LXXXVIII.2 (1999)
On the discrepancy estimate of normal numbers
by
M. B. Levin (Tel-Aviv)
Dedicated to Professor N. M. Korobov on the occasion of his 80th birthday 1. Introduction
1.1. A number α ∈ (0, 1) is said to be normal to base q if in the q-ary expansion of α, α = .d
1d
2. . . (d
i∈ ∆ = {0, 1, . . . , q − 1}, i = 1, 2, . . .), each fixed finite block of digits of length k appears with an asymptotic frequency of q
−kalong the sequence (d
i)
i≥1. Normal numbers were introduced by Borel (1909).
1.1.1. Let (x
n)
n≥1be an arbitrary sequence of real numbers. The quan- tity
(1) D(N ) = D(N, (x
n)
n≥1) = sup
γ∈(0,1]
|#{0 ≤ n < N | {x
n} < γ}/N − γ|
is called the discrepancy of (x
n)
Nn=1, where {x} = x − [x] is the fractional part of x. The sequence {x
n}
n≥1is said to be uniformly distributed (u.d.) in [0, 1) if D(N ) → 0.
1.1.2. It is known that a number α is normal to base q if and only if the sequence {αq
n}
n≥0is u.d. (Wall, 1949). Borel proved that almost every number (in the sense of Lebesgue measure) is normal to base q. In [G], Gal and Gal proved that
D(N, {αq
n}
n≥0) = O((N
−1log log N )
1/2) for a.e. α.
1.2. In [K1] Korobov posed the problem of finding a function ψ with maximum decay, such that
∃α : D(N, {αq
n}
n≥0) ≤ ψ(N ), N = 1, 2, . . .
1991 Mathematics Subject Classification: 11K16, 11K38.
Work supported in part by the Israel Science Foundation Grant No. 0366-172.
[99]
He showed that ψ(N ) = O(N
−1/2) (see [K1]). The lower bound of the discrepancy for the Champernowne and Davenport–Erd˝os normal numbers was found by Schiffer [S]:
D(N, {αq
n}
n≥1) ≥ K/ log N with K > 0, N = 2, 3, . . . For a bibliography on Korobov’s problem see [Po, L1].
1.3. In [L2] we proposed using small discrepancy sequences (van der Corput type sequences and {nα}
n≥0) to construct normal numbers, and announced that
ψ(N ) = O(N
−1log
2N ).
This result is proved below. The estimate of ψ(N ) was previously known to be O(N
−2/3log
4/3N ) (Korobov [K2] for q prime, and Levin [L1] for arbitrary integer q). We note that the estimate obtained cannot be improved essentially, since according to W. Schmidt, 1972 (see [N, p. 24]), for any sequence of reals,
N →∞
lim N D(N )/ log N > 0.
1.4. Let x = [a
0(x); a
1(x), a
2(x), . . .] be the continued fraction expansion of x, with partial quotients a
i(x). For an integer b and Q > 1 let P
a
i(b/Q) denote the sum of all partial quotients of b/Q. Following [P] we prove (see Lemma 3) that there exists an integer sequence b
mand a constant K > 0 with
(2)
X
m r=1X a
i({b
m/q
r}) ≤ Km
3, m = 1, 2, . . .
Theorem 1. Let
(3) α = X
m≥1
1 q
nmX
0≤k<qm
b
mk q
m1 q
mkwhere b
msatisfy (2),
(4) n
1= 0 and n
k= X
1≤r<k
rq
r, k = 2, 3, . . . Then the number α is normal to base q, and
D(N, {αq
n}
n≥0) = O(N
−1log
3N ).
1.5. Let (p
0i,j)
i,j≥1be Pascal’s triangle:
p
0i,1= p
01,i= 1, i = 1, 2, . . . , p
0i,j= p
0i,j−1+ p
0i−1,j, i, j = 2, 3, . . . , and (p
i,j)
i,j≥1be Pascal’s triangle mod 2:
(5) p
i,j≡ p
0i,jmod 2, i, j = 1, 2, . . .
