MT 2016

ANALYSIS I
Webnotes by H.A. Priestley

1
2

Contents

0. Setting the scene 6

1. The Real Number System: Field Axioms 6
1.1. Real numbers and axioms for them 6
1.2. Properties of arithmetic 6
1.3. Powers 9
1.4. Stocktaking so far, and looking further afield: R compared to other systems 9
2. The Real Number System: Order Axioms 9
2.1. Positive numbers; order relations; the real numbers as a ‘number line’ 9
2.2. Properties of the order on R 9
2.3. Interaction of order and inequalities with arithmetic 10
2.4. A useful result: Bernoulli’s Inequality 10
2.5. Modulus (real case) 11
2.6. The Triangle Law and the Reverse Triangle Law 11
3. The Complex Numbers, Briefly 12
4. The Completeness Axiom for the Real Numbers 13
4.1. Upper and lower bounds 13
4.2. Supremum (= least upper bound) 13
4.3. The Completeness Axiom for the Real Numbers 13
4.4. Supremum: examples 14
4.5. Infimum (= greatest lower bound) 14
4.6. Two theoretical examples 14
4.7. Maximum (greatest element) and minimum (least element) 14
4.8. Modulus measures distance 15
4.9. Approximation Property: capturing supremum via approximation 15
√
4.10. Theorem (existence of 2) 15
4.11. Incompleteness of Q 16
4.12. Theorem (existence of nth roots) 16
4.13. Theorem (the Archimedean Property of the Natural Numbers) 16
4.14. Theorem (compare with the well-ordered property of N) 16
4.15. Density properties 17
4.16. R: SUMMING UP 17
5. An aside: countability 17
5.1. Comparing sizes of sets 17
5.2. Finite, countable and uncountable sets 17
5.3. Some familiar sets which are countably infinite 18
5.4. New countable sets from old 18
 3

5.5. Theorem: Q is countable 19

5.6. Theorem: R is uncountable 19
6. Sequences 19
6.1. Real numbers via approximations 20
6.2. The formal definition of a real or complex sequence 20
6.3. Manufacturing new sequences 21
6.4. Tails 21
6.5. Capturing ‘arbitrarily close to’ via ε 21
6.6. Convergence of a real sequence 21
6.7. Tails Lemma 22
6.8. Examples: convergence established directly (but messily) from the definition 22
6.9. Testing a sequence for convergence: remarks and technical tips 23
6.10. Sandwiching Lemma (simple form) 23
6.11. Examples, by sandwiching 24
6.12. More examples (two important limits, employing some useful techniques) 24
6.13. Theorem (uniqueness of limits) 24
6.14. Proposition (limits and modulus) 25
6.15. Limits and inequalities 25
6.16. A general Sandwiching Lemma 25
6.17. Proposition (a convergent sequence is bounded) 26
6.18. Examples: divergence 26
6.19. Infinity 26
6.20. Important examples: limits involving powers 27
6.21. Complex sequences 27
6.22. Theorem (convergence of complex sequences) 27
6.23. Examples: complex sequences 28
7. Subsequences 28
7.1. Subsequences 28
7.2. Proposition (subsequences of a convergent sequence) 29
7.3. Example (using subsequences to establish divergence) 29
8. The Algebra of Limits (AOL) 29
8.1. epsilon-handling 29
8.2. Facilitating the construction of ε-N proofs: making proofs less fiddly 30
8.3. Theorem: (AOL), Part I 30
8.4. Examples:(AOL) Part I 30
8.5. Facilitating the construction of ε-N proofs: useful technical facts 31
8.6. Theorem: (AOL), Part II 31
4

8.7. Example: a typical (AOL) application 32

8.8. Proposition (reciprocals and infinite/zero limits) 32
8.9. More worked examples, (AOL) 32
8.10. A ‘true-or-false?’ worked example 33
8.11. Orders of magnitude 33
8.12. The O and o notation 33
9. Monotonic Sequences 34
9.1. Monotonic sequences: definitions 34
9.2. Monotonic Sequence Theorem 34
9.3. Worked examples: Monotonic Sequence Theorem 35
10. The Bolzano–Weierstrass Theorem, and the Cauchy Convergence Criterion 36
10.1. The Scenic Viewpoint Theorem 37
10.2. Bolzano–Weierstrass Theorem (for real sequences) 37
10.3. Bolzano–Weierstrass Theorem (for complex sequences) 37
10.4. Cauchy sequences 37
10.5. Facts about Cauchy sequences 38
10.6. Theorem (Cauchy Convergence Criterion) 38
11. Convergence of Series 38
11.1. Series: introductory examples 38
11.2. Series: definitions 39
11.3. Convergence of series: first examples 39
11.4. Theorem (terms of a convergent series) 40
P 1
11.5. Example: the harmonic series, k>1 40
k
11.6. Theorem (series of non-negative terms) 41
11.7. Theorem (Comparison Test, simple form) 41
11.8. Examples: simple Comparison Test 41
11.9. Theorem (Cauchy criterion for convergence of a series) 41
11.10. Absolute convergence 41
11.11. Theorem (absolute convergence implies convergence) 42
11.12. A standard example: the series k>1 k −p for p ∈ R
P
42
12. Convergence Tests for Series 42
12.1. Leibniz’ Alternating Series Test 43
12.2. Examples: AST 43
12.3. Comparison Test, simple form (recap of 11.7 and remarks) 43
12.4. Comparison Test, limit form 44
12.5. Examples: Comparison Test (limit form) 44
12.6. D’Alembert’s Ratio Test, for series of strictly positive terms 44
 5

12.7. Ratio Test, preliminary examples 45

12.8. Ratio Test: tips and warnings 45
12.9. Testing for absolute convergence: a corollary to the Ratio Test 12.6 46
12.10. The Integral Test: preamble 46
12.11. Integral Test Theorem 47
12.12. Corollary: the Integral Test 47
12.13. Applications of the Integral Test and Integral Test Theorem 48
12.14. Example: exploiting the existence of Euler’s constant 48
12.15. Definition: rearrangement 49
12.16. Which test? A miscellany of examples in outline 49
12.17. Tips for finding counterexamples 50
12.18. A ‘true-or-false?’ worked example 50
12.19. Postscript: other tests for convergence 50
13. Power Series 51
13.1. Power series 51
13.2. Elementary functions: exponential, trigonometric and hyperbolic functions 51
13.3. AOL for series in general and power series in particular 52
13.4. The radius of convergence of a power series 52
13.5. Proposition on radius of convergence (technical but important) 53
13.6. Examples: radius of convergence 53
13.7. Differentiation Theorem for (real) power series 54
13.8. Some applications of the (real) Differentiation Theorem 55
13.9. Trigonometric functions: what became of π 56
13.10. Looking ahead: the importance of complex power series 56
13.11. The binomial expansion 56
13.12. The logarithmic series 57
6

A summary of the axioms for the real numbers is given in a separate reference
sheet. References [BS ... ] are to the textbook by Bartle and Sherbert.

0. Setting the scene

Introductory remarks on the content, aims and style of the Analysis I course, and how it
relates to school mathematics and to other Prelims courses.

1. The Real Number System: Field Axioms

[BS, Section 2.1]

1.1. Real numbers and axioms for them.

What is a real number? The need for clear assumptions about the real numbers and the
role of these axioms in the course.

Notation: the set of real numbers is denoted R.

We introduce the axioms for R and explore how the familiar rules of arithmetic can be
obtained as consequences of the axioms (and of properties derived from the axioms).
See reference sheet Axioms for the Real Numbers for the list of axioms.
Addition
The operation + of addition and the axioms, A1–A4, for addition.
Multiplication and avoiding collapse
The operation · of multiplication and the axioms, M1–M4, for multiplication. The need
to assume 0 6= 1 (Axiom Z).
Distributive Law
Linking addition and multiplication together via the distributive law D.

1.2. Properties of arithmetic.

Properties (1)–(5) require A1–A4 only, (6)–(8) require M1–M4 only. (9)–(13) together
draw on all the axioms. Here a, b, c, x and y are real numbers.
(1) If a + x = a for all a then x = 0 (uniqueness of zero element).
(2) If a + x = a + y then x = y (cancellation law for addition, which implies uniqueness of
additive inverse of a).
(3) −0 = 0.
(4) − (−a) = a.
(5) − (a + b) = (−a) + (−b).
(6) If a · x = a for all a 6= 0 then x = 1 (uniqueness of multiplicative identity).
(7) If a 6= 0 and a · x = a · y then x = y (cancellation law for multiplication, which implies
uniqueness of multiplicative inverse of a).
(8) If a 6= 0 then 1/ (1/a) = a.
(9) (a + b) · c = a · c + b · c.
 7

(10) a · 0 = 0.
(11) a · (−b) = − (a · b). In particular (−1) · a = −a.
(12) (−1) · (−1) = 1.
(13) If a · b = 0 then either a = 0 or b = 0 (or both). Moreover, if a 6= 0 and b 6= 0 then
1/(a · b) = (1/a) · (1/b).
Full proofs are given below but only a small selection of these will be given in
lectures, to illustrate how derivations from the axioms should look. You are
recommended to work through some of the others by yourself, referring to the
notes only to check that you haven’t cut corners or made correct assertions
you have omitted to validate.

Proof. One step at a time!

(1) We have
x=x+0 by A3
=0+x by A1
=0 by the hypothesis, with a = 0.
(2) We have
y =y+0 by A3
= y + (a + (−a)) by A4
= (y + a) + (−a) by A2
= (a + y) + (−a) by A1
= (a + x) + (−a) by hypothesis
= (x + a) + (−a) by A1
= x + (a + (−a)) by A2
=x+0 by A4
=x by A3.
(3) 0 + 0 = 0 by A3 with a = 0. Hence −0 = 0 by A4 and (2).
(4) We have
(−a) + a = a + (−a) by A1
=0 by A4, and
(−a) + − (−a) = 0 by A4.
Now appeal to (2) (cancellation law for addition).
(5) Exercise on Problem sheet 1.

Claims (6)–(8) do for multiplication what (1)–(3) do for addition, but with inverses neces-
sarily considered only for non-zero a, as M4 requires. The proofs go the same way.

(9) Note that the statement is like Axiom D but with multiplication on the other side. We
have
(a + b) · c = c · (a + b) by M1
=c·a+c·b by D
=a·c+b·c by M1 twice.
8

(10) We have

a·0+0=a·0 by A3
= a · (0 + 0) by A3
=a·0+a·0 by D.

Now appeal to (2) (cancellation law for addition).

(11) We have

(a · b) + (a · (−b)) = a · (b + (−b)) by D
=a·0 by A4
=0 by (10). Also
(a · b) + (−(a · b)) = 0 by A4.

Now appeal to (2).

(12) We have

(−1) · (−1) = −(−1) by (11)

=1 by (4).

(13) Assume for a contradiction that a, b 6= 0 but a · b = 0. Then

0 = (1/a · 1/b) · 0 by (10)

= 0 · (1/a · 1/b) by M1
= (a · b) · (1/a · 1/b) by hypothesis
= ((b · a) · 1/a) · 1/b by M2
= (b · (a · 1/a)) · 1/b by M2
= (b · 1) · 1/b by M4
= b · (1/b) by M3
=1 by M4,

This contradicts Axiom Z. The second assertion has been proved along the way. 

We now have established that, according to the axioms we have set up, arith-
metic behaves as we expect it to. We shall henceforth not spell out uses of
the axioms in such detail, and are ready to revert to more familiar notation.

Notation Henceforth we shall adopt the customary notational shortcuts. We write

a−b in place of a + (−b),

ab in place of a · b,
a/b in place of a · (1/b),
−1
a as alternative notation for 1/a.

We may, thanks to A2 and M2, omit brackets in iterated sums and products, and write
for example a + b + c without ambiguity.
 9

1.3. Powers.
Let a ∈ R \ {0}. As usual, we define a0 = 1. We then define
(
ak+1 = ak · a for k = 0, 1, 2, . . . (an ‘inductive’ or ‘recursive’ definition),
a` = 1/(a−` ) for ` = −1, −2, . . . .
Then in particular a1 = a and a2 serves (as expected) as shorthand for a·a. Problem sheet 1,
Q. 3 asks for a proof of the familiar law of indices.

1.4. Stocktaking so far, and looking further afield: R compared to other systems.
Substitute a set F in place of R in the above, and assume that the operations + and ·,
now defined on F , satisfy the axioms
A1–A4,
M1–M4,
Z,
D.
Then F , or more precisely (F ; +, ·), is a field. So our assumptions about R so far can be
summed up as
(R; +, ·) is a field.
The number systems Q (rational numbers) and C (complex numbers) are also fields. But N
(the set of natural numbers, 0, 1, 2, . . .) and Z (the integers) have weaker arithmetic properties
and are not fields. [In the Prelims Linear Algebra courses, in the definition of a vector space,
the scalars are assumed to be drawn from any field.]

