Está en la página 1de 28

The Riemann-Stieltjes Integral

A. X. Athens
December 2011

Contents
1 Introduction

2 Functions of bounded variation

3 The Riemann-Stieltjes Integral

4 Riemann-Stieltjes measures *

17

5 Path integration

21

A Free groups

27

Introduction

The Riemann-Stieltjes (RS) integral is often used for introductory courses


on analysis because it is adequate for much of real and complex analysis and takes less work to develop than the Lebesgue integral. Its very
construction reveals the meaning of integration along paths in a way
that is lacking in other approaches. While this integral is good for one
dimension of the independent variable (intervals in R and paths in C
or Rn ), its generalization to two or more dimensions is limited to simple geometries like rectangles. This is where measure theory and the
Lebesgue integral become indispensable.
On a finite interval [a, b], the class of Lebesgue integrable functions
includes all RS integrable function. On the other hand, there are situ-

ations where the Lebesgue integral does not exist but the RS integral
provides a useful limit (principal values and improper integrals).
Our treatment will be sufficiently general to allow for real, complex
and vector valued functions. In fact, we permit functions to take values
in an arbitrary Banach space.
Sufficient prerequisites for the main part of this article are provided
by introductory courses on real analysis (excluding integration) and Banach spaces. Section 4 is marked with an asterisk because it assumes a
knowledge of measure theory.

Functions of bounded variation

If < and
= t0 t1 tn =

(1)

then the sequence P = (t0 , t1 , . . . , tn ) is called a partition of the interval


[, ]. When P and Q are partitions of the interval [, ] such that P
is a subsequence of Q, we write P Q and say that Q is a refinement
of P . Now let E be a Banach space and f : [, ] E. If P is the
partition (1) let
Vf (P ) =

n
X

kf (tk ) f (tk1 )k .

(2)

k=1

Although the term kf (tk ) f (tk1 )k makes no contribution to Vf (P )


when tk1 = tk , it is convenient in practice to allow partitions to contain
repeated values. It is easy to deduce that if Q is a refinement of P then
Vf (P ) Vf (Q). The non-negative number
Vf = sup {Vf (P ) : P is a partition of [, ] }

(3)

is called the total variation of f . If Vf < then f is said to have


bounded variation. It is easy to verify that f must be bounded for
f to have bounded variation. Moreover, a function taking values in Rm
has bounded variation if and only if each of its components has bounded
variation and similarly for complex valued functions.
Suppose that f, g : [, ] E have bounded variation and c is a
scalar. Then it is easy to show that
Vf +g Vf + Vg

and Vcf = |c| Vf .


2

(4)

Consequently, the set BV (, , E) of all functions of bounded variation


is a linear subspace of the space of bounded functions on [, ]. We
also define BV (, ) , BV (, , C) to be the set of all complex valued
functions of bounded variation.
Let E, F and G be Banach spaces with a binary operation EF G
such that kxyk kxk kyk for x E and y F. Suppose f BV (, , E)
and g BV (, , F). Then f g : [, ] G where the product is
defined pointwise. Using the triangle inequality, it is straightforward to
show that
Vf g Vf kgk + Vg kf k
(5)
where kf k = sup{kf (t)k : t } as usual.
Example 1 Suppose f : [, ] R is increasing. Then it is easy to
deduce that f has bounded variation with Vf = f () f ().
Example 2 Suppose f : [, ] R is differentiable and f 0 is bounded.
An easy application of the mean value theorem then shows that f has
bounded variation. A similar result is proved for vector valued functions
in Theorem 14 below.
The following results give some insight into the nature of functions of
bounded variation.
Theorem 3 If g BV (, , E) then g has a limit from the left at every
t (, ] and a limit from the right at every t [, ).
Proof. Suppose g does not have a limit from the left at s (, ]. Then
there exists an > 0 such that
kg (t) g (s )k 2

sup
st<s

whenever 0 < s . Putting t1 = 12 ( + s) this allows us to


choose t2 (t1 , s) so that kg (t2 ) g (t1 )k . Similarly we can choose
t3 (t2 , s) so that kg (t3 ) g (t2 )k . Proceeding by induction, we
obtain a sequence t1 < t2 < < s such that kg (tn+1 ) g (tn )k for
all n. This shows that g cannot have bounded variation because
n+1
X

kg (ti ) g (ti1 )k n

i=2

for all positive integers n. A similar argument with respect to limits


from the right completes the proof.
Theorem 4 Let f BV (, , E) and define Jf : [, ] R by
Jf (t) = lim sup {kf (s) f (t)k : s and |s t| } .
0

(a) If > 0 then there exist only finitely many points t such
that Jf (t) .
(b) The map f is continuous at t if and only if Jf (t) = 0.
(c) The set of all discontinuities of f is countable.
The map Jf is called the jump function.
Proof. Suppose there are n points < t2 < t4 < < t2n < such
that Jf (t2k ) for 1 k n. Choose intermediate points
t1 < t2 < < t2n < t2n+1
so that either kf (t2k ) f (t2k1 )k /2 or kf (t2k ) f (t2k+1 )k /2
for each k. Then
Vf

2n+1
X

kf (tk ) f (tk1 )k

k=2

n
2

so that n 2Vf /. This proves (a).


