-1 = 8 q - 1 . ir qdi-f. $ (1- ! )qd1 + ... +ds - P(C).
i=l i=l q i=l q

Together with (4.40), this yields
Sq(C) $ (1- = (1- Dq-P(C)(m+ 1)' _ + 8)),
and, by invoking (4.39), we complete the proof. 0
72 CHAPTER 4
It follows from the results already shown in this section that a point set
constructed by (4.25) is a low-discrepancy point set precisely if p( C) is large.
The problem of maximizing the value of p( C) for a given finite field Fq and
for given integers m 2:: 1 and 8 2:: 1 is an interesting combinatorial question for
vector spaces over finite fields, which is also connected with a classical problem in
algebraic coding theory; see Niederreiter [244, 7], [260]. An explicit construction
of systems C with a large value of p( C) will be given in 4.5.
We now give the description of a general principle for the construction of
(t, s)-sequences in base b. Let the integers s 2:: 1 and b ;::: 2 be given. Then we
choose the following:
(81) A commutative ring R with identity and cardeR) = b;
(82) Bijections 'l/Jr : Zb R for r 2:: 0,
with 1p,JO) = 0 {(iT tl ..U 8ufficipn.tlv larQ"B r:
;. J 'Y )
(83) Bijections 17ij : R Zb for 1 i s and j 2:: 1;
(84) Elements E R for 1 i s, j 2:: 1, and r 2:: O.
For n = 0, 1, ... , let
00
n = Lar(n)b
r
r=O
be the digit expansion of n in base 0, where ar(n) E Zb for r 2:: 0 and ar(n) = 0
for all sufficiently large r. We put
00
= for n 2:: 0 and 1 i s,
;=1
with
(i) (i) )
Ynj = 17ij ejr 'l/Jr(ar(n)) E Zb
for n 2:: 0, 1 i s, and j 2:: 1.
Note that the sum over r is a finite sum, since 'l/Jr(O) = 0 and ar(n) = 0 for all
sufficiently large r. We now define the sequence
( 4.42)
_ ( (1) (8))
Xn - Xn , ... ,X
n
for n = 0,1, ....
To guarantee that the points Xn belong to [8 (and not just to ]8), and also for
the analysis of the sequence (4.42), we need the follovling condition:
(85) For each n 0 and 1 $ i s, we have < b -1 for infinitely many j.
This condition is always tacitly assumed when we consider the sequence (4.42).
A sufficient. condition for (85) is:
(86) 1]ij(O) = 0 for 1 i 8 and all sufficiently large j, and for each 1 $ i $ s
and r 2:: 0, we have = 0 for all sufficiently large j.
NETS AND (t,S)-SEQUENCES 73
Condition (86) guarantees that, for each n 0 and 1 ::; i ::; s, we have
= 0 for all sufficiently large j, and 00 each is given by a finite digit
expansion in base b.
THEOREM 4.35. Suppose that the integer t 0 satisfies the following prop-
erty: For any integers m > t and d
l
, ... ,dB 0 with E:=l d
i
= m - t and any
!fi) E R, 1 ::; j ::; d;, 1 ::; i ::; s, the system of m - t linear equations
m-l
z,. = fP) for 1 ::; j di , 1::; i ::; s
,.=0
l,n the unknowns Zo, ... ,Zm-l over R has exactly b
t
solutions. Then the se-
quence (4.42) is a (t, s)-sequence in base b.
Proof. For integers k 0 and m > t, consider the Xn with kb
ffl
::; n <
(1: + l)b
m
. In this range, the digits ar(n) of n are prescribed for r m, whereas
the ar(n) with 0 ::; l' ::; m - 1 can vary freely over Zb. Let
an elementary interval in base b with AIJ{E) = bt-mj hence L::=1 d
i
= m - t.
Using condition (S5) and proceeding as in the proof of Theorem 4.26, we obtain
that Xn E E if and only if
00
= l1ijl(aij) for 1::; j ::; di ! 1 $;i ::; 8.
,.=0
Since the ?/I,.(a,.(n with r m are given, this reduces to
m-l
I: = fj') for 1 ::; j ::; di , 1::; i ::; 8,
,.=0
with suitable f?) E R. By hypothesis, this system of linear equations in the
unknowns ?/I,.(a,.(n)) , 0 ::; r ::; m - 1, over R has exactly b
t
solutions. Each
solution corresponds to a unique m-tuple Cao(n), ... ,am-l (n)), and each such
m-tuple corresponds to a unique integer n with kb
m
::; n < (k+ l)b
m
. Therefore
the points Xn with kb
m
::; n < (k + l)b
m
form a (t, m, a)-net in base b. The
desired result is thuS shown by Definition 4.2. 0
In the case of a prime power base q, there is an analogue of Theorem 4.28 for
(t, s )-sequences. We must consider systems of vectors obtained from the elements
in (84).
THEOREM 4.36. Let q be a prime power, let R = Fq , and let t 0 be an
integer. If, for each integer m > t, the system c(m) consisting of the vectors
(i) _ ( (i) (i) pm
Cj - CjO "" ,Cj,m-l E q
for 1 i 5, 1 ::; j m,
satisfies p(c(m) m-t, then the sequence (4.42) is a (t,s)-sequence in base q.
74 CHAPTER 4
Proof. We verify the condition in Theorem 4.35 for the given value of t.
Consider integers m > t and d
1
, ... ,dB 0 with 2::=1 di = m - t. From
p(c(m m - t and Definition 4.27, it follows that the coefficient matrix of the
system of m - t linear equations in Theorem 4.35 has rank m - t, and so the
system always has exactly qt solutions. 0
There is also an analogue of Theorem 4.29 for (t, s)-sequences. If the ring R
is chosen as in that theorem, then the elements in (84) can again be written
in the form (4.27).
THEOREM 4.37. Let b = n:=l qv be a product of prime powers qv, let R =
Fq", and let t 0 be an -integer. Suppose that, for each integer m > t and
each v with 1 v h, the system consisting of the vectors
satisfies m - t. Then the sequence (4.42) is a (t, s)-sequence in
base b.
Proof. Proceed in analogy with the proofs of Theorems 4.29 and 4.36. 0
REMARK 4.38. The der Corput sequence in base b (see Definition 3.2)
arises as a special case from the general construction based on (81) - (85). Choose
s = 1 and let the ring R be the residue class ring Z/bZ. Then R and Zb can
be identified, and so all bijections 'l/Jr and 'fJij in (S2) and (S3), respectively,
can be taken to be identity maps. Put c);) = 1 if r = j - 1, and cj;) = 0
otherwise. Then condition (S6) is satisfied, and the sequence (4.42) is the van
der Corput sequence in base b. It is a trivial verification that the condition in
Theorem 4.35 is satisfied with t = O. Thl.'IB the van der Corput sequence in base
b is a (0, I}-sequence in base b (compare aloowith the beginning of 4.1),
4.4. A special construction of nets.
'We discuss a special family of nets arising from the general construction based
on (Nl) - (N4) in 4.3. We choose the base to be a prime power q, and the ring
R to be the finite field Fq with q elements. Let Fq((x-
1
) be the field of formal
Laurent series over Fq in the variable x-I, Thus the elements of Fq((x-
1
)) are
formal Laurent series
00
L= Ltkx-k,
k=w
where w is an arbitrary integer and all E Fq . The discrete exponential val-
uation v on Fq ( x-I)) is defined by v( L) = -w if L # 0 and w is the least
index with i,w # 0, and by v(O) = -06. vVe have v(J) = deg(f) for all nonzero
polynomials f E Fq[xJ. We also note that Fqx-
1
)) contains the field of rational
functions over Fq as a subfield.
For a given s 2, choose f E Fq [xl with deg(f) = m 1 and let
',' ,
F.
NETS AND (t,S)-SEQUENCES 75
9}' . .. ,98 E Fq [X]. Consider the expansions
9i(X) = t E Fq(x-
1
for 1 i 8,
f(x) k=Wi
. where Wi 1 for 1 ::; i 8. Define the elements in (N4) by
(4.43) for 1 i s, 1 j ffi) 0 r S m - 1.
Choose the bijections in (N2) and (N3) arbitrarily. Then the general construc-
tion principle' for nets described in 4.3 yields the point set (4.25) consisting
of qm points in [8. We denote this point set by peg, f)) where we write
g = (91, ... ,9s) E Fq [XJ8 for the s-tuple of polynomials g}, ... ,98' For an
arbitrary h = (hI!" . , h
s
) E Fq [x] s , we define the "inner product"
s
h'g= Lhigi .
i=l
.!:.S usual, we write Jig if f divides 9 E Fq[xJ. In the following, we use the
convention deg(O) = -1.
,DEFINITION 4.39. For f and g as above, we define
8
p(g,f) = s -1 + minLdeg(hi ),
i=1
where the minimum is extended over all nonzero h = (hI, ... ,hs) E Fq [XJ8 with
deg(hi ) < m for 1 i s and f I h g.
LEMMA 4.40. Let
(i) -' ( (i) (i) c Fm
Cj - CjO "'" Cj,m-l . C q
fDr 1 i s, 1 j S m,
where the are given by (4.43). Then, for hij E Fq, 1 i s, 1 j m, we
have
(4.44)
8 1n
"" "\:""'" h (i) - 0 F
m
ijCj - E q
i=1 j=l
if and only if f I h g, where h = (hI!'" ,hs) E Fq[x]S with
m
(4.45) hi(x) = L hijxi-
1
E Fq[x] for 1 SiS s.
j=l
Proof By comparing components, we see that (4.44) is equivalent to
(4.46) for 0 r m - 1.
i=l j=l
76 CHAPTER 4
For 1 i :::;: s, we have
hi(x)gi(X) _ h- __ i-1) (i) -k) _ h- _ (i) -1I:+j-1
() - L tJ;L- L., Uk x - L.t L., 2JUk x
f X j=l k=Wi j=1 k=Wi
m 00
"\:"""' h "\:"""' (i) - r-1
= L., ij L., Ur+jX .
j=1 r=wi-j
Thus, for r 0, the coefficient of x-
r
-
1
in higd f is E.i=l Therefore,
condition (4.46) is equivalent to the following: For 0 r m -1 the coefficient
of x-
r
-
1
in 2.::=1 higd f is O. This means that
1
,h.g=p+L,
where p E and L E Fq((x-
1
) with 1/(L) < -m. The last identity is
equivalent to
hg-pf=Lf.
On the left-hand side, we have a polynomial over F
q
, whereas, on the right-hand
side, we have v(LI) < 0 since v(l) = deg(l) = m. This is only possible if
Lf = 0, Le., if f I h g. 0
COROLLARY 4.41. If C is the system of vectors C)i) in Lemma 4.40 and p( C)
is as in Definition 4.2-7, then p(C) = p(g, f).
Proof. By Definition 4.27, there exist integers db'" ,d" with 0 di m for
1 ::; i 8 and E:=1 di = p(C) + 1 such that the system {C)i) : 1 j di , 1
i ::; 8} is linearly dependent over Fq Then there exist hij E F
q
, 1 :S j ::;
1 i ::; 8, not all zero, such that
" di
"\:"""' h (i) '" F
m
L.,L., ijCj = v E q'
i=l j=1
and, by putting hij = 0 for d
i
< j ::; m, 1 ::; i ::; s, we obtain an identity of the
form (4.44). By Lemma 4.40 it follows that f I h g, where h = (hI, ... ,h/3) =1= 0
with the polynomials hi in (4.45). Hence, from Definition 4.39, we obtain
8 /3
p(g, f) ::; 8 - 1 + L deg(hi ) ::; S - 1 -I- I:(di - 1) = p(C).
i=1 i=l
On the other hand, by Definition 4.39, there exists a nonzero h = (hI, .. . ,h
B
) E
Fq[X]8 with deg(hi ) < m for 1 ::; i s and f I h g such that p(g, J) =
8 -1 -I- deg(hi ). Then, by Lemma 4.40, we obtain (4.44) with the elements I
hij E Fq , 1 i s, 1 ::; j ::; m, determined by (4.45). Thus the system,j
{c)i) : 1 ::; j ::; deg(h
i
) -I- 1, 1 ::; i ::; 8} is linearly dependent over F
q
, and so
8
p(C) I:(deg(hd + 1) - 1 = p(g, f) o
i=1
NETS AND (t,S)-SEQUENCES 77
THEOREM 4.42. The
m - p(g,f).
set peg, is a (t, m, s)-net in base q with t =
Proof This follows from Theorem 4.28 and Corollary 4.41. 0
It is dear from Theorem 4.42 that the larger the value of peg, f), the better
the distribution properties of the point set peg, I). Therefore the number peg, I)
is called the figure of merit. From Corollary 4.41 and the remark following
Definition 4.27, we see that we always have 0 peg, I) m. From (4.31),
Theorem 4.30, and Corollary 4.41, we obtain the discrepancy bounds
(4.47) q
1
q-p(g,/) $ Dj.,(P(g, f)) $ B( B, q)q-p(g,f) (log N)s-l
+ o( q-p(g,J) (log N)O-2),
where the implied constant depends only on q and s. If ?Jij(O) = 0 for 1 $ i $ s,
1 $ j m, then the factor (q -1)/(3q) in the lower bound can be replaced by
(q:- 1)/(2q). .
13 in the case
qis prime and every ?Jij is the identity, map. IT h = (hI,'" ,hs) E Fq[x]1! with
deg(hi ) < m for i $ 8, then we can use (4.45) and the fact that C(q) forms
.1,',
!i a,complete residue system modq to identify h with H = (hij ) E C(q)sxm, and
we put Wq(h) = Wq(H). In analogy with the definition of Rq(C), we then set
( 4.48) R(g,1) ='I::Wq(h),
h
where the sum is over all nonzero h = (hi, . .. ) hI!) E Fq[X]8 with deg(h
i
) < m
for 1 i s and f I h g. It follows from Lemrna 4.40 that R(g, I) = Rq(C),
where C is the system of vectors C;i) in Lemma 4AO. Thus Lemma 4.32 shows
that
(4.49) Dj.,(P(g, f)) $ 1- (1- IJ + R(g, f) $ + R(g, f),
provided that q is prime and every "Iij is the identity map.
On the basis of (4.49), we will prove that the point sets peg, f) are on the
average low-discrepancy point sets. For s 2 and f E Fq[x] with q prime and
deg(f) = m 1, put
Let
Gs(f) = {g = (gl,'" ,gs) E Fq[x]s: gcd(gi, f) = 1 and deg(gi) < m
l$i$s}.
be the mean value of R(g, f) extended over the set Gs(f). Note that
= if!q(f)O, where iJ!q is the analogue of Euler's totient function
78 CHAPTER 4
for the ring Fq[x]. By a formula in [192, Lemma 3.69J, we have
.,.
(4.50) 'Pq(f) = qffi II (1 - q-n
k
),
k=l
where nl, ... ,n.,. are the degrees of the distinct monic irreducible polynomials
over Fq dividing f.
The proof of the following theorem depends on the theory of arithmetic func-
tions on Fq[xJ as developed by Carlitz [36J and on the theory of characters of
Fq(x-
1
)) as developed by Carlitz [37] and Hayes [131J. Henceforth, an arith-
metic function is a real-valued function on the multiplicative semigroup Sq of
monic polynomials over F
q
An arithmetic function E is called multiplicative (re-
spectively, additive) if E(gh) = E(g)E(h) (respectively, E(gh) = E(g) + E(h))
for all B, h E Sq with gcd(g, h) = 1. vVe write L,1) mod f for a sum over all
v E Fq[x] with deg(v) < deg(J), and we write 2:: mod! if the additional condi-
tion gcd(v, f) = 1 is imposed. Furthermore, 2:dl! denotes a sum over all d E Sq
dividing f.
Let X be a fixed nontrivial additive character of Fq. For L E Fqx-
1
)), put
Xq(L) = X(tl)' where tl is the coefficient of X-I in the expression for L. Then
Xq is an additive character of Fq(x-
1
)), that is trivial on Fq[xl. Consequently,
Xq(.j f) is a nontrivial additive character of the residue class ring Fq[x]!(f). For
9 E Fq [xl the orthogonality relations for characters yield
(4.51)
L Xq (VB) = if fig,
'IJ mod f f if J f g.
See, e.g., Car [35, p. 8J for this
THEOREM 4.43. Let q be a prime, let s 2 be an integer, let f E Fq[x] with
deg(f) = m 1, and put N = qm. Then

1 aN d)8_ logN o ((loglogN)2)
Il - N C
q
b + q CqS N + N '
where the implied constant depends only on q and S and where d
2
= 1, dq = 1/ q
for q > 2, C2 = 1/1og4, and
c
q
= -1_1- (q -1 + L csc 1rIZI) for q> 2.
q og q zEC*(q) q.
Proof. We can assume without loss of generality that f is monic. Inserting
the definition of R(g, f) in (4,48) into the expression for M8 U) and interchanging
the order of summation, we obtain
NETS AND (t,S)-SEQUENCES 79
where the sum is over all nonzero b = (h
b
,., j h
s
) E Fq[x]S with deg(h
i
) < m
for 1 ~ i ~ s and where A(b) is the number of g E Gs(f) with fib g, Since
ACO) = ipq(J)8 and Wq(O) = 1, we can write
(4.52)
where the sum is over all h = (hI, ... ,hs) E Fq[x]S with deg(h
i
) < m for
1 ~ i ~ s. For any such h, we have
A(h) = L q-m L Xq(Yh. g)
gEG .. (f) v mod f
by (4.51). By the definition of Wq(h), we can write
8
Wq(b) = n Qq(hi),
i=l
where we use (4.45) to identify hi with (hit, ... ,him) E Cm(q), and we define
Qq(hi ) to be the quantity Qq(hi1!'" ,him) in (3.16) und (3.17). Then we obtain
with
Now
Yq(O, J) = <Pq(f) L Qq(h);
hmod!
thus
(4.53) LA(h)VVq(h) = ~ ~ q ( f ) S ( L Qq(h))S + ~ L Yq(v,J)s,
h h modi v mod!
v,eO
Let J1q be the Mobius function on Sq (see [36] and [192, p. 145]), and note that
J1q is multiplicative. We abbreviate gcd(g, J) by (g, f) in this proof. Then, for
80 CHAPTER 4
fixed v E Fq[x] with 0 deg(v) < m, we obtain
Yq(v, f) = L Qq(h) L Xq (yh9) L tLq(d)
h mod! 9 mod f dl(g,f)
= 2: Qq(h) LJtq(d) L Xq(yh9)
hmod! dlf gmod!
dig
= L Qq(h) L tLq(d) L Xq (yhad)
hmod! dlf amodJld
= L Qq(h) L
hmodf dl! amodd
where, in the last step, we changed d into I/d. Applying (4.51) to the innermost
sum, we obtain
Yq(v,f) = L Qq(h) LtLq( = I> .. L Qq(h}.
hmod! dlf \ dlf hmod!
dlvh dlvh
Now d , vh if and only if d/(d, v) divides hi thus
(4.54)
Y: ( I) - ( f) deg(d) E ( d I)
q v, -;;; Jtq d q q (d, v)' ,
where, for an a E Sq dividing I, we put
If a = I, then
Now let a =1= Ii then
For q = 2, we have
Eq(a, f) = L Qq(h).
hmodf
alh
Eq(a, f) = 1 + L Qq(ab).
bmodfla
b#O
bmodfla
b=l=O
= deg 2- deg(a)-l.
k=O
NETS AND (t,S)-SEQUENCES 81
For q > 2 and for hE Fq[x] with 0 $ deg(h) < m, we have
where sgn(h) is thelea.ding coefficient of h, viewed as an element of C*(q). Since
a is monic, we obtain
2: Qq(ab) = L +U(deg(b),deg(f) -1))
b mod fla b mod fla

deg(f/a)-l
= q-deg(a)-l L q-kqk L (esc -1))
k=O zEC"(q)
= deg ( esc + (q -1) (deg ( f) -1 )q-deg(a)-l
N ,--"-,
= Tq deg ( )q- deg(a) _ qq- deg(a) ,
where we put Tq = c
q
log q for all q and q = (q - 1)/ q for q > 2. The case where
q = 2 is also covered by this formula if we put C2 = O. To include the case where
a = j, we put
{
c
q
if a = j,
oq(a,j) = 0
if a =1= f
Then, for all a E Sq dividing f, we have
(4.55) Eq(a, f) = 1 + Tq deg( ) q- deg(a) - (q - oq(a, f)q- cleg(a)
= 1 + (mTq - q + oq(a, f)q-deg(a) - Tq deg(a)q- deg(a).
Applying this formula with a = dl(d, v) in (4.54), we obtain
Yq(v,j) = LPq()
dlf
. (qdeg(d) + (mT _ c + 0 (_d_ j)) qdegd,v _ T deg (_d_) qdeg((d,v .
\ q q q (d, v) , q (d, v) )'
thus
(4.56) Yq( v, f) = <Pq (I) + (Cq log N - (v, 1)
H
(2) ( j) H(3) ( f)
-Tq q v, + q v,
82
with
CHAPTER 4
f) = 2::,uq
dlf
(v,f) = I'q ( deg ( (/ v) ) qdogd ,
H!3) (v, J) = )q'egd .
In the remainder of the proof, p will always denote a monic irreducible poly-
nomial over Fq. For a nonzero v E Fq[x], let ep(v) be the largest nonnegative
integer such that pep(v) divides v. Now we consider f) for a fixed v 1= O.
Since q{leg((d,u)) is a multiplicative function d, it follows that (v, is a
multiplicative function of f. For any integer k 2:: 1, we have
= qdeg((pie,v _ qdegpk-l,v.
Hence, if ep(v) < k, then = O. If ep(v) h, then
= qde
g
(p1c) _ qdeg(Pk-l) = <pq(pk) (
by (4.50). By multiplicativity, we obtain
( 4.57)
= if f I v,
if f fv.
Next, we consider for a fixed v 1= O. Since deg(d/(d,v)) is an
additive 'and qdegd,v a multiplicative function of d, it follows by induction on
the number of distinct polynomials p dividing f that
(4.58)
f) = L J /pep(J).
pll
For any integer k 1, we have
H(2)(V,ph) = de
g
( ph )qdegpk,V _ de
g
( pk-l )qdegpk-l,V.
q (Pk,V) (pk-l, v)
Hence, if ep(v) k, then = O. If ep(v) < k, then
= deg(p)qep(v)deg(p).
By (4.57) and (4,58), we obtain
f) = L /pep(J),
p
NETS AND (t,S)-SEQUENCES 83
where the sum is over all p satiB:f<;illg the following two conditions: (i) ep(v) <
ep(J); (li) f Ipep(f) divIdes 'l}, Note that (li) means that e
p1
(f) ::; e
p1
(v) for all
monic irreducible pol;ynomialE PI over FC] with PI "I po Thus (i) and (ii) hold
simultaneously if and only if there exists a unique p with ep(v) < ep(f} If this
condition is satisfied, then, with this p, we have
J) = deg(p)qep(v)deg(p)q,q(J Ipep(f),
whereas (v, J) = 0 otherwise.
Now we consider Note that 6
q
(dl(d,v),j) =I- 0 only if q > 2 and
dl(d, v) = f. Since d divides j, however, we have dl(d, v) = f if and only if
d = f and (J, v) = 1. Thus, if q > 2 and (f, v) = 1, then (v, f) = (q - l)lqo
In all other cases, we have (v, f) = O.
For v E Fq[x] with 0 ::; deg(v) < m, we have J) = 0 by (4.57),
so, in view of (4.56) and the formula for (v, we obtain
(1),
with an absolute implied constant. vVe combine this with (4.52) and (4.53) to
obtain
where Jq(v,j) = The formula for shows that
Jq(v,j) = 0(1); hence
(4.59)
Ms(f) = (:)(_Tq)i L Jq(v,f)i+O(q,
v mod! q
v:;i:O
with an implied constant depending only on q and s.
We now consider the sum over v in (4.59) for 1 ::; i ::; s. From the formula
for we obtain
J (.) deg(p)
q v, J = q,q(pep(l)-ep(v)
if there exists a unique p with ep(v) < ep(f), and Jq(v, f) = 0 otherwise. We
put
Hi (f) = L Jq( v, f)i for 1 ::; i ::; s.
vmodJ
v:;i:O
Let g, h E Sq with (g, h) = 1 and let v E Fq[x] with 0 :S deg(v) < deg(gh) be
such that there exists a unique p with ep(v) < ep(gh). Then since ep(gh) > 0, p
84 CHAPTER 4
divides exactly one of 9 and h, and so the sum Hi(gh) can be split up into two
subsums according to these two cases. If pig, say, then v = v1h with VI E Fq[x],
0:::; deg(vl) < deg(g), and Jq(v,gh) = Jq(vlh,gh) = Jq(Vllg). From this, it is
easily seen that Hi(gh) = Hi(g) + Hi(h); hence Hi is additive. For an integer
e 1 and any p, we have
e-I
Hi(pe) = deg(p)i L L iPq(pe-k)-'i,
k=O v modpe
ep(v)=k
For 0 :::; k :::; e -- 1, the number of V E Fq[x] with 0 :::; deg(v) < deg(pe) and
ep(v) = k is equal. to and so
e
( 4.60)
k=l
For i = 1, this yields H
1
(pe) = deg(p)e = deg(pe), thus H
1
(f) = deg(f) = m by
additivity. For i = 2, we obtain
e
H2(pe) = deg(p)2 L q-kde
g
(p)(1_ q- deg(p))-l
k=1
e
:::; 2 deg(p)2 L q-kdeg(p) < 4deg(p)2
q
- deg(p\
k=l
and so, by additivity,
H
2
(f) < 4 L deg(p)2q- deg(p).
pi!
Let n be the number of distinct monic irreducible polynomials over Fq dividing
f, and let PI, ' ., ,Pn be the first n in a list of all monic irreducible polynomials
over Fq ordered by non decreasing degrees. Then
n
H
2
(f) < 4 L deg(pj)2q-de
g
(pj) + 0(1) :::; 4 L r2q-1' Iq(r) + 0(1),
j=1 r=1
where D(n) = deg(Pn) and Iq(r) is the number of monic irreducible polynomials
over Fq of degree r. Since Iq(r) :::; q1' Ir for r 1 by [192, Cor. 3.21], it follows
that
D(n)
(4.61) Fh(f) < 4 L r + 0(1) = 0(D(n)2),
1'=1
On the other hand, we have
n
m = deg(f) L deg(pj) (D(n) - 1)Iq(D(n) - 1).
\ j=1
NETS AND (t,S)-SEQUENCES 85
From [192, Exer. 3.27], we obtain Iq(r) 2:: cqr IT for T 2:: 1 with an absolute
constant c > O. Thus, if D(n) 2:: 2, then m 2:: cqD(n)-1j hence D(n) = 0(1 +
logm), and this holds trivially if D(n) = 1. Combining this with (4.61), we
obtain
(4.62)
For i 2:: 3, it follows from (4.60) that
e e
Hi(pe) deg(p)i L = deg(p)i L q-2kde
g
(p) (1 _ q- de
g
(p))-2
k=1 k=1
< 2deg(p)i(qde
g
(p) _1)-2,
and so, by additivity,
00
Hi(f) < 2 L deg(g)i(qdeg(g) _1)-2 = 2 I:ki(qk _1)-2qk
gESq k=l
deg(g)::;:l
00
42: ki(qk - 1)-1 = 0(1).
1:=1
Now we use this information as well as (4.62) and HI (f) = m in for-
mula (4.59). This yields
1 s logN ((lOglOgN)2). ( 1 )
Ms(f) = NEq(l,f) N +0 .
By (4.55) we have Eq(l,J) = mTq + dq = cq logN + dq , and [35, Prop. VI.ll]
yields = 0(N-
1
1og1og(N + 1)). Hence the desired result follows. 0
It follows from (4.49) and Theorem 4.43 that, if q is prime, every 'rJij is the
identity map, and if s 2:: 2 and f E Fq[x] with deg(f) = m 2:: 1 are fixed, then, as
g runs through Gs(J), we get on the average Div(P(g,J) = O(N-l(logN)s).
We note that, for the number cq in Theorem 4.43, we have the bound
2 7 1
c < - + --- -- for q > 2,
q 1r 510gq qlogq
according to (3.22).
REMARK 4.44. In the special case where f is irreducible over F
q
, q prime, a
result of the same type as Theorem 4.43 can be shown in a much easier fashion.
Put m = deg(f) and let
G:(f) = {g = (91, ... ,9s) E Fq[x]S : deg(9i) < m for 1 i :::; s},
= (I)) L R(g, I).
car s
86 CHAPTER 4
Then card ( (1) = qffiE 1 and, in the same way as in the beginning of the proof
of Theorem 4.43, we obtain
M!(f) = q-m8 L A'(h)Wq(b),
h,eO
where A' (h) is the number of g E (f) with f I h g. For every h in the
range of summation, we have A'(h) = qm(s-l), since s - 1 entries of g can
be prescribed arbitrarily and the remaining entry (for which the corresponding
entry of his =1= 0) is uniquely determined because of the irreducibility of f. Thus,
with N = qm, we obtain
M;(f) = L Wq(b),
h,eO
and so Lemma 3.13 yields
M'(f) = (logN + 1)8 1 if q = 2,
8 N log4 N
I 1 (m 1rlhl m - 1 \ 8
Ms(f) = N - csc-+m---j
q hEC"'(q) q q
1 ((2 7 1) 1) [)
<- -+----- logN+-
N 1r 5 log q q log q q
1
N
1
N
if q > 2.
REMARK 4.45. If f is irreducible over Fq, then the in (4.43) can also
be represented as follows. We first note that the sequence , . .. is a
linear recurring sequence with characteristic polynomial f. Thus it follows from
Theorem A.2 in Appendix A that there exist elements Oil 1 :s; i s, in the
extension field FN of order N = qm such that
ui
i
) = Tr(OW
k
-
1
) for 1 i s and k;:::: 1,
where Tr is the trace function from FN to Fq and where u is a root of fin F
N
.
Hence (4.43) attains the form
= Tr((JiO'r+j-l) for 1 :s; i ::; s, 1 j m, 0 r m - l.
Consequently, if q is prime, f is irreducible over Fq , and the bijections 'l/Jr and "lij
are identity maps, then the construction of the point sets P(g, J) is a special case
ofthe construction in Niederreiter [239, p. 161], where, in the latter construction,
we take f3ij = Oiuj-1 E FN for 1 i s, 1 j m.
In the case where s = 2, there is a connection between the figure of
merit p(g, f) introduced Definition 4.39 and continued fractions for ra-
tional functions over Fq , where q is again an arbitrary prime power. Let
g = (gl, g2) E Fq[xJ2 with gcd(gi' f) = 1 for i = 1,2. Then the con-
dition h . g = hltg1 + h2g2 - 0 mod f in Definition 4.39 is equivalent to
NETS AND (t
7
S)-SEQUENCES 87
hI +h2yi92 :::::: 0 modf, where Yi E Fq[x] is such that glgi = 1 modf. Therefore
the figure of merit is the same for the original g and for the pair (1,9iY2)' Thus
it suffices to consider the figure of merit for pairs g of the form g = (1, g) with
gE Fq[x] and gcd(y, f) = 1. Let
9
f = [Ao;AI,A2'''' ,Azl
be the continued fraction expansion of the rational function 9J I, with partial
quotients An E Fq[x] satisfying deg(An) 1 for 1 :s; n :s; E. Put
(4.63)
K(-fY) = max deg(An) ,

Then we have the following formula.
THEOREM 4.46. If f E Fq[x] with deg(f) = m 2: 1 and g = (1,9) E Fq[xj2
with gOO(y, J) = 1, then
peg, f) = m + 1 - K (7) .
Proof If h = (hI, h
2
) E Fq[xj2 satisfies the conditions in Definition 4.39,
then we must have h2 =1= OJ hence 0 s:; deg(h2) < m. From hl + h29 == 0 modI,
it follows that hI = pf - h
2
y for somep E Fq[xJ, and actually p is chosen so that
deg( hd is minimaL Therefore
peg, f) = 1 + min(deg(h1) + deg(h2)) = 1 + min(deg(pf - h29) + deg(h2)
h
= m + 1 + -:2) + 2 de
g
(h2)),
where the last two minima are extended over all h2 E Fq[x] with 0 s:; deg(h2) < m
and all p E Fq[x]. Now we use notation and results from Appendix B. For
arbitrary h2 andp, as above, there exists n with 1 n I such that deg(Qn-l)
deg(h2) < deg(Qn}. Then
v(j - + 2deg(h2) v(y - + 2deg(Qn-l)
= deg(Qn-l) - deg(Qn) = - deg(An),
with equality if p = Pn-l, h2 = Qn-I. Thus
p(g, f) = m + 1 + min (-deg(An)) = m + 1- K(-f9). 0

It follows from Theorem 4.46 that the desirable pairs g = (1, g) are those with
a small value of K(yJ f). The least possible value of K(g/ f) is 1. For any pre-
scribed m 2: 1, we can obtain j, 9 E Fq [xl with deg(f) = m, gcd(g, f) = 1,
88 CHAPTER 4
and K (g / f) = 1, namely, by choosing partial quotients A
o
,... ,Am with
deg( Ar )- 1 for 1 r m and letting 9 / f be the resulting rational nmc-
tion (.40 can be arbitrary, the simplest choice being Ao = 0). H K(g/ J) = 1,
then, from Theorem 4.46, we obtain peg, f) = m for g = (l,g). With such
a choice of g and I, we obtain a two-dimensional point set P(g, f) satisfying
Div(P(g, f)) = O(N-1logN), according to (4.47). By the general lower bound
mentioned after (3.8), this is the smallest order of magnitude that can be achieved
by the star discrepancy of a two-dimensional point set.
A construction analogous to that of the point sets P(g, f) can also be applied
to sequences. A formal Laurent series L E Fq((x-
1
)) is called irrational if it is
not the expansion of a rational function over F
q
Now let s 1 be a given
dimension and choose irrational Lb' .. ,La E Fq(x-
I
, say
00
L
"" (i) -Ie
i = L-t Uk X
for 1 i s,
where Wi 1 for 1 i s. Then define the elements in (84) by
(i) _ (i)
(4.64) Cjr - Ur+jE Fq for 1 i s, j 2:: 1, r 2:: O.
The bijections in (82) are chosen arbitrarily, subject only to the condition im-
posed in (82). Then the general construction principle for sequences described
in 4.3 yields the sequence (4.42). It remains to guarantee that condition (85)
holds. The following is a sufficient condition for (85) to be satisfied. Here rfi/
denotes again the inverse map of the bijection 'fJij.
LEMMA 4.47. If the bijection817ij are such that 'fJi/(q - 1) is =J: 0 and inde-
pendent of j for all sufficiently large j, then, for each n 0 and 1 i 8, we
have < q - 1 for infinitely many j.
Proof Suppose that for some n and i we had = q - 1 for all sufficiently
large j. Since, for a suitable integer Rn 2:: 0, we have tPr(a1'{n)) = 0 for all
r> R,., it follows from (4.64) and the defhlition of the that
R..
L UnjtPr(a1'(n = ri;/(q - 1) for all sufficiently large j.
1'=0
By hypothesis there exists a nonzero mi E Fq such that 17i/ (q - 1) = mi for all
sufficiently large j. Thus, with a suitable integer jo 2: 1, we have
Rn
I: = mi for all j jo.
1'=0
Hence, if we put Vk = u1
i
ljo for k 2: 0, then the sequence vo, VI, . .. of elements
of Fq satisfies a nontrivial linear recursion and is thus ultimately periodic (com-
pare with Appendix A). Consequently, the sequence ui
i
) j . is ultimately
periodic, and so Li is the expansion of a rational function over F
q
, which is a
contradiction. 0
NETS AND (t,S)-SEQUENCES 89
To illustrate the analysis of these sequences,we consider the simplest case
where 8 = 1. For an irrational L1 E Fqx-
1
as above, let
L1 = [Ao; AI, A2 , . J
be its continued fraction expansion, with partial quotients Ad E Fq [X) , d =
0,1, ... , satisfying deg(Ad) 1 for d 1. Put
K(Ld = sup deg(Ad).

The hypothesis of Lemma. 4.47 is assumed in the COIl.truction to guarantee the
validity of condition (85).
THEOREM 4.48. If K(L
1
) < 00, then the above construction based on (4.64)
with 8 = 1 yields a (t, I)-sequence in base q with t = K(L
1
) - 1.
Proof. By Theorem 4.36, it suffices to verify that, for each integer m > t, the
vectors
(1) _ ( (1) (1 m
Cj - CjO "'" Cj,m_l E Fq ,
1 $j 5:m-t,
arB independent over Fir Suppose that) for some Tn >t, we had
m-t
2: hje)1) = 0 E F:;,
j=1
where not all h
j
E Fq are zero. Then
m-t
L = 0 for 05: r $ m-l.
j=l
With h(x) = E;:-;.t hjx
j
-
1
E Fq[x] we obtain.
hLl = (f u
i1
)x-
k
) = hj f ui
1
)x-
k
+
j
-
1
3=1 ) k=Wl 3=1 k=Wl
m-t 00
= L hj L
j=1 r=Wl-j
and so the coefficient of x-
r
-
1
in hLl is zero for 05: r 5: m-l. Thus v(hL1-p) <
-m for a suitable P E Fq[x]. Since deg(h) 5: m - t - 1, it follows that
(4.65) deg(h) + v(hL
1
- p) < -t - 1 = -K(L
t
).
On the other hand, with notation and results from Appendix B, we see that
there exists a d 1 with deg(Qd-d 5: deg(h) < deg( Qd) and that
deg(h) + v(hLl - p) = 2deg(h) + v( L1 - *) 2deg(Qd-l) + v( L1 -
= deg(Qd-l) - deg(Qd) = - deg(Aa) -K(Ll)'
This is a contradiction to (4.65). 0
90 CHAPTER 4
4.5. A special construction of (t,s)-sequences.
We first consider a prime power base q and we choose the ring R to be the
finite field Fq in the general construction based on (81) - (84) in 4.3. To obtain
suitable elements E Fq in (84), we again use the method of formal Laurent
series employed in 4.4. In the following, we describe the most interesting special
case of a construction introduced in Niederreiter [247].
For a given dimension s 2:: 1, let PI,'" ,Ps E Fq[x] be distinct monic irre-
ducible polynomials over Fq Put ei = deg(Pi) for 1 ::; i ::; s. For 1 ::; i ::; s and
integers j 2:: 1 and 0 ::; k < ei, consider the expansions
(4.66)
Ie 00
X _ I: (i) ( . k ) -r-l c::: F. (( -1
-(-)-. - a J, ,r x '- q x .
P
X J
1 r=O
( i)
the elements c
jr
(841 by
\ / "
(4.67) = a(i)(Q + 1, k, r) E Fq for 1 ::; i ::; 8, j 2:: 1, r 2:: 0,
where j - 1 = Qei + k with integers Q = Q( i, j) and k = k( i, j) satisfying
0::; k < ei. We assume that the bijections 'TJij in (83) are chosen in such a way
that 'TJij (0) = 0 for 1 ::; i ::; s and all sufficiently large j. The bijections in (82)
are selected arbitrarily, subject only to the condition imposed in (82). For each
1 ::; i ::; 8 and r 2:: 0, the elements in (4.67) satisPJ = 0 for all sufficiently
large j. Thus condition (86) holds, and so (4.42) yields a sequence of points
in IS.
THEOREl'4 4.49. The above construction, with the defined by (4.67),
yields a (t, s)-sequence in base q with t = L::=l (ei - 1).
Proof. By Theorem 4.36 and Definition 4.27, it suffices to verify the following
property: For any integer m > L::=l (ei - 1) and any integers d
ll
... ,ds 2:: 0
with 1::; L::=1 di ::; m - L::=l(ei -1), the vectors
(i) (i) (i)) Fm
C
j
= C
jO
j ,Cj,m-l E q
are linearly independent over Fq 8uppose that we have
S di
LLf?)C;i) = 0 E F:;
i=l j=l
for some fP) E Fq , where we can assume without loss of generality that all
d
i
2:: 1. By comparing components, we obtain
8 d.
(4.68) = 0 for 0::; r m-l.
i=l j=l
Consider the rational function
NETS AND (t,8)-SEQUENCES 91
where we used (4.66) and (4.67) in the second identity. In view of (4.68), we
have veL) < -m. If we put Qi = L(ds - l)/eiJ for 1 :5 i :5 s, then a common
denominator of Lis g(x) = n:=lPi(X)Q+l, and 00 Lg is a polynomial. On the
other hand, we have
B S
v(Lg) < -m + deg(g) = -m + L:(Qi + l)ei :5 -m + I)di - 1 + ei) :5 O.
i=l i=1
Thus Lg = 0; hence L = 0, and so
B d. (i) xk(i,j)
p,(x)Q(i,j)+l = o.
The left-hand side is the partial fraction decomposition of a rational function,
and 00 it follows from the uniqueness of the partial fraction decomposition that
all f?) = O. 0
In the general construction of Niedetreiter [247], Pb'" ,Ps can be any pair-
wise relatively prime polynomials over Fq of positive degrees. Furthermore,
in (4.66) we can replace xis /Pi(x)i by xkgij(X)/p,(x)j , where gij E Fq(x] with
gcd(gij,Pi) = 1 for 1 < i :5 s and j 1 and
,lim (jdeg(pi) - deg(Oij)) = 00 for 1 :5 i :5 s.
3-00 .
Now we return to the special case considered in Theorem 4.49, and we note
that, for fixed s and q, the minimum Value of t is obtained by choosing PI, ... ,Ps
as the "first s" mornc irreducible polynomials over F
q
This means that we list all
monic irreducible polynomials over Fq in a sequence according to nondecreasing
degrees, and then we let Pll ... ,PB be the first; s terms of this sequence. With
such a choice for PI, . .. ,Ps, we put
B
(4.69) Tq(s) = 2:(deg(Pi) - 1}.
i=l
The polynomials PI, ... ,Ps are not uniquely determined, but the number Tq(s)
is, of course, well defined. We then obtain the following consequences.
COROLLARY 4.50. For every dimension S 1 and every prime power q,
there exists a (Tq(s), s)-sequence in base q.
COROLLARY <.1, 51. Let b = qv be a p'loduct of prime powers qv' Then,
for every dimension B 1, there exists a (t, s)-sequence in base b with
t = max Tn (s).
l$v$h ,""
Proof. For each v with 1 :5 v :5 h, we choose elements E Fq" for 1 :5 i :S
8, j 1, and r 0 as in (4.67), with Plv,' .. ,Psv E Fq" [xl being selected in such
a. way that
8
I)deg(Piv) -1) = TqJs}.
i=l
92 CHAPTER 4
Then we define the elements E R = n:=l Fq" for 1 :5 i $ 8, j ;;:: 1, and r ;;:: 0
by (4.27). By the proof of Theorem 4.49, the system in Theorem 4.37
satisfies ;;:: m - Tq,,(s) for each integer m > Tq,,(s) and each v with
1 $ v $ h. Thus the condition in Theorem 4.37 holds with t = Tq,,(s);
hence the result. 0
REMARK 4.52. If q is a prime power and 8 is an arbitrary dimension $ q,
then we can choose for Pll'" ,Ps the linear polynomials Pi(X) = x - bi for
1 $ i $ s, where bb'" ,b
s
are distinct elements of Pq. Thus we have Tq(s) = 0
for s :5 q. Furthermore, (4.67) reduces to
= a(i)(j,O,r) for 1 $ i $ s, j;;:: 1, r;;:: 0,
and these elements are obtained from the expansion
= xi (1 _ = x-; t. (r; l)biX-'
= ( r -1'-1
L,; '-1 I x .
r=j-1 J
Thus, for 1 $ i $ s and j ;;:: 1, we have
C\i) = 0
31'
for $ r < j - 1,
\i) = ( r )
C]r . 1 '
)-
for r;;:: j -1,
where we use the convention 0 = 1 E Pq This choice of the yields the (0,8)-
sequences in base q constructed in Niederreiter [244], and, if we specialize further
to q being a prime, then this yields the sequences introduced by Faure [95]. Note
that, as far as the construction of (0, s )-sequences in base q is concerned, the
condition s :5 q on the dimension s is best possible, since Corollary 4.24 shows
that s $ q is a necessary condition for the existence of a (0, s )-sequence in base
q. We obtain the choice of the leading to the van del' Corput sequence in a
prime base q (see Remark 4.38) if we put s = 1 and PI (x) = x.
We now consider in more detail the quantity Tq(s) defined (4.69). Let
Iq(n) be the number of monic irreducible polynomials over Fq of degree n, and
let Jq (n) be the number of monic irreducible polynomials over Fq of degree $ n,
with Jq(O) = O. For given s ;;:: 1, let n = nq(s) be the largest integer with
Jq(n) $ s. Then it follows from the definition of Tq(s) that
nq(s)
(4.70) Tq(s) = L (h - l)Iq(h) + nq(s)(s - Jq(nq(s))).
h=l
The values of Tq(s) for q = 2,3,5 and 1 $ s $ 30 are given in the fonowing
tables. A general uJ?per bound for Tq{s) is shown in Theorem 4.54.
\;
NETS AND (t,S)-SEQUENCES
TABLE 4.1
Values ofT2(s) jar 1 s 30.
s 1 2 3 4 5 6 1 8 9 10 11
T2(S) 0 0 1 3 5 8 11 14 18 22 ,26
s 16 11 18 19 20 21 22 23 24 25 26
T2(S) 48 53 58 63 68 13 18 83 89 95 101
TABLE 4.2
Values oITg(s) for 1 s 30,
I
1 2 3 4 5 6 7 8 9 10 11
0 0 0 1 2 3 5 7 9 11 13
17 18 19 20 21 22 23 24 25 26
__' __

28 31 34 31 40 43 46 49 52 55
TABLE 4.3'
Values of Ts(s) lor 1 s 30.
S 1 2 3 4 5 6 1 8 9 10 11
Ts(s) 0 0 0 0 0 1 2 3 4 5 6
s 16 17 18 19 20 21 22 23 24 25 26
Ts(s) 12 14 16 18 20 22 24 26 28 30 32
LEMMA 4.53. For any prime power q, we have
1
Jq(n) 2: _qn lor all n 2: 1.
n
Proof By Appendix A, we have
Iq(n) = L JL(S)qd for all n 2: 1,
dIn
93
._ ..
12 13 14 15
30 34 38 43
21 28 29 30
107 113 119 125
12 13 14 15
15 11 19 22
27 28 29 30
58 61 64 67
12 13 14 15
7 8 9 10
27 28 29 30
34 36 38 40
where J-L is the Mobius function. This formula immediately yields the lower
bound in the lemma for n = 1,2,3. For n 2: 3 we use induction to obtain
o
94 CHAPTER 4
THEOREM 4.54. Let q be any prime power. Then, for 1 ::; s ~ q, we have
Tq(s) = 0, and, for s > q, we have
Tq(s) < 8(logq 8 + logq logq 8 + 1),
where logq denotes the logarithm to the base q.
Proof. The Jirst part of the theorem was already noted in Remark 4.52. For
S > q, we have
(4.71)
by (4.70). Put
k = llogq S + logq logq sJ + 2.
If either q = 2, Y ~ 4, or q ~ 3, y> 1, then
With y = logq s, we obtain
q logq 8 :2:: logq S + logq logq S + 2 :2:: k
if either q = 2, 8 :2:: 16, or q :2:: 3, S > q. In these cases, it follows that
and so Lemma 4.53 yields
1
Jq(k) ~ kll > s.
By the definition of nq(s), we then see that
nq (s) ~ k - 1 ~ logq S + logq logq S + 1,
and the bound for Tq(s) follows from (4.71). In the remaining case where q = 2,
3 ~ S ::; 15, the bound for Tq(s) is checked directly by using Table 4.1. 0
We now discllilB the important question of fi:ru:Hng a sequence S of points in
IS, whose star discrepancy satisfies
where the constant C
s
is as small as possible. The case where 8 = 1 has already
been considered in 3.1, and so we take 8 :2:: 2. For any prime power q, there
exists a (Tq(s), s)-sequence in base q by Corollary 4.50, and, for such a sequence
S, we have
by Theorem 4.17, where the implied constant depends only on q and 8. If we
optimize in this family of sequences, then we arrive at the value
(4.72)
NETS AND (t,S)-SEQUENCES 95
where the minimum is extended over all prime powers q. Thus, for any 8 ?:: 2,
there exists a sequence S with
(4.73)
where G
s
is given by (4.72) and where the implied constant depends only on s.
For 8 = 2, we have
1
G
2
= G(2,2) = 8(log 2)2 .
For s > 3, let ql(S) be the least even prime power?:: s, let Q2(S) be the least odd
prime power ~ s, and put
G ~ = mID<' G(s, q)qT,,(s) ,
q s
where the minimum is extended over all prime powers q < s. Then
since Tq(s) = 0 for q ~ s. Therefore, for s ~ 3, we obtain
(4.74) G = min ( G' .!. (1- _1_) ( Ql(S) ) s .!. (q2(S) - 1 ) 8)
8 \ ill 81 Ql(S) 21ogql(S))' 31 21ogQ2(s) .
This expresses Gs as a minimum of finitely many numbers. We tabulate the
values of Gil for 2 ~ s ~ 20. We again note that each value of G
s
is obtained
by considering a (Tq(s), s)-sequence in a suitable prime power base q. The ap-
propriate value of Q is listed in Table 4.4. For comparison, we also tabulate the
constants As appearing in the discrepancy bound (3.6) for Halton sequences.
The values of As and Gs in Table 4.4 have been rounded to three significant
digits.
96 CHAPTER 4
TABLE 4.4
Values of As! C
s
for 2 $ s $ 20.
s As C
s
q s
As Cs
q
11 (3.37) . 103 (8.12) . 10-
5
11
2 (6.57) . 10-
1
(2.60) . 10-
1
2 12 (1.68) .10
4
(5.60) . 10-
5
13
3
(8.16) . 10-
1
(1.26) . 10-
1
3 13 (9.06) .10
4
(1.01) . 10-
5
13
4 (1.26) . 10 (8.58) . 10-
2
3 14 (5.06) .10
5
(2.19) . 10-
5
13
5 (2.62) . 10 (2.47) . 10-
2
5 15 (3.02) .10
6
(4.42) . 10-
6
171
i 6 (6.14) 1JP
(1
10-
2
...