2. The Real Number System: Order Axioms

[BS, Section 2.2]

2.1. Positive numbers; order relations; the real numbers as a ‘number line’.
There is a subset P (the (strictly) positive numbers) of R such that, for a, b ∈ R,
P1 a, b ∈ P =⇒ a + b ∈ P;
P2 a, b ∈ P =⇒ a · b ∈ P;
P3 exactly one of a ∈ P, a = 0 and −a ∈ P holds.
We write a < b (or b > a) iff b − a ∈ P and a 6 b (or b > a) iff b − a ∈ P ∪ {0} (the
non-negative numbers).
We subsequently make use of 6 or of <, as convenient.

2.2. Properties of the order on R.

The following properties justify our visualising R as a ‘number line’.
Reflexivity: a 6 a.
Proof. a − a = 0 ∈ P ∪ {0}, by A4.

Antisymmetry: a 6 b and b 6 a together imply a = b.

Proof. If a − b = 0 or b − a = 0, then a = b by properties of addition. Otherwise
P 3 a − b and P 3 b − a = −(a − b) (by properties of addition). Now apply P3.
10

Transitivity: Assume a 6 b and b 6 c. Then a 6 c, and likewise with < in place of 6.

Proof. We have c − a = c + (−a) = c + 0 + (−a) = c + (−b) + b + (−a) = (c − b) + (b − a)
by properties of addition. So the result for < follows from the definition of < and P3.
The proof for 6 is the same except for the need to allow also for the trivial cases
a = b and/or b = c.

Trichotomy: Exactly one of a < b, a = b and b < a holds.

Proof. This follows from P3 and the definition of <.

2.3. Interaction of order and inequalities with arithmetic.

[The arithmetic operations interact as expected with inequalities—indeed the axioms for
order are set up to ensure this.]
The following statements hold.
(1) 0 < 1 (equivalently, 1 ∈ P),
(2) a < b if and only if −b < −a. In particular a > 0 iff −a < 0.
(3) a < b and c ∈ R implies a + c < b + c.
(4) a < b and 0 < c implies ac < bc.
(5) a2 > 0, with equality iff a = 0.
(6) a > 0 iff 1/a > 0.
(7) If a, b > 0 and a < b then 1/b < 1/a.
Claims (2)–(4) also hold with 6 replacing <.

Proof. We use freely the properties of arithmetic established in Section 1.

(1) By trichotomy, exactly one of (i) 1 < 0, (ii) 1 = 0, (iii) 0 < 1 holds. Axiom Z rules
out (ii). Suppose for a contradiction that 1 < 0. Then −1 = 0 + (−1) ∈ P. We deduce
that (−1) · (−1) ∈ P by P2. But 1 = (−1) · (−1) (by 1.2(12)), so 0 < 1 and we have a
contradiction to trichotomy.
(2) By properties of addition,
a < b ⇐⇒ b − a ∈ P
⇐⇒ (−a) − (−b) ∈ P
⇐⇒ (−a) > (−b).
(3) a + c < b + c iff 0 < (b + c) − (a + c) = b − a iff a < b.
(4) a < b and c > 0 implies bc − ac = (b − a)c > 0, by P2.
(5) Note a2 = 0 iff a = 0, by 1.2(13). Assume a 6= 0. Then we have a2 = a · a = (−a) · (−a).
Since either a > 0 or −a > 0 it follows that a2 > 0 by P2.
(6) Assume for contradiction a > 0 but 1/a < 0. Then −1 = −(a · (1/a)) = a · (−1/a) > 0
by P2 which is impossible by (1). Likewise we obtain a contradiction if a < 0 and
1/a > 0.
(7) Use (4) and (6). 

2.4. A useful result: Bernoulli’s Inequality.

Let x be a real number with x > −1 and let n be a positive integer. Then
(1 + x)n > 1 + nx.

[The Mean Value Theorem (from Analysis II) allows one to extend the result by replacing
n by any real number > 1.]
 11

Proof. We shall prove the inequality by induction—note that the inequality is trivially true
when n = 1.
Suppose that, for k ∈ N,
(1 + x)k > 1 + kx
holds for all real x > −1. Then 1 + x > 0 by 2.3(3) and kx2 > 0 as k > 0 and x2 > 0 by
2.3(5).
(1 + x)k+1 = (1 + x) (1 + x)k by definition
> (1 + x) (1 + kx) by hypothesis and 2.3(4) (6 version)
= 1 + (k + 1) x + kx2 by A1–A4
> 1 + (k + 1) x by 2.3.
Hence the result follows by induction. 

2.5. Modulus (real case). [BS 2.2]

The modulus |a| of a ∈ R is defined by


a
 if a > 0,
|a| = 0 if a = 0,

−a if a < 0.

The definition makes sense by P3.

Basic facts about modulus: for any a, b, c,

(1) | − a| = |a|;
(2) |a| > 0;
(3) |a|2 = a2 ;
(4) |ab| = |a||b|;
(5) −|a| 6 a 6 |a|;
(6) if c > 0, then |a| 6 c iff −c 6 a 6 c. if c > 0, then |a| < c iff −c < a < c.

Proof. (1) and (2) are immediate from the definition and the fact that a > 0 iff −a < 0.
We can prove (3) and also (4), by using P3 to enumerate cases; recall too that (−a)(−a) =
a2 for any a.
For (5), note that (
−|a| 6 0 6 a = |a| if a > 0,
−|a| = a < 0 6 |a| if a < 0.

Now consider (6). Assume first that |a| 6 c, Then, by (5) and transitivity of 6, we get
−c 6 a 6 c. Conversely, assume −c 6 a 6 c. Then −a 6 c and a 6 c. Since |a| equals
either a or −a, we obtain |a| 6 c.
The case with < in place of 6 is handled similarly. 

2.6. The Triangle Law and the Reverse Triangle Law. [BS 2.2.3 and 2.2.4] Thw Triangle
Law is also known as the Triangle Inequality.
(1) Let a, b ∈ R. Then
|a + b| 6 |a| + |b|.
(2) Let a, b ∈ R. Then
|a + b| > ||a| − |b|| .
12

Proof. (1): We have

−|a| 6 a 6 |a| and − |b| 6 b 6 |b|.
Adding (see Problem sheet 1, Q2(a)) and using properties of addition we get

−(|a| + |b|) 6 a + b 6 (|a| + |b|).

Now use 2.5(6).

(2): By the Triangle Law,

|a| = |a + b + (−b)| 6 |a + b| + |(−b)| = |a + b| + |b|,

so |a|−|b| 6 |a+b|, and likewise, reversing the roles of a and b, we get |b|−|a| 6 |b+a| = |a+b|.
Now use the fact that |c| is either c or −c always, and apply this with c = |a| − |b|. 

3. The Complex Numbers, Briefly

Properties of the complex numbers are covered in the course Introduction to Complex
Numbers and not in Analysis I. But we shall deal with complex numbers occasionally, and it
is useful to record what does, and does not, hold in relation to arithmetic and inequalities.
We have noted earlier that

• (C; +, ·) is a field.

Here addition and multiplication are defined in the usual way, in terms of real and imaginary
parts, or in the case of multiplication, alternatively via polar representation. The axioms
A1–A4 follow from the corresponding axioms for R. The multiplication axioms M1–M4
are most easily verified using polar coordinates. Axiom Z holds since 0, 1 are real. Axiom
D holds by a straightforward, but tedious, calculation.
We highlight what does not transfer from R to C. The key point to note is that, unlike
R, the complex numbers do not carry a total order relation: we cannot define < on C in
a way which is compatible with arithmetic and such that, for any w, z ∈ C, exactly one of
w < z, w = z or z < w holds. Exercise: Prove this by considering i · i. Therefore

inequalities are off limits unless the quantities being compared are real.

Thus, for z = 3 + 4i√and w = 4 + i we may correctly say that Re z < Re w, that Im z > Im w,
and that |z| = 5 > 17 = |w|.
Both the Triangle Law and the Reverse Triangle Law extend to complex numbers (note
that the modulus of a complex number is real): for z, w ∈ C,

|z + w| 6 |z| + |w| and |z + w| > ||z| − |w||.

[Proofs: Introduction to Complex Numbers notes, Proposition 10.]

 13

4. The Completeness Axiom for the Real Numbers

This section focuses on the order structure of R and exploits its interaction with the
arithmetic structure. In 4.3 we introduce the last of our axioms for R, the Completeness
Axiom. Throughout this section, and beyond, we shall make use of the axioms and results
in Sections 1 and 2, without spelling out the details.
On the basis of our axioms for arithmetic and order, we can’t distinguish between Q and
R. But you would claim they are different:
• in Q there is not a square root for 2 (standard proof, √
not included in lecture); but
• (you almost certainly believe that) there is a number 2 in R.
It will follow from the Completeness Axiom that 2 does have a square root in R (Theorem
4.10). Hence this axiom does distinguish R from Q (see 4.11).
Before presenting the Completeness Axiom we need some order-theoretic preliminaries.

4.1. Upper and lower bounds.

Definitions: Let S ⊆ R and b ∈ R. Then
b is an upper bound of S if s 6 b for all s ∈ S;
b is a lower bound of S if b 6 s for all s ∈ S;
S is bounded above if it has an upper bound;
S is bounded below if it has a lower bound;
S is bounded if it is bounded above and below.

Examples:
(a) R is not bounded above.
(b) Every element of R is an upper bound for the empty set.
(c) { s ∈ R | −4 6 s < 3 }: the numbers 3, 11, 1037 are upper bounds; 2.999999 is not an
upper bound. The set of lower bounds is (−∞, −4].
(d) { 1/n | n = 1, 2, . . . }: b is an upper bound for S iff b > 1 and b is a lower bound iff
b 6 0 (but can you justify this?).
(e) {1}: any b > 1 is an upper bound and any b 6 1 is a lower bound.
(f) Q: no upper bounds and no lower bounds.
(g) Let S := { x ∈ Q | x2 < 2 }. Then certainly S 6= ∅ and, for example, 3 is an upper
bound of S (why?).
Note that we can’t yet in every case here convincingly identify all the bounds.

4.2. Supremum (= least upper bound).

Let S ⊆ R. Then α is the supremum of S, denoted sup S, if
(sup1) s 6 α for all s ∈ S [α is an upper bound of S]
(sup2) s 6 b for all s ∈ S implies α 6 b [α is the least upper bound of S]
Note: sup S is unique if it exists. Why this definition? We’ll go straight to the Completeness
Axiom, then provide examples and commentary.

4.3. The Completeness Axiom for the Real Numbers.

Let S be a non-empty subset of R which is bounded above. Then sup S
exists.
14

Note the necessity for the exclusions here. The empty set has no supremum because it
has no least upper bound ((sup2) fails). A set which is not bounded above cannot have a
supremum because it has no upper bound ((sup1) must fail).

4.4. Supremum: examples.

(a) Consider S = [1, 2). Then 2 is an upper bound. And if b is any upper bound for S then
b > 2: otherwise, b < 2 and b < s := (2 + b)/2 and s ∈ S. So 2 is the least upper bound
of S. Note that sup S ∈/ S.
(b) Consider S = (1, 2]. Here sup S = 2. Note that sup S ∈ S.

4.5. Infimum (= greatest lower bound).

Analogous definitions apply here, by replacing 6 by > in the definitions above. Let S be
a subset of R and α ∈ R. Then α is the infimum of S, written inf S, if
(inf1) α 6 s for all s ∈ S [α is a lower bound for S]
(inf2) if b 6 s for all s ∈ S then b 6 α [α is the greatest lower bound of S]

4.6. Two theoretical examples.

(a) Let ∅ =
6 S ⊆ T and assume that T is bounded above. We claim that S is bounded
above and that sup S 6 sup T .
Proof. Let b be such that t 6 b for all t ∈ T . Then s 6 b for all s ∈ S, so S is indeed
bounded above. By the Completeness Axiom, sup S and sup T exist. Then sup T is an
upper bound for S and hence sup S 6 sup T . 

(b) Let T be non-empty and bounded below. Define S := { −t | t ∈ T }. Then S is

non-empty and bounded above and inf T exists and equals − sup S.
Proof. If b is any lower bound for T then −b is an upper bound for S, so S is bounded
above. Also S is nonempty because T is nonempty. So by the completeness axiom
sup S exists, and −b > sup S since sup S is the least upper bound for S. So b 6 − sup S
for every lower bound b for T .
Furthermore, if t ∈ T then −t ∈ S and so −t 6 sup S and therefore t > − sup S.
Thus − sup S is a lower bound for T , and we have already seen that − sup S is greater
than or equal to every lower bound for T . Hence inf T exists and equals − sup S. 
[This, together with a corresponding argument with sups and infs interchanged, im-
plies that we could equivalently have formulated the Completeness Axiom as ‘Every
non-empty subset of R which is bounded below has an infimum’.]