Every integer n ≥ 0 has a unique digit expansion in base q,
(6) n = X
j≥1
e
j(n)q
j−1with e
j(n) ∈ ∆ = {0, . . . , q − 1}, j = 1, 2, . . . , and e
j(n) = 0 for all sufficiently large j.
Theorem 2. Let
(7) α = X
m≥1
1 q
nmX
0≤n<q2m
1 q
n2m2m
X
i=1
d
i(n) q
iwhere
(8) d
i(n) ≡ X
j≥1
p
i,je
j(n) mod q, d
i(n) ∈ ∆, i = 1, . . . , 2
m, n ∈ [0, q
2m),
(9) n
1= 0 and n
m= X
1≤r<m
2
rq
2r, m = 2, 3, . . . Then the number α is normal to base q and
D(N, {αq
n}
n≥0) = O(N
−1log
2N ).
Remark 1. We use here the sequence of 2
m× 2
mmatrices of Pascal’s triangle mod 2. A similar result is valid for the sequence of m × m matrices of Pascal’s triangle (or m × m matrices of Pascal’s triangle mod p) but with D(N, {αq
n}
n≥0) = O(N
−1log
3N ), where α is denoted by a concatenation of blocks ω
m:
α = .ω
1. . . ω
m. . . , where
ω
m= (d
1(1) . . . d
m(1) . . . d
1(q
m) . . . d
m(q
m)), m = 1, 2, . . . , and
d
i(n) ≡ X
j≥1
p
i,je
j(n) mod q.
Remark 2. Let (σ
i)
i≥1be any sequence of substitutions of the set ∆ = {0, 1, . . . , q − 1}. The proof of Theorem 2 does not change if in (8) we use the functions σ
i(e
i(n)) instead of the functions e
i(n) (see [B], [N, p. 25]).
2. Proof of the theorems. Let m ≥ 1, b, i be integers, 0 ≤ i < m, (b, q) = 1,
α
m= α
m(b) = X
0≤k<qm
bk q
m1 q
mk, (10)
α
mni= [q
2m−i{α
mq
i+mn}]/q
2m−i.
(11)
It is easy to see that {{bn/q
m}q
i} = {bn/q
m−i}, and {α
mq
i+mn} =
bn q
mq
i+
b(n + 1) q
m1 q
m−i+
b(n + 2) q
m1
q
2m−i+ . . .
=
bn q
m−i+
b(n + 1) q
m1 q
m−i+
b(n + 2) q
m1
q
2m−i+ . . . Therefore
(12) α
mni=
bn q
m−i+
b(n + 1) q
m1 q
m−i. Let N ∈ [1, mq
m] be an integer, γ ∈ (0, 1],
(13) A(γ, N, (x
n)) =
#{0 ≤ n < N | {x
n} < γ} for γ > 0,
0 for γ ≤ 0,
and
(14) A(γ, Q, P, (x
n)) = #{Q ≤ n < Q + P | {x
n} < γ}.
Hence and from (10) we obtain
A(γ, N, {α
mq
n}
n≥0) = A(γ, m[N/m], {α
mq
n}
n≥0) (15)
+ A(γ, m[N/m], N − m[N/m], {α
mq
n}
n≥0)
=
m−1
X
i=0
A(γ, [N/m], {α
mq
i+mn}
n≥0) + θm with θ ∈ [0, 1].
Let c = [q
mγ], N
1∈ [1, q
m] and 0 ≤ i < m. From (11) and (13) we deduce
A
c − 1
q
m, N
1, (α
mni)
n≥0≤ A(γ, N
1, {α
mq
mn+i}
n≥0) (16)
≤ A
c + 1
q
m, N
1, (α
mni)
n≥0.
Lemma 1. Let N ∈ [1, mq
m] be an integer , γ ∈ (0, 1], (b, q) = 1. Then (17) A(γ, N, {α
mq
n}
n≥0)
= γN + ε
14m + 3
X
m i=11≤N ≤q
max
iN D(N, {bn/q
i}
n≥0)
, (18) A(γ, mq
m, {α
mq
n}
n≥0) = γmq
m+ 3ε
2m
with |ε
j| < 1, j = 1, 2.
