The effect of competition on the height and length of the forest of genealogical trees of a
large population
M. Ba and E. Pardoux
Colloque JPS, CIRM Marseille
April 17, 2012
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 1 / 26
Introduction
Consider a Galton–Watson branching processXm in continuous time with m ancestors at time t= 0: fork ≥1, the process
Xm jumps
(k→k+ 1, at rate µk; k→k−1, at rate λk.
For any 0≤m<n,Xn(d)= Xm+Xn−m (branching property).
In order to model competition within the population described byXm, we superimpose to each individual a death rate due to competition, equal at timet toγXtm. This gives a global death rate equal at time t toγ(Xtm)2.
Set m= [Nx], µN = 2N+θ,λN = 2N andγN =γ/N, and ZN,x = XN[Nx]. The “total population mass process”ZN converges weakly to the solution of the Feller SDE with logistic drift
x x x 2 p x
Introduction
Consider a Galton–Watson branching processXm in continuous time with m ancestors at time t= 0: fork ≥1, the process
Xm jumps
(k→k+ 1, at rate µk; k→k−1, at rate λk.
For any 0≤m<n,Xn(d)= Xm+Xn−m (branching property).
In order to model competition within the population described byXm, we superimpose to each individual a death rate due to competition, equal at timet toγXtm. This gives a global death rate equal at time t toγ(Xtm)2.
Set m= [Nx], µN = 2N+θ,λN = 2N andγN =γ/N, and ZN,x = XN[Nx]. The “total population mass process”ZN converges weakly to the solution of the Feller SDE with logistic drift
dZtx =
θZtx−γ(Ztx)2
dt+ 2p
ZtxdWt, Z0x =x. (1)
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 2 / 26
Introduction
Consider a Galton–Watson branching processXm in continuous time with m ancestors at time t= 0: fork ≥1, the process
Xm jumps
(k→k+ 1, at rate µk; k→k−1, at rate λk.
For any 0≤m<n,Xn(d)= Xm+Xn−m (branching property).
In order to model competition within the population described byXm, we superimpose to each individual a death rate due to competition, equal at timet toγXtm. This gives a global death rate equal at time t toγ(Xtm)2.
Set m= [Nx], µN = 2N+θ,λN = 2N andγN =γ/N, and ZN,x = XN[Nx]. The “total population mass process”ZN converges weakly to the solution of the Feller SDE with logistic drift
x x x 2 p x
Introduction
This model has been studied by Lambert (2005) and
Pardoux-Wakolbinger (2011). The forest of those m trees is also finite a.s. . and for anyx la solution Zx of equation (1) goes extint in finite timea.s..
On can define the height and the length of the discrete forest of genealogical trees
Hm= inf{t >0, Xtm = 0}, Lm = Z Hm
0
Xtmdt, for m≥1,
as well as the height and the lenght of the continuous “forest of genealogical trees”
Tx = inf{t >0, Ztx = 0}, Sx = Z Tx
0
Ztxdt .
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 3 / 26
Introduction
This model has been studied by Lambert (2005) and
Pardoux-Wakolbinger (2011). The forest of those m trees is also finite a.s. . and for anyx la solution Zx of equation (1) goes extint in finite timea.s..
On can define the height and the length of the discrete forest of genealogical trees
Hm= inf{t >0, Xtm = 0}, Lm = Z Hm
0
Xtmdt, for m≥1, as well as the height and the lenght of the continuous “forest of genealogical trees”
Tx = inf{t >0, Ztx = 0}, Sx = Z Tx
0
Ztxdt
Branching process with polynomial competition
We generalize the above models, replacing the death rate γ(Xtm)2 by γ(Xtm)α,α >0.
Set m= [Nx], µN = 2N+θ,λN = 2N andγN =γ/Nα−1, and ZN,x := X[Nx]N . We show that the process ZN converges weakly to to a Feller SDE with a negative polynomial drift.
dZtx = [θZtx−γ(Ztx)α]dt+ 2p
ZtxdWt, Z0x =x. (2)
We study the height and the length of the genealogical tree, as an (increasing) function of the initial population in the discrete and the continuous model.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 4 / 26
Branching process with polynomial competition
We generalize the above models, replacing the death rate γ(Xtm)2 by γ(Xtm)α,α >0.