4.7. Maximum (greatest element) and minimum (least element).

Assume ∅ 6= S ⊆ R and let s0 ∈ R. We say s0 is the maximum of S, and write
s0 = max S, if
(max1) s0 ∈ S [s0 belongs to S]
(max2) s 6 s0 for all s ∈ S [s0 is an upper bound of S]
If a set S is empty or is not bounded above then max S cannot exist.

Example (sup and max compared): Let ∅ 6= S ⊆ R be such that sup S exists. Then S has
a maximum iff sup S ∈ S and then sup S = max S.

Similarly we say a non-empty set S which is bounded below has a minimum, min S, if
there exists s0 ∈ S such that s0 6 s for all s ∈ S.
 15

We now want to explore the notion of supremum more closely.

4.8. Modulus measures distance.

Given a, x ∈ R we may interpret |x − a| as the distance from x to a.
Important fact: For b > 0, and a, x ∈ R,
|x − a| < b ⇐⇒ a − b < x < a + b.

Proof. 2.5(6) gives |x − a| < b iff −b < (x − a) < b and this holds iff a − b < x < a + b. 

So by considering whether this holds for different values of b we can assess how good an
approximation x is to a.

4.9. Approximation Property: capturing supremum via approximation.

Let S be non-empty and bounded above (so sup S exists). Then, given ε > 0, there exists
sε (in general depending on ε) in S such that
sup S − ε < sε 6 sup S.
√
4.10. Theorem (existence of 2).
There exists a unique positive real number α such that α2 = 2.

Proof. Strategy: let

S := { s ∈ R | s > 0 and s2 < 2 }.
Then we seek to show that S is non-empty and bounded above, so that, by the Completeness
Axiom, α := sup S exists. To show α2 = 2, it will suffice (by trichotomy) to show that each
of (a) α2 > 2 and (b) α2 < 2 leads to a contradiction. We shall use the Approximation
Property to handle (b).

Step 1: S 6= ∅ and S is bounded above, by 2.

To prove this, first note that 1 ∈ S. Now let x > 2, so, by order properties,
x2 = x · x > 4 > 2,
which means x ∈
/ S. Hence 2 > s for all s ∈ S.
Step 2: Assume for contradiction that α2 < 2. Note also that α > 1 > 0 since 1 ∈ S. Let
h > 0. Then
(α + h)2 − 2 = α2 + 2αh + h2 − 2
< α2 − 2 + 3αh if h < α
<0 if h < (2 − α2 )/(3α).
2 − α2
 
Taking h such that 0 < h < min α, we find that α + h ∈ S, contradicting the
3α
assumption that α is the least upper bound.
Step 3: Assume for contradiction that α2 > 2. Let h > 0. Then α − h < α. By the
Approximation Property we can find s ∈ S such that α − h < s. But then
α2 − 2αh < α2 − 2αh + h2 < s2 < 2.
Take h 6 (α2 − 2)/(2α) to get the required contradiction.
16

Finally we must show that α is unique. If also β > 0 and β 2 = 2 then

0 = α2 − β 2 = (α − β)(α + β).
Since α, β > 0 this contradicts properties of order. 

4.11. Incompleteness of Q. The set Q of rationals does not satisfy the Completeness
Axiom with respect to the order it inherits from R. If it did,
T := { q ∈ Q | q > 0 and q 2 < 2 }
would have a supremum in Q. The proof in 4.10 works just as well for T as it does for S.
But we know there is no rational square root of 2.

4.12. Theorem (existence of nth roots). Any positive real number has a real nth root,
for any n = 2, 3, 4, . . ..
[Case of cube root of 2 is an exercise on Problem sheet 2.]

4.13. Theorem (the Archimedean Property of the Natural Numbers).

(i) N is not bounded above.
1
(ii) Let ε > 0. Then there exists n ∈ N such that 0 < < ε.
n
Proof. (i) Assume for a contradiction that N is bounded above. Then sup N exists by the
Completeness Axiom. By the Approximation Property with ε = 1/2, there exists k ∈ N
with
1
sup N − < k 6 sup N.
2
1
But then k + 1 ∈ N and k + 1 > sup N + , a contradiction.
2
For (ii), exploit the fact that 1/ε cannot be an upper bound for N. 

4.14. Theorem (compare with the well-ordered property of N).

(i) Every nonempty subset S of Z which is bounded below has a minimum (least element).
(ii) Every nonempty subset S of Z which is bounded above has a maximum (greatest
element).

Proof. i) We know that inf S exists (by applying the completeness axiom to {−s : s ∈ S} as
in 4.6(b)). So by the approximation property with ε = 1 there is some n ∈ S such that
inf S 6 n < inf S + 1.
It is enough to show that inf S = n, since then inf S ∈ S and so inf S = min S. Assume for
a contradiction that n 6= inf S, so that n > inf S and hence n = inf S + ε where 0 < ε < 1.
By the approximation property again, there exists some m ∈ S such that
inf S 6 m < inf S + ε = n.
Since n > m we have n − m > 0 and so n − m > 1 because n − m is an integer1, so
n > inf S + 1, which contradicts our first inequality for n.
The proof of (ii) is similar. 
1Strictly
speaking we have not proved this property of the integers. It can be proved by observing that a
strictly positive integer is a natural number and using induction to deduce that it must be at least 1.
 17

4.15. Density properties.

(i) Given a, b ∈ R with a < b there exists x ∈ Q such that a < x < b.
(ii) Given a, b ∈ R with a < b there exists y ∈ R \ Q such that a < y < b.
Proof: Exercise (Problem sheet 2, Q. 6).

4.16. R: SUMMING UP.

We assume that R satisfies the axioms set out on the reference sheet Axioms for the Real
Numbers:
• the arithmetic axioms A1–A4, M1–M4, Z, D (as discussed in Section 1);
• the order axioms P1–P3 (as discussed in Section 2);
• the Completeness Axiom (discussed above).
In summary, we assume R is a complete ordered field.
All of real analysis stems from R being a complete ordered field, and no additional
assumptions.

5. An aside: countability

The density properties recorded in 4.15 are important, the more so because, as we shall see
shortly, R is ‘a bigger set’ than Q: the set Q is countable whereas the set R is uncountable.
The objective in Analysis I is not to present a crash course in set theory but to give the
minimum amount of information on countability necessary to distinguish between countable
and uncountable sets in the context of the real numbers. See the supplementary notes on
Countability for an informal account of this topic which goes beyond the Analysis I syllabus.
These notes may be of interest to those who want to go deeper than Section 5 does and in
particular to Maths/Phil students.

5.1. Comparing sizes of sets.

Let A and B be sets. We say A and B are equinumerous (notation A ≈ B) if there is
a bijection f : A → B. Note ≈ has the properties of an equivalence relation.
Given sets A and B we shall write A 4 B if there is an injection f : A → B. Intuitively
this says that B is at least as big as A.

5.2. Finite, countable and uncountable sets.

Let A be a set. We call A finite if either A = ∅ or there exists n ∈ N such that
A ≈ {0, . . . , n − 1} (or equivalently if A ≈ {1, . . . , n}). A set which is not finite is said to be
infinite. [Another way to capture the notion of finiteness is to say that A is finite iff every
injective map from A to A is surjective: the Pigeonhole Principle holds for A.] Any subset
of a finite set is finite.
Note that any non-empty finite subset of R must be bounded above (in fact, it contains a
largest element (why?)). Hence any subset of R which is not bounded above must be infinite.
By the Archimedean property, N is not bounded above, and hence is infinite.
18

We call a set A
• countably infinite if A ≈ N;
• countable if A 4 N;
• uncountable if A is not countable.
[Warning: Some authors use ‘countable’ to mean what we call ‘countably infinite’.]
FACTS: (see supplementary material on countability)
(1) A is countable (that is, A 4 N) iff A is finite or countably infinite.
(2) If A 4 B and B 4 A then A ≈ B.

5.3. Some familiar sets which are countably infinite.

(a) N;
(b) N>0 := N \ {0};
(c) { 2k + 1 | k ∈ N } (the odd natural numbers);
(d) Z;
(e) N × N.
[Note that (c) shows that a countably infinite set may be equinumerous with a proper subset
of itself.]

Proof. (a) is immediate. For (b): the successor function, n 7→ n + 1, is a bijection from N
to N \ {0}. To prove (c), note that the map 2k + 1 7→ k is injective and maps the given set
onto N. Now consider (d). We can define a bijection f from Z to N by
(
−2k if k 6 0,
f (k) =
2k − 1 if k > 0.

We prove (e) by setting up a bijection f : N × N → N. Define f by

f ((m, n)= 2m (2n + 1) − 1.
Injectivity of f : 2m1 (2n1 + 1) = 2m2 (2n2 + 1) implies (by uniqueness of factorisation in N
(assumed)) that 2m1 = 2m2 so that m1 = m2 (make use of laws of indices for the last step)
and 2n1 + 1 = 2n2 + 1, whence n1 = n2 . So (m1 , n1 ) = (m2 , n2 ).
Surjectivity of f : take k ∈ N. Assume first that 2 - (k + 1). Then k is even and so
k = 20 (2n + 1) − 1, for some n ∈ N. Now assume 2|(k + 1). Then there exists m such that
2m | (k + 1) and 2m+1 - (k + 1) (use the fact that { 2m | m ∈ N } is not bounded above
(Problem sheet 2, Q. 2). Then k + 1 = 2m (2n + 1), for some n ∈ N. 

5.4. New countable sets from old.

Let A and B be countable sets. Let f : A → N and g : B → N be injective.
Claim 1: Assume A and B are disjoint. Then A∪B is countable. [Disjointness not essential,
but it simplifies the proof.]

Proof. Define (
2f (x) if x ∈ A,
h(x) =
2g(x) + 1 if x ∈ B.
Then h is an injection from A ∪ B to N. 

Claim 2: A × B is countable.
 19

Proof. Define h : A × B → N by
h((a, b)) = 2f (a)+1 3g(b)+1 for a ∈ A, b ∈ B.
Then uniqueness of factorisation in N>0 implies h is injective. 

5.5. Theorem: Q is countable.

Proof. We write Q as the disjoint union

Q>0 ∪ {0} ∪ Q<0 .
By 5.4, Claim 1, it will be enough to prove that Q>0 (and so, likewise, Q<0 ) is countable.
We can write each element of Q>0 as p/q where p, q ∈ N, p > 0, q > 0 and p/q is expressed
in its lowest terms. Then p/q 7→ 2p 3q is an injection into N. So Q>0 4 N as claimed. 

5.6. Theorem: R is uncountable.

We shall be able to prove this further on in the course using facts concerning decimal
representation of real numbers (see Supplementary note on decimals)—the classic proof due
to Cantor. Meantime, an alternative proof, using nested intervals and the Completeness
Axiom, can be found in [BS 2.5.4].
z

A note on familiar functions

We want our treatment of sequences, and a bit later, of series not to be divorced from
the functions you were introduced to at school and regularly encounter in other courses:
trigonometric functions, exponential functions, logarithms, and general powers. Accordingly
we’d like to involve such functions in examples and exercises. So, for now, we shall take
the existence and the properties of these functions for granted, and use them freely. Later
you will see formal definitions of these various functions and rigorous derivations of their
(familiar) properties.

When we use logarithms these will always be to base e. We adopt the notation log x for
loge x, rather than ln x.
Recall that, for a > 0 and x ∈ R, one defines ax = ex log a .

6. Sequences

[BS, Sections 3.1 and 3.2]

This section covers the rudiments of the theory of convergence of sequences. An ample
supply of worked examples is included in these webnotes. Examples omitted from lectures
are recommended for self-study.
20

6.1. Real numbers via approximations.

Examples:
n digits
z }| {
3 33 333 33 . . . 3
(a) , , , ... , are progressively better approximations to 1/3.
10 100 1000 10n
14 141 1414 14142 √
(b) , , , are well-known approximations to 2.
10 100 1000 10000
(c) [The Archimedean Property revisited] For any ε > 0 there exists N > 1 such that
0 < 1/N < ε. And then 0 < 1/n 6 1/N < ε for all n > N . This says that, apart from
a finite number of terms at the start, the terms of the sequence
1 1 1 1
1, , , , , . . .
2 3 4 5
are all within a distance ε of 0. Here ε can be any positive real number whatsoever.
(d) Consider the string of real numbers
1, −1, 2, −2, 3, −3, 4, −4, . . .
Your intuition tells you there is no limiting value in this case.