P r o o f. Let 0 ≤ i < m, d, d
1, and d
2be integers, d = d
1q
i+ d
2, d
1∈ [0, q
m−i), d
2∈ [0, q
i). By (12) and (13) we get
A
d
q
m, N
1, (α
mni)
n≥0= #
0 ≤ n < N
1bn q
m−i+
b(n + 1) q
m1
q
m−i< d
1q
m−i+ 1 q
m−i· d
2q
i. Consequently,
(19) A(d/q
m, N
1, (α
mni)
n≥0) = T
1(N
1) + T
2(N
1), where
T
1(N ) = #
0 ≤ n < N
bn q
m−i< d
1q
m−i, (20)
T
2(N ) = #
0 ≤ n < N
bn q
m−i= d
1q
m−iand
b(n + 1) q
m< d
2q
i. (21)
Let N
1= N
2q
m−i+ N
3with N
3∈ [0, q
m−i) and N
2∈ [0, q
i). It is easy to see that
T
1(N
1) = T
1(q
m−iN
2) + T
1(N
3).
We see from (20) and (1) that
(22) T
1(N
2q
m−i) = N
2d
1, and
T
1(N
3) = d
1q
m−iN
3+ εN
3D
N
3,
bn q
m−in≥0
with |ε| ≤ 1.
This yields
(23) T
1(N
1) = d
1q
m−iN
1+ ε max
1≤N <qm−i
N D
N,
bn q
m−in≥0
with |ε| ≤ 1.
Now we compute T
2(N ). Let d
0be an integer, d
0≡ d
1b
−1mod q
m−iwith d
0∈ [0, q
m−i), and
Y = {0 ≤ n < N
1| {bn/q
m−i} = d
1/q
m−i}.
Clearly if {bn/q
m−i} = d
1/q
m−i, then bn ≡ d
1mod q
m−i, n ≡ d
0mod q
m−i, and
(24) Y = {d
0+ rq
m−i| 0 ≤ r < N
4} with N
4=
N
1− d
0− 1 q
m−i+ 1.
Combining (21) and (1) we obtain
T
2(N
1) = #
n ∈ Y
b(n + 1) q
m< d
2q
i(25)
= #
0 ≤ r < N
4b(d
0+ 1) q
m+ br
q
i< d
2q
i= N
4d
2q
i+ ε
2N
4D
N
4,
bn q
i+ θ
n≥0
with θ = b(d
0+ 1)/q
m, |ε
2| ≤ 1.
It follows from (1) that for every real θ,
(26) D(N, {x
n+ θ}
n≥0) ≤ 2D(N, {x
n}
n≥0).
By (24) and (25), this yields T
2(N
1) =
N
1+ q
m−i− d
0− 1 q
m−id
2q
i+ 2ε
2max
1≤N ≤qi
N D(N, {bn/q
i}
n≥0) (27)
= N
1d
2q
m+ ε
3+ 2ε
2max
1≤N ≤qi
N D(N, {bn/q
i}
n≥0) with |ε
j| ≤ 1, j = 2, 3.
If N
1= q
m, then N
4= q
i, and N
4D(N
4, {bn/q
i}
n≥0) = 1. Hence and from (25) and (26) we obtain
(28) T
2(q
m) = d
2+ 2ε
4with |ε
4| ≤ 1.
Substituting (23) and (27) into (19), we obtain A(d/q
m, N
1, (α
mni)
n≥0)
= N
1d/q
m+ ε
5(1 + max
1≤N <qm−i
N D(N, {bn/q
m−i}
n≥0) + 2 max
1≤N ≤qi
N D(N, {bn/q
i}
n≥0)) with |ε
5| ≤ 1.
