ELSEVIER

Discrete Time Spatial Models Arising in Genetics,

Evolutionary Game Theory, and Branching Processes
J. RADCLIFFE AND L. RASS
School of Mathematical Sciences, Queen Mary and WesOqeldCollege,
University of London, Mile End Road, London E1 4NS, England
Received 29 December 1995; revised 17 September 1996

ABSTRACT
A saddle point method is used to obtain the speed of first spread of new
genotypes in genetic models and of new strategies in game theoretic models. It is
also used to obtain the speed of the forward tail of the distribution of farthest spread
for branching process models. The technique is applicable to a wide range of models.
They include multiple allele and sex-linked models in genetics, multistrategy and
bimatrix evolutionary games, and multitype and demographic branching processes.
The speed of propagation has been obtained for genetics models (in simple cases
only) by Weinberger [1, 2] and Lui [3-7], using exact analytical methods. The exact
results were obtained only for two-allele, single-locus genetic models. The saddle
point method agrees in these very simple cases with the results obtained by using the
exact analytic methods. Of course, it can also be used in much more general
situations far less tractable to exact analysis.
The connection between genetic and game theoretic models is also briefly
considered, as is the extent to which the exact analytic methods yield results for
simple models in game theory. © Elsevier Science Inc., 1997

1. I N T R O D U C T I O N
A saddle point method has been used to obtain the speed of
propagation for certain continuous time models when the spatial aspect
is described by contact distributions. The main area where the method
has been applied is in the modeling of epidemics. It has also been
applied to contact branching processes.
The result for a simple one-type S ~ I epidemic model on a line was
obtained by Daniels [8]. An application to the spatial birth process is
given in Ref. 9. A rigorous approach to the saddle point method is given
in Ref. 10, in which the speed of propagation for an n-type S ~ I ~ R
model is derived. The results were later extended to cover N-dimen-
sional nonsymmetric contact distributions when the infection matrix is
reducible [11].

MATHEMATICAL BIOSCIENCES 140"101-129 (1997)

© Elsevier Science Inc., 1997 0025-5564/97/$17.00
655 Avenue of the Americas, New York, NY 10010 PII S0025-5564(97)00154-X
102 J. RADCLIFFE AND L. RASS
Rigorous analytical techniques confirmed the speeds of propagation
obtained by the saddle point method; these rigorous techniques were
restricted to the case in which the contact distributions are symmetric
[12-151.
The saddle point method has also been applied to other models for
which an exact analysis is less tractable. These models include epi-
demics allowing a return to the susceptible state and births into the
system [16, 17]. The method gives the speed of first spread of the
forward front of infection. For stochastic contact branching process
models, the speed at which the extreme tail of the distribution function
of farthest spread moves out in any specified direction is given in Ref.
17.
This paper considers discrete time models, which are more appropri-
ate for the modeling of genetics, evolutionary game theory, and certain
branching processes. Again, contact distributions are used to model the
spatial aspect. They represent the distance moved by individuals either
at birth or after reaching maturity prior to the production of the next
generation. This is likely to be more realistic than allowing continual
diffusion during the lifetime of an individual.
A saddle point method can be used for these discrete time models.
Results are obtained for a general model. This is then interpreted in the
context of specific models used in genetics, game theory, and branching
processes, yielding results on the speed of first spread of an allele, a
genotype, or a strategy in an initially stable population, as well as the
speed of the forward tail of the distribution of farthest spread for
branching process models. These results are shown to be consistent with
the results for simple two-allele, single-locus genetic models considered
by Weinberger [1, 2] and Lui [3, 4, 5, 7].
The connection between genetics and game theoretic models is also
briefly considered, as is the extent to which the exact methods yield
results in a simple game theory context.

2. THE SADDLE POINT METHOD

FOR DISCRETE TIME MODELS
Consider a nonspatial discrete time model of the form

X!re+l) .~-fi[X(ra)],
where X~ m) = {x(ra)}i is the value of the ith variable at the mth time
point. The variables represent the proportions of specific types, where
the interpretation of type depends upon the application. In genetics,
types correspond to specific allelles or genotypes; whereas, in evolution-
ary game theory, types refer to individuals playing a given strategy.
DISCRJZE TIME SPATIAL MODELS 103

A spatial aspect that may be continuous or discrete is introduced. In

the continuous-space models, individuals may be at any position r in
RN. For discrete-space models, individuals may be at any site on the
regular lattice Z N. Type j individu als at position r are allowed to move
to a new position s either at birth or upon maturity before giving rise to
the next generation. For spatial models therefore, xim’(s) represents the
value of xi at position s for generation m. Here xi(‘@+ l)(s) depends upon
xjm)(r) and the contact distribution p(r) representing the vector dis-
tance r moved, which may depend upon the type of individual.
Consider a population containing n1 types, which is initially stable
with xi(s) = qi > 0 for i = 1,. ..,n,. Here vi = fi(n) for i = l,.. .,n, with
{q}i = vi for i = 1,. . . , n, and {v}~=O for i=n,+l,...,n. New types
i=n,+l,..., n are introduced into a bounded region B of RN (or Z”‘).
Hence, if this is taken to form part of the zeroth generation, xi(‘)(s) = ni
>Ofor sGB and i=l,..., n,, and X)‘)(S)= ni = 0 for s e B and i = n,
+1 ,. . . , n. Then, for s far from region B, the approximate equations
corresponding to a wide range of models (examples of which are given
in Sections 3, 4, and 5) are given, for the continuous-space model, by

for i=n,+l,..., n, where the rij are nonnegative, with their values and
those of the pii depending on the particular application. For the
discrete-space model, the integral is replaced by a summation over ZN.
For the continuous-space model, we consider the speed of spread of
the forward front for type i in a specified direction, with direction
cosines (Y.Take 5 to be small and positive, and define s(m) so that

xi”‘(u) du = 5.
/ u: a’u 2%
s(m)

For the discrete-space model, we consider the speed of spread of the

forward front for type i in the direction of one of the coordinates of the
lattice (which without loss of generality may be taken to be the first
coordinate). This corresponds to using direction cosines (Y’= (l,O, . . . ,O).
For 6 small and positive, define s(m) to be the smallest integer such
that C o’ tujl> s(mj~i(m)(~)=G5. Define t(m) to be the exact value of the
summation.
In either case, the speed of first spread for type i individuals is
c=lim ,,,Md/ml.
104 J. RADCLIFFE AND L. RASS
The value of c for Equation (1) can be obtained by using the saddle
point method. Let [e*<m)(A)]~and P~(A) be the Laplace transforms of
x}m)(s) and pit(s), respectively. In the continuous case, [L*tm)(A)]i=
fR,~eX*x}")(s)ds and Pi~(A)= fR,~eX~pit(s)ds. In the discrete case, the
integrals are replaced by summations over Z N. Then, from Equation (1),
we obtain
L~(m + 1)(A) = A* ( A)L*(m)(A)
and hence
= [A* ( A)] m
where {A* (X)}ij = TitP?j (X).
Now define e(m)(A)=e*(m)(Aa), Pii(A)=P*(AoD and {A(A)}/t=
{A*(Aa)}~j = TitPi~(Aa) = T~jP/t(A). Note that {L~m)(A)}/ is the Laplace
transform of the integral (summation) of the function x}m)(u) over the
hyperplane a ' u = s. Then
L~m)(A) = [A( A)] mL~°)(A). (2)
The po.(r) are restricted so that Pzt(A) exists in an open neighbor-
hood of A = 0. This implies that A(A) exists and has analytic entries for
all A* < Re(A) < A, where A and A* are the minimum and maximum,
respectively, of the abscissae of convergence of the P~j(A) in the
right-hand and the left-hand half of the complex plane. Each Pit(A) is
taken to be infinite at its abscissae of convergence.
We now consider the continuous and discrete cases separately.
CASE 1: C O N T I N U O U S SPACE