6.2. The formal definition of a real or complex sequence.

Officially, a sequence of real numbers (a real sequence) is an assignment
n 7→ α(n)
of a real number α(n) to each n = 1, 2, . . .. Thus a sequence is a function α : N>1 → R (where
N>1 = N \ {0} = {1, 2, 3, . . .}) and we call α(n) the nth term of the sequence. (Sometimes
later we shall work with sequences with terms labelled by n = 0, 1, 2, . . . instead.)
We shall usually write an in place of α(n) and then say that α defines the sequence
(an ) = (a1 , a2 , a3 , . . .)
also written (an )n>1 — that is, we specify the sequence by its terms. Note that the terms in
order determine the sequence.
A complex sequence (an ) is defined in the same way, but now with the terms an drawn
from the complex numbers, C. This section concentrates on convergence of real sequences.
We defer discussion of complex sequences until 6.21, but note that most of the results we
obtain for real sequences do have analogues for complex sequences. The exceptions will be
those whose statement or proof relies on inequalities involving the terms; recall the comments
in Section 3 about complex numbers and order. We adopt the convention that only in
statements of results applicable only to real sequences do we include the word ‘real’. Thus
‘sequence’ on its own can be read as ‘real or complex sequence’ once 6.21 has been studied.
Examples
(a) Let an = α(n), where α(n) = sin n/(2n + 1). Then the sequence looks like
1 1 1
( sin 1, sin 2, sin 3, . . . ).
3 5 7
(b) Let an = α(n), where α(n) = (−1)n . Then the sequence looks like
(−1, 1, −1, 1, −1, 1, −1, . . . ).
 21

6.3. Manufacturing new sequences. We can form new sequences from given ones ‘a term
at a time’: given sequences (an ) and (bn ), we have sequences (an + bn ), (−an ), (an bn ) and,
provided every bn is non-zero, (an /bn ). Also we can form (can ), for any constant c, and
(|an |).
Examples: let an = (−1)n and bn = 1 for all n. Then

(an + bn ) = (0, 2, 0, 2, 0, 2, . . .);

(−an ) = ((−1)n+1 );
(|an |) = (1, 1, 1, 1, 1, . . .).

The key notion in this section is that of convergence of a sequence (an ). We want to
analyse how the terms an of the sequence behave as n gets ‘arbitrarily large’ and specifically
whether or not the terms approach ‘arbitrarily closely’ some ‘limiting value’ L. For this we
need to convert these informal ideas into precise, formal, ones.

6.4. Tails. As regards the long-run behaviour of the terms of a sequence (an ) as n gets
arbitrarily large we don’t care what the values of the first few terms are, or what the first
10 million terms are, . . . . The notion of a tail of (an ) will allow us to capture this idea.
Given a sequence (an ) and any k ∈ N we can form a new sequence (bn ) by chopping off
the first k terms a1 , . . . , ak of (an ) and relabelling. That is, bn = an+k for all n. We call (bn )
a tail of (an ).

6.5. Capturing ‘arbitrarily close to’ via ε. From 2.5(6) (see also 4.8), we have, for
x, a ∈ R and b > 0,
|x − a| < b ⇐⇒ a − b < x < a + b
and this condition captures the statement that x lies within a distance b of a. In formulating
notions of limits and convergence we are interested in allowing b to be very small—arbitrarily
small—and shall use the customary symbol ε to denote a strictly positive real number playing
this role.

6.6. Convergence of a real sequence. Let (an ) be a sequence of real numbers and let
L ∈ R. Then we say that (an ) converges to L (notation: an → L (as n → ∞)) if

∀ε > 0 ∃N ∈ N ∀n > N |an − L| < ε.

Here N can, and almost always will, depend on ε. Note that we can replace ‘n > N ’ by
n > N and/or ‘|an − L| < ε’ by |an − L| 6 ε in this definition without changing the meaning
(WHY?). However it is crucial that ε should be strictly greater than 0.
When an → L we say that L is the limit of (an ) [in 6.13 we’ll show it must be unique]
and we write
L = lim an or just L = lim an .
n→∞

We say (an ) converges if there exists L ∈ R such that an → L as n → ∞. We say (an )

diverges if it does not converge.
The following technical, but intuitively plausible, result reinforces the notion that conver-
gence, or not, depends only on the long-run behaviour of a sequence.
22

6.7. Tails Lemma. Let (an ) be a sequence.

(i) If (an ) converges to a limit L then each tail of (an ) converges, to the same limit L.
(ii) Assume some tail (bn ) = (an+k ) converges. Then (an ) converges.

Proof. (i) Let k ∈ N and let bn = an+k for n > 1. Assume (an ) converges to L. Then
∀ε > 0 ∃N ∀n > N |an − L| < ε.
This implies |bn − L| = |an+k − L| < ε for all n > N because then n + k > n > N .
(ii) Since (bn ) converges there exists L such that
∀ε > 0 ∃N 0 ∀p > N 0 |bp − L| < ε.
That is, p > N implies |ap+k − L| < ε. Define N = N 0 + k. Then n > N implies that
0

n = p + k where p > N 0 and hence |an − L| < ε. So (an ) converges to L too. 

6.8. Examples: convergence established directly (but messily) from the defini-
tion.
(a) The basic fact that
1/n → 0 as n → ∞
is simply the Archimedean Property in new clothes. To see this note that the Archimedean
Property is the statement that, given ε > 0, there exists N ∈ N such that 1/N < ε.
Then for n > N we have |1/n − 0| = 1/n 6 1/N < ε.
See 6.20 for a discussion of the limiting behaviour of arbitrary powers of n.
1
(b) Let an = 1 + (−1)n √ . We claim an → L where L = 1. Let ε > 0. Then
n
 
n 1  1
 
|an − 1| = (−1) √  < ε ⇐= √ < ε

n n
1
⇐= n > N, where we choose N ∈ N with N > .
ε2
Here , and likewise below, the first ⇐= may be read as ‘if’ and any subsequent ones as
‘and hence if’.
1
(c) Let an = . Let ε > 0. Then
n2−n+1
1
|an − 0| = 2 < ε ⇐= n2 − n + 1 > ε−1
n −n+1
⇐= (n − 21 )2 > ε−1 − 3
4
q
1
⇐= n > N, where N ∈ N and N > 2
+ ε−1 − 34 ,
where we may assume without loss of generality (see 6.9(4)) that ε < 4/3. So an → 0.

n sin n2
(d) Let an = 3 . Let ε > 0. Then
3n − n − 1
n| sin n2 | n
|an − 0| = 3 < ε ⇐= 3 <ε since | sin x| 6 1 for all x
3n − n − 1 3n − n − 1
⇐= n/n3 < ε since n3 > n and n3 > 1
√
⇐= n > N where N > 1/ ε.
So an → 0. Note the need for care in handling inequalities in this example.
 23

6.9. Testing a sequence for convergence: remarks and technical tips. Consider the
situation in which we have a sequence (an ) and a candidate limit L, and we wish to show
that an → L, that is, that it satisfies the ε-N condition in 6.6.
(1) We require N such that
n > N =⇒ |an − L| < ε.
We do NOT need
n > N ⇐⇒ |an − L| < ε.
This means we do not need to find the smallest N possible when establishing conver-
gence. Any N that works will do. This allows us in many cases to simplify calculations
by replacing complicated expressions by simpler ones before trying to write down a
suitable N .
(2) Beginners often give back-to-front arguments when seeking to prove that a sequence

(an ) tends to some limit L. Note the direction of the implication signs in Examples
6.8. Reading downards, we are working towards finding a suitable N . Once such an N
has been identified, the argument can be re-presented, going from bottom to top and
using forward implication
√ signs. Let’s carry this out for the sequence (an ) in 6.8(b),
where an = ((−1)n / n), treating the original presentation as rough work. Take ε > 0.
Choose N ∈ N with N > 1/ε2 . Then
1
n > N =⇒ n > 2
ε
1
=⇒ < ε2
n 
n 1 
 
=⇒ (−1) √  < ε

n
=⇒ |an − 1| < ε.
In the examples in 6.8 most instances of ⇐= could be replaced by ⇐⇒, but not all
can be (see (d)); (1) says that none needs to be.
(3) We have asked that N ∈ N. But it’s good enough to find X ∈ R such that n > X
implies |an − L| < ε. If X exists then we can choose N > X and N ∈ N (since N is not
bounded above, as we proved in 4.13(i)).
(4) The smaller ε is, the greater challenge we have to find a corresponding N in general.
Turning this around, we see that in establishing convergence we may without loss of
generality restrict to values of ε such that ε < 1 (or ε < η, where η is some fixed positive
number). We did in this in 6.8(c)). It’s small values of ε that matter.
(5) Facility with inequalities is a valuable skill!
It is already clear from Examples 6.8(c),(d) that finding an explicit N (or X) for a given
ε > 0 can be tiresome and messy. The following result is elementary. It is useful in two
ways: it allows us
• to simplify ε-N proofs;
• to take advantage of known limits to find the limiting values of other sequences.

6.10. Sandwiching Lemma (simple form). Let (bn ) and (cn ) be real sequences. Assume
cn → 0 and that 0 6 bn 6 cn for all n. Then bn → 0.

Proof. Let ε > 0 and pick N so that |cn − 0| < ε for all n > N . Then, for n > N ,
−ε < 0 6 bn 6 cn = |cn | < ε.
24

and so |bn − 0| < ε for all n > N . 

[What this says is that, given ε, an N that works for (cn ) also works for (bn ).]

6.11. Examples, by sandwiching.

(a) [Example 6.8(c) revisited] Note that (n − 1)2 > 0, so that n2 − n + 1 > n. Hence
1 1
0< 2 6 → 0 as n → ∞.
n −n+1 n
Now apply the Sandwiching Lemma.

(b) 0 6 2−n 6 1/n for all n (by induction) and 1/n → 0. Hence (2−n ) converges to 0 by
the Sandwiching Lemma.
n
(c) Let an = √ . Then an → 1. To prove this, note that
n2 + 1
√ √
n2 + 1 − n n2 + 2n + 1 − n (n + 1) − n 1 1
0 6 |an − 1| = √ 6 √ = √ =√ < .
n2 + 1 n2 + 1 n2 + 1 n2 + 1 n
Now apply the Sandwiching Lemma, once again using the fact that 1/n → 0.
(d) [Example 6.8(d) revisited] The proof given earlier is a sandwiching argument from
scratch. The Sandwiching Lemma can be applied with bn = |an | and cn = 1/n.

6.12. More examples (two important limits, employing some useful techniques).
(a) Let |c| < 1, where c is constant. We claim cn → 0. We can write |c| = 1/(1 + y) where
y > 0. Take ε > 0. By Bernoulli’s inequality,
1 1 1
|c|n = n
6 < < ε if n > N,
(1 + y) 1 + ny ny
where N ∈ N is chosen such that N > 1/(yε).
n
(b) Let an = . We surmise that an → 0. Let ε > 0. Now
2n
n n n
|an − 0| = n = n
= n

(by the binomial theorem)
2 (1 + 1) 1 + n + 2 + ··· + 1
2n n


6 (if n > 2, by retaining the 2
term)
n(n − 1)
<ε provided n − 1 > 2/ε.
So we choose N ∈ N so that N > 1 + 2/ε.
Now we give a significant theoretical result. The proof gives an illustration of working
with the ε-N definition of convergence.

6.13. Theorem (uniqueness of limits). Let (an ) be a sequence and suppose that an → L1
and an → L2 as n → ∞. Then L1 = L2 .

Proof. Suppose L1 6= L2 . Take ε := |L1 − L2 |. Then ε > 0. So ε/2 > 0 and hence there exist
N1 and N2 such that
n > N1 =⇒ |an − L1 | < ε/2,
n > N2 =⇒ |an − L2 | < ε/2.
 25

Then for n > max(N1 , N2 ) we have

|L1 − L2 | = |(L1 − an ) + (an − L2 )|
6 |L1 − an | + |an − L2 | (by Triangle Law)
< ε/2 + ε/2 = |L1 − L2 |
which is the required contradiction. 

The next group of results concerns the interaction of limits with modulus and inequalities.

6.14. Proposition (limits and modulus). Assume that (an ) is a sequence which converges
to L. Then (|an |) converges too, to |L|.

Proof. By the Reverse Triangle Law, 2.6(2),

||an | − |L|| 6 |an − L|.
Now apply the Sandwiching Lemma, or argue directly from the convergence definition. 

6.15. Limits and inequalities. Let (an ) and (bn ) be real sequences.
Preservation of weak inequalities: Assume that an → L and bn → M and that an 6 bn
for all n. Then L 6 M .

Proof. We argue by contradiction. Assume L

M . Then (by trichotomy) L > M . Let
ε := (L − M )/2. We now can find N1 and N2 such that
∀n > N1 |an − L| < ε,
∀n > N2 |bn − M | < ε.
Then for n > max(N1 , N2 ),
L − ε < an 6 bn < M + ε.
Hence L − M < 2ε, a contradiction. 

An example of non-preservation of strict inequalities. Let an = 1/n. Then an > 0 

for all n, and (by the Archimedean Property), an → 0. So lim an > 0 is false.

6.16. A general Sandwiching Lemma. Assume that (xn ), (yn ) and (an ) are real sequences
such that xn 6 an 6 yn for all n. Assume that lim xn = lim yn = L. Then (an ) converges
to L.

Proof. (Outline) Given ε > 0 we can find N such that for all n > N we have |xn − L| < ε
and |yn − L| < ε. Then, for n > N ,
L − ε < xn 6 an 6 yn < L + ε,
so |an − L| < ε. 