16 (1.98) .10
7
(7.80) .

,
7 (1.73) . 10
1
(4.11) . 10-
3
7 17 (1.41) . 10
8
(1.30) . 10-
7
17
8 (5.30) . 10
1
(2.99) . 10-
3
9 18 (1.03) . 10
9
(8.47) . 10-
8
19
9 (1.86) . 10
2
(6.05) . 10-
4
9 19 (8.06) . 10
9
(1.36) . 10-
8
19
10 (7.72) . 10
2
(4.28) . 10-
4
11
Ii 20
(6.62) . 1010 (3.28) . 10-
8
23
To obtain information about the asymptotic behavior of C
s
as S ---+ 00, we
use (4.74) and q1 (8) < 28 to derive
C <.! ( ql (s ) ) s < .! ( s ) 8
,') s! 21ogql(S) 's! log(2s)
Together with Stirling's formula, this yields
lim logCs <-1.
s_oo slog log s -
Thus C
s
--} 0 at a superexponential rate as s ---+ 00. This should be compared
with the observation at the beginning of this chapter that As ---+ 00 at a super-
exponential rate as S -> 00. The sequences satisfying (4.73) have asymptotically
the smallest upper bound for the star discrepancy that is currently known.
The (t, s )-sequences in base q constructed in this section permit a particularly
convenient implementation in the case where q = 2, since then the digits are
obtained by binary arithmetic and the coordinates are dyadic rationals.
For the base q = 2, Sobol' [323J constructed (t, s)-sequences in base 2 for any
dimension !3. Let U (3) denote the least value of t that can be achieved by the
construction of Sobol' for given 8. By Corollary 4.50, the construction in this
section yields a (T2 (s),s)-sequence in base 2 for any s. A comparison with
the formula for U(s) given in [323, Thm. 3.4] shows that T2(S) = U(s) for
1 $ s $7 and T2(S) < U(s) for all 8 8. Therefore, for all dimensions s 8,
the construction in this section yields dyadic sequences having a smaller upper
bound for the star 'discrepancy than the sequences of Sobol' .
NETS AND
97
The construction in this section also has implications for the problem of
determining the number tb(S) in Definition 4.25. By Corollary 4.50, we derive
tq(s) ::; Tq(s) for all prime powers q and all 8. In particular, the first part of
Theorem 4.54 then impliesthattq(s )='O'ror alls $' q. Furthermore, the remarks
following Defuiittorl 4.25 show that tq(s) 2: 1 for's 2: q+ VBince 'Tq(q + 1)= 1,
we obtain tq(q + 1) = 1 for all prime powers q. For an arbitrary base b 2: 2, it
follows from Corollary 4.51 that, ifi the notation of this result, we have
(4.75) tb(S)::; max Tq,,(s) for all s 1.
l:5v:5h
If q denotes the least prime power appearing in the canonical factorization of b
into a product of prime powers, then tb(S) = 0 for alls ::; q. For s > q, it follows
from (4.75) and Theorem 4.54 that
tb(S) < s(logq a + logq logq 8 + 1).
Another exact value of tb{ 8) is known for b = 6. By the remarks following
Definition 4.25, we have ta(s) 2: 1 for 8 2: 3, and so (4.75) shows that t6(3) = 1.
The construction of (t, s )-sequences in this section also yields information
about the existence of nets. Indeed, by combining Lemma 4.22 and Corol-
lary 4.51, we obtain the following result.
COROLLARY 4.55. 'Letb = rC=l qv be a product of prime powers Q'l1 let
s 2: 2, and let t = maxl:5v:5hTq,,(S -1). Then, for every m'2: i, there exists a
(t, m, a)-net in base b.
By the device in Lemma 3.7, the low-discrepancy sequences constructed in
this section can be used to obtain low-discrepancy point sets. For given s 2: 3,
let S be a. sequence of points Xc, Xl, in [B-1 for which
ND"N(S) C,,_1(logN)S-1 + O((logN)S-2) for all N 2: 2,
where C,,-l is given by (4.72) and where the implied constant depends only on 8
(compare with (4.73)). For fixed N 2: 2, let P be the point set consisting of
(n/N,x"..) E JB for n = 0,1, ... ,N -1. Then, by Lemma 3.7, we have
NDN(P) ::; C
s
_
1
(logN)S-1 + O((logN)s-2) ,
where the implied constant depends only on 8. For s = 2, we get low-discrepancy
point sets by using the one-dimensional low-discrepancy sequences 3.1 and
the device in Lemma 3.7.
We briefly describe a method for the construction of (0, s )-sequences in a
prime power base q j which uses the theory of hyperderivatives and was introduced
in Niederreiter [251]> For an integer k ;:::: 0, the kth hyperderivaiive is the Fq-
linear operator H(le) on the polynomial ring Fq[x] defined by H(k) (x
T
) =
for r 2: k and H(k)(x'f') = for 0::; r < k. We again use the general principle for
the construction of (t, s)-sequences based on (81) - (84) in 4.3, with R = Fq
THEOREM 4.56. For s ::; q, let bI, . .. ,b
s
be s distinct elements of Fq For
1 i s, let 9i E Fq[x] with 9i(bi ) =1= O. Define
= [H(j-1}(X
r
9i(X)](b
i
) for 1 ::; i s, j ;:::: 1, r 2: 0,
98 CHAPTER 4
and suppose that the bijections 'fJij are such that 'fJij (0) = 0 for 1 ::; i ::; s and all
sufficiently large j. Then the sequence (4.42) is a (0, s)-sequence in base q.
Proof. It is clear that condition (S6) in 4.3 is satisfied. To prove that the
given sequence is a (0, s )-sequence in base q, it suffices to show by Theorem 4.36
that, for any integers m ~ 1 and d1, ... ,dB ~ 0 with L::=l di, = m, the vectors
(
i) (i) F m
CjO " .. ,Cj,m-l E q
are linearly independent over F
q
Suppose that the columns of the m x m matrix
formed by these row vectors satisfy a linear dependence relation. Then there exist
1o, ... , fm-l E Fq such that
m-l
L fr[H(j-l)(x1"gi(X))]Cb
i
) = 0 for 1 :-:; j::; d
i1
1 ~ i::; s.
1"=0
With f(x) = L:::o
1
l1"x
1"
E Fq[x], this yields
[H(k) (fYi)](b
i
) = 0 for 0::; k ::; d
i
- 1, 1::; i ~ s.
For 1 ::; i ::; 5, this implies by [192, Lemma 6.51] that b
i
is a root of fgi of
multiplicity at least d
i
Since gi(b
i
) i- 0, it follows that bi. is a root of I of
multiplicity at least d
i
. However, deg(J) < m = 1::=1 di" and so we must have
f = 0, I.e., f1" = 0 for 0 ::; r ::; m - 1. 0
Notes.
The basic paper for the theory of (t, m, B)-nets and (t, B)-sequences is Nieder-
reiter [244]. The results of 4.1 and 4.2 and most of the results in 4.3
stem from this paper. For base 2, earlier results are due to Sobol' [323J.
Surveys of the work of Sobol' can be found in Niederreiter [225, 3] and
Sobol' [324]. A construction of (0,2)-sequences in base 2 is also given in Srini-
vasan [333]. For a general background on latin squares, we refer to the book
of Denes and Keedwell [59]. Converses of Theorems 4.28 and 4.36 are shown in
[244, 6]. Note that om definition of p(C) is slightly different from that in [244].
Theorem 4.30, Lemma 4.32, and Theorem 4.33 are due to Niederreiter [261 J.
Theorem 4.34 is a special case of a result in Niederreiter [241], and Theorem 4.37
was shown in Niederreiter [247].
The construction of nets in 4.4 was introduced in Niederreiter [261], whereas
the idea of using formal Laurent series for the construction of nets and (t, s)-
sequences was previously established in Niederreiter [247J. Theorem 4.43 is a
new result. Theorem 4.46 is shown in the same way as a special case that was
considered in Niederreiter [241]. The quantity K(g/f) was studied in detail in
Niederreiter [243]. Calculations of the figure of merit peg, f) have been carried
out in certain special cases, e.g., for g of the form g = (1, x
m
, x
2m
, ... j x(s-l)m);
see Andre, Mullen, and Niederreiter [8] and Mullen and Niederreiter [215]. This
figure of merit also occurs in the context of pseudorandom number genera-
tion (compare with 9.1). The analysis of the sequences constructed in 4.4
NETS AND (t,S)-SEQUENCES 99
leads to interesting connections with problems of diophantine approximation in
Fqx-
1
)), which are further explored by Larcher and Niederreiter [180]. For the
construction in 4.5, we need tables of irreducible polynomials over finite fields;
we refer to [192, Chap. 10] for such tables. An interesting application of (t,8)-
sequences to the numerical solution of integro-differential equations occurs in
Lecot [184]. Fox [107] described a computer implementation of Faure sequences
and compared their efficiency with that of Halton sequences and pseudorandom
numbers; see also Sarkar and Prasad [301] for another comparative study. A
computer implementation of Sobol' sequences was carried out by Antonov and
Saleev [9], and an improved scheme was developed by Bratley and Fox [33J.
5
[Y : () j 'I, _ " ;\' ,'. . ' : , ; \V , . .
Rules,for Numerical
Integration
quasi-Monte Carlo methods for numerical integration that we have discussed
. 'ibased on low-discrepancy point sets and sequences. An inspection of the
ml'll'lI"uU:P error bounds in 2.2 reveals a feature that may be construed as
'Ati7nA4"1i<' of these techniques, namely, that once the integrand is sufficiently
say of bounded variation in the sense of Hardy and 1< '!'aus'P i any
".,.'/"' . ...., ..... regularity of the integrand is not reflected in the order of magnitude
bound. This is in marked contrast 'to classical one-dimensional
methods such as Caussian formulas and Newton-Cotes ruleS, which
tailored to the regularity class of the integrand so that they become more
for more regular integrands,
The quasi-Monte Carlo m.ethods presented in this chapter enjoy the property
," iihe degree of regularity of the integrand is reflected in the order of magnitude
imhliIJIS error bound. To achieve this desirable goal, we must a.ctually assume that
integrand is periodic with period interval js so that the underlying Fourier
makes sense, but this is no serious restriction since a nonperiodic inte-
can always be periodized (see 5.1). The integration rules discussed here
" can be viewed as multidimensional analogues of the one-dimensional trapezoidal
. rule for periodic integrands. Historically, these integration rules first arose in
the special form of the method of good lattice points introduced by Korobov in
, whereas the general class of lattice rules (or lattice methods) was defined
more recently. Particularly in the last few years, intensive research
were devoted to these numerical integration techniques.
In 5.1 we present the fundamentals of the method of good lattice points,
which serve as a stepping stone for the general theory of lattice rules. Theorems
guaranteeing the existence of good parameters in the method of good lattice
points are shown in 5.2, where we also discuss the interesting connections with
continued fractions arising in the two-dimensional case. General lattice rules
with their intriguing group-theoretic and geometric facets are studied in 5.3,
while 5.4 is devoted to existence theorems for good parameters in the context
of general lattice rules.
101
102 CHAPTER 5
5.1. The method of good lattice points.
In 3.1 we mentioned the uniformly distributed sequences 8(70) of fractional
parts Xn = {nz} E jB, n = 0,1, ... ) where z = (Zl,'" ,zs) E lR
s
is such that
1, Zl) ,Zs are linearly independent over the rationals. Discrete versions of
these sequences are obtained if we consider points z with rational coordinates.
If z E QS is such a point and the positive integer N is a common denominator
of its coordinates, then z = N-1g with g E ZS. We are thus led to the point set
(5.1) Xn = {; g } E JB for n = 0, 1, ... ,N - 1.
Note that it is not necessary to consider the points Xn for n ? N, since they just
replicate the points in (5.1). With the points in (5.1), we get the quasi-Monte
Carlo approximation
(5.2)
This approximation is particularly suited for periodic integrands. Let f be
a periodic function on lR
s
with period interval Is (or, equivalently, with period
1 in each of its s variables). Then, first of all, we can drop the fractional parts
in (5.2) to get the simpler form
(5.3)
Furthermore, suppose that f is represented by the absolutely convergent Fourier
series
feu) = I>I(h)e(b' u) for u E ]RB
hez"
with FoUrier coefficients
j(b) = ~ f(u)e(-h. u)du for hE ZS,
11
8
where as usual e(u) = e
2
11'v'=I
u
for u E ]I and X Y denotes the standard inner
product of X, y E ]Rs. Then, since the exact value of the integral in (5.3) is given
by 1(0), we obtain
1 N-l ( ) f 1 N-l ()
N L f ;g -}/B feu) du = N L L j(b)e ;h. g -j(O)
n=O I n=O.hez"
1 N-l ( )
'. N L i(b) L e ;11. g - }(O)
heZ n=O
LATTICE RULES FOR ND:MERlCAL INTEGRATION 103
Now the last inner sum is equal to 0 if h . g =t= 0 mod N and equal to N if
h . g = 0 mod N, and so
(5,4)
f(;g) -h. f(u)du =
where the sum on the right-hand side is over all nonzero h E 1,;s with h g =
o mod N. Thus the integration eITor in (5.3) can be expressed as a sum of certain
Fourier coefficients of f.
According to the s-dimensional Riemann-Lebesgue lemma, f(h) tends to
zero as h moves away from the origin. The rate of convergence of j (h) toward
zero serves as a regularity condition on J. For h E Z, we put
r(h) = ma.x(1, Ihl),
and, for h = (hI! .. . ,hs) E 1,;B, we set
DEFINITION 5.1. Let Ot > 1 and C > 0 be real numbers: Then is
defined to be the class of all continuous periodic functions J on JRB with period
interval Is and with .
li(h)1 Cr(h)-O for all nonzero h E Z;S.
Furthermore, is the class of all f with! E for some C> O.
It is easily seen that if f E then its Fourier series is absolutely convergent
and represents f. An important sufficient condition for the membership of a
periodic function f on ]Rs with period interval fa in the regularity class is the
following. Let Ot > 1 be an integer and suppose that all partial derivatives
a
m1
++
ms
f
ml J:l_.m ..
uU
l
uUs
with 0 mi Ot. - 1 for 1 i s
exist and are of bounded variation on Is in the sense Hardy and Krause; then
! E with a value of C which can be given explicitly (see Zaremba [363]).
A more restrictive sufficient condition is the following: If Ot > 1 is an integer and
all partial derivatives
a
m1
++
ms
f
J:l_.ml
(fUl .. u S
exist and are continuous on ]Rs, then f E with an explicit value of C.
DEFINITION 5.2. For a real number Ot > 1, for g E ZS, and for an integer
N 1, we put
Pa(g, N) = L r(h)-O,
h
where the sum is over all nonzero h E 'Jls with h g == 0 mod N.
104 CHAPTER 5
THEOREM 5.3. For any real numbers Of. > 1 and C > 0, for any g E ZS and
any integer N 1, we have
max Nt L f N
n
g - f f(u)du = CPa(g,N).
N-1 ( )
n=O . ljB
Proof For f E !(C), the bound
follows immediately from (5.4) and Definitions 5.1 and 5.2. Now let fo be the
special function
lo(u) = C L r(h)-Oe(h u) for u E R8.
hez
Then /0 E and
fo (;g) -1. fo(u} du = CPa(g, N)
by (5.4) and Definition 5.2. 0
Theorem 5.3 Shows that, for given a and N, the lattice point g should be
chosen in such away that Pa(g, N) is small. We now introduce a related quantity
that does not depend on the regularity parameter a. The idea for its definition
stems from the observation that the main contributions to the sum defining
Pa(g, N) come from the lattice points h close to O. In the following definition,
we use the set C: (N) introduced in 3.2.
DEFINITION 5.4. For g E ZS, 8 2, and an integer N 2, we put
R(g, N) = I: r(h)-l,
h
where the sum is over aU h E C:(N) with h g == 0 mod N.
Next, we establish a bound for Po(g, N) in terms of R(g, N) in an important
special case. A more general, but somewhat weaker, bound ",lin be proved in
Theorem 5.26. Let (01.) = E:=l m-
O
for CIt > 1 be the lliemann zeta-function.
THEOREM 5.5. Let N 2 be an integer and let g = (911". jUs) E ZS, s 2)
with gcd(gi,N) =1 lor 1 8. Then, for any real 0> 1,'we have
Pafg,N) < + (1 + 2(a)N-
a
)S-1
-It + 2(01.) + 2((a)N
1
-
o
)8 - + 2(0)S,
LATTICE RULES FOR NUMERICAL INTEGRATION 105
Proof. Every h E can be uniquely written in the form h = k + N m with
k E ell(N), mE 1,;8. Thus we can write
(5.5)
Pa(g,N) = L r(Nm)-a +
Now
mEZ"
m#O ,I'
kEC:(N) mEZ",
k:g=:O modN
(5.6) 81 = L r(Nm)-a -1 = (L:: r(N'Tn)-a)
8
-1
mEZ8 mEZ
= (1+ 2 f, (Nm)-a)' ~ 1 = (1 + 2a)N-O)' ~ 1.
Furthermore,
(5.7)
where k = (k1 ; ... j ks ). For k = 0, we have
+ Nm)-a = L:: r(Nm)-a = 1 + 2({a)N-0I.
mEZ mEZ
For 0 < Ikl :::; N /2, we have
mEZ mEZ
00 00
=Ikl-
a
+ L{k+Nm)-a+ I)-k+Nm)-a
m=l m=l
m=l m=l
Thus, in both cases,
L r{k + Nm)-Ot :::; r(k)-a + 2a((a)N-
a
,
mEZ
106 CHAPTER 5
and so from (5.7) we obtain
s
II (r(ki)-O + 2(a)N-
a
)
kEC:(N) i=1
kg=O modN
< Ra(g, N) + 20S(a)B N-OS#{k E Cs(N) : k . g = 0 mod N}
s-1
+ L 2
01
(8-
j
)(0:)8-
j
N-o(s-j)
j=l l$;il <"'<i;$;s kEC,,(N)
with
{l'l)
kg=O moaN
kg=O modN
For fixed 1 $. j $. s - 1 and 1 $. il < ... < ij $. 8, we have
2: r(kiJ-a ... r(kij)-a=
kEC.(N)
kS=O modN
L r(b)-a#{k E C,/N) : k g = 0 mod N and kid = hd for 1 $. d $. i},
hECj(N)
where h = (hI, . .. ,h
j
). The last counting function can be determined explicitly.
If in the congruence
k . g = k
1
g
1
+ ... + kBg
S
= 0 mod N
the values of k
h
,.. . , ki; are prescribed, then 8 - j - 1 of the remaining co-
ordinates of k can be chosen arbitrarily, and the last remaining coordinate of
k is then uniquely determined since gcd(gil N) = 1 for 1 $. i $. s. ThuS the
counting function is given by Ns-j-1. Similarly, we have #{k E Cs(N) : k g =
o modN} = Nil-I. We obtain
L r(kiJ-a ... r(k
i
; )-0 = Ns-j-l L r(h)-a < (1+2(0:))3 Ns-j-\
kEC8 (N) hECj(N)
kg=O modN
and so
LATTICE RULES FOR NUMERICAL INTEGRATION 107
Together with (5.5) and (5.6), this yields
(5.8) Pa(g, N) < Ra(g, N) + (1 + 2((a)N-
a
y, - 1
+
The result of the theorem follows from the ,,!la N) S R(g, N)a. Co
" We infer from Theorem 5.5 that, is then Pa(g, N) is small
for all a > 1. The quantity R(g, N) '" be used to provide an upper bound
for the discrepancy of the point set (5.1. For 8 2 and N 2, we use the
notation in 3.2 to define
(5.9) R1(g,N) = 2: r (h,N)-I,
It!
where we sum over all h E C:(N) with II g ..... O'modN.
THEOREM 5.6. For g E ZS, 8 2, af'd 2, let P be the point
set (5.1). Then " .,",'!( "nil)''.)')'
DN(P) S 1- (1- d + R1(g,N)"5 ; +
Proof. By Theorem 3.10 with M = N and Yn = ng for 0 S n S N - 1, we
obtain
(
1 " s 1 1 N-l ( )
DNfp) < 1 - 1- -) + '"" - '"" e
\ - ... N L- r(h N) N L- N
bEe; (N)' n=O
= 1- (1- ! r +R,(g,N).
The second inequality of the theorem is obtained from r(h, N) 2r(h) for
hE C;(N), which follows, in turn, from sin(-Jrt) 2t for 0 ::; t 0
On the basis of the discrepancy bound in Theorem 5.6, the point set (5,1)
can also be used for the numerical integration of nonperiodic functions of low
regularity, e.g., of bounded variation on js in the sense of Hardy and Krause.
However, the full power of the point set (5.1) is achieved only for periodic in-
tegrands belonging to a function class There are methods for periodization,
i.e., for transforming a sufficiently regular nonperiodic integrand into an inte-
grand belonging to a suitable without changing the value of the integral. A
simple periodization technique is the replacement of a given fiJ.nction f on js by
the function
1 1
!(UI, ... ,us) = 2-
s
L: ... L f(cl + (-1)e
1
Ub . ,Cs + (-lyBus)
el=O es=O
for (Ul
l
... ,us) E Is. This technique may be viewed as an analogue of the
method of antithetic variates described in 1.2, but it is of limited usefulness
108 CHAPTER 5
because of a possible lack of regularity of the periodic extension of ! to RS at
the boundary of Is. More satisfactory techniques are based on suitable changes
of variables or on the device of adding functions to f that are obtained from
Bernoulli polynomials and certain partial. derivatives of f. We refer to Hna and
Wang [145, Chap. 6], Korobov [160, Chap. 1], and Zaremba [365] for detailed
discussions of periodization methods.
Another quantity measuring the quality of lattice points g is the positive
integer introduced in Definition 5.7 below. The idea is that P
a
(g, N) and R(g, N)
will be small if the nonzero lattice p8filts' b. with h . g =0 mod N are rather far
, l
from O.
DEFINITION 5.7. For g E 7/, s 2, and an integer N 2) the figure of
merit peg, N) is defined by
(,.., i\T\
Pvt.,., 'i ;
, . " ..
, - - . " \
where the minimum is extended over all nonzero h E za with h g = 0 modN.
LEMMA 5.8. We always have 1 :s; peg, N) :s; N/2.
"
Proof. If g = (9b .. ' ,98) with gCd(gl,N) = 1, then hgl + g2 - 0 modN
for some h E Z with -N/2 < h :s; N/2j and so p(g,N) ::::; r(h) :s; N/2. If
gcd(gl, N) > 1, then there exists a proper divisor d of N with dgl = 0 modN,
and 80 peg, N) ::::; red) ::::; N/2. 0
REMARK 5.9. It follows from Lemma 5.8 that it suffices to extend the min-
imum in Definition 5.7over all h E C:(N) withh g = 0 mddN. If a nonzero
b E 'L
s
is such tha.t h . g = 0 mod N and peg, N) = reb), then Lemma 5.8
implies that not all coordinates of h are divisible by N. Thus, if we reduce all
coordinates of h modulo N to obtain a pofut ho E C: (N), then ho . g = 0 mod N
and r(ho) ::::; r(h)i hence peg, N) = r(ho)
. The qua.ntities:Pa(g, l'{) irdm below a.nda.bove
in terms of the figure of merit peg, Nt ill detiill, for any g E ZS, 8 2, and any
integer N'? 2; we have '.I . , ' ";
(5.10)
2 < P. ( N) < c(s,a) (1 + logp(g, N))8-l for all a> 1,
peg, N)a - a g, . peg, N)a
where the constant c(s, a) depends only on sand a.The lower bound in (5.10)
is trivial, and the upper bound is a special. case of Theorem 5.34. Similarly, we
have
(5.11)
1 " < R( N"
( N)
- g, } - (' N) ,
p g, .. , p g, "
where the constant c(s) depends only on s. T4e lower bound in (5.11) follows
from Remark 5.9, the upper bound is a special case of Theorem 5.35. For
LATTICE RULES FOR NUMERICAL INTEGRATION
109
the discrepancy of the point set Pin (5.1), we have the bounds
(5.12)
Cl(S) < D (P) < c2(s)(logN)8
peg, N) -.. N - peg, N) ,
where the constants Cl(S) > 0 and C2(S) depend only on s. The lower bound
in (5.12) is a special case of Theorem 5.37 and the upper bound follows from
Theorem 5.6, Lemma 5.8,' and (5.11). We infer from (5.10), (5.11), and (5.12)
that, for a given N ~ 2, the lattice points g E 1.
s
, s ~ 2, which are suitable
for numerical integration-in the sense of yielding a small integration error-are
those for which peg, N) is large. Such a g is informally called a good lattice point
mod N, hence the name "method of good lattice points."
. Good lattice points in large dimensions automatically yield good lattice
points in smaller dimensions. Concretely, for a given dimension s ~ 3, let
g(s) = (91,'" ,9s) E1.
s
,
and, for any dimension t with 2 :::; t < 5, put
(t) _ ( )
g - 91, ,9t E
Then, from Definition 5.7, we immediately obtain
(5.13)
p(g(t) , N) ~ p(g(s) , N) for any integer N ~ 2.
Thus, if g{ s) is an a-dimensional good lattice point mod N, then g( t) is a t-
dimensional good lattice point modN. However, there is no obvious way of
.. constructing higher-dimensional good lattice points from lower-dimensional good
lattice points. In analogy with (5.13), we have
Pa(g(t) , N) :::; Pa(g(s) , N) for any N ~ 2 and a> 1
and
R(g(t) , N) :::; R(g(s) , N) for any N ~ 2.
5.2. Existence theorems for good lattice points.
To guarantee that the method of good lattice points is practicable, we need
results that demonstrate that, for given 8 ~ 2 and N ~ 2, there exist points
g E zs such that peg, N) is large and Pa(g, N) and R(g, N) are small. We start
with the last quantity and show that the average value of R(g, N), taken over a
wen-chosen set of lattice points g, is small. We first note that g is only relevant
modulo N, so that it suffices to take g E Cs(N). We put
Recall that the Euler-Mascheroni constant is given by
'Y = lim ( ~ ~ -lOgn) = 0.577 ....
n---+ 00 L....t m
m=l
110 CHAPTER 5
THEOREM 5.10. For any integers s ~ 2 and N ~ 2, put
Then we have
1
M8(N) = card(G
8
(N)) L R(g, N).
gEGs(N)
M (N) = ~ ( 2 1 N )8 _ 2s1ogN o((loglogN)2)
8 N og +C N + N '
where c = 2-y -log 4 + 1 = 0.768 . .. and where the implied constant depends only
on s.
Proof. Note that card(Gs(N)) = (NY\ where is Euler's totient function.
By the definition of R(g, N), we then aerhre
1
r(h)-l = ;(N)S L A(h)r(h)-l,
hEC;(N)
where A(h) is the number of g E GB(N) with h g = 0 modN. Since A(O) =
;(N)B and 1'(0) = 1, we obtain
(5.14)
1
M8(N) = ;(N)8 A(h)r(h)-l - 1.
hEC,,(N)
For any h, we have
N-l (k )
A(h) = ~ e Nh. g .
gEG.(N) 11:=0
In this proof, we write (m, n) for gcd(m, n). Then
1 N-I
= N I: T(k, N)8
k=O
with
T(k,N) = e ( ~ h g )r(h)-l.
\ hEC(N) gEC(N)
(g,N)=l
LATTICE RULES FOR NUMERICAL INTEGRATION 111
If we put
L(N) = E r(h)-l = L Ihl-
I
,
hEq*(N) hEC-(N)
. :', "
.\ . I
. then T(O, N) = (N)(l + L(N}), and so
. . N-l
(5.15) L A(b)r(b)-l= + L(N8 + L T(k, N)8.
hECs(N) k=l
For fixed 1 k N - 1, we have
T(k,N) = :E r(h)-l :E L JL(d) ,
hEC(N) gEC(N) dl(g,N)
'. , ": '_A (;0(, ; '.' ',' . ,
where JL is the Mobius function and d is restricted to positive divisors; if we also
fix h E C(N), then
Therefore
T(k, N) = LJL( L r(h)-l,
diN hEC(N)
dlkh
Now d divides kh if and only if d/{d, k) divides h; hence
(5.16)
where, for a positive divisor b of N, we put
If b < N, then
L(b, N) = L r(h)-l.
hEC(N)
blh
L(b,N) = 1 + L Ihl-
1
= 1 + I: labl-
1
= 1 +
hEC"(N) aEC"(N/b)
blh
112 CHAPTER 5
If we put L(1) = 0, then this formula for L(b, N) also holds for b = N. Together
with (5.16), this yields
(5.17) T(k,N) = LJL( + L(
diN
= (N) + LJL(
diN
By [226, Lemmas 1, 2J we have, for any integer m 1,
(5.18) L(m) = 210gm+c-1 +c(m) with Ic(m)1 < 4m-
2

Consequently,
= (21ogN +c--l)B(k,N) - 2H(k,N) + V(k,N)
diN
with
B(k,N) =
diN
H(k,N) =
diN .
V(k,N) =
diN
For a prime power pm with 1, we haveB(k,pm) = (pm,k) -(pm-l,k)j .
hence B(k,pm) = pm - pm-l = (Pffl) if pm I k and B(k,pm) = 0 otherwise. For
fixed k, we note that (d, k) is a multiplicative function of d, and 80 B(k, N) is a
multiplicative function of N. Thus, for any positive integers k and N, we obtain
(5.19)
B(k, N) = {cf>(N) if N I
o otherwlse.
In particular, for 1 k N - 1, we have B(k, N) = 0, and so
(5.20) T(k,N) = (N) - 2H(k,N)+ V(k,N) for 1 k N-1.
In the rest of the proof, p will always denote a prime number. For a positive
integer n, let ep(n) be the largest nonnegative integer such that pep(n) divides
n. We now consider H(k, N) for a fixed k. Since (d, k) is a multiplicative and
log(d/{d, k) an additive function of d, it follows by induction on the number of
distinct J:l!ime factors. of N that
(5.21)
H(k, N) = L H(k,pep(N))B(k, N/pep(N)).
piN
LATTICE RULES FOR NUMERICAL INTEGRATION 113
For an integer m 1, we have
ep(k),'then.(pm,k) ...::. pm and (pm-l,k) =pm-l j hence H(k,pm) = o.
H m > ep(k), then (pm,k) = (pm-1,k) = pep(k); hence H(k,pm) = pep(k) logp.
Together with (5.19) and (5.21), this yields
(5.22)
H(k, N) = Lpep(k)>{N/pep(N) logpl
p
where the sum runs over all p satisfying the fonowing two conditions: (i) ep(N) >
ep(k)j (li) N/pep(N) divides k. Note that (li) means that e
pl
(N) e
p1
(k) for all
primes PI =f:. p. Therefore (i) and (li) hold simultaneously if and only if there
exists a unique p with ep(JV) > ep(k). Hence it follows from (5.22) that if there
exists a unique prime p with ep(N) > ep(k), then
H(k, N) = pe
p
(k)4>(N/pe
p
(N logp,
and H(k, JV) = 0 otherwise.
To treat V(k,N), we use the bound for c{m) in (5.18) to obtain
IV(k, N)I 4 I:IJL( I 2 = 4 I: = 0(1)
diN diN
with an absolute implied constant. Combining this with (5.14), (5.15),
and (5.20), we obtain
1 1 N-l( ( 1 ))8
MsCN) = N(1 + L(NS + N k=l 1 - 2J(k, N)+ 0 4>(N) - 1,
where J(k, N) = H(k, N)/>(N). The formula for H(k, N), given above, shows
that J(k, N) = 0(1); hence
(5.23)
M.(N) = L(N))" + (:) (-2)' '%;' J(k,N)i + 0
with an implied constant depending only on s.
We now consider the sum over k in (5.23) for 1 i s. From the formula
for H(k, N), we obtain
114 CHAPTER 5
if there exists a unique prime p with ep(k) < ep(N), and J(k, N) = 0 otherwise.
We put
N-l
Qi(N) = L J(k,N)i for 1 ~ i ::; s,
and we claim that Qi is an additive function. Let NI and N2 be positive integers
with (N
I
, N2 ) = 1. H 1 ~ k ~ NIN2 - 1 is such that there exists a unique p
with ep(k) < ep(NIN2)' then ep(NlN2) > 0, and so p divides exactly one of
Nl and N
2
Consequently, the sum Qi(N
l
N
2
) can be split up into two subsums
according to these two cases. IT p I Nt, say, then k = kIN2 with 1 ::; kl ::; Nl - 1
and J(k, N
1
N2 ) = J(k1N
2
,N
l
N2 ) = J(kl' Nd. From this, it is easily seen that
Qi(N
1
N2 ) = Qi(Nd + Qi(N2 ); hence Qi is additive. For an integer m ;:::: 1 and
any prime p, we have
m-i -1
Qi(pm) = (logp)i L :E (pm-j)-i.
;=0 k=l
ep(k)=j
For 0 ::; j ~ m -1, the number of k with 1 ::; k ::; pm - 1 and ep(k) = j is equal
to (pm-
j
) , and SO
m
(5.24) Qi(pm) = (logp)iL: (pi)1-i.
j=1
For i = 1, this yields Ql(pm) = log(pm)j hence QI(N) = 10gN by additivity.
For i = 2, we obtain
m m
Q2(pm) = (logp)2 Lp-j (1- p-1)-1 ::; 2(10gp)2 LP-i < 4(10gp)2p-l,
j=1
and so, by additivity,
Q2(N) < 4 :2)10gp)2p-l.
piN
j=1
If n is the number of distinct prime factors of N and PI, . .. ,Pn are the first n
prime numbers, then
n
Q2(N) < 4 L(logpt}2ptl + 0(1)
t=1
with an absolute implied constant.. By the prime number theorem, we have
clt log(t + 1) ::; Pt ::; c2tlOg(t + 1) for 1 ::; i ::; n ",lith absOlute constants
cl, C2 > 0, and so . < .
q2(N) = o(t log(t + 1)) = o ((log(n + 12)
t=l t
LATTICE RULES FOR NUMERlCAL INTEGRATION 115
with absolute implied constants. :From the trivial bound log N 2:: n log 2, we see
that
(5.25)
with an absolute implied constant. For i 2:: 3, it follo'ws from (5.24) that
m m
Qi(pm) (logpy: L (Jl)-2 = (logp)i LP-2
j
(1- p-l)-2
j=l j=l
< 2(logp)i(p - 1)-2,
and so, by additivity,
with an implied constant depending only on i.
Now we use this information as well as (5.25) and Ql(N) = logN III for-
mula (5.23). This yields
M (N) = L(N))B _ 2s1ogN O((loglogN)2\ _ 0(_1_)
s N f- N + N) f- <p( N)
. 2s1ogN +O((loglOgN)2)
N N N
with an implied constant depending only on s, where we used (N)-l
O(N-
1
1og1og(N + 1)). The rest follows from (5.18). 0
Theorem 5.10 is an improved version of an earlier result in Niedeneiter [226],
according to which we have
1 ( 7)8
M8(N) < N 210gN + 5
for all 8 2:: 2 and N 2:: 2.
Note that, for any s 2:: 2 and N 2:: 2, there exists agE G
8
(IV) with R(g, N)
Ms(N). It follows then from Theorem 5.6 that, for any s 2:: 2 and N 2:: 2, there
exists agE Gs(N) such that the point set P in (5.1) has discrepancy
where the implied constant depends only on s. Theorem 5.10 shows that the
average order of magnitude of R(g, N) for g E G 8 (N) is N-
I
(log N)8 I but there
is even a general lower bound due to Larcher [175], which says that
for all g E 1:,s, s 2:: 2, and for all IV 2:: 2, where the constant C
s
> 0 depends only
on s.
116 CHAPTER 5
We now consider the question of the existence of lattice points g for which
the figure of merit peg, N) is large. Here the relevant existence theorem is
due to Zaremba [366]. For s 2 and N 2, let E8(N) be the set of all
g = (9b ... ,98) E e8 (N) with 91 = 1. The following auxiliary result from
Zaremba [366, eq. (27)] is basic.
LEMMA 5.l1. For integers s 2, t 2, and N 2, let F8(t,N) be the
number of g E EB(N) with peg, N) t. Then
F (t N) < 2
8
-
1
N8-2t(lOgt)8-1(1 +O(loglog(N + lb(S)))
8, - (s - I)! log t '
where the implied constant depends only on s and where b( s) = 3 for s = 2 and
b(s) = 8 - 1 for 8 3.
It is clear that if to is of the form to = c
8
N/(logN)B-l with a suitable C
8
> 0
depending only on s, then < N3--1 all sulficlently large N, and so
for such N there exists agE Es(N) with peg, N) > to. By a somewhat more
refined argument, Zaremba [366] established the following result.
THEOREM 5.12. For every dimension 8 2 and every sufficiently large in-
teger N, there exists agE Ea(N) such that
(8 -l)!N
peg, N) > (21ogN)B-1'
It follows from (5.10) and Theorem 5.12 that, for every s 2 and N 2,
there exists agE Ea(N) such that
(5.26) PO/(g,N) = O(N-C<(logN)(Ot+l)(8-1) for all 0: > 1,
with an implied constant depending only on s and 0:. Furthermore, Theorems 5.5
and 5.10 show that, for every s 2 and N 2, there exists agE Ga(N) such
that
(5.27)
where the implied constant depends only on s and a and grows exponentially
with s. Disney [66] and Disney and Sloan [67] have developed an approach
that yields the bound in (5.27) for a suitable g E G s (N) depending also on
0:, but with the coefficient of the leading term being (2e/s)Ot8, which decreases
superexponentially as s increases and 0: is fixed. On the other hand, there is
a general lower bound of Sharygin [308J, which shows that Po.(g, N) is always
at least of the order of magnitude N-Ot{log N)8-1. The following two results
of Niederreiter {264] are currently the best general existence theorems for small
values of POt(g, N).
THEOREM 5.13. For any integers 8 2 and N 2 and any real at> 1, the.1'e
exists agE G a( N) with
-P. (g N) = (N-o.(lOgN)Ot{S-lHl ( , (Ot-1)(8-1)
0. J \ ;( N) ) ,
where the implied cJnstant depends only on sand 0:.
Proof. As
LATTICE RULES FOR NUMERICAL INTEGRATION 117
proof of Theorem 5.5, we put
hEC:(N)
hS=O modN
For fixed s,N, and a, consider now
M = M(s,N,a):= L
gEG.(N)
Inserting the definition of Ro. (g, N) and interchanging the order of summation,
we obtain '
hEC:(N) gEGs(N)
hs=O
If, for an h E C;(N), we have h g _ 0 modN, then p(g,N) :::; r(h) by Defi-
nition 5.7. Thus, for every g in the inner sum, we have p(g,N)a-l r(b)'ll-l,
and so
L r(h)-l L 1 = L R(g,N).
hEa:(N) SEG.(N) gEG.(N)
hs=O modN
By Theorem 5.10, we then obtain
(5.28)
For an integer t 2: 2, let K(s, N, t) be thenumber of g E G/j(N) with peg, N) > t.
If we choose t such that K(s, N, t) > 0, then it follows from the definition of M
that
M > t
a
-
1
L Ra(g, N).
sEG.(N)
p(g,Nt
Together with (5.28), this yields
L Ra(g, N) = o (t1-a<fJ(N)/j N-
1
(log N)/j),
gEGs(N)
p(s,Nt
and so there exists agE G s (N) with peg, N) > t and
where the implied constant depends only on s. Since Gs(N) has <fJ(N)/j elements,
we can write
K(s, N, t) = <fJ(N)8 - card( {g E Gs(N) : peg, N) :::; t}).
Now g = (91, ... ,9s) E G 8 (N) satisfies the congruence h . g = 0 mod N if and
only if ihg satisfies the same congruence, where ih is determined by 9191 ==
118 CHAPTER 5
1 modN. Furthermore, gIg is congruent modulo N to a unique element of
Es(N). Thus
(5.30) K(s, N, t) (N)" - (N)Fs(t, N).
If we put
B = c(s)(N)s-1 l B J
Ns-2 and to = (log B)s-l
with a constant c(s) > 0, then to 2:: 2 and B e for all sufficiently large N;
hence to(logtO)s-1 s B for such N. If c(s) is chosen suitably, then it follows
from (5.30) and Lemma 5.11 that
K(s,N,to) cl(s)(N)S
for a cO'nf(tant Cl(S) > 0 "TId all large In particular, we have
K(s, iV, to) > 0 for all sufficiently large N, and so (5.29) is applicable with
t = to. :For agE Gs(N) satisfying (5.29), we then obtain
R,,(g,N) = o( N-"(logN)a(B-')+1 (0-')(0-1).
Since by (5.8) we have
the result of the theorem follows. 0
If we carefully keep track of the constants in the above proof, as is done in
the original paper [2641, then we obtain the bound for POt(g, N) in Theorem 5.13
with the coefficient of the leading term being
2Q (s-1)+10: a
(
. )Q-l
(s-1)!(a-1) ,
which decreases superexponentially as s increases and 0: is fixed. Using (N)-l =
(N-
1
log log ( N + 1) ), we may also write the bound in Theorem 5.13 in the form
PQ(g,N) = O(N-
Q
(logN)Q(s-I)+1 (loglog(N + 1{0I-l)(S-1).
THEOREM 5.14. For any integers s 2 and N 2 and any real 0: > 1, there
exists agE Es(N) with
PQ(g, N) = (N-Q(log N)0I(s-1) (1 + 'T(N) )) if 8 3,
(logN)s-l
N) = 0 ( N-
Ot
(log N)Q + ) ) if s = 2,
where the implied constants depend only on 8 and Ct and 'where 'T(N) denotes the
number of positive H.ivisors of N.
LATTICE RULES FOR NUMERlCAL INTEGRATION 119
Proof For fixed s, N, and a and an integer t 2:: 2, cOIlBider
Het) = H(s,N,a,t):= L Pa(g,N).
geE.(N)
p(g,Nt
For h E 2;6, let D(h,N) be the number of g,E E.(N) with b g = 0 mod A!.
Then, using the definition of Pa(g, N) and the fact that, if peg, N) > t, then a
nonzero h E Z can satisfy h . g == 0 mod N only if reb) > t, we see that
(5.31) H(t} ~ L r(h)-a D(h, N}.
heZ"
r(ht
To determine DCh, N) for any h = (hb'" ,h.) E ,we note that, if g =
(1,g21'" ,g.) E E6(N) is a solution of
h . g = hI + h2g2 + ... + h.ga = 0 mod lV,
then d = gcd( N, ha, . .. ,ha) must divide hI' In this case, we divide the congru-
ence by d and obtain
(5.32)
with m = N/d and gcd(m,ka,.oo ,ka) = 1. If m > 1, then let pB be a prime
power in the canonical factorization of m. Since at least one ki! 2 ~ i ~ s,
is not divisible by p, the number of solutions modpB of (5.32) is pa(a-2). By
combining th'8 solutioIlB by the Chinese remainder theorem, the number of s0-
lutions modrn of (5.32) is ma-a, and this holds for m = 1 as welL Thus
D(h, N) = (N/m)8-1mB-2 = Na-
2
d if d divides hI, and D(h, N) = 0 otherwise.
Using this formula (5.31), we obtain
N(t) ::; N
s
-
2
2: gcd(N, h
a
, .. I ha)r(h)-Ot,
h
where the sum is over all h = (hI. ... ,hs) E zs for which r(h) > t and d =
gcd(N, h
2l
.. ,hs) divides hI' By splitting up the sum according to the value of
d, we obtain
(5.33)
with
H(t) ~ N
s
-
2
L dTd(t)
diN
Td(t) = L r(dh)-a.
neZ
r(dht
We split up the last sum according to the number i of nonzero coordinates of h
and the position of these nonzero coordinates. This yields
(5.34)
120
where
Vi(u) =
(hI,'" ,hi)EZ'
Ih1 .. h.l>u
CHAPTER 5
For integers 1 i 8 and k 1, let be the number of (hI, ... ,hi) E Zi
with 0 < Ihl ... hil :::s; k, and put Ai(O) = O. Then
k>u k>u
< 0 L k-
Ot
-
1
Ai(k).
k>u
In the remainder of the proof, the implied constants in the Landau symbols
a.lways de'Dend only on 8 and o. By induction on i, we see that
and so
Vi(u) = 0(2: k-a(log(k + li-l) = O(ul-(log(u + li-l).
k>u
Thus, from (5.34),
and so, by (5.33),
(5.35)
H(t) = 0(N8-2tl-otOOgt)i-1 Ld
1
-
i
).
i=l diN
For i = 1, we have EdlN d
1
-
i
= T(N), and, for i = 2, we have EdlN d
1
-
i
=
Ed N d-
I
< N 14>( N), where the inequality is obtained by comparing the two
functions at prime powers and using mllitiplicativity. For i 2:: 3, we
have LdlN d
1
-
i
= 0(1). For 8 2:: 3, we thus obtain from (5.35),
Let (8, N, t) be the numbers of g E Es.(N) with: p(g, N) > t. Then L(8, N, t) =
Ns-l - Fs(t, N)i hence if we put
l
c(8)N J
to = (logC(S)N)B-l
LATTICE RULES FOR NUMERlCAL INTEGRATION 121
with a suitable constant c(s) > 0, then to ;::: 2 and L(8, N, to) ;::: cl(s)NB-l for
all sufficiently large N and some constant Cl (8) > 0, by Lemma 5.11. It follows
then from the definition of H(t) and from (5.36) that, for s ;::: 3, there exists a
g E ElI(N) with peg, N) > to and
For 8 = 2, we see, from (5.35), that
H(t) = logt + r(N))).
By choosing to as above and using a similar argument, we obtain the result of
the theorem for 8 = 2. 0
The more detailed performed in (2641 yields the bounds Theo.-
rem 6.14 wlth explicit coefficients of the main terms. These coefficients decrease
superexponentially as s increases and a is fixed: Since 'T( N) has the average
order of magnitude log N by [188, Thm. 6.30], the result of Theorem 5.14 is usu-
ally better than that of TheOrem 5.13. However, there exist sequences of values
of N through which 'T(N) grows faster than any given power of logN (see [188,
p. 164]), and, for such values of N, Theorem 5.13 yields the better result.
The lattice pointsg satisfying the bounds in Theorems 5.13 and 5.14 depend,
in particular, on Ot, so that theSe theorems can guarantee only that Pc:lg, N) is
small for the chosen value of a. However, the proofs of these theorems yield the
additional property that the figure of merit peg, N) is large, and on the basis of
this information it can be shown that P,a(g, N) is small for all f3 > 1 (see [264]).
For s = 2, there is an interesting connection between good lattice points
and continued fractions for rational numbers, which can be used for the explicit
construction of good lattice points. For an integer N ;::: 2, let g = (1, g) E Z2
with gcd(g, N) = 1. Let the rational number giN have the continued fraction
expansion
(5.37)
where the aj are integers with aj 1 for 1 ::; j ::; I and where az = 1 for the
sake of uniqueness (compare with Appendix B). Recall that the convergents to
giN are defined by
p.
_L = [oo;al,a2,'" ,aj] for O::;j::; l.
qj
The integers Pj and qj are uniquely determined if we impose the conditions
qj ;::: 1 and gcd(pj, qj) = 1. Then we have the following explicit formula for the
figure of merit due to Borosh and Niederreiter [30].
122 CHAPTER 5
THEOREM 5.15. If N ~ 2 is an integer and g = (1, g) E Z2 with gcd(g, N) =
1, then
p(g, N) = min qjlqjg - pjNI.
O ~ j < l
Proof. Because of Lemma 5.8, it suffices to extend the minimum in Defi-
nition 5.7 over all1ionzero h = (hI, h
2
) E Z2 with Ihll < N, Ih21 < N, and
h g == 0 modN. for such h, we also have hlh2 i= O. Therefore, we can write
where the minimum is extended over all integers h2 with 1 ~ h2 < N and over all
integers t. If1 ~ h2 < N, then for some j with 0 ~ j < l we have qj ~ h2 < qj+l'
Then, by an inequality in Appendix B, we obtain
hence the result. 0
DEFINITION 5.16. For integers N ~ 2 and 9 with gcd(g,N) = 1, let
K - = max a (
g)
N I ~ J ~ l 3'
where al, ... ,a, are the partial quotients in the continued fraction expansion of
giN given by (5.37).
In the case considered in Theorem 5.15, the figure of merit p(g, N) can be
bounded in terms of K(gIN) , as was shown by Zaremba [361].
THEOREM 5.17. If N ~ 2 is an integer and g = (l,g) E Z2 with gcd(g,N) =
1, then
N < (N)< N
K(gIN) + 2 - P g, - K(gIN)'
Proof. From the theory of continued fractions (see Appendix B), we derive
1 < IlL - Pj I < 1 for 0 ~ j < 1,
qj(qj + qj+l) - N qj - qjqj+l
and so
Nq Nq
_--=3,-- ~ qj!qjg _ pjNI ~ __ 3 for 0 ~ j < l.
qj + qj+l qj+l
Now qj + qj+l = (aHl + l)qj + qj-l ~ (K(g/N) + 2)qj for 0 ~ j < I, and
together with Theorem 5.15 this yields the desired lower bound. If we choose j
with 0 ~ j < 1 and aj+l = K(gIN), then qj+l = (lj+lqj + qj-l ~ K(g/N)qj)
and, in view of Theorem 5.15, we obtain the desired upper bound. 0
From Theorem 5;17, we inferthe principlethat two-dimensional good la.ttice
points g = (l,g) modN are obtained by choosing rational numbers giN with
small partial quotients in their continued' fraction expansion. This principle
leads, in particular, to the following explicit construction based on Fibonacci
numbers. Recall \hat the sequence F
I
, F
2
, of Fibonacci numbers is defined
LATTICE RULES FOR NUMERICAL INTEGRATION 123
recursively by FI = F2 = 1 and Fm ,= Fm - l +F
m
- 2 for m 3. Now let N = Fm
for some m 3 and g = (1, Fm - l ). Since all partial quotients of Fm-I! Fm used
in Definition 5.16 are equal to 1, we have K(Fm-I! Fm) = 1, a.nd so peg, N) is of
the order of magnitude N by Theorem 5.17. In fact, YJe have peg, Fm) = F
m
-
2
according to a formula of Zaremba. [3tH]. From (5.10), we obtain PO/(g, N) =
for all a > 1 with an implied. constant depending only on a, a
result due to Bakhvalov [131. This paper is also the first to contain this explicit
construction of two-dimensional good lattice points. From (5.11), we infer that
R(g,N) = O(N-l(logN)2), where the implied constant is absolute. Note that
these bounds for PQ(g, N) and R(g, N) are the best possible thatcan be obtained.
for any two-dimensional lattice point with any modulus, according to the general
lower bounds for POI (g, N) and R(g, N) due to Sharygin (3081 and Larcher [175],
respectively, which were mentioned earlier in this sectiop..
