Skip to main content
Elsevier Sponsored Documents logoLink to Elsevier Sponsored Documents
. 2010 Sep;78(2-3):123–138. doi: 10.1016/j.tpb.2010.06.002

Evolution and polymorphism in the multilocus Levene model with no or weak epistasis

Reinhard Bürger 1,
PMCID: PMC2965013  PMID: 20561538

Abstract

Evolution and the maintenance of polymorphism under the multilocus Levene model with soft selection are studied. The number of loci and alleles, the number of demes, the linkage map, and the degree of dominance are arbitrary, but epistasis is absent or weak. We prove that, without epistasis and under mild, generic conditions, every trajectory converges to a stationary point in linkage equilibrium. Consequently, the equilibrium and stability structure can be determined by investigating the much simpler gene-frequency dynamics on the linkage-equilibrium manifold. For a haploid species an analogous result is shown. For weak epistasis, global convergence to quasi-linkage equilibrium is established. As an application, the maintenance of multilocus polymorphism is explored if the degree of dominance is intermediate at every locus and epistasis is absent or weak. If there are at least two demes, then arbitrarily many multiallelic loci can be maintained polymorphic at a globally asymptotically stable equilibrium. Because this holds for an open set of parameters, such equilibria are structurally stable. If the degree of dominance is not only intermediate but also deme independent, and loci are diallelic, an open set of parameters yielding an internal equilibrium exists only if the number of loci is strictly less than the number of demes. Otherwise, a fully polymorphic equilibrium exists only nongenerically, and if it exists, it consists of a manifold of equilibria. Its dimension is determined. In the absence of genotype-by-environment interaction, however, a manifold of equilibria occurs for an open set of parameters. In this case, the equilibrium structure is not robust to small deviations from no genotype-by-environment interaction. In a quantitative-genetic setting, the assumptions of no epistasis and intermediate dominance are equivalent to assuming that in every deme directional selection acts on a trait that is determined additively, i.e., by nonepistatic loci with dominance. Some of our results are exemplified in this quantitative-genetic context.

Keywords: Selection, Migration, Recombination, Linkage disequilibrium, Population subdivision, Geographical structure, Quantitative traits, Genotype–environment interaction

1. Introduction

To achieve a proper understanding of the evolutionary dynamics of phenotypic traits, it is essential to study the effects of selection on multiple linked or unlinked loci. Because many species are subdivided into colonies, or demes, and selection varies geographically, the consequences of migration and spatially varying selection need to be taken into account. Each of these aspects has been studied extensively but mostly separately. Multilocus selection and the maintenance of polygenic variation in a panmictic population inhabiting a constant, homogeneous environment have been prime topics of research during the past decades; for reviews or recent treatments of general models, see Karlin (1978), Turelli and Barton (1990), Lyubich (1992), Nagylaki (1992),Zhivotovsky and Gavrilets (1992), Christiansen (1999), Bürger (2000), Kirkpatrick et al. (2002), and Ewens (2004).

Spatially varying selection in subdivided populations was intensively investigated as well, mainly for the single-locus case. A particularly prominent role has been played by the Levene (1953) model, which assumes a finite number of demes, selection within demes, and individuals dispersing independently of their deme of origin. As a consequence, there is no population structure despite geographically variable selection. This property makes it more amenable to mathematical analysis than general migration–selection models. A good overview of the literature can be acquired from the articles of Karlin (1977, 1982) and Nagylaki and Lou (2008), and the pertinent chapters in the books of Nagylaki (1992) and Christiansen (1999).

Work on multilocus selection in subdivided populations is relatively scarce. There is early work by Li and Nei (1974), who showed that even in the absence of epistasis and dominance, migration–selection balance in two demes can maintain linkage disequilibrium (see also Christiansen and Feldman, 1975). Zhivotovsky et al. (1996) used a multilocus Levene model to study the evolution of phenotypic plasticity. Wiehe and Slatkin (1998) investigated a haploid Levene model in which linkage disequilibrium is caused by epistasis. Christiansen (1999) derived sufficient conditions for the protection of gametes in a multilocus context. More recently, Spichtig and Kawecki (2004) and van Doorn and Dieckmann (2006) performed numerical studies on the maintenance of multilocus polymorphism in two demes for arbitrary migration and for the Levene model, respectively. Roze and Rousset (2008) derived recursions for the allele frequencies and for various types of genetic associations in a multilocus infinite-island model. Barton (2010) explored certain aspects of speciation using a generalized, haploid multilocus Levene model that admits habitat preferences.

The notorious complexity of the evolutionary dynamics of multilocus systems as well as the richness of evolutionary phenomena in subdivided populations leave little hope for a general theory combining both aspects. However, considerable progress has been made recently for important limiting or special cases. These include weak migration and weak selection (Bürger, 2009a,b) and the Levene model without epistasis (Nagylaki, 2009b; Bürger, 2009c). If either migration or selection is weak, the evolutionary dynamics are perturbations of relatively simple limiting dynamics which are amenable to mathematical analysis. In both cases, global convergence of trajectories to equilibria in quasi-linkage equilibrium could be proved under natural, quite general assumptions (Bürger, 2009a). If migration is weak, this conclusion requires the assumption of weak epistasis; if selection is weak, then equilibrium states are also spatially quasi-homogeneous.

These results were applied in Bürger (2009b) to study the maintenance of multilocus polymorphism if epistasis is weak or absent and dominance is intermediate. In a panmictic population, polymorphism is impossible under such conditions. For strong migration, however, arbitrarily many recombining loci can be maintained polymorphic if there are at least two demes, and this holds for an open set of parameters. By contrast, for weak migration, the maximum number of loci that can be maintained polymorphic on an open set of parameters equals the number of demes.

Nagylaki (2009b) investigated evolution under the multiallelic multilocus Levene model without epistasis. He demonstrated that geometric-mean fitness, w˜(ρ), depends only on the vector ρ of gene frequencies and is monotone increasing except at equilibria. Therefore, ρ(t) converges generically, i.e., for almost all parameters and initial data, to a gene-frequency equilibrium that is a local maximum of w˜(ρ). In addition, Nagylaki proved global convergence to linkage equilibrium if there are either only two loci or there are multiple loci without dominance. He conjectured that the set of gene-frequency equilibria that are in linkage equilibrium is globally attracting, hence that global convergence to linkage equilibrium occurs always. For deme-independent degree of intermediate dominance (DIDID) he showed that, generically, at most Γ1 diallelic loci can segregate at equilibrium, where Γ denotes the number of demes.

In Bürger (2009c), the equilibrium and stability structure of the diallelic two-locus Levene model with two demes was derived in considerable generality. Epistasis was ignored but dominance admitted. Absence of genotype-by-environment (G×E) interaction was shown to lead to nongeneric, and nonrobust, properties.

In this paper, we shall prove Nagylaki’s conjecture for multiple multiallelic loci under the generic assumption that gene-frequency equilibria are isolated. As a consequence, evolution in the Levene model without epistasis can be fully understood by studying the much simpler gene-frequency dynamics on the linkage-equilibrium manifold for which geometric-mean fitness is monotone increasing along nonconstant solutions. More generally, we establish convergence of trajectories to a stationary point in quasi-linkage equilibrium if epistasis is sufficiently weak. Analogous, and even stronger, results hold if selection acts on haploids. We apply these results to investigate the maintenance of multilocus polymorphism in the Levene model for diallelic nonepistatic loci and, especially, for a quantitative trait that is under linear selection in every deme.

In Section 2, we formulate the multilocus Levene model and summarize the results on the gene-frequency dynamics that are needed subsequently. Section 3 is devoted to the study of convergence to linkage equilibrium in the absence of epistasis. The main result is Theorem 3.1, which states global convergence to stationary points in linkage equilibrium under mild, generic conditions. The proof is based on the simple observation that, at gene-frequency equilibrium, the dynamics reduces to a panmictic dynamics without epistasis. The theorem follows by employing the results of Nagylaki (2009b) on convergence of gene frequencies in the Levene model and those of Kun and Lyubich (1979, 1980) on convergence in the panmictic case without epistasis. Corollary 3.3 is most useful for applications because it formulates several of the conclusions that can be deduced for the full dynamics by analyzing the much simpler gene-frequency dynamics at linkage equilibrium. In Section 3.2, geometric convergence to a unique, globally asymptotically stable equilibrium is established under the assumption of DIDID (Theorem 3.5). Among others, this includes absence of G×E interaction as a special case. In Section 4, the haploid Levene model is studied and geometric convergence to a unique, globally asymptotically stable equilibrium is proved.

The maintenance of multilocus polymorphism is investigated in Section 5. Result 5.1 is a slight extension of Theorem 2.2 in Bürger (2009b) which shows that in the Levene model an arbitrary number of loci can be polymorphic at a globally asymptotically stable equilibrium, and this holds for an open set of parameters. The main results are Theorems 5.3 and 5.5. They apply if the degree of dominance is intermediate and deme independent and show that this assumption considerably restricts the possibility of multilocus polymorphism relative to intermediate dominance that varies among demes. The first theorem establishes that Nagylaki’s (2009b) generic upper bound, Γ1, for the number of segregating loci at equilibrium is assumed on an open set of parameters. The second theorem shows that in the nongeneric case, in which an internal equilibrium with more than Γ1 polymorphic loci exists, there is a manifold of equilibria with generic dimension LΓ+1, where L is the number of loci. This highly degenerate case occurs for instance in the absence of G×E interaction.

Section 6 applies these results to a quantitative-genetic model with linear directional selection in each deme. Corollary 6.3, which summarizes the main results, should serve as a warning when studying models under highly specialized assumptions. It shows that natural assumptions, such as absence of genotype-by-environment interaction, may lead to nongeneric model behavior. Indeed, the analysis of such a simple degenerate model (two diallelic loci, two demes, no dominance and linear fitnesses; see Bürger, 2009c) and the ensuing discussions with Thomas Nagylaki initiated the recent series of papers on multilocus migration–selection models by Nagylaki and the author.

In Section 7, weak epistasis is studied. In Section 7.1, we establish a perturbation result that yields global convergence of trajectories to quasi-linkage equilibrium. In Section 7.2, some of the results of Section 5 on the maintenance of polymorphism are extended to weak epistasis. In Section 8, we recapitulate and discuss our main findings, and mention some open problems.

2. The multilocus Levene model

We briefly introduce the multilocus Levene model and summarize some basic results from Nagylaki (2009b), hereafter abbreviated as N09b, that will be needed. There the model is developed in detail.

We suppose that there are L2 diploid loci with In2 alleles Ain(n) (in{1,In}) at locus n. We designate the set of all loci by L={1,,L} and the set of all demes by G={1,,Γ}, where Γ2 is assumed. The relative size of deme α is denoted by cα>0, hence αcα=1. (Whenever no summation range is indicated, it is assumed to be over all admissible values; here, αG.) We shall consistently use the letters h,l,nL for loci, i,j,k for gametes, and α,βG for demes (see Table 1 for a glossary of symbols). The linkage map is arbitrary, except for the assumption that all recombination probabilities are positive.

Table 1.

Glossary of symbols. For both the Roman and Greek alphabets, uppercase letters precede lowercase ones. For each uppercase or lowercase letter, listing is in order of appearance of the definition in the text. The references are to the equation closest to the definition of each symbol. Thus, (2.1), (2.1)+, (2.1)− refers to (2.1), the text below (2.1), the text above (2.1), respectively. Symbols that occur only in the Appendix are not listed.

