PACIFIC JOURNAL OF MATHEMATICS

Vol. 78, No. 1, 1978

INTEGER MULTIPLES OF PERIODIC

CONTINUED FRACTIONS
T. W. CUSICK

This paper contains much simpler proofs of the results

of Henri Cohen (Acta Arithmetica 26 (197475), 129-148) on
the period length of the continued fraction for Na, where
N is a positive integer and a is a quadratic irrational.

l Introduction* We let [α0, al9 •] denote the simple continued

fraction whose partial quotients are the integers α< (α* > 0 for i > 0).
If a is a quadratic irrational, so that a has a periodic continued
fraction, then we put
# = [K K , K, (*<i, , »u] >
where 60, 6W •••, δm is the nonperiodic part of the continued fraction
and a19 •••, an is the period. We let P{a) = w denote the length of
the period of the expansion of a.
H. Cohen [2] defined the functions
S(N9 n) = sup P(Na)
P(a)=n

for each pair of integers N > 1, w ^ 1. The fact that S(N9 n) is

always finite was already known (see Schinzel [4]).
Let A denote the set of all real quadratic irrationals. Cohen
defined the function
R(N) - sup (S(N, n)/n) = sup (P(Na)/P(a))
n^l aeA

for each integer N > 1, and proved that R(N) is always finite. The
paper of Cohen [2] is devoted to proving various results about S(N, n)
and R(N). In particular, Cohen [2, pp. 141-147] obtained the exact
value of R(N) for infinitely many N and gave a conjecture for the
value of R(N) in all the remaining cases.
Cohen made use of an algorithm given by Mendes France [3] for
computing the continued fraction expansion of Na from the expan-
sion of a, where a is any real number. Cohen [2, §§3 and 4, pp.
132-137] devotes considerable space to showing that if one wants to
use the algorithm of Mendes France [3] in order to study P(Na) for
quadratic irrationals a, then one needs various facts about 2 by 2
matrices with integer entries taken mod JV.
It turns out that the algorithm of Mendes France [3] was already
given by A. Chatelet [1] in a different but equivalent form. The
47
48 T. W. CUSICK

Chatelet formulation of the algorithm has a great advantage as far

as the application to the problems considered by Cohen is concerned;
namely, in the Chέtelet version the algorithm is defined in terms of
2 by 2 matrices with integer entries, so the relevance of these
matrices is immediately apparent. We show below that the results
of Cohen concerning the functions S(N, n) and R(N) can all be
obtained much more simply by using the approach of Chatelet [1].

2* The Chatelet algorithm* For the convenience of the reader,

we give an exposition of the algorithm of Chatelet [1]. Proofs (all
of which are elementary) are omitted; they are given by Chatelet
[1].
We suppose that a — [α0, alf a2, •••] is a real number and that
N > 1 is an integer. We wish to determine the partial quotients of
the continued fraction for Na. We suppose "for simplicity that
infinitely many of the aύ are >N (Chatelet [1, p. 12] considers only
this situation). We may make this supposition with no loss of
generality because S(N, n) depends only on the α* taken mod N (this
is easily verified; see Cohen [2, p. 132]).
We first need the following lemma of Chatelet [1, p. 7] on
matrix factorization. We use the abbreviated notation (α) defined
for each integer a ^ 0 by

this notation was also employed by Chatelet.

LEMMA 1. Any matrix

B
with nonnegative integer entries at least three of which are positive
can be written in one of the four forms

A , (0)A, A(0) , (0)A(0)

where the matrix A is given by

A = Π (ut) , Ui ^ 1 for 1 ^ i <; n .

If P> Q> S and P> R> S, then the integers ut in the

factorization of Lemma 1 are just the successive partial quotients
in the continued fraction for P/R (we need only take care that the
 INTEGER MULTIPLES OF PERIODIC CONTINUED FRACTIONS 49

number of partial quotients is even or odd, as required, by letting

the last partial quotient be 1 if necessary). For example,

The same kind of calculation applies if P> Q > S and P> R> S
do not both hold.
Before continuing, we need the following lemma of Chatelet [1,
pp. 12-15].

