Hans Peter Schlickewei · Wolfgang M. Schmidt · Michel Waldschmidt
Zeros of linear recurrence sequences
Received: 7 August 1998
Abstract. Letf (x) = P0(x)αx0+ · · · +Pk(x)αxk
be an exponential polynomial over a field of zero characteristic. Assume that for each pair
i, jwith
i6=j,
αi/αjis not a root of unity. Define
1=Pkj=0(degPj+1). We introduce a partition of
α0, . . . , αk
into subsets
αi0, . . . , αiki(1
≤i≤m), which induces a decomposition offinto
f =f1+· · ·+fm, so that, for 1
≤i≤m,
αi0: · · · :αiki∈Pki(Q)
, while for 1
≤i6=u≤m, the number
αi0/αu0either is transcendental or else is algebraic with not too small a height. Then we show that for all but at most exp
1(5
1)51solutions
x∈Zof
f (x)=0, we have
f1(x)= · · · =fm(x)=0
.1. Introduction
Let K be a field of zero characteristic, α
1, . . . , α
kbe non-zero elements of K and P
1, . . . , P
knon-zero polynomials with coefficients in K . Consider an exponential polynomial
f (x) = X
k j=0P
j(x)α
jx.
We study the equation
f (x) = 0 ( 1 . 1 )
in x ∈ Z . We suppose that for each pair i, j with i 6= j ,
α
i/α
jis not a root of unity. (1.2)
H. P. Schlickewei: Fachbereich Mathematik, Universität Marburg, Hans-Meerwein-Strasse, Lahnberge, D-35032 Marburg, Germany.
e-mail: hps@mathematik.Uni-Marburg.de
W. M. Schmidt: Department of Mathematics, University of Colorado, Campus Box 395, Boulder, CO 80309-0395, USA. e-mail: schmidt@euclid.colorado.edu
M. Waldschmidt: Université P. et M. Curie (Paris VI), Institut Mathématique de Jussieu, Problèmes Diophantiens, Case 247, 4, Place Jussieu, F-75252 Paris Cedex 05, France. e- mail: miw@math.jussieu.fr
Mathematics Subject Classification (1991): 11B37, 11D61, 11J13
We set
1(f ) = X
k j=0( deg P
j+ 1 ).
It is well known that f ( 0 ), f ( 1 ), . . .
is a linear recurrence sequence of order 1(f ), which is “non-degenerate”. Vice versa, any non-degenerate linear recurrence sequence u
0, u
1, . . .
of elements of K of order q has some representation u
n= f (n) , where f is an exponential polynomial as above satisfying (1.2) and with 1(f ) = q . For more details, cf. e.g. [8].
So in studying (1.1) we study the zeros of linear recurrence sequences. An old conjecture says that the number of solutions x ∈ Z of equation (1.1) is bounded above by a function that depends only upon 1(f ) . Let us briefly review what is known so far in this context
1.
For q = 1, equation (1.1) reduces to a
0α
0x= 0 and clearly there is no solution x at all.
For q = 2, we have one of the following two equations (a
0+ a
1x)α
x0= 0 or a
0α
0x+ a
1α
x1= 0.
In either case, in view of our assumption (1.2) on non-degeneracy, we clearly do not have more than one solution x .
The first non-trivial case is q = 3. Here, Schlickewei [4] proved the con- jecture to be true. His bound has been improved by Beukers and Schlickewei [1]. They showed that for q = 3 equation (1.1) does not have more than 61 solutions.
Now suppose q ≥ 4. In a recent paper [3], Evertse, Schlickewei and Schmidt proved the following: Suppose that in (1.1) the polynomials f
ifor i = 0 , . . . , k are constant. Then equation (1.1) does not have more than exp ( 7 k)
3ksolutions. As in this situation q = k + 1, we see that when the polynomials f
iin (1.1) are all constant, the conjecture is true.
There remains the case when q ≥ 4 and when not all f
i’s are constant.
Now obviously in (1.1) we may suppose without loss of generality that α
0= 1. With this normalization, Schlickewei [5] proved the following:
Suppose that α
1, . . . , α
kare algebraic and that [Q(α
1, . . . , α
k) : Q] ≤ d . Then the number of solutions of equation (1.1) is bounded in terms of q and d only. (A bound was given explicitly). Schlickewei and Schmidt [6] later on established the bound ( 2 q)
35q3d
6q2.
