Infinite Series: The Abel-Dini Theorem and

Convergence Tests
Logan Gnanapragasam
June 10, 2019

1 Introduction
In this paper, we will discuss infinite series. In K. Knopp’s book [3], a proof that there is
no perfect test for convergence is given. To do this, Knopp uses the Abel-Dini Theorem,
which is of interest in its own right. The Abel-Dini Theorem is discussed more fully in T.
H. Hildebrandt’s article [2]. In section 2, we provide a proof of the Abel-Dini Theorem and
discuss some applications. In section 3, we will discuss the results in [3] about convergence
tests.
Fairly little advanced machinery is required to prove the results on convergence tests.
Since this is a topic in real analysis, the reader should expect to see many inequalities. They
should also be familiar with sequences of numbers. Familiarity with sequences of functions
and uniform convergence is recommended but by no means required for the discussion of
convergence tests. We will occasionally use lim sup and lim inf, but the reader does not need
to know all of the equivalent definitions. We will give the lim sup definition that is most
applicable to us below.

1.1 Real Analysis Background

Here we give the definitions and basic theorems on series in real analysis that we will need.
We do not prove any of these results; proofs can be found in most introductory real analysis
books or (advanced) calculus books.
Before we discuss series, we define lim sup and lim inf.

Definition 1 (lim sup and lim inf). Given a sequence {an }∞

n=1 , we say that M = lim sup an
n→∞
if

• for each u > M there are only finitely many indices n for which an > u, and

• for each l < M there are infinitely many indices n for which an > l.

Similarly, we say m = lim inf an if

n→∞

• for each l < m there are only finitely many indices n for which an < l, and

1
 • for each u > m then there are infinitely many indices n for which an < u.
We allow lim sup and lim inf to be ±∞, removing the appropriate condition in the standard
definition. For example, we say lim sup an = ∞ if for each l ∈ R there are infinitely many
n→∞
indices n for which an > l. The remaining modifications are left to the reader.
Every sequence has a lim sup and lim inf, and this definition uniquely specifies these
values; see Exercise 1.5.9 and the discussion preceeding it in [1]. We will not prove these
facts.
Now we turn to infinite series. We begin with series of numbers.
∞
X ∞
X
Definition 2 (Convergence of a Series of Numbers). A series an , often written an
n=1 1
when the index of summation is unambiguous, converges if the sequence {Sn }∞n=1 defined
Xn
by Sn = ak converges (as a sequence of numbers). If Sn → S, we say that the sum of
k=1
the series is S.
∞
X
If the starting index of summation is j, so that the series is an , we modify the
n=j
N
X
definition by declaring SN = an . If the starting index is unknown or irrelevant, we will
n=j
X ∞
X
write an instead of an . It is useful to give a name to the sequence associated with
j
the series.
∞
X N
X
Definition 3 (Partial Sums). Given a series an , we say that SN = an is the N th
n=1 1
partial sum of the series. Thus our definition of convergence is that a series converges if
the sequence of partial sums converges.
∞
X ∞
X
Observe that whenever the series an converges, so does an , since the constant
n=1 n=N
N
X −1
an has been subtracted off from every partial sum of the original series. This observation
n=1
justifies the following definition.
∞
X ∞
X
Definition 4 (Tails). If an converges, we say that an is the N th tail of the series.
n=1 n=N

The indexing in this definition differs from the definition given in some other texts. Some
∞
X
authors call an the N th tail. The difference is not significant as long as the theorems
n=N +1
are stated correctly, but the reader should be aware of this distinction.

2
 The following result follows immediately from the fact that the sequence of partial sums
converges if and only if it is Cauchy.
X∞
Theorem 1 (Cauchy Convergence Criterion). The series ak converges if and only if for
1
m
X
each  > 0, there is N such that ak <  whenever m ≥ n ≥ N .
k=n
From this, one can prove the basic comparison test for series of nonnegative terms. Also,
by applying the triangle inequality, one finds that absolute convergence implies convergence.
For the statements of these results, see Theorems 6.11 and 6.17 in [1].
On occasion we will have something to say about uniform convergence of a series of
functions. The definition of uniform convergence of a sequence of functions can be found in
Section 7.1 of [1].
Definition 5 (Uniform Convergence of a Series of Functions). Given a sequence of real-
∞
X
∞
valued functions {fn }n=1 defined on a set S, we say that the series of functions fn
n=1
N
X
converges uniformly on a set W ⊆ S if the sequence of partial sums SN = fn , which
n=1
are also real-valued functions defined on S, converges uniformly on W .
We also have an analogous result to the Cauchy Convergence Criterion, which follows
from the fact that a sequence of functions converges uniformly if and only if the sequence is
uniformly Cauchy (Theorem 7.7 in [1]).
Theorem 2 (Cauchy Convergence Criterion for Functions). Given a sequence of functions
∞
X
fk : S → R, fk converges uniformly on S if and only if for each  > 0, there is N such
1
m
X
that fk (x) <  for all x ∈ S whenever m ≥ n ≥ N .
k=n