We impose the additional condition that, for any real 01, 0 z with
A* < 0~ < Oz < A, there exists a kij(y), which is integrable, with IPij(O +
iy)l <~ kij(Y) for 01 ~<0 ~ 02. This condition ensures that [L~m)(A)]i is
absolutely integrable over the line A = 0 + iy for -oo < y < ~ for any
A* < Re(O)< A. The 0 chosen will depend upon rn and hence is
denoted by O(m).
Because the absolute integrability allows a change of order of inte-
gration, the following equation is obtained, provided O(m) > 0:

f•
-u" ,~'u ~ s ( m )
x}m>(u) du

1 f~ f O ( m ) + i oo - A w , (rn~: , x
= ~ t t e L i "(a) dAdw
Z.'l'gl J s ( m ) J o ( m ) _ i~
1 c O ( m ) + ioo . . , -
= -a-'--=l A-'e-^stm)|[A(A)]mL(°)(A)}idA.
~q'l'l J o ( m ) - i~
DISCRETE TIME SPATIAL MODELS 105
For simplicity, the discussion is confined to the case where (Yij) is
nonreducible, so A(A) is nonreducible. The extension to the reducible
case can be obtained by using the approach in Ref. 11.
First, consider the dominant term in [A(A)]mL(°)(A) for A = 0 real
with A* < 0 < A. Let p(0) and E(0) be the Perron-Frobenius root and
corresponding idempotent of A(0). Then

A ( 0 ) mL(°)(0) = [ p( O)] mE( 0)L(°)(0).

From Lemmas 1 and 4 in Ref. 11, 0(A) can be defined as the

eigenvalue of A(A) with largest real part [with corresponding idempo-
tent E()0] for A in an open neighborhood of 0, for any A* < 0 < A. In
this neighborhood, p(A) is analytic and

e-;~s(m){[A( A)] mE(°)( A)}i =e-XS(m)+ml°g[°(;O]{E(A) L(°)( A)}i.

Take g(A) = - As(m)+ ma(A), where a(A) = log[ p(A)]. In the follow-
ing lemma, Re[g(A)] is shown to have a saddle point on the real line.
The O(m) used in the integration is taken as this saddle point.
LEMMA 1
Re[g( A)] has a saddle point on the real line at A= 0 where 0 satisfies
the relation a' ( O) = [s( m ) / m ].
Proof Consider the matrix A(O + iy) for y • O. Its (ij)th entry is
yijPiy( O + iy) and I'YijPij( O + iy)[ <<.Yi "P~(0) with strict inequality if "Yij --A:
0. Hence from Lemma 2 in Ref. 10, (p'~O + iy)l < p(O) for y ~: 0 and any
A* < 0 < A. This implies that Re[a(O +/y)] < a(O) and hence Re[g(O +
iy)] < g(O) for y ~ 0.
Now, for y small,

Re[g(O + iy)] -_ R e [[ g ( O ) + i y g ' ( O ) - 2 gmy-

2 )j=g(O
"(O ) -- Y'~2d'( 0 ).

Hence necessarily a"(0)>~ 0 for all A* < 0 < A. Note that g(O)---,oo as
0 1' A and as 0 J, A*. Hence g(O) is a convex function for all real 0 such
that A* < 0 < A and has a minimum at a point such that g'(O) = 0; that
is, such that a'(O)= [s(m)/m]. This point is the saddle point of g(h).
Therefore 0 = O(m) is taken to be the real value satisfying a'(O)=
[s(m)/m]. Because lim m _.~[s(m)/m] = c, O(m) will also tend to a limit
00 as m --->% with a'(O o) = c.
It is assumed that there is a limit c with 00 > 0 and, for simplicity,
that A(00) has distinct eigenvalues. The next step is to find an approxi-
mation to ~ for rn large. To do this, we require the following theorem,
the proof of which is given in the Appendix.
106 J. R A D C L I F F E A N D L. RASS

THEOREM 1
Given any 8 * > 0 there exists an M, E, and 8, with 0 < ~ < I and
0 < 8 < 8", such that
pO(m)e-gt°tm)l( 2zr1
O(m) eg[O(m,+iy]_g[o(m,]lE
× _80(m)+iy t [ O(m)+iy]L(°)[O(m) +iYl}idy [
<Em
for m > M.
Hence, for any positive 8 and for m sufficiently large,

~'0( m )e -St o(m)l= ~ {E[ O( m)] L(°)[ O(m)]}i

[8 O(m) eg[O(m)+iyl_g[o(m)ldy
x j_ o ¥

Now g[ O(m) + iy]- g[ O(m)] = iyg'[O(m)]- ~y2g"[O(m)] for y small.

Also g'[O(m)] = 0 from Lemma 1. Hence

~O(m)e -gt°(m>l= ~-~ {E[ 0(m)] L(°)[ O(m)]]iL~se-l/z'2'~a"t°(m'ldy

{E[O(m)]L(°)[O(m)]}i 1
~/2~rd'[O(m)] vr-m "

Therefore we obtain
s(m) a[O(m)]= 1 log(m)
m O(m) 20(m) m
and hence

c = lim s ( m ) = a ( 0 o )
m~o m 00 '

where 00 is such that a'(Oo)= c.

Define f()0 = [a(A)/A]. Note that f'(O o) -- [a'(0 o ) - f(Oo)]/O o = O.
Also

f"(Oo) = d'(Oo)-2f'(OO)oo = ">-O- ~' do' ( O ° )

Therefore f(O o) is the minimum of f()O.

DISCRETE TIME SPATIAL MODELS 107

Hence c = inf,> o[a(A)/A], provided that this is positive. When the

infimum of f(A) is not positive, then s ( m ) / m does not tend to a
positive limit as m ~ ~. This implies that, if you move at any positive
speed in the specified direction, you will eventually cross the forward
front. The speed of the front is therefore taken to be zero.
Thus the speed of propagation obtained by using the saddle point
method is

where a(A) = log{ p[A(A)]} with {A( A)}0 = 7ijPij(A).

CASE 2: DISCRETE SPACE
Inverting the Laplace transform in Equation (2), for a suitable choice
of O(m) with 0 < O(m) < A, gives

E xm)(u)
a : {'}1 = v

"tr --

"~-2-~ L : [ °(m)+'Y]V{L(m)[O(m ) + & ] }i dY

= f2,/r
_ ' I r l 7r e -[O(m)+iy]v{Pm[O(m) + iy]L(o)[o(m)+iY])idY"

Hence

~(m) = Y'~ x~m)(u)

u: {u}1~ s(m)
1 ~ e -[e(m)+iy]s(m)
{pm[o(m)+~]L(°)[O(m)+~]}idY"

As in Case 1 for continuous space, A = O(m) is taken to be the saddle

point of g(A) = - As(m)+ ma(A), where a(A) = log[ p(A)] and O(A) is
the eigenvalue of A(A) with largest real part with corresponding idem-
potent E(A). Again, A(A) is taken to have distinct eigenvalues when
A = 0 0 = limm -~® O(m).
We now state the theorem equivalent to Theorem 1. The proof
follows in an almost identical manner, except that a = ~r and the
component 11 is not required. The proof is therefore omitted.
108 J. RADCLIFFE AND L. RASS
THEOREM 2
Given any 6" > 0 there exists an M, ~, and 6, with 0 < ~ < 1 and
0 < 8 < 8", such that

[1 _e_O(m)]e_gtO(m)l;~(m)_ 2_~ f 8 rt~l _~ - ~,-O(m)l

- - - - ~ .-] ,e g[O(m)+iy]-g[O(m)]
- 8 [1 - e -°(m)-'y ]
1

X {E[ O(m) + ~1 L(°'[ O(m) + iyl}idyI< ~:m

I
f o r m > M.
Proceeding in a similar manner to Section 2, we obtain
lim m_.~o[s(m)/m] = a(Oo)/O o and hence c = max{0,inf~> 0[a(A)/A]},
where a(A) = log{ o[A(A)]}. This concludes Case 2.
The result obtained for the speed of first spread is the same for both
cases (1 and 2) and is summarized in the following theorem.
THEOREM 3
The speed of propagation obtained by using the saddle point method is

c = m a x [ 0 ' i n' fx a>( -0~ ]

where a(3,) = log{ 0[A(A)]} with {A( A)}o. = yijP~j(A).

