2

           

           3

             7

               10

! "                 16

#            19

$      26

% &  '    32

( ) '    35

* +,        + -    sc .   35


     

   
"      G / 0  '  12  3 ((
 0 4/           12 5  / /4  
    6  "              
     G        + 4' 7   
    1+/ $
2 "           
             6'  162 "      /
   + )       8)9   ' :       
      8 ;<  5  .=  *
!> 6    G   
    )                
                 
1?)2  1" &0   % 2 "          G / 
        '      /      
  /4   G                   .  
      / /        
@    0           G    C  )
x ∈ [0, 1]   x 


∞
anx
(0.1) x= , anx ∈ {0, 1, 2}.
n=1
3n

&  Nx   n /  anx = 1        Nx = ∞    

  anx  "       G : [0, 1] → R   0 

1

Nx −1
1 anx
(0.2) G(x) := + .
2 Nx 2 n=1 2n

A            ;
 >  x  / 
 
"   C         [0, 1] /     ;
 >
      
     x ∈ C ;anx ∈ {0, 2}>     ;
>
4  

1
anx
∞
(0.3) G(x) = .
2 n=1 2n

" /               C 

  
  /    C0 := [0, 1] 0    C1 ⊇ C2 ⊇ ... ⊇ Ck ⊇
...  C0  / @     C1      8   8 
 ( 12 , 23 )  C0  "   C2        C1  

     3

 ( 19 , 29 )  ( 79 , 89 )   Ck    2k  B    

  Ck  Ck+1                
 4  Ck  "      

∞
(0.4) C= Ck .
k=0

&  C 1           C  

(0.5) C ◦ := C\C 1 , I ◦ := [0, 1]\C.

)  I            I ◦ 

Figure 1: The graph of the Cantor function G. This graph is sometimes called
”Devil’s staircase”.

1 Singularity, measurability and representability

by absolutely continuous functions
The well-known properties of the Cantor function are collected in the following.

Proposition 1.1.
1.1.1. G is continuous and increasing but not absolutely continuous.
1.1.2. G is constant on each interval from I ◦ .

     4

1.1.3. G is a singular function.

1.1.4. G maps the Cantor set C onto [0, 1].

Proof. It follows directly from (0.2) that G is an increasing function, and moreover
(0.2) implies that G is constant on every interval J ⊆ I ◦ . Observe also that if
x, y ∈ [0, 1], x = y, and x tends to y, then we can take ternary representations (0.1)
so that
min{n : |anx − any | = 0} → ∞.
Thus the continuity of G follows from (0.2) as well. Since the one dimensional
Lebesgue measure of C is zero,
m1 (C) = 0,
the monotonicity of G and constancy of G on each interval J ⊆ I ◦ imply Statement
1.1.3. One easy way to check the last equality is to use the obvious recurrence
relation
2
m1 (Ck ) = m1 (Ck−1)
3
for closed subset Ck in (0.4). Really, by (0.4)
 2 k
m1 (C) = lim m1 (Ck ) = lim = 0.
k→∞ k→∞ 3
It still remains to note that by (0.3) we have 1.1.4 and that G is not absolutely
continuous, since G is singular and nonconstant.

Remark 1.2. Recall that a monotone or bounded variation function f is called

singular if f  = 0 a.e.

By the Radon-Nikodym theorem we obtain from 1.1.1 the following.

Proposition 1.3. The function G cannot be represented as

 x
G(x) = ϕ(t)dt
0

where ϕ is a Lebesgue integrable function.

In general a continuous function need not map a measurable set onto a measur-
able set. It is a consequence of 1.1.4 that the Cantor function is such a function.

Proposition 1.4. There is a Lebesgue measurable set A ⊆ [0, 1] such that G(A) is
not Lebesgue measurable.

     5

In fact, a continuous function g : [a, b] → R transforms every measurable set

onto a measurable set if and only if g satisfies Lusin’s condition (N):

(m1 (E) = 0) ⇒ (m1 (g(E)) = 0)

for every E ⊆ [a, b] [Sa, p.224].

Let Lf denote the set of points of constancy of a function f , i.e., x ∈ Lf if
f is constant in a neighbourhood of x. In the case f = G it is easy to see that
LG = I ◦ = [0, 1]\C.

Proposition 1.5. Let f : [a, b] → R be a monotone continuous function. Then

the following statements are equivalent
1.5.1. The inverse image f −1 (A) is a Lebesgue measurable subset of [a, b] for
every A ⊆ R.
1.5.2. The Lebesgue measure of the set [a, b]\Lf is zero,

(1.6) m1 ([a, b]\Lf ) = 0.

Proof. 1.5.2 ⇒ 1.5.1. Evidently, Lf is open (in the relative topology of [a, b]) and f
is constant on each component of Lf . Let E be a set of endpoints of components of
Lf . Let us denote by f0 and f1 the restrictions of f to Lf ∪ E and to [a, b]\(Lf ∪ E),
respectively.
f0 := f |Lf ∪E , f1 := f |[a,b]\(Lf ∪E) .
If A is an arbitrary subset of R, then it is easy to see that f0−1 (A) is a Fσ subset of
[a, b] and
f1−1 (A) ⊆ [a, b]\Lf .
Since f −1 (A) = f0−1 (A)∪f1−1 (A), the equality m1 ([a, b]\Lf ) = 0 implies that f −1 (A)
is Lebesgue measurable as the union of two measurable sets.
1.5.1 ⇒ 1.5.2. The monotonicity of f implies that f1 is one-to-one and

(1.7) (f0 (Lf ∪ E)) ∩ (f1 ([a, b]\(Lf ∪ E))) = ∅.

Suppose that (1.6) does not hold. For every B ⊆ R with an outer measure m∗1 (B) > 0
there exists a nonmeasurable set A ⊆ B. See, for instance, [Oxt, Chapter 5, Theorem
5.5]. Thus, there is a nonmeasurable set A ⊆ [a, b]\(Lf ∪ E). Since f1 is one-to-one,
equality (1.7) implies that f −1 (f (A)) = A, contrary to 1.5.1.

Corollary 1.8. The inverse image G−1 (A) is a Lebesgue measurable subset of [0, 1]
for every A ⊆ R.

     6

Remark 1.9. It is interesting to observe that G(A) is a Borel set for each Borel
set A ⊆ [0, 1]. Indeed, if f : [a, b] → R is a monotone function with the set of
discontinuity D, then

f (A) = f (A ∩ D) ∪ f (A ∩ Lf ) ∪ f (A ∩ E) ∪ f (A\(D ∪ Lf ∪ E)), A ⊆ [a, b],

where E is the set of endpoints of components of Lf . The sets f (A ∩ D), f (A ∩ Lf )

and f (A ∩ E) are at most countable for all A ⊆ [0, 1]. If A is a Borel subset of [a, b],
then f (A\(D ∪ Lf ∪ E)) is Borel, because it is the image of the Borel set under the
homeomorphism f [a,b]\(D∪L ∪E) .
f

A function f : [a, b] → R is said to satisfy the Banach condition (T1 ) if

m1 ({y ∈ R : card(f −1(y)) = ∞}) = 0.

Since a restriction G|C ◦ is one-to-one and G(I ◦ ) is a countable set, 1.1.2 implies

Proposition 1.10. G satisfies the condition (T1 ).

N.Bary and D.Menchoff [BaMe] showed that a continuous function f : [a, b] → R

is a superposition of two absolutely continuous functions if and only if f satisfies
both the conditions (T1 ) and (N). Moreover, if f is differentiable at every point of
a set which has positive measure in each interval from [a, b], then f is a sum of two
superpositions,
f = f1 ◦ f2 + f3 ◦ f4
where fi , i = 1, ..., 4, are absolutely continuous [Bar]. Thus, we have

Proposition 1.11. There are absolutely continuous functions f1 , ..., f4 such that

G = f1 ◦ f2 + f3 ◦ f4 ,

but G is not a superposition of any two absolutely continuous functions.

Remark 1.12. A superposition of any finite number of absolutely continuous func-

tions f1 ◦ f2 ◦ ... ◦ fn is always representable as q1 ◦ q2 with two absolutely continuous
q1 , q2 . Every continuous function is the sum of three superpositions of absolutely
continuous functions [Bar]. An application of Proposition 1.10 to the nondifferen-
tiability set of the Cantor function will be formulated in Proposition 7.1.

     7

2 Subadditivity, the points of local convexity

An extended Cantor’s ternary function Ĝ is defined as follows


 0 if x < 0
Ĝ(x) = G(x) if 0 ≤ x ≤ 1


1 if x > 1.

Proposition 2.1. The extended Cantor function Ĝ is subadditive, that is

Ĝ(x + y) ≤ Ĝ(x) + Ĝ(y)

for all x, y ∈ R.

This proposition implies the following corollary.

Proposition 2.2. The Cantor function G is a first modulus of continuity of itself,

i.e.,
sup |G(x) − G(y)| = G(δ)
|x−y|≤δ
x,y∈[0,1]

for every δ ∈ [0, 1].

