a

r
X
i
v
:
0
7
0
7
.
2
5
0
5
v
1


[
m
a
t
h
.
N
T
]


1
7

J
u
l

2
0
0
7
Primitive Divisors in Arithmetic Dynamics
JOSEPH H. SILVERMAN
Abstract. Let (z) Q(z) be a rational function of degree at
least 2 satisfying (0) = 0 and 0 =

(0) Z. Let Q have in-

nite orbit under iteration of and write
n
() = A
n
/B
n
as a frac-
tion in lowest terms. We prove that for all but nitely many n 0,
the numerator A
n
has a primitive divisor, i.e., there is a prime p
such that p | A
n
and p A
i
for all i < n. More generally, we prove
an analogous result when is dened over a number eld and 0 is
a periodic point for .
Introduction
Let A = (A
n
)
n1
be a sequence of integers. A prime p is called a
primitive divisor of A
n
if
p | A
n
and p A
i
for all 1 i < n.
The Zsigmondy set of A is the set
Z(A) =
_
n 1 : A
n
does not have a primitive divisor
_
.
A classical theorem of Bang [1] (for b = 1) and Zsigmondy [25] in
general says that if a, b Z are integers with a > b > 0, then
Z
_
(a
n
b
n
)
n1
_
is a nite set.
Indeed, they prove that Z
_
(a
n
b
n
)
n1
_
contains no n > 6, which is a
strong uniform bound. This useful result has been extended and gener-
alized in many ways, for example to more general linear recursions, to
number elds, and to elliptic curves, see [4, 5, 7, 8, 11, 12, 18, 20, 21].
In this note we prove a Bang-Zsigmondy result for sequences associ-
ated to iteration of rational functions. For ease of exposition, we state
here a special case of our main result for dynanmical systems over Q.
See Theorem 5 for the general statement.
Date: March 30, 2008 (Draft 1).
1991 Mathematics Subject Classication. Primary: 11B37; Secondary: 11G99,
14G99, 37F10.
Key words and phrases. arithmetic dynamical systems, primitive divisor, Zsig-
mondy theorem.
Research supported by NSA H98230-04-1-0064 and NSF DMS-0650017.
1
2 JOSEPH H. SILVERMAN
Theorem 1. Let (z) Q(z) be a rational function of degree d 2.
Suppose that
(0) = 0 and 0 =

(0) Z.
Let Q be a point with innite orbit under iteration of . For
each n 1, write

n
() =
A
n
B
n
Q
as a fraction in lowest terms, where
n
denotes the nth iterate of .
Then the dynamical Zsigmondy set Z
_
(A
n
)
n0
_
is nite.
Remark 2. Rice [19] investigates primitive divisors in the case that
(z) Z[z] is a monic polynomial and Z. (See also [9] for some sim-
ilar results.) For example, Rice proves that if (z) = z
d
, if 0 is preperi-
odic for , and if Z is not preperiodic for , then Z
_
(
n
())
n0
_
is
nite. Our Theorem 5 is a generalization of this result to rational maps
over number elds under the assumption that

(0) = 0 and 0 is peri-

odic. (The assumption that

(0) = 0 is not essential, see Remark 12,

and presumably the same is true of the periodic versus preperiodic
assumption.)
A key tool in the proof of Theorem 1 is a dynamical analog [23] of
Siegels theorem [22, IX.3.1] for integral points on curves. Continuing
with the notation from Theorem 1, an elementary property of dynam-
ical canonical heights [24, 3.4] says that
lim
n
log max
_
|A
n
|, |B
n
|
_
d
n
=

h

() > 0.
A deeper result, proven in [23] as a consequence of Roths theorem,
implies that
lim
n
|A
n
|
d
n
=

h

() > 0, (1)
and an estimate of this sort is needed to prove Theorem 1.
Of course, there are many situations in which it is easy to prove
that (1) holds, for example if (z) Z[z] and Z. In such cases the
exact determination of the Zsigmondy set often becomes an elementary
exercise, see Example 11 and some of the examples in [9, 19].
Remark 3. It is enlightening to restate the assumptions in Theorem 1 in
dynamical terms. The assumption that (0) = 0 says that 0 is a xed
point of . (More generally, it suces to assume that 0 is a periodic
point, see Theorem 5.) Similarly, the assumption that