Part (b) is obvious from the definition of the jump function. To
proveS(c), let Dn be the set of all t for which Jf (t) 1/n. Then

D = n=1 Dn is the set of all discontinuities of f . Since D is a countable


union of finite sets, D is countable.
Not all continuous functions have bounded variation. For instance,
the reader can verify that

0
for
t=0
f (t) =
t cos(1/t) for 0 < t 1
has Vf = . Recall that a function f : [, ] E is said to be piecewise continuous if it is continuous everywhere except for a finite number of points tk say where it has a limit f (tk ) from below for
4

tk > and a limit f (tk +) from above for tk < . We say f is piecewise
C 1 if f has a piecewise continuous derivative. For a function defined on
an unbounded interval such as [a, ), we allow only a finite number
of discontinuities in any finite subinterval. We shall see later that any
piecewise C 1 function on a bounded interval has bounded variation.

2.1

The Variation Function

We now elaborate the notation (3) as follows. If [, ] I where I is a


real interval and f : I E then
Vf (, ) , sup {Vf (P ) : P is a partition of [, ] } .
With fixed, we put Vf (t) , Vf (, t) for t . The map Vf is
called the variation function of f .
Theorem 5 Let < < and f : [, ] E. If f BV (, , E) and
f BV (, , E) then f BV (, , E).
Proof. Let P be a partition of [, ]. Insert the point as necessary to
obtain a refinement Q of the form
= s0 sm = = t0 tn = .
Setting Q1 = {s0 , . . . , sm } and Q2 = {t0 , . . . , tn } we have
Vf (P ) Vf (Q)
m
n
X
X
=
kf (sj ) f (sj1 )k +
kf (tk ) f (tk1 )k
j=1

k=1

= Vf (Q1 ) + Vf (Q2 )
Vf (, ) + Vf (, )
as required.
Theorem 6 Let < and f BV (, , E).
(a) If t then Vf (, ) = Vf (, t) + Vf (t, ).
(b) The variation function Vf is increasing.
(c) If f is continuous on the right at [, ) then so is Vf .
5

The formulation and proof of the left sided version of (c) is left to the
reader. Thus, if f is continuous at [, ] then so is Vf .
Proof. To prove (a), let P be a partition of [, t] and Q be a partition
of [t, ]. Then
Vf (P Q) = Vf (P ) + Vf (Q)
which, by considering suprema, shows that Vf (, ) Vf (, t)+Vf (t, )
and Vf (, ) Vf (, t) + Vf (t, ). This establishes (a) and (b) is an
immediate consequence.
The non-trivial part (c) will now be proved. Suppose f is continuous
on the right at . If we can deduce that
lim Vf (, t) = 0

t+

(6)

then it will follow by (a) that Vf is continuous on the right at . To obtain


a contradiction, suppose the limit is non-zero. Since Vf is increasing,
there is a > 0 such that
Vf (, t) >

for all

< t .

Given < s , we can choose a partition = t0 < t1 < < tn = s


of [, s] such that
n
X
kf (tk ) f (tk1 )k > .
(7)
k=1

By continuity of f on the right, it is possible to choose s1 with <


s1 < t1 so that (7) holds with t0 = replaced by s1 . Consequently,
Vf (s1 , s) > . This procedure may be repeated to yield < s2 < s1
such that Vf (s2 , s1 ) > and hence Vf (s2 , s) > 2. After N steps we
obtain Vf (sN , s) > N . This contradicts the hypothesis of bounded
variation, thereby establishing (6) as required.
Lemma 7 Let g : [, ] R be a function of bounded variation. Then
there exist unique real-valued functions and on [, ] such that
g = + g()

and

Vg = + .

(8)

The functions and are increasing and non-negative with


() = 0 = ().
If g is continuous at t [, ] then so are and .
6

(9)

Thus, a real valued function has bounded variation if and only if it can
be expressed as a difference of increasing functions.
Proof. Adding and subtracting the two equations (8) leads to
2 = Vg + g g ()

and

2 = Vg g + g ()

as the only possible choice for and . These functions satisfy (9).
Given s < t , we have by Theorem 6 that
2 (t) 2 (s) = Vg (s, t) + g (t) g (s) 0.
Similar reasoning shows that is increasing. Finally, if g is continuous
at t then so is Vg by Theorem 6 and this makes and continuous at
t by their definitions.

The Riemann-Stieltjes Integral

We now develop the Riemann-Stieltjes integral from first principles. Although the reader may be familiar with the Riemann integral, our development will be complete and self contained.
Given > 0, we define a -partition of [, ] to be a partition
P : = t0 t1 tn =

(10)

such that
|P | , max (tk tk1 ) .
1kn

We call an n-tuple c = (c1 , . . . , cn ) satisfying


tk1 ck tk

for

1kn

(11)

an interstitial vector with respect to P .


