Functions of Bounded Variation

S. Kumaresan
School of Math. & Stat.
University of Hyderabad
Hyderabad 500 046
kumaresa@gmail.com

The starting point for the study of functions of bounded variation is the desire to define
the length of a curve. Let γ : [a, b] → RN be a continuous function. We call such a function
as a curve in RN . If we write x(t) := γ(t) = (x1 (t), . . . , xN (t)), then x(t) may be considered
as the position vector of a particle at the instant t. We wish to “compute” the length of the
curve γ. Since we know how to compute the length of a line segment, we approximate the
length of γ by the length of a polygonal path which approximates γ.
Let P := {a = t0 < t1 < · · · < tn = b} be a partition of [a, b]. Then we have a
polygonal
Pn path which is built out of line segments joining x(tj−1 ) to x(tj ). Its length is
j=1 kx(tj ) − x(tj−1 )k. We shall denote this length by L(γ, P ). Our intuition says that if
we keep refining the partition, we shall get better approximation to the length of γ. This
suggests that we define

Length of γ ≡ L(γ) := lub {L(γ, P ) : P is a partition of [a, b]}.

It may happen that L(γ) does not exist in R. If L(γ) exists, we say that the curve γ is
rectifiable and call L(γ) as its length.
Let us keep the notationPas above. Fix 1 ≤ i ≤P N . Note that |xi (t)| ≤ kx(t)k. Hence
it follows that L(xi , P ) := j |xi (tj ) − xi (tj−1 )| ≤ j kx(tj ) − x(tj−1 )k. Hence, if we set
L(xi ) := lub {L(xi , P ) : P is a partition of [a, b]}, then L(xi ) ≤ L(γ). Thus if γ is rectifiable,
then L(xi ) is finite for each i.
Conversely, if each L(xi ) is finite, then γ is rectifiable. We defer the proof of this state-
ment towards the end of this article. We also show (at the end) that if γ is continuous
Rb
differentiable, then γ is rectifiable and that L(γ) = a kγ 0 (t)k dt.
Thus it behooves us to study the special case of real valued functions f on [a, b] and
discuss the so-called length L(f ).
Definition 1. Let f : [a, b] → R be given. Let P := {a = x0 , x1 , . . . , xn = b} be a partition of
[a, b]. The variation of f over [a, b] with respect to the partition P is defined by
n
X
V (f, [a, b], P ) := |f (xj ) − f (xj−1 )| .
j=1

If the interval [a, b] is understood, we simplify the notation as V (f, P ).

1
 We insert a point x ∈ (xj−1 , xj ) to form a new partition Q = {x0 , . . . , xj−1 , x, xj , . . . , xn }.
In the sum for V (f, P ) we replace the term |f (xj − f (xj−1 )| by |f (xj − f (x)| + |f (xj − f (x)|.
This shows that V (f, P ) ≤ V (f, Q).
This leads us to conclude that V (f, P ) ≤ V (f, Q) if Q is a refinement of P .

Definition 2. The variation f on [a, b] is defined by

V (f, [a, b]) := lub {V (f, [a, b], P ) : P is a partition of [a, b]}.

We say that f is of bounded variation on [a, b] if V (f, [a, b]) is finite.

Remark 3. If f is of bounded variation on [a, b], then f is bounded on [a, b]. Let x ∈ [a, b]
and consider the partition P := {a, x, b}. Then we have

|f (a) − f (x)| ≤ |f (x) − f (a)| + |f (b) − f (x)| = V (f, P ) ≤ V (f, [a, b]).

By triangle inequality we obtain

|f (x)| = |f (x) − f (a) + f (a)| ≤ |f (x) − f (a)| + |f (a)| ≤ V (f, [a, b]) + |f (a)| .