It is an interesting question as to whether the orders of magnitude of peg, N),
PQ(g, N), and R(g, N) achieved by the above construction can be obtained for
any integer N 2:: 2 in the two-dimensional case. This question leads to the
consideration of the quantity
(5.38)
KN = for N
gcd(g,N)=l
It is clear from the above arguments that, if there exists an absolute constant C
such that KN ::; C for all N 2, then our question has an affirmative answer.
In this context, we draw the attention to Zaremba's conjecture from [365, p. 76],
which amounts to suggesting that KN ::; 5 for all N 2. The constant 5 cannot
be replaced by a smaller one, since K54 = K
150
= 5 (these are the only known.
values of KN that are equal to 5). Borosh calculated KN for 2 ::; N 10
4
(see [225, p. 989]), and Knuth [154, p. 548J extended this calculation to all
N < 2 . 10
6
, the result being that Zaremba's conjecture is valid in this range.
Borosh and Niederreiter [30] conjectured that KN 3 for all sufficiently large
N. Zaremba's conjecture has been verified for some sequences of values of N.
It is trivial that KN = 1 if and only if N is a Fibonacci number. Nieder-
reiter [238] established the following results by constructive proofs: if N = 2
m
or 3
m
, m 1, then KN ::; 3, and KN = 2 for infinitely many m; if N = 5
m
,
m 2:: 1, then KN ::; 4, and KN ::; 3 for infinitely many m.
Consider again a point g = (1, g) E ?} with gcd(g, N) = 1 and N 2. Let the
continued fraction expansion of giN be given by (5.37), Then the discrepancy
DN(P) of the two-dimensional point set P in (5.1) satisfies
l
(5.39) N DN(P) ::; L aj + 1.
j=l
This is shown by adapting the method in the proof of Theorem 3.3 for rational z
and bJ:: using an analogue of Lemma 3.7 for the discrepancy DN(P); also compare
124 CHAPTER 5
with Niederreiter [225, pp. 1024-1025J. The quantity
S(.!L) := Eaj
N . 1
J=
is connected with K(g/N) in Definition 5.16 by the inequality S(g/N) <
cK (g / N) log N with an absolute constant c > 0, since we have l :::; clog N.
Together with (5.39), this yields
NDN(P) = logN)
with an absolute implied constant, and a slightly more refined bound is obtained
by adapting the method in the proof of Corollary 3.5. In analogy with (5.38),
we define
SN = min S(!L) for N 2.
gEZ N
gcd(g,N)=l
Then we have SN CKN logN for all N 2. It is conjectured that SN -
O(log N) with an absolute implied constant, and this conjecture would, of course,
follow from the validity of Zaremba's conjectUl"e. On the other hand, it is easily
seen that SN is at least of the order of magnitude log N (see Niederreiter [245]).
For sequences of values of N for which SN = O(log N), such as for the sequence
of Fibonacci numbers, we obtain two-dimensional point sets P with DN(P) =
O(N-110gN). Note that by a remark following (3.8), this is the least possible
order of magnitude for the discrepancy of a two-dimensional point set. The best
general bound on SN that is currently known is
SN = O( </>(;,) (logN) 10glog(N + 1)) = O((logN)(loglogN)2),
which was shown by LarcheI' [173].
We now return to the of an arbitrary dimension 8
2. Korobov [159] suggested the use" of special lattice points of the form
(1, g, g2, ... ,gS-1) E ,zs to restrict the number of candidates that must be in-
spected in the search for good lattice points modN. It turns out that, at least
for prime moduli N, it is possible to prove existence theorems for these special
lattice points, which are basically of the same quality as the existence theorems
shown earlier in this section. We include the proof of the following theorem as
an example.
THEOREM 5.18. For any integer s 2 and any prime N, put
N-1
ms(N) = L R(1,g,g2, ... ,gs-1), N).
9=0
Then we have
LATTICE RULES FOR NUMERICAL INTEGRATION 125
Proof. Inserting the definition of R(g, N) into
interchanging the order of summation, we obtain
expression for ms(N) and
1
m.(N) \" N L a(h)r(h)-l,
heC;(N)
where a(h) is the number of integers g with 0:$ g:$ N -1 and
h (l,g,g2, ... ,g8-1) = 0 modN.
If we put h = (h
ll
.. , ,ha), then the last congruence can be written as
8
I: hig
i
-
1
= 0 mod N.
i=l
For h E C: (N), this is a nonzero polynomial congruence of degree :$ 8 - 1 in the
unknown g, and, since N is a prime, we conclude that a(h) :$ 8 - L Hence we
m$(N) :$ 8 -; 1 L r(h)-l < 8 -; 1 ( L r(h)-l) s
heC;(N) heC(N)
8-1( )$ 8-1
= J:l L Ihl-
1
+ 1 < ]V (2 log N + 1)B,
hEC"(N)
o
It follows from Theorem 5.18 that, for any dimension s ;::: 2 and any prime
N, there exists a lattice point g E ZS of the form g = (1,g,g2, ... ,gs-l) for
which
8-1
R(g,N) < !il(2 log N + 1)8,
We also have results guaranteeing the existence of lattice points g of this form
for which Pog, N) is small and peg, N) is large, provided that N is a prime;
compare with Hua and Wang [145, Chap. 7]. For 8 = 2, the earlier existence
theorems in this section yield good lattice points g mod N of the form g = (1, g)
for any integer N ;::: 2. For 8 = 3 and any prime power N, a result showing the
existence of small values of R(g, N) for g = (1, g, g2) was proved by Larcher [171],
and a result showing the existence of large values of p(g, N) for g = (1, g, g2)
was established by Larcher and Niederreiter [179].
5.S. General lattice rules and their classification.
For the point set XO,Xl,.,. ,XN-l in (5.1), consider the corresponding residue
classes Xn +Zs = (nj N)g+Z8, 0 :$ n ~ N -1, in the additive group IRs jZB. These
residue classes form a finite cyclic subgroup of IRs jZB generated by (ljN)g+ ZS.
If the point set in (5.1) is viewed in this way, then the following generalization
is obvious. Let L jZS be any finite subgroup of IRB jZB and let Xn + Zs with
Xn E [8 for 0 ~ n ~ N - 1 be the distinct residue classes making up the group
126 CHAPTER 5
Lj'ZB. The s-dimensional lattice rule L is then given by the quasi-Monte Carlo
approximation
(5.40)
The set {xo,Xl,'" ,XN-l} is called the node set ofthe lattice rule L. If we want
to emphasize that the node set has cardinality N, then we refer to an N -point
lattice rule.
The name "lattice rule" stems from a geometric interpretation of the above
approach. If we view L = + .2;B) as a subset of ]RB, then L is an
s-dimensionallattice. Here an s-dimensional lattice is meant to be a discrete ad-
ditive subgroup of RS not contained in any proper linear subspace of IRs. Equiva-
lently, an s-dimensionallattke is obtained by taking s linearly independent vee"-
tors Yl,'" ,Ys E :IRs and forming the set L = {E:=l aiYi : ai E Z for 1 ::; i s}
of all Z-linear combinations. The lattices corresponding to la.ttice rules have the
property that they contain ZS; a lattice containing .2;s is called an s-dimensional
integmtion lattice. If we start from an a-dimensional integration lattice L, then
the corresponding node set is determined as the intersection L n P, which is
always a finite set.
REMARK 5.19. The point set in (5.1) leads to a lattice rule, but it need not
be an N-point lattice rule because of a possible repetition of points. However,
it is an easy exercise to show that this point set yields an N-point lattice rule if
and only if g = (gl, ... ,g8) satisfies gcd(gl, ... ,9s, N) = 1.
The error in the approximation (5.40) can be analyzed in the same way as for
the method of good lattice points. The following definition and the subsequent
lemma are basic. Definition 5.22 is an analogue of Definition 5.2.
DEFINITION 5.20. The dual lattice Ll.. of the. s-dimensional integration
tice L is defined by
L1.. = {h E 7/ : h x E Z for all x E L}.
LEMMA 5.21. Let XQ,XI, ... ,XN-l be the nodes oj an s-dimensional N-
point lattice rule L, and let hE ZS. Then
N-l {N
L e(hXn) =
. 0
n=O
if hELl..,
if h , L1...
Proof. IT we put A = L/Z
B
and Xh{X + ZS) = e(b . x) for all x E L, then Xb
is a well-defined character of the additive group A. We have
N-l
L e(b Xn) = L Xh(a),
n=O
LATTICE RULES FOR NUMERICAL INTEGRATION 127
and the last sum, being a character sum. for the finite abelian group A, is equal.
to N if Xh is trivial. and equal. to zero if )(h is nontrivial. Furthermor8j Xh is
trivial. if and only if h . x E Z for all x E L, i.e., if and only ,if h E LJ...., 0
" I
DEFINITION 5.22. For a real number a > 1 and for a lattice rule L, we
Pa(L) = Lr(b)-a,
h
where the sum is over aU nonzero h E LJ....
THEOREM 5.23. For any real numbers Q > 1 and C > 0 and for an s-
dimensional lattice rule L with node set {Xo,Xl1'" ,XN-l}, we have
I
N-l I
max N
l
L f(Xn) - f feu) du = CPa(L).
/Ee:.(C) n=O Ji
s
Proof. Any f E is represented by its absolutely convergent Fourier
series
feu) = L j(h)e(b. u) for u E ]R8.
hEZS
Then
1 N-l f 1 N-l
N L f(xn) - irs feu) du = N L L I(h)e(h. Xn) - 1(0)
n=O I n=O hEZs
N-l
= L j(b.) L e(hxn ) = 2: j(h)
hEZ n=O hEL.L
h#O
by Lemma 5.21. The bound
I !(x.) - h. f(U)dul " CP.(L)
follows now from Definitions 5.1 and 5.22. The proof is completed by using the
special function fo E as in the proof of Theorem 5.3. 0
Theorem 5.23, which is due to Sloan and Kachoyan [316], shows that, for
given a, the integration lattice L should be chosen in such a way that Pa{L) is
smalL The following definition is analogous to Definition 5.4.
DEFINITION 5.24. For an s-dimensional lV-point lattice rule L with s 2
and N 2, we put
R(L) = 2: r(h)-I,
hEE(L)
where E(L) = C;(N) n LJ....
128 CHAPTER 5
REMARK 5.25. By the proof of Lemma. 5.21, the dual group (or character
group) of LIZs is 'ZlIL1... Thus, since LIZs has order N, so has ZSIL1... It
follows that Nh E L1.. for all h E Z"; hence L1.. contains (NZ)". Furthermore,
L1.. I(NZ)S has order N8 IN = Ns-l. Since C
8
(N) is a complete system of
representatives for ZS I(NZ)'\ the set CB(N) n L1.. has Ns-l elements, and so
card(E(L)) . Ns-l - 1. In particular, if s ~ 2 and N ~ 2, then .E(L) is
nonempty.
THEOREM 5.26. Let L be an s-dimensio.nal N-point lattice rule with 8 ~ 2
and N ~ 2. Then, for any real Ot > 1, we have
+ ~ G) 2"('-;) a}'-; (1 + 2(aW N(l-a)(,-j) .
In particular, if either s = 2 0.1' Ot ~ 2, then
with an implied constant depending only on s and a.
Proof. Every h E ZB can be uniquely represented in the form h = k + Nm
with k E CB(N), m E Z". Since L1.. ;? (NZ)8 (see Remark 5.25), we have h E L1..
if and only if k E L1... Separating the cases where k = () and k f= 0, we obtain
(5.41) Pa{L) = L r(Nm)-a + L L r(k + Nm)-a =: 8
1
+ 8
2
.
mEZ" kEE(L) mEZ2
.mO
As in the proof of Theorem 5.5, we see that
8
2
< Ra(L) + 2as(a)" N-
as
card(C,,(N) n L1..)
,,-1
+ L2
a
("-;)((a)"-;N-
a
(s-j) L L r(kiJ-a .. r(ki;)-a,
;=1 1:5h <<i;:5" kEE(L)
where k = (k1; . ,k,,) and
Ra(L) = L r(k)-a.
kEE(L)
Now card(Cs(N) n L1..) = NIJ-1, as shown in Remark 5.25. Furthermore, for
fixed 1 ~ j :$ s - 2 and 1 ~ i
l
< ... < i j ~ 8, we have
L r(kiJ-a ... r(kij)-a
kE(L}
LATTICE RULES FOR NUMERlCAL INTEGRATION 129
where h = (hI, . .. , h
j
). A trivial
Ns-jj thus
bound for the last counting function is
~ r(kiJ-a ...
kEE(L)
For j = s - 1 and fixed 1 :$ i
1
< ' .. < i,,-l :$ 5, let i be the integer determined
by {il ! ... ,is-d U {i} = {I, 2, ... ,3}. Then
'" (k )-a (k )-a ~ r(k,t)a (N\, a R (L)
L.J r il ... r i a-l = ~ r(k)a:$ "2) a, .
kEE(L) kEE()
Altogether, we obtain
8
2
< (1 + (a)s)ROl(L) + 2as(a)S N-
as
+
a
-
1
+ I: ( ~ ) 2
01
{a-
i
)(a)8-
i
(1 + 2(a))j N(l-Ol)(ll-j).
j=l J
In view of {5.6} and Ra(L) :$ R(LY\ this yields the
first part of the theorem, and second part is an immediate consequence.
For arbitrary 8 ~ 2 and a > 1, we can derive an inequality of the type
occurring in the second part of Theorem 5.26 by the following argument. As in
the proof of Theorem .5.5, we obtain
Now
II
82 :s; 2:
kE() i=l
g(r(k;)-U + 2U(a)N-
U
) = i ~ l r(k,)-U g (1 + (a) Crt")) U)
:s; r(k)-C1 + (a))8,
and so
82 ::; (1 + (a))'" RcL).
Together with (5.6) and (5Al), this yields
(5A2) Po(L) ::; (1 +- ((a))8 Ra(L) + (1 + 2(a)N-
Cl
)8 - 1
::; (1 + (a))8 R(L)C< + (1 + 2(a)N-QyJ - 1
= (1 + (a))8 R(L)O + O(N-
a
)
with an implied constant depending only on 8 and a.
For an s-dimensional N-point lattice rule L with 8 ~ 2 and N ;:::: 2, we define
in analogy with (5.9) the quantity
(5.43) RI(L) = L r(h,N)-l.
hEE(L)
The following discrepancy bound was shown by Niederreiter and Sloan [273].
130 CHAPTER 5
THEOREM 5.27. Let P be the node set of an s-dimensional N -point lattice
rule L with s ~ 2 and N ~ 2. Then
Proof If P = {Xo,Xl,'" ,xN-d, then, from the fact that the group LjZS
has order N, it follows that NXn E ZS for 0 S n S N - 1. Thus we can apply
Theorem 3.10, which yields
where we applied Lemma 5.21 in the second step. The second inequality in the
theorem is obtained as in the proof of Theorem 5.6. 0
A fundamental classification theorem for lattice rules is due to Sloan and
Lyness [317]. This result is based on the structure theory for finite abelian
groups.
THEOREM 5.28. For every s-dimensional N -point lattice rule with N ~ 2,
the node set consists exactly of all fractional parts
with integers 0 S ki < ni for 1 SiS r,
where the integer r with 1 S r S s and the integers nI,'" ,nr ~ 2 with nHl
dividing ni for 1 sis r - 1 and nl ... nr = N are uniquely determined. FUr-
thermore, the vectors gl, ... ,gr E ZS are linearly independent, and, for each
1 SiS r, the coordinates of gi and ni are relatively prime.
Proof. Let A = LjZB be the finite abelian group corresponding to the lattice
rule L. By the primary decomposition theorem, A is the direct sum of cyclic
groups of prime-power order. Let nl be the maximum order of elements of A.
Then all orders of elements of A divide n1, and, from the primary decomposition,
we obtain a decomposition A = C
1
EEl A21 where 0 1 is a cyclic group of order
nl. If n2 is the maximum order of elements of A2 , then n2 divides nI, and A2
has a cyclic direct summand of order n2. Continuing in this way, we obtain a
decomposition A = 0 1 EEl EEl Cr, where Ci is a cyclic group of order ni ~ 2
for 1 SiS T, and nHl divides ni for 1 'S i S r - L A comparison of orders
yields N = nl'" nr . The number T and the orders nl, ... ,n
r
of the direct
summands in this decomposition are uniquely determined by the orders of the
direct summands in the primary decomposition, and the latter orders are, in
turn, uniquely determfued by A.
For 1 S i 'S T, let Ci E :IRs be such that Ci + ,Zs is a generator of C
i
. Then
niCi E ZS; thus Ci\= (l/ni)gi for some gi E ZS. Since Ci +,Zs has order ni,
LATTICE RULES FOR NUMERICAL INTEGRATION 131
the coordinates of gi and ni are relatively prime (compare with 5.19).
From A = 01 EB ... EB an it follows that the node set of L is as indicated in
the theorem. Ifcl,"" Cr were linearly dependent, then 0 could be written
88 a nontrivial linear combination of CI, . ,Cr with rational coefficients. By
clearing denominators, we E;=l iiCi = 0 with integers iI, ... ,j,. satisfying
gcd(iI, ... ,jf') = 1. This yields the identity .
r
Lidci +Z8) = zs
i=l
in the group A, and A = C1 EB EB Or implies that ni divides ji for 1 i r.
Then, however, nf' ;::: 2 divides iI, ... ,i,., a contradiction. Thus Cl, ,c,.! and
80 gl, ... ,g,., are linearly independent. In particular, it follows that r s. 0
DEFINITION 5.29. The integer r in Theorem 5.28 is called the rank of the
lattice rule and the integers 'Ill, . ,n,. in Theorem 5.28 are called the invariants
of the lattice rule.
If we choose a node set (5.1) with g - (gIl'" ,g8) satisfying
gcd(g1!'" ,ga,N) = 1, then we obtain a lattice rule of rank 1 (compare with
Remark 5.19). If we consider the Cartesian product of 8 trapezoidal rules in (1.2)
for periodic integrands, then we get a lattice rule with node set
This lattice rule has rank s and invariants ni = m for 1 i :::; s.
For every s-dimensional lattice, there exists a lattice basis, i.e., a set
{b
1
, ... ,b
a
} of linearly independent vectors such that the lattice consists ex-
actly of all Z-linear combinations of bI, ... ,b
s
. The s x s matrix B with rows
bI, ... ,b
s
is called a generator matN of the lattice. A lattice has more than
one generator matrix, since any matrix of the form U B with U being an s x s
unimodular matrix with integer entries is also a generator matrix. However,
for any lattice L, the absolute value I det{B) I of the determinant of a generator
matrix B of L is invariant, and this number is called the determinant of L and
is denoted by det(L). Geometrically, det(L) is the volume of the parallelepiped
spanned by the vectors in a lattice basis of L. The following result is due to
Sloan and Kachoyan [316].
THEOREM 5.30. If L is an integration lattice yielding an N -point lattice rule,
then det(L) = l/N. If B is a generator matrix of L, then (BT)-l ,is a generator
matrix of the dual lattice Ll. and det(L..L) = N.
Proof. If the dimension of Lis s, then, for every positive integer k, the cube
[0, k)8 contains exactly k
S
N points of L. On the other hand, the geometric
interpretation of det(L) shows that, as k -+ 00, the number of points of L in
[0, k)S is asymptotically equal to k
8
/ det(L). A comparison implies that det(L) =
liN.
132 CHAPTER 5
If bI, ... ,b" are the rows of Band al, ... ,a" are the rows of (BT)-l, then
8.i . h
j
= 6
ij
for 1 i, j s. Thus 8.i E L1.. for 1 i B by Definition 5.20. If
h E L1.. is arbitrary, then
"
h = hBT(BT)-l = I)h. b
i
)8.i.
i=l
Since h b
i
E Z for 1 i s, h is a Z-linear combination of 81, ... ,a". Hence
these vectors form a lattice basis of L1.. and (BT)-l is a generator matrix of L1...
Furthermore,
D
If L is an integration lattice, then any generator matrix of L 1. has integer en-
tries. It follows from the theory of integer matrices (see Newman [216, Chap. 21)
that LJ.. has a unique generator matrix in Hermite normal form, i.e., a lower
triangular matrix with positive diagonal elements and with the elements of any
column lying in the least residue system modulo the diagonal element. Conse-
quently, the number of distinct s-dimensional N-point lattice rules is equal to
the number vs(N) of s x s integer matrices in Hermite normal form with deter-
minant N. By [216, p. 20], vs(N) is a multiplicative arithmetic function of N
and is thus determined by its values at prime powers pk, which are given by
s-1 k+i 1
k rrp -
v,,(p ) = i _ 1 .
i=l P
Compare also with Lyness and S0revik [200] for this formula.
For an s-dimensionallattice rule L, the d-dimensional projection "Ira : ]RB _
R
a
with 1 d 8 defined by
1I"d(UI,'" ,Us) = (Ub" ,Ud)
induces a group homomorphism CPd : L/Z8 _ lRd jZd given by
CPd(X +- ZS) = 1I"d{X) + za . for all x E L.
The image of CPa is a finite subgroup of lRd /za and hence corresponds to a
d-dimensional lattice rule La.. The nodes of La are obtained by omitting the
last s - d coordinates of each node of L and by dropping repeated nodes. Since
LalZd is isomorphic to a factor group of L/Z8, it follows that, if L has rank r and
invariants nl,'" ,n"" then La has rank t min(r, d) and invariants ml, ... ,mt
with mi dividing ni for 1 i t (see Sloan and Lyness [317, Thm. 5.1]). In
particular, if No. is the cardinality of the node set of La" then
Nd ; mi :.';'fYit nl'" nt
The lattice rule Lis called projection regular if Nd = nl'" nmin(r,a) for 1 d s.
An obviously equivalent condition is Na = nl'" nd for 1 d r. A de-
tailed study of prGjection-regu!ar lattice rules was carried out by Sloan and
LATTICE RULES FOR NUMERICAL INTEGRATION 133
Lyness [318]. One of the main results of this paper is the following character-
ization of projection-regular lattice rules: L is projection regular if and only if
the vectors gi = (g?) 9 , g ~ " , 1 ~ i ~ r, in Theorem 5.28 can be chosen in
such a way that g ~ i) = 1 for 1 ~ i ~ r and g[j) = 0 for 1 $; j <' i ~ r. Further-
more, projection-regular lattice rules can be expressed in a uniquely determined
canonical form.
DEFINITION 5.31. For any s-dimensional N-point lattice rule L with s ~ 2
and N ~ 2, the figure of merit peL) is defined by
peL) = minr(h),
h
where the minimum is extended over all nonzero h E L1-.
LEMMA 5.32. We always have 1 :5; peL) ::; nl1 where nl is the first invariant
oIL.
Proof. Since the invariants n2, . .. ,n
r
are divisors of nl, it follows from the
description of the node set of L in Theorem 5.28 tha.t the coordinates of all
points of L are rationals with denominator '11.1. Therefore Ll COiltaillB (n1Z)"'.
In particular, we have he = (nil 0, ... J 0) E L1-; hence p(L) ::; r{ho) = nl. 0
REMARK 5.33. If L has ran...lt ~ 2, then it follows from Lemma 5.32 that
peL) < N/2, and, if L has rank 1, then the same inequality is obtained from
Lemma 5.8. Thus it can be shown as in Remark 5.9 that it suffices to extend
the minimum in Definition 5.31 over all h E E(L), where E(L) is as in Defini-
tion 5.24.
The quantities Pa(L) and R(L) can be bounded in terIllil of the figure of
merit p(L). The lower bound
(5.44)
is trivial, and the bound
(5.45)
for all a > 1
1
R(L) ~ p(L)
follows from Remark 5.33. The corresponding upper bounds were established by
Sloan and Kachoyan [316] and Niederreiter and Sloan [273], respectively.
THEOREM 5.34. For any s-dimensional N -point lattice rule L with s ~ 2
and N ~ 2, we have
Pa(L) ~ c(s, a)p(L)-cr (1 + logp(LB-l for all a> 1,
where the constant c(s, a) depends only on sand a.
Proof. Put 10 = (-00,0], h = (0,00), and, for d = (d
1
, ... ,dB) E {O, 1 p,
define
134 CHAPTER 5
Then we can write
Pa(L) = L S(d)
d
with SCd) = L r(h)-a.
hEQ(d)
We now consider SCd) for fixed d. Let J- be the largest integer with 21' < peL);
we can assume that J- 0, since the case where peL) = 1 is trivial. For q =
(q2,'" ,qs) E zs-l with qi 1 for 2 :s; i :s; s, we define .
M(q) = {(h
2
, ,hs) E zS-l : 2
qi
-
1
:s; r(h
i
) < 2
q
for 2 :s; i :s; s}.
Then we may write Sed) as the sum over all permissible q of
S(d,q) =
hEQ(d)
(h2,'" ,h,,)EM(q)
Case 1. q2 + ... + qs :s; J- + s - 1. Put
2-tJ2-oo-qs+s-lp(L) > 1,
K(q, b) = {h E Z : :s; r(h) < (b + for integers b O.
We claim that, if q = (q2,'" ,qs) belonging to Case 1 and an integer b 0 are
given, then there exists at most one h = (hI"" , hs ) E Q(d) such that hI E
K(q,b) and (h2 , ,hs) E M(q). Suppose that h' = (hi, ... =J:. h are two
points satisfying all these conditions. :s; r(ht}, r(hD < (b +
and hb hi E 1d1 i hence r'(h
i
- hi) < 'For 2 :s; i :s; s, we have 2
q
,-1 :s; r(h
i
),
r(hD < 2
q
and E 1d.,i hence r(hi - hD :s;}q.-I. Therefore
S
r(h - h') = II r(h
i
- hD < = peL).
i=l
On the other hand, h - hi E LJ.. and h -h' =I- 0; thus r(h - h') peL) by
Definition 5.31. This contradiction proves the claim.
Consider the contribution to Sed, q) arising from those h = (hI, ... ,hs) E
Q(d) with (h2 ! ,hs) E M(q) for which hI E K(q, b). We trivially have
r(h) peL) if b = 0, and, for b 1, we obtain
r(h) .. +q,,-8+l = bp(L).
Summing over b, we obtain
S(d,q) $ p(L)-a(l + = (1 H(ap(L)-,
Since there are choices for q ht Case 1, it follows that the sum Sled) of
those S( d, q) for which q belongs to Case 1 satisfies
(5.46) Sled) ::; (1 +0)) p{L)-a 1)
< CI(S, o)p(L)-a (1 + 10gp(L))S-1 .
LATTICE RULES FOR NUMERlCAL INTEGRATION 135
Case 2. q2 + ... + qs > I-' + s - 1. Choose integers V2, ,Vs with 0 Vi < qi
for 2 i sand V2 + ... + Vs = p,. For every a = (a2,'" ,as) E zS-l with
o ai < 2
Q
,-II,-1 for 2 S; i S; 8, let
M(q,a) = {(hz" .. ,hs) E za-I : 2(1-1 + ai211, r(h
i
} < 2(10'--1 + (aj + 1)211.

Then M(q) defined above is the disjoint union of the sets M(q, a). Put
K(b) = {h E Z :2b r(h) 2b+ 1} for integers b O.
We claim that, if q = (q2,'" ,qs) belonging to Case 2, a = (a2,'" ,as) satisfying
the restrictions above, and an integer b 0 are given, then there exists at most
one h " (h
ll
... ,h
ll
) E Q(d) such that hI E K(b) and (h
21
... ,hs) E 2kf(q, a).
Suppose that h' = . .. =f h are two points satisfying all these conditions.
,E n j hf'!nce ) = 1., and, for 2 <:::: iss, we have
r(h
i
- hD 2
11
,. Therefore
8
r(h - hi) = II r(h
i
- hD 2
112
+"'+1
8
= < peL).
i=1
On the other hand, h - h' E L.1. and h - h' =f OJ thus r(h - h') p(L) by
Definition 5.31. This contradiction proves the claim.
Since there are n:=22q-II;-1 = 2Q2+"+qs-fA-s+l choices for tt, it follows
that, if b is given as above, then there are at most points h =
(hI, ... ,hs) E Q(d) such that hI E K(b) and (h
2
, . ,hs) E M(q). Therefore
S(d, q) (1 + 2-
Ot
f.: b-
a
)
. b=l
= 2-
Ot
J.t2(1-a)(q2+ .. +qs-J.t-
s
+l) (1 + 2-Ct(a))
(2Ot + (a))p(L)-Ot2(1-0t)(q2+"+Qa-J.t-S+l) ,
where we used 2: peL) in the last step. If S2(d) is the sum of those S(d,q)
for which q belongs to Case 2, then
00
:::; (2Ot + (a)p(L)-Ot(J.L + 1)8-2 L 2(1-Ot)k(k + 1)8-2
k=l
:::; c2(s,a)p(L)-Ot (1 + logp(L)S-2.
136 CHAPTER 5
Together with (5.46) and Sed) = Sled) + S2(d), this yields
Sed) ::; cg(s, a)p(L)-a (1 + 10gp(L)t-
1
.
Since there are 2
8
choices for d, the desired result follows. 0
THEOREM 5.35. For any s-dimensional N -point lattice rule L with s ~ 2
and N ~ 2, we have
R(L) < p ( ~ ) CO!2) a-\2(lOgN)8 + 3(logN)B-l).
Proof Put 10 = (-N/2, 0], II = (O,N/2], and, for d = (d
l
, ... ,d
s
) E
{O, l}s, define
Then we can vnite
R(L) = LS(d)
d
with Sed) = L r(b)-l.
heQ(d)
We now consider Sed) for fixed d. Let w be the smallest integer with ~ > N/2
and let Jl be the largest integer with 2#' < peL); we can assume that Jl ;?:: 0, since
the case where p( L) = 1 can e?8ily be dealt with by using a trivial bound on
R(L). For q = (CJ2,'" ,qB) E Z ~ - l with 1 ::; CJi ::; w for 1 ::; i ::; 8, let M(q) be
as in the proof of Theorem 5.34 and define
S(d,q) =
hEQ(d)
(h2 ... ,hs)eM(q)
We distinguish two cases 88 in the proof of Theorem 5.34. In Case 1, we use the
same argument as in that proof, but now it suffices to consider integers b with
0::; b < IN/2J. This yields
1 ( LN/2
j
-1
1
) 1
(5.47) S(d,q)::; peL) 1 + tr b < peL) (1 +logN),
where we used [224, Lemma 3.7] in the last step. In Case 2, it suffices to consider
integers b with 0 ::; b ::; LN/4J, but otherwise the argument in the proof of
Theorem 5.34 can again be applied. This yields
LATTICE RULES FOR NUMERlCAL INTEGRATION 137
where the last inequality is valid as long as N =F 4. By (5.47) this bound can
also be used Case 1. Since there are w
s
-
1
choices for q, we obtain
j
< p(i) + log N) ,
and there are 2
8
possibilities for d; hence
1
R(L) < peL) (2w)s-1(3 + 21ogN),
provided tha.t N =F 4. Since the definition of w implies tha.t 2
W
N, we obtain
the desired bound on R(L) for N f:. 4. For N = 4, we 1100 the trivial bound
R(L) 2: r(h)-l = (t r(h)-l)' -1 = G), -1.
hEC;(4) h=-l
Since peL) 2 by Remark 5.33, it is easily seen that R(L) again satisfies the
inequality the theorem. 0 .
We can combine Theorems 5.27 and 5.35 to obtain an upper bound for the
discrepancy of the node set P of an a-dimensional N-point lattice rule L with
s 2 and N 2, namely,
There is also a lower bound for DN(P) in terms of the figure of merit, which was
shown by NiedeITeiter and. Sloan [273].
LEMMA 5.36. Let to, tit ... ,tN-l E IRs and suppose that there exist an h =
(hI,'" ,ha) E ZS with E:=llhil 2 and a f) E [0,1) such that {h t n} = f)
for 0 n N - 1. Then the point set P consisting of the fractional parts
{to}, {tl},'" ,{tN-il satisfies
where m is the number of nonzero coordinates of h.
Proof. Consider the interval J = n:=l Ji fB with
= (:hi'
if hi > 0,
Ji = (0,1) if hi = 0,
(
8+1 e )
Ji = 1+-
h
-,1+-
h
mi mi
if hi < O.
138 CHAPTER 5
8 8+1
- <hiti < -- if hi > 0,
m m
hiti = 0 if hi = 0,
() 8+1
hi + - <hiti < hi + -- if hi < 0,
m m
and so
8 8
L hi + (} < h ' t < I: hi + e + L
Thus {h . t} =f. (} for t E J; hence no point {t
n
} is in J. Consequently,
o
THEOREM 5.37. The discrepancy of the node set P of an s-dimensional N-
point lattice rule L with s 2 and N 2 satisfies
with C2 = 4, Ca = 27, and C
8
= + 1)8 - 1) for s 4.
Proof. By Definition 5.31, exists a nonzero h = (h
b
.,. ,hs) E LJ.. with
r(h) = peL). If P = {Xo,XI,. then hXn E Z for 0:$ n:$ N -1, and
so an application of Theorem 3.16 with tn = Xn for 0 :$ n N - 1 and (} = 0
yields.
2
1 - ((11" + 1)8 - 1) p(L)DN{P).
7r
This is the result for s 4. If E:=llhi ! 2, then we apply Lemma 5.36
with tn = Xn for 0 5:,n'S:N-1 and (} = 0 to obtain
(5.48)
1
DN(P) s8p(L)'
If :L:=1 Ihil = 1, then, for some i with 1 :$ i :$ s, we have hi = 1 and hj = 0
for all j =f. i. From h . Xn E Z it follows then that the ith coordinate of each
Xn is O. This yields DN(P) = 1, and so (5.48) holds trivially. From (5.48), we
obtain the desired result for s = 2 and s = 3. 0
5.4. Existence theorems for efficient lattice rules.
Let L be _an s-dimensional N -point lattice rule with s 2 and N 2, and let
r be its raD.k and nl, .. , ,nr its invariants. The following result from Nieder-
reiter [262] is an analogue of Lemma 5.32 and shows that the first invariant nl
must be large for ad efficient lattice rule.
LATTICE RULES FOR NUMERICAL INTEGRATION 139
THEOREM 5.38. The following bounds are valid:
(i) pateL) (1 + 2((a)n
I
at
r -1 > 2s(a)
nI
at for all a> 1;
(ii) R(L) c(s)nlllog(N/nl)' where c(s) > 0 depends only on s;
(iii) Div(P) 1- (1- nil)" nIl, where P is the node set of L.
Proof. (i) Since Ll.. ;? (nIZ)/l by the proof of Lemma 5.32, we see that, for
a> 1,
Pat(L) L r(h)-at - 1 = (.E r(n1h)-at) /l - 1
hE(nlZ)8 hEZ
= (1 + 2 t,(n1h)-')' -1 = (1 + 2(<nj')' - 1 > 2s(a)n,'.
(ii) From Ll.. 2 ('11IZ)B, we obtain E(L) ;? C;(N) n ('11lZ)B in the notation
of Definition 5.24. The elements of this intersection are exactly all points 'T11Jl
with h E C;(N/'11l). Therefore
R(L) .E r(n1h)-1 - 1 = (. L r(nIh)-I) 8 -1
hECII(N/nt} hEC(N/nd
= (1 + '1111 L Ih
l
-
I
) 8 - 1 c(s)'11
1
1
1og(N/'11I)'
\ hEC*(N/nt}
(iii) By the proof of Lemma 5.32, the coordinates of all points of P are
rationals with denominator '111. Thus the desired result follows from Theorem
3.14. 0
Existence theorems for efficient lattice rules of rank 1 have already been
shown in 5.2. These theorems can be used to obtain results for higher ranks by
a method of extending lattice rules.
LEMMA 5.39. For s 2, let a rank r with 1 r s and invariants
nl, ... ,n". 2 with niH dividing ni for 1 :::; i :::; r - 1 be given, and let Ll be an
a-dimensional nl-point lattice role of rank 1 generated by gl = (g11), ... ,gi
s
) E
zs with gcd(g?), nl) = 1. Then there exists an s-dimensional lattice rule L with
rank r and invariants nl, ... j nr such that the node set of L contains the node
set of L
1
.
Proof. We may assume that r 2. We construct L by putting
gi = ... E ZS for 2 i r
with = 0 for 1 j i-I and ni) = 1 and letting the nodes of L
be all fractional parts
{
". k }
.E
i=l
with 0 :::; k
i
< 'T1i for 1 i r.
140 CHAPTER 5
Consider the mapping 'I/J on the finite abelian group A = (Z/n1Z)EB'" EB(Z/nrZ)
defined by
Then 'I/J is a surjective group homomorphism. We claim that the kernel of 'I/J is
trivial, i.e., that
(5.49)
implies that k
i
- 0 mod ni for 1 $ i $ r. By comparing the first coordi-
nates in (5.49), we derive = 0 mod 1; k1gi
1
) == 0 mod nl,
and thus kl = 0 mod nl since gcd(gil) j nl) = 1. Thus (5.49) reduces to
= 0 modZ
s
, and, comparing second coordinates, we obtain
k2 = 0 mod n2' By continuing in this manner, we establish the claim. Conse-
quently, L/,L,8 is isomorphic to A, and so L has rank r and invariants nl, ... ,n
r

Clearly, the node set of L contains the node set of L
1
. 0
LEMMA 5.40. For s ;::: 2, let L1 L2 be two s-dimensional integration lat-
tices, and let N1 and N2 be the number of nodes of the corresponding lattice roles
Ll and L
2
, respectively, with NI ;::: 2. Then
if L1 is a lattice role of rank 1 generated by a point g E G
s
(N
1
).
Proof. (i) From L1 L2, we obtain Lf Lt, and so the desired inequality
follows immediately from the definition of the figure of merit.
(ii) Use Lt ;2 Lt and the definitions of Pa(L
I
) and P
a
(L
2
).
(ill) N-ote that NI divides' N2l'since Ld'Zl is a subgroup of order NI of
the group L
2
/Z
B
of order N2 Every h E Cs(N.},) can be written in the form
.h = k + N
1
m with k E eB(Nl ), m E Mil := [-N2 /(2N1), N2/(2Nt}]11 n ZII.
LATTICE RULES FOR NUMERlCAL INTEGRATION
Since N1m E Lt by Remark 5.25, we have h E Lt if and only if kELt. Also
using Lt ;? we see that
(5.50) R(L
2
) $
r(h)-1 - 1
hEC8 (N:a) nLt
$ L: r(Nlm)-l- 1 + L: L r(k + N
1
m)-1
mEM. keE(Ll) meMs
with the notation in Definition 5.24. Now
(5.51)
Furthermore)
T:= . L L r(k+N1m)-1 = L iI( L r(ki + N1m)-1),
kEE(Lt} mEM. kEE(Lt} ;=1
where k = (k
I
, ... ,ka). By proceeding as in the argument following (5.7), we
obtain
( ) 1 ) 1 ,2 eN'}.
L..J rk+N1m- $r(k- +N1ogN
. 1 1
for k E C(Nd,
and so
For fixed 1 :::; j $ s - 1 and 1 $ i
l
< ... < i
j
$ s and any k E E(L
l
), we have
and so
142 CHAPTER 5
Since card(E(Ld) = N:-
l
- 1 by Remark 5.25, it follows from (5.52) that
8 () ( N ) 8-j 28 ( N ) 8
T < R(Ld ~ ; 1 + log N: + Nl 1 + log N: '
and, together with (5.50) and (5.51), this yields (iii).
(iv) If Ll is as in (iv), then, for fixed 1 ::; j ::; s -1 and 1 ::; i
1
< ... < i
j
::; s,
we obtain, as in the proof of Theorem 5.5,
L r(k
i
J-
1
... r(k
ij
)-1 ::; N:-j-l L r(h)-l < N:-j-l(2 log Nl + l)j.
kEE(L1 ) hECj(N1 )
Using this in (5.52), we derive
2
8
( N ) 8 1 8-1 ( ) ( N ) s-j
T < R(Ll ) + N 1 + log N
2
+ N L ~ 2 log N
2
+ 2 (2 log Nl + 1)j
1 1 1 j=1 J 1
1
= R(L1 ) + Nl ((2 log N2 +3)8 - (210gNl + 1)8),
and, together with (5.50) and (5.51), this yields (iv). 0
THEOREM 5.41. For s ;:::: 2, let a rank r with 1 ::; r < s and invariants
nl, . . . ,n
r
;:::: 2 with niH dividing ni for 1 ::; i ::; r - 1 be given. Then, for
every a> 1, there exists an s-dimensionallattice rule L with this rank and these
invariants such that
where the implied constant depends only on s and a.
Proof. By Theorem 5.13, there exists a gl E G
8
(nl) such that P
a
(gl, nd
satisfies the bound in Theorem 5.13 with N = nl. Let LI be the corresponding
s-dimensional nl-point lattice rule of rank 1 and let L be as in Lemma 5.39.
Then Ll ~ L, and so Pa(L) ::; Pa(L1) = Pa(gI, nl) by part (ii) of Lemma 5.40.
By the construction of gl, we get the desired result. 0
THEOREM 5.42. For s ;:::: 2, let a rank r and invariants nl, ... ,n
r
be given
as in Theorem 5.41. Then, for every a> 1, there exists an s-dimensionallattice
rule L with this rank and these invariants such that
where r(nl) is the number of positive divisors of nl and where the implied con-
stants depend onlywn s and a.
LATTICE RULES FOR NUMERICAL INTEGFLl\TION 143
Proof. Proceed as in the proof of Theorem. 5.41, but UB Theorem 5.14 :instead
of Theorem 5.13. 0
The remarks following Theorem 5.14 apply, with the necessary changes hav-
ing been made, to Theorems 5.41 and 5.42 as well. An.lilogml existence theorems
for large values of p( L) and small values of R( L) fOT lattice ,["llles L ",lith prescribed
rank and invariants can be deduced from parts (i), (iii), and (iv) of Lemma 5.40
and from the corresponding existence theorems concerning peg, N) and R(g, N)
in 5.2.
Theorems 5.41 and 5.42 have been obtained by a method that is in a sense
indirect, since, in the proof, we first consider efficient lattice rules of rank l.
There is also a direct method of establishing existence theorems for efficient
lattice rules in which we immediately look at lattice rules of the desired rank.
For a dimension 8 ;:::: 2, let a rank 'I' with 1 S; 'I' S; s invariants nl! ... , nr ;:::: 2
with niH dividing ni for 1 S; i S; 'I' - 1 be We put
Zi = {g E Z : 0 S; 9 < ni and gcd(g, ni) = 1} for 1 S; i S; 1'.
Let C = (8; nl, ... ,nr). be the family of all a-dimensional lattice rules L for
the node set COlilSlits
with integers 0 S; k
i
< ni for 1 S; i S; '1',
where the gi have the special form gi = (gil), .. . ,gi
s
)) with gii) = 0 for 1 j
i-I and E Zi for i S; j s. It follows from the proof of Lemma 5.39 that
each L E has rank 'I' and invariants nl, ... ,n,.. Let
M(C) = L R(L)
j LEe
be the avera.ge value of R(L) as L nms through C. Note that card(C) =
(ni)B-Hl. Inserting the definition of R(L) and interchanging the order of
summation, we obtain
1
M(C) = I: A(h)r(h)-l - I,
bECa(N)
where N = nl'" nr and A(h) is the number of L E C with h E Li.. -We now
invoke Lemma 5.21 to express A(h) in terms of exponential sums, and we use
the special form of the node set of each L E C to obtain
for h = (hI, ... ,hs) E es(N). Therefore
A(h) 1 nl-l n
r
-l S ( 1 min(j,r) (
-=-I:'''I:II I: - e
bEC.(N) r(h) N k
1
=O kr=O j=1 hEC(N) r(h) i=l gEZi
))
144 CHAPTER 5
The innermost sum is calculated by the method following (5.15), which yields
The resulting expression for M(.) is rather complica.ted. Even in the simplest
higher-rank case where r = 2, very ela.borate number-theoretic arguments are
needed to deal with this expression. The following result is obtained in Nieder-
reiter [259] for this case.
THEOREM 5.43. For every 8 2 2 and any prescribed invariants nt, n2 2 2
with n2 dividing n1, we have
M(l,) < C
ll
((lOg N)lI + log N)
\ N fl'l /
with a constant C
lI
depending only on 8. In particular, there exists an s-
dimensional lattice rule L of rank 2 with invariants n1 and n2 such that
R(L) < +
By combining this result with Theorem 5.27, we see that there always exists
a lattice rule of rank 2 with prescribed invariants fQr which the node set has a
discrepancy of an order of magnitude close to that of the lower bound inpart(ili)
of Theorem 5.38. Similarly, by virtue of ,Theorem 5.26 or (5.42), we get a value
of Po:(L) of an order of magnitude close to that of the lower bound in part (i) of
Theorem 5.38.
Efficient lattice rules can also be obtained by considering suitable s-
dimensional lattice rules of maximal rank r = 8, with copy roles offering the
most promising option. H L is an s-dimensional lattice rule and k 2 2 is an in-
teger, then the k
8
-copy of L is the lattice rule with integration lattice ,,-lL, Le.,
the lattice L scaled by a factor k-
1
The kll_ropy of L may also be described as
the lattice rule obtained by partitioning f8 into k
8
cubes of, side k-
1
and apply-
ing a properly scaled version of the lattice rule L to each of these smaller cubes.
Thus the use of copy rules may be viewed as a special deterministic version of
the Monte Carlo technique of stratified sampling (compare with 1.2). In the
following theorem, we collect the basic information on copy rules.
THEOREM 5.44. If L is an s-dimensional N -point lattice rule and k 2 2
is an integer, then k-
1
L is a kilN-point lattice role of rank 8 with dual lattice
(k-
1
L)..L = kL..L.
Proof. The number of nodes of k-
1
L is equal to the number of points of
the integration lattice L in (O,k)lJ, which is kIJN. Furthermore, k-
1
L contains
(k-
1
Z)3, and the group (k-
1
Z)8/ZJJ is the direCt sum of 3 cyclic groups of order
k. Hence (k-
1
Z)8 has rank 8, and so has k-
1
L by [317, Thm. 3,3]. Finally,
from the definition of the dual lattice, we immediately obtain that (k-
1
L)..L
contains kL.l... Npw ZB/(k-1L).l.. has order kilN by Remark 5.25, and, from
LATTICE RULES FOR NUMERICAL INTEGRATION 145
zs LJ.. kLJ.. and Remark 5.25, we see that ZS/(kL1.) also has order k
S
N.
Therefore (k-
1
L)1. = kLJ... 0
On the basis of the identity (k-1L)J.. = kLJ.., corresponding quantities for
k-
1
L and L can be related. FOl"instance, we have
Pa(k-1L) = Lr(kh)-a for any a> 1,
h
where the sum is over all nonzero h E L1. I and this expression is closely connected
with Pa(L). This connection can be effectively exploited in the case where L
is a lattice rule of rank 1. Disney and Sloan [68] calculated the mean value of
P
a
(k-
1
L) when L ranges over all s-dimensional N-point lattice rules of rank 1
generated by a point g E G 8 (N), where N is a prime number, and showed that
. for k = 2 this mean value is asymptotically s.ma.ller than the corresponding mean
vjUue of Pa(L) for a comparable This suggests that there are
lattice rules. of' fw;Uc yiEtld. smaller bounds. than lattice rules of
rank 1 with nodes.. ,.
of pr1:ndpa! probleT'Gs the ?xea is to find gener?,l con.structions of
efficient lattice rrues. Except for the case where 8 = 2, in which the connection
with continued fractions can be used (see 5.2), no explicit constructions of
infinite families of efficient lattice rules are known. Some attempts have been
made by Hua and Wang [144] and Zinterhof [370] for lattice rules of rank 1,
but the results fall far short of the levels that are set by the existence theorems.
Thus, at present, we still must resort to computer searches to filld efficient lattice
rules for dimensions s 3. The search for good lattice points mod N may
proceed by two strategies, namely, an elimination method in which the points
g mod N with a small value of peg, N) are systematically eliminated, and a
random search method in which points g mod N are chosen at random and a
good lattice point mod N is obtained after sufficiently many trials. Sometimes
the search is restricted to points of the special form (1, g, 9
2
, . ,g8-1) proposed
by Korobov [159]. Extensive tables of good lattice points were compiled by
Maisonneuve [203J, and these were complemented more recently by the tables in
the book of Hua and Wang [145] and in the papers of Bourdeau and Pitre [31] and
Haber [123]. The search for efficient lattice rules of higher rank is based on similar
strategies. The most ambitious search project involving higher-rank lattice rules
was undertaken by Sloan and Walsh [321] and yielded many practically useful
lattice rules of rank 2.
Notes.
The method of good lattice points was introduced by Korobov [158], and further
pioneering papers on the subject include Bakhvalov [13]' Hlawka [135], and Ko-
robov (159]. Expository accounts ofthis method can be found in the books of Ko-
robov [160J and Hua and Wang [145] and in the survey articles of Zaremba [365]
and Niederreiter [225J; see Niederreiter [251J for an update of the latter paper.
