Noname manuscript No.

(will be inserted by the editor)

Global stability of discrete-time competitive

population models

Stephen Baigent · Zhanyuan Hou

Received: date / Accepted: date

Abstract We develop practical tests for the global stability of interior fixed
points for discrete-time competitive population models. Our method consti-
tutes the extension to maps of the Split Lyapunov method developed for di↵er-
ential equations. We give ecologically-motivated sufficient conditions for global
stability of an interior fixed point of a competitive map of Kolmogorov form,
and show how these conditions can simplify when a carrying simplex is known
to exist. We introduce the concept of a principal reproductive mode, which
is linked to a left eigenvector for the dominant eigenvalue of a positive ma-
trix, which in turn is linked to a normal vector, at an interior fixed point,
to a hypersurface of vanishing weighted-average growth, and also to a normal
to the carrying simplex when present. A connection with permanence is also
discussed. As examples of applications, we first take two well-understood pla-
nar models, namely the Leslie-Gower and the May-Oster models, where our
method confirms global stability results that have previously been established,
through, for example, properties of monotone maps. We also apply our meth-
ods to establish new global stability results for 3-species competitive systems
of May-Leonard type, giving detailed descriptions of the parameter ranges for
which the models have globally stable interior fixed points.

S Baigent
Department of Mathematics, UCL, Gower Street, London WC1E 6BT
Tel.: +44 207 679 4593
E-mail: [email protected]
Z. Hou
School of Computing, Faculty of Life Sciences and Computing, London Metropolitan Uni-
versity, 166-220 Holloway Road London N7 8DB. E-mail: [email protected]
2 Stephen Baigent, Zhanyuan Hou

1 Introduction

There is a large body of literature describing mathematical techniques for

identifying global stability of fixed points of di↵erential equation models of in-
teracting populations, but the corresponding literature for discrete-time mod-
els is significantly smaller. Here we describe a general test for global stability
of a coexistence state that can be successfully applied to a wide range of
discrete-time competitive population models, and we provide examples of the
application of the test to some standard models, including some whose global
stability properties are unresolved.

We restrict our attention to population models described by Kolmogorov sys-

tems, which includes the vast majority of classic population models. These
Kolmogorov systems are defined by per-capita growth rates that are at least
bounded and continuous for non-negative populations. As a consequence, if
a species is initially absent then it remains so for all time; immigration or
emmigration is not modelled.

Thus, let us assume that the model has N interacting species with population
densities xi 0 for i 2 IN := {1, 2, . . . , N } and write x = (x1 , . . . , xN )T
(a column vector). We set C = RN + for the nonnegative first orthant, where
R+ = [0, 1). We use the natural numbers N = {0, 1, . . .} for the model time
units. Let f = (f1 , . . . , fN )T : C ! RN be bounded and continuous in each N
components. Then the Kolmogorov model that we consider is

xi (t + 1) = Ti (x(t)) = xi (t)fi (x(t)), i 2 IN , t 2 N. (1)

For each initial state x 2 C, equation (1) generates a forward orbit O+ (x) =
{x(t)}t2N with x(t) denoting the population state reached from the initial
state x = x(0) after t time steps. Thus our models are restricted to those
where the growth or decline does not impact back on the environment, and
those for which the environment is constant, and not, for example, subject to
seasonal changes or random disturbances.

As with many studies of the global dynamics of population models, our ap-
proach is based upon a Lyapunov function. In fact, the Lyapunov function that
we choose has been previously as an entropic asymmetric distance function in
continuous-time population models [16, 43], and also as part of average Lya-
punov function in the study of permanence of population models [19, 20]. We
recall that a population model is said to exhibit permanence if all populations
remain bounded and whenever all populations are present, they remain present
for all future time with densities eventually above some positive number. In
some respects, our method extends the average Lyapunov function approach,
a technique which is usually applied only over a finite time interval, and ex-
tends it over all forward time to extract long term properties of a permanent
system.
Global stability of discrete-time competitive population models 3

The framework that we use is as follows. The model (1) is assumed to have
a unique interior fixed point: p 2 C˚ = (0, 1)N . Let v be a given positive
QN
vector and V (x) = i=1 xvi i ; this is the Lyapunov function we employ that
has also been widely used as an average Lyapunov function [20, 15]. The vector
v is similar to that identified as Reproductive Value in classic Leslie matrix
models (e.g. [7, 39]) in that it is a left eigenvector associated with the dominant
eigenvalue of an irreducible nonnegative matrix. Then for x(0) = x, and x(t) 2
O+ (x),
V (x(t)) = ⇤(t, x)V (x),
⇣ P ⌘
t
where ⇤(t, x) = exp vT k=0 ln(x(k)) . Notice that ⇤(t, x) is nondecreasing
along the forward orbit O+ (x) when the sum
t
X
⇥(t, x) = vT ln(x(k))
k=0

is nondecreasing. Thus a necessary and sufficient condition for V to be non-

decreasing along O+ (x) is that vT ln(x(t)) 0 for each t 2 N. Consequently,
knowledge of the omega limit set of a point x 2 C˚ under (1) can be obtained by
studying the scalar function ⇥(t, x). An important feature of the expression
⇥(t, x), is the weighting in the sum by the vector v, and an ecologically-
motivated choice of v is the key to the method’s success, as we explain
later.
The key property that our models satisfy is:

Property 1: Given any x 2 C and O+ (x) = {x(t)}t2N there exists a t0

(which may depend on x) such that vT ln(x(t)) 0 for all t t0 .

Property 1 states that an orbit eventually enters a certain region of phase space
where the Lyapunov function V then is nondecreasing. As we shall demon-
strate, such a property can often be used to determine omega limit sets of
interior orbits and even global convergence to an interior fixed point. A major
part of this paper is developing sufficient conditions for Property 1 above to
hold. Defining ↵ = D[p] 1 v, the vector ↵ will be called by us the principal re-
productive mode. The function '(x) := vT ln f (x) = ↵T D[p] ln f (x) appearing
in Property 1 will be called by us the principal component of the reproductive
rate since it is a scalar that provides a measure of the component of the per-
capita growth of the whole population in the direction of the dominant mode
↵. As we will also show, for competitive Kolmogorov systems that possess a
carrying simplex (a manifold of codimension 1 that attracts all nonzero orbits,
as described in section 7 below), the principal reproductive mode points in
the same direction of the normal to the carrying simplex at the interior fixed
point.
From a practical point of view, our results say the following principle:
4 Stephen Baigent, Zhanyuan Hou

For a bounded population model with unique coexistence state p, if the

principal component of the reproductive rate is eventually positive, the
population converges to p.

The positiveness of the principal reproductive mode ↵ is guaranteed when

the model is strongly competitive (as are all the models studied here), but it
is in principle possible to have models where ↵ has negative components to
reflect that some species may sometimes be detrimental to overall population
growth. The price to be paid for this generalisation is that permanence of
the system (1) also has to be proved before the Lyapunov function V can be
used (see [4] for the continuous-time Lotka-Volterra case). When there is a
carrying simplex that lies entirely in a region where the principal component
of the reproductive rate is positive except at the interior fixed point where
it is zero, all nonzero orbits converge to the unique interior fixed point. This
motivates that the normal to the carrying simplex at the fixed point should be
parallel to the normal to the hypersurface of zero principal component of the
reproductive rate which is in turn parallel to a weighted eigenvector ↵ by the
interior fixed point associated with the dominant eigenvalue of the inverse of
the linearisation of the model at the fixed point. A summary of the geometry
of our method is shown in figure 1.

It is worth noting here also that a system of the form x(t + 1) = F(x(t)) which
is also permanent in the first orthant could also be treated by our methods here
by defining fi (x) = Fi (x)/xi for xi > 0. This observation opens our method
up to other application areas such as mathematical genetics where many of
the well-known models contain terms that are not entirely per-capita rates,
such as for genetic recombination.

What is quite surprising to us is the sheer number of discrete-time popula-

tion models to which our method can be successfully applied to show global
attractiveness of the interior fixed point. The principle does not depend upon
a particular form for the per-capita growth functions, although some models
(e.g. those with exponential functions) inevitably give rise to principal com-
ponent of the reproductive rate that are easier to work with. The principle
cannot be applied as it stands when there are periodic orbits present in the
system, and this would include chaotic systems.

As practical demonstrations of their wide applicability to population models,

we apply our methods to establish global stability in the case of 3 species for
two well-known competitive models, namely the Leslie-Gower model and the
May-Oster models. These take the form

bxi
x0i = (i = 1, 2, 3) (Leslie-Gower)
1 + (Ax)i
x0i = xi exp(r(1 (Ax)i )) (i = 1, 2, 3) (May-Oster).
Global stability of discrete-time competitive population models 5

Principal)reproduc.ve)
rate)component)is)
posi.ve)in)shaded)
))) region)

Principal)reproduc.ve)
α
mode))))))is)normal)to)both)
α the)surface)of)zero)
p) principal)reproduc.ve)rate)
(R))and)also)the)a:rac.ng)
carrying)simplex)(S))at)the)
coexistence)state)p.)

S)
Carrying)simplex)

Fig. 1 A summary of our method. All trajectories eventually end up in the shaded region
where the principal reproductive rate component is positive, and then converge to the coex-
istence fixed point p. Both the surface of zero principal reproductive rate and the carrying
simplex (when it exists) have a outward normal at p that points along the vector ↵, the
principal reproductive mode.

where xi , x0i are the current and next generation population densities of species
i respectively,
0 1r, b > 0 constants and A is a positive 3 ⇥ 3 matrix given by
1↵
A = @ 1 ↵ A , and ↵, > 0 reflect the strength of competition.
↵ 1
For these models we establish:

Theorem 1 (Global stability for the 3-species May-Leonard Leslie-

Gower model) Suppose that b > 1, 0 < ↵ + < 2 and either
b 4↵2 + 4 2
4↵ 4 ↵ + 1 < 3 ↵2 + 2
↵ ↵ +1 ,
or
3b(1 2↵ 2 + 3↵ ) 5(1 ↵ ) + 7↵ ↵2 2
.
Then the Leslie-Gower model has a unique fixed point that is globally asymp-
totically stable in the interior of the first orthant.

Theorem 2 (Global stability for the 3-species May-Oster model)

Suppose that r 2 (0, 1), ↵ + < 2 and either
3r(1 + ↵2 + 2
↵ ↵ ) < (2 ↵ )(1 + ↵ + ),
or
r(5 ↵2 2
5(↵ + ) + 7↵ ) (2 ↵ )(1 + ↵ + ).
Then the May-Oster model has a unique fixed point that is globally asymptot-
ically stable in the interior of the first orthant.
6 Stephen Baigent, Zhanyuan Hou

Local asymptotic stability of interior fixed points in these models (indeed their
N species versions) were studied by Roeger in a series of papers [34, 33, 32]
and recently global stability of the N species periodic Leslie-Gower model
was studied [37] using contraction mapping techniques and also the planar
May-Oster in [38]. An advantage of our approach is that it yields a detailed
description of parameter values sufficient for global stability. A disadvantage
over [37] is that our method cannot directly deal with periodic orbits. However,
it might be possible to overcome this shortcoming by working with powers of
the map T.

