LXXXVIII.4 (1999)
Number of representations related to a linear recurrent basis
by
Jean Marie Dumont†, Nikita Sidorov (St. Petersburg) and Alain Thomas (Marseille)
0. General results and notations. We first present a general state- ment on the number of representations related to a linear recurrent basis.
Let r be an integer, r ≥ 1, and let a
1, . . . , a
rbe reals. We consider a sequence (G
k)
k≥0such that for any k ≥ r,
G
k= a
1G
k−1+ . . . + a
rG
k−r.
Let P (X) = X
r−a
1X
r−1−. . .−a
r; we will assume the following hypothesis to be satisfied: there is only one root of P (X) with the maximum modulus;
it is a real φ > 1 and has multiplicity 1. Thus, the limit γ = lim
k→∞
(G
kφ
−k) exists, and we assume it to be positive.
Now let q be an integer, q ≥ 2, and consider the alphabet A
q= {0, 1, . . . . . . , q − 1}. We denote an infinite word on A
qby ε
0ε
1. . . We define f (x), the number of q-representations of a real x in the basis (G
k), as
f (x) = # n
ε
0ε
1. . . : X
∞ i=0ε
iG
i= x o
.
Since G
k∼ γφ
kwith γ > 0 and φ > 1, these sequences are such that ε
i≡ 0 for i large enough. For any x ∈ R, the set of x
0< x such that f (x
0) 6= 0 is finite, and we can define the summation function
Φ(x) = X
x0<x
f (x
0).
1991 Mathematics Subject Classification: Primary 11A67; Secondary 11B75, 60E05, 60E10.
†Prof. Dumont passed away on September 18, 1997.
The second author was supported in part by the French foundation Pro Mathematica.
[371]
We will prove in Section 2 the existence of a unique periodic function H with period 1 such that
Φ(x) = x
logφqH(log
φx) + o(x
logφq), x → ∞.
The function H is continuous, a.e. differentiable and Φ(x) x
logφqas x →
∞, or, equivalently, H(x) > 0 for any x ∈ R.
In Section 3 we give an upper bound for the remainder term, depending on the constant δ such that f (x) = O(x
δ), and in Section 4 a lower bound for the Lipschitz exponent of H will be obtained under the same hypothesis.
Section 5 concerns properties of H for some cases of φ. If φ is an irrational Pisot number, then H is not differentiable on a certain continuum of points.
For φ ∈ N and G
k= φ
kwith q not divisible by φ we have the same result.
On the contrary, in the case when φ divides q, the function H is absolutely continuous.
In Section 6 we study the general case of an integer base of the form G
k= d
k, i.e. a radix base. The formula
Φ(x) = x
αH(log
dx) + O(x
β)
holds with α = log
dq and β = log
d[(q −1+d)/d], and we show that in a gen- eral situation the order of the remainder term R(x) := Φ(x) − x
αH(log
dx) cannot be improved. In the case of a radix base we completely reduce the problems of finding the best possible exponent of the remainder term and the minimal Lipschitz exponent for the function H to estimating f (n). More precisely, it is proved that
(1) δ
f:= inf{δ : R(x) = O(x
δ)} = inf{δ : f (n) = O(n
δ)}, (2) the Lipschitz exponent for H equals α − δ
fif δ
f> α − 1.
Thus, it proves to be sufficient to estimate f (n). We present the best possible estimates in the following cases:
(i) d divides q, (ii) d = 2, q = 5, (iii) q ≤ d
2.
For instance, in the simple case d = 2, q = 3 we prove that the exact order of the remainder term is x
β0with β
0= log
2√5+12.
Note that in the case q = d
mfor some m ≥ 1, we give an explicit asymptotic formula for f (n).
Besides, we associate with any radix base with q > d a finite set of 0-1 matrices M
0, M
1, . . . , M
d−1so that the growth of f (n) can be represented by means of the growth of the quantity kM
ε0M
ε1. . . M
εkk, where n = P
ki=0
ε
id
iis the ordinary d-ary expansion of a positive integer n. Note that certain
problems closely connected with the study of Mahler-type sequences were
considered in the recent work [DF].
The case of the Fibonacci sequence, i.e. G
k= F
kdefined by F
0= 1, F
1= 2 and F
k= F
k−1+ F
k−2, was under consideration in [S]. With the alphabet A
2= {0, 1}, f (n) is the number of representations of n as a sum of distinct Fibonacci numbers for any n ∈ N. Here the formula for the summation function is
Φ(x) = x
ΛH(log
λx) + O(x
1/2) with λ = (1 + √
5)/2 ' 1.62 and Λ = log
λ2 ' 1.44. The remainder term O(x
1/2) cannot be replaced by o(x
1/2); the function H belongs to the Lip- schitz class with exponent Λ − 1/2 ' 0.94. Note that the precise maximal order of the function f itself is f (x) = O(x
1/2) (see [Pu]).
Another example of such type is G
k= λ
k: the above formula holds with the same function H.
We would also like to mention that if one studies the number g(n) of representations of a positive integer n in a d-radix base with the set of all nonnegative integer digits, then the old result of N. G. Bruijn [Br] states that the asymptotic order of log g(n) is log
2n, i.e. g(n) grows faster than any polynomial. The authors express their gratitude to P. Flajolet for indicating this result.