3. SELECTION-MIGRATION MULTIPLE ALLELE

GENETIC MODELS
Genetic models are discussed in Ref. 18. Discrete time contact
models have been used in genetics to describe the spatial spread of a
gene when there are two alleles present. The n allele extension of this
model is discussed in Section 3.1. An exact analysis of the two-allele
model is given in Refs. 1-5, 19, and 20.
A selection-migration model with sex-linked locus, again when there
are two alleles present, has been analyzed by Lui [6, 7]. There is, again,
an n allele analogue, which is considered in Section 3.2.
More-general models that allow the migration to depend upon the
genotype can be postulated. Such models are formulated in Sections 3.3
and 3.4.
The saddle point method is applied to all the models in this section
to give the speed of first spread of a new allele (genotype) in a
population that is initially stable. For the simple case with two alleles,
the results obtained are exactly the results obtained by using the exact
analytic methods of the authors referred to above.
DISCRETETIME SPATIAL MODELS 109
The saddle point method can be applied in a fairly obvious way to
more-complex models including several loci. We do not pursue this
further, because the extension is clear.

3.1. A MULTIPLE ALLELE MODEL

The dynamics of the model are described in Ref. 1 for two alleles. A
habitat that has constant carrying capacity at all points of R N or all
points of the N-dimensional lattice Z N is considered. Consider the
composition of the alleles for generation m at position s prior to
mating. Let x!m)(s) be the proportion of allele Ai, for each i = 1.... , n.
Random mating occurs at each position s, the number of offspring
being unaffected by the genotype. Survival to maturity is governed by
the fitness matrix W and reaping occurs to reduce the population to the
carrying capacity of the habitat. Migration then takes place, with p(r)
being the density function (probability) in the continuous (discrete)
space model corresponding to migration by a vector distance r in R N
(zN). A new generation is then produced as before, and the preceding
generation dies. Define wij = {W}ij.
The equations describing the continuous-space process are given
below. Equivalent equations are obtained for the discrete-space process.

x~m+ 1'(s) = f n f i[x(~'(r) l P ( s - r ) dr, (4)

where
x ~ ) ( r ) {Wx(~)(r) }i
fitx(m)(r)] = [x(m,(r)],Wx(m,(r)

for i = 1..... n. Here [x('n)(r)]i = x~m)(r).

Consider the spread of a single new allele when introduced into a
bounded region B of the habitat by mutation or immigration, this being
taken to form part of the zeroth generation. Prior to this occurrence,
the population contained n I alleles in stable equilibrium, there being a
proportion ~/i of allele A i for i = 1 .... ,n 1 at all positions s. Here
7 / / = { w * - l l } i / ( I ' W * - l l ) > 0 for i = 1..... nl, where W* is the part of
the fitness matrix relating to the first n 1 alleles. Note that, when there is
a stable interior equilibrium for the n 1 allele model, necessarily it is
unique and W* is nonsingular with W * - 11 > 0. Let the new allele be
A , , where n = n 1 + 1. Define r/n = 0 and let {7/}/= r/i for i < n r Then the
zeroth generation has x(°)(s) -= 7/for s ~ B.
Consider a position s in the forward region, far from the region B
where the new allele A . was introduced. Then x(m)(s) is small and
x~m)(s) is close to r/i for i = 1..... n 1. In the forward front, Equation (4)
110 J. RADCLIFFE AND L. RASS
for i = n may be approximated by

X(D + 1)(S) = fRNp( S - - r ) ( fn(~) + j=l[ I

0{X}j ] x = [ x}m)(r)-~j]
) dr.

(5)

Now,

[ dfn(x) ] = / 0{Wr/}n j • n,
[ 0{x}y Jx~n [ 7/'W~/ j=n.

Hence we obtain the following approximate equations, which are valid

in this forward region:

- f x(~m)(r){Wr/}n
x~m+I)(S) --JR N ~'Wr/ p ( s - r ) dr. (6)

Equation (6) is just a special case of Equation (1) when a single new
type is introduced. The equation has n = n 1 +1, 3,nn= {W'0}n/(r/'W~7)
and p,n(s) = p(s). If we consider the speed of first spread c in a specific
direction for the new allele An, then, from Theorem 3,

c = max(0, inf l ° g [ P ( h ) ] +l°g(3,) ) (7)

X>0 A '

where P(A) is the Laplace transform of the projection of the migration

distribution p(s) in the specified direction and 3' = {W~}n/(ffWT/). The
same result is obtained for the discrete-space process.
Suppose that there are several new alleles. Let c i be the speed of
spread if only a single allele Ai is introduced into the original system.
Its speed of spread is given by Equation (7) with 3' =
nl nl n1
Y'~jffilWijT~j/~.,kffilY'~]ffil"l'~kWjkT~j.NOW suppose that we take n = n 1 + k and
introduce k new alleles An1 + 1,..-, An into the system at the same time.
Let c = maX(nl+l) < r ~ n Cr; then the approximation in the forward front
will be valid only for alleles A i with c i = c; that is, with fastest speed.
This corresponds to alleles Ai, from i = n I + 1,..., n, for which Z,~L lwij~lj
is a maximum. The value of c is the speed at which_a disturbance to the
equilibrium is first felt, this disturbance being caused by the relatively
fittest new allele(s).
DISCRETE TIME SPATIAL MODELS 111
The case n = 2 and n 1 = 1 was considered in Refs. 1-5, 19, and 20.
Exact analytic methods were used, the problem being split into four
cases. The case when Wl~ = wÂ2 = w22 is excluded, because, in this case,
fxt")(s)ds -- fx(°)(s)ds for all m. The first case has Wll ~ W12 ~ W22, with
at least one inequality strict, so the fitness of the heterozygote is
intermediate between the homozygote fitnesses. The second (and third)
cases corresponded to the heterozygote fitness being greater than (or
less than) both homozygote fitnesses. The fourth case has w22 ~ w12
Wll , with at least one inequality strict, so the heterozygote is again
intermediate, but the homozygote corresponding to allele A 1 is more fit
than the homozygote corresponding to the new allele A 2.
In the first case, the speed of propagation was obtained, provided Wll
and w22 are not equal and

1
(8)

The speed of propagation was obtained for all values of the fitness
parameters covered by the second case. In the fourth case, allele 2 dies
out uniformly in R N for the given initial conditions that x(2°)(s)= 0 for
s e~ B for some bounded region B. For the third case, no expression was
given for the speed of propagation; nor was any way given of obtaining
it. However, note that allele 2 dies out uniformly in R N if, in addition to
the condition that x(2°)(s) = 0 for s ~ B, it is also true that

Wll -- WI2
Wll d" W22 --2w12
for all s ~ B. The values of c obtained in the first and second cases
when n = 2 correspond to our general results for n alleles.