It is possible to prove the Weierstrass M -Test (Theorem 7.9 in [1]) with the triangle in-
equality and the Cauchy Convergence Criterion for numbers and the Cauchy Convergence
Criterion for functions.
While the Cauchy Convergence Criterion is important, it is not always easy to apply.
There are several tests that can be used to determine whether a series converges. The first
is the standard limit comparison test, not stated here (see Theorem 6.12 in [1]). This isn’t
always easy to apply because we need to come up with a series to compare to. Two of the
best known convergence tests which do not require us to find a comparison series are the
ratio and root tests.
∞
X ak+1
Theorem 3 (Ratio Test). If ak is a series of positive terms, u = lim sup , and
1 k→∞ a k
ak+1
l = lim inf , then the series converges if u < 1 and diverges if l > 1. Otherwise no
k→∞ ak
conclusions can be drawn.

3
 ∞
X 1
Theorem 4 (Root Test). If ak is a series of nonnegative terms and r = lim sup (ak ) k ,
1 k→∞
then the series converges if r < 1 and diverges if r > 1. If r = 1 then no conclusions can be
drawn.
It is possible to prove that if all the an are positive, then
ak+1 1 1 ak+1
lim inf ≤ lim inf (ak ) k ≤ lim sup (ak ) k ≤ lim sup ,
k→∞ ak k→∞ k→∞ k→∞ ak
which shows that if the ratio test gives the convergence/divergence of a particular series
then so does the root test (i.e., the root test is better than the ratio test). However, the
X∞
root test can also be inconclusive; both the divergent series 1 and the convergent series
1
∞
X 1
2
are series on which the root test gives r = 1. To understand why r = 1 is an issue,
1
n
we need to examine the proof of the root test. For convergence when r < 1, the series is
X
compared to a geometric series xn with r < x < 1. For divergence when r > 1, one

observes that |an | ≥ 1 infinitely often, so an 6→ 0 and hence the series diverges. However,
since the geometric series has radius of convergence 1, when r = 1 we can’t compare to a
geometric series, but we also cannot say that the terms are bigger than 1 in absolute value
infinitely often.
Inconclusiveness is not specific to tests relying on comparison to geometric series. As an
example, we briefly discuss Raabe’s Test.
∞
X an+1
Theorem 5 (Raabe’s Test). If ak is a series of positive terms for which → 1 and
1
an
 
an+1
n 1− → L as n → ∞, then the series converges if L > 1 and diverges if L < 1. If
an
L = 1 then no conclusions can be drawn.
For a more complete discussion of Raabe’s Test, along with a proof, see the subsection
“Raabe’s Test” in Section 6.2 of [1]. The proof involves comparing the series to the p-series
for a well-chosen p. As we will see in Section 2.1, the p-series converges if and only if p > 1.
When L = 1, we cannot compare to the p-series and hence the test is inconclusive.
Notice that all of our tests rely on comparison. In Section 3, to prove that there is “no
perfect test for convergence”, we will show that no series can be used in the comparison test
to find the convergence/divergence of all other series.

2 Abel-Dini Theorem
In this section, we prove the Abel-Dini Theorem and discuss some of its corollaries. Unless
otherwise stated, all series have positive terms.
The proof will be very similar to the proof in [2], but there are some differences. Our
first step is to prove a result in the case that the original series converges.

4
 ∞ ∞
X X an
Lemma 1. If ak converges, then with Tn as the nth tail, 1+α
converges if and only
1 1
Tn
if α < 0.

The proof of this result given in [2] requires Bernoulli’s inequality to show that for x ≥ 0
and 0 < β < 1, xβ = (1+(x−1))β ≤ 1+β(x−1). The standard proof of Bernoulli’s inequality
that we are aware of requires taking derivatives; however, the proof of Lemma 1 that we
present here only requires an inequality that can be derived from the elephant-teacup
formula, or the difference of powers formula.

Inequality 1. For any integer m > 0 and real number v > 0,

1
1
vm ≤ m
(v − 1) + 1.

