NECESSARY AND SUFFICIENT CONDITIONS FOR

RIEMANN AND RIEMANN-STIELTJES INTEGRABILITY

Looking back over the notes already posted on this and related subjects, I felt that their organization
did not emphasize the things that are important on the subject of the caption. Hence these notes, which
will initially reprise1 things that are already treated in the others, but with different emphasis. The material
on almost-everywhere continuity as a necessary and sufficient condition for integrability will be new, in the
sense that both my notes and the textbook rely on Lebesgue(-Stieltjes) integration to establish this condition,
while these notes will give a completely elementary treatment.

1. Integrability Characterized by a Limit Condition.

Z b
Theorem: In order that f(t) dg(t) exist, it is sufficient that V [g; a, b] < ∞ and that
a

 
k(Γ)
X 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
(Γ,⊇)  
j=1

when the integral is taken in the sense of (I), or

 
k(Γ)
X 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
mesh(Γ)→0  
j=1

when the integral is taken in the sense of (II).

Proof. Let Γ : a = t0 < · · · < tk = b and ∆ : a = s0 < · · · < sn = b be partitions of [a, b], with sample
points {τj } and {σi} respectively. If ∆ ⊇ Γ then for each tj there is a unique i(j) for which si(j) = tj . If we
break up the Riemann-Stieltjes sum R[f, g; ∆, σ] into subsums for i(j − 1) + 1 < i ≤ i(j), we find

X
i(j)
X
i(j)
f(τj )[g(tj ) − g(tj−1 )] − f(σi )[g(si ) − g(si−1 )] = [f(τj ) − f(σi )][g(si ) − g(si−1 )]
i=i(j−1)+1 i=i(j−1)+1

and since τj and all the σi’s with i(j − 1) + i < i ≤ i(j) belong to [tj−1 , tj ], we see that

X
i(j)
X
i(j)
|f(τj )[g(tj ) − g(tj−1 )] − f(σi )[g(si ) − g(si−1 )]| ≤ |f(τj ) − f(σi )| · |g(si ) − g(si−1 )|
i=i(j−1)+1 i=i(j−1)+1

X
i(j)
≤ ωj (f; Γ) · |g(si ) − g(si−1 )| ≤ ωj (f; Γ) · V [g; tj−1, tj ] .
i=i(j−1)+1

Adding these up over j = 1, . . . , k and applying the triangle inequality gives

X
k
|R[f, g; Γ, τ ] − R[f, g; ∆, σ]| ≤ ωj (f; Γ) · V [g; tj−1, tj ] . ($)
j=1

1
“Reprise,” indeed! Shamelessly copy from the earlier set of notes, via the wonders of modern text-editing technology, that’s what
“reprise” is. Alas, this is just the technology that led to garbling of an earlier set of notes.

1 Math 501:01 Fall 1999

 (1) (1) (2) (2)
Given two partitions Γ1 = {tj }kj=1 and Γ2 = {tj }kj=1 of [a, b], let ∆ = Γ1 ∪ Γ2 . Then given choices of
sample points τ1 and τ2 for Γ1 and Γ2 respectively, and any choice of sample points σj for ∆, aplying ($)
twice gives

|R[f, g; Γ1, τ1 ] − R[f, g; Γ2 , τ2 ]| ≤ |R[f, g; Γ1, τ1 ] − R[f; g; ∆, σ]| + |R[f, g; ∆, σ] − R[f, g; Γ2, τ2 ]
X X
(1) (2)
k k
(1) (1) (2) (2)
≤ ωj (f; Γ1 ) · V [g; tj−1, tj ] + ωj (f; Γ2 ) · V [g; tj−1 , tj ] .
j=1 j=1