(0) Z says
that 0 is p-adically nonrepelling for every prime p. This is equivalent
to the assertion that 0 is in the p-adic Fatou set for every prime p.
Primitive Divisors in Arithmetic Dynamics 3
Remark 4. The rst natural question one asks about the Zsigmondy
set of a sequence is whether it is nite. Theorems 1 and 5 give an
armative answer for certain sequences dened by iteration of rational
maps on P
1
. Assuming that the Zsigmondy sets under consideration
are nite, it is also natural to ask for explicit upper bounds for
#Z(A) and max Z(A).
Naturally one hopes to nd bounds that depend only minimally on
the sequence. For example, Zsigmondys original theorem says that
for integers a > b > 0, we have max Z(a
n
b
n
) 6. A recent deep
result of Bilu, Hanrot and Voutier [4] extends this to the statement
the max Z(L) 30 for any nontrivial Lucas or Lehmer sequence L.
In this paper we are content to prove the niteness of certain dynami-
cal Zsigmondy sets. We leave the question of explicit and/or uniform
bounds as a problem for future study.
The material in this article is divided into two sections. In Section 1
we state and prove our main theorem via a sequence of lemmas, some
of which may be of independent interest. Section 2 discusses variants
of our main theorem and raises questions, makes conjectures, and in-
dicates directions for further research.
1. A dynamical Zsigmondy theorem
In this section we state and prove our main theorem concerning prim-
itive divisors in sequences dened by iteration of certain types of ratio-
nal functions. We start by recalling that primitive divisors in number
elds are most appropriately dened using ideals, rather than elements.
Denition. Let K be a number eld and let A = (A
n
)
n1
be a se-
quence of nonzero integral ideals. A prime ideal p is called a primitive
divisor of A
n
if
p | A
n
and p A
i
for all 1 i < n.
The Zsigmondy set of A is the set
Z(A) =
_
n 1 : A
n
does not have a primitive divisor
_
.
We also recall some basic denitions from dynamical systems.
Denition. Let (z) K(z) be a rational function of degree at least 2,
which we may view as a morphism : P
1
K
P
1
K
. A point P
1
(

K)
is periodic for if
n
() = for some n 1. The smallest such n
is called the -period of . Similarly, we say that is preperiodic
if
m+n
() =
m
() for some n 1 and m 0. Equivalently, is
preperiodic if its -orbit
_
, (),
2
(), . . .
_
is nite. A point that
4 JOSEPH H. SILVERMAN
is not preperiodic, i.e., that has innite -orbit, is called a wandering
point.
If is periodic with -period n, then the multiplier of is the quan-
tity

() = (
n
)

() =
n1

i=0

i
()
_
.
(If some iterate
i
() is equal to , then a dierent formula must be
used, see for example [24, 1.3 and Exercise 1.13].)
Theorem 5. Let K be a number eld with ring of integers R and
let (z) K(z) be a rational function of degree d 2. Let K be a
periodic point for whose multiplier

() satises
0 =

() R.
Let K be a wandering point, i.e., a point with innite -orbit, and
for each n 1, write the ideal
_

n
()
_
= A
n
B
1
n
as a quotient of relatively prime integral ideals. (If
n
() = , we
set A
n
= (1) and B
n
= (0).) Then the dynamical Zsigmondy set
Z
_
(A
n
)
n1
_
is nite.
Proof. The proof of Theorem 5 is structured as a series of lemmas that
provide the necessary tools.
Lemma 6. If Theorem 5 is true when is a xed point of , then it
is true when is a periodic point of .
Proof. Suppose that has -period k 2, so
k
() = and no smaller
iterate of xes . For each 0 i < k we consider the subsequence
_

nk+i
()
_
= A
nk+i
B
1
nk+i
for n = 0, 1, 2, . . . .
We claim that these subsequences have very few common prime divi-
sors. More precisely, dene the set of good primes P = P
,,
to be
primes satisfying the following two conditions:
(A) has good reduction at p. (See [24, Chapter 2] for the denition
and basic properties of maps with good reduction.)
(B)
i
() (mod p) for all 0 i < k.
It is clear that P contains all but nitely many primes. Now suppose
that some p P divides terms in dierent subsequences, say
p | A
nk+i
and p | A
mk+j
for some 0 j < i < k. (2)
Note that the good reduction assumption means that ( mod p)
n
=
(
n
) mod p, i.e., reduction modulo p commutes with composition of
Primitive Divisors in Arithmetic Dynamics 5
(see [24, Theorem 2.18]). Hence if p is a prime of good reduction for ,
then we have
p | A
n

n
() (mod p).
So we can rewrite assumption (2) as

nk+i
()
mk+j
() (mod p).
Suppose rst that nk+i > mk+j. Since 0 j < i k by assumption,
this implies that n m. We compute

nk+i
() (mod p)
=
(nk+i)(mk+j)
_

mk+j
()
_

(nk+i)(mk+j)
() (mod p)
=
ij
_
(
k
)
nm
()
_
=
ij
() since
k
() = .
But 0 < i j < k, so this contradicts Property (B) of P.
Similarly, if nk + i < mk + j, then m > n (since i > j), so we have