The proof of Propositions 2.1 and 2.2 can be found in Timan’s book [Tim, Section
3.2.4] or in the paper of Josef Dobos̆ [Dob].
It is well-known that a function ω : [0, ∞) → [0, ∞) is the first modulus of
continuity for a continuous function f : [a, b] → R if and only if ω is increasing,
continuous, subadditive and ω(0) = 0 holds.
A particular way to prove the subadditivity for an increasing function
ω : [0, ∞) → [0, ∞) with ω(0) = 0 is to show that the function δ −1 ω(δ) is de-
creasing [Tim, Section 3.2.3]. The last condition holds true in the case of concave
functions. The following propositions show that this approach is not applicable for
the Cantor function.
Let f : [a, b] → R be a continuous function. Let us say that f is locally concave-
convex at point x ∈ [a, b] if there is a neighbourhood U of the point x for which f |U
is either convex or concave. Similarly, a continuous function f : (a, b) → R is said
to be locally monotone at x ∈ (a, b) if there is an open neighbourhood U of x such
that f |U is monotone.

Proposition 2.3. The function x−1 G(x) is locally monotone at point x0 if and only
if x0 ∈ I ◦ .

     8

Proposition 2.4. G(x) is locally concave-convex at x0 if and only if x0 ∈ I ◦ .

Propositions 2.3 and 2.4 are particular cases of some general statements.

Theorem 2.5. Let f : [a, b] → R be a monotone continuous function. Suppose

that Lf is an everywhere dense subset of [a, b]. Then f is locally concave-convex at
x ∈ [a, b] if and only if x ∈ Lf .

For the proof we use of the following.

Lemma 2.6. Let f : [a, b] → R be a continuous function. Then [a, b]\Lf is a

compact perfect set.

Proof. By definition Lf is relatively open in [a, b]. Hence [a, b]\Lf is a compact subset
of R. If p is an isolated point of [a, b]\Lf , then either it is a common endpoint of two
intervals I, J which are components of Lf or p ∈ {a, b}. In the first case it follows
by continuity of f that I ∪ {p} ∪ J ⊆ Lf . That contradicts to the maximality of the
connected components I, J. The second case is similar.

Proof of Theorem 2.5. It is obvious that f is locally concave-convex at x for

x ∈ Lf . Suppose now that x ∈ [a, b]\Lf and f is concave-convex in a neighbourhood
U0 of the point x. By Lemma 2.6 x is not an isolated point of [a, b]\Lf . Hence,
there exist y, z in [a, b]\Lf and δ > 0 such that

z < y, (z − δ, y + δ) ⊂ U0 ∩ [a, b], (z, y) ⊂ Lf .

Since [a, b]\Lf is perfect, we can find z0 , y0 for which

z0 ∈ (z − δ, z) ∩ ([a, b]\Lf ), y0 ∈ (y, y + δ) ∩ ([a, b]\Lf ).

Denote by lz and ly the straight lines which pass through the points (z0 , f (z0 )),
z+y
2
, f z+y
2
and (y0 , f (y0 )), z+y 2
, f z+y
2
, respectively. Suppose that f is in-
creasing. Then the point (z, f (z)) lies over lz but (y, f (y)) lies under ly (See Figure
2). Hence, the restriction f |(z0 ,y0 ) is not concave-convex. This contradiction proves
the theorem since the case of a decreasing function f is similar.

Theorem 2.7. Let f : (a, b) → [0, ∞) be an increasing continuous function and let
ϕ : (a, b) → [0, ∞) be a strictly decreasing function with finite derivative ϕ (x) at
every x ∈ (a, b). If

(2.8) m1 (Lf ) = |b − a|,

then the product f · ϕ is locally monotone at a point x ∈ (a, b) if and only if x ∈ Lf .


     9

Figure 2: The set of points of constancy is the same as the set of points of convexity.

For a function f : (a, b) → R and x ∈ (a, b) set

f (x + ∆x) − f (x − ∆x)
SDf (x) = lim
∆x→0 2∆x
provided that the limit exists, and write

Vf := {x ∈ (a, b) : SDf (x) = +∞}.

Using the techniques of differentation of Radon measures (See [EvGa, in particular

Section 1.6, Lemma 1]), we can prove the following.

Lemma 2.9. Let f : (a, b) → R be continuous increasing function. If equality (2.8)

holds, then Vf is a dense subset of (a, b)\Lf .

Proof of Theorem 2.7. It is obvious that ϕ · f is locally monotone at x0 for

every x0 ∈ Lf . Suppose that ϕ · f is monotone on an open interval J ⊂ (a, b) and
x0 ∈ ((a, b)\Lf ) ∩ J. Since Lf is an everywhere dense subset of (a, b), there exists
an interval J1 ⊆ J ∩ Lf . Then ϕ · f |J1 is a decreasing function. Consequently, ϕ · f |J
is decreasing too. By Lemma 2.9 we can select a point t0 ∈ Vf ∩ J. Hence, by the
definition of Vf

SDϕ(x)f (x)|x=t0 = ϕ (t0 )f (t0 ) + ϕ(t0 )SDf (t0) = +∞.

This is a contradiction, because ϕ · f |J is decreasing.


     10

Remark 2.10. Density of Lf is essential in Theorem 2.5. Indeed, if A is a closed

subset of [a, b] with nonempty interior IntA, then it is easy to construct a continuous
increasing function f such that Lf = [a, b]\A and f is linear on each interval which
belongs to IntA. Theorem 2.7 remains valid if the both functions f and ϕ are
negative, but if f and ϕ have different signs, then the product f · ϕ is monotone
on (a, b). Functions having a dense set of constancy have been investigated by
A.M.Bruckner and J.L.Leonard in [BrLe]. See also Section 6 in the present work.

3 Characterizations by means of functional equa-

tions
There are several characterizations of the Cantor function G based on the self-
similarity of the Cantor ternary set. We start with an iterative definition for G.
Define a sequence of functions ψn : [0, 1] → R by the rule

 1 1
 ψn (3x)
 if 0 ≤ x ≤

 2 3




1 1 2
(3.1) ψn+1 (x) = if <x<

 2 3 3




 1 1

 + ψn (3x − 2) if 2 ≤ x ≤ 1.
2 2 3
where ψ0 : [0, 1] → R is an arbitrary function. Let M[0, 1] be the Banach space of
all uniformly bounded real–valued functions on [0, 1] with the supremum norm.

Proposition 3.2. The Cantor function G is the unique element of M[0, 1] for
which

 1 1

 G(3x) if 0≤x≤

 2 3




1 1 2
(3.3) G(x) = if <x<

 2 3 3






 1 + 1 G(3x − 2) if
2
≤ x ≤ 1.
2 2 3
If ψ0 ∈ M[0, 1], then the sequence {ψn }∞
n=0 converges uniformly to G.

     11

Proof. Define a map F : M[0, 1] → M[0, 1] as


 1 1

 f (3x) , 0≤x≤ ,

 2 3




1 1 2
F (f )(x) = , <x< ,

 2 3 3






 1 + 1 f (3x − 2) , 2 ≤ x ≤ 1.
2 2 3
Since
1
F (f1 ) − F (f2 ) ≤ f1 − f2 
2
where f1 , f2 ∈ M[0, 1] and  ·  denotes the norm in M[0, 1], F is a contraction
map on complete space M[0, 1] and, consequently, by the Banach theorem F has an
unique fixed point f0 , f0 = F (f0 ), and that ψn → f0 uniformly on [0, 1]. It follows
from the definition of G, that (3.3) holds. Hence, F (G) = G and by uniqueness
f0 = G.

It should be observed here that there exist several iterative definitions for the
Cantor ternary function G. The above method is a simple modification of the cor-
responding one from Dobos̆’s article [Dob]. It is interesting to compare Proposition
3.2 with the self-similarity property of the Cantor set C.
Let for x ∈ R
1 1 2
(3.4) ϕ0 (x) := x, ϕ1 (x) := x + .
3 3 3

Proposition 3.5. The Cantor set C is the unique nonempty compact subset of R
for which

(3.6) C = ϕ0 (C) ∪ ϕ1 (C)

holds. Further if F is an arbitrary nonempty compact subset of R, then the iterates

Φk+1 (F ) := ϕ0 (Φk (F )) ∪ ϕ1 (Φk (F )), Φ0 (F ) := F,

convergence to the Cantor set C in the Hausdorff metric as k → ∞.

Remark 3.7. This theorem is a particular case of a general result by Hutchinson

about a compact set which is invariant with respect to some finite family contraction
maps on Rn [Hut].

     12

In the case of a two-ratio Cantor set a system which is similar to (3.3) was found
by W.A.Coppel in [Cop]. The next theorem follows from Coppel’s results.

Proposition 3.8. Every real–valued F ∈ M[0, 1] that satisfies

 x  F (x)
(3.9) F = ,
3 2

 x F (x)
(3.10) F 1− =1− ,
3 2

1 x 1
(3.11) F + =
3 3 2
is the Cantor ternary function.

The following simple characterization of the Cantor function G has been sug-
gested by D.R.Chalice in [Ch].

Proposition 3.12. Every real–valued increasing function F : [0, 1] → R satisfying

(3.9) and

(3.13) F (1 − x) = 1 − F (x)

is the Cantor ternary function.

Remark 3.14. Chalice used the additional condition F (0) = 0, but it follows from
(3.9).

The functional equations, given above, together with Proposition 3.5 are some-
times useful in applictions. As examples, see Proposition 4.5 in Section 4 and Propo-
sition 5.1 in Section 5.