It is now evident that each of the conditions of the theorem will imply |R[f, g; Γ1, τ1 ] − R[f, g; Γ2 , τ2 ]| < 
Pk(Γ)
provided that both Γ1 and Γ2 contain a partition Θ for which Γ ⊇ Θ implies j=1 ωj (f; Γ) · V [g; tj−1 , tj ] <
Pk(Γ)
/2 (for integrals in sense (I)), and similarly if δ > 0 is so small that mesh(Γ) < δ ⇒ j=1 ωj (f; Γ) ·
V [g; tj−1 , tj ] < /2 (for integrals in sense (II)), respectively. These conditions thus imply that the filter base
of R(f, g; δ)’s is Cauchy, thus convergent, or that the filter base of R(f, g; δ)’s is Cauchy, thus convergent,
and that guarantees the existence of the integrals in sense (I) or (II), respectively, that is asserted by the
theorem.
From this point of view the standard existence theorem (Wheeden & Zygmund, (2.24), p. 27) appears
as a
Z b
Corollary: If f is continuous on [a, b] and V [g; a, b] < ∞, then f(t) dg(t) exists in the sense of (II)
a
(and thus a fortiori in the sense of (I)).
Indeed, from the estimate

X
k(Γ)
X
k
ωj (f; Γ) · V [g; tj−1, tj ] ≤ max{ωj (f; ∆) : 1 ≤ j ≤ k} · V [g; tj−1, tj ]
j=1 j=1

= max{ωj (f; Γ) : 1 ≤ j ≤ k} · V [g; a, b]

and the uniform continuity of f on [a, b], it is clear that the r. h. s. of this inequality can be made arbitrarily
small simply by taking the mesh of Γ suffciently small, since that will make all the ωj (f; Γ)’s arbitrarily
small in view of the the uniform continuity of f.
The sufficiency theorem just proved is quite robust; the proof works for vector-valued integrators and
integrands. But the theorem is hard to run backward;2 the converse theorem is more delicate, works only
with real-valued functions, and gives a result that is really satisfactory only for definition (I). However, the
converse does make it possible, in the case of real-valued integrands and integrators, to reduce the study of
integrators of bounded variation to that of monotone integrators, and thus to enable some order-theoretic
arguments that could not be formulated in general.
Z b
Theorem: In order that f(t) dg(t) exist for given real-valued f(t) and given real-valued g(t) with
a
V [g; a, b] < ∞, it is necessary that
 
Xk 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
(Γ,⊇)  
j=1

for integrals in the sense of (I); or, when g(t) is continuous,

 
X k 
lim ωj (f; Γ) · V [g; tj−1 , tj ] = 0
mesh(Γ)→0  
j=1

2
Sed revocare gradum superasque evadere ad auras/ Hoc opus, hic labor est. Aen. vi. 126

2 Math 501:01 Fall 1999

for integrals in the sense of (II).
Proof. Let Γ : a = t0 < · · · < tk = b be a partition of [a, b], and let  > 0 be given. It is always possible
to make “bad choices” {σj } and {τj } of sample points for Γ. Namely, one can always3 choose σj and τj so
that (
sgn[g(tj ) − g(tj−1 )]
sgn[f(σj ) − f(τj )] =
{possibly 0 if f is constant on the cell}
|f(σj ) − f(τj )| > ωj (f; Γ) −  .
Then putting the two relations together gives

[f(σj ) − f(τj )][g(tj ) − g(tj−1 )] ≥ ωj (f; Γ) |g(tj ) − g(tj−1 )| −  · |g(tj ) − g(tj−1 )|

and summing these gives, using V [g; tj−1, tj ] to overestimate |g(tj ) − g(tj−1 )|, the relation
 
X
k Xk
R[f, g; Γ, σ] − R[f, g; Γ, τ ] ≥ ωj (f; Γ) · |g(tj ) − g(tj−1 )| −   V [g; tj−1, tj ]
j=1 j=1

X
k
ωj (f; Γ) · |g(tj ) − g(tj−1 )| ≤ |R[f, g; Γ, σ] − R[f, g; Γ, τ ]| +  · V [g; a, b] .
j=1

We would love to replace the |g(tj ) − g(tj−1 )| on the l. h. s. of the last displayed inequality by V [g; tj−1, tj ].
To do this we need another estimate, whose proof fortunately turns out to be obvious enough to leave to the
reader:
 