Set m= [Nx], µN = 2N+θ,λN = 2N andγN =γ/Nα−1, and ZN,x := X[Nx]N . We show that the process ZN converges weakly to to a Feller SDE with a negative polynomial drift.
dZtx = [θZtx−γ(Ztx)α]dt+ 2p
ZtxdWt, Z0x =x. (2)
We study the height and the length of the genealogical tree, as an (increasing) function of the initial population in the discrete and the continuous model.
Results
Both IE[supmHm]<∞ and IE[supxTx]<∞ ifα >1, while Hm → ∞as m→ ∞and Tx → ∞ as x→ ∞ a. s. ifα≤1.
Both IE[supmLm]<∞ andIE[supxSx]<∞ ifα >2, whileLm → ∞ as m→ ∞andSx → ∞ as x→ ∞ a. s. ifα≤2.
This necessitates to define in a consistent way the population processes jointly for all initial population sizes, i. e. we will need to define the two–parameter processes {Xtm, t ≥0,m≥1}and {Ztx, t ≥0,x>0}.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 5 / 26
The model with asymetric effect of the competition
The description of the process (Xtm,t≥0) is valid for one initial conditionm, but it is not sufficiently precise to describe the joint evolution of {(Xtm,Xtn), t≥0}, 1≤m<n.
Modelize the effect of the competition in a asymmetric way. The idea is that the progenyXtm of them “first” ancestors should not feel the competition due to the progenyXtn−Xtm of then−m “additional”
ancestors which is present in the populationXtn.
We order the ancestors from left to right, this order being passed to their progeny.
The model with asymetric effect of the competition
The description of the process (Xtm,t≥0) is valid for one initial conditionm, but it is not sufficiently precise to describe the joint evolution of {(Xtm,Xtn), t≥0}, 1≤m<n.
Modelize the effect of the competition in a asymmetric way. The idea is that the progenyXtm of them “first” ancestors should not feel the competition due to the progenyXtn−Xtm of then−m “additional”
ancestors which is present in the populationXtn.
We order the ancestors from left to right, this order being passed to their progeny.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 6 / 26
The model with asymetric effect of the competition
The description of the process (Xtm,t≥0) is valid for one initial conditionm, but it is not sufficiently precise to describe the joint evolution of {(Xtm,Xtn), t≥0}, 1≤m<n.
Modelize the effect of the competition in a asymmetric way. The idea is that the progenyXtm of them “first” ancestors should not feel the competition due to the progenyXtn−Xtm of then−m “additional”
ancestors which is present in the populationXtn.
We order the ancestors from left to right, this order being passed to their progeny.
Order killing
The individual placed at the position i at time t dies becuse of competition at rate γ[Li(t)α−(Li(t)−1)α],Li(t) is the number of alive individuals at timet, who are located at his left on the planar tree.
11 18
1 2 3 4 5
Figure: Bijection between binary trees and exploration processes
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 7 / 26
Discret model in the asymetric competition picture
{Xtm, t ≥0} is a continuous timeZ+–valued Markov process, which evolves as follows. Xtm jumps to
(k+ 1, at rateµk;
k−1, at rateλk+γ(k−1)α. The above description specifies well the evolution of the
twoparametersprocess{Xmt , t ≥0,m≥0}.
For α= 1,{Xtm, m≥1}is a Markov chain for fixed t
For α6= 1,{Xtm, m≥1}is not a Markov chain for fixed t. The conditional law of Xtn+1 given Xtn differs from its conditional law given (Xt1,Xt2, . . . ,Xtn).
However, {X·m, m≥0} is a Markov chain with values in the space D([0,∞);Z+), which starts from 0 atm= 0.
Discret model in the asymetric competition picture
{Xtm, t ≥0} is a continuous timeZ+–valued Markov process, which evolves as follows. Xtm jumps to
(k+ 1, at rateµk;
k−1, at rateλk+γ(k−1)α. The above description specifies well the evolution of the
twoparametersprocess{Xmt , t ≥0,m≥0}.
For α= 1,{Xtm, m≥1}is a Markov chain for fixed t
For α6= 1,{Xtm, m≥1}is not a Markov chain for fixed t. The conditional law of Xtn+1 given Xtn differs from its conditional law given (Xt1,Xt2, . . . ,Xtn).