The next result shows us that any convergent sequence has a special property, that of
being bounded. The result will be useful in some technical proofs later, and also, in its
contrapositive form, provides a way to show that certain sequences fail to converge.
 26

6.17. Proposition (a convergent sequence is bounded). Assume (an ) converges. Then

∃M ∈ R ∀n |an | 6 M.
(This says that (an ) is bounded, meaning that its set { an | n = 1, 2, . . . } of terms is
bounded.)
If (an ) is not bounded then (an ) diverges.

Proof. Assume an → L. Take ε = 1. Then there exists N such that

n > N =⇒ |an − L| < 1
=⇒ |an | 6 |L| + 1 (by Triangle Law).
(This says (an ) has a bounded tail.) Hence
∀n |an | 6 M := max{|a1 |, . . . , |aN −1 |, |L| + 1}.

The second statement is just the contrapositive. 

6.18. Examples: divergence.

(a) Unboundedness implies divergence by 6.17. So (2n ) diverges.
(b) Unboundedness not necessary for divergence. Let an = (−1)n . Let ε = 1 and let
L be any real number. Assume there exists N such that n > N implies |(−1)n − L| < 1.
Then
n > N =⇒ −1 < (−1)n − L < 1.
Taking n = 2N gives L > 0 and taking n = 2N + 1 gives L < 0, so we have a
contradiction to the assumption that the sequence converges to L, and this holds for
every L. So the sequence must diverge.
This example also shows that (|an |) convergent need not imply (an ) convergent, So
Proposition 6.14 does not have a converse.

n2
(b) Let an = (−1)n . Then (an ) diverges.
n2 + 1
The result is plausible, but this is an awkward example to handle slickly from first
principles. We’ll give a proof later, once we have developed an efficient method.

6.19. Infinity. Let an be a sequence of real numbers. We say ‘an tends to infinity’ and write
an → ∞ as n → ∞ if
∀M ∈ R ∃N ∈ N ∀n > N an > M.
Similarly we write bn → −∞ if
∀M ∈ R ∃N ∈ N ∀n > N bn < M.
(Here we tend to think of M as being a very large positive/negative number.)
 Here the symbol ∞ provides a convenient notational shorthand. Do NOT treat ∞ as
though it were a real number. Remember that in our convergence definition in 6.6 we
demanded that the limit L belong to R. See further discussion of infinite limits in Section 7.

Examples
(a) Let an = n2 − 1000. Then an → ∞. To prove >0
√ this let M ∈ R . Then an > M for all
( choose N ∈ N such that N > M + 1000—possible by 4.13(i).
n > N if we
n if n is odd,
(b) Let an =
0 if n is even.
 27

If it were true that an → ∞, then (taking M = 1) we could find N such that n > N
implies an > 1. But a2N = 0 and we have a contradiction.

6.20. Important examples: limits involving powers.

(a) Powers of n. In this example we assume the familiar properties of the exponential
function exp and the natural logarithm log; recall that these functions will be defined
and their properties proved later in the year. For α a real number we define nα to be
exp(α log n). We have
(i) if α < 0, then nα → 0 as n → ∞;
(ii) if α > 0, then nα → ∞ as n → ∞.
Proof. (i) Write β = −α so that nα = 1/nβ . Let ε > 0 and assume ε < 1. We have
1/nβ < ε ⇐⇒ nβ > 1/ε ⇐⇒ β log n > − log ε ⇐⇒ n > e− log ε/β .
[Here, by properties of exponentials and logs, it’s just as easy to get ⇐⇒ as to get ⇐=.]
(ii) Note that to check that nα → ∞ it suffices to consider the case when M > 0.
Then
nα > M ⇐⇒ α log n > log M ⇐⇒ n > elog M/α .
This gives the result. 
(b) Powers with exponent n. Let c be a positive constant.
(i) If c < 1 then (cn ) converges to 0.
(ii) If c = 1 then (cn ) converges to 1.
(iii) If c > 1 then cn → ∞.
Proof. For (i), see Example 6.12(a); (iii) can be handled in a similar way, while (ii) is
immediate. Alternatively consider log cn . 

6.21. Complex sequences. As noted already, the definition of convergence and much of
the theory of convergence of real sequences carry over in the obvious way to sequences of
complex numbers. In particular we say a sequence (zn ) converges to L (where now L ∈ C) if
∀ε > 0 ∃N ∈ N ∀n > N |zn − L| < ε.
Moreover, the limit is unique if it exists. Uniqueness is proved as in 6.13.
Recall the remarks in Section 3 concerning complex numbers.
• Inequalities: don’t try to write w < z when w, z ∈ C, not both real !
• The Triangle and Reverse Triangle Laws do hold for complex numbers.
• Simple sandwiching is valid in the following form: suppose (wn ) and (zn ) are complex
sequences such that |wn | 6 |zn | and zn → 0, then wn → 0. [Note the moduli ! ]

6.22. Theorem (convergence of complex sequences). Let (zn ) be a sequence of com-

plex numbers and write zn = xn + iyn , so (xn ) and (yn ) are real sequences. Then (zn )
converges if and only if (xn ) and (yn ) both converge.

Proof. The proof of uniqueness is exactly the same as in the real case.
Now consider the first assertion. For =⇒ use the fact that |xn | 6 |zn | and |yn | 6 |zn |
and sandwiching. The proof of ⇐ is a definition-chase and left as an exercise (ideas from
8.3 are useful). 
28

6.23. Examples: complex sequences.

√ n−i √
(a) Let zn = n2 + 1/(n + i). Then zn = √ . The real sequences (n/ n2 + 1) and
√ n2 + 1
2
(−1/ n + 1) both converge, the former to 1 and the latter to 0. So zn → 1 + 0 · i = 1.
(b) Let zn = in /n. Then |zn | = 1/n and hence zn → 0.
(c) Let zn = in . Then the sequence is
i, −1, −i, 1, i, −1, . . . and Re zn = 0, −1, 0, 1, 0, −1, . . . .
Thus (Re zn ) does not converge and so (zn ) cannot converge.
(d) (cos n) and (sin n) diverge (see Problem Sheet 4). More generally it is poossible to show
that (cos nθ) and (sin nθ) diverge for 0 < θ < 2π. Hence if z ∈ C and |z| = 1 then (z n )
diverges except when z = 1.

7. Subsequences

This short section introduces the notion of a subsequence of a (real or complex) sequence.
The results we obtain here usefully enlarge our armoury of techniques for establishing con-
vergence/divergence. Deeper results involving subsequences are given in Section 10.

When we say that (an )n>1 is a sequence, or that an → L as n → ∞, the symbol n is a

‘dummy’, standing for a strictly positive integer. We are allowed to replace n here with any
other symbol so long as we are consistent, so we might write (am )m>1 or (ak )k>1 or (ar )r>1 to
mean exactly the same as (an )n>1 , while am → L as m → ∞ or ak → L as k → ∞ or ar → L
as r → ∞ mean exactly the same as an → L as n → ∞. On the other hand it is worth
remembering that it makes maths much easier to read when we use the traditional symbol
n or letters close to (or similar to) n in the alphabet in these circumstances, especially if we
want to use the shorthand (an ) or an → L, dropping the subscript n > 1 or the phrase ‘as
n → ∞’.

7.1. Subsequences. Let (an )n>1 be a (real or complex) sequence. Informally, a subsequence
of (an )n>1 is a sequence (br )r>1 whose terms are obtained by taking infinitely many terms
from (an )n>1 , in order. So, for example, if (an ) = (1, 2, 3, 4, . . .) then:
(2, 4, 6, . . .) is a subsequence of (an ) (it is the sequence (a2n )),
(2, 22 , 24 , . . .) is a subsequence of (an ) (it is the sequence (a2n )),
(6, 4 . . .) is not a subsequence of (an ) (terms not in correct order),
(2, 4, 0, 0, . . .) is not a subsequence of (an ) (not all terms are terms of (an )),
(1, 2, 3, 4, . . . , 2015) is not a subsequence of (an ) (not a sequence).

Formally, a subsequence (br )r>1 of the sequence (an )n>1 is defined by a map f : N → N
such that f is strictly increasing (meaning that r < s implies f (r) < f (s)), so that
br := anr , where nr = f (r).
Expressing this another way, we have a infinite sequence of natural numbers
n1 < n2 < n3 < . . .
 29

and the sequence (br ) = (anr ) has terms

an1 , an2 , an3 , . . . .
Points to note:
• Being formal: if (an ) is defined by a function α : N → R (or C) then the subsequences
of (an ) are defined by the functions of the form α ◦ f , where f : N → N is strictly
increasing.
• The variable r used to label the terms of (br ) is a dummy variable; any other variable
name (except n) would have done equally well.
• r 6 nr for all r ∈ N (proved by induction on r).

7.2. Proposition (subsequences of a convergent sequence).

Let (an ) be a sequence.
(i) Assume (an ) converges to L. Then every subsequence (anr ) of (an ) converges, to the
same limit L.
(ii) Assume (an ) has subsequences which converge to limits L and M where L 6= M . Then
(an ) does not converge,

Proof. (i) Take ε > 0. Choose N ∈ N such that n > N implies |an − L| < ε for all n > N .
In particular, nr > N implies |anr − L| < ε. Since r 6 nr this holds whenever r > N . Thus
we have proved that
∀ε > 0 ∃N ∈ N ∀r > N |anr − L| < ε.

(ii) is just the contrapositive of (i) (using uniqueness of limits 6.13). 

7.3. Example (using subsequences to establish divergence). Take

n2
 
n
an = (−1) .
n2 + 1
Then (a2n ) converges to 1 and (a2n+1 ) converges to −1. By 7.2(ii), (an ) does not converge.

8. The Algebra of Limits (AOL)

In Section 6 we concentrated on the limit definition and familiarisation examples, and

explored the relationship between limits and order. We did not consider how limits interact
with arithmetic operations. We now remedy this omission. This section is unashamedly
technical, with lots of ε-N proofs. It supplies machinery needed for working correctly with
limits of sequences, in this course and beyond. [Only a selection of the proofs will be presented
in lectures.]

8.1. epsilon-handling. Recall the definition of an → L as n → ∞:

∀ε > 0 ∃N ∈ N ∀n > N |an − L| < ε.
If we want to prove an → L as n → ∞:–
• we must take an arbitrary ε > 0 and show there exists a corresponding N .
If we know an → L as n → ∞:–
• we can make use of the convergence definition with any ε that suits us.
30

8.2. Facilitating the construction of ε-N proofs: making proofs less fiddly.
(1) As noted earlier, in proving convergence it is sufficient to consider, for example, ε < 1;
it is only small values of ε that we need to consider.
(2) (an ) converges to L if, for some constant K > 0,
∀η > 0 ∃N ∈ N ∀n > N |an − L| < Kη.
Here it is crucially important that K is constant—it must not involve n.
Proof. Take ε > 0. Apply the given condition with η := ε/K to get the standard
condition for convergence. 
(3) Remember that (an ) converges provided it has a convergent tail; recall the Tails Lemma,
6.7. [We needn’t even fuss if we have a sequence whose first few terms are not defined.]
We’ll split the AOL results into two groups. Here are the ones which are easier to prove.

8.3. Theorem: (AOL), Part I.

Assume that (an ) and (bn ) are real or complex sequences and assume that (an ) converges
to L and (bn ) converges to M . Then the following hold as n → ∞.
(i) (constant) If an = a (constant) for all n then an → a.
(ii) (addition) an + bn → L + M .
(iii) (scalar multiplication) can → cL for any constant c.
(iv) (subtraction) an − bn → L − M .
Not strictly algebra, but recalled here for convenience:
(v) an → L as n → ∞ implies |an | → |L| as n → ∞.

Proof. (i) Immediate from convergence definition.

(ii) If c = 0 use (i). Now assume c 6= 0 and let ε > 0. Apply the convergence definition
with ε replaced by ε/|c| to find N such that n > N implies |an − L| < ε/|c|. Then
ε
n > N =⇒ |can − cL| = |c||an − L| < |c| = ε.
|c|
(iii) Choose N1 and N2 so that
ε ε
n > N1 =⇒ |an − L| < and n > N2 =⇒ |bn − M | < .
2 2
Then, by the Triangle Law, n > N := max(N1 , N2 ) implies
ε ε
|(an + bn ) − (L + M )| = |(an − L) + (bn − M )| 6 |an − L| + |bn − M | < + = ε.
2 2
(iv) Use (iii) and (ii) (with c = −1).
(v) Use the Reverse Triangle Law; see 6.14. 