1
After the present paper was written, the second author [7] settled this conjecture.
We denote by Q the algebraic closure of Q in K (this is the field of algebraic elements in K ). We define an equivalence relation on the set K
×of non-zero elements of K by the condition
z
1∼ z
2⇐⇒ z
1/z
2is algebraic . This relation induces a partition of
α
0, . . . , α
k: α
0, . . . , α
k=
[
m i=1α
i0, . . . , α
iki,
where, for 1 ≤ i ≤ m ,
α
i0: · · · : α
iki∈ P
ki(Q),
while for 1 ≤ i 6= u ≤ m , the number α
i0/α
u0is transcendental. Accord- ingly, f is decomposed into
f = f
1+ · · · + f
m, (1.3) with
f
i(x) = P
i0(x)α
ix0+ · · · + P
iki(x)α
ikxi(1 ≤ i ≤ m).
We prove
Theorem 1.1. Suppose we have (1.2). Define 1 = 1(f ) and F (1) = exp 1( 5 1)
51.
Then for all but at most F (1) solutions x ∈ Z of (1.1), we have
f
1(x) = · · · = f
m(x) = 0 . ( 1 . 4 ) Our result, in other words, says that the only case when the conjec- ture possibly could fail to be true arises from the algebraic case, i.e. when α
0, . . . , α
kare in Q . Moreover we shall see that the conjecture would follow from the special case where α
0, . . . , α
kare algebraic and each α
i/α
jhas a small height. Actually our method of proof gives a result of the type stated in the Theorem also under the assumption that the quotients α
i/α
uare not transcendental but have logarithmic height bounded away from zero (for more details, see the final remark in Sect. 6).
We mention that our proof was inspired by a similar result for q = 3 by Beukers and Tijdeman [2]. They showed:
Let α and β be non-zero elements of K . Suppose that α , β and α/β are not roots of unity. Let a and b be non-zero elements of K . Suppose that the equation
aα
x+ bβ
x= 1
has at least 4 solutions x ∈ Z . Then α and β are algebraic.
Our proof uses a recent result of Schlickewei and Schmidt [6] on polynomial
exponential equations.
2. Heights
Let K be a number field of degree d . Write M(K) for the set of places of K . For v ∈ M(K) , let | |
vbe the valuation which extends either the standard absolute value of Q , or if v|p for a rational prime p , let | |
vbe the valuation with |p|
v= p
−1. Write d
vfor the local degree [K
v: Q
p] and define the absolute value k k
vby
k k
v= | |
dvv/d.
Let n ≥ 1 and let α = (α
0, . . . , α
n) 6= (0, . . . , 0) be a point in K
n+1. We then put
kαk
v= max
kα
0k
v, . . . , kα
nk
vand we define the homogeneous height as H (α) = Y
v∈M(K)
kαk
v.
Since it depends only on the class α = (α
0: · · · : α
n) of α in P
n(Q), we also denote it by H (α) . Let
h(α) = h(α
0: · · · : α
n) = log H(α)
be the homogeneous logarithmic absolute height of α ∈ P
n(Q) We shall also need the inhomogeneous absolute heights
H
in(x) = H(1 : x
1: · · · : x
n) and
h
in(x) = h(1 : x
1: · · · : x
n) = log H
in(x) of x = (x
1, . . . , x
n) ∈ Q
n. Further, for x ∈ Q, we set
H
in(x) = H ( 1 : x) and h
in(x) = h ( 1 : x) = log H
in(x).
Given D ∈ N and h / > 0, we will use the fact that the set of elements α ∈ Q
×with
deg α ≤ D and h
in(α) ≤ h /
is finite.
3. Algebraic linear recurrence sequences
The results in this section are consequences of the Subspace Theorem.
Lemma 3.1. Let m ≥ 1 and 0 be a finitely generated subgroup of (Q
×)
mof rank r ≥ 0. Then the solutions z = x ∗ y = (x
1y
1, . . . , x
my
m) of z
1+ · · · + z
m= 1 (3.1) with z ∈ 0 , y ∈ Q
mand
h
in(y) ≤ 1
4m
2h
in(x) (3.2)
are contained in the union of at most
exp ( 4 m)
4m(r + 1 ) proper subspaces of Q
m.
