LXXIV.4 (1996)
Ergodic properties of generalized L¨ uroth series
by
Jose Barrionuevo (Mobile, Ala.), Robert M. Burton (Corvallis, Ore.),
Karma Dajani (Utrecht) and Cor Kraaikamp (Delft)
1. Introduction
1.1. L¨uroth series. In 1883 J. L¨ uroth [Lu] introduced and studied the following series expansion, which can be viewed as a generalization of the decimal expansion. Let x ∈ (0, 1]. Then
(1) x = 1
a
1+ 1
a
1(a
1− 1)a
2+ . . . + 1
a
1(a
1− 1) . . . a
n−1(a
n−1− 1)a
n+ . . . , where a
n≥ 2, n ≥ 1. L¨ uroth showed, among other things, that each irra- tional number has a unique infinite expansion (1), and that each rational number has either a finite or a periodic expansion.
Underlying the series expansion (1) is the operator T
L: [0, 1] → [0, 1], defined by
(2) T
Lx :=
1 x
1 x
+ 1
x −
1 x
, x 6= 0; T 0 := 0
(see also Figure 1), where bξc denotes the greatest integer not exceeding ξ.
For x ∈ [0, 1] we define a(x) := b1/xc + 1, x 6= 0; a(0) := ∞ and a
n(x) = a(T
n−1x) for n ≥ 1. From (2) it follows that T
Lx = a
1(a
1− 1)x − (a
1− 1), and therefore
x = 1 a
1+ 1
a
1(a
1− 1) T
Lx
= 1
a
1+ 1
a
1(a
1− 1)a
2+ . . . + T
Lnx
a
1(a
1− 1) . . . a
n(a
n− 1) .
Since a
n≥ 2 and 0 ≤ T
Lnx ≤ 1 one easily sees that the infinite series from (1) converges to x.
1991 Mathematics Subject Classification: 28D05, 11K55.
[311]
Recently S. Kalpazidou, A. and J. Knopfmacher introduced and studied in [K
31,2] the so-called alternating L¨uroth series. For each x ∈ (0, 1] one has
x = 1
a
1− 1 − 1
a
1(a
1− 1)(a
2− 1) + . . .
+ (−1)
n+1a
1(a
1− 1) . . . a
n−1(a
n−1− 1)(a
n− 1) + . . . , where a
n≥ 2, n ≥ 1. Dynamically the alternating series expansion is gen- erated by the operator S
A: [0, 1] → [0, 1] defined by
(3) S
Ax := 1 +
1 x
−
1 x
1 x
+ 1
x, x 6= 0, and S
A0 := 0, i.e. S
Ax = 1 − T
Lx (see also Figure 1).
Fig. 1
L¨ uroth series have been extensively studied; for further reference we mention here papers by H. Jager and C. de Vroedt [JdV], T. ˇ Sal´at [Sa], and books by J. Galambos [G], O. Perron [Pe] and W. Vervaat [V].
In [JdV] it was shown that the stochastic variables a
1(x), a
2(x), . . . are
independent with λ(a
n= k) = 1/(k(k − 1)) for k ≥ 2, and that T
Lis
measure preserving with respect to Lebesgue measure (
1) and ergodic (
2).
Using similar techniques analogous results were obtained in [K
31,2] for the operator S
Afrom (3). In fact, much stronger results can be obtained easily, not only in the case of the (alternating) L¨ uroth series, but also in a more general setting.
1.2. Generalized L¨uroth series. Let I
n= (l
n, r
n], n ∈ D ⊂ N = {0, 1, 2, . . .}, be a finite or infinite collection of disjoint intervals of length L
n:= r
n− l
n, such that
(4) X
n∈D
L
n= 1 and
(5) 0 < L
i≤ L
j< 1 for all i, j ∈ D, i > j.
The set D is called the digit set. Usually such a digit set is either a finite or infinite set of consecutive positive integers, see also the examples at the end of this section.