X k
  X
k
ωj (f; Γ) · V [g; tj−1, tj ] − |g(tj ) − g(tj−1 )| ≤ 2 · kfk∞ · V [g; a, b] − |g(tj ) − g(tj−1 )| .
j=1 j=1

If we add these estimates, we get

X
k
ωj (f; Γ) · V [g; tj−1, tj ] ≤
j=1
 
X
k
2 · kfk∞ · V [g; a, b] − |g(tj ) − g(tj−1 )| + |R[f, g; Γ, σ] − R[f, g; Γ, τ ]| +  · V [g; a, b] . (#)
j=1

If the integrals in the respective senses exist, then as Γ gets fine (for integrals in sense (I)) or as mesh(Γ) → 0
(for integrals in sense (II), assuming that the integrator g is continuous), we see that all three terms on the
r. h. s. of the estimate (#) can be made arbitrarily small: bounded variation (and continuity of g if we’re
dealing with integrals in sense (II)) take(s) care of the first, the Cauchy condition on the Riemann-Stieltjes
sums takes care of the second, and the factor  in the last term was in our hands when we started (V [g; a, b]
is a constant). This proves the converse theorem.
As a corollary of this result, we can now see that in many cases we can reduce our considerations to the
case of monotone (increasing) integrators.
Z b Z b
Corollary: If V [g; a, b] < ∞ and f dg exists in sense (I), then so does f(t)d(V [g; a, t]), and
a a
conversely. The same assertion is true for integrals in sense (II) if the integrator g is continuous.
3
One can always interchange the choices so that the sign comes out right—something that would not be possible for vector-valued,
or even complex-valued functions; this converse theorem is thus pretty much restricted to the scalar-valued situation.

3 Math 501:01 Fall 1999

 Proof. If we temporarily denote the function t 7→ V [g; a, t] by v, then since it is a monotone increasing
function for which v(a) = 0, its total variation is the same as itself: V [v; a, t] = v(t). But the interval-
additivity of the total variation implies that for any a ≤ tj−1 < tj ≤ b, the relation

V [v; tj−1 , tj ] = v(tj ) − v(tj−1 ) = V [g; a, tj ] − V [g; a, tj−1] = V [g; tj−1 , tj ]

holds. It follows that the conditions

 
k(Γ)
X 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
(Γ,⊇)  
j=1

when the integral is taken in the sense of (I), or

 
k(Γ)
X 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
mesh(Γ)→0  
j=1

when the integral is taken in the sense of (II), are exactly the same for the integrator g and for the integrator
v. Thus in the situations in which these conditions are logically equivalent to the existence of the integral—
with no additional restrictions in the case of integrals in sense (I), and with the additional assumption of
continuity of the integrator in the case of integrals in sense (II)—the integrability of f with integrator g is
logically equivalent to its integrability with integrator t 7→ V [g; a, t].
The importance of this theorem and corollary is that they make it possible to reduce the study of Stieltjes
integrability against integrators of bounded variation to the case where the integrator is monotone. It is
obvious that we can get sufficient conditions for Stieltjes integrability of f for an integrator g of bounded
variation by getting sufficient conditions for monotone integrators—that follows from the linearity of the
integral in “g” and the fact that g is a function of bounded variation if and only if it is a difference of
monotone functions. But the theorem just proved tells us that if g(t) is of bounded variation and f is
integrable dg in sense (I), then f is also integrable dV [g; a, t] in sense (I); consequently, necessary conditions
for integrability w. r. t. monotone integrators will also be necessary for integrators of bounded variation,
with no additional hypotheses in the case of sense (I), but with the additional hypothesis of continuity of g
if limits are taken in sense (II).4
Note, incidentally, that the relations

1 1
P [g; a, t] = {V [g; a, t] + [g(t) − g(a)]} and N [g; a, t] = {V [g; a, t] − [g(t) − g(a)]}
2 2

tell us also that f is g(t)-integrable if and only if it is P [g; a, t]- and N [g; a, t]-integrable.
We should not take leave of the sufficient conditions for integrability
 
X k 
lim ωj (f; Γ) · V [g; tj−1 , tj ] = 0
mesh(Γ)→0  
j=1

for integrals in the sense of (II), or

 
Xk 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
(Γ,⊇)  
j=1

4
In some sense, it is this result—in which the continuity of the integrator does not have to be assumed—that justifies the definition
of the “generalized Stieltjes integral” of sense (I).

