The Devil’s Staircase

Recall the usual construction of the Cantor set: C0 = [0, 1], C1 = [0, 13 ] ∪ [ 23 , 1], and in
general Cn is a disjoint union of 2n closed intervals, each of length 3−n , constructed from
Cn−1 by deleting the open-middle-third of each of the 2n−1 intervals constituting Cn−1 .
ThenTthe total length of the intervals in Cn is ( 32 )n . The Cantor set C is the intersection
C = n≥0 Cn .
Set gn = ( 23 )n 1Cn ; that is,
(
( 32 )n , x ∈ Cn
gn (x) = .
0, x∈/ Cn
The function gn is discontinuous only at the 2n+1 points at the boundaries of the intervals
making up Cn . This is a finite set, and so gn is Riemann integrable. So we may define
fn : [0, 1] → R as follows: Z x
fn (x) = gn (t) dt.
0
From the Fundamental Theorem of Calculus, we know that the functions fn are Lipschitz
R1
continuous. Note also that fn (0) = 0, while fn (1) = 0 gn (t) dt. This integral can be
calculated as Z 1
gn (t) dt = ( 23 )n 0 1Cn (t) dt = ( 32 )n · length(Cn ) = 1.
R1
0
So fn is a continuous function with fn (0) = 0 and fn (1) = 1 for each n. Notice that fn is
constant on [0, 1] − Cn , and is linear with slope ( 32 )n on the intervals making up Cn . So fn
is monotone increasing. Here is the graph of f2 .

F IGURE 1. The graph of f2 .

1
2

Let I be any one of the 2n closed intervals that make up Cn . Then gn (x) = ( 32 )n for all
x ∈ I, while gn+1 (x) is equal to ( 32 )n+1 = 32 gn (x) for x in the first third or last third of I, and
equal to 0 in the middle interval. It follows that
Z Z
gn (t) dt = gn+1 (t) dt = 2−n . (∗)
I I

Note also that gn (x) = gn+1 (x) for x ∈

/ Cn . Hence, if a is any endpoint of an interval in Cn ,
Z a Z a
gn (t) dt = gn+1 (t) dt. (∗∗)
0 0

It follows by integrating (∗) that

fn+1 (x) = fn (x), x∈
/ Cn . (∗ ∗ ∗)
Now, the CantorSset C is closed, and so any point x ∈ / C is contained in a small open
interval in C = n≥0 Cn . Therefore there is some N ∈ N such that x ∈ Cnc for all n ≥ N ,
c c

and by the above we have fn (x) = fN (x) for all n ≥ N . This shows that the sequence of
functions fn (x) converges pointwise on C c . But it’s even better than that.
Let x ∈ Cn , and let I = [a, b] be the interval in Cn containing x. Applying (∗∗), we have
Z x  Z x 
   
|fn (x) − fn+1 (x)| =  [gn (t) − gn+1 (t)] dt =  [gn (t) − gn+1 (t)] dt
  
0
Z ax
≤ |gn (t) − gn+1 (t)| dt
a
Z b
≤ |gn (t) − gn+1 (t)| dt.
a

Again, we have |gn (t) − gn+1 (t)| = ( 23 )n+1 − ( 23 )n on the first and last third of [a, b], while it
equals ( 32 )n on the middle third. The difference is therefore bounded above by ( 32 )n+1 on
the interval which has length 3−n , and so the above integral is bounded by 3−n ( 32 )n+1 ; that
is, we have proved that
3 −n
|fn (x) − fn+1 (x)| ≤ · 2 < 2−n+1 , x ∈ Cn .
2
Combining this with (∗ ∗ ∗), we have |fn (x) − fn+1 (x)| < 2−n+1 for all x. It follows that the
sequence {fn } is uniformly Cauchy, and therefore converges uniformly to a limit function
f . The functions fn are continuous, and so the uniform limit function f is continuous.
/ Cn the sequence {fk (x)}∞
Also, for x ∈ k=n is constant, and therefore fn (x) = f (x). But fn
is constant on Cnc . Finally, since fn (0) = 0 and fn (1) = 1 for all n, we have f (0) = 0 and
f (1) = 1. We have therefore proved the following:
fn → f uniformly, where f is continuous, f 0 (x) = 0 for x ∈ C c , and f (0) = 0 and f (1) = 1.
A little additional thought shows that the limit function f is monotone increasing. It
is called the Cantor Function or the Devil’s staircase. Its graph is shown in Figure 2. It
can actually be described in simple terms. Here is an algorithm for calculating f (x) for
x ∈ [0, 1].
• Express x in base 3, [x]3 . (Choose the representation that does not end in 1111 . . .)
 3

F IGURE 2. The graph of the Devil’s staircase.

• If [x]3 contains any 1s, with the first 1 being at position n: [x]3 = 0.x1 x2 . . . xn−1 1xn+1 . . .,
replace the number with T (x) in ternary [T (x)]3 = 0.x1 x2 . . . xn−1 2. Otherwise, if
[x]3 contains no 1s (i.e. x ∈ C), then T (x) = x.
• Since T (x) is an even ternary number (i.e. in C), we can relabel all 2s as 1s. The 0-1
string that remains, re-interpretted in base 2, is f (x).
For example, 1/4 in base 2 is 0.0202020 . . . (isn’t that surprising? 1/4 ∈ C!). Hence,
f [(1/4)]2 = 0.01010101 . . . = 1/3. On the other hand, [1/5]3 = 0.012101210 . . ., hence
[T (1/5)]3 = 0.02, and so [f (1/5)]2 = 0.01, which is 1/4 in base 2; hence, f (1/5) = 1/4.
In general, a non-constant monotone increasing function that is continuous, differentiable
almost everywhere with derivative 0, is called singular. These functions really point out
the necessity of the continuity (and definition everywhere) of the derivative of f in order
for the Fundamental TheoremR of Calculus to make sense. (After all, the derivative of f is
0 almost everywhere, and so f 0 6= f in this case.)
For other fun examples of singular functions, lookup the Wikipedia entries on Minkowski’s
question mark function, and Kolmogorov’s circle map.
One final note: for those who are physically inclined, the 1998 Nobel Prize in Physics was
awarded for the discovery and explanation of what is known as the fractional quantum
Hall effect. This phenomenon is a physical manifestation of a singular function.