Symbol Reference Definition
Ain(n) (2.1)- Allele at locus n
Aˆ (3.4)+ Region of attraction of ρˆ
b (5.17) (b1,,bΓ)T
bα (5.17)+ w¯ˆαnu22,α(n)
cα (2.1)- Proportion of zygotes in deme α
D (3.4)- Vector of all linkage disequilibria in an L-locus system
Di,α (3.4)- Linkage disequilibrium in gamete i in deme α
Dˆ (3.4)+ D evaluated at gene-frequency equilibrium
F (2.14) lnw˜
f (5.14)- (f(1),,f(L))T
f(n) (5.14) 2ϑ(n)p(n)+(12ϑ(n))(p(n))2
G (2.1)- Set of all demes
G(n) (2.17)- Covariance matrix for locus n
h (2.1)- Locus index
In (2.1)- Number of alleles at locus n
I (6.8)- Number of alleles at locus n if independent of n
in (2.1)- Allelic index at locus n
i (2.1)- Gamete (i1,,iL)
J (2.3)- Number of gametes
Jn (7.3)- Set of alleles present at equilibrium at locus n
j (2.1)- Index for gamete
jn (2.4)- Allelic index at locus n
k (2.1)- Index for gamete
L (2.1)- Number of loci
L (2.1)- Set of all loci
l (2.1)- Locus index
mα (6.7)+ Backward migration rate
n (2.1)- Locus index
O (7.3)+ Order symbol
pi (2.1)- Frequency of gamete i
pin(n) (2.1)- Frequency of allele Ain(n)
p (2.3)- Vector of gamete frequencies
pjhin(hn) (2.10) Frequency of gamete Ajh(h)Ain(n)
p(n) (2.17)- (p1(n),,pIn(n))T (only Section 2)
p(n) (5.1)- Frequency of allele A1(n) in the diallelic case
pˆ (3.4)+ p at gamete-frequency equilibrium
Ri,jk (2.1)+ Probability that gamete i is produced by haplotypes jk
RJ (2.3)- J-dimensional Euclidean space
rmin (3.8)+ Smallest two-locus recombination frequency
sα (6.2) Directional selection coefficient in deme α
sij,α (7.1) Epistasis coefficients
t (2.10)+ Time in generations
U (5.9)+ Open set of parameters
uinjn,α(n) (2.3)+ Fitness contribution of Ain(n)Ajn(n) in deme α
u¯α(n) (2.6) Mean fitness contribution of locus n in deme α
uin,α(n) (2.7) Marginal fitness contribution of Ain(n) in deme α
uin,α(n) (4.2a) Fitness contribution of Ain(n) in deme α (only Section 4)
V (5.10)- L×Γ matrix with entries Vnα
Vnα (5.10)- Entry of matrix V
Wjk (3.1) Averaged fitness of jk at gene-frequency equilibrium
W¯α(z) (5.15) Mean fitness in deme α as a function of z
Wα(X) (6.2) Fitness of individuals with trait value X in deme α
W (7.3)+ Open set of parameters
wij,α (2.1)- Fitness of genotype ij in deme α
wi,α (2.1) Marginal fitness of gamete i in deme α
wi,α (4.1)- Fitness of gamete i in deme α (only Section 4)
w¯α (2.1) Mean fitness in deme α
w˜ (2.13) Geometric-mean fitness
w¯ˆα (3.1)- w¯α(ρˆ)
X (6.1) Value of quantitative trait
Xij,α (6.1) nγinjn,α(n)
xˆ (5.11)+ (c1/w¯ˆ1,,cΓ/w¯ˆΓ)T
z(n) (5.14) f(n)(p(n))
z (5.14)+ (z(1),,z(L))T

α (2.1)- Deme index
β (2.1)- Deme index
Γ (2.1)- Number of demes
γinjn,α(n) (6.1)+ Contribution of Ain(n)Ajn(n) to trait value
γinjn(n) (6.6) Deme-independent contribution of Ain(n)Ajn(n) to trait value
γα (6.6) Genotype-independent contribution to trait value
ΔJ (2.3)- Simplex in RJ
Δ (2.15)- Difference operator
δij (2.17)- Kronecker delta
ϑinjn(n) (3.6) Degree of intermediate dominance
ϑinjn(n) (6.5) Degree of intermediate dominance at trait level
ϑ(n) (5.7) Degree of intermediate dominance in the diallelic case
ϑ (5.9)+ (ϑ(1),,ϑ(L))T
ϑ (6.15)- ϑ12
ϑin(n) (6.8) Degree of intermediate dominance of Ain(n) at the trait level
ϵ (7.1a)- Small positive parameter
Λ (2.11) {pΔJ:ρ=ρ}
Λˆ (2.18) {pΔJ:ρ=ρˆ}
Ξ (2.12) Linkage-equilibrium manifold
ξinjn(n) (2.8a) Averaged relative fitness of Ain(n)Ajn(n)
ξˆinjn(n) (2.18)+ ξinjn(n) at gene-frequency equilibrium
ξin(n) (2.8b) Averaged relative fitness of Ain(n)
ξˆin(n) (2.18)+ ξin(n) at gene-frequency equilibrium
ξ¯(n) (2.8c) Averaged mean fitness at locus n
ξ¯ˆ(n) (2.18)+ ξ¯(n) at gene-frequency equilibrium
ξin(n) (4.3) Averaged relative fitness of Ain(n) (only Section 4)
ρ (2.3)- Vector of allele frequencies
ρ (5.1) Vector of allele frequencies (diallelic case)
ρˆ (2.18)- ρ at gene-frequency equilibrium
Σ0 (7.1b)+ Set of equilibria without epistasis
Σw(ϵ) (7.1b)+ Set of equilibria with epistasis
ϕin(n) (6.11) Allele specific factor of fitness excess
Ω (2.3) Gene-frequency space
Ω (5.1) Gene-frequency space (diallelic case)

(2.2) Value of quantity in next generation
(2.3)- Subset
T (2.3)- Transposition of a vector
int (5.9)+ Topological interior of a set
rank (5.12) Rank of a matrix
dim (5.11)+ Dimension of a linear subspace
ker (5.11)+ Kernel of a linear map or matrix
(n) (2.17)+ Gradient operator

Throughout, we assume the Levene model with soft selection, which means that population regulation by selection occurs within demes. This assumption induces frequency-dependent selection. Because in the Levene model migration rates are independent of the deme of origin, there is no population structure, and gamete and gene frequencies before selection are deme independent (Levene, 1953; Nagylaki, 1992). We denote the frequency of gamete i, which carries allele Ain(n) at locus n, by pi, and the frequency of allele Ain(n) by pin(n). Let wij,α be the fitness of the diploid genotype ij in deme α. Then the marginal fitness of gamete i and the mean fitness of the population in deme α are

wi,α=jpjwij,αandw¯α=i,jpipjwij,α, (2.1)

respectively. Further, let Ri,jk denote the probability that a parent of genotype jk produces a gamete i during meiosis. Because there is soft selection, adult dispersal, and random mating within demes, the gamete frequencies evolve according to (N09b, Eq. (2.42))

pi=j,k,αRi,jkpjpkcαwjk,α/w¯α, (2.2)

where the prime, ′, signifies the next generation. We note that these recursions are obtained whether dispersion precedes recombination or not (Bürger, 2009a). Moreover, they admit the classical interpretation that intrademic selection is followed by random mating in the entire population (cf. Nagylaki and Lou, 2008).

The state space is the simplex ΔJRJ of probability vectors of length J, where J=nIn is the number of gametes. We write p=(p1,,pJ)TΔJ for the vector of gametic frequencies. The vector consisting of all gene frequencies pin(n) (for every n and every in) is denoted by ρΩ, where

Ω=nΔIn (2.3)

is the space of gene frequencies, or the gene-frequency space, for short.

For the rest of this section, we assume absence of epistasis. Then we can assign fitness contributions to single-locus genotypes. We denote the contributions at locus n in deme α by uinjn,α(n), where in and jn refer to the alleles carried by the genotype ij at locus n, and assume uinjn,α(n)=ujnin,α(n)0. Thus, we posit that the fitness of genotype ij is given by

wij,α=nuinjn,α(n). (2.4)

An easy calculation shows that the mean fitness in deme α becomes

w¯α=nu¯α(n), (2.5)

where

u¯α(n)=in,jnuinjn,α(n)pin(n)pjn(n)=inuin,α(n)pin(n), (2.6)

and

uin,α(n)=jnuinjn,α(n)pjn(n) (2.7)

is the fitness contribution of allele Ain(n) in deme α. Importantly, w¯α=w¯α(ρ) depends only on the vector ρ of gene frequencies, but not on the vector p of gamete frequencies.

We introduce the compact notation

ξinjn(n)=αcαw¯αuinjn,α(n), (2.8a)
ξin(n)=jnξinjn(n)pjn(n)=αcαw¯αuin,α(n), (2.8b)
ξ¯(n)=inξin(n)pin(n)=αcαw¯αu¯α(n), (2.8c)

and note that ξinjn(n), ξin(n), and ξ¯(n) are nonnegative functions of ρ, and

nξ¯(n)=1. (2.9)

With this notation, the dynamics of gene frequencies (Eq. (2.48) in N09b) can be written as

pin(n)=pin(n)ξin(n)+h:hnjhpjhin(hn)ξjh(h), (2.10)

where pjhin(hn) is the frequency of Ajh(h)Ain(n). Clearly, every solution p(t) of (2.2) generates a solution ρ(t) of (2.10), but ρ cannot be inferred from (2.10) if only ρ, and not p, is known.

We define

Λ={pΔJ:ρ=ρ} (2.11)

as the set of gametic frequencies at gene-frequency equilibrium, or the set of gene-frequency equilibria for short, and

Ξ={pΔJ:pi=pi1(1)piL(L)for every gamete i} (2.12)

as the linkage-equilibrium manifold. Further, let

w˜(ρ)=α[w¯α(ρ)]cα (2.13)

denote the geometric-mean fitness and define

F(ρ)=lnw˜(ρ)=αcαlnw¯α(ρ). (2.14)

The following result will play a fundamental role. It depends crucially on the fact that w˜ and F are functions only of ρ rather than of p.

Result 2.1 N09b —

In the absence of epistasis, the dynamics (2.2) of gamete frequencies has the following properties:

  • (a)

    The single-generation change Δw˜(ρ)=w˜(ρ)w˜(ρ) satisfies Δw˜(ρ)0, and equality holds only at gene-frequency equilibrium. The function F(ρ) shares these properties.

  • (b)
    The internal gene-frequency equilibria are the stationary points of w˜ . These are obtained as the critical points of F and satisfy
    ξin(n)=ξ¯(n)for every n and every in. (2.15)

Statement (a) about Δw˜ is Theorem 3.1 in N09b. The statement about ΔF is a trivial consequence. Statement (b) is Theorem 3.3 in N09b.

We call a property generic if it holds in an open dense set of full measure.

Remark 2.2

Simple important consequences of statement (a) are (Remark 3.2 in N09b):

  • (i)

    The set ΛΔJ of gene-frequency equilibria is globally attracting, i.e., p(t)Λ as t.

  • (ii)

    If every equilibrium point ρˆ is isolated in the gene-frequency space Ω, as is generic, every ρ(t) converges as t to some ρˆΩ.

  • (iii)

    For generic initial conditions ρ(t) converges to a local maximum of w˜.

If there is linkage equilibrium, i.e., on Ξ, the recursions (2.10) for the gene frequencies simplify to

Δpin(n)=pin(n)αcαw¯α(uin(n)u¯α(n))=pin(n)(ξin(n)ξ¯(n)). (2.16)

In fact, (2.16) can be written as a generalized gradient system, similar to the single-locus selection dynamics in a panmictic population (Bürger, 2000, p. 42). To see this, note that

Fpin(n)=2αcαw¯αuin(n)=2ξin(n),

and define p(n)=(p1(n),,pIn(n))T and G(n) as the In×In covariance matrix with entries 12pin(n)(δinjnpjn(n)). Then, for every locus n, we obtain

Δp(n)=G(n)(n)F, (2.17)

where (n)F=(F/p1(n),,F/pIn(n))T. Although, (2.16) and (2.17) are closed systems on Ω, it should be noted that Ξ is not invariant under the full dynamics (2.2), as this is not even the case in a panmictic population.

Remark 2.3

F(ρ) and w˜(ρ) are Lyapunov functions for (2.16), and for (2.17), and the asymptotically stable equilibria of (2.16) are the local maxima of F(ρ). The internal equilibria of (2.16) are exactly the internal critical points of F(ρ), or of w˜(ρ), on ΔJ, and this holds for every lower-dimensional subsystem. They are precisely the solutions of (2.15). Comparison with Result 2.1 yields that ρˆ is an (asymptotically stable) equilibrium of (2.16) if and only if it gives rise to an (asymptotically stable) gene-frequency equilibrium of (2.2).