Example 4. Any monotone function f : [a, b] → R is of bounded variation and we have

Vab (f ) = |f (b) − f (a)|. Assume that f is increasing. Observe that |f (xj ) − f (xj−1 )| = f (xj )−
f (xj−1 ). Hence the sum V (f, [a, b], P ) is telescopic sum.
X X
|f (xj ) − f (xj−1 )| = f (xj ) − f (xj−1 ) = f (b) − f (a).
j j

Example 5. Let f : [a, b] → R be Lipschitz, say, |f (x) − f (y)| ≤ L |x − y|. Then f is of

bounded variation on [a, b]. Observe that, for any partition, we have
X X
|f (xj ) − f (xj−1 )| ≤ L(xj − xj−1 ) = L(b − a).
j j

Example 6. Let f : [a, b] → R be continuously differentiable (that is, f 0 (x) exists on [a, b] and
f 0 is continuous on [a, b]). Then f is of bounded variation.
Since f 0 is continuous on [a, b], there exists M > 0 such that |f 0 (x)| ≤ M . For, x, y ∈ [a, b],
mean value theorem tells us that there exists z between x and y such that f (x) − f (y) =
f 0 (z)(x − y). It follows that |f (x) − f (y)| ≤ M |x − y|. Thus, f is Lipschitz. The result follows
from the last example.

Example 7. (We now give an example of a bounded function which is not of bounded variation.
sin(1/x) x 6= 0 2
Let f (x) = for x ∈ [0, 1]. Let x0 = 0 and xj = (n−j)π for 0 < j < n. We
0 x=0
then have
X n
|f (xj ) − f (xj−1 )| = 2n.
j=1

Hence F is not of bounded variation.

2
Example 8. The function f (x) = x sin(1/x) for 0 < x ≤ 2/π and f (0) = 0 is continuous,
bounded but is not of bounded variation. (Draw the graph of this function to see why.)
Let
2 2 2 2 2
x0 = 0 < x1 = , < x2 = < x3 = · · · < x2n = < x2n+1 =
(2n + 1)π (2n)π (2n1)π 2π π
be a partition of the domain. Then we have
n
X 2 2 2 2
|f (xj ) − f (xj−1 )| = + + + ··· +
π 2π 3π (2n + 1)π
j=1
2n+1
!
2 X 1
= .
π k
k=1
P 1
Since the harmonic series k k is divergent, it follows that the function is not of bounded
variation.
Proposition 9. If f, g are of bounded variation on [a, b] so are f + g and f − g.

Proof. Observe that we have

n
X X X
|f (xj ) ± g(xj ) − [f (xj−1 ) ± g(xj−1 )]| ≤ |f (xj ) − f (xj−1 )| + |g(xj ) − g(xj−1 )|
j=1 j j

≤ V (f, [a, b]) + V (g, [a, b]).

Definition 10. Total variation of f : [a, b] → R, a function of bounded variation is the function
F defined as follows:

F (x) := V (f, [a, x]) ≡ lub {V (f, [a, x], P ) : where P is a partition of [a, x]}.

Theorem 11. Let f : [a, b] → R be of bounded variation and F its total variation. Then
(i) |f (x) − f (y)| ≤ |F (x) − F (y)| for a ≤ x < y ≤ b.
(ii) F and F − f are increasing on [a, b] and
(iii) Vab (f ) ≤ Vab (F ).

Proof. The idea is very simple. If x < y, then any partition P = {x0 , . . . , xn = x} of [a, x]
gives rise to a partition Q of [a, y] in a natural way: Q = {a = x0 , . . . , xn = x, y}. Hence we
obtain
n
X n
X
|f (xj ) − f (xj−1 )| ≤ |f (xj ) − f (xj−1 )| + |f (y) − f (x)|
j=1 j=1

≤ F (y).

If we take the LUB of this inequality over all the partitions of [a, x] we obtain

F (x) ≤ F (x) + |f (y) − f (x)| ≤ F (y).

3
(i) follows from this.
If ∅ 6= A ⊂ B ⊂ R, then we know lub A ≤ lub B. Given any partition P of [a, x], let
Q denote the partition of [a, y] as in (i). Then we have V (f, [a, x], P ) ≤ V (f, [a, y], Q).
If we take the LUB of this inequality over all the partitions of [a, x], we get F (x) ≤
lub {V (f, [a, y], Q)} ≤ F (y). Hence it follows that F is increasing on [a, b].
To show that F − f is increasing, we use the inequality of (i):

f (y) − f (x) ≤ |f (y) − f (x)| ≤ F (y) − f (x).

Therefore, F (x) − f (x) ≤ F (y) − f (y). Thus (ii) is proved.

Let P be a partition of [a, b]. Using (i), we obtain
X X
|f (xj ) − f (xj−1 )| ≤ |F (xj ) − F (xj−1 )| ≤ V (F, [a, b]).
j j

If we now take the LUB of this inequality over all partitions of [a, b], we obtain V (f, [a, b]) ≤
V (F, [a, b]).