3.2. A MULTIPLE A L L E L E MODEL WITH SEX-LINKED LOCUS

A two-allele model with sex-linked locus was considered by Lui [6, 7].
The dynamics are similar to that of Section 3.1 except that there are
separate male and female genetic outputs and the migration distribu-
tions for males and females may differ. Let y/(')(s) and z~m)(s) be the
proportions of the allele A i in the female and male outputs at location
s in generation m prior to mating. The male is haploid for the X
chromosome and therefore only has n genetic types, the ith being ,4 i.
The female is diploid and so has n ( n + 1 ) / 2 genotypes A i A j for
1 ~<i ~<j ~<n. Random mating occurs between males and females at
position s, the number of offspring produced being unaffected by the
genotype of the parents. The survival fitness of the male A i genotype is
112 J. RADCLIFFE AND L. RASS
v i = {v}i and of the female A i A j genotype is wi/= {W}i/. Reaping occurs
to maintain the densities of males and females. Migration now occurs,
the density functions for the distance r moved being p(r) for females
and q(r) for males. A new generation is then produced as before and
the preceding generation dies. The equations for the continuous-space
model are given below. There is an equivalent discrete-space model.

y~i'n +')(s) = y~m)(r){Wz(m)(r)}i + z~m)(r){Wy(m)(r) }i'p( s - r ) dr,

(2(y(m)(r))'Wz(m)(r)) (9)
z~- + 1)(s ) = fR N [v,y(m)(r)
viY~m)(r) ] q ( s - r ) dr,

for i = l,...,n.
Consider the spread of an allele A~ when introduced into a bounded
region B of a habitat by mutation or immigration, this being taken to
form part of the zeroth generation. Prior to this occurrence, the
population contained ( n - 1) alleles in stable equilibrium, there being
proportions ~i (~i) for females (males) of allele A i for i = 1.... , n - 1 at
all positions s. Here

{[W* d/ag(v* ) + d/ag(v* )W* ] -11}i

>0,
7//= {I'[W* diag(v* ) + d/ag(v*)W* 1-11}

{[W* +diag(v*)W*digg(v*)-l]-ll)i>0, (10)

~'i= {l'[W* + d i a g ( v * ) W * d i a g ( v * ) - l ] - l l }

for i = l ..... n - l , where W* and v* are the parts of the fitness

matrices relating to the first ( n - 1) alleles. Define ~, = 0 and ~, = 0.
Let {~}i =~i and {~}i = ~i- Then the zeroth generation has y(°)(s)-
and z(°)(s) = ~ for s ~ B.
Consider a position s in the forward region, far from the region B
where the new allele A~ was introduced. Then y~m)(s) and z~m)(s) are
small, y/tm)(s) is close to ~i, and z!m)(s) is close to ~i for i = 1..... n - 1.
From Equations (9), the following approximate equations may be ob-
tained, which are valid in this forward region:

y~m+ a)(S) = --jR y~m)(r) {W~"}~ + z~m)(r) {Wr/} ~ p ( s - r) dr,

(11)
z(m+l)(s)- f
-JRu
v"Y~(~
)(r)- "s-r") d r .
(vr/) q(
DISCRETE TIME SPATIAL. MODELS 113

Note that Equations (11) are just a special case of Equation (1) with
n replaced by 2n, n, = 2n -2, and x$z? I(s) = y:“)(s) and x$;)(s) =
zim)(s). Also pi 2n_l(r) = p(r), pi 2n(rl = q(r) for i = (2~2- 1),2n, and
YZ”_l zn-1 ={wkI,/(2rlrWs), Yz:-l,zn =MI~J(29’WS), Y2n,2n-1=

u, /<v’q> and Y~“,~,, = 0. Hence we obtain the result that the speed of
first spread in a specific direction of allele A, among both males and
females is c, where

log{y,,-,,,,-,[P(A)/21}+log[l+\/1+48Q<A)/P(A)]
A ,

where P = y2n_1,2n~2n,2n_1 /Y&_~,~~_~

and f’(A) and Q(A)are the
two-sided Laplace transforms of the projection of the migration distri-
butions p(s) and q(s) in the specified direction.
This result corresponds to the result obtained by Lui [6, 71 for the
case n = 2. Lui again considered four cases, corresponding to females
having the heterozygote intermediate, superior and inferior; the het-
erozygote intermediate case is split into the cases when the homozygote
for allele 2 is more fit (or less fit> than the homozygote for allele 1. The
speed of propagation was obtained for all values of the parameters in
the superior case and with the restriction in the intermediate case with
w22 > wll that (w12 - wI1) 2 wIl(w22 - w12)/2w12.
Note that we could consider the spread of several new alleles. This
can be treated in a similar fashion to Section 3.1.

3.3. A MULTIPLE ALLELE MODEL WITH MIGRATION DEPENDENT

UPON GENOTWE

Consider dynamics similar to that of Section 3.1 but where the

density function for the migration depends upon the genotype. We now
consider the proportion of genotype A,A, at position s in generation m
prior to mating, which is denoted by 2x$“)(s) for i # j and x~~~(s). Let
{X(“)(S))~~= {Xcm)<,>)li= x:7)(s). Note that the matrix X@“)(s)is symmet-
ric and that, at posrtron s in generation m, the proportion of allele A, is
given by {X(m)(s)l}i. The density function for migration of genotype
AiAj by position vector r is pii( The equations describing the process
are

1 Wij{Xc”‘(r)l},{X’“‘(r)l}j
xi? + l)(‘) = p)(s) / RN lfX(m)(r)~(m)(r)l pij(s-r)dr, (12)
114 J. RADCLIFFE AND L. RASS
where

Wij{x(m)(r) 1 } i { X ' m ) ( r ) 1}y

k(ra)(s) = ~.t Z fR N l'X(m)(r)$rcx(m)(r)l Pij(s-r)dr"
J

Consider a stable population with alleles A 1..... An - 1" Let 2,/i j > 0,
i ~ j , and 7/ij > 0, i = j be the proportions of genotype AiA j at all
positions s ~ R N. A new allele occurs by mutation in (or immigration
into) a bounded region B, so the zeroth generation is taken to have
x!°)(s) = 7//j for i < n and j < n and s ~ B. Also x~°)(s) = 0 for i = n or
j = n and s ~ B. Let 7//be the proportion of allele A i and let {7/*}i = r/i
for i < n. Then r/* = W * - 11/1'W*- 11 > 0. Hence

~liwijwj {W*- 111'W* -1}ij{W*}i j

~lij = ~7*'W*r/* = l'W*-ll

where W* is the part of the fitness matrix relating to the first (n - 1)

alleles.
Define ~Tn= 0, and let r/ be the vector with {~7}i= r/i. Note that
n--1
r/i = Ej=I~ij, Then, in the forward front, Equation (12) for i < n may be
approximated by
n--1
x!7+1)(s) = E q~ilf_l,,qit(s-r)x~m)(r) dr, (13)
l=1 "1¢

where ~bit = winrli/rfWrl and qit(r) = Pin(r). Here both q~it and qil(r)
depend only on i and not on I. Equation (13) is identical with Equation
(1) if we take xi+(m)l(S)= x!m)(s) for i = 1,... ,(n -- 1), replace Yu by ~bil and
Pil by qil, and let n 1 = 1.
In this case, A(A), as defined in Section 2, has identical columns. The
maximum eigenvalue is therefore equal to the sum of the entries of the
common column vector. Hence,
n-1
Win~i
p[A(A)] = ~] Pin(A) w--TW"~ ='yP(A),
i=I

where 3, = { W ~ } n / w ' W v / and P(A) = Y'.~-~aiPin(A), where a i =

WinTli/{WT}}n. Note that P(A) is the Laplace transform of the projection
in the specified direction of a density p(r), which is a weighted function
of migration densities; that is, p(r) - Zi=
n- i1aiPin(r). Hence the speed of
- -

first spread can be given in the simple form

c = max{O, inf l°g[ P(A)] + l ° g ( ' ) )

x>0 A "
DISCRETE TIME SPATIAL MODELS 115
Note that, if we let x(~m)(s)be the proportion of allele A n (at position
s and in generation m), then by summing over i in Equation (13), we
obtain

n
x(m + l)(s) = fRNTX~m)(r)p(s _ r) dr.