2 Notation and previous results

2.1 Notation and standing assumptions

Let X be a metric space and T : X ! X be continuous. For a given x 2 X one

typically studies the sequence O+ (x) := {Tt (x) : t 2 N} (N = {0, 1, 2, . . .})
which is referred to as the forward orbit of T through x. A set A ⇢ X is
T invariant if T(A) = A. The set of fixed points of T⇣ is Fix(T) = {x
⌘ 2X :
T1 S1 k (W) denote
T(x) = x}. For a given set W ⇢ X let !(W) := n=0 k=n T
the omega limit set of W. When O+ (x) has compact closure in X , !(x) is a
nonempty, compact and invariant set. An invariant set A ⇢ X is said to be
an attractor if there exists an open U A for which !(U ) = A. A compact
invariant set R ⇢ X is said to be a repellor if there exists a U R such that
for all x 62 R there exists an integer n0 (x) > 0 such that Tk (x) 62 U for all
k n0 (x). If A ⇢ X and U A is open, we say that A is absorbing for U if
for any bounded set B ⇢ U, there exists an ` (which may depend on B) such
that Tk (B) ✓ A for all k `.

Definition 1 The model (1) is called dissipative if there exists a compact

attractor for T, whose basin of attraction is RN
+.

Let R+ = [0, 1), C = RN + denote the nonnegative orthant, with the integer N
representing the number of species, and C˚ denote the interior of C in RN . Here
we focus on a special class of Kolmogorov maps of the form T(x) = D[x]f (x)
where D[x] denotes the diagonal matrix whose elements are xi ij for i, j 2
IN := {1, . . . , N }. Thus the discrete-time model we study is
x(t + 1) = T(x(t)) = D[x(t)]f (x(t)), t 2 N. (2)
We will often use the following partial ordering for vectors u 2 RN : (a) u  v
if ui  vi for i 2 IN , (b) u < v if u  v but u 6= v, and (c) u ⌧ v if
v u 2 C. ˚ Similar orderings are used for matrices, so that, for example, for a
given matrix A, A > 0 means that each element of A is nonnegative, and at
least one element is positive.
Global stability of discrete-time competitive population models 7

We will use the following definition for a competitive map.

Definition 2 (Competitive map) A continuous map T : C ! C is compet-

itive if T(x) < T(y) ) x < y for each x, y 2 C.

Definition 3 (Strongly Competitive map) A continuous map T : C ! C

is strongly competitive if T(x) < T(y) ) x ⌧ y for each x, y 2 C.

A sufficient condition for T to be strongly competitive on a set ⌦ is that

DT 1 (x) 0 for x 2 ⌦.
We place the following standing assumptions on the map f :
1. Assumption A1 f : C ! C˚ is of class at least C 2 .
2. Assumption A2 For each i 2 IN , fi ( ei ) = 1 has a unique solution
⇤
i > 0.
⇤
Thus if there is only one extant species i, its equilibrium density is i.

3. Assumption A3 There is a unique interior fixed point p.

Here we only study stability of coexistence states. The stability of bound-
ary fixed points have been studied in [11–14, 30, 29, 31, 35] and will also be
considered by us in future work.

2.2 Previous work

For planar competitive discrete-time (and continuous-time) models (i.e. two

species models, N = 2), there is a standard approach to the study of stability
that does not rely (directly, at least) upon a global Lyapunov function, namely
monotone systems theory. It turns out that planar competitive systems are
monotone systems when a non-standard orthant ordering is used. By exploiting
this monotonicity, many authors, and in particular Smith [42, 6], have been able
to establish global stability results for both coexistence and partial extinction
states of well-known planar competitive models including the planar Leslie-
Gower and May-Oster models. Since there does not appear to be a suitable
ordering to render 3-species discrete-time competitive models monotone, it
seems likely that a similar approach will not lend itself to general discrete-
time competitive models and alternatives need to be found.
Some earlier ideas appeared in a little-cited 1977 paper, where Fisher and Goh
[9] present a Lyapnuov function approach to global stability of discrete-time
population models. The method described by us here has some similarities with
Fisher and Goh’s approach, particularly that they determine absorbing sets
based upon the sign of the di↵erence V between successive values of a specific
Lyapunov function V which is the sum of polynomial and logarithmic terms.
8 Stephen Baigent, Zhanyuan Hou

They do not use injectivity of the flow map, but are still able to determine
absorbing sets Y which are the union of level sets of the Lyapunov function,
and on which V  0, so that V is nonincreasing on Y . Here we develop a
similar approach, but with a di↵erent Lyapunov function and using that the
flow map is injective on our chosen sets.

Perhaps the most promising geometrical approaches to (2) have been devel-
oped by Franke and Yakubu in a series of papers [10, 13, 12], and also by Kon
in [30]. We first mention the work of Franke and Yakubu, which is primarily
concerned with exclusion principles for discrete-time population models. In
[10] the authors study global attractors of 2-species discrete-time competitive
models, and in particular establish an ecological principle for mutual exclu-
sion of species through the introduction of dominance of one species by the
other. Subsequent papers [12] extended these ideas to more species with the
introduction of the notions of weak and strong dominance.

In a similar, but distinct geometrical approach, Ryusuke Kon has studied

permanence and mutual exclusion principles in discrete-time competitive pop-
ulation models [29–31]. In [30] Kon uncovers a very nice geometrically-derived
exclusion principle based upon the convexity or concavity of the per-capita
growth rate functions. These papers contain sums similar to ⇥, but with equal
weights vi = 1 for i 2 IN , and utilised in quite a distinct fashion.

More recently Riuz-Herrera has used the carrying simplex often found in com-
petitive systems to study exclusion principles [35]. The carrying simplex is a
codimension-one invariant manifold [18, 28, 8] that attracts all nonzero points
in the first orthant. One of its appeals is that it contains all limit sets other
than the origin and hence stability or repulsion need only be checked on the
carrying simplex. In [35], the author provides proofs of results stated in [18]
and uses these to examine the equivalence for 3 species between non-existence
of interior fixed points and exclusion of a species (a form of the Poincaré-
Bendixson theorem on the 2-dimensional carrying simplex). For the latest
results on carrying simplicies for discrete-time systems see [27].

Employing an approach that has found success for continuous-time Lotka-

Volterra models, Wang and Zhengyi [46] assume diagonal dominance of the
interaction matrix and use a Lyapunov function to show global stability of an
N -species non autonomous May-Oster model under the assumption of strong
persistence. In particular, they deduce from known permanence results for the
2-species autonomous case, sufficient conditions for global stability (these have
also been shown by many other authors, e.g. [5, 6, 42], and are also derived here
as an application of our own methods in examples 1 and 2 later).

Recently Sacker [37] used dynamical reduction to demonstrate global stability

in a multi-species periodic Leslie-Gower model. Theorem 3.5 in [37] confirms
that in the autonomous case, the Leslie-Gower model with sufficiently weak
competition (to ensure that the mapping that he constructs is a contraction)
Global stability of discrete-time competitive population models 9

globally attracts the interior of the first orthant. Other models, including those
of May-Oster type are also considered in an earlier paper [38].
For di↵erential equations, recent geometrical approaches include the Split-
Lyapunov method introduced by Zeeman and Zeeman [47] for competitive
systems and extended to more general Lotka-Volterra systems in [22, 4].
The main objective of this paper is to link the geometrical ideas of those ap-
plied by Franke, Yakube, Kon and others, while also placing emphasis on injec-
tivity of maps, with the method of the Split Lyapunov function for continuous-
time Kolmogorov systems developed by E. C. Zeeman, M. L. Zeeman in [47],
Hou and Baigent [22, 4, 23]. A secondary objective is to demonstrate how exis-
tence of a carrying simplex, plus some knowledge of its geometry, can be used
to identify globally stable fixed points.

3 Forward invariance of sets under maps

For a continuous-time dynamical system generated by a smooth di↵erential

equation on RN forward invariance of a compact set B ⇢ RN is often obtained
by showing that the vector field points outwards nowhere on @B, so that
points on the boundary @B either stay in @B or move into B̊ and stay there
for all subsequent times. In discrete-time systems, even if the map T maps
all boundary points into B̊, then some points in B̊ may be mapped outside of
B, and then B is not forward invariant. For example, for the one-dimensional
model x0 = T (x) = 16(x 1)2 (x 2)2 + 32 and B = [1, 2] ⇢ R+ . Then
T (1) = 3/2 2 B̊, T (2) = 3/2 2 B̊ but T (3/2) = 5/2 62 B.
The following theorem is fundamental [24] (Lemma 2.1). Here X is a metric
space, and T : X ! X is continuous.

Theorem 3 Let U ⇢ X be an open set with compact closure and suppose that
V ⇢ X is open and forward invariant under T, and U ⇢ V. If O+ (x) \ U = 6 ;
for every x 2 V, then O+ (U ) is compact, forward invariant and absorbing for
V.

As pointed out by Kon [29], this gives a test for dissipativity of T. Here, we may
take V = C. If there is a compact set K ⇢ C such that O+ (x) \ K̊ 6= ; for every
x 2 C then O+ (K) is compact and absorbing for C and hence T is dissipative.
Thus, if B ⇢ RN is a compact set for which T(@B) ⇢ B then B may itself not
be forward invariant, but the typically larger compact set ⌦ = O+ (B) certainly
is. Theorem 3 is certainly useful for showing that a discrete dynamical system
is dissipative, but it is not always straightforward to determine explicitly the
absorbing set ⌦ = O+ (B). To obtain forward invariance based upon movement
of boundary points, which is typically much more practical to implement, we
will appeal to the recent result of Sacker [36] which shows that injectivity
10 Stephen Baigent, Zhanyuan Hou

of the map (which is immediate for the time-1 flow map of continuous-time
systems) is what is missing:

Theorem 4 (Sacker) Let D ⇢ Rn be a bounded subset and T : D ! Rn

be continuous. Suppose that T : D̊ ! Rn is injective and T(@D) ⇢ D. If
C
D = Rn \ D has no bounded components then T(D) ⇢ D.

Our strategy will be to utilise theorem 4 to determine explicitly a compact set

K that absorbs C. Properties of omega limit sets will then be examined via a
Lyapunov function restricted to K.
For an e↵ective application of theorem 4, we need sufficient conditions for T
to be injective. From the general results of Wang and Jiang (Theorem 4.1 in
[45]) we know that if T : C ! C is competitive and a local homeomorphism
on U ✓ C then T : U ! T (U ) is a global homeomorphism from U onto T(U ).
Thus for competitive maps T we need only check that det DT(x) 6= 0 for
x 2 U to conclude that T is injective on U .
Note that the injective condition is very restrictive. Our next result removes
this restriction and the boundedness requirement of D.

Theorem 5 Let D ⇢ Rn with D̊ a connected set and D̊ = D. Let T : D ! Rn

be continuous. Assume that there is a p0 2 D̊ such that T(p0 ) 2 D̊ and no
interior points of D are mapped to the boundary of D under T:
1
8x 2 @D, T (x) = {y 2 D : T(y) = x} \ D̊ = ;. (3)
Then T(D) ⇢ D and T(D̊) ⇢ D̊.

