On the rational approximations to the powers of an algebraic number

Solution of two problems of Mahler and Mendès France

Pietro Corvaja Umberto Zannier

Abstract. About fifty years ago Mahler [M] proved that if α > 1 is rational but not an integer and
if 0 < l < 1, then the fractional part of αn is > ln apart from a finite set of integers n depending on
arXiv:math/0403522v1 [math.NT] 30 Mar 2004

α and l. Answering completely a question of Mahler, we show (Thm. 1) that the same conclusion
holds for all algebraic numbers which are not d-th roots of Pisot numbers. By related methods
we also answer a question of Mendès France, characterizing completely the quadratic irrationals α
such that the continued fraction for αn has period length tending to infinity (Thm. 2).

§1 Introduction. About fifty years ago Mahler [M] proved that if α > 1 is rational but not an
integer and if 0 < l < 1, then the fractional part of αn is > ln apart from a finite set of integers n
depending on α and l. His proof used a p-adic version of Roth’s Theorem, as in previous work by
Mahler and especially by Ridout.
At the end of that paper Mahler
√ pointed out that the conclusion does not hold if α is a suitable
algebraic number, as e.g. 21 (1 + 5); of course a counter-example is provided by any Pisot number,
i.e. a real algebraic integer α > 1 all of whose conjugates different from α have absolute value < 1
(note that rational integers > 1 are Pisot numbers according to our definition). Mahler also added
that “It would be of some interest to know which algebraic numbers have the same property as...”
the rationals in his theorem.
Now, it seems that even replacing Ridout Theorem with the modern versions of Roth’s The-
orem, valid for several valuations and approximations in any given number field, the method of
Mahler does not lead to a complete solution to his question.
One of the objects of the present paper is to answer completely Mahler’s question; our methods
will involve a suitable version of the Schmidt Subspace Theorem, which may be considered as a
multi-dimensional extension of the mentioned results by Roth, Mahler and Ridout. We state at
once our first theorem, where as usual we denote by ||x|| the distance of the real number x from
the nearest integer.
Theorem 1. Let α > 1 be a real algebraic number and let 0 < l < 1. Suppose that ||αn || < ln for
infinitely many natural numbers n. Then some power αn is a Pisot number. In particular, α is an
algebraic integer.
We remark that the conclusion is best-possible, since, if it is true, then conversely we have
kαn k ≪ ln for all n in a suitable arithmetic progression, where l is the maximum absolute value
of the conjugates of α different from α. Here, Mahler’s example with the Golden ratio is typical.
Also, the conclusion is not generally true without the assumption that α is algebraic; see for
this the Appendix.
The present application of the Subspace Theorem seems different from previous ones, and
occurs in Lemma 3 below. Related methods actually enable us to answer as well a question raised
by Mendès France about the length of the periods of the continued fractions for αn , where α is
now a quadratic irrational; this appears as Problem 6 in [MF]. We shall prove, more generally:
Theorem 2. Let α > 0 be a real quadratic irrational. If α is neither the square root of a rational
number, nor a unit in the ring of integers of Q(α) then the period of the continued fraction for
αn tends to infinity with n. If α is the square root of a rational number the period length of the
continued fraction for α2n+1 tends to infinity. If α is a unit, the period of the continued fraction
for αn is bounded.
Clearly, if α is the square root of a rational number, then the continued fraction for α2n is
finite(1) , so Theorem 2 gives a complete answer to the problem of Mendès France.
(1)
its length tends to infinity by a result of Pourchet, proved in greater generality in [CZ]

1
 The main tool in the proof of both theorems is the following new lower bound for the fractional
parts of S-units in algebraic number fields. We first need a definition:
Definition. We call a (real) algebraic number α a pseudo-Pisot number if:
(i) |α| > 1 and all its conjugates have (complex) absolute value strictly less then 1;
(ii) α has an integral trace: TrQ(α)/Q (α) ∈ Z.
Of course, pseudo-Pisot numbers are “well approximated” by their trace, hence are good
candidates for having a small fractional part compared to their height. The algebraic integers
among the pseudo-Pisot numbers are just the usual Pisot numbers. We shall prove
×
Main Theorem. Let Γ ⊂ Q be a finitely generated multiplicative group of algebraic numbers, let
×
δ ∈ Q be a non zero algebraic number and let ǫ > 0 be fixed. Then there are only finitely many
pairs (q, u) ∈ Z × Γ with d = [Q(u) : Q] such that |δqu| > 1, δqu is not a pseudo-Pisot number and

0 < kδquk < H(u)−ǫ q −d−ǫ . (1.1)

Note again that, conversely, starting with a Pisot number α and taking q = 1, u = αn for
n = 1, 2, . . . produces an infinite sequence of solutions to 0 < kquk < H(u)−ǫ for a suitable ǫ > 0.

The above Main Theorem can be viewed as a Thue-Roth inequality with “moving target”, as
the Theorem in [CZ], where we considered quotients of power sums with integral roots instead of
elements of a finitely generated multiplicative group. The main application of the Theorem in [CZ]
also concerned continued fractions, as for our Theorem 2.