Let E, F, and G be Banach spaces over the same base field k = R
or k = C. Suppose given a bilinear operation E F G such that
kxyk kxk kyk

and

(x) y = (xy) = x (y)

(12)

for all k, x E, y F. Specific cases of interest are multiplication of


scalars k k k, scalar multiplication of vectors k E E and the
inner product E E R when E is a real Hilbert space. Now let
f : [, ] E and g : [, ] F.
7

We shall refer to
S (f, g, P, c) ,

n
X

f (ck )[g(tk ) g(tk1 )]

k=1

as a Riemann-Stieltjes sum or more briefly an RS-sum. The simpler


notation S(P, c) will also be used when f and g remain fixed throughout
an argument or discussion. There exists at most one vector L G
satisfying the following condition:
For each > 0 then there exists a > 0 such that
kS (f, g, P, c) Lk <
whenever |P | and c is an interstitial vector.
When the limit exists, we say f is RS-integrable with respect to g and
call L the RS-integral. When there is no ambiguity, we shall say more
briefly that f is integrable with respect to g and call L the integral.
R
R
When it exists, the RS-integral is denoted by f (t)dg(t) or f dg.
Choosing F = R and g(t) = t gives the usual Riemann integral.
Given that f is integrable with respect to g, it is obvious that the
following condition holds:
For each > 0 there exists a > 0 such that
kS (Q, cQ ) S (P, cP )k

(13)

whenever P and Q are -partitions of [, ] and cP , cQ are


corresponding interstitial vectors.
We shall refer to the latter condition as the Cauchy condition for the
existence of the RS-integral. To establish its sufficiency, let Pn be a
partition of [, ] with |Pn | 1/n and cn be an interstitial vector with
respect to Pn for each positive integer n. Then the Cauchy condition
together with the completeness of G implies that Sn = S (Pn , cn ) converges to a limit L G as n . Let > 0 and choose > 0 so
that (13) holds. If P is a partition with |P | and c is an interstitial
vector then
kS (P, c) Lk kS (P, c) Sn k + kSn Lk
and letting n shows that kS (P, c) Lk as required.
The following proposition states some obvious properties of the RSintegral. Lemma 10 below is a partial converse of (b).
8

Proposition 8 Let E, F, and G be Banach spaces, all real or all complex, and E F G be a bilinear operation satisfying (12). Let and
be real numbers with < .
(a) Let g : [, ] F. Then the set RS (g) = RS (, , g) of all RSintegrable functions f : [, ] E with respect to g is closed under
R
linear combinations and the RS-integral f 7 f dg is a linear
operation of RS (g) G.
(b) If < < and f RS (, , g) then f belongs to both RS (, , g)
and RS (, , g), and
Z
Z
Z
f dg =
f dg +
f dg.

(c) Let f : [, ] E. Then the set G of all functions g : [, ] F


such that f RS (g) is closed under linear combinations and the
R
RS-integral g 7 f dg is a linear operation of G G.
(d) If f : [, ] E belongs to RS (g) for some g : [, ] F then
Z





f dg sup kf (t)k Vg .


t
Let f : [, ] E and g : [, ] F. When E = F, we can form
R
RS-sums S(g, f, P, c) and the integral g df as before. Otherwise, if
E 6= F, we define
S (g, f, P, c) ,

n
X

[f (tk ) f (tk1 )]g(ck )

k=1

R
and take g df to the the limit of these sums when it exists. Accordingly, there are two possibilities to consider: f RS (g) and g RS (f ).
The following result deals with this situation.
Theorem 9 Let f : [, ] E and g : [, ] F. Then f RS (g) if
and only if g RS (f ). In either case,
Z
Z
f dg +
g df = f () g () f () g () .

Proof. For any partition P and interstitial vector c we have


n
X

S (f, g, P, c) =

f (ck ) [g(tk ) g(tk1 )]

k=1

= f (cn ) g () f (c1 ) g ()

n1
X

[f (ck+1 ) f (ck )] g(tk )

k=1

= f () g () f () g ()

[f (c1 ) f ()] g ()
+

n1
X

[f (ck+1 ) f (ck )] g(tk )

k=1


+ [f () f (cn )] g ()
= f () g () f () g () S (g, f, Q, d)
where Q = {, c1 , . . . , cn , } and d = (, t1 , . . . , tn1 , ). The proof is
completed by observing that |Q| 2 |P | then employing a routine -
argument.
Lemma 10 Suppose f : [, ] E and g : [, ] F are such that
f RS(, , g) and f RS(, , g) for some < < . If f is
continuous at and g is bounded on a neighborhood of or vice versa
then f RS(, , g) with
Z

f dg =

Z
f dg +

f dg.

Proof. Let f be continuous at and g be bounded on a neighborhood


of . We shall deduce that f RS(, , g). An application of Theorem 9
then gives g RS(, , f ) to complete the proof.
First, take 0 > 0 and M > 0 so that kg(t)k M for |t | 0 .

10

Now let > 0 and choose 0 < 0 such that:




Z


S(P 0 , c0 )

for |P 0 |
f
dg



Z




00 00
f dg
S(P , c )

kf (t) f ()k
M

for

|P 00 |

for

|t |

Let P = (t0 , . . . , tn ) be a -partition of [, ] and c be an interstitial


vector. We have tm1 < tm for some m so that
P 0 = (t0 , . . . , tm1 , )

and P 00 = (, tm , . . . , tn )

are -partitions of [, ] and [, ] respectively. If cm , we put


c0 = (c1 , . . . , cm )

and c00 = (, cm+1 , . . . , cn )

to obtain
S(P, c) = S(P 0 , c0 ) + S(P 00 , c00 ) + [f (cm ) f ()][g(tm ) g()]
and hence



Z
Z




f dg
f dg 3.
S(P, c)

Minor adjustments for cm > complete the proof.