1
The reader should note that this is Bernoulli’s inequality with β = .
m
1
Proof of Inequality 1. Let µ = v m . Then µ > 0 and the inequality we wish to show is
equivalent to m(µ − 1) ≤ µm − 1. There are three cases to consider: µ = 1, µ < 1, and
µ > 1.
If µ = 1 then we have equality, since m(µ − 1) = 0 = µm − 1.
In the other two cases, we apply the elephant-teacup formula. We do the case that µ < 1.
Since µ < 1, we have µk < 1 for all integers k ≥ 1 and hence
µm − 1
m > µm−1 + µm−2 + · · · + 1 = .
µ−1

Since µ − 1 < 0 it follows that m(µ − 1) < µm − 1. The case µ > 1 differs only in that certain
inequalities are reversed and hence is left to the reader.
Before proving the lemma, we state one more fact. For fixed v satisfying 0 < v < 1, the
function t 7→ v −t is monotonically increasing. For the proof, write v −t = e−t log v and apply
that ex is monotonically increasing and − log v > 0.
Now we can prove Lemma 1. The proof involves writing the terms in the original series
as differences of tails and applying the Cauchy Convergence Criterion.
∞
X
Proof of Lemma 1. Since ak converges, for n large enough we have Tn < 1, so Tn ∈ (0, 1).
1
an an
Hence for α > 0 one finds 0 < ≤ 1+α . Thus by the basic comparison test it suffices
Tn Tn
∞
X bn
to show that diverges. For this, we write bn = Tn − Tn+1 and apply the Cauchy
1
Tn
Convergence Criterion. Since the tails form a decreasing sequence, we have
m m m
X bk X Tk − Tk+1 1 X 1 Tm+1
= ≥ Tk − Tk+1 = (Tn − Tm+1 ) = 1 − .
k=n
Tk
k=n
T k Tn
k=n
Tn Tn

5
 Tm+1 1
Now for each (fixed) n, there is m ≥ n, m dependent upon n, such that < . To see
Tn 2
this, observe that the denominator is positive and fixed while the numerator decreases to 0.
1
Now we apply the Cauchy Convergence Criterion with  = . For any N , let n = N and
2
Tm+1 1
then take m ≥ n such that < . Then we have
Tn 2
m
X bk Tm+1 1
≥1− > .
k=n
Tk Tn 2
m
X bk 1
In other words, there is no N such that < whenever m ≥ n ≥ N . Hence by the
k=n
Tk 2
Cauchy Convergence Criterion, the series diverges.
∞
X bn
We now want to show that converges for each α < 0. We will instead prove
1
Tn1+α
∞
X bn
that 1 converges for each integer m ≥ 0. This will give the desired result, because
1−
1 T m
n
1 bn bn
for any α < 0 we can take m so large that α ≤ − . Then we have 0 ≤ 1+α ≤ 1 ,
m Tn 1−
m
Tn
∞
X bn
whence 1+α
converges by the basic comparison test.
1
Tn
bn Tn − Tn+1
Let m be fixed. We can rewrite the mth term in the series, 1 , as 1 =
1− 1−
Tn m Tn m
1
Tn − Tn+1 1
Tnm . Now we can apply Inequality 1 in the form 1 − u ≤ m(1 − u m ), obtaining
Tn
 
1   1  1 1 1
bn m Tn+1 m Tn+1 m
m m
1 = Tn 1− ≤ Tn m 1 −  = mTn − mTn+1 .
1−
m
Tn Tn
Tn
∞
X 1 1
The convergence of the telescoping series mTnm − mTn+1
m
implies the convergence of
1
∞
X bn
1 by the basic comparison test.
1−
1 Tn m
The reason why we include this proof rather than the proof using Bernoulli’s inequality is
that this proof only requires basic algebra and somewhat intuitive facts about convergence.
The ingredients of the proof are the elephant-teacup formula, the increasing property of
ex , properties of limits, the Cauchy Convergence Criterion, and the basic comparison test.
Moreover, the proof also contains a proof that the function f defined by
∞
X an
f (α) = 1+α
, α<0
1
Tn

6
is continuous in α, because each term in the sum is continuous in α and the series converges
1
uniformly on each of the intervals Sm = −∞, − by the Weierstrass M -Test with Mn =
m
an
1 . The continuity of f is not particularly important in the material on convergence
1−
m
Tn
tests, but it is an interesting result to obtain.
We are now ready to prove the Abel-Dini Theorem.
∞ ∞
X X bn
Theorem (Abel-Dini). If bk diverges, then with Sn as the nth partial sum,
1 1
Sn1+α
converges if and only if α > 0.

The proof involves taking the divergent series and forming a convergent series with the
1
appropriate tails, then applying Lemma 1. We will want the nth tail of the series to be ,
Sn
1 1 Sn+1 − Sn bn+1
so the nth term of the series is an = − = = .
Sn Sn+1 Sn Sn+1 Sn Sn+1
Before we begin the proof, we record another fact. For fixed t, the function defined for
positive v by v 7→ v −t is increasing if t < 0 and decreasing if t > 0. To see this, again
write v −t = e−t log v and notice that log v is increasing, so −t log v is increasing if t < 0 and
decreasing if t > 0; since ex is increasing, we find that e−t log v increases if t < 0 and decreases
if t > 0. If t = 0 then the function is constant.
∞
X
Proof of the Abel-Dini Theorem. Suppose bn diverges. Define
1

bn+1 1 1
an = = − .
Sn Sn+1 Sn Sn+1
∞
X 1 1 1
Then the series ak converges, because its nth partial sum is − and → 0 as
1
S1 Sn Sn
∞
X 1
n → ∞. By the same reasoning, the nth tail is Tn = ak = . By Lemma 1 the series
n
Sn
∞
X an
1+α
converges precisely when α < 0. Now we have
1
Tn

an bn bn+1
1+α
= Sn1+α = −α .
Tn Sn Sn+1 Sn Sn+1
∞
X bn+1 1
By replacing α with −α and switching the direction of α < 0, one finds α
con-
1
Sn S n+1
verges if and only if α > 0.
1
This is refered to in [2] as the “Pringsheim modification” of the result.