The system (3.9)+(3.13) is a particular case of the system that was applied in the
earlier paper of G. C. Evans [Ev] to the calculation of the moments of some Cantor
functions. In this interesting paper Evans noted that (3.9) and (3.13) together with
continuity do not determine G uniquely. However, they imply that

1 2 2
(3.15) F (x) = + F x − , ≤ x ≤ 1.
2 3 3
Next we show that a variational condition, together with (3.9) and (3.13), de-
termines the Cantor function G.

     13

Proposition 3.16. Let F : [0, 1] → R be a continuous function satisfying (3.9) and

(3.13). Then for every p ∈ (1, +∞) it holds
 1  1
p
(3.17) |F (x)| dx ≥ |G(x)|p dx,
0 0

and if for some p ∈ (1, ∞) an equality

 1  1
p
(3.18) |F (x)| dx = |G(x)|p dx
0 0

holds, then G = F .

Figure 3: The continuous function F which satisfies ”the two Cantor function equa-
tions” (3.9) and (3.13) but F (x) = 12 (12x − 5)2 on 13 , 12 .

Lemma 3.19. Let f : [a, b] → R be a continuous function and let p ∈ (1, ∞).
Suppose that the graph of f is symmetric with respect to the point a+b
2
; f a+b
2
.
Then the inequality
 b  p
1  a + b 
(3.20) |f (x)| dx ≥ f
p 

|b − a| a 2

holds with equality only for 

a+b
f (x) ≡ f .
2

     14

Proof. We may assume without loss of generality that a = −1, b = 1. Now for
f (0) = 0 inequality (3.20) is trivial. Hence replacing f with −f , if necessary, we
may assume that f (0) < 0. Write

Ψ(x) = f (x) − f (0).

Given m > 0, we decompose [−1, 1] as

A+ := {x ∈ [−1, 1] : |Ψ(x)| > m}, A0 := {x ∈ [−1, 1] : |Ψ(x)| = m},

A− := {x ∈ [−1, 1] : |Ψ(x)| < m}.

b
Set Jm := |Ψ(x) − m|p dx. Since Ψ is an odd function, we obtain
a
 
 p p

Jm = (|Ψ(x)| − m) + (|Ψ(x)| + m) dx + (2m)p dx
A+ ∩[0,1] A0 ∩[0,1]

 
+ (m − |Ψ(x)|)p + (m + |Ψ(x)|)p dx := Jm
+ 0
+ Jm −
+ Jm .
A− ∩[0,1]

Elementary calculation shows that the function

g(y) := (B − y)p + (B + y)p

is strictly increasing on (0, B) for p > 1. Hence,

  
+ p p
(3.21) Jm ≥ 2 |Ψ(x)| dx = |Ψ(x)| dx ≥ mp dx.
A+ ∩[0,1] A+ A+

If Ψ(x) ≡ 0, then A− is nonempty and open, as Ψ(0) = 0. Thus,

 
− p
(3.22) Jm > 2 m dx = mp dx
A− ∩[0,1] A−

for Ψ(x) ≡ 0. Moreover, it is obvious that

  
0 p−1 p
(3.23) Jm = 2 2 m dx ≥ mp dx.
A0 ∩[0,1] A0

Choose m = −f (0). Then, in the case where f (x) ≡ f (0), from (3.21)–(3.23) we
obtain the required inequality

1 1
|f (x)|p > |f (0)|p.
2 −1



     15

Remark 3.24. Inequality (3.20) holds also for p = 1, but, as simple examples show,
in this case the equality in (3.20) is also attained by functions different from the
constant function.

For every interval J ∈ I (where I is a family of components of the open set

◦
I = [0, 1]\C) let us denote by xJ the center of J and by GJ the value of the Cantor
function G at xJ .

Lemma 3.25. If F : [0, 1] → R is a continuous function satisfying (3.9) and (3.13),

then the graph of the restriction F |J is symmetric with respect to the point (xJ , GJ )
for each J ∈ I.

Proof. It follows from (3.13) that

 
1 1 1
F =G = ,
2 2 2

thus we can rewrite equation (3.13) as

  
1 1 1 1
F = F +x +F −x .
2 2 2 2

Consequently a graph of F |( 1 , 2 ) is symmetric with respect to the point 12 , 12 . If J is

3 3
an arbitrary element of I, then there is a finite sequence of contractions ϕi1 , ..., ϕin
such that J is an image of the interval 13 , 23 under superposition ϕi1 ◦ ... ◦ ϕin
and each ϕik , k = 1, ..., n belongs to {ϕ0 , ϕ1 } (see formula (3.4)). Now the desired
symmetry follows from (3.9) and (3.15) by induction.

Proof of Proposition 3.16. Let J ∈ I. If F satisfies (3.9) and (3.13), then

Lemma 3.25 and Lemma 3.19 imply that
 
(3.26) |F (x)| dx ≥ |G(x)|p dx
p
J J

for every p ∈ (1, ∞). Thus, we obtain (3.17) from the condition m1 (C) = 0. Suppose
now that (3.18) holds. Then we have the equality in (3.26) for every J ∈ I. Thus,
by Lemma 3.19
F |J = G|J , J ∈ I.
Since I 0 = ∪ J is a dense subset of [0, 1] and F = G as required.

J∈I

In the special case p = 2 we can use an orthogonal projection to prove Proposition

3.16.

     16

Define a subspace L2C [0, 1] of the Hilbert space L2 [0, 1] by the rule: f ∈ L2C [0, 1]
if f ∈ L2 [0, 1] and for every J ∈ I there is a constant CJ such that

|f (x) − CJ |2 dx = 0.
J

It is obvious that G ∈ L2C [0, 1].

Let us denote by PC the operator of the orthogonal projection from L2 [0, 1] to
L2C [0, 1].

Proposition 3.27. Let f be an arbitrary function in L2 [0, 1] and let J ∈ I be an

interval with endpoints aJ , bJ .
3.27.1 Suppose that f is continuous. Then the image PC (f ) is continuous if and
only if 
1
f (aJ ) = f (bJ ) = f (x)dx
|bJ − aJ | J
for each J ∈ I.
3.27.2. Function f ∈ L2 [0, 1] is a solution of the equation PC (f ) = G if and only
if  
f (x)dx = G(x)dx
J J
for each J ∈ I.

We leave the verification of this simple proposition to the reader.

Now, the conclusion of Proposition 3.16 follows directly from Lemma 3.25 by
usual properties of orthogonal projections. See, for example, [BSU, Chapter VII, 9].

4 The Cantor function as a distribution function

It is well known that there exists an one-to-one correspondence between the set of
the Radon measures µ in R and the set of a finite valued, increasing, right continuous
functions F on R with lim F (x) = 0. Here we study the corresponding measure
x→−∞
for the Cantor function.
Let ϕ0 , ϕ1 : R → R be similarity contractions defined by formula (3.4). Write

lg 2
(4.1) sc := .
lg 3

We let Hsc denote the sc -dimensional Hausdorff measure in R. See [Fal1] for
properties of Hausdorff measures.

     17

Proposition 4.2. There is the unique Borel regular probability measure µ such that
1 1
(4.3) µ(A) = µ(ϕ−1 −1
0 (A)) + µ(ϕ1 (A))
2 2
for every Borel set A ⊆ R. Furthemore, this measure µ coincides with the restriction
of the Hausdorff measure Hsc to C, i.e.,
(4.4) µ(A) = Hsc (A ∩ C)
for every Borel set A ⊆ R.

For the proof see [Fal1, Theorem 2.8, Lemma 6.4].

Proposition 4.5. Let m1 be the Lebesgue measure on R. Then

(4.6) m1 (G(A)) = Hsc (A)
for every Borel set A ⊆ C.

Proof. Write
µ(A) := m1 (G(A ∩ C))
for every Borel set A ⊆ R. If suffices to show that µ is a Borel regular probability
measure which fulfils (4.3). Since C 1 is countable (see formula (0.5)), m1 (G(C 1 )) =
0. Hence, we have
µ(A) = m1 (G(C ◦ ∩ A)), A ⊆ R.
The restriction G|C ◦ : C ◦ → G(C ◦ ) is a homeomorphism and
µ(R) = m1 (G(C ◦ )) = 1.
Hence, µ is a Borel regular probability measure. (Note that G(A) is a Borel set
for each Borel set A ⊆ C. See Remark 1.9.) To prove (4.3) we will use following
functional equations (see (3.9), (3.15)).

x 1
(4.7) G = G(x), 0 ≤ x ≤ 1,
3 2

1 2 2
(4.8) G(x) = + G x − , ≤ x ≤ 1.
2 3 3
Let A be a Borel subset of R. Put
   
1 2
A0 := A ∩ 0, , A1 := A ∩ , 1 .
3 3

     18

Since ϕ−1 −1
0 (x) = 3x and ϕ1 (x) = 3 x −
2
3
, we have

µ(ϕ−1 −1
i (A)) = µ(ϕi (Ai ))

for i = 1, 2. It follows from (4.7) that

2G(x) = G(ϕ−1
0 (x)),
 
for x ∈ 0, 13 . Hence,
1 1 1
µ(ϕ−1 −1
0 (A0 )) = m1 (G(ϕ0 (A0 ))) = m1 (2G(A0 )) = m1 (G(A0 )).
2 2 2
Similarly, (4.7) implies that

2
2G x − = G(ϕ−1
1 (x))
3
2 
for x ∈ 3
, 1 , and we obtain
1 1
µ(ϕ−1 −1
1 (A1 )) = m1 (G(ϕ1 (A1 ))) =
2 2
 
1 2 2
= m1 2G A1 − = m1 G A1 − .
2 3 3
Now using (4.8) we get
 
2 1
m1 G A1 − = m1 − + G(A1 ) = m(G(A1 )).
3 2

The set G(A1 ) ∩ G(A0 ) is empty or contains only the point 12 . Hence we obtain

m1 (G(A0 )) + m1 (G(A1 )) = m1 (G(A0 ∪ A1 )) = µ(A).