Theorem 5.5 is from Niederreiter [264]. By averaging the quantity R1(g, N)
in (5.9) over suitable sets of points g mod N and using the first inequality in
146 CHAPTER 5
Theorem 5.6, improved existence theorems for the discrepancy DN(P) of the
point set Pin (5.1) are obtained; see Niederreiter [239J for primes N and Nieder-
reiter [237] for arbitrary N 2. Algorithms for the determination of good lattice
points modulo powers of 2 were described by Korobov [161]. Computer imple-
mentations of the method of good lattice points are discussed in Genz [113] and
Kahaner [149]. A widely available software implementation is routine DOlGCF
in the NAG library. Several open problems on good lattice points are stated in
Niederreiter [233]. Further existence theorems for good lattice points of the form
(1,g,g2, ... ,g8-1) can be found in Larcher [171] and Temlyakov [343J.
A report on calculations in connection with Zaremba's conjecture is given in
Borosh and Niederreiter [30]. For theoretical results related to Zaremba's conjec-
ture, see Cusick [46], [47], [48J, Hensley [133], and Sander [300]. Cusick [49] con-
sidered a higher-dimensional analogue of Zaremba's conjecture. Temlyakov [344]
proved the optimality of the two-diniensionallattice point g = (1, F m-l) mod F m
for numerical integration a class of functions with bounded mixed deriv0tive.
The first steps in the direction of lattice rules were taken by Frolov [111].
Sloan [314] and Sloan and Kachoyan [315] again approached the subject, and a
systematic theory of lattice rules was developed by Sloan and Kachoyan [316].
For a background on lattices, we refer to the book of Cassels [38J. Lyness [198]
and Sloan and Walsh [320] gave brief surveys of lattice rules from the viewpoints
of generator matrices and classification theory, respectively. Lyness [197] dis-
cussed various ways of assessing the quality of lattice rules. Further results on
ranks and invariants can be found in Sloan and Lyness [317], and a detailed
study of lattice rUles of r8.nk2 iscarned out in LyneSs and Sloan [199]. Struc-
tural results on integration latticeS arid their dual lattices m:e established in
Lyness, Srevik, and Keast [201 J. For lattice niles of'r3.nk 1, . Theorem 5.34 was
shown by Zaremba [365], and in a 'somewhat weaker form, by
Niederieiter . [224] (see also Niederreiter [234]). The metli6d of obtaining the
lower bound for'thediscrepancy' in can also be traced to [234].
Lemmas 5.39 and 5.40 are derived from [262J. Fhrther informa-
tion on copy ruleS, in' oil their be f9und in Sloan and
Lyness [317]. Joe [148] and
Patterson [45] from lattice rules i to rules. ;Analogues of
lattice rilles for integration over JiB Sloan and Os-
born [319J, and Sugihara (336). A generalization oflattice rules for Haar integrals
over compact groups was introduced by Niooerreiter [262].
--"'.,.,
CHAPTER 6
Quasi-Monte Carlo. Methods for
Opti m ization
Another basic problem of numerical analysis to which quasi-Monte Carlo meth-
ods can be applied is glQbal optimization. The standard Monte Carlo method
for finding global 6ptnha; is random search, and it is employed in situations
where the objective function has a low degree of regularity, e.g., in the case of
a nondifferentiable objective function in which the usual gradient methods fail.
The deterministic analogue of random search is the quasi-Monte Carlo method
of quasirandom search. The analysis of quasi-Monte Carlo optimization follows
the same approach as for quasi-Monte Carlo integration: We first establish an
effective error bound in te:n:ns of a suitable quantity depending on the deter-
ministically selected points (in this case, the relevant quantity is the dispersion
rather than the discrepancy), and then we strive to find deterministic point sets
or sequences that make this quantity as small as possible.
Both random search and quasirandom search can be described in a quite
general setting, and this is done in 6.L In the standard case where the objective
function is defined on a bounded subset of a Euclidean space, more concrete
information can be given. Since a straightforward quasirandom search method
is usually inefficient, we also discuss more refined techniques, such as localization
of search. In all these variants of quasirandom search, a basic role is played by
low-dispersion point sets and sequences, which are studied in 6.2.
6.1. General theory of quasirandom search methods.
Let X be a separable tOlpological space and let the objective function f be a
real-valued function Oln X for which we want to calculate a global optimum. It
suffices, of course, to restrict the attention to the case where we are interested
in the global supremum (or maximum) of f. Thus we assume that f is bounded
from above, and we put
m(f) = sup f(x).
xEX
The Monte Carlo method of random search proceeds as follows. Put a probability
measure A on X, take a sequence S of independent A-distributed random samples
Xl, X2, . EX, and use the estimate
(6.1)
148 CHAPTER 6
The sequence mN(f; S), N = 1,2, ... is nondecreasing, and we would expect
that it converges to m(f). In fact, this can be shown to happen with probability
1 if I is continuous and the measure >. is suitable.
THEOREM 6.1. If f is continuous on X and if the measure ,\ is such that
'\(A) > 0 for every non empty open subset A of X, then
lim mN(JjS) = mU) ,\oo-a.e.
N-+oo
Proof. For given c > 0, the nonempty set Ae = {x EX: I(x) > mU) -c:} is
open by the continuity of I. Furthermore, we have mNU; S) > mU) - c if and
only if at least one of Xl, . . . ,X N lies in Ae. The probability of the latter event
is 1 - (1 - >'(AeN by the independence of the samples. Thus, if XOO denotes
the set of all sequences of elements of X, then
>. 00 ( {S E Xoo : m N (J; S) > m(f) -- c for all sufficiently large N})
2:: 1 - (1 ->'(Ae))h for h = 1,2, ....
Letting h -+ 00 and using >'(Ae) > 0, we obtain
>.
00
( {S E Xoo : mN(fj S) > m(f) - c for all sufficiently large N}) = 1.
Applying this with c = 11k, k = 1,2, ... , we get the desired result. 0
In the quasi-Monte Carlo method of quasimndom search, we use a determin-
istic sequence S of points Xl, X2, in X and employ approximation (6.1). We
clearly have
lim mN(f; S) = m(f)
N->oo
whenever f is continuous on X and the sequence S is dense in X. For the
purpose of error analysis, let (X, d) be a bounded metric space, which means
that the metric d satisfies SUP:r:,YEX d(x; y) < 00.
DEFINITION 6.2. If (X, d) is a. bounded metric space and the point set ,p
consists of Xl, . ,x N EX, then the dispersion of P in X is defined by
Note that the boundedness of (X, d) implies that the dispersion dN(Pj X) is
finite. If R(x; r) denotes the closed ball with center x E X and radius T', then
dN(P; X) may also be described as the infimum of all radii r 2:: 0 such that the
balls B(xljr), ... ,B(xNjr) cover X. If P is as in Definition 6.2, then we also
write mN(fi P) for the maximum in (6.1). Let
W(Ji t) = sup I/(x) - f(y) I for t 2:: 0
x,yEX
d(x,y):5t
be the modulus of fOntinuity of f.
QUASI-MONTE CARLO METHODS FOR OPTIMIZATION 149
THEOREM 6.3. If (X, d) is a bounded metric space, t h e n ~ jor any point set
P of N points in X with dispersion dN = dN(Pj X), we'l..ave
m(f) - mN(j;P):5 W(j;dN)'
Proof. Choose e > 0 and let 11 E X be such that fey) > m(J) - e. For some
k with 1 :5 k:5 N, we have
dey, Xk) = min dey, Xn ),
l ~ n ~ N
where Xl, ,XN are the points of P. It follows that dey, XI,:) :5 dN. Further-
more, we have fey) - f(xk;) :5 w(fj dN), and so
which implies the result of the theorem. 0
Theorem 6.3 shows that suitable deterministic point sets for quasirandom
liearch :are tho.<re with small dispersioIL If S is a sequence of elements of Xl
we write dN(Sj X) for the dispersion of the first Nterms of S. For a sequence S,
the error bound W(J;dN(S;X in Theorem 6.3 tends to zero as N -+ 00 if f is
uniformly continuous on X and dN(S; X) -+ 0 as N -+ 00. The latter property
is seen to be equivalent to the denseness of S in X, provided that (X, d) is totally
bounded, which means that, for every e > 0, X can be covered by finitely many
closed balls of radius e. Thus the dispersion may be viewed as a measure for the
deviation from denseness.
To transform low-dispersion point sets and sequences from one domain to
another, the following simple result is convenient.
THEOREM 6.4. Let T : (X, d) -+ (Y, d') be a map from the bounded metric
space (X, d) onto the metric space (Y, d') such that theTe exists a constant L ;:::: 0
with
(6.2) d'(T(x), T(z :5 Ld(x, z) for all x, Z EX.
If P is the point set consisting of Xl, ... ,x N E X and pI is the point set con-
sisting ofT(xd, ... ,T(XN) E Y, then
with equality holding if 'We have equality in (6.2).
Proof. Note first that, from (6.2) and the surjectivity of
(Y, d') is also bounded. Furthermore,
it follows that
dN(P'; Y) = sup min d'(y, T(xn)) = sup min. d'(T(x), T(xn ))
yEY l::;n::;N xEX l:S;n:S;N
~ sup min Ld(x, xn) = LdN(Pj X),
xEX l ~ n ~ N
with equality throughout if we have equality in (6.2). 0
150 CHAPTER 6
In most applications, the domain of the objective function will be a bounded
subset E of a Euclidean space RS. In this case, there are two interesting metrics
on E, namely, the standard Euclidean metric d and the maximum metric d'
defined by
for U = (UI,'" ,us) and v = (Vb o ,VB) in E. We write dN(P; E) for the
dispersion of a point set P in (E, d) and (Pj E) for the dispersion in (E, d'),
and similarly for sequences. Since d'(u, v) d(u, v) Sl/2d'(u, v) for all u, v E
E, it follows from Theorem 6.4 applied to the identity map on E that
Thus the two dispersions have the same order of magnitude, but, in certain
circumstances) one '?Rsler to wor'k with. (P; E), we have the
general lower bound
(6.3)
where).s is the s-dimensional outer Lebesgue measure and "/s = 7r
s
/
2
/r(s/2+ 1 )
is the volume of the s-dimensional unit ball. This is shown by let-
ting Xl,"" XlV E E be the points of P and noting that the balls
B(XI; r); ... ,B(XN; r) with radius r = dN(PjE) cover E. Comparing mea,..
sures, we get N,sr
s
;:::: ).8(E), which leads to (6.3). The analogous bound for
d'N(Pj E) is
(6.4)
REMARK 6.5. If AB(E) > 0, then (6.3) and (6.4) show that dN(P; E) and
E) are at least of the order of magnitude N-1/B. On the other hand, we
can construct point sets for which the dispersion attains this order of magnitude.
Let E be a bounded subset of RS, so that E is contained in some cube [a, b f
For a positive integer k, we partition [a, b 1 into the intervals Ih = [a+h(b-a)/k,
a + (h + l)(b - a)/k) for 0 h k - 2 and Ik-l = [b - (b - a)/k,b]. Then
the cubes n:=l hi' with the hi, 1 i s, running independently through the
integers 0,1, ... , k - 1, form a partition of [a, b]8. For each of those cubes C
having a nonempty intersection with E, we select one point from C n E. In
this way, we arrive at a point set P consisting, say, of Xl, ... ,XN E E. If
X E E is arbitrary, then X lies in some cube Co from above having a nonempty
intersection with E, and also Xn E Co for some n vlith 1 n N. Since
(b - a)/k, we arrive at d'N(P;E) (b - a}/k, Now N k
S
, and
therefore d'N(Pj E) 5 (b - a)N-l/s.
The quasirandom search method already described is a.J.so known as crude
search. This method has a limited usefulness, since its rate of convergence is, in
general, very slow. Its inefficiency can also be seen from the fact that most of the
QUASI-MONTE CARLO METHODS FOR OPTIMIZATION 151
time is spent calculating function values at points that are far from those points
at which the function values are close to m(f). A refinement of crude searCh was
introduced by Niecierreiter and Peart [269) and is called localization of search.
To simplify the exposition,we assume that E = Is with the maximum metric
d/. For B = B(ajc) with a E Is and 6' > 0, letOB': is ....,.. B be the function
defined by 'f! .
9B(X) = a + 8(2x - (1, ... ,1)) for x E jB,
which is the canonical similarity transformation from is onto B. Now let
XI, ... ,XN be points in is, put
and define
IN(fi B) = max(J(a), max !(9B(Xn ).
neS(B)
Furthermore, let x*(B) be one of the points in the set {a}U{gB(Xn) : n E S(B)}
for which
IN(fi B) = f(x*(B.
Now choose a sequence et, e2,." of positive numbers converging monotonically
to zero and define a sequence Bo, BI, ... of baliB (in the maximum metric, Le.,
cubes in the usual geometry) as follows:
Bo = P',
and, for j = 1,2, ... J
if IN(fj Bj ) > IN(f; Bj - 1),
otherwise.
Then the sequence IN(fi B
j
), j = 0,1, ... , is nondecreasing and, for suitably
large j, the value of IN(fj B
j
) is taken as an approximation to m(f). In practice,
we will usually set ej = e
j
for j = 1,2, ... a fixed e satisfying 0 < e < !.
The point set P of initial points Xl, ... ,XN E 1
8
should then be chosen in such
a way that is) < e.
It should be noted that, contrary to what we know about the crude search
method, the convergence of the method of localization of search cannot be guar-
anteed. Problems arise when f has a local maximum very close, but not equal, to
m(f). The difficulty of proving a rigorous convergence theorem for localization
of search occurs even in the context of a stochastic model) as was noted by Solis
and Wets [330, p. 21]. It is advisable to perform the method of localization of
search several times, with different initial point sets and different values of e,
and then to take the maximum function value reached in all these calculations
as the approximate value of m(f).
It may also happen that the sequence IN(f; B
j
), j = 0,1, ... , does not con-
verge to m(f) because the sequence El, e2, . .. converges too quickly to zero.
152 CHAPTER 6
The following procedure of outer iteration may be used to countervail that ten-
dency. We first run through a small number of steps in the localization of search
method, thus obtaining a point xi at which f has the largest computed value.
Then we start a second localization of search method with a ball B(xi; hI) with
hI larger than the radius of the ball searched last. In general, having arrived
after k - 1 outer iterations at the point XA;-l' we start the kth iteration with a
ball B(Xk_1i 15k - I ) with hk - 1 by a constant factor (independent of k) larger than
the radius of the ball searched last. We refer to Niederreiter and Pe&""t [269] for
a further discussion of this procedure.
If prior information on the function f is available, we could also employ
adaptive methods to improve the performance of crude search. One idea is to
use fewer points Xn in those regions where f varies slowly and more points in
those regions where f oscillates strongly. Information on the oscillation of f can
also be obtained from an initial crude search. Another possibility is to adapt the
metric to the behavior of Fbr instance, we may use the "weighted)) maximum
metric
for u = (Ul, ... ,us) and v = (Vll." ,VB) in E, where the positive weight t;
measures the oscillation of j('Ul,' . ,UB ) as a function of Ui. If f had bounded
partial derivatives on E, we could use for Ci the supremum of la! lauil over E.
In general, these suprema will not exist or will not be available numerically, but
we may use a difference scheme based on a coarse grid to get approximate values
for the t;. See Niederreiter [235] for further comments on adaptive methods.
Low-dispersion .... ,n."" ... sets sequences.
We now normalize E to be 18. Instead of dN(p;1
8
) and we simply
write dN(P) and respectively, and similarly for sequences. The following
result provides an easy connection between the dispersion and the discrepancy,
THEOREM 6.6. For any point set P consisting of N points in 1
8
, we have
Proof Let Xl, ... ,XN be the points of P. For given e: with 0 < e < r :=
there exists an X E Is such that d'(x,x,.) > '1' - e for 1 :$ n :$ N. Thus
the ball B(xj r - e:) in the metric d' contains none of the points of P. Now
J = B(x; r - e) n Is is a dosed interval. with AB(J) ;2: (r - e:)8, and so, for the
half-open version J of J, we have
which implies the desired inequality. 0
Thus-every low-discrepancy point set (or sequence) is a low-dispersion point
set (or sequence), but not conversely. If we had a general inequality of the form
DN(P) :$ -with a function 9 satisfying limu->o+ g(u) = 0, then this
QUASI-MONTE CARLO METHODS FOR OPTIMIZATION 153
would lead to the conclusion that every dense sequence in Is is uniformly dis-
tributed in Is, which is obviously incorrect. The constructions of low-discrepancy
. point sets in Chapters 3 and 4 yield, in the a-dimensional case, point sets P with
DN(P) = O(N-l(logN)B-l), and so, from the bound in Thoorem 6.6, we obtain
d'N(P) = O(N-
1
/s(log N)(S-l)/S), where the implied constants depend only on
s. As we have seen in Remark 6.5, we can sometimes achieve a slightly smaller
order of magnitude by a direct inBpaction of the dispersion.
In the case where 8 = 1, the two metrics d and d' are identical, and so
dN(P) = d'N(P). If we order the points Xl,'" ,XN of P such that 0 ::; Xl <
X2 ::; ... ::; XN :::;; 1, then it is easily seen that we have the explicit formula
(6.5)
dN(P) = ma.x( Xl, !(X2 - xd, - X2),'" ,HXN - XN-I), 1 - XN)'
, HX
n
= (2n-l)/(2N) for 1 ::; n :::;; N, then dN(P) = 1/(2N), and (6.4) shows that
\. this is the minimum value of dN(P) for any point set P consisting on N points
,. itt 1. The corresponding question for one-dimenBional sequences is already less
Here -:;i;e first, the lower due to Niederreiter [230],
THEOREM 6.7, For any sequence S of elements of I, we have
- 1 .
lim NdN(S) -1 - = 0.721.,. ,
N->oo og4
Proof. Suppose that there exists a sequence S of elements of I with
- 1
lim NdN(S) < c= -1 4'
N->oo og
Then, for a sufficiently small c > 0, we have
(6.6)
c-c
dN(S) ---yy- for all sufficiently large N.
For such an N, we order the first N terms of S so that 0 Xl ::; X2 . , . X N
1. There are N + 1 steps Xl! Xn+l - Xn for 1 n N -1, and 1- XN, which we
denote by Uo, Ul, ,UN in such a way that Uo ;::: Ul . , . ;::: UN. We note that
dN(S) ;::: Uo/2 by (6.5). The (N + l)th term of S falls at best into the largest step;
hence dN+l(S) ;::: ul/2. Similarly, after having taken into account the (N + l)th
and (N + 2)th term, a step;::: U2 will remain; thus dN +2(S) ;::: U2/2. Continuing
in this way, we get dN+m(S) u
m
/2 for 0 ::; m N, and so, by (6,6),
U < 2(c - c) for 0 m N.
ffl_ N+m
It follows that
N N 1 ( 1 r
2N
dt)
<2(c-c) N+JN t
2(c - c)
= 1 - clog 4 +, ,
iV
which is a contradiction for sufficiently large N. 0
154 CHAPTER 6
The lower bound in Theorem 6.7 is the best possible, according to the fol-
lowing example given by Ruzsa (see Niederreiter [230]). Consider the sequence
S defined by
(6.7)
_ (10g(2n - 3) }
Xn - t log 2
for n = 2,3, ....
A straightforward argument shows that
dN(S)
__ 10gN -log(N -1)
lor N 2,
log 4
and so, for this sequence, we have
lim NdN(S) = I 1
4
,
N-+oo og
It is remarkable that this sequence S of asymptotically minimal dispersion
is not uniformly distributed in i, and so its discrepancy does not satisfy
limN->oo DN(S) = O. At any rate, the problem of constructing a sequence with
asymptotically minimal dispersion can be solved in the ontrdimensional case,
whereas the corresponding problem for the discrepancy is far from being settled
(compare with 3.1). This suggests that questions of irregularities of distribution
are easier to handle for the dispersion than for the discrepancy.
In the multidimensional case, we can determine for every fixed N the mini-
mum value of the dispersion relative to the maximum metric. The follow-
ing result of Sukharev [337] contains the relevant information. The lower bound
for d'w (P) is a slight refinement of (6.4).
THEOREM 6.8. For any point set P of N points in is, we have
Furtherm,ore, for every Nand 8, there .exi8ts a P for which equality holds.
Proof. Let the points of P be Xl,.;. ,'XN E i/
S
and let m be the positive
integer with m
S
:::; N < (m + 1)s. If r = then the balls B(Xn; r),
1 :::; n :::; N, in the metric d' cover is. Consider the pornt set Q consisting of the
(m+ 1)S points (kdm, ... ,ks/m) with the ki running independently through the
set {O, 1, ... ,m}. By the pigeon-hole principle, there exists an n with 1 :::; n :::; N
such that B(Xnj r) contains two distinct points of Q, say ql and Q2. Then it
follows that
hence
1 . 1
= r 2m = 2LNI/sJ
To prove the second part, let N and s be given and choose m as above. Con-
sider the m
S
(hl/(2m), ... ,hs/(2m)) with the hi running independently
QUASI-MONTE CARLO METHODS FOR OPTIMIZATION 155
through the set {I, 3, 5, ... ,2m - l}, and then repeat some of these points to
obtain a point set P with N points. For this P, it is easily seen that
I () 1 1
d
N
P = 2m = 2LNl/8J'
o
The analogous problem of determining the minimum value of dN(P) for fixed
, N is equivalent to a difficult geometric problem, namely, that of finding the most
economical covering of 2i1 by balls (in the Euclidean metric) of equal radius. This
problem has been solved only in the trivial case where s = 1 and in the case where
s = 2, where the solution is given by the circumscribed circles of a hexagonal
, tiling of 22 (see [102, pp. 58-59]).
'Next, we consider the question of constructing low-dispersion sequences rel-
ative to the maximum metric in the multidimensional case. Currently, the best
, on this problem is due to Niederreiter [235].
\',1" THEOREM 6.9. For every s 2:: 1, there exists a sequence S of points in 1&
"JJith
lim = I 1
4
,
og
Proof. Let 8
0
be the one-dimensional. sequence defined in (6.7). Consider
a sequence 8 of points Xl,X2" . E Iii such that, for any positive integer m,
the first mil terms of 8 consist exactly of all points (Ul, ... ,u
iI
) with the Ui
running independently through the set {Xl,." ,x
m
}. The precise order of the
Xn is immaterial. For given N 2:: I, determine m by mil S; N < (m + 1)8
and let y = (Yl,'" ,Ya) E Ia be arbitrary. Then, for each i with 1 S; i S; s,
there exists an integer ni with 1 S; ni S; m such that IYi - xn.l S; dm(80 ),
The point (xnl"" ,xnJ is among the points Xl, ... , XN; thus S; dm(So).
Furthermore, there exists t E I such that
Then the point (t, ... ,t) E Ia has distance 2:: d
m
+1(So) in the metric d' from
all points Xl, ... ,XN. This implies that d'rv(S) 2:: dm +1(So). Consequently, we
have
mdm +1(So) S; N
1
/
s
d'rv(S) < (m + l)dm (So),
and so lim
m
__ = mdm(So) = Ij(log4) yields the desired conclusion. 0
Thus, for every dimension s, there is a sequence S of points in Ia with
d'rv(S) = D(N-l/s). In view of Theorem 6.8, this order of magnitude is the best
possible. Note that the corresponding problem of determining the minimal order
of magnitude for the discrepancy of s-dimensional sequences is unsolved for s 2:: 2
(compare with 3.1). As to the dispersion d'rv(S), there remains the problem of
finding the asymptotically optimal constant in the bound div(S) = O(N-1/s).
More precisely, we want to determine the constant
(6.8)
156 CHAPTER 6
where the infimum is taken over all sequences 8 of points in Is, We infer from
Theorems 6.8 and 6.9 that
1 - 1
- < Dcr) <-.
2 - - 10g4
The value of DUS) is only known in the case where s = 1, where it is equal to
1J(10g4) by Theorem 6.7. In analogy with (6.8), we may define, for any bounded
subset E of ]Rs with As (E) > 0, the constant
D(E) = in{ lim Nl/sd' (8' E)
S N-+oo N"
where the infimum is extended over all sequences 8 of points in E. It appears
to be a rather challenging problem to determine the value of this constant.
We now analyze the dis'Persion some typical low-discreuancy 'Point sets
and sequences to improve on the bound obtained by Theorem in the cases
considered. The following two results for nets and (t, s)-sequences (see Chapter 4)
were shown by Niecierreiter [247].
THEOREM 6.10. For any (t, m, s)-net P in base b, we have
d ~ ( P ) ~ b-L('I'n-t)/sJ ~ b(s-Hi)/s N-
l
/
s
with N = b
m
.
Proof. Write k = L(m - t)/sJ, so that m - t = ks + h with {) :::; h :::; s - 1.
Consider the partition of 1
8
into elementary intenrals in base b of the form
h s
II[ (li b-
k
-l, (ai + 1)b-
k
-
1
) X [aiD-k, (ai + l)b-
k
)
i=l
with integers 0 ~ ai < h
k
+
1
for 1 SiS h and 0 ~ ai < hk for h +- 1 :::; iss. :U
X E IS is arbitrary, then x belongs to a unique interval J of that partition. Since
it follows from Definition 4.1 that J contaks at least one point Xn of
d'(x, Xn) < b-kj thus
d ~ ( P ) S b-
k
~ b{s-Ht)jsb-
m
/
s
= b(s-Hi)js N-
1
/
s
.
THEOREM 6.11. For any (t, s)-sequence 8 in base b
J
'We have
d ~ ( 8 ) < b(sH)/s N-1jfJ for aU N ;?: L
o
Proof. For N < b
fJH
, we have ' d ~ ( 8 ) ~ 1 < b(a+t)/s N-1js. For N ;?: biJ,+t
letm be the largest integer with b
m
~ N,so that, in particular, m > t. By)
Definition 4.2, thefust b
m
terms of 8 form a (t, m, 8)-net base b. Usmgl
Theorem-o.lO and N < b
m
+
1
j we therefore obtain
o
QUASI-MONTE CARLO METHODS FOR OPTIMIZATION 157
Next, we present bounds for the d.iBpersion of Halton sequences and Hammer-
sley point sets (see 3.1). Theorem 6.12 is an analogue of a result of Mitchell [212]
for the dispersion with respect to the Euclidean metric, and Theorem 6.13 was
established by Bayrhamer [17] (see [212] for the "Euclidean" version).
'THEOREM 6.12. For the Halton sequence S paif"l1}'id1e relatively prime
lmses hI, ... ,b
s
, we have
N2::1.
Proof. Fix N 2:: 1, and, for 1 ::; i ::; 8, let Ii be the largest integer with
_ b{' :::; NI/s. Consider the partition of JB into intervals of the form
s
rr[ Cib-;", (Ci +,1)b-;")
i=l
with integers 0 ::; Ci < b{> for 1 :::; i ::; 8. If x E J8 is arbitrary, then x belongs
to a unique interval J of that partition. In the proof of Theorem 3.6, we have
, J1 ',) .? , -'
among :any 01 ,. , , , '-' etrild Oi iJ 011t) m J.
Since N 2:: b{l ... b{a, it follows, in particular, that J contEJins at least one term
Xn of S with 0:::; n:::; N - 1. Now d'(x,Xn) < maxl::;i::;a b-;!;, and so
d' (8) < max b-:'; < N-
1
/
s
max hi
N - l::;i::;1i t l::;i::;a '
where we used b{'+! > NI/s for 1 ::; i ::; 8. 0
THEOREM 6.13. If 8 2:: 2, then, for the N-element Hammersley point set P
the pairwise relatively prime bases bI, ... ,bs-b we have
Proof. For 1 :::; i ::; 8 - 1, let be the largest integer with h{i :::; NIls, and
put m = n:,:; b{'. A given x E Is has a distance:::; m/N from a suitable interval
of the form
J = [em (c+ l)m) srr-
1
[ .b-:h (. l)b-:
fi
)
N' N x Ci 1 ,Ci+
i=l
with integers 0 :::; c < IN/mJ and 0 :::; Ci < b{i for 1 :::; i :::; s - 1. A point
Xn of P lies in J if and only if em :::; n < (e + l)m EL.'i1d the term with index
n of the (8 - I)-dimensional Halton sequence in the bases 01, >" ,b
a
-
1
lies in
n:':;[Cib-;1i, (Ci + l)b-;'i). By the proof of Theorem 3.6, these two conditions
on n are satisfied for exactly one value of n. Vilith this n, we have
d
'( ) m I (m b--li)
X,Xn :::; N ,max'\N' i -,

and the bound for djq (P) follows. 0
158 CHAPTER 6
Theorems 6.11 and 6.12 provide further examples of sequences S of points
in is for which d'rv(S) = D(N-l/S). A sequence S of dyadic points in j2 with
dN(S) = O(N-
1
/
2
) was constructed by Lambert [167]. Good lattice points (see
5.1) can also be used to obtain s-dimensional point sets with a dispersion of
minimal order of magnitude. In fact, it was shown by Niederreiter [240] that,
for every s, there are infinitely many N for which an effective construction of a
g E ZS can be given such that the corresponding point set P in (5.1) satisfies
= O(N-1
/
8).
Notes.
Expository accounts of random search methods can be found in the books of
Evtushenko [92], Rubinstein [297], [298], Torn and Zilinskas [352], and Zhigl-
javsky [368]. Various stochastic models for random search have been analyzed
by Devroye [61], Matyas [208], Pinter [282]' and Solis and Wets [330], among
others.
Quasirandom search methods have been used in numerical practice for a con-
siderable time; see Aird and Rice [6], Artobolevskii et al. [10], [11], Sobol' [325],
and Sobol' and Statnikov [329] for early published accounts. Sukharev [337J car-
ried out the first theoretical analysis of quasirandom search for objective func-
tions satisfying a Lipschitz condition. A general systematic analysm, including
the error bound in Theorem 6.3, was first given in Niederreiter [228J (see also
Niederreiter and McCurley [266]), where the term "dispersion" was also coined.
The dispersion was studied in detail in Niederreiter [230], [235]. For the case of
point sets in is, Sobol' [326] proposed to take into account the dispersion of the
projections ola point set on all lower-dimensional faces of is whena.ssessing the
suitability of a point set for quasirandom search. Sobol' [327], [328] investigated
the dispersion with respect to weighted gl metrics on is.
The lower bound (6.3) for the dispersion was shown in Niederreiter [228],
where an improvement can be found for a tiling domain E, i.e., for a bounded
subset E of lR
S
with As(E) >0 such that lR
s
can be tiled by congruent copies of
E (such as E = jS). An improvement on (6.3) for the case where E is a conve;x: "
body was established by Gritzmann [115]. In the proof of Theorem 6.8, w;e
followed Niederreiter [235]. For the special case of (t, m, s )-nets P in base 2, an
analogue of Theorem 6.10 for the dispersion dN(P) was proved by Sobol' [326].
Bounds for the dispersion of nets with respect to weighted 11 metrics were given
in Sobol' [327], [328]. For s = 2, Peart [280] determined the exact value of dN(P)
when P is an N-element Hammersley point set in base 2 with N being a power
of 2. Again for 8 = 2, LarcheI' [174] gave a formula for d'rv(P) when P is the
N-element point set (5.1) generated by g = (1,g) E zl,. The related concept
of the qispersion of a tw<H:limensional la.ttice was introduced and analyzed by
Larcher [177]. The of the sequence of fractional parts {nz},
1,2, ... , with z E 2
8
was studied by Drobot [71] and Niederreiter [230] in'the/
case and by Niederreiter [235] in the multidimensional case. ' ,
With suitable choices of z we arrive at s-dimensionalsequences S with dN(S) =
O(N-
1
/
B
). General bounds for the dispersion in terms of exponential sums were
\
QUASI-MONTE CARLO METHODS FOR OPTIMIZATION 159
shown by Mitchell [212] and Niooerreiter [230], [240]. A result of Deheuvels [57]
on maximal spacings for multivariate order statistics implies that d ~ (S) is almost
surely of the order of magnitude N-l/B(logN)l/s for random sequences S of
points in ]a.
Reports on comparative studies of the performance of various low-discrepancy
point sets and sequences in quasirandom search methods can be found in
Bandyopadhyay and Sarkar [15] and Niederreiter and Peart [267]. Tests with
Halton sequences were carried out by Niederreiter and McCurley [266]. Nu-
merical experiments for domains other than js were reported in Niederreiter
and Peart [268], and this paper contains also Theorem 6.4 as well as methods
of generating low-dispersion point sets and sequences in special domains such
as balls, spheres, and simplices. Fedorenko [101] compared quasirandom search
methods with other global optimization methods. Applications of quasirandom
search methods to computational physics are discussed in Bandyopadhyay and
'i Sarkar [14] and Khasnabis et aI. [151].
~ ; : : Level sets of the objective function and their connections with quasiran-
-{'.:
,O) dom search were studied by Bayrhamer [16], [17]. A description of localiza-
tion of search in thB setting spaces was given by Niederreiter [235].
Sukharev [338] analyzed sequential search strategies, Le., procedures where the
choice of a sample point depends on the previously calculated sample points and
function values. Applications of quasirandom search to discrete optimization are
discussed in Niederreiter [240].
CHAPTER' 7
i.
Random Numbers a'n'd
Pseudorandom Nombers
Random. Bam,pling is at the heart of the Monte Carlo method, 88 we have seen
in Chapter The success9f a Carlo calculation depi!nds, of course,
on the appropriateness of the model, but also, to a large
extent, on how well the the,f.'X)mputation simulate the
random variables in the model. The specification oftha requirements on random
numbers and the discussion of the testing of these are two, of the
topjca 9f this chapter. ,In .. actual practice, numbers are generated by
a deterministic algorithm that is implemented in the and so we are
. - '-. .
really working .with pseudorandom numbers. Therefore oW," main attentiop in
this and the following chapters will, be . focused on pseudorandom numbers. l'he
emphasis be on methoPs of pseudorandom number generation for which a
theoretical analysis of the structural and statistical properties is possible.
In 7.1 we discuss the desira.ble properties of random numbers and SQme at-
tempts at defining the rather elusive concept of a sequence of random numbers.
Section 7.2 lays the foundations for our treatment of pseudorandom numbers,
gives a brief account of some methods for the generation of nonuniform pseu-
dorandom numbers, and describes randomness tests for uniform pseudorandom
numbers. Some classical methods for the generation of uniform pseudorandom
numbers, such as the linear congruential method, are briefly analyzed in 7.3.
7.1. Random number generation.
The task in random number generation is the fonowing: Given a distribution
function F on R, generate a sequence of real numbers that simulates a sequence
of independent and identically distributed random variables with distribution
function F. Random numbers are needed in a wide range of areas; we have noted
their applications in Monte Carlo methods and, more generally, in simulation
methods, but they are also of basic importance in computational statistics, in the
implementation of probabilistic algorithms, and in related problems of scientific
computing that have a stochastic ingredient. In addition, random numbers are
applied in areas of direct practical interest such as VLSI testing, cryptography,
and computer games.
The requirements on random numbers may be classified into the following
four categories: structural, statistical, complexity-theoretic, and computational.
161
162 CHAPTER 7
Structural requirements refer to aspects such as period length, lattice structure,
and so on. Statistical and complexity-theoretic requirements will be discussed
more fully in the sequel. By computational requirements, we mean features relat-
ing to the computer implementation of random numbers, such as ease of coding,
set-up time, running time, required memory space, and portability. Given the
power of modern computers, the consideration of the efficiency, Le., of the time
and space requirements, of random number generation has a lower priority nowa-
days; we quote from James [146, p. 332]: "This [efficiency] was considered very
important in the early days, but with the kind of computations being performed
now, both the computer time and memory space taken by random number gen-
eration are increasingly insignificant and can almost always be neglected."
It is common to make a distinction between uniform and nonuniform random
numbers. Uniform mndom numbers are random numbers for which the target
distribution is the uniform distribution U on r = r JJ. te" thedistribntir::m
"" .J' ,
function defined by U(t) = 0 for t < 0, U(t) = t for 0 :-:; t :-:; 1, and U(t) = 1 for
t> 1. Nonuniform random numbers (also called random variates) have a target
distribution different from U. Nonuniform random numbers are often generated
by suitably transforming uniform random numbers; see 7.2 for a description of
such methods in the context of pseudorandom number generation.
The statistical requirements imposed on random numbers relate mostly to
distribution and statistical independence properties. Statistical tests for these
properties will be discussed in 7.2 in connection with uniform pseudorandom
numbers. Here we report on efforts of handling the vexing problem of defining the
concept of a sequence of random numbers, a problem which, if explored to its full
consequence, actually leads from the mathematical into the philosophical realm.
A possible formalization of the statistical properties of uniform random numbers
is based on the following extension of the notion of a uniformly distributed
sequence (compare with 2.1).
DEFINITION 7.1. A sequence Xo, Xl, . .. of numbers in I is completely uni-
formly distributed if for every integer 8 ~ 1 the sequence of points (xnJ Xn+l,'" ,
xn+s-d E is, n = 0,1, ... , is uniformly distributed in]s.
Knuth [154, 3.5] proposes a hierarchy of definitions for a sequence of uni-
form random numbers, and, on the lowest level of this hierarchy, is his Definition
Rl, according to which a sequence of numbers in j is random if it is completely
uniformly distributed. The notion of complete uniform distribution attempts to
formalize the requirement that successive random numbers should be statistically
independent. However, this notion does not reflect the intuitively obvious prin-
ciple that the distribution and statistical independence properties of a sequence
of random numbers should be invariant under certain selection rules for subse-
quences. Clearly, we cannot insist that the sequence and all its subsequences be
completely uniformly distributed, as this would .lead to a void concept. Hence
some restrict jon on the admissible selection rules has to be imposed. Definition
R4 of Knuth [154, 3.5] proposes to call a sequence Xo, Xl, .. of numbers in J
random if, for every effective algorithm that specifies a sequence 80, 81, . .. of
distinct nonnegative integers, the sequence x
so
' X
Sl
, is completely uniformly
distributed. \
RANDOM NUMBERS AND PSEUDORANDOM NUMBERS 163
Explicit constructions of completely uniformly distributed sequences are
known, and we refer-to [225,p. 997) for a survey of such constructions. Note that,
according to [163,p. 204,'I'hm.3.13], all sequences of numbem in I
are completely, \lDif(),nnlydWtrjbMted, :w:here;.A is the Lebesgue measure on I. A
classical says that th':HJequence of.fr,actional
parts {on}, n;:: 0,.]" .. c()'IQ.plet.ely,unifornlly distributedJor almost all real
numbers 0> 1 (in,the sense,of Lebesgue measure). Thus, in the indicated sense,
almost all these specific sequeQ.ceB satisfy Knuth's Definition R1 of randomness.
Knuth [154, 3.5] conjectured that these sequences satisfy his much stronger
Definition R4 in the above probabilistic sense, and Niederreiter and Tichy [275]
managed to prove this conjecture by combining several techniques of probabilis-
tic number theory and results on the location of zeros of sparse polynomials.
Concretely, the theorem of Niederreiter and Tichy states the following.
- THEOREM 7.2. The sequenCe"o! jraetionalparts n -0, i, ... , satisfies
Knuth's DefinitionR4 Olmost-all real nu:niikrs 8 with 181> 1
'(in the Lebesgue measure). . - - . -
---\ -. , - ,,'"
The approach to Ti3;ndomness on the notion of -complete uliiIorm di&-
tributiOll can also be pursued'for"Sequences of:digits (see Knuth [154, 3.5]). In
this case, thete are With' the clasSical' theory of normal
numbers (see Knuth [154, 3.5j' and; Kuipem and Niederreiter [163, 1.8]). _
A to a' definition of randomness for finite
strings of digits is due'to Kolrilogorov ,[156J, Chait in [39], 'and Martin-Lof [207].
If b 2 is an integer ahd"UN is a string, Le., a finite sequence, of length N
consisting of elements of Zb -to, 1, ... ,b - I}, then its Kolmogorov complexity
is defined as the shorteSt length ofa program that generates (TN on a Turing
machine. The string (1 N is designated as random if it has the maximum Kol-
mogorov complexity among all strings of lengthN consisting of elements of Zb'
This is, of course, the worst concept of randomness from the viewpoint of practi-
cal random number generation, since it means that the random strings are those
that are the least efficient with regard to computer implementation. On the
other hand, this complexity-theoretic approach is of great relevance for random
bit strings that are used for cryptographic purposes. In these applications, it is
essential that the utilized bit strings have a suffidently high complexity, so that
the algorithms generating the strings cannot be "cracked." The procedures that
have been devised for the generation of cryptographically strong pseudorandom
bit strings have not yet made an impact on random number generation for simu-
lation methods, one reason being that the secrecy aspect in cryptography is not
a concern in simulation studies. On the contrary, it can be argued that, in the
latter area, transparency and reprodudbility of random numbers are virtues.
Efforts have also been made to produce physical random numbers, i.e., to
generate random numbers by physical devices. The sources of random numbers
that have been considered include coin flipping, roulette wheels, white noise,
and counts of emitted particles. Some relevant references are given in [225,
p. 998J. Although the use of physical random numbers may appeal to intuition,
it must be emphasized that these types of random numbers are problematic in
practice. First, since no theoretical analysis sf these random numbers
164 CHAPTER 7
is possible, they must be subjected to extensive statistical tests for randomness
prior to actual use. Furthermore, the simulation studies now performed require
huge amounts of random numbers, and since these must be available for the later
validation of computations, this causes considerable storage difficulties in the
case of physical random numbers. By way of comparison, in the case of standard
algorithms for pseudorandom number generation, just a few parameters and
some seed values must be stored. There is also a more philosophical standpoint
that posits that calculations executed in a computer should be supported as
much as possible by internal routines and should not have to rely on external
sources for auxiliary data such as random numbers.
1.2. Pseudorandom numbers.
Early in the history of the Monte Carlo method, it already became clear that
('truly random" numbers arB fictitious a practical point of view. Therefore
users have resorted to pseudorandom numbers (abbreviated PRN) that can be
readily generated in the computer by deterministic algorithms with relatively few
input parameters. In this way, problems of storage and reproducibility do not
arise. Furthermore, well-chosen algorithms for pseudorandom number generation
can be subjected to a rigorous theoretical analysis. Naturally, the PRN must pass
an assortment of statistical tests for randomness to be suitable for simulation
purposes. It should be clear that a deterministic sequence of numbers cannot
perform well under aU imaginable tests for randomness. Therefore the user of
PRN must be aW8;fe of the specific desirable statistical properties of the random
samples in the current computational project and must PRN that are
known to pass the corresponding Statistical tests.
The stru:;l.dard algorithms for generating sequences of PRN are based on recur-
sive procedures and yield sequences that are ultimately periodic. It is no serious
loss of generality if restrict the. attention to (purely) periodic sequences. A
sequence Yo, Yl, . .. of elements ,,of a nonempty set is (purely) periodic if there
exists a positive integer T 'lin for all n O. We write per(Yn)
for the least value of T, i.e., for least period length.
The desirable properties qf a sequence of PRN may be summarized as follo'!J!lS:
(i) the least period length should be sufficiently large; (ii) it should have little
intrinsic structure (such as lattice structure); (iii) it should have good statistical
properties; (iv) the computational requirements listed in 7.1 should be satisfied
(e.g., the algorithm generating the sequence should be reasonably efficient).
As in the case of random numbers, we distinguish between uniform pseudo-
random numbers and nonuniform pseudorandom numbers. Uniform PRN sim-
ulate the uniform distribution U on 1. Nonuniform PRN are usually generated
by starting from uniform PRN and transforming tllem to.fit a given target dis-
tribution F #- tT. In our we ,will concentrate on uniform PRN: Many
methods have been develdped for PRN hito nonuniform
'_ " 'I', ':' ,', ',. Co .> 1]'11 \ .' _ ,: I, .:
PRN, and, in the following, we briefly discuss the moot commo;n ones.
(i) Inversion method. Let the distribution function F be strictly increasing
and continuous oni]R. Then F has an inverse function F-
1
, which is defined at
RANDOM NUMBERS AND PSEUDORANDOM NUMBERS 165
least on the open interval (0,1). Now take a sequence xo, Xl, . of uniform PM
in (0,1) and define Yn = F-l(x
n
), n = 0,1, ... , as the sequence of no:ounHorm
PRN. IT F does not satisfy the two conditions above, then we may consider
the generalized inverse function G(u) = min{t E R : Fet) u} for u E (0,1),
which exists under. the standard assumption that the distribution fuhction F.
continuous from the right j and we put Yn =G(X;n) for n O.
this method is, in theory, universal, it is in many cases impractical since the
(generalized.) inverse function may not be available in a convenient form.
(ii) Rejection method. Suppose that F has a. density I, Le., F(u) =
I (t) dt for u E It. Find a ''nice'' distribution G with density 9 such that
I(t) cg(t) for all t E 1R with a. constant c 1. Now generate two independent
random samples, namely, x from the uniform distribution U and z from the dis-
tribution G. IT ex> I(z)/g(z), then reject z and try again with other x and z..
This is repeated until finally ex f(z)/g(z), and then the random variate y = z
is the output. Note that
Probe z :s; u
. . ftl f(tl_ " ,., f{t) ,
z18 RCcepted) = J-oo eg(i) dUlt)::;:; i_
oo
Cg(t)g(t)d.t
11'" 1
= - I(t) dt = -F(u),
e -00 e
and, by taking u -+ 00, we see that
Prob(z is accepted) = !.
e
Thus we obtain the conditional probability
-l/.i'( )
Prob(z ::; u i z is accepted) = e J1. 1 U = P(u).
c-
Consequently, the random variate y has the distribution function F. The c0n-
stant c should be as close to 1 as possible so that the acceptance probability is
large.
(iii) Composition method. We regard F as a mixture F = Ej=lPj Fj
distribution functions F
j
on lR and a discrete probability distribution (PI, . .. ,'Pr)
with Ej=lPj = 1 and Pj 0 for 1 j ::; T. Now we first sample j from the
discrete distribution, and then y from the corresponding distribution F
j
The
random variate y has the distribution function F. This method is often applied
in the following form: Partition R into finitely ma.ny intervals h j ,Ir and let
F
j
be the distribution function with support I
j
induced by F.
(iv) Ratio-oj-uniforms method. Suppose that F has a density I, put A =
{(u, v) E IR2 : 0 < u f(V/U)1/2}, and assume that A is contained in some
compact interval [0, a 1 x [b
l
, ba]. Now we generate two independent uniform
random numbers Xl and X2, and we put Zl = aXl and Z2 = bl + (ba - bl )X2-
If (Zl-,_Z2) tf. A, then we try again with other Xl and X2. This is repeated until
166 CHAPTER 7
finally (Zb Z2) E A, and then the random variate y = Z2/ Zl is the output. It can
be shown that y has the distribution function F (see [62, p. 195], [295, p. 66]).
There are quite a few other general techniques for transforming uniform PRN
into nonuniform PRl'J, and there is a great abundance of methods tailored to
special distributions such as normal distributions, beta distributions, gamma
distributions, and so on. An encyclopedic account of methods for random variate
generation is given in the book of Devroye [62].
We now discuss some typical statistical tests for the randomness of uniform
PRN. Acceptable sequences of uniform PRN should pass a variety of such tests.
Let xo, XI, ... be a sequence of numbers in I to be tested for randomness and
let N be a large integer; if the sequence is periodic, then it suffices to take
N per(x
n
).
A. Unilormit1/ test, This is a goodness-of-fit test for em;pirical distribu-
tion of an initial segment xo, Xl,." ,XN-l of the sequence. The test is per-
formed by calculating (or bounding) the discrepancy or the star discrepancy of
XO,Xl, ... ,XN-l If DN(xo" .. ,XN-l) is taken as the test quantity, then this
amounts to performing the two-sided Kolmogorov test in nonparametric statis-
tics, which is a very well--studied procedure. For fixed N and random point
sets Zo, Zl, ... ,ZN-l E I, the distribution of D"N(zo, ... ,ZN-l) is known. Var-
ious formulas for Prob(D"N(zo, ... ,zN-d u), u 1, are given in [220,
5]. For random sequences Zo, Zl,'" of numbers in I, the star discrepancy
Div (zo, ... ,ZN -1) follows the Kolmogorov-Smirnov law
00
lim Prob(NI/2 Dj,,(zo, ... ,ZN-l) u) = 1-2 ""( _1)k+l e-2k
2
u
2
for u> O.
N-+oo L,a
k=l
Furthermore, we have the law of the iterated logarithm
(7.1)
11
'm (2N)1/2 Df.,(zo, ... ,ZN-l) = 1
(
)1/2
"xoo-a.e.,
N -'>00 log logN "
where A is the Lebesgue measure on I (see Chung [41]). In a related test known
as the frequency test, we partition I into finitely many subintervals, count the
number of terms among xo, Xl, , . ,XN-l falling into the Various subintervals,
and perform a X
2
test on these data. '
B. Gap test. Let J be a fixed proper subinterval of I. If, for some n 2 0, we
have xn+i J for 0 ::s; j ::s; k - 1, butxn+k E ,J, then we refer to a gap of length
k. These gap lengths are geometrically distributed with parameter 1- Al (J). In
the practical implementation of the gap test, we choose a positive integer h and
cqunt the\llumber gaps oflengths 0, 1,. .. ,h - 1, and ;;::: h until a. large total
number of gaps is reached. we' cipply a X
2
test to these data.
C. Run test.;A segment ofthe sequence satisfying Xn < Xn+1 < , .. < Xn+k-l,
but X
n
-l 2 Xn and Xn+k-l Xn+k, is called a run-up of length k. We count
the number of of lengths 1,2, ... ,h, and 2 h + 1 in xo, Xl,'" ,XN-l'
RANDOM NUMBERS AND PSEUDORANDOM NUMBERS 167
Since adjacent ru.n..s are not independent, we cannot use a straightforward X
2
test for these data. A more complicated statistic given in [154, 3.3.2] must be
wmputed. A similar test can be pe:rformed for nms-down.
D. Permutation test' We choose an integer 8 ? 2 and consider the a-tuples
(zn, Xn+b"" Xn + ...... l) , n :::::: 0,1, ... ,N ":"':1. There are a!'possible relative order-
ings among the entries of such an s.;;tuple, and these orderings are equiprobable.
We determine the frequency of each ordering and apply a X2 test to these data,
using the probability 1/ sf for each ordering. Alternatively, we may consider the
maximal deviation of these frequencies from the expected number of occurrences.