Lemma 11 Let g : [, ] F and f0 RS(g). Suppose f (t) = f0 (t)
for all but finitely many points t [, ] and g is continuous at each t
R
R
where f (t) 6= f0 (t). Then f RS(g) and f dg = f0 dg.
This result is important for practical reasons. For instance, if f (t) is
only defined for < t < and g(t) is continuous at t = and t =
then we consider f to belong to RS(g) provided that f = f0 on (, )
R
R
for some f0 RS(g) and put f dg = f0 dg.
Proof. Once established for a single point of difference, the stated
result follows by a simple induction argument. We assume henceforth
that f (t) = f0 (t) except for t 6= and consider < < , leaving it to
the reader to make minor adjustments for = and = .
11

Given a partition
P : = t0 < t1 < < tn =
we have tk1 < tk for some k whence
S(f, g, P, c) S(f0 , g, P, c) = [f (ck ) f0 (ck )][g(tk ) g(tk1 )]
+[f (ck+1 ) f0 (ck+1 )][g(tk+1 ) g(tk )]
for any interstitial vector c. From this, we get
kS(f, g, P, c) S(f0 , g, P, c)k
kf () f0 ()k {kg(tk ) g(tk1 )k + kg(tk+1 ) g(tk )k}
and the desired result follows by letting |P | 0.
Theorem 12 Let f : [, ] E be a bounded function with only finitely
many discontinuities. Let g : [, ] F have bounded variation and be
continuous at each point where f is discontinuous. Then f is integrable
with respect to g.
In particular, a bounded function on a compact interval with only finitely
many discontinuities is Riemann integrable. This result together with
Theorem 9 shows that BV(, , E) R(, , E), the set of all Riemann
integrable functions [, ] E.
Proof. It suffices by Lemma 10 to prove this result for at most one discontinuity of f occurring at one of the endpoints. We assume henceforth
that f is continuous on [, ) and g is continuous at , leaving it for the
reader to make the adjustments needed for the other case.
Consider a partition P 0 of [, ] and a refinement
P : = t0 < t1 < < tn = .
Then P 0 can be written
P 0 : = ti0 < ti1 < < tim =
with i0 < i1 < < im . Let c0 be an interstitial vector with respect to
P 0 and extend it to an interstitial vector c with respect to P so that
c = (c1 , . . . , cn )

and c0 = (cj1 , . . . , cjm )


12

for some 1 j1 < < jm n. Put


c = (cj1 , . . . , cj1 , cj2 , . . . , cj2 , . . . , cjm )
where the term cjk occurs ik ik1 times. Then
S(P 0 , c0 ) =

m
X

f (cjk )[g(tik ) g(tik1 )] =

n
X

f (ci )[g(ti ) g(ti1 )]

i=1

k=1

so that
S(P, c) S(P 0 , c0 ) =

n
X

[f (ci ) f (ci )][g(ti ) g(ti1 )].

(14)

i=1

Now put M = sup kf k and let > 0. Since g is continuous at , we


can choose 0 < < so that Vg (, ) /M . Since f is uniformly
continuous on [, ], there exists a > 0 such that
kf (t) f (s)k

for s, t

with

|t s| .

If |P 0 | and = tk P then (14) gives


kS(P, c) S(P 0 , c0 )k

k
X

kf (ci ) f (ci )k kg(ti ) g(ti1 )k

i=1

n
X

kf (ci ) f (ci )k kg(ti ) g(ti1 )k

i=k

Vg (, ) + M Vg (, ) (Vg + 1).

(15)

Now let P 0 and P 00 be partitions of [, ] with |P 0 | , |P 00 | < . Let c0 and


c00 be an interstitial vectors with respect to P 0 and P 00 . Then we can
choose P so that P 0 P 00 P and P . By further refinement of P as
necessary, we can choose an interstitial vector c with respect to P which
extends both c0 and c00 . We then obtain via (15)
kS(P 0 , c0 ) S(P 00 , c00 )k
kS(P, c) S(P 0 , c0 )k + kS(P, c) S(P 00 , c00 )k
2(Vg + 1)
as required.
13

Theorem 13 Every Riemann integrable function is bounded.


Proof. Consider a function f : [, ] E and a partition
P : = t0 < t1 < < tn = .
Let c be an interstitial vector and c0 be another one which differs from
c only at one index i. Then


n
X



0
0
kS(P, c ) S(P, c)k = (tk tk1 ) [f (ck ) f (ck )]


k=1

= (ti ti1 ) kf (c0i ) f (ci )k .