Furthermore, let I
∞:= [0, 1] \ S
n∈D
I
n, L
∞:= 0 and define the maps T, S : [0, 1] → [0, 1] by
T x :=
( x − l
nr
n− l
n, x ∈ I
n, n ∈ D, 0, x ∈ I
∞, Sx :=
( r
n− x r
n− l
n, x ∈ I
n, n ∈ D, 0, x ∈ I
∞. Define for x ∈ Ω := [0, 1] \ I
∞= S
n∈D
I
n, s(x) := 1
r
n− l
nand h(x) := l
nr
n− l
n, in case x ∈ I
n, n ∈ D, s
n= s
n(x) :=
s(T
n−1x), T
n−1x 6∈ I
∞,
∞, T
n−1x ∈ I
∞, and
h
n= h
n(x) :=
h(T
n−1x), T
n−1x 6∈ I
∞, 1, T
n−1x ∈ I
∞. For x ∈ (0, 1) such that T
n−1x 6∈ I
∞, one has
T x = s(x)x − h(x) = s
1x − h
1.
(
1) That is, λ(T
L−1(A)) = λ(A) for every Lebesgue measurable set A.
(
2) That is, λ(A M T
L−1A) = 0 ⇒ λ(A) ∈ {0, 1}.
Inductively we find x = h
1s
1+ 1
s
1T x = h
1s
1+ 1
s
1h
2s
2+ 1
s
2T
2x
= . . . (6)
= h
1s
1+ h
2s
1s
2+ . . . + h
ns
1. . . s
n+ 1
s
1. . . s
nT
nx.
Since Sx = 1 − T x = −s
1x + 1 + h
1, for x ∈ Ω, one finds
(7) x = h
1+ 1
s
1− Sx s
1.
Continuing in this way we obtain an alternating series expansion (see also [K
31,2]). Figure 1 illustrates the case D = N \ {0, 1}, I
n:= (1/n, 1/(n − 1)].
Now let ε = (ε(n))
n∈Dbe an arbitrary, fixed sequence of zeroes and ones.
We define the map T
ε: [0, 1] → [0, 1] by
(8) T
εx := ε(x)Sx + (1 − ε(x))T x, x ∈ [0, 1], where
ε(x) :=
ε(n), x ∈ I
n, n ∈ D, 0, x ∈ I
∞. Let ε
n:= ε(T
εn−1x),
s
n= s
n(x) :=
s(T
εn−1x), T
εn−1x 6∈ I
∞,
∞, T
εn−1x ∈ I
∞, and h
ndefined similarly. By (6) and (7) one finds that
x = h
1+ ε
1s
1+ (−1)
ε1s
1T
εx
= h
1+ ε
1s
1+ (−1)
ε1s
1h
2+ ε
2s
2+ (−1)
ε2s
2T
ε2x
= . . .
= h
1+ ε
1s
1+ (−1)
ε1h
2+ ε
2s
1s
2+ (−1)
ε1+ε2h
3+ ε
3s
1s
2s
3+ . . . + (−1)
ε1+...+εn−1h
n+ ε
ns
1. . . s
n+ (−1)
ε1+...+εns
1. . . s
nT
εnx.
For each k ≥ 1 and 1 ≤ i ≤ k one has s
i≥ 1/L > 1, where L = max
n∈DL
n, and |T
εkx| ≤ 1. Thus,
(9)
x − p
kq
k= T
εkx s
1. . . s
k≤ L
k→ 0 as k → ∞, where
p
kq
k= h
1+ ε
1s
1+ (−1)
ε1h
2+ ε
2s
1s
2+ (−1)
ε1+ε2h
3+ ε
3s
1s
2s
3+ . . . (10)
+ (−1)
ε1+...+εk−1h
k+ ε
ks
1. . . s
kand q
k= s
1. . . s
k. In general p
kand q
kneed not be relatively prime (see also Section 3.1). Let ε
0:= 0, then for each x ∈ [0, 1] one has
(∗) x =
X
∞ n=1(−1)
ε0+...+εn−1h
n+ ε
ns
1. . . s
n.
For each x ∈ [0, 1] we define its sequence of digits a
n= a
n(x), n ≥ 1, as follows:
a
n= k ⇔ T
εn−1x ∈ I
k,
for k ∈ D ∪ {∞}. The expansion (∗) is called the (I, ε)-generalized L¨uroth series (GLS ) of x. Notice that for each x ∈ [0, 1] \ I
∞one finds a unique ex- pansion (∗), and therefore a unique sequence of digits a
n, n ≥ 1. Conversely, each sequence of digits a
n, n ≥ 1, with a
n∈ D ∪ {∞} and a
16= ∞ defines a unique series expansion (∗). We denote (∗) by
(11) x =
ε
1, ε
2, ε
3, . . . , ε
n, . . . a
1, a
2, a
3, . . . , a
n, . . .
