Complete right tail asymptotic for the density of branching

processes with fractional generating functions

Anton A. Kutsenko
University of Hamburg, MIN Faculty, Department of Mathematics, 20146 Hamburg, Germany; email:
[email protected]
arXiv:2502.08477v1 [math.PR] 12 Feb 2025

Abstract
The right tail asymptotic series consisting of attenuating exponential terms are derived for
the densities of Galton-Watson processes with fractional probability generating functions.
The frequencies in the exponential factors form fractal structures in the complex plane. We
discuss conditions when the asymptotic series converges everywhere. The obtained right tail
asymptotic is compared with the standard integral representation of the density and with
the complete left tail asymptotic.
Keywords: Galton-Watson process, left and right tail asymptotic, Schröder and
Poincaré-type functional equations, Karlin-McGregor function, Fourier analysis

1. Introduction
The analysis of the tail asymptotic of the (integral of) density of the martingale limit
of supercritical Galton–Watson processes was initiated in [1]. The first asymptotic terms
in the right tail were found in [2] for the logarithm and clarified without the logarithm in
[3]. In addition to the mentioned papers, one can look at [4], [5], [6], [7], and references
therein. The current paper considers a special but fairly broad class of rational probability-
generating functions. In this case, it is possible to write the complete right tail asymptotic
series explicitly. Moreover, sometimes, these series converge to the density of the martingale
limit everywhere. We compare the series with two universal formulas for the density: the
complete left tail asymptotic series and the Fourier integral representation. Recently, in [8],
the universal right-tail bound for the integral of the density is expressed in terms of a simple
decaying exponent. Our right-tail asymptotic series also consists of decaying exponents, but
their form is quite complex - often, the distribution of frequencies has a fractal structure in
the complex plane. Thus, starting from the second asymptotic term (the first main frequency
is almost always real), the right tail asymptotic contains a lot of attenuating oscillations.
Sometimes, the presence and absence of oscillations in tail asymptotics may attract special
attention; see, e.g., [9], [10], [11].
The methods we use involve analytical techniques based on contour integration and
residue theory for functions with non-trivial distribution of poles in the complex plane.
Generally, these ideas can be used in the Fourier analysis of multiple functional iterations.
Preprint submitted to Elsevier February 13, 2025
Such iterative mappings are widely used in practice. A nice introduction to the mathematical
theory of iterative mappings is given in [12]. At the same time, the history of such mappings
begins with classical works [13], [14], [15], and [16].
We consider a simple Galton-Watson branching process in the supercritical case with the
minimum family size 1 - the so-called Schröder case. The Galton–Watson process is defined
by
XXt
Xt+1 = ξj,t , X0 = 1, t ∈ N ∪ {0},
j=1

where all ξj,t are independent and identically-distributed natural number-valued random
variables with the probability-generating function

G(z) := Ez ξ .

We assume that it is a ratio of two polynomials

P (z)
G(z) = , P (z) = p1 z + p2 z 2 + ... + pN z N , Q(z) = q0 + q1 z + ... + qM z M , M ⩾ 1. (1)
Q(z)
The probability of the minimum family size is 0 < r < 1, where we denote
p1
r := . (2)
q0
In our case, p0 = 0. The case of non-zero extinction probability p0 /q0 ̸= 0 can usually
be reduced to the supercritical case with the help of Harris-Sevastyanov transformation.
The corresponding information about branching processes is given in, e.g., [1] and [17]. We
assume that P and Q have no common factors. All Taylor coefficients at z = 0 of G(z)
are non-negative because they represent the probabilities of the offspring distribution. The
generating function satisfies also the identity G(1) = 1, and

E := G′ (1) < +∞, (3)

where E is the expectation of the offspring distribution. Then one may define the martingale
limit W = limt→+∞ E −t Xt , the density of which can be expressed as a Fourier transform of
some special function
Z +∞
1 z
p(x) = Π(iy)e−iyx dy, where Π(z) = lim G ◦ {z
... ◦ G}(1 + t ), (4)
2π −∞ t→+∞ | E
t

see, e.g., [18]. By definition, it is seen that Π satisfies the so-called Poincaré-type functional
equation
G(Π(z)) = Π(Ez), Π(0) = 1, Π′ (0) = 1. (5)
Since the Fourier integral in (4) is quite complex, any expansion and asymptotic analysis of
p(x) is welcome.
2
 To understand what Π(z) looks like, we should start with the structure of the Julia set for
the generating function G(z). The filled Julia set consists of points z in the complex plane
whose orbits are bounded under multiple actions of G(z). Since all the Taylor coefficients
of G(z) at z = 0 are non-negative and G(1) = 1, we conclude that at least the unit ball
{z : |z| ⩽ 1} belongs to the filled Julia set. Moreover, all the points that belong to the open
component J0 of the filled Julia set containing the open unit ball satisfy

