...
Figure 3.1: The pruning process, starting from explosion time A defined in (3.32).
Consider the ascension time (or explosion time):
A=inf©
θ∈Θψ, σθ< ∞ª
, (3.32)
where we use the convention inf; =θ∞. The following Theorem gives the distribution of the ascension time A and the distribution of the tree at this random time. Recall that θ¯=ψ−1(ψ(θ))is defined in (3.10).
Theorem 3.24([AD12a]). Letψbe a critical branching mechanism satisfying Assumptions 1 and 2.
1. For allθ∈Θψ, we haveNψ[A>θ]=θ¯−θ.
2. Ifθ∞<θ<0, underNψ, we have, for any non-negative measurable functionalF, Nψ[F(TA+θ0,θ0≥0)|A=θ]=ψ0( ¯θ)Nψh
F(Tθ0,θ0≥0)σ0e−ψ(θ)σ0i . 3. For allθ∈Θψ, we haveNψ[σA< +∞|A=θ]=1.
In other words, at the ascension time, the tree can be seen as a size-biased critical Lévy tree. A precise description of TA is given in [AD12a]. Notice that in the setting of [AD12a], there is no need of Assumption 2.
3.3 The growing tree-valued process
Special Markov Property of pruning
In [ADV10], the authors prove a formula describing the structure of a Lévy tree, conditionally on theθ-pruned tree obtained from it in the (sub)critical case. We will give a general version of this result. From the measure of marks, M in (3.30), we define a measure of increasing marks by:
M↑(d x,dθ0)=X
i∈I↑
δ(xi,θi)(d x,dθ0), (3.33)
with
I↑=n
i∈Iske∪Inod;M(;,xi ×[0,θi])=1o .
The atoms(xi,θi)fori∈I↑correspond to marks such that there are no marks ofM on;,xi with aθ-component smaller thanθi. In the case of multipleθj for a given nodexi∈Br∞(T), we only keep the smallest one. In the caseΠ=0, the measureM↑ describes the jumps of a record process on the tree, see [AD11] for further work in this direction. Theθ-pruned tree can alternatively be defined usingM↑instead ofM as forθ≥0:
Λθ(T,M)=n
x∈T, M↑(;,x×[0,θ])=0o . We set:
Iθ↑=n
i∈I↑,xi∈Lf(Λθ(T,M))o
=n
i∈I↑,θi<θ and M↑(;,xi×[0,θ])=0o and fori∈Iθ↑:
Ti=T\T;,xi={x∈T, xi∈ ;,x},
whereTy,x is the connected component ofT\{x}containing y. For i∈Iθ↑,Ti is a real tree, and we will considerxi as its root. The metric and mass measure onTi are the restriction of the metric and mass measure ofT onTi. By construction, we have:
T =Λθ(T,M)~i∈I↑
θ(Ti,xi). (3.34)
Now we can state the general special Markov property.
Theorem 3.25(Special Markov Property). Letψbe a branching mechanism satisfying Assump-tions 1 and 2. Letθ>0. Conditionally onΛθ(T,M), the point measure:
Mθ↑(d x,dT0,dθ0)=X
i∈Iθ↑
δ(xi,Ti,θi)(d x,dT0,dθ0)
underPψr0 (or underNψ) is a Poisson point measure onΛθ(T,M)×T×(0,θ]with intensity:
mΛθ(T,M)(d x)³
2βNψ[dT0]+ Z
(0,+∞)Π(d r)re−θ0rPψr(dT0)´
1(0,θ](θ0)dθ0. (3.35) Proof. It is not difficult to adapt the proof of the special Markov property in [ADV10] to get Theorem 3.25 in the (sub)critical case by taking into account the pruning times θi and the w-tree setting; and we omit this proof which can be found in [ADH12b]. We prove how to extend the result to the super-critical Lévy trees using the Girsanov transform of Definition 3.18.