. Since ε
n= ε(a
n), n ≥ 1, we might as well replace (11) by (12) x = [ a
1, a
2, a
3, . . . , a
n, . . . ]
No new information is obtained using (11) instead of (12). However, we will see in Section 3 that it is sometimes adventageous to use (11) instead of (12).
Finite truncations of the series in (∗) yield the sequence p
n/q
nof (I, ε)- GLS convergents of x. We denote p
n/q
nby
p
nq
n=
ε
1, ε
2, ε
3, . . . , ε
na
1, a
2, a
3, . . . , a
n.
Examples. 1. Let I
n:= (1/n, 1/(n − 1)], n ≥ 2. In case ε
n= 0 for n ≥ 2, one gets the classical L¨ uroth series, while ε
n= 1 for n ≥ 2 yields the alternating L¨ uroth series.
2. For n ∈ N, n ≥ 2, put I
i= (i/n, (i + 1)/n], i = 0, 1, . . . , n − 1. In case ε(i) = 0 for all i, T
εyields the n-adic expansion. In case n = 2 and ε(0) = 0, ε(1) = 1, T
εis the tent map.
See also Sections 3.2 and 3.3 for more intricate examples.
2. Ergodic properties of generalized L¨ uroth series. Let Ω be as in Section 1.2, B be the collection of Borel subsets of Ω, and λ be the Lebesgue measure on (Ω, B). Let (I, ε) be as in the previous section, viz. I = (I
n)
n∈Dsatisfies (4) and (5), while ε = (ε(n))
n∈Dis a sequence of zeroes and ones.
We have the following lemma.
Lemma 1. The stochastic variables a
1(x), a
2(x), . . . , corresponding to
the (I, ε)-GLS operator T
εfrom (8) are i.i.d. with respect to the Lebesgue
measure λ, and
λ(a
n= k) = L
kfor k ∈ D ∪ {∞}.
Furthermore, (I
n)
n∈Dis a generating partition.
P r o o f. Define for (k
1, . . . , k
n) ∈ D
n, n ≥ 1, the so-called fundamental intervals of order n by
(13) ∆
εk1...kn:= {x ∈ Ω : a
1(x) = k
1, . . . , a
n(x) = k
n}.
Let p
n/q
n∈ Q be defined as in (10) (and recall that the s
i’s, h
i’s and ε
i’s are uniquely determined by k
1, . . . , k
n), then obviously one has
x ∈ ∆
εk1...kn⇔ ∃y ∈ [0, 1] such that x = p
nq
n+ (−1)
ε1+...+εns
1. . . s
ny.
Thus ∆
εk1...kn
is an interval with p
n/q
nas one endpoint, and having length 1/(s
1. . . s
n). Therefore,
λ(∆
εk1...kn) = λ(a
1= k
1, . . . , a
n= k
n) = 1 s
1. . . s
n. Since
s
i= 1
r
ki− l
ki= 1
L
ki, i = 1, . . . , n, one finds
λ(∆
εk1...kn) = λ(a
1= k
1, . . . , a
n= k
n) = Y
n i=1L
ki.
The independence of the a
n(x)’s and the equality λ(a
n= k) = L
kfor k ∈ D ∪ {∞} now easily follow from
X
ki∈D
L
ki= 1 for all n ≥ 1 and all 1 ≤ i ≤ n.
That (I
n)
n∈Dis a generating partition is immediate from (9).
Theorem 1. The (I, ε)-GLS operator T
εfrom (8) is measure preserving with respect to Lebesgue measure and Bernoulli.
P r o o f. For any k
1, . . . , k
n∈ D, n ≥ 1, T
ε−1∆
εk1...kn= [
k∈D
∆
εkk1...knis a disjoint union of fundamental intervals of order n + 1, so that λ(T
ε−1∆
εk1...kn) = X
k∈D
λ(∆
εkk1...kn) = L
k1. . . L
knX
k∈D
L
k= L
k1. . . L
kn= λ(∆
εk1...kn).
