A CANTOR FUNCTION CONSTRUCTED BY

CONTINUED FRACTIONS
F. HERZOG AND B. H. BISSINGER

1. Introduction. The well known Cantor function was treated in a

paper by Hille and Tamarkin [5 J1, who collected together a few prop-
erties of that function, which is often used as an example in the theory
of functions of a real variable. In the definition of this function, the
representation of the numbers of the closed interval [0, l ] in the
arithmetic scales of radices 3 and 2 (ternary and binary scales of
notation) was used. Gilman [3] generalized the above function by
replacing the radices 3 and 2 by ce and /3, respectively, where a and /?
are integers such that a >/3 > 1 and a — 1 is divisible by /3 — 1. He thus
obtained a whole class of functions of the Cantor type.
The present paper gives an example of a function of the same type,
based on the representation of numbers as simple continued fractions.
While in the main our function has the same properties as those
treated in [5] and [3], there are also some interesting differences be-
tween them, to some of which we shall call attention as the occasion
arises. Since some of the results of [S] are a special case of those of
[3] (with a = 3, /3 = 2), it will in most cases suffice to compare our re-
sults with those of [3 ].
2. The set E. Let 0 < # < 1 . If the positive integers au a2, a3, • • •
are successively the denominators (partial quotients) appearing in the
expansion of x in a simple continued fraction, we shall write
1 1 1
# = — — — = [ai, G2, 03, • • • J.
ai + a2 + az + - - •
In the case of a rational x, 0 <x < 1, that is, of a finite simple continued
fraction, we have two such expansions, namely, x = {#i, • • -, a w -i, an}
= {#1, * * • , dn-iy an — l> l } , where a n ^ 2 . Both of these expansions
will be admitted in this paper, also the expansion 1 = {1}.
Throughout the paper we shall use the letter c (with or without
subscript) to denote positive integers.
We now define the set E as the set consisting of x — 0 and of all x,
0 < x < l , which can be expanded in a simple continued fraction none of
whose denominators equals unity. In other words, E contains x = 0
and all finite and infinite simple continued fractions of the form
Received by the editors May 3, 1946.
1
Numbers in brackets refer to the bibliography at the end of the paper.

104
 A CANTOR FUNCTION 105

x={ci+l, C2+I, - • • }. The above wording is to imply that x

= { c i + 1 , • • • , Cn + 1} does belong to E although it can also be writ-
ten a s x = { c i + 1 , • • • , Cn, l } .
The set complementary to E with respect to the interval [0, 1 ] will
be denoted by D. Thus D consists of x = 1 and of all open intervals
with end points £ and £', where 2

£ = {d + 1, c2 + 1, • • • , <V-i + 1, 1}
(1) = {d + 1, C2+ 1, • • • , C n _ i + 2 } ,

Such an interval will be called a Dn-interval in the following, (For ex-

amples of P o i n t e r vals, see the table at the end of the paper.) There
is one J9i-interval, (1/2, 1), but infinitely many jD n -intervals for each
n ^ 2 . I t is easily seen t h a t any two D n -intervals (with equal or un-
equal values of n) do not overlap and do not even have an end point
in common.
The points of E (except for x = 0) can be divided into two classes,
ER and Ej. ER is to contain the rational # £ E , other than zero, and
Ei the irrational ^ £ £ It is clear from the above that x £ ER if and
only if x is the end point of a 2} n -interval (except x = l, which does
not belong to E a t all). x = 0 will be considered as belonging neither
to ER nor to Ej.
We shall now derive a few of the main properties of the set E.
(2.1) E is a closed set.
Consider D as the union of the interval l / 2 < # r g l and all open
D n -intervals, n = 2, 3, • • • . All these intervals are open sets with re-
spect to [O, 1 ]. Thus D is an open set and E is a closed one.
(2.2) E is a perfect set.
The following three relations, referring to the cases # = 0, XÇZER,
X G E J , respectively, show that E is perfect. (In each case m—»+«>
through integral values.)

