on his 70th birthday
ON THE METRICAL THEORY OF A PECULIAR CONTINUED FRACTION EXPANSION
MARIUS IOSIFESCU and GABRIELA ILEANA SEBE
Consider the transformationT of the unit intervalI= [0,1]dened as T(x) =1[0,1/2)(x) x
1−x+1[1/2,1](x)1−x
x , x∈I,
that was studied by Ito [7], especially from the point of view of ergodic theory.
Our work should be seen as complementary to his. First, we take up the Gauss problem forT in the manner of [6]. This amounts to the asymptotic behaviour ofm T−n(A)
asn→ ∞for probability measuresmλ(Lebesgue measure on theσ-algebraBI of Borel subsets ofI) and setsA∈ BI. Second, we are concerned with the sequence(εn)n∈N+, N+={1,2, . . .},of{0,1}-valued random variables on(I,BI), whereεn=ε1◦Tn−1, n∈N+, withε1=1[1/2,1]andT0(x) =x, x∈I. Theεn, n∈N+, occur in a peculiar continued fraction expansion of irrationals in I, allowing to recover the so-called Lehner continued fraction expansion see [2]
and [5]. We also consider the random variableszn, n∈N+, dened recursively byzn= 1 + (1−2εn)/(εn+ 1/zn−1)with an arbitrarily prescribedz0. We prove that for anyn∈N+ the conditional probabilityλ(εn+1= 0|ε1, . . . , εn)is equal to (zn+ 1)/(zn+ 2) when z0 = 0. Also, when z0 is arbitrary, (zn)n∈N+ is a [0,∞)-valued Markov process with transition probability function
P(z, A) = 1 z+ 21A
1
z+ 1
+z+ 1
z+ 21A(z+ 1), z∈[0,∞], A∈ B[0,∞), that has as stationary distribution the σ-nite, innite measure with density 1/(z+ 1)on[0,∞).
AMS 2000 Subject Classication: Primary 11K55;
Secondary 47B38, 60J05, 60J10.
Key words: Continued fraction expansion, Gauss' problem, Perron-Frobenius ope- rator, Markov chain, Markov process.
REV. ROUMAINE MATH. PURES APPL., 53 (2008), 56, 465477
1. PREREQUISITES
Consider, cf. Ito [7], the transformationT of the unit interval I = [0,1]
dened as
T(x) =1[0,1/2)(x) x
1−x +1[1/2,1](x)1−x
x , x∈I,
with 1A standing for the indicator function of the set A, and the inverse branches V0 and V1 ofT dened by
V0(x) = x
x+ 1, 0≤x <1, and
V1(x) = 1
x+ 1, x∈I,
that are modular transformations. This suggests matrix representations for them, namely,
A0 =
1 1 0 1
0 + 1·x
1 + 1·x =V0(x)
and
A1 =
1 1 1 0
1 + 0·x
1 + 1·x =V1(x)
.
Write1[1/2,1] =ε1 and for any x∈ I deneεn(x) =ε1 Tn−1(x)
,x ∈I, n∈N+={1,2, . . .}, withT0 the identity transformation.
It is easy to check that
(1) x= ε1(x) + (1−ε1(x))T(x)
1 +T(x) = 1−ε1(x) +2ε1(x)−1
1 +T(x) , x∈I.
Hence
(2) x= 1−ε1(x) + 2ε1(x)−1
1 + 1−ε1(T(x)) +2ε1(T(x))−1 1 +T2(x)
= 1−ε1(x) + 2ε1(x)−1 2−ε2(x) + 2ε2(x)−1
1 +T2(x)
=· · ·
= 1−ε1(x) + 2ε1(x)−1 2−ε2(x) + 2ε2(x)−1
2−ε3(x) +
... +2εn(x)−1 1 +Tn(x) for any x ∈ I and n ∈ N+ such that Tn(x) 6= 0. This expansion can be continued indenitely if and only if x ∈ Ω = the set of irrationals in I while
for any x ∈ I \Ω there exists m = m(x) ∈ N+ such that Tn(x) = 0 for all n≥m,and the expansion terminates. Clearly, theεn, n∈N+, can be viewed as {0,1}-valued random variables on (I,BI), where BI is the Borel σ-algebra inI.