When f is unbounded, we can always choose i and c0i so that the last
expression exceeds 1.
Recall that a function f : [, ] E is said to be piecewise continuous if the limit from below (or left limit)
f (t) , lim f (t h)
h0+

exists for each t (, ], the limit from above (or right limit)
f (t+) , lim f (t + h)
h0+

exists for each t [, ), and the two limits are equal at all but finitely
many points t (, ). We observe that f need not be continuous on
the whole of [, ] to have a piecewise continuous derivative. Indeed

0 if t < 0
f (t) =
(16)
1 if t 0
has f 0 (t) = 0 for t 6= 0 but has a discontinuity at t = 0. We say f
is piecewise continuously differentiable (or piecewise C 1 ) if f is
continuous on [, ] and has a piecewise continuous derivative.
For Riemann integrable f : [, ] E we define
Z

f (t) dt ,

f (t)d( + t).

f (t) dt =

14

Fundamental theorem of calculus


(a) Let f : [, ] E be Riemann integrable. Then
Z t
F (t) =
f (s) ds

defines a continuous function F : [, ] E. Furthermore, if f is


continuous at t then F is differentiable at t with F 0 (t) = f (t).
(b) If f : [, ] E is piecewise C 1 then
Z
f 0 (t) dt = f () f ().

The continuity of f is essential to (b). For instance, (16) has


Z 1
f 0 (t) dt = 0 while f (1) f (1) = 1.
1

Proof. From statements (b) and (d) of Proposition 8, we have


Z s



kf k |s t|
kF (s) F (t)k =
f
(
)
d

which shows that F is continuous. Suppose now that f is continuous


at t. Let > 0 and choose > 0 so that
kf (s) f (t)k for

|s t| .

Then we have
Z s




kF (s) F (t) (s t)f (t)k = [f ( ) f (t)]d

t

|s t|
for |s t| . This shows that
F (s) F (t)
f (t)
st

as

st

to complete the proof of (a). Statement (b) is an immediate consequence


when f 0 is continuous on [, ]. The piecewise C 1 case follows by applying Lemma 10.
15

Theorem 14 If g : [, ] F is piecewise C 1 then


Z
Vg =
kg 0 (t)k dt < .

Proof. Using Theorems 5 and 6, it suffices to prove the special case


where g is C 1 on the whole of [, ]. For any partition t0 t1 tn
of [, ] we have


n
n Z ti

X
X


g 0 (t) dt
kg(ti ) g(ti1 )k =


ti1
i=1
i=1
n Z ti
X
kg 0 (t)k dt

i=1

Z
=

ti1

kg 0 (t)k dt.

R
This shows that Vg kg 0 (t)k dt. Now let > 0. By uniform continuity of g 0 , there exists a 0 > 0 such that
kg 0 (s) g 0 (t)k

for

|s t| 0 .

Since kg 0 (t)k is Riemann integrable, we can choose 0 < 0 so that




Z
n
X



0
0
kg (ci )k (ti ti1 )
kg (t)k dt


i=1

whenever t0 t1 tn is a -partition and c is an interstitial


vector. For any such choice we have
Z
n
X
kg 0 (t)k dt +
kg 0 (ci )k (ti ti1 )

i=1


n Z ti

X


=+
g 0 (ci ) dt

ti1

i=1



Z
n Z ti
n ti
X

X




0
0
0
g (t) dt
+
[g (ci ) g (t)] dt +



ti1

ti1
i=1

i=1

+ ( ) +

n
X

kg (ti ) g (ti1 )k

i=1

(1 + ) + Vg .
16

R
This shows that kg 0 (t)k dt Vg to complete the proof.
We now come to a very important result regarding path integrals and
their use in complex analysis and differential geometry.
Theorem 15 Let g : [, ] F be piecewise C 1 and f : [, ] E be
piecewise continuous. Then f RS (g) with
Z

f (t) dg(t) =

f (t)g 0 (t) dt.

Proof. The integrals both exist by Theorems 12 and 14. Using (b) of
Proposition 8, it suffices to establish equality for the special case where
f is continuous and g is C 1 on the whole of [, ]. Let > 0 and choose
> 0 so that
kf (s) f (t)k for |s t| .
Then for any -partition P and interstitial vector c we have


Z




0
f (t)g (t) dt
S(f, g, P, c)

n

Z
X



0
f (ci )[g(ti ) g(ti1 )]
=
f (t)g (t) dt

i=1

Z
Z
n
X

ti



=
f (ci )
g 0 (t) dt
f (t)g 0 (t) dt


ti1

i=1

n Z ti
X



=
[f (ci ) f (t)]g 0 (t) dt


t
i1
i=1
Z
n
X ti

kf (ci ) f (t)k kg 0 (t)k dt Vg


i=1

ti1

as needed to complete the proof.

Riemann-Stieltjes measures *

A function g : R C is said to have bounded variation on R if its


total variation
Vg , sup Vg (, )
(17)
<

17

is finite. It is a routine exercise to verify that the set BV(R) of all such
functions constitutes a vector space with (17) as a seminorm. The total
variation is strictly a seminorm since Vg = 0 whenever g is constant.
Every g BV(R) is bounded since
|g(t)| |g(t) g(0)| + |g(0)| Vg + |g(0)| .
Theorem 6 can be extended as follows. The easy verification is left to
the reader.
Proposition 16 Let g BV(R) and R.
(a) Vg (, t) increases to a finite limit Vg (, ) as t .
(b) Vg (t, ) increases to a finite limit Vg (, ) as t .
(c) Vg (, t) + Vg (t, ) whenever < t < .
(d) The function Vg (t) , Vg (, t) is increasing and bounded.
(e) If g is continuous on the left (or right) at then so is Vg .
Given f Cc (R) and g BV(R), Lemma 12 allows us to define
Z

f dg =
R

f dg

by taking [, ] to contain the support of f . Proposition 8 shows that


this defines a bilinear function Cc (R) BV(R) C satisfying
Z



f dg kf k Vg for f Cc (R), g BV(R).

