5.3 Analysis of the branching process
5.3.3 Proof of Theorem
We can make the condition of Theorem 5.3.5 a little easier to check for our case by observing that our thinning process narrows down the types of sequences ¯s we need to consider.
In the context of our thinned branching process, the requirement thatpn↓0 implies that the sequence ¯s must satisfy sn→1. It also should not be surprising that any finite number of leading terms of the sequence (sn) don’t matter so that it is enough for us to check this condition as n→ ∞. Let us summarize these arguments in a lemma:
Lemma 5.3.6. For a thinned branching process driven by thinned distributions with pgfs {fn}n>0, e < 1 if and only if there exists a sequence {sn}n>0, sn ∈ [0,1) and sn → 1 such that for some N
fn(sn+1)6sn, ∀n>N (5.4) Proof:
We begin by arguing that there is no loss of generality in considering only {sn} with sn →1. If there does in fact exist ¯s such that ¯f(¯s)6¯s, then it is necessary thatsn →1. To see why, note that fn(s)> fn(0) for all s ∈[0,1]. Therefore if fn(sn+1) 6sn for alln, then sn >fn(0) for alln. Sincepn →0, then fn(0)→1 so sn→1 also.
The fact that s → 1 is necessary also implies that if there does not exist ¯s satisfying ¯
f(¯s) 6 ¯s with sn → 1, then there cannot exist any other such sequence ˜s ∈ I∞ satisfying ¯
f(˜s)6s˜.
To see how the tail condition 5.4 implies the full condition on ¯s in Theorem 5.3.5, note that we can start from sN and then construct the initial elements of the sequence s1, . . . , sN−1 satisfying condition 5.3 by setting sN−1 =fN−1(sN)∈(0,1).
Now letting g(s, α) be the pgf of the offspring distribution, the thinned distribution f(s, p, α) for a fixed thinning probability p may be written g(1− p+ps, α) so that the survival criteria in Lemma 5.3.6 is that there exists{s } ,s ∈[0,1) for all nwith s →1
such that
g(1−pn+pnsn+1, α)6sn, for all n
It will be convenient to work on a simpler approximating function for g(·). For future reference we introduce the notation
h(s) := (−log(s))α (5.5)
The approximation we will work with is
g(1−p+ps)∼1−C·h(1−p+ps) = 1−C(−log(1−p+ps))α
holding for allC > 0 ass→1. Clearly, any function converging to 0 can be used in place of h for this approximation, but the reason for our particular choice will become clear shortly.
The survival criteria is now
1−Ch(1−pn(1−sn+1))6sn, some sequence sn→1
To establish the possibility of (sn)n>0 satisfying this criteria, we will make use of the following
deterministic lemma:
Lemma 5.3.7. For brevity, say that (pn)n>0 satisfies (?) if there exists b ∈(0,1) such that:
− n
X
k=1
γ−klogpk→ −logb
Fix someγ >1and suppose that(pn)n>0 is a sequence∈(0,1)which converges to 0. Note: in
what follows below the sequences(an)depend on the constantC but we suppress this notation for simplicity.
1. If (pn)n>1 satisfies (?), then for every C > 0 there exists a sequence (an)n>0 ∈ (0,1]
also converging to 0 such that pn>C a γ n
2. If for some C > 0 there exists a sequence (an)n>0 ∈ (0,1] also converging to 0 such
that pn>C a γ n
an+1 for all n >1, then (pn)n>1 satisfies (?).
Proof: For the direct portion we show that a sequence (an)n>1 satisfying the theorem can
be constructed. Observe that for the (n+ 1)th terman+1 to satisfy the stated condition, we
must have:
an+1 >C aγn pn Iterating this n times we see that
an+1 >Cγ+γ 2+...+γn aγ n 1 Qn k=1p γ(n−k) k
Now assume without loss of generality that C <1 so that the term involving C is bounded above by 1 (ifC >1 then it can be absorbed into the pk’s by pk 7→pk/C and the rest of the argument goes through). If we show that aγ1n/Qn
k=1p
γ(n−k)
k → 0, then this guarantees that we will be able to pick (an) satisfying the stated inequality with an ∈ (0,1) for all n and an →0.
Take logs and set the initial valuea1 =b− where∈(0, b). Then
log a γn 1 Qn k=1p γ(n−k) k =γnlogb− n X k=1 γ(n−k)logpk =γn logb− n X k=1 γ−klogpk ! Since Pn k=1γ −klogp
k→logb then the quantity inside the parenthesis eventually stays neg- ative and the entire RHS → −∞, which implies that the original ratio →0.
