January 28, 2012

Abel Partial Summation Formula

First some notation: For x R let [x] denote the greatest integer less than or equal to x. Thus, for
example, [3.1] = 3 and [1.7] = 2. We introduce the fractional part of x: {x} := x [x]. Thus,
for example, {3.1} = 0.1 and {1.7} = 0.3. Observe that 0 {x} < 1.
Let {cn } be a sequence of complex numbers and f (x) a complex-valued function defined for
x R+ . We assume f has a continuous derivative in R+ . Define for x R+
X
C(x) =
cn
1nx

Thus, for example, C(3.1) = c1 + c2 + c3 .

The following is an algebraic identity that can be easily checked. Fix x with k x < k + 1, then
X
X
cn f (n) =
cn f (n)
1nx

1nk

C(k)f (k)

C(n) (f (n + 1) f (n))

(1)

1nk1

Now
X

C(n) (f (n + 1) f (n))

1nk1

f 0 (t) dt

1nk1

n+1

C(n)

1nk1

n+1

C(t)f 0 (t) dt since C(t) = C(n), n t < n + 1

C(t)f 0 (t) dt

=
1
x

C(t)f 0 (t) dt

C(t)f 0 (t) dt

Now
Z

x
0

C(t)f (t) dt = C(k)

k

f 0 (t) dt = C(k)f (x) C(k)f (k) = C(x)f (x) C(k)f (k)

Putting these results (1) and (2) together we obtain the Abel partial summation formula
X

Z
cn f (n) = C(x)f (x)

C(t)f 0 (t) dt

1nx

Examples:
1. Let
Hx =

X 1
n
1nx

Choose f (x) = 1/x and cn = 1, then

X

C(x) =

1nx

1 = [x].

(2)

By the Abel partial summation formula

Z x
1
[x]
[t] 2 dt
+
Hx =
x
t
1
Z x
1
x {x}
(t {t}) 2 dt
=
+
x
t
1
Z x
{t}
{x}
+ log x
= 1
dt
x
t2
1
Rx
dt converges to a limit as x . This is so because
The integral 1 {t}
t2
Z x
 Z x
Z x

{t} 
1
1

|{t}| 2 dt
dt
dt

2
2
t
t
1
1
1 t

(3)

and the last integral is convergent as x . Thus write

Z
Z
Z x
{t}
{t}
{t}
dt =
dt
dt
2
2
t
t
t2
1
x
1
By a similar estimate as in (3) we see that for large x
Z
{t}
1
dt = O( )
2
t
x
x
Thus we have shown
1
Hx = log x + + O( ), x
x
where

{t}
dt.
t2
1
The constant is called Eulers constant. Its approximate value is = 0.5772156649 . . ..
It is an unsolved problem to prove that is irrational. From the representation
for one can derive the series expansion


X
1
1
=1
log(1 + )
n
n+1
n=1
Z

=1

This is a very slowly convergent series: Summing the first 10,000 terms gives
0.577266. . . .
2. Consider the sum
Sx =

X log n
n
1nx

Choose
cn = 1, f (x) =
2

log x
x

so that
C(x) = [x], f 0 (x) =

1 log x
x2

The summation formula tells us

Z x
1 log t
log x
[t]

Sx = [x]
dt
x
t2
1
Z x
1 log t
log x

(t {t})
dt
= (x {x})
x
t2
1
Z x
Z x
1 log t
1 log t
log x
{t}
= log x + O(
)
dt +
dt
x
t
t2
1
1
Now use the fact that
Z
1

and that

1
1 log t
dt = log x (log x)2
t
2
x

Z
1

1 log t
log x
dt =
2
t
x

to conclude that
Z x
Z
1 log t
1 log t
log x
{t}
dt
=
{t}
+
O(
), x ,
t2
t2
x
1
1
and hence

1
log x
Sx = (log x)2 + c1 + O(
), x
2
x
where c1 is a constant given by
Z
1 log t
{t}
c1 =
dt.
t2
1

3. Let
Sx =
Show that

X (log n)2
n
1nx

1
(log x)2
Sx = (log x)3 + c2 + O(
), x .
3
x

4. Assume f = f (x) is a continuously

R x differentiable function of x with f (x) 0 as
x
know that
R 0. Since f (x) f (1) = 1 f (t) dt, and limx f (x) exists,
R we
0
f (t) dt exists (and equals f (1)). We further require that 1 |f (t)| dt < .
1
We then have
Z x
X
f (n) =
f (t) dt + f + O(|f (x)|) + o(1)
(4)
1nx

where

{t}f 0 (t) dt.

f = f (1) +
1

Remark: If f is nonnegative, we can eliminate the o(1) term and simply have an
error of O(f (x)).
Proof: In Abel partial summation choose cn = 1, then
Z x
X
[t]f 0 (t) dt
f (n) = [x]f (x)
1

1nx

(t {t}) f 0 (t) dt
= (x {x}) f (x)
1
Z x
Z x
0
{t}f 0 (t) dt {x}f (x)
tf (t) dt +
= xf (x)
1

Since |{t}f 0 (t)| |f 0 (t)| we have as x

Z
Z
0
|
{t}f (t) dt|
|f 0 (t)| dt = o(1) as x
x

since the integral 1 |f 0 (t)|dt exists. If f (x) is nonnegative then we dont need
the absolute values and we get an error of the order O(f (x)).
Observe that the first three examples are special cases of (4).
(a) Choose f (x) = 1/x for example 1.
(b) Choose f (x) = log x/x for example 2.
(c) Choose f (x) = (log x)2 /x for example 3.