4 Math 501:01 Fall 1999

for integrals in the sense of (I), without pointing out that they are inherited by products and uniform limits.
In fact, for bounded f and h we have for any t and u in the interval of integration

|f(t)h(t) − f(u)h(u)| = |f(t)| · |h(t) − h(u)| + |h(u)| · |f(t) − f(u)|

≤ kfk∞ · |h(t) − h(u)| + khk∞ · |f(t) − f(u)|

so for any partition Γ we have

ωj (f · h; Γ) ≤ kfk∞ · ωj (h; Γ) + khk∞ · ωj (f; Γ)

and therefore
   
X
k Xk  Xk 
ωj (f · h; Γ) · V [g; tj−1, tj ] ≤ khk∞ · ωj (f; Γ) · V [g; tj−1, tj ] + kfk∞ · ωj (h; Γ) · V [g; tj−1 , tj ]
   
j=1 j=1 j=1

so if the limits on the r. h. s. are zero (in the sense appropriate to integrals in sense (I) or (II) respectively)
then so is the one on the r. h. s., which is a sufficient condition for integrability of f · h. Similarly, if for every
 > 0 one can approximate h uniformly on [a, b] by an f for which one of those limit conditions is satisfied,
then for every  > 0 we shall have

|h(t) − h(u)| ≤ |h(t) − f(t)| + |f(t) − f(u)| + |f(u) − h(u)| ≤ 2 ·  + |f(t) − f(u)|
ωj (h; Γ) ≤ 2 ·  + ωj (f; u)
X
k X
k
ωj (h; Γ) · V [g; tj−1, tj ] ≤ 2 ·  · V [g; a, b] + ωj (f; Γ) · V [g; tj−1, tj ]
j=1 j=1

so the same limit condition is satisfied for h (and h will be integrable dg). It is somewhat less than satisfactory
to have this sufficient condition for integrability inherited unless one knows it is a necessary condition,
however, and that brings us right back either to using the more general sense (I) definition of integrability
or—if we need the filters with countable bases provided by integrals in sense (II)—to being restricted to
continuous integrators. {However, this sufficient condition makes some sense in vector settings.}
It is an easy exercise for the reader to show that these sufficient conditions for integrability are also
inherited by |f|. Thus, because of the relations

1 
f ∨h= (f + h) + |f − h|
2
1 
f ∧h= (f + h) − |f − h| ,
2
the functions satisfying them form a vector lattice of functions (a vector space of functions closed under
taking pointwise suprema and infima of finite subsets).

2. Integrability Characterized by a Continuity Condition.

For this material one needs only the very beginnings of Lebesgue measure theory on the real line. Having
defined the length of an open interval U = (c, d) as m(U ) = d − c as any sensible person would, one needs
to know that
(1) A null set, or set of (Lebesgue) measure zero, is a set A ⊆ R with the property that for anyS >0
therePexists a countable (finite or countably infinite) family of open intervals {Uj } for which A ⊆ j Uj
and j m(Uj ) < . (The intervals may overlap, and in fact they usually must overlap.) It is clear that
the empty set is such a set, and that any subset of a set of measure zero is also a set of measure zero.
S
(2) Less obvious but true nonetheless: if {Ak } is a countable family of sets of measure zero, then A = k Ak
is a set of measure zero. The reason is that for given  > 0 one can find for each k a family of open