We designate equilibrium points of (2.16) by ρˆΩ. Every ρˆ gives rise to the set

Λˆ={pΔJ:ρ=ρˆ}Λ (2.18)

of gene-frequency equilibria of (2.2). We write ξˆinjn(n), ξˆin(n), and ξ¯ˆ(n) if ξinjn(n), ξin(n), and ξ¯(n), respectively, are evaluated at ρˆ.

We will need the following two assumptions on the gene-frequency equilibria:

The equilibrium pointsρˆ of (2.16) are isolated in Ω. (2.19)
ξˆinjn(n)>0 for every n and every in and jn. (2.20)

Clearly, (2.20) is satisfied if no single-locus genotype is lethal everywhere. A simple consequence is ξ¯ˆ(n)>0. We recall that we also assume

All recombination probabilities are positive. (2.21)

These three conditions hold generically.

3. Convergence to linkage equilibrium

In the absence of epistasis, generic global convergence to ΞΛ, hence to a stationary point in linkage equilibrium, was proved in N09b if there are two multiallelic loci (Theorem 4.6) or if the number of loci is arbitrary and there is no dominance (Theorem 4.14). Theorem 4.13 in N09b states (local) asymptotic stability of ΞΛ for multiple multiallelic loci. However, as Professor Nagylaki informed me, the induction proof of this theorem has a gap that may require extensive calculations to fill. (The gap does not affect the validity of the proof of his Theorem 4.14 for no dominance.) Motivated by this failure, I found a simple alternative method that is more general and yields global convergence to linkage equilibrium for the general multiallelic multilocus model without epistasis. It is based on a simple observation and employs the results in N09b on the gene-frequency dynamics (summarized in Result 2.1) as well as the result of Kun and Lyubich (1979, 1980) on convergence to equilibrium in the panmictic multilocus model without epistasis. Throughout this section, we assume absence of epistasis.

3.1. The general case

The crucial observation is the following. Because there is no epistasis, w¯α is a function of the gene frequencies only. For a given gene-frequency equilibrium ρˆΩ, we write w¯ˆα=w¯α(ρˆ) and obtain from (2.4) and (2.8a):

αcαwjk,αw¯ˆα=αcαw¯ˆαnujnkn,α(n)=nξˆjnkn(n). (3.1)

On Λˆ, every ξˆjnkn(n) is constant because every w¯ˆα is. Therefore, on Λˆ, the dynamics (2.2) of gamete frequencies becomes

pi=j,kRi,jkpjpkWjk, (3.2)

where

Wjk=nξˆjnkn(n) (3.3)

is constant. Hence, (3.2) is equivalent to a panmictic multilocus selection dynamics without epistasis. It describes the (panmictic) evolution of linkage disequilibria under selection. Global convergence of trajectories to an equilibrium point in linkage equilibrium now follows from Theorems 1 and 2 in Kun and Lyubich (1979); for a detailed treatment in English, see Section 9.6 in Lyubich (1992). (Application of their result requires the assumptions (2.20) and (2.21).) Because the equilibria of the gene-frequency dynamics are isolated, every ρ(t) converges as t to some ρˆΩ. Hence, for every solution p(t) of (2.2), there exists a ρˆ such that the limit points of p(t) are contained in Λˆ (Remark 2.2). Now LaSalle’s invariance principle (LaSalle, 1977, p. 10) yields the following theorem:

Theorem 3.1

If there is no epistasis and the assumptions (2.19), (2.20), (2.21) are satisfied, then every trajectory p(t) of the dynamics (2.2) of gamete frequencies converges to an equilibrium point that is in linkage equilibrium. Therefore, ΛΞ is the global attractor for solutions of (2.2).

Remark 3.2

Theorem 3.1 holds more generally if every trajectory ρ(t) converges to an equilibrium point ρˆ. However, if a manifold of equilibria exists in Ω, convergence has not been proved. A proof would require a generalization of the argument that establishes Theorem 9.6.3 in Lyubich (1992), which seems very challenging.

Because we have shown convergence to linkage equilibrium under generic conditions, the full dynamics of gamete frequencies as well as the equilibrium and stability structure can be determined by studying the much simpler gene-frequency dynamics (2.16). To formulate the result precisely, we denote the vector of all linkage disequilibria in the L-locus system by D=(Di,α), where Di,α is the linkage disequilibrium defined in Eq. (4.47) of N09b. Then we can write

p=(ρ,D). (3.4)

Corollary 3.3

Under the assumptions of Theorem 3.1 the following hold:

  • (a)

    pˆ=(ρˆ,Dˆ) is an equilibrium of (2.2) if and only if ρˆ is an equilibrium of (2.16) and Dˆ=0, i.e., if and only if pˆ=ΛˆΞ.

  • (b)

    pˆ is an asymptotically stable equilibrium of (2.2) if and only if ρˆ is an asymptotically stable equilibrium of (2.16). This equivalence also holds for globally asymptotically stable equilibrium points.

  • (c)

    If F(ρ) is concave and ρˆ is an isolated equilibrium of (2.16) that is either stable or internal, then (ρˆ,0) is the globally asymptotically stable equilibrium of (2.2).

Proof

  • (a)

    This is an immediate consequence of Theorem 3.1 and Remark 2.3.

  • (b)

    For given ρˆ, (ρˆ,0) is an asymptotically stable equilibrium of the dynamics (3.2) on Λˆ because Kun and Lyubich (1979) proved convergence to D=0 with the help of a Lyapunov function (see, e.g., Theorem 5.7 in LaSalle, 1977). In view of Remark 2.3, the statement about asymptotic stability is then an immediate consequence of Theorem 3.1 because asymptotic stability of ρˆ implies isolation. If the region of attraction of ρˆ with respect to (2.16) is Aˆ, the region of attraction of pˆ with respect to (2.2) is {pΔJ:ρAˆ}. This yields the statement on global asymptotic stability.

  • (c)

    Corollary 3.8 in N09b and the subsequent remark establish uniqueness of a stable gene-frequency equilibrium and its global stability for the gene-frequency dynamics. Convergence to linkage equilibrium follows from Theorem 3.1.  □

Corollary 3.3(c) applies in particular if there is no dominance (cf. Theorem 4.14 in N09b). For diallelic loci, it applies whenever fitness contributions at every locus are sublinear, i.e., if for every n and every α,

u12,α(n)12(u11,α(n)+u22,α(n)) (3.5)

holds. This assumption, meaning that there is either no dominance or the beneficial allele is (partially) dominant, implies that every w¯α is concave (Bürger, 2009c). Hence, F is concave.

Remark 3.4

Corollary 3.3(c) does not apply if fitness contributions within loci are multiplicative. Then an unstable internal equilibrium may exist, and two monomorphic equilibria may be simultaneously stable (see the remark following the proof of Theorem 4.8 in Bürger, 2009c).

3.2. Deme-independent degree of intermediate dominance

Here, we make an assumption about the genetic architecture. Following N09b, we say there is deme-independent degree of intermediate dominance (DIDID) if

uinjn,α(n)=ϑinjn(n)uinin,α(n)+ϑjnin(n)ujnjn,α(n) (3.6a)

holds for constants ϑinjn(n) such that

0ϑinjn(n)1andϑjnin(n)=1ϑinjn(n) (3.6b)

for every α, every n, and every pair in,jn. In particular, ϑinin(n)=12.

Obviously, DIDID covers complete dominance or recessiveness (ϑinjn(n)=0 or =1 if injn), and no dominance (ϑinjn(n)=12), but not multiplicativity. We also note that DIDID includes the biologically important case of absence of genotype-by-environment interaction. In general, F(ρ) is not concave under DIDID. Nevertheless, Theorem 3.14 in N09b establishes that under DIDID there exists exactly one stable gene-frequency equilibrium (point or manifold), and it is globally attracting. If an internal gene-frequency equilibrium exists, it is globally asymptotically stable.

Let us assume DIDID and that ρˆ is an internal gene-frequency equilibrium. It can be shown that

ξˆinjn(n)=ξ¯ˆ(n) (3.7)

holds for every n and for every in and jn (Appendix A). Therefore, (3.3) and (2.9) yield Wjk=nξ¯ˆ(n)=1. Hence, on Λˆ, the dynamics of gamete frequencies reduces to that under a pure recombination process,

pi=j,kRi,jkpjpk, (3.8)

for which convergence to linkage equilibrium is well known (Geiringer, 1944). Of course, it is sufficient to consider internal gene-frequency equilibria because, otherwise, restriction to the subsimplex that supports ρˆ yields the result. Under pure recombination, the linkage-equilibrium manifold Ξ is globally attracting at a uniform geometric rate. For generic initial conditions the rate of approach is 1rmin, where rmin is the smallest two-locus recombination rate (see Lyubich, 1992; Nagylaki, 1993; Nagylaki et al., 1999). Therefore, the assumption rmin>0, which is weaker than (2.21), is sufficient to establish convergence to linkage equilibrium. We summarize these results as follows.

Theorem 3.5

Suppose there is no epistasis and the assumptions (3.6), (2.19), and rmin>0 are satisfied. Then there exists a unique gene-frequency equilibrium ρˆ such that pˆ=(ρˆ,0)ΛΞ is a globally asymptotically stable equilibrium point of (2.2).

Remark 3.6

Because convergence of (3.8) to the linkage-equilibrium manifold Ξ occurs at a geometric rate (1rmin), under the assumptions of Theorem 3.5 an equilibrium pˆ=(ρˆ,0) of (2.2) is hyperbolic if ρˆ is a hyperbolic equilibrium of (2.16).

It would be interesting to identify the dominance patterns that lead to (3.7), hence, to (3.8). The set of these dominance patterns is a proper subset of all dominance patterns because for the internal equilibrium computed in Theorem 4.3 in Bürger (2009c), in general, the first locus does not satisfy (3.7).

4. The haploid Levene model

Here we consider a species that is haploid but reproduces sexually with recombination. General properties of the multilocus dynamics with viability selection in a panmictic population of haploids were derived by Kirzhner and Lyubich (1997). Various aspects of the one-locus haploid Levene model were studied by Strobeck (1979); see also Nagylaki and Lou (2008). Wiehe and Slatkin (1998) and Barton (2010) investigated haploid multilocus Levene models with certain forms of epistasis. As in the above-treated diploid case, and as in Wiehe and Slatkin (1998) and in one of Barton’s (2010) models, we assume that the life cycle consists of viability selection, dispersal, and recombination. Moreover, soft selection is assumed.

All unexplained notation is as in Section 2. In haploids, the constant fitness wi,α is assigned to gamete i in deme α. The mean fitness in deme α is simply w¯α=αwi,αpi. Because recombination may occur between haplotypes originating from different demes, the recursion relations for the gamete frequencies are given by

pi=j,k,α,βRi,jkcαpjwj,αw¯αcβpkwk,βw¯β. (4.1)

From now on we assume absence of epistasis and set

wi,α=nuin,α(n), (4.2a)
u¯α(n)=inuin,α(n)pin(n), (4.2b)
w¯α=iwi,αpi=nu¯α(n). (4.2c)

Furthermore, we define

ξin(n)=αcαw¯αuin,α(n) (4.3)

and observe that, with ξ¯(n) as in (2.8c), (2.9) continues to hold, i.e., 

nξ¯(n)=1. (4.4)

Without epistasis we have

αcαwi,αw¯α=αcαw¯αnuin,α(n)=nξin(n), (4.5)

and (4.1) simplifies to

pi=j,kRi,jkpjpknξjn(n)lξkl(l). (4.6)

Essentially the same, sometimes slightly simpler, calculations as in Sections 2 and 3 of N09b prove that the dynamics of gene frequencies has the form (2.10) and that Result 2.1 and Remarks 2.2 and 2.3 hold. An argument analogous to that yielding Corollary 3.8 in N09b shows that there exists exactly one stable gene-frequency equilibrium (point or manifold), and it is globally attracting. If an internal gene-frequency equilibrium exists, it is globally asymptotically stable.