mk+j
() (mod p)
=
(mk+j)(nk+i)
_

nk+i
()
_

(mk+j)(nk+i)
() (mod p)
=
ji+k
_
(
k
)
mn1
()
_
=
ji+k
() since
k
() = .
This is again a contradiction of Property (B), since 0 < j i + k < k.
We have now proven that for primes p P, at most one of the
subsequences
(A
nk+i
)
n0
, i = 0, 1, . . . , k 1,
has a term divisible by p.
We are assuming that Theorem 5 is true if is a xed point. It
follows that for each 0 i < k, the Zsigmondy set Z
_
(A
nk+i
)
n0
_
is
nite, since (A
nk+i
)
n0
is the sequence associated to the map
k
, the
initial point
i
(), and the xed point of
k
. Note that the integrality
condition on the multiplier is still true, since by denition

() = (
k
)

() =

k ().
It follows that we can nd a number N so that for all n N and
all 0 i < k there is a prime ideal p
n,i
satisfying
p
n,i
| A
nk+i
and p
n,i
A
mk+i
for all 0 m < n.
6 JOSEPH H. SILVERMAN
In other words, for a xed i, the ideals p
n,i
are primitive divisors in
the subsequence (A
nk+i
)
n0
. Increasing N if necessary, we may assume
that p
n,i
P for all n and all i, since the complement of P is nite.
It is now clear that for n > N and 0 i < k, the prime p
n,i
is a
primitive divisor of A
nk+i
in the full sequence (A
m
)
m0
. This is true
because it is a primitive divisor in its own subsequence, and we proved
above that it does not divide any of the terms in any of the other
subsequences.
We next reduce to the case = 0, which will simplify our later
computations.
Lemma 7. It suces to prove Theorem 5 under the assumption that
= 0.
Proof. Let f(z) = z + . Then we have

n
() = f
1
_

n
()
_
= (f
1
f)
n
_
f
1
()
_
.
Hence replacing by f
1
f and replacing by f
1
() allows us
to replace wtih f
1
() = 0. Note that there is no problem with the
multiplier condition, since one easily checks that [24, Proposition 1.9]

() =
f
1
f
_
f
1
()
_
. =
f
1
f
(0).

So we are now reduced to the case that = 0 is a xed point of (z).

This means that we can write in the form
(z) =
a
1
z + a
2
z
2
+ + a
d
z
d
b
0
+ b
1
z + b
2
z
2
+ + b
d
z
d
, (3)
and without loss of generality we may assume that all a
i
and b
i
are in
the ring of integers R of K. Further, since deg = d, we know that at
least one of a
d
and b
d
is nonzero, and also b
0
= 0.
The multiplier at 0 is =

(0) = a
1
/b
0
, and by assumption we
have R with = 0, so a
1
= 0. We next prove an elementary,
but useful, lemma that bounds how rapidly the p-divisibilty of A
n
can
grow, where recall that A
n
is the integral ideal obtained by writing
_

n
()
_
= A
n
B
1
n
as a quotient of relatively prime integral ideals. In order to state the
result, we rst set some notation.
Denition. Assuming that we have xed a representation (3) for as
a quotient of polynomials with integral coecients, for each prime p
of R we let

p
= ord
p
a
1
.
Primitive Divisors in Arithmetic Dynamics 7
We note that for all p,

p
= ord
p
+ ord
p
b
0
ord
p
b
0
,
since by assumption the multiplier =

(0) = a
1
/b
0
is integral. We
then dene
s
p
= min{s 0 : ord
p
A
s
>
p
}.
(If no such s exists, we set s
p
= .) We call s
p
the generalized rank of
apparition at p. Notice that if
p
= 0, which is true for almost all p,
then s
p
is the usual rank of apparition, i.e., the index of the rst term
in the series that is divisible by p.
Lemma 8. With notation as above, the valuation of A
k1
has the fol-
lowing properties:
If k s
p
, then ord
p
A
k1

p
. (4)
If k > s
p
, then ord
p
A
k
= ord
p
A
k1
+ ord
p
. (5)
(Recall that =

(0) = a
1
/b
0
is the multiplier of at 0.)
Proof. We note that (4) is true by the denition of s
p
, so we only need
to prove (5). The assumption that k > s
p
tells us
ord
p
A
k1
>
p
= ord
p
a
1
ord
p
b
0
. (6)
For notational convenience, we let =
k1
(). From (6) we see
that ord
p
A
k1
> 0, so ord
p
() = ord
p
A
k1
. Thus
ord
p
> ord
p
a
1
ord
p
b
0
,
so when we evaluate the numerator and denominator of (), the lowest
degree terms have the strictly smallest p-adic valuation. Hence the
ultrametric triangle inequality gives
ord
p
(A
k
) = ord
p
_
()
_
= ord
p
(a
1
+ a
2