§2. Proofs. We shall use the following notations: let K be a number field, embedded in C, Galois
over Q. We denote by MK (resp. M∞ ) be the set of places (resp. archimedean places) of K; for
each place v we denote by | · |v the absolute value corresponding to v, normalized with respect to
K; by this we mean that if v ∈ M∞ then there exists an automorphism σ ∈ Gal(K/Q) of K such
that, for all x ∈ K
d(σ)
|x|v = |σ(x)| [K:Q]
where d(σ) = 1 if σ(K) = K ⊂ R and 2 otherwise (note that d(σ) is now constant since K/Q is
Galois). Non-archimedean absolute values are normalized accordingly, so that the product formula
holds and the absolute Weil height reads
Y
H(x) = max{1, |x|v }.
v∈Mk

For a vector x = (x1 , . . . , xn ) ∈ K n and a place v ∈ MK we shall denote by kxkv the v-norm of x:

kxkv := max{|x1 |v . . . , |xn |v },

and by H(x) the projective height

Y
H(x) = max{|x1 |v , . . . , |xn |v }.
v∈MK

We begin by proving the Main Theorem. First of all notice that, by enlarging if necessary the
multiplicative group Γ, we can reduce to the following situation: Γ ⊂ K × is the group of S-units:

Γ = OS× = {u ∈ K : |u|v = 1 for all v 6∈ S}

2
of a suitable number field K, Galois over Q, with respect to a suitable finite set of places S,
containing the archimedean ones and stable under Galois conjugation.
Our first lemma can be easily deduced from a theorem of Evertse, which in turn was obtained
as an application of the already mentioned Subspace Theorem; we prefer to give a proof for
completeness.
Lemma 1. Let K, S be as before, σ1 , . . . , σn distinct automorphisms of K, λ1 , . . . , λn non zero
elements of K, ǫ > 0 be a positive real number, w ∈ S be a distinguished place. Let Ξ ⊂ OS× be an
infinite set of solutions u ∈ OS× of the inequality

|λ1 σ1 (u) + . . . + λn σn (u)|w < max{|σ1 (u)|w , . . . , |σn (u)|w } · H(u)−ǫ .

Then there exists a non trivial linear relation of the form

a1 σ1 (u) + . . . + an σn (u) = 0, ai ∈ K,

satisfied by infinitely many elements of Ξ.

Proof. Let Ξ be as in the Lemma. Going to an infinite subset of Ξ we may assume that |σ1 (u)|w =
max{|σ1 (u)|w , . . . , |σn (u)|w } for all the involved u’s. Let us consider, for each v ∈ S, n linear forms
Lv,1 , . . . , Lv,n in n variables x = (x1 , . . . , xn ) as follows: put

Lw,1 (x) := λ1 x1 + . . . + λn xn

while for (v, i) ∈ S × {1, . . . , n}, with (v, i) 6= (w, 1), put Lv,i (x) = xi . Note that the linear
d
forms Lv,1 , . . . , Lv,n are indeed linearly independent. Now put x = (σ1 (u), . . . , σd (u)) ∈ OS× and
consider the double product
n
YY |Lv,i (x)|v
.
i=1
kxkv
v∈S

By multiplying and dividing by |x1 |w = |σ1 (u)|w , and using the fact that the coordinates of x are
S-units, we obtain the equality
n
YY |Lv,i (x)|v
= |λ1 σ1 (u) + . . . + λn σn (u)|w · |σ1 (u)|−1
w · H(x)
−n
.
kxkv
v∈S i=1

Since u ∈ Ξ, we have

|λ1 σ1 (u) + . . . + λn σn (u)|w · |σ1 (u)|−1 −ǫ

w < H(u) ,

then
n
YY |Lv,i (x)|v
< H(x)−n H(u)−ǫ .
i=1
kxk v
v∈S

The height of the point x = (σ1 (u), . . . , σd (u)) is easily compared with the height of u by the
estimate
H(x) ≤ H(u)[K:Q]
hence the above upper bound for the double product also gives
n
YY |Lv,i (x)|v
< H(x)−n−ǫ/[K:Q] .
kxkv
v∈S i=1

3
Now, an application of the Subspace Theorem in the form given in e.g. [S, Theorem 1D′ ] gives the
desired result.
Our next tool is a very special case of the so called unit-equation-theorem, proved by Evertse
and van der Poorten-Schlickewei. It also rests on the Subspace Theorem:
Lemma 2. Let K, S, σ1 , . . . , σn be as before, ǫ > 0 a real number a1 , . . . , an be non zero elements
of K. Suppose that Ξ ⊂ OS× is an infinite set of solutions to the equation

a1 σ1 (u) + . . . + an σd (u) = 0.

Then there exist two distinct indices i 6= j, two non zero elements a, b ∈ K × and an infinite subset
Ξ̄ ⊂ Ξ of Ξ such that for all u ∈ Ξ̄

aσi (u) + bσj (u) = 0.