0 = lim {w},

{ci + 1, c2 + 1, • • • , cn+ l} = lim {ci + 1, c2 + 1, • • • , cn + 1, m),

m
{ci + 1, c2 + 1, • • • } = Km {ci + 1, c2 + 1, • • • , cm + l } .
m

2
If w = l the numbers in (1) are to be interpreted as meaning £= {l} = 1 and
$ ' - {2} =1/2. We also remark t h a t £ < £ ' (£>£') when «is even (odd). See [8, p. 41,
Theorem 8].
106 F. HERZOG AND B. H. BISSINGER [February

(2.3) E has the power of the continuum.

E contains as a subset the set of all infinite simple continued frac-
tions { Ê I + 1 , £2+1, * ' * } with each Ct = l or 2.
(2.4) E is of Lebesgue measure zero.
This is a result due to F. Bernstein [l, p. 429, Remark]. (See also
[4, p. 334, corollary].)
(2.5) E is nowhere dense in [0, 1 ].
The complementary set D is by (2.4) of measure one, hence every-
where dense in [0, l ] . It is also open in that interval by (2.1). Hence
(see [2, p. 64, definition]) Eis nowhere dense in [0, l ] .
Comparing our set E with the corresponding sets Pa$ s=P of [3], we
see that they agree in the five properties, derived above. The subsets
ER and Ei of E correspond, respectively, to the subsets P~+P+ and
P° of P. (See [3, p. 434].) However, while P is symmetric about
x = l / 2 , E has no such symmetry. A more essential difference is that
the intervals dPk of [5, p. 255] for fixed p are finite in number and of
equal length, while in our case we have infinitely many Z>n-in ter vals
for fixed n ^ 2 and their lengths are not equal.
3. Definition and main properties of <£(#)• We shall first define our
function for # £ 2 ? by the following three equations, which apply to
the cases # = 0, XÇZER, xt-Ei, respectively.
(2a) 0(0) = 0.
(2b) x = {ex + 1, c2 + 1, • • • , cn + 1} : <t>(x) » \ch c2i • • • , cn).
(2c) x = {d + 1, c2 + 1, • • • } : <l>(x) = {ch c2, • • • }.
Before extending this definition to values of # £ D , we first define
<£(1) = 1 and show that the values of <t>(x) at both end points of each
/^„-interval are equal to one another. This is true for the J9i-interval
(1/2, 1), since 0(1/2) =<£( {2} ) = {1} = 1 = 0 ( 1 ) , and for a D n -interval
with w ^ 2 w e see from (2b) that, when £ and J' are given by (1), we
have
0 ( 0 = {Cl> C2, — • , <V_i + 1} = [d, C2t • • • , Cn-l, 1} = <K£')-

We are now able to define <t>(x) for any interior point of a jD n -interval
as the common value of <j>(x) a t the end points of that interval. (For
examples, see the table at the end of the paper.) The definition of 0(x)
in [O, 1 ] is thus complete.
We shall now derive some of the more important properties of <j>(x).
(3.1) <j>(x) is constant in every (closed) Dn-interval. The union of the
intervals of constancy of 4>(x) is a set of measure one.
19471 A CANTOR FUNCTION 107