2
+ + a
d

d
)
ord
p
(b
0
+ b
1
+ + b
d

d
)
= ord
p
(a
1
) ord
p
(b
0
)
= ord
p
() + ord
p
(a
1
/b
0
)
= ord
p
(A
k1
) + ord
p
().
This completes the proof of Lemma 8.
Next we recall the denition and basic properties of the canonical
height associated to .
8 JOSEPH H. SILVERMAN
Lemma 9. The canonical height associated to is the function

h

:
P
1
(

K) P
1
(

K) dened by the limit

() = lim
n
1
d
n
h
_

n
()
_
.
It satises, and is characterized by, the two following properties:

() = h() + O(1) for all

K. (7)

_
()
_
= d

() for all

K. (8)
The O(1) constant in (7) depends on , but is independent of .
The values

h

() are nonnegative, and

() > 0 has innite -orbit.

Proof. See [6], [10, B.4], or [24, 3.4].
We next apply the main result of [23] to show that log N
K/Q
A
n
grows
like a constant multiple of d
n
. We observe that the proof of [23, The-
orem E] requires some sort of nontrivial theorem on Diophantine ap-
proximation such as Roths theorem, so despite its simple statement,
the following lemma conceals the deepest part of the proof of Theo-
rem 5. For our purposes, we require the general number eld version
proven in [23], but see [24, 3.8] for a more leisurely exposition of the
same result over Q.
Lemma 10. For every > 0 there is an n() so that for all n n()
we have
(1 )d
n

()
1
[K : Q]
log N
K/Q
A
n
(1 + )d
n

(). (9)
(More generally, the estimate (9) is valid as long as 0 is not an excep-
tional point for , or equivalently, if 0 is not a totally ramied xed
point of
2
.)
Proof. The assumption that the multiplier
0
() is nonzero implies
in particular that 0 is not an exceptional point of . Hence we can
apply [23, Theorem E] to deduce that for any xed embedding of K
into C, we have
lim
n

n
(), 0
_
d
n
= 0. (10)
Here is a logarithmic distance function on P
1
(C) which, in our situ-
ation, reduces to
(, 0) = 1 + log
_
max{||, 1}
||
_
= 1 + log max{1, ||
1
}.
Primitive Divisors in Arithmetic Dynamics 9
(See [23, 3].) Substituting this into (10), we obtain the equivalent
statement
lim
n
log max{1, |
n
()|
1
}
d
n
= 0. (11)
Note that (11) is true for every embedding of K into C, where the
dierent embeddings correspond to dierent choices of the absolute
value that appears in the numerator.
In general, if K

and if we write the ideal () as a quotient

of relatively prime integral ideals () = AB
1
, then the (normalized
logarithmic) height of is given by the formula
h() =
1
[K : Q]
_
log N
K/Q
B+

vM

K
[K
v
: Q
v
] log max{1, ||
v
}
_
. (12)
See for example [14, 3.1]. We apply (12) with =
n
()
1
, so B =
A
n
. Since h() = h(
1
) for any nonzero , this gives
h
_

n
()
_
=
1
[K : Q]
log N
K/Q
A
n
+

vM

K
[K
v
: Q
v
]
[K : Q]
log max{1, |
n
()|
1
v
}. (13)
We now divide both sides of (13) by d
n
and let n . The limit
formula (11) tells us that the sum over the archimedean absolute values
goes to 0. On the other hand, the left-hand side of (13) is exactly the
limit that denes the canonical height. Hence we obtain