7
 From here, to obtain the result we apply the basic comparison test. Since the Sn are a
−α
monotonically increasing sequence of positive numbers, we have Sn−α ≥ Sn+1 if α > 0 and
−α −α
Sn ≤ Sn+1 if α ≤ 0. Hence we have
bn+1 bn+1
α
≥ α
, α>0
Sn Sn+1 Sn+1 Sn+1
∞
X bn+1
so by the basic comparison test 1+α converges if α > 0. On the other hand, we have
1
Sn+1

bn+1 bn+1
α
≤ α
, α≤0
Sn Sn+1 Sn+1 Sn+1
∞ ∞
X bn+1 X bn
so 1+α diverges if α ≤ 0. Hence converges if and only if α > 0, and since
1
Sn+1 2
Sn1+α
∞
b1 X bn
adding 1+α does not affect the convergence or divergence of the series, 1+α
converges
S1 1
S n
if and only if α > 0.
∞
X bn
Once again, the function g(α) = 1+α
defined for α > 0 is continuous in α. This can
1
S n
be shown by comparison with the series defining f in Lemma 1.
Lemma 2. If fn and gn are sequences of functions defined on a set S, 0 ≤ gn ≤ fn for all
∞
X X∞
n ≥ 1, and fn converges uniformly on S, then gn converges uniformly on S.
1 1

Proof. We will apply the Cauchy Convergence Criterion for functions. Let  > 0 be arbitrary;
X∞
since fn converges uniformly on S, there is K such that for all k ≥ j ≥ K we have
1
k
X k
X k
X
fn (x) = fn (x) <  for all x ∈ S. Hence for all k ≥ j ≥ K we have gn (x) =
n=j n=j n=j
Xk k
X ∞
X
gn (x) ≤ fn (x) <  for all x ∈ S, so gn converges uniformly on S.
n=j n=j 1

From this it is possible to deduce that g is continuous. Define sequences of functions

bn+1 an
fn , gn : (0, ∞) → R by fn (α) = α = 1−α , with an as defined in the proof of the Abel-
Sn Sn+1 Tn
bn+1
Dini Theorem, and gn (α) = 1+α . These are positive and satisfy 0 ≤ gn ≤ fn on (0, ∞),
Sn+1
∞  
X 1
and in the proof of Lemma 1 we proved that fn converges uniformly on , ∞ for
1
m
∞  
X 1
each positive integer m, so gn converges uniformly on , ∞ for each positive integer
1
m

8
m. Since the gn are all continuous, the series defines a continuous function on (0, ∞). Hence
∞
X bn
1+α
defines a continuous function of α. Adding in the first term, which is continuous in
2
S n
α on (0, ∞), we find that g is continuous on (0, ∞). As was the case with f , the continuity
of g is also not used in our discussion of comparison tests, but it is interesting to obtain.

2.1 Applications
We now discuss some applications of the Abel-Dini Theorem. The results about convergence
tests do not depend on the results in this section, so the reader can skip to Section 3.
The Abel-Dini Theorem yields the convergence properties of the p-series, which is often
examined via the integral test (see the proof of Theorem 6.9 in [1]). Indeed, let bk = 1. The
∞
X 1
Abel-Dini theorem applies and gives that 1+α
converges if and only if α > 0.
1
n
∞
X 1
The convergence properties of can also be proven from the Abel-Dini
1
n(log n)1+α
∞
1 X
0
Theorem. The previous paragraph shows that if bk = , the series b0k diverges. Setting
k 1
n ∞
X 1 X 1
Sn = , the Abel-Dini Theorem gives that converges if and only if α > 0. To
k=1
k 1
nSn1+α
get the result with logarithms we will apply the limit comparison test.
It is possible to show (Exercise 6.2.24 of [1]) that the sequence {cn }∞ n=1 defined by Sn =
cn +log n is positive and decreasing and hence convergent; its limit γ is the Euler-Mascheroni
log n cn
constant. So =1− → 1 as n → ∞. Hence
Sn Sn
1  1+α
nSn1+α log n
1 = → 1 as n → ∞,
n(log n)1+α S n

∞
X 1
so the limit comparison test gives that converges if and only if α > 0. Notice
2
n(log n)1+α
that we have to replace the lower index of summation with a 2 because when n = 1 the
denominator is zero.
The usual proof of both of these results uses the integral test. The benefit to the Abel-
1
Dini method is that we didn’t have to show that f (x) = is decreasing for large
x(log x)1+α
x. For series of the form
X 1
,
n(log n)(log log n) · · · (log log · · · log n)1+α

where the starting index is chosen so that all terms are well defined, the the Abel-Dini
Theorem does not apply so easily, as it’s not immediately clear how to get terms of the form
log · · · log n. The following way of extending this result is due to [2].