... ◦ G}(z) → 0 for z ∈ J0 .

|G ◦ {z (6)
t

This is because z = 0 is the unique attracting point inside the open unit ball, and G is a
contraction mapping here; see details in, e.g., [19]. Due to the definition of Π(z), see (4)
and (5), the structure of J0 in the vicinity of 1 determines the behavior of Π(z). Roughly
speaking, the “zoomed” and transferred to z = 0 boundary of J0 near z = 1 divides the
complex plane into two parts with different behavior of Π(z). This fact is illustrated in Fig.
1. Since G(z) is a polynomial in this example, the right part corresponds to the domain
where Π(z) → ∞ when z → ∞. If G(z) is a fractional function, the behavior of Π(z) on the
right part of the complex plane can be different. Moreover, in both cases, the boundary can
consist of many connected components. However, the left part always contains the half plane
{z : Re z < 0} since the unit ball belongs to the Julia set. The “zoomed” and “shifted”
unit ball is exactly the left half-plane. One of the main characteristics here is the critical
angle ϑ, see Fig. 1, which can be defined as

ϑ = sup{φ : {1 − reiκ } ⊂ J0 for |κ| < φ and r → +0}. (7)

Since J0 contains the unit ball, we have ϑ ⩾ π. For all the examples I tested, ϑ > π strictly.
I cannot find in the literature the proof of the statement that ϑ > π for any reasonable G,
but it seems that it is not difficult to adapt the reasoning from, e.g., [20] involving the second
derivative G′′ (1) > 0 to obtain the proof. We skip this because it can be cumbersome and,
maybe, already done somewhere.
The first two figures in Fig. 1 are computed with the help of this site1 . I am not an
owner of this resource, but, in my opinion, this is one of the best places on the internet
where you can see the visualization of general concepts of holomorphic dynamics. Also, this
resource contains utilities for the visualization of Julia sets for rational functions, not only
for the polynomials, see this place2 .
The convergence of orbits (6) is exponentially fast, namely, as rt , or faster if one of the
orbit points is already 0, since G′ (0) = r and G(0) = 0, see (1) and (2). This allows us to
define an analytical function

Φ(z) = lim r−t G

| ◦ {z
... ◦ G}(z), z ∈ J0 . (8)
t→+∞
t

It is seen from the definition that Φ(z) satisfies the so-called Schröder-type functional equa-
tion
Φ(G(z)) = rΦ(z), Φ(0) = 0, Φ′ (0) = 1. (9)
3
 Π(z) → ∞
z→∞
ϑ 0

Π(z) → 0
z→∞

Figure 1: The filled Julia set (black area) for G(z) = 0.9z + 0.1z 6 and its zoom at z = 1. “Zoomed and
shifted” Julia set divides C onto two sectors with different behavior of Π(z).

4
The functional equations (5) and (9) allow us to construct a periodic function very useful in
the analysis of Π(z). One-periodic Karlin-McGregor function, see, e.g., [21], [22], and [23] is
defined by
K(z) = r−z Φ(Π(−E z )). (10)
Taking into account the periodicity of K(z) and the information about the domains of
definition of functions Π(z) and Φ(z) discussed above, it is not difficult to see that K(z) is,
at least, defined in the strip | Im z| < ϑ/(2 ln E). This fact also gives the rate of attenuation
of Fourier coefficients
+∞
πϑ
X
K(z) = θn e2πinz , |θn | ⩽ Cε e−( ln E −ε)|n| (11)
n=−∞

for any ε > 0 and some constants Cε > 0. Using (10), we obtain

Π(−z) = Φ−1 (z logE r K(logE z)). (12)

Since K(z), as an analytic periodic function, is uniformly bounded in the strip | Im z| <
(ϑ − ε)/(2 ln E) for any ε > 0, and Φ(z) ∼ z for z → 0, from (12), we deduce that

ϑ−ε
|Π(−z)| ⩽ Dε |z|logE r , | arg z| < (13)
2
for any ε > 0 and some Dε > 0. Recall that logE r < 0 because 0 < r < 1. Thus, Π(z) is
decaying to 0 when z → ∞ is inside the corresponding cone. Everything is ready to write
the right tail asymptotic expansion.
Theorem 1.1. If ϑ > π and logE r < −1 then
X
p(x) ∼ − Res(Π, ωα )e−ωα x , x → +∞, (14)
α

where {ωα } are all the poles of Π, arranged in ascending order of their real parts. Any ωα
can be written in the form

ωα = E n Π−1 (G−1 −1
i1 ◦ ... ◦ Gim (zj )), (15)

where n > m ⩾ 1 are some integer numbers, G−1 −1

ik is a branch of multi-valued function G ,
and zj is a zero of Q(z). The residue Res(Π, ω) can be expressed through the values and
derivatives of Π (and G) at points having the form (15).