() = lim
n
1
[K : Q]
log N
K/Q
A
n
d
n
.
This limit is equivalent to the statement (9) that we are trying to prove,
which completes the proof of Lemma 10.
We have assembled all of the tools needed to complete the proof of
Theorem 5. Our goal is to show that A
n
has a primitive prime divisor
for all suciently large n. Without loss of generality, we may assume
that
n > max{s
p
: p | a
1
and s
p
< }. (14)
Suppose that p is a prime that divides A
i
for some i < n. We consider
three cases.
Case A.
p
= 0
Then by denition s
p
is the smallest s such that p | A
s
, so in particu-
lar i s
p
. Repeated application of Lemma 8 (starting with k = i + 1)
10 JOSEPH H. SILVERMAN
yields
ord
p
A
i+j
= ord
p
A
i
+ j ord
p
for all j 0.
Taking j = n 1 i, we nd in particular that
ord
p
A
n1
ord
p
A
i
> 0,
so p | A
n1
.
Case B.
p
> 0 and s
p
<
Note that
p
> 0 is equivalent to p | a
1
, so (14) and the assumption
that s
p
< imply that n > s
p
. We apply Lemma 8 repeatedly,
starting from k = s
p
+ 1, to obtain
ord
p
A
sp+j
= ord
p
A
sp
+ j ord
p
, valid for all j 0.
Taking j = n 1 s
p
, which is permissible since n > s
p
, yields
ord
p
A
n1
= ord
p
A
sp
+ (n 1 s
p
) ord
p
ord
p
A
sp
1.
So in this case we again nd that p | A
n1
.
Case C.
p
> 0 and s
p
=
In this case we are not able to conclude that p | A
n1
, but from the
denition of s
p
we know that A
n
is not highly divisible by p. More
precisely, we know that
ord
p
A
n

p
= ord
p
a
1
. (15)
In order to prove that A
n
has a primitive prime divisor, we need to
prove that the product
E
n
:=

primes p that divide

one of A
0
, A
1
, . . . , A
n1
p
ord
p
An
(16)
is strictly smaller than A
n
. According to the three cases, every prime p
in the product dening E
n
either divides A
n1
, or else
p
> 0 and s
p
=
. Hence E
n
satises
E
n

p|A
n1
, sp<
p
ord
p
An

p>0, sp=
p
ord
p
An
. (17)
The second product in (17) comes from Case C, so we can bound it
using (15). Substituting this into (17) yields
E
n

p|A
n1
, sp<
p
ordp An
(a
1
). (18)
Primitive Divisors in Arithmetic Dynamics 11
We break the remaining product into two pieces and consider each
in turn:

p|A
n1
, sp<
p
ordp An
=

p|A
n1
pa
1
p
ordp An

p|A
n1
p|a
1
sp<
p
ordp An
. (19)
For the rst product in the right-hand side of (19), we have
p
= 0, so
the fact that p | A
n1
implies that s
p
n1. It follows from Lemma 8
that ord
p
A
n
= ord
p
A
n1
. (Note that p a
1
implies that ord
p
= 0,
since = a
1
/b
0
.) Hence the rst product in (19) satises

p|A
n1
pa
1
p
ordp An
=

p|A
n1
pa
1
p
ordp A
n1
. (20)
Next consider the second product in the right-hand side of (19).
Applying Lemma 8 and using the fact that n s
p
, we see that
ord
p
A
n
= ord
p
A
n1
+ ord
p
.
Hence

p|A
n1
p|a
1
sp<
p
ordp An
=

p|A
n1
p|a
1
sp<
p
ordp A
n1
+ordp

_

p|A
n1
p|a
1
p
ordp A
n1
_
(). (21)
Substituting (20) and (21) into (19) and then (19) into (18) yields
E
n
| A
n1
() (a
1
). (22)
As noted earlier, the ideal A
n
has a primitive prime divisor if and
only if A
n
is strictly larger than E
n
. We compute
log N
K/Q
A
n
log N
K/Q
E
n
log N
K/Q
A
n
log N
K/Q
(A
n1
a
1
) from (22),
[K : Q]

()
_
(1 )d
n
(1 + )d
n1
_
log N
K/Q
(a
1
)
from Lemma 10, valid for n n(),
=
1
4
[K : Q]

()d
n
log N
K/Q
(a
1
) taking =
3d 4
4d + 4
.
This is clearly positive if n is suciently large. More precisely, we have
proven that A
n
has a primitive prime divisor if n satises the following
three conditions:
(i) n > max{s
p
: p | a
1
and s
p
< }, i.e., n satises (14).
(ii) n n
_
3d4
4d+4
_
, where n() is the quantity dened in Lemma 10.
(iii) n > log
d
_
4 log N
K/Q
(a
1
)
[K:Q]

h()
_
.
12 JOSEPH H. SILVERMAN
Therefore A
n
has a primitive prime divisor for all suciently large n,
so the Zsigmondy set Z
_
(A
n
)
n1
_
is nite.
Example 11. It is easy construct specic examples, and even families
of examples, for which one can nd the full dynamical Zsgimondy set
by an elementary calculation. We illustrate with one such example,
see [9, 19] for others.
Let (z) = z
2
+ z, let = 0, and let Q with > 0. Then for
all n 1 we have gcd(A
n1
, B
n1
) = 1 and
A
n
B
n
=
A
2
n1
+ A
n1
B
n1
B
2
n1
,
so A
n
= A
2
n1
+ A
n1
B
n1
and B
n
= B
2
n1
. In particular,
A
0
| A
1
| A
2
|
so A
n
has a primitive prime divisor if and only if there is a prime p | A
n
with p A
n1
. But from A
n
= A
n1
(A
n1
+B
n1
), this means that A
n
has a primitive prime divisor if and only if A
n1
+ B
n1
> 1. This is
true for every n 1, so it follows that Z
_
(A
n
)
n1
_
= .
2. Questions and Speculations
It is interesting to ask to what extent one can relax or modify the
hypotheses in Theorem 5. We discuss this question in a series of re-
marks.
Remark 12. The condition that the multiplier