Since the collection {∆
εk1...kn: n ≥ 1, k
i∈ D} generates B, it follows
from [W, Theorem 1.1, p. 20] that λ is T
ε-invariant. From Lemma 1, viz.
λ(∆
εk1...kn) = Y
n i=1λ(∆
εki), we conclude that ([0, 1], B, λ, T
ε) is a Bernoulli system.
R e m a r k s. 1. The Bernoullicity of the L¨ uroth operator T
Lwas already noticed by P. Liardet in [Li].
2. From the fact that T
εis Bernoulli, and therefore ergodic, one can draw a great number of easy consequences, using Birkhoff’s Ergodic Theorem. See also [JdV] and [K
32]. We mention here some results:
For almost every x the sequence (T
εnx)
∞n=0is uniformly distributed over [0, 1]. Furthermore,
n→∞
lim 1 n
n−1
X
k=0
T
εkx = 1
2 , lim
n→∞
n−1Y
k=0
T
εkx
1/n= 1 e a.e.
and
n→∞
lim (a
1. . . a
n)
1/n= e
ca.e., where c = P
k∈D
L
klog k (
3).
Define the map T
ε: [0, 1] × [0, 1] → [0, 1] × [0, 1] by (14) T
ε(x, y) :=
T
ε(x), h(x) + ε(x)
s(x) + (−1)
ε(x)s(x) y
. Notice that for
x =
ε
1, ε
2, ε
3, . . . , ε
n, . . . a
1, a
2, a
3, . . . , a
n, . . .
one has
T
ε(x, 0) =
ε
2, ε
3, . . . , ε
n, . . . a
2, a
3, . . . , a
n, . . .
,
ε
1a
1,
where
ε
1a
1= (−1)
ε0h
1+ s
1s
1= h
1+ s
1s
1. Now
T
ε2(x, 0) =
ε
3, ε
4, . . . , ε
n, . . . a
3, a
4, . . . , a
n, . . .
,
ε
2, ε
1a
2, a
1, where
ε
2, ε
1a
2, a
1= (−1)
ε0h
2+ ε
2s
2+ (−1)
ε0+ε2h
1+ ε
1s
1s
2= h
2+ ε
2s
2+ (−1)
ε2h
1+ ε
1s
1s
2.
(
3) In case 0 ∈ D we put e
c:= 0.
Putting T
εn(x, 0) =: (T
n, V
n), n ≥ 0, where T
n= T
εnx =
ε
n+1, ε
n+2, . . . a
n+1, a
n+2, . . .
, n ≥ 0, and
V
n=
ε
n, ε
n−1, . . . , ε
1a
n, a
n−1, . . . , a
1, n ≥ 1, V
0:= 0, we see inductively that
V
n+1= h
n+1+ ε
n+1s
n+1+ (−1)
εn+1s
n+1V
n.
As in the case of the continued fraction we will call T
n= T
εnx the future of x at time n, while V
n= V
n(x) is the past of x at time n (see also [K]).
We have the following theorem.
Theorem 2. The system ([0, 1] × [0, 1], B × B, λ × λ, T
ε) is the natural extension of ([0, 1], B, λ, T
ε). Furthermore, ([0, 1] × [0, 1], B × B, λ × λ, T
ε) is Bernoulli.
P r o o f. For any two vectors (k
1, . . . , k
n) ∈ D
n, (l
1, . . . , l
m) ∈ D
mone has
∆
εk1...kn× ∆
εl1...lm= T
εm(∆
εlm...l1k1...kn× [0, 1]).
Since {∆
εk1...kn× ∆
εl1...lm: k
i, l
j∈ D, 1 ≤ i ≤ n, 1 ≤ j ≤ m, n, m ≥ 1}
generates B × B, it follows that _
m≥0
T
εm(B × [0, 1]) = B × B.
Now, for any ∆
εk1...kn× ∆
εl1...lmone has
T
ε−1(∆
εk1...kn× ∆
εl1...lm) = ∆
εl1k1...kn× ∆
εl2...lm. Thus,
λ × λ(T
ε−1(∆
εk1...kn× ∆
εl1...lm)) = λ(∆
εl1k1...kn)λ(∆
εl2...lm)
= λ(∆
εk1...kn)λ(∆
εl1...lm)
= λ × λ(∆
εk1...kn× ∆
εl1...lm).
