The k-parent spatial Lambda-Fleming-Viot process as a stochastic measure-valued model for an expanding population

(1)

HAL Id: hal-03154662

https://hal-univ-paris.archives-ouvertes.fr/hal-03154662

Preprint submitted on 3 Mar 2021

HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.

The k-parent spatial Lambda-Fleming-Viot process as a stochastic measure-valued model for an expanding

population

Apolline Louvet

To cite this version:

Apolline Louvet. The k-parent spatial Lambda-Fleming-Viot process as a stochastic measure-valued model for an expanding population. 2021. �hal-03154662�

(2)

The k-parent spatial Lambda-Fleming-Viot process as a stochastic measure-valued model for an expanding population

Apolline Louvet ^∗,1

1MAP5, Université de Paris, CNRS, 45 rue des Saints-Pères, 75270 Paris Cedex 06, France.

Abstract

We model spatially expanding populations by means of a spatial Λ-Fleming Viot process (SLFV) with selection : thek-parent SLFV. We fill empty areas with type 0 "ghost" individuals, which have a strong selective disadvantage against "real" type 1 individuals. This model is a special case of the SLFV with selection introduced in [19, 22] : natural selection acts during all reproduction events, and the fraction of individuals replaced during a reproduction event is constant equal to 1. Letting the selective advantagek of type 1 individuals over type 0 individuals grow to +∞, and without rescaling time nor space, we obtain a new model for expanding populations, the∞-parent SLFV.

This model is reminiscent of the Eden growth model [13], but with an associated dual process of potential ancestors, making it possible to investigate the genetic diversity in a population sample.

In order to obtain the limitk→+∞of thek-parent SLFV, we introduce an alternative construction of thek-parent SLFV adapted from [38], which allows us to couple SLFVs with different selection strengths.

Running headline: A spatial Lambda-Fleming-Viot process for expanding populations

Keywords: spatial Lambda-Fleming-Viot process, range expansion, duality, genealogies, population genetics, limiting process

MSC 2020 Subject Classification: Primary: 60G57, 60J25, 92D10,Secondary: 60J76, 92D25

1 Introduction

Population expansions are common events occuring at all biological scales. The growth of a population in a new environment generates interfaces with distinctive features [27, 30] and specific patterns of genetic variation [23, 24, 26], both being a consequence of the stochasticity of reproduction at the front, where local population sizes are small. The models which are used to study expanding populations can be divided in two main categories : growth models, mostly used to investigate the front features, and models coming from population genetics, which are more suited to study genetic diversity patterns.

(4)

Experimental approaches suggest that the dynamics of fronts of real expanding populations belongs to the universality class of the Kardar-Parisi-Zhang (KPZ) equation introduced in [30] (see e.g [27]).

It has been conjectured (and demonstrated in the case of the solid-on-solid (SOS) growth model [6]) that many growth models generate similar interfaces. One of these models is the Eden growth model, initially introduced on a lattice in [13]. Under this model, each node of the lattice is either occupied or empty. At each time step, an empty node with at least one occupied neighbour is chosen, and becomes occupied. There exist alternative update rules for this model [28], as well as off-lattice variants (see e.g [39]). While this model can be used to study the growth of an expanding population, it is less suited to study genetic diversity patterns.

Conversely, models used in population genetics are generally associated with tools allowing one to investigate these patterns. The analysis of the genetic diversity of a population often goes through modelling the ancestral lineages of a subset of individuals, and studying how these lineages coalesce into common ancestors [15]. However, most classical population genetics models assume that populations have constant sizes and that individuals are uniformly distributed over the area of interest. Therefore, they appear at first ill-suited to model a population during an expansion event. One way to overcome this consists in filling empty areas with "ghost" individuals, which can reproduce but have a very strong selective disadvantage against "real" individuals [12, 25]. Under this framework, the reproduction of

"ghost" individuals can be interpreted as a local extinction of real individuals.

Using this idea, it is possible to model a population expansion as the spread of a genetic type favoured by natural selection. Such a question was already studied by means of different models including a stochastic component, mostly in one dimension (see e.g [3, 18, 33]). The most classical one is based on the Fisher-KPP equation [21, 32], in which stochasticity is introduced through a Wright- Fisher noise term. If 0≤p(t, x) ≤1 represents the proportion of individuals of the favoured type at locationx ∈R at timet≥0, thenp(t, x) solves the stochastic Fisher-KPP equation if forx∈R and t >0,

∂p

∂t(t, x) = m

2∆p(t, x)dt+s0p(t, x)(1−p(t, x)) + s1

p_ep(t, x)(1−p(t, x))W(dt, dx) (1) where W is a space-time white noise and p_e an effective population density. In one dimension, the stochastic Fisher-KPP equation exhibits travelling wave solutions [35], which describe how does the advantageous type spreads through space. However, Eq. (1) has no solution in higher dimensions.

Many variants of the deterministic version of the Fisher-KPP equation have been studied, including versions with individuals having different motilities [8, 9], different growth rates (see e.g [11]), other diffusion kernels, or other choices for the nonlinearity [5].