Proof. Suppose ω is a pole of Π. Then E −n ω is a regular point for some large n, since Π(z)
is analytic for small z. By (5), we have

Π(E t E −n ω) = G ... ◦ G}(E −n ω).

| ◦ {z
t

5
 ... ◦ G}(E −n ω) should be a pole of G, which is a zero of
Thus, for some t < n the value |G ◦ {z
t
Q. Denoting such minimal t as m and the zero as zj , we obtain (15). It is seen also that the
residue can be computed in terms of derivatives Π and G at the points of the form (15).
Note that if deg P > deg Q and ω is a pole of Π then E s ω is a pole of Π as well for any
s ∈ N, since Π(Ez) = P (Π(z))/Q(Π(z)) and if Π(z) = ∞ then Π(Ez) = ∞ as well. This
means that n can be greater than m + 1 in (15).
The mapping G−1 is invertible and contracting in a neighborhood of z = 1, since G(1) = 1
and G′ (1) = E > 1. Thus, we can find open neighborhoods of 1, denote them as U1 ⊂ V1 ,
such that G : U1 → V1 and G−1 −1
0 : V1 → U1 . Here, G0 is the corresponding branch of G
−1

analytic in V1 . The function Π−1 is also analytic in some open neighborhood of z = 1, since
Π(0) = 1 and Π′ (0) = 1. Decreasing, if necessary, the size of V1 (and U1 ), we may assume
that Π : U0 → V1 and Π−1 : V1 → U0 , where U0 is some open neighborhood of 0. Thus

EΠ−1 (G0−1 (z)) = Π−1 (z), z ∈ U1 , (16)

because Π(Ez) = G(Π(z)). Hence, without loss of generality, we may assume that G−1 i1 ◦ ... ◦
G−1 (z
im j ) ∈
̸ U 1 in (15), since, otherwise i1 should be zero and we can shorten the composition
by (16). In turn, this means that Π−1 (G−1 −1
i1 ◦ ... ◦ Gim (zj )) ̸∈ U0 and, therefore, under the
n
mentioned assumptions |ωα | > RE in (15) for some constant R > 0 depending on the size
of U0 . Since |ωα | > RE n , n > m, and the number of branches G−1 i is finite, we deduce that
any bounded domain contains only a finite number of the poles. In other words, the set of
poles of Π has no finite condensation points. Thus, we can choose arbitrary large t > 0 such
that the line t + iR does not contain any poles of Π. Due to (4) and the Cauchy’s residue
theorem, we have
Z Z Z
1 −zx 1 −zx 1
p(x) = Π(z)e dz − Π(z)e dz + Π(z)e−zx dz
2πi iR 2πi t+iR 2πi t+iR
Z
X
−ωα x 1
=− Res(Π, ωα )e + qt (x), qt (x) := Π(z)e−zx dz. (17)
Re ω <t
2πi t+iR
α

We have ϑ > π and, hence, both infinite tails of the path t + iR must belong to the cone,
where (13) is fulfilled. Using (13), we obtain
Z
−tx e t e−tx ,
|qt (x)| ⩽ e Dt |z|logE r dy ⩽ D (18)
t+iR

for some constants Dt and D e t , since logE r < −1 the integral converges. Identities (17) and
estimate (18) show that asymptotic series (14) is true.
Remark 1. While (15) is a complete right tail asymptotic series for p(x), in many cases,
we cannot replace ∼ with =. For example, if Q(z) ≡ 1 and G(z) = P (z) is a polynomial of
β
degree deg G ⩾ 2 then p(x) has a super-exponential attenuation of the order e−x , x → +∞
with some β > 1. The function Π(z) is entire and has no poles. It is still attenuating in the
left part of C, see Fig. 1. However, due to Π(Ez) = G(Π(z)), it has a super-exponential
6
grows of the order exp(z logE deg G ) in the right part of Fig. 1. To obtain the super-exponential
attenuation of its Fourier transform p(x) for x → +∞, one may apply the method of steepest
descent. Research on the right tail asymptotic for p(x) in this case is more or less finished
in [3] - at least, for the first asymptotic term.
Remark 2. Let us return to the case Q(z) ̸= const. If deg P ⩽ deg Q + 1 - we write it
as deg G ⩽ 1 - then Π(z) has no exponential growth in the right part of the complex plane,
see the reasoning set out in the previous remark. Hence, taking arbitrary large t in (17), we
can make qt (x) to be arbitrarily small. Thus, we expect that ∼ in (14) can be replaced with
= in this case.
Remark 3. For the Schröder case, research on the left-tail asymptotic was initiated in
[18], significantly advanced in [2], and finished in [19]. The corresponding result from [19] is
formulated in terms of the Karlin-McGregor function K(z), see (10).
Theorem 1.2. If ϑ > π and logE r < −1 then