LEMMA 2. Suppose δ and d are any positive integers such that

§d •=• N, and suppose k is any integer such that 0 ^ k< d. Given
any matrix

P Q
~\, PS-QR=±1, —^ t f - 1 , Q^N-1
R S] R S

with nonnegative integer entries, there exist unique nonnegative

integers A, B, C, D, ^ and dx with δλdx — N and a unique integer
kx with 0 ^ kλ < dlf such that the following matrix identity holds:

kTP A BTδ,
(1)
0 d R S C D]\β

The integers A, B, C, D are determined by

A = δ XδP -kR) , B = N-\(δP - kR)k, + (δQ - kS)δ1) ,

1
C = dRδϊ , D = dN-\kJEt + δβ) .

The integers δlf dγ and kx are determined by the conditions δ^ = N,

0 ^ kλ < dx and

δ, = (δP - kR, dμ) where μ = (δ, R) ,

k,R + δβ = 0 mod δ ,
k,(δP - kR) + δt(δQ - i S ) Ξ θ modtf.

Later on we shall mainly be interested in the following corollary,

which is proved by taking P — a, Q = R = 1, £ = 0 i n Lemma 2.

COROLLARY. Suppose δ and d are any positive integers such

that δd — N, and suppose k is any integer such that 0 ^ k < d.
Given any integer a ^ N — 1, there exist unique nonnegative integers
A, Bf C, D, δt and eZj. with δtdt = N and a unique integer kt with
0 ^ &] < dx such that the following matrix identity holds:
50 T. W. CUSICK

δ -fcΊΓα 1 A BJδ1 -&Γ

0 dJLl θ C D±0 d
The integers A, B. C, D are determined by
A = δτ\δa - k) , B = N-\(δa
1 1
C = eZSr , 2? - diV" ^ .

Tfce integers δlf dx and kt are determined by the conditions δj ^ — Nι

0 ^ kx < dλ and

! = 0 mod δ

*1

Now we can describe the algorithm for finding the partial

quotients of Na, as follows: We divide the partial quotients
α0, α,, of a into blocks, each of which begins with an α* > N
followed by other αέ's which are ^N (we can assume α0 > N with-
out loss of generality). We denote the ΐth block by
so 6ίί} > N and JV for 2 ^ j ^ n(i) .

For each block, we compute the matrix product

and define

ί = l,2f

Starting with δ = iSΓ, d — 1, & = 0 in Lemma 2, we use (1) to

define successively integers Aif Bif Ci9 Dif δi9 dt and kt (i — 1, 2, •)>
as follows:

(2)

In this way we obtain a sequence of matrices Mt with entries

0
5 f , Cif Di. By Lemma 1, we can factor each of these matrices
1 as follows:
(3) if
 INTEGER MULTIPLES OF PERIODIC CONTINUED FRACTIONS 51

Thus we obtain a sequence of nonnegative integers

ι
(4) u[ >, , < \ } , u?\ ,<>2), , u?\ , uί%)f . . . .
We modify this sequence, if necessary, by replacing every triple
α, 0, b by the single integer a + δ, and repeating this until a se-
quence of only positive integers is obtained. This new sequence is
precisely the sequence of partial quotients for Na.

REMARK. In the discussion of this algorithm given by Mendes

France [3], the sequence corresponding to (4) may contain some
members equal to — 1 in addition to some members equal to 0. This
is because Mendes France does not make the simplifying assumption
that infinitely many of the partial quotients at of a are > N, as we
did at the beginning of this section.
From now on, it will be convenient to make the following even
stronger

ASSUMPTION. Suppose that all of the partial quotients of a =

[α0, aίf •] satisfy α< ^ 2N.
As we remarked earlier, this assumption can be made with no
loss of generality in the study of the functions S(N9 n) and R(N).
The assumption means that the blocks of partial quotients
mentioned above are all of length one, so in (2) we have P, = α ^ ,
Q. = jRt = l, Si = 0 for i = 1, 2, -. Also, by Lemma 2 Corollary,
the integers δt, di9 kt in (2) are determined recursively as follows:

(5) δo = N, do = l , ko = O;

{6 ) δt = ( ^ - A - i - A?*-!, rfi-J for i ^ 1

(7 ) kt = 0 mod δ^, for i ^ 1

( 8 )
-A/^-i^-i-fci-Λs! mod-^- for i^l.