Theorem 12 (Jordan). Let f : [a, b] → R. Then f is of bounded variation iff there exist
increasing functions F and G such that f = F − G.

Proof. Take G = F − f where F is the total variation of f .

Theorem 13. Let f : [a, b] → Rn be C 1 . Then we have

Z b
L(f ) := V (f, , [a, b]) = f 0 (t) dt.
a

Proof. Let P = {x0 , x1 , . . . , xn } be a partition of [a, b]. By the fundamental theorem of

calculus, we have
Z xj Z xj
0
kf (xj ) − f (xj−1 )k = f (t) dt ≤ f 0 (t) dt.
xj −1 xj −1

Summing this inequality over j we obtain

X Z b
V (f, [a, b], P ) ≡ kf (xj ) − f (xj−1 )k ≤ f 0 (t) dt.
j a

Rb
Taking the LUB over all partitions, we get V (f, [a, b]) ≤ a kf 0 (t)k dt.
To get the reverse inequality, let ε > 0 be given. Since f 0 is continuous on [a, b], it
is uniformly continuous on [a, b]. Hence there exists δ > 0 such that for s, t ∈ [a, b] with
|s − t| < δ, we have kf 0 (s) − f 0 (t)k < ε. We subdivide [a, b] into N subintervals of equal
length (b − a)/N in such a way that (b − a)/N < δ. Let P = {x0 , x1 , . . . , xN } be the partition.
Observe that for t ∈ [xj−1 , xj ], we have kf 0 (t) − f (xj )k < ε. Hence for such t, we have

f 0 (t) ≤ f 0 (xj ) + ε. (1)

4
We integrate this inequality on [xj−1 , xj ] and obtain
Z xj Z xj
0
f (t) dt ≤ f 0 (xj ) dt + ε(xi − xj−1 )
xj−1 xj−1
0
= f (xj ) (xj − xj−1 ) + ε(xi − xj−1 )
Z xj
 0
f (t) + f 0 (xj ) − f 0 (t) dt + ε(xi − xj−1 )

=
xj−1
Z xj Z xj
0
 0
f (xj ) − f 0 (t) dt + ε(xi − xj−1 )

= f (t) dt +
xj−1 xj−1
Z xj Z xj
0
f 0 (xj ) − f 0 (t) dt + ε(xi − xj−1 )


≤ f (t) dt +
xj−1 xj−1
Z xj Z xj
f 0 (t) dt + f 0 (xj ) − f 0 (t) dt + ε(xi − xj−1 )


≤
xj−1 xj−1
0
≤ f (xj ) − f (xj−1 ) + 2ε(xi − xj−1 ).
Summing over j, we arrive at
Z b
f 0 (t) dt ≤ V (f, P ) + 2ε(b − a) ≤ V (f, [a, b]) + 2ε(b − a).
a
Rb
Since ε > 0 is arbitrary, it follows that a kf 0 (t)k dt ≤ V (f, [a, b]) and hence we conclude
Rb
that a kf 0 (t)k dt = V (f, [a, b]) = L(f ).

We now prove that γ = (x1 , . . . , xn ) is rectifiable if each xi is of “rectifiable” which is

same saying that each xi is of bounded variation. This followsP from the equivalence of the
1 norm on RN : kxk := N
standardP(Euclidean) norm with the L 1 1 |xi |. Note that we have
N
P
kxk1 = j |xj | ≤ j kxk ≤ N kxk for x ∈ R . (In fact, using Cauchy-Schwartz, we have
√
kxk1 ≤ N kxk.) On the other hand, since kxk21 ≥ kxk2 , we have kxk ≤ kxk1 .
Now if L(xi ) = Li , for any partition P of [a, b], we have
X X
L(γ, P ) := kx(tj ) − x(tj−1 )k ≤ kx(tj ) − x(tj−1 )k1
j j
N
XX
= |xi (tj ) − x(tj−1 )|
j i=1
N X
X
= |xi (tj ) − x(tj−1 )|
i=1 j
X
= L(xi , P )
i
X
≤ Li .
i

This establishes our claim that γ is rectifiable iff each component function xi of γ is “recti-
fiable” or is of bounded variation.