p(x) = xα V1 (x) + xα+β V2 (x) + xα+2β V3 (x) + ..., x > 0, (19)

with α := −1 − logE r > 0, β := − logE r > 1, and

+∞ 1
− ln x θn∗m e2πinz
X Z
Vm (x) = Km ( ), Km (z) = κm 2πin+m ln r
, θn∗m = K(x)m e−2πinx dx,
ln E n=−∞
Γ(− ln E ) 0
(20)
where κm are Taylor coefficients of

Φ−1 (z) = κ1 z + κ2 z 2 + κ3 z 3 + .... (21)

This Theorem is proven in [19] under the assumption that G(z) is entire, but it is easy to
extend the result to rational G(z) and even more big classes of G(z) regular at z = 1. Using
Stirling’s approximation, see, e.g., [24],
r  z   
2π z 1
Γ(z) = 1+O , | arg z| < π − ε,
z e z
we, for some constant B > 0, obtain
r
4π 2 n2 +(ln r)2 m2 2
π |n|
− mlnlnEr ln − ln E
(ln E)2
 
2πin + m ln r Be
Γ − ⩾ p . (22)
ln E 4π 2 n2 + (ln r)2 m2
The constants Cε > 0 in (11) can be chosen such that
+∞
πϑ
X
K(z) =m
θn∗m e2πinz , |θn∗m | ⩽ (Cε )m e−( ln E −ε)|n| , ε > 0, m ⩾ 1. (23)
n=−∞

The standard estimates for Taylor coefficients of analytic functions show |κm | ⩽ K m , see
(21), for some K > 0 depending on the convergence radius for Φ−1 (z). Combining these
7
estimates with (22) and (23), and using the conditions ϑ > π and ln r < 0 of Theorem 1.2,
we obtain
κm θn∗m
2πin+m ln r
⩽ Ae−αm ln(m+|n|)−β|n| for some A, α, β > 0. (24)
Γ(− ln E )

Thus, the convergence of the Fourier series in (20) is exponentially fast, and the convergence
of the series (19) is super-exponentially fast.
In the next section, we compare three formulas (4), (14), and (19) in a series of examples.
We focus on the cases deg P ⩽ deg Q + 1 for which (14) should converge to p(x) everywhere
(x > 0). Also, when G−1 (z) admits closed-form expression, (14) demonstrates the fastest
way to compute p(x). However, if deg P > deg Q + 1 then (14) does not coincide with p(x)
- the maximum that it gives is an asymptotic expansion for large x. At the same time, the
other two formulas (4) and (15) are universal, but the exact behavior of p(x), x → +∞, is
invisible from them.

2. Examples
For most examples, Embarcadero Delphi Rad Studio Community Edition and the li-
brary NesLib.Multiprecision is used. This software provides a convenient environment for
programming and well-functioning basic functions for high-precision computations. All the
algorithms related to the article’s subject, including efficient parallelization, are developed
by the author (AK).
2.1. General remarks on computational procedures.
For the numerical results, we need to compute many functions. Let us discuss numerical
schemes for some of them. From (4), we have
z
Π(z) = lim Πt (z),
t→+∞ | ◦ {z
Πt (z) := G ... ◦ G}(1 +
Et
). (25)
t

Denoting the second Taylor term at z = 1 as

G(z) = 1 + E(z − 1) + (z − 1)2 G1 (z), (26)

we obtain
z2
 
z
Πt+1 (z) = Πt z + t G1 (1 + t+1 ) . (27)
E E
Numerical schemes based on (27) demonstrate exponentially fast convergence since E > 1.
Differentiating the composition in (25), it is not difficult to obtain a convenient identity for
the numerical computation of the derivative Π′ (z).
From (8), we have

Φ(z) = lim Φt (z), Φt (z) = r−t G

| ◦ {z
... ◦ G}(z). (28)
t→+∞
t

8
Denoting the second Taylor term at z = 0 as

G(z) = rz + z 2 G0 (z), (29)

we obtain
Φt+1 (z) = Φt (z) + rt−1 Φt (z)2 G0 (rt Φt (z)). (30)
Numerical schemes based on (30) demonstrate exponentially fast convergence since 0 < r <
1.
From (9), we have
Φ−1 (rz) = G(Φ−1 (z)). (31)
Substituting Taylor expansions

