Stability of The Replicator Equation For A Single Species With A Multi-Dimensional Continuous Trait Space

ARTICLE IN PRESS
Journal of Theoretical Biology 239 (2006) 273–288

www.elsevier.com/locate/yjtbi
Stability of the replicator equation for a single species with a

multi-dimensional continuous trait space
Ross Cressmana,,1, Josef Hofbauerb, Frank Riedelc
a
Department of Mathematics, Wilfrid Laurier University, Waterloo, Ontario, N2L 3C5, Canada
b
Department of Mathematics, University College London, London WC1E 6BT, UK
c
Department of Economics, University of Bonn, Adenauerallee 24, 53113 Bonn, Germany
Received 13 April 2005; received in revised form 27 July 2005; accepted 28 July 2005
Available online 24 October 2005
Dedicated to the memory of John Maynard Smith
Abstract
The replicator equation model for the evolution of individual behaviors in a single species with a multi-dimensional continuous trait
space is developed as a dynamics on the set of probability measures. Stability of monomorphisms in this model using the weak topology
is compared to more traditional methods of adaptive dynamics. For quadratic fitness functions and initial normal trait distributions, it is
shown that the multi-dimensional continuously stable strategy (CSS) of adaptive dynamics is often relevant for predicting stability of the
measure-theoretic model but may be too strong in general. For general fitness functions and trait distributions, the CSS is related to
dominance solvability which can be used to characterize local stability for a large class of trait distributions that have no gaps in their
supports whereas the stronger neighborhood invader strategy (NIS) concept is needed if the supports are arbitrary.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: Adaptive dynamics; CSS; NIS; Replicator equation; Neighborhood superiority; Strategy dominance; Measure dynamics; Weak topology
1. Introduction encounters that determine their payoffs (i.e. their fitness or