One can always decompose the integral into a linear combination of no


more that four terms, each of which is an integral like (18), but involves
only real functions.
For each g BV(R), the Riesz representation theorem furnishes a
unique regular Borel measure g such that
Z
Z
f dg =
f dg for f Cc (R).
(18)
R

18

Depending on g, this measure is positive, signed or complex with


|g | = |g | (R) Vg <
and (18) is the restriction to Cc (R) of the integral on L 1 (g ). We call
g the Riemann-Stieltjes measure associated with g.
Let : R R be a bounded increasing function. Then
< () () <
where
() , lim (t) = sup
t

and () , lim (t) = inf .


t

Consequently
V = () () = sup inf < .
It follows immediately from the definition of the RS-integral that gives
a positive linear functional on Cc (R) so that is a positive measure.
Choosing an increasing sequence {fn } in Cc (R) satisfying
[n,n] fn [n1,n+1]

for n Z+

we obtain
Z
(n) (n)

fn dg (n + 1) (n 1)
R

and so the monotone convergence theorem gives


(R) = () () = V .
Lemma 7 can be restated as follows and proved in much the same way
as before.
Proposition 17 Given a real function g BV(R) and R, there
exist unique real-valued functions and such that
g =+

and

Vg = + .

(19)

Moreover, and are both increasing and bounded. If g is continuous


at t R then so are and .
19

This together with the preceding observation shows that a real function
has bounded variation if and only if it can be expressed as the difference
of two increasing functions of bounded variation.
Since a complex valued function of bounded variation decomposes
into a pair of real functions of bounded variation, it follows that any
g BV(R) has limits g() as t and g() as t . By
taking = g() in (19) we obtain the canonical decomposition of
a real function of bounded variation in which and are both nonnegative with () = 0 = ().
The Riesz representation theorem carries these decomposition across
to Riemann-Stieltjes measures. Thus if g = + c where and are
real functions of bounded variation then g = + c . In particular,
if g = with and increasing then g = .
Theorem 18 Let g : R R have bounded variation.
(a) The set of all discontinuities of g is countable.
(b) g () = g(+) g() for R
(c) g (, ) = g() g(+) for , R, <
In particular, if g is continuous at then g () = 0. Using the additivity
of measures, we get
g [, ] = g () + g (, ) + g () = g(+) g()
R
which need not be the same as 1 dg = g() g() unless g is continuous at both and .
Proof. Statement (a) follows easily from (c) of Theorem 4. By the
preceding observations, it suffices to establish (b) and (c) for the case
of monotone increasing g. In this case, the RS-integral is a positive
Rb
Rb
functional so that a dg a dg whenever in RS(a, b, g).
Take {fn } to be a decreasing sequence in Cc (R) satisfying
[1/2n,+1/2n] fn [1/n,+1/n]
Since

R
R

fn dg =

R +1/n
1/n

for n 1.

fn dg we have




 Z




1
1
1
1
g +
g

fn dg g +
g
2n
2n
n
n
R

20

and (b) follows by the monotone convergence theorem. To prove (c) we


choose an increasing sequence {fn } in Cc (R) satisfying
[a+1/n,1/n] fn [a+1/2n,1/2n]
so that




 Z



1
1
1
1
g
fn dg g
g +

g +
n
n
2n
2n
R

then apply the monotone convergence theorem.


Applications like Fourier analysis sometimes involve absolutely integrable functions of bounded variation, i.e. g L 1 (R) BV(R). The
following observation is useful in this regard.
Lemma 19 If g L 1 (R) BV(R) then
lim g(t) = 0 = lim g(t).

Proof. Suppose g BV(R) with C = 12 |g()| > 0. Then it is possible


to choose < so that |g(t)| C for t whence
Z
Z
|g(t)| dt
|g(t)| dt C( )

for . This shows that g is not absolutely integrable and a similar


argument leads to the same result when g() 6= 0.

Path integration

Throughout this section E, F and G are Banach spaces over the same
base field k = R or k = C. A bilinear operation E F G is assumed
with the following properties:
(a) kxyk kxk kyk for all x E, y F
(b) (x)y = (xy) = x(y) for all k, x E, y F

21

5.1

Paths

Let , R with < . A continuous map g : [, ] F is called


a path. The variable t in g(t) is said to be a parameter and the
interval [, ] on which the path is defined is the parameter interval.
The image set {g} , {g (t) : t } is called the trace of g. When
{g} Y for a subset Y of F, we say g is a path in Y . The initial
point of the path is g () and its final point is g (). The path is
closed if g () = g () and open otherwise. We say the path is simple
if g (s) 6= g (t) whenever s < t < or < s < t . A rectifiable
path is a path having bounded variation. We define the length of a
path g to be its total variation and denote it by Lg rather than Vg as in
previous sections. A rectifiable path has finite length by definition.
Because of Corollary 22 below, the most important type of path in
practice is a piecewise C 1 map g : [, ] F. Theorem 14 gives
Z

Lg =

kg 0 (t)k dt <

for any such path.


