You are on page 1of 9

Functional Analysis and Its Applications, Vol. 38, No. 2, pp.

79–87, 2004
Translated from Funktsional nyi Analiz i Ego Prilozheniya, Vol. 38, No. 2, pp. 1–11, 2004
Original Russian Text Copyright  c by M. O. Avdeeva

On the Statistics of Partial Quotients


of Finite Continued Fractions ∗
M. O. Avdeeva
Received February 16, 2003

Abstract. We refine the remainder estimate in the asymptotic formula, earlier obtained in a joint
paper with V. A. Bykovskii, for Arnold’s problem about Gauss–Kuzmin statistics.
Key words: continued fraction, partial quotient, convergent, Gauss–Kuzmin statistics.

Introduction
The following problem was posed by Arnold in [1, p. 17] (see also [2]).
“I. Consider the integer points (p, q) lying in the positive quarter of the circle of radius N , i.e.,
such that p2 + q 2  N 2 , p > 0, and q > 0. Let us expand each rational number α = p/q in a
continued fraction (all these fractions are finite). Let us count the numbers of ones, twos, threes,
etc., in the set of elements of all these fractions and find the corresponding frequencies, which
depend on N . Now let N be very large. Is it true that these numbers will be close to the Gaussian
probabilities  
1 1
pk = log 1 + ?”
log 2 k(k + 2)
This problem is closely related to the metric theory of Diophantine approximations, dynamical
systems, and many other areas of mathematics.
One formally represents the expansion of a rational number r in a finite continued fraction
of length s = s(r) with integer part t0 = [r] and partial quotient t1 , . . . , ts (which are positive
integers), ts  2, in the form
r = [t0 , t1 , . . . , ts ]. (0.1)
Let Hk (N ) be the total number of partial quotients ti (for all indices i, 1  i  s(r)) coinciding
with a positive integer k (for all fractions r = α in the problem posed above). Then the number
∞
H(N ) = Hk (N )
k=1
is the sum of lengths of the continued fractions of the above-mentioned rational numbers r.
In the present paper, we prove that
 
Hk (N ) 1
= pk + O (0.2)
H(N ) log N
for large N uniformly with respect to k. This refines the answer given to Arnold’s question in the
paper [3], where the weaker remainder estimate O((log N )−1/2 ) was obtained.
By definition, we set
Ai /Bi = [ti , . . . , ts ] (1  i  s), (0.3)
where Ai is an integer, Bi is a positive integer, and gcd(Ai , Bi ) = 1. Let sx (r) be the number of
indices i in the expansion (0.1) for which Bi  xAi . Since we always have 0 < Bi  Ai , one can
assume that x ∈ (0, 1]. It readily follows from (0.3) that the difference (k = 1, 2, 3, . . . )
s(k) (r) = s1/k (r) − s1/(k+1) (r) (0.4)

This work was financially supported by the RFBR (project No. 04-01-97000).

0016–2663/04/3802–0079 2004
c Plenum Publishing Corporation 79
is the number of partial quotients ti equal to k in the expansion (0.1) and


s(r) = s1 (r) = s(k) (r). (0.5)
k=1
For a positive parameter R → ∞, we set

Nx (R) = sx (a/d),
where the sum is taken over all positive integers a and d such that a2 + d2  R2 . Then, replacing
the letter N used at the beginning of the paper by R, we obtain
H(R) = N1 (R), Hk (R) = N1/k (R) − N1/(k+1) (R).
The asymptotic formula (0.2) is an immediate consequence of the following main result of the
paper.
Theorem. One has
3
Nx (R) = log(1 + x)R2 log R + O(R2 )
π
uniformly with respect to x ∈ (0, 1].
The author thanks V. A. Bykovskii for useful advice and attention.