Proof. This is a variation on Proposition A of [6]. In that proposition there was a distinction between three kinds of solutions:
i) Solutions where some y
i= 0, i.e., some z
i= 0. These clearly lie in m subspaces.
ii) Solutions where each y
i6= 0 and where h
in(x) > 2m log m. These were called large solutions in [6] and it was shown in (10.4) of that paper that they lie in the union of fewer than
2
30m2(21m
2)
rproper subspaces.
iii) Solutions where each y
i6= 0 and where h
in(x) ≤ 2m log m. These were called small solutions in [6]. Here we argue as follows. We have h
in(y) ≤ (2m log m)/(4m
2) < log 2 by (3.2). Then each component has h
in(y
i) < log 2, which is H
in(y
i) < 2. Since y
i∈ Q
×, we have y
i= ± 1. The equation (3.1) now becomes
±x
1± x
2± · · · ± x
m= 1. (3.3) The group 0
0generated by 0 and the vectors (± 1 , · · · , ± 1 ) contains no more than r multiplicatively independent elements. By Proposition 2.1 of [6], the solutions of (3.3) lie in the union of not more than
exp ( 4 m)
3m· 2 (r + 1 )
proper subspaces of Q
m.
Combining our estimates we obtain m + 2
30m2( 21 m
2)
r+ exp ( 4 m)
3m· 2 (r + 1 )
< exp ( 4 m)
4m(r + 1 ) . ut Corollary. Let q > 1 and let 0 be a finitely generated subgroup of (Q
×)
qof rank r ≥ 0. Then the solutions of
z
1+ · · · + z
q= 0 (3.4) where z = x ∗ y with x ∈ 0, y ∈ Q
qand
h (y) ≤ 1 4q
2h (x) are contained in the union of fewer than
exp ( 4 q)
4q(r + 1 )
( 3 . 5 ) proper subspaces of the space given by (3.4).
Proof. This is just the homogeneous version of Lemma 3.1. We apply Lemma 3.1 with m = q − 1. One needs also to consider the possible solu- tions with z
q= 0. But they lie in one subspace, and 1 is absorbed in (3.5) since q > m. ut
Lemma 3.2. Let α ∈ Q
×be given with h
in(α) > 0. Let a ∈ Q
×. Then there is a u ∈ Z such that
h
in(aα
x−u) ≥ 1
4 h
in(α)|x|
for x ∈ Z.
Proof. This is the case r = n = 1 of Lemma 15.1 in [6]. ut Agreement. We define the degree of the zero polynomial as −1.
Lemma 3.3. Consider an equation
P
0(x)α
0x+ · · · + P
k(x)α
kx= 0 ( 3 . 6 ) where (α
0, . . . , α
k) ∈ (Q
×)
k+1and, for 0 ≤ j ≤ k, P
jis a non-zero polynomial of degree d
j≥ 0 with algebraic coefficients. Write
1 = X
kj=0
(d
j+ 1 ), D = max
0≤j≤k
d
j. Suppose that 1 ≥ 3,
0≤i,j
max
≤kh (α
i: α
j) ≥ h /
where 0 < h / ≤ 1 and set
E = 16 1
2D/ h /, t = exp ( 5 1)
41+ 5 E log E.
Then there are tuples
P
0(`), . . . , P
k(`)6= ( 0 , . . . , 0 ) ( 1 ≤ ` ≤ t)
of polynomials where deg P
j(`)≤ d
j(0 ≤ j < k , 1 ≤ ` ≤ t ) and deg P
k(`)<
d
kfor ` = 1 , . . . , t , such that every solution x ∈ Z of (3.6) satisfies P
0(`)(x)α
0x+ · · · + P
k(`)(x)α
kx= 0 ( 3 . 7 ) for some ` .
Proof. Suppose u ∈ Z and set y = x + u. Then (3.6) may be rewritten as P
0(y − u)α
0−uα
0y+ · · · + P
k(y − u)α
−ukα
yk= 0 ,
which is the same as
P e
0(y)α
0y+ · · · + P e
k(y)α
yk= 0 , ( 3 . 8 ) with P e
j(Y ) = P
j(Y − u)α
−uj( 0 ≤ j ≤ k).