reproductive success). A primary objective of these
Dynamical systems on the set of probability measures theoretical models is the characterization, in terms of static
over a continuous trait space have been developed as one payoff/fitness comparisons, of static conditions (and
means to predict the evolution and stability of distributions distributions that satisfy the conditions) that imply
of individual behaviors in a biological species (Bomze, dynamic stability. Such conditions then allow practitioners
1990, 1991; Oechssler and Riedel, 2001, 2002). These of evolutionary game theory to describe the outcome of the
systems generalize the well-known replicator equation evolutionary process without a detailed analysis of the
approach of dynamic evolutionary game theory (Hofbauer underlying dynamical system.
and Sigmund, 1998; Cressman, 2003) when the trait space For instance, when there are n pure strategies (so trait
is finite (i.e. when there are a finite number of pure space is finite) and there are random pairwise encounters
strategies) and individuals interact in random pairwise that contribute additively to fitness, the static equilibrium
concept of an evolutionarily stable strategy defined by
Corresponding author. Tel.: +519 884 1970; fax: +529 884 9738. Maynard Smith (1982) (i.e. a strategy for which, whenever
E-mail addresses: rcressma@wlu.ca (R. Cressman), all individuals use this strategy, the population cannot be
j.hofbauer@ucl.ac.uk (J. Hofbauer), friedel@uni-bonn.de (F. Riedel). invaded by a rare mutant under the influence of natural
1
This research was initiated while R.C. was a Fellow at the Collegium selection) has been a huge success since the resultant
Budapest. Partial support from the German Science Foundation heuristic static conditions have a clear biological basis. We
(Transregio ‘‘Governance and the Efficiency of Economic Systems’’),
will refer to this case as the matrix game model (Meszéna et
from UCL’s Centre for Economic Learning and Social Evolution (ELSE)
and from a NSERC of Canada Individual Discovery Grant is also al., 2001) and the equilibrium concept as a matrix-ESS
acknowledged. (Vincent and Cressman, 2000; Cressman and Hofbauer,
0022-5193/$ - see front matter r 2005 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jtbi.2005.07.022
ARTICLE IN PRESS
274 R. Cressman et al. / Journal of Theoretical Biology 239 (2006) 273–288
2005) since payoffs are given through an n n payoff particular interest in proving these results is the technique
matrix.2 of iterated elimination of strictly dominated strategies that
For continuous trait spaces, an alternative means to is borrowed from classical game theory and also used in the
predict the evolutionary outcome is the adaptive dynamics stability analysis of evolutionary game theory applied to
method that has generated an enormous literature (see matrix games (e.g. Samuelson and Zhang, 1992) but seems
Abrams, 2001 and the references therein) since the phrase not to have been used previously in the adaptive dynamics
was introduced by Hofbauer and Sigmund (1990). This literature.
method is especially useful when the resident biological The main purpose of this paper is then to examine
species is monomorphic (i.e. when all individuals in the whether static extensions of the one-dimensional adaptive
population are using the same strategy) and there is a one- dynamics concepts continue their relevance for the measure
dimensional continuous trait space. Here, adaptive dy- dynamic model when the trait space is multi-dimensional.
namics predicts stability of a monomorphic equilibrium if, We begin in Section 2 by briefly developing the replicator
for all other monomorphisms that are small perturbations equation on a continuous trait space and introducing
of this equilibrium, trait substitution through nearby essential notation used throughout the paper concerning
mutations is only successful when this substitution moves this measure dynamics and the underlying fitness functions.
the population closer to the equilibrium. Mathematically, Sections 3 and 4 provide valuable insight by fully analysing
the adaptive dynamics of mutation and trait substitution is the replicator equation in the special case of quadratic
modeled here by the canonical equation (Marrow et al., fitness functions and normal distributions (Section 3) and
1996), a one-dimensional dynamical system (see Section 4) then relating these results to potential static extensions of
whose stable equilibria are characterized by the static the CSS concept to multi-dimensional adaptive dynamics
convergence stable conditions of Christiansen (1991) (also in Section 4 (see Theorems 5 and 6 there). Section 5
called the m-stability concept in Taylor, 1989). Combined considers the general case of arbitrary fitness functions and
with a further condition (often called the ESS or probability distributions in the multi-dimensional setting.
uninvadability condition) that guards against the mono- Unfortunately, our results here do not give as thorough a
morphism being successfully invaded by an evolving static characterization of stability for the replicator
dimorphism through a process now referred to as evolu- equation as that available through the one-dimensional
tionary branching (Doebeli and Dieckmann, 2000), we analysis of Cressman and Hofbauer (2005). Although we
obtain the solution concept of continuously stable strategy are able to obtain necessary conditions related to adaptive
(CSS) introduced by Eshel (1983). dynamics for stability of monomorphisms in large classes
From our perspective, adaptive dynamics and its of measure dynamic models as well as sufficient conditions
canonical equation are approximate descriptions of how for others, an exhaustive classification is beyond the reach
the mean of the distribution of individual behaviors evolves of our current techniques. The final section discusses these
and do not adequately model the spread of the distribution. shortcomings as it summarizes the positive aspects of our
In this paper, we use the replicator equation with a classification.
continuous trait space to model the evolution of the
probability distribution (i.e. probability measure) of
individual behaviors. The CSS and/or convergence stability 2. The replicator dynamics on the space of probability
conditions are then heuristic tools that at best can suggest measures
when the distribution will evolve to a monomorphism (i.e.
to a Dirac delta distribution in measure theoretic terminol- The probability measure dynamics is the extension of the
ogy). In fact, Cressman and Hofbauer (2005) have shown replicator equation originally defined for matrix games
the relevance of the CSS concept (and the closely related with a finite trait space (Taylor and Jonker, 1978). In
concept of a neighborhood invader strategy (NIS) of general, individuals are assumed to play a strategy s in a
Apaloo, 1997) for stability of monomorphisms in the fixed trait space S and the population state is given by a
measure dynamics of a one-dimensional continuous trait probability measure P on a measurable space ðS; BÞ. If
space. Specifically, a non-CSS monomorphism is unstable A 2 B, PðAÞ is interpreted as the proportion of individuals
in the measure dynamic and, conversely, a CSS is in the population who are using strategies in the set A.
dynamically stable if the initial distribution of individual For our model of a multi-dimensional continuous trait
behaviors is close to the CSS and satisfies an additional space, S will be a Borel subset of Rn and B will be the Borel
technical requirement concerning the strategies present in subsets of S (i.e. the s-algebra of the Borel sets of Rn
the population (i.e. the support of this distribution).3 Of intersected with S and so P is a Borel measure).4 Let DðSÞ
denote the set of probability measures with respect to
2
The matrix-ESS terminology will help avoid confusion with the term ðS; BÞ. Since P is a Borel measure, there is a unique
ESS as it has been used in the literature with continuous trait spaces.
3 4
See Section 5 for further details of this technical requirement that In fact, S will typically have further topological properties such as
successful mutant monomorphisms in the adaptive dynamics approach are being convex and open (or closed with nonempty interior). The Borel
available for trait substitution. Without this requirement, the stronger NIS subsets of Rn form the smallest s-algebra containing the open subsets of
conditions are needed to guarantee dynamic stability. Rn .
ARTICLE IN PRESS
R. Cressman et al. / Journal of Theoretical Biology 239 (2006) 273–288 275
(relatively) closed subset of S, called the support of P, such with finite support fx1 ; . . . ; xm g, we can take -neighbor-
that the measure of its complement is 0 but every open set hoods in the weak topology to be of the form
that intersects it has positive measure (Royden, 1988).
fQ 2 DðSÞ : jQðB ðxi ÞÞ P ðfxi gÞjo 8i ¼ 1; . . . ; mg,
The measure dynamics on DðSÞ (see Eq. (1) below) is
given in terms of the fitness (or expected payoff) pðs; PÞ ¼ where B ðxÞ is the open ball of radius centered at x. In
pðds ; PÞ of an individual using strategy s when the particular, two monomorphisms dx1 and dx2 are within of
population is in state P. Here, for a given s 2 S, ds denotes each other if and only if the Euclidean distance between
the Dirac delta measure that assigns unit mass to fsg. We these points is less than . In the following all topological
will assume throughout that the fitness pðs; PÞ is given notions in DðSÞ are taken for this weak topology, unless
through a continuousR real-valued payoff function p : S otherwise stated.
S!R by pðs; PÞ ¼ S pðs; yÞPðdyÞ. In particular, matrix
games that assume random pairwise interactions and a 2.1. The fitness function pðs; yÞ
finite trait space S may be put in this form.5 The mean
payoff to a random Rindividual in the population with state For the multi-dimensional continuous trait space, we
P is then pðP; PÞ S pðds ; PÞPðdsÞ. assume S is the closure of an open connected subset of Rn
We assume the replicator equation Eq. (1) describes how that contains the origin in its interior. In fact, we often
the population state evolves (i.e. its solutions define assume S is star-shaped with respect to the origin (i.e. if
trajectories Pt in DðSÞ). x 2 S, then so does the line segment joining 0 to x). We are
Z particularly interested in the stability of the monomorph-
dP
ðAÞ ¼ ðpðds ; PÞ pðP; PÞÞPðdsÞ. (1) ism d0 . To this end, consider the Taylor expansion of pðx; zÞ
dt A
about ð0; 0Þ 2 R2n
Heuristically, this dynamic increases the probability of
those sets of strategies A that have a higher expected payoff pðx; zÞ ¼ pð0; 0Þ þ r1 p x þ r2 p z þ 12 ½x ðr211 pÞx
than the mean payoff to a random individual in the þ 2x ðr212 pÞz þ z ðr222 pÞz þ h:o:t:;
population. It has been shown (Oechssler and Riedel, 2001)
that there is a unique solution that satisfies this dynamics where for i 2 f1; 2g, ri p is the gradient vector of p at the
for all positive t given any initial probability measure P0 origin in the ith variable (e.g. ðr1 pÞk ¼ ðqpðx; zÞ=qxk Þ
with compact support when pðs; yÞ is continuous.6 Here A jðx;zÞ¼ð0;0Þ Þ) and r2ij p is the n n matrix with entries the
is a Borel subset of S and dP=dt at time t is defined to be appropriate second-order partials.
limh!0 ðPtþh Pt Þ=h with respect to the variational norm Each monomorphism is a rest point of Eq. (1). Their
(i.e. limh!0 kdP=dtðAÞ ðPtþh Pt Þ=hðAÞk ¼ 0 where k k stability in the weak topology requires the monomorphism be
is the variational norm as in Oechssler and Riedel, 2001). a NE of the payoff function restricted to the game with
Furthermore, the support of Pt is the same as for P0 for all nearby strategies (e.g. 0 is such a NE if pðx; 0Þppð0; 0Þ for all
tX0. A population state P is an equilibrium of Eq. (1) if x sufficiently close to 0). Since 0 is an interior point, this
and only if pðds ; P Þ pðP ; P Þ ¼ 0 for all s 2 suppðP Þ. implies the gradient r1 p is the zero vector and x ðr211 pÞxp0
Our primary aim in this paper is the investigation of the for all x 2 Rn . In fact, we will assume the symmetric Hessian
convergence and stability properties (especially related to matrix r211 p is negative definite throughout to avoid technical
monomorphic equilibrium population states ds ) of the issues. That is, we assume 0 is a strict NE of the restricted
measure dynamics (1). Heuristically, dynamic stability of game.7 Since r211 p is symmetric, we can diagonalize it by an
P refers to the question whether Pt stays close and/or orthogonal transformation and then all diagonal entries d k
evolves to P if the initial P0 is chosen appropriately in are negative. Furthermore,
pffiffiffiffiffi a change of variables (that
DðSÞ. From Oechssler and Riedel (2002), it is clear that the replaces xk with d k xk and takes the payoffs with respect
answers to the stability question depend critically on the to these new variables) allows us to assume r211 p ¼ 2I
concept of closeness of probability measures (i.e. on the where I is the n n identity matrix. Without loss of
topology used for the space of Borel probability measures), generality, the fitness function can then be written in these
when the trait space is not a finite subset of Rn . new coordinates as
We feel the weak topology captures best the essence of pðx; zÞ ¼ pð0; 0Þ þ r2 p z x x þ x Bz
evolutionary convergence in our biological systems. This
1
topology will mostly be applied to neighborhoods of þ z ðr222 pÞz þ h:o:t. ð2Þ
2
monomorphic P . In general, for a probability measure P
(i.e. r211 p ¼ 2I and r212 p ¼ B).
5
In fact, Bomze and Pötscher (1989) argue that the existence of such a
7
pðs; yÞ for an arbitrary trait space S means the evolutionary game can be The condition pðx; 0Þopð0; 0Þ is related to the concepts of uninvad-
interpreted as being based on pairwise interactions. It is only the form of ability and evolutionarily stable strategy (ESS) as used in adaptive
pðs; PÞ that is important to us, not whether players are competing pairwise. dynamics (Marrow et al., 1996; Vincent et al., 1996). We especially avoid
6
If pðs; yÞ is not continuous or P0 does not have compact support, one this latter terminology since the ESS description is overused in the
must be careful that the desired integrals are defined. The latter concern is literature and may have different interpretations for different readers. On
discussed further for the normal distributions of Section 3. the other hand, strict NE seems to have a universally accepted meaning.
ARTICLE IN PRESS
In terms of the Taylor expansion, the replicator equation Theorem 2. For any initial positive semidefinite matrix Cð0Þ,
(1) can then be rewritten as the solution of Eq. (7) is given by
Z Z Z
dP CðtÞ ¼ Cð0ÞðI þ 2Cð0ÞtÞ1 . (8)
ðAÞ ¼ ðpðx; zÞ pðy; zÞÞPðdxÞPðdyÞPðdzÞ, (3)
dt S S A
Note that Eq. (8) is well defined for all tX0.
where the integrand is given by Alternatively, one can write the solution in the following
pðx; zÞ pðy; zÞ ¼ ðy xÞ ½x þ y Bz þ h:o:t:. (4) way.
In particular, the constant and linear terms in the fitness CðtÞ ¼ OT DðtÞO,
function are irrelevant for the dynamic analysis.
where O is an orthogonal matrix such that OCð0ÞOT ¼ D
for some diagonal matrix D (and OT denotes the transpose
3. The replicator equation with normal distributions and
of O) and DðtÞ is the diagonal matrix with entries
quadratic fitness functions
Dii
Dii ðtÞ ¼ . (9)
In this section, we analyse the replicator equation when 1 þ 2tDii
the higher-order terms are ignored in Eq. (4) and the initial In particular, the covariance matrix CðtÞ converges to the
probability measure P0 is the (multivariate) normal zero matrix, and CðtÞ ¼ ð1=2tÞI þ Oð1=t2 Þ as t ! 1
distribution Nðm; CÞ with mean vector m 2 Rn and whenever the initial condition Cð0Þ is positive definite. By
covariance matrix C 2 Rnn . From Section 2.1, we may inserting solution (8) into Eq. (6), the mean evolves
assume fitness has the form of the quadratic function according to the time-dependent linear differential
pðx; zÞ ¼ x x þ x Bz (5) equation
n
for x; z 2 S ¼ R and B an n n matrix. dm
We proceed as follows. The first step is to show that the ¼ Cð0ÞðI þ 2Cð0ÞtÞ1 ðB 2IÞmðtÞ. (10)
dt
class of normal distributions is forward invariant under the
After changing the time-scale ðt þ 1 ¼ e2t Þ this differential
replicator equation. Therefore, the infinite-dimensional
equation becomes asymptotically autonomous
measure dynamics is reduced to a finite-dimensional system
of n þ n2 ordinary differential equations for the mean and dm
¼ ðB 2IÞmðtÞ þ RðtÞmðtÞ (11)
covariance matrix. These facts are stated in Theorem 1 dt
below where it is also apparent that the dynamics of the with exponentially decreasing remainder term
covariance matrix does not depend on the mean vector. RðtÞ ¼ ð2Cð0Þ IÞðI þ 2ðe2t 1ÞCð0ÞÞ1 ðB 2IÞ. There-
The next step is to obtain the explicit solution (Theorem 2) fore the eigenvalues of B 2I determine the asymptotic
for the evolution of the covariance matrix. Substitution of behavior of mðtÞ and imply the following theorem (see its
this solution into the dynamics for the mean results in a proof in the Appendix for further details).
system of linear differential equations with time varying
coefficients. The stability analysis of this system for the
Theorem 3. Consider the replicator equation (1) restricted to
equilibrium d0 (i.e. for the limit of the normal distributions
the class of normal distributions with quadratic fitness
Nð0; CÞ as C approaches the zero matrix) is summarized in
functions as in Eq. (5).
Theorems 3 and 4 in terms of the matrix B.
Theorem 1. The class of normal distributions is forward 1. If every eigenvalue of the matrix B 2I has negative real
invariant under the replicator dynamics (1). Assume that the part, then d0 attracts all normal distributions P0 of the
initial distribution is normal, P0 ¼ Nðm; CÞ for a mean form Nðm; CÞ with positive definite symmetric covariance
vector m 2 Rn and a covariance matrix C 2 Rnn . Then the matrix C (i.e. CðtÞ ! O and mðtÞ ! 0 as t ! þ1).
solution of the replicator dynamics starting at P0 is given by Furthermore, d0 is asymptotically stable with respect to
Pt ¼ NðmðtÞ; CðtÞÞ, where the mean and the covariance all such P0 with initial covariance matrix C orthogonally
matrix solve the initial value problem similar to a positive diagonal matrix D (i.e. OCOT ¼ D)
m0 ðtÞ ¼ CðtÞðB 2IÞmðtÞ (6) satisfying minfDii =Djj j 1pi; jpngX for some 40.
2. If some eigenvalue of B 2I has positive real part, then d0
C 0 ðtÞ ¼ 2CðtÞCðtÞ. (7) is unstable: for all positive definite C there are m
arbitrarily close to 0 such that along the solution with
with mð0Þ ¼ m and Cð0Þ ¼ C.
initial P0 ¼ Nðm; CÞ one has kmðtÞk ! 1 as t ! 1.8
The proof of Theorem 1 (and most other theorems) is in
the Appendix. Since the dynamical system (7) for the It is important to note that, in part 1 of Theorem 3, we
covariance matrix does not depend on the mean vector, we do not state that d0 is asymptotically stable with respect to
may study this system of differential equations on its own. all P0 of the form Nðm; CÞ with positive definite symmetric
The following theorem gives the explicit solution which can
8
be easily verified. kmk is the Euclidean norm of m 2 Rn .
ARTICLE IN PRESS
covariance matrix C. In fact, this asymptotic stability isons developed in this section continue to be based on a
requires stronger assumptions on B. quadratic payoff function pðx; yÞ.
To see this, we consider degenerate normal distributions The canonical equation (13) from adaptive dynamics for
Nðm; CÞ with a positive semidefinite covariance matrix C the evolution of a (monomorphic) population with mean m
with m in the range of C (i.e. Gaussian distributions through mutation and trait substitution was developed by
confined to a linear subspace of Rn ). Then Eq. (11) is Dieckmann and Law (1996). Following Meszéna et al.
replaced by (2001) (see also Leimar, 2005), this takes the form (in our
notation)
dm
¼ PðB 2IÞmðtÞ þ PRðtÞmðtÞ, (12)
dt m0 ðtÞ ¼ 12uðmÞNðmÞCðmÞr1 pðx; mÞx¼m . (13)
where P ¼ limt!1 2tCð0ÞðI þ 2tCð0ÞÞ1 .9 Then P is a Here uðmÞ and NðmÞ are positive real-valued functions
projection matrix (i.e. P ¼ PT ¼ P2 ) since its eigenvalues giving the mutation rate and the equilibrium population
are 0 and 1 only. size respectively at mean m. These can be ignored in
Note that two (possibly degenerate) normal distributions analysing the limiting behavior of the canonical equation
are close in the weak topology if and only if their means since they do not affect the evolutionary trajectory but only
and covariances are close in the Euclidean metric. Since the speed of evolution along this trajectory (and so
there are normal distributions with a positive definite 1
2 uðmÞNðmÞ will be deleted from now on). More impor-
covariance matrix arbitrarily close to these degenerate tantly, the covariance matrix CðmÞ (which now charac-
ones, stability requires every eigenvalue of PðB 2IÞ to terizes the expected mutational effects in different
have negative real part. This condition implies the directions from m and does affect its evolution) is assumed
following result on instability. to depend only on m. In the earlier formulation of adaptive
dynamics by Hofbauer and Sigmund (1990) the symmetric
Theorem 4. If B 2I is not negative semidefinite (i.e. and positive definite matrix CðmÞ comes from a Rieman-
x Bx42kxk2 for at least one x 2 Rn ) then d0 is unstable nian metric on the trait space.
for Eq. (1) restricted to normal distributions. To compare Eq. (13) to the replicator dynamics, assume
m ¼ 0 is a strict NE in the interior of the trait space as in
It is instructive to compare Theorems 3 and 4 in the case Section 2.1. Then r1 pðx; 0Þjx¼0 ¼ 0 and so m ¼ 0 is an
where every eigenvalue of the matrix B 2I has negative interior equilibrium of Eq. (13). With the same change of
real part but B 2I is not negative semidefinite (in variables that led to Eq. (2) in Section 2.1, the canonical
particular, B 2I cannot be a symmetric matrix). In the equation becomes
proof of Theorem 4 we construct rank one matrices Cð0Þ
and means mð0Þ such that mðtÞ ! 1 as t ! þ1, i.e. there m0 ðtÞ ¼ CðmðtÞÞðB 2IÞmðtÞ. (14)
is divergence for normal distributions concentrated on m ¼ 0 is called convergence stable (with respect to CðmÞ) if
certain lines through 0. By continuous dependence on it is asymptotically stable under Eq. (14).
initial conditions in Eqs. (6) and (7), there are positive A quick glance at Eq. (6) shows the canonical equation is
definite covariance matrices Cð0Þ~ and means mð0Þ~ close by quite closely related to the evolution of the mean for
~
such that kmðtÞk=k ~
mð0Þk becomes arbitrarily large for some normal distributions under Eq. (1) with quadratic fitness
t. Hence d0 is unstable even for normal distributions with functions. When CðmðtÞÞ ¼ cðtÞI for some positive function
positive definite covariance matrix. However, in this case, cðtÞ40; the two dynamical systems have the same
~ will eventually converge to 0 after this long excursion
mðtÞ trajectories for the mean although the mean evolves much
away by Part 1 of Theorem 3.1. slower under the replicator equation through the change in
This phenomenon—attractivity without stability— time-scale (given by t þ 1 ¼ e2t ) as the covariance ap-
requires a multi-dimensional trait space. proaches the zero matrix. In general, the only difference
mathematically is that Eq. (14) is an autonomous system of
differential equations whereas Eq. (6) is not. As we will see,
4. Multi-dimensional adaptive dynamics, the canonical
this difference has important consequences for multi-
equation and CSS
dimensional trait space on how convergence stability is
related to dynamic stability of Eq. (6) where the covariance
As stated in the Introduction, a main purpose of this
matrix CðtÞ is given explicitly in Theorem 2.
paper is to examine the relevance (for the dynamic stability
For a one-dimensional trait space, convergence stability
of the replicator equation with multi-dimensional contin-
is independent of the choice of CðmÞ. That is, m ¼ 0 is
uous strategy space) of static extensions of the one-
asymptotically stable with respect to the canonical Eq. (14)
dimensional stability conditions developed by adaptive
for one choice of positive variance as a function of m if and
dynamics (e.g. the CSS and NIS concepts). The compar-
only if it is for any other choice. In fact, a one-dimensional
9
Inserting Cð0Þ ¼ PCð0Þ in Eq. (10) leads to Eq. (12) instead of Eq. (11) strict NE that is convergence stable is called a Continuously
with PRðtÞ exponentially decreasing when restricted to m in the range Stable Strategy (CSS), a concept introduced by Eshel
of C. (1983). Furthermore, m ¼ 0 is a CSS if and only if d0 is
ARTICLE IN PRESS
asymptotically stable under (6).10 Unfortunately, this where