() be nonzero is not
essential, although some condition is required to avoid trivial coun-
terexamples such as (z) = z
d
with = 0. It is not hard to rework the
proof of Theorem 5 and show that it remains true provided that the
periodic point is not an exceptional point and that a certain gener-
alized multiplier of at is in the ring of integers of K. For example,
suppose that = 0 is a xed point. Then Theorem 5 is true if we
require that can be written in the form
(z) =
a
r
z
r
+ a
r+1
z
r+1
+ + a
d
z
d
b
0
+ b
1
z + b
2
z
2
+ + b
d
z
d
with all a
i
and b
i
integral, with a
r
= 0, with a
r
/b
0
integral, and with
1 r < d.
Remark 13. Presumably Theorem 5 remains true if we only assume
that is preperiodic, i.e., has a nite -orbit, rather than assuming
that is periodic. However, some care is needed in order to generalize
the argument, so we have not pursued it here. We observe that Rices
main result [19, Theorem 1.1] for (z) Z[z] and Z permits
Primitive Divisors in Arithmetic Dynamics 13
to be preperiodic, albeit in a situation where one can classify quite
precisely all situations where there exist preperiodic points.
Remark 14. Suppose that we drop the requirement in Theorem 5 that
the multiplier be in the ring of integers of K. For concreteness, we
return to the setting of Theorem 1, so (z) Q(z) satises (0) = 0,
but we no longer require that

(0) Z. As usual, we write

n
(0) =
A
n
/B
n
in lowest terms. Fix a prime p and let r 1 be the smallest
integer so that p | A
r
. In order to prove a Zsigmondy-type result, we
would like to bound the extent to which high powers of p can divide
later terms in the sequence, similar to the estimates proven in Lemma 8.
Unfortunately, no such estimates exist.
In dynamical terms, assuming that has good reduction at p, the
conditions p | A
r
and
r
(0) 0 (mod p) are equivalent, and both mean
that
r
(0) is p-adically close to the xed point 0. If 0 is a nonrepelling
xed point, then once an iterate of 0 is p-adically suciently close to 0,
we get full control over how p-adically close to 0 future iterates become.
This is the content of Lemma 8. However, if 0 is p-adically repelling,
i.e., if

(0) is not p-integral, then no matter how p-adically close some

iterate comes to 0, eventually the iterates move away, and then there is
no easy way to control how close they may be when they later return
to a small p-adic neighborhood of 0.
Of course, for proving Zsigmondy results, we are all right unless
there is a prime p for which some iterate is p-adically tremendously
close to 0. The hope that this does not happen leads us to make the
following conjecture.
Conjecture 15. Let K/Q be a number eld, let (z) K(z) be a
rational map of degree d 2 having 0 as a nonexceptional xed point,
and let K be a wandering point. Then for all primes p and all
> 0, there are only nitely many n 0 such that
ord
p

n
() d
n
. (23)
The analog of Conjecture 15 is true in the classical multiplicative
group and elliptic curve cases, and indeed it is true over the comple-
tion K
p
and can be proven by a purely local calculation in the formal
group. In the classical cases one gets an exponentially stronger estimate
of the form
ord
p

n
() C(, , p)n.
So one might ask if Conjecture 15 holds in the stronger form that there
exists a C > 0 such that there are only nitely many n 0 satisfying
ord
p

n
() Cn.
14 JOSEPH H. SILVERMAN
In view of the classical results, it is tempting to try to prove Conjec-
ture 15 via a similar local calculation. However, Rob Benedetto (private
communication) has sketched an argument using ideas from [2, 3] which
suggests that the conjecture is false if we repace Q with Q
p
, and indeed
it is false even if d
n
is replaced by a much faster growing function. Thus
in the dynamical setting over number elds, if Conjecture 15 is true,
then the proof will of necessity involve some sort of global argument.
We also note that over a number eld, ord
p

n
() cannot grow much
faster than O(d
n
), since
ord
p

n
() log Np log NA
n
[K : Q]h
_

n
()
_
= [K : Q]

n
()
_
+ O(1) = d
n
[K : Q]