E. Serial correlation. This is a rather weak test for the interdependence
'between Xn and Xn+h, where h ? 1 is a given integer. The test is performed by
calculating the serial cofTelationcoefficient
, " " . 2
~ (h) MN(xnxy>+h) - MN(Xn)
UN = MN(x!) _ MN{Xn)2 '
where .J.'if
N
( un) = N-l 4:;;':; Un is the me8.n valne of Uo) 1lL; , "' 1J N ,-1 and where
-<Ve assume that' the dElDominator is nonzero. If Xn and Xn+h are almost inde-
pendent, then I u ~ ) I is vef"j small. It is a deficiency of this test that the converse
does not necessarily hold.
F. Serial test. This is a more severe test for the statistical independence of
successive PRN and is a multidimensional version of the uniformity test in A.
The serial test can also be viewed as a 'Way of measuring the deviation from
the property of complete uniform distribution in Definition 7.1. It is therefore
closely connected with Knuth's Definition Rl of randomness, stated in 7.1. For
a fixed dimension s ? 2, put
Xn = (Xn,Xn+b ... ,xn+s-d E PJ for n = 0,1"" ,N-1.
Then consider the discrepancy
(7.2) D ~ ) := D N{Xo, ... ,XN -1)
or the star discrepancy
(7.3)
D
*(s) D* ( )
N := NXo,,XN-l
For fixed N and random point sets ZO,Zl,." ,ZN-l E Is, we have the following
result of Kiefer [152] on the distribution of the star discrepancy: For every E > 0,
there exists a constant c > 0 depending only on E and s such that
Prob(N
l
/
2
D'N(zo, ... ,zN-d :S u) ? 1 - Ce-(2-e)u
2
for all u? O.
For random sequences ZO,Zl,... of points in Is, the behavior of
D'N(zo, ... ,ZN-l) is governed by the s-dimensional version of the law of the
iterated logarithm in (7.1), namely,
(7.4) lim (2N)1/2 DN(zo, ... , ZN -1) = 1 X,o -a.e.,
N-+oo (log log N)1/2
168 CRl\PTER 7
where A is now the Lebesgue measure on III (see [152]). This law serves as a
benchmark for the orders of magnitude of and The a-dimensional
serial test has an added significance for numerical analysis, since upper bounds
on yield error bounds for quasi-Monte Carlo integrations using the nodes
Xo I I XN -1 (compare with 2.2). Moreover, further quantities of statistical
relevance can be bounded in terms of D;?). For instance, by the method in
Niederreiter [233]! we obtain
for the serial correlation coefficient cr};). Some authors have also considered the
distribution of the nonoverlapping s-tuples Xns = (xns! Xns+I, ... ,Xns+lI-l) E
Is, n = 0, 1, .... However, it should be noted that usually this does not make any
dilIcTBIice, for, the standard case where xo, Xl,'" is pei'ioalc and = per(x
n
},
the point sets KG, X17 .. ,XN -1 and KG, Xs , ... ,X(N -I)s are identical whenever
ged(s, N) = 1, since then the indices ns, n = 0,1, ... ,N - 1, run modulo N
through the set {O, 1, ... ,N -l}. There is also an a-dimensional frequency test,
which operates like the frequency test mentioned under A, namely, by starting
from a partition of I into finitely many subintervals, counting the number of
terms among KG, Xl, ... ,XN-l falling into the various subintervals, and applying
a X
2
test to these data.
G. Spectral test. This test is applied to periodic sequences Xo, Xl, , where
all Xn are rationals 'f}Tith common denominator M. If per(xn) = T, then, for
b. = (hI,'" ! hs ) E za, we define the exponential sums
Ideally, we should have X(h) = 1 if h = 0 modM and X(h) = otherwise.
'The deviation from this behavior is a measure for the nonrandomness of the
given sequence of PRN. The difficulty here is to find a convincing quantitative
formulation of this idea. Coveyou and MacPherson [44] used an intuitive analogy
with the theory of harmonic oscillation; in the case of linear cong11lential PRN,
this proposal has a nice geometric interpretation (see 7.3).
1.3. Classical generators.
A classical and still very popular method for the generation of uniform PRN is
the linear congruential method introduced by Lehmer [187]. As the parameters in
this we choose a large integer M, an integer.a with 1 ::s; a <M
gcd(a,lvJ). 1, and an integer C E ZM ={O, 1, ... ,At! - I}. ThenWf3
select an initial value Yo EZM and gene'rate a sequence Yo, Yl, ... E ZM by the
recursion-
(7.5) 1/n+1 = aYn + c modM for n = 0,1, ....
RANDOM NUMBERS AND PSEUDORANDOM NUMBERS 169
From this sequence, we derive the linear congruential pseudorandom numbers
(7.6) Xn = E 1= [0,1) for n = 0,1, ....
" h'
In, this context, M is referred to as the modulUs and a as multiplier. The
., ' "I' -'-'l' .
choice of the modulus'is customarily accorded with the word length of the ma.-
chine, typical values being M = 2
32
or the Mersenne prime M = 2
31
- 1. For
high-precision calculations, larger values such as M = 2
48
have' also been used.
A common distinction is between the homogeneous case where c = 0, also called
the multiplicative congroential method, and the inhomogeneous case where c : 0,
also called the mi:ced congroential method. Since linear congruential method
has been exterisivelytreated in book (see, e.g., [154], [295]), we limit our
discussion to those aspects that are related to other material in these lecture
'r" ...'., '
' .
i i ,It is dear from' (7.S) and, (7.6) tliat thesequence(xO;x1, . .. andYol Y1, ...
'both periodic and = per(Yn)::; M.' AD. elementary number-
theoretic argument shows that pe:r(xn) = M if and oniy if gcd(c,lvI) = 1,
a' ., 1 modp for every pfPne p a= 1 mod4 whenever 4 di-
vides M [154, 3:2J)). The folloWing are three standard that are
considered in practicafimplementations of the linear congruential method (in all
three cases per(xn) is reasOnably clOse to M):
(i) M is prime, a is a primitive root (i.e., the multiplicative order
of a modulo M is M - 1), c = 0, and Yo =1 O. Note that (7.5) implies that
Yn = anyo modM for n = 0,1, ... , and so per(xn) = M - 1.
(ii) M is a. power of 2, a = 5 mod 8, and c is odd. From the above criterion,
we get per(xn) = M.
(iii) M is a power of 2, a _ 5 mod8, c = 0, and Yo is odd. Then) again,
Yn = an Yo mod M for n = 0,1, ... , and so per(xn) equals the multiplicative
order of a modulo M, which yields per(xn) = MI4 by an elementary argument
(see [295, Thm. 2.2]).
The performance of linear congruential PRN under the uniformity test is
easy to describe in the above three cases if we apply the test to the full pe-
riod. Put T = per(xn) and first consider case (ii). Then the point set
xo, Xl,'" ,XT-I is equal to the set of rationals in I with denominator M,
and so DT(xo, ... ,xT-d = 11M by Theorem 2.6. In case (i) the point set
Xo, Xl, ... ,XT-l is equal to the equidistant point set 11M, 21M, ... ,(M -1)/M,
and so DT(xo, ... ,XT-l) = liN! by Theorem 2.6. In case (iii), the point set
xo, XI, . . ,XT -1 is equal to the set of all rationals in I of the form b / M with an
integer b = Yo mod 4, and so DT(xo, . .. ,XT-l) = 31M by Theorem 2.6. Thus,
in all three cases, we have DT(xo, ... ,XT-l) = O(M-
1
) with an absolute im-
plied constant. The situation is less trivial if we consider the uniformity test for
parts of the period, Le., for 1 ::; N < T. Results of Niederreiter [221], [222] show
170 CHAPTER 7
that, in our three cases, we have
for 1:::; N < T
with an absolute implied constant and that this bound is, in general, the best
possible, apart from the logarithmic factor.
Historically, the distribution of the s-tuples Xn = (Xnl Xn+l, .. . ,Xn+s-l) E
IS, n = 0,1, ... , with s ~ 2 was first considered in the context of stochastic
models. Franklin [108J studied the dynamics of the recursion xn+l = {axn + 8},
n = 0,1, ... , with an integer a > 1, a fixed 8 E I, 8.-nd a random initial value
Xo E I, as well as the dynamics of the associated sequence Xc, Xl, ... of s-
tuples. Although the sequence Xc, Xl, ... can never be uniformly distributed in
Is, Franklin could prove that, for almost all Xo E I (in the sense of Lebesgue
measure), the sequence is "asymptotically" uniformly distributed in fa as the
,"''+-ON,'''' a tends to infinity. For the CB88 where e = corresponding to the
multiplicative congruential method, a quantitative refinement was established
by Ermakov [91J. We refer also to [225, 1l] for a brief discussion of these
results.
The more important question of the performance of a concrete sequence
Xc, Xl, . .. of linear congruential PRN under the s-dimensional serial test was
explored much later. To arrive at the crucial observation in a more leisurely
way, we consider, for now, only case (i). Then Yn = anyo mod M for all n ~ 0,
and so, for Yn = (Yn, Yn+l,'" ,Yn+B-l) E ZIt, we obtain
where g = (1, a, a
2
, . .. ,as-I) E ZS. Since a is a primitive root modulo M,
gcd(yo, M) = 1, and T = per(xn) = M - 1, the integers anyo, n = 0,1, ... ,
T - 1, run modulo M through the set {I, 2, ... ,M - I}. Now Xn = M-1Yn
for n ~ 0 by (7.6), and so the point set Xc, Xl, ... ,XT-l is equal to the point
set consisting of the fractional parts {(n/M)g}, n = 1,2, ... ,M - 1. This is
a point set of the type (5.1), except that the point corresponding to n = 0 is
missing (which is, of course, insignificant for large M). Therefore, by the theory
developed in 5.1, the distribution behavior of the point set Xc,XI, ... ,XT-l,
and thus the performance of the full period of the sequence Xo, Xl, . .. under
s-dimensional serial test, is governed by the quantities measuring the suitability
of g as a good lattice point modM. In particular, for the discrepancy n!;)
the point set Xc, Xl, ... ,XT-lJ we obtain the following bound, which is a slight
variant of a result of Niederreiter [223], [224].
THEOREM 7.3. In case (i) of the linear congruential m ~ t h o d , we have
n
(s) < 8 + 1 M ( )
T - M _ 1 + M- 1 Rl g, M
for every s ~ 2,
where Rl (,g, Nt) is defined by (5.9) and where g = (1, a, a
2
, . , a
B
-
1
) E ZE.
Proof. Let P be the point set consisting of {(nj}J)g}, n = 0,1, ... ,NI - 1,
and let be obtained from P by deleting O. For every J E J (see Defini-
\
RANDOM NUMBERS AND PSEUDORANDOM NUMBERS 171
tion 2.2), we have A(J; P*) = A(J; P) - c(J) with c(J) = 0 or 1, and so
IA(J; P*) - (M-:: l)..\s(J)1 = IA(Jj P) - c(J) - M..\s(J) + ..\s(J)1
MDM(P) + l..\s(J) - c(J)1 MDM(P) + 1.
Consequently,
TD5;) = (M - I)DM-l(P*) MDM(P) + 1.
We can bound DM(P) by Theorem 5.6, and this yields
T D5;) M (1 - (1 - s + Rl (g, M)) + 1 s + 1 + M Rl (g, M). 0
We may use the bound R
1
(g, M) M) (see the proof of Theorem 5.6),
and then there arises the problem of the existence of a primitive root a modulo
the prime M for which R( (1, a, a
2
, . ,as-I), M) is small. This problem can be
treated by the averaging technique in the proof of Theorem 5.18, although now
mean is over the primitive roots M. This leads to
the following theorem of Niederreiter [224].
THEOREM 7.4. In case (i) of the linear congruential method, we have, for
every s 2:: 2,
D5;) = O(M-:-I(logM)Sloglog(M + 1))
for an average multiplier a, where the implied constant depends only on s;
In analogy with the theory of good lattice points, the quality of a multi-
plier can also be assessed by an appropriate figure of merit. The following is a
specialization of Definition 5.7.
DEFINITION 7.5. For every dimension s 2:: 2 and any integers M 2:: 2 and
a, define the figure of merit p(s)(a,M) by p(s)(a,M) = p(g,M) with g =
(1,a,a
2
, ... ,as-I) E ZS.
In view of Theorem 7.3 and (5.11), the discrepancy can be bounded
in terms of p(S) (a, M) in case (i) of the linear congruential method. For this
case, the following result concerning large figures of merit was shown by Nieder-
reiter [224]: For every s 2:: 2 and every prime M, there exists a primitive root a
modulo M with
(7.7)
(S)( M) CaM
p a, > (logM)a-lloglog(M + 1)'
where the constant C
a
> 0 depends only on s.
If s = 2, then the quantities considered above, such as p(s)(a, M), are con-
nected with continued fractions (compare with 5.2). In case (i), the favorable
multipliers a for the two-dimensional serial test are the primitive roots a modulo
M for which the number K(a/M) introduced in Definition 5.16 is small. This
then leads to the consideration of a quantity analogous to that in (5.38), namely,
QM = for primes M,
172 CHAPTER 7
where the minimum is extended over all primitive roots a modulo M. In partic-
ular, there is the open problem of whether there exists an absolute constant C
such that QM ::; C for all primes M. From Theorem 5.17 and (7.7) with s = 2,
we obtain QM = O((logM)loglog(M + 1)) with an absolute implied constant.
Let us return to the crucial observation in the treatment of case (i), above,
namely, that the point set xo, Xl, ... ,XT-l E [8 with T = per(xn) is basically
the same as a point set (5.1) with a suitable g E ZS. This brings us directly to
the lattice structure of linear congruential PRN, i.e., to the phenomenon that the
points Xn = (xn' Xn+l, . .. ,xn+s-d obtained from a sequence xo, Xl,'" of linear
congruential PRN are bound to form a regular lattice pattern. We refer to 5.3 for
the basics of the theory of lattices in lR
s
. For s ~ 2, put Xs = {Xo, Xl, . " ,XT-l},
where, again, T = per(xn). Since the points X
n1
0 ::; n ::; T -1, are distinct, Xs
is the same as the point set consisting of Xo, Xl, ... ,XT-l. For an integer N ~ 1
and a multiplier a, let Ls(a, N) be the s-dimensionallattice with lattice basis
b
1 ( 2 s-l)
1 = N 1, a, a ,... ,a ,
b
i
= ei for 2 ~ i ::; s,
where2i is the ith vector in the standard ordered basis ofR
s
. Note that Ls(a, N)
contains ZS and is thus an integration lattice in the sense of 5.3. The following
result describes the lattice structure in the three standard cases of the linear
congruential method.
THEOREM 7.6. For every 8 ~ 2, we have
Xs = Ls(a, lvJ) n (0, 1)B in case (i),
Xs = (Xo + Ls(a,M)) nIs in case (ii),
M
Xs = (Xo + LB (a, 4")) n JB in case (iii).
Proof. Incase (i), we have already soon that Xs 5;:;: Ls(a, M) n 1
8
, and, sinf,e
no Xn has a zero coordinate, we getX
s
~ Ls(a, M) n (0, l)s. On the other han4,
Ls(a,l\1) has determinant 11M, and so Theorem 5.30 shows that card(Ls(a, lVI)n
JB) = .M. Thus caxd(Ls(a, M) n (0, 1)8) ::; M -1; hence Xs = Ls(a, M) n (0, l)S
since card(Xs) = M - 1.
In general, we note that if, for fixed n ~ 0 we use induction on i and the
recursion (7.5), we then obtain
(7.8) Yn+i - Yi = ai(Yn - Yo) modlVI for i = 0,1, ....
Therefore
(7.9) x ~ - Xo = (Yn - Yo) 1r (1, a, a
2
, ... ,as-I) mod ZS for all n ;;::: 0,
and 00 always Xs ~ (xo + Ls(a, M)) n IS. In case (ii), both sets have cardinality
M, and so they ~ e equal. In case (iii), we have Yn = Yo mod4 for all n > 0:
,:,
RANDOM NUMBERS AND PSEUDORANDOM NUMBERS 173
hence (7.9) can be written as
_Yn-Yo 1 (1 2 8-1) dZ8 c. _11 0
Xn - Xo = 4 M 14 ' a, a , :., ,a mo tor WJ. n 2: .
Thus X8 (x:o+Ls(a, Af)jnI, and, sUlce both sets have cardiruility M/4, they
are equal. 0 '
In the abowproof, we have noted that' Xs is always contained in (Xo +
L.(a, M n I. Thus all points Xn lie on a grid, i.e., a shifted lattice, with the
determinant .of the lattice being 11M. In case (iii), we can even get a determi-
nant 41M. At any rate, the determinant is independent of Sj we get what is
called a "coarse" lattice structure. Theorem 7.6 can be used to give bounds on
the discrepancy Dr;) of the point set X., which are analDgous to the bound in
Theorem 7.3 (compare with Niederreiter [224], [234] for such bounds).
The lattice structure of linear congruential PRN provides a basis for lattice
tests that can be applied to these PRN. Strictly speaking, "lattice tests" is a
misnomer, since these are riot statistical but: rather methods of check-
ing structural properties of linear cong.ruerrtial PRN. Typicalcxa.mples Hxe the
following (see [154, 3.3.4], (295, Chap. 2] for a fuller description):
Ll. Number of hyperplanes. Determine the minimum number of parallel
hyperplanes on which all points Xn lie. This number should be as large as
possible.
L2. Distance between hyperplanes. Determine the maximum distance be-
tween adjacent hyperplanes, taken .over all families of parallel hyperplanes with
the property that they contain all points Xn. This distance should be as small
as possible. In certain cases, such as case (ii), this can be viewed as a geometric
version of the spectral test (see test G in 7.2).
L3. Ba,sis test. Consider the lattice Ls(a, M) in cases (i) and (ii) and the
lattice Ls(a, M14) in case (iii). Determine a "reduced" lattice basis of this lattice,
such as a Minkowski-reduced lattice basis, and calculate the ratio of the lengths
of the shortest and the longest basis vector. This ratio should be as close to 1
as possible.
Below, we list some deficiencies of the linear congruential method, which are
of relevance when comparing this method with other methods to be discussed
later:
(a) The modulus M, and therefore per(x
n
), are bounded in terms of the word
length of the machine. For instance, with a 32-bit processor we have per(xn) :::;
M:::; 2
32
, unless we are willing to use costly multiple-precision arithmetic.
(b) There is too much regularity in sequences of linear congruential PRN,
which stems from the simple nature of the generation algorithm. For instance,
in case (i), we have Theorem 7.4, which states that Dt>_l = O(M-l(logM)8
log loge M + 1) for most choices of multipliers. Compare this with the law of the
iterated logarithm in (7.4), according to which the discrepancy of M -1 random
points in Is should have an order of magnitude around M-
1
/2. Also, some long-
in sequences of linear congruential PRN have recently been
174 CHAPTER 7
discovered; we mention the papers of de Matteis and Pagnutti [58J, Eichenauer-
Herrmann and Grothe [84], and Percus and Percus [281]. These regularities can
create misleading results in simulations where "true randomness" is essential.
( c) The coarse lattice structure of linear congruential PRN is a strong element
of nonrandomness and can be disruptive in simulations; see, e.g., Eichenauer and
Lehn [78] and Ripley [295, pp. 55-59] for illustrations.
The idea of employing higher-order linear recursions for pseudorandom num-
ber generation is used in the multiple-recursive method. Let p be a large prime,
let k 2:: 2 be an integer that serves as the order of the linear recursion, and choose
ao, ... ,ak-l E Zp with ao # O. Then we generate a sequence Yo, Yl1" . E Zp by
the recursion
k-l
(7.10)
Yn+k = L a,Yn+l modp for n = 0,1, ... ,
1=0
where Yo, ... ,Yk-l are initial values that are not all zero. We obtain multiple-
recursive pseudorandom numbers by the normalization
Y11, I
Xn=- E for n = 0,1, ....
p
The sequences Xo, Xl,'" and Yo, Yll'" are both periodic and per(x11,) = per(y11,)'
The k-tuples (Y11" Y11,+l,' .. ,Y11,+k-l) E Z;, n = 0,1, ... , are all # 0 and cannot
be repeated within the period. Therefore perCYn) ~ pk - 1. The upper bound
pk_ 1 can be achieved as follows. View ao, ... ,ak-l as elements of the finite
field Fp of order p and let
Ie-I
J(x) = xle - L a,x
l
E Fp[x]
1=0
be the characteristic polynomial of the recursion (7.1O). Now assume that J is
a primitive polynomial over Fp in the sense of Definition A.l in Appendix A.
Then per(Yn} = pie - 1, and so T = per(x11,) = pie - 1.
Let Xc, Xl, . be a sequence of multiple-recursive PRN with primitive char-
acteristic polynomial. Then, for every dimension s ~ k, the distribution of the
points Xn = (x11"Xn+l"" ,X11,+a-l) E la, 0 ~ n ~ T - 1, or, equivalently, of
the points Y11, = (Yn, Y11,+l, ... ,Yn+a-l) E Z;, 0 ~ n ~ T - 1, can be described
explicitly. For C E Z;, let A( c) be the number of integers n with 0 ~ n ~ T - 1
and Y 11, = c. Then we have
(7.n)
A(c) = {pk-a if C # 0,
pk-a - 1 iJ c = o.
For the--proof, we observe that, since T - ph - 1, the k-tuples
(Yn, Y11,+l, ,Yn+k-r), 0 ~ n ~ T - 1, run exactly through all nonzero points
in Z;. Therefore.(1( c) is equal to the number of nonzero b E Z; that have c
RANDOM NUMBERS AND PSEUDORANDOM NUMBERS 175
as the s-tuple of their first 8 coordinates. For c f. 0, we can have all possible
combinations of elements of Zp in the remaining k - 8 coordinates of bi hence
A( c) = pk-a. For e = 0, we must exclude the possibility that all remaining k - 8
coordinates of bare 0; hence A(e) = pk-8 - 1.
Formula (7.11) establishes. an almost perfect equidistd1;>ution of the points
Yn! and so of the points Xn. Nevertheless, multiple-recursive PRNhav:e a
weakness. Note that the coordinates of all points Xn are rationals with denomi-
nator p. Therefore the star discrepancy D;Ja) = DN(Xo, ... ,XN-t} satisfies
n;J8) 1 - (1 - ) 8 for aU 8 1 and N 1
by Theorem 3.14. In particular, with N = T = pk -1, we obtain that is
at least of the order of magnitude T-1/k) which is too large for k 3 (compare
with (7.4. The underlying reason for this deficiency is, of course, that compared
to the least period length the discretization afforded by multiple-recursive PRN
is too coarse.
Notes.
Expository accounts of random and pseudorandom number generation are given
in the books of Bratley, Fox, and Schrage [34], Dagpunar [50], DeSk [53], De-
vroye [62], Kalos and Whitlock [150], Knuth [1541, and Ripley [295], and in the
recent survey articles of Anderson [7], James [146], L'Ecuyer [186], and Nieder-
reiter [251J. Refinements of Theorem 7.2 were obtained by Goldstem [114],
Niederreiter and Tichy [276], and Tichy [350]. Interesting work on complete uni-
form. distribution of sequences of digits was recently done by Flajolet, Kirschen-
hofer, and Tichy (105J, [106] and Kirschenhofer and Tichy [153], among others.
Further information on the complexity-theoretic approach to randomness can be
found in Chaitin [40], Kolmogorov and Uspenskii [157], Schnorr [304], and Us-
penskii, Semenov, and Shen' [355]. See Kranakis [162] for methods of generating
cryptographically strong pseudorandom bit strings.
Further results on the uniformity test and the serial test for linear congru-
entia! PRN, also for parts of the period, can be found in Levin [189}, Nieder-
reiter [225], [231], [234], and Shparlinskii [313]. Detailed investigations of the
two-dimensional serial test for the full period were carried out by Dieter [63] and
Niederreiter [224]. Exact calculations of the discrepancy of linear congruential
PRN were discussed by Bhavsar et aI. [27J in the one-dimensional case, and by
Afflerbach and Weilbacher [5] in the multidimensional case. Further existence
theorems for large figures of merit are shown in Larcher &!."1.d Niederreiter [179]
and Niederreiter [224]; see also Larcher [171J for a related result. Various prob-
lems on figures of merit and on the discrepancies D!;) for linear congruential PRN
are posed in Niederreiter [233]. Fishman and Moore [104] performed an exhaus-
tive search for good multipliers with the prime modulus M = 2
31
-1, whereas
Fishman [103] undertook an exhaustive search with the modulus M = 2
32
and a
partial search with M = 2
48
A systematic search algorithm for optimal multipli-
ers respect to the two-dimensional serial test was developed by Borosh and
176 CHAPTER 7
Niederreiter [30], and a different algorithm for this purpose is due to Denzer and
Ecker [60]. Applications of linear congruential and other PRN to quasi-Monte
Carlo integration were considered by Niederreiter [237].
The lattice structure of linear congruential PRN was noted by
MarsagUa [204], and detailed investigations of the lattice structure were later
done by Beyer [25], Beyer, Roof, and Williamson [26], Marsaglia [205], and
Ripley [294]. The formulation of Theorem 7.6 stems from [294]. The proof of
Theorem 7.6 shows that the formula for case (ii) holds whenever per(xn) = M.
Sometimes the structure of Xs cannot be described by a single grid, and then
a union of grids is needed. A recent paper on the lattice structure is Affler-
bach [2]. In the implementation of lattice tests, an algorithm of Dieter [64] for
calculating a shortest nonzero vector in a. la.ttice is often useful. See Afflerbach
and Grothe [3] and Eichenauer-Herrmann and Grothe [85] for recent work on
latt.ice tests,
The uniformity test and the serial test for multiple-recursive PRN were stud-
ied by Niederreiter [2221, [227] and Shparlinskii [312J, and the lattice structure
was investigated by Dieter [65] and Grube [119]. Interesting methods for pseu-
dorandom number generation that are not discussed in these lecture notes in-
clude the combining of generators (see, e.g., Collings [43], L'Ecuyer [185], and
Wichmann and Hill [358], [359J), shuffling techniques (see, e.g., MacLaren and
Marsaglia [202]), and the use of cellular automata (see, e.g., Wolfram [360]).
CHAPTER 8
Nonlinear Congruential
Pseudorandom Numbers
To overcome some of the deficiencies of the linear oongruentiai method, such as
the coarse lattice structure, new methods. for the generation of uniform PRN
have recently been designed and analyzed. The basic idea is to consider recur-
sions other than the Hn.e8.l' 'recursion used for the generation of linear
congruentiai PRN. The result is that, with suitable nonlinear recursions, the
coarse lattice structme can be broken up.
The general nonlinear congruential method is described in 8.1, and the re-
sulting PRN are analyzed by their lattice structme and by the serial test. In
this section, we also briefly discuss the quadratic oongmential method, which
has a somewhat longer history than the general nonlinear congruential method.
A particularly promising type of nonlinear congruential method is the inversive
congruential method treated in 8.2. Here the nonlinearity is achieved by rising
multiplicative inversion in modular arithmetic. Inversive congruential PRN with
a prime modulus are, in a sense, optimal with regard to the lack of a lattice struc-
ture, and they behave much better under the serial test than linear congmential
PRN.
8.1. The general nonlinear congruential method.
The general framework for the methods to be discussed in this chapter is the
following. Choose a. large positive integer lvI, called the modulus, and define
a congruential generato'!' modulo M to be a sequence Yo, Yl, ... of elements of
ZM = {O, 1, ... ,M - I}, that is generated by a congruential recursion (of arbi-
trary order and with an arbitrary feedback function). A very special instance
of this is the sequence generated by the recursion (7.5) in the linear congruen-
tial method. From a congruential generator modulo M, we derive a sequence
XO,Xl, ... of uniform PRN by the normalization Xn = Yn/1vI for n = 0, 1, ....
In the general first-order congruential method, the congruential generator
modulo M is obtained by an arbitrary first-order (or one-step) recursion
(8.1) Yn+l = f(Yn) mod M for n = 0, 1, ... ,
where f is an integer-valued function on Z}/!. In practice, f and the initial
value Yo are chosen in such a way that the sequence Yo, Yl, . .. is (purely) pe-
riodic ana per(Yn) is large. We clearly have per(xn) = per(Yn) ~ M. We can
177
178 CHAPTER 8
always achieve per(Yn) - M by starting from a periodic sequence Yo, Y!, ...
with per(Yn) = M and {Yo, YI,'" ,YM-d = ZM and defining f(Yn) = Yn+l for
o n M - 1. In practical terms, however, this particular example is useless,
since we want a function f that is computable without the prior knowledge of the
sequence Yo, Yl, . . .. Among first-order congruential methods, we can distinguish
the linear congruential method and nonlinear congruential methods, which are all
remaining methods. As in the linear congruential method, the most convenient
moduli for first-order congruential methods are primes and powers of 2.
Now let M be a prime modulus and let us write M = p. In this case, we
can identify Z M with the finite field Fp of order p. Consider a periodic sequence
Yo, Yll of elements of Fp with per(Yn) = p. Then the map n E Fp Yn E F
p
,
like any self-map of a finite field, can be represented by a uniquely determined
polynomial 9 E Fp[x] with d:= deg(g) < p. In other words, we have
(8.2) Yn = g(n) E Fp for n = 0,1, ... ,
where n is also viewed as an element of Fp. If {Yo, Yll . ,Yp-d = FPl
then 9 is a permutation polynomial of Fp! i.e., a polynomial over Fp with
{g(O),g(l), ... ,g(p -I)} = Fp. Conversely, if the sequence Yo,Y!, ... is given
by (8.2) with a permutation polynomial 9 of Fp! then the sequence is periodic
with per(Yn) = p and {yo! YI,' .. ! Yp-I} = Fp. We now restrict the attention
to this case. Note that the sequence Yo, Yl, ... can also be generated by (8,1)
with a suitable I, as was shown in the previous paragraph. The degree d of tlle
polynomial 9 plays an important role in the theory. Since 9 is a
polynomial of Fp , we must have d? 1. If d = 1, then the sequence Yo, Yl, ...
be generated by (8.l) with f being a monic linear polynomial, and so we get
linear congruential method with modulus p and multiplier a = 1 (by the
this is a very bad choice for the multiplier in .the linear congruential methqd)"\
If d > 1, then by a result from the theory of permutation polynomials (see [192;
Cor. 7.5]) d cannot divide p - 1, and so 3 d p - 2; note that this implies, in
particular, that p ? 5. If d > 1, then we speakof the sequence Yo, Yll'" as .
(first-order) nonlinear congruential generator modulo p, and the corresponding
PRN Xn = Yn/P E I, n = 0,1, ... , are called nonlinear congruential
dom numbers (with modulus p). These PRN were.first proposed by Eichenauet;
Grothe, and Lehn [76], and the viewpoint based on permutation polynomials
was introduced by Niederreiter [248].
The following basic definition makes sense for any congruential generator
modulo a prime p. For s ? 1, we view F; as an s-dimensional vector space over
Fp.
DEFINITION 8.1. For given s ? 1, a congruential generator Yo, Yl"" modulo
the prime p passes the a-dimensional lattice test if the vectors Yn - Yo, n =
1,2, ... , span F;, where
-_ Yn = (Ynl Yn+b ... ,Yn+s-l) E F; for n = 0, 1, ....
For comparison, we first consider the a-dimensional lattice test for a linear
congruential generator Yo, Yb' .. obtained from (7.5) with a prime modulus p.
NONLINEAR CONGRUENTIAL PSEUDORANDOM NUMBERS 179
From (7.B), we infer that
Yn - Yo = (11'11. - Yo)(1,a,a
2
, ,aB-I) for 11. = 1,2, ...
in the vector space F;. This shows that the vectors Y n - Yo, n = 1, 2, ... , span a
linear subspace of F; of dimension at most 1. Consequently, a linear congruential
generator with prime modulus can pass the s-dimensional lattice test at most
for s = 1. This is another instance of the weakness of linear congruential PRN
with regard to lattice structure.
The situation is quite different for nonlinear congruential generators modulo
p. Here the importance of the number d, introduced above, becomes apparent.
The theorem below was shown by Eichenauer, Grothe, and Lehn [76J; our proof
follows Niooerreiter [248].
THEOREM B.2. A nonlinear congroential generator modulo p passes the s-
dimensional lattice test if and only if s d.
Proof. Since passing the s-dimensionallattice test implies passing any lower-
dimensional lattice test, it suffices to show that a nonlinear congruential gener-
ator Yo, Yl, . .. modulo p passes the a-d.imensional lattice test and does not pass
the (d + 1 )-diroensionallattice test. Equivalently, if we put
OJ = (Yi+l - 1/j, Yj+2 - Yj,' .. ,Yj+p-I - Yj) E pC-I for j = 0, 1, ...
and Us is the s x (p - 1) matrix with rows 110, tll, .. ,Us-I, then it suffices to
show that rank(Ud) = rank(Ud+l) = d (consider the columns of Ud and Ud+l).
Let b/
c
be the k-fold iterate of the difference operator At'll. = tn+! - tn defined
on a sequence to, it, .. .. For i = 1,2, ... , we have
n=O n=O
where the last step follows from (B.2) and deg(g) = d. Therefore
d "(d)
2)-1)d-
J
Uj = 0;
j=O J
hence lid is a linear combination of 110, ... ,lid-I, and so rank(Ud+1) d. Fur-
thermore, Ad
Yn
= Adg(n) =1= 0 and = 0 for all n 2:: OJ thus Yo, Ylt . .. is
a linear recurring sequence with minimal polynomial (x - l)d+1 (compare with
Appendix A). With
Yj = (Yj, Yj+l,'" ,Yj+d) E F;+1,
we claim that Yo, y}, ... ,Yd are linearly independent over Fp. Suppose that there
exist b}, ... ,bd E Fp , not all zero, such that 'L1=o bjYj = 0. By comparing
180 CHAPTER 8
components, we see that
Ii,
I: bjg(n + j) = 0 for n = 0,1, ... ,d.
j=O
This means that the polynomial h(x) = 2:1=0 bjg(x+j) has at least d+l distinct
roots in Fp. Now deg(h) $ a, however, and so h is the zero polynomial. From
this, we obtain
d
L bjYn+j = 0 for n = 0,1, ... ;
j=O
thus the sequence Yo, Yb . .. satisfies a linear recursion of order $ d, which is
a contradiction. From the linear independence of Yo, Yb ... ,Y d" we deduce the
Hnear independence -Yo, ... ,Yd.-Yo, and since the transposes ofthe latter
vectors are the first d columns of Ud+b we obtain rank(Ua+d = d. follows
that UQ, .. ,Ud-l are linearly independent over FpI and so rank(Ud) = d. 0
Since d ~ p - 2, it follows from Theorem 8.2 that a nonlinear congruential
-generator modulo p cannot pass the s-dimensionallattice test for 8 ~ p-l. From
the viewpoint of the lattice test, a large value of d is preferable. The polynomial
9 determined by (8.2) can be written as
and so d is equal to the largest integer k $ p - 2 for which
p-l
L nP-1-kYn ~ 0 modp.
n=O
In particular, d has the maximal value p - 2 if and only if
p-l
(8.3) L nYn ~ 0 modp.
n=O
This was used by Eichenauer and Niederreiter [81 J to construct the following
example. Consider the periodic sequence Yo, Yl ... E Zp with per(Yn) = p,
which is given by Yn = n for 0 $ n $ p- 3, Yp-2 = p-l, and Yp-l = p- 2. Then
it is easily checked that {8.3} is satisfied, and so this nonlinear congruential
generatQr modulo p passes the s-dimensional lattice test for all 8 $ p - 2 by
Theorem 8.2. On the other hand, it is clear that this generator has extremely
bad statistical properties (consider, e.g' j the distribution of pairs (Yn,Yn+1) or
the run test). Therefore the lattice test is rather weak and should only be used
for the elimination of bad generators.
We now study the behavior of the full period of a sequence xo, Xl, . .. of non-
linear congruential PRN with modulus p under the uniformity test and the serial
\
NONLINEAR CONGRUENTIAL PSEUDORANDOM NUMBERS 181
test. Since in the full period we find exactly all rationals in I with denominator
p, we obviously get Dp(xo, ... ,Xp-l) = lip. To assess the performance under
the serial test, we consider the discrepancy D ~ ) defined by (7.2). We have the
following result of Niederreiter [249] in which the number d is again important.
THEOREM 8.3. For nonlinear congroential PRN withmodulu.s p,we have
D ~ 8 ) ::; 1 - (1 - ~ ) 8 + (d _ 1)p-l/2 (11'; logp + 1.72) 8 for 2::; s ::; d.
Proof. If Yo, Yl, ... E Fp is the corresponding nonlinear congruential gener-
ator modulo p, then we have shown in the proof of Theorem 8.2 that this is
a linear recurring sequence 'with minimal polynomial (x - l)d+1. For a given
s with 2 ::; s ::; d, put Yn = (Yn, Yn+b'" , Yn+s-d for n ~ O. For a fixed
h = (hl, ... ,hI!) E C:(p), we consider

h Yn = E hiYn+i-l
i=l
as an element of Fl" Suppose that there would exist abE Fp such that
,
B
L h"Yn+i.-l = b for all n ~ O.
i=l
Then, by applying the difference operator fj, on both sides, it follows that the
sequence Yo, 1/1, . . satisfies a nontrivial linear recursion of order ::; s, which
is a contradiction. Thus, in conjunction with (8.2), we see that there exists a
polynomial G E Fp[x] with 1 ::; deg(G) ::; d < p such that h Yn = G(n) for all
n ~ O. From Weil's bound for character sums (see [192, Thm. 5.38]), we then
obtain
Therefore the hypothesis of Corollary 3.11 is satisfied with B = (d _1)p1/2, and
the result of the theorem follo'il'lS. 0
Further results in [249] show that Theorem 8.3 is, in general, the best possible,
in the sense that D ~ 8 ) can be of an order of magnitude at least p-l/2 and that
the bound cannot hold for dimensions s > d. Similar theorems for parts of the
period are also available in [249]. Note that the bound in Theorem 8.3 is only
useful if d is at most of a somewhat smaller order of magnitude than pl/2.
A special first-order congruential method that has received some attention
is the quadmtic congruential method proposed by Knuth [154, 3.2.2]. Here we
use (8.1) with the modulus M = 2, Ot ~ 2, and with f being the quadratic
polynomial f(x) = ax2 + bx + c, where a, b, c E ZM. For the corresponding
sequence Xo, Xl!'" of quadmtic congruential pseudomndom numbers, we have
per(xn)=_Y if and only if a is even, b - a + 1 mod4, and c is odd, according
182 CHAPTER 8
to a result in [154, 3.2.2]. In the following discussion, we always assume that
per(xn) = M.
The lattice structure of quadratic congruential PRL'J was analyzed by
Eichenauer and Lehn [79]. For a given dimension s 2:: 2, the set of all nonoverlap-
ping s-tuples Xns = (xns , Xns+l,'" ,Xns+s-l) E IS, n = 0,1, ... ,M - 1, is the
same as the intersection of J8 with a union of 2""-7] grids with explicitly known
shift vectors and lattice bases. Here 1} and ware determined by gcd(s, M) = 2
17
,
gcd(a, M) = 2;3, and w = - {3 + l)J,1}).
Eichenauer-Herrmann and Niederreiter [88] investigated the performance of
quadratic congruential PRN under the two-dimensional serial test. Let Dr; be
the discrepancy defined by (7.2) for s 2:: 2 and the full period, and let {3 be as
above. Then, for s = 2, we have
I ",)2
T
5
This is the best possible in the sense that, if {3 ::; a - 2 and b == 1 mod 2;3+1, then
D(s) > 1 2/3/
2
M-
1
/
2
for all s 2:: 2.
M - (n + 2)V2
The upper bound for suggests that it is reasonable to choose the parameter
a in such a way that (3 = 1, Le., that a = 2 mod4. Then the criterion for
per(xn) = M implies that b = 3 mod 4, and, in this case, we have the lower
bound
for all s 2:: 2.
The resulting order of magnitude of is in line with the probabilistic law (7.4)
for random sequences.
8.2. The inversive congruential method.
In the inversive congruential method, we use the recursion (8.1) with a simple
class of functions f, which involve multiplicative inversion in modular arithmetic.
As usual, convenient moduli are primes and powers of 2. We first consider
the case of a prime modulus M = p 2:: 5. For c E ZPl we define C E Zp by
cC = 1 modp if c i= 0 and c = 0 if c = O. Now we choose parameters a, b E Zp
with a i= 0 and an initial value Yo E Zp. Then the sequence Yo, Yb ... E Zp
generated by the recursion
(8.4) Yn+! = l1/fjn + b modp for n = 0,1, ...
is called an inversive congruential generator modulo p, and the corresponding
PRN Xn = Yn/P E I, n = 0,1, ... , are called inversive congnumtial pseudoran-
dom numbers (with modulus p). These PRN were introduced by Eichenauer and
Lehn [78]. Since a i= 0, the congruence (8.4) can be solved uniquely for Yn if Yn+l
is given, and so the sequence Yo, Yll ... is periodic and per(xn) = per(Yn) ::; p.
NONLINEAR CONGRUENTIAL PSEUDORANDOM NUMBERS 183
The following sufficient condition for per(xn) = p was shown in [78]. We again
identify Zp with the finite field Fp! and we recall the definition of a primitive
polynomial. over Fp from Appendix A.
THEOREM 8.4. 1/ a, bE Fp are such that x2-bx-a is a primitive polynomial
over Fp,then the sequence xo, Xl, of inversive congruential PRN with modulus
p satisfies per(xn) = p.
Proof The sequence Co, Cl,'" E Fp defined by Co = 0, Cl = 1, Cn+2 =
0Cn+1 + 0Cn for n ;::: 0 is a maximal period sequence in Fp (see Appendix A);
hence it is periodic with per(Cn) = p2 - L Let"Y be a root of X2 - bx - a in F
p
2.
Then, by Theorem A.2 in Appendix A, there exists a (J E Fp2 with (J =1= 0 such
that
(8.5) en = 8"(n + f)P",/pn for n = 0,1, ....
, Since Co = 0, (8.5) shows that f)p-l = -1. Now suppose that we had em =
for some m with 1 ::; m ::; p. Then (8.5) yields ,/,(p-l)m = _f}l-p = L Again,
by (8.5), we then get cn+(p- = en for n = 0, 1, ... ; thus pe:r(en) :::; (p-l)m ::;
(p - l)p < p2 - 1, a contradiction. Thus Cn =1= 0 for 1 :::; n :::; p.
Let Yo, YlI'" be generated by (8.4) and consider first the case where Yo = 0.
Then, using Cn =1= 0 for 1 ::; n ::; p, we obtain by induction on n that Yn = CnCn+1
for 0::; n ::; p. Consequently, we have Yn =1= 0 for 1 ::; n ::; p-l, and so per(Yn)
p; hence per(xn) = per(Yn) = p. In particular, {Yo! Yb'" ,Yp-l} . Zp. If we
have an arbitrary initial value Yo, then the sequence Yo, Yl, ... is a shifted version
of the sequence with initial value 0, and so again per(xn) = per(Yn) = p. 0
We henceforth assume that a, b E Fp have been chosen in such a way that
x2-bx-a is a primitive polynomial over Fp. Since per(Yn) = p, we are then in the
framework developed in 8.1 for prime moduli. Inversive congruential generators
modulo p behave very well under the lattice test. The following result of Nieder-
reiter (248] represents a slight improvement on an earlier theorem of Eichenauer,
Grothe, and Lehn [76] ..
THEOREM 8.5. An inversive congruential generator modulo p passes the s-
dimensional lattice test Jor all 8 ::; (p + 1) /2.
Proof. If 9 E Fp[x] is as in (8.2), then, by Theorem 8.2, it suffices to show
that d = deg(g) (p + 1)/2. Put Wn = YnYn+! - byn - a for n 0; then
Wn is represented by a. polynomial over Fp of degree 2d. From (8.4), it follows
that Wn = 0 for p - 1 values n E F
p
, namely, for those with Yn i= O. Thus
we must have p - 1 2d. Now d = (p - 1)/2 is impossible by a property of
permutation polynomials of Fp stated in 8.1, and so d 2: (p+l)/2. The fact that
d = (p - 1)/2 is impossible can alsobe seen directly as follows. If d = (p - 1)/2,
then g(x)2 = ajx
j
with all aj E Fp and ap-l =1= O. In Fp, however, we have
p-1 p-1 1'-1 p-l p-l
0= 2:
n2
= = :Lg(n)2 = Laj 2:
nj
= -ap _!,
n=O n=O n=O j=O n=O
which a contradiction. 0
184 CHAPTER 8
The following strong nonlinearity property of inversive congruential genera-
tors modulo p was shown by Eichenauer-Herrmann [82]. Here we view F; as an
affine space.
THEOREM 8.6. For s ~ 2 and every inver'sive congruential generator
Yo, YI, .. modulo p, any hyperplane in F; contains at most s of the points
Yn = (Yn! Yn+b ... ,Yn+s-I) with 0 :$ n :$ p - 1 and Yn'" Yn+s-2 =I o.
Proof. Let Uo, U1, . .. E Fp be generated by (8.4) with Uo = 0 and put
d
j
=-auj E Fp for j ~ O. From {uo,Ul, ... ,Up-I} = Fp! it follows that
do! db ... ,dp-I are distinct. Define 'I/J( n) = an + b E Fp for n E Fp and let 'l/Jj be
the jth iterate of the function 'I/J, with 'l/J(n) = n. By induction on j, we have
. n-d
'l/J'(n) = Uj J
n - d
j
-
1
(8.6) for I :$ j :$ p - 1
and whenever n =I d
i
for 0 ~ i :$ j - 1. Since the
can assume that s < p. We have
hury,,oln11 is trivial for s ~ p, we
(8.7) {Yn: 0 :$ n:$ p - 1} = (('l/J(n), 1/i(n), ... , 'l/JlJ-l(n)) : 0 :$ n :$ p - I},
and the condition Yn' .. Yn+s-2 =I 0 amounts to the condition n =J: d
i
for 0 :$
i :$ s - 2 in the second set. Now let a hyperplane H in F; be defined by
2:;=1 bjZj = bo with bo, bl , ... 1 bs E Fp and (bI, ... ,bs ) =I O. Then (8.6) shows
that, for n =I d
i
, 0 :$ i :$ s - 2, we have ('l/J(n), 'l/Jl(n), ... , 'l/J8-1 (n) E H if and
only if
2:
8 n - d
j
-
1
bIn + bjUj-l d = bo-
n- . 2
j=2 .1-
Clearing denominators, we see that this is equivalent to h(n) = 0, where the
polynomial hover Fp is given by
8 S 8
h(x) = (b1x - bo) II (x - dj - 2 ) + L bjUj_I (x - dj - 1) II (x - di - 2 ).
j=2 j=2 ;=2
i#j
Since deg(h) :$ s, the result of the theorem follows if we can show that h is
not the zero polynomial. If h were the zero polynomial, then, by looking at the
coefficient of xS, we would get hI = O. Furthermore, for 2 :$ k :$ s, we would
obtain
8
0= h(dk - 2 ) = bk
Uk-l(dk-2 - dk - 1) n (dk - 2 - di - 2 )i
i=2
i#k
hence bk = 0, a contradiction to (bI, ... ,b
s
) =J: 0, 0
Note- that any s points 'in F;' define a hyperplane or a lower-dimensional
affine subspace, and so Theorem 8.6 is optimal in the sense that there do exist
hyperplanes in FJ that contain exactly s of the points Y n considered in this
NONLINEAR CONGRUENTIAL PSEUDORANDOM NUMBERS 185
theorem. We remark also that the condition Un' .. Yn+s-2 =tf 0 in Theorem 8.6
elimina.tes exactly 8 - 1 of the points Y n'
SUice inversive oongruentiai PRN with modulus p are, in particular, nonlinear
oongruentiai PRN with modulus p, we could hope to deaiwith the serial test for
inversive oongruential PRN by simply referring to Theorem 8.3. However, for
these PRN, we have d :? (p+ 1) /2 by the proof of Theorem 8.5, and so the bound
in Theorem 8.3 becomes useless. Therefore the case of inversive congruential
PRN must be treated by a specially adapted method. The following isa slight
variant of a result of Niederreiter [252].
THEOREM 8.7. For inversive congruential P RN with modulus p, we have
n(s) 1 _ (1 _ !) iii + (28 -2 + s - 1) + L 72) iii
p \ P pl/2 P 11"2
for 8 2:: 2.
Proof. Let Yo, Yl, . .. be the corresponding inversive oong;ruential generator
modulo p, and, for a given 8 2, put Yn = (Yn,Yn+h ... ;Yn+liI-d forn O.
Since the theorem is trivial for s P, we can assUme that 8 < p. For a fixed
h = (hI,'" , hs ) E C:(p), we use (8. to obtain
Let k be the least index with hk =I O. If k = s, then, since 'if;B-l is a permutation
of Fp , we get S(h) = O. For k < s, we simplify the writing by introducing the
additive character xCc) = e(c/p} for c E Fp. We have
iii iii iii
= = (1,bk-l(n)) I
;=1 j=k j=k
and, using 'if;h-l(n) as a new summation variable, we derive
Now we apply (8.6) for n =I d
i1
0 i s - k - 1, which yields
The last expression is a rational function Q/R of n with Q E Fp[x] and R(x) =
- dj - 1 ) E Fp[xJ. Therefore
IS (h) I s - k +
(Q(n))
L- X R(n)
nEFp

186 CHAPTERB
The rational function Q / R has at most 8 - k finite poles, each of multiplicity
1, and, since deg(Q) = deg(R) + 1, it has a pole at the point at infinity of
multiplicity 1. Thus a well-known bound of Weil [357] (see also Stepanov [335J
for an elementary proof) yields
L x ( ~ i : D II ~ 2(s - k)pl/2 ~ (23 - 2)pl/2.
nEFp
R(n);i:O
Altogether, we obtain
IS(h)1 ::; (2s - 2}pl/2 + s - 1 for all hE C:(p).