8.4. Examples:(AOL) Part I.

n
(a) 2−n + 1000n−5 + √ → 1 as n → ∞, by (AOL) (addition) and earlier examples.
n2 + 1
(b) Let an = n−2 + (−1)n . Then (an ) diverges.
Proof. Argue by contradiction. Note (n−2 ) converges. If (an ) converged, then ((−1)n )
would converge, by (AOL) (subtraction). 
 31

8.5. Facilitating the construction of ε-N proofs: useful technical facts.

(1) Assume an → 0 and (bn ) is bounded. Then an bn → 0 as n → ∞. (Easy exercise on
Problem sheet 3.)
(2) Assume an → L and L 6= 0. Then there exists N such that n > N implies |an | > |L|/2
(and in particular an 6= 0) and hence
1 2
∃N ∈ N ∀n > N <
|an | |L|
(so (1/|an |) has a bounded tail).
Proof. Apply the convergence definition with ε = |L|/2 to find N such that
1
n > N =⇒ |an − L| < |L|.
2
By the Triangle Law,
1
n > N =⇒ |L| = |L − an + an | 6 |an − L| + |an | < |L| + |an |. 
2

8.6. Theorem: (AOL), Part II. Assume that (an ) and (bn ) are (real or complex) se-
quences and assume that (an ) converges to L and (bn ) converges to M . Then the following
hold as n → ∞.
(vi) (product) an bn → LM .
(vii) (reciprocal) If M 6= 0, then 1/bn → 1/M .
(viii) (quotient) an /bn → L/M if M 6= 0.
[In (vii) and (viii), we may need to restrict to a tail of, respectively, (1/bn ) and (an /bn ) to
get well-defined terms. This will always be possible by 8.5(2).]

Proof. Consider (vi). By arithmetic and the Triangle Law,

|an bn − LM | = |(an − L)(bn − M ) + L(bn − M ) + (an − L)M |
6 |(an − L)(bn − M )| + |L||bn − M | + |an − L||M |.
Fix ε < 1 (no loss of generality) and choose N1 , N2 such that
n > N1 =⇒ |an − L| < ε and n > N2 =⇒ |bn − M | < ε.
Then
n > N := max(N1 , N2 ) =⇒ |(an − L)(bn − M )| < ε2 < ε.
Hence
n > N =⇒ |an bn − LM | < ε(1 + |M | + |L|).
Since 1 + |M | + |L| is a positive constant, we are done, by 8.2(3).
Now consider (vii). Given ε > 0, we can find N1 such that n > N1 implies |bn − M | < ε
and we can find N2 such that n > N2 implies |bn | > |M |/2 (and so also bn 6= 0), by 8.5(2)
applied to (bn ). Then
 
1 1  |M − bn | 2
n > max(N1 , N2 ) =⇒  −  =
 6 ε.
bn M |bn ||M | |M |2
This gives 1/bn → 1/M as n → ∞ (using 8.2(2) again).
To obtain (viii), combine (vi) and (vii). 
32

8.7. Example: a typical (AOL) application.

[A simple example of this type will be discussed in lectures before the technical proofs of
the (AOL) results.]
n2 + n + 1
Let an = . Then
3n2 + 4
 
2 1 1 
1 1

n 1+ + 2
1  1 + n + n2 
 
1 n n 1 1+0+0 1
an = ·  =  → = ,
4

3 4 3 3 1+0 3
n2 1 + 2 1+ 2
3n 3n
by the fact that 1/n → 0 as n → ∞ and the (AOL) results (constant, scalar multiplication,
product, quotient).

8.8. Proposition (reciprocals and infinite/zero limits). Let (an ) be a sequence of

positive real numbers. The following are equivalent:
(a) an → ∞ as n → ∞;
(b) 1/an → 0 as n → ∞.
The proof is a simple definition-chase. Results in 6.20 are special cases.

8.9. More worked examples, (AOL).

(a) [A general fact worth noting] Suppose that
p(x) = xk + ak−1 xk−1 + · · · + a1 x + a0 and q(x) = x` + b`−1 x`−1 + · · · + b1 x + b0
are polynomials. Then 
p(n) 1 if k = `,

→ 0 if ` > k,
q(n) 
∞ if k > `.

Proof. Write p(n)/q(n) in the form nk−` [· · · ]. The quantity in square brackets tends to
1, by the (AOL) results for constants, scalar multiplication, product and quotient. If
k 6 ` the required result then follows by (AOL) for product. For the case k > `, note
q(n)/p(n) → 0, by interchanging the roles of p and q above. Then use Proposition 8.8,
noting that q(n)/p(n) is positive for large n. 
(b) [A useful limit] Let a > 1 and m ∈ N>0 . Then nm /an → 0 as n → ∞. [The sequence
(n2−n )n>1 considered in 6.7(b) is a special case.]
Proof. Write a = 1 + b, where b > 0. Then, by the binomial theorem,
nm nm
= .
1 + · · · + nk bk + · · · + bn


(1 + b)n
All the terms in the denominator are positive so if we drop all but one of these we make
the fraction bigger. We elect to retain the term with k = m + 1, where we assume
n > m + 1. We have
 
n n! n(n − 1) · · · (n − k + 1)
= = .
k (n − k)!k! k!
Then
nm nm
 
(m + 1)!
06 6 .
(1 + b)n bm+1 n(n − 1) · · · (n − m)
The expression on the right-hand side tends to 0: the first term is a constant, and the
second tends to 0 by (a). Now use simple sandwiching. 
 33

8.10. A ‘true-or-false?’ worked example. Let (an ) be a real sequence such that an → ∞
as n → ∞. To decide, with a proof or counterexample as appropriate, whether an /bn → ∞
under each of the following assumptions.
(a) Assume that (bn ) is a bounded sequence with bn 6= 0. Consider an = n and bn = (−1)n .
Then an /bn 6→ ∞.
(b) Assume that (bn ) is a bounded sequence with bn > 0 for all n (or for n > k for some
k). Then an /bn → ∞ is true. There exists K (constant) such that if 0 < bn 6 K, and
so also 1/bn > 1/K, for a tail of (bn ). Since an → ∞, given M , we have an > M K for
large n and hence an /bn > (M K)/K = M .
(c) Assume that (bn ) is a sequence which converges to L > 0. Then an /bn → ∞. Note that
(bn ) is bounded because it is convergent and that bn > 0 for large n by 8.5(2). Now
appeal to (b).
 Don’t expect AOL results to work when infinite limits come into play. Note also Problem
sheet 3, Question 5.

8.11. Orders of magnitude. When looking for a candidate limit for a given sequence (an ),
an obvious strategy is to look for the dominate components in an , which can be expected
to dictate the sequence’s behaviour in the long term. Usually for this we need to appreciate
the relative magnitudes of the terms as n becomes large.
Examples
n3 − 107 n
(a) Let an = . For large n, the dominant term in the numerator of an is n3
n5 + 6n + 1
and that in the denominator is n5 . So we expect the sequence to behave like (n−2 ), and
to have limit 0. Indeed, this is what exactly what an argument using the Algebra of
Limits formalises.

(b) We proved in 6.12(b) by a proper ε-N -argument that 2−n n → 0 as n → ∞.

 You might try to say simply that 2n grows ‘much faster’ than does n, and hence that
an → 0 as n → ∞. But this is MUCH too imprecise to constitute a proof.

√
(c) Let an = sin(nn )/ n. The oscillatory sine function does not assist in getting conver-
n
gence of an to 0 but because | sin x| 6 1 for all x, the
√ rapid growth of n is an irrelevance.
To get the limit, just use sandwiching: |an | 6 1/ n → 0 as n → ∞.
As a rule of thumb, when it comes to the behaviour of functions f (x) for large x:

trig functions & constants < logarithms

< polynomials < positive exponentials & hyperbolic functions.
More precisely:
• |cos n| 6 1 and |sin n| 6 1 for all n.
• For any rational number q > 0, log n/nq → 0 as n → ∞ [discussion of case q = 1 given
as an example in next section].
• For any a > 1 and polynomial p then p(n)/an → 0 as n → ∞.

8.12. The O and o notation. Let (an ) and (bn ) be real or complex sequences. We write
an = O(bn ) as n → ∞ if there exists c such that for some N
n > N =⇒ |an | 6 c|bn |.
34

Note that, if an = O(bn ) and bn → 0 as n → ∞, then an → 0 as n → ∞ too. We write

an = o(bn ) as n → ∞ if an /bn is defined and
an
→0
bn
as n → ∞.

Examples
(a) n + 106 = O (n2 ) is true but n2 = O(n) is false;
(b) n = o (n2 ) is true;
(c) sin n = O (1) is true;
(d) by 8.9(b) when a > 1 and m ∈ N we have nm = o (an ) as n → ∞.
The symbol ≈, used to indicate ‘approximately the same size as’, is not precise (how close 
an approximation is intended?) and so ≈ is outlawed in this course. O, by contrast, has a
precise meaning.
The O notation is useful in particular for simplifying calculations while maintaining rigour.
Given a complicated an , consideration of dominant terms in the expression defining an ) can
help us a simpler expression bn with an = O(bn ).

9. Monotonic Sequences

So far, in order to prove a sequence converges, we have had to identify a candidate limit
at the outset. There is a very important class of real sequences which can be guaranteed to
converge to a limit L ∈ R, or to tend to ∞ or to −∞.

9.1. Monotonic sequences: definitions. Let (an ) be a real sequence.

• (an ) is monotonic increasing if an 6 an+1 for all n;
• (an ) is monotonic decreasing if an > an+1 for all n;
• (an ) is monotonic if it is either monotonic decreasing or monotonic increasing.
[Some authors use the term ‘monotone’ instead of ‘monotonic’.]

9.2. Monotonic Sequence Theorem. Let (an ) be a real sequence.

(i) Assume (an ) is monotonic increasing. Then (an ) converges if and only if it is bounded
above (that is, there exists a finite constant M such that an 6 M for all n).
(ii) Assume (an ) is monotonic decreasing. Then (an ) converges if and only if it is bounded
below.

Proof. We prove (i). Assume (an ) is bounded above. Then S := { an | n ∈ N} is non-empty

and bounded above. By the Completeness Axiom, sup S exists. We shall prove an → sup S.
Take ε > 0. By the Approximation Property for sups (4.8), there exists N such that
sup S − ε < aN 6 sup S.
But then
n > N =⇒ sup S − ε < aN 6 an 6 sup S =⇒ |an − sup S| < ε.

For the converse we use the fact that any convergent sequence is bounded (6.17).
For (ii), note that (an ) is monotonic decreasing and bounded below iff (−an ) is monotonic
increasing and bounded above. 
 35

Points to note
• A real sequence which has a tail which is monotonic increasing converges iff it is bounded
above.
• A real sequence which is monotonic increasing and not bounded above tends to ∞.
Proof. Given M there exists N such that aN > M . But then n > N implies an > M ,
since (an ) is monotonic increasing. 

9.3. Worked examples: Monotonic Sequence Theorem.

(a) An exercise revisited: on Problem sheet 1, Q.5 you were asked to prove that the
sequence (an ) with
 1 n
an = 1 +
n
is monotonic increasing and bounded above by 3. Hence the Monotonic Sequence
Theorem implies that there exists L such that
 1 n
1+ → L 6 3.
n
Also L > an for every n so that, in particular, L > 2. In fact the limit is e, but we need
more information about the exponential function to prove this than we currently have
available.
(b) Let c be a non-negative real constant. Let (an ) be defined by
 
1 c
a1 = 1, an+1 = an + (n > 1).
2 an
(Note that an+1 is defined so long as an 6= 0. A trivial induction
√ shows that this holds
for all n.) We claim that (an ) is convergent to a limit L = c. The proof proceeds in
stages:
(1) Find possible limit(s): IF ∃ L ∈ R such that an → L, then
an+1 → L (tail),
 
1 c 1 c
an + → L+ (by (AOL) (scalar multiple, sum,quotient)).
2 an 2 L
Hence, by uniqueness of limits,
1 c
L= L+ .
2 L
√
This gives L2 = c. Therefore the only possibilities
√ for L are ± c. Since an > 0
for all n, we must have L > 0 so L = − 2 is ruled out.
(2) Compare terms to the candidate limit: Consider
   2
2 1 2 1 1 c
an+1 − c = an + 2c + 2 − c = 2 an − > 0.
4 an 4an an
√
Hence an > c for n > 2, so we conjecture it is monotonic decreasing, heading
downwards to its potential limit.
In this example Stage (2) is particularly simple. Normally an inductive proof would
be needed at this point.
36

(3) Test for monotonicity:

 
1 c 1
a2n − c .


an − an+1 = an − an + =
2 an 2an

By (2), we have an − an+1 > 0 for n > 2.

(4) Putting the pieces together: The sequence (an ) is bounded below by 0 and
(an )n>2 is monotonic decreasing by (3). By the Monotonic
√ Sequence Theorem,
an → L for some (finite) limit L. By (1), L must be c.
√
This example proves the √existence of c. Where this argument differs from the fiddly
one used earlier to show 2 exists is in the availability now of the Algebra of Limits.
Observe that both the new proof and the old one use the Completeness Axiom.