For a proof of this lemma, see [S, Chap. 4].

Our last lemma is the key of the proof of the Main Theorem; its proof is once more based on
the Subspace Theorem.
Lemma 3. Let K, S be as before, k ⊂ K ∩ R a (real) subfield of K, of degree d over Q, δ ∈ K ×
a nonzero element of K. Let ǫ > 0 be given. Suppose we have an infinite sequence Σ of points
(q, u) ∈ Z × (OS× ∩ k) such that |δqu| > 1, δqu is not a pseudo-Pisot number and

0 < kδquk < H(u)−ǫ q −d−ǫ . (2.1)

Then there exists a proper subfield k ′ ⊂ k, an element δ ′ ∈ k × and an infinite subsequence Σ′ ⊂ Σ

such that for all (q, u) ∈ Σ′ , u/δ ′ ∈ k ′ .
Note that Lemma 3 gives a finiteness result in the case k = Q, since the rational field Q
admits no proper subfields.
Proof of Lemma 3. Let us suppose that the hypotheses of the Lemma are satisfied, so in particular
Σ is an infinite sequence of solutions of (2.1). We begin by observing that, by Roth’s theorem [S,
Theorem 2A], in any such infinite sequence u cannot be fixed; so we have H(u) → ∞ in the set
Σ. Let H := Gal(K/k) ⊂ Gal(K/Q) be the subgroup fixing k (note that H contains the complex
conjugation). Let {σ1 , . . . , σd } (d = [k : Q]) be a (complete) set of representatives for the left
cosets of H in Gal(K/Q), containing the identity σ1 . Each automorphism ρ ∈ Gal(K/Q) defines
an archimedean valuation on K by the formula
d(ρ)
|x|ρ := |ρ−1 (x)| [K:Q] (2.2)

where as usual | · | denotes the usual complex absolute value. Two distinct automorphisms ρ1 , ρ2
define the same valuations if and only if ρ−1
1 ◦ ρ2 is the complex conjugation. Note that in this
case ρ1 , ρ2 coincide on k. Let now (q, u) ∈ Σ be a solution of (2.1) and let p ∈ Z be the nearest
integer to δqu. Then for each ρ ∈ Gal(K/Q) we have, with the notation of (2.2),
d(ρ) d(ρ)
kδquk [K:Q] = |δqu − p| [K:Q] = |ρ(δ)ρ(qu) − p|ρ . (2.3)

Let, for each v ∈ M∞ , ρv be an automorphism defining the valuation v according to the rule (2.2):
|x|v := |x|ρv ; then the set {ρv | v ∈ M∞ } represents the left cosets of the subgroup generated by
the complex conjugation in Gal(K/Q). Let Si , for i = 1, . . . , d, be the subset of M∞ formed by
those valuations v such that ρv coincides with σi on k; note that S1 ∪ . . . ∪ Sd = M∞ . We take the

4
product of the terms in (2.3) for ρ running over the set {ρv | v ∈ M∞ }: this corresponds to taking
the product over all archimedean valuations. Then we obtain

Y d Y
Y
|ρv (δ)ρv (qu) − p|v = |ρv (δ)σi (qu) − p|v .
v∈M∞ i=1 v∈Si
P
By (2.3) and the well-known formula v∈M∞ d(ρv ) = [K : Q] it follows that

d Y
Y
|ρv (δ)σi (qu) − p|v = kδquk. (2.4)
i=1 v∈Si

Now, let us define for each v ∈ S a set of d + 1 linearly independent linear forms in d + 1 variables
(x0 , x1 , . . . , xd ) in the following way: for an archimedean valuation v ∈ Si (i = 1, . . . , d) put

Lv,0 (x0 , x1 , . . . , xd ) = x0 − ρv (δ)xi

and for all v ∈ S \ M∞ or 0 < j ≤ d put

Lv,j (x0 , x1 , . . . , xd ) = xj .

Plainly the forms Lv,0 , . . . , Lv,d are independent for each v ∈ S. Finally, let x ∈ K d+1 be the point

x = (p, qσ1 (u), . . . , qσd (u)) ∈ K d+1 .

Let us estimate the double product

d
YY |Lv,j (x)|v
. (2.5)
j=0
kxkv
v∈S

Using the fact that Lv,j (x) = q · σj (u) for j ≥ 1 and that the σj (u) are S-units, we obtain, from
the product formula,
YY d Y Y d
|Lv,j (x)|v ≤ |q|v = |q|d . (2.6)
v∈S j=1 v∈M∞ j=1

Since the coordinates of x are S-integers, the product of the denominators in (2.5) is H(x)d+1 ;
then in view of (2.4) and (2.6)
d
YY |Lv,j (x)|v
≤ H(x)−d−1 · |q|d · kδquk ≤ H(x)−d−1 (qH(u))−ǫ ,
j=0
kxk v
v∈S

the last inequality being justified by the fact that (q, u) ∈ Σ, hence (2.1) holds. Since H(x) ≤
|q| · |p| · H(u)d and |p| ≤ |δqu| + 1 ≤ |q| · |δ| · H(u)d + 1 ≤ |q|H(u)2d for all but finitely many pairs
(q, u) ∈ Σ (recall that H(u) → ∞ for (q, u) ∈ Σ), we have |q|2 H(u)3d ≥ H(x) so the last displayed
inequality gives
d
YY |Lj,v (x)|v
≤ H(x)−d−1−(ǫ/3d) .
j=0
kxk v
v∈S