Hence the speed of spread of allele A n can be derived from this single
equation and is the common speed of spread of the genotypes, c, given
above.
3.4. A MULTIPLE ALLELE MODEL WITH SEX-LINKED LOCUS AND
MIGRATION DEPENDENT UPON GENOTYPE
The dynamics of this model are similar to Section 3.3, except that the
male and female genetic outputs are separate and may have different
migration distributions. As in Section 3.2, the male is haploid for the X
chromosome and the female is diploid. Consider the proportion of the
different genotypes at position s in generation m for each sex prior to
mating. For the females, let 2y!7)(s) be the proportion of genotype
A i A j when i ~ j and let y~m)(s) be the proportion of genotype A i A i.
Also let z~m)(s) be the proportion of genotype A i for the males. The
migration density and fitness for females of genotype A i A / are pij(r)
and wij. The corresponding migration density and fitness for males of
genotype A i are qi(r) and v i. Let {Y(m)(s)}ij = y~)(s). Also let {z(m)($)}i
= z!m)(s). The equations describing the process are

y~7+l)(s) = 1 wij[{y(m)(r)l}i{z(m)(r)}j +{Y(m)(r)l}y{z(m)(r)}i]

k(mS(s) f: 2 [ze~)(r) ]'W[y(m)(r) ] 1
×piy(s - r) dr, (14)

z~m+O(S) : 1 ," vi{y(m)(r)l}i

b(m)(s) JRN [v'y(m)(r)l] q i ( s - r ) dr,

for j < i and i = 1..... n, where

k(m)(s) = ~. E fR Nwiy[{Y(m)(r)l}itz(m)(r)}j +{Y(m)(r)l}jtz(m)(r)}i]

• j 2[z(m)(r)]'W[y(m)(r)]l

× p i j ( s - r) dr,
f vi{Y(m)(r)l}i
b(m)(s) = t~. JnN[v'y(m)(r)l] q i ( s - r ) dr"
116 J. R A D C L I F F E A N D L. RASS

Consider a stable population with alleles A 1..... An-1 with 2r/q > 0,
i :~ j, and ~/i:> 0, i = j being the proportion of females of genotype
A~Ai at all positions s ~ R N. The proportion of males of genotype A~ at
all positions s in R N is ~i. A new allele occurs by mutation in (or
immigration into) a bounded region B, so the zeroth generation is taken
to have y!°)(s) = ~Tij and z~°)(s) = ~i for i < n and j < n and s ~ B. Also
yi<°)(s) = 0 for i = n or j = n or both, and z~°)(s) = 0 when s ~ B. Let */i
be the proportion of allele A i and let {7/*}i = ~7i and {~*}i = ~i for i < n.
Then

~, diag(v*)n*
----- V,t~ , ,

71" = [w* aiag(v*) + a / a g ( v * ) W * ] - l l

r [w* diag(v* ) + diag(v* )W* 1 - 11
Hence, for i < n and j < n,

w~(v? + vT){[w* diag(v*) + diag(v*)W* ]-11},

{[W* diag(v* ) + d/ag(v* )W* ] - ll}j
~ij =
1' [W* diag(v* ) + a/ag(v*)W* ] -11

where W* and v* are the parts of the fitness matrices W and v relating
to the first (n - 1 ) alleles.
Define r/n = 0 and ~'n = 0, and let 7/ and ~" be the vectors with
{r/}i = 7//and { ~'}i -- ~'i. Then in the forward front, Equations (14) may be
approximated by
1
Y/~+l)(s) = 2~"Wr/

× fRNwin[{7/1i{zfm)(r)
}n+ {Yf~'(r)1}n{~"}i]Pin(s - r)dr,

z~m+l)(s) : ~ fnvn{yCm)(r)l}nqn(s--r)dr.

As in Section 3.3, a simple way of obtaining the speed of first spread

of allele A n is to rewrite these equations in terms of y~m)(s)=
y'n= lY/(nm)(S), t o give

y(m + I)(S)= ')/11f/(m)(r)P~'I(S-r) dr + V~ fRz<:)(r)

P~2(s-r)dr,
R

g(m+ I,(S) = Y21 fR/(m'(r)P~I(S--r) dr,

DISCRETE TIME SPATIAL MODELS 117

where

P~l(r) = ~ Win~i
~/11 = 2 ~ , , W r I , i= l (--~n'Pin(F)' Y12 = 2br,W~,

win(n}i , . v.
P~2(r) = ~-~'Pi,tr), Y2, = v'-~'
i=1

and

p~l(r) = qn(r).

Hence

c=max(O, inf
A>O
log[p(A)]}
A '

where

= g1 [ yl,P~,l(A) + V/y21P~'12(A) +4Y21 ylZP~'z(A) P~'I(A) ],

with Pi~(A) the Laplace transform of the projection of p~.(r) in the

specified direction.

4. EVOLUTIONARY GAME THEORY MODELS

In this section, we consider two models in evolutionary game theory.
For a general discussion of evolutionary game theory, see Refs. 21-23.
Spatial models have been considered in Refs. 24 and 25, with diffusion
terms being used to model the spatial aspect. In contrast, we use
contact distributions to describe the spatial spread.
The first model consists of a single population in which individuals
can play n possible strategies. This turns out to be analogous to the
model in Section 3.1 except that the matrix W is now a payoff matrix
that is not in general symmetric. The second model is that of a bimatrix
game. The results of the saddle point method are applied to give the
speed of spread of new strategies in populations that were in equilib-
rium.
The connection between the first model and the genetic model of
Section 3.1 is briefly discussed, as is the extent to which the exact
analytic methods of Weinberger [1, 2] and Lui [3] yield results for the
two-strategy model in game theory.
118 J. RADCLIFFE AND L. RASS
4.1. A SINGLE-POPULATION EVOLUTIONARY GAME THEORY MODEL
A setup analogous to the model of Section 3.1 is considered. Each
individual in the population plays one of n strategies, with x!m)(s) now
denoting the proportion in the mth generation at position s that plays
strategy i. The survival to maturity now depends on a payoff matrix A,
which "is not in general symmetric. The model is described by the
equations

x}m)(r)[{Ax(m)(r)}i + k]
x}m+l)(s)=JRN[X(m)(r)],[Ax(m)(r)+k]P(s-r)dr, (15)

for i = 1,...,n.
Here k is a large positive constant representing the common back-
ground fitness. This is a spatial version of the model described on page
133 of Ref. 23. Let Wij ~- {W}ij = {A}ij + k. Provided k > m a x ( - {A}ij)
then wij > 0 for all i and j. When W is substituted into Equations (15),
they become identical with Equations (4). Note that, for the model
described in Section 3.1, the fitness matrix W is symmetric. For the
present model, however, the matrix W is not in general symmetric.
Consider the spread of a new strategy S n. This is introduced into a
population in stable equilibrium, with strategies S 1.... , Snl being played
by proportions ~ ..... */nl of the population, respectively, at all points s
of R N. The expression for 7// is identical with that obtained for the
corresponding genetic model of Section 3.1. The speed of first spread c
of new strategy Sn is given by Equation (7). Results for several new
strategies are identical to those obtained for genes in Section 3.1.
Exact methods to obtain the speed of propagation for the genetic
model of Section 3.1 have been used for the case n = 2 and are
discussed briefly in that section. Equivalent results using exact methods
may also be obtained for the game theory model in the general
nonsymmetric case.
Because n = 2, we need consider only the second equation of Equa-
tions (4) with W nonsymmetrie, because x[m)(s) = 1 -- x~2m)(s). This may
be written in the form

(16)

where

w ~ x 2 + w21x(1 - x )
g(x) --
w22x 2 + ( w l 2 + w l)x(1- x) + w 1(1- x)
DISCRETE TIME SPATIAL MODELS 119
Consider the solutions to g(x)= x for 0 ~< x ~< 1. In all cases, x = 0
and x = 1 are solutions. There is an additional solution x*, where

W21 -- Wll )
x* = [(w2, - w H ) + ( w 1 2 - w 2)]

Define .~ = x* if 0 < x* < 1; otherwise define $ = 1.