However, {X·m, m≥0} is a Markov chain with values in the space D([0,∞);Z+), which starts from 0 atm= 0.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 8 / 26
Discret model in the asymetric competition picture
{Xtm, t ≥0} is a continuous timeZ+–valued Markov process, which evolves as follows. Xtm jumps to
(k+ 1, at rateµk;
k−1, at rateλk+γ(k−1)α. The above description specifies well the evolution of the
twoparametersprocess{Xmt , t ≥0,m≥0}.
For α= 1,{Xtm, m≥1}is a Markov chain for fixed t
For α6= 1,{Xtm, m≥1}is not a Markov chain for fixed t. The conditional law of Xtn+1 given Xtn differs from its conditional law given (Xt1,Xt2, . . . ,Xtn).
However, {X·m, m≥0} is a Markov chain with values in the space D([0,∞);Z+), which starts from 0 atm= 0.
Discription of jointly laws
For arbitrary 0≤m<n, letVtm,n:=Xtn−Xtm,t ≥0. Conditionally upon {X·`, `≤m}, and given that Xtm =x(t),t ≥0,{Vtm,n, t≥0}is a Z+–valued time inhomogeneous Markov process starting from V0m,n=n−m, whose time–dependent infinitesimal generator {Qk,`(t), k, `∈Z+}is given by
Q0,`(t) = 0, ∀`≥1, and for any k ≥1, Qk,k+1(t) =µk,
Qk,k−1(t) =λk+γ(x(t) +k−1)α, Qk,`(t) = 0, ∀`6∈ {k−1,k,k+ 1}.
Moreover we have
hVm,n,Xmi= 0.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 9 / 26
The continuous Model
{Ztx, t≥0,x≥0}is such that for each fixed x >0,{Ztx,t ≥0} is continuous process, solution of the SDE (2).
For any 0<x <y,{Vtx,y :=Zty−Ztx, t ≥0} solves the SDE dVtx,y =
θVtx,y−γ
(Ztx+Vtx,y)α−(Ztx)α dt+ 2 q
Vtx,ydWt0, V0x,y =y−x,
{Wt0, t ≥0} is independent from the Brownian motion W which drives the SDE (2) for Ztx.
{Z·x, x≥0}is a Markov process with values in C([0,∞),IR+), starting from 0 atx = 0.
Convergence result
Let {Z˜tN,x, t ≥0,x ≥0} be the process such that for fixedx,
{Z˜tN,x, t ≥0} is the linear interpolation of{ZtN,x, t ≥0} between its jumps.
Theorem As N → ∞,
{Z˜tN,x, t≥0,x≥0} ⇒ {Ztx, t ≥0,x≥0}
in D([0,∞);C([0,∞);IR+)), equipped with the Skohorod topology of the space of c`adl`ag functions of x , with values in the space C([0,∞);IR+) equipped with the topology of locally uniform convergence.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 11 / 26
Height of the discrete tree
Theorem
If0< α≤1, then
sup
m≥1
Hm = +∞ a. s.
Ifα >1, then
IE
sup
m≥1
Hm
<∞.
Proof of the Theorem[Height of the discrete tree] for α > 1
Let H1m= inf{s ≥0;Xsm = 1}.
Proposition
For α >1,λ= 0,∀ m≥1,IE(H1m) is given by IE(H1m) =
m
X
k=2
1 γ(k−1)α
∞
X
n=0
µn γn
1
[k(k+ 1)· · ·(k+n−1)]α−1 <∞.
Moreover we have
Hm ≤H1m+GH12+
G
X
i=1
Ti,
where G is geometric variable with parameter λ+µλ andTi is exponential with mean 1/(λ+µ).
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 13 / 26
Proof of the Theorem[Height of the discrete tree] for α ≤ 1
We have Xtα,m ≥Xt1,m, for all m≥1, t≥0, a. s..
{Xtm, t ≥0} is the sum of mmutually independent copies of {Xt1, t ≥0}. The result follows from the fact that: for allt >0 IP(H1 >t)>0 and IP(Hm<t) = (1−IP(H1 >t))m.
Length of the discrete tree
Time change ofXm: Amt :=
Z t
0
Xrmdr, ηtm = inf{s >0; Ams >t}. Um :=Xm◦ηm, and Sm = inf{r >0;Urm= 0}. We haveLm =Sm since Sm =RHm
0 Xrmdr.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 15 / 26
Length of the discrete tree
Xtm=m+P1 Z t
0
µXrmdr
−P2 Z t
0
[λXrm+γ(Xrm−1)α]dr
,
Utm =m+P1(µt)−P2
Z t
0
λ+γ(Urm)−1(Urm−1)α dr
. On the interval [0,Sm),Utm ≥1, and consequently we have
m−P2 Z t
0
λUrm+γ(Urm−1)α−1 dr
≤Utm
≤m+P1
Z t
0
µUrmdr
−P2
Z t
0
hγ
2(Urm−1)α−1 i
dr
.