In view of (7), we can define integers ί, (i = 1, 2, •••) by

Under our assumption, it is a simple matter to verify that the

algorithm described by Cohen [2, §2] is the same as the Chatelet
algorithm described above. The formulas (5), (6), (7), (8) above cor-
respond to Cohen [2, formulas (1), p. 130]. Cohen's δt corresponds
to δ ί + 1 above, Cohen's df corresponds to dt above, Cohen's ct cor-
responds to δtdt — hi and Cohen's (cjδt)"1 corresponds to —t < + 1 defined
in (9).
52 T. W. CUSICK

We close this section with the following lemma, which we need

later on.

LEMMA 3. In the sequence of identities (2), we have (8i~19 δt) = 1

for each i — 1, 2, .

Proof. Suppose that for some i, a prime p divides (<?*_!, δt).

Then by (6) p divides δ^^^ — k^19 so p divides k^ = δ^i-\
Hence either p divides 8^t or p divides ί^. But in the latter case
r 1
we have p divides 2Ϋ8rΛ = d^ (from (6)) and (*,_„ iS/ (δi_Λ-2)~ ) = 1
(from (8) and (9)), so that p divides δ^2 also. Hence if p divides
(£<_!, δj, then p divides <5y for every j <; ί; but this is a contradic-
tion, since δ1 = 1 by (5) and (6).

3* Upper bounds for S(N, n) and R(N). In this section we

use our previous work to give a much simpler proof of certain upper
bounds on S(N, n) and R(N) given by Cohen [2, Theorem 4.3, p.
136].
For each rational number x = [α0, alf , αw], αw ^ 2, we use
Cohen's [2, p. 129] notation L(%) to denote the number of partial
quotients in that continued fraction expansion of x which has an
odd number of partial quotients; thus L(x) = n + 1 if n is even and
L(x) — n + 2 if n is odd.
Now suppose JV > 1 is a given integer and

(10) a = [60, &!, , δw, e19 , c%] = [α0, α1? α2, •]

is a given quadratic irrational for which the Assumption of §2 holds.
It is easily verified that the Assumption implies that At > Bt > Dt
and At > Ct > Di for each matrix Mt (i = 1, 2, •) in (2). Hence
(see the remarks after Lemma 1) in the factorization (3) of Mt each
uψ is positive, so the unmodified sequence (4) is the sequence of
partial quotients of Na. In fact, the sequence (4) is just the
sequences of partial quotients of the rational numbers AJCt (i =
1,2, •••) taken in order, where the continued fraction expansion
used for each AJCi is the one with an odd number of partial
quotients (this is because the determinant of each Mt is — 1, so the
corresponding factorization given by Lemma 1 has an odd number
of matrices).
In is clear from the work of §2 that the sequence of triples
(αif kif d^ (i = 0, 1, 2, •) is eventually periodic, and thus the sequence
of rational numbers AJCi (i = l, 2, •) is also eventually periodic. Say
Am+ι/Cm+1, , Am+r/Cm+r is the period of the latter; then the length
of the period of Na is given by
 INTEGER MULTIPLES OF PERIODIC CONTINUED FRACTIONS 53

(11) P(Na) = ±L(Am+i/Cm+i) .

Our next lemma shows that P(Na) can also be expressed in terms
of LOcJdi).

LEMMA 4. Suppose p and q are two relatively prime positive

integers. Define p*, 0 < p* < q, by pp* Ξ= — 1 mod q. Then L(p/q) =
Up*/q).