For the genetics model with n = 2, x~2m)(s)satisfies Equation (16) and
g(x) is defined as above but has w12 = w21. Weinberger [1, 2] and Lui [3]
proved results for the asymptotic speed of propagation for the genetic
model in terms of a general function g(x) that satisfied certain condi-
tions. A sufficient set of conditions is as follows:
(1) g(x) has continuous bounded derivatives for x ~ [0,1].
(2) g(0) = 0, g(1) = 1, and g($) = $.
(3) g(x) > x for x ~ (0, $) and g(x) < x for x ~ ($, 1).
(4) g ( x ) / x is a nonincreasing function of x for x ~ (0, $).
(5) 0 < g'(x) ~ g'(O) for x ~ [0, :~].
(6) There exists a D > 0 such that g'(O)(x - Dx 2) < g(x) <. g'(O)x for
x ~ [0,11.
For the genetic model, four cases were identified and are given in
Section 3.1. The conditions on g(x) were satisfied for all values of
wll,w22 and the common value of w12 and w21 in Case 2. In Case 1, a
certain condition on the parameters [given by Equation (8)] was re-
quired for the conditions on g(x) to be satisfied. In these cases, the
asymptotic speed of propagation was obtained and corresponded to the
speed c given by Equation (7).
For the game theory model, there also are four cases. The trivial case
when w22 =w12 and w11=w21 is excluded, because, in this case,
fx[m)(s)ds =- fx<2°)(s)ds for all m. Case 1 has w22 ~ W12 and w21 >I Wll
with at least one inequality strict. Case 2 has w12 > w22 and w21 > W~l.
Case 3 has w22 > w12 and W~l > w21. Case 4 has w H >i w21 and w12 >/w22,
with at least one inequality strict.
For Case 2 of the game theory model with general nonsymmetric W,
the conditions on g(x) can be shown to be satisfied for all values of the
w,7. F o r Case 1, the conditions are satisfied provided W21(W12 "t- W21)
W1~(W22 + W2~). Note that this reduces to the condition given by Equa-
tion (8) for the genetic model in the symmetric case when w12 = w2~.
Hence the asymptotic speed of propagation can be obtained by exact
methods for these cases. Note that Strategy 2 is introduced into a
bounded region of R N, so x~2°)(s)= 0 for s ~ B for some bounded region
120 J. RADCLIFFE AND L. RASS
B. In Case 4, as in the genetic model, Strategy 2 will die out uniformly
in R N. In Case 3, Strategy 2 will also die out and not spread in any
direction when, in addition, x~°)(s) < ~ for all s ~ B.
When n = 2, the speed of first spread obtained by the saddle point
method for the evolutionary game theory model agrees with the results
for the speed of propagation obtained by exact methods.
4.2. A BIMATRIX GAME
Consider two different populations that are in conflict. Individuals in
Population 1 can play strategies S~ ..... St, and individuals in Population
2 can play strategies T1,...,T ~. Let x!m)(s) and y~m)(s) denote the
proportions in Populations 1 and 2 in the mth generation at position s
that play strategies i and j, respectively. The survival to maturity in
Populations 1 and 2 now depends on two payoff matrices, A and B,
which are not in general symmetric. The model is described by the
equations

. x~m)(r)[{Ay(m)(I)}i + k~]
x~m+1'(S) = JR N [x'm'(r)]ttAy(m'(r) Jr kl] P l ( s - r ) dr'
. y(m)(r)[{Bx(m)(r)}j Jr k2]
y(m+ 1)(S) = JRs [Y(m)(r)]' [Bx(m)(r) + k2] p 2 ( s _ r ) dr ' (17)

for i = 1 ..... t and j = 1,...,n. Here k 1 and k 2 are large positive

constants representing the common background fitness of individuals in
Populations 1 and 2, respectively. Let {V}ij= {A}i/+ kl and {W}i. = {B}ij
+ k 2. Then provided k 1 > max(-{A}i) and k 2 > max(-{B}ij), (V}ij > 0
and {W}ij > 0 for all i and j. Equations (17) can be written as

r x~m)(r){Vy(m)(r)}i
x!m+I)(S)=JRN ~ Pl(s-r)dr'

r y(m)(r){WX(m)(r)}j
y(m+l)(s ) = JRN ~ P2(s_r)dr ' (18)

f o r i = l ..... t a n d j - - 1 ..... n.
Initially, Population 1 plays strategies S 1..... S5 and Population 2
plays strategies T1,...,Tnc The populations are m equilibrium, with
proportions {0"}i --0* playing strategy S i (i--1 ..... t 1) and {~/*}j--~7
playing strategy T/ ( j = 1,...,n 1) at all points s,.where~ 1'0"= 1 and
1'7" =1. Then 0* and ~j* satisfy 0* =gi(O*,~*) and ~j* =hj(O*,~*),
where gi(O*, 71") = (O*{V*~*}i/O*'V*71*) and hi(O*, 71") =
DISCRETE TIME SPATIAL MODELS 121
(7/p{W*0*}j/r/*'W*0*). Here iV*}0. = {A*}ij + k 1 and {W*}0 ={B*}0. +
k2, where A* and B* are the payoff matrices for Populations 1 and 2
corresponding to the first t I and n I strategies, respectively. Thus
{0"} i = {V* - 11}i//(l'V* -11) and {,/* }j = {W*- 11}/./(rW* -11).
Let t = t 1 + 1 and n = n I + 1. Suppose that Population 1 introduces a
new strategy S t, but Population 2 continues to play the existing strate-
gies. Define {0} i = 0* for i = 1.... t 1 and {0}t = 0. Also let {,/}i = r/* for
i = 1 .... n 1 and {7/}~= 0 .
The approximate equation obtained from Equations (18) in the
forward front is

x~m+ 1)($) = ~fttfRNPl(r)x~m)(s _ r) dr, (19)

where 3~, = {V~/}t / ( 0 ' V~/).

The speed of first spread of strategy St in a specific direction is

c l = m a x ( O , inf log[Pl(A)]+log('Y"))
~>0 A '

where PI(A) is the Laplace transform of the projection of the migration

distribution pl(s) in the specified direction.
Similarly, suppose that Population 2 introduces a new strategy T~,
while Population 1 continues playing strategies S 1..... St1. The approxi-
mate equation obtained from Equations (18) in the forward front is

y~m +1)(s ) = ~n,fRnp2(r)Y~nm)(s _ r) dr, (20)

where ~n~ = {W0}n/(17'W0). The speed of first spread of strategy Tn in a

specific direction is

where P2(A) is the Laplace transform of the projection of the migration

distribution p2(s) in the specified direction.
Now consider the case when Populations 1 and 2 each introduce a
new strategy, S t and T, respectively. As in Section 3.1, Equation (19)
will be valid only if c I >/c2, in which case the speed of first spread of
strategy S t is c 1. Equation (20) is similarly valid only if c 2 >/Cl, in which
case the speed of first spread of strategy Tn is c 2. The value of
122 J. RADCLIFFE AND L. RASS
c = max(c1, c 2) gives the speed at which a disturbance to the equilibrium
is first felt, this disturbance being caused by the relatively more advan-
tageous strategy.
The introduction of several new strategies may be treated in a similar
fashion.

5. B R A N C H I N G PROCESS MODELS
The saddle point method can be applied to a wide range of discrete
time branching process models. Two processes are given in this section
as illustrations. The first is a multitype Galton-Watson process and the
second is a demographic branching process. A discussion of multitype
and demographic branching processes is given in Ref. 26.
In branching process models, a system of nonlinear integral equa-
tions is obtained for the probability that the farthest spread in a specific
direction at discrete time m is at least s, given that, at time zero, there
is one individual in the population who is at the origin and is of ith type
(or age or a combination of both) for all possible types i. This is
denoted by x!m)(s). Here x!°)(s) = 0 for s t> 0, and x~°)(s) = 1 for s < 0.
Because x~°)(s) is not zero outside a bounded region of R, the
branching models do not satisfy the conditions imposed in Section 2 to
prove Theorems 1 and 2. Note however that the actual condition used
in the proofs of these theorems requires that D°)(0) has entries that
exist for Re(O) > 0 and are uniformly bounded for 01 ~<Re(O) ~< 02 for
any 0 < 01 < 02 < 0o.
For the case N - - 1 , to ensure that these conditions hold when
establishing the speed of spread in the positive direction, we need only
require that each x~°)(s) is bounded and that x!°)(s)= 0 for s >t A for
each i and some finite A.
This latter condition is met for the branching process models; hence
the results of Section 2 may still be applied.