Proof Suppose T(D) 6⇢ D. Then there is a q 2 D such that T(q) 62 D. If

q 62 D̊, by D̊ = D and the continuity of T, we can always choose a point
q0 2 D̊ such that both kq0 qk and kT(q0 ) T(q)k are sufficiently small
so that T(q0 ) 62 D. Without loss of generality, we assume that q 2 D̊ with
T(q) 62 D. Since D̊ is connected, there is a continuous curve c ⇢ D̊ connecting
p0 to q. Since T(p0 ) 2 D̊ but T(q) 62 D, by continuity of c and T there is
r 2 c ⇢ D̊ such that T(r) 2 @D. This contradicts (3). Therefore, we must
have T(D) ⇢ D.
As T(D̊) ⇢ T(D) ⇢ D, if T(D̊) 6⇢ D̊ then there is y 2 D̊ such that x = T(y) 2
@D so T 1 (x) \ D̊ =
6 ;, a contradiction to (3). Hence, we also have T(D̊) ⇢ D̊.

4 Population Permanence

A key question in demographic studies is which, and how many, species can
coexist in a community. This does not ask for convergence to a steady pop-
ulation state, but rather that the population avoids the boundary of C, thus
Global stability of discrete-time competitive population models 11

allowing for more complicated dynamical attractors, such as attracting limit

cycles or deterministic chaos. A popular way of modelling this survival of all
species is to introduce the notion of permanence:

Definition 4 The system (1) is said to be permanent if it is dissipative and

˚
if T has an attractor contained in C.

Thus for a permanent population model, the boundary @C is repelling. This

repulsion leads to the well-known fact that a permanent system must have an
interior fixed point p.

Studies of permanence are wide-ranging, and include non-autonomous and

stochastic population models [41, 21,40,1].

Focussing on relations to our method, we highlight [20] where the authors

study the permanence of di↵erence equations of the form x0 = D[x] exp(f ).
The average Lyapunov function used in [20] is the same as that used here to
QN
show global stability, namely V (x) = i=1 xvi i with suitably chosen v. The
authors show
V (Tm (x)) = ↵(m, x)V (x)
⇣P ⌘
m 1 T j
where ↵(m, x) = exp j=0 v f (T (x) . Since our f corresponds to exp(f (x))
we may equate ↵(m, x) = ⇥(m, x). Using the average Lyapunov approach, the
authors in [20] show that if

0 1
m m
X1
V (T (x))
sup = sup exp @ vT f (Tj (x))A > 1
m 1 V (x)) m 1 j=0

for each x 2 S := K \ @C then the system x0 = D[x] exp(f ) is permanent.

In fact, using results from Hutson and Moran [24], Hofbauer et al. note that
(x) > 1 need only be checked for each x 2 [s2S !(s). Hence we note that
permanence requires only properties on the boundary of C, but to prove global
convergence to interior fixed points later (which of course implies permanence),
we will need to utilise properties of the Lyapunov function on all of the com-
pact attractor K and for all sufficiently large time. That the properties of T
on S = K \ @C are enough to show permanence follows from the fact that
boundary regularity allows boundary properties to partially extended into the
interior.

Not dealt with in [20] or it seems the related literature is how a choice of
the vector v might be motivated by the ecology. Our analysis below suggests
that a good candidate for v is D[p]↵ where ↵ is the principal reproductive
mode.
 12 Stephen Baigent, Zhanyuan Hou

5 Global attraction of interior fixed points

QN
Fix v 0 and define V (x) := i=1 xvi i , (x) := V (f (x)). For an orbit
O+ (x) 3 x(t) we will use the shorthand Vt := V (x(t)) and t = (x(t)) so
that Vt+1 = t Vt . Note that V 0 with equality only on the boundary @C and
that Vt+1 > Vt when t > 1. For any vector u = (u1 , . . . , uN )T 2 C, ˚ we shall
T
use the shorthand ln u for (ln u1 , . . . , ln uN ) . We will find it more convenient
to work with '(x) := ln (x) so that Vt+1 > Vt when 't > 0. Let us take some
x 2 C˚ and suppose that O+ (x) is bounded, so that the numbers Vt , t are
also bounded. The following theorem, an elaboration of LaSalle’s invariance
theorem, provides a test for global convergence to an interior fixed point p. It
does not require the results of section 3 for application.

Theorem 6 Under the general assumptions A1-A3 for system (2), assume
that the following conditions hold.
(i) There is a compact set K ⇢ C such that
˚ 9t1
8x(0) = x 2 C, 0, 8t t1 , x(t) 2 K.

(ii) There is a v 2 C˚ such that either

8x 2 K, '(x) = vT ln (x) 0 (4)

or
8x 2 K, '(x) = vT ln(x)  0. (5)

(iii) Under (5) the set S = {x 2 @C : '(x)  0} contains no invariant set.

˚ there is a constant c > 0 such that

Then for each x 2 C,
1 1
!(x) ✓ K \ ' (0) \ V (c).

In addition, if x 2 K \ (' 1 (0) \ {p}) implies T(x) 62 ' 1

(0) \ {p}, then the
interior fixed point p is globally attracting.
Q N vi
Proof With V (x) = i=1 xi , (x) = V (f (x)), Vt = V (x(t)) and t =
(x(t)), we⇣have V0 =⌘V (x(0)) = ⇣V (x), 0 = ⌘(x(0)) = (x) and Vt+1 = t Vt
Qt Pt
and Vt = k=0 k V0 = exp k=0 '(xk ) V0 . Suppose (4) holds. Then
Vk is monotone nondecreasing with k. By the boundedness of O+ (x), Vk is
bounded for k 2 N. Therefore, limk!1 Vk = c for some c > 0. Let y 2 !(x)
so that xtk ! y as k ! 1 for some sequence tk ! 1 as k ! 1. Since
limk!1 Vtk = c (all subsequences of the nondecreasing sequence Vk have
the same limit c), continuity of V shows that V (y) = c. This shows that
!(x) ⇢ V 1 (c). Now let O+ (y) be the forward orbit through y. Since !(x)
is invariant, c = V (y(t + 1)) = (y(t))V (y(t)) = (y(t))c, so that since
Global stability of discrete-time competitive population models 13

c > 0, (y(t)) = 1 for each t 2 N. This shows that y 2 ' 1

(0) and hence
!(x) ⇢ ' 1 (0). Thus !(x) ⇢ ' 1 (0) \ V 1 (c).
Alternatively suppose (5) holds. Now Vk is monotone nonincreasing in k so
there is a c 0 such that limk!1 Vk = c and so !(x) ⇢ V 1 (c). If c = 0,
then by the definition of V , V 1 (c) = @C so !(x) ⇢ @C and since k  1 for
k 0 we see that (y)  1, i.e. '(y)  0 for all y 2 !(x). This leads to
!(x) ⇢ S = {x 2 @C : '(x)  0}, a contradiction to condition (iii) as !(x)
is invariant. Therefore, we must have c > 0. Then by the same reasoning as
before, we have !(x) ⇢ K \ ' 1 (0) \ V 1 (c).
Now suppose y 2 ' 1 (0) \ {p} implies T(y) 62 ' 1 (0) \ {p}. We show that
!(x) = {p}, so that limt!1 x(t) = p. Assume !(x) 6= {p} and let y 2 !(x) \
{p}. Then, as !(x) ⇢ ' 1 (0) \ V 1 (c) ⇢ ' 1 (0), we have y 2 ' 1 (0) \ {p}
so T(y) 62 ' 1 (0) \ {p}. On the other hand, however, the invariance of !(x)
ensures that T(y) 2 !(x) ⇢ ' 1 (0). Thus, we must have T(y) = p. This shows
that !(x) = T(!(x)) = {p}, a contradiction to our supposition. Therefore,
!(x) = {p}.

6 Split-Lyapunov stability of interior fixed points

Up to now, and in particular for theorem 6, we have not specified how the
vector v used to construct the Lyapunov function V is to be chosen. Now
we turn to the discrete-time version of the Split Lyapunov method which is a
method for choosing v introduced for competitive Lotka-Volterra di↵erential
equations in [47] and developed further for general Lotka-Volterra systems in
[22, 4] and for general Kolmogorov di↵erential equations in [23].
Since we deal with strongly competitive models we may utilise the Frobnius-
Perron theorem for positive matrices such as DT(p) 1 0. Let u0 be a
positive left eigenvector in the 1-dimensional eigenspace associated with the
eigenvalue of DT(p) 1 of largest modulus µ0 > 0 say.
Recall the last condition ensuring convergence of theorem 6, namely that if
T (D0 \{p}) ⇢ D˚+ then '(T(x)) 0 for x 2 ' 1 (0) with equality only if x = p.
Hence ' T is minimised on ' 1 (0) at x = p which implies that there is a  2 R
such that r'(p) + DT(p)T r'(p) = 0. Thus for stationarity the normal
r'(p) must be a left eigenvector of DT(p). Note that since '(x) = vT ln f (x),
D0 = ' 1 (0) in theorem 7 is an (N 1)-dimensional surface with p 2 D0 . At
p, we compute a normal to this surfaces at p:
N
X rfi (p)
r'(p) = vi = vT Df (p) = ↵T (DT(p) I),
i=1
fi (p)

where v = D[p]↵. Hence if the last condition ensuring convergence of theorem

6 is to be obtained, we must take v = D[p]↵ where ↵ is a left eigenvector of
14 Stephen Baigent, Zhanyuan Hou

DT(p); ↵ is then also normal to D0 = ' 1 (0) at p. This provides a useful

constraint on practical choices of v. However, it does not suggest which of the
eigenvectors of DT(p) should be chosen. In section 7 we show that a suitable
choice for strongly competitive maps, at least in the case where there is a
carrying simplex, is that ↵ should coincide with the normal to the carrying
simplex at the interior fixed point p, which in turn corresponds to an eigenvec-
tor in the 1-dimensional eigenspace associated with the dominant eigenvector
of DT(p) 1 .
For the choice v = D[p]↵ we define V, , ' as above and take D+ = {x 2 C :
'(x) > 0}. The aim is to apply forward invariance theorems with D = D+ to
establish that D+ is forward invariant. First, for Sacker’s theorem, we require
(i) D to be bounded and T to be injective on D˚+ . Next (ii) the set C \ D+ ,
i.e. {x 2 C : '(x)  0}, must have no bounded components. Finally, perhaps
the most difficult part we must show that (iii) T(@D+ ) ⇢ D+ . If (i)-(iii) hold
true then by theorem 4 the set D+ is forward invariant, and if '(x) 0 then
'(Tn x) 0 for all n 2 N.

Theorem 7 Under the general assumptions A1-A3 for system (2), suppose
that T is dissipative. Let p be an interior fixed point of T and ↵ a positive
left eigenvector of DT(p) associated with a real eigenvalue µ 6= 0. Define the
function ' : C ! R by '(x) := ↵T D[p] ln f (x) and the sets D+ := {x 2 C :
'(x) > 0}, D := {x 2 C : '(x) < 0} and D0 := {x 2 C : '(x) = 0}. Suppose
that (i) D+ is bounded and T is injective on D˚+ , (ii) D has no bounded
components in C, (iii) T(D0 [ (D+ \ @C)) ⇢ D+ , and (iv) T has no invariant
set in @C \ D .
˚ !(x) ✓ D0 \ V
Then for each x 2 C, 1
(c) for some c > 0.
Moreover, if T(D0 \ {p}) ⇢ D˚+ then !(x) = {p} for all x 2 C.
˚

Proof Let K = D+ . Then, by conditions (i)–(iii) in theorem 4 and the con-

clusions of theorem 4, K is forward invariant. Given any x 2 C,˚ there are two
possibilities: (a) x(t) = T (x) 62 K for any t 2 N or (b) x(⌧ + t) 2 K for some
t

⌧ 2 N and all t 2 N.
In case (b) we immediately have '(x(t)) > 0. Thus by theorem 6, !(x) ⇢
D0 \ V 1 (c). In case (a) we have '(x(t)) < 0 for all t 2 N, so that by theorem
6 !(x) ⇢ D0 \ V 1 (c). Finally, if T (D0 \ {p}) ⇢ D˚+ , then since !(x) ✓ D0 is
an invariant set and T (D0 \ {p}) \ D0 = ;, we must have !(x) = {p}.