Assume thatψis super-critical. Fora>0, we shall writeΛθ,a(T,M)=πa(Λθ(T,M))for short. According to (3.34) and the definition of super-critical Lévy trees, we have that for any a>0, the truncated treeπa(T)can be written as:
πa(T)=Λθ,a(T,M)~i∈I↑
θ, Hxi≤a
³πa−Hxi(Ti),xi´
3.3. The growing tree-valued process and we have to prove thatPi∈I↑
θδ(xi,Ti,θi)(d x,dT0,dθ0)is conditionally onΛθ(T,M)a Pois-son point measure with intensity (3.35). Since a is arbitrary, it is enough to prove that the point measureMa, defined by
Ma(d x,dT0,dθ0)= X non-negative measurable functional onT. Let
B=Nψ£
F(Λθ,a(T,M)) exp(− 〈Ma,Φ〉)¤ .
Thanks to Girsanov formula (3.22) and the special Markov property for critical branching mechanisms, we get:
Thanks to the Girsanov formula and (3.29), we get:
and thanks to (3.7), we get:
G(h,x,θ)= Using (3.39) withF replaced byF R gives:
Nψh
exp(−〈Ma,Φ〉)F(Λθ,a(T,M))i
=B=Nψ£
F(Λθ,a(T,M))R(Λθ,a(T,M))¤ . This implies thatMa is, conditionally onΛθ,a(T,M), a Poisson point measure with intensity (3.36). This ends the proof.
3.3. The growing tree-valued process An explicit construction of the growing process
In this section, we will construct the growth process using a family of Poisson point measures.
Letψ be a branching mechanism satisfying Assumptions 1 and 2. Letθ∈Θψ. According to (3.20) and (3.7), we have:
Nψθ[T ∈ •]=2βNψθ[T ∈ •]+ Z
(0,+∞)Π(d r)re−θrPψrθ(T ∈ •). (3.40) LetT(0)∈Twith root;. Forq∈Θψand q≤θ, we set:
T(0)q =T(0) and m(0)q =mT(0).
We define the w-trees grafted on T(0) by recursion on their generation. We suppose that all the random point measures used for the next construction are defined on T under a probability measureQT(0)(dω).
Suppose that we have constructed the family((T(k)q ,m(n)q ), 0≤k≤n,q∈Θψ∩(−∞,θ)). We write
T(n)= G
q∈Θψ,q≤θ
T(n)q .
We define the (n+1)-th generation as follows. Conditionally on all trees from generations smaller thann,(T(k)q , 0≤k≤n, q∈Θψ∩(−∞,θ)), let
Nθn+1(d x,dT,d q)= X
j∈J(n+1)
δ(xj,Tj,θj)(d x,dT,d q) be a Poisson point measure onT(n)×T×Θψ with intensity:
µn+1θ (d x,dT,d q)=m(n)q (d x)Nψq[dT]1{q≤θ}d q.
Forq∈Θψ andq≤θ, we set
J(n+1)q =©
j∈J(n+1), q<θj
ª
and we define the treeT(nq+1)and the mass measurem(nq +1) by:
T(nq+1)=T(n)q ~j∈Jq(n+1)(T j,xj) and m(nq +1)= X
j∈J(nq+1)
mTj(d x).
Notice that by construction, (T(n)q ,n∈N) is a non-decreasing sequence of trees. We set Tq the completion of∪n∈NT(n)q , which is a real tree with root ;and obvious metricdTq, and we define a mass measure onTq bymTq=P
n∈Nm(n)q .
For q∈Θψ and q <θ, we consider Fq the σ-field generated by T(0) and the sequence of random point measures (1{q0∈[q,θ]}Nθ(n)(d x,dT,d q0),n∈N). We setNθ=P
n∈NNθn. The backward random point process q7→1{q≤q0}Nθ(d x,dT,d q0) is by construction adapted to the backward filtration(Fq,q∈Θψ∩(−∞,θ]).
The proof of the following result is postponed to Section3.3.
Theorem 3.26. Let ψbe a branching mechanism satisfying Assumptions 1 and 2. UnderQψθ:=
Nψθ[dT(0)]QT(0)(dω), the process
³³
Tq,dTq,;,mT¯q´
,q∈Θψ∩(−∞,θ]´
is aT-valued backward Markov process with respect to the backward filtration Fθ=(Fq,q ∈ Θψ∩(−∞,θ]). It is distributed as((Tq,mTq),q∈Θψ∩(−∞,θ])underNψ.