Suppose our assertion is true for (3.8), with polynomials P e
0(`), . . . , P e
k(`)(1 ≤ ` ≤ t). Thus every solution of (3.8) satisfies
P e
0(`)(y)α
0y+ · · · + P e
k(`)(y)α
ky= 0 for some `. But then x = y − u satisfies (3.7) with
P
j(`)(X) = P e
j(`)(X + u)α
ju( 0 ≤ j ≤ k).
We therefore may make a change of variables x 7→ y = x + u.
We may suppose that h (α
0: α
ι) ≥ h / for some ι in the range 1 ≤ ι ≤ k . Write P
j(X) = a
j0+a
j1X +· · ·+ a
j,djX
dj. Pick u according to Lemma 3.2 with h (a
0,d0α
y−u0: a
ι,dια
ιy−u) ≥
14h /|y| . Writing P e
j(Y ) = P
j(Y − u)α
j−u= b
j0+ b
j1Y + · · · + b
j,djY
dj, we have b
0,d0= a
0,d0α
0−u, b
ι,dι= a
ι,dια
ι−u, so that
h(b
0,d0α
0y: b
ι,dια
yι) ≥ 1
4 h /|y|. (3.9)
The equation (3.8) may be written as b
00+ b
01y +· · ·+b
0,d0y
d0α
y0+· · ·+ b
k0+ b
k1y +· · ·+ b
k,dky
dkα
ky= 0 .
Some coefficients may be zero; omitting the zero coefficients, we rewrite this as
b
000y
v00+ · · · + b
0,d0y
d0α
y0+ · · · + b
k00y
vk0+ · · · + b
k,dky
dkα
ky= 0.
Let q be the total number of (non-zero) coefficients here, and consider the following vectors in q -space:
x = b
000α
0y, . . . , b
0,d0α
0y, . . . , b
0k0α
yk, . . . , b
k,dkα
ykw = y
v00, . . . , y
d0, . . . , y
vk0, . . . , y
dk. Our equation becomes
z
1+ · · · + z
q= 0 ( 3 . 10 ) with z = x ∗ w = (x
1w
1, . . . , x
qw
q). Hence x lies in the group 0 of rank r ≤ 2 generated by (α
0, . . . , α
0, . . . , α
k, . . . , α
k) and (b
000, . . . , b
0,d0, . . . , b
0k0, . . . , b
k,dk) . Further
h (x) ≥ h (b
0,d0α
0y: b
ι,dια
ιy) ≥ 1 4 h /|y|
by (3.9). On the other hand, h (w) ≤ D log |y| . Therefore when
|y | ≥ 2E log E, (3.11)
so that |y| ≥ ( 32 q
2D/ h /) log ( 16 q
2D/ h /) by q ≤ 1 , then
|y| > ( 16 q
2D/ h /) log |y|, and
h(w) ≤ D log |y| < h /
16 q
2|y | = 1 4 q
21
4 h /|y| ≤ 1 4 q
2h(x).
By the corollary, for such y , we have z contained in the union of exp (4q)
4q· 3
< exp (51)
41proper subspaces of the space (3.10). Consider such a subspace c
1z
1+
· · · + c
qz
q= 0 (where (c
1, . . . , c
q) is not proportional to (1, . . . , 1)).
Taking a linear combination of this and (3.10) we obtain a non-trivial relation c
10z
1+· · ·+c
q−0 1z
q−1= 0. But this means exactly that y satisfies a non-trivial equation
Q e
0(y)α
0y+ · · · + Q e
k(y)α
ky= 0, (3.12) where deg Q e
j≤ d
j(0 ≤ j < k ) and deg Q e
k< d
k.
There are not more than 5 E log E values of y where (3.11) is violated. For fixed y, and since 1 = P
(d
j+ 1) ≥ 3, there will certainly be polynomials Q e
0, . . . , Q e
kas above (deg Q e
j≤ d
j(0 ≤ j < k ) and deg Q e
k< d
k) with (3.12). ut
Remark. When h / ≥ exp −(51)
41we have t ≤ exp (51)
51.