Therefore µ(A) satisfies (4.3).

Corollary 4.9. The extended Cantor function Ĝ is the cumulative distribution

function of the restriction of the Hausdorff measure Hsc to the Cantor set C.

This description of the Cantor function enables us to suggest a method for the
proof of the following proposition. Write

Ĝh (x) := Ĝ(x + h) − Ĝ(x)

for each h ∈ R. The function Ĝh is of bounded variation, because it equals a

difference of two increasing bounded functions.

     19

Proposition 4.10. [HiTa] Let V ar(Ĝh ) be a total variation of Ĝh . Then we have
sup V ar(Ĝh ) = 2
0≤h≤δ

for every δ > 0.

This holds because sets C ∩ (C ± 3−n ) have finite numbers of elements for all
positive integer n.

Remark 4.11. If F is an absolutely continuous function of bounded variation and

Fh (x) := F (x + h) − F (x), then
(4.12) lim V ar(Fh ) = 0.
h→0

Really, if F is absolutely continuous on [a, b], then F  exists a.e. in [a, b], F  ∈ L1 [a, b],
and
b
V ar(F ) = |F (x)| dx.
a

Approximating of F by continuous functions, for which the property is obvious, we
obtain (4.12).

In fact, Proposition 4.10 remains valid for an arbitrary singular function of

bounded variation.

Theorem 4.13 Let F : R → R be a function of bounded variation with singular

part ϕ. Then the limit relation
lim sup(V ar(Fh )) = 2V ar(ϕ)
h→0

holds.

The theorem is an immediate adaptation of the result that was proved by Norbert
Wiener and R. C. Young in [WY].

5 Calculation of moments and the length of the

graph
In the article [Ev], G.C.Evans proved the recurrence relations for the moments
 1
xn Gα (x)dx
0

     20

where Gα is a ”Cantor function” for a α-middle Cantor set. In the classic middle-
third case Evans’s results can be written in the form.

Proposition 5.1. Let n be a natural number and let Mn be a moment of the form
 1
Mn = xn G(x)dx.
0

Then the following relations hold

1 1
n n−1
(5.2) 2Mn = + n+1 2n−k Mk ,
n+1 3 − 1 k=0 k

1
  n−1
n k+1 n
(5.3) (1 + (−1) )Mn = + (−1) Mk
n + 1 k=0 k

n
for all positive integers n where k
are the binomial coefficients.

It follows immediately from G(x) + G(1 − x) = 1 that M0 = 12 . Hence, (5.2)

can be used to compute all moments Mn . Let µC be the restriction of the Hausdorff
measure Hsc to the Cantor set C (see Corollary 4.9). Now set
 1
mn := xn dµC (x).
0

Proposition 5.4. The following equality holds


n
n+1
k
2n−k mk
k=0
(5.5) mn+1 = ,
3n+1 − 1
for every natural n and we have

n−1 n
(5.6) 2mn = (−1)k mk ,
k
k=0

n−1 
k n+1
(5.7) (n + 1)mn = (−1) mk
k=0
k

for every odd n.


     21

Proof. First we prove (5.5).

The Cantor set C can be defined as the intersection ∩∞ k=0 Ck , see (0.4). Each Ck
consists of 2k disjoint closed intervals Cki = [aik , bik ], i = 1, ..., 2k . The length of Cki is
3−k and
µC (Cki ) = G(bik ) − G(aik ) = 2−k .
Let Lk be a set of all left-hand points of intervals Cki ⊆ Ck , i.e.

Lk = {aik : i = 1, ..., 2k }.

It is easy to see that L0 = {0} and

 
1 2 1
(5.8) Lk = Lk−1 ∪ + Lk−1
3 3 3
for k ≥ 1. Observe that
 1  1
n
x dµC (x) = xn dG(x)
0 0

because the integrand is continuous. Hence we have

l
mk = lim (ξi)n [G(xj+1 ) − G(xj )]
l→∞
j=1

where 0 = x1 < ... < xl < xl+1 = 1 is any subdivision of [0, 1] with maxj (xj+1 −xj ) →
0 as l → ∞ and ξj ∈ [xj , xj+1 ]. Subdivide [0, 1] into l = 3k equal intervals [xj , xj+1]
and take ξj = xj . Then
1
n
mn = lim k x .
k→∞ 2
x∈L k

The last relation and (5.8) imply that

 
1


mn = lim k  xk + xn 
k→∞ 2
1
x∈ 3 Lk−1 x∈ 23 + 13 Lk−1

   
1 1 1
n 1 1 1

= · n lim  k−1 x + · n lim  k−1 (x + 2)n 
2 3 k→∞ 2 x∈L
2 3 k→∞ 2 x∈L
k−1 k−1


1 1

n
1 1 n
= · n mn + · n mp 2n−p ,
2 3 2 3 p=0 p

and we obtain (5.5).


     22

In order to prove (5.6) we can use (5.3), because

 1 
1 1
(5.9) mn+1 = x dG(x) = x G(x)0 −(n+1)
n+1 n+1
xn G(x)dx = 1−(n+1)Mn .
0 0

Suppose that n is even, then it follows from (5.3) and (5.9) that

n−1 n
2(1 − mn+1 ) = 2(n + 1)Mn = 1 + (n + 1) (−1)k+1 Mk
k=0
k

n + 1 n
n−1
=1+ (−1)k+1 (1 − mk+1 ).
k=0
k+1 k
n+1 n n+1
Hence from k+1 k
= k+1
and (1 − m0 ) = 0 we get

n−1 
k+1 n+1
2(1 − mn+1 ) = 1 + (−1) (1 − mk+1 )
k=0
k+1

n 

n 
k n+1 k n+1
=1+ (−1) (1 − mk ) = 1 + (−1)
k k
k=1 k=0

n 
n+1
− (−1)k mk .
k=0
k
Observe that


n  
k n+1 n+1 n+1
(−1) = (1 − 1) − (−1)n+1 = 1
k=0
k n+1

because n is even. Consequently,

n
2(1 − mn+1 ) = 2 − (−1)k mk .
k=0

The last formula implies (5.6). The proof of (5.7) is analogous to that of (5.6).

Remark 5.10. The measure µC is frequently referred to as the Cantor measure. In

mn
the proof of (5.5) we used an idea from [HiTa]. By (5.5) with τn = n!2n we obtain


1 τn τn−1 τ0
τn+1 = + + ... + .
2(3n+1 − 1) 1! 2! (n + 1)!

     23

The asymptotics of the moments mn was determinated in [GoWi]. Write

2πim
ν(m) := 1 −
lg 3
for m ∈ Z.

Theorem 5.11.Let mn be the nth moment of the Cantor measure µC , then

(5.12) mn ∼ 21/2−3sc /2 n−sc exp(−2H(n)), n → ∞,

where

1

1 −ν(m)
+∞
H(x) := 2 (1 − 21−ν(m) )ζ(ν(m))Γ((ν(m))x1−ν(m)
2πi m∈Z m
m=0

and
lg 2
sc = .
lg 3

Note that H(x) is a real-valued function for x > 0 and periodic in the vari-
able log x. See also [Fis] for an example of an absolutely continuous measure with
asymptotics of moments containing oscillatory terms.
The behavior of the integrals
 1  1
λ
(5.13) I(λ) := (G(x)) dx, E(λ) = exp(λG(x)) dx
0 0

was described in [GorKu]. It was noted that I extends to a function which is analytic
in the half-plane Re(λ) > −sc and E extends to an entire function. The following
theorem is a particular case of the results from [GorKu].

Theorem 5.14.
5.14.1 For Re(λ) > −sc , the function I obeys the formula
 1
λ
(3 · 2 − 1)I(λ) = 1 + (1 + G(x))λ dx,
0

and for all λ ∈ C the function E obeys the formula

3E(2λ) = eλ + (eλ + 1)E(λ)


     24

5.14.2 For all natural n ≥ 2, we have

1


n 2k−1 − 1 B(k)
I(n) = −
n + 1 k≤n k 3 · 2 k−1 −1 n−k+1

where the primes mean that summation is over even positive k and B(k) are Bernoulli
numbers
5.14.3 For λ → ∞, we have

λsc I(λ) = Φ(log2 λ) + O(λ−sc −1 ),

λsc E(−λ) = Φ(log2 λ) + O(λ−sc −1 )

where Φ is the function analytic in the strip |Im(z)| < π/(2 lg 2),

+∞
2  2πin   2πin 
Φ(z) = Γ sc + ζ sc + exp(−2πinz).
−∞
3 lg 2 lg 2 lg 2

Proposition 5.15. The equality

 1 a 
∞ a
ax
(5.16) e dG(x) = exp cosh k
0 2 k=1 3

holds for every a ∈ C.