1. The Heilbronn Correspondence


Recall that if a number r has the continued fraction expansion (0.1), then
Pi /Qi = [t0 , t1 , . . . , ti ] (1.1)
is called the ith convergent of r. Here it is assumed that Pi is an integer, Qi is a positive integer,
and gcd(Pi , Qi ) = 1. We point out that the continued fraction expansion of Pi /Qi has the canonical
form
(a) (1.1) if ti  2;
(b) Pi /Qi = [t0 , t1 , . . . , ti−1 + 1] if ti = 1.
We need the following well-known relations (see [4]):
Qi /Qi−1 = [ti , ti−1 , . . . , t1 ], Pi−1 Qi − Pi Qi−1 = (−1)i . (1.2)
Let d be a positive integer. By Ω(d) we denote the set of all d-quadruples, i.e., quadruples
ω = (m, m , n, n ) of positive integers for which
mn + m n = d, 1  m < m, 1  n < n , gcd(m, m ) = gcd(n, n ) = 1.
Heilbronn [5] constructed a bijective mapping of Ω(d) onto the set
{(a; i) | 1  a  d/2, gcd(a, d) = 1, 1  i < s(a/d)}
by the rule
ω → (a(ω); i), (1.3)
where the number a = a(ω) and the index i are determined by the expansions
n/n = [ti , . . . , t1 ], m/m = [ti+1 , . . . , ts ], (1.4)
d/a = [t1 , . . . , ts ]. (1.5)
For each ω ∈ Ω(d), we define an integer n̄ = n̄(ω) by the conditions
n̄n ≡ 1 (mod n ), −n/2 < n̄  n/2.
In conclusion of this section, note that the set Ω(d) is empty only for d = 1, 2, 3, 4, 6. Moreover,
1 < a(ω) < d/2
for all other values of d and each quadruple ω ∈ Ω(d).
80
2. Auxiliary Lemmas
Let θ ∈ (0, 1/2] and
Ωθ (d) = {ω ∈ Ω(d) | a(ω)  θd},  θ (d) = {ω ∈ Ω(d) | −θn (ω) < n̄(ω)  θn (ω)}.

 θ (d) do not coincide. However, they do not differ from each other
In general, the sets Ωθ (d) and Ω
“very much,” which underlies the proof of the following assertion playing an important role below.
Let #M be the number of elements in a set M .
Lemma 1. One has
#Ωθ (d) = #Ω θ (d) + O(d)
uniformly with respect to θ ∈ (0, 1/2].
Proof. Let
Qi = n , Qi−1 = n, Qi /Pi = [t1 , . . . , ti ]
in (1.2). Then it follows from the second equation in (1.2) that
(−1)i−1 Pi n ≡ 1 (mod n ).
Since t1  2, it follows that −n/2 < (−1)i−1 Pi  n/2 and n̄ = (−1)i−1 Pi . Thus n/|n̄| = [t1 , . . . , ti ].
Therefore, |n̄|/n is a convergent of the fraction a/d = [0, t1 , . . . , ts ], and moreover (see [4]),
 
a  
 n − |n̄| < 1 .
d  n

With regard to the conditions imposed on ω (a(ω)  θd for Ωθ (d) and −θn (ω) < n̄(ω)  θn (ω) for
 θ (d)), this implies that the total number of d-quadruples in the set-theoretic symmetric difference

 θ (d)) \ (Ωθ (d) ∩ Ω
(Ωθ (d) ∪ Ω  θ (d))
does not exceed the number of elements in the set
  
  1  1
ω ∈ Ω(d)  |n̄(ω) − θn (ω)| <  or |n̄(ω) + θn (ω)| <  . (2.1)
n (ω) n (ω)
For coprime positive integers n and n , all integer solutions of the equation mn + m n = d can be
represented as
m = m0 + nt, m = m0 − n t (t ∈ Z),
where (m0 , m0 ) is some given solution. The condition 0 < m < m is equivalent to the double
inequality
m0 − m0 m0
< t < .
n + n n
Since
m0 m0 − m0 d
− =   ,
n n + n n (n + n)
it follows that the number of elements in the set (2.1) does not exceed
(±)  d

1+   ,

n (n + n)
1<n<n <d

where the symbol (±) indicates that the sum is taken over the pairs (n, n ) such that
1 1
|n̄ − θn | < or |n̄ + θn | < .
n n
Since n  2, these inequalities imply that
1 1
|n̄ − θn | < or |n̄ + θn | < .
2 2
81
Each of these inequalities determines the number n̄ (and hence n) for a given n in at most one
way. Therefore, the desired quantity can be estimated by the sum
  d

2 1 +  2 = O(d).

(n )
1<n <d

 θ (d) differ from each other in O(d) elements. This proves Lemma 1.
Hence the sets Ωθ (d) and Ω
Let Φx (d; P ) be the number of all pairs (ω; n̄) with ω ∈ Ω(d) and n̄ ∈ Z for which
m (ω)  xm(ω), (2.2)
P P
− n (ω) < n̄  n (ω), (2.3)
d d
n̄ · n(ω) ≡ 1 (mod n (ω)). (2.4)
We point out that n̄ is an independent integer variable here.
Lemma 2. As P → ∞,

gcd(a,d)=1
sx (a/d) = Φx (d; P ) + O(P + d)
1aP

uniformly with respect to x ∈ (0, 1].