Remark 1 Note that the choice of nonzero eigenvalues of DT(p) is irrelevant to

the proof of theorem 7 as long as it has a corresponding positive left eigenvector
↵. However, if we are aiming at global stability of p, we implicitly require a real
eigenvalue µ of DT(p) to satisfy 0 < |µ| < 1. Since DT(p) = I + D[p]Df (p),
if ↵ is a positive left eigenvector of DT(p) associated with an eigenvalue µ
then it is also a positive left eigenvector of D[p]Df (p) associated with the
Global stability of discrete-time competitive population models 15

eigenvalue µ 1. Moreover, for p to be stable, we must have µ 1 2 ( 2, 0).

This remark applies to all the onward results.

Next, we apply theorem 5 for global attraction to an interior fixed point.

Theorem 8 Under the general assumptions A1-A3 for system (2), suppose
that T is dissipative. Let p be an interior fixed point of T and ↵ a positive
left eigenvector of DT(p) associated with a real eigenvalue µ 6= 0. Define the
function ' : C ! R by '(x) := ↵T D[p] ln f (x) and the sets D+ := {x 2 C :
'(x) > 0}, D := {x 2 C : '(x) < 0}, D0 := {x 2 C : '(x) = 0} and
D = C˚\ D+ . Suppose that (i) D is connected, (ii) T(D) \ D =
6 ;, (iii) for each
x 2 D0 , {y 2 C : T(y)) = x} \ D = ;, and (iv) T has no invariant set in
@C \ D .
˚ !(x) ✓ D0 \ V
Then for each x 2 C, 1
(c) for some c > 0.
Moreover, if T(D0 \ {p}) ⇢ D˚+ then !(x) = {p} for all x 2 C.
˚

Proof Let K = D. Then, by conditions (i)–(iii) and theorem 5, K is forward

invariant. The rest of the proof is the same as that of theorem 7.

In addition, the the conditions of the following theorem that uses a tangent
hyperplane approximation to the carrying simplex are sometimes easier to
satisfy.

Theorem 9 Suppose that p is an interior fixed point of T, ↵ is a positive left

eigenvector of DT(p) associated with a real eigenvalue µ 6= 0, and the function
'(x) = ↵T D[p] ln f (x) is convex. Let ⇠(x) = ↵T (x p) and K = {x 2 C :
⇠(x)  0}. Then if T(K) ⇢ K and ⇠(T(x)) < ⇠(x) for all x 2 C \ {p} with
˚
⇠(x) 0, the fixed point p globally attracts C.

Proof We first claim that !(x) ⇢ K for all x 2 C. As a matter of fact, this
is obvious for x 2 K since T(K) ⇢ K. For x 2 C \ K, we have ⇠(x) > 0 so
⇠(T(x)) < ⇠(x). If there is an integer m > 0 such that ⇠(Tm (x))  0, then
Tm (x) 2 K and the positive invariance of K implies that Tk (x) 2 K for all
k m so !(x) ⇢ K. Otherwise, we have ⇠(T m (x)) > 0 for all m > 0 so that
⇠(T m (x)) is decreasing. Then there is a c 0 such that limm!1 ⇠(Tm (x)) = c,
i.e. !(x) ⇢ ⇠ 1 (c). If c > 0 then for any y 2 !(x), we have ⇠(y) = c > 0 so
⇠(T(y)) < ⇠(y) = c. But the invariance of !(x) implies that T(y) 2 !(x) so
⇠(T(y)) = ⇠(!(x)) = c, a contradiction. Therefore, c = 0 so !(x) ⇢ ⇠ 1 (0) ⇢
K.
˚ Suppose !(x) 6= {p} for some
Next, we show that !(x) = {p} for all x 2 C.
˚ Then !(x) \ {p} =
x 2 C. 6 ;. Taking y 2 !(x) \ {p} ⇢ K, we have ⇠(y)  0
so either ⇠(y) < 0 or ⇠(T(y)) < ⇠(y) = 0. As both y and T(y) are in !(x),
we can always find an integer m > 0 such that Tm (x) is close enough to y or
T(y) so that ⇠(Tm (x))  0, i.e. Tm (x) 2 K. By the positive invariance of K
16 Stephen Baigent, Zhanyuan Hou

we have Tk (x) 2 K for all k m. Since the function '(x) is convex, the set
D = D [ D0 is also convex. As the plane defined by ⇠(x) = 0 is the tangent
plane of D0 at p, we must have K \ {p} ⇢ D+ . Then, from theorem 6 (ii), we
obtain !(x) ⇢ K \ D0 = {p}, a contradiction to the supposition !(x) 6= {p}.
˚
Hence, !(x) = {p} for all x 2 C.

Based on theorem 7, theorem 8 or theorem 9, we are now in a position to

deal with the global asymptotic stability of p, i.e. !(x) = {p} for all x 2 C˚
and
8" > 0, 9 > 0, 8x 2 B (p) \ C,˚ 8n 1, Tn (x) 2 B" (p).
Here Br (p) is the open ball centred at p with a radius r > 0.

Fig. 2 An illustration of the asymptotic stability argument in theorem 10 of the interior

fixed point p.

Theorem 10 (Split Lyapunov Stability) Assume that p is globally at-

tracting by theorem 7, theorem 8 or theorem 9. Then p is globally asymptoti-
cally stable if one of the following conditions is met:
(a) Each eigenvalue µ of DT(p) satisfies |µ| < 1.
(b) Under the conditions of theorem 7 or theorem 8, there exist ⇢ > 0 and an
integer m > 0 such that T m B⇢ (p) ✓ D+ [ D0 .
(c) Under the conditions of theorem 9, there exist ⇢ > 0 and an integer m > 0
such that T m B⇢ (p) ✓ K.

Proof Under condition (a), p is locally stable. Then the conclusion follows
from this and the global attraction of p.
Global stability of discrete-time competitive population models 17

Now suppose condition (b) holds. We need only show the local stability of
p. Restricted to a small open ball centred at p, D0 and V 1 (`) = {x 2 C˚ :
V (x) = `} for any ` > 0 close to V (p) are (N 1)-dimensional surfaces (see
figure 2). From the proof of theorem 7 or theorem 8 we know that D+ \ C˚
is forward invariant. Thus, x 2 D+ \ C˚ implies x(t) = Tt (x) 2 D+ \ C˚ for
all t 2 N so that Vt+1 (x) = t Vt (x) Vt (x), Vt (x) tends to V (p) = `⇤
monotonically (since x(t) ! p) as t ! 1. If x 2 D+ \ C˚ but x 6= p, the
condition T(D0 \ {p}) ⇢ D+ implies that Vt (x) strictly increases to `⇤ as
t ! 1. This shows that D0 \ V 1 (`⇤ ) = {p}. Moreover, for any ` 2 (0, `⇤ )
close enough to `⇤ , the two surfaces D0 and V 1 (`) intersect with each other
near p so that the closed set A` = {x 2 C˚ : '(x) 0, V (x) `} is forward
invariant. Now for any given " > 0, we can choose ` 2 (0, `⇤ ) close enough to
`⇤ so that A` ⇢ B" (p) (see figure 2). Then, by (ii) and continuity of T, we can
choose 2 (0, min{⇢, "}) sufficiently small such that

8t 2 Im , Tt B (p) ⇢ B" (p); Tm B (p) ⇢ A` .

Hence, by forward invariance of A` , we have Tt B (p) ⇢ B" (p) for all t =

1, 2, . . . so p is stable.
Suppose (c) holds. Then, since (K \ C)˚ ⇢ (D+ [ {p}) and T(K) ⇢ K, the above
reasoning under (b) is still valid after the replacement of D+ by K.

6.1 Ultra-bounded population models

Some models of Kolmogorov form satisfy

Tk (x) = xk fk (x) ! 0 for each k as |x|1 ! 1 in C. (6)

Definition 5 We will call a map T : C ! C˚ satisfying (6) ultra-bounded.

An example is the May-Oster model Tk (x) = xk exp(rk (1 (Ax)k ) where

ri > 0 and A 0. Using simplex coordinates R = |x|1 and uk = xk /R we
have that Tk (R, u) = Ruk fk (Ru) ! 0 as R ! 1. Thus, for each u 2 , the
parameterised curve u = S{T (R, u) : R 2 [0, 1)} is a bounded and closed
curve in C. The set TC = u2 u is, by construction, a compact absorbing
set for C. The boundary of TC in C is envelope of all the curves u as u varies
over and it is precisely the set of points x 2 C such that det DT(x) = 0.
The map T is thus an injective map from TC into into itself.
Thus we have the following result for ultra-bounded models.

Theorem 11 Let the model T(x) = D[x]f (x) be ultra-bounded and compet-
itive. Let p be an interior fixed point of T and ↵ a positive left eigenvec-
tor of DT(p) associated with a real eigenvalue µ 6= 0. Define the function
' : C ! R by '(x) = ↵T D[p] ln f (x) and the sets D+ = {x 2 C : '(x) > 0}
18 Stephen Baigent, Zhanyuan Hou

and D0 = {x 2 C : '(x) = 0}. Then p is globally asymptotically stable if one

of the following conditions holds:
(i) TC ⇢ D+ [ {p}.
(ii) D˚+ [ {p} ⇢ TC is absorbing for C.
˚

Proof Under either (i) or (ii), the conclusion follows directly from theorem 10.

We now illustrate the Split Lyapunov method with a couple of examples.

Example 1 Planar May-Oster model.