(c) Engineering a recurrence relation: In (b), we started from a sequence given by a

recurrence relation, and used the recurrence relation to identify a possible limit for our
sequence. Here we illustrate how we can sometimes use this idea when we do not have
a recurrence relation given. We consider an = log n/n where log denotes loge . We shall
assume properties of the log function.
Certainly an > 0, so (an ) is bounded below. To show (an ) is monotonic decreasing we
look at the gradient of the function f (x) = log x/x. We have
1 log x
f 0 (x) = − 6 0 (for x > e).
x2 x2
So the function f is decreasing, and in particular f (n + 1) 6 f (n), for n > 3. Applying
the Monotonic Sequence Theorem, there exists L > 0 such that
log n
an = → L.
n
To find L we elect to relate a2n to an (note that we have no useful relationship between
log(n + 1) and log n so it’s not helpful to compare an+1 with an ). We have
log 2n log 2 + log n
a2n = = .
2n 2n
Now we let n → ∞ on both sides: (a2n ) is a subsequence of (an ) and so converges
to L. By (AOL) results, the right-hand side tends to 0 + L/2. By uniqueness of limits,
L = L/2, so L = 0. We have proved that
log n
→ 0.
n

10. The Bolzano–Weierstrass Theorem, and the Cauchy Convergence

Criterion

In this section we obtain important general results about sequences which are not mono-
tonic. Since monotonic real sequences behave so well, the following theorem provides welcome
information.
 37

10.1. The Scenic Viewpoint Theorem. Let (an ) be a real sequence. Then (an ) has a
monotonic subsequence.

Proof. We consider the set V = { k ∈ N | m > k =⇒ am < ak }. This is the set of “scenic
viewpoints” (also known as “peaks”)—given k ∈ V , looking towards infinity from a point at
height ak , no higher point would impede our view.
Case 1: V is infinite. Then the elements of V can be enumerated as k1 < k2 < . . . .
Then (akr ) is a subsequence of (an ) and
r > s =⇒ kr > ks =⇒ akr < aks
that is, (akr ) is monotone decreasing.
Case 2: V is finite. Let N be such that every element of V is < N . Then m1 = N
is such that m1 ∈/ V , so there exists m2 > m1 with am2 > am1 . Since m2 ∈
/ V , there
exists m3 > m2 such that am3 > am2 . Proceeding in this way we can (inductively)
generate a monotonic increasing sequence (amk ). 

10.2. Bolzano–Weierstrass Theorem (for real sequences). Let (an ) be a bounded real
sequence. Then (an ) has a convergent subsequence.

Proof. By the Scenic Viewpoint Theorem 10.1, (an ) has a monotonic subsequence which is
also bounded. By the Monotonic Sequence Theorem, this subsequence converges. 

Neither the Monotonic Sequence Theorem nor the Scenic Viewpoint Theorem extends to
complex sequences, but the Bolzano–Weierstrass Theorem does.

10.3. Bolzano–Weierstrass Theorem (for complex sequences). Let (zn ) be a bounded

sequence in C. Then (zn ) has a convergent subsequence.

Proof. Write zn = xn + iyn . Let M be such that |zn | 6 M for all n > 1. Then by the
definition of modulus in C, we have |xn | 6 M and |yn | 6 M , for all n > 1. Therefore by
the real BW Theorem, (xn )n>1 has a convergent subsequence, say (xnr )r>1 . Let wr := ynr ,
for r > 1. Then the subsequence (wr )r>1 , being a subsequence of the bounded sequence
(yn )n>1 , is itself bounded. Choose a subsequence (wrs )s>1 of (wr )r>1 which is convergent.
Then (ynrs )s>1 converges and so does (xnrs )s>1 , since it is a subsequence of a convergent
sequence. Finally, (znrs )s>1 is convergent since the sequences of real and imaginary parts
converge. 

10.4. Cauchy sequences. The idea behind Cauchy sequences is that if the terms of a
sequence (an ) are ultimately arbitrarily close to one another, then there should be a value
L to which they must converge. This would be valuable to know, since up till now we have
needed to identify a candidate limit in advance (except for bounded monotonic sequences,
where we may be able to find the limit from a given, or constructed, recurrence relation).
Definition. Let (an ) be a real or complex sequence. Then (an ) is a Cauchy sequence if
it satisfies the Cauchy condition:
∀ε > 0 ∃N ∈ N ∀m, n > N |an − am | < ε.
 It is not sufficient in the Cauchy condition just to consider adjacent terms, that is, only
to consider m = n + 1.
38

10.5. Facts about Cauchy sequences.

(1) A Cauchy sequence (an ) is bounded.
Proof. Choose N such that m, n > N implies |an − am | < 1. Thus (considering m = N )
we get |an | 6 |aN | + 1 for n > N . Now
|an | 6 max(|a1 |, . . . , |aN −1 |, |aN | + 1). 
[This argument is very like the proof in 6.17 that a convergent sequence is bounded.]
(2) A convergent sequence is a Cauchy sequence.
Proof. Assume an → L. Take ε > 0 and choose N such that k > N implies |ak − L| <
ε/2. Then, using the Triangle Law, m, n > N implies
ε ε
|an − am | = |(an − L) − (am − L)| 6 |an − L| + |am − L| < + = ε. 
2 2
(3) Let (an ) be a Cauchy sequence and assume that (an ) has a subsequence (anr ) which
converges, to L say. Then (an ) converges to L.
Proof. Take ε > 0 and pick N so that
ε
m, n > N =⇒ |an − am | < .
2
0
Since anr → L, there exists N ∈ N such that
ε
r > N 0 =⇒ |anr − L| < .
2
0
Fix r such that r > max(N, N ). Then nr > N (since nr > r). Then applying Cauchy
condition with m = nr ,
ε ε
n > N =⇒ |an − L| 6 |an − anr | + |anr − L| < + . 
2 2

10.6. Theorem (Cauchy Convergence Criterion). Let (an ) be a (real or complex) se-
quence. Then
(an ) is convergent ⇐⇒ (an ) is a Cauchy sequence.

Proof. =⇒ : By 10.5(2).
⇐=: By 10.5(1) and the BW Theorem, 10.2 or 10.3, (an ) has a convergent subsequence.
Now appeal to 10.5(3). 

Applications of Theorem 10.6 will come later.

11. Convergence of Series

11.1. Series: introductory examples.

(a) Infinite geometric progressions: for r 6= 1,
1 − rn+1
1 + r + r2 + · · · + rn =
1−r
and, if |r| < 1,
1
lim 1 + r + r2 + · · · + rn =


.
n→∞ 1−r
 39

(b) Decimal expansions: for

1 3 3 3
= 0·333333333333333 · · · , meaning + + + ··· .
3 10 100 1000
Decimal representations exist for all real numbers. See Supplementary Notes for an
account, and a proof via decimals that R is uncountable.
1 1 1
(c) e = 1 + + + + · · · by definition.
1! 2! 3!
(d) Functions given by (or defined by?) infinite sums: for example,
1 1 1
ex = 1 + x + x2 + x3 + · · · .
1! 2! 3!

11.2. Series: definitions. Let (ak ) be a real or complex sequence. Let

n
X
s n = a1 + a2 + · · · + an = ak (n > 1).
k=1
P P
Then we say that the series k>1 ak converges (or as shorthand that the series ak
converges)
P∞ if the sequence (sn ) of partial sums P converges. If sn → s, then we write
a
k=1 k = s. If (s n ) fails to converge, then we say k>1 ak diverges.
A trivial but important observation:
an = sn − sn−1 for n > 2;

Note on notation: We have two sequences in play at the same time here: the sequence
(ak ) of the terms
P of the series and the sequence (sn ) of partial sums. To avoid writing
(erroneously!) nn=1 an , we have introduced a new (dummy) variable k to label the terms.

11.3. Convergence of series: first examples.

(a) Geometric series: Let z ∈ C. Let ak = z k , so
n

 z(1 − z )
if z 6= 1,
s n = a1 + · · · + an = 1−z
n if z = 1.
Hence k>1 z k converges if |z| < 1 and fails to converge if |z| > 1 (recall 6.23). (Note:
P
it might P
have been more natural here to start from k = 0 rather than k = 1 and to
consider k>0 z k .)
(b) A telescoping series: Let ak = 1/(k(k + 1)). Then
n n 
X 1 X 1 1 
sn = = −
k=1
k(k + 1) k=1 k k + 1
 1 1 1 1 1  1
= 1− + − + ··· + − =1− → 1.
2 2 3 n n+1 n+1
P 1
Hence k>1 converges.
k(k + 1)
Note that we are working here with finite sums of real numbers and all the normal
rules of arithmetic apply, in particular associativity of addition. This is always the case
when we work with partial sums.
40

(c) Let ak = (−1)k . Then (

0 if n is even,
sn =
−1 if n is odd.
(−1)k diverges. But, naively,
P
Hence
X∞
(−1)k = (−1) + 1 + (−1) + 1 + (−1) + 1 + . . .
k=1
= ((−1) + 1) + ((−1) + 1) + · · · = 0 + 0 + · · · = 0.
This is clearly wrong. The error lies in the implicit assumption that infinite sums satisfy 
the same arithmetic properties as finite sums do. They don’t, in general.
P
Deciding whether a series k>1 ak converges requires us to decide whether the sequence
Pn

(sn ) = k=1 ak converges. We can bring everything we know about convergence of se-
quences to bear on this problem. This leads immediately to a clutch of results sufficiently
important to be recorded as theorems.

11.4. Theorem (terms of a convergent series).

P
(i) Assume k>1 ak converges. Then ak → 0 as k → ∞.
P
(ii) A sufficient condition for k>1 ak to diverge is that ak 9 0.

Proof. (i) Assume sn → s. For n > 2 we have an = sn − sn−1 → s − s = 0.

(ii) is just the contrapositive of (i). 

1 P
11.5. Example: the harmonic series, . Let
k>1
k
1 1 1
sn = 1 + + + · · · + (n > 1).
2 3 n
We claim (sn ) is not a Cauchy sequence, so (sn ) does not converge and therefore
X1
diverges.
k>1
k

Proof. Consider
1 1
|s2n+1 − s2n | = + · · · + n+1
+12n 2
1
> n+1 · (2n+1 − 2n ) ((smallest term) × (number of terms))
2
1
= .
2
Therefore (sn ) is not Cauchy, so fails to converge (by 10.6)). 
P P 1
This example shows also that ak → 0 does not imply k>1 ak converges: k>1 provides 
k
a counterexample.
Later (12.13(c)) we’ll see that sn grows about as fast as log n, in that sn − log n tends to
a finite constant.

A particularly amenable class of series will be those whose partial sum sequences are
monotonic, thanks to the Monotonic Sequence Theorem.
 41

11.6. Theorem (series of non-negative terms). Assume ak is real. Then

(sn ) is monotonic increasing if and only if ak > 0 (for k > 2).
P
Moreover, if ak is non-negative, then k>1 ak converges if and only if its partial sum sequence
(sn ) is bounded above.

Proof. By the fact that sn − sn−1 = an (n > 2) and Monotonic Sequence Theorem, 9.2. 

11.7. Theorem (Comparison P Test, simple form). P Assume 0 6 ak 6 Cbk , where

P C is
a positive constant.
P Then b
k>1 k convergent implies a
k>1 k convergent. Also k>1 ak
divergent implies k>1 bk divergent

Proof. Let
sn = a1 + · · · + an and tn = b1 + · · · + bn .
Then sn 6 Ctn for all n. Since (tn ) converges, it is bounded above, by T say. Hence sn 6 T
for all n, and Theorem 11.6 applies. 

11.8. Examples: simple Comparison Test.

(a) Take ak = 1/k 2 and bk = 1/(k(k + 1)). Then 0 6 ak 6 2bk and we proved earlier that
P P 1
k>1 bk converges. Hence k>1 2 converges.
k
P
(b) Let ak = 1/k! and bk = 1/(k(k P + 1)). Then 0 6 ak 6 bk and hence the series k>1 1/k!
converges by comparison with P k>1 1/(k(k + 1)). Hence the series defining e converges
(note this series is not exactly k>1 ak because the latter has one fewer term at the
front but this does not affect convergence).

So far in considering series we have restricted attention to

• series whose partial sums we can compute explicitly (for example, geometric or tele-
scoping series);
• series of non-negative terms, whose partial sum sequences are monotonic increasing.

In general, given a series k>1 ak we won’t be able to calculate sn = nk=1 ak explicitly.

P P
Indeed we may want, as with e, to use the sum of a series to define a number and in such
cases we won’t have a convenient formula for the partial sums. The Cauchy Criterion, 10.6,
provides a way of testing a sequence for convergence without knowing a candidate limit. We
now apply this to the partial sum sequence of a series.

11.9. Theorem (Cauchy criterion for convergence of P a series). Let (ak ) be a real or
complex sequence with partial sum sequence (sn ). Then k>1 ak converges if and only if

∀ε > 0 ∃N ∈ N ∀n > m > N |am+1 + · · · + an | = |sn − sm | < ε.