() + O(1).
This, combined with Benedettos p-adic example that does grow more
rapidly, is a further indication that a proof of Conjecture 15 will require
a global argument.
Remark 16. If we change the assumptions of Theorem 5 to allow
to be a wandering point, then Zsigmondy-type theorems appear to be
more dicult to prove. Note that when stripped to their essentials,
proofs of Zsigmondy-type theorems have two main ingredients:
(1) Prove that the sequence A
n
grows very rapidly.
(2) Prove that once a prime p divides some term in the sequence,
it cannot divide later terms to an extremely high power.
Condition 1 is a global condition, and it is true for both preperiodic
and wandering . This may be proven using [23, Theorem E] as we did
in the proof of Lemma 10. Indeed, the proof of Lemma 10 only uses
the assumption that is not a totally ramied xed point of
2
. (If
is a totally ramied xed point of
2
, then A
n
need not grow, as shown
by the example (z) = z
d
and = 1/b for any integer b 2.)
Thus the diculty in proving a Zsigmondy-type theorem for wander-
ing is Condition 2. In the classical case of the multiplicative group
or an elliptic curve, the p-divisibility of later terms in the sequence can
be described quite precisely via properties of the formal group of the
group variety. Roughly speaking, if A
n
is the rst term divisible by p,
then
ord
p
(A
nm
) = ord
p
(A
n
) + ord
p
(m),
and no other terms are divisible by p. However, in the dynamical
setting there is no analogous uniform result, and indeed it is easy
to construct examples in which ord
p
(A
nm
) is arbitrarily much larger
than ord
p
(A
n
). For example, let
= 0, = p, and (z) = z
2
pz + p
e
.
Primitive Divisors in Arithmetic Dynamics 15
Then
ord
p
(A
0
) = ord
p
() = 1 and ord
p
(A
1
) = ord
p
(p
e
) = e.
Remark 17. As noted in Remark 3, the assumption in Theorem 5 that
the multiplier of at the periodic point is integral is equivalent to
the assumption that is in the p-adic Fatou set of for every prime p.
Thus in trying to generalize Theorem 5 to wandering , one might
make the assumption that is p-adically Fatou for every p. Even more
specically, one might ask if Theorem 5 is true for wandering when
has everywhere good reduction, since good reduction implies empty
Julia set [24, Theorem 2.17], so every point is Fatou.
Remark 18. Our main theorem is really about primes p such that the
reduction modulo p of a wandering point coincides with a given
periodic point . This is a natural question, but it is not the most
natural generalization of the classical divisibility sequences associated
to the mulitplicative group or to elliptic curves. In the classical case,
one studies the order of modulo p for varying primes ideals p, which
means the order of mod p as a torsion point. The dynamical analog
of torsion points are preperiodic points, so the most natural analog is
to look for primitive prime divisors of

n
() or, more generally,
m+n
()
m
().
We formulate two primitive divisor conjectures, analogous to Theo-
rem 5, corresponding to these situations.
Conjecture 19. Let K be a number eld, let (z) K(z) be a rational
function of degree d 2, and let K be a -wandering point. For
each n 1, write the ideal
_

n
()
_
= A
n
B
1
n
as a quotient of relatively prime integral ideals. Then the dynamical
Zsigmondy set Z
_
(A
n
)
n1
_
is nite.
Note that as discussed in Remark 16, there is no problem with the
growth of NA
n
. However, just as in the case of p-adically repelling
periodic points described in Remark 14, there is a potential problem of
primes reappering in the sequence to very high powers. For example,
if we take (z) = p z + p
e
z
2
and = 0, then
() = p and
2
() = (p) = p
e+2
.
In order to state the second conjecture, we need to dene the Zsig-
mondy set of a doubly indexed sequence.
16 JOSEPH H. SILVERMAN
Denition. Let
(A
m,n
)
m1
n0
be a doubly indexed sequence of ideals. We say that p is a primitive
prime divisor of A
m,n
if
p | A
m,n
and p A
i,j
for all i, j with 0 i < m or 1 j < n.
The we dene the Zsigmondy set Z(A
m.n
) to be the set
_
(m, n) : n 1, m 0, and A
m,n
has no primitive divisors
_
.
Conjecture 20. Let K be a number eld, let (z) K(z) be a rational
function of degree d 2, and let K be a -wandering point. For
each n 1 and m 0, write the ideal
_