Graphic interpretation. We associate a graded graph with any sequence (G
k) of the type described above and any q ≥ 2 in the following way. Let the zero level of the graph contain a single vertex. Next, the nth level with n ≥ 1 consists of the numbers of the form { P
ni=0
ε
iG
i: 0 ≤ ε
i≤ q − 1}. From any vertex exactly q edges descend, and each edge is parametrized by one of the digits 0, . . . , q − 1 (normally the leftmost one corresponds to 0). Thus, any finite path (ε
n, ε
n−1, . . . , ε
0) in this graph corresponds to one of the representations described above. We call this graph the representation graph associated with the sequence (G
n) and the positive integer q ≥ 2. Note that in a meaningful situation (i.e. when the function f assumes values greater than 1), the graph is not a tree. For the Fibonacci case the combinatorial and ergodic properties of the representation graph are studied in detail in [AlZ] and [SV]. The representation graph for the radix case with d = 2, q = 3 is depicted in the last section (see Figure 1). Note that this last graph appeared for the first time in the master thesis [Bo] in connection with the study of the celebrated de Rham curve introduced in [Rh].
1. Some preliminary lemmas about representations in the basis (G
k). We put
Y (ε
0, ε
1, . . .) :=
X
∞ i=0ε
iφ
−i,
where ε
iassumes the integer values between 0 and q − 1 with probability
1/q each. Let F
Ybe the distribution function of the random variable Y ; its continuity and purity is a consequence of the theorem stating that a series of discrete stochastically nonconstant random variables is continuous and either absolutely continuous or singular (the “Law of Pure Types”, see [JW, Theorem 35]).
We first give, in Lemma 1.1 below, some estimates for G
kand P
ki=0
ε
iG
i. Let the roots of P (X) be denoted as φ
1= φ, φ
2, . . . , φ
d0, and let α
ibe the order of multiplicity of φ
i. We suppose
φ > |φ
2| ≥ |φ
i| ∀i ∈ {3, . . . , d
0},
α
2≥ α
i∀i ∈ {3, . . . , d
0} such that |φ
2| = |φ
i|, γ = lim G
kφ
−k> 0.
We write g(x) h(x) if g(x) = O(h(x)) and h(x) = O(g(x)). We set
% = max(|φ
2|, 1), α
0=
α
2− 1, |φ
2| > 1, α
2, |φ
2| = 1, 0, |φ
2| < 1 (% = 1 and α
0= 0 if P (X) has degree one),
ϕ(k) = |φ
2|
kk
α2−1, ψ(k) = %
kk
α0(ψ(k) is defined in such a way that ψ(k) P
ki=0
ϕ(i)).
Lemma 1.1. There exist two positive constants K
1and K such that (i) |G
k− γφ
k| ≤ K
1ϕ(k),
(ii) for any k ∈ N and an infinite word ε
0ε
1. . . on the alphabet A
q,
X
ki=0
ε
iG
k−i− X
∞ i=0ε
iγφ
k−i≤ Kψ(k).
P r o o f. (i) is a direct consequence of the root expression of G
k, i.e.
G
k= X
i
X
j<αi
λ
i,jφ
kik
j, and of the general assumptions.
(ii) We have X
ki=0
ε
i(G
k−i− γφ
k−i) = O
X
ki=0
ϕ(k − i)
and
X
∞ i=k+1φ
k−i= O(1), hence both expressions are O(ψ(k)).
The following lemma gives a relation between the functions Φ and F
Y.
Lemma 1.2. Let x ∈ R and k ∈ N be such that x < γφ
k, let x
k= x − Kψ(k) and x
0k= x + Kψ(k). For k large enough we have
q
k+1F
Yx
kγφ
k≤ Φ(x) ≤ q
k+1F
Yx
0kγφ
k, (i)
Φ(x
k) ≤ q
k+1F
Yx γφ
k≤ Φ(x
0k).
(ii)
P r o o f. (i) Let ε
0ε
1. . . , x and k be such that P
∞i=0
ε
iG
i< x < γφ
k, and j be such that ε
j6= 0. Since G
i∼ γφ
i, the G
iare positive for i > i
1and we deduce from the hypothesis that
γφ
k> G
j−
i1
X
i=0
q|G
i| > γφ
j−1where the last inequality is valid for j large enough. In other words, if k is large enough, then ε
j= 0 for any j > k, whence
Φ(x) = # n
ε
0ε
1. . . ε
k: X
k i=0ε
iG
i< x o
= q
k+1m
qn
(ε
0, ε
1, . . .) : X
k i=0ε
iG
k−i< x o
,
where m
qis the uniform product measure on the compactum Q
∞0
{0, . . . . . . , q − 1}.
On the other hand, we have, with y = x
kor x
0k, F
Yy γφ
k= m
q(ε
0, ε
1, . . .) : X
∞ i=0ε
iφ
−i< y γφ
k, whence (i) follows by Lemma 1.1(ii).
(ii) We can replace x by x
k, and hence x
0kby x, in the second inequality in (i). Indeed, x
kalso satisfies the condition x
k< γφ
k. Thus, we obtain the first inequality in (ii).
We obtain the second one noting that Φ(x
0k) ≥ #
n
ε
0ε
1. . . ε
k: X
ki=0
ε
iG
i< x
0ko
and using the same method as for (i).