(5)

Other models are more individual-based, and are adaptations of classical population genetics models, such as the Moran model [12, 18, 25] or the stepping-stone model [2, 3, 37]. They generally require to divide the space into subunits called demes, to which reproduction events are limited and which are connected by migration.

In this article, we will focus instead on a "reproduction event-based" model allowing us to keep the spatial continuum : the spatial Λ-Fleming Viot process, or SLFV [4, 14]. Its main feature is that it models reproduction events affecting whole areas rather than reproduction individual by individual, by means of a Poisson point process of reproduction events.

The original version of the SLFV does not account for the presence of a selectively favoured genetic type, but it can be modified in order to incorporate selection : see [22] for different forms of fixed selection mechanisms, and [7, 10, 31] for ways to introduce fluctuating selection. Our approach will be based on a version of the SLFV with selection introduced in [22] and rigourously constructed in [19]. Most of the work on the SLFV with selection involved investigating scaling limits under different forms of weak selection (see also [16, 17]). However, in our case, since the selectively disadvantaged individuals do not actually exist, the selection can be considered as very strong. Therefore, we shall consider a different limit, when selection goes stronger and stronger, and neither time nor space are rescaled. The limiting model we shall obtain will be close to the off-lattice Eden growth model, hence we can expect it to generate interfaces similar to the ones observed in real expanding populations.

Moreover, and contrary to the off-lattice Eden growth model, a dual process of potential ancestors is also associated to it, giving us tools to investigate the genetic diversity patterns observed in an expanding population. This model therefore constitutes a new model for expanding populations, which naturally appears as the limit of other well-known processes, and which seems promising in order to investigate both the front features and the genetic diversity patterns of an expanding population. In this work, we will focus on the area occupied by the population, but future works will include genetic diversity inside the expanding population, using for instance tracer dynamics [12, 25].

1.1 The k-parent SLFV and its dual

1.1.1 The k-parent SLFV

All the random objects introduced in this section will be defined over some probability space (Ω,F,P).

Before presenting the processes we will consider, we need to introduce some notation.

Let d≥1. LetCc(R^d) be the space of all continuous and compactly supported functions R^d→R, let C¹(R) be the space of all continuously differentiable functions on R, letC_b¹(R) be the space of all bounded functions R→ R that are C¹ and whose first derivative is also bounded, and let B(R^d) be

(6)

the space of all measurable functionsR^d→R.

We start by introducing the state space over which the variant of the SLFV with selection we consider is defined. LetM^f_λ be the space of all measuresM onR^d×{0,1}such that for allf ∈Cc(R^d),

Z

R^d×{0,1}

f(x)M(dx, dκ) =^Z

R^d

f(x)dx.

In other words, M^f_λ is the space of all measures on R^d× {0,1} whose marginal over R^d is Lebesgue measure. By a standard decomposition theorem (see e.g [29], p.561), for all M ∈ Mf_λ, there exists ω:R^d→[0,1] measurable such that

M(dx, dκ) = ((ω(x)δ₀(dκ) + (1−ω(x))δ₁(dκ))dx. (2) Such aω is not unique, but defined up to a Lebesgue null set. The state space we consider is the set M_λ of all measuresM ∈M^f_λ such that there exists a measurable functionω :R^d→ {0,1}(instead of ω:R^d→[0,1]) satisfying (2).

We endowM^f_λ andM_λ with the topology of vague convergence. Moreover, letDM_λ[0,+∞) (resp.

D

Mf_λ[0,+∞)) denote the space of all càdlàgM_λ-valued paths (resp. M^f_λ-valued paths), endowed with the standard Skorokhod topology.

Let M ∈ M_λ, and let ω :R^d→ {0,1} be a measurable function satisfying Eq. (2). The function ω can be interpreted as the indicator function of a measurable set E ⊂R^d corresponding to the area occupied by what will be called "type 0" individuals, whileR^d\E corresponds to the area occupied by

"type 1" individuals. We will consider that type 0 individuals correspond to the "ghost" individuals mentionned in the introduction, and type 1 individuals to the "real" individuals. Therefore, type 0 individuals have a strong selective disadvantage against type 1 individuals, andE corresponds to the area not yet invaded by the real population (up to a Lebesgue null set). In all that follows, any ω:R^d→ {0,1} such that (2) is true will be called a density of M, and the notation ω_M will be used to denote an arbitrarily chosen density of M.

For all f ∈Cc(R^d),F ∈C¹(R) and ω:R^d→ {0,1} measurable, we set :

hω, fi:=^Z

R^d

f(x)ω(x)dx and we define the function ΨF,f ∈C_b(M_λ) as :

∀M ∈ M_λ,ΨF,f(M) :=F Z

R^d×{0,1}

f(x)1{κ=0}M(dx, dκ)

!

(3)

(7)

=F Z

R^d

f(x)ωM(x)dx

=F(hω_M, fi).