Proof. First, consider the case 0 < P  d/2. Note that the pair (Ai+1 , Bi+1 ) (see the intro-
duction) coincides with (m, m ) for 1  i < s(a/d) (see the second expansion in (1.4)). Using the
Heilbronn correspondence (1.3), we obtain


gcd(a,d)=1
sx (a/d) = #{ω ∈ ΩP/d (d) | m (ω)  xm(ω)} + O(d).
1aP

The remainder estimates the number of pairs (A1 , B1 ) that are not taken into account. Using
Lemma 1, we obtain the assertion of Lemma 2 in the case under consideration.
One can readily see that
d d
= [t1 , t2 , . . . , ts ], = [1, t1 − 1, t2 , . . . , ts ]
a d−a
whenever 1  a  d/2 and gcd(a, d) = 1. Moreover,
 a a
sx k + = sx
d d
for each integer k. Using these observations and the mod n (ω)-periodicity of the solutions n̄ of the
congruence (2.4), one extends the assertion of Lemma 2 to the other values P > d/2.

3. Application of Estimates of the Kloosterman Sums


For a positive integer q and an integer l, set

1 if l ≡ 0 (mod q),
δq (l) =
0 if l ≡
 0 (mod q).
Let Q be an arbitrary integer, and let a positive integer P range from 1 to q inclusively. For an
integer w ∈ (Q, Q + P ], we set

1 if Q < w  Q + P ,
∆q (w; P, Q) =
0 if Q + P < w  Q + q.
82
This is the characteristic function distinguishing the integers belonging to the half-interval (Q, Q +
P ] in the set of integers belonging to (Q, Q + q]. Using the discrete Fourier transform modulo q,
we obtain the expansion
  
kw
∆q (w; P, Q) = ∆q (k; P, Q) exp 2πi (3.1)
q
−q/2<kq/2

with coefficients

Q+P  
q (k; P, Q) = 1 kl
∆ exp −2πi .
q q
l=Q+1

If k ≡ 0 (mod q), then


∆ q (0; P, Q) = P .
q (k; P, Q) = ∆ (3.2)
q
Summing the geometric progression, for k ≡ 0 (mod q) we obtain

q (k; P, Q) = 1 1 − exp(−2πikP/q) exp(−2πik(Q + 1)/q).


∆ (3.3)
q 1 − exp(−2πik/q)
Equations (3.2) and (3.3) readily imply the well-known inequality

q (k; P, Q)|  1
|∆
if k = 0,
(3.4)
1/|k| if 0 < |k|  q/2.
By definition, the expression

q  
mu + nv
Sq (m, n) = δq (uv − 1) exp 2πi
q
u,v=1

for integer m and n is the Kloosterman sum modulo q. It satisfies the estimate

Sq (m, n)  q 1/2+ε gcd(m, q) (3.5)
ε

for each ε > 0 (see [6]). Since Sq (m, n) = Sq (n, m), we can replace m by n on the right-hand side
in (3.5). If m ≡ n ≡ 0 (mod q), then
Sq (m, n) = Sq (0, 0) = ϕ(q), (3.6)
where ϕ(q) is the Euler function, i.e., the number of integers k ∈ (0, q] such that gcd(k, q) = 1.
Lemma 3. Let ε be an arbitrary positive number. Then
 P1 P2
δq (uv − 1) = ϕ(q) 2 + Oε (q 1/2+ε )
q
Q1 <uQ1 +P1
Q2 <vQ2 +P2

for any Q1 , Q2 , P1 , P2 ∈ R such that 0 < P1 , P2  q.


Proof. One can readily see that it suffices to consider the case of integer Q1 and Q2 and positive
integer P1 and P2 . Using the expansion (3.1) and denoting the desired sum by Dq (P1 , P2 ; Q1 , Q2 ),
we obtain
Q
1 +q Q
2 +q

Dq (P1 , P2 ; Q1 , Q2 ) = δq (uv − 1)∆q (u; P1 , Q1 )∆q (v; P2 , Q2 )


u=Q1 +1 v=Q2 +1

= q (m; P1 , Q1 )∆
∆ q (n; P2 , Q2 )Sq (m, n).
−q/2<m,nq/2

83
Isolating the term with m = n = 0 in the last sum and using relations (3.1)–(3.6), we see that
P1 P2 
q
1 1/2+ε/2
Dq (P1 , P2 ; Q1 , Q2 ) − ϕ(q)  q gcd(m, q)
q2 ε mn
m,n=1
√  1  1
 q 1/2+ε/2 log(1 + q) · t 
 q 1/2+ε/2 log2 (1 + q) √ .
mt t
t|q 1m tq t|q