An application of the Subspace Theorem in the form given e.g. in [S, Theorem 1D′ ] implies the
existence of a hyperplane containing infinitely many points x = (p, qσ1 (u), . . . , qσd (u)). We then
obtain a non trivial linear relation of the form

a0 p + a1 qσ1 (u) + . . . + ad qσd (u) = 0, ai ∈ K, (2.7)

5
satisfied by infinitely many pairs (q, u) ∈ Σ. Our next goal is to prove the following:
Claim: there exists one such non-trivial relation with vanishing coefficient a0 , i.e. one in-
volving only the conjugates of u.
We prove the Claim. Rewriting the above linear relation, if a0 6= 0, we obtain
a1 ad
p=− qσ1 (u) − . . . − qσd (u). (2.8)
a0 a0

Suppose first that for some index j ∈ {2, . . . , d}, σj (a1 /a0 ) 6= aj /a0 ; then by applying the auto-
morphism σj to both sides of (2.8) and subtracting term-by-term from (2.8) to eliminate p, we
obtain a linear relation involving only the terms σ1 (u), . . . , σd (u). Such a relation is non trivial
since the coefficient of σj (u) becomes σj (a1 /a0 ) − aj /a0 . Hence we have proved our Claim in this
case.
Therefore, we may and shall assume that aj /a0 = σj (a1 /a0 ) for all j, so in particular all
coefficients aj /a0 are nonzero. Let us then rewrite (2.8) in a simpler form as

p = q(σ1 (λ)σ1 (u) + . . . + σd (λ)σd (u)) (2.9)

with λ = −(a1 /a0 ) 6= 0. Suppose now that λ does not belong to k. Then there exists an
automorphism τ ∈ H with τ (λ) 6= λ. (Recall that H is the subgroup of Gal(K/Q) fixing k
and that {σ1 = id, σ2 , . . . , σd } is a complete set of representatives of left cosets of H). By applying
the automorphism τ to both sides of (2.9) and subtracting term-by-term from (2.9) to eliminate
p, we obtain the linear relation

(λ − τ (λ))σ1 (u) + (σ2 (λ)σ2 (u) − τ ◦ σ2 (λ)τ ◦ σ2 (u)) + . . . + (σd (λ)σd (u) − τ ◦ σd (λ)τ ◦ σd (u)) = 0.

Note that τ ◦ σj coincides on k with some σi . Note also that since τ ∈ H and σ2 , . . . , σd 6∈ H,
no τ ◦ σj with j ≥ 2 can belong to H. Hence the above relation can be again written as a linear
combination of the σi (u); in such expression the coefficient of σ1 (u) will remain λ − τ (λ) and will
therefore be 6= 0, so we obtain a non trivial relation among the σi (u), as claimed.
Therefore we may and shall suppose that λ ∈ k and write (2.9) in the simpler form

p = q · Trk/Q (λu) = Trk/Q (qλu). (2.10)

After adding −δqu to both sides in (2.9), recalling that (q, u) is a solution to (2.1) and that σ1 is
the identity, we obtain

|p − δqu| = |(λ − δ)qσ1 (u) + qσ2 (λ)σ2 (u) + . . . + qσd (λ)σd (u)| < q −d−ǫ H(u)−ǫ ≤ q −1 H(u)−ǫ .

In particular,
|(λ − δ)u + σ2 (λ)σ2 (u) + . . . + σd (λ)σd (u)| < q −1 H(u)−ǫ . (2.11)
We want to apply Lemma 1. We distinguish two cases:
First case: λ = δ (in particular δ ∈ k).
In this case the algebraic number qδu = qλu has an integral trace. Since by assumption it
is not a pseudo-Pisot number, the maximum modulus of its conjugates |σ2 (qλu)|, . . . , |σd (qλu)| is
≥ 1. This yields

max{|σ2 (u)|, . . . , |σd (u)|} ≥ q −1 max{|σ2 (λ)|, . . . , |σd (λ)|}−1 .

Hence from (2.11) we deduce that infinitely many pairs (q, u) ∈ Σ satisfy the inequality of Lemma
1, where w is the archimedean place associated to the given embedding K ֒→ C, n = d − 1,
λi = σi+1 (λ) and with σ2 , . . . , σd instead of σ1 , . . . , σn . The conclusion of Lemma 1 provides what
we claimed.