Proof. Write  1
Φ(a) := eax dG(x).
0

It follows from (3.9)–(3.11) that

x 
1 2 x 1 1
G = G(x), G + = + G(x).
3 2 3 3 2 2
Hence
 1/3  1  1 x
ax
ax ax
Φ(a) = e dG(x) + e dG(x) = e 3 dG
0 2/3 0 3
 1   
a( 23 + x3 ) 2 x 1 1 ax 1 1 a( 23 + x3 )
+ e dG + = e dG(x) +
3 e dG(x)
0 3 3 2 0 2 0
1 2
a a
a a
= (1 + ea 3 )Φ = e 3 cosh Φ .
2 3 3 3

     25

Consequently, by successive iteration we obtain

a a a a a a a
Φ(a) = exp + + ... + n cosh cosh ...cosh k Φ k .
3 9 3 3 9 3 3
Since Φ(x) → 1 as x → 0, the last formula implies (5.16).

1
Observe that the function Φ, Φ(a) = eax dG(x), is the exponential generating
0
function of the moment sequence {mn }.
 1 The
2πinx
next result follows from (5.16) and shows that Fourier coefficients µ̂n , µ̂n :=
0
e dµC (x), of µC do not tend to zero as n → ∞.

Proposition 5.17. [HiTa]. For all integers n,


∞ 
πin 2πn
µ̂n = e cos .
j=1
3j

Corollary 5.18. [HiTa]. Let k be a positive integer and let n = 3k . Then

∞ 
2π
µ̂n = − cos ν
= const = 0.
ν=1
3


∞
2π
Remark 5.19. The infinite product cos 3ν
converges absolutely because
ν=1

2π π π2
2
|1 − cos | ≤ 2 sin < 2 ,
3ν 3ν 9ν

∞
2π
and hence cos 3ν
= 0.
ν=1

In the paper [WW] Wiener and Wintner proved the more general result similar
to Corollary 5.18. The study of the Fourier asymptotics of Cantor type measures
has been extended in [Str1, 2, 3], [LaWa], [HuLa].

Proposition 5.20. The length of the arc of the curve y = G(x) between the points
(0, 0) and (1, 1) is 2.

Remark 5.21. The detailed proof of this proposition can be found in [Dar1]. In
fact, this follows rather easily from 1.1.4. Probably the length of the curve y = G(x)
was first calculated in [HiTa].

     26

The next theorem is a refinement of Proposition 5.20.

Theorem 5.22. Let F : [0, 1] → R be a continuous, increasing function for which

F (0) = 0 and F (1) = 1. Then the following two statements are equivalent.
5.22.1 The length of the arc y = F (x), 0 ≤ x ≤ 1, is 2.
5.22.2 The function F is singular.

This theorem follows from the results of M.J.Pelling [Pel]. See also [DoMa].

6 Some topological properties

There exists a simple characterization of the Cantor function up to a homeomor-
phism. If Φ : X → Y and F : X → Y are continuous functions from topological
space X to topological space Y , then F is said to be (topologically) isomorphic to Φ
if there exist homeomorphisms η : X → X and ψ : Y → Y such that F = ψ ◦ Φ ◦ η
[RoFu, Chapter 4, §1, Section 1]. If the last equality holds with ψ equal the identity
function, then F and Φ will be called the Lebesgue equivalent functions (Cf. [GNW,
Definition 1.1]).

Proposition 6.1. The Cantor function G is isomorphic (as a map from [0, 1] into
[0, 1]) to a continuous monotone function q : [0, 1] → [0, 1] if and only if the set of
constancy Lg is everywhere dense in [0, 1] and

(6.2) g({0, 1}) = {0, 1} ⊆ [0, 1]\Lg .

Lemma 6.3. Let A, B be everywhere dense subsets of [0, 1] and let f : A → B be

an increasing bijective map. Then f is a homeomorphism and, moreover, f can be
extended to a self-homeomorphism of the closed interval [0, 1].

Proof. It is easy to see that

0 = lim
x→0
f (x) = 1 − lim
x→1
f (x)
x∈A x∈A

and
lim f (x) =
x→t
lim f (x)
x→t
x∈A∩[0,t] x∈A∩[t,1]

for every t ∈ (0, 1). Thus there exists the limit

f˜(t) := lim
x→t
f (x), t ∈ [0, 1].
x∈A

     27

Obviously, f˜ is strictly increasing and f˜(t) = f (t) for each t ∈ A. Suppose that

f˜− (x0 ) := x→x lim f˜(x) := f˜+ (x0 )

lim f˜(x) < x→x
0 0
x∈[0,x0 ) x∈(x0 ,1]

for some x0 ∈ (0, 1). Since B is a dense subset of [0, 1], there is b0 ∈ B such that

f (x0 ) = b0 ∈ (f˜− (x0 ), f˜+ (x0 )).

Let b0 = f (a0 ), where a ∈ A. Now we have the following:

(i) a0 = x0 , because b0 = f˜(x0 ),
(ii) a0 ∈/ [0, x0 ), because b0 > f˜− (x0 ),
(iii) a0 ∈/ (x0 , 1], because b0 < f˜+ (x0 ),
and this yields a contradiction. Reasoning similary, we can prove the continuity of
f˜ for the points 0 and 1. Hence, f˜ is a continuous bijection of the compact set [0, 1]
onto itself and every such bijection is a homeomorphism.

Proof of Proposition 6.1. It follows directly from the definition of isomorphic

functions that (6.2) holds and

(6.4) Clo(Lg ) = [0, 1],

if G is isomorphic to a continuous function g : [0, 1] → [0, 1].

Suppose now that g : [0, 1] → [0, 1] is a continuous monotone function for which
(6.2) and (6.4) hold. We may assume, without loss of generality, that g is increasing.
It follows from Lemma 2.6, that a set [0, 1]\Lg is compact and perfect. Moreover,
(6.2) and (6.4) imply that [0, 1]\Lg is a nonempty nowhere dense subset of [0, 1].
Hence, there exists an increasing homeomorphism η : [0, 1] → [0, 1] such that

η(C) = [0, 1]\Lg .

In fact, there is an order preserving homeomorphism η 0 : C → [0, 1]\Lg [Al, Chapter

4, §6, Theorem 25]. It can be extended on each complementary interval J ∈ I as a
linear function. The resulting extension η is strictly increasing and maps [0, 1] onto
[0, 1]. Hence, by Lemma 6.3, η is a homeomorphism.
It is easy to see that a set of constancy of g ◦ η coincides with I ◦ , the set of
constancy of G. Set  
gI ◦ := g(J), GI ◦ := G(J)
J∈I J∈I

(g(J) and G(J) are one-point sets for every J ∈ I).

The maps I  J → g(J) ∈ gI ◦ and I  J → G(J) ∈ GI ◦ are one-to-one and
onto. Hence the map

Ψ◦ : gI ◦ → GI ◦ , Ψ◦ (g(J)) = G(J), J ∈ I

     28

is a bijection. It follows from the definition of G and (6.4) that

Clo(gI ◦ ) = Clo(GI ◦ ) = [0, 1].

Moreover, since g and G are increasing functions, the function Ψ◦ is strictly increas-
ing. Hence, by Lemma 6.3, Ψ◦ can be extended to a homeomorphism ψ : [0, 1] →
[0, 1]. It is easy to see that
G(x) = ψ(g(η(x)))
for every x ∈ I ◦ . Since I ◦ is a everywhere dense subset of [0, 1], the last equality
implies that
G = ψ ◦ g ◦ η.
Thus g and G are topologically isomorphic.

Let f : [0, 1] → R be a continuous function whose total variation is finite. Then,

by the theorem of Bruckner and Goffman [BrGo], f is Lebesgue equivalent to some
function with bounded derivative. As a corollary we obtain

Proposition 6.5. There exists a differentiable function f with an uniformly bounded

derivative f  such that G and f are Lebesgue equivalent.

We now recall a notion of set of varying monotonicity ([GNW, Definition 3.7]).

Let f : [a, b] → R. A point x ∈ (a, b) is called a point of varying monotonicity of
f if there is no neighborhood of x on which f is either strictly monotonic or constant.
We also make the convention that both a and b are points of varying monotonicity
for every f : [a, b] → R.
Let us denote by Kf the set of points of varying monotonicity of f . Then,
as was shown by Bruckner and Goffman in [BrGo], m1 (f (Kf )) = 0 if and only if
f is Lebesgue equivalent to some continuously differentiable function. For every
homeomorphism ψ : [0, 1] :→ [0, 1] we evidently have

Kψ◦G = C, ψ(G(C)) = [0, 1].

Thus, we obtain.

Proposition 6.6. If F is topologically isomorphic to the Cantor function G, then

F is not continuously differentiable.

Let GC be the restriction of G to the Cantor set C. As it is easy to see, GC is

continuous, closed but not open. For example we have
      
1 1 1
GC 0, ∩ C = G 0, = 0, .
2 2 2

     29

However, GC is weakly open in the following sence.

Proposition 6.7. Let A be a nonempty subset of C. If the interior of A is nonempty

in C, then the interior of GC (A) is nonempty in [0, 1], or in other words

(6.8) (IntC A = ∅) ⇒ Int[0,1] G(A) = ∅) .

Proof. Suppose that A ⊆ C and IntC A = ∅. Since C ◦ is a dense subset of C, there

is an interval (a, b) such that both a and b are in C ◦ and

A ⊇ (a, b) ∩ C.

It is easy to see that

(6.9) (x < t < y) ⇒ (G(x) < G(t) < G(y))

for all x, y in [0, 1] and every t ∈ C ◦ . Hence

(6.10) (G(a), G(b)) = G((a, b) ∩ C) ⊆ G(A) = GC (A).