G(z) = rz + r2 z 2 + r3 z 3 + ..., Φ−1 (z) = κ1 z + κ2 z 2 + κ3 z 3 + ... (32)

into (31) and using initial condition κ1 = 1, one can express κj through rj , step by step.
However, κj grows exponentially. It makes sense to compute aj κj instead of κj , where a is
small enough. For this, one may set κ1 = a instead of κ1 = 1.
When n → +∞, the Fourier coefficients θn∗m should be divided by the small values of the
Gamma function, see (20) and (24). A good strategy is to compute the normalized Fourier
coefficients Z 1
2πny ∗m
e θn = K(x − iy)m e−2πinx dx, n ⩾ 1, y ⩾ 0, (33)
0
∗m
and use the symmetry θ−n = θn∗m . The parameter y in (33) should be chosen close enough
to the edge of the strip where K is defined, see above (11). Taking appropriate a and y, it
is not difficult to achieve accurate and stable computations of the ratio in (20), namely

κm θn∗m
 
m 2πny ∗m 2πin + m ln r
= exp ln(a κm ) + ln(e θn ) − ln Γ(− ) − m ln a − 2πny .
Γ(− 2πin+m
ln E
ln r
) ln E
(34)
If necessary, one may also try to compute (bK(z)) with some small b instead of K(z)m ,
m

but we didn’t have such a need. In addition, note that the computation of ln Γ is much more
stable than Γ.
From (5), we obtain
Π−1 (G(z)) = EΠ−1 (z), (35)
which is, up to initial conditions and constants, similar to (9). Thus, for the computation
of Π−1 , one may use fast algorithms similar to those for Φ(z), see (28)-(30).
2.2. Example 1.
Let us consider the probability generating function

z + z2
G(z) = . (36)
3−z

9
 Figure 2: A fragment of the filled Julia set (black area) for G defined in (36).

The filled Julia set for G(z) is given in Fig. 2.

There are two branches of the inverse function
√ √
−1 −1 − z + z 2 + 14z + 1 −1 −1 − z − z 2 + 14z + 1
G0 (z) = , G1 (z) = . (37)
2 2
The probability of the minimal family size and the expectation are
1
r = G′ (0) = , E = G′ (1) = 2. (38)
3
There is one pole z = 3 for G. We can enumerate all the poles of Π, see (15), with the help
of the following scheme. The first set of “primary” poles consists of

ω0,0 = 2Π−1 (3), ωj,0 = 2n+1 Π−1 ◦ G−1 −1

jn ◦ ... ◦ Gj1 (3), j ⩾ 1, (39)

where
j = {jn ...j1 }2 (40)
is the dyadic representation of j. Note that the last element is always jn = 1. Hence, we
do not count poles twice because 2Π−1 (G−1 −1
0 (z)) = Π (z). Due to Π(2z) = G(Π(z)) - if
Π(z) = ∞ then Π(2z) = ∞ as well, other poles have the form

ωj,k = 2k ωj,0 , k ⩾ 1, (41)

again, in accordance with (15). The distribution of ωj,k is plotted in Fig. 3. It resembles
the structure of the zoomed Julia set near z = 1. Such a phenomenon is similar to that of
10
 0 3000

(a) a fragment of the filled Julia set from Fig. 2 (b) Distribution of poles, see (39) and (41)

Figure 3: (a) The filled Julia set (black area) from Fig. 2 near z = 1 zoomed in about 1000 times; (b)
Distribution of poles of Π(z), see (4), in complex plane, for G(z) defined in (36).

the distribution of zeros of Π(z), see, e.g., [25].

Using Π(2z) = G(Π(z)) with G defined in (36), one can find the expression for the
residue
−24
Res(Π, ωj,0 ) = ′ ωj,0 , Res(Π, ωj,k ) = (−2)k Res(Π, ωj,0 ). (42)
Π( 2 )
Everything is ready to compute pa (x) by RHS of (14), where we use about 5000 “primary”
poles ωj,0 , j ⩽ 5000, and ωj,k generated by the “primary” poles, all satisfy |ωj,k | ⩽ 5000. For
the computation of exact p(x) by (4), we use the interval of integration y ∈ [−2 · 105 , 2 · 105 ]
divided by 2 · 107 points in the trapezoidal rule. For the computation of pb (x) given by RHS
in (19), we use about 30 terms Vj (x), j = 1, ..., 30 for each of which we use about 50 (or 100
with the symmetry) Fourier coefficients. For the computation of Π and Φ, we use about 150
iterations. The comparison of p(x), pa (x), and pb (x) is given in Fig. 4. They are almost
coincide. In this case, the computation of pa (x) is fastest. However, if deg P > deg Q + 1 in
(1), we cannot expect that pa (x) will coincide with p(x) = pb (x). In (36), deg P = deg Q + 1
and everything is OK, see Remark 2 above.
Let us discuss the computation of the coefficients κj , see (21). From (31), we have
z
(3 − Φ−1 (z))Φ−1 ( ) = Φ−1 (z) + Φ−1 (z)2 , (43)
3
which leads to
N
X −1
N −1 −1
κN = −(3 − 1) κj κN −j (1 + 3−j ), κ1 = 1. (44)
j=1
11
 p, pa , pb
1