Using (16) and (15) we get A(γ, N
1, {α
mq
mn+i}
n≥0)
= γN
1+ ε
6(2 + max
1≤N <qm−i
N D(N, {bn/q
m−i}
n≥0) + 2 max
1≤N ≤qi
N D(N, {bn/q
i}
n≥0)) with |ε
6| ≤ 1, and
A(γ, N, {α
mq
n}
n≥0) = θm + X
m i=1γ[N/m]
+ ε
72m + 3 X
mi=1
1≤N ≤q
max
iN D(N, {bn/q
i}
n≥0)
= γN + ε
84m + 3 X
m i=11≤N ≤q
max
iN D(N, {bn/q
i}
n≥0)
,
where |ε
j| ≤ 1, j = 7, 8. Assertion (17) is proved. Assertion (18) follows analogously from (22) and (28).
Lemma 2. Let j ≥ 1, 1 ≤ N ≤ q
j, (b, q) = 1, and a
i(x) be partial quotients of {x}. Then
N D(N, {bn/q
j}
n≥0) ≤ X
a
i(b/q
j).
For the proof of this well-known theorem, see for example [N, p. 26].
Lemma 3. There exists a constant K > 0 and integers c
m∈ [0, q
m) such that
X
m r=1X a
i({c
m/q
r}) ≤ Km
3, m = 1, 2, . . .
P r o o f. According to [P, p. 2144] there exist constants K
qsuch that X
1≤c≤qr, (c,q)=1
X a
i(c/q
r) ≤ K
qq
rr
2, r = 1, 2, . . .
Therefore
(29) X
1≤c≤qm, (c,q)=1
X
m r=1X a
i({c/q
r})
= X
m r=1q
m−rX
1≤c≤qr, (c,q)=1
X a
i(c/q
r) ≤ X
m r=1q
mK
qr
2≤ K
qq
mm
3. Let φ(q
m) = #{1 ≤ c ≤ q
m| (c, q) = 1} and K = K
qq/φ(q). It is known that φ(q
m) = q
m−1φ(q). Now the assertion of Lemma 3 follows from (29).
Corollary. Let 1 ≤ N ≤ mq
m. Then
(30) A(γ, N, {α
m(b
m)q
n}
n≥0) = γN + O(m
3).
The statement follows from (1), (2), (10), and Lemmas 1–3.
Applying (3) and (10) we get
{αq
nm+n} = {α
m(b
m)q
n} + θq
n−mqmwith 0 < θ < 1 and 0 ≤ n < mq
m. Hence and from (13) we have, for N ∈ [1, mq
m],
A(γ − 1/q
m, N − m, {α
m(b
m)q
n}
n≥0) ≤ A(γ, N, {αq
nm+n}
n≥0)
≤ A(γ, N, {α
m(b
m)q
n}
n≥0).
By using (30) and (14), we obtain
(31) A(γ, n
m, N, {αq
n}
n≥0) = γN + O(m
3) with 1 ≤ N ≤ mq
m. Similarly, from (18) we deduce that
(32) A(γ, n
m, mq
m, {αq
n}
n≥0) = γmq
m+ O(m).
End of the proof of Theorem 1. For every N ≥ 1 there exists an integer k such that N ∈ [n
k, n
k+1). By (4) this yields
(33) N = n
k+ R with 0 ≤ R < kq
k, N > (k − 1)q
k−1, k ≤ 2 log
qN.
Applying (4), (14) and (31)–(33) we obtain A(γ, N, {αq
n}
n≥0) =
k−1
X
r=1
A(γ, n
r, rq
r, {αq
n}
n≥0) + A(γ, n
k, R, {αq
n}
n≥0)
=
k−1
X
r=1
(γrq
r+ O(r)) + γR + O(k
3)
= γN + O(k
3) = γN + O(log
3N ).
Thus, by (1), the theorem is proved.
Proof of Theorem 2. In [So] Sobol’ proposed the use of Pascal’s triangle mod 2 to construct small discrepancy sequences (see also [F], [N]). Here we use Pascal’s triangle mod 2 to construct normal numbers.
Let P
nbe a sequence of a 2
n× 2
nmatrices such that P
1=
1 1 1 0
, . . . , P
n+1=
P
nP
nP
n0
, . . .