For the rest of this section we assume (2.19). The results about the gene-frequency dynamics ensure that every ρ(t) converges to some ρˆΩ. Hence, for every solution p(t) of (4.1), there exists a ρˆ such that the limit points of p(t) are contained in Λˆ. (For generic initial conditions solutions are attracted by the set Λˆ that is generated by the (unique) maximum ρˆ of geometric-mean fitness w˜.) For a given internal gene-frequency equilibrium ρˆ, the equilibrium condition (2.15) together with (4.4) yields

nξˆin(n)=nξ¯ˆ(n)=1. (4.7)

Therefore (4.6) reduces to

pi=j,kRi,jkpjpk, (4.8)

which is the well-known dynamics of gamete frequencies under a pure recombination process. For (4.8) convergence to linkage equilibrium is well known and requires only rmin>0 instead of (2.21) (see Section 3.2). Thus, we have proved the following result:

Theorem 4.1

Suppose there is no epistasis and the assumptions (2.19) and rmin>0 are satisfied. Then there exists a unique gene-frequency equilibrium ρˆ such that pˆ=(ρˆ,0)ΛΞ is a globally asymptotically stable equilibrium point of (4.1).

An analog of Remark 3.6 applies. We also note that the haploid model is not equivalent to the diploid model without epistasis and multiplicative fitnesses within loci. Instead, it is equivalent the diploid model with additive fitnesses within gametes and multiplicative fitnesses between them (see Remark 3.4 and Kirzhner and Lyubich, 1997).

5. Maintenance of polymorphism

In this section, we study the maintenance of multilocus polymorphism in the Levene model. In addition to assuming absence of epistasis, we assume that the degree of dominance is intermediate at every locus and in every deme. Thus, overdominance and underdominance are excluded. In a panmictic population, no polymorphism is possible in the absence of epistasis if there is intermediate dominance (Bürger, 2009b, Proposition 3.2 and Corollary 3.4). Although the Levene model lacks population structure, the following result shows that, nevertheless, it harbors the potential for extensive multilocus polymorphism under such conditions:

Result 5.1

Assume an arbitrary number of multiallelic loci, Γ2 demes, and let all recombination rates be positive and fixed. Then there exists a nonempty open set of parameters such that for every parameter combination in this set, there is a unique, internal, asymptotically stable equilibrium point pˆ of the dynamics (2.2). This equilibrium is in linkage equilibrium and it is globally asymptotically stable.

Essentially, this result was proved in Bürger (2009b, Theorem 2.2, Remark 2.3 (iii), Remark 2.4). There, only convergence to quasi-linkage equilibrium was shown. Theorem 3.1 demonstrates convergence to linkage equilibrium. More generally, Theorem 2.2 in Bürger (2009b) shows that Result 5.1 holds for arbitrary ergodic, i.e., irreducible and aperiodic, migration patterns, and not only for the Levene model. Such extensive polymorphism can occur if selection is weak relative to migration and recombination. The constructive proof in Bürger (2009a,b) requires balancing selection and a certain form of average overdominance across demes that can be achieved only if the direction of selection and the degree of dominance vary among demes.

If, however, the degree of dominance is not only intermediate within demes, but there is DIDID, generically, at most Γ1diallelic loci can segregate (Proposition 3.18 in N09b). If, in addition, selection is sufficiently weak, then no polymorphism can be maintained (Proposition 2.6 and Remark 2.7 in Bürger, 2009b). Obviously, these results are in sharp contrast to Result 5.1 and underline that dominance plays an important role in maintaining genetic variation in a subdivided population.

For the rest of this section, we assume diallelic loci. Our main aim is to prove that the above mentioned upper bound Γ1 is indeed assumed on an open set of parameters. As a consequence, such polymorphic equilibria are structurally stable. We also show that if LΓ and if, as is nongeneric but may occur under additional constraints on the parameters, there exists an internal equilibrium, then it is a manifold.

For diallelic loci, we write p(n) for the frequency of A1(n), and 1p(n) for the frequency of A2(n). The vector of gene frequencies ρ can then simply be represented as

ρ=(p(1),,p(L))TΩ=[0,1]L, (5.1)

where the definition of Ω is modified. The condition (2.15) for internal gene-frequency equilibria reduces to

ξ1(n)ξ2(n)=αcαw¯α(u1,α(n)u2,α(n))=0for every n, (5.2)

where the fitness contributions of the two alleles at locus n in deme α become

u1,α(n)=p(n)u11,α(n)+(1p(n))u12,α(n) (5.3a)

and

u2,α(n)=p(n)u12,α(n)+(1p(n))u22,α(n). (5.3b)

The contribution to mean fitness of locus n in deme α is

u¯α(n)=(p(n))2u11,α(n)+2p(n)(1p(n))u12,α(n)+(1p(n))2u22,α(n). (5.4)

The allele-frequency dynamics, (2.16) or (2.17), simplifies to

Δp(n)=p(n)(1p(n))αcαw¯α(u1,α(n)u2,α(n)) (5.5a)
=12p(n)(1p(n))F(ρ)p(n). (5.5b)

Remark 5.2

From (5.3), we obtain

u1,α(n)u2,α(n)=p(n)(u11,α(n)u12,α(n))+(1p(n))(u12,α(n)u22,α(n)). (5.6)

Therefore, u11,α(n)u12,α(n)u22,α(n) and u11,α(n)>u22,α(n) yield u1,α(n)u2,α(n)>0 if 0<p(n)<1. If u1,α(n)u2,α(n)>0 for every α, then(5.5a) and Remark 2.3 imply that allele A1(n) becomes fixed. This confirms the intuition that, in the absence of overdominance and underdominance, an allele that is the best in every deme is fixed. Hence, a locus can be polymorphic only if each of the alleles is the best in at least one deme.

For two alleles, also the definition of DIDID simplifies. There is DIDID if for every nL, there exist constants ϑ(n) such that for every αG,

u12,α(n)=ϑ(n)u11,α(n)+(1ϑ(n))u22,α(n) (5.7a)

holds, where

0ϑ(n)1. (5.7b)

If (5.7) holds, the condition (5.2) for an internal gene-frequency equilibrium simplifies to

ξ11(n)ξ22(n)=αcαu11,α(n)u22,α(n)w¯α=0for every nL, (5.8)

because

u1,α(n)u2,α(n)=[ϑ(n)+p(n)(12ϑ(n))](u11,α(n)u22,α(n)), (5.9)

and ϑ(n)+p(n)(12ϑ(n))>0 if 0<p(n)<1. We shall write ϑ=(ϑ(1),,ϑ(L))T[0,1]L. Therefore, the dynamics of gamete frequencies (2.2) is uniquely determined by the linkage map and by the parameters c1,,cΓ1, ϑ, u11,α(n) and u22,α(n), where αG and nL.

Our main result in this section is the following.

Theorem 5.3

Assume LΓ1 diallelic loci with DIDID, i.e.,  (5.7). Let ϑ[0,1]L be arbitrary but fixed. Then there exists an open nonempty set U of parameters such that for every parameter combination in U, there is a unique, internal, asymptotically stable equilibrium point pˆ of (2.2). This equilibrium is in linkage equilibrium and globally attracting. The set U is independent of the choice of the recombination rates.

Together with Proposition 3.18 in N09b, this result implies that under DIDID and if LΓ, at most Γ1 diallelic loci can be maintained polymorphic for an open set of parameters. The proof of the above theorem is based on

Proposition 5.4

Under the assumptions of Theorem 5.3 there exists an open, nonempty subset U of the parameter space such that for every element in Uthere is an isolated, internal, asymptotically stable equilibrium point ρˆ of the gene-frequency dynamics (5.5a). It is the unique solution of (5.8) in intΩ=(0,1)L.

The basic idea for the proof of Proposition 5.4 is to find parameter combinations such that (12,,12)TRL is an equilibrium of the gene-frequency dynamics (5.5a), and then to show that every sufficiently small perturbation of the parameters still yields an (isolated) internal equilibrium. The proof of this proposition is rather technical and constructs the set U, which is the same in Theorem 5.3 and Proposition 5.4. A more detailed formulation of the proposition and its proof are given in the Appendix.

Proof of Theorem 5.3

Statements (a) and (b) of Corollary 3.3 show that for every ρˆ in the open set U of Proposition 5.4, pˆ=(ρˆ,0) is the desired unique and internal equilibrium point of (2.2). Because the proof of Proposition 5.4 is based entirely on the gene-frequency dynamics (5.5a), the construction of U is independent of the linkage map.  □

Now we formulate and prove our second main result of this section.

Theorem 5.5

Assume LΓ diallelic loci with DIDID. If there exists an internal equilibrium ρˆ of the gene-frequency dynamics (5.5a), then there is a manifold of equilibrium points containing it. Generically, this manifold has dimension LΓ+1.

Proof

Let V=(Vnα) denote the L×Γ matrix with entries Vnα=u11,α(n)u22,α(n). By assumption, there exists an internal equilibrium ρˆ of the gene-frequency dynamics (5.5a). If we write w¯ˆα=w¯α(ρˆ), (5.8) informs us that

αVnαcαw¯ˆα=αcαw¯ˆα(u11,α(n)u22,α(n))=0,nL, (5.10)

holds. In matrix form this reads

Vxˆ=0, (5.11)

where xˆ=(c1/w¯ˆ1,,cΓ/w¯ˆΓ)T. Generically, we therefore have dim(kerV)=1 and, because LΓ,

rankVT=rankV=min{L,Γ}1=Γ1. (5.12)

Although scalar multiples of xˆ also solve (5.11), scalar multiples of (w¯ˆ1,,w¯ˆΓ)T do not give rise to further solutions of (5.10) because, by Result 2.1, geometric-mean fitness w˜ is maximized, hence constant, at gene-frequency equilibria.

Therefore, it is sufficient to determine the dimension of the solution space of the system

w¯α(ρ)=w¯ˆα,αG, (5.13)

for the given vector (w¯ˆα). This is a system of Γ quadratic equations in the L unknowns p(n). We transform it into a system of Γ linear equations in the L unknowns p(n) by using the nonlinear transformation f=(f(1),,f(L))T defined by

z(n)=f(n)(p(n))=2ϑ(n)p(n)+(12ϑ(n))(p(n))2. (5.14)

This is the specification of (3.20) in N09b for diallelic loci. By Lemma 3.1 in Nagylaki (2009a), every f(n) is a homeomorphism on [0,1]. Denoting z=(z(1),,z(L))T, we obtain from (3.23) in N09b by a simple calculation,

w¯α(ρ)=W¯α(z)=nu22,α(n)+n(u11,α(n)u22,α(n))z(n)=nu22,α(n)+nVnαz(n), (5.15)

which is linear in z for every αG (cf. Lemma 3.13 in N09b). Therefore, the system (5.13) of quadratic equations in ρ is equivalent to the linear system

W¯α(z)=w¯ˆα,αG, (5.16)

in z of the same dimension. This can be written as

VTz=b, (5.17)

where b=(b1,,bΓ)T and bα=w¯ˆαnu22,α(n). Because (5.17) has the nontrivial solution zˆ=f(ρˆ), standard linear algebra shows that the dimension of its solution space, hence that of (5.13), generically equals

dim(kerVT)=LrankVT=L(Γ1)=LΓ+1. (5.18)

Here, the second equality follows from (5.12) which holds generically.  □

Because under the conditions of the above theorem a manifold of gene-frequency equilibria exists, it cannot be inferred that convergence to linkage equilibrium occurs although this seems likely. The reason is that the application of Theorem 3.1 requires isolated gene-frequency equilibria or at least a proof that every ρ(t) converges (to a single point); cf. Remark 3.2.

6. A quantitative-genetic model

We apply the above results and those of N09b to a simple quantitative-genetic model and discuss their implications. To this end, we consider a quantitative trait that is determined additively (i.e., without epistasis) by L2 multiallelic loci. Intermediate dominance and genotype-by-environment (G×E) interaction are admitted. Let the multilocus genotype ij have the trait value

X=Xij,α=nγinjn,α(n) (6.1)

in deme α, where γinjn,α(n) is the effect on the trait in deme α assigned to the single-locus genotype Ain(n)Ajn(n). To avoid degeneracy, we posit γinin,α(n)γjnjn,α(n) if injn.