2. Main result
Theorem 2.1. There exists a unique periodic function H with period 1,
well defined on the whole real line, such that
Φ(x) = x
logφqH(log
φx) + o(x
logφq), x → ∞.
This function is positive, continuous and a.e. differentiable.
P r o o f. With any x we associate the integer k = k(x) such that γφ
k−1≤ x < γφ
k: it is
k(x) =
log
φx
γ
+ 1.
We apply Lemma 1.2(i) and note that x/(γφ
k) − x
k/(γφ
k) and x
0k/(γφ
k) − x/(γφ
k) tend to 0 as x → ∞ and k = k(x). By the uniform continuity of the distribution function, F
Y(y + t) − F
Y(y) tends to 0 uniformly in y as t → 0. So, we deduce from Lemma 1.2(i) that
Φ(x) = q
k(x)+1F
Yx
γφ
k(x)+ o(q
k(x)+1).
Thus, we obtain the desired formula with H(log
φx) = q
k(x)+1−logφxF
Yx
γφ
k(x).
Periodicity of H. Clearly H(1 + log
φx) = H(log
φ(φx)) = H(log
φx).
Uniqueness of H. The above formula and the periodicity of H imply (2.1) H(log
φx) = lim
k→∞
H(log
φ(xφ
k)) = lim
k→∞
Φ(xφ
k)(xφ
k)
− logφq. Positivity. We deduce from the definition of H that
H(log
φx) ≥ q
1−logφγF
Y(1/φ).
The number F
Y(1/φ) is positive, as there exists k
0such that Y (ε
0, ε
1, . . .) <
1/φ if ε
0= ε
1= . . . = ε
k0= 0.
Continuity. The function H could be discontinuous only at x such that log
φ(x/γ) is an integer; nevertheless, if we replace in the definitions of H and k(x) the condition γφ
k−1≤ x < γφ
kby γφ
k−1/ √
φ ≤ x < γφ
k/ √ φ (for instance), the formula of Theorem 2.1 will remain valid. Since log
φ(x/γ) and log
φ(x √
φ/γ) cannot be both integers, the (unique) function H is everywhere continuous.
Differentiability. The continuity of the distribution function F implies that it is a.e. differentiable, thence H is a.e. differentiable. The proof is complete.
3. Improvement of the remainder term. Note first that since f (n) ≤
Φ(n+1), we have f (n) = O(n
logφq). So, the sequence f (n) cannot grow more
rapidly than a polynomial. Henceforward we assume that a constant δ > 0
is such that
(I) f (x) = O(x
δ) for x ∈ R
+, and
(II) #{x
0: x
0∈ [x, x + 1[, f (x
0) 6= 0} is uniformly bounded for x ∈ R
+(
1).
For this situation the main formula of the previous section can be improved.
Theorem 3.1. Let % and α
0be defined as in Section 1. We have Φ(x) = x
αH(log
φx) + O(x
δ0(log x)
α0)
with α = log
φq and δ
0= δ + log
φ%.
P r o o f. We use Lemma 1.2(ii) with k = k(x) = [log
φ(x/γ)] + 1. By hypothesis (II), the number of reals y ∈ [x
k, x
0k] such that f (y) 6= 0 is O(x
0k− x
k); and by hypothesis (I) we have in this case f (y) = O((x
0k)
δ) = O(x
δ). We deduce that both Φ(x) − Φ(x
k) and Φ(x
0k) − Φ(x) are O(x
δψ(k)), whence
q
k(x)+1F
Yx
γφ
k(x)= Φ(x) + O(x
δψ(k(x))).
It remains to note that ψ(k(x)) = %
k(x)(k(x))
α0= O(x
logφ%(log x)
α0).
Theorem 3.1 improves Theorem 2.1 iff α > δ
0. We will give a suitable estimate of the order of f (n) in a certain natural case.
Proposition 3.2. Let f (x) be the number of q-representations of x in a certain nonnegative integer base (G
k), with q = [φ] + 1. If φ > (1 + √
5)/2 and φ 6= 2, then f (n) = O(n).
P r o o f. By Lemma 1.1, there exists a constant K
0such that G
k− [φ]
k−2
X
i=0
G
i≥ γφ
k− [φ]γ
k−2
X
i=0
φ
i− K
0ψ(k)
> γφ
k− [φ]γ φ
k−1φ − 1 − K
0ψ(k)
> γφ
k−1φ
2− φ − [φ]
φ − 1 − K
0ψ(k).
The assumptions on φ imply that (φ
2− φ − [φ])/(φ − 1) > 0, whence there is a constant k
1such that
(3.1) G
k> [φ]
k−2
X
i=0
G
i, G
k−1> 0 and G
kis nondecreasing for k ≥ k
1.
(
1) The second condition is satisfied, for instance, if G
k∈ N, or if G
k= φ
kwith φ
being a Pisot number (see [Bu]).
Choosing a constant C satisfying the relation f (n) ≤ Cn for any n <
G
k1, we prove by induction the validity of this relation for n ∈ N. Namely, suppose f (n
0) ≤ Cn
0for any n
0≤ n − 1 and let k be such that G
k≤ n <
G
k+1. There exists also an integer ε ≥ 0 such that (3.2) εG
k−1+ [φ]
k−2
X
i=0
G
i< n ≤ (ε + 1)G
k−1+ [φ]
k−2
X
i=0
G
i.