Here x = (x, y)T and T = (f, g)T where
⇣ ⌘
T(x)T = xer(1 x ↵y) , yes(1 x y)
. (7)

This map is the class of maps studied for 0  r, s  1 and ↵, > 0 by Smith
[42] and also in [25]. Smith showed that (i) if 0 < ↵ < 1 < then e1 = (1, 0)T
attracts all points not on the y-axis, (ii) if < 1 < ↵ then e2 = (0, 1)T attracts
all points not on the x-axis, (iii) if ↵, < 1 then there is a unique interior
fixed point p that attracts the interior of R2+ , and (iv) if 0 < r, s < 1 and
↵, > 1 then there is a C 1 separatrix which partitions R2+ into the basin of
attractions of e1 , e2 , and \{0} is the stable manifold of p. In other words the
situation is exactly analogous to the well-known planar di↵erential equation
competition model.
We restrict attention to 0 < r, s  1.
This model is ultra-bounded and TC is a compact absorbing set for C = R2+ .
The function ' in this model is actually linear: '(x) = rv1 (1 x ↵y) +
sv2 (1 y x). The determinant det DT(x) = 0 when (x) := (1 rx)(1
sy) rs↵ xy = 0 which has one branch in C when ↵ > 1 and two branches
in C when ↵ < 1. When ↵ > 1, the region C \ TC is convex (so that
the relative boundary of TC in C is the graph of a convex function). When
↵ < 1, the region TC is convex (its relative boundary in C is the lower branch
of the hyperbola, which is the graph of a concave function). The set D+ is the
triangular region below the hyperplane D0 = ' 1 (0) in C.
⇣ ⌘
When ↵ < 1 and < 1, the system has an interior fixed point p = 11 ↵↵ , 11 ↵ .
As ✓ r(1 x y↵) ◆
e (1 rx) er(1 x y↵) rx↵
DT(x, y) = ,
es(1 y x ) sy es(1 y x ) (1 sy)
✓ ◆
rp1 rp1 ↵ @(f,g)
we have DT(p) = I ⌦ where ⌦ = = D[p] @(x,y) (p). We find
sp2 sp2
that det ⌦ = rsp1 p2 (1 ↵ ) and trace ⌦ = rp1 + sp2 . Moreover,

trace⌦ 2 4 det ⌦ = (rp1 +sp2 )2 4rsp1 p2 (1 ↵ ) = (rp1 sp2 )2 +4rsp1 p2 ↵ > 0

Global stability of discrete-time competitive population models 19

αβ < 1 αβ > 1
δ>0

δ>0 TC TC
p D0 = ϕ −1 (0)
T (D0 )
D+ δ>0

Fig. 3 The areas TC for the model (7) when (a) ↵ < 1 and (b) ↵ > 1. The map T is
injective within the absorbing set TC and D0 [ D+ ⇢ TC.

so that the eigenvalues ± of ⌦ satisfy 0 < ± < trace⌦ < 2. Hence the
eigenvalues of DT(p), 1 ± , lie in ( 1, 1) when ↵, < 1 and we have local
stability. Now we obtain global stability.
From the local stability of p, we need only show the global attraction of p
by theorem 7. Note that D0 is given by '(x) = 0, v = D[p]↵ where ↵ is a
positive left eigenvector of ⌦ corresponding to an eigenvalue 2 ( 2, 0), so
↵ is a normal vector of D0 at p (↵ = (↵1 , ↵2 )T should not be confused with
the parameter ↵). Thus,

r↵1 p1 s ↵ 2 p2 = ↵ 1 , (8)
r↵↵1 p1 s↵2 p2 = ↵2 , (9)

so that eliminating the and writing z = ↵1 /↵2 we have r↵p1 z 2 + (sp2

rp1 )z s p2 = 0. There is a positive root
p
4(↵ 1)rsp1 p2 + (rp1 + sp2 )2 + rp1 sp2
z= .
2↵rp1
↵2 1 1
Since ↵ < 1 we find z < 1/↵. Similarly we find ↵1 = z < and hence

↵1 1
< < . (10)
↵2 ↵
20 Stephen Baigent, Zhanyuan Hou

From the definition of ', '(p) = 0, (8) and (9) we may rewrite '(x) =
[↵1 (x p1 ) + ↵2 (y p2 )].
We first show that T(D0 \ {p}) ⇢ D+ , i.e. '(T(x)) > 0 for x 2 D0 \ {p}. For
this purpose, we use (x, y(x)) to parameterise D0 and compute

(x) = ↵1 x exp (r(1 x ↵y(x))) + ↵2 y(x) exp (s(1 y(x) x)) ↵ 1 p1 ↵ 2 p2

= ↵1 x exp ( r((x p1 ) + ↵(y(x) p2 )))
+↵2 y exp ( s((y(x) p2 ) + (x p1 ))) ↵ 1 p1 ↵ 2 p2
✓ ◆
↵1
= ↵1 x exp r(x p1 )(1 ↵ )
↵2
✓ ◆
↵1
+(↵1 p1 + ↵2 p2 ↵1 x) exp s(x p1 )( ) ↵ 1 p1 ↵ 2 p2
↵2
= ↵1 x exp ( r(x p1 )$1 )
+(↵1 p1 + ↵2 p2 ↵1 x) exp ( s(x p1 )$2 ) ↵ 1 p1 ↵ 2 p2

where $1 = 1 ↵↵2 ↵ and $2 =

1 ↵1
↵2 . Note that $1 > 0, $2 < 0 from (10) and
r↵1 p1 $1 = s↵2 p2 $2 from (8) and (9). As (x) = '(T(x, y(x)))/ , we need
only show that (x) < 0 for x 2 [0, ↵1 p1↵+↵
1
2 p2
] with x 6= p1 . Di↵erentiation of
(x) gives
0 r(x p1 )$1 s(x p1 )$2
(x) = ↵1 (1 r$1 x)e [↵1 + s$2 (↵1 p1 + ↵2 p2 ↵1 x)]e .

Then 0 (p1 ) = r↵1 $1 p1 s↵2 $2 p2 = 0. Let `(x) = ↵1 (1 r$1 x)

[↵1 + s$2 (↵1 p1 + ↵2 p2 ↵1 x)]. Then `(0) = s$2 (↵1 p1 + ↵2 p2 ) > 0 and
`( ↵1 p1↵+↵
1
2 p2
) = r$1 (↵1 p1 + ↵2 p2 ) < 0. As `(x) is linear and `(p1 ) = 0, we
have `(x) > 0 for x 2 [0, p1 ) and `(x) < 0 for x 2 (p1 , ↵1 p1↵+↵ 1
2 p2
]. Note also
that ↵1 (1 r$1 x) > 0 for x 2 [0, p1 ) and [↵1 + s$2 (↵1 p1 + ↵2 p2 ↵1 x)] < 0
for x 2 (p1 , ↵1 p1↵+↵
1
2 p2
]. Then, from the expression of 0 (x) we obtain
0 s(x p1 )$2
(x) > `(x)e , x 2 [0, p1 ),

0 r(x p1 )$1 ↵ 1 p1 + ↵ 2 p 2 i
(x) < `(x)e , x 2 (p1 , .
↵1
Therefore, 0 (x) and `(x) have the same sign for each x. This shows that
(x) < (p1 ) = 0 for all x 2 [0, ↵1 p1↵+↵
1
2 p2
] with x 6= p1 so T(D0 \ {p}) ⇢ D+ .
Next, we assume that 1 > ↵ s > 0 and 1 > r > 0. Then, as D0 is
given by ↵1 (x p1 ) + ↵2 (y p2 ) = 0 in C, it has end points ( ↵1 p1↵+↵ 1
2 p2
, 0)
↵1 p1 +↵2 p2
and (0, ↵2 ). As the lower branch of the curve = 0 in C has end points
( 1r , 0) and (0, 1s ), from (10) we have
↵ 1 p 1 + ↵ 2 p2 p2 1 ↵ 1 p 1 + ↵ 2 p2 p1 1
< p1 +  ; < p2 +  .
↵1 r ↵2 ↵ s

Thus, T is injective on D+ . On the x-axis, T1 (x, 0) = xer(1 x)

has maximum
er 1 /r < 1/r at x = 1/r; on the y-axis, T2 (0, y) = yes(1 y)
has maximum
Global stability of discrete-time competitive population models 21

es 1 /s < 1/s at y = 1/s. Thus, T(@D+ ) ⇢ D+ . By Sacker’s theorem, T(D+ ) ⇢

D+ . Clearly, there is no invariant set on either x-axis with x 1/r or y-axis
with y 1/s. Then all the conditions of theorem 7 are met and p attracts C. ˚

7 Simplifications that utilise a carrying simplex

The carrying simplex ⌃ [17, 18, 35, 8, 28, 3, 26, 44] is an invariant manifold that
is a common feature of many continuous- and discrete-time competitive sys-
tems. ⌃ is a Lipschitz manifold of codimension one that attracts all nonzero
orbits and is unordered in the sense that no two points in ⌃ may be ordered
with respects to the standard ordering which use the first orthant as a cone.
The most useful property of the carrying simplex for us here, and the one
exploited by Zeeman and Zeeman in [47] to study global stability of interior
fixed points of competitive Lotka-Volterra systems, and later in [22] for the
case of boundary fixed points, is that ⌃ contains all limit sets of nontrivial
orbits. Consequently, ⌃ may be used as the absorbing set K in the results
developed above.
We recall that D+ = {x 2 C : '(x) > 0}, D = {x 2 C : '(x) < 0} and
D0 = {x 2 C : '(x) = 0}, where '(x) = ↵T D[p] ln f (x) with ↵ being a
positive left eigenvector of DT(p).
We use the following definition of convexity or concavity of the carrying sim-
plex [47]. Let ⌃+ and ⌃ denote the two components of C \ ⌃ that lie above
and below ⌃ respectively. We call ⌃ convex, flat, or concave if for all x, y 2 ⌃,
the interior of the line segment xy lies in ⌃ , ⌃, or ⌃+ respectively.

Lemma 1 Suppose that (2) has a carrying simplex ⌃ and p 2 ⌃ is an interior

fixed point. Let ↵ be a left eigenvector of DT(p) associated with a real eigen-
value µ 6= 0 and let '(x) = vT ln(f (x)) where v = D[p]↵. Then D0 = ' 1 (0)
is tangent to ⌃ at p.

Proof ⌃ globally attracts C \ O. The manifold = {x 2 C : ↵T (x p) = 0} is

mapped into a sequence of manifolds Mt = Tt , t 2 N and limt!1 Mt = ⌃.
For each t 2 N let n(t) be normal to Mt at p, with n(0) = ↵. Then n(t + 1) =
cof(DT(p))n(t) is normal to Mt+1 at p (where cof(P ) denotes the cofactor
matrix of P , and cof(P ) = (det P )P 1 when P is nonsingular). Since ↵ is a left
eigenvector of DT(p) it is also a left eigenvector of (det DT(p))DT(p) 1 =
cof(DT(p)), so ↵ normal to Mt at p for all t 2 N. This implies that ↵ belongs
to the normal cone to ⌃ at p. Now we use that ⌃ is di↵erentiable at p [26] so
that the normal at p to ⌃ is in the direction of ↵.