We assume that the trait is under linear directional selection in each of the Γ1 demes, i.e., 

Wα(X)=1sαX, (6.2)

where the range of possible values X and the selection coefficients sα0 are constrained such that Wα(X)>0 for every αG. Therefore, we define the single-locus contribution of Ain(n)Ajn(n) to fitness by

uinjn,α(n)=1Lsαγinjn,α(n). (6.3)

Then, as in (2.4),

wij,α=Wα(Xij,α)=nuinjn,α(n). (6.4)

A glance at (6.3) reveals that, for given L, Γ, and selection coefficients sα, (6.3) establishes a one-to-one correspondence between the single-locus fitness contributions uinjn,α(n) and the single-locus trait effects γinjn,α(n).

Therefore, this model of linear selection on a quantitative trait is as general as the nonepistatic model introduced in Section 2 and studied in Sections 3 and 5. Obviously, no generality would be lost by setting sα=1 for every α. However, we refrain from doing so. Moreover, it is immediate from (6.3) that no or intermediate dominance at the trait level is equivalent to no or intermediate dominance, respectively, at the fitness level. The dominance relations are reversed in demes with sα>0.

In analogy to (3.6), we say there is DIDID on the trait level if there exist constants ϑinjn(n) such that

γinjn,α(n)=ϑinjn(n)γinin,α(n)+ϑjnin(n)γjnjn,α(n) (6.5a)

and

0ϑinjn(n)1andϑjnin(n)=1ϑinjn(n) (6.5b)

for every α, every n, and every pair in,jn. Trivially, the relation (6.3) transforms DIDID on the trait level to DIDID on the fitness level, and vice versa; the constants ϑinjn(n) are the same for trait and fitness.

Genotype-by-environment interaction is absent on the trait level if there exist constants γinjn(n) and γα such that

γinjn,α(n)=γinjn(n)+γα (6.6)

holds for every α, every n, and every pair in,jn. Because we assume intermediate dominance at the trait level, (6.6) implies the existence of constants ϑinjn(n) satisfying (6.5b) such that

γinjn(n)=ϑinjn(n)γinin(n)+ϑjnin(n)γjnjn(n). (6.7)

Therefore, (6.6) implies (6.5). Summarizing, there is no G×E interaction on the trait level if and only if there is DIDID and (6.6) holds for all homozygous single-locus effects. A further simple consideration establishes equivalence of absence of G×E interaction at the trait and the fitness level.

An important consequence of this one-to-one relation between fitness and trait value is that results that hold generically for the model defined in terms of the fitness contributions uinjn,α(n) also hold generically for this model of a quantitative trait under linear selection. Especially, Result 5.1 and Theorems 5.3 and 5.5 apply. Thus, if the number of demes is at least two, for an open set of parameters a globally asymptotically stable equilibrium exists such that

  • (i)

    an arbitrary number of multiallelic loci is polymorphic;

  • (ii)

    up to Γ1 diallelic loci are polymorphic if there is DIDID.

Parameter combinations yielding (i) can be constructed by applying the procedure in Proposition 2.1 in Bürger (2009b) with m2=1m1=1c2=c1 and substituting sαγinjn,α(n) for the parameters uinjn,α(n) used there (here, mα is the backward migration rate into α). In fact, it is easy to show that this is possible if γinin,α(n)=γinin(n), i.e., in the absence of interactions between homozygous effects and environment.

An unresolved question concerns the number of alleles that can be maintained at a locus if there is DIDID. In N09b (Corollary 3.9) it is shown that if I¯=1LnIn>Γ, then generically no internal gene-frequency equilibrium exists. Therefore, if the number of alleles is the same at every locus (In=I for every n), the number of demes is a generic upper bound on the number I of alleles per locus. Clearly, a generalization of (ii) to multiallelic loci would be desirable.

Example 6.1

We specialize the model (6.1)–(6.4) by assuming absence of G×E interaction, i.e., (6.6). Therefore, (6.5) holds and we define

ϑin(n)=jnϑinjn(n)pjn(n). (6.8)

From (2.7) and (2.6) we obtain by straightforward calculations employing (6.3), (6.5) and (6.6):

uin,α(n)=1Lsαγαsα(γinin(n)ϑin(n)+jnϑjnin(n)γjnjn(n)pjn(n)), (6.9a)
u¯α(n)=1Lsαγα2sαjnϑjn(n)γjnjn(n)pjn(n), (6.9b)

and

uin,α(n)u¯α(n)=sαϕin(n), (6.10)

where

ϕin(n)=jn(2ϑjn(n)ϑjnin(n))γjnjn(n)pjn(n)ϑin(n)γinin(n). (6.11)

Therefore, the condition (2.15) for internal gene-frequency equilibria becomes

ϕin(n)αcαsαw¯α=0 (6.12)

for every n and in. We conjecture that, at least generically, for every n, ϕin(n)0 holds for at least one in (in fact, for all but one). If this is true, then

αcαsαw¯α=0 (6.13)

is a necessary and sufficient condition for the existence of an internal gene-frequency equilibrium in the absence of G×E interaction.

In the absence of dominance, this is easy to prove because we have ϑin(n)=ϑinjn(n)=12 for every n and every in,jn, hence

ϕin(n)=12(jnγjnjn(n)pjn(n)γinin(n)). (6.14)

Obviously, for each n, ϕin(n)=0 can hold for at most one in. Thus, (6.13) is the condition for an internal gene-frequency equilibrium.

Example 6.2

Here, we specialize to diallelic loci and use the notation introduced in Sections 2 and 5, e.g., p(n)=p1(n) and ϑ=ϑ12. By Corollary 3.3, the equilibrium and stability structure can be determined by studying the relatively simple system (5.5a), where a straightforward calculation yields

u1,α(n)u2,α(n)=sα[γ22,α(n)γ12,α(n)+p(n)(2γ12,α(n)γ11,α(n)γ22,α(n))]. (6.15)

In the absence of G×E interaction, i.e., if (6.6) holds, (6.15) simplifies to

u1,α(n)u2,α(n)=sα(γ22(n)γ11(n))[ϑ+p(n)(12ϑ)]. (6.16)

Because ϑ+p(n)(12ϑ)>0 if 0<p(n)<1, the condition (5.2) for an internal gene-frequency equilibrium again collapses to the single equation (6.13) in the L unknowns p(n).

Because w¯α is quadratic in p(n), (6.13) is equivalent to a single polynomial equation (of degree 2Γ2) in the L variables p(n). Therefore, if an internal solution ρˆ exists, there is a manifold of solutions of (generic) dimension L1. For more than two demes, this exceeds the dimension LΓ+1 derived in Theorem 5.5. As a consistency check, we observe that the matrix V defined in the proof of Theorem 5.5 has the entries Vnα=sαϕ1(n)/(1p(n)) because ϕ1(n)=(1p(n))(γ22(n)γ11(n))[ϑ+p(n)(12ϑ)]. Therefore, it has rank 1 and a calculation as in (5.18) yields L1. Thus, if Γ>2, the equilibrium structure in the absence of G×E interaction is even more highly degenerate than under DIDID alone.

The above examples demonstrate that (6.13) is a necessary and sufficient condition for an internal gene-frequency equilibrium if there is no G×E interaction and either there is no dominance or loci are diallelic. In the absence of G×E interaction, an internal equilibrium cannot exist if all selection coefficients sα have the same sign (for a more general statement, see Remark 5.2). If, for instance, sα>0 for every α, then the gamete with the largest genotypic value has maximum fitness and becomes fixed (cf. N09b, Proposition 3.15). If the selection coefficients sα vary in sign, it is always possible to choose the relative demes sizes cα such that (6.13) has an internal solution ρˆ, hence a manifold of solutions. In fact, there is an open set of such parameters combinations (c1,,cΓ1). For other deme sizes, there is at least one isolated, asymptotically stable equilibrium. At such a stationary state, at most Γ1 loci can be polymorphic.

We summarize the main results of this section.

Corollary 6.3

Assume diallelic loci and let all recombination rates be positive and fixed. Further, assume the selection model given by (6.1)–(6.4), i.e., linear directional selection on a quantitative trait that is determined additively by L diallelic loci exhibiting intermediate dominance.

  • (a)

    There is an open subset of the full parameter space for which a globally asymptotically stable internal, hence fully polymorphic, equilibrium exists.

  • (b)

    If there is DIDID, then, generically, at most Γ1 loci can be polymorphic at an equilibrium. An internal equilibrium exists for an open subset of parameters if and only if LΓ1 . If LΓ and an internal equilibrium exists, then it is a manifold of equilibria which, generically, has dimension LΓ+1.

  • (c)

    If there is no G×E interaction, then there is an open subset of parameters for which a unique internal equilibrium exists. If an internal equilibrium exists, it is the manifold of equilibria that solves Eq. (6.13). Generically, its dimension is L1.

  • (d)

    Every equilibrium is in linkage equilibrium.

Among others, this result demonstrates that biologically reasonable but special assumptions can lead to nongeneric, and nonrobust, model behavior. Therefore, caution is necessary when generalizing conclusions obtained from specific models.

A detailed and rather complete analysis of the diallelic two-locus case in two demes was performed in Bürger (2009c).

7. Weak epistasis

Several of our results can be extended to weak epistasis. By weak epistasis we mean that there is a small ϵ>0, such that the fitness scheme has the form

wij,α=nuinjn,α(n)+sij,α, (7.1a)

where

|sij,α|<ϵ for every i,j, and α. (7.1b)

First, we generalize Theorem 3.1 and show that global convergence to quasi-linkage equilibrium occurs under weak epistasis. Then we briefly point out some applications for the maintenance of polymorphism. Throughout, we assume (2.19)–(2.21).

7.1. Convergence

As noted by Nagylaki et al. (1999), the assumption of hyperbolicity of every equilibrium is not robust to small perturbations because limit sets need not change continuously. What has good behavior under perturbations is the set of chain-recurrent points, which contains the limit sets of all orbits (Conley, 1978; Akin, 1993). For the Levene model without epistasis, we have the following:

Lemma 7.1

In the case of no epistasis, the only chain-recurrent points of (2.2) are its equilibria.

Proof

By Result 2.1, the Lyapunov function F takes only finitely many values on the set Λ of gene-frequency equilibria. Therefore, Theorem 3.16 in Akin (1993) shows that every chain-recurrent point is contained in Λ. By our general assumption (2.19), Λ has finitely many components Λˆ. Since, by (3.1), on Λˆ the dynamics reduces to the panmictic multilocus dynamics (3.2) with no epistasis, Lemma 2.2 of Nagylaki et al. (1999) yields the assertion.  □

To formulate the main result of this section, we introduce the following notation. We write Σ0ΔJ for the set of equilibria of (2.2) if there is no epistasis (ϵ=0), and we write Σw(ϵ)ΔJ for the set of equilibria for a fitness scheme wij,α that satisfies (7.1) with ϵ>0.

Theorem 7.2

Let fitness contributions uinjn,α(n) be given, such that without epistasis every equilibrium of (2.2) is hyperbolic. Further, let all recombination probabilities be fixed (and positive), and let ϵ>0 be sufficiently small. Then for every set of fitnesses wij,α satisfying (7.1) the following holds:

  • (a)

    The set of equilibria Σ0 contains only isolated points, as does the set of equilibria Σw(ϵ) . As ϵ0 in (7.1), each equilibrium in Σw(ϵ) converges to the corresponding equilibrium in Σ0.

  • (b)

    In the neighborhood of each equilibrium point in Σ0, there exists exactly one equilibrium point in Σw(ϵ) . The stability of each equilibrium in Σw(ϵ) is the same as that of the corresponding equilibrium in Σ0 ; i.e., each pair is either asymptotically stable or unstable.

  • (c)

    Every solution p(t) of (2.2) converges to one of the equilibrium points in Σw(ϵ).