For all f ∈C_c(R^d), we denote the support off bySupp(f), and for allR ∈R^∗+, we set :

Supp^R(f) :={y∈R^d:∃x∈Supp(f),||y−x|| ≤ R}

and VR:= Vol(B(0,R)).

In other words, VR is the volume of a ball of radius R, and Supp^R(f) is the set of all points which are at a distance of at mostR of a point in the support off.

For all ω :R^d → {0,1},R ∈ R^∗+ and x∈ R^d, we define the functions Θ⁺x,R(ω) :R^d → {0,1} and Θ⁻x,R(ω) :R^d→ {0,1}by :

Θ⁺_x,R(ω) :=1B(x,R)^c×ω+1B(x,R), Θ⁻x,R(ω) :=1B(x,R)^c×ω.

Θ⁺_x,R(ω) corresponds to filling the ballB(x,R) with type 0 individuals (or equivalently, emptying the ball B(x,R) of all real individuals), while Θ⁻_x,R(ω) can be interpreted as filling the ballB(x,R) with type 1 individuals. Notice that if M ∈ M_λ, then Θ⁺x,R(ω_M)∈ M_λ and Θ⁻x,R(ω_M)∈ M_λ.

We now introduce the operator which will be used to define the specific SLFV with selection we will consider as the solution to a well-posed martingale problem. Let k ∈ N\{0,1}, and let µ be a σ-finite measure on R^∗+ such that

Z ∞ 0

R^dµ(dR)<+∞. (4)

LetL^k_µ be the operator acting on functions of the form ΨF,f withf ∈C_c(R^d) andF ∈C¹(R), defined the following way. Let f ∈C_c(R^d) and F ∈C¹(R). Then, for allM ∈ M_λ,

L^k_µΨF,f(M) :=^Z

R^d

Z ∞ 0

Z

B(x,R)^k

1 V_R^k×









k

Y

j=1

ω_M(y_j)



×F(hΘ⁺_x,R(ω_M), fi) +



1−

k

Y

j=1

ω_M(y_j)



×F(hΘ⁻x,R(ω_M), fi)

− F(hω_M, fi)



dy1...dy_kµ(dR)dx.

(8)

In Section 5, it is shown that this operator is well-defined, and that it can be rewritten as

L^k_µΨF,f(M) =^Z ^∞

0

Z

Supp^R(f)

Z

B(x,R)^k

1 V_R^k ×^h

k

Y

j=1

ωM(yj)×F(hΘ⁺_x,R(ωM), fi) + (1−

k

Y

j=1

ω_M(y_j))×F(hΘ⁻_x,R(ω_M), fi)

−F(hω_M, fi)ⁱdy1...dykdxµ(dR).

Intuitively, an interpretation of this operator in terms of reproduction events is the following.

Whenever a reproduction event affects the ball B(x,R), k positions y₁, ..., y_k are sampled inside the ball, and we take k individuals occupying each one of these positions. Since the density of type 0 individuals ωM is {0,1}-valued, we can consider that all the individuals occupying the position y1

(resp. y2, ..., yk) are of type 1−ωM(y1) (resp. 1−ωM(y2), ..., 1−ωM(yk)). If ^Q^k_j=1ωM(yj) = 1, then all the individuals are of type 0, and we fill the ball B(x,R) with type 0 individuals. Conversely, if 1−^Q^k_j=1ω_M(y_j) = 1, then at least one individual is of type 1, and this time we fill the ballB(x,R) with type 1 individuals. Since type 0 individuals model "ghost" individuals, they are supposed to have a selective disadvantage against "real" type 1 individuals, hence the exclusion of the casek= 1 which would not give any advantage to type 1 individuals. Moreover, kcan be interpreted as measuring the strength of the selective advantage of "real" individuals against "ghost" individuals, or in other words, the capacity of "real" individuals to invade an empty environment.

Ifk= 2,L²_µis the operator introduced in [19] to define and characterize the "selection part" of the SLFV with selection, in the special case for which there are no neutral events and all reproduction events have an impact of u = 1. Their proof of the existence and uniqueness of the D

Mf_λ[0,+∞)- valued solution to the martingale problem associated to L^k_µ can easily be extended to the casek≥2, by restricting the martingale problem to an increasing sequence of compact subsets ofR^d converging toR^d. In Section 2, we will show that this unique solution is in factDM_λ[0,+∞)-valued if the initial value belongs toM_λ.

Theorem 1. Let k≥2, and let µ be a σ-finite measure on (0,+∞) satisfying condition (4). For all M⁰∈ M_λ, there exists a uniqueDMλ[0,+∞)-valued process (Mt)t≥0 such that M0 =M⁰ and, for all F ∈C¹(R) and f ∈Cc(R^d),

ΨF,f(M_t)−ΨF,f(M₀)− Z t

0

L^k_µΨF,f(M_s)ds

t≥0

is a martingale. Moreover, the process (Mt)t≥0 is Markovian, and the corresponding semigroup is Feller.