Proof. We assume p < q with no loss of generality. Let p/q =

[0, Λ, •,/»], Λ ^ 2, and define pjq, = [0, Λ, , /«] for 1 ^ i ^ w.
1
First suppose that p/q < 1/2. We have pnqn-.x — pn-ιqn = ( — I)*" , so
p g ^ == ( — ly^modq. Thus if w is even, then qn^ = p*. But
as
(12) = L = [0,/„/.-„ ••-,/,]

so L(p/q) = L(p*/q) if w is even.

Now define p', 0 < ^' < q, by pp' = 1 mod q (so p' + p* = g). It
is easy to see that pf\q < 1/2 if and only if n is odd. Thus when
n is odd we have pf — q%_19 so by (12)

^ * — 1 — fl*-1 — Γ0 1 f —I f ... fl
Q Q

Hence L(p/q) = L(p*/q) = n + 2. Similar arguments take care of the

case p/q > 1/2, so the lemma is proved.

COROLLARY. For each i = 1, 2, , L(At/Ci) = L(kjdi).

Proof. We take p = ( 3 ^ ^ . ! - fc^OSr1 - Λ and g =

Ciβ Then p* = fc4δrΛ by (8) and p*/g =fci/cίi,so the corollary follows
from the lemma.
It follows from Lemma 4 Corollary and (11) that

(13)

We have from (8) and (9)

where
£ ^ A V I
Thus if we define sets Γ(m) for each positive integer m by T(m) =
{(mw m2): mlf m2 positive integers such that (mw m2) = 1 and m ^ = m}
(so that if m has A; ^ 0 distinct prime divisors, then T(m) has 2k
members), then (using Lemma 3 and (14)) we see that all possible
54 T. W. CUSICK

values of kjdi9 one for each of the different pairs ku dίf are con-
tained in the set
1:ab m divides N a δ 6T
ΛrΓ
(ab)" <
% = > ( > ) ^> fa -%)
\ ab=/ A )
Note that Ct(N) will contain repeated elements.
1
LEMMA5. The set C^N) has f(N) = N JJ (1 + p' ) elements,
where the product is taken over all distinct primes p which divide
N.

Proof. Define W{m) for positive integers m by W(l) = 1,

fc
WΉίW Pi*) — 2 , where the p/s are distinct primes and the α/s are
positive integers. It follows from the principle of inclusion and
exclusion that

and the left-hand side is just the number of elements in Cλ(N).

It turns out that the set Ct(N) is the same as the set

C2(N) = ί — : e divides N, 0 ^ α' < N/c, a' = a mod (N/c),

I JV/c
where a lies exactly once in each residue class mod
(N/c) such that (a, c) = 1 is possible! .

LEMMA 6. The sets C^N) and C2(N) are identical for each

Proof. Cohen [2, Proposition 3.4, p. 134] proved that the number
of elements in C2(N) is the number f(N) of Lemma 5. It is easily
seen that the map from C2(N) —• C^N) given by

w
w hh ee rr ee

N/c N(c(a', N/c))-1 ab = c(a', N/c)

is into and one-to-one. Since Ct(N) and C2(N) have the same number
of elements, this proves the lemma.

THEOREM 1. For each n ^ 1 and each N > 1, we have

(15) S(N, n)^n Σ L(u) , R(N) £ Σ I'M .

If N — p% s ^ 1, /or α prime p, then the latter estimate becomes

 INTEGER MULTIPLES OF PERIODIC CONTINUED FRACTIONS 55

(16) &) Σ ( Σ 4
< \p8/ i=o \p

Proof. We consider the number a, with a period of length n,

given by (10). We have already seen that the periodicity of the
sequence of triples (aif kt, dt) (i = 1, 2, ) leads to the formulas (11)
and (13) for P(Na). Evidently the period of the (aif kif cQ is as
f8
long as possible if each of the n a in the period of a occurs with
each of the possible different pairs kit d^ thus by Lemma 5 the
longest possible period length for the (aif kif dt) is nf(N). This fact
and (13) lead at once to the estimates (15). The estimate (16) follows
8
because it is easy to see that the set C^p ) is made up of the
s 8 ι 8 s 8 1 8 1
p + p ~ numbers i/p (0 ^ i < p ) and i/p ' (0 ^ i < p " ).
Theorem 1 is the same as Cohen's Theorem 4.3 and Corollary 4.4
[2, pp. 136-137]; but the states his estimates in terms of CZ(N)
instead of