Since (1) is associated with a modular transformation with matrix rep- resentation Aε1(x), equation (2) can be also written as
(3) x= tn(x) +Tn(x)un(x)
rn(x) +Tn(x)sn(x), x∈I, n∈N, where
rn(x)sn(x) tn(x) un(x)
=
n
Y
i=1
Aεi(x), x∈I.
In particular, r0(x) = 1, r1(x) = 1, s0(x) = 0,s1(x) = 1, t0(x) = 0, t1(x) = ε1(x),u0(x) = 1,u1(x) = 1−ε1(x),x∈I. As
Aεn =
1 1 εn 1−εn
, n∈N+, we have
rn sn
tn un
=
rn−1 sn−1
tn−1 un−1
1 1
εn 1−εn
, hence
(4) rn=rn−1+εnsn−1, sn=rn−1+ (1−εn)sn−1
for n ∈N+. Thus, the sequence (rn)n∈N+ of positive integers is clearly non- decreasing. It converges to ∞ for x ∈Ωsince in the sequence(εn(x))n∈N+ = (ε1 Tn−1(x)
n∈N+ the value 1 occurs innitely often while rn+1 > rn if εn(x) = 1.Consequently, by (3) and taking into account that
(5) |rnun−sntn|=
det
n
Y
i=1
Aεi
= 1, n∈N+, we have
x− tn(x) +un(x) rn(x) +sn(x)
= 1−Tn(x)
(rn(x) +sn(x)) (rn(x) +Tn(x)sn(x)), hence
x= lim
n→∞
tn(x) +un(x) rn(x) +sn(x) for any x∈Ω.
Remark. Let a : I → R be any real-valued function dened on I. We actually have
x= lim
n→∞
tn(x) +a(x)un(x)
rn(x) +a(x)sn(x), x∈Ω, since
x− tn(x) +a(x)un(x) rn(x) +a(x)sn(x)
= |a(x)−Tn(x)|
(rn(x) +a(x)sn(x)) (rn(x) +Tn(x)sn(x)). The importance of the special case a(x) ≡ 1 lies in the fact that the set of convergents
tn(x) +un(x)
rn(x) +sn(x), n∈N+,
is the union of the sets of the regular continued fraction and mediant conver- gents of x. See Ito [7, p. 562].
To conclude this section, let us note that by (2) we have (6) y= 2−ε1(y−1)+ 2ε1(y−1)−1
2−ε2(y−1)+ 2ε2(y−1)−1 2−ε3(y−1) +
... +2εn(y−1)−1 1 +Tn(y−1) for anyy∈[1,2]andn∈N+such thatTn(y−1)6= 0. This allows to recover the so-called Lehner continued fraction expansion. Recall that any number y ∈ [1,3/2)∪(3/2,2) has a unique innite continued fraction expansion of the form
b0+ e1 b1+ e2
b2+· · ·
:= [b0;e1/b1, e2/b2, . . .],
where (bn, en+1) is equal to either (1,1) or (2,−1), n ∈ N. Cf. Dajani and Kraaikamp [2]. See also Iosifescu and Kraaikamp [5, Subsection 4.4.2]. Clearly, (6) is of this kind, so that using results in the references above we can state the results below.
Proposition 1. Almost everywhere in Ωwe have
n→∞lim 1 n
n
X
k=1
1 2−εk=1
2, lim
n→∞
pn
(2−ε1)· · ·(2−εn) = 2, lim
n→∞
1 n
n
X
k=1
εk= 0.
The last equation above implies that the asymptotic theoretical frequency of value 1, resp. 0, in the sequence (εn)n∈N+ is 0, resp. 1.
An explanation of this asymmetry, will be given later in Section 2.
2. THE PERRON-FROBENIUS OPERATOR AND GAUSS' PROBLEM
Daniels [3] showed that the transformation T has a σ-nite, innite in- variant measure µ with density 1/xon (0,1], 0 at x = 0, and the dynamical system(I,BI, T, µ) is ergodic.