Since cylinders of the form ∆
εk1...kn
× ∆
εl1...lm
generate the σ-algebra B × B, it follows that T
εis measure preserving with respect to Lebesgue measure. Thus, T
εis the natural extension of T
ε(see [R] for details). Since T
εis Bernoulli it is an exercise to show that T
εis Bernoulli (see also [B]).
Corollary 1. For almost all x the two-dimensional sequence (15) T
εn(x, 0) = (T
n, V
n), n ≥ 0,
is uniformly distributed over [0, 1] × [0, 1].
P r o o f. Denote by A that subset of [0, 1] for which the sequence (T
n, V
n)
∞n=0is not uniformly distributed over [0, 1] × [0, 1]. It follows from Lemma 1 and the definition of T
εthat for all x, y, y
∗∈ [0, 1] one has
|T
εn(x, y) − T
εn(x, y
∗)| < L
n, n ≥ 0,
and we see that (T
εn(x, y) − T
εn(x, y
∗))
∞n=0is a null-sequence. Hence, if A :=
A × [0, 1], then for every pair (x, y) ∈ A the sequence T
εn(x, y), n ≥ 0, is not uniformly distributed over [0, 1] × [0, 1]. Now, if A had, as a subset of [0, 1], positive Lebesgue measure, so would A as a subset of [0, 1] × [0, 1]. However, this is impossible in view of Theorem 2.
The partition ξ = {I
k× [0, 1]}
k∈Dis a generator for T
ε, which implies that the entropy h(T
ε) of T
εequals h(T
ε, ξ) (see also [W], p. 96). There- fore,
h(T
ε) = − X
k∈D
L
klog L
k.
Now let (I
k)
k∈Dand (I
k∗)
k∈Dbe two partitions of [0, 1] satisfying (4) and (5), and suppose that L
k= L
∗kfor k ∈ D. Furthermore, let ε = (ε
k)
k∈Dand ε
∗= (ε
∗k)
k∈Dbe two arbitrary sequences of zeroes and ones. It follows at once from Ornstein’s Isomorphism Theorem (see [W], p. 105) and Theorem 2 that T
εand T
ε∗are metrically isomorphic. We conclude this section with the following theorem, which gives a concrete isomorphism.
Theorem 3. Let (I
k)
k∈Dand (I
k∗)
k∈Dbe two partitions of [0, 1], sat- isfying (4) and (5). Suppose that L
k= L
∗kfor k ∈ D. Furthermore, let ε = (ε
k)
k∈Dand ε
∗= (ε
∗k)
k∈Dbe two sequences of zeroes and ones. Let T
εand T
ε∗be defined as in (8). Finally, define Ψ : [0, 1] × [0, 1] → [0, 1] × [0, 1]
by Ψ
ε
1, ε
2, . . . a
1, a
2, . . .
,
ε
0, ε
−1, . . . a
0, a
−1, . . .
:=
ε
∗(a
1), ε
∗(a
2), . . . a
1, a
2, . . .
,
ε
∗(a
0), ε
∗(a
−1), . . . a
0, a
−1, . . .
. Then Ψ is a measure preserving isomorphism.
P r o o f. Since almost every x ∈ [0, 1] has unique (I, ε)-, (I
∗, ε
∗)-GLS expansions, it follows that Ψ is injective. Now, for any cylinder ∆
εk∗1...kn×
∆
εl∗1...lm
,
∆
εk1...kn× ∆
εl1...lm= Ψ
−1(∆
εk∗1...kn× ∆
εl1∗...lm)
and
(λ × λ)(∆
εk∗1...kn× ∆
εl1∗...lm) = L
∗k1. . . L
∗knL
∗l1. . . L
∗lm= L
k1. . . L
knL
l1. . . L
lm= (λ × λ)(∆
εk1...kn× ∆
εl1...lm)
= (λ × λ) Ψ
−1(∆
εk∗1...kn× ∆
εl1∗...lm).
This shows that Ψ is measurable and measure preserving.