4. A specialization-type argument
Lemma 4.1. Let K be a number field, D , N , M , L
1, . . . , L
Mnon-negative integers, A
1, . . . , A
Nhomogeneous polynomials in K [X
0, . . . , X
k] , each of degree ≤ D and B
λµ(1 ≤ λ ≤ L
µ, 1 ≤ µ ≤ M) homogeneous polynomials in Q[X
0, . . . , X
k]. Assume that there exists α ∈ P
k(K) such that
(i) A
1(α) = · · · = A
N(α) = 0 and
(ii) for each µ = 1, . . . , M, there exists λ ∈ {1, . . . , L
µ} with B
λµ(α) 6= 0.
Then there exist elements e α
0, . . . , e α
kin K , algebraic over Q , not all of which are zero, which generate an extension K e = K(e α
0, . . . ,e α
k) of K of degree e K : K
≤ D
kand such that the point e α = (e α
0: · · · : e α
k) ∈ P
k( K) e satisfies
(i)
aA
1(e α) = · · · = A
N(e α) = 0 and
(ii)
afor each µ = 1 , . . . , M , there exists λ ∈ { 1 , . . . , L
µ} with B
λµ(e α) 6= 0.
Proof. Given homogeneous polynomials Q
1, . . . , Q
Nin K[X
0, . . . , X
k], we write
Z(Q
1, . . . , Q
N) ⊂ P
k(K)
for the set of zeros in P
k(K) of the ideal (Q
1, . . . , Q
N) in K[X
0, . . . , X
k] generated by Q
1, . . . , Q
N.
Let Y be an absolutely irreducible component of Z(A
1, . . . , A
N) ⊂ P
k(K) containing α. Consider the Zariski closed subset
F =
\
M µ=1Z(B
1µ, . . . , B
Lµ,µ)
of P
k(K). By assumption α is not in F . Hence Lemma 4.1 is a consequence of the following statement:
Let A
1, . . . , A
Nbe homogeneous polynomials in K[X
0, . . . , X
k], each of degree ≤ D . Let Y be an irreducible component of dimension δ of
Z(A
1, . . . , A
N)
and F a Zariski closed subset of P
k(K) such that Y \F is not empty. Then there exists an element e α = (e α
0: · · · : e α
k) in Y \ F whose components e α
0, . . . ,e α
kare algebraic over Q and such that we have
K(e α
0, . . . , e α
k) : K
≤ D
k−δ.
Since Y is absolutely irreducible and not contained in F , we have dim(Y ∩ F ) ≤ δ − 1. Pick linear forms L
1, . . . , L
δwith coefficients in K and in sufficiently general position such that
Z(L
1) ∩ · · · ∩ Z(L
δ) ∩ F ∩ Y = ∅ and such that moreover
Z(L
1) ∩ · · · ∩ Z(L
δ) ∩ Y
is a non-empty finite set which does not contain more than D
k−δpoints. Let γ = (γ
0: · · · : γ
k) be one of its elements. One at least among γ
0, . . . , γ
kis non-zero, say γ
0. Put e α
i= γ
i/γ
0. Then our construction implies that e α = ( 1 : e α
1: · · · : e α
k) = γ lies in Y \ F . Since our linear forms L
ias well as the polynomials A
1, . . . , A
Nhave coefficients in K , it follows that for any K -embedding σ of K(e α
1, . . . ,e α
k) in K we have
( 1 : σe α
1: · · · : σe α
k) ∈ Z(L
1) ∩ · · · ∩ Z(L
δ) ∩ Y.
Since moreover the right hand side has cardinality ≤ D
k−δ, we may conclude that in fact e α
1, . . . ,e α
kare algebraic over K and that
[K(e α
1, . . . , e α
k) : K ] ≤ D
k−δ. ut Here is a consequence of Lemma 4.1.
Lemma 4.2. Let k be a non-negative integer, p , S , T , d
1, . . . , d
Spositive in- tegers and h / a positive real number. For 1 ≤ s ≤ S, let C
s= (C
1s, . . . , C
ps) be a p-tuple of homogeneous polynomials in Q[X
0, . . . , X
k], each of de- gree d
s. For 1 ≤ t ≤ T , let D
t= (D
1t, . . . , D
pt) be a p -tuple of homo- geneous polynomials in Q[X
0, . . . , X
k] , with deg D
1t= · · · = deg D
pt. Let α
0, . . . , α
kbe non-zero elements of K and α = (α
0, . . . , α
k) ∈ K
k+1. Denote by V the subspace of K
pspanned by C
1(α), . . . , C
S(α) . Assume that for each t = 1, . . . , T , we have D
t(α) 6∈ V .