Let L1(µ) and L∞(µ) denote the usual L1- and L∞-Banach spaces on (I,BI, µ). In what follows we parallel the discussion in [6], where a similar situation has been considered.
According to the general theory (cf. [1, p. 33]), the Perron-Frobenius operator UT,µ of T underµtakes L1(µ) into itself and satises the equation (7)
Z
I
(UT,µh)gdµ= Z
I
(g◦T)hdµ for any h∈L1(µ) andg∈L∞(µ). Hence
(8)
Z
I
UT,µn h gdµ=
Z
I
(g◦Tn)hdµ for any n∈N+. Let νh(A) =R
Ahdµ for h∈L1(µ) and A∈ BI. Then UT,µh can be expressed as a Radon-Nikodym derivative, namely, d(νh◦T−1)/dµ. In our case we have
Z
T−1([a,b])
hdµ= Z b+1b
a a+1
hdx x +
Z a+11
1 b+1
hdx x
for any 0 ≤ a < b ≤1. After the changes of variable x = y+1y and x = y+11 , respectively, the right-hand member of the last equation becomes
Z b a
h y
y+ 1
y+ 1 y
dy (y+ 1)2 +
Z b a
h 1
y+ 1
(y+ 1) dy (y+ 1)2
= Z b
a
h y
y+ 1 1
y+ 1µ(dy) + Z b
a
h 1
y+ 1 y
y+ 1µ(dy). It follows that
UT,µh(y) = 1 y+ 1h
y y+ 1
+ y
y+ 1h 1
y+ 1
forµ-almost ally∈I and anyh∈L1(µ). Clearly,UT,µ1 = 1even if1∈/ L1(µ). Actually,UT,µhmakes sense for any functionh:I →Cbut, of course, without satisfying (7) for all of them.
The Gauss problem for the transformationT amounts to the asymptotic behaviour ofm(T−n(A))asn→ ∞for probability measuresmλ(Lebesgue
measure onBI) and sets A∈ BI. It follows from (8) by takingg= 1A that Z
A
UT,µn hdµ= Z
T−n(A)
hdµ.
Hence
(9) m T−n(A)
= Z
A
UT,µn fdµ for any n∈Nand A∈ BI, where
f = dm dµ = dm
dλ / dµ dλ, that is, f(x) =xdmdλ(x) for almost allx∈I.
Therefore, the asymptotic behaviour of m(T−n(A)) depends on the asymptotic behaviour ofUT,µn asn→ ∞.By Proposition 1.1.3 in [1] the trans- formationT is conservative. SinceT is conservative and ergodic, by [1, Exercise 2.2.1, p. 61], we have
(10) lim
n→∞
1 n
n−1
X
k=0
h◦Tk = 0 µ-a.e.,
hence λ-a.e., for any h ∈ L1(µ). In particular, for h = 1A with A ∈ BI and µ(A)<∞ we get
(11) lim
n→∞
1 n
n−1
X
k=0
1A◦Tk= 0 bothµ-a.e. and λ-a.e.
By dominated convergence, this clearly implies
(12) lim
n→∞
1 n
n−1
X
k=0
m T−k(A)
= Z
I n→∞lim
n−1
P
k=0
1A◦Tk
n pdλ= 0 for any A∈ BI such thatµ(A)<∞.This shows that if the limit
(13) lim
n→∞m T−n(A) exists, then it should be 0; and if so then, from (9),
0 = lim inf
n→∞
Z
A
UT,µn fdµ≥ Z
A
lim inf
n→∞ UT,µn fdµ, hence
lim inf
n→∞ UT,µn f = 0 µ-a.e.
on any A∈ BI such thatµ(A)<∞. This actually means that lim inf
n→∞ UT,µn f = 0 λ-a.e. in [0,1].
In any case, one can assert that we always have lim inf
n→∞
1 n
n−1
X
k=0
UT,µk f = 0 λ-a.e. inI.
Recall that f above is dened byf(x)≡xdmdλ(x)for almost allx∈I.
Coming back to equation (11), let A∈ BI be such that µ(A) <∞. Let Ac =I\A. Since1A+ 1Ac = 1, we have
1 n
n−1
X
k=0
1A◦Tk+ 1 n
n−1
X
k=0
1Ac◦Tk= 1.