9
 ∞
X 1
We begin by attempting to generalize the result that 1+α
converges if and
1
n(log n)
only if α > 0, just as the Abel-Dini Theorem generalizes the p-test. A nice generalization
∞ ∞
X X bn
would be that if bk diverges, then with Sn as the nth partial sum,
1 1
Sn (log Sn )1+α
converges if and only if α > 0, but this is unfortunately not true. The “if” direction is true
(that is, if α > 0 then the series converges), but in some cases the “only if” direction fails.
n
An example given in [2] is a series whose nth partial sum is Sn = nn ; the details are not
discussed here.
To shed light on the issue, we attempt to prove the result to understand what additional
assumptions need to be made. Consider the series whose kth term is b0k = log Sk − log Sk−1 ,
X∞
where we declare S0 = 1. Then b0k diverges because the nth partial sum is log Sn . Hence
1
X b0n X log Sn − log Sn−1
by the Abel-Dini Theorem, = converges if and only if
(log Sn )1+α (log Sn )1+α
α > 0, where we take the starting index large enough that the denominator is never zero.
Recalling that bn = Sn − Sn−1 , the quotient of the nth terms in the original and new series
is
bn Sn−1 Sn−1
1− −1
Sn (log Sn )1+α Sn − Sn−1 Sn Sn
= = = .
log Sn − log Sn−1 Sn Sn Sn−1
Sn log log log
(log Sn )1+α Sn−1 Sn−1 Sn
x−1 x−1
It is possible to show that is strictly increasing on (0, 1), that lim+ = 0, and
log x x→0 log x
x−1
that lim = 1; to prove these facts, differentiate or use L’Hôpital’s rule. Since the terms
x→1 log x
Sn−1
bn in the original series are all positive, ∈ (0, 1) so the ratio satisfies
Sn
Sn−1
−1
Sn bn log Sn − log Sn−1
≤ 1 and hence ≤ ;
Sn−1 Sn (log Sn )1+α (log Sn )1+α
log
Sn
∞
X bn
thus converges if α > 0 by the basic comparison test.
1
Sn (log Sn )1+α
Sn−1
The issue in the “only if” direction is that could tend to 0. If it does, then the
Sn
Sn−1
−1
Sn
limit of the sequence is also 0, so the limit comparison test (which would show
Sn−1
log
Sn
that either both series converge or both series diverge) does not apply. On the other hand,
Sn−1 δ−1
if there is a δ > 0 for which > δ for n large, then setting  = , we have  > 0 and
Sn log δ

10
(whenever n is large)

Sn−1
−1
Sn log Sn − log Sn−1 bn log Sn − log Sn−1
≤ ≤ 1 and hence  · ≤ ≤ ,
Sn−1 (log Sn )1+α Sn (log Sn )1+α (log Sn )1+α
log
Sn
∞
X bn Sn−1
so 1+α
converges if and only if α > 0. The condition that > δ > 0 can
1
Sn (log Sn ) Sn
bn bn
be rewritten as 1 − > δ or as < 1 − δ < 1. Hence we have shown the following:
Sn Sn
∞
X
Lemma 3. If bn diverges and Sn is the nth partial sum, and if there is δ > 0 so that
1
∞
bn bn X bn
< 1−δ for n large (that is, if the ratio is bounded away from 1), then
Sn Sn 1
Sn (log Sn )1+α
converges if and only if α > 0.
3
With bn = 1 the hypotheses are satisfied (take r = , and then the inequality is satisfied
4
for n ≥ 2). This yields the logarithm result which we proved earlier by using the Euler-
Mascheroni constant. ∞
X bn
In [2] the author states the result for 1+α
and says that it is
1
Sn (log Sn )(log log Sn )
possible to continue this reasoning to obtain the general result. The proof of the general
result is not conceptually different from the proof of this result, so we will give the general
result.
∞
X
Theorem (p-test for Logarithms). If bn diverges and Sn is the nth partial sum, and if
1
bn
there is r < 1 so that < r for large n, then
Sn
X bn
Sn (log Sn )(log log Sn ) · · · ((log)k−1 Sn )((log)k Sn )1+α

converges if and only if α > 0. Here (log)j denotes the logarithm applied j times, so
(log)0 (x) = x and (log)j+1 (x) = log(logj (x)) for j ≥ 1.
∞
X (log)k Sn − (log)k Sn−1
Proof. We compare the terms in the series to those in k S )1+α
, declaring
1
((log) n
S0 = 1; the Abel-Dini Theorem shows that this series converges if and only if α > 0, since
(log)k Sn − (log)k Sn−1 are the terms of a divergent series with nth partial sum (log)k Sn . The