Proof

Parts (a) and (b) are consequences of the implicit function theorem and the Hartman–Grobman theorem, and follow immediately from Theorem 4.4 in Karlin and McGregor (1972). The only issue that remains to be shown in (b) is that perturbed equilibria do not leave ΔJ (cf. Remark 2.3 in Nagylaki et al., 1999). This follows from the explicit characterization

pin(n)=0orξin(n)=ξ¯(n) (7.2)

of equilibria if ϵ=0; see (2.16). Let pˆ be some equilibrium for ϵ=0, and denote the set of alleles present at locus n by Jn={in:pˆin(n)>0}. The face of ΔJ determined by pi=0 if inJn is invariant under the dynamics (2.2) for every ϵ>0, as can be seen from the alternative representation

pi=αcα(piwi,αw¯αDi,α), (7.3)

where Di,α is a measure of linkage disequilibrium in gamete i in deme α (see Eq. (4.47) in N09b). Therefore, the equilibrium pˆ, which is hyperbolic for ϵ=0, persists in this face for ϵ>0.

(c) Upon replacing the reference to Lemma 2.2 in the proof of Theorem 2.3 in Nagylaki et al. (1999) by one to the above Lemma 7.1, their proof applies unaltered. However, our assumption (7.1) is slightly weaker than their assumption (2.1). Our assumption is uniform in the sense that we first choose ϵ>0 and then admit all parameter combinations wij,α satisfying (7.1), whereas following their formulation, we first had to fix a set of sij,α and then ϵ>0. Because Corollary 32 in Akin (1993), which is used in the proof of Theorem 2.3 in Nagylaki et al. (1999), admits this greater generality, their proof indeed applies.  □

Remark 7.3

  • 1.

    Hyperbolicity of equilibria of (2.2) with ϵ=0 is a generic property. For a single panmictic deme, this was proved by Nagylaki et al. (1999, Appendix A). In general, (2.2) shows that the Jacobian (at any point) is simply a linear combination of single-deme Jacobians.

  • 2.

    With weak epistasis, equilibria may be in linkage disequilibrium, but deviate from Ξ only to order O(ϵ).

  • 3.

    If we define w˜(p)=α[w¯α(p)]cα if ϵ>0, then Δw˜(p)>0 holds if p is bounded away from the set of gene-frequency equilibria Λ.

We conjecture that an equilibrium pˆ=(ρˆ,0) of (2.2) with ϵ=0 is hyperbolic if ρˆ is a hyperbolic equilibrium of (2.16). In the presence of selection, in general, this cannot be inferred from the results or proofs in Lyubich (1992) because it has not been shown that convergence to linkage equilibrium occurs at a geometric rate. If there are only two loci, or if there is DIDID, or if selection acts on haploids, then Theorem 4.6 in N09b, or Remark 3.6, or the remark following Theorem 4.1, respectively, establish the conjecture. It can also be established for three diallelic loci (Bürger, unpublished).

7.2. Maintenance of polymorphism

With the help of Theorem 7.2, several of the results on the maintenance of polymorphism can be generalized to weak epistasis. Of course, Result 5.1 generalizes to weak epistasis. The only difference then is that equilibria are not necessarily in linkage equilibrium but in quasi-linkage equilibrium. In fact, this was already proved in Bürger (2009b, Theorem 2.2 and Remark 2.4). Proposition 3.18 in N09b and Theorem 5.3 have the following generalization.

Theorem 7.4

Assume weak epistasis and diallelic loci with DIDID.

  • (a)

    If LΓ, at most Γ1 loci can be maintained polymorphic for an open set of parameters.

  • (b)

    If LΓ1 and ϑ[0,1]L in (5.7) is arbitrary but fixed, then there exists an open nonempty set W of fitness schemes satisfying (7.1) and of deme proportions (c1,,cΓ1), such that for every parameter combination in W, there is aunique, internal, asymptotically stable equilibrium point pˆ of (2.2). This equilibrium is in quasi-linkage equilibrium and globally attracting. The set W is independent of the choice of the recombination rates.

Proof

  • (a)

    Proposition 3.18 in N09b shows that generically, intΛ=. Because, in the proof of Theorem 7.2(b), it was shown that boundary equilibria persist in the face where they exist for ϵ=0, Theorem 7.2 yields the assertion.

  • (b)

    This is a simple consequence of Theorems 5.3 and 7.2.  □

Also statements (a) and (b) in Corollary 6.3 can be generalized to weak epistasis. In (d), quasi-linkage equilibrium can be stated.

8. Discussion

The analysis of multilocus systems is greatly simplified, and often only feasible, if linkage equilibrium or, at least, quasi-linkage equilibrium can be assumed (e.g. Karlin and Liberman, 1979; Turelli and Barton, 1990; Christiansen, 1999; Bürger, 2000). For the classical multilocus selection model, it has long been known that global convergence to linkage equilibrium occurs if there is no (additive) epistasis (Kun and Lyubich, 1979, 1980; Lyubich, 1992). If either epistasis or selection is weak, generic convergence to a stationary point in quasi-linkage equilibrium has been proved (Nagylaki, 1993; Nagylaki et al., 1999). The latter results can be generalized to subdivided populations as follows. If migration and epistasis are weak (relative to recombination) or if selection is weak (relative to migration and recombination), then global convergence to a stationary point in quasi-linkage equilibrium occurs generically (Bürger, 2009a). However, if migration is moderately strong, stable linkage disequilibrium may persist in the absence of epistasis (Li and Nei, 1974).

For the multilocus Levene model without epistasis, and under the generic assumption that all equilibria are isolated, Nagylaki (2009b) proved global convergence to an equilibrium point in linkage equilibrium if there are two (multiallelic) loci or if there is no dominance. He conjectured global convergence for multiple multiallelic loci, even if equilibrium points are not isolated. In Section 3, we establish this conjecture under the generic assumption of isolated gene-frequency equilibria (Theorem 3.1). For the haploid Levene model, a slightly stronger result is demonstrated in Section 4.

Whenever convergence to an equilibrium point in linkage equilibrium occurs in the Levene model without epistasis, the analysis of the model simplifies greatly. Then it is sufficient to study the system of recursion equations (2.16), which describes the dynamics of gene frequencies under the assumption of linkage equilibrium. Corollary 3.3 summarizes the most important conclusions that can be drawn from investigating (2.16) and its Lyapunov function F=lnw˜(2.14). Further useful properties of the gene-frequency dynamics may be found in Section 3 of N09b.

With the help of perturbation methods (Karlin and McGregor, 1972; Nagylaki et al., 1999), the convergence result Theorem 3.1 can be extended to weak epistasis. Then equilibria are not necessarily in linkage equilibrium, but they are in quasi-linkage equilibrium (Theorem 7.2). Therefore, many results that hold without epistasis can be extended to weak epistasis. For an example, see Theorem 7.4.

In Section 5, we apply our convergence results to investigate the maintenance of multilocus polymorphism. Overdominance and underdominance (within demes) are excluded. It has been shown previously that, if there is intermediate dominance in every deme, arbitrarily many multiallelic loci can be fully polymorphic on an open set of parameters (Result 5.1). In the proof, this open set was constructed by assuming, among others, that at each locus and in each deme, the fitter alleles are partially dominant. Because an internal equilibrium requires that the direction of selection is different in at least two demes, the proof of this result invokes G×E interaction. In N09b it was shown that for DIDID, generically, the number of segregating loci is strictly less than the number Γ of demes. Hence, some form of G×E interaction is necessary to maintain Γ or more loci polymorphic. Theorem 5.3 complements Nagylaki’s result and shows that with DIDID, Γ1 loci can indeed be maintained at a stable equilibrium for an open set of parameters. If LΓ and, as is nongeneric, an internal equilibrium exists, Theorem 5.5 establishes that a manifold of equilibria exists. Generically, its dimension is LΓ+1. If there is DIDID and selection is sufficiently weak, then no polymorphism is possible in the Levene model without epistasis.

Of course, DIDID itself is a nongeneric property. Perturbations of isolated equilibrium points can be studied using, for instance, Karlin and McGregor’s (1972) method of small parameters which requires that equilibria are hyperbolic. Thus, if there is an asymptotically stable equilibrium under DIDID, there will be an asymptotically stable equilibrium in its neighborhood if the deviation from DIDID is small. A much stronger result can be inferred from Lemma 7.1 and the proof of Theorem 7.2. If all equilibria are hyperbolic when there is DIDID, under small deviations from DIDID, the global dynamics will remain qualitatively unchanged.

In Section 6, these results are applied to a quantitative trait that is determined by L additive loci (i.e., without epistasis but with intermediate dominance). If the trait is under linear directional selection in every deme, the resulting model is equivalent to the general nonepistatic model studied in this paper. Hence, the results of Section 5 concerning the maintenance of variation apply unaltered. Interestingly, under the assumption of no G×E interaction on the trait level, there is an open set of parameters such that an internal equilibrium exists. However, because absence of G×E interaction implies DIDID, this is a manifold of equilibria. In fact, in this case, its dimension is L1 instead of LΓ+1. Thus, if there are more than two demes, the assumption of no G×E interaction in the Levene model with linear selection, leads to a highly degenerate equilibrium structure. In such a case, an arbitrarily small perturbation of the parameters that introduces G×E interaction, may drastically change the equilibrium structure and the dynamics (see Bürger, 2009c). As already mentioned in the Introduction, these results (Corollary 6.3) should serve as a warning when studying models under highly specialized assumptions, even if they are biologically well motivated.

The fact that the maximum possible number of polymorphic loci may be constrained by the number of demes demonstrates that, even in the absence of epistasis and in the presence of linkage equilibrium, one cannot simply extrapolate single-locus results to multiple loci. It would be of considerable interest to find out for which genetic architectures and migration patterns such constraints occur.

It remains an unresolved problem how many alleles at how many loci can be maintained segregating under the assumption of DIDID or under the stronger assumption of no G×E interaction. Theorem 3.14 and Corollary 3.9 in N09b imply that if there is DIDID and the number of alleles is the same at each locus, then Γ is a generic upper bound on the number of segregating alleles per locus.

Eventually, one would like to quantify how frequent polymorphism is in the Levene model without epistasis, and how this depends on the selection scheme and the dominance relations. Numerical results for two diallelic loci and two demes show that for arbitrary intermediate dominance, the volume of the parameter space in which a full polymorphism is maintained varies between about 12% (for very weak selection) and about 16% (for very strong selection). Under the assumption of sublinear fitnesses (see (3.5) and Corollary 3.3(c)), these numbers increase to 22% and 36%, respectively. In the absence of G×E interaction, they are 0% and 43%, respectively (Bürger, 2009c). With increasing number of alleles or loci, the volume of parameter space in which a stable internal equilibrium exists will certainly decrease rapidly, at least if the parameter range is unconstrained. Still, the set of parameter combinations for which a significant fraction of loci can be maintained polymorphic may be quite large.

Therefore, we expect that spatially varying selection has the potential to maintain considerable multilocus polymorphism, especially if locally beneficial alleles tend to be partially dominant. Properties of the genetic architecture, such as the dominance pattern or the presence or absence of (certain forms of) G×E interaction, may greatly constrain this potential. In view of this, it seems worthwhile to study phenotypic plasticity and the evolution of genetic architecture in a spatial context (see e.g. Zhivotovsky et al., 1996; de Jong, 1999; Otto and Bourguet, 1999; de Jong and Gavrilets, 2000; van Doorn and Dieckmann, 2006).

Whereas DIDID plays an important role in the Levene model in constraining the amount of polymorphism, this is not necessarily so for general migration patterns. Peischl (2010) studied a single-locus model with two demes and general migration. He proved that, if there is DIDID, three alleles can be maintained at a stable equilibrium. His numerical computations suggest that more than three alleles can be maintained for an open set of parameters. His results have interesting interpretations in the context of the coexistence of specialists and generalists. Further, it is yet unknown whether DIDID also plays a prominent role in the Levene model with epistasis, for instance, if there is spatially varying stabilizing selection on a quantitative trait. This reconfirms the conviction that studying the role of genetic architecture for the maintenance of polymorphism in a spatial context seems to be a worthwhile enterprise.