This follows at once from the definition of <£(#) for # £ D and from
(2.4).
(3.2) <j>(x) is a nondecreasing function of x in [0, l ] ; moreover, when
0^x<x' ^ 1 then <t>(x) ^#(#')> the equality holding if and only if x and
xr lie in the same {closed) Dn-interval.
On account of (3.1), it suffices to consider the case in which x and
x' belong to E. Since by (2a) the case x = 0 is trivial we may assume
furthermore t h a t 0<x<x' ^ 1 / 2 . Now let x and x' be written in the
form (2b) or (2c) and assume that they agree in the first m — 1 de-
nominators, but no further (m ^ 1). Then when m is even (odd) either
the mth denominator of x' is larger (smaller) than that of x, or x' (x)
has only m — 1 denominators. (See [8, pp. 40-41, Theorems 7 and 8].)
The definitions (2b) and (2c) of (j>(x) are such that these same rela-
tions will hold true for the simple continued fractions for <f>(x) and
<j>(x'), so t h a t in general 4>{x) <<f>(x'). And it is easily seen that the
only exceptional case, leading to 4>{x) =<£(#'), is t h a t in which x and
x' are the smaller and larger, respectively, of the numbers £ and £'
(with n*£2) of (1). This completes the proof.
(3.3) The function y—<f>(x) assumes every value of y of the interval
O ^ y ^ l . Moreover, if y is rational and 0 < ; y ^ l then y is assumed by
4>(x) in one whole Dn-interval, while y = 0 and every irrational y,
0 <y < 1, is assumed at one point x.
This follows a t once from the definitions (2a), (2b) and (2c) of the
function (j>(x) and its definition for xÇzD.
(3.4) <t>(x) is continuous in [0, l ] .
A function ƒ(x) which is defined and monotone in a closed interval
[a, b] is continuous in t h a t interval if and only if it assumes every
value between f (a) and ƒ(&). Thus (3.4) follows from (3.2) and (3.3).
(3.5) For every positive}^, the\-variation (see [2, p. 511, Definition])
of <j>{x) in [0, l ] is equal to unity \ <f>{x) is not absolutely continuous.
We shall show the first part of (3.5) by proving the more general
theorem below, whose hypotheses are satisfied for 4>{x) by (3.2) and
(3.1). The second part of (3.5) follows from the first, but could be
directly concluded from the fact t h a t fl<j>f{x)dx = 0, but<£(l) — 0(0) = 1.
T H E O R E M 1. If f(x) is monotone* in the closed interval [a, b] and if
the intervals of constancy of f(x) cover [a, b ] except for a set of measure
zero, then the ^variation of f{x) in [a, b] equals \f(b)—f(a)\ for every
positive X.
Let f(x) be nondecreasing. (If ƒ(#) is nonincreasing consider the
3
Continuity of f(x) is not assumed.
108 F. HERZOG AND B. H. BISSINGER [February

function —fix).) Let X > 0 be given. We denote the intervals of con-

stancy of fix), of which there are at most denumerably many, in
some order by I,- (j = l, 2, • • • ) and their lengths by |l,-|. Since
22/1-^1 ^"""fl we can in any case choose k such that ]C*-i|^l >&
— a —X. We now replace these k intervals Ij by intervals J/* such that
the end points of ! ƒ are inner points of Ij (i = l, 2, • • • , k) and
such that we still have ]C*-i| Ij*\ >b — a—\. Let these k inter-
vals I f be denoted in their natural order (from left to right) by
(xi, x 2 ), (#3, Xi)$ • • • > (^2*-i, #2*). We know then that/(#2*-i) s=f(x2i) for
i = l, 2, • • • , jfe. The complementary intervals (a, #i), (x2, #3), • • • ,
(ff2*-2, #2*-i), (#2*> &) have a combined length less thanX, while (with
x0 = a,X2k+i:=b)

E [/(*tw) - ƒ(*«) ] = ƒ(») - ƒ(<*)•

This completes the proof of Theorem 1.
(3.6) The curve y=<f>(x) is of arc length4 2.
This will be proved by the following theorem. Its hypotheses are
satisfied by <j>(x) on account of (3.2) and (3.5).
THEOREM 2. If fix) is monotone1 in the closed interval [a, 6] and if
its \-variation in [a, b] equals \f(b) —f(a) \ for every positive X, then the
arc length4 of the curve y —fix) is b — a+ \f(b) —fia) |.
L e t / ( x ) be nondecreasing. T h a t the curve y—fix) is rectifiable fol-
lows from the fact t h a t ƒ(x) is of bounded variation. Since the length
of any inscribed polygon can obviously not exceed b — a + ƒ (6) —ƒ (a)
it suffices to show the existence of an inscribed polygon of length
greater than 6 —a+ƒ(&)—ƒ (a) —€ for every given positive €. By the
hypothesis, we can select a set of intervals (xi> x 2 ), (#3> au), • • • ,
ix2h-i, X2h) with Xj<Xj+i C/ = l, 2, • • • , 2h—l) such that
h €