5 Math 501:01 Fall 1999

 S P
 S P
intervals {Uk,j } for which Ak ⊆ j Uk,j and j
k
; then A ⊆ k,j Uk,j and k,j m(Uk,j ) < 
m(Uk,j ) <
2
(the order of summation
S being unimportant in a series of nonnegative terms). Contrapositively stated:
if a countable union k Ak ⊆ R is not a set of measure zero, then at least one of the Ak ’s must fail to
be a set of measure zero.
Theorem: Let f be a bounded real-valued function on the interval [a, b]. If the set of discontinuities
of f has Lebesgue measure zero, then f is Riemann-integrable in sense (II) on the interval [a, b], i.e., the
Z b
integral f(t) dt exists in sense (II).
a
Proof. Consider the oscillation function of f:
ω(f; t) = inf{sup{|f(s) − f(u)| : s, u ∈ B(t; δ) ∩ [a, b]} : δ > 0} .
It is easy to see that every set F (η) = {t ∈ [a, b] : ω(f, t) ≥ η} is closed.5 Indeed, if t0 is a limit point of F (η)
then every open ball B(t0 , ρ) contains points t for which sup{|f(s) − f(u)| : s, u ∈ B(t; δ) ∩ [a, b]} ≥ η no
matter how small δ > 0 may be. If we take δ < ρ − |t − t0 |, then we see that there are points s, t ∈ B(t; δ) ⊆
B(t0 , ρ) for which |f(s) − f(u)| > η −  for every 0 <  < η, so inf{sup{|f(s) − f(u)| : s, u ∈ B(t0 ; ρ) ∩ [a, b]} :
ρ > 0} must be at least η. This puts t0 ∈ F (η).
The set of points of discontinuity of f is given by
∞
[ 1
{x ∈ [a, b] : ω(f; t) > 0} = {x ∈ [a, b] : ω(f; t) ≥ }.
k
k=1

Since the union has measure zero by hypothesis, each of the (closed and bounded, therefore compact) uniands
F (1/k) has measure zero, and so every set F (η) has measure zero (the sets get bigger as η gets smaller).
Now
P let  > 0 be given. Since m(F S ()) = 0, one can find a countable family of open intervals {Ui }
with i m(Ui ) <  for which F () ⊆ i Ui , and since F () is compact one can take the family {Ui }m i=1 to
be finite, discarding unneeded sets if necessary: the remaining Ui ’s cover F (). On the other hand, the set
[a, b] \ F () is relatively open, and so each point x ∈ [a, b] \ F () is contained in an open set Vx for which
sup{|f(s) − f(u)| : s, u ∈ Vx ∩ [a, b]} < . The family {Ui }m i=1 ∪ {Vx : x ∈ [a, b], x ∈
/ F ()} is then an
open cover of [a, b]. Let δ > 0 be a Lebesgue number of this cover, so that every set of diameter < δ is
contained in some single element of this cover. Suppose Γ : a = t0 < · · · < tk = b is a partition of [a, b] with
mesh(Γ) < δ. Then each interval [tj−1 P, tj ] of the partition is contained in some single Ui 6or Vx . For each
fixed indexP i, the sum of the lengths {(tj − tj−1 ) : [tj−1, tj ] ⊆ Ui } cannot exceed m(Ui ). It follows that
the sum {ω(f; [tj−1 , tj ]) · (tj − tj−1 ) : [tj−1, tj ] ⊆ Ui } cannot exceed 2 · kfk∞ · m(Ui ), and therefore that
the sum
( )
X [ X
ω(f; [tj−1 , tj ]) · (tj − tj−1 ) : [tj−1, tj ] ⊆ Ui ≤ 2 · kfk∞ · m(Ui ) ≤ 2 · kfk∞ ·  ,
i i

even allowing for the fact that some intervals [tj−1 , tj ] may be contained in more than one Ui . On the other
hand, if an interval [tj−1 , tj ] ⊆ Vx , then ω(f; [tj−1 , tj ]) <  and therefore

X X
k(Γ)
{ω(f; [tj−1 , tj ]) · (tj − tj−1 ) : ∃Vx ⊇ [tj−1 , tj ]} ≤  · (tj − tj−1 ) = (b − a) ·  .
j=1

Therefore, even allowing for the possibility that some intervals may be counted more than once in the first
sum and that some may be included in both sums, we have the estimate

X
k(Γ)
ωj (f; Γ) · (tj − tj−1 ) < 2 · kfk∞ ·  + (b − a) ·  = [2 · kfk∞ + (b − a)] ·  .
j=1

5
This is equivalent to the observation that ω(f;t) is an upper-semicontinuous function of t∈[a,b].
6
This is an obvious statement, but requires a bit of proof; the reader should try to provide it.