6
 Second case. λ 6= δ.
In this case the first term does appear in (2.11). Since we supposed |qδu| > 1, we have

max{|σ1 (u)|, . . . , |σd (u)|} ≥ |u| > |δ|−1 · q −1

so we can again apply Lemma 1 (with the same place w as in the first case, n = d, λ1 = (λ−δ), λ2 =
σ2 (λ), . . . , λd = σd (λ)) and conclude as in the first case.
This finishes the proof of the Claim, i.e. a relation of the kind (2.7) without the term a0 p is
satisfied for all pairs (q, u) in an infinite set Σ̄ ⊂ Σ.
We can now apply the unit theorem of Evertse and van der Poorten-Schlickewei in the form of
Lemma 2 which implies that a non trivial relation of the form aσj (u) + bσi (u) = 0 for some i 6= j
and a, b ∈ K × is satisfied for (q, u) in an infinite subset Σ′ of Σ̄. We rewrite it as
a
−σi−1 · (σi−1 ◦ σj )(u) = u.
b
Then for any two solutions (q ′ , u′ ), (q ′′ , u′′ ) ∈ Σ′ , the element u := u′ /u′′ ∈ k is fixed by the
−1
automorphism σi−1 ◦ σj 6∈ H, thus u belongs to the proper subfield k ′ := K <H,σi ◦σj > of k. In
other words, if we let (q ′ , δ ′ ) be any solution, then infinitely many solutions are of the form (q ′′ , v)
for some v of the form v = u · δ ′ with u ∈ k ′ as wanted.
Proof of the Main Theorem. As we have already remarked, we can suppose that the finitely
generated group Γ ⊂ Q̄× is the group of S-units in a number field K, where K is Galois over Q
containing δ and S is stable under Galois conjugation as in lemmas 1,2,3.
Suppose by contradiction that we have an infinite set Σ ⊂ Z × OS× of solutions (q, u) to the
inequality (2.1). Let us define by induction a sequence {δi } ⊂ K, an infinite decreasing chain Σi of
infinite subsets of Σ and an infinite strictly decreasing chain ki of subfields of K with the following
properties:
For each natural number n ≥ 0, Σn ⊂ (Z × kn ) ∩ Σn−1 , kn ⊂ kn−1 , kn 6= kn−1 , and all but
finitely many pairs (q, u) ∈ Σn satisfy the inequalities |δ0 · · · δn · qu| > 1 and

kδ0 · · · δn · quk < q −d−ǫ H(u)−ǫ/(n+1) . (2.12)

We shall eventually deduce a contradiction from the fact that the number field K does not
admit any infinite decreasing chain of subfields. We proceed as follows: put δ0 = δ, k0 = K ∩ R
and Σ0 = Σ. Suppose we have defined δn , kn , Σn for a natural number n. Applying Lemma 3 with
k = kn , δ = δ0 · · · δn we obtain that there exists an element δn+1 ∈ kn , a proper subfield kn+1 of
kn and an infinite set Σn+1 ⊂ Σn such that all pairs (q, u) ∈ Σn+1 verify u = δn+1 v with v ∈ kn+1 .
Now, since for v ∈ K, H(δn+1 v) ≥ H(δn+1 )−1 H(v), we have in particular that for almost all
v ∈ K, H(δn+1 v) ≥ H(v)(n+1)/(n+2) ; then all but finitely many such pairs satisfy

kδ0 · · · δn δn+1 · qvk < q −1−ǫ H(v)−ǫ/(n+2) ,

and the inductive hypothesis is fullfilled.

The contradiction is then obtained as noticed above, concluding the proof of Main Theorem.

To prove Theorem 1 we also need the following

Lemma 4. Let α be an algebraic number. Suppose that for all n in an infinite set Ξ ⊂ N, there
exists a positive integer qn ∈ Z such that the sequence Ξ ∋ n 7→ qn satisfies

log qn
lim =0 and TrQ(α)/Q (qn αn ) ∈ Z \ {0}
n→∞ n

7
(the limit being taken for n ∈ Ξ). Then α is either the h-th root of a rational number (for some
positive integer h) or an algebraic integer.
Proof. It is essentially an application of Lemma 1, so it still depends on the Subspace Theorem. Let
us suppose that α is not an algebraic integer. Let K be the Galois closure of the extension Q(α)/Q
and let h be the order of the torsion group K × . Since Ξ is an infinite subset of N, there exists
an integer r ∈ {0, . . . , h − 1} such that infinitely many elements of Ξ are of the form n = r + hm.
Let us denote by Ξ′ the infinite subset of N composed of those integers m such that r + hm ∈ Ξ.
Let σ1 , . . . , σd ∈ Gal(K/Q), where d = [Q(αh ) : Q], be a set of automorphisms of K giving all the
embedding Q(αh ) ֒→ K. In other terms, if H ⊂ Gal(K/Q) is the subgroup fixing Q(αh ), the set
{σ1 , . . . , σd } is a complete set of representatives of left cosets of H in Gal(K/Q). This proves that
if d = 1 then α is a h-th root of a rational number. Suppose the contrary and suppose morevoer
that α is not an algebraic integer; we try to obtain a contradiction. Since α is not an algebraic
integer, there exists a finite absolute value w of K such that |α|w > 1. Let H ′ ⊂ Gal(K/Q) be the
subgroup fixing Q(α), so that we have the inclusions H ′ ⊂ H ⊂ Gal(K/Q), corresponding to the
chain Q(αh ) ⊂ Q(α) ⊂ K.
For each i = 1, . . . , d, let Ti ⊂ Gal(K/Q) be a complete set of representatives for the set of
automorphisms coinciding with σi in Q(αh ), modulo H ′ . In other words the elements of Ti , when
restricted to Q(α), give all the embedding of Q(α) ֒→ K whose restrictions to Q(αh ) coincide with
σi . Also T1 ∪ . . . ∪ Td is a complete set of embeddings of Q(α) in K. Then we can write the trace
of αn (n = r + hm) as
d
!
X X
r+hm
TrQ(α)/Q (α )= τ (α ) σi (αh )m = λ1 σ1 (αh )m + . . . + λd σd (αh )m ,
r