(3) X ) ( * 2 , - - X2i-l) < —>

(4) E [ƒ(***) - /(*2,--i)] > fib) - fia) - 4 *

Let Pt- be the point (a\-, ƒ(#;)), i = 1, 2, • • • , 2&, and form the polygon
4
If ^ / ( a ) is of bounded variation in [a, b] then the two definitions of arc length
by Peano and by Jordan and Scheeffer (see [6, p. 338]) are equivalent for functions
f(x) which have the additional property that for each x, a<x<b, either fix—) ^f(x)
^ / ( * + ) or f(x-)^f(x)^f(x+). (See [7, pp. 300-302].) Monotone functions obvi-
ously have this property. The definition used in the proof of Theorem 2 is Peano's
definition.
19473 A CANTOR FUNCTION 109

P0P1P2 • • • P * J W i > where 6 P 0 = (a, f (a)) and P 2M .i = (&,ƒ(&)). By (3)

the projections of the segments P0P1, P2P3, • • •, P2A-2 Pin-i, PuPth+i
on the #-axis have a combined length greater than b — a — e/2; and
by (4) the projections of the segments P1P2, P3P4, • • • , Pih-iP2h on
the y-axis have a combined length greater than f(b) —jr(a) — e/2. This
completes the proof of Theorem 2.
The function <£(#) has the properties, which were derived in this
section, in common with the functions ü)ap(x) z=œ(x) of [3]. In fact,
properties (3.5) and (3.6) were proved simultaneously for all of these
functions by Theorems 1 and 2, respectively. I t is to be observed,
however, t h a t according to (3.3) our curve y=4>(x) has all rational
numbers y, 0 < y ^ l , as ordinates of its horizontal segments, while
the functions o)ap(x) of [3] have as such only certain rational numbers
y y 0 <y < 1, namely, those of the form y = rn/l3r, r and m being positive
integers and j3 being the radix, used for the values of y.
4. The derivatives of <£(#). The problem of determining the deriva-
tives of the functions o)ap(x) of [3] was dealt with by Gilman [3,
pp. 438-442], who obtained conditions for the existence of the deriva-
tive on the left and t h a t on the right a t the points of the set Pa$.
The various possibilities are listed on p. 442 of [3]. The corresponding
problem for our function cj>(x) seems to be more difficult. Except for
one case (see footnote 7 below), we shall obtain the values of the four
derivatives of <t>(x) at all points x of [0, 1 ].
We introduce first the following notations. In this section x 0
= { c i + 1 , C2 + I, • • • } will always denote a fixed number such that
XotEiE, #0 7^0. If the simple continued fraction for XQ has at least n
denominators we shall denote the nth. complete quotient of x 0 by
Cn+l+sn, where sn = 0 if x0 has only n denominators and otherwise
(5) Sn = {Cn+l + 1, Cn+2 + 1, • • • }.
In either case we have snÇzE, so that 0 ^ s w ^ l / 2 , and we conclude
from [8, p. 27, Rule A] t h a t
(6) Xo = {Ci + 1, C2 + 1, • • • , Cn_i + l,Cn+ 1+ Sn}.
I t follows readily from (2a), (2b) and (2c) in conjunction with (5) that
(7) <K#o) = {ch c2, • • • , cn_i, cn + *($»)}•
We denote the mth convergent of the simple continued fraction for x0
by Pm/Qm and t h a t for #(x 0 ) by 6 pm/qm (m = l, 2, • • • ) and obtain
6
If Xi—a the polygon is to start with Pi and if X2h — b it is to end with P 2 J».
6
As usual we p u t P _ i = £ _ i = Q0 = go = 1 and P 0 = p o Œ Q-i — Q-i •• 0.
110 F. HERZOG AND B. H. BISSINGER [February