As n < G
k+1, the representations of n in the alphabet {0, 1, . . . , [φ]} satisfy ε
i= 0 for i ≥ k+1. If ε
k= 0, we have ε
k−1≥ ε+1. With each representation of n satisfying ε
k6= 0, we associate a representation of n − G
k, replacing the digit ε
kby ε
k− 1. With each representation of n such that ε
k= 0, we associate a representation of n−(ε+1)G
k−1, replacing ε
k−1by ε
k−1−(ε+1).
Thence
f (n) ≤ f (n − G
k) + f (n − (ε + 1)G
k−1)
≤ C(n − G
k) + C(n − (ε + 1)G
k−1), and we obtain f (n) ≤ Cn from inequalities (3.1) and (3.2).
Remark. The above proposition works, for instance, in the case of φ being a Pisot or Salem number, because in that case α > 1 ≥ δ
0= δ. Besides, considering the example introduced in [Fr], namely, G
k= 3G
k−1+ 2G
k−2+ 3G
k−4with positive initial values and q = 4, we make sure that α > δ
0, although φ is neither Pisot nor Salem number.
4. Lipschitz exponent for H. We assume the same hypotheses as in Section 3 to be satisfied, and also suppose that the constant α = log
φq is greater than δ
0= δ + log
φ%.
Theorem 4.1. There exists a constant C > 0 such that
|H(t
0) − H(t)| ≤ C|t
0− t|
α−δ0|log |t
0− t||
α0for |t
0− t| ≤ 1/2.
We first prove
Lemma 4.2. For x ≥ 1, (i) Φ(x + 1) − Φ(x) = O(x
δ), (ii) Φ(x) = O(x
δ+1).
P r o o f. The first inequality follows from the hypothesis on f , and the second one is deduced from the first with the help of the inequality
Φ(x) − Φ(1) ≤ X
[x]n=1
(Φ(n + 1) − Φ(n)).
P r o o f (of Theorem 4.1). As H is periodic, we can confine ourselves to 1 ≤ t < 2 and t < t
0≤ t + 1/2. We have φ
t0− φ
t< φ
3. Furthermore, there exists a nonnegative integer h such that φ
2≤ (φ
t0− φ
t)φ
h< φ
3. We set x = φ
t+hand x
0= φ
t0+h; then φ ≤ x < x
0< x
2. From Theorem 3.1 we get
|Φ(x) − x
αH(log
φx)| = O(x
δ0(log x)
α0) and the same inequality for x
0, whence
|H(t
0) − H(t)| = |H(log
φx
0) − H(log
φx)|
≤ (x
0)
−α(Φ(x
0) − Φ(x)) + |(x
0)
−α− x
−α|Φ(x) + O(x
δ0−α(log x
0)
α0).
We have 0 < x
0−x < φ
3, whence we deduce from Lemma 4.2(i) that the first term is O((x
0)
δ−α); it is also O(x
δ−α), as, by hypothesis, δ − α < − log
φ% ≤ 0.
The second term, in view of the mean value theorem and Lemma 4.2(ii), is O(x
−α−1Φ(x)) = O(x
δ−α). Finally, as δ ≤ δ
0and 1 < x < x
0< x
2,
|H(t
0) − H(t)| = O(x
δ0−α(log x)
α0).
It remains to estimate x = φ
t+h. By the mean value theorem, we have φ
tlog φ < φ
t0− φ
tt
0− t < φ
t0log φ < φ
t+1log φ,
whence, by the definition of h, there exist positive constants C
1and C
2such that
C
1t
0− t < x < C
2t
0− t , which completes the proof of Theorem 4.1.
Corollary 4.3. If φ is a Pisot number , then H belongs to the Lipschitz class with exponent α − δ, where α = log
φq and f (x) = O(x
δ).
5. Properties of F
Yfor some cases of φ Lemma 5.1. Let
Y (ε e
0, ε
1, . . .) := Y − EY = X
∞ i=0ε
i− q − 1 2
φ
−i.
Then the characteristic function of e Y is given by the formula F
Ye(t) =
Y
∞ k=0sin(qφ
−kt/2) q sin(φ
−kt/2) .
P r o o f. e Y being a sum of independent random variables, its charac-
teristic function is the product of their characteristic functions, which we
compute using the well-known expression for the Dirichlet kernel D
n(x) = P
nk=−n
e
ikx(see, e.g., [Ru]).
Proposition 5.2. If φ is an irrational Pisot number , then F
Ye(t) does not tend to 0 as t → ∞, and thus, Y is singular.
P r o o f. Let t
N= 2πφ
N. Following [E], we will show that F
Ye(t
N) 6→ 0 as N → ∞. Let kxk := min{|x − n| : n ∈ Z}; we have
F
Ye(t
N) = Y
∞ k=0sin(πqφ
N −k) q sin(πφ
N −k) =
Y
N j=1sin(πqφ
j) q sin(πφ
j) ·
Y
∞ k=0sin(πqφ
−k) q sin(πφ
−k)
= ±F
Ye(2π) · Y
N j=1sin(πqkφ
jk) q sin(πkφ
jk) ,
whence in view of the fact that φ is an irrational Pisot number, there exists an irrational θ ∈ (0, 1) such that kφ
jk ≤ θ
jfor j ≥ N
0. Hence
|F
Ye(t
N)| ≥ |F
Ye(2π)| ·
Y
∞ j=N0+1sin(πqθ
j) q sin(πθ
j) ·
N0
Y
j=1
sin(πqkφ
jk) q sin(πkφ
jk) > 0,
as |F
Ye(2π)| > 0 by the irrationality of φ.