correspondence does not extend beyond one dimension as ! !
illustrated by the following example. a 14 42 m1
.
b 39 16 m2
Example 1. Consider the two-dimensional trait space R2
with quadratic fitness function (5) and Substitution into Eq. (14) yields
! ! ! !1 !
0 b m01
B¼ , a2 ab 2 1 14 42 a
c 0 ¼
m02 ab b2 1=2 2 39 16 b
where b; c are both positive. By Theorem 3, d0 is attractive 2
! ! !
under Eq. (1) for the class of normal distributions 1 a ab 1 10 a
pffiffiffiffiffiif and ¼
only if the eigenvalues of B 2I, l1;2 ¼ 2 bc, are 202 ab b2 10 1 b
both negative, i.e., bco4. Also, by Theorem 4, d0 is !
unstable for this class if B 2I is not negative semidefinite, a 2 þ b2 a
¼ .
i.e., jb þ cjo4. 202 b
Example 1a. Suppose CðmÞ is the constant symmetric
matrix Thus,
! !
1 a a0 a 2 þ b2 14 42 a
C¼ ¼
a 1 b0 202 39 16 b
for all m where a is a fixed parameter satisfying jajo1 so and so m ¼ 0 is unstable for Eq. (14) since the trace of this
that C is positive definite. Then m ¼ 0 is globally last 2 2 matrix is positive. 11 On the other hand, d0 is
asymptotically stable under Eq. (14) if and only if every globally asymptotically stable for the replicator dynamics
eigenvalue of CðB 2IÞ has negative real part. restricted to the normal distributions in the sense of
With b ¼ 1=4 and c ¼ 4, the eigenvalues of B 2I are Theorem 3, Part 1.
1; 3 and so d0 is attractive for the replicator dynamics
restricted to the normal distributions if jajo1. On the other The two choices of explicit parameters in Examples 1a
hand, the sum of the eigenvalues of CðB 2IÞ is the trace and 1b above illustrate two methods to extend the one-
4 þ 17a=4 of dimensional CSS conditions to multi-dimensions. The
! more common approach (Meszéna et al., 2001) is to
2 þ 4a 1=4 a consider m 2 Rn a multi-dimensional CSS if it is a strict
.
2a þ 4 a=4 2 NE that is convergence stable with respect to any choice of
constant positive definite symmetric covariance matrix C.
Thus, some eigenvalue has positive real part if 16=17oao1 Translating m to the origin, this is equivalent to asserting
and so m ¼ 0 is not attractive for the canonical equation CðB 2IÞ has every eigenvalue with negative real part for
(14) with constant covariance matrix parameterized by all choices of C. Hines (1980b); Cressman and Hines (1984)
these a. (see also Leimar, 2005) show this is true if and only if
The mathematical reason for this difference between B 2I is negative definite, which in the above example is
properties of the replicator equation and the canonical the condition jb þ cjo4. The negative definiteness of B
equation (see Theorem 6 below) is that B 2I is not 2I is also equivalent to (multi-dimensional) m-stability
negative definite for b ¼ 14 and c ¼ 4 since jb þ cj44. introduced by Lessard (1990).
Furthermore, even if B 2I were negative definite, the In another approach, Leimar (2005) considers a more
asymmetry of B implies there is a continuous choice CðmÞ restrictive notion of CSS by allowing all continuously
depending on m for which m ¼ 0 is unstable. varying CðmÞ that are positive definite but not necessarily
symmetric, a condition Leimar called absolute convergence
Example 1b. For an explicit example of this latter stability (see also the concept of a Darwinian demon in
phenomenon, take b ¼ 1 and c ¼ 12 (so B 2I is negative Leimar, 2001). He then showed this condition is equivalent
definite) with to B 2I being symmetric and negative definite, a similar
! result as that illustrated in Example 1b. The essential
a2 ab 42