Notice the Theorem in particular entails that(Tq,dTq,;,mT¯q)is a w-tree for allq. We shall use the following lemma.
Lemma 3.27. Let ψ be a branching mechanism satisfying Assumptions 1 and 2. Let K be a measurable non-negative process (as a function ofq) defined onR+×T×Twhich is predictable with respect to the backward filtrationFθ. We have:
Qψθ This means that the predictable compensator ofNθ is given by:
µθ(d x,dT,d q)=mTq(d x)Nψq[dT]1{q∈Θψ,q≤θ}d q.
Notice this construction does not fit in the usual framework of random point measures as the support at time q of the predictable compensator is the (predictable backward in time) random setTq×T×Θψ.
Proof. Based on the recursive construction, we have:
Qψθ Now, by construction, we have that:
Tq=T(n)q ~j∈J(n)q ( ˜Tj,xj)
3.3. The growing tree-valued process
It can be noticed that the mapq7→Tq is non-decreasing càdlàg (backwards in time) and that we have, for j∈ ∪n∈NJ(n), xj∈Tθj: Tθj−=Tθj~(T j,xj). In particular, we can recover the random measureNθ from the jumps of the process (Tq,q∈Θψ∩(−∞,θ]). This and the natural compatibility relation ofNθ with respect toθgives the next Corollary.
Corollary 3.28. Letψbe a branching mechanism satisfying Assumptions 1 and 2. Let(Tθ,θ∈Θψ) be defined underNψ. Let
N =X
j∈J
δ(xj,Tj,θj)
be the random point measure defined as follows:
• The set{θj;j∈J}is the set of jumping times of the process(Tθ,θ∈Θψ): for j∈J,Tθj−6=Tθj.
with respect to the backward left-continuous filtrationF=(Fθ,θ∈Θψ)defined by:
Fθ=σ((xj,T j,θj);θ≤θj)=σ(Tq−;θ≤q).
More precisely, for any non-negative predictable processK with respect to the backward filtrationF, we have:
Nψ
·Z
N(d x,dT,d q)K³
q,Tq,Tq−
´¸
=Nψ
·Z
µ(d x,d T,d q)K³
q,Tq,Tq~(T,x)´
¸ . (3.41) Remark 5. Notice that Assumption 2 is assumed only for technical measurability condition, see Remark2. We conjecture that this results holds also if Assumption 2 is not in force.
As a consequence, thanks to property 3 of Theorem 3.24, we get, with the convention sup; =θ∞, that:
A=sup{θj,j∈J andσj= +∞} with σj=mTj(T j).
Proof of Theorem3.26
By construction, it is clear that the process (Tq,q ∈Θψ∩(−∞,θ]) is a backward Markov process with respect to the backward filtration (Fq,q ∈Θψ∩(−∞,θ]). By construction this process is càglàd in backward time. Since the process (Tq,q ∈Θψ) is a forward càdlàg Markov process, it is enough to check that forθ0∈Θψ, such thatθ0<θ, the two dimensional marginals(Tθ0,Tθ)and(Tθ0,Tθ) have the same distribution.
Replacing ψby ψθ0, we can assume that θ0=0 and 0<θ. We shall decompose the big treeT0 conditionally on the small tree Tθ by iteration. This decomposition is similar to the one which appears in [AD07] or [Voi10] for the fragmentation of the (sub)critical Lévy tree, but roughly speaking the fragmentation is here frozen but for the fragment containing the root.
We set T(0)=Tθ and m˜(0)=mTθ, so that (T(0),m(0)) and (T(0), ˜m(0)) have the same distribution. Recall notationM↑from (3.33) as well as (3.34):T0=T(0)~i∈I↑,1
θ (Ti,xi), where we writeIθ↑,1=Iθ↑ and whereP1=P
i∈Iθ↑,1δ(xi,Ti,θi) is, conditionally onT(0), a Poisson point measure with intensity:
ν1(d x,dT0,d q)=m˜(0)(d x)³
2βNψ[dT0]+ Z
(0,+∞)Π(d r)re−qrPψr(dT0)´
1(0,θ](q)d q.