Now we are going to study the same problem for the case of integral φ =: d. We assume that G
k= d
k.
Proposition 5.3. Let d be an integer. Then the function F
Yis abso- lutely continuous if and only if d divides q.
P r o o f. Let, within this proof,
G
ν(u) := sin νu ν sin u . Consider the following cases.
1. q = d
m. Here
G
q(u) =
m−1
Y
j=0
G
d(d
ju), hence,
F
Ye(t) = Y
∞ k=0G
q(d
−kt/2) =
m−1
Y
j=0
Y
∞k=0
G
d(d
−k· d
jt/2)
, i.e.
Y = U e
− 1 2 , 1
2
∗ U
− d 2 , d
2
∗ . . . ∗ U
− d
m−12 , d
m−12
,
where U [a, b] denotes the uniform distribution on [a, b]. Therefore, the ran- dom variable e Y is absolutely continuous, and supp e Y =
−
2(d−1)q−1,
2(d−1)q−1.
2. Let now q be not divisible by d. Here, as in Proposition 5.2, taking t
N= 2πd
N, we have
F
Ye(t
N) = ± Y
∞ j=1sin(πqd
−j) q sin(πd
−j) ,
and the assertion follows from the fact that sin(πqd
−j) 6= 0 for any j ≥ 1 by the hypothesis on q and d.
3. Finally, if q = d
mp with p ≥ 2 and p not divisible by d, then G
q(d
−kt/2) = G
dm(d
−kt/2) · G
p(d
−k· d
mt/2),
and the statement for this case follows from items 1, 2, as e Y is the convo- lution of the absolutely continuous distribution and the singular one, and, thus, it is itself absolutely continuous.
As a consequence, we obtain the result making Theorem 2.1 more precise.
Proposition 5.4. If φ is an irrational Pisot number or equals a positive integer d and in this latter case G
k= d
k, and d does not divide q, then F
Yis singular , and H is not differentiable on a certain continuum of points.
The question of for which bases the function H is differentiable every- where seems to be difficult and remains open. The only sufficient condition that we know is G
k= d
kand d divides q (see above).
Example. Let d = 2, q = 6. Here e Y is the convolution of the uniform distribution U [−1/2, 1/2] and the singular one for the case “d = 2, q = 3”
studied below in detail. Thus, the function f for this case has no asymptotics;
this can be shown directly by proving that f (2
k) ∼
16·3
k, and f (2
k+2
k−1) ∼ 3
k−1, which in turn implies that there is no limit of the ratio f (n)/n
log23as n → ∞, though this ratio is bounded.
6. Case of a radix base
6.1. General facts. Let d ≥ 2, q ≥ 2 be integers, and let, as above, for any integer n,
f (n) = # n
ε
0ε
1. . . : 0 ≤ ε
i≤ q − 1, X
∞ i=0ε
id
i= n o
. We also define, for any x ∈ R \ {0},
ϕ(x) = X
n≥0
f (n)x
n. Expanding the infinite product Q
k≥0
(1 + x
dk+ x
2dk+ . . . + x
(q−1)dk), we find that it is equal to ϕ(x); hence
ϕ(x) = (1 + x + x
2+ . . . + x
q−1)ϕ(x
d).
Since f (n) is zero for n < 0, we obtain
(6.1) X
n∈Z
f (n)x
n= X
n∈Z
f (n)(x
dn+ x
dn+1+ . . . + x
dn+q−1).
Let {x} = x − [x] (the fractional part of x) for any x ∈ R. So, it follows from formula (6.1) that
(6.2) f (dn + r) =
[(q−1)/d]−1
X
i=0
f (n − i), r > d{(q − 1)/d},
[(q−1)/d]
X
i=0
f (n − i), r ≤ d{(q − 1)/d},
for every n ∈ Z and every r = 0, 1, . . . , d − 1. Furthermore, if [(q − 1)/d] = 0, the first case in (6.2) turns into f (dn + r) = 0 for r > q − 1. We deduce from (6.2) by induction
f (n) ≤ n
β, β = log
dq − 1 + d d
,
and by Theorem 3.1, the following corollary holds (as % = 1 and α
0= 0).
Proposition 6.1.
Φ(x) = x
αH(log
dx) + O(x
β) with α = log
dq and β = log
d[(q − 1 + d)/d].
Remarks. 1. We always have α − 1 ≤ β < α, with equality iff d divides q. In the case q ≤ d, we have β = 0 and it is easy to see that the function f is d-automatic in the sense of [Co]: (f (n))
n≥0is a fixed point, beginning with the letter 1, of the substitution
0 → 0 . . . 0 (d times 0),
1 → 1 . . . 10 . . . 0 (q times 1, d − q times 0).
The a.e. differentiability of H is a consequence of Theorem 2.1. Note that this substitution is not primitive, as for a primitive substitution it was proved in [DuT] that H is nowhere differentiable.
Below we will suppose q > d.