Finally, let (x, y) ∈ [0, 1] × [0, 1] with x =
ε
1, ε
2, . . . a
1, a
2, . . .
, y =
ε
0, ε
−1, . . . a
0, a
−1, . . .
. Then
Ψ T
ε(x, y) =
ε
∗(a
2), ε
∗(a
3), . . . a
2, a
3, . . .
,
ε
∗(a
1), ε
∗(a
0), ε
∗(a
−1), . . . a
1, a
0, a
−1, . . .
= T
ε∗Ψ (x, y),
therefore Ψ is a measure preserving isomorphism.
3. Applications and examples
3.1. Approximation coefficients and their distribution. As before let I = (I
n)
n∈Dbe a partition of [0,1] which satisfies (4) and (5), and let ε = (ε(n))
n∈Dbe a sequence of zeroes and ones. Putting q
k= s
1. . . s
k, it follows from (9) and Corollary 1 that for a.e. x the approximation coefficients θ
n, n ≥ 0, defined by
θ
ε,n= θ
ε,n(x) := q
nx − p
nq
n, n ≥ 0,
have the same distribution as T
εnx, n ≥ 0. Viz., for a.e. x the sequence (θ
ε,n)
nis uniformly distributed on [0,1].
In fact, for many partitions (I
n)
n∈Dmore information on the distribution of the θ
n’s can be obtained by a more careful definition of q
n. As an example we will treat here the case of the classical L¨ uroth series, and all other GLS expansions with the same partition (I
n)
n∈D(see also the examples at the end of Section 1.2).
In this case
s
n= s(T
εn−1x) = 1
1 an−1
−
a1n
= a
n(a
n− 1), h
n= a
n− 1, and
h
n+ ε
ns
1. . . s
n= a
n− 1 + ε
na
1(a
1− 1)a
2(a
2− 1) . . . a
n(a
n− 1) = 1
a
1(a
1− 1) . . . (a
n− ε
n) . Therefore it is more appropriate to put
q
1= a
1−ε
1, q
n= a
1(a
1−1)a
2(a
2−1) . . . a
n−1(a
n−1−1)(a
n−ε
n), n ≥ 2,
and we see
(16) θ
n(x) = a
n− ε
na
n(a
n− 1) T
εnx, n ≥ 1.
We have the following theorem.
Theorem 4. Let (I
n)
n∈Dbe the L¨uroth partition, that is, I
n:= (1/n, 1/(n − 1)] for n ≥ 2, and let ε(n) ∈ {0, 1} for n ≥ 2. Then for a.e. x and for every z ∈ (0, 1] the limit
N →∞
lim 1
N #{1 ≤ j ≤ N : θ
j(x) < z}
exists and equals F
ε(z), where F
ε(z) :=
b1/zc+1−ε(b1/zc+1)
X
k=2
z
k − ε(k) + 1
b1/zc + 1 − ε(b1/zc + 1) , 0 < z ≤ 1.
P r o o f. Let z ∈ (0, 1]. From (15) and (16) it follows that (17) θ
n< z ⇔ (T
n, V
n) ∈ Ξ(z) =
[
∞ k=2Ξ
k(z), where
Ξ
k(z) :=
0, k(k − 1) k − ε(k) z
∩ [0, 1]
× ∆
k, k ≥ 2.
For k ≥ 2 we have the following two cases (of which the first one might be void).
(A) 2 ≤ k ≤ b1/zc + 1 − ε(b1/zc + 1). In this case Ξ
k(z) =
0, k(k − 1) k − ε(k) z
× ∆
k.
(B) k > b1/zc + 1 − ε(b1/zc + 1). In this case Ξ
k(z) = [0, 1] × ∆
k. From (A) and (B) one finds, that
(λ × λ)(Ξ(z))
=
b1/zc+1−ε(b1/zc+1)
X
k=2
(λ × λ)
0, k(k − 1) k − ε(k) z
× ∆
k+ 1
b1/zc + 1 − ε(b1/zc + 1)
=
b1/zc+1−ε(b1/zc+1)
X
k=2
z
k − ε(k) + 1
b1/zc + 1 − ε(b1/zc + 1) .
The theorem at once follows from Corollary 1.