5.1. A MULTITYPE GALTON-WATSON PROCESS

There are n types. Each individual lives for one discrete time
interval, giving birth at that time to offspring, with all individuals
contributing offspring to the process independently. The probability
that there are tl .... ,t n offspring of types 1 , . . . , n of a type i parent is
gi(t I ..... tn). The density function for the distance r in a specified
direction of all the offspring from the parent of type i is pi(r). Let Um
be the position of the individual farthest from 0 in the ruth generation
in the given direction, and let x}m)(s)= P [Urn > s given that there is
one type i individual at position 0 in the 0th generation].
By conditioning on what occurs at the end of the first discrete time
DISCRETE TIME SPATIAL MODELS 123
interval, we obtain the following system of equations describing this
model:

x.(m+
i 1)(S) . .fRt~
. .~ ~.~gi(tl,...,tn) 1- [1 -- x~m)(r)] ' Pi(S r)dr,
tn
for i = 1,..., n. Note that x~°)(s) = 1 if s < 0 and x~°) = 0 if s > 0.
In the forward tail of the distribution of farthest spread, these
equations may be approximated by:

x}m+l)(s)= fg~t 1 "'" ~.~gi(tl ..... tn)[ ~ t j x } m ) ( r ) ] p i ( s - r ) dr

tn [j=l 1
n
= E l z i j f x } m ) ( r ) p i ( s - - r ) dr,
j~l R

where /zq is the expected number of offspring of type j from a type i

parent.
Let {A(A)}q =/zijP/(A), where P/(A) is the Laplace transform of pi(r).
Then the speed of spread of the forward tail, c, is given by Theorem 3
with this matrix A(A). A simplification occurs when Pi(A) is the same for
all i with common value P(A). In this case, c is given by Equation (7)
with ~, = p(/.0, where { ~}ij = [tl'ij"
5.2. A DEMOGRAPHIC BRANCHING PROCESS MODEL
Consider an age-dependent branching process model, with xi(s) now
measuring the probability that the farthest spread at time m in a
specified direction is at least s, given that there was one individual
initially of age i at position 0. An individual of age i has a probability/3~
of dying before reaching age i + 1 with /3k = 1, so an individual cannot
live for k + 1 years or more. If the individual survives to age i + 1, then
the probability of t offspring is gi(t) and, conditional on t, the density
function of the positions r 1..... rt of these offspring in the specified
direction relative to the parent is pi(r 1..... rt) , with common marginal
density function pi(r). Again, using a conditioning argument, we obtain
the following equations describing the process:
x ! m + l ) ( S ) = ( 1 -- ~ i ) ~-,gi(t)
t

x 1 - [ 1 - (m)

× fR,E [1--x(°m)(s- r j ) l P i ( r l ' " " r t ) d r l ' " " d r t '

for i = 1,..., k. Note that x~"~)l(S) -= O.
124 J. RADCLIFFE AND L. RASS
In the forward tail of the distribution of farthest spread, these
equations may be approximated by

X}m+ 1)($) = ( 1- /3i)[X~+~(S) q" I,Lifx(om)(r)pi(s--r)dr ],

for i = 1..... k, where /z/ is the expected number of offspring of an

individual of age (i + 1). Let P/(A) be the Laplace transform of pi(r).
Define {A( A)}il = (1 -/3i_ 1)/zi- 1P/- I(A) for i = 1..... k + 1, {A( A)}i,i+ 1 =
(1-/3/-1) for i = 1 .... ,k, and {A(A)}ij= 0 for j ~ : l or (i +1). Then the
speed of spread of the forward tail, c, is given by Theorem 3 with this
matrix A(A).

6. CONCLUSION
The saddle point method is a powerful tool. It can be applied to a
wide range of biological models in which time and space may be
discrete or continuous. It has already been used for the analysis of
continuous time models of epidemics. In this paper, we have derived the
saddle point approximation for discrete time models. For our results to
be applicable to such discrete time models, the approximate equations
in the forward front must be linear, as in Equation (1), with nonnegative
coefficients y/j. Some care is needed when these equations in the
forward front yield a reducible system, when the original system of
equations is nonreducible. In such cases, the equations are valid only in
the forward front for a new type, where no connected new type can
spread at a faster speed.
Many contact models that fit into the format described can be
considered. We have shown the wide range of applicability of the
method by applying it to several contact models in genetics, evolution-
ary games, and branching processes. Many variations and more-complex
models can be written for which the method is still applicable and for
which the speed of first spread can be obtained by the saddle point
method described in this paper.

APPENDIX
Proof of Theorem 1. Take M 0 such that 10(m)- 001 < ( 0 0 / 2 ) for
m > M 0. Then, for m > M0, write

1 [O(m)+i~O(m)
--e- As(m)-g[O(m)]
O(m)e-g[°(m)]~ = ~ "O(m)- i~ )t

{[A( A)]mL(°)( A)}idA = I 1 d- I 2 -I- 13 d- I4,

DISCRETE TIME SPATIAL MODELS 125
where

1
II=2-~fl>aJly O(.m) e_iys(m) 1
O(m)+iy {p[O(m)] }m
× {{A[ 0 ( m ) +/y] }'nL(°)[ O(m) + iY]}idY,
O(m) 1
12 = 2-~ fs< [yl<.aO(m)+ iy e-iys(m) tp [ O(m)]} m

X {{A[ O(m) +/y] }'L(°)[ Ofm) + iYl},dy,

1 fl. O(m) 1
I3 = )'w~ ,..l<80(m)+iye-'Y'('){P[O(m)]}"

× {{A[ O(m) + iy]m _ e m logtp[o(m)+iy]}

×E[ O(m) + iy] L(°)[ 0 ( m ) +/y] }}i dy,

1 fly O(m) egtO(m)+iy]_gtO(m)l
I4 = )-~ i<~ O(m) +/y

× {El O(m) + iy] L(°)[ O(m) + iy] }i dy.

Now

dO(m)e_g[O(m)l~_~__~f8_80(m)+iy
O(m) egte(m)+~y]_gto(m)]
× {E[ O(m) +/y] L(°)[ O(m) + iy] }i dr

< II11+ 1121+ 1131.

Hence the proof consists of showing that we can choose cz so that

1/11, 1121, and 1131 can each be made less than era/3 for m sufficiently
large and some 0 < E < 1.
First, consider 11. Take 1 < k ~< 3 / 2 with kO < A, and let m/> M 0 so
that O(m) ~ [0o/2, kOo].The conditions on the x)°)(s) imply that I{Z(°)(x
-12/j.')}/t:~h for all j and some h and for x~[Oo/2,kOo]. In addition,
the conditions on the contact distributions state that there exist func-
tions kiy(y) that are integrable with IAij(x + iy)l ~ %jkii(y) for all i,j,
and x ~[Oo/2,kOo]. From the integrability, for any d > 0 there exists
an a* such that yi/kij(y).<< d for lyl I> a * . Take. p =
minx~tOo/2,kOolp(x). The continuity and positivity of p(x) imply that
p>O.
126 J. RADCLIFFE ANDL. RASS
Then for o~ >/a*,

illl < 2_~ flyI p-m E y#kiy(Y){(dll , ) m - i

>a j
IL'@
h nl,]m 1
= E ~ ' ~ YiYflyI> ? kit(Y) dy.
j

Take d = p/2(n - n 1) and find the corresponding a*. Then

h
1Ill < E -~p "Yiy2-mflyl> akU(Y) dy
J

for c~ >~ c~*.

Now choose a > a* such that

flyI> ,~kij(y) dy < 3hY.Ty

1rp #

for all i,j.