11
quotient of the nth terms in the series is
bn bn
Sn (log Sn ) · · · ((log)k−1 Sn )((log)k Sn )1+α Sn (log Sn ) · · · ((log)k−1 Sn )
=
(log)k Sn − (log)k Sn−1 (log)k Sn − (log)k Sn−1
((log)k Sn )1+α
Sn − Sn−1
= .
Sn (log Sn ) · · · ((log) Sn )[(log)k Sn − (log)k Sn−1 ]
k−1

Now we can rewrite this quotient as a product (henceforth referred to as the product of
index n) by multiplying by 1 several times:

Sn − Sn−1 log Sn − log Sn−1 (log)k−1 Sn − (log)k−1 Sn−1

··· .
Sn [log Sn − log Sn−1 ] log Sn [log log Sn − log log Sn−1 ] (log)k−1 Sn [(log)k Sn − (log)k Sn−1 ]
Now we examine each factor to obtain bounds on the entire product when n is large. The
jth factor in the product of index n is
(log)j−1 Sn−1 (log)j−1 Sn−1
1 − −1
(log)j−1 Sn − (log)j−1 Sn−1 (log)j−1 Sn (log)j−1 Sn
= = .
(log)j−1 Sn (log)j−1 Sn−1
 
(log)j−1 Sn [(log)j Sn − (log)j Sn−1 ]
log log
(log)j−1 Sn−1 (log)j−1 Sn
In the discussion before Lemma 3, we saw that there is N1 such that if n > N1 then the first
factor in the product of index n,
Sn−1
−1
Sn
,
Sn−1
log
Sn
lies in the interval [, 1], where  is a fixed positive real number dependent only upon the
sequence {Sn }∞ n=1 (that is,  is not dependent upon the index n). Now we will show that for
each j such that 2 ≤ j ≤ k, there is Nj so that if n > Nj then the jth factor in the product
1
of index n lies between and 1:
2
(log)j−1 Sn−1
−1
1 (log)j−1 Sn
<  < 1.
(log)j−1 Sn−1

2
log
(log)j−1 Sn
Once this is known, it follows that if n > max {N1 , . . . , Nk } then the quotient of the nth
terms of the series, which is exactly the product of index n, lies in (2−k+1 , 1), whence

(log)k Sn − (log)k Sn−1 bn

2−k+1  <
((log)k Sn )1+α Sn (log Sn )(log log Sn ) · · · ((log)k−1 Sn )((log)k Sn )1+α
(log)k Sn − (log)k Sn−1
< ,
((log)k Sn )1+α

12
 ∞
X bn
so k−1 S )((log)k S )1+α
converges if and only if α > 0 by comparison.
1
Sn (log S n ) · · · ((log) n n
We prove the bound on the jth factor of index n, where 2 ≤ j ≤ k. First, notice that for
j−1 S
n large we have (log)j−1 Sn > (log)j−1 Sn−1 > 0 and hence (log) n−1
(log)j−1 Sn
∈ (0, 1), so

(log)j−1 Sn−1
−1
(log)j−1 Sn
 < 1.
(log)j Sn−1

log
(log)j Sn
(log)j−1 Sn−1
This proves part of the result. Now I claim that → 1 as n → ∞. If we can
(log)j−1 Sn
show this, then by sequential continuity we would have
(log)j−1 Sn−1
−1
(log)j−1 Sn
 →1
(log)j−1 Sn−1

log
(log)j−1 Sn
(log) j−1
Sn−1
1 (log)j−1 Sn
−1
as n → ∞, so <   whenever n is sufficiently large.
2 (log)j−1 Sn−1
log (log)j−1 Sn
(log)j−1 Sn−1
We will show that → 1 as n → ∞ by induction on j. For the base case, we
(log)j−1 Sn
write
Sn−1
log
log Sn−1 log Sn log Sn−1 − log Sn Sn
= + =1+ .
log Sn log Sn log Sn log Sn
Sn−1
For n sufficiently large, we have 0 < 1 − r < < 1, so log SSn−1 is bounded. On the other
Sn n

log SSn−1 log SSn−1

hand, log Sn → ∞ as n → ∞, so n
→ 0. Thus it follows that 1 + n
→ 1 as n →
log Sn log Sn
(log)j0 −1 Sn−1
∞, which proves the base case. For the inductive step, we assume that →1
(log)j0 −1 Sn
(log)j0 Sn−1
as n → ∞. We need to show that → 1. We proceed as was done in the base
(log)j0 Sn
case, writing
(log)j0 −1 Sn−1
 
log
(log)j0 Sn−1 (log)j0 Sn−1 − (log)j0 Sn (log)j0 −1 Sn
= 1 + = 1 + .
(log)j0 Sn (log)j0 Sn (log)j0 Sn
(log)j0 −1 Sn−1
 