(9)

Definition 2 (Definition of the k-SLFV). Let k≥2, let µ be a σ-finite measure on (0,∞) satisfying (4), and let M⁰ ∈ M_λ. Then, the k-parent spatial Λ-Fleming-Viot process (or k-SLFV process) with initial condition M⁰ associated to µis the unique solution to the martingale problem (L^k_µ, M⁰) stated in Theorem 1. In particular, the k-SLFV is a strong Markov process with càdlàg paths a.s.

By extension, if ω : R^d → {0,1} is measurable, we will define the k-SLFV process with initial density ω associated to µ to be the k-SLFV process with initial condition M⁰ associated to µ, with M⁰∈ M_λ of density ω.

Intuitively, the k-SLFV process can be constructed in the following way. Let M⁰ ∈ M_λ, and let µ be a σ-finite measure on (0,∞) satisfying (4). Moreover, let Π be a Poisson point process on R×R^d×(0,+∞) with intensitydt⊗dx⊗µ(dr). Initially, the k-SLFV is equal toM⁰. The dynamics of the k-SLFV process (M_t)t≥0 is then as follows. If (t, x,R) ∈Π, a reproduction event happens at time tin the ball B(x,R). We samplek types according to the type distribution in the ballB(x,R) at the timet−. We interpret these types as the types ofk potential "parents". With probability

1 V_R^k

Z

B(x,R)^k





k

Y

j=1

ω_M_t−(y_j)



dy₁...dy_k,

thektypes sampled are 0, so the kpotential parents are of type 0. In this case, all the individuals in the ballB(x,R) die, the k-th potential parent (of type 0) fills the ball B(x,R) with its descendants, which means that we set :

∀z∈B(x,R), ω_M_t(z) = 1. Conversely, with probability

1− 1 V_R^k

Z

B(x,R)^k





k

Y

j=1

ω_M_t−(y_j)



dy₁...dy_k,

at least one of the ktypes sampled is 1. As in the other case, all the individuals in the ball B(x,R) die, but this time the first potential parent to be of type 1 fills the ball B(x,R) with its descendants, which amounts to setting

∀z∈B(x,R), ω_M_t(z) = 0.

Note that the position of the parent which actually reproduces is then uniformly distributed over the closure of the region {y ∈ B(x,R) : ωMt−(y) = 0}. The value taken by the density out of the ball B(x,R) at timet is not affected by this reproduction event. We repeat this for each (t, x,R)∈Π.

This construction can be made rigourous using arguments adapted from [38], and will be used in

(10)

Section 2 to complete the proof of Theorem 1.

Remark 3. Thek-SLFV process is a special case of the general definition of an SLFV with selection in [22], with impact parameter u= 1, selection parameter s= 1, and selection function F :x→x−x^k. Remark 4. The condition (4) on µmatches the standard condition for the existence of the SLFV [4].

It comes from the fact that a pointx∈R^d is affected by a reproduction event at rate :

Z

R^d

Z +∞

0 1y∈B(x,R)µ(dR)dy=^Z ^+∞

0

VRµ(dR)∝ Z +∞

0

R^dµ(dR).

Remark 5. Since the density ω_M is only defined up to a Lebesgue null set, the type of individuals present in a given positiony ∈R^d cannot be uniquely defined. Therefore, even though intuitively we can first sample parental positions, and deduce parental types from ωM, we cannot formally sample positions in order to sample parental types.

A particularly interesting feature of this model is that there exists a dual process of potential ancestors associated to it, which follows the locations of the potential ancestors of a set of individuals.

In other words, the genetic diversity in a sample of the population can be determined by going backwards in time, and reconstructing the genealogical tree of the sample. Fork= 2, the dual process is analogous to the Ancestral Selection Graph (ASG) [34, 36], but with a spatial structure.

1.1.2 The k-parent ancestral process

All the new objects introduced in relation with the dual process will be defined on a new probability space (Ω,F,P). As before, we let µ be a σ-finite measure on (0,+∞) satisfying condition (4), and we let ←−

Π be a Poisson point process onR×R^d×(0,+∞) with intensitydt⊗dx⊗µ(dR).

Let M_p(R^d) denote the set of all finite point measures onR^d, equipped of the topology of weak convergence. For all Ξ =^P^l_i=1δξi ∈ M_p(R^d), for allx∈R^d and R>0, we define

Ix,R(Ξ) ={i∈J1, lK:||x−ξ_i|| ≤ R}

and S^R(Ξ) ={x∈R^d:∃i∈J1, lK:||x−ξ_i|| ≤ R}.

In other words,Ix,R(Ξ) is the set of all the indices of the points in Ξ which are at distance at mostR of x, while S^R(Ξ) is the set of all the points inR^d which are at distance at mostRof a point of Ξ.

Definition 6. Let Ξ⁰ ∈ M_p(R^d). The k-parent ancestral process (Ξt)t≥0 associated to µ (or equiv- alently to ←−

Π) and with initial condition Ξ⁰ is the M_p(R^d)-valued Markov jump process defined as follows.

(11)

• First, we setΞ0 = Ξ⁰.