4* Periodicity properties of matrices* For each integer N > 1,

we define a multiplicative group Γ(N) of 2 by 2 unimodular matrices
with integer entries by
a 6Ί
: ad — be — ± 1 , b = c ΞΞ a — d = 0 mod N\
c d]

The same notation is used by Cohen [2, p. 132]; for any 2 by 2

unimodular matrix M, he also defined [2, p. 135] λo(iV, M) to be the
smallest positive integer such that MX°{N>M) belongs to Γ(N).
We can associate a unimodular matrix M with the quadratic
irrational a given in (10) as follows:
a 61
(17) M = foXc,) (β J - J , say
c d
We call M the matrix of a or the matrix of the period clf , cn
or the matrix of the continued fraction [a19 •••, an]. We have

(18) 2L - [αlf -..,(*.] and A = [αlf , α._J

c a
(here 6 = 1, d = 0 if n = 1). In view of (18), we see that this
definition of the matrix of a is the same as the one given by Cohen
[2, p. 129].
Cohen showed the relevance of λo(iSΓ, M) to the study of S(N, n)
and R{N). The role of λo(iSΓ, M) is made really clear by the use of
the Ch&telet algorithm of §2. Before exploring this further, we
need the following lemmas.
56 T. W. CUSICK

LEMMA 6. Let

a b~ a{M) b{M)~\
(19) M= ad — be = ( — l)n = ε , say ,
c d d(M)\ '
be a unimodular matrix with integer entries. Define sequences r(i)
and s(i) by
r(l) = 1, r(2) = a + d, , r{i) = (α + ώ)r(i - 1) - εr(i - 2) (i ^ 3)

8(1) - l/2(α + d), 8(2) = l/2(α + d)2 - ε, . ,

s(i) = (α + d)β(i - 1) - εs(i - 2) (i ^ 3) .
/or eαcfc positive integer k
a(Mk) = β(fc) + l/2(α - d)r(jk) , δ(ΛΓ*) = br(k) ,
c(Mfc) = cr(fc) , d(Mk) - β(fc) - l/2(α - d)r(k) .

Proof. This is Lemma 5.4 of Cohen [2, p. 139].

LEMMA 7. Let p be a prime and suppose M is given by (19).

Define
D = D(M) = (a - df + 46c = (α + 4e = (α + d)2 + 4 ( -
Then Mλ belongs to Γ(N) (i.e., X0(p% M) divides λ) for the value of
λ given in the following tables, where (D/p) is a Jacobi symbol:
(a) ifp>2

+1
(?) - (f) = o \p/
l/2n(p — 1) odd P-KP -1) impossible P-\P + 1)
l/2n(p - 1 ) even l/2p-(P - 1) p l/2ί>«-ι(ί> + 1)

(b) i/p =

2) = - 1 (i.e., ΰ s δ m o d 8 )
p/
s=1 s =2 s^ 3

» odd 3 6 3 2'-2
8 3
» even 3 3 32 -
 INTEGER MULTIPLES OF PERIODIC CONTINUED FRACTIONS 57

if D ΞΞ 0 mod 8 , λ = 2
8

8 1
i/ JD s 4 mod 8 , λ = 2 /or ί = = 2" /or * ^ 2

Proo/. This is Theorem 5.3 of Cohen [2, pp. 137-138].

Our next lemma shows how X0(p, M)f where p is prime and M
is defined by (17), is related to periodicity properties of the algorithm
(2). We here confine ourselves to the case N = p, p prime, because
the results are simplest in that case.