By the inductive hypothesis and by sequential continuity, we have log → 0,
(log)j0 −1 Sn
(log)j0 +1 Sn−1
and since (log)j0 Sn → ∞ we have → 1, completing the inductive step.
(log)j0 +1 Sn

13
 As we described before, this shows that the jth factor in the product of index n lies
1
between and 1 when n is sufficiently large, yielding the bound on the product of index
2 ∞
X (log)k Sn − (log)k Sn−1
n, which shows that the series in the theorem and k S )1+α
either both
1
((log) n
converge or both diverge. Therefore,
X bn
Sn (log Sn )(log log Sn ) · · · ((log)k−1 Sn )((log)k Sn )1+α

converges if and only if α > 0, as claimed.

3
By taking bn = 1, the hypotheses are satisfied with r = and n ≥ 2. Although this proof
4
is possibly not as clear as the one using the integral test, the result obtained is more general.
Unfortunately the proof requires more than just “elementary” facts; we needed L’Hôpital’s
rule, and sequential continuity played a major role in the estimates. However, this doesn’t
relate directly to the material on convergence, so if the reader is lost in this material they
can still continue.

3 Convergence Tests
We are now ready to discuss convergence tests. These results are all due to [3], in which the
results are attributed to A. Pringsheim. Once again, all series have positive terms.
The first of our results has to do with comparison tests. We have so far used only the
basic comparison test. A generalized version can be stated with lim sup and lim inf.
∞
X ∞
X
Theorem 6 (Limit Comparison Test, lim sup and lim inf version). Let an and bn be
1 1
series of positive terms. Then:
∞ ∞
an X X
• if 0 < lim inf then the convergence of an implies the convergence of bn , and
n→∞ bn
1 1

∞ ∞
an X X
• if lim sup < ∞, then the divergence of an implies the divergence of bn .
n→∞ bn
1 1

Proof. We use the characterization of lim sup given in the section on real analysis, along
with the basic comparison test.
an
For the first statement, choose l ∈ R so that 0 < l < lim inf . Then there is N so that
n→∞ bn
∞ ∞
an X X
if n > N , then > l. Hence an > lbn > 0, so if an converges then so does lbn and
bn 1 1
X∞
hence so does bn .
1

14
 an
For the second statement, choose u ∈ R so that lim sup < u; then since all terms in
n→∞ bn
an
both series are positive, u > 0. Now there is N so that for n > N , we have < u, and
bn
∞
X ∞
X ∞
X
hence an < ubn . If an diverges, then ubn diverges, whence bn diverges.
1 1 1

This is more general, but we still need a convergent or divergent series to compare to.
The question is whether there is a convergent series that works with the comparison test to
give the convergence of all other convergent series, or a divergent series that works with the
comparison test to give the divergence of all other divergent series.
X∞
To restate the question, is there a convergent series cn of positive terms with the
1
∞
X cn
property that whenever cn is a convergent series of positive terms, lim inf
e > 0? If so,
1
n→∞ cn
e
∞
X cn
then a series an converges if and only if lim inf
> 0; the “if” direction is guaranteed by
n→∞ an
1
the limit comparison test, and the “only if” direction would be guaranteed by the assumption
X∞
on cn .
1
∞
X
The Abel-Dini Theorem shows that there is no such series. Indeed, if cn has the
1
desired properties, then it converges. Consider
∞ ∞
X X c
cn =
e √n
1 1
Tn
∞
X ∞
X
where Tn is the nth tail of ck ; the series cn converges by Lemma 1, but the ratio of
e
1 1
∞
cn √ cn X
terms is = Tn → 0 as n → ∞, so lim inf = 0. Hence cn does not satisfy the
√cn n→∞ cn
Tn
e 1
conditions. ∞
X
For divergent series, the question is whether there is a divergent series dn with the
1
∞
X dn
property that whenever den diverges, lim sup is real. The Abel-Dini Theorem once
n→∞ d en
1
X∞
again shows that there is no such series. Indeed, if dn has the desired properties, then it
1
diverges. Consider
∞ ∞
X X dn
den =
1 1
Sn