0.5

0 0.5 1.0 1.5 2.0 2.5 x

Figure 4: Comparison of three formulas (4), (14), and (19) for the computation of the density p in case of
the probability generating function (36).

Setting κ1 = a with a = 1/5.3 instead of κ1 = 1 we compute aj κj , which is more accurate

than the direct computation of κj because of their exponential growth, see the discussion
after (32). We use y = 2.9 in (33) to compute the normalized Fourier coefficients. The
integrals in (33) are computed with 106 uniformly distributed nodes in the trapezoidal rule.
2.3. Example 2.
Let us consider the probability-generating function
4(z + z 2 )
G(z) = . (45)
9 − z2
The filled Julia set for G(z) is given in Fig. 5.
There are two branches of the inverse function
p p
−2 + 9z(z + 4) + 4 −2 − 9z(z + 4) + 4
G−1
0 (z) = , G−1
1 (z) = . (46)
4+z 4+z
The probability of the minimal family size and the expectation are
4 7
r = G′ (0) = , E = G′ (1) = . (47)
9 4
There are two poles z = 3 and z = −3 for G. We can enumerate all the poles of Π, see (15),
with the help of the following scheme. The first set of poles consists of
 n+1
7 −1 7
ω0,0 = Π (3), ωj,0 = Π−1 ◦ G−1 −1
jn ◦ ... ◦ Gj1 (3), j ⩾ 1, (48)
4 4
and the second set
 n+1
7 7
ω0,1 = Π−1 (−3), ωj,1 = Π−1 ◦ G−1 −1
jn ◦ ... ◦ Gj1 (−3), j ⩾ 1, (49)
4 4
where
j = {jn ...j1 }2 (50)
12
 Figure 5: The filled Julia set for G defined in (45).

is the dyadic representation of j. In contrast to the previous example, if ω is a pole of Π,

then 7ω/4 is not a pole because G(∞) ̸= ∞. The distribution of ωj,k is plotted in Fig. 6. It
resembles the structure of the zoomed Julia set near z = 1.
Using Π(7z/4) = G(Π(z)) with G defined in (45), one can find the expression for the
residue
−14 7
Res(Π, ωj,0 ) = 4ωj,0 , Res(Π, ωj,1 ) = 4ωj,1 . (51)
′
Π( 7 ) ′
Π( 7 )
While the distribution of the poles has some certain structure, the distribution of residues
is mostly chaotic. We plot some of them in Fig. 7.
Everything is ready to compute pa (x) by RHS of (14), where we use about 2500 poles
ωj,0 , j ⩽ 2500, and the same amount of ωj,1 : among them we select only those that satisfy
|ωj,k | ⩽ 5000. For the computation of exact p(x) by (4), we use the interval of integration
y ∈ [−2 · 105 , 2 · 105 ] divided by 2 · 107 points in the trapezoidal rule. For the computation of
pb (x) given by RHS in (19), we use about 40 terms Vj (x), j = 1, ..., 40 for each of which we
use about 60 (or 120 with the symmetry) Fourier coefficients. In comparison to the previous
example, we should increase the number of harmonics to achieve the same accuracy about
10−4 on the interval x ∈ (0, 5). If we take x ∈ (0, 3), as it is plotted in Fig. 8, then a much
smaller number of harmonics is acceptable. For the computation of Π and Φ, we use about
150 iterations. The comparison of p(x), pa (x), and pb (x) is given in Fig. 8. They are almost
identical. In this case, the computation of pa (x) is fastest. However, if deg P > deg Q + 1
in (1), we cannot expect that pa (x) will coincide with p(x) = pb (x). In (45), deg P = deg Q

13
 0 3000

(a) a fragment of the filled Julia set from Fig. 5 (b) Distribution of poles, see (48) and (49)

Figure 6: (a) The filled Julia set (black area) from Fig. 5 near z = 1 zoomed in about 1000 times; (b)
Distribution of poles of Π(z), see (4), in complex plane, for G(z) defined in (45).

and everything is OK, see Remark 2 above.