i=1 τ ∈Ti

r
P
where λi = τ ∈Ti τ (α ). Note that not all the coefficients λi can vanish, since then the trace
would also vanish. Since the trace of qn αn is integral, we have |TrQ(α)/Q (αn )|w ≤ |qn |−1
w ; then,
since log qn = o(n), for every ǫ < log |α|w / log H(α) and sufficiently large m ∈ Ξ′ we have
|λ1 σ1 (αhm ) + . . . + λd σd (αhm )|w ≤ |qr+hm |−1
w < |α
hm
|w · H(αhm )−ǫ .
Applying Lemma 1 we arrive at a non-trivial equation of the form
a1 σ1 (αh )m + . . . + ad σd (αh )m = 0
satisfied by infinitely many integers m. An application of the Skolem-Mahler-Lech theorem leads
to the conclusion that for two indices i 6= j some ratio σi (αh )/σj (αh ) = (σi (α)/σj (α))h is a root
of unity. But then it equals 1, since σi (α)/σj (α) lies in K × (and by assumption h is the exponent
of the torsion group of K × ). Then σi coincides with σj on Q(αh ). This contradiction concludes
the proof.
Proof of Theorem 1. Let us suppose that the hypotheses of Theorem 1 are satisfied, so kαkn < ln
for infinitely many n. Then, either α is a d-th root of an integer, and we are done, or kαn k = 6 0
for all n > 1. In this case, taking any ǫ < − log l/ log H(α) we obtain that for infinitely many n
0 < kαn k < H(αn )−ǫ
and of course the sequence αn belongs to a finitely generated subgroup of Q̄× . Then our Main
Theorem, with q = 1, u = αn , implies that infinitely many numbers αn are pseudo Pisot number,
in particular have non-zero integral trace. By Lemma 4, αn is an algebraic integer, so it is a Pisot
number as wanted.

§3 Proof of Theorem 2. We recall some basic facts about continued fractions. Let α > 1 be a
real irrational number. We will use the notation
α = [a0 , a1 , . . .],

8
where a0 , a1 , . . . are positive integers, to denote the continued fraction for α. We also let ph , qh , for
h = 0, 1 . . . , be the numerator and denominator of the truncated continued fraction [a0 , a1 , . . . , ah ],
so that, by a well known fact, for all h we have
ph 1
−α ≤ 2 . (4.1)
qh qh · ah+1
Also, we have the recurrence relation, holding for h ≥ 2,

qh+1 = ah+1 qh + qh−1 . (4.2)

 
a b
Let T = ∈ GL2 (Z) be a unimodular matrix and let [b0 , b1 , . . .] be the continued fraction
c d
of
aα + b
T (α) := .
cα + d
Then there exists an integer k such that for all large h, bh = ah+k .
Consider now the case of a real quadratic irrational number α > 0, and let α′ be its (algebraic)
conjugate. The continued fraction of α is eventually periodic; it is purely periodic if and only if
α > 1 and −1 < α′ < 0; we call such a quadratic irrational reduced. In this case the period of −1/α′
is the “reversed” period of α. For every quadratic irrational number there exists a unimodular
transformation T ∈ GL2 (Z) such that T (α) is reduced (this is equivalent to saying that the
expansion of α is eventually periodic). Since the transformation x 7→ −1/x is also unimodular, it
follows that in any case α and α′ have a period of the same length. We can summarize these facts
as follows:
Facts. Let α be a real quadratic irrational, α′ its conjugate. The continued fraction development
of α is eventually periodic. The quadratic irrationals
1 1
|α|, |α′ |, ,
α α′
have periods of the same length.
We begin by proving the easier part of Theorem 2, namely:
If α > 0 is a unit in the quadratic ring Z[α] then the period for the continued fraction for αn
is uniformely bounded.
Let us begin with the case of odd powers of a unit α > 1, with α′ < 0, i.e. a unit of norm −1.
Let us denote by tn = αn + α′n = TrQ(α)/Q (αn ) the trace of αn . Then for all odd integers n, the
number αn satisfies the relation (αn )2 − tn αn − 1 = 0, i.e.
1
αn = tn +
αn
hence its continued fraction is simply [tn ], and has period one.
We shall now consider units αn of norm 1, so including also even powers of units of norm −1.
Suppose α > 1, with 0 < α′ < 1 is such a unit, so α′ = α−1 . Denoting again by tn the trace of αn ,
we see that the integral part of αn is tn − 1, so we put a0 (n) = tn − 1. We have
1 − αn αn − 1
(αn − a0 (n))−1 = (1 − α′n )−1 = = .
(1 − α′n )(1 − αn ) tn − 2
For sufficiently large n, its integral part is 1. So put a1 (n) = 1 (at least for sufficiently large n).
Then  n −1  n −1
α −1 α − 1 − tn + 2 tn − 2
−1 = = .
tn − 2 tn − 2 1 − α′n