CðmÞ ¼ , properties for this specific example are that A 14 39 16
ab b2 has an eigenvalue of positive real part and that AT ðB 2IÞ
is positive definite.
10
The condition for asymptotic stability in both dynamics is that the
11
only entry b 2 of the matrix B 2I is negative (i.e. bo2). Here we ignore CðmÞ is only positive semidefinite. Positive terms can be added to the
the degenerate case with b ¼ 2. Similarly, the above definition of CSS diagonal of CðmÞ to make it positive definite, at least for ma0. If these
ignores the possibility that a non-strict NE can be a CSS, a situation we terms are of the form oða2 þ b2 Þ (i.e. they go to zero faster than a2 þ b2 ),
also view as degenerate. they will not affect the instability of m ¼ 0 under Eq. (14).
ARTICLE IN PRESS
By the following definition, we follow the first perspec- On the other hand, if B is not symmetric, none of the
tive of Lessard (1990); Meszéna et al. (2001) in this paper. statements are equivalent by Examples 1a and 1b. We then
have the following theorem.
Definition 1. m 2 Rn is a multi-dimensional CSS if it is a
strict NE and convergence stable with respect to any Theorem 6. Suppose B is not symmetric and m ¼ 0 is a strict
choice of constant positive definite symmetric covariance NE. Then
matrix C.
(1) d0 is attractive and asymptotically stable (as in Theorem
That is, we define a strict NE to be a CSS if B 2I is 3, Part 1) under Eq. (1) for the class of normal
negative definite, i.e., x Bxo2kxk2 for all xa0. There are distributions if and only if every eigenvalue of B 2I
several reasons for this choice. Not only is mutational has negative real part.
covariance near a monomorphic equilibrium assumed to be (2) m ¼ 0 is a CSS if and only if B 2I is negative definite.
constant in most treatments of adaptive dynamics (Vincent (3) m ¼ 0 is not absolutely convergence stable.
et al., 1993; Metz et al., 1996; Meszéna et al., 2001), it is
also a standard assumption in earlier treatments of Remark. By Theorem 6 and the discussion following
evolution of strategy distributions from game-theoretic Theorem 4, the multi-dimensional CSS concept emerges
models where payoff functions are often assumed bilinear by requiring asymptotic stability with respect to all positive
as in pðx; yÞ ¼ x By (Hines, 1980b).12 For us, another semidefinite covariance matrices with one-dimensional
important reason is that this definition of CSS is the most range (i.e. a line through the origin). By Theorems 3 and
relevant condition for dynamic stability of the general 4 restricted to each such line, we have an independent proof
replicator equation (1) analysed in the following section of the result (Lessard, 1990; Meszéna et al., 2001) that the
(see Theorem 12 there). multi-dimensional CSS concept is equivalent to the one-
The above example and/or the proof of Theorem 3 show dimensional CSS conditions for each line through the
that dynamic stability depends critically on the evolution of origin.
the covariance matrix CðmðtÞÞ and/or CðtÞ. However, if B
2I is symmetric, all concepts are equivalent. That is, we 5. Stability of monomorphisms under the replicator equation
have the following.
The explicit analysis of the replicator dynamics in
Theorem 5 (Symmetric B). Suppose B is symmetric. The Section 3 relies heavily on the assumptions the payoff
following four statements are equivalent.13 function pðx; yÞ is quadratic and the initial population is
normally distributed with mean m. Such normal distribu-
(1) d0 is attractive under Eq. (1) for the class of non- tions are one way to model aggregate individual mutations
degenerate normal distributions. for a monomorphic population at m. An underlying
(2) d0 is Lyapunov stable under Eq. (1) for the class of all assumption of the replicator equation (1) is that each
normal distributions. individual reproduces offspring with its same trait and at a
(3) m ¼ 0 is a CSS. rate equal to its fitness. If this reproduction of clones is
(4) m ¼ 0 is a strict NE and absolutely convergence stable. subject to small shocks from m that are independent of
each other, the Central Limit Theorem can be used to
The reason for this equivalence is that, for symmetric conclude the initial traits will be normally distributed after
matrices, negative definiteness is equivalent to all eigenva- such a shock. So here mutation is treated indirectly by
lues being negative. The stability assertion in statement 2 considering variations of initial conditions.
follows from the Fundamental Theorem of Natural Selection In the remainder of the paper, we consider other initial
(see Oechssler and Riedel, 2002, Theorem 1, or Cressman distributions (that can also be given a mutational
and Hofbauer, 2005, Section 4.2). Specifically, mean fitness interpretation) and arbitrary payoff functions. We will be
~
pðP; PÞ is a Lyapunov function for Eq. (1) for the most interested in convergence and stability properties of
~ yÞ ¼ x x þ x By y y
equivalent payoff function pðx; monomorphisms dm under Eq. (1) for initial distributions
~ yÞ ¼ pðy;
which is symmetric (i.e. pðx; ~ xÞ) and P ¼ d0 is a whose support is close to m to reflect the adaptive
~
maximizer of pðP; PÞ if and only if B 2I is negative dynamics assumption that trait substitution involves only
definite. nearby mutations. This means we cannot simply use the
weak topology on DðRn Þ (or on DðSÞ for that matter) since
12
See however Hines (1980a) where it is shown that non-constant P may be close to dm in the weak topology and still have
variances arise quite naturally and can play an important role in the support all of Rn (e.g. a normal distribution with mean 0
stability analysis. Effects of evolving (co)variances are important in and small variance is close to d0 in the weak topology).
models of quantitative genetics as well (e.g. Bürger, 2000) where the mean That is, for stability with respect to the weak topology, one
strategy dynamics is similar to the canonical equation, although here
variances are again sometimes assumed constant (Abrams, 2001). must also account for possibly large payoff effects of
13
These equivalences ignore degenerate possibilities. For example, we (admittedly rare) interactions involving an individual
ignore situations where relevant eigenvalues have zero real part. whose strategy is far from that of the monomorphic
ARTICLE IN PRESS
population. To avoid this problem, we concentrate instead pðx; yÞ up to quadratic terms is given by Eq. (5). To repeat,
on convergence to dm in the weak topology for initial we make the following assumption.
distributions whose support is close to m (see the concept
Assumption. m ¼ 0 is a local strict NE (i.e. pðx; 0Þopð0; 0Þ
of neighborhood attractivity in the following definition). It
for all x sufficiently close (but not equal) to 0) and
is important to note that neighborhood attractivity also
the Taylor expansion of pðx; yÞ about ð0; 0Þ is pðx; yÞ ¼
requires m to be in the support of these initial distributions
x x þ x By up to second-order terms.
since the replicator dynamics (1) in a continuous trait space
shares the same property of its analog for a finite number
of traits; namely, supp Pt ¼ supp P0 for all tX0 (Bomze, 5.1. Dynamic stability, neighborhood superiority and NIS
1991). The main advantage of this concept is that the
Taylor expansion of the payoff function about the In this section, we seek the strongest type of attractivity
monomorphism can be used (specifically, up to quadratic possible by taking Q ¼ Q0 :¼fP 2 DðSÞ : 0 2 supp Pg. Con-
terms) as a good approximation. sider the trajectory in DðSÞ for initial P0 with support f0; xg
and x 2 S. The replicator equation (1) for such P0 is
Remark. Alternatively, we could modify the weak topol- dP
ogy by defining P close to P if P is close to P in the weak ðf0gÞ ¼ Pðf0gÞðpðd0 ; PÞ pðP; PÞÞ
dt
topology and the support of P is close to that of P . In fact, ¼ Pðf0gÞPðfxgÞ½ðpð0; 0Þ pðx; 0ÞÞPðf0gÞ
when the trait space is one dimensional, this is the
approach taken by Eshel and Sansone (2003) who called þ ðpð0; xÞ pðx; xÞÞPðfxgÞ
this the maximal shift topology. However, every mono- ¼ Pðf0gÞPðfxgÞ½x xPðf0gÞ
morphism P ¼ dm is then automatically stable in this þ ðx x x BxÞPðfxgÞ þ h:o:t. ð15Þ
modified weak topology since P is close to dm if and only if
its support is close to m .14 However, Pt cannot converge to If x x x Bxo0 for some xa0, then dP=dtðf0gÞo0 if
dm in this new topology since the support of Pt does not x is sufficiently close to 0 and PðfxgÞ is sufficiently close to
converge to fm g. That is, if dm is neighborhood attractive, 1. Since 0 is in the interior of S, we may assume x 2 S.
then it immediately enjoys stability properties under this Thus, Pt does not converge to d0 and so d0 is not attractive
new topology but convergence must still be taken with for such dimorphic P0 .
respect to the weak topology. For this reason, we prefer That is, neighborhood attractivity of d0 for Q0 implies B
not to formally define a modified weak topology. I is negative definite.15 This negative definiteness condition is
similar to the CSS condition of adaptive dynamics (see
Definition 2. Let Q be a set of probability distributions, Definition 1 of Section 4). In fact, it is more closely related to
whose support contains that of P , that is forward the static condition called a neighborhood invader strategy
invariant under Eq. (1). P is attractive for Q if Pt (NIS) (McKelvey and Apaloo, 1995; Apaloo, 1997).16 The
converges to P in the weak topology for every P0 2 Q. trait 0 is NIS if it has higher payoff against all nearby
P ¼ dm is neighborhood attractive for Q if, for some monomorphic populations dx than the expected payoff of x.
neighborhood U of m in Rn , Pt converges to P in the That is, we have the following definition.
weak topology for every P0 2 Q with supp P0 U. Definition 3. m 2 Rn is a NIS if
Clearly, neighborhood attractivity depends critically on pðm ; xÞ4pðx; xÞ (16)
the choice of Q. For instance, in the trivial case that dm is for all other x 2 Rn in a neighborhood of m .
the only distribution in Q with support close to fm g, dm is
neighborhood attractive by default. Sections 5.1 and 5.2 From the Taylor expansion of a general payoff function
examine the attractivity properties of Definition 2 for two pðx; yÞ about ð0; 0Þ, 0 is NIS if
choices of Q that are more important. In either case, it is r2 p x þ 12x ðr222 pÞx4r1 p x þ r2 p x þ 12½x ðr211 pÞx
assumed that Q contains measures whose support is a small
closed neighborhood of m ¼ 0 in the Euclidean topology. þ 2x ðr212 pÞx þ x ðr222 Þx þ h:o:t.
As shown in Lemma A.1 in the Appendix, the attractivity Since 0 is in the interior of S, r1 p x ¼ 0 and so the NIS
concept in Definition 2 then requires at a minimum that m condition is that
be a NE locally (i.e. pðm ; m ÞXpðx; m Þ for all x near m ,
see also Alós-Ferrer and Ania, 2001). To avoid technical x ðr211 pÞx þ 2x ðr212 pÞxo0.
complications, we again assume throughout Section 5 that
15
m is in fact a local strict NE as determined by the second- Throughout Section 5, we again ignore degenerate possibilities. Thus,
order Taylor expansion of p. Furthermore, we assume that here we assume B I is not negative semidefinite. There is a partial
converse as well; namely, if B I is negative definite (i.e. x x x Bx40
the trait space has been parameterized in such a way that for all x), then d0 attracts all dimorphic P with support containing 0.
m ¼ 0 is in the interior of S and the Taylor expansion of 16
An NIS is also known as a good invader (Kisdi and Meszéna, 1995)
and as satisfying (multi-dimensional) m -stability (Lessard, 1990). For a
14
This statement is no longer true for stability of a non-monomorphic one-dimensional trait space, Eshel and Sansone (2003) proved the NIS
distribution P (for dimorphisms, see Cressman, 2005). condition is necessary for neighbourhood attractivity of d0 .
ARTICLE IN PRESS
By reparameterizing our trait space so that p is given by Theorem 8. If d0 is neighborhood superior and P0 has
Eq. (2) (i.e. r211 ¼ 2I and r212 ¼ B), the local strict NE 0 compact support sufficiently close to 0 and containing 0, then
is NIS if and only if B I is negative definite. Further- d0 is an o-limit point of Pt .19
more, from Eq. (16) combined with the fact that
pð0; 0Þ4pðx; 0Þ for all other x 2 Rn in a neighborhood of In the special case of symmetric payoff functions (i.e.
0, B I is negative definite if and only if 0 strictly pðx; yÞ ¼ pðy; xÞ), we have the following characterization,
dominates all other nearby x in the two-strategy game similar to Theorem 2 of Oechssler and Riedel (2002) and
based on the trait space f0; xg. This game-theoretic Theorem 4 in Cressman and Hofbauer (2005).
characterization that Theorem 9. Suppose pðx; yÞ is a symmetric payoff function
pð0; zÞ4pðx; zÞ (in particular, B is symmetric).20 Then d0 is neighborhood
attractive for the set Q0 if and only if d0 is neighborhood
for all z 2 f0; xg is important for comparison to the superior.
analogous characterization of the CSS condition in Section
5.2 (see Eq. (17) there). Remark. Attempts to extend Theorem 9 to general pðx; yÞ
Another game-theoretic characterization with attractiv- have an interesting history. Oechssler and Riedel (2002)
ity consequences is given in terms of the following conjecture that Theorem 9 remains true for the weak
definition introduced by Cressman (2005) for continuous topology when p is not symmetric (see their concept of
strategy spaces. Neighborhood superiority is closely con- evolutionarily robust). Eshel and Sansone, 2003 provide a
nected to the concept of evolutionarily robust (Oechssler proof of Theorem 9 for general pðx; yÞ if the trait space is
and Riedel, 2002) (also called locally superior with respect one dimensional (although we have been unable to follow
to the weak topology by Cressman and Hofbauer, 2005) all the details of this proof). If Q is taken as the set of all
whereby pðP ; PÞ4pðP; PÞ for all P sufficiently close to P measures with Pðf0gÞ40, Theorem 9 was proven by Bomze
in the weak topology. (1990) for pðx; yÞ ¼ fðxÞ þ fðyÞ, by Oechssler and Riedel
(2002, Theorem 3) (see also Cressman, 2005, Theorem 1)
Definition 4. The monomorphism P ¼ dm is neighborhood for general pðx; yÞ.
superior if, for all other P with support sufficiently close to
m , pðP ; PÞ4pðP; PÞ.17 5.2. Dynamic stability, dominance solvability and CSS
The following result summarizes the above discussion.
Section 5.1 illustrates the importance of the static NIS
Theorem 7. The following four statements are equivalent concept for stability of the replicator dynamics (1). We now
under our above Assumption for Section 5. turn to the relevance of the CSS concept. From the
adaptive dynamics perspective, dynamics (15) models trait
(1) d0 is neighborhood superior. substitution from x to 0 in one step, whereas the canonical
(2) 0 is an NIS. equation is built on the premise that mutation and trait
(3) 0 strictly dominates all other nearby strategies x in the substitution is a gradual process whereby x evolves to 0 in a
two-strategy game based on the trait space f0; xg. sequence of many steps.
(4) B I is negative definite. For a one-dimensional trait space, the canonical
equation requires all traits between 0 and x be available
The only non-obvious implication in the above Theorem for substitution and so we now assume the support of P0
is that the first statement is implied by any one of the other contains this interval. The heuristic condition (Eshel, 1983)
three statements. This proof is Theorem 1 in Cressman for a strict NE at m ¼ 0 to be a CSS amounts to replacing
(2005). An NIS need not be neighborhood superior if inequality (16) with
quadratic terms in the Taylor expansion do not determine pðy; xÞ4pðx; xÞ (17)
the NIS conditions. Oechssler and Riedel (2002) provide
whenever y is close to x and between 0 and x. The Taylor
the counterexample pðx; yÞ ¼ ðx yÞ4 2x4 with a one-
expansion of pðx; yÞ about ð0; 0Þ now yields a local strict
dimensional trait space (see also Eshel and Sansone, 2003).
NE satisfies Eq. (17) if and only if bo2 where b is the only
Intuitively, a neighborhood superior P should be entry of the 1 1 matrix B in Eq. (5).
neighborhood attracting since P has a higher than average Cressman and Hofbauer (2005) were able to use an
payoff at every nearby population state P.18 Unfortu- iterated strategy domination argument to show that, for
nately, we are only able to prove the following partial result any jbjo2, d0 is attractive for initial distributions P0 whose
for general payoff functions. support is a (sufficiently small) interval S containing 0.
17 19
Cressman (2005) analysed this concept for any P with finite support. d0 is attractive for Q0 if and only if d0 is the unique o-limit point of Pt
18
For a finite trait space S, this intuition is equivalent to the notion of an for all such P0 2 Q0 .
20
evolutionarily stable strategy (i.e. a matrix-ESS) by Maynard Smith We may assume pðx; yÞ has the form pðx; yÞ ¼ x x þ x By y
(1982). Furthermore, it is well-known a matrix-ESS is locally asympto- y þ h:o:t: since the terms pð0; 0Þ þ r2 p y and y y (cf. Eq. (2)) are
tically stable for the replicator equation on a finite trait space. irrelevant for the replicator equation.
ARTICLE IN PRESS
Specifically, they showed the game with the continuum of for nonzero x (i.e. B 2I is negative definite). The
traits in S is strictly dominance solvable (see Definition 5 condition kBko2 also has an interesting connection to
below) to the trait 0. By this process, each trait x 2 S that is the Cournot tatonnement process of Moulin (1984). To see
strictly dominated by another trait y 2 S is eliminated and this clearly, let us ignore the non-quadratic terms in Eq. (5).
then each remaining trait that is strictly dominated (in the For our single-species model, the Cournot process is the
reduced game with the resultant trait space) by another sequence of best replies yiþ1 arg maxfpðx; yi Þ : x 2 Rn g
remaining trait is eliminated, etc. If every trait except 0 is which is given by yiþ1 ¼ 12 Byi . This discrete-time tatonne-
eventually eliminated by this countable process, standard ment process converges to 0 if and only if the eigenvalues
techniques extended from finite trait space (Samuelson and of 12 B all have modulus less than 1. For symmetric B, this is
Zhang, 1992) show d0 is attractive for the above initial equivalent to kBko2.
distributions with respect to the replicator equation under When kBko2 but B is not symmetric, the proof of
this iterated elimination of strictly dominated strategies Theorem 11 still shows that the games restricted to all trait
(Cressman and Hofbauer, 2005) (see also Heifetz et al. spaces S Rn that are star-shaped about 0 and sufficiently
(2003) and the proof of Theorem 10 below). close to 0 are strictly dominance solvable if kBko2 (and so
In this section, we extend this argument to a multi- d0 is still neighborhood attractive for Q ). However, there
dimensional setting, starting with the concept of domi- are other choices of B with kBk42 for which the associated
nance solvability similar to that introduced by Moulin games are strictly dominance solvable (see Section 5.3).
(1984). The reason for this is that we do not need to use Euclidean
distance as in the proof of Theorem 11 (where the nested
Definition 5. The game with compact trait space S is sequence of trait spaces are disks). The essential inequality
strictly dominance solvable to x 2 S if there is a countable there is that, for nonzero y0 ,
nested sequence of closed subsets Si in S with Siþ1 S i
and S 0 ¼ S satisfying y0 Bzo2y0 y0 (18)
(i) for every iX0 and every x 2 Si nS iþ1 , there exists a y 2 for all z in the disk fzjz zpy0 y0 g. These regions can be
S i such that pðy; zÞ4pðx; zÞ for all z 2 Si replaced by others that are compact and star-shaped. For
T
(ii) 1
i¼0 S i ¼ fx g.
instance, if D is a positive definite symmetric matrix, we
will have strict dominance solvability if y0 Bzo2y0 y0
Theorem 10. If the game with compact trait space S Rn is for all z in the disk z 2 fzjz Dzpy0 Dy0 g. Since Eq. (18)
strictly dominance solvable to x 2 S, then Pt converges to is linear in z, we can restrict to the boundary
dx in the weak topology for each initial distribution P0 with fzjz Dz ¼ y0 Dy0 g. The following theorem then gener-
full support S. alizes Theorem 11 to non-symmetric B.
The most straightforward application of this theory to