Let us discuss the computation of the coefficients κj , see (21). From (31), we have
4z
(9 − Φ−1 (z)2 )Φ−1 ( ) = 4(Φ−1 (z) + Φ−1 (z)2 ), (52)
9
which leads to
  N −1 −1  N −1 N −2 j N −j−1 
4 X X 4 X
κN = 4 −4 4 κj κN −j + κj κk κN −j−k , κ1 = 1, (53)
9 j=1 j=1
9 k=1

where the second double sum is zero for N = 2. Setting κ1 = a with a = 1/4.1 instead of
κ1 = 1, we compute aj κj , which is more accurate than the direct computation of κj because
of their exponential growth, see the discussion below (32). We use y = 4 in (33) to compute
the normalized Fourier coefficients. The integrals in (33) are computed with 106 uniformly
distributed nodes in the trapezoidal rule.
2.4. Example 3.
Let us consider the probability-generating function
2z
G(z) = . (54)
(3 − z)(2 − z)
The (filled) Julia set for G is almost the whole complex plane except for some sparse real
sets. There are two branches of the inverse function
√ √
−1 5z + 2 − z 2 + 20z + 4 −1 5z + 2 + z 2 + 20z + 4
G0 (z) = , G1 (z) = . (55)
2z 2z
14
 Figure 7: First 14 · 104 residues (51) are plotted in the complex plane.

p, pa , pb
1

0.5

0 0.5 1.0 1.5 2.0 2.5 x

Figure 8: Comparison of three formulas (4), (14), and (19) for the computation of the density p in case of
the probability generating function (45).

The probability of the minimal family size and the expectation are
1 5
r = G′ (0) = , E = G′ (1) = . (56)
3 2
There are two poles z = 2 and z = 3 for G. We can enumerate all the poles of Π, see (15),
with the help of the following scheme. The first set of poles consists of
 n+1
5 −1 5
ω0,0 = Π (2), ωj,0 = Π−1 ◦ G−1 −1
jn ◦ ... ◦ Gj1 (2), j ⩾ 1, (57)
2 2

and the second set

 n+1
5 5
ω0,1 = Π−1 (3), ωj,1 = Π−1 ◦ G−1 −1
jn ◦ ... ◦ Gj1 (3), j ⩾ 1, (58)
2 2

where
j = {jn ...j1 }2 (59)
15
is the dyadic representation of j. Using Π(5z/2) = G(Π(z)) with G defined in (54), one can
find the expression for the residue
−10 15
Res(Π, ωj,0 ) = 2ω , Res(Π, ωj,1 ) = 2ω . (60)
Π′ ( 5j,0 ) Π′ ( 5j,1 )
The poles (57), (58) and the residues (60) are all real. Hence, in contrast to the previous two
examples, there are no oscillations in representation (14). Everything is ready to compute
pa (x) by RHS of (14), where we use about 2500 poles ωj,0 , j ⩽ 2500, and the same amount
of ωj,1 . For the computation of exact p(x) by (4), we use the interval of integration y ∈
[−2 · 105 , 2 · 105 ] divided by 2 · 107 points in the trapezoidal rule. For the computation of
pb (x) given by RHS in (19), we use about 30 terms Vj (x), j = 1, ..., 30 for each of which we
use about 50 (or 100 with the symmetry) Fourier coefficients. For the computation of Π and
Φ, we use about 150 iterations. The comparison of p(x), pa (x), and pb (x) is given in Fig.
8. They are almost identical. In this case, the computation of pa (x) is fastest. However, if
deg P > deg Q + 1 in (1), we cannot expect that pa (x) will coincide with p(x) = pb (x). In
(54), deg P = deg Q − 1 and everything is OK, see Remark 2 above.
p, pa , pb
1

0.5

0 0.5 1.0 1.5 2.0 2.5 x

Figure 9: Comparison of three formulas (4), (14), and (19) for the computation of the density p in case of
the probability generating function (54).

Let us discuss the computation of the coefficients κj , see (21). From (31), we have
z
(3 − Φ−1 (z))(2 − Φ−1 (z))Φ−1 ( ) = 2Φ−1 (z), (61)
3
which leads to
 −1  N −1 N −2 N −j−1 
2 X κj κN −j X κj X
κN = 2 − −5 + κk κN −j−k , κ1 = 1, (62)
3N −1 j=1
3j j=1
3j k=1

where the second double sum is zero if N = 2. Setting κ1 = a with a = 1/2 instead of
κ1 = 1 we compute aj κj , which is more accurate than the direct computation of κj because
of their exponential growth, see the discussion after (32). We use y = 3 in (33) to compute
the normalized Fourier coefficients. The integrals in (33) are computed with 106 uniformly
distributed nodes in the trapezoidal rule.
16
Acknowledgements
This paper is a contribution to the project S1 of the Collaborative Research Centre TRR
181 ”Energy Transfer in Atmosphere and Ocean” funded by the Deutsche Forschungsge-
meinschaft (DFG, German Research Foundation) - Projektnummer 274762653.