6 Math 501:01 Fall 1999

Since the r. h. s. of this inequality is a constant multiple of  and the only thing assumed about the partition
was that mesh(Γ) < δ, we have shown that
 
k(Γ)
X 
lim ωj (f; Γ) · V [id[a,b] ; tj−1, tj ] =0
mesh(Γ)→0  
j=1

Z b
which is sufficient for the existence of the Riemann integral f(t) dt (whose integrator is the identity
a
function id[a,b] , also known as “g(t) = t”).
Now for the converse.
Theorem: Let f be a bounded real-valued function on the interval [a, b]. If the set of discontinuities of
f does not have Lebesgue measure zero, then f is not Riemann-integrable in sense (I) on the interval [a, b],
Z b
i.e., the integral f(t) dt fails to exist even in (the more general) sense (I).
a

Proof. Again, the set of points of discontinuity of f is given by

∞
[ 1
{x ∈ [a, b] : ω(f; t) > 0} = {x ∈ [a, b] : ω(f; t) ≥ },
k
k=1

and since the union by hypothesis does not have measure zero, some among the (closed and bounded,
therefore compact) uniands F (1/k) must fail to have measure zero. There must therefore be a set F (η) S and
a number
P  > 0 with the property that if {U i } is a countable family of open intervals for which F (η) ⊆ i Ui ,
then i m(Ui ) ≥ . Let Γ : a = t0 < · · · < tk = b be a partition of [a, b]; then for any δ > 0, the open
cells {(tj−1, tj )}kj=1 together with the δ-balls {B(tj ; δ)}kj=0 cover [a, b] and a fortiori cover F (η). The sum
of the lengths of the open cells {(tj−1 , tj )}kj=1 that have nonempty intersection with F (η), together with the
lengths of the δ-balls centered on the points of Γ, must therefore be ≥ :
X
{(tj − tj−1 ) : F (η) ∩ (tj−1 , tj ) 6= ∅} + 2(k + 1)δ ≥  .

Since k is determined by the choice of Γ and δ > 0 is in our hands,

X
{(tj − tj−1 ) : F (η) ∩ (tj−1 , tj ) 6= ∅} ≥  .

If F (η) ∩ (tj−1 , tj ) 6= ∅ then there exists some ball centered on a point of F (η) that is contained in (tj−1 , tj ),
and therefore ω(f; [tj−1 , tj ]) ≥ η. It follows that
X
{ω(f; [tj−1 , tj ]) · (tj − tj−1 ) : F (η) ∩ (tj−1 , tj ) 6= ∅} ≥ η · 

and so certainly
X
k
ω(f; [tj−1 , tj ]) · (tj − tj−1 ) ≥ η · 
j=1

holds for every partition of [a, b]. The necessary condition

 
k(Γ)
X 
lim ωj (f; Γ) · (tj − tj−1 ) = 0
(Γ,⊇)  
j=1

for the Riemann-integrability of f (in sense (I)) can thus never be satisfied: all such sums will be ≥ η · .