(cn + 1 + 5„)P B _! + 2 V 2
# 0 -s , •

,0. («. + 1 + *»)e-l + Cn-2

(8) r l
, N [Cn + <t>(sn) \p»-l + pn-2
<t>(Xo) = -r
[cn + <t>(sn)\qn-i + qn-2
from (6) and (7). If now x is any number such t h a t # £ E , x?*0, xj^x^
and (at least) the first n — 1 denominators of the simple continued
fraction of x agree with those of Xo we may write in analogy to (6)
(9) x = {a + 1, c2 + 1, • • • , cw-i + 1, c + 1 + s}9
where, as in (6), c is a positive integer and s £ E , so t h a t 0Ss^l/2.
Formulas analogous to (7) and (8) hold also for x. By the use of the
formulas (8) for x0 and x and the relation P„__i<2n-2 —Pw_2(?n-i
=£ w _ig n _ 2 —p n -.2q n -i = ( — l ) w we then obtain the following formula:
A(ff, xo) ss [<£(*) — <K*o)]/(# — #o)
_ C — Cn + <l>(s) — <£(sn) (Cn+ I + S n )Qn-l + Ôn-2
(10) C — Cn + S — Sn [cn + 0($n)kn-l + ^ - 2

(g + 1 + * ) & - ! + Qn-2
[C + 0(s) ]gW-l + Çn-2
We shall also need the following inequalities. If #£2?, # ^ 0 , we can
write x in the form (9) with n = 1, t h a t is, x = { c + 1 + s } with positive
integral c, and s(EE; by (7), <£(#) = {c+$(s)} a n d hence
(11) * ( » ) / * = (c + 1 + s)/[c + <Ks)l
sothat<l>(x)/x>l.Thereîore,4(x)/x^(c+l+s)/(c+s)^(c+l)/cè2.
In view of 0 Ss g 1/2 we conclude now that <£(#)/* ^ ( c + 1 + s ) / ( c + 2 s )
à ( c + 3 / 2 ) / ( c + l ) è l + l/4c. We thus obtained for x £ E , x^0t
(12) 1 < 1 + l/4c g <f>{x)/x ^ 1 + 1 / ^ 2 ,

where c + 1 is the first denominator of the simple continued fraction

for x. We now write the second fraction on the right of (10) in the form

f N Qn~l Cn + 1 + Sn + tn-l
(13) }
qn-\ Cn + <j>(sn) + <t>(tn-l)
where
tn l S
(1A) ~ {Cn~1 + 1» * ' * » ^2 + 1, Cl + 1} = Ön-2/Ön-l,
m
0(*n-l) = \Cn-h * ' , C2$ Ci\ = ^n-Vtfn-l»
I
947l A CANTOR FUNCTION 111