• Then, for all (t, x,R)∈←−

Π, if Ix,R(Ξt−)6=∅ and if we write

Ξt−=

Nt−

X

i=1

δ_ξi t−,

we sample kpoints y1, ..., yk independently and uniformly at random in B(x,R), and we set

Ξt:=

Nt−

X

i=1

δ_ξi

t−− ^X

i∈Ix,R(Ξt−)

δ_ξi t−+

k

X

j=1

δ_y_j.

In other words, we remove all the atoms of Ξt− sitting in B(x,R), and we add k atoms at locations that are i.i.d and uniformly distributed over the ball B(x,R).

This process is well-defined, sinceN_t is stochastically bounded by the number (Y_t^k)t≥0 of particles in a Yule process withk children and with individual branching rate^R₀^∞VRµ(dR)<+∞(see [19] for a proof in the casek= 2, which can be generalized to the casek≥2).

The k-parent ancestral process solves a martingale problem that we now introduce. For all F ∈C_b¹(R) and f ∈ B(R^d), we define the function ΦF,f :M_p(R^d)→Rby :

∀Ξ∈ M_p(R^d),ΦF,f(Ξ) =F Z

R^d

f(x)Ξ(dx)=F(hΞ, fi).

We now define the operator G_µ^k on the set of functions of the form ΦF,f, which will be at the basis of the martingale problem satisfied by (Ξt)t≥0. Let F ∈ C_b¹(R) and f ∈ B(R^d), then for all Ξ =^P^l_i=1δξi ∈ M_p(R^d), we set :

G_µ^kΦF,f(Ξ) :=^Z

R^d

Z +∞

0

Z

B(x,R)^k



1x∈S^R(Ξ)× 1 V_R^k ×F



hΞ, fi − ^X

i∈I_x,R(Ξ)

f(xi) +

k

X

j=1

f(yj)





−1x∈S^R(Ξ)× 1

V_R^kF(hΞ, fi)

#

dy₁...dy_kµ(dR)dx

This operator is well defined. Indeed, for all Ξ∈ M_p(R^d), by Fubini’s theorem,

|G_µ^kΦF,f(Ξ)| ≤ Z +∞

0

Z

S^R(Ξ)

Z

B(x,R)^k2× 1

V_R^k × ||F||∞dy1...dykdxµ(dR)

≤2||F||_∞× Z +∞

0 Vol(S^R(Ξ))µ(dR)

≤2||F||_∞×Ξ(R^d)× Z +∞

0

VRµ(dR)

<+∞

(12)

by Condition (4).

Proposition 7. LetΞ⁰ ∈ M_p(R^d), and let(Ξt)t≥0be thek-parent ancestral process of initial condition Ξ⁰ associated to µ. Then, for all F ∈C_b¹(R) and for all f ∈ B(R^d), the process

ΦF,f(Ξt)−ΦF,f(Ξ0)− Z t

0

G_µ^kΦF,f(Ξs)ds

t≥0

is a martingale.

Intuitively, thek-parent ancestral process records the locations of the potential ancestors of a given sample of individuals. However, because densities are only defined up to a Lebesgue null set, it is not possible to assign uniquely a type to an individual located atx∈R^dlooking at the value of the density at this point. Therefore, as in [19], in order to give a duality relation between thek-parent SLFV and the k-parent ancestral process, we will need to consider a distribution of sampling locations, rather than fixed locations.

More specifically, for all l≥1 andx₁, ..., x_l∈(R^d)^l, we define :

Ξ[x₁, ..., x_l] :=

l

X

i=1

δxi ∈ M_p(R^d).

If Ψ is a density function on (R^d)^l, let µ_Ψ be the law of the random point measure^P^li=1δ_X_i, where (X1, ..., X_l) is sampled according to Ψ. If M ∈ M_λ and Ξ =^P^li=1δxi ∈ M_p(R^d), we set :

D(M,Ξ) := ^Y^l

i=1

ωM(xi).

Notice that for all l∈N^∗ and for all density function Ψ on (R^d)^l,

Z

M_p(R^d)

D(M,Ξ)µΨ(dΞ) =^Z

(R^d)^lΨ(x1, ..., xl)







l

Y

j=1

ωM(xj)







dx1...dxl

=^Z

(R^d×{0,1})^lΨ(x1, ..., xl)







l

Y

j=1

10(κj)







M(dx1, dκ1)...M(dxl, dκl)

does not depend on the choice of a densityωM of M.

A straightforward adaptation of the proof of Proposition 1.7 in [19] to the case k≥2 leads to the following proposition.

Proposition 8. Let k≥2. LetM⁰ ∈ M_λ, let l∈N^∗, and let Ψ be a density function on (R^d)^l. Let µ be a σ-finite measure on (0,+∞) satisfying (4). Let (Mt)t≥0 be the k-SLFV with initial condition

(13)

M⁰ associated to µand (Ξt)t≥0 be the k-parent ancestral process associated toµ. Then, for all t≥0, Z

Mp(R^d)E[D(Mt, ξ)|M₀ =M⁰]µΨ(dξ) =E[D(M⁰,Ξt)|Ξ0∼µΨ].