References
[1] Harris T.E.: Branching processes. Ann. Math. Stat. 41, 474-494 (1948)
[2] Biggings, J. D., Bingham, N.H.: Large deviations in the supercritical branching process. Adv. Appl.
Prob. 25, 757-772 (1993)
[3] Fleischmann, K., Wachtel, V.: On the left tail asymptotics for the limit law of supercritical Gal-
ton–Watson processes in the Böttcher case. Ann. Inst. H. Poincaré Probab. Stat. 45, 201–225 (2009)
[4] Bingham, N. H., Doney R. A.: Asymptotic properties of supercritical branching processes. I: The Galton-
Watson process. Adv. Appl. Probab. 6, 711–731 (1974)
[5] Bingham, N. H., Doney R. A.: Asymptotic properties of supercritical branching processes. II:
Crump–Mode and Jirina processes. Adv. Appl. Probab. 7, 66–82 (1975).
[6] De Meyer, A.: On a theorem of Bingham and Doney. J. Appl. Probab. 19, 217–220 (1982).
[7] Wachtel, V.I., Denisov, D.E., Korshunov, D.A.: Tail asymptotics for the supercritical Galton-Watson
process in the heavy-tailed case. Proc. Steklov Inst. Math. 282, 273–297 (2013).
[8] Fernley, J., Jacob, E.: A universal right tail upper bound for supercritical Galton–Watson processes
with bounded offspring. Stat. Probabil. Lett. 209, 110082 (2024)
[9] B. Derrida, C. Itzykson, and J. M. Luck, “Oscillatory critical amplitudes in hierarchical models”. Comm.
Math. Phys. 94, 115–132 (1984)
[10] O. Costin and G. Giacomin, “Oscillatory critical amplitudes in hierarchical models and the Harris
function of branching processes”. J. Stat. Phys. 150, 471–486 (2013)
[11] B. Derrida, S. C. Manrubia, D. H. Zanette. “Distribution of repetitions of ancestors in genealogical
trees”. Physica A 281, 1-16 (2000)
[12] Milnor, J.: Dynamics in One Complex Variable. Univ. Press, Princeton NJ (2006)
[13] Schröder, E.: Ueber iterirte Functionen. Math. Ann. 3 (2), 296–322 (1870)
[14] Koenigs, G.: Recherches sur les intégrales de certaines équations fonctionnelles. Annales Sci. de l’École
Normale Sup. de Paris 1, 3-41 (1884)
[15] Poincaré, H.: Sur une classe nouvelle de transcendantes uniformes. J. Math. Pures Appl., Ser. 4 6,
313-365 (1890)
[16] P. Fatou: Memoire sur les equations fonctionnelles. Bull. Soc. Math. Fr. 47, 161-271; 48 33-94, 208-314
(1919)
[17] Bingham, N.H.: On the limit of a supercritical branching process. J. Appl. Probab. 25, 215–228 (1988)
[18] Dubuc, S.: La densité de la loi-limite d’un processus en cascade expansif. Z. Wahrsch. Verw. Gebiete
19, 281-290 (1971)
[19] Kutsenko, A. A.: Complete Left Tail Asymptotic for the Density of Branching Processes in the Schröder
Case. J. Fourier Anal. Appl. 30, 39 (2024).
[20] Kutsenko, A.A.: Approximation of the number of descendants in branching processes. J. Stat. Phys.
190, 68 (2023)
[21] Karlin, S., McGregor, J.: Embeddability of discrete-time branching processes into continuous-time
branching processes. Trans. Am. Math. Soc. 132, 115-136 (1968)
[22] Karlin, S., McGregor, J.: Embedding iterates of analytic functions with two fixed points into continuous
groups. Trans. Am. Math. Soc. 132, 137-145 (1968)
[23] Dubuc, S.: Étude théorique et numérique de la fonction de Karlin-McGregor. J. Anal. Math. 42, 15–37
(1982)
[24] Whittaker, E.T., Watson, J.N.: A Course of Modern Analysis. Fifth Edition by Moll, V.H. (ed).
Cambridge Univ. Press, Cambridge (2021)
17
[25] Kutsenko, A.A.: An entire function connected with the approximation of the golden ratio. Am. Math.
Month. 127, 820-826 (2020).

Notes
1
https://www.marksmath.org/visualization/polynomial julia sets/
2
https://www.marksmath.org/visualization/rational julia sets/

18