when n > 1, and to =<j>(to) = 0. Obviously, / w _i£ E for n à 1. Applying the

"outer" estimates of (12) to <t>(sn) and 0(/ n -i) and using 0 ^$ n +*«-i ^1»
we see t h a t the second fraction in (13) is always bounded between 1
and 2. Treating the third fraction on the right of (10) similarly, we
obtain from (10)
(15) Gn{Qn-i/qn-iY S A(x, xo) ^ 4Gn(<2n_1/<Z«-i)2,
where
(16) Gn s (c - cn + 4(s) - <t>(sn))/(c - cn + s - sn).
We shall now state our results, concerning the derivatives of <£(x).
We shall use the notations Z}_, Z>~, D+, D+ for the lower and upper
left-hand, lower and upper right-hand derivatives, respectively. If all
four are equal we shall use the notation #'(x). We remark that, in
investigating the behavior of the difference quotient A (x, x0) as x—»x0,
where x 0 G2£, we may restrict also x to values of JE, since for x £ I ) ,
A(x, XQ) lies between A (J, x0) and A(£', xo), where £ and £' are the end
points of the X>M-interval to which x belongs.
(4.1) When xGD then <t>'(x)=Q.
This follows a t once from (3.1).
(4.2) 0'(O) = 1.
By (2a) we have to show that A(x, 0) = $ ( x ) / x - * l , as x—»0+. We
write x in the form x = { c + 1 + s} (see above). When x—»0 + , c—»+<*>,
while 0 ^ 5 ^ 1 / 2 ; thus our statement follows from (11).
(4.3) When x 0 G ER and x 0 is the right-hand end point of a Dn-interval
(n ^ 2) we have D-<p(x0) =Z>-<£(x0) = 0, whileD+<t>(x0) = D+<t>{xo) = (Q/q)2>
where Q and q are the denominators of the reduced fractions for x 0 and
0(x o ), respectively. When x 0 w /Ae left-hand end point of a Dn-interval
(n^l) the above relations hold with the indices — and + interchanged.
(For examples, see the table at the end of the paper.)
L e t x 0 = {^i+l, £2 + 1, • • • ,£« + l } and assume that Xo is the right-
hand end point of an interval of constancy of #(x), t h a t is, that n is
even. The first statement of (4.3) follows a t once from (3.1). To in-
vestigate lim A(x, Xo) as x—»x0+, we may restrict x to numbers of E
in the interval x 0 < x < { c i + 1 , c2 + l, • • • » Cn-i+1, cn+2}. Such an
x can be written in the form (9) with c = cn and sDZE, s5*0. By (6)
we have in this case sw = 0. Substituting these values of c and sn in
(10) and using the reduction formulas for the Qm and qm (see [8, p. 28,
(6) and (7)]), we obtain
N 4>(S) Qn (Cn + 1 + *)Ön-X + Qn-2
A(X, Xo) = z r
s qn [cn + <t>(s) \qn~i + qn-2
112 F. HERZOG AND B. H. BISSINGER [February

As x-*Xo + , we have 5—»0+ and hence, by (4.2), A(x, x0)—>(Qn/qn)2

= (Q/<Z)2, since x0 = Pn/Qn and 4>(xo)=pn/qn. The proof for the left-
hand end point of an interval of constancy of <£(#) requires only the
changing of a few symbols in the above proof.
The result of (4.3) represents an interesting difference between our
function cj>(x) and the functions ù)ap(x) of [3]. For the latter the deriv-
ative a t the points of the corresponding sets P~~ and P+ is zero on one
side, but +<*> on the other side. See [3, p. 438, Theorem 3].
Before considering the case x0ÇzEi, we shall prove the following
lemma, concerning the fractions Qn/qn, which in this case are defined
for all positive integral n.
LEMMA. When Xo£zEi the fractions Qn /qn form an increasing sequence
which approaches a finite limit L>1 or +<*>, according to whether the
infinite series
(17) E l/c,
converges or diverges.
We have Ql/q1 = 1 + 1/ci. Let j> 1. Then by (14), (Qj/qù/iQj-i/ui-ù
= (qj-i/qj)/(Qj-i/Qj) =<t>(ti:)/tj, and hence, by the "inner" estimates
in (12),
1 + 1/4*, ^ (Qi/qù/(Q*-i/qi-Ù â 1 + IA/.
Multiplying over j from j = 2 to j = nf we obtain for w = l, 2, • • •
I C - i ( 1 + V 4 ^ ) < Q n / 2 n ^ r C f - i ( l + l A y ) . From these inequalities the
lemma follows easily.
(4.4) Letx0*={ci + l,c2+l, • • • } £ # / .
(i) If the Cj are bounded then 4>'(x0) = +oo.
(ii) If the Cj are unbounded but the series (17) diverges, then1 D~0(# o )
=#+0(xo)==+°o.
(iii) If the series (17) converges then D~4>(x0) and D+4>(x0) are finite
numbers greater than unity, while Z}_0(xo) =^+<K#o) = 0 .
Let x £ E , xT^O, XT^X^ and let the simple continued fractions for x
and Xo agree in the first n — 1 denominators but no further. We can
then 8 write x in the form (9) with CT^CU and make use of (IS). We ob-
7
No statement is being made about P_0(#o) and D+(f>(xo) in this case. In fact, this
is the only problem, concerning the derivatives of <t>(x), not answered in this section.
8
There is one exceptional case : the convergents of x0t that is, the numbers
x= { c i + 1 , C2+I, • • • t c „ _ i + l } , w = 2, 3, • • • , cannot be written in the form (9)
with cj^Cn and s£E. A moment's consideration shows that their omission (as well as
that of # = 0) from the values of x in A(x, x0) has no influence on the behavior of
A(x, XQ) as #-*#o— or #-»#o+.
Ï9471 A CANTOR FUNCTION 113