By applying the standard estimates


 1
log2 (1 + q)  q ε/4 , √  q ε/4
ε
t|q
t ε

(for the second estimate, see [7]), we obtain the assertion of Lemma 3.
Remark. Assuming that P2 in Lemma 3 is an arbitrary positive number and splitting (Q2 , Q2 +
P2 ) into subintervals of length  q, we obtain an asymptotic formula with remainder
Oε (q 1/2+ε + q −1/2+ε P2 ).
Lemma 4. Let q > 1. Then, as P → ∞,
q−1  
 
1 log(1 + x) ϕ(q)
δq (nn̄ − 1) =2 P + Oε (q −1/2+ε + P q −3/2+ε )
q + xn x q2
n=1 |n̄|P

uniformly with respect to x ∈ (0, 1] (for every ε > 0).


Proof. Using the Abel transform
       
a(n)b(n) = a(k) b(N ) + a(k) (b(n) − b(n + 1))
1nN 1kN 1n<N 1kn

and taking the remark to Lemma 3 into account, we obtain the following expression for the sum in
question:
  
 1   1 1
δq (uv − 1) + δq (uv − 1) −
q + x(q − 1) q + xn q + x(n + 1)
1u<q 1n<q−1 1u<n
|v|P |v|P
  
2P 1
= ϕ(q) + Oε q 1/2+ε + P q −1/2+ε
q q + x(q − 1)
  2P n

1 1

−1/2+ε
+ ϕ(q) 2 + Oε (q 1/2+ε
+ Pq ) −
q q + xn q + x(n + 1)
1n<q−1

2ϕ(q)  1
q−1
= P + Oε (q −1/2+ε + P q −3/2+ε ).
q2 q + xn
n=1
It remains to notice that

q−1  q    
1 dα 1 log(1 + x) 1
= +O = +O .
q + xn 0 q + xα q x q
n=1
This proves Lemma 4.
Lemma 5. Under the assumptions of Lemma 4,
 δq (nn̄ − 1) log(1 + x) ϕ(q) P
2 2
=2 3
arctg + Oε (q −5/2+ε )
(q + xn)(q + n̄ ) x q q
1n<q
|n̄|P

uniformly with respect to x ∈ (0, 1].


84
Proof. It suffices to prove the assertion of the lemma for positive integers P . Representing the
desired sum in the form
   
1 δq (nn̄ − 1)
q 2 + n̄2 q + xn
|n̄|P 1n<q

and making the Abel transform with respect to |n̄|, we obtain the following expression with the
help of Lemma 4:
    
δq (nn̄ − 1) 1
q + xn q2 + P 2
|n̄|P 1n<q


P −1     δq (nn̄ − 1)  
1 1
+ −
q + xn q 2 + N 2 q 2 + (N + 1)2
N =1 |n̄|N 1n<q
 
log(1 + x) ϕ(q) −1/2+ε −3/2+ε 1
= 2 2
P + Oε (q + Pq ) 2
x q q + P2

P −1   
log(1 + x) ϕ(q) −1/2+ε −3/2+ε 1 1
+ 2 N + Oε (q + Nq ) −
x q2 q 2 + N 2 q 2 + (N + 1)2
N =1
log(1 + x) ϕ(q) 
P
1
=2 2
+ Oε (q −5/2+ε ).
x q q + N2
2
N =1
It remains to notice that

P  P    
1 dα 1 1 P 1
= + O 2 = arctg + O 2 .
q + N2
2
0
2
q +α 2 q q q q
N =1
This completes the proof of Lemma 5.

4. Proof of the theorem


Let R be an arbitrary positive integer greater than unity. Applying Lemma 2, we see that

gcd(a,d)=1 a 
Nx∗ (R) = sx = Φx (d; R2 − d2 ) + O(R2 ).
d
a2 +d2 R2 1d<R

The last sum is modulo O(R2 ) the number of integer quintuples (m, m , n, n , n̄) such that
mn + m n  R, 1  m < xm, 1  n < n , gcd(m, m ) = gcd(n, n ) = 1,

|n̄|  (R/(mn + m n))2 − 1 · n , n̄n ≡ 1 (mod n ).
Setting n = q for convenience, we obtain the relation
 