Equivalently, for all t≥0,

EM⁰



 Z

(R^d)^lΨ(x₁, ..., x_l)







l

Y

j=1

ω_M_t(x_j)







dx₁...dx_l



=^Z

(R^d)^lΨ(x₁, ..., x_l)E_Ξ[x

1,...,xl]





Nt

Y

j=1

ω_M0(ξ_t^j)



dx₁...dx_l.

1.2 Construction of the ∞-parent SLFV

Let us now introduce the limit of thek-parent SLFV whenk→ ∞ somewhat informally. We will call it the∞-parent spatial Λ-Fleming Viot process, or∞-parent SLFV. It will be constructed rigourously in Section 2 using an alternative construction of thek-parent SLFV inspired by [38], but we now give an intuitive idea of its definition.

Let µ be a σ-finite measure on (0,+∞) satisfying Condition (4), and let Π be a Poisson point process on R×R^d×(0,+∞) with intensity dt ⊗dx⊗µ(dR). Let also M⁰ ∈ M_λ. We start the

∞-parent spatial Λ-Fleming Viot process (M_t^∞)t≥0, or ∞- parent SLFV, at M₀^∞ = M⁰. Then, if (t, x,R) ∈ Π, as before, we consider that a reproduction event occurs in the ball B(x,R) at time t. However, this time we do not sample a finite number of potential parents. Instead, we look at the value of the integral

Z

B(x,R)

1−ωM_t−^∞(z)dz,

which amounts to sampling an infinite number of potential parents over the ball B(x,R) and looking at the proportion of them which are of the "existing" type (i.e, type 1).

If^R_B(x,R)1−ω_M_t−^∞(z)dz= 0, we consider that the parent which reproduces is of type 0, and we set :

∀z∈B(x,R), ω_M_t^∞(z) = 1.

Note that in this case, the "parent" which reproduces was "sampled" at a location which is uniformly distributed over the ball B(x,R).

Conversely, if^R_B(x,R)1−ω_M_t−^∞(z)dz6= 0, there is a non negligible number of individuals of type 1 inB(x,R). We impose that it is one of them which reproduces, and in such a way that its offspring invades the whole region. That is, we set :

∀z∈B(x,R), ωM_t^∞(z) = 0.

(14)

Again, note that in our interpretation, the location of the parent which actually reproduces is uniformly distributed over the closure of the region {y∈B(x,R) :ω_M_t−^∞(y) = 1}.

As for thek-parent SLFV, the∞-parent SLFV is solution to a martingale problem. However, and in contrast with the case of thek-parent SLFV, the condition (4) on µwill not be sufficient to ensure that this solution is unique. Instead, we will need the following stronger condition.

Definition 9. Let a_d>0 such that the minimal number of d-dimensional balls of radius 1 needed to cover the border of an hypersphere of radiusnin ddimensions is bounded from above bya_d×n^d−1 for every n≥1. Aσ-finite measure µ onR^∗+ is said to satisfy Condition (5) if it satisfies Condition (4), and if there exists R>0 such that

+∞

X

n=1

Z nR

(n−1)R(R+r)^dµ(dr)

!

a_d×n^d−1+ 1<+∞. (5)

Examples ofσ-finite measures µon R^∗₊ satisfying Condition (5) are the following : 1. Measures µon R^∗+ having a bounded support.

2. Measures µon R^∗+ of the form α×(1 +r)^−3d−1dr, with α >0.

We define the operatorL^∞ on functions of the form ΨF,f whereF ∈C¹(R) andf ∈Cc(R^d) in the following way. For allM ∈ M_λ, we set :

L^∞_µΨF,f(M) :=^Z ^+∞

0

Z

Supp^R(f)

"

δ₀ Z

B(x,R)(1−ω_M(z))dz

!

×F(hΘ⁺x,R(ω_M), fi) + 1−δ0

Z

B(x,R)(1−ωM(z))dz

!!

×F(hΘ⁻x,R(ωM), fi)

−F(hω_M, fi)

#

dxµ(dR).

Note that if δ0

R

B(x,R)(1−ωM(z))dz = 1, then for all y ∈B(x,R) except possibly on a Lebesgue null set,

Θ⁺_x,R(ω_M)(y) =ω_M(y).

In other words, the ballB(x,R) is already completely void of "existing" individuals, and filling it with

"ghost" individuals does not change anything. Therefore, we also have

L^∞_µ ΨF,f(M) =^Z ^+∞

0

Z

Supp^R(f) 1−δ₀ Z

B(x,R)(1−ω_M(z))dz

!!

×^hF(hΘ⁻_x,R(ωM), fi)−F(hω_M, fi)ⁱdxµ(dR).