For the converse portion, suppose that the contrapositive is not true. That is, suppose −Pn
k=1γ
−klogp
k → ∞but that there exists (an)n>0 such that pk >C·
aγk ak+1
Rearranging this we have the relation
ak 6
pkak+1
C
1/γ
Now starting at a1 and expanding the right-hand side recursively n times we obtain
a1 6 (pγ1−1pγ2−2· · ·pγ−n n )a γ−n n+1 Cγ−1+γ−2+···+γ−n
SinceC < 1 thenγ−1+· · ·+γ−n6γ/(γ−1) for allnso the denominator is bounded above by C0 = Cγ/(γ−1). If C > 1 then this quantity is bounded by C0 = C1/γ. Combining this
with the fact that an <1 for all n, the inequality simplifies to
a1 6 Qn k=1p γ−k k C0 However, −Pn k=1γ −klogp k → ∞ implies Qnk=1pγ −k
k → 0, thus sending n → ∞ above
implies a1 = 0 and we have a contradiction.
From there it is only a short hop to the result for our approximation sequence:
Lemma 5.3.8. Again for brevity, say that (pn)n>0 satisfies (?) if there exists b∈(0,1)such
that: − n X k=1 γ−klogpk→ −logb
1. If (pn)n>1 satisfies (?), then for every C > 0 there exists a sequence (sn)n>0 ∈ [0,1)
converging to 1 such that 1−C·h(1−pn(1−sn+1))6sn for all n>1.
2. If for some C > 0 there exists a sequence (sn)n>0 ∈ [0,1) converging to 1 such that
1−C·h(1−pn(1−sn+1))6sn for all n >1, then (pn)n>1 satisfies (?).
It is sufficient to show the relevant properties for sequences (sn)n>0 satisfying the easier
inequality
1−C·hpn(1−sn+1)
iα
6sn for all n >1 (5.6) since, by the identity 1−x6−log(x), x > 0 and the facts thatpn, sn∈[0,1) the following inequality holds for all n:
1−C·h(1−pn(1−sn+1))61−C·
h
pn(1−sn+1)
iα
Inequality (5.6) implies that we want to study sequences (sn)n>0, sn ∈ [0,1) for all n such that:
pn>C·
(1−sn)1/α 1−sn+1
, for all n>1 (5.7)
Now applying the previous lemma with the map 1−sn=an gives the result.
Finally we need to tighten the result above concerning the approximating functionh(·) to the pgfg(·). The key tool is a Tauberian theorem giving the behavior of the Laplace-Stieltjes transform of a heavy-tailed random variable.
Theorem 5.3.9. ([16], Theorem 8.1.6) Let F(·)be the CDF of some probability distribution and φ(·) its Laplace-Stieltjes transform. Then for 0 6α <1 and a slowly-varying function `, the following are equivalent
(1) 1−φ(s)∼sα`(1/s), s↓0 (2) 1−F(x)∼ `(x)
xαΓ(1−α), x→ ∞
The relationship between the Laplace-Stieltjes transform φ(·) and the probability gen- erating function g(·) is φ(−logs) = f(s). Also, for our offspring distributions the function `(x) is a constantCα >0 for allα∈(0,1). So making the maps7→ −(logs), we can rewrite
Corollary 5.3.10. Letg(·)be the pgf of an offspring distribution satisfying 5.2 from Theorem 5.3.3. Then there exists C >0 such that
1−g(s)∼C·h(s), s→1
Where h(·) is defined in 5.5.
This implies the following simple Lemma:
Lemma 5.3.11. Given an offspring distribution satisfying 5.2, then there exist constants C0, C1 >0 such that
1. There exists s0 such that 1−g(s)6C0·h(s) for all s > s0
2. There exists s1 such that 1−g(s)>C1·h(s) for all s > s1
Proof: Solving (1) for C0 gives
1−g(s)
h(s) 6C0 for all s > s0
Now by Corollary 5.3.10, there exists C > 0 such that (1−g(s))/h(s) → C. Thus setting C0 > C implies the result. The proof for (2) is exactly analogous.
Combining Lemma 5.3.11 and Lemma 5.3.8, extracting the {pn}n>0 criterion from the
CHAPTER 6 Future directions
We discuss some potential extensions of the two works presented above, and some ideas for new work in unrelated areas.