It follows from (11) that
(14) lim
n→∞
1 n
n−1
X
k=0
1Ac◦Tk= 1 bothµ-a.e. and λ-a.e.
This means that both µ-almost and λ-almost all orbits (x, T(x), T2(x), . . .), x∈Ω, hitAcwith relative frequency asymptotical to1. Also, similarly to (11), we have
(15) lim
n→∞
1 n
n−1
X
k=0
m T−k(Ac)
= 1.
Hence, if the limit
(16) lim
n→∞m T−n(Ac) exists, then it should be 1.
Note that (14) makes more precise the assertion following Proposition 1 concerning the asymptotic relative frequency of value 0 in the sequence (εn)n∈N+.
We conjecture that both limits (13) and (16) do not exist, so that (12) and (15) appear to be the solution of Gauss' problem for T.
3. AN INFINITE ORDER CHAIN AND A MARKOV CHAIN Let us dene
zn= sn rn
, n∈N.
It follows from (4) that
(17) zn= (1−εn)zn−1+ 1
εnzn−1+ 1 = 1 + 1−2εn εn+ 1/zn−1
,
hence
zn= 1 + 1−2εn
εn+ 1
1 + 1−2εn−1
εn−1 +
... + 1
1 + 1−2ε2
ε2+ 1
1 +(1−2ε1)z0
ε1z0+ 1
for any n ∈N+ with z0 = s0/r0 = 0.(Further on, values z0 6= 0 will be also considered.)
Alternatively, for anyn∈N+ we have (18) zn+ 1 = (zn−1+ 1) + 1
1−εn+εn(zn−1+ 1) = 1 εn+ 1−2εn
1 + (zn−1+ 1) , hence
zn+ 1 = 1
εn+ 1−2εn
1 + 1
εn−1 +
...+ 1
ε1+ 1−2ε1
1 + (z0+ 1) .
Let us note that by (17) and (18) we can assert that zn= anz0+bn
cnz0+dn
, with
an bn
cn dn
:=
n
Y
i=1
1−εi 1 εi 1
, n∈N+, and
zn+ 1 = a0n(z0+ 1) +b0n c0n(z0+ 1) +d0n,
with
a0n b0n c0n d0n
:=Aε1· · ·Aεn, n∈N+.
We shall now show that the dependence structure of (εn)n∈N+ can be described in terms of the zn, n ∈ N+. First, we dene the cylinder sets or fundamental intervals. For any n ∈ N+ and any ik ∈ {0,1}, 1 ≤ k ≤ n, consider the n-tuple i(n) := (i1, . . . , in). Let I(i(n)) denote the set of
x ∈ Ω for which εk(x) = ik, 1 ≤ k ≤ n. Remark that all n-tuples are ad- missible, to mean that all I(i(n)), n ∈ N+, are non-empty sets. These sets are called cylinder sets or fundamental intervals. For example, I(0) = 0,12
∩ Ω, I(1) = 1
2,1
∩ Ω, I(0,0) = 0,13
∩ Ω, I(0,1) = 1
3,12
∩ Ω, I(1,1) =1
2,23
∩Ω, I(1,0) =2
3,1
∩Ω, I(0,0,0) = 0,14
∩Ω, I(0,0,1) = 1
4,13
∩ Ω, I(0,1,1) = 1
3,25
∩ Ω, I(0,1,0) = 2
5,12
∩ Ω, I(1,1,0) = 1
2,35
∩Ω, I(1,1,1) =3
5,23
∩Ω, I(1,0,1) =2
3,34
∩Ω, I(1,0,0) =3
4,1
∩Ω, etc. We shall prove that any I i(n)
is the set of irrationals from a certain interval with rational end points.
Proposition 2. For any n∈N+ a cylinder set I i(n)
is either the set of irrationals in the interval with end points
tn−1
rn−1 and tn−1+un−1/2 rn−1+sn−1/2
when in= 0 or the set of irrationals in the interval with end points tn−1+un−1/2
rn−1+sn−1/2 and tn−1+un−1
rn−1+sn−1
when in = 1. Here the rk, sk, tk, uk, k ∈ N+, are computed using the rules given in Section 1.