Thus we can apply Corollary 3.11 with B = (2s - 2)pl/2 + s -- 1, and this yields
the desired result. 0
The upper bound in Theorem B.7 has the order of magnitude p-l/2{logp)B.
This is, in general, the best possible up to the logarithmic factor, as the following
result of Niederreiter [255] demonstrates. Note that (p2 - 1)/2 is the total
number of primitive polynomials x
2
- bx - a over Fp by [192, Thm. 3.5].
THEOREM 8.8. Let p ~ 5 be a prime and let 0 < t < 1. Then there are more
than Ap(t)(p2 - 1)/2 primitive polynomials x
2
- bx - a over Fp such that, for
the corresponding inversive congrneniial PRN with modu.lus p, we have
D(s) > t p-l/2 for all s ~ 2,
P 21r + 4,
where the numbers Ap(t) satisfy
1 - t
2
lim Ap(t) = -4 2 > 0 for all t.
p .... oo - t
A comparison between the linear congruential method and the inversive con-
gruential method with prime modulus reveals that the latter has at least the
following three advantages:
(i) Inversive congruential PRN are vastly superior with respect to lattice
structure. Once a, b E Fp have been chosen such that x
2
- bx - a is primitive over
F
p
, then Theorem 8.5 guarantees that the lattice test is passed for a very large
range of dimensions, and Theorem 8.6 ensures strong nonlinearity properties. In
contrast, for the linear congruential method, laborious calculations are needed
to find multipliers that yield a nearly optimal lattice structure even for a modest
range of dimensions. For instance, an inversive congruential generator modulo
p = 2
31
- 1 passes the s-dimensional lattice test for all 8 ~ 2
30
, whereas, for the
linear congruential method with this modulus, it is already difficult to guarantee
a nearly optimal lattice structure for all s .s; 10 (compare with Fishman and
Moore [10-4]).
(ii) Inversive congruential PRN show a better behavior under the serial test.
Theorems 8.7 and 8.8 entail that, for a positive fraction of the possible pairs
NONLINEAR CONGRUENTIAL PSEUDORANDOM NUMBERS 187
(a, b) E F; of parameters, and, in particular, on the average, D ~ B ) has an order of
magnitude betweenp-l/2 andp-l/2(1ogp)B for every s ~ 2. This is in reasonable
accordance with the law of the iterated logarithm in (7.4). On the other hand,
in case (i) of the linear congruential method with prime modulus p and least
period lengthp-l, we have on the average D ~ ~ l =D(p-l(logp)8log1og(p+l))
for every s ~ 2 by Theorem 7.4. Thus sequences of inversive oongruential PRN
display more irregularity in their distribution and, in this sense, model truly
random numbers more closely than linear congruential PRN.
(iii) The inversive oongruential method allows a wide choice of parameters, all
of which lead to guaranteed and comparable structural and statistical properties.
It suffices to choose a, b E Fp such that X2 - bx - a is primitive over Fp, and then
the properties in Theorems 8.5, 8.6, and 8.7 hold without any further restriction.
This feature can be of great practical value in various parallelized simulation
techniques in which many parallel streams of PRN are needed. In case (i) of the
linear congruential method with prime modulus p, just choosing any primitive
root a modulo p is certainly not enough. Rather, the multiplier a must be
I:ltJlec[;ea very this selection process is a
task.
A computational issue that arises in connection with the recursion (8.4) is
the efficient calculation of c for given c E Fp. One method is based on the
observations that c = cP-
2
for all c E Fp and that c
p
-
2
can be calculated
with O(logp) multiplications in Fp by using the standard square-and-multiply
technique (see [193, p. 347]). A second method calculates c for c =1= 0 by using the
Euclidean algorithm with the integers c and p, and this algorithm will terminate
after O(logp) steps (see [154, 4.5]).
The inversive congruential method can also, be used with a composite modulus
M. Let GM be the set of c E ZM with gcd(c;:M") = 1. For c E GM, let c be the
unique element of GM with cC _ 1 modM. With an initial value Yo E,GM and
parameters a E G M and b E Z M, a sequence Yo, Yl, ... E G M is generated by the
recursion
(8.8)
Yn+l = aYn + b mod M for n = 0, 1, ....
The integers a, b must be chosen in such a way that each Yn is guaranteed to be
in GM. In this case, the sequence Yo, Yl, ... is called an inversive congruential
generator modulo M, and the corresponding PRN Xn = Yn/M E I, n = 0,1, ... ,
are called inversive congruential pseudorandom numbers (with modulus M). In
the case of greatest practical interest, namely, when M is a power of 2, these PRN
were introduced by Eichenauer, Lehn, and Topuzoglu [80]. Since gcd(a, M) = 1,
the sequence Yo, Yl, ... generated by (8.8) is periodic and per(xn) = per(Yn) S
card(GM ) = (M).
Now let M be a power of 2. Then a = Yo = 1 mod 2, and we have Yn E G M
for all n ~ 0 if and only if b = 0 mod 2. From the above general bound, we obtain
per(xn) S M/2. The following criterion for per(xn) = M/2 was established by
Eichenauer, Lehn, and Topuzoglu [80].
188 CHAPTER 8
THEOREM 8.9. If M 2:: 8 is a power of 2, then a sequence Xo, Xl, .. ' of
inversive congruential PRN with modulus M satisfies per(xn) = M/2 if and
only if a = 1 mod 4 and b == 2 mod 4.
Proof. Ifper(xn) = M/2, then {YO,YI, ... ,Y(M/2)-tl = GM, and so we can
assume that Yo = 1. If we consider the sequence Yo, YI, ... modulo 4, then it has
least period length 2; hence Y2 == 1 mod 4. If we consider this sequence modulo
8, then it has least period length 4; hence Y2 1 mod 8, and so Y2 _ 5 mod 8.
Since c = c mod 8 for c E GM
, it follows from (8.8) that
Y2 = a( a + b) + b = 1 + (a + l)b mod 8,
and so (a + l)b = 4 mod 8. This implies that a = 1 mod 4 and b = 2 mod 4.
Conversely, suppose that a = 1 mod 4 and b = 2 mod 4 and consider first
the case where Yo = L For M = 8, it is checked by the above .arguments that
per(Yn) = 4. Now NJ = 20< with Ot 2:: 4. Define the sequence Co, el, ... E ZM
by Co = Cl = 1 and
(8.9) Cn+2 == bCn+l + aCn modM for n = 0,1, ....
Then Cn = 1 mod 2, and so Cn E G M for all n 2:: 0. By induction on n, we obtain
(8.10) Yn = CnCn+l mod M for all n 2:: O.
With the integer matrix
- we see, from (8.9), that
( ) = A ) modM
all n 2:: 0,
and so
(8.11)
( ) = An ( i ) mod M for all n 2:: O.
By induction on h 2:: 4 and using a = 1 mod 4, b = 2 mod 4, it is shown that
for all h 2:: 4
with suitable IE Z. By applying (8.12) with h = Ot + 2 and also using (8.11),
we see that Cn+M = en for all n 2:: 0, and so Yn+M = Yn for all n 2:: 0 by (8.10).
Thus divides lvI, and we already know thatper(Yn) :5 M/2. Hence,
to prove tha.t per(Yn) = M/2, it suffices to show that per(Yn) > M/4. If we
had per(Yn} :5 M/4, then YM/4 = Yo = 1, and so C(M/4)+1 = CM/4 by (8.10).
From (8.11), howeter, with n = M/4 and (8.12) with h = a, we obtain the
NONLINEAR CONGRUENTIAL PSEUDORANDOM NUMBERS 189
contradiction C(M/4)+1 = CM/4 + (M/2) modM, and so per(Yn) = M/2 is proved
if Yo = 1. In particular, {YO,Y1, ... ,Y(M/2)-d = GM. IT we have an arbitrary
initial value Yo E GM, then the sequence YO,Yl, ... is a shifted version of the
sequence with initial value 1, and so again per(xn) = per(lIn) = M/2. 0
According to a result of Eichenauer-Herrmann et aI. [87], the lattice structure
of inversive congruential PRN with modulus M = 2
a
, a ~ 3, and with a =
1 mod 4 and b = 2 mod 4 can be described as follows. For a given dimension s ~
2, the set of all nonoverlapping s-tuples Xns = (xns , Xns+b' .. j xns+s-d E J8,
n = 0, 1, ... ,(M/?) - 1, is the same as the intersection of JB with a union of
2"'-'1-
1
grids with explicitly known shift vectors and lattice bases. Here 'fJ and
ware determined by gcd(s, 2
a
-
1
) = 2'1 and w = maxUHa + I)J, 'fJ + 1}.
Again for M = 2
a
, a ~ 3, and a = 1 mod 4, b == 2 mod 4, the behavior of in-
versive congruential PRN with modulus M under the two-dimensional serial test
was studied by Niederreiter [252]. IT nt>/2 is the discrepancy defined by (7.2) for
s ~ 2 and the full period, then, for s = 2, we have D ~ / 2 = O(M-1/2(10gM)2)
-vvith an absolute implied constant. Eichenauer-Herrmann Niederreiter [891
have shown that this bound is the best possible in the sense that, for a positive
fraction of all possible pairs (a, b) of parameters, the discrepancy D<j;2 is at least
of the order of magnitude M-
1
/
2
for all s ~ 2.
Notes.
The s-dimensionallattice test in Definition 8.1 can be viewed as a special. case
of a lattice test proposed by Marsaglia [205]. For more information on permuta-
tion polynomials, we refer to [192, Chap 7]. Quadratic congruential PRN with
arbitrary modulus were studied by Eichenauer and Lehn [79]. The least pe-
riod length of sequences of inversive congruential PRN with an arbitrary prime
power modulus was discussed by Eichenauer-Herrmann and Topuzoglu [90], and
Eichenauer-Herrmann [83J treated the two-dimensional serial test for odd prime
power moduli. The proof of Theorem 8.8 depends on distribution properties of
the values of Kloosterman sums over finite fields; see Niederreiter [256] for a
detailed study of such properties. Parameters a, b E Fp such that x
2
-"bx - a is a
primitive polynomial over Fp are tabulated in Grothe [118] for some large primes
such as p = 2
31
- 1. A generalization of the inversive congruential method, which
is based on higher-order recursions, was introduced by Eichenauer et a1. [77].
CHAPTER 9
Shift-Register Pseudora dorn
Numbers
Pseudorandom numbers generated by higher-order linear recursions have already
been considered briefly in 7.3 in the form of multiple-recursive PRN. However,
the method we 'used there to transform a linear recurring sequence into a sequence
of uniform PRN 1 namely normalization, is not quite satisfactory, and much better
lli'B 3;v-ailable pmpose. The bask of 1S to
employ a small prime modulus p (such as p = 2) for the generation of the linear
recurring sequence and then take certain collections of terms of this sequence
as digits of the PRN to be constructed. This has the advantages that modular
arithmetic must be performed only with respect to a small modulus and that
there is much flexibility as far as the discretization of the PRN is concerned.
There are two methods of this type that have received much attention in the
literature, namely, the digital multistep method and the GFSR (for "generalized
feedback shift register") method. The PRN generated by these two methods have
comparable properties, and these PRN are collectively known as shift-register
pseudorandom numbers. One common feature of these PRN is that they tend
to have strong uniformity properties, which can be made explicit in terms of
the theory of nets described in Chapter 4. The reference to shift registers in the
terminology stems from the fact that linear recurring sequences in finite fields can
be generated by simple switching circuits called (linear feedback) shift registers
(compare with [192, 8.1]).
In 9.1 we discuss the digital multistep method, and in 9.2 the GFSR
method. In the latter section, we also perform a comparative analysis of lin-
ear congruential PRN and shift-register PRN.
9.1. The digital multistep method.
Let p be a small prime (usually p = 2), let k ~ 2 be an integer, and generate a
kth-order (or k-step) linear recurring sequence Yo, Yll .. . E Zp by
(9.1)
k-l
Yn+k == L aZYn-l-1 modp for n = 0,1, ... ,
1=0
where Yo, . .. ,Yk-l are initial values that are not all zero. The integer coefficients
ao, . . _ ~ , ak-l in (9.1) are chosen in such a way that, ifthey are viewed as elements
191
192 CHAPTER 9
of the finite field F
I'
, then the characteristic polynomial
k-1
f(x) = xk - L a,x
l
E FI'[x]
1=0
ofthe recursion (9.1) is a primitive polynomial over FI" We then have per(Yn) =
pk _ 1 (compare with 7.3 and Appendix A).
In the digital multistep method due to Tausworthe [342], the sequence
Yo, Yl, ... is transformed into a sequence xo, Xl, . of uniform PRN in the fol-
lowing way. Choose an integer m with 2 ~ m :5 k and put
m
(9.2) Xn = LYmn+i-1p-i for n = 0,1, ....
j=1
In other words, the numbers Xn are obtained by splitting up the sequence
Yo, Yl, .. , into consecutive blocks of length m and then interpreting each block
as the digit expansion in base p of a number in I = [0,1). The numbers Xn are
called digital multistep pseudorandom numbers, or digital k-step pseudorandom
numbers if we want to emphasize the value of k.
LEMMA 9.1. The sequence XO,Xl, defined by (9.2) is periodic with
pk -1
per(xn) = d( k 1)'
gc m,p -
Proof. Put T = pk_l and d = gcd(m, T). Then, from (9.2) and per(Yn) = T,
we get Xn+(T/d) = Xn for all n ~ 0; thus Xo, xl, ... is periodic and Tl = per(xn)
divides Tid. From Xn+T
1
= Xn for all n ~ 0 and (9.2), we infer that
Ymn+j-l+mTl = Ymn+j-l for all n ~ 0 and 1 :5 j :5 m;
hence Yn+mT
1
= Yn for all n ~ O. This implies that T divides mT1 ; thus Tid
divides TI , and so Tl = Tid. 0
We henceforth assume that gcd(m,pk - 1) = 1, and we thus guarantee by
Lemma 9.1 that the sequence xo, Xl,." of digital k-step PRN satisfies per(xn) =
p.!c _ 1 = T. To study the behavior of these PRN under the uniformity test and
the serial test, we consider, for a given dimension s 2:: 1, the points
(9.3) Xn = (xn' Xn+b." ,xn+a-d E IS for n = 0, 1, ... ,T - 1.
For small dimensions, Le., for 8 :5 kim, the distribution of these points can be
described exactly. Note that the coordinates of all points Xn are rationals with
denominator pm.
THEOREM 9.2. Let s ~ kim and let t E IS be such that all coordinates of
t are rationals wUh denominator pm. If Z (t) is the number of integers n with
o :5 n :5 T - 1 and Xn -:- t, then
Z(t) = {pk-ma if t =J. 0,
pk-ms -1 if t = O.
SHIFT-REGISTER PSEUDORANDOM NUMBERS 193
Proof. The sequence Yn = (Yn,YnHI'" ,Yn+ms-1) E z;:a, n = 0,1, ... ,
satisfies per(Yn) = T, and, from gcd(m, T) = 1, it follows tha.t the points Ymf'ld
o ~ n :5 T - 1, form a. rearrangement of the points Yn, 0 :5 n :5 T - 1.
Since rns :5 k, formula. (7.11) can be applied with s replaced by ms, and the
same formula. also holds for the counting function referring to the points Y mfU
o ~ n :5 T - 1. From (9.2) and (9.3), we see tha.t Xn = t if and only if Ymn = c
for a. suitable C E z;:a depending only on t, with c = 0 if and only if t = 0, and
so we obtain the desired result. 0
If D;(8) denotes, 88 in (7.3), the star discrepancy of the point set in (9.3),
then from Theorem 9.2 we obtain
(9.4)
by a. straightforward argument; see Niederreiter [241 J for the details.
For 8 > kim, the distribution of the points in (9.3), and thus the
s-dimensional serial test, can be analyzed by means of the theory of nets in
Chapter 4. The only provision we must make is that we must add the point 0
to the point set (9.3). Note that f has a root (;0
in the extension field Fq of Fpl where q = pk. Consider the system
(9.5)
C
1
= {a(i-l)m+j-l : 1 :5 i :5 8, 1:5 j :5 m}
of elements of Fq. If we view Fq as a k-dimensional vector space over FPl then
we can assign to C
1
the number p(Cd according to Definition 4.27.
DEFINITION 9.3. For s > kim, the figure of merit r(a)(f, m) depending on
s, the characteristic polynomial f, and m is defined by
REMARK 9.4. The value of r(s)(j, m) does not depend on the choice of the
root a of f. Note that f as a primitive polynomial over F',u is, in particular,
irreducible over Fp. The residue class field Fp[xJl(J) can be viewed as a vector
space over Fp. Then a subsystem {a(i-1)m+
j
-l : 1 :5 j ::; d
i
, 1 :5 i :5 s}
of C
1
is linearly independent over Fp if and only if the corresponding system
{x(i-1)m+
j
-1 + (I) : 1 ::; j :5 d
i
, 1 ::; i ::; s} of residue classes is linearly
independent over P
p
. Now it suffices to observe that the latter statement does
not depend on a:.
THEOREM 9.5. For 8 > kim, the ph points 0, xo, xI, ... ,XT-l form a
(t, k, s)-net in base p with t = k - r(B) (I, m).
Proof. We first show that the pk points in the theorem can be obtained by the
general construction of (t; k, s )-nets in base p leading to (4.25). In the notation
of this construction, we obtain, from (9.2) and (9.3), that; for 1 ::; n :5 T and
1 :5 i ::; s, we have
(i) _
Ynj - Ym(n+i-l)+j-l for 1 :5 j ::; m,
Y ~ 1 = 0 for m < j ::; k.
194 CHAPTER 9
If Tr denotes the trace from Fq to Fp, then, by Theorem A.2 in Appendix A,
there exists a () E F; such that
Yn = Tr(8a
n
) for n = 0, 1, ....
It follows then that, for 1 n T, 1 i s, and 1 j m, we have
= Tr(8a
mn
a(i-l)m+i -l).
Put rn = (}a
mn
and let {,BOl' .. , ,Bk-l} be a basis of Fq over Fp. Then
k-l
rn = L br(rn),Br
1'"=0
k-l
= L brC"Yn) 'ft(Pra(i-l)m+i-l)
r=O
for 1 n T, 1 i s, and 1 j m. From gcd(m, T) = 1, it follows that
am is a generator of the cyclic group F;, and so the 'Yn, 1 n T, run exactly
through F;. If we now put 'Yo = 0, then the k-tuples (bo(rn), ... ,bk-l(rn)),
o n T, run exactly through F;. Therefore we get the point set in the
theorem if, in the construction of (t, k, s)-nets in base p leading to (4.25), we
choose R = Fp and the bijections in (N2) and (N3) to be identity maps, and if,
for 1 i s and 0 r k - 1, we put
= Tr(,Bra(i-l)m+i-l) for 1 j m,
c(i) = 0 for m < J' k.
Jr
Note that the inclusion of rO = 0 amounts to the inclusion of 0 in the point set.
L 0 {
(i) 1 < ' < 1 < . < k} 'th (i) (i) (i) Fk
et = C
j
: _ z _ s, _ J _ WI C
j
= c
jO
"'" cj,k-l E p .
Then, by virtue of Theorem 4.28, it suffices to show that p(O) = r(a)(j, m). Let
0' = {C}i) : 1 i s, 1 j m} and note that c}i) = 0 for 1 i s and
m < j k. We claim that p(O) = min(m,p(O')). This is trivial if p(C') m.
If pC G') > m, then, from the linear dependence of cll) , ... ,c!,;), = 0 over
F
p
, it follows that p(C) mj however, any system {C}i) : 1 j d
i1
1 i s}
with di 0 for 1 i s and E:=l di .= m is linearly independent over Fp!
and so p(C) = m. Next we claim that p(C') = P(Ol)' Consider a subsystem
{C;i) : 1 j d
i
, 1 s} of C' with 0 d
i
m for 1 i s. It suffices
to prove that this subsystem satisfies a linear dependence relation over Fp if and
only if the corresponding subsystem {a(i-l)m+
j
-l : 1 j d
il
1 i s} of
C
l
satisfies the same linear dependence relation. Suppose that we have
8 di
t= L b}i) c}i) = 0 with all b}i) E Fp.
i=l j=l
SHIFT-REGISTER PSEUDORANDOM NUMBERS 195
By comparing components and using the formula for the we see that the
relation above is equivalent to
o d.
L I: 0 for 0 :5 ". :5 k - 1.
i=1 j=1
By the Fp-linearity of the trace, this is, in turn, equivalent to
Tr(.8
r
tt. b),),,('-I)=+;-1) = 0 for 0:5".:5 k - 1.
Since {Po, ... ,Pk-d is a basis of Fq over Fp and the trace is surjective, the last
condition holds if and only if
B d.
L L b;i)CL(i-l)m+j-l = 0,
i=1 j=1
and the claim is verified. Altogether we have
peG) = min(m, p(G' = min(m, p(Gl )) = ".(0)(1, m).
o
COROLLARY 9.6. For digital k-step PRN, we have, for 8 > kim,
(
1 1) -r < D*(o) _ O( 0-1 -r)
2" - 2p P - T - ". P
with". = ".(0)(1, m), where the implied constant depends only on p and s,
Proof. In the proof of Theorem 9.5, we have shown that the points
O,Xo,Xl,'" ,XT-l form a point set of the type in (4.25). Therefore the
lower bound for the star discrepancy of the point set consisting of
Xo, Xl! ... ,XT-l follows from Theorem 4.30 and Remark 4.31. The upper bound
is obtained from Theorems 4.5 and 4.6 and from the argument in the proof of
Theorem 7.3, which takes care of the deletion of 0 from the net. 0
Theorem 9.5 and Corollary 9.6 are due to Niederreiter [244]. The upper
bound in Corollary 9.6 can be made explicit by the results in 4.1 (compare also
with [244, Thm. 9.4]). It is evident from Theorem 9.5 and Corollary 9.6 that the
larger the figure of merit, the better the distribution properties of the points X
n
.
The following theorem, which is a special case of a result of Niederreiter [246],
shows that large figures of merit can always be achieved by a suitable choice of f.
THEOREM 9.7. For any prime p and any 8, k, m with s > kim, there exists
a primitive polynomial lover Fp with deg(l) = k and
(0)(1) . (ll (pk -1) J)
r ,m 2:: mm m, ogp f 1)( ) 1 '
\ms - m + 1 B-
logp denotes the logarithm to the base p.
196 CHAPTER 9
Proof. Put
d -llO (Pk - 1) J
- gp (ms - l)(m + 1)8-1 .
If d ~ 0, then there is nothing to prove, so we can assume that d ;::: 1. It suffices
to show that there exists a primitive polynomial over Fp of degTee k for which the
corresponding system C
1
in (9.5) satisfies p(Gl) ;::: d. Fix an s-tuple (db'" I d
8
)
of integers with 0 ~ d
i
~ m for 1 ~ i ~ s and L::=l d
i
=. d. The system
{x(i-l)m+i-l + (j) : 1 ~ j ~ d
il
1 ~ i ~ s} in Fp[xJl(f) is linearly dependent
over Fp if and only if there exist b)i) E Fp, 1 ~ j ~ dill::; i ~ s, which are not
all zero, such that I(x) divides
8 d.
L l: b)i)x(i-l)m+j-l.
i=l j=l
There are less than pd choices for the b)i) I and, for each such choice, there exist
at most (ms - l)/k primitive polynomials f over Fp of degree k that divide
the above polynomial. Furthermore, an upper bound for the number of 8-tuples
(d
l
, . .. ,d
8
) satisfying the above conditionS is (m + 1) 8-1. Thus it follows from
Remark 9.4 that the number Ak(d) of primitive polynomials over Fp of degree k
with peG
l
) ~ d - 1 satisfies
From the definition of d, we obtain Ak(d) < (Pk - l)/k, which is the total
number of primitive polynomials over Fp of degree k by [192, Thm. 3.5]. Con-
sequently, there exists a primitive polynomial lover Fp with deg(f) = k that is
not counted by Ak(d), and for this f we have p(G
l
) ~ d. 0
To obtain large values of r(s)(j, m), we first choose m to be maximal, i.e., we
take m = k. Then r(s)(f, k) = p(C
1
) by Definition 9.3, and we arein the case
where s > k / m if 8 > 2. In the remainder of this section, we assume that m = k.
The following special case of a result of Niederreiter [246J provides information
on the average order of magnitude of the star discrepancy n;'(B) of the point set
in (9.3).
THEOREM 9.B. If s ;::: 2, m = k, and gcd(k, T) = 1, then, for digital k-step
PRN, we have, on the average,
with an implied constant depending only on p and s, where the average is taken
over all primitive polynomials lover Fp of degree k.
Proof In the proof of Theorem 9.5, we have shown that the point set P
consisting of 0, X{) , Xl! .. ,XT-l can be obtained by the construction leading
to (4.25), with the system C = {eji) E F; : 1 ::; i ~ s, 1 ~ j ~ k} given in that
SHIFT-REGISTER PSEUDORANDOM NUMBERS 197
proof. Thus it follows from Lemma 4.32 that
D;"(P) ~ ; + Rp(C)
with N = pic. By the argument in the proof of Theorem 7.3, we obtain T D ~ ( ) ~
N Div(P) + 1, and so
(9.6) D;(B) ~ 8; 1 + (1 + ~ ) Rp(C).
Note that C depends on I, so we write C = CJ. We have
Rp(GJ) = L Wp(H),
HEBf
where Bf is the set of all nonzero H = (h
ij
) E C(P)BXIc with
(9.7)
We now consider the average value R of Rp(Cf) taken over the set Q of all
primitive polynomials! over Fp of degree k. Since card(Q) = l/>(T)/k, we obtain
For a fixed H =/: 0, the condition HE Bf is described by (9.7), and, by the proof
of Theorem 9.5 and Remark 9A, condition (9.7) is equivalent to
B Ie
LLhijx(i-l)k+j-l - 0 mod/ex).
i=l j=l
On the left-hand side, we have a nonzero polynomial of degree ~ ks - 1, and so
there are at most s - 1 polynomials f E Q satisfYing this condition. Therefore
R ~ (Sl/>(;i
k
L Wp(H) = O(4)(T)-l(logT)S+l)
H#D
by Lemma 3.13, and the result of the theorem follows from 1>(T)-l =
O(T-lIog log T) and (9.6). 0
Since we have assumed that m = k, the maximum possible value of r(s)(f, k)
is k. If res) (j, k) = k, then the point set in Theorem 9.5 forms a (0, k, B)-net
198 CHAPTER 9
in base p. However, by Corollary 4.21, a (0, k, s)-net in base p can only exist if
s :s; p + 1. Therefore, for s p + 2, we must have '1(8)(1, k) < k.
We are led to the following computational problem: Given s 2, k 2, and
the prime p, find a primitive polynomial lover Fp of degree k for which '1(8)(1, k)
is large. For practical implementations, the case where p = 2 is of the greatest
interest. A straightforward argument shows that we have '1(8)(1, k) r(V) (I, k)
whenever s :s; v. Therefore, to obtain large values of r(8)(f, k) for some s 2,
we first must make sure that we achieve large values of r(2)(f, k). The treatment
of the latter problem is facilitated by the formula given in Theorem 9.9, below.
Using the notation introduced in (4.63), we put L(f) = K(f(x)Jx
k
).
THEOREM 9.9. For any chamderistic polynomial f 01 degree k, we have
'1(2) (f, k) = k + 1 - L(f).
Proof. We have 7,(2)(1, k) """ p(G
l
), where, by SA, ,,/s can take G
1
=
{x(i-l)k+j-l + (f) : 1 :::; i :::; 2, 1 :::; j :s; k}. Note that, for 0 :::; db d
2
:::; k with
d
1
+ d
2
> 0, the subsystem {X(i-1)k+
i
-1 + (f) : 1 :s; j :s; d
i
, 1 :s; i :s; 2} of G
1
is
linearly dependent over Fp if and only if there exist h, hI E Fp[x] not both zero
with h1(X) + h(x)x
k
== 0 modf(x) and deg(hl) :s; d
1
-1, deg(h) :::; d
2
-1. Then
we must have h # and h1(x) + h(x)x
k
= h2 (x)/(x) with deg(h2 ) = deg(h),
and the congruence is equivalent to h1(x) - h
2
{x)f(x) = 0 modxk. If we put
g = (1,j) E Fp[x]2, then the above argument shows that peed = p(g,xk),
where the latter quantity is defined in Definition 4.39. The desired formula for
r(2)(f, k) now follows from Theorem 4.46. 0
It is a consequence of Theorem 9.9 that r(2)(I,k) is large if and only if L(f)
is small. The problem of finding primitive polynomials f over Fp for which L(f)
is small can be solved on the basis of results in Niederreiter [243]. The nec-
essary computations were carried out for the case where p = 2 by Mullen and
Niederreiter [215]. Polynomials f over F2 with L(f) = 1 can be characterized
completely (see [243]), and they are rarely irreducible, let alone primitive. How-
ever, if we consider the next larger value 2, then primitive polynomials lover F2
of degree k with L(J) = 2 can be found for each k in the range 3 :s; k 64. A
table of such polynomials is given in [215]. It is conjectured that such polynomi-
als exist for all k 3. For s 3, the search for primitive polynomials f over F2
of degree k with large r(B)(J,k) proceeds by finding many f with L(J) :s; 2 and
then maximizing the value of r(8)(j, k) among these. This program was carried
out by Andre, Mullen, and Niederreiter [8] for dimensions s = 3,4,5 and degrees
k :s; 32.
9.2. The generalized feedback shift-register (GFSR) method.
The genemlized feedback shift-register (GFSR) method due to Lewis and
Payne [190] is another technique for transforming the sequence Yo, Yb." in 9.1
into a sequence of uniform PRN. Again, let p be a small prime, let k 2 be an
integer, and define the sequence Yo, Yb'" E Zp by the recursion (9.1) with prim-
itive characteristic polynomial f over Fp and with initial values Yo, . .. ,Yk-l that
are not all zero. Th\:ln we choose an integer m 2 and integers h
1
, ... ,hm 0,
SHIFT-REGISTER PSEUDORANDOM NUMBERS 199
and we put
m
(9.8) " -j
Xn = L...J Yn+hjP
for n = 0,1, ....
j=l
The numbers Xn belong to the interval I and are called GFSR p8eu.domn-
dom numbers. From per(Yn) = pk - 1, it immediately follows that per(xn) =
pk -1 = T. A convenient implementation of GFSR PRN is based on the vectors
Yn = (Yn+hll'" j Yn+h,.J, n = 0,1, .... It follows from (9.1) that the Yn can be
generated by the vector recursion
Ie-I
Yn+k = L a'Yn+l modp for n = 0, 1, ... ,
1=0
and, if Y
n
is interpreted as a block of digits in base p, then we obtain Xn due
to (9.8).
In the treatment of digital multistep PRN in 9.1, we have seen that the
cases where 8 ::::; kim and 8 > kIm must be distinguIshed. The
distinction for GFSR PRN is given in the following definition. Let a again denote
a root of f in the field Fq! where q = pk, and let 8 ~ 1 be a given dimension.
DEFINITION 9.10. For GFSR PRN, we have the low-dimensional case if the
ms elements a
i
-
Hhj
, 1 S; i S; 8, 1 S; j ::::; m, of Fq are linearly independent over
Fp and the high-dimensional case otherwise.
As in Remark 9.4, it is verified that this distinction does not depend on the
choice of the root a. Note that, since Fq is a vector space over Fp of dimension
k, the low-dimensional case can only arise if 8 ::::; kIm. We now consider the
points
(9.9) Xn = (Xn, Xn+1. ... , xn+s-d E IS for n = 0, 1, ... ,T - 1.
In the low-dimensional case, we have the following analogue of Theorem 9.2.
THEOREM 9.11. Let t E IS be such that all coordinates of t are rationals
with denominator pm. If Z (t) is the nu.mber of integers n with 0 $ n ::::; T - 1
and Xn = t, then, in the low-dimensional case of the GFSR method, we have
Z(t) _ {pk-ms if t i= 0,
- pk-ms - 1 f t 0
1, =.
Proof. If Y
n
is as above, then we have Xn t if and only if
(Y
n
, Y
n
+1,' .. ,Yn+s-d = c for a suitable C E Z;:s depending only on t, with
c = 0 if and only if t = O. If C = (Cij), 1 ::::; i ::::; s, 1 ::::; j ::::; m, is viewed as
an element of F;"s and if we use the formula for Yn in terms of the trace stated
in the proof of Theorem 9.5, then we see that Z(t) is equal to the number of n
with 0 :S n ::::; T - 1 and
Tr(Oanai-l+
hi
) = Cij for 1 :S i ::::; s, 1:S j ::::; m.
200 CHAPTER 9
Since the Ban, 0 n T -1, run exactly through F; , it follows that Z (t) equals
the number of (3 E F; with
(9.10) 'I'r({3a
i
-1+h
j
) = Cij for 1 i s, 1 j m.
Since the system {ai-l+
hj
: 1 i s, 1 j m} is linearly independent
over Fp, it can form the first ms elements of an ordered basis of Fq over Fp.
Let {17b'" ,7]k} be the dual basis of this ordered basis (see Appendix A), and
write (3 = b'17l with all bl E Fp. Then (9.10) states that the coefficients bt,
1 l rna, are uniquely determined by the Cij, with all these b, = 0 if and only
if c = 0, while the remaining bl can be arbitrary. This leads to the formula for
Z (t) in the theorem. 0
H D;(8) denotes the star discrepancy of the point set in (9.9), then Theo-
rem 9.11 shows that, in the low-dimensional case, we have the formula in (9.4).
In the high-dimensional case, we proceed in analogy with the case where s > kIm
in 9.1. Consider the system
02 = {a
i
-1+h
j
: 1 i s, 1 j S m}
of elements of Fq, view Fq as a k-dimensional. vector space over Fp, and let p(02)
be given by Definition 4.27.
DEFINITION 9.12. In the high-dimensional case of the GFSR method, the
figure of merit r(8)(j, H) depending on 8, the characteristic polynomialJ, and
the m-tuple H = (hi, ... ,hm ) of parameters hI,'" ,hm is defined by
r(B)(J,H) = min(m,p(C
2
)).
REMARK 9.13. As in Remark 9.4, we see that r(B)(f,H) is independent of
the choice of a, since p(C
2
) may also be defined by considering the system
{x
i
-1+hj + (J): 1 SiS s,1 S j m} of residue classes in Fp[xJl(f).
THEOREM 9.14. In the high-dimensional case of the GFSR method, the pk
points O,X{),x}, ... ,X'T'-l form a (t, h, s}-net in base p with t = k - r(B)(j,H).
Proof. We proceed as in the proof of Theorem 9.5, and we again use the
notation in the general construction of (t, k, 8 )-nets in base p leading to (4.25).
From and (9.9), we obtain that, for 1 n STand 1 SiS s, we have
( i)
Ynj = Yn+i-!+h
j
for 1 S j m,
= 0 for m < j k.
Continuing as in the proof of Theorem 9.5 and using the simpler expression
"In = (Jan, we arrive, for 1 i S and 0 r k - 1, at the formulas
= 'I'r(,Brai-1+hj) for 1 S j :::; m,
= 0 for m < J' k.
JT
The proof is then completed as in the case of Theorem 9.5. 0
SHIFT-REGISTER PSEUDORANDOM NUMBERS 201
COROLLARY 9.15. For GFSR PRN, we have, in the high-dimensional case,
(
1 . 1)' -r < D*(8) O( 8-1 -r)
2"- 2p P - T = l' P
with l' = r(8)(j,H), where the implied co'natant depends only on p and s.
Proof. For the proof, proceed as in the proof of Corollary 9.6. 0
Theorem 9.14 and Corollary 9.15 are due to Niederreiter [244J. The upper
bound in Corollary 9.15 can be made explicit by the results in 4.1 (compare
also with [244, Thm. 9.4]). Theorem 9.14 and Corollary 9.15 imply that a larger
figure of merit guarantees better distribution properties of the points X n . The
following result of Niederreiter [253] shows that, under reasonable conditions, a
large value of r(8)(j, H) can be obtained by a suitable choice of H.
THEOREM 9.16. Letp be a prime, let k a 2 and ma > k, and let f
be a primitivepolynomal o;per Fp: of degree k. Then there exists a choice of
H = (hI,'" ,h
m
) ,such that .
1'(8)(1, H) min(m, Lk - (8 - 1) logp(m + 1)J),
where logp denotes the logarithm to the base p.
Proof It suffices to show that there exist nonzero 11, . .. ,1m E Fq such that
the system C = {ai-1,'/j : 1 8, 1 :::; j m} of elements of the vector space
Fq over Fp satisfies
(9.11) p(C) Lk - (8 - 1) logp(m + 1)J =: d.
The reason is that a, being a root of f, generates the cyclic group F; j hence,
if 11, .. ' ,1m E F; as above have been found, then, for suitable integers
hI, ... ,hm 0, we have 1j = a
hj
for 1 :::; j :::; m, and so the result of the theorem
follows from Definition 9.12 with H = (hI,' .. ,h
m
). H d :::; 0, then (9.11) holds
trivially. For d 1, let G(d) be the number of m-tuples C'Yb. " ,1m} of elements
of F; such that the system {a
i
-
1
1j : 1 j d
i
, 1 i :::; s} is linearly dependent
over Fp for some 8-tuple (db'" ,d
8
) of integers with 0 :::; di m for 1 S i s
and 2::=1 d
i
= d. Fix such an a-tuple (db ... ,dB) and let M(db ... j dB) be
the number of m-tuples C'Yb ... ,1m) of elements of F; such that the system
{a
i
-
1
1j : 1 :::; j :::; d
i
, 1:::; i s} is linearly dependent over Fp. Let
8 m
i=l j=l
be a linear dependence relation with fixed Cij E Fp not all zero and Cij = 0 for
j > c4. With rJj = it can be written as
(9.12)
m
L = O.
j=l
202 CHAPTER 9
Since k 8, the system {a
i
-
1
1 i 8} is linearly independent over F
p
,
and so TJj =1= 0 for at least one j. It follows that the number of (,1, ... "m)
satisfying (9.12) is at most (q - l)m-l. Since the number of choices for the Cij
is pd - 1, we obtain that
M(dI, ... ,d
8
) (pd - 1)(q - l)m-1.
The number of (d
1
! ,d
s
) with 0 S d
i
m for 1 i 8 and 2::=1 d
i
= d is at
most (m + 1)8-1, and so
G(d) (m + 1),,-1(Pd - 1)(q _ l)m-l.
From the definition of din (9.11), it follows that
G(d) S (m + 1)8-1 (em + ql}'l-1 - 1) (q - l)m-l < (q - l)m.
Thus there exists an m-tuple Crl,'" !')''m) of elements of F; that is not counted
by G(d). For this m-tuple, we have peG) d, and so (9.11) holds. 0
As in 9.1, we now consider again the special case where m = k. The following
is a special case of a result of Niederreiter [244], which provides information on
the average order of magnitude of the star 'discrepancy D;(8) of the point set
in (9.9).
THEOREM 9.17. If f is a primitive polynomial over Fp of degree k s 2
and if m = k, then, for GFSR PRi'J, we have, on the average,
with an implied constant depending only on p and B, where the average is taken
over all H = (hI, ... ,hie) with 0 h
j
S T - 1 for 1 j k.
Proof. By the argument leading to (9.6), we obtain
(9.13) D;(8) 8; 1+ )Rp(G).
Here C depends on H, so we write G = CH' We have
Rp(CH )= L Wp(E) ,
EEBH
where BH is the set of all nonzero E = (eij) E C(p)Bxk with
(9.14)
8 k
L L eijOti-Hhj = O.
i=1 j=1
We now consider the average value taken over the set Q of all
= ,hk) with hj T":" 1 for j k.'Then
R = T-
k
L Rp(CH) = T-
k
L L Wp(E),
\ HEQ HeQ EEBH
SHIFT-REGISTER PSEUDORANDOM NUMBERS 203
and, by interchanging the order of summation, we obtain
(9.15)
R=T-
k
LWp(E) L l.
E=;O HEQ
EEBK
For fixed E =j:. 0, the inner sum in (9.15) is equal to the number of H -
(hb ... ,hk) E Q that satisfy (9.14). We write (9.14) in the form
k
L fJjQ!h
j
= 0
j=l
8
with 1}j = L eija
i
-
1
i=l
for 1 ~ j ~ k.
Since k ~ s, the system {a'i-l : 1 ~ i ~ s} is linearly independent over F
p
, and
so fJj =j:. 0 for at least one j. From this, it follows that the number of H E Q
satisfying (9.14) is at most Tk-l. Thus, from (9.15),
1
R ~ ; L Wp(E) = O(T-l(logT)B),
Ei:O
where we applied Lemma 3.13 in the second step. The result of the theorem now
follows from (9.13). D
For m = k, the maximum possible value ofr(B)(f,H) is k. Ifr(8)(j,H) = k,
then the point set in Theorem 9.14 forms a (0, k, a)-net in base p. However, by
Corollary 4.21, a (0, k, s)-net in base p can only exist if s ~ p + 1. Thus, for
s ~ p + 2, we must have rCs) (j, H) < k; in the special case where p = 2, this was
proved in Tezuka [346] by a different method.
The results in this chapter demonstrate that shift-register PRN tend to have
strong uniformity properties and small discrepancy, and, at least in the latter
respect, they are thus similar to linear congruential PRN (compare with 7.3).
There are, however, the following advantages of shift-register PRN over linear
congruential PRN.
(i) Shift-register PRN can be generated by fast algorithms, since, in the
standard case where p = 2, we use only binary arithmetic rather than modular
arithmetic with a very large modulus.
(ii) The least period length of sequences of shift-register PRN is not bounded
in terms of the word length of the machine. Note that, in the standard case
where p = 2, the least period length is 2k - 1 and that the value of k can be
chosen far beyond the word length. The value of k is, in principle, limited only
by memory restrictions.
(iii) Shift-register PRN are preferable with regard to lattice structure. It is a
consequence of Theorems 9.5 and 9.14 that the digital multistep method and the
GFSR method with well-chosen parameters yield PRN with very little lattice
structure. We have, of course, the trivial lattice structure, which stems from the
fact -that all generated PRN are rationals with fixed denominator pm.
204 CHAPTER 9
Notes.
Proofs of elementary facts about digital multistep PRN, such as least period
length and low-dimensional distribution properties, and of results on the unifor-
mity test, even for parts of the period, can also be found in Lidl and Nieder-
reiter [193, Chap. 7]. A matrix-theoretic approach to elementary properties of
shift-register PRN was presented by Marsaglia and Tsay [206]. Further results
on the serial test, even for parts of the period and for arbitrary irreducible
characteristic polynomials over F
p
, are proved in Niederreiter [246] for digital
multistep PRN and in Niederreiter [242J for GFSR PRN. Upper bounds for the
(star) discrepancy in terms of the figure of merit that are weaker than those in
Corollaries 9.6 and 9.15 where shown by Niederreiter [241], [242], and, in the spe-
cial case where p = 2, for the GFSR method by Tezuka [345]. Theorem 9.9 was
mst proved in Niederreiter [241]. A notion of "k-distribution" for shift-register
PRN was studied, e.g., by Fushimi and Tezuka [112], but this property is weaker
than the net property established in Theorems 9.5 and 9.14, since it amounts
to considering the distribution of the points Xn in elementary cubes rather than
in arbitrary elementary intervals, as in the case of nets. Historical remarks and
further references on shift-register PRN can be found in Niederreiter [257].
CHAPTER 10
Pseudorandom Vector Generation
The task in random vector generation is to produce a sequence of vectors (or
points) that simulates a sequence of independent and identically distributed ran-
dom vector variables with a given multivariate distribution. Random vectors are
becoming more bec8,use the toward parallelization in sci-
entific computing. Typical applications of random vectors arise in parallelized
probabilistic algorithmsj in parallel Monte Carlo and simulation methods, and in
multivariate statistics. If the generation proceeds by a deterministic algorithm,
then we speak of pseudorandom vectors (abbreviated PRV). We discuss here only
the case of uniform pseudorandom vectors where the target distribution is the
uniform distribution on jk, k 2: 2. One possibility of generating k-dimensional
Uniform PRY is to derive them from uniform pseudorandom numbers xo, Xl,'"
by formulas such as
However, it seems to be preferable to generate uniform PRY directly, and such
methods form the subject of this chapter.
Section 10.1 is devoted to the matrix method, which is a multidimensional
analogue of the multiplicative congruential method for generating uniform pseu-
dorandom numbers (see 7.3). The matrix method inherits some of the draw-
backs of the linear congruential method, such as the coarse lattice structure. In
10.2 we consider nonlinear methods for uniform pseudorandom vector genera-
tion. This subject is still in its infancy, and so our report on it will be rather
brief.
10.1. The matrix method.
For given k ~ 2, the generation" of k-dimensional uniform PRY by the matrix
method proceeds as follows. '1-1e choose a large positive integer M, caned the
modulus, and a k x k matrix A with entries from ZM' Then we generate a
sequence Zo, Zl, ... of row vectors in Zj& by starting from an initial vector Zo =f 0
and using the recursion
(10.1) Zn+l == znA mod 111/ for n = 0,1, ....
205
206 CHAPTER 10
The sequence Zo, Z1, . .. is referred to as a matrix generator modulo M. From
this sequence, we derive the uniform PRY
(10.2)
1 k
Un = M Zn E I for n = 0, 1, ... ,
which are called uniform pseudorandom vectors generated by the matrix method.
We can guarantee that the sequences zo, Zt, ... and 110, Ut, . are (purely)
periodic if we assume that the matrix A is nonsingular modulo M, i.e., that
gcd(det(A) , M) = 1.
We first consider the case where M = p is a prime modulus. Then we can
view (10.1) as a vector recursion over the finite field Fp and A as a matrix over
Fp. We assume that A is nonsingular, so that A belongs to the general linear
group GL(k, Fp). Note that then per(u
n
} = per(zn) :::; pk - 1 since there are
pk _ 1 nonzero vectors F;. Theorem 10,2, below) characterizes the cases in
which the upper bound is attained.
LEMMA 10.L If the characteristic polynomial of A is irreducible over F
p
,
then per(u
n
) is equal to the order 01 A in the group CL(k, Fp) for every Zo =f:. 0
. Fk
m p'
Proof From (10.1) we obtain in F;,
(10.3) Zn = zoAn for n = 0, 1, ....
Thus, if m is the order of A in GL(k, Fp), then Zn+m = Zn for all n 0, and so
per(zn) divides m. The characteristic polynomial I(x) = xk - L:7:01 alx
l
is also
the minimal polynomial of A under our hypothesis; hence
k-1
Ak . La,A'.
1=0
Premultiplying by Zn and using (10.3), we see that
k-l
Zn+k = L aZZn+1 for n = 0, 1, ....
1=0
Since 210 =I- 0, there exists a j with 1 :::; j :::; k such that the jth component of 210
is =f:. 0. If Yn is the jth component of Zn, then
lc-l
Yn+k = L a'Yn+l for n = 0, 1, ....
[=0
Since Yo =f:. 0, we get = ord(f) by [192, Thm. 8.28], where ord(f) is the
least positive integer h such that I(x) divides xh - 1, and, from the fact that
I is the mlmmal polynomial of A, we deduce ord(f) = m. Now per(zn) is also
a period length of the sequence Yo, Yl, ... ; thus per(Yn) = m implies that m
divides per(zn). per(u
n
) = per(zn) = m. 0
PSEUDORANDOM VECTOR GENERATION 207
THEOREM 10.2. }lor A E GL(k, Fp) the following properties are equivalent:
(i) per(Un) = pk - 1 for every 2:0 -# 0 in Z;i
(ii) A has order ph - 1 in GL(k, Fp);
(iii) The characteristic polynomial of A is primitive over Fp.
Proof (i) => (ii). If m is the order of A in GL(k, Fp), then the argument
following (1O.3), which holds for any A E GL(k, Fp), shows that per(u
n
) divides
m, and so m ~ pk - 1. We now prove m ::; ph - 1 for any A E GL(k,F
p
).
If 9 E Fp[x] is the minimal polynomial of A, then m is equal to the order of
x + (9) in the group U of units of the residue class ring Fp[x]/(g). Therefore
m ::; card(U) ::; pdeg(g) - 'I ::; pk - 1.
(ii) => (iii). If m = pk - 1, then, from the last chain of inequalities, it follows
that deg(g) = k and m = card(U) = pk - 1, so that Fp[x]j(g) is a field and
x + (g) generates U. Hence g, which is then also the characteristic polynomial
of A, is primitive over Fp since it has x + (g) as a root in Fp[x]j(g).
(iii) => (i). If the characteristic polynomial f of A is primitive over F
pJ
then
it is irreducible over Fp and equal to g. The order of x + (g) in U is then the
same as the order of a root of f in F ~ " J which is pk - 1. Thlli3 in = pk_ 1 by
first part of the proof, and the rest follows from Lemma 10.1. 0
We henceforth assume that the characteristic polynomial of A is primitive
over F
p
, so that per(Un) = per(zn) = pk - 1 = T by Theorem 10.2. Then
zo, 211, ,ZT -1 run exactly through all nonzero vectors in F;. For the PRY
Un, this means that 110, lib," , liT-l run exactly through all points p-1a with
a nonzero a E Z;, and so
D;;'(no, . .. , UT-I) = 1 - (1 _ p-l)k.
The number M(k,p) of k x k matrices over Fp with a primitive characteristic
polynomial over Fp is given by
(10.4)
(
Ie) k-l
M(k,p) = if> p k - 1 II (pk - pi),
j=l
a result derived by Niederreiter [254] from a formula of Reiner [292J. If f is a
fixed primitive polynomial over Fp of degree k, then it follows from the theory
of rational forms of matrices that a k x k matrix A over Fp has the character-
isticpolynomial f if and only if A = SBS-l, where S E GL(k, Fp) and B is
the companion matrix of f. Furthermore, if an A E GL(k, Fp) with a prim-
itive characteristic polynomial over Fp has been found, then all such matrices
(with repetitions) are obtained by forming SAhS-l with S E GL(k, Fp) and an
integer h satisfying 1 ::; h < pk - 1 and gcd(h,pk - 1) = 1. This provides a
way of explicitly constructing matrices with primitive characteristic polynomial
over Fp. A probabilistic algorithm for producing such matrices proceeds by ran-
domly selecting a k x k matrix A over Fp and then testing whether A belongs
to GL(k, Fp) and has order pk - 1 (compare with Theorem 10.2). This is done
by checking whether Apk_l = E and A(pk-1)/r -# E for every prime divisor r of
pk --1, where E E GL(k, Fp) is the identity matrix.