Recall that a subset B of topological space X is residual in X if X\B is of the

first category in X.

Proposition 6.11. If B is a residual (first category) subset of [0, 1], then G−1
C (B)
is residual (first category) in C.

Proof. Let B be a residual subset of [0, 1]. Write

K1 := [0, 1]\C ◦ , W := G−1 (B).

It is sufficient to show that W ∩ C ◦ is residual in C.

Since G is one–to–one on C ◦ ,

(6.12) K1 ∪ C ◦ = [0, 1] = G(K1 ) ∪ G(C ◦ ),

and

(6.13) C ◦ ∩ K1 = ∅ = G(C ◦ ) ∩ G(K1 ).

Moreover, we have

(6.14) W ∩ C ◦ = G−1 (G(W ) ∩ G(C ◦ )) = G−1 ◦

C (G(W ) ∩ G(C )).

     30

It follows immediately from (6.12) and (6.13) that

G(W ) ∩ G(C ◦ ) = G(W ) ∩ ([0, 1]\G(K1 )).

The set G(K1 ) is countable. Thus [0, 1]\G(K1) is residual and G(W ) ∩ G(C ◦ ) is
residual too, as an intersection of residual sets. Hence, there is a sequence {On }∞
n=1
for which

∞
◦
(6.15) G(W ) ∩ G(C ) ⊇ On
n=1

and each On is a dense open subset of [0, 1]. Since GC is continuous on C ◦ , (6.14)
and (6.15) imply that

∞
◦
W ∩C ⊇ G−1
C (On ),
n=1

where each G−1

C (On ) is open in C. Suppose that we have

IntC (C\G−1
C (On0 )) = ∅

for some positive integer n0 . Then by Proposition 6.7 we obtain

Int[0,1] ([0, 1]\On0 ) = ∅

−1
contrary to the properties of {On }∞
n=1 . Consequently GC (On ) is dense in C for each
positive integer n. It follows that W ∩ C ◦ is residual in C. For the case where B
is of the first category in [0, 1] the conclusion can be obtained by passing to the
complement.

Remark 6.16. In a special case this proposition was mentioned without proof in
the work by J.A.Eidswick [Eid].

Proposition 6.17. Let B ⊆ C. Then B is a everywhere dense subset of C if and

only if GC (B) is a everywhere dense subset of [0, 1].

Proof. Since GC is continuous, the image GC (B) is a dense subset of [0, 1] for every
dense B. Suppose that
Clo[0,1] (GC (B)) = [0, 1]
but
C\CloC (B) = ∅.
GC is a closed map, hence

(6.18) GC (CloC (B)) ⊇ Clo[0,1] (GC (B)) = [0, 1].


     31

As in the proof of Proposition 6.7 we can find a two-point set {a, b} ⊆ C ◦ such that
(6.10) holds with A = C\CloC (B). Taking into account implication (6.9) we obtain
that
G(x) ∈/ (G(a), G(b))
for every x ∈ CloC (B), contrary to (6.18).

Remark 6.19. If (X, τ1 , τ2 ) is a space with two topological structures τ1 and τ2 ,

then one can prove that the condition
(Intτ1 (A) = ∅) ⇔ (Intτ2 (A) = ∅) ∀A ⊆ X
is equivalent to
(Cloτ1 (A) = X) ⇔ (Cloτ2 (A) = X) ∀A ⊆ X.

The formulations and proofs of Propositions 6.7, 6.10, 6.17 can be easily carried
over to the general case of functions which are topologically isomorphic to G. In the
rest of this section we discuss a new characterization for such functions.
We say that a subset A of R has the Baire property if there is an open set
U ⊆ R such that the symmetric difference A∆U, A∆U = (A\U) ∪ (U\A), is of first
category in R.

Theorem 6.20. The Cantor function G is isomorphic (as a map from [0, 1] into
R) to a continuous monotone function f : [0, 1] → R with {0, 1} ⊆ ([0, 1]\Lf ) if and
only if the inverse image f −1 (A) has the Baire property for every A ⊆ R.

Proof. Let f : [0, 1] → R be a function which is topologically isomorphic to G

and let A ⊆ R. Then it follows from Proposition 6.1 that the set of constancy Lf
is everywhere dense in [0, 1]. Reasoning as in the proof of Proposition 1.5 we can
prove that f −1 (A) is the union of a Fσ subset of [0, 1] with some nowhere dense set.
Since the collection of all subsets of [0, 1] having the Baire property forms σ-algebra
[Oxt, Theorem 4.3] we obtain the Baire property for f −1 (A).
Suppose now that f : [0, 1] → R is a continuous monotone function, {0, 1} ⊆
([0, 1]\Lf ), but f is not isomorphic to G. Then using Proposition 6.1 we see that
there is an open interval (a, b) ⊆ ([0, 1]\Lf ). Let B be subset of (a, b) which does
not have the Baire property. It is easy to see that B = f −1 (f (B)). Thus, there
exists a set A such that f −1 (A) does not have a Baire property.

Remark 6.21. The existence of a subset of the reals not having the Baire property
depends on the axiom of choice. In fact, from the axiom of determinateness it follows
that every A ⊆ R is Lebesgue measurable (Cf. Propositions 1.4, 1.5) and has the
Baire property. See, for example, [Kan] and [Kl].

     32

7 Dini’s derivatives
We recall the definition of the Dini derivatives. Let a real–valued function F be
defined on a set A ⊆ R and let x0 be a point of A. Suppose that A contains some
half-open interval [x0 , a). The upper right Dini derivative D + F of F at x0 is defined
by
F (x) − F (x0 )
D + F (x0 ) = lim sup .
x→x0 x − x0
x∈(x0 ,a)

We define the other three extreme unilateral derivatives D+ F , D − F and D− F sim-

ilarly. (See [Br] for the properties of the Dini derivatives).
Let C be the Cantor ternary set and G the Cantor ternary function.
We first note that the upper Dini derivatives D + G and D − G are always +∞
or 0.
We denote by M the set of points where G fails to have a finite or infinite
derivative. It is clear that M ⊆ C.
Since G satisfies Banach’s (T1) condition (see Proposition 1.10), the following
conclusion holds.

Proposition 7.1. The set G(M) has a (linear Lebesgue) measure zero.

Proof. See Theorem 6.2 in Chapter IX of [Sa].

lg 2
Corollary 7.2. Let Hsc be a Hausdorff measure with sc = lg 3
. Then we have

Hsc (M) = 0.

Proof. This follows from Propositions 7.1 and 4.5.

Write
W := {x ∈ (0, 1) : D− G(x) = D+ G(x) = 0}.
In the next theorem we have collected some properties of Dini derivatives of G which
follow from [Mor].

Proposition 7.3.
7.3.1. The set G(W ) is residual in [0, 1].
7.3.2. For each λ ≥ 0 the set {x ∈ [0, 1) : D+ G(x) = λ} has the power of the
continuum.

     33

Remark 7.4. However, the set

{x ∈ [0, 1) : D + G(x) = λ}
is void whenever 0 < λ < ∞.

Corollary 7.5. The set W ∩ C is residual in C.

Proof. The proof follows from 7.3.1 and Proposition 6.11.

Remark 7.6. This corollary was formulated without a proof in [Eid].

Let x be a point of C. Let us denote by zx (n) the position of the nth zero in
the ternary representation of x and by tx (n) the position of the nth digit two in this
representation.
The next theorem was proved by J.A.Eidswick in [Eid].

Theorem 7.7. Let x ∈ C ◦ and let

3zx (n)
(7.8) λx := lim inf .
n→∞ 2zx (n+1)
Then
(7.9) λx ≤ D+ G(x) ≤ 2λx .
Furthermore, if lim zx (n + 1) − zx (n) = ∞, then D+ G(x) = λx .
n→∞
Similarly, if
3tx (n)
(7.10) µx := lim inf ,
n→∞ 2tx (n+1)
then
(7.11) µx ≤ D− G(x) ≤ 2µx ,
with the equality on the right if lim tx (n + 1) − tx (n) = ∞.
n→∞

Corollary 7.12. If x ∈ C ◦ , then G (x) = ∞ if and only if λx = µx = ∞.

Remark 7.13. This Corollary implies that the four Dini derivatives of G agree at
x0 ∈ (0, 1) if and only if either x ∈ I ◦ or λx = µx = ∞.

Corollary 7.14. If x ∈ C ◦ , then D− G(x) = D+ G(x) = 0 if and only if

λx = µx = 0.

     34

The following result is an improvement of 7.3.2.

Theorem 7.15. [Eid] For each µ ∈ [0, ∞] and λ ∈ [0, ∞], the set

{z ∈ C : D− G(z) = λ and D+ G(z) = µ}

has the power of the continuum.

Corollary 7.16. For each µ ∈ [0, ∞] and λ ∈ [0, ∞] there exists a nonempty,
compact perfect subset in {z ∈ C : D− G(z) = λ and D+ G(z) = µ}.

Proof. Since G is continuous on [0, 1], each of the Dini derivatives of G is in a Baire
class two [Br, Chapter IV, Theorem 2.2]. Hence, the set {z ∈ C : D− G(z) = λ and
D+ G(z) = µ} is a Borel set. It is well-known that, if A is uncountable Borel set in
a complete separable topological space X, then A has a nonempty compact perfect
subset [Kur, §39, I, Theorem 0].