Theorem 12. Suppose there exists a positive definite sym-
our multi-dimensional setting is through the following
metric matrix D such that, for all ya0,
theorem when B is symmetric (and the trait space is
parameterized so that the payoff function has form (5) up
ðy DyÞðBT y D1 BT yÞo4ðy yÞ2 . (19)
to quadratic terms). A set S Rn is called star-shaped
about 0 if it contains the line segment from 0 to x for every Then the games restricted to all compact trait spaces S Rn
x 2 S. Hence for n ¼ 1, S is an interval containing 0. Let that are star-shaped about 0 and sufficiently close to 0 are
Q be the set of all probability measures in DðSÞ whose strictly dominance solvable and d0 is neighborhood attractive
support is star-shaped about 0. for Q in the replicator equation (1). If d0 is neighborhood
Theorem 11. Suppose B is symmetric. The games restricted attractive in this sense, then 0 is CSS.
to all compact trait spaces S Rn that are star-shaped about
0 and sufficiently close to 0 are strictly dominance solvable if Remark. If D is chosen as a (positive) multiple of the
and only if kBko2.21 Furthermore, if kBko2, then d0 is identity matrix in Theorem 12, then Eq. (19) is equivalent
neighborhood attractive for the replicator equation (1) for to kBT yko2kyk. Since kBk ¼ kBT k, Theorem 12 shows
Q . that the statement of Theorem 11 is valid when kBko2 and
B is non-symmetric.
Notice that 0 is a CSS if kBko2 since we then have
x ðB 2IÞxokBkkxk2 2kxk2 o0 Conversely, if B is symmetric with kBk42, then BT y ¼
21
ly for some y and jlj42. Since ðy DyÞðy D1 yÞXðy yÞ2
Here kBk is the operator norm of B (i.e. kBk supkxk¼1 kBxk where for all positive definite symmetric matrices D, we have ðy
kxk is the Euclidean length of x). We ignore the degenerate possibility
kBk ¼ 2. Strict dominance solvability for all games with trait space S DyÞðBT y D1 BT yÞ44ðy yÞ2 and so Theorem 12 does not
sufficiently close to x is closely related to the concept of locally strictly expand the set of payoff functions with B symmetric for
dominance solvable defined by Moulin (1984). which we have a proof that d0 is neighborhood attractive.
ARTICLE IN PRESS
5.3. Two-dimensional trait space In fact, the exact condition for strict dominance
solvability to 0 is unknown for non-symmetric B. It is also
Using Theorems 11 and 12 on dominance solvability, we an open question whether the CSS condition is sufficient
are able to determine a large class of 2 2 matrices B for for neighborhood attractivity of d0 for Q .
which d0 is neighborhood attractive in the replicator
equation (1) for Q . For this, we use the fact that every 6. Discussion
2 2 matrix is orthogonally similar to a matrix of the form
As stated in the Introduction, we feel the adaptive
a b
B¼ . (20) dynamics model to predict stability of monomorphisms by
c a
emphasizing the evolution of the population mean strategy
That is, there is a rotation O such that OT BO has this form. misses the effects of the spread of the distribution of
Note that such a transformation does not affect the form of individual behaviors. The replicator equation on a
the fitness function (2) nor the symmetry of B. continuous trait space is our preferred method to include
these effects. The basic issue we consider is then whether
Theorem 13. If B is given by Eq. (20), then the games
the static CSS and NIS concepts for monomorphic stability
restricted to all compact trait spaces S R2 that are star-
of one-dimensional adaptive dynamics predict stability of
shaped about 0 and sufficiently close to 0 are strictly
the replicator equation when generalized to multi-dimen-
dominance solvable if
sions. An initial obstacle to analysing this issue is that
jbj þ jcjo2ð2 jajÞ. (21) universally accepted static extensions are not agreed upon
In this case, d0 is neighborhood attractive for Q . in the adaptive dynamics approach since stability of the
canonical equation now depends on the relative rates
If B is symmetric, Eq. (21) becomes jaj þ jbjo2 which is mutations occur in different directions (in technical terms,
the condition of Theorem 11 since kBk ¼ jaj þ jbjo2. On on the mutations’ covariance matrix).
the other hand, by Theorem 13, there are non-symmetric One assumption is that the covariance matrix will evolve
B’s for which d0 is neighborhood attractive but kBk42. 0 b
very slowly (if at all) and so can be taken as essentially
For instance,
neighborhood attractive for B ¼ 0 0 if
0d0 bis constant (Vincent et al., 1993), a method that has also been
jbjo4 but 0 0 ¼ jbj. Finally, it is straightforwardpffiffiffiffiffi to used effectively much earlier in the matrix game model
confirm that condition pffiffiffiffiffi (21) implies the eigenvalues a bc (Hines, 1980a). With arbitrary (but fixed) covariance,
of B satisfy ja bcjo2. Thus the best reply Cournot monomorphic stability with respect to both the canonical
tatonnement process (Moulin, 1984) again converges to equation and to the potential evolution of dimorphisms
zero although it is no longer true that kxtþ1 kokxt k. leads to the CSS conditions in each direction through the
monomorphism (Meszéna et al., 2001). We take this as our
5.3.1. Summary for two-dimensional trait space