7 Math 501:01 Fall 1999

 Corollary: The following three conditions on bounded f : [a, b] → R are logically equivalent:
Z b
(1) f(t) dt exists in sense (I);
a
(2) The set of points of discontinuity of f has Lebesgue measure zero;
Z b
(3) f(t) dt exists in sense (II).
a
Proof. We have just shown that (1) ⇒ (2), and the immediately preceding theorem showed (2) ⇒ (3).
The implication (3) ⇒ (1) was known as soon as the filter bases of Riemann sums corresponding to the two
senses in which the integral exists were defined.
We could try to adapt these ideas to Riemann-Stieltjes integration. The sufficient condition for integra-
bility adapts quite easily. Let g : [a, b] → R be an integrator with V [g; a, b] < ∞. We can make the following
tentative definitions:
(1) A g-null set is a set A ⊆ R with the property that for any  > 0 S there exists a countable
P (finite or
countably infinite) family of open intervals {(aj , bj )} for which A ⊆ j (aj , bj ) and j V [g; aj , bj ] < ,
where g has been extended to be defined on all of R by setting it equal to g(a) to the left of a and equal
to g(b) to the right of b. (Again the intervals may overlap.) It is clear that the empty set is such a set,
and that any subset of a g-null set is also g-null.
S
(2) If {Ak } is a countable family of g-null sets, then A = k Ak is also g-null. Again the reason is
that for given  > 0 one can find for each k a family of open intervals {(ak,j , bk,j )} for which Ak ⊆
S P  S P
j (ak,j , bk,j ) and j V [g; ak,j , bk,j ] < k ; then A ⊆ k,j (ak,j , bk,j ) and k,j V [g; ak,j , bk,j ] <  (the
2
order
S of summation being unimportant in a series of nonnegative terms). So if a countable union
k kA ⊆ R is not g-null, then at least one of the A k ’s must fail to be g-null.
Proposition: Let f be a bounded real-valued function on the interval [a, b] and g be a real-valued
function of bounded variation on the same interval. If the set of discontinuities of f is g-null, then f is
Z b
Riemann-Stieltjes integrable dg in sense (II) on the interval [a, b], i.e., the integral f(t) dg(t) exists in
a
sense (II).
Proof. Again the set of points of discontinuity of f is given by
∞
[ 1
{x ∈ [a, b] : ω(f; t) > 0} = {x ∈ [a, b] : ω(f; t) ≥ },
k
k=1

and so every set F (η) must be g-null. Let  > 0 be given; Pagain we can find a finite family of open intervals
m
{(ai , bi)}m
i=1 whose union contains F () and for which i=1 V [g; ai, bi] < . Again the set [a, b] \ F () is
relatively open, and so each point x ∈ [a, b] \ F () is contained in an open set Vx for which sup{|f(s) − f(u)| :
s, u ∈ Vx ∩ [a, b]} < . The family {(ai, bi)}m i=1 ∪ {Vx : x ∈ [a, b], x ∈ / F ()} is an open cover of [a, b]. Let
δ > 0 be a Lebesgue number of this cover, so that every set of diameter < δ is contained in some single
element of this cover. Suppose Γ : a = t0 < · · · < tk = b is a partition of [a, b] with mesh(Γ) < δ. Then
each interval P[tj−1, tj ] of the partition is contained in some single (ai , bi) or single Vx . For each fixed index
i,
P the sum {V [g; tj−1, tj ] : [tj−1, tj ] ⊆ (ai , bi)} cannot exceed V [g; ai, bi]. It again follows that the sum
{ω(f; [tj−1 , tj ]) · V [g; tj−1, tj ] : [tj−1 , tj ] ⊆ (ai , bi)} cannot exceed 2 · kfk∞ · V [g; ai, bi], and therefore that
the sum
( )
X [ Xm
ω(f; [tj−1 , tj ]) · V [g; tj−1, tj ] : [tj−1 , tj ] ⊆ (ai , bi) ≤ 2 · kfk∞ · V [g; ai, bi] ≤ 2 · kfk∞ ·  .
i i=1

If an interval [tj−1 , tj ] ⊆ Vx , then ω(f; [tj−1 , tj ]) <  and therefore

X X
k(Γ)
{ω(f; [tj−1 , tj ]) · V [g; tj−1 , tj ] : ∃Vx ⊇ [tj−1 , tj ]} ≤  · V [g; tj−1, tj ] = V [g; a, b] ·  .
j=1