208 CHAPTER 10
The analysis of the structural and statistical properties of uniform PRY pro-
ceeds in analogy with that for uniform PRN. If the Un are given by (10.2) with
M = p, then, for a given integer s :2: 2, we consider the points
(10.5)
Because of per(vn ) = ph - 1 = T, it suffices to take 0 :::; n :; T - 1. The lattice
structure of these points was determined by Afflerbach and Grothe [4]; compare
with case (i) of Theorem 7.6 for the analogous result for linear congruential PRN.
For the terminology used here, we refer to 5.3.
THEOREM 10.3. If the characteristic polynomial of A is primitive over F
p
,
then for s :2: 2 we have
where L is the ks-dimensionallattice with generator matrix
o
'0
o E
and where E is the k x k identity matrix over IR.
Proof. Note that L contains Zks and is thus an integration lattice. For any
n :2: 0, we obtain, using (10.5), (10.2), and (1O.3),
_ 1 ( ). _ 1 2 s-1
Vn = - Zn,Zn+!,'" ,Zn+s-1 = -(zn,znA,znA "., ,znA )
p p
= zn(p-lE p-
1
A p-1A2 ... p-IA
S
-
1
) modZ
ks
,
andso
{O,Vo,V}, ... ,vT-d ~ Lnlks.
The set on the left-hand side has pk elements. Now det(L) = p-k; hence card(Ln
IkS) = pk by Theorem 5.30, and the desired result follows. 0
The fact that det{L) = p-k is independent of s means that we again have
a coarse lattice structure, as in the case of the linear congruential method. In
analogy with the s-wmensional serial test for uniform pR,N, we now consider the
s-dimensional serial test for uniform PRY. In the present context, this means that
we study the discrepancy D ~ ) of the point set consisting of Yo, VI,." ,VN-l E
Iks. We define
(10.6)
where the sum is over all ks-dimensional column vectors h E Cks(p), which,
when represented as a concatenation of the k-dimensional column vectors
(10.7)
PSEUDORANDOM VECTOR GENERATION
s-1
L A i ~ == 0 modp.
i=O
209
Furthermore, r(b,p) is defined as in the beginning of 3.2. The following is a
slight variant of a result of Niederreiter [254] for the full period.
THEOREM 10.4. If the characteristic polynomial of A is primitive over FpI
then for s ~ 2 we have
Df;) 5 ~ + (1 + ~ ) ( ~ + R(S) (A,P).
Proof. If P = {O, Vo, VI, ... ,VT-l}, then by Theorems 3.10 and 10.3 and
Lemma 5.21, we obtain
DT+l(P) 5 ks + Lr(b,p)-l,
'P h.
where the sum is over all b E Cks(P) n L1. and where L1. is the dual lattice of
the lattice L in Theorem 10.3. If h is a concatenation of ho, hI,'" ,fis-I, then
the form of L shows that hE L1. if and only if (10.7) holds. Therefore
DT+l(P) ~ ks + R(s)(A,p).
p
It remains to observe that, as in the proof of Theorem 7.3, we obtain
T D ~ ) 5 (T + l)DT +l(P) + 1. o
Note that R(s)(A,p) has some similarity with the quantity R
1
(L) defined
in (5.43), which can, in turn, be bounded in terms of the quantity R(L) intro-
duced in Definition 5.24. Pursuing this analogy further, we define a figure of
merit similar to that in Definition 5.31 by putting
p(s)(A,p) = min r(2h),
h
where the minimum is extended over all h E Cks(p) for which (10.7) holds. It is
easily seen that we always have 2 ::; p(s)(A,p) 5 2pk. By arguments similar to
those in the proof of Theorem 5.35, it is shown that
R(S)(A) (21og2p)kS + 3(21og2p)ks-l.
,p < (log2)ks-lp(s)(A,p) ,
compare with [254]. Together with Theorem lOA, this implies that p(s)(A,p)
should be large to make the upper bound for the discrepancy Df;) as small as
possible. In [254] there is also a lower bound of the form
D(s c(k,s)
T - p(s)(A,p)
210 CHAPTER 10
with a constant c(k, s) > 0 depending only on k and s. Furthermore, it is shown
in [254] that, for every prime p and for every k ~ 2 and s ~ 2, there exists a
k x k matrix A over Fp with a primitive characteristic polynomial over Fp such
that
(&) (A ) Cl (k, S )pk
P ,p > (logp)ks-qog log(p + 1)
with a constant cl(k, s) > 0 depending only on k and s.
We now study the average order of magnitude of R(s)(A,p) for fixed p,k,
and s. We use the convention that an eigenvector of A corresponding to the
eigenvalue a is a k-dimensional row vector c =1= 0 with cA = ac.
LEMMA 10.5. Let A be a k x k matrix over Fp whose characteristic polyno-
mial is irreducible over Fp. If (-rI, ... ,'Yk) E F; is an eigenvector of A when A
is considered as a matrix over Fq with q = pk, then {11, , .. ,'Yk} is a basis of Fq.
over Fp.
Proof. Let a E Fq be the eigenvalue corresponding to the given eigenvector.
Then a is a root of the characteristic polynomial of A; hence {I, a, . .. ,a
k
-
1
} is
a basis of Fq over Fp- We have
(10.8)
Furthermore, there exists a j with 1 S j S k such that "'Ij =1= o. If /3 E Fq is
arbitrary, then, for some f E Fp[x] with deg(J) < k, we have /3"'1;1 = f(a). By
forming a suitable linear combination of the identities in (10.8), we obtain
Now we compare the jth components in this identity. Since f(A) is a matrix
over Fp , we obtain on the left-hand side a linear combination of "'11, . .. , "'Ik with
coefficients from Fp, while, on the right-hand side, we obtain f(Othj = /3. Thus
we have shown that "'11, . .. ,"'Ik span Fq over Fp. 0
THEOREM 10.6. Let a prime p and integers k ~ 2 and s ~ 2 be given. Then
the average value R of R(S) (A,p), taken over all k x k matrices A over Fp with
a primitive characteristic polynomial over F
p
, satisfies
s -1 (2 7)kS
R < (pk _ 1) :; logp + 5"
Proof. If Q is the indicated range of matrices A, then card(Q) = M(k,p) in
the notation of (lOA), and so
R = M ( ~ ) L R(s)(A,p).
,p AEQ
Inserting the definition of R(s)(A,p) from (10.6), interchanging the order of sum-
mation, and using the fact that every primitive polynomial over Fp is irreducible
over FPl we obtain
(10.9)
PSEUDORANDOM VECTOR GENERATION
R < 1 L N(k,p,h),
- M(k,p) heC" () r(h,p)
lea p
211
where N(k,p, h) is the number of k x k matrices A over Fp with irreducible
characteristic polynomial over F
p
, which satisfy
8-1
(10.10) = 0 E F;.
i=O
Here we have interpreted the congruence (10.7) as an identity in the vector space
F; and we have represented h E C
ks
(P) as a concatenation of the column vectors
ho, hI, ... ,h8 - 1 E F;. It follows from Lemma 10.5 that every matrix counted by
N(k,p, h), when considered as a matrix over Fq with q = ph:, has an eigenvector
(1, '1'2), .. , '1'k) E F!: and {I, '1'2, ... ,1'd is a basis of Fq over Fl" In addition to
hE Cks(P), we now also fix a row vector e = (1,1'2, ... ,1'k) E such that
{I, 1'2, ... ,'1'd is a basis of Fq over Fp, and we count the k x k matrices A over
Fp with irreducible characteristic polynomial over Fp, which satisfy (10.10) and
have c as an eigenvector. If A is such a matrix, then let Of E Fq be the eigenvalue
corresponding to the eigenvector c. Premultiplying (1O.1O) by c, we derive
(10.11)
8-1
2: Olichi = O.
i=O
Now hi '# 0 for some i with 0 i 8 - 1, and then the sca.lar Chi is '# 0, since
{I, '1'2, ... ,'1'k} is a basis of Fq over Fp. Thus (10.11) is a nonzero polynomial
equation for OlE Fq of degree 8 -1, and so (10.11) has at most 8 -1 solutions
Ot E F
q
. For each solution a the matrix A is uniquely determined from cA = Ole,
since {I, '1'2, ... ,1'k} is a basis of Fq over Fp.
Let G be the Galois group of Fq over Fp. We let G act componentwise on
the vector space F;. Let (J' E G and let c E F: be as above. If the k x k matrix
A over Fp has the eigenvector erc when considered as a matrix over Fq, then let
{3 E Fq be the corresponding eigenvalue. Applying er-
1
to (erc)A = {3(ac), we
get cA = (a-
1
{3)c. Thus, if we assume in addition that A has an irreducible
characteristic polynomial over Fp and satisfies (10.10), then A is already obtained
in the counting procedure above by considering the eigenvector c. Therefore an
vectors in the orbit of c under G yield the same solutions A, and there are at
most s - 1 solutions A from each orbit.
The above argument shows that N (k, p, h) is at most s - 1 times the number
of different orbits that are obtained from the action of G on the vectors c. There
are exactly
k-l
II (pk - pi)
j=l
212 CHAPTER 10
choices for vectors c = (1, 12, . .. "k) with the property that {I, 12, . .. "d
is a basis of Fq over Fp. We claim that each orbit contains exactly Ii;' distinct
elements. Let 0"1 be the Frobenius automorphism (see [192, p. 75]) that generates
the cyclic group G, and suppose that a{ c = O"t c for some 0 ::; 1 < j < k. Then
a{-lc = c; hence, if m is the least positive integer with crrc = c, then m < k and
m divides k. This shows that 1, 12, . .. ,'k all lie in the proper subfield Fpm of
Fq, a contradiction to {I, 12, ... ,'Ic} being a basis of Fq over Fp. Thus the orbit
of c under G consists exactly of the distinct elements a{c, j = 0,1, ... ,k -1. It
follows that the number of different orbits is exactly
and so we have
k-1
s-l
rr
k .
N(k,p, b) S -k- (P - pJ).
j=l
Together with (10.4) and (10.9), this yields
8-1
R ~ cp(pk _ 1) hEtr.(P) r(h,p)-l,
and the desired boUnd on R is obtained from [224, Lemma 2.3]. 0
Theorem 10.6 was proved by Niederreiter [254] and shows, in particular, that
for any given p, k, 5, there are matrices A as in Theorem 10.6 for which
with an implied constant depending only on k and s. For such A, the quantity
R(8)(A,p) is of a smaller order of magnitude than the term kslp in the upper
bound for Df;) in Theorem 10.4. Note that kelp is essentially the discretization
error, since it follows from Theorem 3.14 that
D(S) > 1- (1- !)kS
T - p'
which is ks/ p in first approximation. For these PRV, it thus seems to be reason-
able to consider, instead of D!;), a discrete version of the discrepancy in which
the discretization error does not playa role. This will be discussed further in
1O.2.
We now consider the case where M is a prime power modulus, say M = pO:
with a prime p and an integer a :2: 2. We assume that A is nonsingular modulo
M, Le., that det(A) ~ 0 modp. Then the sequences ZO,Zl,'" and Do, Ul,'"
generated by (10.1) and (10.2) are periodic. The follovnng is an upper bound
on the least period length of these sequences which also holds in the case where
a = 1 (compare whh the remark prior to Lemma 10.1).
PSEUDORANDOM VECTOR GENERATION 213
LEMMA 10.7. If M = po. with a prime p and an integer a ~ 1 and if the
k x k matrix A is nonsingular modulo M, then per(Un) ~ (Pk - l)pO!-l .
Proof. It suffices to show that there exists an integer m with 1 ~ m ~
(Pk -1 )po.-l such that Am == E mod M, where E is the k x k identity matrix, for
then Zm = zoAm = Zo modM, and so per(u
n
) = per(2l
n
) ::; m ~ (ph: _l)po.-l.
We proceed by induction on a. The case where a = 1 was settled in the first
part of the proof of Theorem 10.2. Under the induction hypothesis for some
a ~ 1, we have Am = E+po. B for some m with 1 ~ m ~ (pk _l)po.-l and some
integer matrix B. Since E and B commute, we can apply the binomial theorem
to obtain
Since 1 ~ mp ~ (ph: - l)po., the induction is complete. 0
In the case where a = 1, we have already seen that the upper bound pk - 1
Lemma 10.7 can CI :2: 2, Grothe,
and Lehn [86] have given a construction of matrices such that the bound for
per(Un) in Lemma 10.7 is again attained. We mention only a simple special
case of this construction. Start from a matrix B E GL(k, Fp) with a primitive
characteristic polynomial over Fp and view B as an integer matrix. Then, by
Theorem 10.2,
.
for some k x k integer matrix C. Now define the k x k matrix A with entries
from Z M with M = pOt by
A == B + p(BC - B2) modJVJ.
Then, for any initial vector Zo . 0 modp, we have per(u
n
) = (ph: - l)po.-l.
Further work must be done on the matrix method with composite moduli to
assess structural and statistical properties.
10.2. Nonlinear methods.
We present some recent attempts at extending the nonlinear congruential meth-
ods for pseudorandom number generation in Chapter 8 to the case of p s e u d ~
random vector generation.
The general first-order nontinear method for the generation of k-dimensional
uniformPRV proceeds as follows. Let p be a large prime and let Fq be the finite
field with q = pk elements. We generate a sequence /'0, /'1, ... E Fq by selecting
an initial value /'0 and using the recursion
(10.12) 'Yn+1 = 1/J(')'n) for n = 0,1, ... ,
where the map 1/J : Fq -- Fq is chosen in such a way that the sequence /'0,/'1, ...
is periodic with per(-Yn) = q. Such maps 7/J always exist; compare with the simple
214 CHAPTER 10
argument following (B.1). Now view Fq as a k-dimensional vector space over Fp
and . let {,Bl' ... ,(3k} be a basis of Fq over Fp. If Tr denotes the trace from Fq to
F
p
, then we derive a sequence of uniform PRY by putting
1 . k
lin = -(Tr(,Bnn)," . , Th(,Bk'Yn)) E I for n = 0,1, ....
p
(10.13)
This proposal is due to Niederreiter [263] .
THEOREM 10.B. The sequence 110, Ub'" defined by (10.13) is periodic with
perC un) = pk. Over the full period this sequence runs exactly through all the pk
points of the form p-1a with a E Z;.
Proof. Since the sequence 'Yo, 'Yl, ... is periodic with per ( 'Yn) = q = pk, it is
clear that the sequence 110, lib'" is periodic with period length q. To prove that
q is the least period length, we show that 110, 0b ... ,Uq-l are distinct. Suppose
we had U
m
=Ll
n
lor some m and n with 0 ::; m < n .::; q - 1. Then Tr(j3j'Ym) =
Tr(,Bj'Yn) for 1 ::; j ::; k. From the Fp-linearity of the trace and from the fact that
{,Bl,'" ,,Bk} is a basis of Fq over Fp , it follows then that Tr(Ctbm - 'Yn)) = 0
for all a E Fq By the surjectivity of the trace, this implies 'Ym = 'Yn, which is
impossible. The second part of the theorem is obtained by noting that every Un
is of the form p-la for some a E Z; and that 110, Ul, ... ,Uq-l are distinct, as
we have just proved. 0
A special first-order nonlinear method is the inversive method proposed in
Niederreiter [25B}. This is a natural extension of the inversive congruential
method discussed in 8.2. We use the recursion (10.12) with a simple class
of maps 'I/J. Let p and Fq be as above. For'Y E Fq we define;y E Fq by'Y = 'Y-
1
if
'Y =1= 0 and 'Y = 0 if 'Y = O. Then we choose parameters Ct, j3 E Fq with a =1= 0 and
generate a sequence 'Yo, 'Yll ... E F
q
. by selecting an initial value 'Yo and using the
recursion
(10.14) 'Yn+l = Ct'Yn + ,B for n = 0, 1, ....
The uniform PRY 110,01, ... are then obtained by (1O.13). As in Theorem 8.4,
we have the following sufficient condition for per(u
n
) = pk.
THEOREM 10.9. If a,,B E Fq are such that x
2
-,Bx - a is a primitive poly-
nomial over F
q
, then the sequence no, UI, ... defined by (10.13) and (10.14) is
periodic with per(nn) = pk. Furthermore, over the full period, this sequence runs
exactly through all the pk points of the form p-lawith a. E z;.
Proof. By Theorem 10.8 it suffices to show that the sequence 'YO, 'Yl, . .. gen-
erated by (1O.14) is periodic with perbn) = q. This is demonstrated by the
same arguments as the proof of Theorem 8.4; obviously, the fields Fp and Fp2
in that proof have to be changed into Fq and Fq2, respectively. D
In 1O.1 we briefly described the s-dimensional serial test for sequences of
k-dimensional PRY. We now apply this test to a sequence 110, Ub.'. of PRY
generatoo by the inversive method. For a given 8 ~ 2, we define the points
Vn = (\Un, Un+l, ... ,un+s-d E Iks for n = 0, 1, ....
PSEUDORANDOM VECTOR GENERATION 215
The s-dimensional serial test amounts to considering the discrepancy nt) -
DN(vo, ... ,vN-d. Note that by Theorem 3.14 we have
Dt) 2:: 1 - (1 - ~ ) ks for all N 2:: 1.
This is a large discretization error, which, as we have already seen in 1O.1, can
dominate the other terms in upper bounds for nt). Thus it seems more appro-
priate here to consider the discrete discrepancy Et) of Vo, ... ,VN-l, which is
defined by replacing the family .::r in Definition 2.2 by the family of all subinter-
vals of Iks of the form
fi[ai, bi)
i=l P P
with ai, bi E Z for 1 ~ i ~ ks. The discrete discrepancy satisfies a bound as in
Theorem 3.10, but without the term corresponding to the discretization error.
Q ,8 are as for E5;) with N = q, the
full period, analogues of Theorems 8.7 and 8.8 can be established. The upper
bound for the discrete discrepancy has the form E ~ s ) = O(q-1/2(logp)k8) for
2 ~ s < p, with an implied constant depending only on k and s. There are
(q2 _ 1)/2 parameter pairs (a, (3) satisfying the condition in Theorem 10.9,
and, for a positive proportion of these, we have E ~ s ) 2:: Cq-1/2 for all s 2:: 2,
where C > 0 is an absolute constant. The proofs require deep results from the
theory of character sums over finite fields.
Notes.
Matrix generators have long been studied in the theory of linear modular sys-
tems (see [192, Chap. 9]). In the context of pseudorandom vector generation,
the first detailed analysis was performed by Tahmi [339J, who investigated peri-
odicity properties for arbitrary moduli. Independently, the matrix method was
proposed by Grothe [ll6], [ll7], who showed the equivalence between (i) and
(iii) in Theorem 10.2. Another approach was pursued by Niederreiter [236J and
Niki [279], but their method was later found to be equivalent to the matrix
method (see [254]). The classical work of Franklin [109J can be viewed as an
analysis of a stochastic model for the matrix method.
In the proofs of Lemma 10.1 and Theorem 10.2, we have noted that the
order of A in GL(k, Fp) is equal to ord(g), where 9 is the minimal polynomial
of A. Detailed information on ord(g) is contained in [192, Chap. 3]. Further
work on the lattice structure of PRY generated by the matrix method, including
computational results for specific matrices A, is presented in Afflerbach and
Grothe [4] and Grothe [ll8]. In particular, the latter dissertation contains tables
of matrices A with primitive characteristic polynomial over Fp for some large
primes such as p = 224 - 3 and p = 2
31
- 1. For the matrix method, bounds on
D ~ ) for parts of the period have been given in Niederreiter [236J for s = 1 and
in Nie<:l_erreiter [254J for s 2:: 2.
216 CHAPTER 10
Theorems 10.8 and 10.9 have been shown in Niederreiter [263]. The sequences
1'0, 1'1, . .. E Fq considered in a first-order nonlinear method are uniformly dis-
tributed in Fq; see, e.g., Niederreiter and Shiue [271], [272J for an investiga-
tion of this property in the context of recursions. Higher-order recursions in
place of (10.12) could also be contemplated. The transition from a sequence
1'0, 1'1, . .. E Fq generated by such a recursion to a sequence of uniform PRY
would then again be achieved by (10.13).
Expository accounts of random and pseudorandom vector generation are
given in the books of Dagpunar [50], Deak [53], and Devroye [62]. The sur-
vey article of L'Ecuyer [186] contains a discussion of the matrix method. The
uses of pseudorandom vectors in parallel Monte Carlo calculations are explained,
e.g., in Bhavsar and Isaac [28]. For a recent discussion of the problems of ran-
dom number and vector generation for parallel processors, we refer to Eddy [75J.
The interesting concept of a pseudorandom tree for parallel computations was
introduced by Fredrickson et al. [110]; see also Anderson [7] and Halton [126].
APPENDIX A
in ite Fields an
Sequences
Linear Recurring
We collect some basic facts about finite fields and linear recurring sequences,
which are used in various parts of these lecture notes. A suitable reference for
this material is the book of Lidl and Niederreiter [192].
For any prime power q, all finite fields with q elements are isomorphic, and
SOW2 can speak of i;he finite field F ~ with q elements of q). q a
power of the prime p, then Fq is a simple extension of its prime subfield Fp, and
the latter field can be identified with the residue class field of Z modulo p. The
field Fq with q = pm, m ~ 1, is usually constructed as a factor ring Fp[xl/(f),
where f is an irreducible polynomial. over Fp of degree m. The multiplicative
group F; of nonzero elements of Fq is cyclic.
H k is a positive integer, then every irreducible polynomial over Fq of degree k .
has a root in the extension field Fqle of F
q
Conversely, every defining element of
the extension Fqle over F
q
, Le., every element of Fqk that does not belong to any
proper subfield F of Fql< with F ~ Fq, is a root of some irreducible polynomial
over Fq of degree k. This leads to the formula
Iq(k) = ~ LIt( ~ ) qd
dlk
for the number Iq(k) of monic irreducible polynomials over Fq of degree k, where
It is the Mobius function and the smn is over all positive divisors d of k. The
following class of polynomials is of particular importance.
DEFINITION A.I. A monic polynomial over Fq of degree k ~ 1 is called a
primitive polynomial over Fq if it has a root in Fqk that generates the cyclic
group F;1c'
Since every generator of F;k is a defining element of Fqk over Fq, it follows
that every primitive polynomial over Fq is irreducible over Fq. A useful map
from Fqk to Fq is the trace
k-l
Tr(a) = L a
qj
for a E Fqk.
j=O
The trace is Fq-linear and surjective. The following notion is connected with the
trace: __ For every ordered basis {al, .. ' ,ad of Fqk over F
q
, there is a uniquely
217
218 APPENDIX A
determined dual basis, i.e., an ordered basis {,B!, ... , ,Bk} of Fqk over Fq such
that, for 1 :::; i,j :::; k, we have Tr(ai,Bj) = 0 if i =I j and Tr(ai,Bj) = 1 if i = j.
A sequence Yo, Yb'" of elements of Fq is called a (kth-order) linear recurring
sequence if it satisfies a linear recursion
k-l
Yn+k = L aZYn+l for n = 0,1, ...
l=O
with constant coefficients ao, ... ,ak-l E Fq A kth-order linear recurring se-
quence Yo, Yb'" is uniquely determined by the recursion and by the initial val-
ues Yo, . .. , Yk-l. A linear recurring sequence is always ultimately periodic, and
it is (purely) periodic if ao =I O. The polynomial f(x) = xk - 2:7:0
1
alx
l
E Fq[x]
is called the characteristic polynomial of the recursion and also a characteristic
polynomial of the linear recurring sequence. The constant polynomial 1 is also
viewed as a characteristic polynomial of the zero sequence.
THEOREM A.2. If the kth-order linear recurring sequence Yo, Y!, ... has an
irreducible characteristic polynomial f over Fq and a E Fqk is a root of f, then
there exists a uniquely determined e E Fqk such that
Yn = Tr(ea
n
) for n = 0, 1, ....
If Yo, ... , Yk-l are not all zero, then () =I O.
If Yo, Yl, . .. is a kth-order linear recurring sequence with a primitive charac-
teristic polynomial over Fq and with initial values that are not all zero, then we
speak of a maximal period sequence, and we have per(y",,) = qk - 1.
Let Fq ( (x -1 be the field of formal Laurent series over Fq in the variable X-I
(compare with 4,4). Then, with an arbitrary sequence Yo, Y1, . " of elements
of Fq! we associate its generating function 2::=0 Ynx-n-1 E Fq((x-
1
. If f E
Fq [x J is monic, then the sequence Yo! Yl! . .. is a linear recurring sequence with
characteristic polynomial f if and only if
~ -n-l _ g(x)
~ Y n x - f(x)
n=O
with g E Fq(x] and deg(g) < deg(J). This follows immediately by considering the
coefficients in the Laurent series expansion of f (x) 1::=0 Ynx-n-1. The rational
generating function of a linear recurring sequence has a uniquely determined
reduced form him with a monic mE Fq[x] and with hE Fq[x] , deg(h) < deg(m),
and gcd( h, m) = 1. The polynomial m is called the minimal polynomial of
the linear recurring sequence. It follows from this definition that the minimal
polynomial of a linear recurring sequence divides any characteristic polynomial
of the linear recurring sequence. For a linear recurring sequence that is not the
zero sequence, the degree of its minimal polynomial can thus be interpreted as
the least order of a linear recursion satisfied by the sequence.
APPENDIX B
Co ti ued Fraction
Connections with continued fractions appear in several parts of these lecture
notes. We provide a brief summary of the theory of continued fractions for real
numbers and formal Laurent series. All the results on continued fraction expan-
sions of real numbers that we need can be found in LeVeque [188, Chap. 9]. The
theory of fractions for formal,s analogous (see,
e.g., [192, pp. 235-237] and [250]), except for a result on best approximations,
which we prove in Theorem B.1, below.
Every irrational number z has a uniquely determined infinite continued frac-
tion expansion
where the partial quotients ai, i = 0,1, ... , are integers with ao = lzJ and ai ~ 1
for i ~ 1. For i ~ 0, the ith convergent Ti to z is defined by
The rationals Ti are also obtained by the following algorithm. Define
P-2 = 0, P-l = 1, Pi = aiPi-l + Pi-2
q-2 = 1, q-l = 0, qi = aiqi-l + qi-2
for i ~ 0,
for i ~ 0.
Then, for i ~ 0, we have Ti = Pi/qil and the fractions Pi/qi are in reduced form,
Le., gCd(Pil qi) = 1. Moreover, we have 1 = qo :::; ql < q2 < ... and
I
Pi I 1
z--<--
qi qiqi+l
for i ~ O.
The continued fraction expansion of a rational number is finite, but it is not
unique since we have the identity [ao; aI, ... ,at J = [ao; al, ... ,a/-I, al - 1,1 J
if al > 1. However, the continued fraction expansion of a rational z becomes
unique if we write it in the form
with al = 1 and 1 ~ 1.
219
220 APPENDIX B
The integers aD, . .. ,a, satisfy ai ~ 1 for 1 :::; i :::;l and are again called the partial
quotients. For 0 :::; i :::; 1, the ith convergent to z is again defined as the rational
number [aD; all'" ,ai], which can be represented by the reduced fraction Pi/qi,
where Pi and qi are obtained by the same recursions as in the case of irrational
z. We have 1 = qo :::; ql < q2 < ... < ql and
1 < Iz _ Pi I < _1_
qi(qi + qi+d - qi - qiqi+1
for 0:::; i :::; l - 1.
Furthermore, pdqi is a best approximation to z in the sense that, if h is any
integer satisfying 1 :::; h < qi+1 for some i with :::; i :::; l - 1, then
!hz - h! ~ !qiz - Pi! for all integers b.
Let F be an arbitrary field and let F( (X-I)) be the field of formal Laurent
series over F in the variable . Every L E F (( X -1)) has a unique continued
fraction expansion
L = [Ao; Al , A'l , ... ],
where the partial quotients Ai, i = 0,1, ... j are polynomials over F and
deg(Ai) ~ 1 for i ~ 1. This expansion is finite if L represents a rational function
over F and is infinite otherwise. If the contmued fraction expansion of L is bro-
ken off after the term Ai, i ~ 0, we get a rational function, which is called the
ith convergent to L. The polynomials Pi and Qi over F are defined recursively
by
P-2 = 0, P-1 = 1, Pi = AiPi - 1 + Pi - 2 for i ~ 0,
Q-2 = 1, Q-l = 0, Qi = Ai Qi-1 + Qi-2 for i ~ 0,
where these recursions can be used as long as Ai exists. For such i, we have
[Ao; AI, ... ,Ail = Pi/Qi, gcd(Pi' Qi) = 1, and
i
deg(Qi) = L deg(Ar) if i 2: L
r=l
For rational L, we now interpret deg(Ai) = deg(Qi) = 00 whenever Ai and Qi
do not exist.
We introduce the discrete exponential valuation v on F((x-
1
)), which extends
the degree function on F[x]. For a nonzero L E F((x-
1
)), we put
00
v(L) = -w if L = L tkx-
k
and tw i= 0,
k=w
where all til: E F, and we set v(o) = -00. Then we have
for i 2: o.
The following stroIl'g result on best approximations holds in this setting.
CONTINUED FRA.CTIONS 221
THEOREM B.l. Let L E F((x-
1
)) be given. If hE F[x] with 0:::; deg(h) <
deg(Qi+l) for some i ~ 0, then
Proof. Suppose that for some bE F[xJ we have
Then
v(hL - b) < v( L - ~ : ) + v(h) = - deg(Qi) - deg(Qi+l) + deg(h) < 0,
and so it follows from Lemma 3 in [250J that h = E::
j
CrQr with Cr E F[x)
and deg(Cr ) < deg(Ar+d for j ::; l' ::; m8J1d C
j
=f 0) and that
On the other hand, from the above, we get v(hL - b) < - deg(Qi), and so we
must havej ~ i. Since deg(h) < deg(Qi+l), we cannot havej ~ i+lj thusj = i
and h = CiQi. This implies.
deg(Ci ) - deg(Qi+d = v(hL - b) < - deg(Qi) - deg(Qi+l) + deg(h)
= deg(Ci ) - deg(Qi+l),
which is a contradiction. 0
Bibliography
[1] V. G. ADLAKHA, A Monte Carlo technique with quasirandom points for the stochastic
shortest path problem, ArneI'. J. Math. Management Sci., 7 (1987), pp. 325-358.
[2] L. AFFLERBACH, The sub-lattice structure of linear congruential random number gen-
erators, Manuscripta Math., 55 (1986), pp. 455-465.
[3] L. AFFLERBACH AND H. GROTHE, C{}2cutation of Minkowski-reduced latiic8
Computing, 35 (1985), pp. 269-276.
[4] ---, The lattice structure of pseudo-random vectors generated by matrix generators,
J. Compo App1. Math., 23 (1988), pp. 127-131.
[5] L. AFFLERBACH AND R. WEILBACHER, The exact determination of rectangle discrep-
ancy for linear congruential pseudomndom numbers, Math. Comp., 53 (1989), pp. 343-
354.
[6] T. J. AIRD AND J. R. RICE, Systematic search in high dimensional sets, SIAM J.
Numer. Anal., 14 (1977), pp. 296-312.
[7) S. L. ANDERSON, Random number generators on vector supercomputers and other ad-
vanced architectures, SIAM Rev., 32 (1990), pp. 221-251.
[8] D. A. ANDRE, G. L. MULLEN, AND H. NIEDERREITER, Figures of merit for digital
multistep pseudorandom numbers, Math. Comp., 54 (1990), pp. 737-748.
[9) I. A. ANT()NOV AND V. M. SALEEV, An economic method of computing LP.,.-sequences,
Zh. Vychisl. Mat. i Mat. Fiz., 19 (1979), pp. 243-245. (In Russian.)
[10) I. I. ARTOBOLEVSKII, M. D. GENKIN, V. K. GRINKEVICH, I. M. SOBOL', AND R. B.
STATNIKOV, Optimization in the theory of machines by an LP-search, Dokl. Akad. Nauk
SSSR, 200 (1971), pp. 1287-1290. (In Russian.)
[11] 1. I. ARTOBOLEVSKII, V. 1. SERGEYEV, I. M. SOBOL', AND R. B. STATNIKOV, On the
computer aided formulation of mechanical optimum design problems, DokL Aka.d. Nauk
SSSR, 233 (1977), pp. 567-570. (In Russian.)
[12] H. BABOVSKY, F. GROPENGIESSER, H. NEUNZERT, J. STRUCKMEIER, AND B. WIESEN,
Application of well-distributed sequences to the numerical simulation of the Boltzmann
equation, J. Compo Appl. Math., 31 (1990), pp. 15-22.
[13] N. S. BAKHVALOV, Approximate computation of multiple integrals, Vestnik Moskov.
Univ. Ser. Mat. Melfi. Astr. Fiz. Khim., 4 (1959), pp. 3-18. (In Russian.)
[14] T. BANDYOPADHYAY AND P. K. SARKAR, A quasi random sear'ch technique for func-
tional jitting of detector responses, Pramfula, 24 (1985), pp. 643-655.
[15) ---, Implementation and comparative study of random sequences for nonlinear least
square data jitting, Pramana, 28 (1987), pp. 97-107.
[16) W. BAYRHAMER, Some remarks on quasi-random optimization, in Nondifferentiable
Optimization: Motivations and Applications, Sopron, 1984, Lecture Notes in Economics
and Math. Systems, Vol. 255, Springer, Berlin, 1985, pp. 305-309.
[17] Quasi-zufiillige Suchmethoden der globalen Optimierung, dissertation, Univer-
223
224 BIBLIOGRAPHY
sitat Salzburg, Salzburg, Austria, 1986.
[18] J. BECK, Irregularities of distribution and combinatorics, in Surveys in Combinatorics
1985, Glasgow, 1985, London Math. Soc. Lecture Note Series, Vol. 103, Cambridge
University Press, Cambridge, UK, 1985, pp. 25-46.
[19] --, A two-dimensional van Aardenne-Ehrenfest theorem in irregularities oj distri-
bution, Compositio Math., 72 (1989), pp. 269-339.
[20] J. BECK AND W. W. L. CHEN, Irregularities of Di-ainbution, Cambridge University
Press, Cambridge, UK, 1987.
[21] R. BEJIAN, Sur certaines suites presentant une Jaible discrepance a l'origine, C. R.
Acad. Sci. Paris Ser. A, 286 (1978), pp. 135-138.
[22] --, Minoration de ia discrepance d'une sui.te quelconque sur T, Acta Arith., 41
(1982), pp. 185-202.
[23] R. BEJIAN AND H. FAURE, Discrepance de la suite de van der Corput, C. R. Acad.
Sci. Paris Ser. A, 285 (1977), pp. 313-316.
[24] V. BERGSTROM, Einige Bemerkungen zur Theone der diophantischen Approximatio-
nen, Fysiogr. Salsk. Lund. Forh., 6 (1936), no. 13, pp. 1-19.
[25] VI. A Lattice struct'J/fe and reduced bases of mndom vectors genemted
linear recurrences, in Applications of Number Theory to Numerical Analysis, S. K
Zaremba, ed., Academic Press, New York, 1972, pp. 361-370.
[26] W. A. BEYER, R. B. ROOF, AND D. WILLIAMSON, The lattice structure of multiplica-
tive congruential pseudo-random vectors, Math. Comp., 25 (1971), pp. 345-363.
[27] V. C. BHAVSAR, U. G. GUJAR, J. D. HORTON, AND L. A. LAMBROU, Evaluation of
the discrepancy of the linear congruential pseudo-random number sequence, BIT, 30
(1990), pp. 258-267.
[28] V. C. BHAVSAR AND J. R. ISAAC, Design and analysis of parallel Monte Carlo algo-
rithms, SIAM J. Sci. Statist. Comput., 8 (1987), pp. 873-895.
[29] M. BLUMLINGER AND R. F. TICHY, Bemerhngen zu einigen Anwendungen
gleichtJerteilter i<oigen, Sitzungsber. Oaterr. Akad. Wiss. Math.-Natur. Kl. II, 195
(1986), pp. 253-265.
[30] I. BOROSH AND H. NIEDERREITER, Optimal multipliers for pseudo-random number
generation by the linear congruential method, BIT, 23 (1983), pp. 65-74.
[31] M. BOURDEAU AND A. PITRE, Tables of good lattices in four and five dimensions,
Nwner. Ma.th., ~ 1 ' 7 (1985), pp. 39-43.
[32] E. BRAATEN AND G. 1NELLER, An improved.Zow-discrepancy sequence for multidimen-
sional quasi-Monte Carlo integration, J. Compo Physics, 33 (1979), pp. 249-258.
[33] P. BRATLEY AND B. L. Fox, Algorithm 659: Implementing Sobol's quasirandom se-
quence genemtor, ACM Trans. Math. Software, 14 (1988), pp. 88-100.
[34] P. BRATLEY, B. L. Fox, AND L. E. SCHRAGE, A Guide to Simulation, 2nd ed.,
Springer, New York, 1987.
[35] M. CAR, Sommes de carres dans Fq[X], Dissertationes Ma.th. (Rozprawy Mat.), 215,
1983.
[36] L. CARLITZ, The arithmetic of polynomials in a Galois field, Amer. J. Math., 54
(1932), pp. 39-50.
[37] --, The singular series for sums of squares of polynomials, Duke Math. J., 14
(1947), pp. 1105-1120.
[38] J. W. S. CASSELS, An Introduction to the Geometry of Numbers, Springer, Berlin,
1971.
[39] G. J. CHAITIN, On the length of programs for computing finite binary sequences, J.
Assoc. Comput. Ma.ch.., 13 (1966), pp. 547-569.
[40] --, Information, Randomness & Incompleteness. Papers on Algorithmic Informa-
tion Theory, World Scientific Publishers, Singapore, 1987.
[41 J K. L. CJIUNG, An estimate concerning the Kolmogoroff limit distribution, Trans. Arner.
Math. Soc., 67 (1949), pp. 36-50.
[42] T. COCHRANE, On a trigonometric inequality of VinogrndDv, J. Number Theory, 27
(1987), pp. 9-16. \
BIBLIOGRAPHY 225
[43] B. J. COLLINGS, Compound random number generators, J. Amer. Sta.tist. Assoc., 82
(1987), pp. 525-527.
[44] R. R. COVEYOU AND R. D. MACPHERSON, Fourier analysis oj uniform random number
generators, J. Assoc. Comput. Mach., 14 (1967), pp. 100-119.
[45] R. CRANLEY AND T. N. L. PATTERSON, Randomization of number theoretic methods
for multiple integration, SIAM J. Numer. Anal., 13 (1976), pp. 904-914.
[46) T. W. CUSICK, Continuants with bounded digits, Mathematika., 24 (1977), pp. 166-172.
[47] --, Continuants with bounded digits - II, Mathematika., 25 (1978), pp. 107-109.
[48] --, Continuants with bounded digits - III, Monatsh. Math., 99 (1985), pp. 105-109.
[49] --, Products 0/ simultaneous approximations oj rational numbers, Arch. Math.
(Basel), 53 (1989), pp. 154-158.
[50] J. DAGPUNAR, Principles of Random Variate Generation, Clarendon Press, Oxford,
UK,1988.
[51] P. J. DAVIS AND P. RABINOWITZ, Methods of Numerical Integration, 2nd ed., Academic
Press, New York, 1984.
[52J 1. DEAK, Multidimensional integration and stochastic programming, in Numerical Tech-
niques for Stochastic Optimization, Yu. Ermoliev and R. J.-B. Wets, eds., Springer,
Berlin, 1988, pp. 187-200.
[53] --, Random Number Generators and Simulation, Akademiai Kiad6, Budapest, Hun-
gary, 1989.
[54] L. DE CLERCK, A 'P'f'OOf of Niederreiter's conjedure conceT'Y!.ing eITor bounds for
Monte Carlo integration, Adv. in Appl. Ma.th., 2 (1981), pp. 1-6.
[55] --, De exacte berekening van de sterdiscrepantie van de rijen van Hammersley in
2 dimensies, Ph.D. thesis, Ka.tholieke Universiteit Leuven, Leuven, Belgium, 1984.
[56] --, A method jor exact calculation of the stardiscrepancy oj plane sets applied to
the sequences 0/ Hammersley, Monatsh. Math., 101 (1986), pp. 261-278.
[57] P. DEHEUVELS, Strong bounds for multidimensional spacings, Z. Wahrsch. Verw. Geb.,
64 (1983), pp. 411-424. .
[58] A. DE MATTEIS AND S. PAGNUTTI, Parallelization of random number generators and
long-range correlations, Numer. Math., 53 (1988), pp. 595-608.
[59] J. DENES AND A. D. KEEDWELL, Latin Squa1'eS and Their Applications, Academic
Press, New York, 1974.
[60] V. DENZER AND A. ECKER, Optimal multipliers for linear congruential pseudo-random
number generators with prime moduli, BIT, 28 (1988), pp. 803-808.
[61] L. P. DEVROYE, Progressive global random search 0/ continuous functions, Math.
Programming, 15 (1978), pp. 330-342.
[62] --, .Non-UniJorrn Random Variate Generation, Springer, New York, 1986.
[63] U. DIETER, Pseudo-random numbers: The exact distribution 0/ pairs, Math. Comp.,
25 (1971), pp. 855-883.
[64] --, How to calculate shortest vectors in a lattice, Math. Camp., 29 (1975), pp. 827-
833.
[65] --, Probleme bei der Erzeugung gleichverteilter ZuJallszahlen, in Zufallszahlen und
Simulationen, L. Afflerbach and J. Lehn, eds., Teubner, Stuttgart, Germany, 1986,
pp.7-20.
[66] S. DISNEY, Error bounds for rank 1 lattice quadrature rules modulo composites,
Monatsh. Math., 110 (1990), pp. 89-100.
[67] S. DISNEY AND 1. H. SLOAN, Error bounds Jor the method. of good lattice points, Math.
Comp., 56 (1991), pp. 257-266.
[68] _. -, Lattice integration rules of maximal rank, SIAM J. Numer. Anal., 29 (1992),
pp. 566-577.
[69] M. DRMOTA, Applications oj uniformly distributed functions and sequences in statistic
ergodic measuring techniques, J. Compo Appl. Math., 31 (1990), pp. 57-62.
[70] M. DRMOTA AND R. F. TICHY, DeterminisUsche Approximation stochastisch-
ergodischer Signale, Messen-Steuern-Regeln, 32 (1989), pp. 109-113.
[71] V. DROBOT, On dispersion and Markov constants, Acta Math. Hungar., 47 (1986),
226 BIBLIOGRAPHY
pp.89-93.
[72J Y. DUPAIN, Discrepance de 1a suite ({na}), a = (1 + ,,;5)/2, Ann. Inst. Fourier, 29
(1979), no. 1, pp. 81-106.
[73J Y. DUPAIN AND V. T. SOS, On the discr-epancy of (na) sequences, in Topics in Classical
Number Theory, Budapest, 1981, Colloq. Math. Soc. Janos Bolyai, Vol. 34, North-
Holland, Amsterdam, 1984, pp. 355-387.
[74J R. ECKHARDT, Stan Ulam, John von Neumann, and the Monte Carlo method, in
Stanislaw Ulam 1909-1984, Los Alamos Sci. No. 15 Special Issue, Los Alamos National
Laboratory, Los Alamos, NM, 1987, pp. 131-137.
[75J W. F. EDDY, Random number generators for parallel processors, J. Compo Appl. Math.,
31 (1990), pp. 63-71.
[76J J. EICHENAUER, H. GROTHE, AND J. LEHN, Marsaglia's lattice test and non-linear
congruential pseudo random number generators, Metrika, 35 (1988), pp. 241-250.
[77J J. EICHENAUER, H. GROTHE, J. LEHN, AND A. TOPUZOGLU, A multiple 'recursive non-
linear congruential pseudo random number generator, Manuscripta Math., 59 (1987),
pp. 331-346.
[781 J. EICHENAUER AND J. LEHN, A non-linear congruential pset!do 'random number gen-
erator, Statist. Papers, 27 (1986), pp. 315-326.
[79J ---, On the structure of quadratic congruential sequences, Manuscripta Math., 58
(1987), pp. 129-140.
[80] J. EICHENAUER, J. LEHN, AND A. TOPUZOGLU, A nonlinear congruential pseudoran-
dom number generator with power of two modulus, Math. Comp., 51 (1988), pp. 757-
759.
[81J J. EICHENAUER AND H. NIEDERREITER, On Marsaglia's la-tiice test for pseudorandom
numbers, Manuscripta Math., 62 (1988), pp. 245-248.
[82J J. EICHENAUER-HERRMANN, Inversive congruential pseudorandom numbers avoid the
planes, Ma.th. Comp., 56 (1991), pp. 297-301.
[83J ---, On the discrepancy of inversive congruential pseudorandom numbers with prime
power modulus, Manuscripta Math., 71 (1991), pp. 153-161.
[84J J. EICHENAUER-HERRMANN AND H. GROTHE, A remark on long-range correlations
in multiplicative congruential pseudo random number generators, Numer. Math., 56
(1989), pp. 609-611.
[85J ---, Upper bo'unds for the Beyer ratios of linear congruential generators, J. Compo
Appl. Math., 31 (1990), pp. 73-80.
[86] J. EICHENAUER-HERRMANN, H. GROTHE, AND J. LEHN, On the length of pseu-
dorandom vector sequences generated by matrix generators, Math. Comp., 52 (1989),
pp. 145-148.
[87J J. EWHENAUER-HERRMANN, H. GROTHE, H. NIEDERREITER, AND A. TOPUZOGLU, On
the lattice structure of a nonlinear generator with modulus 2"'} J. Compo AppL Ma.th.,
31 (1990), pp. 81-85.
[88J J. EICHENAUER-HERRMANN AND H. NIEDERREITER, On the discrepancy of quadratic
congruential pseudorandom numbers, J. Compo Appl. Math., 34 (1991), pp. 243-249.
[89J ---, Lower bounds for the discrepancy of 'inversive congrueniial pseudorandom num-
bers with power of two modulus, Ma.th. Comp., 58 (1992), pp. 775-779.
[90J J. EICHENAUER-HERRMANN AND A. TOPUZOGLU, On the period length of congruential
pseudorandom number sequences generated by inversions, J. Compo Appl. Math., 31
(1990), pp. 87-96.
[91J S. M. ERMAKOV, Note on pseudorandom sequences, Zh. Vychisl. Ma.t. i Mat. Fiz., 12
(1972), pp. 1017-1082. (In Russian.)
[92J YU. G. EVTUSHENKO, Numerical Optimization Techniques, Springer, New York, 1985.
[93J H. FAURE, Suites Ii faible discrepance dans T
8
, Publ. Dep. Math., Universite de Limo-
ges, France, 1980.
[94J ---, Discrepances de suites associees a un systeme de numeration (en dimension
un), Bull. Soc. Math. France, 109 (1981), pp. 143-182.
[95J ---, Discrepanqe de suites associees a. un systeme de nl!.meration (en dimension 8),
BIBLIOGRAPHY 227
Acta Arith., 41 (1982), pp. 337-351.
[96] --, Etude des reates pour les suites de van der CO'l'put generolisees, J. Number
Theory, 16 (1983), pp. 376-3fjL1.
[97] --, Lemme de Bohl pour lea suites de van der Corput genemlisees, J. Number
Theory, 22 (1986), pp. 4-20.
[98] --, On the star-discrepancy of generalized Hammersley sequences in two dimen-
sions, Monatsh. Math., 101 (1986), pp. 291-300.
[99] --, Using permutations to reduce discrepancy, J. Compo Appl. Math., 31 (1990),
pp.97-103.
[100] --, Good permutations for extreme discrepancy, J. Number Theory, to appear.
[101] R. P. FEDORENKO, Minimization of nonsmooth functions, Zh. Vychisl. Mat. i Mat.
Fiz., 21 (1981), pp. 572-584. (In Russian.)
[102] L. FEJES T6TH, Lagerungen in der Ebene, auf der Kugel und im Raum, 2. Aufi.,
Springer, Berlin, 1972.
[103J G. S. FISHMAN, Multiplicative congruential random number generators with modulus
213: An exhaustive analysis for fJ = 32 and a partial analysis for fJ = 48, Math. Comp.,
54 (1990), pp. 331-344.
[104J G. S. FISHMAN AND L. R. MOORE HI, An exhaustive analysis of multiplicative congru-
ential random number generators with modulus 2
31
_1, SIAM J. Sci. Statist. Comput.,
7 (1986), pp. 24-45; Erratum, ibid., 7 (1986), p. 1058.
[105] P. FLAJOLET, P. KmSCHENHOFER., AND R. F. TICHY, Deviation.,; frOln uniformity in
random strings, Probab. Theory Related Fields, 80 (1988), pp. 139-150.
[106] --, Discrepancy of sequences in discrete spaces, in Irregularities of Partitions,
G. Halasz and V. T. S6s, eds., Springer, Berlin, 1989, pp. 61-70.
[107] B. L. Fox, Algorithm 647: Implementation and relative efficiency of quasirandom -
sequence generators, ACM 'frans. Math. Software, 12 (1986), pp. 362-376.
[108] J. N. FRANKLIN, Deterministic simulation of random processes, Math. Comp., 17
(1963), pp. 28-59.
[109] --, Equidistribution of matm-power residues modulo one, Math. Comp., 18 (1964),
pp. 5 6 ~ 5 6 8 .
[110] P. FREDRICKSON, R. HIROMOTO, T. L. JORDAN, B. SMITH, AND T. WARNOCK,
Pseudo-random trees in Monte Carlo, Parallel Comput., 1 (1984), pp. 175-180.
[111] K. K. FROLOV, On the connection between quadrature formulas and sublattices of the
lattice of integral vectors, Dok!. Akad. Nauk SSSR, 232 (1977), pp. 40--43. (In Russian.)