multi-dimensional CSS concept. On the other hand, if the
Take pðx; zÞ ¼ x x þ x Bz and B ¼ ac ba . relative rates of mutation are not constant but can change
For B symmetric, 0 is at different points along the evolutionary path to have their
(i) NIS if and only if ao1 and jbjoj1 aj, most extreme effect (Leimar, 2001, 2005), much stronger
(ii) CSS if and only if ao2 and jbjoj2 aj, stability conditions than being a CSS in each direction are
(iii) dominance solvable if and only if jaj þ jbjo2. needed in the adaptive dynamics approach (see Section 4).
In light of the above discussion, the analytic results of
For B non-symmetric, 0 is Section 3 are quite surprising. By Theorem 3 there, when
individual behaviors are initially normally distributed and
(i) NIS if and only if ao1 and jb þ cjo2j1 aj, fitnesses are approximated by their second-order Taylor
(ii) CSS if and only if ao2 and jb þ cjo2j2 aj, expansions about a monomorphism, then even the weaker
(iii) dominance solvable if jbj þ jcjo2ð2 jajÞ. CSS conditions are too strong (see Theorem 3 for the
precise statement) to characterize stability since the
Condition (21) is equivalent to strict dominance solva- distribution’s covariance evolves slowly to having equal
bility when B is non-symmetric and bcX0 (i.e. b and c have effect in all directions. Although this result raises legitimate
the same sign). To see this, Theorem 11 applied to the one- concerns about current adaptive dynamics approaches for
dimensional trait space S ¼ ftð1; 1Þ : jtjp1g yields the multi-dimensional trait space, the presence of strategies at
necessary condition j2a þ b þ cjo4 for strict dominance the tails of the normal distribution does not match the
solvability and to the one-dimensional trait space usual assumption that mutations only occur near the
S ¼ ftð1; 1Þ : jtjp1g yields j2a b cjo4. Thus, j2ajþ monomorphic equilibrium.
pffiffiffiffiffiffiffiffiffiffiffiffiffi
jb þ cjo4. On the other hand, kBk ¼ 1 þ b2 for B ¼ For this reason, Section 5 concentrates on behavioral
1 b
distributions with compact support contained in a small
b 1pffiffiand
ffi
so the game is strictly dominance solvable if
neighborhood of a monomorphism m .22 In this setting
jbjo 3 even though j2aj þ jbj þ jcj may be arbitrarily
pffiffiffi 22
close to 2 þ 2 344. For technical reasons, this support must also include m .
ARTICLE IN PRESS
R
and with no more restrictions on the support of the initial components are uncorrelated, xi ðxj mj ÞPðdxÞ ¼ 0 for
distribution, we show in Section 5.1 that the multi- iaj. Substitution of these results into the last equation for
dimensional NIS is the most relevant concept for stability m0i yields
of dm under the replicator equation (Theorems 7–9), X
n
generalizing results of Eshel and Sansone (2003); Cressman m0i ¼ 2mi V i þ V i Bik mk .
(2005). When distributions are restricted to those whose k¼1
supports are compact intervals in each direction from m , That is, in vector notation, we have Eq. (6).
we show in Section 5.2 the relevance of the CSS conditions When the covariance matrix is not diagonal, the
by clarifying the relationship between CSS and strategy following proof uses moment generating functions. As this
dominance in the multi-dimensional model. As explained method has potential use in other contexts as well, we start
there, stability of the measure dynamics via strategy with a general exposition here.
domination corresponds to stability of the Cournot Take a probability measure P with mean m and
tatonnement process with a continuum of strategies covariance matrix C and a vector l 2 Rn . Define the
(Moulin, 1984), a discrete-time dynamic whereby rational Laplace transform
decision makers choose the optimal strategy in the next Z
time period given current population behavior. This Lðl; PÞ ¼ expðlT xÞPðdxÞ,
connection continues the long tradition of classical game-
theoretic methods providing valuable insight into the and its logarithm
eventual outcome of behavioral evolution under the
Mðl; PÞ ¼ log Lðl; PÞ.
replicator equation (and vice versa).
Finally, Section 5.3 summarizes how our results apply to L can be extended to finite signed measures in a
two-dimensional trait spaces, an important special case straightforward way. It is well known (and follows
that highlights the added difficulties that arise when trait immediately through differentiation under the integral)
space has an extra degree of freedom compared to the that
analysis of Cressman and Hofbauer (2005) where trait
q
space is one dimensional. Mðl; PÞ ¼ mi
qli l¼0
Appendix and

q2
Mðl; PÞ ¼ C ij .
Proof of Theorem 1. The proof that the class of normal qli qlj l¼0
distributions is invariant under replicator dynamics is in Thus, M generates the mean through the gradient and the
Oechssler and Riedel (2001) for n ¼ 1. A similar method covariance matrix through its Hessian. Therefore, we can
works for many dimensions and it establishes the system of obtain differential equations for the mean and the
ordinary differential equations (6) and (7) at the same time. covariance matrix by differentiating the moment generat-
In the following, a different proof by using moment ing function M. Normal distributions Nðm; CÞ are char-
generating functions is given. acterized by Mðl; Nðm; CÞÞ ¼ lT m þ 12 lT Cl.
First, we provide some intuition for the proof of Eq. (6). Let PðtÞ be a trajectory of replicator dynamics in the
Assume that all Pt are normally distributed with mean mðtÞ following. It is useful to associate with the measures PðtÞ
and diagonal covariance matrix with entries V i ðtÞ. Note the probability measures Pl ðtÞ as given by
that pðx; Pt Þ ¼ x x þ x BmðtÞ and pðPt ; Pt Þ ¼ mðtÞ Z
mðtÞ V ðtÞ 1 þ x BmðtÞ, where 1 denotes the summing 1
Pl ðtÞðGÞ ¼ expðlT xÞPðtÞðdxÞ.
vector ð1; 1; . . . ; 1Þ. Then the differential equation for the Lðl; PðtÞÞ G
ith component of the mean is (omitting the time variable t) Note that
Z
m0i ¼ xi ðpðx; PÞ pðP; PÞÞPðdxÞ d Lðl; P0 ðtÞÞ
Mðl; PðtÞÞ ¼ .
Rn dt Lðl; PðtÞÞ
Z !!
X n
¼ 3
xi þ xi ðxi mi ÞBik mk Þ PðdxÞ By definition of replicator dynamics
Rn k¼1 Z
!! d 1
XZ X
n Mðl; PðtÞÞ ¼ expðlT xÞ½pðx; PðtÞÞ
þ xi x2j þ ðxj mj ÞBjk mk Þ PðdxÞ dt Lðl; PðtÞÞ
n
jai R k¼1 pðPðtÞ; PðtÞÞPðtÞðdxÞ
mi ðm m V 1Þ. Z
1
¼ expðlT xÞpðx; PðtÞÞPðtÞðdxÞ
Lðl; PðtÞÞ
R The third moment of a normal random variable is
pðPðtÞ; PðtÞÞ
x3i PðdxÞ ¼ m3i þ 3mi V i . The covariance of xi and xi mi
is equal to the variance V i . As we have assumed that the ¼ pðPl ðtÞ; PðtÞÞ pðPðtÞ; PðtÞÞ.
ARTICLE IN PRESS
From this, we get the following relations for the mean and ~ is diagonal with entries
where DðtÞ
the covariance matrix:
ai D11
D~ ii ðtÞ ¼ and ai ¼ 1.
q ai þ e2t Dii
m0i ðtÞ ¼ pðPl ðtÞ; PðtÞÞ ,
qli l¼0 The system of differential equations (24) is asymptoti-

q 2 cally autonomous with exponentially decreasing remainder
0
C ij ðtÞ ¼ pðPl ðtÞ; PðtÞÞ . term. Therefore, one can invoke standard results (see, for
ql ql
i j l¼0
example, Coddington and Levinson, 1955, Chapter 3,
We will now apply these relationships to the case of Problem 35) to conclude that the eigenvalues of G (i.e.,
normal distributions. As is well known (and can be seen via those of B 2I) determine the asymptotic behavior of mðtÞ
completing a square), if P ¼ Nðm; CÞ, then Pl ¼ if the positive definite covariance matrix Cð0Þ is fixed. In
Nðm Cl; CÞ. Thus, for the quadratic payoff function particular, if one eigenvalue of G has positive real part, the
pðx; yÞ ¼ x Ax þ x By, mean diverges for most initial conditions (Part 2) and m ¼
0 is asymptotically stable if every eigenvalue of G has
pðPl ; PÞ ¼ ðm ClÞT Aðm ClÞ
negative real part (first statement of Part 1).
Xn
For the second statement of Part 1, we do vary Cð0Þ and
þ Aij C ij þ ðm ClÞT Bm.
i;j¼1
so need to work a little more using our uniform estimate in
the initial covariances. By our assumption on the
Since this is quadratic in l, the set of normal distributions eigenvalues of Cð0Þ, we have ai =ðai þ e2t Þpb=ðb þ e2t Þ with
is invariant, and by comparing with ðd=dtÞMðl; PðtÞÞ ¼ b ¼ 1= 1. By the variations of constants formula, the
lT m0 ðtÞ þ 12 lT C 0 ðtÞl, we get the desired differential solution of Eq. (24) can be written as
equations Z t
~ ¼ expðGtÞnð0Þ
nðtÞ ~ ~
expðGðt sÞÞDðsÞG ~ ds.
nðsÞ
m0 ðtÞ ¼ CðtÞð2A þ BÞmðtÞ, (22) 0
As G is stable, there exist constants K; Z40 such that