8 Math 501:01 Fall 1999

So we have the estimate

X
k(Γ)
ωj (f; Γ) · V [g; tj−1, tj ] < 2 · kfk∞ ·  + V [g; a, b] ·  = [2 · kfk∞ + V [g; a, b]] ·  ,
j=1

and since the r. h. s. of this inequality is a constant multiple of  and the only thing assumed about the
partition was that mesh(Γ) < δ, we have shown that
 
k(Γ)
X 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
mesh(Γ)→0  
j=1

Z b
which is sufficient for the existence of the Riemann-Stieltjes integral f(t) dg(t) in sense (II).
a

If the integrator g is discontinuous, then the necessity of this “almost-everywhere-continuity” condition

will unfortunately fail. The counterexample is again the choice f = χ(c,b] and g = χ[c,b] , where a < c < b.
Here f is discontinuous at c and the variation of g over any open interval containing c is 1, but the integral in
sense (I) manages to exist anyway. On the other hand, there is no obstacle to adapting the proof of necessity
when g is continuous on [a, b]:
Proposition: Let f be a bounded real-valued function on the interval [a, b] and g a continuous function
of bounded variation on the same interval. If the set of discontinuities of f fails to be g-null, then f is not
Z b
Riemann-Stieltjes-integrable dg in sense (I) on the interval [a, b]: the integral f(t) dg(t) fails to exist in
a
sense (I).
Proof. Again, the set of points of discontinuity of f is given by
∞
[ 1
{x ∈ [a, b] : ω(f; t) > 0} = {x ∈ [a, b] : ω(f; t) ≥ },
k
k=1

and since the union by hypothesis is not g-null, some among the (closed and bounded, therefore compact)
uniands F (1/k) must fail to be g-null. There must therefore be a set F (η) and a number S  > 0 with
the
P property that if {(ai , b i )} is a countable family of open intervals for which F (η) ⊆ i (ai , bi ), then
i V [g; ai , b i ] ≥ . Let Γ : a = t0 < · · · < tk = b be a partition of [a, b]; then for any δ > 0, the open
cells {(tj−1 , tj )}kj=1 together with the δ-balls {B(tj ; δ)}kj=0 cover [a, b] and a fortiori cover F (η). The sum of
the variations V [g; ai, bi] corresponding to the the open cells {(tj−1 , tj )}kj=1 that have nonempty intersection
with F (η), together with the variations V [g; tj − δ, tj + δ] over the δ-balls centered on the points of Γ, must
therefore be ≥ :

X X
k
{V [g; tj−1, tj ] : F (η) ∩ (tj−1 , tj ) 6= ∅} + V [g; tj − δ, tj + δ] ≥  .
j=0

Since k is determined by the choice of Γ, δ > 0 is in our hands, and t 7→ V [g; a, t] is continuous at each tj ,
X
{V [g; tj−1, tj ] : F (η) ∩ (tj−1 , tj ) 6= ∅} ≥  .

If F (η) ∩ (tj−1 , tj ) 6= ∅ then there exists some ball centered on a point of F (η) that is contained in (tj−1 , tj ),
and therefore ω(f; [tj−1 , tj ]) ≥ η. It follows that
X
{ω(f; [tj−1 , tj ]) · V [g; tj−1, tj ] : F (η) ∩ (tj−1 , tj ) 6= ∅} ≥ η · 

9 Math 501:01 Fall 1999

and so certainly
X
k
ω(f; [tj−1 , tj ]) · V [g; tj−1, tj ] ≥ η · 
j=1

holds for every partition of [a, b]. The necessary condition

 
k(Γ)
X 
lim ωj (f; Γ) · V [g; tj−1, tj ] =0
(Γ,⊇)  
j=1

for the Riemann-Stieltjes integrability of f dg in sense (I) can thus never be satisfied: all such sums will be
≥ η · .

So can necessary and sufficient conditions be found? The answer is yes, but one has to “normalize” the
discontinuities of g in some way: the customary way is to make g continuous from the right. While it is
possible to carry out the details without using the notion of Lebesgue-Stieltjes measures (or [regular] Borel
measures on the line), there really is no reason to avoid measure-theoretic machinery at this point, and so
we need to look at how those machines are built.

10 Math 501:01 Fall 1999