R e m a r k s . 1. Although the map x → (1/x) mod 1 generating the con- tinued fraction is not piecewise linear, which leads to complications in esti- mations, a similar result as in Theorem 4 was obtained for continued frac- tions (see also [BJW]).
2. If ε(n) = 0, n ≥ 2 (the classical L¨ uroth case) (
4), the distribution function F
εreduces to
F
L(z) =
b1/zc+1
X
k=2
z
k + 1
b1/zc + 1 , 0 < z ≤ 1.
Furthermore
F
A(z) =
b1/zc
X
k=2
z
k − 1 + 1
b1/zc , 0 < z ≤ 1;
see also Figure 2. Notice that F
A≤ F
ε≤ F
L.
Fig. 2
We have the following corollary.
(
4) From now on the classical (resp. the alternating) L¨ uroth case will be indicated by
a subscript L (resp. A).
Corollary 2. Let (I
n)
n∈Dbe the L¨uroth partition and let ε(n) ∈ {0, 1}
for n ≥ 2. Then there exists a constant M
εsuch that for a.e. x,
N →∞
lim 1 N
X
N i=1θ
ε,i= M
ε.
Moreover , M
εcan be calculated explicitly, and M
A≤ M
ε≤ M
L, where M
A= 1 −
12ζ(2) = 0.177533 . . . and M
L=
12(ζ(2) − 1) = 0.322467 . . .
P r o o f. By definition M
εis the first moment of F
εand thus M
ε= R
10
(1 − F
ε(x)) dx.
R e m a r k s . 1. From Corollary 2 it follows that among all (I, ε)-GLS expansions with I the L¨ uroth partition the alternating L¨ uroth series has the best approximation properties.
2. The presentation of Corollary 2 suggests that by choosing ε = (ε(n))
n≥2appropriately, each value in the interval [M
A, M
L] = [0.177533 . . . , 0.322467 . . .] might be attained. This is certainly incorrect, as the following example shows. Let ε = (ε(n))
n≥2be given by ε(2) = 1 and ε(n) = 0 for n ≥ 3, and let ε
∗= (ε
∗(n))
n≥2be given by ε
∗(2) = 0 and ε
∗(n) = 1 for n ≥ 3. A simple calculation yields that M
ε= M
L− 1/8 = 0.197467 . . . and M
ε∗= M
A+ 1/8 = 0.302533 . . . ; we thus see that M
ε< M
ε∗and from this one can easily deduce that there does not exist a sequence ε
[= (ε
[(n))
n≥2of zeroes and ones for which M
ε[∈ [M
ε, M
ε∗]. Some further investigations even suggest that the set
Υ := {M
ε: ε(n) ∈ {0, 1} for n ≥ 2}
is a Cantor set.
3.2. Jump transformations. Let T
εbe a (I, ε)-GLS operator with digit set D, and let a ∈ D. For x ∈ Ω, put
n
a= n
a(x) := min
n≥1
{a
n(x) : a
n(x) = a}
(and n
a= ∞ in case a
n(x) 6= a for all n ≥ 1). Define the jump transforma- tion J
a: Ω → Ω by
(18) J
ax :=
T
εnax, n
a∈ N, 0, n
a= ∞.
Jump transformations were first studied by H. Jager [J] for the particular case that T
εx = 10x (mod 1). Jager showed that such jump transformations are stronly mixing. Here, in the more general setting, we have a stronger result.
Theorem 5. Let T
εbe an (I, ε)-GLS operator with digit set D. For each
a ∈ D the corresponding jump transformation J
a, as defined in (18), is an
(I
∗, ε
∗)-GLS operator , with
I
∗= {∆
a1...an: n ≥ 1, a
n= a and a
i6= a for 1 ≤ i ≤ n − 1}
and for each ∆
a1...an∈ I
∗the corresponding value of ε is given by ε
∗(∆
a1...an) = ε(a
1) + . . . + ε(a
n) (mod 2).
3.3. β-expansions and pseudo golden mean numbers. For an irrational number β > 1 the β-transformation T
β: [0, 1] → [0, 1] is defined by
T
βx = βx (mod 1)
(see also [FS] for further references). Clearly, β-transformations do not be- long to the class of GLS-transformations. However, in some cases there exists an intimate relation between both types of transformations, as the following example shows.