8. Suppose M is given by (19), and let p be a prime

LEMMA
which does not divide the entry c in M. Define

p 0 1 -Af
A= (0 £ k ^ p - 1) .
0 1 β pj

Suppose

p 0 Ά B
(20)
0 1 C D
for some n and some P in A. Then Mn is in Γ(p) if and only if

Lo ij

Proof. Suppose M* is in Γ(p), but P does not have the form

asserted in the lemma, i.e.,
1
P = r -* for some A;.
LO p

Then (20) gives

n n
pa(M ) pb(M Y\ YA -JcA + pBl
n n
AM ) d(M )\ ~[c -fcC + pDJ '
so c(Mn) = C, whence p divides C, since p divides c(Mn) because Mn
is in Γ(p). But this means p also divides d(Mn) = ~kC + pD, which
contradicts
a(Mn)d{M«) - b(Mn)e(Mn) = ± U α(MΛ)rf(ilίw) mod p .
Now suppose P does have the form asserted in the lemma. Then
(20) gives
58 T. W. CUSICK

Γpa(M*) pb(Mn) ΓpA B

ipC D
so c(Mn) = pC, whence p divides c(Mn). By Lemma 6, b(Mn) = br(n)
and c(Mn) = cr(n); now p divides r(ri) since p does not divide c, so
also p divides b{Mn). It also follows from Lemma 6 that p divides
a(Mn) — d(Mn). Hence P is in Γ(p), and the proof of Lemn^a 8 is
complete.
Now we are in a position to give the exact value for R(p), p
prime. We use the abbreviated notation

of Cohen [2, p. 141].

THEOREM 2. We have 22(2) = F(2) L = 5. If p is an odd

prime, then
(21) R(p) - + 1 if p = I mod 4 ,
(22) R(p) = mod 4 .

Proof. Suppose the Assumption of §2 holds and suppose a is a

quadratic irrational with period length n and continued fraction ex-
pansion given by (10). We saw in §2 that under these conditions
the sequence of identities (2) holds with the matrices Mt there equal
to (di-i) (i = 1, 2, •••)• We saw in §3 that the sequence of triples
(ai9ki9di) (i - 0, 1, 2, •••) is eventually periodic. Let us suppose
that a, with period length n, has been chosen so that the period
length of the sequence of triples (aif kif dt) is maximal, say equal to
r. Since each c* in the period of a given in (10) can be associated
with at most p + 1 different pairs ki9 di (namely, those correspond-
ing to the p + 1 matrices in the set Δ of Lemma 8), we have r <J
n(p + 1).
Suppose r = n(p + 1) does occur, and that the sequence of r
identities

p 0
I+ί
0
(23)
•>I+r p 0"
0 +r D I+r. 0 1

is a typical periodic part of the sequence (2) (of course, this means
§I+r = p, fcI+r = 0, dI+r — 1, as indicated in (23)). If we multiply on
 INTEGER MULTIPLES OF PERIODIC CONTINUED FRACTIONS 59

the right in the first equation of (23) by (α / + 1 ), (α /+8 ), •••, (aI+r^) in

order, and after the ith such multiplication use equation i + 1 of
(23) f or 1 ^ i <; r - 1, we obtain