Then there exist non-zero algebraic elements e α
0, . . . ,e α
kin K such that e α = (e α
0, . . . , e α
k) ∈ Q
k+1has the following properties. The subspace V e of K
pspanned by C
1(e α), . . . , C
S(e α) has dim V e = dim V . Further, for each t = 1, . . . , T , we have D
t(e α) 6∈ V e . Furthermore, for 0 ≤ i, j ≤ k , we have
( e α
i/e α
j= α
i/α
jif α
i/α
jis algebraic,
h (e α
i: e α
j) ≥ h / if α
i/α
jis transcendental .
Proof. Let K denote a number field containing all coefficients of C
is(1 ≤ i ≤ p , 1 ≤ s ≤ S ) and all algebraic elements of K which belong to the set α
i/α
j; 0 ≤ i, j ≤ k . We shall prove the existence of e α = (e α
0, . . . , e α
k) ∈ K
k+1satisfying the desired properties together with an upper bound for the degree of the number field K e = K(e α
0, . . . ,e α
k) , namely
e K : K
≤ D
kwith D = p max
1≤s≤S
d
s.
Define r = dim V . Since D
t(α) is not in V , we have V 6= K
p, hence r < p . Denote by
A
1, . . . , A
Jthe set of (r + 1 ) × (r + 1 ) minors of the p × S
matrix
C
1, . . . , C
S.
Each of these polynomials A
1, . . . , A
Jis homogeneous of degree
≤ (r + 1 ) max
1≤s≤S
d
s≤ D.
Also, for 1 ≤ t ≤ T , denote by
B
1t, . . . , B
Ltthe set of (r + 1 )×(r + 1 ) minors of the p × (S + 1) matrix
C
1, . . . , C
S, D
t. Further, let
A
J+1, . . . , A
Ndenote the set of polynomials α
iX
j− α
jX
iwhere (i, j) runs over the set of pairs with 0 ≤ i, j ≤ k for which α
i/α
jis algebraic. Furthermore, denote by
B
T+1, . . . , B
Mthe set of polynomials X
0, . . . , X
k, and βX
i− X
j, where (i, j) runs over the set of pairs with 0 ≤ i, j ≤ k for which α
i/α
jis transcendental, while β runs over the (finite) set of algebraic elements of K for which
[K(β) : K] ≤ D
kand h
in(β) ≤ h /.
By assumption the point α ∈ K
k+1satisfies
A
1(α) = · · · = A
N(α) = 0 , B
µ(α) 6= 0 for T + 1 ≤ µ ≤ M,
and for each µ = 1, . . . , T , there exists λ ∈ {1, . . . , L} such that B
λµ(α) 6=
0.
From Lemma 4.1 we deduce that there exists e α ∈ Q
k+1such that K(e α
0, . . . ,e α
k) : K
≤ D
k, A
1(e α) = · · · = A
N(e α) = 0 , B
µ(e α) 6= 0 for T + 1 ≤ µ ≤ M,
and for each µ = 1, . . . , T , there exists λ ∈ {1, . . . , L} such that B
λµ(e α) 6=
0. This e α then satisfies all desired properties. ut
We apply Lemma 4.2 to exponential polynomials.