2. In the general case, the sequence (f (n))
n≥0is d-regular in the sense of Allouche and Shallit [AllS]. Indeed, by (6.1) we are in the situation of Becker’s Theorem 2 (see [Be]). The authors wish to express their thanks to J. P. Allouche for indicating the last-mentioned paper.
6.2. Relation between the order of f and the order of the remainder term.
In this subsection we are going to show that in the radix case it suffices to
present the best possible upper bound of the order of f for obtaining the
best possible order of the remainder term.
We first rewrite formula (6.2) by means of some transition matrices M
r, r = 0, 1, . . . , d − 1. Let M
rbe the a × a matrix, where a is the minimal integer such that a ≥ (q − 1)/(d − 1). For i = 1, . . . , a, let
λ
r,i= i + d − r − 2
d , λ
0r,i=
q−1d
, {λ
r,i} +
q−1d
< 1 −
1d,
q−1
d
+ 1, otherwise.
Then the ith row of M
ris defined as (M
r)
i,j=
1, [λ
r,i] + 1 ≤ j ≤ [λ
r,i] + [λ
0r,i], 0, otherwise.
This definition makes sense if the integer [λ
r,i] + [λ
0r,i] is at most a for any i;
or, equivalently, if it is less than a + 1. Now we deduce from the definition of λ
0r,ithat [λ
r,i]+[λ
0r,i] ≤ (a+d−2)/d+q/d. The inequality (a+d−2)/d+q/d <
a + 1 is valid, because it is equivalent to a > (q − 2)/(d − 1).
We denote by v(n) the vector
f (n)
.. .
f (n−a+1)
, for any n ≥ 0.
Lemma 6.2. Any integer n = ε
kd
k+ . . . + ε
0with 0 ≤ ε
i≤ d − 1 satisfies
v(n) = M
ε0. . . M
εk
1 0 .. . 0
.
P r o o f. It is sufficient to show that v(n) = M
ε0v(n
0) with n
0such that n = dn
0+ ε
0. Let i ∈ {1, . . . , a} and let i
0, i
00be such that i + d − ε
0− 2 = di
0+ i
00, 0 ≤ i
00≤ d − 1. We have n + 1 − i = d(n
0− i
0) + (d − 1 − i
00) whence by (6.2),
f (n + 1 − i) =
[(q−1)/d]−1
X
k=0
f (n
0− i
0− k), d − 1 − i
00> d{(q − 1)/d},
[(q−1)/d]
X
k=0
f (n
0− i
0− k), d − 1 − i
00≤ d{(q − 1)/d}.
This coincides with the product of the ith row of M
ε0by v(n
0).
Lemma 6.3. The matrix M = M
0+ . . . + M
d−1is primitive, and its Perron–Frobenius eigenvalue equals q.
P r o o f. Let
M
1 .. . 1
=
x
1.. . x
d
.
We have x
j= d
q − 1 d
+ #
i : 0 ≤ i ≤ d − 1, {λ
i,j} +
q − 1 d
≥ d − 1 d
. Since i 7→ {λ
i,j} is a bijection from {0, 1, . . . , d − 1} to
0d
,
1d, . . . ,
d−1d, we have
x
j= d
q − 1 d
+ d
q − 1 d
+ 1 = q.
We see that q is an eigenvalue of M , and the other eigenvalues have moduli at most q.
For M to be primitive, it suffices that the entries M
i,i−1, M
i,iand M
i,i+1are positive for any i (because this implies (M
a)
i,j> 0 for any i and j).
We consider the ith row of M . We have [λ
0r,i] ≥ 1 for any r and i; hence (M
r)
i,[λr,i]+1= 1. The set J
i= {j : M
i,j> 0} is the union, for r = 0, . . . , d − 1, of {j : (M
r)
i,j> 0}, the latter being an interval in N containing [λ
r,i] + 1. As [λ
r+1,i] − [λ
r,i] is 0 or −1, J
iis also an interval.
It is sufficient to check that [λ
d−1,i]+1 < i < [λ
0,i]+[λ
00,i] for i = 1, . . . , a, except [λ
d−1,1] + 1 = 1 and [λ
0,a] + [λ
00,a] = a. The first inequality is equivalent to λ
d−1,i< i − 1, thence to i > 1. For the second, we first deduce from the definition of [λ
0,i] that [λ
0,i] + [λ
00,i] is at least λ
0,i+ (q − 1)/d − (1 − 2/d), thence it is > i iff i < (q − 1)/(d − 1), i.e. iff i < a.
We obtain easily the two last-mentioned relations.
Proposition 6.4. Let R(x) be the remainder term in Proposition 6.1.
Then
inf{δ : R(x) = O(x
δ)} = inf{δ : f (n) = O(n
δ)}.
P r o o f. We first compute H(log
dx) for integral x. By (2.1), it is
k→∞
lim
Φ(xd
k) x
logdqq
k.
Let n be an integer such that 0 ≤ n < xd
k; we have n = md
k+ r with 0 ≤ m < x, 0 ≤ r < d
k. Let r = P
k−1i=0
ε
id
ibe the canonical representation of r in base d. From Lemma 6.2 we get
v(n) = M
ε0. . . M
εk−1v(m).
Summing for 0 ≤ r < d
k, 0 ≤ m < x, we obtain
Φ(xd
k) .. .