Hence IIl1<~(~) 1 1 m for this ot and for m>~M o.
Next consider 13. Because the eigenvalues of A(00) are distinct, from
Dieudonn6 [27], page 248, there exists a neighborhood IA - 00l< 80 for
which A ( A ) h a s distinct eigenvalues /.~I(A).... ,/~n_nl(A) with p(A)--
/Zl(A), where Re[ p(A)] > I/z,(A)l for s t> 2. Also/~s(A), and the entries of
the corresponding idempotent E,(A), are continuous functions of A for
all s and IA - 00l< 80.
Let E(A) = El(A). In the neighborhood of 0, define
m

K(;O=(1}( X)]

From the continuity of the idempotents and the entries of L<°)(A), there
exists a D such that t{Es(A)Lt°)(A)}il < D for IA - 001 < 8o and s >/2. In
addition, I ~s(A)l < Re[ p(A)] < p[Re(A)] and hence I/z~(A)/o[Re(A)]I < 1
for s >i 2. Because max, ~ 21/zs(A)/p[Re(A)]l ~ max~l/~s(0o)l P(0o) < 1
as A ~ 00, there exists a positive 81< 80 and y < l such that
I/z~(A)/p[Re(A)]I ~<y for s >/2 and IA - 00l< 61. Hence IK(A)I ~<ym(n
- n 1 - 1 ) D for any m ~ M o and I A - 01< B1.
DISCRETE TIME SPATIAL MODELS 127

Choose MI>_.M o such that IO(m)-Ool<<.(81/vr2). Then 1131~<

(1/2~)(n-nt _ - l ) D S y m for 8 2 < 8 2 - ½ 8 2 = ½ 8 ? . T a k e 8 =
m i n { 8 * , ( 8 1 / ¢ 2 ) , [ 2 ~ r / ( 3 ( n - n 1-1)D)]}. Then 1131 < ~1y m for this 8
and for m > / M 1.
Finally, for the a and 8 previously obtained, consider the integral 12.
Define A*j(x) = sup8 ,~ lyl,~ ~ IAo(x + iy)l. W h e n Yo = 0, then Ao(x + iy)
-- 0 and hence A*j(x) = 0 • If Yij ~: 0, then A*.(x) j < Ao(x) • In this case,
lim x_~ oo[A*j(x)/Ao(x)] = [A*j(Oo)/Aij(Oo)] < 1. H e n c e there exists a
positive /3 < 1 and an M 2 t> M 1 such that {A'~j[O(m)]/Ao[O(m)]} </3
for m >/M2 and all i,j such that 3'o ~ 0. H e n c e A*[0(m)]</3A[O(m)]
for all m i> M 2.
Now, because M 2 > / M o, we have ~{L(°)[0(m)+/Y]}jl < h for all j and
all m > / M o. Therefore

2ha r
lI2l ~< --~- p [ 0( m)]- m{(A* [ 0(re)l) ml}i

h a p [ O(m)]-m/3m{(A[ O(m)])ml}i

for m >~M 2.
But A[0(m)] has distinct eigenvalues for m 1> M 1 and hence for
m >i M 2, and the idempotent Es[0(m)] tends to E~(0) as m --,oo for all s,
so the entries are bounded. Therefore

n~.~n1 ~l~s[O(m)] m
P[O(m)]-m{(A[O(m)])ml}i<i=1 p[O(m)] {IEs[0(m)]l }i
n-- n 1

< {IEs[0(m)]ll}i
S=I
<F
for some constant F.
H e n c e 1121 ~< (hotF/'n')fl m for m I> M 2. Take ~b =½(1 + /3).
1 m
Then JI21 < ~ b p r o v i d e d m t> M3, w h e r e M 3 = m a x [ M 2 ,
log(3h aF/'tr)/log( ~b/ / 3 )].
The theorem then follows by taking • = max(½, y, ~b) and M = M3.

REFERENCES
1 H.F. Weinberger, Asymptotic behavior of a model in population genetics. In
NonlinearPartialDifferentialEquations and Applications. J. Chadam, Ed., Lecture
Notes in Mathematics, Vol. 648, Springer, New York, 1978, pp. 47-98.
2 H.F. Weinberger, Long-time behavior of a class of biological models. S/AM Z
Math. Anal. 13:353-396 (1982).
128 J. RADCLIFFE AND L. RASS
3 R. Lui, A nonlinear integral operator arising from a model in population
genetics, I: monotone initial data. SIAM J. Math. Anal. 13(6):913-937 (1982).
4 R. Lui, A nonlinear integral operator arising from a model in population
genetics, II: initial data with compact support. S/AMJ. Math. Anal. 13(6):938-953
(1982).
5 R. Lui, Existence and stability of travelling wave solutions of a nonlinear
integral operator. J. Math. Biol. 16:199-220 (1983).
6 R. Lui, Biological growth and spread modeled by systems of recursions, I:
mathematical theory. Math. Biosci. 93:269-295 (1989).
7 R. Lui, Biological growth and spread modeled by systems of recursions, II:
biological theory. Math. Biosci. 93:297-312 (1989).
8 H.E. Daniels, The deterministic spread of a simple epidemic. In Perspectives in
Probability and Statistics: Papers in Honour of M. S. Bartlett, J. Gani, Ed.,
distributed for Applied Probability Trust by Academic Press, London, 1975 pp.
689-701.
9 H.E. Daniels, The advancing wave in a spatial birth process. J. Appl. Probability
14:689-701 (1977).
10 J. Radcliffe and L. Rass, Saddle-point approximations in n-type epidemics and
contact birth processes. Rocky Mt. J. Math. 14(3):599-617 (1984).
11 J. Radcliffe and L. Rass, Reducible epidemics: choosing your saddle. Rocky Mt.
J. Math. 23(2):725-752 (1993).
12 D . G . Aronson, The asymptotic speed of propagation of a simple epidemic. In
Nonlinear Diffusion, W. E. Fitzgibbon and H. F. Walker, Eds., Research Notes in
Mathematics 14, Pitman, London, 1977, pp. 1-23.
13 O. Diekmann, Run for your life: a note on the asymptotic speed of propagation
of an epidemic. J. Differential Equations 33(1):58-73 (1979).
14 J. Radcliffe and L. Rass, The asymptotic speed of propagation of the determin-
istic nonreducible n-type epidemic. J. Math. Biol. 23:341-359 (1986).
15 J. Radcliffe and L. Rass, The asymptotic behavior of a reducible system of
nonlinear integral equations. Rocky Mt. J. Math. 26(2):731-752 (1996).
16 J. Radcliffe and L. Rass, Spatial branching and epidemic processes. In Mathe-
matical Population Dynamics: Analysis of Heterogeneity, Vol. 1, O. Arino, D. E.
Axelrod, M. Langlais, and M. Kimmel, Eds., Wuerz, Winnipeg, 1995, pp.
147-170.
17 J. Radcliffe and L. Rass, Multitype contact branching processes. In Proceedings
of First World Conference on Branching Processes, Springer Lecture Notes in
Statistics No. 99, 1995, pp. 169-179.
18 W.J. Ewens, MathematicalPopulation Genetics. Springer-Verlag, Berlin, Heidel-
berg, New York, 1979.
19 P. Creegan and R. Lui, Some remarks about the wave speed and traveling wave
solutions of a nonlinear integral operator. J. Math. Biol. 20:59-68 (1984).
20 O. Diekmann and H. G. Kaper, On the bounded solutions of a nonlinear
convolution equation. Nonlinear Anal. Theory Appl. 2(6):721-737 (1978).
21 J. Maynard Smith, Evolution and the Theory of Games. Cambridge University
Press, 1982.
22 E. C. Zeeman, Dynamics of the evolution of animal conflicts. J. Theor. Biol.
89:249-270 (1981).
23 J. Hofbaeur and K. Sigmund, The Theory of Evolution and Dynamical Systems.
London Mathematical Society Student Texts 7, Cambridge University Press,
1988.
DISCRETE TIME SPATIAL MODELS 129
24 V . C . L . Hutson and G. T. Vickers, Traveling waves and dominance of ESS's. J.
Math. Biol. 30:457-471 (1992).
25 G.T. Vickers, Spatial patterns and ESS's. J. Theor. Biol. 140:129-135 (1989).
26 C . J . Mode, Multitype Branching Processes Theory and Applications. American
Elsevier, New York, 1971.
27 J. Dieudonn6, Foundations of Modem Analysis. Chelsea, New York, 1969.