serve that X-*XQ means t h a t #—><*>, while c in (9) may assume any
positive integral value, other than s n , and 5 in (9) any value of E.
Finally, if c>cn we have x>x0 (x<x0) when n is even (odd), while if
c<cn the situation is reversed.
In the case (i), the series (17) diverges, hence, by the lemma,
Qn/qn—>+°°. Therefore, by (15) it will suffice to show that Gn is
greater than a positive constant for all values of n, c and 5 (see above).
Since the Cj are bounded and in view of (5), there exists a 5, 0 < 8
< l / 4 , such t h a t ô < ^ n < 1/2 — 5 for all w, and hence by (3.2) an e,
0 < € < l / 2 , such that €<<£(sn) < 1 —e for all n. Thus the denominator
of Gn (see (16)) is numerically less than \c — cn\ + 1 / 2 , while its nu-
merator is numerically greater than \c — cn\ — (1 — e). In view of
\c — cn\ è l , this shows that Gn>2e/3.
In the case (ii) we still have Qn/qn~>+°°. Hence in view of (15),
we have only to show the existence of two sequences of numbers in E,
approaching x0l one from the left and one from the right, such that Gn
for all members of these sequences is greater than a positive constant.
These sequences are obtained by putting c==cn + l and s = sn for each
tty in which case (16) gives Gn = 1. And for even (odd) n we have ap-
proach from the right (left).
In the case (iii), Qn/qn approaches a finite limit L>\ (see the
lemma). In order to show t h a t D~(j)(x0) and D+4>(xo) are finite, it
suffices by (15) to show t h a t Gn is bounded above by a constant for
all fi, cand 5. But it is easily seen that the denominator of Gn in (16) is
numerically greater than or equal to \c — cn\ —1/2, its numerator is
numerically less than or equal to \c — cn\ + 1 and hence G n ^ 4 . In
order to show t h a t Z>~^(x0) and D + 0(^ o ) are greater than unity, we
use the method used above in case (ii). In fact, the same sequences,
defined by c = cn + l, s = s n , hence Gw = l, will suffice here. 9 Finally,
in order to show t h a t D^<j>(x0) =D+(j)(xo) = 0 , we remark that by (3.2)
these derivatives could not be less than zero and hence the method of
case (ii) can again be used, only t h a t here (see (15)) we have to choose
c and s in such a way t h a t Gn—»0. Since, in case (iii), £w—> °° and hence,
by (5), s*-->0 + , we may put (for sufficiently large n) c = cn — l and
5 = 1/2, so that, by (16), G n = # ( S n ) / ( l / 2 + s n ) - > 0 , as n-± <x>. This com-
pletes the proof of (4.4).
We remark t h a t the cases (i), (ii) and (iii) of (4.4) divide the set Er
into three subsets, each of which has the power of the continuum.