Φ(xd
k− a + 1)
= M
k
Φ(x) .. . Φ(x − a + 1)
.
By Lemma 6.3, q
−kM
ktends to a matrix with positive entries; let α
1. . . α
abe its first row, and β
i= P
aj=i
α
jfor any i. Then H(log
dx) = x
−α(α
1Φ(x) + . . . + α
aΦ(x − a + 1)) (6.3)
= x
−α(β
1Φ(x) − β
2f (x − 1) − . . . − β
af (x − a + 1)).
We have
R(x) = Φ(x) − x
αH(log
dx)
= (1 − β
1)Φ(x) + β
2f (x − 1) + . . . + β
af (x − a + 1).
Now 1 − β
1is zero, otherwise R(x) would have the same order as Φ(x). The hypothesis q > d implies a ≥ 2, whence
f (x − 1) R(x) f (x − 1) + . . . + f (x − a + 1).
This proves Proposition 6.4 for x integer.
It is also true for any x real: let n be the minimal integer greater than or equal to x; we have R(n) − R(x) = n
αH(log
dn) − x
αH(log
dx) and, using Theorem 4.1, we conclude that it is O(x
δ) for any δ such that f (n) = O(n
δ).
Remark. Some upper and lower bounds for f (n) depending on the transition matrices under consideration are given in [D, p. 160, 161, 165].
6.3. Lipschitz exponent for H. Now we will prove that the same situation takes place also for the Lipschitz exponent of the function H. Let δ
f= inf{δ : f (n) = O(n
δ)}, and γ
H= sup{γ : |H(t
0) − H(t)| = O(|t
0− t|
γ)}.
Corollary 6.5. If δ
f> α − 1, then γ
H= α − δ
f. P r o o f. For any ε > 0 we have, by Theorem 4.1,
|H(t
0) − H(t)| |t
0− t|
α−δf−ε.
Let x be an integer, t
0= log
d(x + 1) and t = log
dx; we deduce from (6.3) that
(x + 1)
αH(t
0) − x
αH(t) ≥ α
1f (x), (x + 1)
αH(t
0) ≥ α
1Φ(x + 1), and
H(t
0) − H(t) = ((x + 1)
−α− x
−α)(x + 1)
αH(t
0) + x
−α((x + 1)
αH(t
0) − x
αH(t))
≥ − αx
−α−1α
1Φ(x + 1) + x
−αα
1f (x).
As Φ(x) x
αand f (x) ≥ x
δf−εfor infinitely many x, we deduce
|H(t
0) − H(t)| x
δf−α−ε|t
0− t|
α−δf+ε, whence α − δ
f− ε ≤ γ
H≤ α − δ
f+ ε.
Remark. This corollary cannot be applied in the case when d divides
q: by Proposition 6.8 (see below) we have δ
f= α − 1. Moreover, if q = d
m,
then by Proposition 6.7 below we also have H(t) = constant and γ
H= ∞.
6.4. The case of d dividing q; reduction. In this subsection we give an arithmetic version of Proposition 5.3. Let us first study the radix case with q = d
min more detail. Namely, we will give a precise asymptotic formula for f in this case. Let f
m(d)(n) denote the number of representations of n in base d with digits 0, 1, . . . , d
m− 1. Whenever it is clear what d is considered, we write simply f
m(n).
Lemma 6.6. If m ≥ 2, n ≥ 0 and d ≥ 2, then f
m(n) =
[n/d]
X
j=0
f
m−1(j).
P r o o f. Fix n ∈ N. Let ε
0ε
1ε
2. . . be a representation of n in base d with digits 0, 1, . . . , d
m− 1. We consider an integer j ≤ [n/d] and a se- quence ε
00ε
01ε
02. . . which is the representation of j in base d with digits 0, 1, . . . , d
m−1− 1 defined as ε
0i:= [ε
i/d]. The correspondence ε
0ε
1ε
2. . . 7→
(j, ε
00ε
01ε
02. . .) is one-to-one, because the ε
00i:= ε
i− dε
0iare just the digits of the canonical representation of n − dj in base d.
Proposition 6.7. If q = d
m, then f
m(n) = d
−m(m−1)/2(m − 1)! n
m−1+ O(n
m−2), n → ∞, and
H(t) ≡ d
−m(m−1)/2m! = q
−(m−1)/2m! .
P r o o f. Note first that f
2(n) = n/d + O(1) (a straightforward conse- quence of Lemma 6.6). Moreover, by induction and the above lemma, we have, for m ≥ 3,
f
m(n) =
[n/d]
X
j=0
d
−(m−1)(m−2)/2(m − 2)! j
m−2+ O(j
m−3)
= d
−(m−1)(m−2)/2(m − 1)!
n d
m−1+ O(n
m−2).
This clearly implies the first part of the proposition. We deduce the second one from the relation
Φ(x) = X
n<x
f
m(n) = d
−m(m−1)/2m! x
m+ O(x
m−1) and (2.1).
Let f
u∗(n) denote the number of representations of a nonnegative integer
n in fixed base d with digits 0, 1, . . . , u − 1.
Proposition 6.8. The following relation holds:
f
du∗(n) =
[n/d]
X
j=0
f
u∗(j).
The exact order of f
du∗(n) is n
δwith δ = log
du.