C 0 ðtÞ ¼ 2CðtÞACðtÞ (23)
k expðGtÞkpK expðZtÞ for all t. It follows that
which reduces in the special case A ¼ I to Eqs. (6) and
~
knðtÞkpK ~
expðZtÞknð0Þk
(7). & Z t
b
þ K expðZðt sÞÞ ~
kGkknðsÞk ds
Proof of Theorem 3. As a first step, we perform an 0 b þ e2t
orthogonal transformation as in the analysis of the
or
dynamics of the covariance matrix. Let C ¼ Cð0Þ be the
Z t
initial covariance matrix, and choose an orthogonal matrix b
~
expðZtÞknðtÞkpKk ~
nð0Þk þ K ~
kGk expðZsÞknðsÞk ds.
O and a diagonal matrix D with decreasing entries 0 b þ e2t
D11 X Dnn 40 such that OCOT ¼ D. We know from
the discussion after Theorem 2 that CðtÞ ¼ OT DðtÞO for By applying Gronwall’s lemma (Coddington and Levinson,
DðtÞ diagonal with Dii ðtÞ ¼ 1=ðD1 ii þ 2tÞ. From this, it is 1955, Chapter 1, Problem 1) to the function
straightforward to show that mini;j Dii ðtÞ=Djj ðtÞ ¼ ~
f ðtÞ ¼ expðZtÞknðtÞk, we obtain
Dnn ðtÞ=D11 ðtÞ and that Dnn ðtÞ=D11 ðtÞXDnn =D11 . Hence, the Z t
b
set of covariance matrices with bounded ratios of ~
f ðtÞpKknð0Þk exp KkGk ds
2s
eigenvalues mini;j Dii =Djj ¼ Dnn =D11 X40 is forward in- 0 bþe
Z 1
variant under Eq. (7). b
pKknð0Þk
~ exp KkGk ds
Define the transformed mean vectors nðtÞ ¼ OmðtÞ. As O 0 b þ e2s
is orthogonal, nðtÞ and mðtÞ have the same distance to zero ~
¼ Kknð0Þkð1 þ bÞKkGk=2 .
and it suffices to study the stability behavior of nðtÞ. nðtÞ
solves the differential equation The constant L ¼ Kð1 þ bÞKkGk=2 does not depend on the
initial conditions Cð0Þ and mð0Þ. Thus, we get the estimate
n0 ðtÞ ¼ Om0 ðtÞ ¼ OCðtÞðB 2IÞmðtÞ
~
knðtÞkpL ~
expðZtÞknð0Þk
¼ OOT DðtÞOðB 2IÞmðtÞ
which establishes asymptotic stability in Part 1.
¼ DðtÞOðB 2IÞOT nðtÞ .
At the same time, this estimate shows that d0 attracts all
Let G ¼ OðB 2IÞOT . G has the same eigenvalues as B trajectories that start at some Nðm; CÞ with a positive
2I since they are similar matrices. definite covariance matrix. However, since the number b ¼
In the second step, we perform a suitable time change by 1= 1 tends to infinity as tends to zero, our argument
setting sðtÞ ¼ ðe2t 1Þ=2D11 : Let nðtÞ
~ ¼ nðsðtÞÞ. Then, does not establish asymptotic stability for the set of all
positive definite initial covariance matrices (which requires
n~ 0 ðtÞ ¼ GnðtÞ ~
~ DðtÞG ~
nðtÞ, (24) stronger assumptions on B). &
ARTICLE IN PRESS
Proof of Theorem 4. Let B~ ¼ B 2I and P ¼ xxT be a Hofbauer, 2005, p. 53). Hence d0 is the only o-limit point
symmetric rank one matrix (which arises from any initial of Pt and hence Pt ! d0 .
normal distribution concentrated on the line through the This completes the proof that d0 is neighborhood attractive
nonzero vector x 2 Rn , i.e. the kernel of Cð0Þ is perpendi- if d0 is neighborhood superior. The converse follows from the
cular to x). Then stability of Eq. (12) implies stability of the stability analysis of Eq. (15) and Theorem 7. &
matrix PB~ ¼ xxT B~ which in turn implies 0XtrðPBÞ ~ ¼
T ~ T ~ T ~ 23
trðxx BÞ ¼ trðx BxÞ ¼ x Bx. Since this is true for all x, Proof of Theorem 10. It is sufficient to show by induction
B~ ¼ B 2I is negative semidefinite. & on i that limt!1 Pt ðSnS i Þ ¼ 0 for all iX1. Given
x0 2 SnS1 , there exists a y0 2 S such that pðy0 ; zÞ4
The following Lemma generalizes a familiar statement pðx0 ; zÞ for all z 2 S. By continuity of p, there are open
from finite games, a part of the folk theorem of evolutionary neighborhoods Uðx0 Þ and Uðy0 Þ of x0 and y0 , respectively
game theory, see Cressman (2003); Hofbauer and Sigmund so that
(1998), to infinite games. pðy; zÞ pðx; zÞXK40
Lemma A.1. For a compact metric space S, let the initial P0 for all x 2 Uðx0 Þ; y 2 Uðy0 Þ and z 2 S. Since suppðPt Þ ¼ S,
have full support S and Pt ! P in the weak topology. Then P both Pt ðUðx0 ÞÞ and Pt ðUðy0 ÞÞ are positive. From Eq. (1), an
is a Nash equilibrium. application of the quotient rule yields
Z "Z Z
Proof. Denote sðx; PÞ ¼ pðx; PÞ pðP; PÞ. If Pt ! P, then d Pt ðUðy0 ÞÞ 1
¼ ðpðy; zÞ
sðs; Pt Þ ! sðs; PÞ for each s. Suppose that sðs; PÞ40 for dt Pt ðUðx0 ÞÞ Pt ðUðx0 ÞÞ2 S Uðx0 Þ Uðy0 Þ
some s 2 S. By continuity, there is a compact set A S #
(with nonempty interior, and hence P0 ðAÞ40) such that pðx; zÞÞPt ðdxÞPt ðdyÞ Pt ðdzÞ
sðs; PÞX240 forR all s 2 A. Thus, for each s 2 A, there is a
t
ts 40 such that 1t 0 sðs; Pt Þ dtX40 for all tXts . Pt ðUðy0 ÞÞ
XK .
R t Lemma 2 in Bomze (1991), dPt =dP
By R 0 ðxÞ ¼ Pt ðUðx0 ÞÞ
exp 0 sðx; Pt Þ dt Xet for tXtx . Hence Pt ðAÞ ¼ A dPt =
dP0 ðxÞP0 ðdxÞ ! 1 as t ! 1 by Fatou’s lemma (Royden, Thus limt!1 Pt ðUðy0 ÞÞ=Pt ðUðx0 ÞÞ ¼ 1 and, in particular,
1988). Since Pt ðAÞp1, this is a contradiction. & Pt ðUðx0 ÞÞ converges to 0. Since SnS 1 is compact, it is
covered by finitely many Uðx0 Þ and so limt!1
Proof of Theorem 8. The mapping s : ðx; PÞ ! pðx; PÞ Pt ðSnS 1 Þ ¼ 0.
pðP; PÞ is jointly continuous in x (Euclidean topology) and Now assume limt!1 Pt ðSnS i Þ ¼ 0. As above, for every
P (weak topology). As d0 is neighborhood superior, we x0 2 S i nS iþ1 , there exists a y0 2 S i such that pðy0 ; zÞ4
have sð0; PÞ40 for all Pad0 whose support is within an 0 pðx0 ; zÞ for all z 2 Si . By continuity of p, there are open
ball of 0. If d0 is not a limit point in the weak topology of neighborhoods Uðx0 Þ and Uðy0 Þ of x0 and y0 , respectively
an initial P0 with such support, then there is an open so that
neighborhood of d0 that includes no Pt for all t sufficiently pðy; zÞ pðx; zÞXK40
large. The set of all P outside this neighborhood with
support within the 0 ball is compact in the weak topology. for all x 2 Uðx0 Þ; y 2 Uðy0 Þ and z 2 S i . Let k max
By continuity in P, we have sð0; PÞXk40 for all such fjpðy; zÞ pðx; zÞj : x; y; z 2 Sg. Then
P and some k. By continuity in x, we can find some 1 40 Z "Z Z
d Pt ðUðy0 ÞÞ 1
such that we have sðx; PÞXk=240 for all jxjp1 and all ¼ ðpðy; zÞ
such P. But this implies
dt Pt ðUðx0 ÞÞ Pt ðUðx0 ÞÞ2 SnSi Uðx0 Þ Uðy0 Þ
#
P0 ðtÞðUÞ k pðx; zÞÞPt ðdxÞPt ðdyÞ Pt ðdzÞ
X ,
PðtÞðUÞ 2
Z "Z Z
for the ball U ¼ fx 2 S : jxjp1 g. Then PðtÞðUÞ " 1, 1
þ ðpðy; zÞ
which is a contradiction. & Pt ðUðx0 ÞÞ2 Si Uðx0 Þ Uðy0 Þ

Proof of Theorem 9. From Oechssler and Riedel (2002, pðx; zÞÞPt ðdxÞPt ðdyÞ Pt ðdzÞ
Theorem 1), we know that mean fitness pðP; PÞ is a
Lyapunov function, so that t 7!pðPt ; Pt Þ is monotonically Pt ðUðy0 ÞÞ
X½kPt ðSnS i Þ þ KPt ðS i Þ
increasing. For any initial P0 with compact support close to Pt ðUðx0 ÞÞ
0 and containing 0, Theorem 8 implies that pðPt ; Pt Þ ! 40
pð0; 0Þ as t ! 1. Neighborhood superiority of d0 implies 0
is a local NE and pð0; 0Þ4pðP; PÞ for all Pad0 with for t sufficiently large since limt!1 Pt ðSnSi Þ ¼ 0 and
support close to 0 (cf. the remark in Cressman and limt!1 Pt ðS i Þ ¼ 1. The proof continues as above to yield
limt!1 Pt ðS i nSiþ1 Þ ¼ 0 and so limt!1 Pt ðSnSiþ1 Þ ¼ 0 ¼
23 ~ is the trace of PB.
Here trðPBÞ ~ limt!1 Pt ðSnS i Þ þ limt!1 Pt ðSi nS iþ1 Þ ¼ 0 for all iX1.
ARTICLE IN PRESS
Therefore, limt!1 Pt ðSnS i Þ ¼ 0 and so Pt converges Proof. If BT y0 ¼ 0, there is nothing to prove so assume
weakly to d0 . & BT y0 a0. The result may be shown by using Lagrange pffiffiffiffi
multipliers or by the following algebraic method. Let D
Proof of Theorem 11. Suppose that kBk42. By the be the positive definite square root of D. Then
symmetry of B, there exists a (nonzero) eigenvector x0 pffiffiffiffi1 pffiffiffiffi 2
such that its corresponding eigenvalue is kBk. Consider ðy0 BzÞ2 ¼ ðy0 B D DzÞ
the star-shaped compact trait spaces S Rn of the form pffiffiffiffi1 T pffiffiffiffi 2
fcx0 : jcjpg for some 40. That is, S is the line segment ¼ ð D B y0 DzÞ
pffiffiffiffi1 pffiffiffiffi1 pffiffiffiffi pffiffiffiffi
through the origin joining x0 to x0 . We claim no x 2 S pð D BT y0 D BT y0 Þ ð Dz DzÞ
can be eliminated through strict domination by another
y 2 S if is sufficiently small. To see this, we parameterize ¼ ðBT y0 D1 BT y0 Þ ðz DzÞ
our trait space so that it is a subset of R (i.e. parameterized ¼ ðBT y0 D1 BT y0 Þ ðy0 Dy0 Þ.
by a scalar x) with x0 ¼ 1. We then have Bx0 ¼ bx0 where
It is straightforward to verify z0 satisfies the require-
jbj42. From Eq. (4), we have
ments. &
pðx; zÞ pðy; zÞ ¼ ðy xÞ½x þ y bz þ h:o:t:
Proof of Theorem 12. By Lemma A.2, a straightforward
where the higher-order terms are at least of degree two in generalization of the relevant parts of the proof of
the variables x; y; z. Since jxjp; jyjp and jbj42, when Theorem 11 yields the first result. Now suppose 0 is not
is sufficiently small the expression x þ y bz þ h:o:t: is CSS. Then B 2I is not negative definite and so there is a
positive for all x; y 2 S by either choosing z as or nonzero x such that x ðB 2IÞx40. Take S as the one-
appropriately and negative for all x; y 2 S by choosing dimensional bounded interval in the direction x that
the alternate z. That is, for all x; y 2 S, there is a z 2 S with includes 0 in its interior. Note that S is star-shaped. By
pðx; zÞ pðy; zÞX0 and so x 2 S cannot be strictly domi- Cressman and Hofbauer (2005), d0 is not neighborhood
nated by any y 2 S. attractive for Q . &
Now suppose kBko2. Let d maxz2S kzk. Suppose x0 2
S and kx0 k is close to d. From Eq. (4) with h 13. Suppose
Proof of Theorem i jbj þ jcjo2ð2 jajÞ. Take
y0 ¼ ð1 Þx0 2 S, the matrix D ¼ jajþjcj 0 24
0 jajþjbj in (19). The left-hand side
pðy0 ; zÞ pðx0 ; zÞ ¼ ½ð2 Þkx0 k2 x0 Bz þ h:o:t: of Eq. (19) is then