(15)

In Section 5, we show that this operator is well-defined, even if µ satisfies Condition (4) rather than Condition (5). If µ satisfies Condition (5), then the associated martingale problem can be used to define and fully characterize the∞-parent SLFV. If µsatisfies only Condition (4), then the∞-parent SLFV is still solution to the martingale problem, but we no longer know whether this solution is unique, as stated in the following theorem. Therefore, we will provide in Section 2 a construction of the∞-parent SLFV which does not rely on the martingale problem, and works even ifµonly satisfies Condition (4).

Theorem 10. Let ω : R^d → {0,1}, let M⁰ ∈ M_λ, and let µ be a σ-finite measure on (0,+∞) satisfying Condition (4). Then, the ∞-parent SLFV with initial condition M⁰ associated to µ defined in Section 2.2 is a solution to the martingale problem for (L^∞_µ , M⁰).

Moreover, if µ satisfies Condition (5), the martingale problem associated to (L^∞_µ, M⁰) is well- posed, and the ∞-parent SLFV with initial condition M⁰ associated to µ is the unique solution to it in DM_λ[0,+∞).

1.3 Dual of the ∞-parent SLFV

As for thek-parent SLFV, the ∞-parent SLFV also has a dual process of potential ancestors.

Let E^c be the set of Lebesgue measurable and connected subsets of R^d whose Lebesgue measure is finite. Let E^cf be the set of all finite unions of elements of E^c. If E ∈ E^cf can be written as E = ∪^l_i=1Eⁱ where for all 1 ≤ i ≤ l, Eⁱ ∈ E^c, we let m(E) = m(E¹, ..., E^l) be the measure on R^d defined bym(E)(dx) :=1x∈Edx, and we set :

M^cf :={m(E) :E∈ E^cf}.

Definition 11 (∞-parent ancestral process). Let µ be aσ-finite measure on (0,+∞) satisfying Con- dition (5). Let ←−

Π be a Poisson point process on R+×R^d×(0,+∞) with intensity dt⊗dx⊗µ(dR), defined on the probability space (Ω,F,P).

LetΞ⁰ =m(E₀¹, ..., E₀^l)∈ M^cf. Then, theM^cf-valued∞-parent ancestral process(Ξ^∞_t )t≥0 with initial condition Ξ⁰ associated to µ(or equivalently to ←−

Π) is defined in the following way.

First, we set Ξ^∞₀ = Ξ⁰. Then, if for all t≥0, we write Ξ^∞t as Ξ^∞t =m(Et),

(16)

then for all (t, x,R)∈←−

Π, if Et−∩B(x,R) has a non zero Lebesgue measure,

Ξ^∞_t =m(Et−∪B(x,R)).

Moreover, this process is Markovian.

We will show that this process is well-defined in Section 3.

Remark 12. Note that the case Et−∩B(x,R) = B(x,R) is equivalent toEt−∪B(x,R) = Et−, and hence does not correspond to a visible jump of (Ξ^∞t )t≥0.

For all M ∈ M_λ with density ω and for all Ξ =m(E)∈ M^cf, we set :

De(M,Ξ) :=δ0

Z

E(1−ω(x))dx

.

If we know the value of D^e(M,Ξ) for all Ξ ∈ M^cf, since ω is {0,1}-valued, we know the value of ω everywhere up to a Lebesgue null set, and so we have completely characterized M. Therefore, the following duality result shows that the solution to the martingale problem associated toL^∞_µ is unique.

Proposition 13. Let µ be a σ-finite measure on (0,+∞) satisfying Condition (5). Let M⁰ ∈ M_λ, and let(M_t^∞)t≥0 be a solution to the martingale problem associated to (L^∞_µ, δ_M0). Then, for allt≥0 and for all E⁰∈ M^cf,

EM⁰

h

D(Me _t^∞, m(E⁰))ⁱ=E_m(E0)

h

D(Me ⁰,Ξ^∞_t )ⁱ,

where (Ξ^∞t ) is the∞-parent ancestral process of initial conditionm(E⁰)associated toµ. Equivalently, for everyt≥0, if ω_t etω₀ are {0,1}-valued densities of M_t^∞ and M⁰,

E

δ0

Z

E⁰(1−ωt(x))dx

=E_m(E0)

δ0

Z

Et

(1−ω0(x))dx

.

1.4 Structure of the paper

In Section 2, we construct the∞-parent SLFV rigorously, by introducing a coupling between a sequence ofk-parent SLFV processes with the same initial conditions. We also show the first part of Theorem 10, i.e, that the∞-parent SLFV is a solution to the martingale problem associated to L^∞_µ. In Section 3, we first demonstrate that the∞-parent ancestral process is well defined, and we then show that it can be characterized as the unique solution to a specific martingale problem. Section 4 is devoted to the proof of the duality relation between the ∞-parent SLFV and the∞-parent ancestral process stated

(17)

in Proposition 13. The second part of Theorem 10 is then a direct consequence of Proposition 13.

Section 5 contains technical lemmas used throughout the paper.