Nx∗ (R) = δq (nn̄ − 1)G∗x (q, n, n̄; R) + O(R2 ),
1q<R 1n<q
|n̄|R

where G∗x (q, n, n̄; R) is the number of pairs of positive integers (m, m ) such that
qR
1  m  xm, mq + m n  , gcd(m, m ) = 1.
q 2 + (n̄)2
The quantity Gx (q, n, n̄; R) is determined by the same conditions except for the condition that m
and m are coprime. This quantity coincides with the number of integer points in the triangle with
vertices    
qR xqR R
(0, 0), , , ,0 .
(q + xn) q 2 + n̄2 (q + xn) q 2 + n̄2 q 2 + n̄2
85
Therefore,
 
x qR2 R
Gx (q, n, n̄; R) = (area) + O(perimeter) = +O .
2 (q + xn)(q 2 + n̄2 ) q + |n̄|
Applying the second Möbius inversion formula, we obtain

G∗x (q, n, n̄; R) = µ(t)Gx (q, n, n̄; R/t).
1tR/q
Since  −1
 µ(t) ∞
 ∞
µ(t) 1 6
= + O(Q−1 ) = + O(Q−1 ) = + O(Q−1 )
t2 t2 t2 π2
1tQ t=1 t=1
for Q  1, it follows that
3x qR2
G∗x (q, n, n̄; R) = + W,
π (q + xn)(q 2 + n̄2 )
2
where
 R R R+q
W   log .
t(q + |n̄|) q + |n̄| q
1tR/q
We finally obtain
3x 2   q
Nx∗ (R) = R δq (nn̄ − 1) + V,
π2 (q + xn)(q 2 + n̄2 )
1q<R 1n<q
|n̄|<R

where
  R R+q
V  δq (nn̄ − 1) log + O(R2 ).
q + |n̄| q
1q<R 1n<q
|n̄|<R
Since 
δq (nn̄ − 1)  1,
1n<q
it follows that
   
R 1
V R· log 1 + ·
q q + |n̄|
1q<R |n̄|<R
    R   
R dα R
R· log 1 + · =R 2
log 1 +
q 0 q+α q
1q<R 1q<R
 R    ∞
R log2 (1 + β)
R log2 1 + dα = R2 2
dβ  R2 .
0 α 1 β
Thus
3x 2 
Nx∗ (R) = R q · Ux (R; q) + O(R2 ),
π2
1q<R
where  1
Ux (R; q) = δq (nn̄ − 1) .
(q + xn)(q 2 + n̄2 )
1n<q
|n̄|<R
Applying Lemma 5, we obtain
6  ϕ(q) R
Nx∗ (R) = 2
log(1 + x)R 2
2
arctg + O(R2 ).
π q q
1q<R

Since  
R π q
arctg = +O ,
q 2 R
86
it follows that
3  ϕ(q)
Nx∗ (R) = log(1 + x)R2 + O(R2 ).
π q2
1q<R
Recall that (see [7])
 µ(t)
ϕ(q) = q .
t
t|q
Therefore,
 ϕ(q)  1  µ(t)  µ(t)  1  µ(t)   1

= = · =
q2 q t t tq  t2 q
1q<R 1q<R t|q 1q<R 1tq  <R 1t<R 1q  <R/t
 µ(t)  R
 

µ(t)

6
= log + O(1) = log R + O(1) = 2 log R + O(1).
t2 t t2 π
1t<R t=1
Hence
18
Nx∗ (R) = log(1 + x)R2 log R + O(R2 ),
π3
and we finally obtain
     18  2  2 
∗ R R R R
Nx (R) = Nx = 3
log(1 + x) log + O 2
t π t t t
1tR 1tR
  
18 1 3
= 3 log(1 + x) 2
R2 log R + O(R2 ) = log(1 + x)R2 log R + O(R2 ).
π t π
1tR
This completes the proof of the theorem.

References
1. V. I. Arnold, Continued Fractions [in Russian], MCCME, Moscow, 2000.
2. V. I. Arnold, Arnold’s Problems [in Russian], Fazis, Moscow, 2000, Problem 1993-11(C).
3. M. O. Avdeeva and V. A. Bykovskii, Solution of Arnold’s Problem on the Gauss–Kuz  min
Statistics [in Russian], Preprint, Vladivostok, Dal  nauka, 2002.
4. A. Ya. Khinchin, Continued Fractions [in Russian], Nauka, Moscow, 1961.
5. H. Heilbronn, “On the average length of a class of finite continued fractions,” In: Number
Theory and Analysis, Plenum Press, New York, 1969, pp. 87–96.
6. P. Sarnak, Some Applications of Modular Forms, Cambridge University Press, Cambridge,
1990.
7. I. M. Vinogradov, Foundations of the Theory of Numbers [in Russian], Nauka, Moscow, 1972.

Khabarovsk State Pedagogical University


Khabarovsk Division of the Institute of Applied Mathematics,
Far Eastern Division of RAS
e-mail: mariya@iam.khv.ru

Translated by A. I. Shtern

87

You might also like