It is easy to prove by induction that P
nis the 2
n×2
nupper left-hand corner of Pascal’s triangle (5), and P
nis a triangular-type matrix. The following lemma is proved in [BH] for Pascal’s triangle, and it is clearly valid also for Pascal’s triangle mod 2.
Lemma 4. The determinant of any n × n array taken with its first row along a row of ones, or with its first column along a column of ones in Pascal’s triangle, written in rectangular form, is one.
From (7) we have
(34) {αq
nm+2mn+k} = .d
k+1(n)d
k+2(n) . . . d
2m(n)d
1(n + 1) . . . Let 1 ≤ k, i ≤ 2
mand
(35) α
ki(n) = [{αq
nm+2mn+k}q
i]/q
i. It is easy to see that
(36) α
ki(n)
=
.d
k+1(n) . . . d
k+i(n) if k + i ≤ 2
m, .d
k+1(n) . . . d
2m(n)d
1(n + 1) . . . d
k+i−2m(n + 1) otherwise.
Lemma 5. Let m, k, i, B, f be integers, 1 ≤ i, k ≤ 2
m, B ∈ [0, q
2m−i), f ∈ [0, q
i). Then
A(f /q
i, Bq
i, q
i, (α
ki(n))
n≥0) = f + 2ε with |ε| < 1.
P r o o f. Case 1. Let k + i ≤ 2
m, c
j∈ ∆ = {0, 1, . . . , q − 1} (j = 1, . . . , i).
We examine the system of equations
(37) d
k+j(n) = c
j, j = 1, . . . , i, n ∈ [Bq
i, (B + 1)q
i).
According to (8) this system is equivalent to the system of i congruences X
1≤ν≤2m
p
k+j,νe
ν(n + Bq
i) ≡ c
jmod q, j = 1, . . . , i, n ∈ [0, q
i).
Applying (6) we see that e
ν(n + Bq
i) = e
ν(n) + e
ν(Bq
i), ν = 1, 2, . . . , and
(38) X
1≤ν≤i
p
k+j,νe
ν(n)
≡ c
j− X
i<ν≤2m
p
k+j,νe
ν(Bq
i) mod q, j = 1, . . . , i, with n ∈ [0, q
i). It follows from Lemma 4 that
(39) |det(p
k+j,ν)
1≤j,ν≤i| = 1.
For any c
1, . . . , c
ithe system (38) has a unique solution (e
1(n), . . . , e
i(n)), and consequently there exists a unique n
0∈ [Bq
i, (B + 1)q
i) satisfying (37).
From (36) and (37) we see that the set {α
ki(n) | n ∈ [Bq
i, (B + 1)q
i)}
coincides with {j/q
i| j ∈ [0, q
i)}. Hence and from (14) we have (40) A(f /q
i, Bq
i, q
i, (α
ki(n))
n≥0) = f.
Case 2. Let k + i > 2
m, l
1= 2
m− k. As in (37) and (38), the system of equations
d
k+j(n) = c
j, j = 1, . . . , l
1, (41)
d
j(n + 1) = c
j+l1, j = 1, . . . , i − l
1, (42)
with n ∈ [Bq
i, (B + 1)q
i), is equivalent to the systems of congruences
(43) X
1≤ν≤i
p
k+j,νe
ν(n)
≡ c
j− X
i<ν≤2m
p
k+j,νe
ν(Bq
i) mod q, j = 1, . . . , l
1,
(44) X
1≤ν≤i
p
j,νe
ν(n + 1)
≡ c
j+l1− X
i<ν≤2m+1
p
j,νe
ν((B + [(n + 1)/q
i])q
i) mod q, where j = 1, . . . , i − l
1and n ∈ [0, q
i).
Let n = n
1+ n
2q
l1with n
1∈ [0, q
l1) and n
2∈ [0, q
i−l1). It is evident that e
ν(n) = e
ν(n
1) for ν = 1, . . . , l
1.