Acknowledgments

The author is most grateful to Professor Thomas Nagylaki for extensive conversation and sharing of unpublished results. Many thanks go to Tom Nagylaki, Nick Barton, and an anonymous reviewer for perceptive comments on previous versions of this paper. This work was supported by a grant of the Vienna Science and Technology Fund (WWTF) and by grant P21305 of the Austrian Science Fund (FWF).

Appendix A. Proof of Eq. (3.7)

Because we assume DIDID (3.6), (2.8) yields

ξinjn(n)=ϑinjn(n)ξinin(n)+ϑjnin(n)ξjnjn(n) (A.1)

and

ξin(n)=ξinin(n)jnϑinjn(n)pjn(n)+jnϑjnin(n)pjn(n)ξjnjn(n). (A.2)

For every locus n, we introduce the nonnegative In×In matrices M(n), D(n), P(n), and S(n) as follows:

M(n)=D(n)P(n)+S(n), (A.3)

where

[D(n)]injn=ϑjnin(n), (A.4)

P(n) is diagonal with

[P(n)]inin=pin(n), (A.5)

and S(n) is diagonal with

[S(n)]inin=jnϑinjn(n)pjn(n). (A.6)

A simple calculation shows that M(n) is stochastic:

jn[M(n)]injn=(jnϑjnin(n)pjn(n)+jnϑinjn(n)pjn(n))=jn(ϑjnin(n)+ϑinjn(n))pjn(n)=1. (A.7)

Because we consider only internal equilibria, we have pjn(n)>0 for every h and jn. Hence, P(n) is nonsingular. Furthermore,

[S(n)]ininϑinin(n)pin(n)=12pin(n)>0, (A.8)

whence S(n) is also nonsingular.

Lemma A.1

If ρˆ is a gene-frequency equilibrium and M(n)=M(n)(ρˆ) is nonsingular, then

ξˆinjn(n)=ξ¯ˆ(n) holds for every in and jn. (A.9)

Proof

Let x(n)=(ξˆ11(n),,ξˆInIn(n))T. Then, in vector form, (A.2) reads

ξˆin(n)=[S(n)x(n)]in+[D(n)P(n)x(n)]in=[M(n)x(n)]in, (A.10)

where ξˆin(n), x(n), and M(n) all depend on ρˆ. Since the internal gene-frequency equilibria ρˆ are precisely the solutions of (2.15), they are precisely the solutions of

M(n)x(n)=ξ¯ˆ(n)1(n), (A.11)

where 1(n)=(1,,1)TRIn. Because M(n) is stochastic, x(n)=ξ¯ˆ(n)1(n) is a solution of (A.11). If M(n) is nonsingular, for any given ξ¯(n), it is the unique solution. Therefore, ξˆinin(n)=ξ¯ˆ(n) is satisfied for every in. The assertion now follows from (A.1) and (3.6b).  □

Our aim now is to show that M(n) is nonsingular. We write Un for the n×n matrix with all entries 1 and recall that a real matrix B is skew-symmetric if B+BT=0.

Lemma A.2

Let Bn be a skew-symmetric n×n matrix and Cn(x)=Bn+xUn . Then det(Cn(x))0 for every n1 and x0.

Proof

We begin by deriving a recursive formula for det(Cn(x)). If we first subtract row n from all other rows of Cn(x) and then (the resulting) column n from all other columns, we obtain the matrix

(b1nB˜n1bn1,nbn1bn2bn,n1x), (A.12)

where the entry (k,j) of the matrix B˜n1 is

b˜kj=bkjbnjbkn. (A.13)

Because Bn+BnT=0, a simple calculation yields B˜n1+B˜n1T=0. Clearly, if x=0, the determinant of the matrix (A.12) is simply det(Bn). Developing the determinant of (A.12) with respect to the last row (or column), it follows immediately that

det(Cn(x))=xdet(B˜n1)+det(Bn). (A.14)

Now we use the fact (Godsil, 1993, pp. 113–116) that for every skew-symmetric matrix Bn,

det(B2n1)=0, (A.15a)
det(B2n)=(Pf(B2n))2 (A.15b)

holds, where

Pf(B2n)=12nn!σS2nsgn(σ)i=1nbσ(2i1),σ(2i) (A.16)

denotes the Pfaffian of the matrix B2n, and S2n is the permutation group of a set of 2n elements. Applying (A.15) to (A.14), we obtain

det(Cn(x))={(Pf(Bn))2if n is even ,x(Pf(B˜n1))2if n is odd , (A.17)

which proves the lemma.  □

Lemma A.3

For arbitrary number In of alleles, arbitrary matrix D(n) of dominance parameters satisfying (3.6b), and arbitrary positive matrix P(n), the matrix M(n) is nonsingular.

Proof

We shall prove det(M(n))>0. For the proof, we omit the locus superscripts (n) and assume that all matrices are n×n. Because S is nonsingular, (A.3)–(A.7) show that

M=(DPS1+I)S, (A.18)

where I is the n×n identity matrix. By assumption (3.6b), D can be written as D=B+12U, where B is skew-symmetric. Therefore, Lemma A.2 shows that det(D)0. Because this holds for arbitrary dimension, every principal minor (of arbitrary order) of D is also nonnegative. Since PS1 is a diagonal matrix with positive elements on the diagonal, we have det(DPS1)=det(D)det(PS1)0. Because this holds for every dimension and because PS1 is diagonal, the principal minors of DPS1 are the products of the corresponding principal minors of D and PS1 (the latter being just products of the corresponding diagonal elements). Hence, all principal minors of DPS1 are nonnegative.

It is another well-known fact (e.g. Gantmacher, 1986, p. 99) that the coefficient of xk of the characteristic polynomial of a square matrix can be written as the sum of all principal minors of order nk. As a consequence, det(I+DPS1), which equals the characteristic polynomial of DPS1 evaluated at 1, can be written as 1 plus the sum of all principal minors of order 1. Since by the above reasoning all of them are nonnegative, we have

det(M)=det(DPS1+I)det(S)det(S)>0, (A.19)

which finishes the proof.  □

Lemmas A.1 and A.3 show that Eq. (3.7) holds for every internal gene-frequency equilibrium if there is DIDID.

Appendix B. Proof of Proposition 5.4

Throughout, we assume two alleles per locus and DIDID, i.e., (5.7). Let ML,Γ denote the set of all L×Γ matrices with strictly positive entries. For given alleles i,j at locus n, we denote by Uij=(uij,α(n))ML,Γ the matrix containing all fitness contributions uij,α(n), where uij,α(n)>0 and nL and αG. Further, let

G={1,,Γ1} (B.1)

and denote

cC={(c1,,cΓ1)T:cα>0,αG, and αGcα<1}. (B.2)

Then, cΓ=1αGcα>0 and (c1,,cΓ)TintΔΓ.

For any given matrix A=(a1,,aΓ)ML,Γ, where aαRL (αG) is a column vector, we write

A=(a1,,aΓ1)ML,Γ1

for the matrix obtained by omitting the last column (which requires Γ2). For any given vector a, a>0 means that every component of a is positive, by which we always mean strictly positive.

We shall need the following parameter sets:

PD={(c,U11,U22):cC,U11,U22ML,Γ+}, (B.3a)
PD={(c,U11,U22):cC,U11ML,Γ1+,U22ML,Γ+}. (B.3b)

Because we assume (5.7), the gene-frequency dynamics (5.5a) is completely specified by ϑ[0,1]L and (c,U11,U22)PD.

Our aim is to prove Proposition 5.4, which here we formulate more precisely.

Proposition B.1

If 1LΓ1 and ϑ[0,1]L . Then there is an open, nonempty subset UPD, such that for every (c,U11,U22)U an isolated, internal, asymptotically stable equilibrium point of the gene-frequency dynamics (5.5a) exists. It is the unique solution ρˆ of (5.8) in intΩ=(0,1)L.

The proof is based on three lemmas.

Lemma B.2

ρˆ=12=(12,,12)T is an equilibrium of the gene-frequency dynamics (5.5a) if and only if

αcαu11,α(n)u22,α(n)wα=0for every nL, (B.4)

where

wα=w¯α(12)=14n[(1+2ϑ(n))u11,α(n)+(32ϑ(n))u22,α(n)]. (B.5)

Proof

Because we assume (5.7), the internal equilibria of (5.5a) are precisely the solutions of (5.8). Substitution of ρˆ=12 and (5.7) into (5.4) and (2.5) yields (B.5).  □

We denote u11,Γ=(u11,Γ(1),,u11,Γ(L))TRL.

Lemma B.3

Let ϑ[0,1]L . There exists an open subset OPD such that for every (c,U11,U22)O a unique solution u11,Γ of (B.4) exists. It satisfies u11,Γ>0.

Proof

Let

λn=1cΓαGcαwα(u11,α(n)u22,α(n)), (B.6)

xn=u11,Γ(n), x=n(1+2ϑ(n))xn, and μ1=n(32ϑ(n))u22,Γ(n). Then we obtain wΓ=14(x+μ1) from (B.4), and (B.5) yields

cΓwΓ(xnu22,Γ(n))=cΓλn. (B.7)

Hence,

xn=u22,Γ(n)+14λn(x+μ1),nL. (B.8)

We solve this system for (xn). Let λ=n(1+2ϑ(n))λn and μ2=n(1+2ϑ(n))u22,Γ(n).

Multiplying (B.8) by (1+2ϑ(n)) and summing over all n, we obtain

x=μ2+14λ(x+μ1). (B.9)

If λ4, (B.9) has the unique solution

x=λμ1+4μ24λ. (B.10)

Substituting (B.10) into (B.8), we obtain

xn=u22,Γ(n)+λn4λ(μ1+μ2)=u22,Γ(n)+4λn4λlu22,Γ(l), (B.11)

which exists and is uniquely determined if λ4. This yields the desired solution u11,Γ=(x1,,xn)T. It satisfies u11,Γ>0 if

|4λn4λ|lu22,Γ(l)<u22,Γ(n)for every nL. (B.12)

It remains to show that (B.12) holds on an open set of parameters. Because |λ|3n|λn|, we have λn/(4λ)1 whenever n|λn|1. The latter is easy to achieve by choosing the cα, αG, sufficiently small. As a consequence and because u22,Γ(n)>0 for every n, the estimate (B.12) can be realized for every ϑ and every (U11,U22) by choosing a sufficiently small ϵ>0 such that maxαGcα<ϵ. Choosing (U11,U22) from an arbitrary, but bounded, open subset MML,Γ1+×ML,Γ+, we can find a uniformly small ϵ>0. Thus, we can choose the desired set O=Cϵ×M, where Cϵ=(0,ϵ)Γ1 for sufficiently small ϵ>0. Thus, for every ϑ[0,1]L, there exists an open bounded set O and a vector u11,Γ with the desired properties.  □

Remark B.4

u11,Γ depends smoothly on (c,U11,U22) and on ϑ.

Remark B.5

The set O can be chosen such that it applies to every ϑ[0,1]L. This follows from the proof because 14n(u11,α(n)+u22,α(n))wα34n(u11,α(n)+u22,α(n)) and, therefore, the critical ϵ can be chosen independently of ϑ. Hence, the left-hand side in (B.12) can be made small uniformly in ϑ[0,1]L.

Next we show that small perturbations of u11,Γ yield equilibria ρˆ close to 12. For this, we need the Hessian matrix of F=lnw˜, evaluated at ρintΩ. We denote it by H(ρ)ML,L, i.e., 

Hnl(ρ)=2F(ρ)p(n)p(l). (B.13)

A critical point ρ of F is isolated if H(ρ) is invertible. Because we need to check regularity of H at critical points, we calculate H(ρ). Partial differentiation of (2.14), using (2.5) and (5.4), yields

Fp(n)=2αcαw¯α(u1,α(n)u2,α(n)). (B.14)

Differentiating this with respect to p(l), we obtain straightforwardly from (5.3) and (5.4):

Hnl(ρ)=2αcαp(l)(u1,α(n)u2,α(n)w¯α)=2αcα[δnlu11,α(n)2u12,α(n)+u22,α(n)w¯α2(u1,α(n)u2,α(n))(u1,α(l)u2,α(l))w¯α2], (B.15)

where δnl=1 if n=l, and δnl=0 otherwise.