P r o o f. We obtain this relation in the same way as in Lemma 6.6. Thence the function f
du∗(n) itself is the summation function Φ
∗u, and Theorem 2.1 yields its precise asymptotic order.
Examples. 1. For any d, f
2(d)(dn) = n + 1, f
3(d)(d
2n) = (dn + 2)(n + 1)/2, etc.
2. If d = 2, then f
3(4n) = (n + 1)
2, f
3(4n + 2) = (n + 1)(n + 2); f
4(8n) = P
nk=0
(2k + 1)
2= (n + 1)(2n + 1)(2n + 3)/3, f
4(8n + 2) = f
4(8n) + (n + 1)
2, f
4(8n+4) = f
4(8n+2)+(n+1)(n+2), f
4(8n+6) = f
4(8n+4)+(n+1)(n+2), f
4(8n + 8) = f
4(8n + 6) + (n + 2)
2, etc.
Conclusion. Thus, we can reduce a radix case with d dividing q to the case when q > d and q not divisible by d. In turn, for the latter case it suffices to find the best possible estimate for f in order to solve the summation problem completely. In the present paper we will study some basic examples and plan to return to this problem in subsequent work.
For d dividing q, as shown in Proposition 6.8, the maximal order of f is exactly O(n
β). It is curious that such a situation can arise even if d does not divide q.
6.5. The precise maximal order of f in the case q ≤ d
2. We associate with any integer n the integer n
0= [n/d]. Let r := [(q − 1)/d], and s :=
d{(q − 1)/d}.
Proposition 6.9. Let q ≤ d
2. We have two cases.
(1) f (n) = O(n
logd(r+1)) for r ≤ s.
(2) f (n) = O n
logdr+√
r2+4s+4 2
for r > s.
Both maximal orders are precise.
P r o o f. We consider both cases simultaneously. From (6.2) we have
f (n) =
X
r−1i=0
f (n
0− i), n ∈ {s + 1, . . . , d − 1} mod d, X
ri=0
f (n
0− i), n ∈ {0, . . . , s} mod d.
Let I
kbe the interval {d
k−d, . . . , d
k+1−1}, k ≥ 0. We denote by X
kthe
maximum of f (n) for n ∈ I
kand n ∈ {0, . . . , s} mod d, by Y
kits maximum
for n ∈ I
kand n ∈ {s + 1, . . . , d − 1} mod d. The hypothesis q ≤ d
2implies r ≤ d − 1 and hence, given n ∈ I
k, the integers n
0− i belong to I
k−1for any i ∈ {0, . . . , r}, and at most s + 1 of them belong to {0, . . . , s} mod d (or at most r + 1 if r ≤ s). For any k ≥ 1, we obtain
X
k≤ (s + 1)X
k−1+ (r − s)Y
k−1if r ≥ s + 1, Y
k≤ (s + 1)X
k−1+ (r − s − 1)Y
k−1if r ≥ s + 1, X
k≤ (r + 1)X
k−1if r ≤ s.
Now we suppose r ≥ s + 1 (hence r ≥ 1 and s ≤ d − 2). We consider the integers
n
k=
sd + rd
2+ . . . + sd
k−1+ rd
k, k even, rd + sd
2+ . . . + sd
k−1+ rd
k, k odd,
m
k=
n
k+ r, k even, n
k− 1, k odd, with n
0= 0. They satisfy the relations
f (n
k) = (s + 1)f (n
k−1) + (r − s)f (m
k−1), f (m
k) = (s + 1)f (n
k−1) + (r − s − 1)f (m
k−1),
whence we obtain by induction X
k≤ f (n
k) and Y
k≤ f (m
k) for any k ≥ 0.
Actually, we have equality (as n
kbelongs to I
kand to {0, . . . , s} mod d, we cannot have f (n
k) > X
k).
From the first relation we obtain the expression for f (m
k−1) in terms of f (n
k) and f (n
k−1) and, bringing it into the second relation, we deduce that for k ≥ 1,
f (n
k+1) = rf (n
k) + (s + 1)f (n
k−1) with f (n
0) = 1 and f (n
1) = r + 1, hence
f (n
k)
r + √
r
2+ 4s + 4 2
k. As k = log
dn + O(1) for n ∈ I
k, we see that
f (n) = O n
logdr+√
r2+4s+4
2
is the precise maximal order of f (n).
For r ≤ s, we obtain in the same way f (n) = O(n
logd(r+1)).
Remark. It is possible to show that the estimate f (n) = O(n
β) can be
improved for any d and q such that d{(q − 1)/d} < [(q − 1)/d]. We restrict
ourselves to mentioning a better estimate for the case d = 2 and arbitrary
odd q ≥ 3. In the same way as before it may be shown that X
k≤
q + 3 4
X
k−1+
q + 1 4
· q − 1 2
X
k−2, whence f (n) = O(n
βe), where
β = log e
2 q+34
+ q
q+34
2+ 2(q − 1)
q+14
2 < β.
Note that for q = 5 this estimate yields f (n) = O(n
log2(√3+1)) = O(n
1.450), while below we will show that the best possible estimate is approximately O(n
1.344) (see Proposition 6.16).
As a corollary of the previous proposition, we obtain the corresponding result for some natural case.
Corollary 6.10. For d = 2, q = 3 the precise maximal order of f is f (n) = O n
log2√5+1 2