Corollary 1 Suppose that (2) has a carrying simplex ⌃ and that D0 = ' 1 (0)
defines a hypersurface that separates D+ and D in C: C = D0 [ D+ [ D . If
⌃ ⇢ D+ [ {p}, then p attracts C.˚
22 Stephen Baigent, Zhanyuan Hou

Proof We first show that, for any x 2 (⌃ \ {p}) \ C, ˚ limt!1 Tt (x) = p and
V (x) < V (p). Indeed, since (⌃ \ {p}) \ C˚ is invariant and (⌃ \ {p}) ⇢ D+ , we
have x(t) = Tt (x) 2 ⌃ \ {p} so Vt+1 (x) = t Vt > Vt (x) for all integers t 0.
As Vt (x) is increasing and bounded, there is a c > 0 such that

t
Y
lim Vt+1 (x) = lim iV (x) = c,
t!1 t!1
i=0

which implies limt!1 t = 1, i.e. limt!1 '(x(t)) = '(!(x)) = 0, so !(x) ⇢

D0 . As !(x) ⇢ ⌃ and ⌃ \ D0 = {p}, we must have !(x) = {p}. Hence,
limt!1 x(t) = p and V (x) < limt!1 Vt (x) = V (p).
Next, we show that limt!1 x(t) = p for any fixed x 2 C. ˚ For any " > 0,
by continuity of T there is a 2 (0, ") such that TB (p) ⇢ B" (p). As ⌃ is
compact, D+ is open in C and ⌃ \ {p} ⇢ D+ , for this there is a ⇢ 2 (0, )
such that ⌃⇢ = [u2⌃ B⇢ (u) satisfies ⌃⇢ \ D ⇢ B (p). Since !(x) ⇢ ⌃, for
this ⇢ there is an integer K 0 such that x(t) 2 ⌃⇢ for ¯ all t K. If x(t) 2 D+
for sufficiently large t then the same reasoning as in the previous paragraph
shows that limt!1 x(t) = p. Otherewise, there are infinitely many t > K such
that x(t) 2 ⌃⇢ \ D ⇢ B (p), so that x(t + 1) 2 ⌃⇢ \ B" (p) and Vt+1 (x)
minu2B" (p) V (u) = `("). If x(t + 1) 2 ⌃⇢ \ D then x(t + 2) 2 ⌃⇢ \ B" (p) so
Vt+2 (x) `("); if x(t + 1) 2 ⌃ \ (D+ [ D0 ) then Vt+2 (x) Vt+1 (x) `(").
In all cases, we derive that Vt (x) `(") for all large t. As lim"!0 `(") = V (p),
by letting t ! 1 and then " ! 0 we see that V (!(x)) V (p). As !(x) ⇢ ⌃
and V (y) < V (p) for all y 2 ⌃ \ {p}, we must have !(x) = {p}.

With the help of lemma 1 and corollary 1 we are able to prove the following
result.

Corollary 2 Suppose that (2) has a carrying simplex ⌃, an interior fixed

point p, and that D0 = ' 1 (0) defines a hypersurface that separates D+ and
D in C: C = D0 [ D+ [ D . If D is a convex set and the carrying simplex
˚
⌃ is convex, then p attracts C.

Proof From the definition of ' we know that ↵ is a positive left eigenvector of
DT(p) associated with a real eigenvalue µ = 1+ 6= 0 and of D[p]Df (x) asso-
PN
ciated with . Also, by lemma 1, r'(p) = i=1 ↵i pi Df i (p) T
fi (p) = ↵ D[p]Df (p) =
↵T is normal to the carrying simplex ⌃ at x = p. Since r'(p) is also a nor-
mal vector of D0 at p, and both ⌃ and D are convex with D0 as part of the
boundary of D , the set ⌃ \ {p} is below the hyperplane r'(p)T (x p) = 0
whereas D is above the plane. Thus, ⌃ ⇢ D+ [ {p}. Then the conclusion
follows from corollary 1.

Example 2 Planar Leslie-Gower model

Global stability of discrete-time competitive population models 23

Consider the planar Leslie-Gower model

✓ ◆
T ↵x y
T(x) = , . (11)
1 + x + ay 1 + y + bx

Cushing et al. [6] showed that if (a) ↵, < 1 then e0 = (0, 0)T is globally
asymptotically stable on R2+ , (b) ↵ > 1, < 1 then e1 = (↵ 1, 0)T is globally
asymptotically stable on intR2+ , (c) ↵ < 1, > 1 then e2 = (0, 1)T is
2
globally asymptotically stable on intR+ . When ↵ > 1, > 1, e0 is a repeller
and there are 4 distinct cases: When (a) b(↵ 1) > 1, ↵ 1 > a( 1)
then e1 is asymptotically stable on intR2+ and e2 is a saddle, (b) b(↵ 1) <
1, ↵ 1 < a( 1) then e2 is asymptotically stable on intR2+ and e1
is a saddle, (c) when b(↵ 1) < 1, ↵ 1 > a( 1) then the interior
fixed point p is globally asymptotically stable on intR2+ and (d) when b(↵
1) > 1, ↵ 1 < a( 1) then the interior fixed point p is a saddle.
Here we are concerned with the case ↵, > 1. An interior fixed point has
coordinates pT = ( a(1 1 )+↵
ab
1
, 1+b(1
1 ab
↵)
) and so is feasible, since we are
assuming ↵ > 1, > 1 when either (a) a < ↵ 11 < 1b or (b) 1b < ↵ 11 < a. As
we will show, in the case (a) p is globally asymptotically stable in the interior
of R2+ .
The sets D+ , D0 and D are defined by
1 1
'(x) = ↵1 p1 ln ↵(1 + x + ay) + ↵2 p2 ln (1 + y + bx) > 0,

' = 0 and ' < 0 respectively. The equation '(x) = 0 in R2+ for D0 determines
dy
that y is a function of x. From ' = 0 we have 'x + 'y dx = 0 so

dy 'x ↵1 p1 (1 + y + bx) + b↵2 p2 (1 + x + ay)

= = < 0.
dx 'y a↵1 p1 (1 + y + bx) + ↵2 p2 (1 + x + ay)

Further,

d2 y (1 ab)2 ↵1 p1 ↵2 p2 (↵1 p1 + ↵2 p2 )(1 + y + bx)(1 + x + ay)

= > 0.
dx2 [a↵1 p1 (1 + y + bx) + ↵2 p2 (1 + x + ay)]3

This shows that the function y(x) determined by ' = 0 is strictly convex, so
D is convex.
The geometry of the carrying simplex ⌃ for this model has recently been
investigated [2]. Define = (1 + a( 1))(1 + b(↵ 1)) ↵ . Then in [2] it was
shown that ⌃ is convex when > 0, concave when < 0 and a straight line
when = 0. In the case (a), 1 + b(↵ 1) < and 1 + a( 1) < ↵ and hence
= (1 + b(↵ 1))(1 + a( 1)) ↵ < 0, so that in case (a) the carrying
simplex is convex. Hence, from corollary 2 we conclude that in case (a) the
interior fixed point p is globally attracting.
24 Stephen Baigent, Zhanyuan Hou

For global asymptotic stability, we need only check that the eigenvalues of
DT(p) lie in ( 1, 1). A straightforward computation gives DT(p) = I ⌦
where ✓ ◆
p1 ap1
↵ ↵
⌦= bp2 p2 .

p1
Then each eigenvalue of ⌦ satisfies 0 < < trace⌦ = ↵ + p2 . From 1 + b(↵
1) < and (a) we have 1 b(↵ 1) < ( 1)(1 ab) so p2 < 1
< 1.
Similarly, from 1+a( 1) < ↵ and (a) we have ↵ 1 a( 1) < (↵ 1)(1 ab)
so p↵1 < ↵↵ 1 < 1. Thus, 2 (0, 2) and each eigenvalue 1 of DT(p) lies in
( 1, 1).

Example 3 3-species May-Leonard Leslie-Gower models.

We take
0 1
1↵
bxi
Ti (x) = , i = 1, 2, 3, A = @ 1 ↵ A , ↵, > 0. (12)
1 + (Ax)i
↵ 1

Note that for b < 1, the map T is a contraction with O as the unique fixed
point. Thus we will assume b > 1 in what follows.
The map T has an unique interior fixed point at p = ((b 1) 1 , (b 1) 1 , (b
1) 1 ) where = 1 + ↵ + . We will prove global asymptotic stability of p in
C˚ (theorem 1 given in section 1) via theorem 9.
From the definition of T we obtain
b b b
DT(x) = diag[ , , ]
1 + (Ax)1 1 + (Ax)2 1 + (Ax)3
bx1 bx2 bx3
diag[ , , ]A.
(1 + (Ax)1 ) (1 + (Ax)2 ) (1 + (Ax)3 )2
2 2

It was shown (for example, in [27]) that T is a competitive map, that T is

injective on C and that (12) has a carrying simplex, but its geometry is not
known. Now we determine parameter values for which p is asymptotically
stable through showing that DT(p) has eigenvalues of modulus less than one.
From the expression of DT(x) we see that DT(p) = I 1b D[p]A, which has
eigenvalues
p
1 2b(↵ + + 1) + (b 1)(↵ + 2) 3(b 1)|↵ |
1 = , 2,3 = ± i.
b 2b(↵ + + 1) 2b(↵ + + 1)

Then | 2,3 |2 < 1 if and only if (1 b 1 )[(↵ )2 + (1 ↵)(1 )] < (↵ + +

1)(2 ↵ ). Under the conditions b > 1, ↵  1,  1 with ↵ + < 2, we
have

(↵ )2 +(1 ↵)(1 ) (↵+ +1)(2 ↵ ) = 2↵(↵ 1)+2 ( 1)+↵ 1 < 0.

Global stability of discrete-time competitive population models 25

Thus, p is asymptotically stable when b > 1, ↵  1,  1 with ↵ + < 2.

˚
The rest is to show by theorem 9 that p attracts C.
The matrix DT(p) has a left eigenvector ↵T = (1, 1, 1) with associated eigen-
value µ1 = 1/b. The function '(x) = ↵T D[p] ln f (x) becomes

(b 1)
'(x) = [3 ln b ln(1 + (Ax)1 ) ln(1 + (Ax)2 ) ln(1 + (Ax)3 )].
↵+ +1
2
For any distinct x, y 2 C, we can check that dds'2 (sx + (1 s)y) > 0 for all
s 2 [0, 1] so that '(x) is convex. For the function

⇠(x) = ↵T (x p) = kxk1 ✓0 , ✓0 = 3(b 1)/(↵ + + 1),

define K = {x 2 C : ⇠(x)  0}. If we can prove that ⇠(T(x)) < ⇠(x) holds for
all x 2 C \{p} with ⇠(x) 0, then p attracts C˚ by theorem 9. We are therefore
working with a general ✓ ✓0 so as to show that for each ✓ ✓0 points on
the plane kxk1 = ✓ move below this plane.
For this purpose, we define the function ✓ (x) = ✓ kxk1 for a fixed ✓ ✓0 .
Then ⇧✓ = ✓ 1 (0) defines a plane with normal (1, 1, 1)T passing through the
point ( ✓3 , ✓3 , ✓3 )T . We need only show that ✓ (T(x)) > 0 for all x 2 ⇧✓ \ C \ {p}
for all ✓ ✓0 . We thus find the extrema of

3
X bxi
✓ (T(x)) =✓
i=1
1 + (Ax)i

on x 2 ⇧✓ \ C which written in terms of extrema of the function

b✓u1 b✓u2 b✓u3

✓ (u) =✓
Cu1 + Au2 + Bu3 Bu1 + Cu2 + Au3 Au1 + Bu2 + Cu3

on u 2 3 := {x 2 C : kxk1 = 1}, where C = ✓ + 1, A = ✓↵ + 1, B = ✓ + 1.