o lj Lc DjLo
for some A, B, C, D, with
(24) M = (α7)(α7+1) (α^J .
p+1
It follows from Lemma 8 that M is in Γ(p), so r = w(p + 1)
is possible if and only if there exists a matrix M of form (24) such
that λo(p, Af) = p + 1. By Lemma 7, X0(p, Λf) = p + 1 is possible for
p = 2 and for any p = 3 mod 4, but not for p ^ l mod 4. An easy
calculation shows that λo(2, M) = 3 for ikf = (3)% (w = 1, 2, •)• Thus,
by (13), S(2, w) = 5n for all w and R(2) = 5. For any p = 3 mod 4,
it is also possible to find M such that X0(p, M) — p + 1, but only
when % is odd (by Lemma 7). In fact, if n is odd we can take
M = (α)(2p) (2p) (w — 1 factors (2p)), where a is defined by α Ξ
z+£modp; here z — uΛ vi is any generator of the group of numbers
x + iy, x and # integers, with norm ± l m o d p (this group has
2(p + 1) elements and φ(2(p + 1)) generators, where φ is Euler's
function). A proof that this choice of M satisfies λo(p, M) = p + I
was given by Cohen [2, pp. 142-143] (note that there is an incorrect
factor of 1/2 in the congruence defining a [2, p. 142]). Thus, by
(13), we have S(p, n) = (F(p) + l)n whenever p Ξ 3 mod 4 and n is
odd. Since always S(p, n) ^ (F(p) + ΐ)n by Theorem 1, this proves
(21).
If p = 1 mod 4, then by Lemma 7 the largest possible value of
r is wp, and this attained if and only if p divides D(M). Hence
S(p, ri) ^ nF(p); equality actually holds here for n even because
r = np when M = (a)(2p) (2p) (n — 1 factors (2p)), where a
satisfies a2 + 4 = mod p. This is stated without proof by Cohen [2,
p. 145]. A proof using (13) can easily be given by considering the
sequence of triples (aif kif dt) which arises from (2) for this choice
of M. It turns out that each of the p — 1 pairs (kif dt) = (&, p) with
1 ^ k ^ P — 1 occurs n times among the triples (aif ki9 d^) in a period,
and the remaining p triples in the period have the form (2p, 0, 1)
or (2p, 0, p), except for one triple (1, 0, p) Thus we have (22), and
this completes the proof of Theorem 2.

REMARK. Theorem 2 shows that the estimate (16) of Theorem

1 holds with equality when s = 1 and p = 2 or p = 3 mod 4. As
Cohen [2, Corollary 6.5, p. 144] remarked, the only other cases in
which (16) holds with equality are those in the following theorem.
60 T. W. CUSICK

THEOREM 3. Let p be a prime such that p s 3 mod 4. Then for

each s ^ 1 we have

(25) R(ps) = F(p )

Ifp = 7 mod 12, then also
(26) R(2p8) = F(2p8) + F(2p*~ι) + F(ps)
We also have JB(4) = 14 and 22(6) = 28.

Proof. First suppose p Ξ= 3 mod 4. By a generalization of the

argument used in the proof of Theorem 1 to establish (21), we see
that (25) holds if, for each odd n, we can find a matrix M of form
(24) such that λo(p , M) = p'~\p + 1). In fact, the matrix M used
for the case * = 1 in the proof of (21) also suffices for any s > 1
(see Cohen [2, pp. 142-143]).
Now suppose p = 1 mod 12. In this case we see that (26) holds
if, for each odd n, we can find a matrix M of form (24) such that
\(2p% M) — 3p*~\p + 1). It is easy to deduce the existence of such
a matrix (see Cohen [2, p. 144]) from the existence of M with
X0(p% M) = p-\p + 1).
Finally, we evaluate 22(4) and 22(6) by special arguments similar
to the one used to show 22(2) = 5 in the proof of Theorem 1.

5* Concluding remarks* In the final part of his paper, Cohen

[2, §§7 and 8, pp. 144-147] gave several conjectures, including
conjectures for the exact values of S(N, n) when n is even and N
is arbitrary, and for the exact values of 22(i\O when N is arbitrary.
These conjectures can certainly be approached via the Chatelet
algorithm as described above, but it seems that considerable calcula-
tion might be necessary in order to make progress. We do not go
into these questions here.

REFERENCES

1. A. Chatelet, Contribution a la theorie des fractions continues arithmetiques, Bull.

Soc. Math. France, 4 0 (1912), 1-25.
2. H. Cohen, Multiplication par un entier d'une fraction continue periodique, Acta
Arith., 2 6 (1974-75), 129-148.
3. M. Mendes France, Sur les fractions continues limitees, Acta Arith., 2 3 (1973),
207-215.
4. A. Schinzel, On some problems of the arithmetical theory of continued fractions,
I, Acta. Arith., 6 (1961), 394-413; II, Ibid. 7 (1962), 288-298.

Received October 8, 1977.

STATE UNIVERSITY OP NEW YORK AT BUFFALO

BUFFALO, NY 14214