Length of the discrete tree
Xtm=m+P1 Z t
0
µXrmdr
−P2 Z t
0
[λXrm+γ(Xrm−1)α]dr
,
Utm =m+P1(µt)−P2
Z t
0
λ+γ(Urm)−1(Urm−1)α dr
.
On the interval [0,Sm),Utm ≥1, and consequently we have m−P2
Z t
0
λUrm+γ(Urm−1)α−1 dr
≤Utm
≤m+P1
Z t
0
µUrmdr
−P2
Z t
0
hγ
2(Urm−1)α−1 i
dr
.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 16 / 26
Length of the discrete tree
Theorem Ifα≤2, then
sup
m≥0
Lm=∞ a. s..
Ifα >2, then
IE
sup
m≥0
Lm
<∞.
Height of the continuous tree
Consider again {Ztx, t≥0}solution of (2). We have Theorem
If0< α <1,0<IP(Tx =∞)<1if θ >0, while Tx <∞ a. s. if θ= 0.
Ifα= 1, Tx <∞a. s. if γ ≥θ, while 0<IP(Tx =∞)<1ifγ < θ.
Ifα >1, Tx <∞ a. s.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 18 / 26
Height of the continuous tree
Theorem
It α≤1, then Tx → ∞a. s., as x → ∞.
Ifα >1, thenIE[supx>0Tx]<∞.
Proof of of Theorem 5 for α > 1
We first need to establish some preliminary results on SDEs with infinite initial condition. Let f :IR+→IR be locally Lipschitz and such that
x→∞lim
|f(x)|
xα = 0. (3)
Theorem
Let α >1,γ >0and f satisfy the assumption (3). Then there exists a minimal X ∈C((0,+∞);IR) which solves
(dXt = [f(Xt)−γ(Xt)α]1{Xt≥0}dt+dWt;
Xt → ∞, as t →0. (4)
Moreover, if T0 := inf{t>0, Xt = 0}, thenIE[T0]<∞.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 20 / 26
Proof of the Theorem for α > 1
The processYtx :=p
Ztx solves the SDE dYtx =
θ
2Ytx −γ
2(Ytx)2α−1− 1 8Ytx
dt+dWt, Y0x =√ x.
By a well–known comparison theorem, Ytx ≤Utx, whereUtx solves dUtx =
θ
2Utx−γ
2(Utx)2α−1
dt+dWt, U0x =√ x.
The result follows from the previous Theorem.
Proof of the Theorem for α < 1
The result is equivalent to the fact that the time to reach 1, starting from x >1, goes to∞ asx → ∞. But whenZtx ≥1, a comparison of SDEs for various values of α shows that it suffices to consider the caseα= 1. But in that case,Tn is the maximum of the extinction times ofn mutually independent copies of Zt1, hence the result.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 22 / 26
Length of the continuous tree
Theorem
Ifα≤2, then Sx → ∞ a. s. as x → ∞.
Ifα >2, thenIE[supx>0Sx]<∞.
Proof of the Theorem for α > 2
Time change ofZx: At =
Z t
0
Zsxds, η(t) = inf{s >0, As >t}.t≥0, and Utx =Zx◦η(t) The processUx solves the SDE
dUtx =
θ−γ(Utx)α−1
dt+ 2dWt, U0x =x. (5) Let τx := inf{t >0, Utx = 0}. It follows from the above that
η(τx) =Tx, henceSx =τx. The result follows for α >2.
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 24 / 26
Proof of the Theorem for α ≤ 2
It suffice to consider the case α= 2. In that case, we have Utx =e−γtx+
Z t
0
e−γ(t−s)[θds+ 2dWs], hence
Sx = inf
t >0, Z t
0
eγs(θds+ 2dWs)≤ −x
, which clearly goes to infinity, as x→ ∞.
Thanks!
M. Ba and E. Pardoux (Colloque JPS, CIRM Marseille)Marseille, 17Avril 2012 April 17, 2012 26 / 26