9
By the use of more careful estimates it is not difficult to show that D~<f>(x0)
—D+<f>(xo) —2L2> where £ = lim„ (ön/ffn). The details of the proof of this fact are not
given here.
114 F. HERZOG AND B. H. BISSINGER [February

We have thus exhibited a continuum of points at which 10 <j>f(x) = + 0 0

(case (i)) and also one at which neither the left-hand nor the right-
hand derivative exists (case (iii)). These two subsets correspond to
the types I and IV, respectively, of [3, p. 442].
The methods used in this section can also be applied to the problem
of whether the function <f>{x) satisfies a Lipschitz condition: <$>{xf)
-<j>{x)SM(x'-x)v for 0 ^ x < x ' g l . (Compare [3, p. 437, Theorem
2].) We mention without proof that $(x) satisfies a Lipschitz condi-
tion of order p = log b/log a = .546, where <z = {2, 2, • • • } =2 1 / 2 — 1
and b=<t>(a)=: {l, 1, • • • } = ( 5 1 / 2 - l ) / 2 , but that 4>(x) does not sat-
isfy a Lipschitz condition of any higher order. The "best" value for
the constant M, if ? = log 6/log a, is 10"/2 = 1.758.
5. Table. The following table, which we give for the convenience
of the reader, contains a few examples of P n -intervals, which we de-
note here by (x%9 x 2 ); it gives the value of <j>(x) for each of these in-
tervals and the left-hand (right-hand) derivative at its left (right)
end point. The values of x and $(x) are written as simple continued
as well as ordinary fractions.
TABLE

Z> «-interval = (xi, #2)

<}>{x) for D+<f>(x2)
n
Xi x% Xl^X^Xi D~<t>(xi)D+<f>(x2)

1 {2} - 1 / 2 {1}-1 {1}=1 4 —

2 {3} =1/3 {2, 2} =2/5 {2} =1/2 9/4 25/4

2 {4} « 1 / 4 {3, 2} =2/7 {3} =1/3 16/9 49/9

2 {5} =1/5 {4, 2} =2/9 {4} =1/4 25/16 81/16

3 {2, 2, 2} =5/12 {2,3} =3/7 {1.2}-2/3 16 49/9

3 {2,3,2} =7/16 {2,4} =4/9 {1,3} =3/4 16 81/16 1

3 {3,2,2} =5/17 {3,3} =3/10 {2, 2} =2/5 289/25 4

4 ! {2,2,3} =7/17 {2,2,2,2} =12/29 {1,1,2} =3/5 289/25 841/25

10
In this connection, see [5, p. 258, footnote 2 ] . We also note that all quadratic
irrationalities between 0 and 1 fall under case (i).
19471 A CANTOR FUNCTION 115

BIBLIOGRAPHY
1. F . Bernstein, Über eine Anwendung der Mengenlehre aufein aus der Theorie der
sàkularen Störungen herrührendes Problem, M a t h . Ann. vol. 71 (1912) pp. 417-439.
2. C. Carathéodory, Vorlesungen über réelle Funktionen, Leipzig and Berlin, 1927.
3. R. E. Gilman, A class of f unctions continuous but not absolutely continuous,
Ann. of M a t h . (2) vol. 33 (1932) pp. 433-442.
4. F . Herzog and B. H . Bissinger, A generalization of BoreVs and F. Bernstein1 s
theorems on continued fractions, Duke M a t h . J. vol. 12 (1945) pp. 325-334.
5. E. Hille and J. D. Tamarkin, Remarks on a known example of a monotone con-
tinuous function, Amer. M a t h . Monthly vol. 36 (1929) pp. 255-264.
6. E. W. Hobson, The theory of functions of a real variable, vol. 1, Cambridge, 1927.
7. G. Kowalewski, Grundzüge der Differential» und Integralrechnung, Leipzig and
Berlin, 1928.
8. O. Perron, Die Lehre von den Kettenbrüchen, Leipzig and Berlin, 1929.
M I C H I G A N S T A T E COLLEGE AND
LEOMINSTER, M A S S .