[112J M. FUSHIMI AND S. TEZUKA, The Is-distribution of the generalized feedback shift register
pseudorandom numbers, Comm. ACM, 26 (1983), pp. 516-523.
[113J A. GENZ, Testing multidimensional integration routines, in Tools, Methods and
Langua.ges for Scientific and Engineering Computation, B. Ford, J. C. Rault, and
F. Thomasset, eds., North-Holland, Amsterdam, 1984, pp. 81-94.
[114] M. GOLDSTERN, Eine Klasse vollstandig gleichverteilter Folgen, in Zahlentheoretische
Analysis II, E. Hlawka, ed., Lecture Notes in Math., VoL 1262, Springer, Berlin, 1987,
pp.37-45.
[115J P. GRITZMANN, Finite Packungen und Uberoeckungen, Habilitationsscmift, Univer-
sitat-Gesamthochschule Siegen, Siegen, Germany, 1984.
[116J H. GROTHE, Matrixgeneratoren zur Erzeugung gleichverteilter Zu/allsvektoren, in Zu-
faUszahlen und Sirnulationen, L Afflerbach and J. Lehn, eds., Teubner, Stuttgart, Ger-
many, 1986, pp. 29-34.
[117] --, Matrix generators for pseudo-random vector generation, Statist. Papers, 28
(1987), pp. 233-238.
(118] --, Matrixgeneratoren zur Erzeugung gleichverteilter Pseudozujallsvekioren, Dis-
sertation, Teclmische Hochschule Darmstadt, Darmstadt, Germany, 1988.
[119] A. GRUBE, Mehrfach rekursiv-erzeugte Pseudo-ZuJallszahlen, Z. Angew. Math. Mech.,
53 (1973), pp. T223-T225.
[120] B. GYlRES, Doubly ordered linear rank statistics, Acta Math. Acad. Sci. Hungar., 40
(1982), pp. 55-63.
228 BIBLIOGRAPHY
[121] ---, Linear rank statistics generated by uniformly distributed sequences, in Nonpara-
metric Statistical Inference, Budapest, 1980, Colloq. Math. Soc. Janos Bolyai, Vol. 32,
North-Holland, Amsterdam, 1982, pp. 391-400.
[122] S. HABER, On a sequence of points of interest for numerical quadrature, J. Res. Nat.
Bur. Standards Sect. B, 70 (1966), pp. 127-136.
[123] ---, Parameters for integrating periodic junctions of several variables, Math. Comp.,
41 (1983), pp. 115-129.
[124J M. HALL, Combinatorial Theory, 2nd ed., John Wiley, New York, 1986.
[125] J. H. HALTON, On the efficiency of certain quasi-random sequences of points in eval-
uating multi-dimensional integrals, Numer. Math., 2 (1960), pp. 84-90; Berichtigung,
ibid., 2 (1960), p. 196.
[126] ---, Pseudo-random trees: Multiple independent sequence generators for parallel and
branching computations, J. Compo Physics, 84 (1989), pp. 1-56.
[127] J. H. HALTON AND G. B. SMITH, Algorithm 247: Radical-inverse quasi-random point
sequence [G5], Comm. ACM, 7 (1964), pp. 701-702.
[128J J. H. HALTON AND S. K. ZAREMBA, The extreme and L2 discrepancies of some plane
sets, Monatsh. Math., 73 (1969), pp. 316-328.
[129] J. M. HAMMERSLEY, Monte Carlo methods for solving mutiivariable problems, Ann.
New York Acad. Sci., 86 (1960), pp. 844-814.
[130] J. M. HAMMERSLEY AND D. C. HANDSCOMB, Monte Carlo Methods, Methuen, London,
1964.
[131J D. R. HAYES, The expression of a polynomial as a sum of three irreducibles, Acta
Arith., 11 (1966), pp. 461-488.
[132J P. HELLEKALEK, Regularities in the distribution of special sequences, J. Number The-
ory, 18 (1984), pp. 41-55.
[133] D. HENSLEY, The distribution of badly approximable rationals and continuants with
bounded digits. II, J. Number Theory, 34 (1990), pp. 293-334.
[134] E. HLAWKA, Jilunktionen von beschriinkter Variation in der Theorie der
Gleichverteilung, Ann. Mat. Purl:!. Appl., 54 (1961), pp. 325-333.
[135] ---, Zur angeniiherten Berechnung mehrfacher Integrale, Monatsh. Math., 66 (1962),
pp. 140-151.
[136] ---, Discrepancy and uniform distribution of sequences, Compositio Math., 16
(1964), pp. 83-91.
[137J ---, Discrepancy and Riemann integration, in Studies in Pure Mathematics,
L. Mirsky, ed., Academic Press, New York, 1971, pp. 121-129.
[138] --, Gleichverteilung auf Produkten von Sphiiren, J. Reine Angew. Math., 330 (1982),
pp.1-43.
[139] ---, Beitrage zur Theorie der Gleichverteilung und ihren Anwendungen I-V,
Sitzungsber. Osten. Akad. Wiss. Math.-Natur. Kl. II, 197 (1988), pp. 1-94, 95-120,
121-154, 209-259, 261-289.
[140] ---, Niiherungsformeln zur Berechnung von mehrfachen Integralen mit Anwendun-
gen auf die Berechnung von Potentialen, Induktionskoeffizienten und Losungen von
Gleichungssystemen, in Number-Theoretic Analysis, E. Hlawka and R. F. Tichy, eds.,
Lecture Notes in Math., Vol. 1452, Springer, Berlin, 1990, pp. 65-111.
[141] E. HLAWKA, F. FIRNEIS, AND P. ZINTERHOF, Zahlentheoretische Methoden in der
Numerischen Mathematik, Oldenbourg, Vienna, 1981.
[142] J. HORBOWICZ, Criteria for uniform distribution, Indag. Math. (N.S.), 43 (1981),
pp. 301-307.
[143J J. HORBOWICZ AND H. NIEDERREITER, Weighted exponential sums and discrepancy,
Acta Math. Hungar., 54 (1989), pp. 89-97.
[144] L. K. HUA AND Y. WANG, On uniform distribution and numerical analysis (Number-
theoreJic method) I-III, Sci. Sinica, 16 (19'73), pp. 483-505; 17 (1974), pp. 331-348; 18
(1975f, pp. 184-198.
[145] ---, Applications of Number Theory to Numerical Analysis, Springer, Berlin, 1981.
[146J F. JAMES, A revifw of pseudoTandom number generators, Computer Physics Comm.,
BIBLIOGRAPHY 229
60 (1990), pp.
[141] S. JOE, The generation of lattice points for numerical multiple integration, J. Compo
Appl. Math., 26 (1989), pp. 321-331.
[148] --, Randomization of lattice ndes for numerical multiple integration, J. Compo
Appl. Math., 31 (1990). pp. 299-304.
[149] D. KAHANER, SourceJJ of information on quadrature software, in Sources and Devel-
opment of Mathema.tical Software, W. R. Cowell, ed., Prentice-Hall, Englewood Cliffs,
NJ, 1984, pp. 134-164.
[150] M. H. KALOS AND P. A. WHITLOCK, Monte Carlo Methods, Vol. 1: Basics, John
Wiley, New York, 1986.
[151] B. K. KHASNABIS, T. BANDYOPADHYAY, P. K. SARKAR, AND G. MUTHUKRISHNAN,
Spectrum measurement of neutrons from 40 MeV alpha particle bombardment of tanta-
lum, Indian J. Physics, 61A (1981), pp. 93-100.
[152] J. KIEFER, On large deviations of the empiric d.f. of vector chance variables and a Zaw
of the iterated logarithm, Pacific J. Math., 11 {1961}, pp.
[153] P. KIRSCHENHOFER AND R. F. TICHY, Some distribution properties ofO,l-sequences,
Manuscripta. Ma.th., 54 (1985), pp. 205-219.
[154J D. E. KNUTH, The Art of Computer Programming, Vol. 2: Seminumerical. Algorithms,
2nd ed., Addison-Wesley, Reading, MA, 1981.
[155J J. F. KOKSMA, Een algemeene stelling uit de theorie der gelijkrnatige verdeeling modulo
1, Mathematica. B (Zutphen), 11 (1942/1943), pp. 7-11.
[156] A. N. KOLMOGOROV, Three approaches to the definition of the concept "quantity of
information," Problemy Peredachl Informasii, 1 (1965), no. 1, pp. 3-11. (In Russian.)
[151] A. N. KOLMOGOROV AND V. A. USPENSKH, Algorithms and randomness, TeoI'. Veroy-
atnost. i Primenen., 32 (1981), pp. 425-455. (In Russian.)
[158] N. M. KOROBOV, The approximate computation of multiple integrals, Dohl. Akad.
Nauk SSSR, 124 (1959), pp. 1207-1210. (In. Russian.)
[159] --, Properties and calculation of optimal coefficients, Dok!. Akad. Nauk SSSR, 132
(1960), pp. (In Russian.)
[160] --, Number-Theoretic Methods in Approximate Analysis, Fizmatgiz, Moscow, 1963.
(In Russian.)
[161] --, On the computation of optimal coefficients, Dokl. Akad. Nauk SSSR, 267 (1982),
pp. (In Russian.)
[162] E. KRANAKIS, Primaliiy and Cryptography, John Wiley, Chichester, UK, 1986.
[163] L. KUIPERS AND H. NIEDERREITER, Uniform Distribution of Sequences, John Wiley,
New York, 1974.
[164] M. LACZKOVICH, Equidecomposability and discrepancy; A solution of Tarski's circle-
squaring problem, J. Reine Angew. Math., 404 (1990), pp. 77-111.
[165] J. P. LAMBERT, Quasi-Monte Carlo, low-discrepancy sequences, and ergodic transfor-
mations, J. Compo Appl. Math., 12/13 (1985), pp. 419-423.
[166] --, Quasi-random sequences for optimization and numerical integration, in Nu-
merical Integration: Recent Developments, Software and Applications, P. Keast and
G. Fairweather, eds., NATO Adv. Sci. Inst. Ser. C, Vol. 203, Reidel, Dordrecht, the
Netherlands, 1987, pp. 193-203.
[167] --, A sequence well dispersed in the unit square, Proe. Amer. Math. Soc., 103
(1988), pp. 383-388.
[168] --, Quasi-random sequences in numerical practice, in Numerical Mathematics Sin-
gapore 1988, R. P. Agarwal, Y. M. Chow, and S. J. 'Wilson, eds., Internat. Series of
Numer. Math., VoL 86, Birkhauser, Basel, 1988, pp. 273-284.
[169] B. LAPEYRE AND G. PAGES, Families de suites a discrepance faible obtenues par
iteration de transformations de [0,1], C.R. Acad. Sci. Paris Ser. I Math., 308 (1989),
pp. 507-509.
[170] B. LAPEYRE, G. PAGES, AND K. SAB, Sequences with low discrepancy. Generalisation
and application to Robbins-Monro algorithm, Statistics, 21 (1990), pp. 251-272.
[171] G. LARCHER, Optimale Koeffizienten bezuglich zusammengesetzter Zahlen, Monatsh.
230 BIBLIOGRAPHY
Math., 100 (1985), pp. 127-135.
[172] --, Ober die isotrope Diskrepanz von Folgen, Arch. Math. (Basel), 46 (1986),
pp. 240-249.
[173] --, On the distribution of sequences connected with good lattice points, Monatsh.
Math., 101 (1986), pp. 135-150.
[174] --, The dispersion of a special sequence, Arch. Math. (Basel), 47 (1986), pp. 347-
352.
[175] --, A best lower bound for good lattice points, Monatsh. Math., 104 (1987), pp. 45-
5l.
[176] --, On the distribution of s-dimensional Kronecker-sequences, Acta Arith., 51
(1988), pp. 335-347.
[177J --, A new extremal property of the Fibonacci ratio, Fibonacci Quart., 26 (1988),
pp. 247-255.
[178] --, On the distribution of the multiples of an s-tuple of real numbers, J. Number
Theory, 31 (1989), pp. 367-372.
[179] G. LARCHER AND H. NIEDERREITER, Optimal coefficients modulo prime powers in the
three-dimensional case, Ann. Mat. Pura Al'l'L, 155 (1989), Pl'. 299-315.
[180] --, Kronecker-type sequences and nonarchimedean diophantine approximations, in
preparation.
[181] C. LECOT, Low discrepancy sequences for solving the Boltzmann equation, J. Compo
Appl. Math., 25 (1989), Pl'. 237-249.
[182] --, A direct simulation Monte Carlo scheme and uniformly distributed sequences
for solving the Boltzmann equation, Computing, 41 (1989), pp. 41-57.
[183] --, An algorithm for generating low discrepancy sequences on vector computers,
Parallel Comput., 11 (1989), pp. 113-116.
[184] --, A quasi-Monte Carlo method for the Boltzmann equation, Math. Comp., 56
(1991), pp. 621-644.
[185] P. L'ECUYER, Efficient and portable combined random number generators, Comm.
ACM, 31 (1988), pp. 742-749,774.
[186] --, Random numbers for simulation, Comm. ACM, 33 (1990), no. 10, pp. 85-97.
[187] D. H. LEHMER, Mathematical methods in large-scale computing units, in Proc. 2nd
Sympos. on Large-Scale Digital Calculating Machinery, Cambridge, MA, 1949, Harvard
University Press, Cambridge, MA, 1951, pp. 141-146.
[188] W. J. LEVEQUE, Fundamentals of Number Theory, Addison-Wesley, Reading, MA,
1977.
[189) M. B. LEVIN, On the choice of parameters in generators of pseudorandom numbers,
Dok!. Akad. Nauk SSSR, 307 (1989), pp. 529-534. (In Russian.)
[190] T. G. LEWIS AND W. H. PAYNE, Generalized feedback shift register pseudorandom
number algorithm, J. Assoc. Comput. Mach., 20 (1973), pp. 456-468.
[191] P. LIARDET, Regularities of distribution, Compositio Math., 61 (1987), pp. 267-293.
[192] R. LIDL AND H. NIEDERREITER, Finite Fields, Addison-Wesley, Reading, MA, 1983.
[193] --, Introduction to Finite Fields and Their Applications, Cambridge University
Press, Cambridge, UK, 1986.
[194] J. LINHART, Approximation of a ball by zonotopes using uniform distribution on the
sphere, Arch. Math. (Basel), 53 (1989), pp. 82-86.
[195] A. LUBOTZKY, R. PHILLIPS, AND P. SARNAK, Hecke operators and distributing points
on the sphere. I, Comm. Pure Appl. Math., Suppl., 39 (1986), 8149-S186.
[196] --, Hecke operators and distributing points on 8
2
. II, Comm. Pure Appl. Math., 40
(1987), pp. 401-420.
[197] J. N. LYNESS, Some comments on quadrature rule construction criteria, in Numerical
Integration III, H. Brafl and G. Hfunmerlin, OOs., Internat. 8eries of Numer. Math., Vol.
85, ll!rkhauser, Basel, 1988, pp. 117-129.
[198] --, An introduction to lattice rules and their generator matrices, IMA J. Numer.
Anal., 9 (1989), pp. 405-419.
[199] J. N. LYNESS AN\1 I. H. SLOAN, Some properties of rank-2 lattice rules, Math. Comp.,
BIBLIOGRAPHY 231
53 (1989), pp. 627-637.
[200] J. N. LYNESS AND T. S0REVIK, The number of lattice rules, BIT, 29 (1989), pp. 527-
534.
[201] J. N. LYNESS, T. S0REVIK, AND P. KEAST, Notes on integration and integersublattices,
Math. Comp., 56 (1991), pp. 243-255.
[202J M. D. MACLAREN AND G. MARSAGLIA, Uniform random number generotors, J. Assoc.
Comput. Mach., 12 (1965), pp. 83-89.
[203] D. MAISONNEUVE, Recherche et utilisation des "bons treillis." Progmmmation et
resultats numer'iques, in ApplicatioDS of Number Theory to Numerical Analysis, S. K.
Zaremba, ed., Academic Press, New York, 1972, pp. 121-201.
[204] G. MARSAGLIA, Random numbers fall mainly in the planes, Proc. Nat. Acad. Sci.
U.S.A., 61 (1968), pp. 25-28.
[205] --, The structure of linear con!J'f"'lJ,ential sequences, in ApplicatioDS of Number The-
ory to Numerical Analysis, S. K. Zaremba, ed., Academic Press, New York, 1972,
pp. 249-285.
[206] G. MARSAGLIA AND L.-H. TSAY, Matrices and the structure of random number se-
quences, Linear Algebra Appl., 67 (1985), pp. 147-156.
[207] P. MARTIN-LoF, The definition of random sequences, Inform. Control, 9 (1966),
pp. 602-619.
[208] J. MATYAS, Random optimization, Automat. Remote Control, 26 (1965), pp. 244-251.
[209] H. G. MEIJER, The discrepancy of a g-adic sequence, Math.) 30 (1968), pp. 54-
66.
[210] N. METROPOLIS, The beginning of the Monte Carlo method, in Stanislaw Ulam 1909-
1984, Los Alamos Sci. No. 15 Special Issue, Los Alamos National Laboratory, Los
Alamos, NM, 1987, pp. 125-130.
[211] N. METROPOLIS AND S. M. ULAM, The Monte Carlo method, J. Amer. Statist. Assoc.,
44 (1949), pp. 335-341.
[212] R. A. MITCHELL, EfTOr estimates arising from certain pseudorandom sequences in a
quasi-random search method, Math. Comp., 55 (1990), pp. 289--297.
[213] B. J. T. MORGAN, Elements of Simulation, Chapman and Hall, London, 1984.
[214] S. MOTTA, G. Russo, H. MOOCK, AND J. WICK, A number-theoretical convergence
proof of a point approximation of a space homogeneous transport equation, Matematiche
(Catania), 41 (1986), pp. 161-178.
[215] G. L. MULLEN AND H. NIEDERREITER, Optimal characteristic polynomials for digital
multistep pseudorandom numbers, Computing, 39 (1987), pp. 155-163.
[216] M. NEWMAN, Integral Matrices, Academic Press, New York, 1972.
[217] H. NIEDERREITER, Discrepancy and convex programming, Ann. Mat. Pura Appl., 93
(1972), pp. 89-91.
[218] --, Methods for estimating discrepancy, in ApplicatioDS of Number Theory to Nu-
merical Analysis, S. K. Zaremba, ed., Academic Press, New York, 1972, pp. 203-236.
[219] --, Application of diophantine approximations to numerical integration, in Dio-
phantine Approximation and Its Applications, C. F. Osgood, ed., Academic Press, New
York, 1973, pp. 129--199.
[220] --, Metric theorems on the distribution of sequences, in Proc. Sympos. Pure Math.,
Vol. 24, American Mathematical Society, Providence, RI, 1973, pp. 195--212.
[221] --, On the distribution of pseudo-random numbers generoted by the linear congru-
ential method. II, Math. Comp., 28 (1974), pp. 1117-1132.
[222] --, On the distribution of pseudo-random numbers generated by the linear congru-
ential method. III, Math. Comp., 30 (1976), pp. 571-597.
[223] --, Statistical independence of linear congruential pseudo-random numbers, Bull.
Amer. Math. Soc., 82 (1976), pp. 927-929.
[224] --, Pseudo-random numbers and optimal coefficients, Adv. in Math., 26 (1977),
pp. 99--181.
[225] --, Quasi-Monte Carlo methods and pseudo-random numbers, Bull. Amer. Math.
Soc., 84 (1978), pp. 957-1041.
232 BIBLIOGRAPHY
[226] ---, Existence of good lattice points in the sense of Hlawka, Monatsh. Math., 86
(1978), pp. 203-219.
[227] ---, Statistical tests for Tausworthe pseudo-random numbers, in Probability and Sta-
tistical Inference, W. Grossmann, G. C. Pflug, and W. Wertz, eds., Reidel, Dordrecht,
the Netherlands, 1982, pp. 265-274.
[228J ---, A quasi-Monte Carlo method for the approximate computation of the extreme
values of a junction, in Studies in Pure Mathematics (To the Memory of Paul Thran),
Birkhauser, 1983, pp. 523-529. (Submitted in 1977.)
[229] ---, Applications des corps finis aux nombres pseudo-aleatoires, Sem. Theo. Nombres
Bordeaux 1982-1983, Exp. 38, Univ. de Bordeaux I, Talence, France, 1983.
[230] ---, On a measure of denseness for sequences, in Topics in Number Theory,
Budapest, 1981, Colloq. Math. Soc. Janos Bolyai, Vol. 34, North-Holland, Amsterdam,
1984, pp. 1163-1208.
[231] --, Optimal multipliers for linear congruential pseudo-random numbers: The dec-
imal case, in Statistics and Probability, J. MogyorOdi, I. Vincze, and W. Wertz, eds.,
Reidel, Dordrecht, the Netherlands, 1984, pp. 255-269.
The
mity test, SIAM J. Sci. Statist. Comput., 5 (1984), pp. 798-810.
[233] ---, Number-theoretic problems in pseudorandom number generation, in Proc. Sym-
pos. on Applications of Number Theory to Numerical Analysis, Kyoto, 1984, Lecture
Notes No. 537, Research Inst. of Math. Sciences, Kyoto University, Kyoto, Japan, 1984,
pp. 18-28.
[234] ---, The serial test for pseudo-random numbers generated by the linear congruential
method, Numer. Math., 46 (1985), pp. 51-68.
[235] ---, Quasi-Monte Carlo methods for global optimization, in Proc. Fourth Pannonian
Sympos. on Math. Statistics, Bad Tatzmannsdorf, 1983, Reidel, Dordrecht, the Nether-
lands, 1986, pp. 251-267.
[236] ---, A pseudomndom vector generator based on finite field arithmetic, Math. Japon.,
31 (1986), pp. 759'-774.
[237] ---, Multidimensional numerical integration using pseudorandom numbers, Stochas-
tic Programming 84, Part I, A. Prekopa and R. J.-B. Wets, eds., Math. Programming
Study, VoL 27, North-Holland, Amsterdam, 1986, pp. 17-38.
[238] --, Dyadic fractions with small parti.al quotients, Monatsh. Math., 101 (1986),
pp. 309'-315.
[239] --, Low-discrepancy point sets, Monatsh. Math., 102 (1986), pp. 155-167.
[240] ---, Good lattice points for quasirandom search methods, in System Modelling and
Optimization, A. Prekopa, J. Szelezsan, and B. Strazicky, eds., Lecture Notes in Control
and Inform. Sci., Vol. 84, Springer, Berlin, 1986, pp. 647-654.
[241] ---, Pseudozufallszahlen und die Theorie der Gleichverteilung, Sitzungsber. OsterI'.
Akad. Wiss. Math.-Natur. Kl. II, 195 (1986), pp. 109'-138.
[242] --, A statistical analysis of generalized feedback shift register pseudorandom number
generator's, SIAM J. Sci. Statist. Comput., 8 (1987), pp. 1035-1051.
[243] ---, Rational junctions with partial quotients of small degree in their continued
fraction expansion, Monatsh. Math., 103 (1987), pp. 269'-288.
[244) ---, Point sets and sequences with small discrepancy, Monatsh. Math., 104 (1987),
pp. 273-337.
[245] ---, Continued fractions with small partial quotients, in Proc. 1986 Nagasaki Sym-
pos. on Number Theory, Y. Morita, ed., T6hoku University, Sendai, Japan, 1987, pp. I-
ll.
[246) ---, The serial test for digital k-step pseudorandom numbers, Math. J. Okayama
Unl"., 30 (1988), pp. 93-119.
[247] ---, Low-discrepancy and low-dispersion sequences, J. Number Theory, 30 (1988),
pp.51-70.
[248] --, Remarks on nonlinear congruential pseudorandom numbeT3, Metrika, 35 (1988),
pp. 321-328. \
BIBLIOGRAPHY 233
[249] --, Statistical independence of nonlinear congruential pseudorandom numbers,
Monatsh. Math., 106 (1988), pp. 149-159.
[250] --, Sequences with almost perfect linear complezity profile, in Advances in Cryptol-
ogy - EUROCRYFT '87, D. Chaum and W. L. Price, ads., Lecture Notes in Computer
Science, Vol. 304, Springer, Berlin, 1988, pp. 37-51.
[251.] --, Quasi-Monte Carlo methods for multidimensional numerical integration, in Nu-
merical Integration III, H. Bra.6 and G. Hammerlin, eds., Internat. Series of Numer.
Math., Vol. 85, Birkhiiuser, Basel, 1988, pp. 151-171.
[252] --, The serial test for con!J'f"lJf!fltial pseudorandom numbers generated by inversions,
Math. Comp., 52 (1989), pp. 135-144.
[253] --, Pseudorandom numbers generated from shift register sequences, in Number-
Theoretic Analysis, E. Hlawka. and R. F. Tichy, OOs., Lecture Notesin Math., Vol. 1452,
Springer, Berlin, 1990, pp. 165-117.
[254J --, Statistical independence properties of pseudorandom vectors produced by matrix
generators, J. Compo Appl. Math., 31 (1990), pp. 139-151.
[255] --, Lower bounds for the discrepancy of inversive congruential pseudorandom num-
bers, Math. Comp., 55 (1990), pp. 217-287.
[256] --, The distribution of values of Kloosterman sums, Arch. Math. (Basel), 56 (1991),
pp. 270-277.
[257] --, Recent trends in random number and rondom vector generation, Ann. Oper.
Res., 31 (1991), pp. 323-345.
[258] --, Finite fields and their applications, in Contributions to General Algebra 7,
Vienna, 1990, Teubner, Stuttgart, Germany, 1991, pp. 251-264.
[259] --, The existence oj efficient lattice rules for multidimensional numerical integra-
tion, Math. Comp., 58 (1992), pp. 305-314, S7-S16.
[260] --, A combinatorial problem for vector spaces over finite fields, Discrete Ma.th., 96
(1991), pp. 221-228.
[261J --, Low-discrepancy point sets obtained by digital constructions over finite fields,
Czechoslovak Math. J., 42 (1992), pp. 143-166.
[262] --, Lattice rules for multiple integration, in Proc. Workshop on Stochastic Opti-
mization, Neubiberg, 1990, Springer Lecture Notes, to appear.
[263] --, Nonlinear methods for pseudorandom number and vector generation, in Simu-
lation and Optimization, G. Pflug and U. Dieter, eds., Lecture Notes in Economics and
Math. Systems, Vol. 374, Springer, Berlin, 1992, pp. 1
L
15-153.
[264] --, Improved error bounds for lattice rules, submitted for publication.
[265J H. NIEDERREITER AND J. HORBOWICZ, Optimal bounds for exponential sums in terms
of discrepancy, Colloq. Math., 55 (1988), pp. 355-366.
[266] H. NIEDERREITER AND K. MCCURLEY, Optimization of functions by quasi-random
search methods, Computing, 22 (1919), pp. 119-123.
[267] H. NIEDERREITER AND P. PEART, A comparative study oj quasi-Monte Carlo methods
for optimization of functions of several variables, Caribbean J. Math., 1 (1982), pp. 27-
44.
[268] --, Quasi-Monte Carlo optimization in general domains, Caribbean J. Math., 4
(1985), pp. 67-85.
[269] --, Localization oJ search in quasi-Monte Carlo methods for global optimization,
SIAM J. Sci. Statist. Comput., 7 (1986), pp. 660-664.
[270] H. NIEDERREITER AND C. P. SCHNORR, Local randomnes3in polynomial random num-
ber and random function generators, SIAM J. Comput., to appear.
[211] H. NIEDERREITER AND J.-S. SHIUE, Equidistribution of linear recurring sequences in
finite fields, Indag. Math., 80 (1977), pp. 397-405.
[272] --, Equidistribution oj linear recurring sequences in finite fields, II, Acta Arith., 38
(1980), pp. 197-20'7.
[273] H. NIEDERREITER AND I. H. SLOAN, Lattice rules for multiple integration and discrep-
ancy, Math. Comp., 54 (1990), pp. 303-312.
[274] H. NJEDERREITER AND R. F. TICHY, Beitriige zur Diskrepanz beziiglich gewichteter
234 BIBLIOGRAPHY
Mittel, Manuscripta Math., 42 (1983), pp. 85--99.
[275] -'--, Solution of a problem of Knuth on complete uniform distribution of sequences,
Mathematika, 32 (1985), pp. 26-32.
[276] --, Metric theorems on uniform distribution and approximation theory, in Journees
Arithmetiques de 1985, Asterisque, Vols. 147-148, Soc. Math.
France, Paris, 1987, pp. 31!f-323.
[277] H. NIEDERREITER, R. F. TICHY, AND G. TURNWALD, An inequality for differences of
distribution jv.nctions, Arch. Math. (Basel), 54 (1990), pp. 166-172.
[278] H. NIEDERREITER AND J. M. WILLS, Diskrepanz und Distanz von MafJen bezuglich
konvexer und Jordanscher Mengen, Math. Z., 144 (1975), pp. 125-134; Berichtigung,
ibid., 148 (1976), p. 99.
[279] N. NIKI, Finite field arithmetics and multidimensional uniform pseudorandom num-
bers, Proc. Inst. Statist. Math., 32 (1984), pp. 231-239. (In Japanese.)
[280] P. PEART, The dispersion of the Hammersley sequence in the unit square, Monatsh.
Math., 94 (1982), pp. 249-261.
[281] O. E. PERCUS AND J. K. PERCUS, Long range correlations in linear congruential
genemcDT3,J Comp, Physics, 77 (1988), pp. 267-269,
[282] J. PINTER, On the convergence and computational efficiency of mndom search opti-
mization, Meth. Oper. Res., 33 (1979), pp. 347-362.
[283] P. D. PROINOV, Note on the convergence of the general quadrature process with positive
weights, in Constructive Function Theory '77, Blagoevgrad, 1977, Bulgar. Acad. ScL,
Sofia, 1980, pp. 121-125. (In Russian.)
[284] --, Upper bound for the error of a general quadrature process with positive weights,
C.R. Acad. Bulgare Sci., 35 (1982), pp. 605-608. (In Russian.)
[285] --, Generalization of two results of the theory of uniform distribution, Proc. Amer.
Math. Soc., 95 (1985), pp. 527-532.
[286] ---, Numerical integration and approximation of differentiable jv.nctions, C.R. Acad.
Bulgare ScL, 38 (1985), pp. 187-190.
[287] --, Numerical integration and approximation of differentiable jv.nctions, II, J.
Approx. Theory, 50 (1987), pp. 373-393.
[288] --, Discrepancy and integration of continuous jv.nctions, J. Approx. Theory, 52
(1988), pp. 121-131.
[289] --, Integration of smooth jv.nctions and cp-discrepancy, J. Approx. Theory, 52
(1988), pp. 284-292.
[290] P. D. PROINOV AND G. KH. KIROV, A quadrature formula for approximate integration
of jv.nctions of class cr[o, 1], C.R. Acad. Bulgare Sci., 36 (1983), pp. 1027-1030. (In
Russian.)
[291] L. RAMSHAW, On the discrepancy of the sequence formed by the multiples of an irra-
tional number, J. Number Theory, 13 (1981), pp. 138-175.
[292] L REINER, On the number of matrices with given characteristic polynomial, Illinois J.
Math., 5 (1961), pp. 324-329.
[293] R. D. RICHTMYER, The evoJ.uation of definite integrals, and a quasi-Monte-Carlo
method based on the properties of algebraic numbers, Report LA-1342, Los Alamos
Scientific Laboratory, Los Alamos, NM, 1951.
[294] B. D. RIPLEY, The lattice st'7'Ucture of pseudo-random number generators, Proc. Roy.
Soc. London Ser. A, 389 (1983), pp. 197-204.
[295] --, Stochastic Simulation, John Wiley, New York, 1987.
[296] K. F. ROTH, On irregularities of distribution, Mathematika, 1 (1954), pp. 73-79.
[297] R. Y. RUBINSTEIN, Simulation and the Monte Carlo Method, John Wiley, New York,
1981.
[298] --, Monte Carlo Optimization, Simulation and Se'nsitivity of Queuing Networks,
John-Wiley, New York, 1986.
[299] H. J. RYSER, Combinatorial Mathematics, Mathematical Association of America,
Washington, DC, 1963.
[300] J. W. SANDER, (),? a conjecture of Zaremba, Monatsh. Math., 104 (1987), pp. 133-137.
BIBLIOGRAPHY 235
[301J P. K. SARKAR AND M. A. PRASAD, A comparative study of pseudo and quasi random
sequences for the solution of integral equations, J. Comp. Physics, 68 (1987), pp. 66-88.
[302] W. M. SCHMIDT, Metrical theorems on fractional parts of sequences, Trans. Amer.
Math. Soc., 110 (1964), pp. 493-518.
[303] --, Irregularities of distribution. VII, Acta Arith., 21 (1972), pp. 45-50.
[304] C. P. SCHNORR, Zufiilligkeit und Wahrscheinlichkeit, Lecture Notes in Math., Vol. 218,
Springer, Berlin, 1971.
[305] J. SCHOISSENGEIER, On the discrepancy of (na), Acta Arith., 44 (1984), pp. 241-279.
[306] --, Uber die Diskrepanz der Folgen (na), in Zahlentheoretische Analysis, E. Hlawka,
ed., Lecture Notes in Math., Vol. 1114, Springer, Berlin, 1985, pp. 148-153.
[307] --, On the discrepancy of (na), II, J. Number Theory, 24 (1986), pp. 54-64.
[308] I. F. SHARYGIN, A lower estimate for the e1T'Or of quadrature formulas for certain
classes of functions, Zh. Vychisl. Mat. i Mat. Fiz., 3 (1963), pp. 370-376. (In Russian.)
[309] J. E. H. SHAW, A quasirandom approach to integration in Bayesian statistics, Ann.
Statist., 16 (1988), pp. 895-914.
[310] S. Z. SHI, Optimal uniform distributions generated by M-sequences, Acta Math. Sinica,
22 (1979), pp. 123-128. (In Chinese.)
[311J --, Estimate of e1T'Or for quadrature of a multidimensional continuous function,
Math. Numer. Siniea, 3 (1981), pp. 360-364. (In Chinese.)
[312] I. E. SHPARLINSKII, The distribution of the fractional parts of recurrent sequences, Zh.
Vychlsl. Mat. i Mat. Fiz., 21 (1981), pp. 1588--1591. (In Russian.)
[313J --, A sequence of pseudorandom numbers, Avtomat. i Telemekh., 7 (1988), pp. 185-
188. (In Russian.)
[314] I. H. SLOAN, Lattice methods for multiple integration, J. Compo Appl. Math., 12/13
(1985), pp. 131-143.
[315J I. H. SLOAN AND P. KACHOYAN, Lattices for multiple integration, in Math. Program-
ming and Numerical Analysis Workshop, Canberra, 1983, Proc. Centre Math. Analysis
Austral. Nat. Univ., Vol. 6, Australian National University, Canberra, Australia, 1984,
pp. 147-165.
[316] --, Lattice methods for multiple integration: Theory, e1T'Or analysis and examples,
SIAM J. Numer. Anal., 24 (1987), pp. 116-128.
[317J I. H. SLOAN AND J. N. LYNESS, The representation of lattice quadrature rules as
multiple sums, Math. Comp., 52 (1989), ~ p . 81-94.
[318J --, I,.attice rules: Projection regularity and unique representations, Math. Comp.,
54 (1990), pp. 649-660.
[319J I. H. SLOAN AND T. R. OSBORN, Multiple integration over bounded and unbounded
regions, J. Comp. AppL Math., 17 (1987), pp. 181-196.
[320] I. H. SLOAN AND L. WALSH, Lattice rules - classification and searches, in Numerical
Integration III, H. BraS and G. Hiimmerlin, eds., Internat. Series of Numer. Math., Vol.
85, Birkhiuser, Basel, 1988, pp. 251-260.
[321J --, A computer search of rank-2 lattice rules for multidimensional quadrature,
Math. Comp., 54 (1990), pp. 281-302.
[322} 1. M. SOBOL', An accurate e1T'Or estimate for multidimensional quadrature formulae
for the functions of the class Sp, Dokl. Akad. Nauk SSSR, 132 (1960), pp. 1041-1044.
(In Russian.)
[323] --, The distribution of points in a cube and the approximate evaluation of integrals,
Zh. Vychisl. Mat. i Mat. Fiz., 7 (1967), pp. 784-802. (In Russian.)
[324] --, Multidimensional Quadrature Formulas and Haar Jilunctions, Izdat. "Nauka,"
Moscow, 1969. (In Russian.)
[325] --, On the systematic search in a hypercube, SIAM J. Numer. Anal., 16 (1979),
PI'. 790-793; Erratum, ibid., 16 (1979), p. 1080.
[326] --, On an estimate of the accuracy of a simple multidimensional search, Dokl.
Akad. Nauk SSSR, 266 (1982), pp. 569-572. (In Russian.)
[327] --, On functions satisfying a Lipschitz condition in multidimensional problems of
computational mathematics, Dokl. Akad. Nauk SSSR, 293 (1987), PI'. 1314-1319. (In
236
[328]
[329]
[330]
[331]
[332]
[333]
[334]
[335]
[336]
[337]
[338]
[339]
[340]
[341]
[342]
[343]
[344]
[345]
[346]
[347J
[348]
[349]
[350]
BIBLIOGRAPHY
Russian.)
--, On the search for extreme values of functions of seveml variables satisfying a
geneml Lipschitz condition, Zh. Vychisl. Mat. i Mat. Fiz., 28 (1988), pp. 483-491. (In
Russian.)
I. M. SOBOL' AND R. B. STATNIKOV, LP-search and problems of optimal design, in
Problems of Random Search, Vol. 1, Izdat. "Zinatne," Riga, 1972, pp. 117-135. (In
Russian.)
F. J. SOLIS AND R. J .-B. WETS, Minimization by random search techniques, Math.
Oper. Res., 6 (1981), pp. 19-30.
V. T. S6s, On strong irregularities of the distribution of {no} sequences, in' Studies in
Pure Mathematics (To the Memory of Paul Turan), Birkhauser, Basel, 1983, pp. 685-
700.
--, Irregularities of partitions: Ramsey theory, uniform distribut'ion, in Surveys
in Combinatorics, E. K. Lloyd, ed., London Math. Soc. Lecture Note Series, Vol. 82,
Cambridge University Press, Cambridge, UK, 1983, pp. 201-246.
S. SRINIVASAN, On two-dimensional Hammersley's sequences, J. Number Theory, 10
(1978), pp. 421-429.
H. STEGBUCHNER, Numerische Quadmtur glatter Funktionen mit Gleichverteilungs-
methoden, Rend. Mat. (7), 2 (1982), pp. 593-599.
S. A. STEPANOV, On estimating mtional trigonometric sums with prime denominator,
Tmdy Mat. Inst. Steklov., 112 (1971), pp. 346-371. (In Russian.)
M. SUGIHARA, Method of good matrices for multi-dimensional numerical integrations
- An extension of the method of good lattice points, J. Camp. AppL Math., 17 (1987),
pp. 197-213.
A. G. SUKHAREV, Optimal stmtegies of the search for an extremum, Zh. Vychisl. Mat.
i Mat. Fiz., n (1971), pp. 910-924. (In Russian.)
--, Best sequential search stmtegies for findir.g an extremum, Zh. Vychisl. Mat. i
Mat. Fiz., 12 (1972), pp. 35-50. (In Russian.)
E. A. D. E. TAHMI, Contribution au:z: genemteurs de vecteurs pseuao-aleatoires, These,
Univ. Sci. Techn. H. Boumedienne, Algiers, 1982.
R. J. TASCHNER, Eine zahlentheoretische Methode .rur Bestimmung von Niihernngslo-
sungen einer Anfangswertaufgabe, Sitzungsber. Osterr. Akad. WiBB. Math.-Natur. Kl.
n, 189 (1980), pp. 277-284.
--, Eine zahlentheoretische Methode zur Besiimmung von Niiherungslosungen
nichtlinearer Gleichungssysteme, Sitzungsber. Osten-. Alq;,d. Wiss. Math.-Natur. Kl.
II, 189 (1980), pp. 285-289.
R. C. TAUSWORrHE, Random numbers genemted by linear recurrence modulo two,
Math. Comp., 19 (1965), pp. 201-209.
V. N. TEMLYAKOV, On reconstruction of multivariate periodic functions based on their
values at the knots of number-theoreii.cal nets, Anal. Math., 12 (1986), pp. 287-305.
(In Russian.)
--, Error estimates for quadmture formulas fot classes of functions with a bounded
mixed derivative, Mat. Zametki,46 (1989), pp. 128-134. (In Russian.)
S. TEZUKA, On the discrepancy of GFSR pseudorandom numbers, J. Assoc. Comput.
Mach., 34 (1987), pp. 939-949.
--, On optimal GFSR pseudomndom number generators, Math. Comp., 50 (1988),
pp. 531-533.
A. THOMAS, Discrepance en dimension un, Ann. Fac. Sci. Toulouse Math. (5), 10
(1989), pp. 369-399.
R. F. TICHY, Zur Diskrepanz beziiglich gewichteter Mittel, Manuscripta Math., 37
(1982), pp. 393-413.
--, Uber eine zahlentheoretische Methode ruT numerischen Integration und ruT Be-
handlu;'g von Integralgleichungen, Sitzungsber. Osten-. AkBd. Wiss. Math.-Natur. Kl.
II, 193 (1984), pp. 329-358.
--, Ein metrif{cher Satz ii.ber vollstiindig gleichuerteilte Folgen, Acta Arith., 48
[351]
[352]
[353]
[354]
[355J
[356]
[357]
[358J
[359]
[360]
[3tH]
[362]
[363]
[364]
[365]
[366]
[367]
[368]
[369]
[370]
[371]
BIBLIOGRAPHY 237
(1987), pp. 197-207.
--, Random points in the cube and on the sphere with applications to numerical
analysis, J. Compo AppL Math., 31 (1990), pp. 191-197.
A. TORN AND A. ZILINSKAS, Global Optimization, Lecture Notes in Computer Science,
Vol. 350, Springer, Berlin, 1989.
G. A. TOTKov, On the convef9ence of multidimensional quadrature formulae, C.R.
Acad. Bulgare Sd., 37 (1984), pp. 1111-1174. (In Russian.)
N. TRENDAFILOV, Uniformly distributed sequences and vector-valued integration, C.R.
Acad. Bulgare ScL, 40 (1987), no. 10, pp. 23-25.
V. A. USPENSKII, A. L. SEMENOV, AND A. KH. SHEN', Can an individual sequence
of zeros and ones be random?, Uspekbi Mat. Nauk, 45 (1990), no. 1, pp. 105-162. (In
Russian.)
J. G. VAN DER CORPUT, Verteilungs/uriktionen I, II, Nederl. Akad. Wetensch. Proe.
Ser. B, 38 (1935), pp. 813-821, 1058-1066.
A. WElL, On some exponential sums, Proe. Nat. Acad. Sci. U.S.A., 34 (1948), pp. 204-
207.
B. A. WICHMANN AND I. D. HILL, An efficient and portable pseudo-random number
generator, Appl. Statist., 31 (1982), pp. 188-100; Corrections, ibid., 33 (1984), p. 123.
--, Building a random-number genemtor, Byte, 12 (1987), no. 3, pp. 127-128.
S. WOLFRAM, Random sequence generation by cellular automata, Adv. App1. Math., 7
(1986), pp. 123-159.
S. K. ZAREMBA, Good lattice points, discrepancy, and numerical integration, Ann. Mat.
Pura Appl., 73 (1966), pp. 293-317.
--, Good lattice points in the sense of ffiawka and Monte Carlo integration,
Monatsh. Math., 72 (1968), pp. 264-269.
--, Some applications of multidimensional integration by parts, Ann. Polon. Math.,
21 (1968), pp. 85-96.
--, La discrepance isotrope et l'-integration numerique, Ann. Mat. Pura Appl., 87
(1970), pp. 125-136.
--, La methode des "bons treillis" pour Ie calcul des integrates multiples, in Appli-
cations of Number Theory to Numerical Analysis, S. K. Zaremba, ed., Academic Press,
New York, 1972, pp. 39-119.
--, Good lattice points modulo composite numbers, Monatsh. Math., 78 (1974),
pp. 446-460.
--, Computing the isotropic discrepancy of point sets in two dimensions, Discrete
Math., 11 (1975), pp. 79-92.
A. A. ZHIGLJAVSKY, Theory of Global Random Search, Kluwer Academic Publishers,
Dordrecht, the Netherlands, 1991.
P. ZINTERHOF, tiber einige Abschatzungen bei der Approximation von Funktionen mit
Gleichverteilungsmethoden, Sitzungsber. Osterr. Akad.. \Viss. Math.-Natur. Kl. II, 185
(1976), pp. 121-132.
--, Gratis lattice points for multidimensional integration, Computing, 38 (1987),
pp.347-353.
P. ZINTERHOF AND H. STEGBUCHNER, Trigonomdrische Approximation mit Gleich-
verteilungsmethoden, Studia Sci. Math. Hungar., 13 (1978), pp. 273-289.
Index
adaptive methods, 152
antithetic variates, 8, 107
basis test, 173
bounded metric space, 148
Cartesian product
of integration rule, 1
of trapezoidal rule, 2
central limit theorem, 5
characteristic polynomial, 218
completely uniformly distributed,
162
composition method, 165
congruential generator, 177
convergent, 219, 220
copy rule, 144
crude search, 150
curse of dimensionality, 2
determinant of lattice, 131
diaphony, 22
digital multistep method, 192
digital multistep pseudorandom
numbers, 192
discrepancy, 13, 21
discrete, 215
extreme, 14
isotropic, 17, 21
star, 14
discrete discrepancy, 215
discretization error, 42
239
dispersion, 148
dual basis, 218
dual lattice, 126
elementary interval, 48
extreme discrepancy, 14
Fibonacci numbers, 122
figure of merit, 108, 133, 171, 193,
200, 209
first-order congruential method, 177
first-order nonlinear method, 213
formal Laurent series, 88, 98
fractional part, 26, 28
frequency test, 166, 168
gap test, 166
generalized feedback shift-register
(GFSR) method, 198
generalized feedback shift-
register (GFSR) pseudo-
random numbers, 199
generalized Halton sequence, 44
generalized Hammersley point set,
44
generalized van der Corput se-
quence, 25
generator matrix, 131
good lattice point, 109, 158
grid, 173
Halton sequence, 29, 44, 45, 157, 159
240
Hammersley point set, 31, 45, 157,
158
Hermite normal form, 132
integration lattice, 126
invariants, 131
inversion method, 164
inversive congruential generator,
182, 187
inversive congruential method, 182
inversive congruential pseudoran-
dom numbers, 182, 187
inversive method, 214
irregularities of distribution, 24, 33,
154
isotropic discrepancy, 17, 21
Koksma's inequality, 18, 21
Koksma-Hlawka inequality, 19
Kolmogorov complexity, 163
Kolmogorov test, 166
Kolmogorov-Smirnov law, 166
latin square, 61
lattice, 126
lattice basis, 131
lattice methods, 101
lattice rule, 101, 126
lattice structure, 172, 208
lattice test, s-dimensional, 178
lattice tests, 173
law of the iterated logarithm, 166-
168
linear congruential method, 168
linear congruential pseudorandom
numbers, 169
linear recurring sequence, 218
localization of search, 151
low-discrepancy point set, 23
low-discrepancy sequence, 23
matrix generator, 206
matrix method, 205
maximal period sequence, 218
maximum metric, 150
method of good lattice points, 101
\
INDEX
minimal polynomial, 218
mixed congruential method, 169
modulus of continuity, 18, 20, 148
Monte Carlo estimate, 4-6
Monte Carlo method, 2, 3
multiple recursive method, 174
multiple-recursive pseudorandom
numbers, 174
multiplicative congruential method,
169
multiplier, 169
mutually orthogonal, 62
mutually orthogonal squares, 62
net, 46
node set, 127
nonlinear congruential generator,
178
nonlinear congruential methods, 178
nonlinear congruential pseudoran-
dom numbers, 178
nonuniform pseudorandom num-
bers, 164
nonuniform random numbers, 162
nonuniformity, 22
normal numbers, 163
orthogonal latin squares, 6(}-62
outer iteration, 152
partial quotients, 219, 220
periodic, 164
periodization, 107
permutation polynomial, 178
permutation test, 167
physical random numbers, 163
point set, 14
generalized Hammersley, 44
Hammersley, 31, 45, 157, 158
low-discrepancy, 23
primitive polynomial, 217
projection regular, 132
pseudorandom numbers, 3, 164-168
pseudorandom vectors, 3, 205-208
quadratic congruential method, 181
quadratic congruential pseudoran-
dom numbers, 181
quasi-Monte Carlo approximation,
10
quasi-Monte Carlo method, 3, 9
quasirandom points, 23
quasirandom search, 148
quasirandom sequences, 23
radical-inverse function, 24
random number generation, 3, 161
random search, 147, 158
random variates, 162
random vector generation, 3, 205
rank, 131
ratio-of-uniforms method, 165
rejection method; 165
run test) 166
sequence, 14
completely uniformly
distributed, 162
generalized Halton, 44
generalized van der Corput, 25
Halton, 29, 45, 157, 159
linear recurring, 218
low-discrepancy, 23
maximal period, 218
periodic, 164
quasirandom, 23
(t, s)-, 46
uniformly distributed, 13
van der Corput, 25, 45, 46, 74,
92
sequence, (t, B )-, 46
serial correlation, 167
serial correlation coefficient, 167
serial test, 167, 208
shift-register pseudorandom num-
bers, 191
simulation methods, 6
spectral test, 168
standard deviation, 5
star discrepancy, 14
statistical tests for randomness, 166
stochastic simulation, 6
INDEX
stratified sampling, 7, 144
strong law of large numbers, 4
totally bounded, 149
trace, 217
trapezoidal rule, 1
241
uniform pseudorandom numbers,
164
uniform pseudorandom vectors, 205
uniform random numbers, 162
uniformity test, 166
uniformly distributed, 13
van der Corput sequence, 25, 45, 46,
74,92
variance, 4
variance reduction, 7
variation, 19, 22
Zaremba's conjecture, 123