Lemma 20 Let : [a, b] R be continuous and injective. Then is
strictly monotone and a homeomorphism [a, b] [, ] where
= min{ (a) , (b)}

and

= max{ (a) , (b)}.

Proof. Suppose (a) < (b) and let a s < t b. To obtain a


contradiction, suppose (s) (t). Then (s) > (t) by injectivity
so that (a) < (s) > (t) or (s) > (t) < (b). In either case,
the intermediate value theorem yields two distinct points and such
that () = () contrary to the injectivity of . Therefore is strictly
increasing. If (a) > (b) then is strictly increasing by the foregoing
argument so that is strictly decreasing.
All that remains is to establish the continuity of = 1 . By what
has already been proved, and are either both strictly increasing or
both strictly decreasing. Assume is increasing. Let < and
put c = () so that a < c b. Let > 0 with c a and take
= (c) (c ) > 0. Then the monotonicity of gives
c < (t) c

for < t

22

from which it follows that is continuous on the left. Continuity on the


right is established by similar reasoning. Finally, if is decreasing then
is increasing so that (t) = ()1 (t) is continuous.
Lemma 21 Let g : [, ] F be a rectifiable path and be a homeomorphism of [a, b] onto [, ] for some a < b. Then:
(a) {g } = {g}
(b) g is a rectifiable path with Lg = Lg
(c) If f RS (g) then f RS (g ) with
Z b
Z
f ( (t)) dg ( (t)) = sgn()
a

f (t) dg (t)

where sgn() = 1 if is increasing and sgn() = 1 if is decreasing.


Proof. Statement (a) is obvious. Applying Lemma 20, we shall assume
is strictly increasing, leaving the reader to modify our argument for
the case of decreasing .
If a = t0 tn = b is a partition of [a, b] then
= (t0 ) (tn ) =
is a partition of [, ] so that
n
X

kg ( (ti )) g ( (ti1 ))k Lg .

i=1

This gives Lg Lg and the opposite inequality follows immediately


by writing g = (g ) 1 .
To prove (c), let > 0 and choose > 0 so that


Z




f dg
(20)
S (f, g, P, c)

whenever P is an -partition of [, ] and c is an interstitial vector. By


uniform continuity of , there exists a > 0 such that
| (s) (t)|
23

for

|s t| .

(21)

Now let a = t0 tn = b be a -partition of [a, b] and s = {si }


be an interstitial vector. Then P = {(ti )} is an -partition of [, ]
by (21) and c = {(si )} is an interstitial vector so that (20) gives
n

Z
X



f ((si ))[g((ti )) g((ti1 ))]
f dg .


i=1

This establishes (c) to complete the proof.


Lemma 21 shows that Riemann-Stieltjes integrals are invariant with
respect to a change of parameter t = (s) by an increasing homeomorphism. On the other hand, when is decreasing, the change of parameter
t = (s) simply changes the sign of the integral. If g : I F is a rectifiable path and : J I is a homeomorphism of compact intervals
then the path g is called a reparameterization of g. The reparameterization is said to be orientation preserving or have the same
orientation as g if is increasing. If is decreasing, we say g has
the opposite orientation to g. As in the lemma, we define

1 if is increasing
sgn () =
.
1 if is decreasing
Let g1 : [1 , 1 ] F and g2 : [2 , 2 ] F be rectifiable paths satisfying
g1 (1 ) = g2 (2 ). Put = 1 and = 1 + 2 2 . Then

g1 (t)
if t 1
g (t) =
g2 (t 1 + 2 ) if 1 t
defines a rectifiable path g : [, ] F such that {g} = {g1 } {g2 } and
Lg = Lg1 + Lg2 . This construction can be extended to any finite number
of rectifiable paths g1 , . . . , gn such that the final point of gi is the same
as the initial point of gi+1 for 1 i n 1. The resulting path is called
the concatenation of g1 , . . . , gn and denoted by (g1 , . . . , gn ).

5.2

Integration along Paths

Let g : [, ] F be a rectifiable path and f : {g} E be continuous.


Then f (g) = f g is continuous and hence integrable with respect to g
by Theorem 12. The integral
Z
Z
Z
f = f (y) dy ,
f (g (t)) dg (t)
(22)
g

24

is called the path integral of f along g and satisfies



Z


f (y) dy Lg kf k
g

(23)

by Proposition 8 where
kf kg , sup kf (y)k = sup kf (g(t))k .
t

y{g}

As an immediate corollary to Theorem 15, we have:


Corollary 22 Let g : [, ] F be piecewise C 1 and f : {g} E be
continuous. Then
Z
Z
f (y) dy =
f (g (t)) g 0 (t) dt
g

and

Z
Z


f (y) dy kf k
g

kg 0 (t)k dt.

Theorem 23 Let g : [, ] F be a rectifiable path and fn : {g}


E be continuous for each n. If the sequence {fn } converges uniformly
on {g} then its pointwise limit is a continuous function f : {g} E and
Z
Z
fn f as n .
g

Proof. As the image of a compact interval under a continuous function,


the trace {g} is a compact subset of F. The sequence {fn } therefore
converges uniformly to a continuous limit by a standard topological argument. From (23) we have
Z
Z


fn f Lg kfn f k

in which the RHS vanishes as n .