The matrix P
mis triangular. Hence
p
k+j,ν= 0 with ν > 2
m− k − j = l
1− j.
The system (43) is equivalent to the following system of congruences:
(45) X
1≤ν≤l1
p
k+j,νe
ν(n
1)
≡ c
j− X
i<ν≤2m
p
k+j,νe
ν(Bq
i) mod q, j = 1, . . . , l
1, where n
1∈ [0, q
l1) and n
2∈ [0, q
i−l1).
Applying (39) with i = l
1shows that this system has a unique solution with (e
1(n
1), . . . , e
l1(n
1)). Consequently, there exists a unique solution n
1= n
01∈ [0, q
l1) satisfying (45).
By (41) and (43) we obtain
(46) {(d
k+1(n
1+n
2q
l1+Bq
i), . . . , d
k+l1(n
1+n
2q
l1+Bq
i)) | 0 ≤ n
1< q
l1}
= {(c
1, . . . , c
l1) | c
j∈ ∆, j = 1, . . . , l
1}.
Now we examine the system (44) with n
1= n
01the solution of (45).
Case 2.1. Let n
01≤ q
l1− 2. Bearing in mind that
e
ν(n + 1) = e
ν(n
01+ 1 + q
l1n
2) = e
ν(n
01+ 1) + e
ν(q
l1n
2), we deduce from (44) that
X
l1<ν≤i
p
j,νe
ν(q
l1n
2)
≡ c
j+l1− X
1≤ν≤l1
p
j,νe
ν(n
01+ 1) − X
i<ν≤2m
p
j,νe
ν(Bq
i) mod q with j = 1, . . . , i − l
1and 0 ≤ n
2< q
i−l1.
Applying Lemma 4 we obtain a unique solution for this system with (e
l1+1(q
l1n
2), . . . , e
i(q
l1n
2)).
By (42) and (44) we get
(47) {(d
1(n
01+ n
2q
l1+ Bq
i+ 1), . . . , d
i−l1(n
01+ n
2q
l1+ Bq
i+ 1)) | 0 ≤ n
2< q
i−l1} = {(c
l1+1, . . . , c
i) | c
l1+j∈ ∆, j = 1, . . . , i − l
1}.
Let
(48) F = {d
k+1(n) . . . d
2m(n)d
1(n + 1) . . . d
k+i−2m(n + 1) |
0 ≤ n
1< q
l1− 1, 0 ≤ n
2< q
i−l1, n = n
1+ n
2q
l1+ Bq
i}, and
(49) g
ν= d
k+ν(q
l1− 1 + Bq
i), ν = 1, . . . , l
1. From (46) and (47) we have
(50) F = {(c
1, . . . , c
i) | c
j∈ ∆, j = 1, . . . , i, (c
1, . . . , c
l1) 6= (g
1, . . . , g
l1)}
and
(51) #F = q
i− q
i−l1.
Case 2.2. Let n
01= q
l1− 1, n
2∈ [0, q
i−l1− 2] and n = n
01+ n
2q
l1. Then
e
ν(n
01+ 1) = 0 for 1 ≤ ν ≤ l
1and e
ν(n + 1) = e
ν((n
2+ 1)q
l1) for l
1< ν ≤ i.
The system (44) is equivalent to the following system of congruences:
(52) X
l1<ν≤i
p
j,νe
ν((n
2+ 1)q
l1)
≡ c
j+l1− X
i<ν≤2m
p
j,νe
ν(Bq
i) mod q, j = 1, . . . , i − l
1, with 0 ≤ n
2≤ q
i−l1− 2.
For n
2∈ [0, q
i−l1− 2] we have the q
i−l1− 1 distinct vectors of (e
l1+1((n
2+ 1)q
l1), . . . , e
i((n
2+ 1)q
l1)).
Using Lemma 4 and by (52) we obtain for n
2∈ [0, q
i−l1− 2] the q
i−l1− 1 distinct vectors of (c
l1+1, . . . , c
i).