Recall that M is the set of points at which G fails to have a finite or infinite
derivative and sc = lg 2
lg 3
. The next result was proved by Richard Darst in [Dar2].

Theorem 7.17. The Hausdorff dimension of M is s2c , or in other words

(7.18) dimH M = s2c .

Remark 7.19. It is interesting to observe that the packing and box counting
dimension are equal for M and C

(7.20) dimP M = dimB M = dimP C = dimB C.

The proof can be found in [DeLi], see also [Fal2] and [LXD] for the similar question.

The following proposition was published in [Dar2] with a short sketch of the
proof.

Proposition 7.21. The equality

dimH (G(M)) = sc

holds.

     35

An equality which is similar to (7.18) was established in [Dar3] for more general
Cantor sets. Some interesting results about differentiability of the Cantor functions
in the case of fat symmetric Cantor sets can be found in [Dar1] and [Den]. The
Hausdorff dimension of M was found by J.Morris [Mo] in the case of nonsymmetric
Cantor functions.

8 Lebesgue’s derivative
The so-called Lebesgue’s derivative or first symmetric derivative of a function f is
defined as
f (x + ∆x) − f (x − ∆x)
SDf (x) = lim .
∆x→0 2∆x
It was noted by J.Uher that the Cantor function G has the following curious property.

Proposition 8.1. [Uh] G has infinite Lebesgue’s derivative, SDG(x) = ∞, at every

point x ∈ C\{0, 1}, and SDG(x) = 0 for all x ∈ I ◦ .

Remark 8.2. It was shown by Buczolich and Laczkovich [BuLa] that there is no
symmetrically differentiable function whose Lebesgue’s derivative assumes just two
finite values. Recall also that if f is continuous and has a derivative everywhere
(finite or infinite), then the range of f  must be a connected set.

From the results in [BuLa] we obtain

Theorem 8.3. For all sufficiently large b the inequality

G(x + b∆x) − G(x − b∆x)
(8.4) lim sup <b
∆x→0+ G(x + ∆x) − G(x − ∆x)
holds for all x ∈ C\{0, 1}.

Remark 8.5. It was shown in [BuLa] that (8.4) is true for all b ≥ 81 and false if
b = 4. Moreover, as it follows from [BuLa] (see, also [Thom, Theorem 7.32]) that
Theorem 8.3 implies Proposition 8.1.

9 Hölder continuity, distortion of Hausdorff di-

mension, sc-densities
lg 2
Proposition 9.1. The function G satisfies a Hölder condition of order sc = lg 3
,
the Hölder coefficient being not greater than 1. In other words
(9.2) |G(x) − G(y)| ≤ |x − y|sc

     36

for all x, y ∈ [0, 1]. The constants sc and 1 are the best possible in the sense that the
inequality
|G(x) − G(y)| ≤ a|x − y|β
does not hold for all x, y in [0, 1] if either β > sc or β = sc and a < 1.

Proof. Using Lemma from [GorKu] we obtain the inequality

 1  sc
(9.3) t ≤ G(t) ≤ tsc
2
for all t ∈ [0, 1]. Since G is its first modulus of continuity (see Proposition 2.2), the
last inequality implies (9.2). Moreover it follows from (9.3) that sc is a sharp Hölder
exponent in (9.2). By a simple calculation we have from formula (0.2) that
G(3−n ) = 3−nsc
for all positive integers n. Thus, the Hölder coefficient 1 is the best possible in
(9.2).

Remark 9.4. It was first proved in [HiTa] that G satisfies the Hölder condition with
the exponent sc and coefficient 2. R.E.Gilman erroneously claimed (without proof
and for a more general case) that both constants sc and 2 are the best possible [Gil].
In [GorKu] a similar to (9.3) inequality does not link with the Hölder condition for
s
G. Observe also that 12 c is the sharp constant for (9.3) as well as the constant 1.
See Figure 4.

Proposition 9.5. Let x be an arbitrary point of C. Then we have

lg |G(x) − G(y)|
lim inf = sc .
y→x
y∈C
lg |x − y|

Proof. It suffices to show that

lg |G(x) − G(y)|
lim inf ≤ sc .
y→x
y∈C
lg |x − y|
2
Choosing |x − y| = 3n
for n → ∞ we obtain the result.

Let x be a point of the Cantor ternary set C. Then x has a triadic representation

∞
αm
(9.6) x=
m=1
3m

     37

where αm ∈ {0, 2}. Define a sequence {Rx (n)}∞n=1 by the rule


inf{m − n : αm = αn , m > n} if ∃m > n : αm =  αn
Rx (n) :=
0, if ∀m > n : αm = αn ,

i.e., Rx (n) = 1 ⇐⇒ (αn = αn+1 );

Rx (n) = 2 ⇐⇒ (αn = αn+1 ) & (αn+1 = αn+2 )

and so on.

1 sc
Figure 4: The graph of G lies between two curves y1 (x) = xsc and y2 (x) = 2
x .

Theorem 9.7. [DMRV] For x ∈ C we have

lg |G(x) − G(y)|
lim = sc
y→x
y∈C
lg |x − y|

if and only if
Rx (n)
lim = 0.
n→∞ n

This theorem and a general criterion of constancy of linear distortion of Hausdorff

dimension (see [DMRV]) imply the following result.

     38

Theorem 9.8. There exists a set M ⊆ C such that Hsc (M) = 1 and
1
dimH (G(A)) = dimH (A)
sc
for every A ⊆ M.

Remark 9.9. In the last theorem we can take M as

M = G−1 (SN2 )

where SN2 is the set of all numbers from [0, 1] which are simply normal to base 2.
See, for example, [Ni] for the definition and properties of simply normal numbers.

Let Θ∗sc (t) and Θs∗c (t) denote the upper and lower sc -densities of µC at t ∈ R,
i.e.,

|Ĝ(t + x) − Ĝ(t − x)|

Θ∗sc (t) := lim sup
x→0 |2x|sc
x=0

|Ĝ(t + x) − Ĝ(t − x)|

Θs∗c (t) := lim inf .
x→0
x=0
|2x|sc
It can be shown that
Θ∗sc (t) < Θs∗c (t)
for every t ∈ C, i.e., sc -density for measure µC does not exist at any t at the Cantor
set C. However, the logarithmic averages sc -density does exist for many fracfals
including C.

Theorem 9.10. For Hsc almost all x ∈ C the logarithmic average density
 T
sc 1 Ĝ(x + e−t ) − Ĝ(x − e−t )
A (x) := lim dt
T →∞ T 0 |2e−t |sc

exists with

sc 1
A (x) = sc |x − y|−sc dG(x)dG(y) = 0, 6234...
2 log 2
|x−y|≥ 31

For the proof see [Fal1, Theorem 6.6] and also [BeFi].

     39

The interesting results about upper and lower sc -densities of µC can be found in
[FHW]. For x with representation (9.6) we define τ̂ (x) and τ (x) as


∞
τ̂ (x) := lim inf αm+k 3−m ,
k→∞
m=1

τ (x) := min(τ̂ (x), τ̂ (1 − x)).

Theorem 9.11. [FHW] 9.11.1 For all x ∈ C,

Θs∗c (x) = (4 − 6τ (x))−sc .

9.11.2 For all x ∈ C,



 2−sc if x ∈ C 1
−sc
Θ∗sc = 2

 (2 + τ (x))−sc if x ∈ C ◦ .
3
9.11.3 For all x ∈ C ◦ ,

9(Θ∗sc (x))−1/sc + (Θs∗c (x))−1/sc = 16.

9.11.4 sup{Θ∗sc (x) − Θs∗c (x) : x ∈ C} = 4−sc ,

where the supremum can be attained at {x ∈ C ◦ : τ (x) = 0}, and
−sc −sc
∗sc sc 3 5
inf{Θ (x) − Θ∗ : x ∈ C} = − ,
2 2
 
where the infimum can be attained at x ∈ C ◦ : τ (x) = 14 .
9.11.5 For Hsc -almost all x ∈ C,

Θs∗c (x) = 4−sc , Θ∗sc (x) = 2 · 4−sc .

Remark 9.12. The general result similar to 9.11.5 can be found in the work Salli
[Sal].

Acknowledgements. The research was partially supported by grants from the

University of Helsinki and by grant 01.07/00241 of Scientific Fund of Fundamental
Investigations of Ukraine.

     40

References
[Al] CDEFGHIJKLMNO
           

 PNIHOJ 8QJRHJ8 *%% [P.S.Aleksandrov, ”Introdaction to Set The-
ory and General Topology”. Izdat. ”Nauka”, Moskow, 1977.]

[Bar] N.Bary, Memoire sur la representation finie des fonctions continues.

Math.Ann., 103, (1930), 185–248 and 598–653.

[BaMe] N.Bary et D.Menchoff, Sur l’integrale de Lebesgue-Stieltjes et les fonctions

absolumant continues. Ann. Math. Pura e Appl., 5, (1928), 19–54.

[BeFi] T.Bedford and A.M.Fisher, Analogues of the Lebesgue density theorem for
fractal sets of reals and integer. Proc. London Math. Soc., 64, (1992), 95–
124.

[Br] A.M.Bruckner, ”Differentiation of Real Functions”. Lecture Notes in Math-

ematics 659, Springer–Verlag, Berlin, 1978.

[BrGo] A.M.Bruckner and J.L.Goffman, Differentiability through change of vari-

ables, Proc. Amer. Math. Soc., 61, (1976), 235–241.