4 ½2kx0 k2 kx0 kkBzk ðay1 þ cy2 Þ2 ðby1 þ ay2 Þ2
2 þ
jaj þ jcj jaj þ jbj
kx0 k2
4 ½2 kBk40 ½ðjaj þ jcjÞðy1 Þ2 þ ðjaj þ jbjÞðy2 Þ2
4 2
for all z 2 S. That is, y0 strictly dominates x0 . By continuity which is less than or equal to
of p, y0 strictly dominates x for all x near x0 . Define Aða; bÞ
ðjay1 j þ jcy2 jÞ2 ðjby1 j þ jay2 jÞ2
with 0paobod as those elements of S in the annulus þ
fy 2 S : d bpkykpd ag. Each Aða; bÞ is compact and jaj þ jcj jaj þ jbj
non-empty by the definition of d and the fact S is star- ½ðjaj þ jcjÞðy1 Þ2 þ ðjaj þ jbjÞðy2 Þ2 .
shaped. Thus, for some d40, each x 2 Að0; dÞ is strictly
By considering the cases where jy1 jpjy2 j and jy1 jXjy2 j
dominated and so we can take S 1 ¼ fy 2 S : kykpd dg.
separately, it is straightforward to show that this last
We now iterate this argument and obtain our sequence S i
expression is increasing in jcj for fixed y1 ; y2 and jbj. Thus
of nested sets as the intersection of S with a disk centered at
we can replace jcj by 2ð2 jajÞ jbj. Furthermore, for
0 whose radius tends to 0 as i ! 1.
fixed jbj and y2 ¼ ky1 for k40, the right-hand side of
We next show d0 is attractive if kBko2. By the argument
Eq. (19) minus this last expression has an absolute maximum
used in the preceding paragraph, we can apply Theorem 10
when k ¼ 1. For k ¼ 1, the value is zero and so the games
and conclude that Pt converges to d0 in the weak topology
are strictly dominance solvable by Theorem 12. &
whenever P0 has star-shaped support sufficiently close to 0.
That is, d0 is neighborhood attractive for Q . &
References
The following lemma is used in the proof of Theorem 12
below. Abrams, P., 2001. Modelling the adaptive dynamics of traits involved in
inter- and intraspecific interactions: an assessment of three methods.
Lemma A.2. If D is a positive definite symmetric matrix, the Ecol. Lett. 4, 166–175.
Alós-Ferrer, C., Ania, A.B., 2001. Local equilibria in economic games.
maximum of y0 Bz subject to fzjz Dz ¼ y0 Dy0 g is
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiqffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi Econ. Lett. 70, 165–173.
y0 Dy0 BT y0 D1 BT y0 and this occurs at z0 ¼
h i
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
24
If jaj þ jcj ¼ 0, then B ¼ 00 b0 . Take D ¼ 0 0
jbj for 40. If jbjo4,
y0 Dy0 = BT y0 D1 BT y0 D1 BT y0 if BT y0 a0. there is an 40 such that (19) holds.
ARTICLE IN PRESS
Apaloo, J., 1997. Revisiting strategic models of evolution: the Kisdi, É, Meszéna, G., 1995. Life histories with lottery competition in a
concept of neighborhood invader strategies. Theoret. Popul. Biol. 52, stochastic environment: ESSs which do not prevail. Theoret. Popul.
71–77. Biol. 47, 191–211.
Bomze, I., 1990. Dynamical aspects of evolutionary stability. Monatsh. Leimar, O., 2001. Evolutionary change and Darwinian demons. Selection
Math. 110, 189–206. 2, 65–72.
Bomze, I., 1991. Cross entropy minimization in uninvadable states of Leimar, O., 2005. Multidimensional convergence stability and the
complex populations. J. Math. Biol. 30, 73–87. canonical adaptive dynamics. In: Dieckmann, U., Metz, J.A.J.
Bomze, I., Pötscher, B., 1989. Game Theoretic Foundations of Evolu- (Eds.), Elements of Adaptive Dynamics. Cambridge University Press,
tionary Stability. Lecture Notes in Economics and Mathematical Cambridge in press.
Systems, vol 324. Springer, Berlin. Lessard, S., 1990. Evolutionary stability: one concept, several meanings.
Bürger, R., 2000. The Mathematical Theory of Selection, Recombination, Theoret. Popul. Biol. 37, 159–170.
and Mutation. Wiley, New York. Marrow, P., Dieckmann, U., Law, R., 1996. Evolutionary dynamics of
Christiansen, F.B., 1991. On conditions for evolutionary stability for a predator–prey systems: an ecological perspective. J. Math. Biol. 34,
continuously varying character. Am. Nat. 138, 37–50. 556–578.
Coddington, E.A., Levinson, N., 1955. Theory of Ordinary Differential Maynard Smith, J., 1982. Evolution and the Theory of Games. Cambridge
Equations. McGraw-Hill, New York. University Press, Cambridge.
Cressman, R., 2003. Evolutionary Dynamics and Extensive Form Games. McKelvey, R., Apaloo, J., 1995. The structure and evolution of
MIT Press, Cambridge, MA. competition-organized ecological communities. Rocky Mount.
Cressman, R., 2005. Stability of the replicator equation with continuous J. Math. 25, 417–436.
strategy space. Math. Soc. Sci. 50, 127–147. Meszéna, G., Kisdi, E., Dieckmann, U., Geritz, S.A.H., Metz, J.A.J.,
Cressman, R., Hines, W.G.S., 1984. Correction to the appendix of three 2001. Evolutionary optimisation models and matrix games in the
characterizations of population strategy stability. J. Appl. Probab. 21, unified perspective of adaptive dynamics. Selection 2, 193–210.
213–214. Metz, J.A.J., Geritz, S.A.H., Meszéna, G., Jacobs, F.J.A., van Heerwaar-
Cressman, R., Hofbauer, J., 2005. Measure dynamics on a one- den, J.S., 1996. Adaptive dynamics a geometrical study of the
dimensional continuous trait space: theoretical foundations for consequences of nearly faithful reproduction. In: van Strien, S.J.,
adaptive dynamics. Theoret. Popul. Biol. 67, 47–59. Verduyn Lunel, S.M. (Eds.), Stochastic and Spatial Structures of
Dieckmann, U., Law, R., 1996. The dynamical theory of coevolution: a Dynamical Systems. Amsterdam, North-Holland, pp. 183–231.
derivation from stochastic processes. J. Math. Biol. 34, Moulin, H., 1984. Dominance-solvability and Cournot stability. Math.
579–612. Soc. Sci. 7, 83–102.
Doebeli, M., Dieckmann, U., 2000. Evolutionary branching and Oechssler, J., Riedel, F., 2001. Evolutionary dynamics on infinite strategy
sympatric speciation caused by different types of ecological interac- spaces. Econ. Theory 17, 141–162.
tions. Am. Nat. 156, S77–S101. Oechssler, J., Riedel, F., 2002. On the dynamic foundation of evolutionary
Eshel, I., 1983. Evolutionary and continuous stability. J. Theoret. Biol. stability in continuous models. J. Econ. Theory 107, 223–252.
103, 99–111. Royden, H.L., 1988. Real Analysis, third ed. Macmillan, New York.
Eshel, I., Sansone, E., 2003. Evolutionary and dynamic stability in Samuelson, L., Zhang, J., 1992. Evolutionary stability in asymmetric
continuous population games. J. Math. Biol. 46, 445–459. games. J. Econ. Theory 59, 363–391.
Heifetz, A., Shannon, C., Spiegel, Y., 2003. What to maximize if you must, Taylor, P.D., 1989. Evolutionary stability in one-parameter models under
mimeo. weak selection. Theoret. Popul. Biol. 36, 125–143.
Hines, W.G.S., 1980a. Strategy stability in complex populations. J. Appl. Taylor, P.D., Jonker, L., 1978. Evolutionarily stable strategies and game
Probab. 17, 600–610. dynamics. Math. Biosci. 40, 145–156.
Hines, W.G.S., 1980b. Three characterizations of population strategy Vincent, T.L., Cressman, R., 2000. An ESS maximum principle for matrix
stability. J. Appl. Probab. 17, 333–340. games. Theoret. Popul. Biol. 58, 173–186.
Hofbauer, J., Sigmund, K., 1990. Adaptive dynamics and evolutionary Vincent, T.L., Cohen, Y., Brown, J.S., 1993. Evolution via strategy
stability. Appl. Math. Lett. 3, 75–79. dynamics. Theoret. Popul. Biol. 44, 149–176.
Hofbauer, J., Sigmund, K., 1998. Evolutionary Games and Population Vincent, T.L., Van, M.V., Goh, B.S., 1996. Ecological stability, evolutionary
Dynamics. Cambridge University Press, Cambridge. stability and the ESS maximum principle. Evol. Ecol. 10, 567–591.

Stability of The Replicator Equation For A Single Species With A Multi-Dimensional Continuous Trait Space

Uploaded by

Copyright:

Available Formats

Stability of The Replicator Equation For A Single Species With A Multi-Dimensional Continuous Trait Space

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Stability of The Replicator Equation For A Single Species With A Multi-Dimensional Continuous Trait Space

Uploaded by

Copyright:

Available Formats

ARTICLE IN PRESS

Journal of Theoretical Biology 239 (2006) 273–288

Stability of the replicator equation for a single species with a

1. Introduction encounters that determine their payoffs (i.e. their ﬁtness or

asymptotically stable under (6).10 Unfortunately, this where

The most straightforward application of this theory to

As G is stable, there exist constants K; Z40 such that

You might also like