2 The ∞-parent SLFV

2.1 Alternative construction of the k-parent SLFV

In order to construct the∞-parent SLFV rigorously, we start by introducing an alternative construction of thek-parent SLFV, based on a variant of its dual. It relies on the sampling of parental locations along with reproduction events, and is an adaptation of the concept ofparental skeleton presented in Section 2.3.1 of [38].

In all that follows, letµbe aσ-finite measure on (0,+∞) satisfying Condition (4). LetU =B(0,1)^N, and let ˜ube the law of a sequence of i.i.d random variables (P_n)n≥1 uniformly distributed overB(0,1).

We will call an element of U a sequence of potential parents. Let us now extend the Poisson point process Π considered earlier by adding to each event a countable sequence of locations of potential parents. Indeed, let Π^c be a Poisson point process onR×R^d×(0,+∞)×U with intensity

dt⊗dx⊗µ(dR)⊗u˜(d(p_n)n≥1).

Then for all (t, x,R,(pn)n≥1)∈Π^c,

• as before, t can be interpreted as the time at which the reproduction event occurs, and we can seeB(x,R) as being the area affected by the reproduction event.

• For all n≥1,x+R ×p_n is uniformly distributed over the ballB(x,R), and can be interpreted as the location of the n-th potential parent sampled, if at least n potential parents have to be sampled.

We start by defining the variant of the k-parent ancestral process, on which the alternative construction of thek-parent SLFV is based.

Definition 14(Quenchedk-parent ancestral process). Letk≥2, letΞ⁰∈Mp(R^d), and let˜t≥0. The k-parent ancestral process (Ξ^Π_k,t^c^,^˜^t,Ξ⁰)t≥0 associated to Π^c, started at time ˜t and with initial condition Ξ⁰ is the M_p(R^d)-valued Markov jump process defined as follows.

• First, we setΞ^Π_k,0^c^,^˜^t,Ξ⁰ = Ξ⁰.

• Then, for all (t, x,R,(p_n)n≥1)∈ Π^c such that t≤t, recalling that for˜ Ξ =^P^li=1δ_ξ_i ∈ M_p(R^d),

(18)

Ix,R(Ξ) ={i∈J1, lK:||x−ξ_i|| ≤ R}, if

Ix,R(Ξ^Π_k,(˜^c^,_t−t)^˜^t,Ξ⁰

−)6=∅, then for all 1≤l≤k, we set

yl:=x+R ×pl

and Ξ^Π_k,^c_˜_t−t^,^˜^t,Ξ⁰ := Ξ^Π_k,(˜^c^,_t−t)−^˜^t,Ξ⁰ − ^X

x⁰∈Ix,R(Ξ^Πc,^˜^t,Ξ0

k,(˜t−t)−)

δ_x⁰ +

k

X

l=1

δ_y_l.

It is straightforward to check that this process has the same distribution as the k-parent ancestral process associated to µ and with initial condition Ξ⁰. Its interest is twofold. First, conditionally on Π^c, (Ξ^Π_k,t^c^,^˜^t,Ξ⁰)t≥0 is completely deterministic. Moreover, if for all Ξ =^P^li=1δ_x_i ∈ M_p(R^d), we denote the set of atoms of Ξ by

A(Ξ) :={x_i :i∈J1, lK},

then the process satisfies the following property, which will be useful in the coupling that we will introduce later.

Lemma 15. Let 2 ≤k ≤ k⁰, let Ξ⁰ ∈ M_p(R^d), let ˜t ≥0, and let Π^c be a Poisson point process on R×R^d×(0,+∞)×U with intensity dt⊗dx⊗µ(dR)⊗u(d(p˜ n)n≥1).

Then, for all t≥0,

A(Ξ^Π_k,t^c^,^t,Ξ^˜ ⁰)⊆A(Ξ^Π_k⁰^c_,t^,^t,Ξ^˜ ⁰).

In particular, for all t≥0 and x∈R^d,

A(Ξ^Π_k,t^c^,^t,δ^˜ ^x)⊆A(Ξ^Π_k⁰^c_,t^,^t,δ^˜ ^x).

Remark 16. Since A(Ξ) is a set, if there exists i6=j such that x_i =x_j, then x_i appears only once in A(Ξ).

Intuitively, the idea behind this lemma is the following. Since the coupled k-parent andk⁰-parent ancestral processes are based on the same extended Poisson point process of reproduction events, their evolutions are determined by the same reproduction events. Moreover, sincek⁰ ≥k, all the potential parents which are involved in the dynamics of thek-parent ancestral process are also potential parents for thek⁰-ancestral process. Therefore, we can consider that thek-parent ancestral process is embedded

The k-parent spatial Lambda-Fleming-Viot process as a stochastic measure-valued model for an expanding population

HAL Id: hal-03154662

https://hal-univ-paris.archives-ouvertes.fr/hal-03154662

The k-parent spatial Lambda-Fleming-Viot process as a stochastic measure-valued model for an expanding

population

Apolline Louvet

To cite this version:

The k-parent spatial Lambda-Fleming-Viot process as a stochastic measure-valued model for an expanding population

Contents

1 Introduction

2 The ∞-parent SLFV