Proof. If x∈I i(n)
and in = 0, thenTn−1(x)∈I0 while if x∈I i(n) and in= 1,thenTn−1(x)∈I1.By (3) we have
x= tn−1(x) +Tn−1(x)un−1(x) rn−1(x) +Tn−1(x)sn−1(x)
with 0 ≤Tn−1 <1/2 in the rst case and1/2≤ Tn−1(x) ≤1 in the second one, which leads to the assertion in the statement.
Corollary. The Lebesgue measure of a cylinder I i(n) is λ I(i(n))
= 1
2 (rn−1+sn−1/2) (rn−1+δ(in,1)sn−1). Proof. This follows by direct computation using (5).
Proposition 3. For any n∈N+ we have λ(εn+1 = 0|ε1, . . . , εn) = 1 +zn
2 +zn
and
λ(εn+1= 1 |ε1, . . . , εn) = 1 2 +zn.
Proof. The equations are simple consequences of the above corollary and equations (4). Clearly,
λ(I(ε1, . . . , εn)) = 1
2 (rn−1+sn−1/2) (rn−1+δ(εn,1)sn−1) and
λ(I(ε1, . . . , εn,0)) = 1
2 (rn+sn/2)rn
= 1
(rn−1+εnsn−1+ (rn−1+ (1−εn)sn−1/2) 2 (rn−1+εnsn−1),
where thernandsn, n∈N+, are computed according to equations (4). Hence λ(εn+1 = 0 |ε1, . . . , εn) = λ(I(ε1, . . . , εn,0))
λ(I(ε1, . . . , εn))
= (rn−1+sn−1/2) (rn−1+δ(εn,1)sn−1)
(rn−1+εnsn−1+ (rn−1+ (1−εn)sn−1)/2) (rn−1+εnsn−1)
= (2 +zn−1) (1 +δ(εn,1)zn−1)
(2 (1 +εnzn−1) + 1 + (1−εn)zn−1) (1 +εnzn−1)
= 2 +zn−1
3 + (1 +εn)zn−1
. Finally, remark that for all n≥2 we have
(zn>1)≡(εn= 0)≡(zn−1=zn−1), (zn<1)≡(εn= 1)≡
zn−1 = 1 zn −1
. Hence
2 +zn−1
3 + (1 +εn)zn−1
=
2 +zn−1
3 +zn−1 = 1 +zn
2 +zn if εn= 0 2 + 1
zn −1 3 + 2
zn
−2
= 1 +zn
2 +zn if εn= 1
for all n≥2 while for n= 1the above equation holds trivially as z0 = 0 and z1 = 1.The proof is complete.
Proposition 3 shows that the sequence(εn)n∈N+ is a{0,1}-valued innite order chain, cf. Iosifescu and Grigorescu [4, Section 5.5]. On the other hand, the sequence (zn)n∈N+ on (I,BI, λ) is a Q+-valued Markov chain starting at z1 = 1 with the following transition mechanism: from state z ∈ Q+ the
only possible transitions are to states 1/(z+ 1) and z+ 1 with probabilities 1/(z+ 2)and (z+ 1)/(z+ 2),respectively.
More generally, we can also consider anR+-valued Markov process with the same transition mechanism. Clearly, the transition operator U of such a process can be expressed as
U h(z) = z+ 1
z+ 2h(z+ 1) + 1 z+ 2h
1 z+ 1
, z∈R+= [0,∞), for any bounded measurable function on R+.
Proposition 4. The innite σ-nite Gauss' measure γ on R+ whose density is 1/(x+ 1), x∈R+, is invariant for the R+-valued Markov process (zn)n∈N+.
Proof. The assertion amounts to the validity of the equationR
R+U hdγ= R
R+hdγ for any boundedγ-integrable function onR+. This is easy to check as Z
R+
U hdγ= Z ∞
0
f(z+ 1) z+ 2 dz+
Z ∞ 0
f(1/(z+ 1))
(z+ 1) (z+ 2)dz:=I1+I2. Making the changes of variable z+ 1 =w and1/(z+ 1) =w, respectively, in the two integrals above we obtain
I1= Z ∞
1
f(w) w+ 1dw=
Z ∞ 1
fdγ and
I2= Z 1
0
wf(w) (1 + 1/w)
dw w2 =
Z 1 0
fdγ, and the proof is complete.