9
Again for sufficiently large n, the integral part of the above number is tn − 2, so put a2 (n) = tn − 2.
Now

tn − 2 tn − 2 − tn + tn α′n + 2 − 2α′n 1 + α′2n − 2α′n

− (t n − 2) = = = 1 − α′n
1 − α′n 1 − α′n 1 − α′n

where we used the equation satisfied by α′n to simplify the numerator. Then
−1
αn − 1

tn − 2 ′n −1
− a 2 (n) = (1 − α ) =
1 − α′n tn − 2

and the algorithm ends giving the continued fraction expansion, valid for sufficiently large n,

αn = [a0 (n), a1 (n), a2 (n)] = [tn − 1, 1, tn − 2].

We now prove our last lemma:

Lemma 5. Let α > 1 be a real quadratic number, not the square root of a rational number. Let
a0 (n), a1 (n), . . . be the partial quotients of the continued fraction for αn . Then either α is a Pisot
number, or for every i with i 6= 0,
log ai (n)
lim = 0. (4.3)
n→∞ n
If α > 1 is the square root of a rational number, then (4.3) holds provided the limit is taken over
odd integers n.
Proof. We argue by contradiction. Let h ∈ N, h > 0 be the minimum index i such that (4.3)
does not hold. Then for a positive real ǫ and all n in an infinite set Ξ ⊂ N

ah (n) > eδn . (4.4)

Here it is meant that Ξ contains only odd integers if α is the square root of a rational number.
On the other hand, since (4.3) holds for all i = 1, . . . , h − 1, we have, in view of the recurrence
relation (4.2), that the denominators qh−1 (n) of ph−1 (n)/qh−1 (n) = [a0 (n), a1 (n), . . . , ah−1 (n)]
satisfy
log qh (n)
lim = 0. (4.5)
n→∞ n
δ
Put ǫ = . By the vanishing of the above limit we have in particular
2 log H(α)
δn
qh−1 (n)1+ǫ ≤ e 2

for sufficiently large n ∈ Ξ̄. In view of (4.1) we have, for all such n,

ph−1 (n) 1 1
αn − < 2 δn
< 3+ǫ
qh−1 (n) qh−1 (n) · e qh−1 (n) · eδn/2

which can be rewritten as

kqh−1 (n)αn k < qh−1 (n)−2−ǫ · (H(αn ))−ǫ .

An application of our Main Theorem, with u = αn (for n ∈ Ξ), gives the conclusion that for large
n the algebraic number qh−1 (n)αn is pseudo-Pisot. Now, taking into account (4.5), our Lemma 4
implies that α is an algebraic integer (it cannot have a vanishing trace, since otherwise it would be

10
the square root of a rational number, which we excluded). This in turn implies that αn is a Pisot
number. Since α is quadratic irrational and αn is not a rational, α itself is a Pisot number.
We can now finish the proof of Theorem 2. Let α be a real quadratic irrational number; let
us treat first the case that α is not the root of a rational number. Suppose that for all positive
integers n in an infinite set Ξ ⊂ N, the period of the continued fraction of αn has the same length
r. We would like to prove that α is a unit.
Let us first show that, by the above mentioned Facts, we can reduce to the case where

α>1 and α > |α′ |. (4.6)

In fact, after replacing if necessary α by ±1/α we can suppose that α > 1. Observe now that if
α = −α′ then α is the square root of a rational number, which we have excluded. Then |α| 6= |α′ |
(since we cannot have α = α′ ). If now α > |α′ | we are done; otherwise |α′ | > α > 1; in this case
we replace α by ±α′ and obtain (4.6) as wanted.
n
So, from now on, let us suppose (4.6) holds for α. Then for all n ≥ n0 (α) we have αn > α′ + 2
so there exists an integer kn such that

αn − kn > 1, −1 < (αn − kn )′ < 0,

so αn − kn is reduced. Then we have for all n ∈ Ξ,

αn − kn = [a0 (n), a1 (n), . . . , ar−1 (n)]

where a0 (n), . . . , ar−1 (n) are positive integers and the period (a0 (n), . . . , ar−1 (n)) is the same as
the one for αn (still for n in the infinite set Ξ). Suppose first that α is not a Pisot number. We
would like to derive a contradiction. Our Lemma 5 implies that all partial quotients ai (n) satisfy
n
(4.3), including a0 (n) = ar (n). The algebraic numbers αn − kn and α′ − kn satisfy, for n ∈ Ξ, the
algebraic equation

pr−1 (n)x + pr−2 (n)

x = [a0 (n); a1 (n), . . . , ar−1 (n), x] = ,
qr−1 (n)x + qr−2 (n)

which can be also written as

qr−1 (n)x2 + (qr−2 (n) − pr−1 (n))x − pr−2 (n) = 0.