[BrLe] A.M.Bruckner and J.L.Leonard, On differentiable functions having an ev-

erywhere dense set of intervals of constancy, Cand. Math. Bull., 8 (1965),
73–76.

[BSU] Yu.M. Berezansky, Z.G.Sheftel, G.F.Us, ”Functional analysis. Volume II”.

Birkhäuser Veglar, Basel, 1996.

[BuLa] Z.Buczolich and M.Laczkovich, Concentrated Borel measures. Acta. Math.

Hungar., 57, (349–362), 1991.

[Can] G.Cantor, De la puissance des ansembles parfaits de points.Acta Math. 4,

(1884), 381–392. Reprinted in G.Cantor, Gezammelte Abhandlungen math-
ematischen und philosophischen Inhalts. E.Zermelo (ed.), Springer-Verlag,
New York, 1980.

[Ch] D.R.Chailice, A characterization of the Cantor function. Amer. Math.

Monthly, 98, (1991), N3, 255–258.

[Cop] W.A.Coppel, An interesting Cantor set. Amer. Math. Monthly, 90, (1983),
N7, 456–460.

[Dar1] R.B.Darst, Some Cantor sets and Cantor functions. Mathematics Magazine,
45, (1972), 2–7.

     41

[Dar2] R.B.Darst, The Hausdorff dimension of the nondifferentiability set of the

Cantor function is [ln 2/ ln 3]2 . Proc. Amer. Math. Soc., 119, N1, (1993),
105–108.

[Dar3] R.B.Darst, Hausdorff dimension of sets of non-differentiability points of

Cantor functions. Math. Proc. Camb. Phil. Soc., 117, (1995), 185–191.

[DeLi] F. M. Dekking and Wenxia Li, How smoth is a devil’s staircase? Fractals,
11, 1, (2003), 101–107.

[Den] T.P.Dence, Differentiable points of the generalized Cantor function. Rocky

Mountain J. Math., 9, 2, (1979), 239–249.

[Dob] J.Dobos̆, The standard Cantor function is subadditive. Proc. Amer. Math.
Soc., 124, (1996), N11, 3425–3426.

[DMRV] O.Dovgoshey, O.Martio, V.Ryazanov, M.Vuorinen, Linear distortion of

Hausdorff dimension and Cantor’s function. to appear in Real Anal. Ex-
change.

[DoMa] O.Dovgoshey and O.Martio, The arc length evaluation and Lebesgue decom-
position. Preprint 379, (2004), University of Helsinki.

[Eid] J.A.Eidswick, A characterization of the nondifferentiability set of the Cantor

function. Proc. Amer. Math. Soc., 42, 1, (1974), 214–217.

[Ev] G.C.Evans, Calculation of moments for a Cantor–Vitali function. Amer.

Math. Monthly, 64, 8, (1957), 22–27.

[EvGa] Lawrence C. Evans and Ronald F. Gariepy ”Measure Theory and Fine Prop-
erties of Functions”. CRC Press, Boca Raton, 1992.

[Fal1] Kenneth Falconer, ”Techniques in Fractal Geometry”. John Wiley, New

York, 1997.

[Fal2] Kenneth J. Falconer, One-side miltifractal analysis and points of non-

differentiability of devil’s staircases. Math. Proc. Cambridge Philos.Soc.,
136, 1, (2004), 167–174.

[FHW] De-Jun Feng, Su Hua, Zhi-Ying Wen, The Pointwise Densites of the Cantor
Measure. J.Math. Anal. Appl. 250, (2000), 692–705.

[Fis] H.-J. Fisher, On the paper: ”Asymptotics for the moments of singular dis-
tributions”. J. Approx. Theory, 82, 3, (1995), 362–374.

     42

[Fle] Julian F.Fleron, A note on the history of the Cantor set and Cantor func-
tion. Mathematics Magazine, 67, (1994), 136–140.

[Gil] R.E.Gilman, A class of functions continuous but not absolutely continuous.

Annals of Math. (2), 33, (1932), 433–442.

[GNW] Casper Goffman, Togo Nishiura, Daniel Waterman, ”Homeomorphisms in

Analysis”. Mathematical Survey and Monographs, Vol. 54. American Math-
ematical Society, Providence, R.I., 1998.

[GoWi] William Goh and Jet Wimp, Asymptotics for the moments of singular dis-
tributions. J. Approx. Theory, 74, 3, (1993), 301–334.

[GorKu] E. A. Gorin and B. N. Kukushkin, Integrals related to the Cantor function.

St. Petersburg Math. J., 15, 3, (2004), 449–468.

[Haw] T.Hawkins, ”Lebesgue’s Theory of Integration”. Chelsea Publishing Co.,

Madison, WI, 1975.

[HiTa] E.Hille and J.D.Tamarkin, Remarks on known example of a monotone con-

tinuous function. Amer. Math. Monthly, 36, (1929), 255–264.

[HuLa] Tian-You Hu and Ka-Sing Lau, Fourier asymptotics of Cantor type measures
at infinity. Proc. Amer. Math. Soc., 130, 9, (2002), 2711–2717.

[Hut] J.E.Hutchinson, Fractals and self–similarity. Indiana Univ. Math. J., 30, 5,
(1981), 713–747.

1S2 TU VJKNOGW
          

PNIHOJ 8QJRHJ8 *(! [V.G. Kanoveı̌, ”The Axiom of Coice and the Ax-
iom of Determinateness”. Izdat. ”Nauka”, Moskow, 1984.]

[Kl] Eugene M. Kleinberg, ”Infinitary Combinatorics and the Axiom of Determi-

nateness”. Lecture Notes in Mathematics, 612, Springer-Verlag, Berlin–New
York, 1977.

[Kur] K.Kuratowski, ”Topology, Volume 1”. Academic Press, New York and Lon-
don, 1966.

[LaWa] K. S. Lau and J. Wang, Mean quadratic variations and Fourier asymptotics
of self-similar measures. Monat. Math., 115, (1993), 99–132.

[LXD] Wenxia Li, Dongmei Xiao and F. M. Dekking, Non-differentiability of devil’s

staircaces and dimentios of subsets of Moran sets. Math. Proc. Cambridge
Philos. Soc., 133, 2, (2002), 345–355.

     43

[Mo] J.Morris, The Hausdorff dimension of the nondifferentiability set of a non-

symmetric Cantor function. Rocky Mountain J. Math., 32, 1 (2002), 357–
370.

[Mor] A.P.Morse, Dini derivatives of continuous functions. Proc. Amer. Math.

Soc., 5, N1, (1954), 126–130.

[Ni] Ivan Niven, ”Irrational numbers”. Wiley and Sons, Inc., New York, 1956.

[Oxt] J.Oxtoby, ”Measure and Category”. Springer-Verlag, New York, 1971.

[Pel] M.J.Pelling, Formular for the arc-length of curve in Rn . Amer. Math.

Montly, 84, 6 (1977), 465–467.

[RoFu] TEXNYFZK [\]RHI
 !" #      $   %


 PNIHOJ 8QJRHJ8 *%% [V.A.Rohlin, D.B.Fuks, ”A Beginnig
Course in Topology: Geometrical Chapters”. Izdat. ”Nauka”, Moskow,
1977.]

[Sa] S.Saks, ”Theory of the Integral”. Second revised edition, Dover Publications,
INC., 1964.

[Sal] A.Salli , Upper density properties of Hausdorff measures on fractals. Ann.

Acad. Sci. Fenn. Ser AI Math. Dissertationes, 55, 1985.

[Str1] R. Strichartz, Fourier asymptotics of fractal measures. J. Functional Anal.,

89, (1990), 154–181.

[Str2] R. Strichartz, Self-similar measure and their Fourier transform I. Indiana

University Math. J., 39, (1990), 797–817.

[Str3] R. Strichartz, Self-similar measure and their Fourier transform II. Trans.
Amer. Math. Soc., 336, N 1, (1990), 335–361.

[Thom] B.C.Thomson, ”Symmetric Properties of Real Function”. Marcel Dekker,

New York, 1994.

[Tim] A.F.Timan, ”Theory of Approximation of Functions of a Real Variable”. In-

ternational Series of Monographs in Pure and Applied Mathematics, Vol.34.
Pergamon Press, Oxford, 1963.

[Uh] J.Uher, Symmetric continuity implies continuity. Rel. Anal. Exchange, 13,
(1987/88), 36–38.

[WW] N. Wiener and A. Wintner, On singular distributions. J. Math. Phy., 17,

(1939), 233–246.

     44

[WY] Norbert Wiener and R. C. Young, The total variation of g(x + h) − g(x).
Trans. Amer. Math. Soc., 35, N 1, (1933), 327–340.

Oleksiy Dovgoshey and Vladimir Ryazanov

Institute of Applied Mathematics
and Mechanics, NAS of Ukraine,
74 Roze Luxemburg str.,
Donetsk, 83114 UKRAINE
email: dovgoshey@iamm.ac.donetsk.ua
ryaz@iamm.ac.donetsk.ua

Olli Martio
Department of Mathematics and Statistics,
P.O. Box 68 (Gustaf Hällströmin katu 2b)
FIN – 00014 University of Helsinki
FINLAND
email: martio@cc.helsinki.fi

Matti Vuorinen
Department of Mathematics,
FIN – 20014 University of Turku
FINLAND
email: vuorinen@csc.fi