This result allows us to interchange sums with integrals as follows.
Corollary 24 Let g : [, ] F be a rectifiable path
Pand fn : {g} E
be a continuous function for each P
n. If the series n=1 fn (y) converges

uniformly over y {g} then f = n=1 f is continuous and


(
)
Z X

Z
X
fn =
fn .
g

n=1

n=1

25

5.3

Chains

In complex analysis and differential geometry, one often needs to integrate a function not just along a single path, but over a finite collection
of paths. One way to do this is by concatenating two or more paths
as explained above. If g = (g1 , . . . , gn ) is a concatenation of rectifiable
paths in F and f : {g} E is continuous then we have
Z
n Z
X
f (y) dy
f (y) dy =
g

i=1

gi

by Proposition 8. We now consider a more general approach which does


not require explicit reparameterization of individual paths.
Let K be a compact subset of F and G be the collection of all rectifiable paths lying in K. If g G , then a linear map g : C (K, E) G
is given by
Z
g (f ) = f (y) dy.
g

This map satisfies


kg (f )k Lg kf kK
where kf kK is the sup-norm of f over K. Since C (K, E) forms a Banach
space with respect to the sup norm, each g is a bounded linear operator
with kg k Lg . As explained in Appendix A, the map g 7 g extends
uniquely to a homomorphism Z hG i L where Z hG i is the free Abelian
group on G and L is the Banach space consisting of all linear maps
C (K, E) G. The elements of Z hG i are formal linear combinations
k1 g1 + + kn gn with ki Z and gi G . We call one of these objects
a chain or a cycle when each gi is a closed path. The homomorphic
extension of g 7 g gives
Z
n
n
X
X
(k1 g1 + + kn gn ) (f ) =
ki (gi ) (f ) =
ki
f dgi
(24)
i=1

i=1

gi

for ki Z and gi G . Given a rectifiable path g : [, ] K and


a homeomorphism : [a, b] [, ], we have (g ) = (g) where
= 1 if preserves orientation and = 1
R otherwise. The whole
point of chains is to allow for expressions like g1 g2 f which should be
interpreted as
Z
Z
Z
f=
f
f.
g1 g2

g1

26

g2

This is useful as an algebraic tool because the paths which arise in practice are typically concatenations of lines and arcs each parameterized in
one direction or the other.

Free groups

Let G be an Abelian group. We write the group operation additively,


denote the identity element by 0, and the inverse of g by g. If n is
a positive integer then ng denotes the n-fold sum g + + g. We also
define 0g , 0 and (n)g , (ng) so that
(n)g = (ng) = n(g)

for n Z, g G.

These definitions ensure that


(m + n)g = mg + ng

and n(g + h) = ng + nh

whenever m, n Z and g, h G. The set nG , {ng : g G} is a


subgroup of G for each n Z as is hgi , {ng : n Z} for each g G.
Given a non-empty set S, let Z hSi be the set of all maps k : S Z
such that k(x) = 0 for all but finitely many x. Then Z hSi forms an
Abelian group with respect to pointwise addition
(k + h)(x) = k(x) + h(x).
The zero element is 0(x) = 0 and the inverse of k is (k)(x) = k(x).
This group is called the free Abelian group on S. For a finite set
S = {x1 , . . . , xn }, the free Abelian group on S may be written as
Z hx1 , . . . , xn i , Z hSi .
Each x S is uniquely associated with the map x Z hSi given by
x (t) = 1 if t = x and x (t) = 0 otherwise. We can thereby identify S
with the subset {x : x S} of Z hSi. Each non-zero element of Z hSi
can thus be written uniquely as a finite sum k1 x1 + + kn xn where
0 6= ki Z and xi are distinct elements of S. This the reason for the
notation Z hSi which indicates that S is a set of generators of Z hSi.
The following result from group
L theory provides the underlying formalism for chains. The symbol
xS hxi in (b) stands for the direct
sum of the subgroups hxi.
27

Proposition 25 Let S be a non-empty set.


(a) If G is an Abelian group and f : S G then there exists a unique
extension of f to a homomorphism Z hSi G.
L
(b) Z hSi u xS hxi
(c) If S is a finite set {x1 , . . . , xn } then
(k1 , . . . , kn ) 7 k1 x1 + + kn xn
is an isomorphism of Zn onto Z hx1 , . . . , xn i.
In path integration theory, we are given a bilinear operation E F G
between Banach spaces. Given a compact subset K of F, we take S to
be the set of all rectifiable paths in K and G to be the Banach space
of all bounded linear maps C (K, E) R G. For each g S, a bounded
linear map g G is given by g (f ) = g f dg. Statement (a) gives
(k1 g1 + + kn gn ) (f ) =

n
X

ki (gi ) (f ) =

i=1

n
X
i=1

Z
ki

f dgi
gi

for k1 , . . . , kn Z and g1 , . . . , gn S. A linear combination


k1 g1 + + kn gn Z hSi
is called a chain. Given a rectifiable path g : [, ] K and a homeomorphism : [a, b] [, ], we have (g ) = (g) where = 1 if
is orientation preserving and = 1 otherwise.

28

También podría gustarte