Now we invoke (5.7) and evaluate Hnl(ρ) at ρ=12. This gives u11,α(n)2u12,α(n)+u22,α(n)=(12ϑ(n))(u11,α(n)u22,α(n)) and w¯α=wα. Therefore, (B.4) implies

αcαu11,α(n)2u12,α(n)+u22,α(n)wα=0. (B.16)

Applying (5.9) with ρ=12 to the second term in (B.15), the entries of the Hessian matrix evaluated at 12, Φ=H(12), become

Φnl=αcα(u11,α(n)u22,α(n))(u11,α(l)u22,α(l))(wα)2. (B.17)

Now we have all ingredients to formulate and prove our last lemma.

Lemma B.6

Let 1LΓ1 and ϑ[0,1]L . Then there exists an open subset O1O such that for every (c,U11,U22)O1 and corresponding solution u11,Γ>0 of (B.4), the Hessian Φ is invertible.

Proof

Let C denote the Γ×Γ diagonal matrix with entries cα along the diagonal, and C12 its square root, which exists and is invertible because cα>0 for every α. Moreover, we set

eα(n)=u11,α(n)u22,α(n)wα (B.18)

and define E as the matrix with these entries, i.e., E=(eα(n))ML,Γ. Then we can write

Φ=ECET=(EC1/2)(EC1/2)T. (B.19)

Because (B.4) informs us that αcαeα(n)=0 for every n, generically, E has rank min(Γ1,L). If ΓL+1, then, generically, rankE=L. Because rankMMT=rankM is valid for every matrix MML,Γ, we obtain rankΦ=rank(C1/2E)=rankE, where the latter equality holds because multiplication by a nonsingular matrix leaves the rank unchanged. Therefore, Φ is (generically) invertible if and only if ΓL+1.

Finally, if Q denotes the preimage of the open subset of ML,Γ of matrices of rank L under the continuous map (U11,U22)E, then we can choose O1=(Cϵ×Q)O, where O is as in the proof of Lemma B.3. Therefore the set O1 is open and nonempty.  □

Remark B.7

In the absence of DIDID, this Hessian matrix is of the form Φ=DETCE, where D is a nonnegative diagonal matrix that vanishes if there is DIDID; cf. (B.15). Then Φ is generically invertible whenever L1 and Γ2. In fact, the proof can be extended to arbitrary dominance relations, thus giving a very different proof of Result 5.1 that does not assume weak selection. However, the generalizations of Lemmas B.3 and B.6 require considerable work.

Proof of Proposition B.1

We choose (c,U11,U22) and the corresponding u11,Γ according to Lemma B.6, i.e., (c,U11,U22)O1. Clearly, (c,U11,U22)PD. By Lemma B.2, 12 is an internal equilibrium of (5.5a), and by Lemma B.6 the Hessian H(12)=Φ is invertible. Therefore, 12 is an isolated equilibrium of (5.5a). Since (B.15), together with (5.3), (5.4), (2.5) and Remark B.4, demonstrates that H(ρ) is smooth in ρ, the Implicit Function Theorem shows that there exists an open neighborhood UPD of (c,U11,U22) for which there is a unique, isolated gene-frequency equilibrium ρˆ close to 12. Theorem 3.14 in N09b yields global asymptotic stability.  □

References

  1. Akin E. Amer. Math. Soc; Providence, RI: 1993. The General Topology of Dynamical Systems. [Google Scholar]
  2. Barton N.H. What role does natural selection play in speciation? Phil. Trans. R. Soc. B. 2010;365:1825–1840. doi: 10.1098/rstb.2010.0001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Bürger R. Wiley; Chichester: 2000. The Mathematical Theory of Selection, Recombination, and Mutation. [Google Scholar]
  4. Bürger R. Multilocus selection in subdivided populations I. Convergence properties for weak or strong migration. J. Math. Biol. 2009;58:939–978. doi: 10.1007/s00285-008-0236-5. [DOI] [PubMed] [Google Scholar]
  5. Bürger R. Multilocus selection in subdivided populations II. Maintenance of polymorphism for weak or strong migration. J. Math. Biol. 2009;58:979–997. doi: 10.1007/s00285-008-0237-4. [DOI] [PubMed] [Google Scholar]
  6. Bürger R. Polymorphism in the two-locus Levene model with nonepistatic directional selection. Theor. Popul. Biol. 2009;76:214–228. doi: 10.1016/j.tpb.2009.07.002. [DOI] [PubMed] [Google Scholar]
  7. Christiansen F.B., Feldman M. Subdivided populations: a review of the one- and two-locus deterministic theory. Theor. Popul. Biol. 1975;7:13–38. doi: 10.1016/0040-5809(75)90003-9. [DOI] [PubMed] [Google Scholar]
  8. Christiansen F.B. Wiley; Chichester: 1999. Population Genetics of Multiple Loci. [Google Scholar]
  9. Conley C. vol. 38. Amer. Math. Soc; Providence, RI: 1978. Isolated invariant sets and the Morse index. (NSF CBMS Lecture Notes). [Google Scholar]
  10. de Jong G. Unpredictable selection in a structured population leads to local differentiation in evolved reaction norms. J. Evol. Biol. 1999;12:839–851. [Google Scholar]
  11. de Jong G., Gavrilets S. Maintenance of genetic variation in phenotypic plasticity: the role of environmental variation. Genet. Res. 2000;76:295–304. doi: 10.1017/s0016672300004729. [DOI] [PubMed] [Google Scholar]
  12. Ewens W. second ed. Springer; Berlin: 2004. Mathematical Population Genetics. I. Theoretical Introduction. [Google Scholar]
  13. Gantmacher F.R. Springer; Berlin: 1986. Matrizentheorie. [Google Scholar]
  14. Geiringer H. On the probability theory of linkage in Mendelian heredity. Ann. Math. Stat. 1944;15:25–57. [Google Scholar]
  15. Godsil C.D. Chapman and Hall; New York: 1993. Algebraic Combinatorics. [Google Scholar]
  16. Karlin S. Gene frequency patterns in the Levene subdivided population model. Theor. Popul. Biol. 1977;11:356–385. doi: 10.1016/0040-5809(77)90018-1. [DOI] [PubMed] [Google Scholar]
  17. Karlin S. Theoretical aspects of multilocus selection balance I. In: Levin S.A., editor. Studies in Mathematics. vol. 16. The Mathematical Association of America; Washington: 1978. pp. 503–587. (Studies in Mathematical Biology). [Google Scholar]
  18. Karlin S. Classification of selection-migration structures and conditions for a protected polymorphism. Evol. Biol. 1982;14:61–204. [Google Scholar]
  19. Karlin S., Liberman U. Representation of nonepistatic selection models and analysis of multilocus Hardy-Weinberg equilibrium configurations. J. Math. Biol. 1979;7:353–374. doi: 10.1007/BF00275154. [DOI] [PubMed] [Google Scholar]
  20. Karlin S., McGregor J. Polymorphism for genetic and ecological systems with weak coupling. Theor. Popul. Biol. 1972;3:210–238. doi: 10.1016/0040-5809(72)90027-5. [DOI] [PubMed] [Google Scholar]
  21. Kirkpatrick M., Johnson T., Barton N. General models of multilocus selection. Genetics. 2002;161:1727–1750. doi: 10.1093/genetics/161.4.1727. [DOI] [PMC free article] [PubMed] [Google Scholar]
  22. Kirzhner V., Lyubich Yu.I. Multilocus dynamics under haploid selection. J. Math. Biol. 1997;35:391–408. doi: 10.1007/s002850050058. [DOI] [PubMed] [Google Scholar]
  23. Kun L.A., Lyubich Yu.I. Convergence to equilibrium under the action of additive selection in a multilocus multiallelic population. Sov. Math. Dokl. 1979;20:1380–1382. [Google Scholar]
  24. Kun L.A., Lyubich Yu.I. Convergence to equilibrium in a polylocus polyallele population with additive selection. Probl. Inform. Transmiss. 1980;16:152–161. [Google Scholar]
  25. LaSalle J.P. Stability theory for difference equations. In: Hale J.K., editor. Studies in Ordinary Differential Equations. vol. 14. Mathematical Association of America; Washington, DC: 1977. pp. 1–31. (Studies in Mathematics). [Google Scholar]
  26. Levene H. Genetic equilibrium when more than one ecological niche is available. Amer. Natur. 1953;87:331–333. [Google Scholar]
  27. Li W.-H., Nei M. Stable linkage disequilibrium without epistasis in subdivided populations. Theor. Popul. Biol. 1974;6:173–183. doi: 10.1016/0040-5809(74)90022-7. [DOI] [PubMed] [Google Scholar]
  28. Lyubich Yu.I. Springer; Berlin, Heidelberg, New York: 1992. Mathematical Structures in Population Genetics. [Google Scholar]
  29. Nagylaki T. Springer; Berlin, Heidelberg, New York: 1992. Introduction to Theoretical Population Genetics. [Google Scholar]
  30. Nagylaki T. The evolution of multilocus systems under weak selection. Genetics. 1993;134:627–647. doi: 10.1093/genetics/134.2.627. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Nagylaki T. Polymorphism in multiallelic migration–selection models with dominance. Theor. Popul. Biol. 2009;75:239–259. doi: 10.1016/j.tpb.2009.01.004. [DOI] [PubMed] [Google Scholar]
  32. Nagylaki T. Evolution under the multilocus Levene model. Theor. Popul. Biol. 2009;76:197–213. doi: 10.1016/j.tpb.2009.07.003. [DOI] [PubMed] [Google Scholar]
  33. Nagylaki T., Hofbauer J., Brunovský P. Convergence of multilocus systems under weak epistasis or weak selection. J. Math. Biol. 1999;38:103–133. doi: 10.1007/s002850050143. [DOI] [PubMed] [Google Scholar]
  34. Nagylaki T., Lou Y. The dynamics of migration–selection models. In: Friedman A., editor. Tutorials in Mathematical Biosciences IV. vol. 1922. Springer; Berlin, Heidelberg, New York: 2008. pp. 119–172. (Lect. Notes Math.). [Google Scholar]
  35. Peischl S. Dominance and the maintenance of polymorphism in multiallelic migration–selection models with two demes. Theor. Popul. Biol. 2010;78:12–25. doi: 10.1016/j.tpb.2010.03.006. [DOI] [PubMed] [Google Scholar]
  36. Roze D., Rousset F. Multilocus models in the infinite island model of population structure. Theor. Popul. Biol. 2008;73:529–542. doi: 10.1016/j.tpb.2008.03.002. [DOI] [PubMed] [Google Scholar]
  37. Spichtig M., Kawecki T.J. The maintenance (or not) of polygenic variation by soft selection in a heterogeneous environment. Amer. Natur. 2004;164:70–84. doi: 10.1086/421335. [DOI] [PubMed] [Google Scholar]
  38. Strobeck C. Haploid selection with n alleles and m niches. Am. Nat. 1979;113:439–444. [Google Scholar]
  39. Turelli M., Barton N.H. Dynamics of polygenic characters under selection. Theor. Popul. Biol. 1990;38:1–57. [Google Scholar]
  40. van Doorn G.S., Dieckmann U. The long-term evolution of multilocus traits under frequency-dependent disruptive selection. Evolution. 2006;60:2226–2238. [PubMed] [Google Scholar]
  41. Wiehe T., Slatkin M. Epistatic selection in a multi-locus Levene model and implications for linkage disequilibrium. Theor. Popul. Biol. 1998;53:75–84. doi: 10.1006/tpbi.1997.1342. [DOI] [PubMed] [Google Scholar]
  42. Zhivotovsky L.A., Feldman M.W., Bergman A. On the evolution of phenotypic plasticity in a spatially heterogeneous environment. Evolution. 1996;50:547–558. doi: 10.1111/j.1558-5646.1996.tb03867.x. [DOI] [PubMed] [Google Scholar]
  43. Zhivotovsky L.A., Gavrilets S. Quantitative variability and multilocus polymorphism under epistatic selection. Theor. Popul. Biol. 1992;42:254–283. doi: 10.1016/0040-5809(92)90015-l. [DOI] [PubMed] [Google Scholar]

RESOURCES