Let us note that if we consider the transformationτ of (0,∞) dened as τ(x) =
( (1−x)/x if x∈(0,1), x−1 if x≥1,
thenzi=τ(zi+1) for anyi∈Nwhile, as easily checked, the Perron-Frobenius operatorUτ,γ ofτ underγ has the same analytical expression asU, but acting now on functions h∈L1(γ). Clearlyτ(x) =τ(1/x), x∈(0,∞), and
T(x) =
1
τ(x) if 0< x≤1/2, τ(x) if 1/2≤x≤1.
Note also that τ was considered more than 45 years ago in another context by Parry [8, p. 153].
Coming back to Proposition 4, we note that the existence of a stationary measure dierent from γ of the R+-valued Markov process (zn)n∈N+ is an open question. Such a measure would be a τ-invariant one, leaving just the possibility of a λ-singular measure. Cf. [1, p. 45] on unicity of a λ-absolutely continuous invariant measure.
Concerning both Perron-Frobenius operators UT,µ and U =Uτ,γ, we re- mark that they are operators dealt with in dependence-with-complete-con- nection theory. Cf. [4], especially Chapters 1 and 5. No known result from that theory can be used in the present case to derive convergence of UT,µn h or Unhasn→ ∞for certain functionsh. Note that by Proposition 3 the proba- bility λ(εn+1 = 0) that εn+1 takes on value 0 is equal to Unh(0), n ∈ N, where
h(z) = z+ 1
z+ 2, z∈[0,∞).
This follows from the equations
λ(εn+1= 0|ε1, . . . , εn) =λ(zn+1>1 |zn) =U1(1,∞)(zn) =zn+ 1
zn+ 2, n∈N+, λ(ε1 = 0) = 1
2 =h(0) =
= z0+ 1 z0+ 2
with z0 = 0.
It follows from Proposition 1 that
n→∞lim 1 n
n
X
k=1
λ(εk= 1) = lim
n→∞Eλ 1 n
n
X
k=1
εk
!
= 0, hence
n→∞lim 1 n
n
X
k=1
λ(εk= 0) = 1.
We conjecture that, moreover, we even have
n→∞lim λ(εk= 0) = 1.
Acknowledgement. Marius Iosifescu gratefully acknowledges support from Contract 2-CEx-06-11-97 of the Romanian Authority for Research.
REFERENCES
[1] J. Aaronson, An Introduction to Innite Ergodic Theory. Amer. Math. Soc., Providence, RI, 1997.
[2] K. Dajani and C. Kraaikamp, The mother of all continued fractions. Colloq. Math. 84/85 (2000), 109123.
[3] H.E. Daniels, Processes generating permutation expansions. Biometrika 49 (1962), 139149.
[4] M. Iosifescu and . Grigorescu, Dependence with Complete Connections and its Applica- tions. Cambridge Univ. Press, Cambridge, 1990.
[5] M. Iosifescu and C. Kraaikamp, Metrical Theory of Continued Fractions. Kluwer, Dor- drecht, 2002.
[6] M. Iosifescu and C. Kraaikamp, Metric properties of Denjoy's canonical continued frac- tion expansion. Tokyo J. Math. 31 (2008), 495510.
[7] S. Ito, Algorithms with mediant convergents and their metrical theory. Osaka J. Math.
26 (1989), 557578.
[8] W. Parry, Ergodic properties of some permutation processes. Biometrika 49 (1962), 151156.
Received 3 April 2008 Romanian Academy
Gheorghe MihocCaius Iacob Institute of Mathematical Statistics and Applied Mathematics
Calea 13 Septembrie nr. 13 050711 Bucharest 5, Romania
miosifes@acad.ro and
Politehnica University of Bucharest Department of Mathematics I
Splaiul Independenµei 313 060042 Bucharest, Romania
gisebe@mathem.pub.ro