Lemma 4.3. Let k ≥ 1 be an integer, h / a positive real number, d
0, . . . , d
knon-negative integers and α
0, . . . , α
knon-zero elements of K satisfying (1.2). For 0 ≤ j ≤ k , let
P
j(X) =
dj
X
i=0
a
ijX
ibe a non-zero polynomial in K[X] of degree d
j. Define
f (x) = X
k j=0P
j(x)α
xjand denote by N the set of solutions x ∈ Z of the equation f (x) = 0. Let E be a finite subset of Z . Assume that for each x ∈ E we are given a subset I (x) of
(i, j) ; 0 ≤ i ≤ d
j, 0 ≤ j ≤ k for which X
(i,j)∈I (x)
a
ijx
iα
jx6= 0. (4.1)
Then there exist non-zero algebraic elementse α
0, . . . , e α
kof K and there exist polynomials P e
0, . . . , P e
kwhich are not all zero,
P e
j(X) =
dj
X
i=0
e a
ijX
i(0 ≤ j ≤ k),
with algebraic coefficients e a
ij, and with the following properties:
deg P e
j≤ d
j( 0 ≤ j ≤ k) (4.2) X
kj=0
P e
j(x)e α
xj= 0 for all x ∈ N , (4.3) X
(i,j)∈I (x)
e a
ijx
ie α
jx6= 0 for each x ∈ E, (4.4)
and, for 0 ≤ i, j ≤ k,
( e α
i/e α
j= α
i/α
jif α
i/α
jis algebraic,
h (e α
i: e α
j) ≥ h / if α
i/α
jis transcendental . (4.5)
Proof. We fix an ordering of the set I =
(i, j) ; 0 ≤ i ≤ d
j, 0 ≤ j ≤ k, a
ij6= 0 and we denote by p the number of elements in this set. Also we write N = {n
1, . . . , n
S} (recall that N is finite) and E = {x
1, . . . , x
T} . For 1 ≤ s ≤ S, we define C
sas the p-tuple composed of the polynomials n
isX
jnsfor (i, j) ∈ I . For 1 ≤ t ≤ T , let D
tbe the p -tuple composed of the
polynomials (
x
tiX
xjtfor (i, j) ∈ I ∩ I (x
t) 0 for (i, j) ∈ I \ I (x
t).
From the definition of N we deduce that the dimension r of the vector space V spanned by C
1(α), . . . , C
S(α) satisfies r < p. According to (4.1), for each t = 1 , . . . , T we have D
t(α) 6∈ V . Therefore Lemma 4.3 follows from Lemma 4.2. ut
Remark. Let K denote the field generated over Q by all algebraic elements which belong to the set
α
i/α
j; 0 ≤ i, j ≤ k . The proof of Lemma 4.3 also yields an upper bound for the degree of the number field K e = K(e α
0, . . . , e α
k) , namely
e K : K
≤ 1 max
x∈N
|x|
kwith 1 = d
1+· · ·+d
k+k+1. One may prove a variant of Lemma 4.3 where (4.3) holds only for some subset N
0of N with Card N
0/ Card N ≥ 1 /(k+ 1 ) but with the estimate
e K : K
≤ 1 min
x∈N0
|x|
k.
5. Dividing exponential polynomials
Let α
0, . . . , α
kbe given non-zero elements of K satisfying (1.2) and P
0, . . . , P
kbe polynomials with coefficients in K , possibly zero. Consider the exponential polynomial
f (x) = X
k j=0P
j(x)α
jx.
We set
1(f ) = X
kj=0 Pj6=0
( deg P
j+ 1 ).
Thus 1(f ) = 0 precisely when P
0= · · · = P
k= 0. When g(x) =
X
k j=0Q
j(x)α
jxis another exponential polynomial with the same frequencies (α
0, · · · , α
k), we write g ≺ f if deg Q
j≤ deg P
jfor 0 ≤ j ≤ k . We write g f if g ≺ f and 1(g) < 1(f ).
Lemma 5.1. Suppose g ≺ f and g 6= 0. Then there is an exponential polynomial
r(x) = R
0(x)α
0x+ · · · + R
k(x)α
kxwith r f such that
f (x) = r(x) + cx
ng(x) for some c in K
×and some n ≥ 0.
Proof. With f and g written as above, set n = min
0≤j≤k Qj6=0
( deg P
j− deg Q
j).
We may suppose n = deg P
0− deg Q
0. When
P
0= c
aX
a+ c
a−1X
a−1+ · · · , Q
0= d
bX
b+ d
b−1X
b−1+ · · · , where now a = b + n , set c = c
a/d
band
r(x) = f (x) − cx
ng(x).
If again r(x) = R
0(x)α
0x+ · · · + R
k(x)α
xk, we have R
0(X) = P
0(X) − (c
a/d
b)x
nQ
0(X),
so that deg R
0< deg P
0. Also deg R
j≤ max ( deg P
j, n+ deg Q
j) ≤ deg P
j, so that r f . ut
Consider an exponential polynomial f (x) =
X
k j=0P
j(x)α
xjwhere α
0, . . . , α
kare non-zero algebraic elements in K satisfying (1.2).