Let
G = {(g
1, . . . , g
l1, d
1((n
2+ 1)q
l1+ Bq
i), . . . , d
i−l1((n
2+ 1)q
l1+ Bq
i)) | 0 ≤ n
2≤ q
i−l1− 2}.
From (42), (44) and (52) we find that #G = q
i−l1− 1, and from (46) and (48)–(51) that #(F ∪ G) = q
i− 1. Hence and from (36) the set {α
ki(n) | n ∈ [Bq
i, (B + 1)q
i− 2]} coincides with q
i− 1 distinct values of j/q
iwith j ∈ [0, q
i). By (14) we get
A(f /q
i, Bq
i, q
i, (α
ki(n))
n≥0) = f + 2ε with |ε| < 1.
Hence and from (40) we have the assertion of Lemma 5.
Corollary 1.
(53) A(γ, Bq
i, q
i, {αq
nm+2mn+k}
n≥0) = γq
i+ 4ε with |ε| < 1.
P r o o f. Analogously to (16), from (14) and (35) we have A
f − 1
q
i, Bq
i, q
i, (α
ki(n))
n≥0≤ A(γ, Bq
i, q
i, {αq
nm+2mn+k}
n≥0)
≤ A((f + 1)/q
i, Bq
i, q
i, (α
ki(n))
n≥0) with f = [γq
i]. By using Lemma 5 we obtain (53).
Corollary 2. Let 1 ≤ N < 2
mq
2m. Then
A(γ, n
m, N, {αq
n}
n≥0) = γN + 5qε2
2mwith |ε| < 1, (54)
A(γ, n
m, 2
mq
2m, {αq
n}
n≥0) = γ2
mq
2m+ 5ε2
mwith |ε| < 1.
(55)
P r o o f. Let N
0= [N/2
m], N
00= N − 2
mN
0, N
0= P
2m−1i=0
b
iq
iwith b
i∈ ∆,
(56) N
0= 0, N
j= X
j−1i=0
b
2m−iq
2m−i, j = 1, 2, . . . , B
i= N
2m−i−1/q
i.
It is evident that B
i(i = 1, 2, . . .) are integers, and N
00∈ [0, 2
m). As in (15) we see from (14) that
A(γ, n
m, N, {αq
n}
n≥0) = εN
2+
2m
X
k=1
A(γ, N
0, {αq
nm+2mn+k}
n≥0), and
A(γ, N
0, {αq
nm+2mn+k}
n≥0)
=
2m
X
i=1
A(γ, N
i−1, b
2m−iq
2m−i, {αq
nm+2mn+k}
n≥0)
=
2
X
m−1 i=0A(γ, N
2m−i−1, b
iq
i, {αq
nm+2mn+k}
n≥0)
=
2
X
m−1 i=0b
X
i−1 B=0A(γ, N
2m−i−1+ Bq
i, q
i, {αq
nm+2mn+k}
n≥0).
Using (56) we have A(γ, n
m, N, {αq
n}
n≥0)
= ε2
m+
2m
X
k=1 2
X
m−1i=0 b
X
i−1 B=0A(γ, (B
i+ B)q
i, q
i, {αq
nm+2mn+k}
n≥0).
Applying (53) we obtain
A(γ, n
m, N, {αq
n}
n≥0) = ε2
m+
2m
X
k=1 2
X
m−1i=0 b
X
i−1 B=0(γq
i+ 4ε
i) = γN + 5qε
12
2mwith |ε
1| ≤ 1.
Assertion (54) is proved. We prove (55) analogously.
End of the proof of Theorem 2. For every N ≥ q there exists an integer k such that N ∈ [n
k, n
k+1). By (9), this yields N = n
k+R with 0 ≤ R < 2
kq
2k, N ≥ 2
(k−1)q
2k−1, 2
k≤ 2 log
qN. Applying (9), (13), (14), (54) and (55) we obtain
A(γ, N, {αq
n}
n≥0) =
k−1
X
m=1
A(γ, n
m, 2
mq
2m, {αq
n}
n≥0) + A(γ, n
k, R, {αq
n}
n≥0)
=
k−1
X
m=1