Thus, we need only show that ✓ (u) > 0, except u1 = u2 = u3 when ✓ = ✓0 ,
under the conditions of theorem 1. But note that if for some ✓1 > 0 and u 2 C
we have ✓1 (u) 0, then ✓2 > ✓1 implies ✓2 (u) > ✓1 (u) 0 since each of
A, B, C are increasing with ✓. Hence we need only concern ourselves with the
case ✓ = ✓0 . From now on we drop the ✓ subscript on the understanding that
✓ = ✓0 .
We now show that if (T(x)) > 0 for x 2 @⇧✓ \ C then (T(x)) 0 for
all x 2 ⇧✓ \ C where the equality holds only for x = p. Indeed, the set
⇧✓ \ C consists of line segments passing through p with two ends in @⇧✓ \ C.
Pick up any one such line segment with two points x1 , x2 2 @⇧✓ \ C such
that the line segment is given by x(s) = sx1 + (1 s)x2 for s 2 [0, 1] and
x(s⇤ ) = p for some s⇤ 2 (0, 1). We need only show that (T(x(s))) > 0
 26 Stephen Baigent, Zhanyuan Hou

h(s)
for s 2 [0, 1] with s 6= s⇤ . To this end, we write (T(x(s))) = H(s) where
Q3
H(s) = i=1 (1 + (Ax(s))i ) > 0 and
h(s) = ✓H(s) b(x(s))1 (1 + (Ax(s))2 )(1 + (Ax(s))3 )
b(x(s))2 (1 + (Ax(s))1 )(1 + (Ax(s))3 )
b(x(s))3 (1 + (Ax(s))1 )(1 + (Ax(s))2 ).
Then h(s) is a polynomial of degree at most three. As h(s⇤ ) = ✓b3 b3 kpk1 = 0,
h0 (s⇤ ) = (✓ 2p1 )b2 (1, 1, 1)A(x1 x2 ) b3 (1, 1, 1)(x1 x2 )
2 3
= [b p1 (↵ + + 1) b ](kx1 k1 kx2 k1 )
= [b2 (b 1) b3 ](✓ ✓) = 0,
and h(0) > 0 and h(1) > 0 by (T(x)) > 0 for x 2 @⇧✓ \ C, s = s⇤ is a
repeated root of h so h(s) > 0 for s 2 [0, 1] with s 6= s⇤ .
The first step is to find conditions for which (u) > 0 for u 2 @ 3. For
example, when u3 = 0 we have
b✓u1 b✓u2
(u1 , u2 , 0) = ✓ .
Cu1 + Au2 Bu1 + Cu2
b✓
At the vertex u = (1, 0, 0)T , (u) = ✓ C > 0 if b < C. Similarly when
T b✓
u = (0, 1, 0) , (u) = ✓ C > 0 when b < C. Thus b < C is a necessary
condition for (u) > 0 on @ 3 . On a boundary line u3 = 0 when u2 > 0, we
may introduce X = u1 /u2 and consider the function
b✓X b✓
p(X) = ✓ .
CX + A BX + C
Lemma 2 For a1 , a2 , a3 , b1 , b2 , b4 , c1 > 0,
a1 X b1
c1 > 0, 8X 0, (13)
a2 X + a3 b2 X + b3
when a1  a2 c1 , b1 < b3 c1 and either (i) or (ii) below holds:
(i) (c1 a2 b3 + c1 a3 b2 a 1 b3 b1 a2 )2 < 4a3 b2 (c1 a2 a1 )(c1 b3 b1 ),
(ii) a2 b1 + a1 b3 a 3 b2 c 1 a2 b3 c1  0.

Proof The expression on the left hand side of (13) can be written as
X 2 (a2 b2 c1 a1 b2 ) + X(a3 b2 c1 + a2 b3 c1 a2 b1 a 1 b3 ) + a 3 b3 c 1 a 3 b1
.
(a2 X + a3 )(b2 X + b3 )
This is positive when
X 2 (a2 b2 c1 a1 b2 ) + X(a3 b2 c1 + a2 b3 c1 a 2 b1 a 1 b3 ) + a 3 b 3 c 1 a3 b1 > 0.
From the cases X = 0 and X > 0 large we require a1  a2 c1 and b1 < b3 c1 . If
(ii) holds then the coefficients of X and X 2 are nonnegative so the inequality
holds for X 0. Under (i) the polynomial has a positive minimum value.
Global stability of discrete-time competitive population models 27

Applying lemma 2 with a1 = b, a2 = C, a3 = A, b1 = b, b2 = B, b3 = C, c1 = 1,

we require b < C and either 0 < 4AB(C b)2 (C 2 + AB 2bC)2 = (C 2
AB)(AB (C 2b)2 ) = (C 2 AB)(4b(C b) (C 2 AB)) or 2bC C 2 AB  0
for > 0 on the boundary line u3 = 0 of 3 . The same conditions are needed
on the other boundary lines. Note that 2bC AB C 2 < C 2 AB since b < C.
Thus, C 2 AB  0 implies 2bC AB C 2 < 0. In summary, we have

Lemma 3 The function (T(x)) > 0 on the boundary of ⇧✓ \ C if and only

if C > b and at least one of the following inequalities holds:

(C 2 AB)(AB (C 2b)2 ) > 0, (14)

2
2bC C AB  0. (15)

3(b 1)
First consider the condition C > b. This translates to 1+✓0 > b. Thus 1+↵+ >
b 1 which requires
↵ + < 2.
Now

C2 AB = ✓0 ((2 ↵ ) + (1 ↵ )✓0 )) ,
2 2
AB (C 2b) = (b 1)(↵ + + 1) [3 ↵2 + 2
↵ ↵ +1
2 2
b 4↵ + 4 4↵ 4 ↵ + 1 ],
2 2
2bC C AB = (b 1)(↵ + + 1) [3b(1 2↵ 2 + 3↵ )
5(1 ↵ ) + ↵2 + 2
7↵ ].

Thus when ↵ + < 2 we have ↵ < 1 and so C 2 > AB. Hence when ↵ + <2
and

b 4↵2 + 4 2
4↵ 4 ↵ + 1 < 3 ↵2 + 2
↵ ↵ +1

inequality (14) of lemma 3 is satisfied. Similarly when ↵ + < 2 and

3b(1 2↵ 2 + 3↵ ) 5(1 ↵ ) + 7↵ ↵2 2

inequality (15) of lemma 3 is satisfied.

This establishes theorem 1 given in section 1. Figure 4 illustrates the region in
(b, ↵, ) space satisfying the conditions of theorem 1 where the interior fixed
point of Leslie-Gower model (12) is globally stable.

Example 4 3-species May-Leonard May-Oster models

The Competitive May-Oster Model is an analogue of the continuous time May-
Leonard system for the May-Oster model. The model has been studied by
Hofbauer, Hutson and Jansen [20] particularly in the context of permanence,
28 Stephen Baigent, Zhanyuan Hou

Fig. 4 Region in (b, ↵, ) space satisfying the conditions of theorem 1 where the interior
fixed point of Leslie-Gower model (12) is globally stable

and also by Roeger in terms of local asymptotic stability, limit cycles and
heteroclinic cycles [32]. The map is defined by
0 1
1↵
Ti (x) = xi exp r(1 (Ax)i ), i 2 I3 , A = @ 1 ↵ A . (16)
↵ 1

There is a unique interior fixed point p that satisfies Ap = 11 = (1, 1, 1)T , i.e.
1
p = 1+↵+ 11, if (↵, ) 6= (1, 1). It is clear that all orbits are bounded.
We easily compute

DT(x) = diag[er r(Ax)1

, er r(Ax)2
, er r(Ax)3
](I rD[x]A).

At x = p we have
r
DT(p) = I r diag(p)A = I A.
1+↵+
Let be an eigenvalue of A and ✓0 = 3(1 + ↵ + ) 1 . It is easy to see that
A has a positive eigenvector ↵T = (1, 1, 1) and the corresponding eigenvalue
is = 1 + ↵ + . Hence DT(p) has a positive (left) eigenvector ↵T = (1, 1, 1)
associated with eigenvalue µ = 1 r.
Clearly, det DT(x) = 0 if and only if det(I D[rx]A) = 0. If the matrix
norm kD[rx]Ak1 < 1 then det(I D[rx]A) 6= 0 and DT(x) 1 0, so that
T is strongly competitive. Indeed in kxk1 < 1/r we have kD[rx]Ak1 < 1
Global stability of discrete-time competitive population models 29

provided that ↵, < 1. This shows that T(x) is competitive and injective on
{x 2 C : 0  kxk1 < 1/r} when ↵, < 1.
P3
The function '(x) = ↵T D[p] ln f (x) becomes '(x) = ↵+r +1 j=1 (1 (Ax)j ) =
r↵T A(p x)
↵+ +1 = r(✓0 kxk1 ). For simplicity, we drop the factor r from the ex-
pression of ' without loss of generality. Then the sets D+ , D0 , D are defined
by ' > 0, ' = 0, ' < 0 respectively.
If we can find appropriate conditions such that '(T(x)) > 0 for all x 2 D0 \{p}
and T has no invariant set in @C \ D , then theorem 7 takes e↵ect so p is
globally attracting. The global asymptotic stability of p then follows from a
simple check that the eigenvalues of DT(p) satisfy |µ| < 1.
That T has no invariant set in @C \ D follows from the analysis of two-
dimensional May-Oster model in example 1 in section 6.
The three eigenvalues of DT(p) are
p
2(↵ + + 1) r(2 ↵ ) 3r|↵ |
µ1 = 1 r, µ2,3 = ±i .
2(↵ + + 1) 2(↵ + + 1)
Then 0 < µ1 < 1 since r 2 (0, 1). Clearly, |µ2,3 |2 < 1 if and only if
4(↵ + + 1)(2 ↵ )
r< .
(2 ↵ )2 + 3(↵ )2
But the numerator minus denominator gives 8↵(1 ↵)+8 (1 )+4(1 ↵ ) >
0, so we require
4(↵ + + 1)(2 ↵ )
r<1<
(2 ↵ )2 + 3(↵ )2
for asymptotic stability of p.
Our task is now to derive conditions for '(T(x)) > 0 for all x 2 D0 \ {p}.
From '(x) = ✓0 x1 x2 x3 we have

'(T(x)) = ✓0 x1 er r(Ax)1
x2 er r(Ax)2 x3 er r(Ax)3
x1 x2 x3
= ✓0
e r+r(Ax)1 e r+r(Ax)2 e r+r(Ax)3
x1 x2 x3
✓0 .
1 r + r(Ax)1 1 r + r(Ax)2 1 r + r(Ax)3
Pn xi
Set %(x) = ✓0 i=1 1 r+r(Ax)i . Then '(T(x)) > 0 if we can show that
%(x) > 0 for x 2 D0 \ {p}. Note that by introducing b = 1, C = r✓0 + 1 r,
A = ↵r✓0 + 1 r and B = r✓0 + 1 r, we may reduce % in u coordinates to
the form of in the previous example.
To apply lemma 3 to %, we need to translate the conditions of inequalities (14)
and (15) in terms of r, ↵, . The inequality C > b translates to r✓0 + 1 r >
1 which simplifies to ↵ + < 2 as in the previous example. Similarly we
find that it is always the case that C 2 > AB and so that in addition to
30 Stephen Baigent, Zhanyuan Hou

↵+ < 2 we need either AB (C 2)2 > 0 or AB 2C + C 2 0. But

2 2
r ( 2 (↵+ ) +↵+ 3r ( ↵ ↵( +1)+( 1) +1 ))
AB (C 2)2 = (↵+ +1)2 and AB 2C +
2
r (r ( ↵ ↵(5 7 ) ( +5)+5) (↵+ +1)(2 ↵ ))
C2 = (↵+ +1)2 . Hence we conclude that
the May-Oster model has a globally attracting interior fixed point when either
3r(1 + ↵2 + 2
↵ ↵ ) < (2 ↵ )(1 + ↵ + )
or
r(5 ↵2 2
5(↵ + ) + 7↵ ) (2 ↵ )(1 + ↵ + ).

Figure 5 shows the regions in (r, ↵, ) space where global stability for the
May-Oster model is obtained.