Assume
α
0, . . . , α
k= [
m i=1α
i0: · · · : α
ikiis a partition of
α
0, . . . , α
kand define
f
i(x) = P
i0(x)α
ix0+ · · · + P
iki(x)α
ikxi( 1 ≤ i ≤ m) so that
f (x) = f
1(x) + · · · + f
m(x).
Suppose further, for 1 ≤ i 6= u ≤ m, 0 ≤ j ≤ k
iand 0 ≤ v ≤ k
u,
h
in(α
ij/α
uv) ≥ 1 . ( 5 . 1 ) From (1.2) we deduce
1(f ) = 1(f
1) + · · · + 1(f
m).
Set
1 = 1(f ) Lemma 5.2. Define
F (1) = exp 1( 5 1)
51.
Then for all but at most F (1) solutions x ∈ Z of f (x) = 0, we have f
1(x) = · · · = f
m(x) = 0 . ( 5 . 2 ) Proof. The lemma is non-trivial only when m ≥ 2 and at least two of f
1, . . . , f
mare non-zero, so that 1 ≥ 2. We now proceed by induction on 1 . When 1 = 2 and m ≥ 2, we have in fact f (x) = aα
10x+ bα
x20with ab 6= 0 and h
in(α
10/α
20) ≥ 1, so that α
10/α
20is not a root of 1.
There can be at most one zero x of f , for if f (x) = f (y) = 0, then (α
10/α
20)
x= (α
10/α
20)
y= −b/a, so that (α
10/α
20)
x−y= 1 hence x = y since α
10/α
20is not a root of 1.
Now assume 1 ≥ 3. In the induction step we apply Lemma 3.3 with h / = 1. The condition max
0≤i,j≤kh (α
i: α
j) ≥ 1 is satisfied because m ≥ 2.
Any x ∈ Z with f (x) = 0 satisfies a relation f
(`)(x) = 0
for some ` in the range 1 ≤ ` ≤ t where t = exp ( 5 1)
51and each f
(`)6= 0 has f
(`)f . By Lemma 5.1 we have, for 1 ≤ ` ≤ t
f (x) = r
(`)(x) + c
(`)x
n(`)f
(`)(x) with r
(`)f . Write out
f
(`)(x) = f
1(`)(x) + · · · + f
m(`)(x), r
(`)(x) = r
1(`)(x) + · · · + r
m(`)(x) with
f
i(`)(x) = P
i(`)0(x)α
ix0+ · · · + P
ik(`)i(x)α
xiki, r
i(`)(x) = R
i(`)0(x)α
ix0+ · · · + R
(`)iki(x)α
ikxiand
f
i(x) = r
i(`)(x) + c
(`)x
n(`)f
i(`)(x). (5.3)
By induction, and since f
(`)f and r
(`)f , hence 1(f
(`)) < 1(f ), 1(r
(`)) < 1(f ) , we see that all but at most F (1− 1 ) solutions of f
(`)(x) = 0 have
f
1(`)(x) = · · · = f
m(`)(x) = 0 , ( 5 . 4 ) and similarly all but at most F (1 − 1 ) solutions of r
(`)(x) = 0 have
r
1(`)(x) = · · · = r
m(`)(x) = 0 . ( 5 . 5 ) But (5.3), (5.4) and (5.5) imply (5.2). Taking the sum over ` in 1 ≤ ` ≤ t , we see that all but at most
2 tF (1 − 1 ) ≤ exp 1 + ( 5 1)
51+ (1 − 1 )( 5 1)
51−5≤ F (1) solutions of f (x) = 0 have (5.2). ut
6. Proof of Theorem 1.1
Assume that the assumptions of Theorem 1.1 are satisfied. Let E be a set of more than F (1) solutions of (1.1). Assume that for each x in E there is an index i = i(x) in the range 1 ≤ i ≤ m such that f
i(x)(x) 6= 0.
We apply Lemma 4.3 with h / = 1. We produce algebraic elements e α
0, . . . ,e α
kand polynomials with algebraic coefficients P e
0, . . . , P e
ksatis- fying (4.2), (4.3), (4.4) and (4.5). The exponential polynomial
f (x) e = X
k j=0P e
j(x)e α
jxcan be written
f (x) e = f e
1(x) + · · · + f e
m(x) where, for 1 ≤ i ≤ m ,
f e
i(x) =
ki
X
j=0