m+n
()
m
()
_
= A
m,n
B
1
m,n
(24)
as a quotient of relatively prime integral ideals. Then the dynamical
Zsigmondy set Z
_
A
m,n
_
is nite.
Remark 21. In the classical Zsigmondy theorems for the multiplicative
group and for elliptic curves, every prime divides some term of the
sequence, but this is (probably) not true for (most) dynamically dened
sequences, and indeed in some cases one can prove that
_
p : p | A
n
for some n
_
is a set of density 0, see [13, 16, 17]. On the other hand, it is clear that
every prime divides at least one term in the doubly indexed dynamical
sequence A
m,n
dened by (24). In particular, if has good reduction
at p, the p | A
m,n
if and only if the orbit of mdoulo p has a tail of
length m and a cycle of length n.
References
[1] A. S. Bang. Taltheoretiske Undersogelser. Tidsskrift Mat., 4(5):7080, 130137,
1886.
[2] R. L. Benedetto. Examples of wandering domains in p-adic polynomial dynam-
ics. C. R. Math. Acad. Sci. Paris, 335(7):615620, 2002.
[3] R. L. Benedetto. Wandering domains in non-Archimedean polynomial dynam-
ics. Bull. London Math. Soc., 38(6):937950, 2006.
[4] Y. Bilu, G. Hanrot, and P. M. Voutier. Existence of primitive divisors of Lucas
and Lehmer numbers. J. Reine Angew. Math., 539:75122, 2001. With an
appendix by M. Mignotte.
[5] G. D. Birkho and H. S. Vandiver. On the integral divisors of a
n
b
n
. Ann.
of Math. (2), 5(4):173180, 1904.
[6] G. S. Call and J. H. Silverman. Canonical heights on varieties with morphisms.
Compositio Math., 89(2):163205, 1993.
[7] R. D. Carmichael. On the numerical factors of the arithmetic forms
n

n
.
Ann. of Math. (2), 15(1-4):3048, 1913/14.
Primitive Divisors in Arithmetic Dynamics 17
[8] G. Everest, G. Mclaren, and T. Ward. Primitive divisors of elliptic divisibility
sequences. J. Number Theory, 118(1):7189, 2006.
[9] A. Flatters. Arithmetic dynamics of rational maps. preliminary manuscript,
July 9, 2007.
[10] M. Hindry and J. H. Silverman. Diophantine Geometry: An Introduction, vol-
ume 201 of Graduate Texts in Mathematics. Springer-Verlag, New York, 2000.
[11] P. Ingram. Elliptic divisibility sequences over certain curves. J. Number Theory,
123(2):473486, 2007.
[12] P. Ingram and J. H. Silverman. Uniform estimates for primitive divisors in
elliptic divisibility sequences. preprint, December 2006.
[13] R. Jones. The density of prime divisors in the arithmetic dynamics of quadratic
polynomials, 2006. ArXiv:math.NT/0612415.
[14] S. Lang. Fundamentals of Diophantine Geometry. Springer-Verlag, New York,
1983.
[15] J. Lubin. Formal ows on the non-Archimedean open unit disk. Compositio
Math., 124(2):123136, 2000.
[16] R. W. K. Odoni. The Galois theory of iterates and composites of polynomials.
Proc. London Math. Soc. (3), 51(3):385414, 1985.
[17] R. W. K. Odoni. On the prime divisors of the sequence w
n+1
= 1 +w
1
w
n
.
J. London Math. Soc. (2), 32(1):111, 1985.
[18] L. P. Postnikova and A. Schinzel. Primitive divisors of the expression a
n
b
n
in algebraic number elds. Mat. Sb. (N.S.), 75 (117):171177, 1968.
[19] B. Rice. Primitive prime divisors in polynomial arithmetic dynamics. Integers
(electronic), 7:A26, 2007. 16 pages.
[20] A. Schinzel. Primitive divisors of the expression A
n
B
n
in algebraic number
elds. J. Reine Angew. Math., 268/269:2733, 1974.
[21] J. H. Silverman. Wieferichs criterion and the abc-conjecture. J. Number The-
ory, 30(2):226237, 1988.
[22] J. H. Silverman. The Arithmetic of Elliptic Curves, volume 106 of Graduate
Texts in Mathematics. Springer-Verlag, New York, 1992.
[23] J. H. Silverman. Integer points, Diophantine approximation, and iteration of
rational maps. Duke Math. J., 71(3):793829, 1993.
[24] J. H. Silverman. The Arithmetic of Dynamical Systems, volume 241 of Grad-
uate Texts in Mathematics. Springer-Verlag, New York, 2007.
[25] K. Zsigmondy. Zur Theorie der Potenzreste. Monatsh. Math. Phys., 3(1):265
284, 1892.
E-mail address: jhs@math.brown.edu
Mathematics Department, Box 1917 Brown University, Providence,
RI 02912 USA