15
 ∞
X ∞
X
where Sn is the nth partial sum of dn ; the series den diverges by the Abel-Dini The-
1 1
∞
dn X
orem, but the ratio of terms is dn
= Sn → ∞ as n → ∞, so dn does not satisfy the
Sn 1
conditions.
With very little effort, the Abel-Dini Theorem has allowed us to show that there is no
series which is “perfect” for the comparison test. Less formally, there is no series that
converges (resp. diverges) slower than any other convergent (resp. divergent) series.
What about other tests? One conceivable test is to have a sequence {pn }∞n=1 of positive
X ∞
numbers such that for any series an of positive terms, if an pn → 0, then the series
1
converges, and if there is α > 0 such that for all n, an pn ≥ α, then the series diverges.
We give two examples to show that there are sequences that satisfy the first condition and
sequences that satisfy the second condition, so that the conditions are not untenable. If
pn = n2 , then the first condition is satisfied because if n2 an → 0 then there is C such that
∞ ∞
C X X C
n2 an ≤ C for all n, whence an ≤ 2 , so that an converges by comparison to 2
. If
n n=1 n=1
n
pn = 1, then the second condition is satisfied because if an ≥ α > 0, then the series diverges
X∞
by comparison to α.
n=1
Since in both cases we used comparison, we should expect that the sequences fail one of
the properties. In the case that pn = 1, we have the harmonic series as a counterexample,
∞
1 X 1 1
since an pn = → 0 but diverges. In the case that pn = n2 , we have an = 3 as a
n 1
n n2
∞
1 X 1
counterexample; for all n ≥ 1 we have an pn = n 2 ≥ 1 > 0, but 3 converges.
1 n2
Now the question is whether there is a sequence that satisfies both conditions. Notice
that even if can we find such a sequence, it will not give the convergence/divergence of every
series. In particular, given the sequence {pn }∞
n=1 , define the terms of a series by

1 1
a2k = , a2k+1 = .
2kp2k p2k+1
1
Then a2k p2k = → 0 but a2k+1 p2k+1 = 1 > 0. Hence an pn 6→ 0 but also there is no α > 0
2k
such that for all n, an pn ≥ α, so the test would be inconclusive.
In any case, we do not have to worry about the test being inconclusive because there is
no such sequence. We will prove this by using the Abel-Dini Theorem. Indeed, suppose that
the sequence {pn }∞n=1 satisfies the desired conditions. Take any α > 0 and consider the series
α
whose nth term is an = . Since an pn = α > 0, the series must diverge if the test works.
pn
∞
X an
With Sn as the nth partial sum, the Abel-Dini Theorem shows that diverges. But
1
Sn

16
 ∞
an α X an
pn = → 0 since Sn → ∞, so if our test worked then would converge. Hence
Sn Sn 1
S n
{pn }∞
n=1 does not satisfy the desired conditions.
We have now seen that there is no perfect comparison test for convergence. We close by
discussing a result in [3] relating convergent and divergent series.

Theorem 7. For any monotonely decreasing sequence {n }∞ n=1 tending to 0, there is a con-
∞
X ∞
X
vergent series cn and divergent series dn with the property that cn = n dn .
1 1

The reader should not be surprised to learn that we will apply the Abel-Dini Theorem.
1
Proof. Let pn = if n ≥ 1, p0 = 0. Then pn increases monotonically to ∞. Now the
n
∞
X
series pn − pn−1 has nth partial sum pn , and since this increases monotonically to ∞,
n=1
∞
X pn − pn−1
the series diverges. The Abel-Dini Theorem applies, and we find that di-
n=1
pn
∞ ∞
X pn+1 − pn X pn+1 − pn
verges. Reindexing gives that diverges, and hence diverges.
n=0
pn+1 n=1
pn+1
∞
pn+1 − pn pn n+1 pn+1 − pn X
Now = 1− = 1− . Therefore, with dn = , we have that dn
pn+1 pn+1 n pn+1 1
  ∞
n+1 X
diverges. But n dn = 1 − n = n − n+1 , so with cn = n dn , the series cn is
n 1
telescoping and hence convergent, since the nth partial sum is 1 − n+1 → 0 as n → ∞.

4 Conclusion
The Abel-Dini Theorem offers a generalization of the p-test. While the p-test is usually
proved via the integral test and hence requires improper integrals, the proof of the Abel-
Dini Theorem presented here requires almost no results in advanced calculus. We only
needed to carefully use inequalities. The theorem is also widely applicable. We saw that it
could be used to generalize the p-test for logarithms. The advantage to this proof is that we
did not have to verify the hypotheses of the integral test, which requires that the function
involved is increasing, and we also did not need the theory of improper integrals. Moreover,
once we started examining convergence tests, the Abel-Dini Theorem gave immediately that
no series could be totally effective with the comparison test. Convergence tests, in general,
are proved by comparison to a convergent or divergent series. Indeed, the ratio, root, and
Raabe’s tests are all ways of using the comparison test indirectly, since they compare a given
series to a geometric series or p-series. Since there is no perfect series for comparison, the
work done here strongly suggests that there cannot be a perfect convergence test.

17
References
[1] G. B. Folland. Advanced Calculus. Prentice Hall, Upper Saddle River, NJ, 2002.

[2] T. H. Hildebrandt. Remarks on the Abel-Dini Theorem. The American Mathematical

Monthly, 49(7):441–445, 1942.

[3] K. Knopp. Theory and Application of Infinite Series, pages 298–305. London: Blackie,
1951.

18