Fori∈Iθ↑,1, we define the sub-tree ofTi: T˜i=
n
x∈Ti;M↑(xi,x×[0,θi])=0o .
Since Ti is distributed according to Nψ (or to Pψri for some ri >0), using the property of Poisson point measures, we have that conditionally onT0andθi, the treeT˜i is distributed as
3.3. The growing tree-valued process Λθi(T,M)underNψ(or underPψri) that is the distribution ofT˜i isNψθi[dT](orPψriθi(dT)), thanks to the special Markov property. Furthermore we haveTi=T˜i~i0∈Iθ,i↑,2(Ti0,xi0)where
X
i0∈Iθ,i↑,2
δ(xi0,Ti0,θi0)
is, conditionally onT(0) andT˜i a Poisson point measure on T˜i×T×(0,θ]with intensity:
mT˜i(d x)³
2βNψ(dT0)+ Z
(0,+∞)Π(d r)re−qrPψr(dT0)´
1[0,θi)(q)d q.
Thus we deduce, using again the special Markov property, that:
N˜θ1(d x,dT,d q)= X
i∈I↑,1
δ(xi, ˜Ti,θi)(d x,dT,d q)
is conditionally onT0 a Poisson point measure onT(0)×T×Θψ with intensity:
µ˜1(d x,dT,d q)=m˜(0)q (d x)Nψq[dT]1[0,θ)(q)d q, withm˜(0)q (d x)=m˜(0)(d x). We setT(1)=T(0)~i∈I↑,1
θ ( ˜Ti,xi)for the first generation tree and forq∈[0,θ]:
m˜(1)q (d x)= X
i∈Iθ↑,1
mT˜i(d x)1[0,θi)(q).
See Figure 3.2 for a simplified representation. We get that (T(1)θ , (m(1)q ,q ∈[0,θ]),T(0),mT(0)) and(T(1), ( ˜m(1)q ,q∈[0,θ]),T(0), ˜m(0))have the same distribution.
Furthermore, by collecting all the trees grafted onT(1), we get that T =T(1)~i0∈Iθ↑,2(Ti0,xi0),
where Iθ↑,2= ∪i∈I↑,1
θ Iθ,i↑,2 and where
P2= X
i0∈Iθ↑,2
δ(xi0,Ti0,θi0)
is, conditionally on (T(1), ( ˜m(1)q ,q∈[0,θ]),T(0), ˜m(0)) a Poisson point measure onT(1)×T×
(0,θ]with intensity:
ν2(d x,dT,d q)=m˜(1)q (d x)³
2βNψ(dT0)+ Z
(0,+∞)Π(d r)re−qrPψr(dT0)´
1[0,θ](q)d q.
Notice that:
T(1)={x∈T0;M↑(;,x×[0,θ])≤1} and m˜(1)θ (d x)+m˜(0)(d x)=1T(1)(x)mT0(d x). (3.42)
Figure 3.2: The tree T0, T(0), and a tree Ti and its sub-tree T˜i belonging to the first generation treeT(1)\T(0).
Then we can iterate this construction, and by taking increasing limits we obtain that the pair((∪n∈NT(n)θ ,P
n∈Nm(n)θ ),T0)has the same distribution as(T0,T(0)), where:
T0=n
x∈T0;M↑(;,x×[0,θ])< +∞o
and m˜0(d x)=1T0(x)mT0(d x).
To conclude, we need to check first that the completion ofT0isT0, or asT0 is complete that the closure ofT0 as a subset ofT0 is exactlyT0 and then thatmT0(T0c)=0.
Notice thatM↑has less marks thanM. Then Proposition 1.2 in [AD07] in the case when β=0or an elementary adaptation of it in the general framework of [Voi10], gives there is no loss of mass in the fragmentation process. This implies that, ifψis (sub)critical, then:
mT0({x∈T0;M(;,x×[0,θ])= ∞}=0. (3.43) Then, if ψis super-critical, by considering the restriction of T0 up to level a, πa(T0), and using a Girsanov transformation from Definition3.18 withθ=θ∗ and (3.43), we deduce that (3.43) holds forπa(T0). Sincea is arbitrary, we deduce by monotone convergence that (3.43)
3.4. Application to overshooting