In view of (4.5), the coefficients of this equation have logarithms bounded by o(n), whence the
n
logarithmic heights of x = αn − kn and x′ = α′ − kn are (o(n)) for n ∈ Ξ. But then we would
have
n
log |αn − α′ | = o(n)
which is clearly impossible.
We have then proved that α is a Pisot number, so it is an algebraic integer and verifies |α′ | < 1.
But now the quadratic irrational ±1 α′ also satisfies (4.6), so the same reasoning implies that 1/α ,
′

hence 1/α, is also an integer. This proves that α is a unit as wanted.p

The method is exactly the same in the second case, when α = a/b is the square root of a
positive rational number a/b. Here too, by replacing if necessary α with 1/α, we reduce to the
case α > 1. In this case the pre-period has length one, so (αn − kn )−1 is reduced, where kn is the
integral part of αn . Under the hypothesis that the period length of αn remains bounded for an
infinite set of odd integers, we can still apply our Lemma 5 and conclude as before.

Appendix
11
 In this section we show that the word “algebraic” cannot be removed from the statement of
Theorem 1. A little more precisely, we prove that: There exists a real number α > 1 such that
||αn || ≤ 2−n for infinitely many positive integers n, but αd is not a Pisot number, no matter the
positive integer d.
Our assertion will follow from a simple construction which we are going to explain. For a set
I of positive real numbers and a positive real t, we put I t := {xt : x ∈ I}.
We start by choosing an arbitrary sequence {βn }n≥1 of real numbers in the interval [0, 1/2];
we shall then define inductively integers b0 , b1 , . . . and closed intervals I0 , I1 , . . . of real numbers,
of positive length and contained in [2, ∞). We set I0 = [2, 3], b0 = 1 and, having constructed
b
In , bn we continue as follows. Let bn+1 be a positive integer divisible by n + 1 and such that Inn+1
contains an interval of length > 2; such an integer will certainly exist since by induction every
element of In is ≥ 2 and since In has positive length.
b
Put Bn = b0 · · · bn . Then we choose In+1 to be any subinterval of Inn+1 of the shape
In+1 = [q + βn+1 , q + βn+1 + 2−Bn+1 ],
b
where q = qn is an integer. This will be possible since Inn+1 has length > 2. Clearly q ≥ 2, so In+1
has the required properties.
1/B b
For n ∈ N, let now Jn denote the closed interval InT n . Since In+1 ⊂ Inn+1 , we have Jn+1 ⊂ Jn
for all n and J0 = I0 = [2, 3]. Hence there exists α ∈ ∞ n=0 Jn such that α ≥ 2.
Note that αBn ∈ In ; hence we conclude that: For n ≥ 1 the fractional part {αBn } of αBn lies
between βn and βn + 2−Bn .
To obtain our original assertion, we apply this construction by choosing βn = 0 for even
integers n ≥ 1 and βn = 1/3, say, for odd n. We then have in the first place
||αBn || ≤ {αBn } ≤ 2−Bn ,
for all even integers n ≥ 1. This yields the first required property of α, because Bn → ∞.
Further, suppose that αd were a Pisot number, for some positive integer d. Then we would
have, for some fixed positive λ < 1 and integers m ≥ 1, ||αdm || ≪ λm (by a well-known argument,
used also in the previous proofs of this paper). However this is not possible, because for odd
integers n > d we have that Bn is divisible by d (recall bn is divisible by n for n > 0) and the
fractional part of αBn is between 1/3 and 1/3 + 2−Bn , whence the norm ||αBn || is ≥ 1/6.
This concludes the argument. It will be noted that the construction yields more precise conclu-
sions; also, Theorem 1 obviously implies that any α likewise obtained is necessarily transcendental.

References
[CZ] - P. Corvaja, U. Zannier, On the length of the continued fraction for the ratio of two power
sums, (2003) preprint NT/0401362 at http://www.arXiv.org/ .
[M] - K. Mahler, On the fractional parts of the powers of a rational number (ii), Mathematika 4
(1957), 122-124.
[MF] - M. Mendès France, Remarks and problems on finite and periodic continued fractions, L’En-
seignement Math. 39 (1993) 249-257.
[S] - W. M. Schmidt, Diophantine approximation and diophantine equations. Lecture Notes in
Mathematics 1467, Springer (1991).

P. Corvaja U. Zannier
Dip. di Matematica e Informatica Ist. Univ. Arch. - D.C.A.
Via delle Scienze S. Croce, 191
33100 - Udine (ITALY) 30135 Venezia (ITALY)
[email protected] [email protected]

12