Fig. 5 Region in (r, ↵, ) space (a) satisfying the conditions of theorem 2 where the May-
Oster model (16) is globally stable.

8 Conclusions and discussion

Here we have introduced a method for identifying global stability of interior

fixed points of competitive, discrete-time, population models. In its current
form, our method is restricted to models under parameter ranges where there
are no periodic orbits, but is evidently applicable to a wide range of standard
population models. We have attempted to elucidate our method in terms of
ecological concepts, and in particular have introduced the idea of a principal
reproductive mode, which is similar to Fisher’s reproductive value, but not
bound to age-structured models. We have also introduced the notion of the
Global stability of discrete-time competitive population models 31

principal component of reproductive rate which measures the net reproductive

growth using the principal reproductive mode. The eventual positiveness of the
principal component of reproductive rate is diagnostic of global convergence
to a steady coexistence state. We have also discussed our method in relation
to standard approaches to system permanence.
We have provided basic theorems for testing global attraction and stability
of an interior fixed point in the first orthant, which is a generalisation of
the split Lyapunov method from continuous dynamical systems to discrete
systems. These results have then been refined to incorporate the use of a
carrying simplex, an attracting invariant manifold of codimension one which
is a common feature of competitive population models. We have linked the
principal reproductive mode to the normal direction of the carrying simplex
at the interior fixed point.
The ideas developed in the paper have been demonstrated to work with some
well-known two and three dimensional models. For planar models, the well-
known competitive systems approach is simpler than our method, but for 3
species systems or richer, our method performs well. In particular, we have
extended results on local stability of some May-Leonard models to global re-
sults.
However, there are some issues that we are unable to deal with in this paper
so they need further investigation.
Problem 1 Theorem 10 for global stability requires conditions in addition
to those for global attraction. Are these extra conditions for stability really
necessary? Is stability also guaranteed by the conditions for global attraction
in theorems 7, 8 or 9?
Problem 2 Theorem 5 provides an alternative rule for positive invariance
to Sacker’s rule (theorem 4). Based on this criteria for global attraction and
stability are proved (theorems 8 and 10). Applications of these criteria to
concrete models need to be fulfilled.
Problem 3 For continuous systems possessing a carrying simplex, split Lya-
punov method is also applicable to global repulsion of a fixed point on the
carrying simplex. Is the generalisation of the method from continuous system
to discrete system also possible for global repulsion on the carrying simplex?
Problem 4 Further development of the method for a boundary fixed point of
discrete systems is needed.

References

1. Ackleh, A.S., Salceanu, P.L.: Competitive exclusion and coexistence in an n-species

Ricker model. Journal of Biological Dynamics 9(sup1), 321–331 (2014)
2. Baigent, S.: Convexity of the carrying simplex for discrete-time planar competitive Kol-
mogorov systems. Journal of Di↵erence Equations and Applications (Under Revision)
32 Stephen Baigent, Zhanyuan Hou

3. Baigent, S.: Convexity-preserving flows of totally competitive planar Lotka–Volterra

equations and the geometry of the carrying simplex. Proceedings of the Edinburgh
Mathematical Society 55, 53–63 (2012)
4. Baigent, S., Hou, Z.: Global Stability of Interior and Boundary Fixed Points for
Lotka–Volterra Systems. Di↵erential Equations and Dynamical Systems 20(1), 53–66
(2012)
5. Cushing, J.M., Levarge, S.: Some Discrete Competition Models and the Principle of
Competitive Exclusion. Di↵erence Equations and Discrete Dynamical Systems, Pro-
ceedings of the Ninth International Conference pp. 1–20 (2006)
6. Cushing, J.M., Levarge, S., Chitnis, N., Henson, S.M.: Some Discrete Competition Mod-
els and the Competitive Exclusion Principle. Journal of Di↵erence Equations and Ap-
plications 10(13-15), 1139–1151 (2004)
7. Cushing, J.M., Zhou, Y.: The Net Reproductive Value And Stability in Matrix Popu-
lation Models. Natural Resource Modelling 8(4), 297–333 (1994)
8. Diekmann, O., Wang, Y., Yan, P.: Carrying Simplices in Discrete Competitive Systems
and Age-structured Semelparous Populations. Discrete and Continuous Dynamical Sys-
tems 20(1), 37–52 (2008)
9. Fisher, M.E., Goh, B.S.: Stability in a class of discrete time models of interacting pop-
ulations. Journal of Mathematical Biology 4(3), 265–274 (1977)
10. Franke, J.E., Yakubu, A.A.: Global attractors in competitive systems. Nonlinear Anal-
ysis, Theory, Methods and Applications 16(2), 111–129 (1991)
11. Franke, J.E., Yakubu, A.A.: Mutual exclusion versus coexistence for discrete competitive
systems. Journal of Mathematical Biology 30, 161–168 (1991)
12. Franke, J.E., Yakubu, A.A.: Geometry of Exclusion Principles in Discrete Systems.
Journal of Mathematical Analysis and Applications 168, 385–400 (1992)
13. Franke, J.E., Yakubu, A.A.: Species extinction using geometry of level surfaces. Non-
linear Analysis, Theory, Methods and Applications 21(5), 369–378 (1993)
14. Franke, J.E., Yakubu, A.A.: Exclusion principles for density-dependent discrete pioneer-
climax models. Journal of Mathematical Analysis and Applications 187, 1019–1046
(1994)
15. Garay, B.M., Hofbauer, J.: Robust Permanence for Ecological Di↵erential Equations,
Minimax, and Discretizations. SIAM Journal on Mathematical Analysis 34(5), 1007–
1039 (2003)
16. Goh, B.S.: Global Stability in Many-Species Systems. The American Naturalist
111(977), 135–143 (1977)
17. Hirsch, M.W.: Systems of di↵erential equations which are competitive or cooperative:
III Competing species. Nonlinearity 1, 51–71 (1988)
18. Hirsch, M.W.: On existence and uniqueness of the carrying simplex for competitive
dynamical systems. Journal of Biological Dynamics 2(2), 169–179 (2008)
19. Hofbauer, J.: Saturated equilibria, permanence, and stability for ecological systems.
In: Mathematical ecology (Trieste, 1986), pp. 625–642. World Sci. Publ., Teaneck, NJ
(1988)
20. Hofbauer, J., Hutson, V., Jansen, W.: Coexistence for systems governed by di↵erence
equations of Lotka-Volterra type. Journal of Mathematical Biology 25, 553–570 (1987)
21. Hofbauer, J., Schreibner, S.J.: Robust permanence for interacting structured popula-
tions. Journal of Di↵erential Equations 248(8), 1955–1971 (2010)
22. Hou, Z., Baigent, S.: Fixed point global attractors and repellors in competitive
Lotka–Volterra systems. Dynamical Systems 26(4), 367–390 (2011)
23. Hou, Z., Baigent, S.: Global stability and repulsion in autonomous Kolmogorov systems.
Communications on Pure and Applied Analysis 14(3), 1205–1238 (2015)
24. Hutson, V., Moran, W.: Persistence of Species Obeying Di↵erence Equations. Journal
of Mathematical Biology 15, 203–213 (1982)
25. Jiang, H., Rogers, T.D.: The discrete dynamics of symmetric competition in the plane.
Journal of Mathematical Biology pp. 1–24 (2004)
26. Jiang, J., Mierczyński, J., Wang, Y.: Smoothness of the carrying simplex for discrete-
time competitive dynamical systems: A characterization of neat embedding. Journal of
Di↵erential Equations 246(4), 1623–1672 (2009)
27. Jiang, J., Niu, L., Wang, Y.: On heteroclinic cycles of competitive maps via carrying
simplices. Journal of Mathematical Biology pp. 1–34 (2015)
Global stability of discrete-time competitive population models 33

28. Jiang, J., Wang, Y.: Uniqueness and attractivity of the carrying simplex for discrete-time
competitive dynamical systems. Journal of Di↵erential Equations 186, 1–22 (2002)
29. Kon, R.: Permanence of discrete-time Kolmogorov systems for two species and saturated
fixed points. Journal of Mathematical Biology 48(1), 57–81 (2004)
30. Kon, R.: Convex Dominates Concave: An Exclusion Principle in Discrete-Time Kol-
mogorov Systems. Proceedings of the American Mathematical Society 134(10), 3025–
3034 (2006)
31. Kon, R., Takeuchi, Y.: E↵ect of a Parasitoid on Permanence of Competing Hosts. Viet-
nam Journal of Mathematics 30, 473–486 (2002)
32. Roeger, L.I.W.: Discrete May–Leonard Competition Models III. Journal of Di↵erence
Equations and Applications 10(8), 773–790 (2004)
33. Roeger, L.I.W.: Discrete May-Leonard Competition Models II. Discrete and Continuous
Dynamical Systems B 6(3), 841–860 (2005)
34. Roeger, L.I.W., Allen, L.J.S.: Discrete May–Leonard Competition Models I. Journal of
Di↵erence Equations and Applications 10(1), 77–98 (2004)
35. Ruiz-Herrera, A.: Exclusion and dominance in discrete population models via the car-
rying simplex. Journal of Di↵erence Equations and Applications pp. 1–18 (2011)
36. Sacker, R.J.: A Note: An Invariance Theorem for Mappings. Journal of Di↵erence
Equations and Applications pp. 1–4 (2010)
37. Sacker, R.J.: Global stability in a multi-species periodic Leslie–Gower model. Journal
of Biological Dynamics 5(5), 549–562 (2011)
38. Sacker, R.J., von Bremen, H.F.: Dynamic reduction with applications to mathematical
biology and other areas. Journal of Biological Dynamics 1(4), 437–453 (2007)
39. Samuelson, P.A.: Generalizing Fisher’s ”reproductive value”: linear di↵erential and
di↵erence equations of ”dilute” biological systems. In: Proceedings of the National
Academy of Sciences, pp. 5189–5192 (1977)
40. Schreibner, S.J.: Persistence for stochastic di↵erence equations: a mini-review. Journal
of Di↵erence Equations and Applications 18(8), 1381–1403 (2012)
41. Schuster, P., Sigmund, K., Wol↵, R.: Dynamical systems under constant organization.
III. Cooperative and competitive behavior of hypercycles. Journal of Di↵erential Equa-
tions 32(3), 357–368 (1979)
42. Smith, H.L.: Planar competitive and cooperative di↵erence equations. Journal of Dif-
ference Equations and Applications 3(5-6), 335–357 (1998)
43. Takeuchi, Y.: Global dynamical properties of Lotka-Volterra systems. World Scientific
Publishing Co. Pte. Ltd. (1996)
44. Tineo, A.: On the convexity of the carrying simplex of planar Lotka-Volterra competitive
systems. Applied Mathematics and Computation pp. 1–16 (2001)
45. Wang, Y., Jiang, J.: The General Properties of Discrete-Time Competitive Dynamical
Systems. Journal of Di↵erential Equations 176(2), 470–493 (2001)
46. Wendi, W., Zhengyi, L.: Global stability of discrete models of Lotka–Volterra type.
Nonlinear Analysis, Theory, Methods and Applications 35(8), 1019–1030 (1999)
47. Zeeman, E.C., Zeeman, M.L.: From local to global behavior in competitive Lotka-
Volterra systems. Trans Amer Math Soc 355(2), 713–734 (2003)