Let β > 1 be the positive root of the polynomial X
m−X
m−1−. . .−X −1, where (
5) m ≥ 2. Due to C. Frougny and B. Solomyak [FS] we know that such β’s are Pisot numbers and that the β-expansion d(1, β) is finite, and equals
1 = 1 β + 1
β
2+ . . . + 1 β
m.
Notice that T
βi1 = β
−1+ . . . + β
−(m−i), 0 ≤ i ≤ m − 1, and T
βi1 = 0 for i ≥ m. Furthermore, let
X :=
m−1
[
k=0
(T
βm−k1, T
βm−k−11] × [0, T
βk1]
(see also Figure 3 for m = 4), and define T
β: X → X by T
β(x, y) :=
T
βx, 1
β (bβxc + y)
.
Let Y := [0, 1] × [0, 1/β] and W : Y → Y the corresponding induced trans- formation under T
β, that is
W(x, y) = T
βn(x,y)(x, y),
where n(x, y) = min{s > 0 : T
βs(x, y) ∈ Y }. Clearly one has W(x, y) = T
βk+1(x, y),
where k = k(x) ∈ {0, 1, . . . , m − 1} is such that x ∈ (T
βm−k1, T
βm−k−11].
Notice that W maps rectangles to rectangles; see also Figure 3.
Finally, let T
εbe the (I, ε)-GLS operator with partition I given by (T
βm−i1, T
βm−i−11], 0 ≤ i ≤ m − 1 (see also Figure 3), and ε(n) = 0 for
(
5) For m = 2 one has β = ( √
5 + 1)/2, which is the golden mean. For m ≥ 3 we call
these β’s pseudo golden mean numbers.
Fig. 3
each digit n. Notice that for x ∈ (T
βm−i1, T
βm−(i+1)1], 0 ≤ i ≤ m − 1, one has
(19) T
εx = T
βi+1x.
We have the following lemma.
Lemma 2. Let Ψ : [0, 1] × [0, 1] → Y be defined by Ψ (x, y) := (x, y/β).
Then Ψ is a measurable bijection which satisfies Ψ ◦ T
ε= W ◦ Ψ.
P r o o f. For (x, y) ∈ [0, 1] × [0, 1] let i = i(x) ∈ {0, 1, . . . , m − 1} be such that x ∈ (T
βm−i1, T
βm−(i+1)1]. From (14) it then follows that
T
ε(x, y) =
T
εx, T
βm−i1 + y β
i+1and therefore
(20) Ψ (T
ε(x, y)) =
T
εx, 1
β T
βm−i1 + y β
i+2.
From (19), (20) and the definitions of W and Ψ it now follows that
W(Ψ (x, y)) = Ψ (T
ε(x, y)) for i = 0 and in case i 6= 0 one has
W(Ψ (x, y)) = W
x, y
β
= T
βi+1x, y
β
=
T
βi+1x, 1 β
0 +
1 β
1 + . . . 1 β
1 + 1
β
1
| {z }
i times
+ y β
. . .
=
T
εx, 1
β
2+ . . . + 1
β
i+1+ y β
i+2= Ψ (T
ε(x, y)).
Let % be the measure on Y defined by
%(A) := (λ × λ)(Ψ
−1(A)) for each Borel set A ⊂ Y.
It follows from Lemma 2 and the fact that λ × λ is an invariant measure for T
εthat % is invariant with respect to W, and % = β(λ × λ). Lemma 2 now at once yields the following proposition.
Proposition. The dynamical systems ([0, 1] × [0, 1], λ × λ, T
ε) and (Y, %, W) are isomorphic.
Using standard techniques (see [CFS], p. 21) one obtains the measure µ on X which is invariant with respect to T
β, viz.
µ(A) = β
1 β + 2
β
2+ . . . + m β
m(λ × λ)(A)
for each Borel set A ⊂ X. One also easily shows that (X, µ, T
β) forms the natural extension of ([0, 1], ν, T
β), where ν is the invariant measure with respect to T
β[So]. Projecting µ on the first coordinate of X yields ν; one finds that ν has density h(x), where
h(x) = 1
1 β + 2
β + . . . + m β
mX
x<Tβi1