11.10.
P Absolute convergence. Let (aP k ) be a sequence of real or complex numbers. We
say k>1 ak converges absolutely if k>1 |ak | converges.
P
Note that k>1 |ak | is a series of non-negative terms, to which Theorem 11.6 applies. Thus
the following theorem is very useful.
42

11.11. Theorem (absolute convergence implies convergence). Let (ak ) be a real or

complex sequence. Then
X X
|ak | converges =⇒ ak converges.
k>1 k>1
Pn Pn
Proof. Let sn = k=1 ak and Sn = k=1 |ak |. For n > m,
|sn − sm | = |am+1 + · · · + an | 6 |am+1 | + · · · + |an | = |Sn − Sm |,
by the Triangle Law (as it extends by induction to finite sums). Hence
X
|ak | convergent =⇒ {Sn } is a Cauchy sequence (Cauchy Criterion, easy direction)
k>1
=⇒ {sn } is a Cauchy sequence (from above)
=⇒ (sn ) converges (Cauchy Criterion, harder direction). 
3 2
3k sin (k ) 1
Example Let ak = (−1) Here we do not have ak > 0. But 0 6 |ak | 6 2 . Hence
.
k2 + 1 k
1/k 2 and so k>1 ak converges too.
P P P
k>1 |ak | converges by comparison with

11.12. A standard example: the series k>1 k −p for p ∈ R.

P

−p
P
(a) k>1 k diverges if p 6 1;
Proof. For p 6 0, the terms do not tend to 0. For p = 1 we have the harmonic series.
For 0 < p < 1, use the contrapositive of the simple Comparison Test, with ak = k −1
and b = k −p . 
P k −p
(b) k>1 k converges if p > 2.
Proof. Use simple Comparison Test, comparing with k>1 1/k 2 , which we proved ear-
P
lier is convergent.P 
(c) [Looking ahead] k>1 k −p converges if 1 < p < 2.
Proof deferred until we have available the Integral Test, which allows us to handle
all values of p > 0 in a uniform way; see 12.13(a).
In summary:

k −p diverges if p 6 1 and converges if p > 1.

P

12. Convergence Tests for Series

In this section we shall obtain some very useful tests for convergence/divergence of series.
For testing series of non-negative terms and so for testing for absolute convergence
(from which convergence follows):
• Comparison Test (limit form);
• D’Alembert’s Ratio Test;
• Integral Test.
For testing series (−1)k−1 uk , where uk > 0:
P

• Leibniz’ Alternating Series Test.

 43

We shall present the Alternating Series Test first. It is quite special and there are close
connections with material in the previous section.

(−1)k−1 uk converges if
P
12.1. Leibniz’ Alternating Series Test. The series
(i) uk > 0,
(ii) uk+1 6 uk ,
(iii) uk → 0 as k → ∞.

Proof. Consider the subsequence (s2n ). We have

s2n = u1 − u2 + u3 − u4 + · · · + u2n−1 − u2n
= u1 − (u2 − u3 ) − · · · − (u2n−2 − u2n−1 ) − u2n 6 u1 ,
by (i) and (ii). Also, by (ii) again,
s2(n+1) − s2n = u2n+1 − u2n+2 > 0.
Hence (s2n ) is monotonic increasing and bounded above, so converges, to s say, by the
Monotonic Sequence Theorem. Now consider (s2n−1 ). We then have, using (iii) and (AOL),
s2n−1 = s2n − u2n → s − 0 = s.
Hence (s2n ) and (s2n−1 ) converge to the same limit. Hence (see Problem sheet 4, Q.2(a)),
(sn ) converges. 

12.2. Examples: AST.

k−1 1
P
(a) By AST, k>1 (−1) converges.
k
Here
Pwe have an example of a series which converges but fails to converge absolutely,
since k>1 1/k diverges (recall 11.5).
P k−1 √1
(b) By AST, k>1 (−1) converges. This is a useful series for counterexamples.
k

12.3. Comparison Test, simple form (recap of 11.7 and remarks).

Assume ak , bk are real numbers and that
0 6 ak 6 Cbk for all k, where C > 0 is a constant.
Then (P P
ak convergesif bk converges;
P P
bk diverges if ak diverges.
P
Suppose we wishPto use the Comparison Test to prove that ak is convergent and have
a candidate series bk with which to compare. In practice it can be awkward to get a valid
inequality ak 6 Cbk so we can apply the test: consider for example
k2 + k + 1
ak = .
4k 4 − k 2 − 1
We’d want to try taking bk = 1/k 2 , but valid applications of the Triangle Law (on the
numerator of ak ) and Reverse Triangle Law (on the denominator of ak ) don’t make it very
easy to find a suitable C. Fortunately there’s a version of the Comparison Test which is very
easy to apply, without the need to manipulate inequalities.
44

12.4. Comparison Test, limit form. Let ak and bk be strictly positive and assume that
ak
→ L, where 0 < L < ∞.
bk
Then
P P
ak converges ⇐⇒ bk converges.

Proof. In the limit definition take ε = L/2 and choose N such that, for k > N ,
 
 ak
 − L < 1 L and hence 1 L < ak < 3 L.

 bk  2 2 bk 2
P P
Then, restricting to tails, bk convergent implies that aPk is convergent, by the simple
P
Comparison Test with C = 3L/2. In the other direction, ak convergent implies bk
−1
converges by comparison, because 0 < bk < 2L ak for large k. 

Note: It is crucial that the terms ak and bk are (ultimately) strictly positive and that L is 
non-zero and finite.

12.5. Examples: Comparison Test (limit form).

k2 + k + 1
(a) Take ak = 4 − k2 − 1
. Take bk = k −2 and use (AOL) to get ak /bk → 1/4. Hence
P 4k P
ak converges because bk does.
1
−k 1
(b) Let ak = 1 + k1 . Take bk = . Then
k k
 −k
ak 1
= 1+ → e−1
bk k
P P
(standard limit). So ak diverges by comparison with the divergent series bk .

The Comparison Test in either form is not ‘internal’, in that we have to produce a suitable
series with which to compare. We next obtain some important tests which don’t have this
disadvantage.

12.6. D’Alembert’s Ratio Test, for series of strictly positive terms. Let ak > 0.
Assume that
ak+1
lim exists and equals L.
k→∞ ak
Then
P
06L<1 =⇒ ak converges;
P
L>1 =⇒ ak diverges;
L=1 =⇒ the test gives no result.
(Here it is permissible, and useful, to allow L = ∞ as a possible limit and to treat it as > 1
for the purposes of stating the test.)

Proof. The idea will be to compare with a suitable geometric series.

Case 1: L < 1. Fix K such that L < K < 1 and apply the limit definition with ε = K − L.
Then there exists N ∈ N such that
 
 ak+1  ak+1
k > N =⇒  − L < K − L and hence < L + (K − L) = K.
ak ak
 45

So, for k > N , we get

0 < ak+1 < Kak .
By induction, we obtain, for k > 1,
0 < ak+N < aN K k .
Hence,
P k by the simpleP Comparison Test, comparing with the convergent geometric series
K , the series ak has a convergent tail and so converges.
Case 2: L > 1. Assume first that L is finite. Then pick K such that L > K > 1 and apply
the limit definition with ε = L − K. Then there exists N such that k > N implies
 
 ak+1 
 ak − L < L − K
 

and hence, by the Reverse Triangle Inequality,

ak+1
> L − (L − K) = K,
ak
If L = ∞, then we can, by definition of an infinite limit, certainly find a constant K such
that am+1 /am > K for all suitably large m.
Now, using an argument similar to that in Case 1, we get, for some N ,
ak+N > aN K k for all k > 1.
P
This shows that ak 6→ 0 so ak diverges. 

12.7. Ratio Test, preliminary examples.

P 2k
(a) Consider ak , where ak = . Here we have ak > 0 and
k!
ak+1 k! 2k+1 2
lim = lim k
= lim = L, where L = 0.
ak (k + 1)! 2 k+1
Hence the series converges,
P 1
(b) Don’t forget to take the limit! Consider ak where ak = p , where p > 0. Here
k
we have
  p
ak+1 k
lim = lim = 1.
ak k+1
Hence the Ratio Test, correctly applied, gives no result.
Consider ak = 1/k (the case p = 1). Here we have
ak+1 k
= < 1.
ak k+1
P
Note that we already know that 1/k diverges. Hence having ak+1 /ak < 1,Prather
than lim ak+1 /ak = L < 1, does not give a sufficient condition for a series ak of
positive terms to converge.

12.8. Ratio Test: tips and warnings.

P
 • It is quite possible that lim ak+1 /ak fails to exist: consider for example ak given by
1 1 1 1 1 1
1+1+ + + 2 + 2 + 3 + 3 + ....
3 2 3 2 3 2
(This series does converge: can you prove this by the simple Comparison Test?)
46

• For a series with ‘gaps’, leave out the zero terms and relabel before applying the test.
For example, consider

 1
if k is of the form 2m for some m > 1,
ak = 2k!
0 otherwise,
P
so ak (starting from k = 1) looks like
1 1 1
0 + 1! + 0 + 4! + 0 + 0 + 0 + 8! + 0 + . . . .
2 2 2
m
We cannot apply the Ratio Test directly to k>1 ak but we can apply it to m > 11/2(2 )! .
P P

rk is convergent/divergent by applying the 

P
• We cannot prove that a geometric series
Ratio Test. Why not?

12.9. Testing for absolute convergence: a corollary to the Ratio Test 12.6. Let ak
be non-zero real or complex numbers and assume that
 
 ak+1 
lim   exists and equals L.
k→∞  ak 

(Here we allow L = ∞.) Then

P
L<1 =⇒ ak converges absolutely, and hence converges;
P
L>1 =⇒ ak diverges;
L=1 =⇒ the test gives no result.
P
Proof.
P For the case L < 1 we apply the Ratio Test 12.6, to |ak |. If L > 1, again considering
P
|ak |, we note that the proof given above implies that |ak | 9 0, so ak 9 0 and hence ak
diverges. 

The (corollary to the) Ratio Test is particularly useful for testing series
X X
ck xk (x real) and ck z k (z complex)
for convergence. Here we regard x or z as a variable, so that the series, provided it is
convergent, will define a function. The whole of Section 13 is devoted to such power series
and their properties. There we give further examples of the use of the Ratio Test in important
particular cases.

12.10. The Integral Test: preamble. P Here we shall analyse the the behaviour of the
partial sum sequence (sn ) of a series f (k) by comparing it with the sequence (In ) where
Z n
In = f (x) dx,
1

where f : [1, ∞) → [0, ∞) is a suitable function. We shall need to make use of standard
properties of integrals (integration is treated in Analysis III in Trinity Term). We assume
the following. On a suitable class of integrable functions:
(a) integration preserves 6;
R k+1
(b) k c dx = c, for any constant c;
Rc Rb Rc
(c) intervals slot together: a = a + b for a < b < c.
 47

12.11. Integral Test Theorem. Assume that f is a real-valued function defined on [1, ∞)
with the following properties:
(i) f is non-negative and decreasing;
R k+1
(ii) k f (x) dx exists for each k > 1 [for future reference: this holds if f is continuous].
Let
n
X Z n
sn = f (k) and In = f (x) dx.
k=1 1

Let σn = sn − In . Then (σn ) converges to a limit σ, where 0 6 σ 6 f (1).

Proof. Note that, because f is decreasing, f (k) > f (x) > f (k + 1) for all x ∈ [k, k + 1]. Now,
by properties (a) and (b) above,
Z k+1
f (k + 1) 6 f (x) dx 6 f (k) (k = 1, 2, . . .) .
k

So we have
Z 2
f (2) 6 f (x) dx 6 f (1)
1
Z 3
f (3) 6 f (x) dx 6 f (2)
2
··· ··· ···
Z n
f (n) 6 f (x) dx 6 f (n − 1).
n−1

Add these inequalities and use property (c) to get

Xn n−1
X Z n n
X
f (r) − f (1) = f (k + 1) 6 f (x) dx 6 f (k) − f (n).
r=1 k=1 1 k=1

Then
n
X Z n
0 6 f (n) 6 f (k) − f (x) dx 6 f (1),
k=1 1

so that 0 6 σn 6 f (1). Also

Z n+1
σn+1 − σn = f (n + 1) − f (x) dx 6 0.
n

Therefore (σn ) is monotonic decreasing and bounded below, and hence converges to a limit σ,
where 0 6 σ 6 f (1). 

 Some sources work with a different choice of monotonic sequence.

12.12. Corollary: the Integral Test. Assume that (as in 12.11), f : [1, ∞) → [0, ∞) is
R k+1 P
monotonic decreasing and such that k f (x) dx exists for each k. Then f (k) converges
if and only if (In ) converges.

Proof. This is immediate from elementary properties of limits and the fact that (sn − In )
converges. 
48

12.13. Applications of the Integral Test and Integral Test Theorem. Here we assume
not just properties of integrals but also how to evaluate standard integrals.
P −p
(a) k (p > 0), definitively. Take f (x) = x−p . Then f is non-negative and decreasing
on [1, ∞). We have
 1 1

 −p+1 n
(n−p+1 − 1) if p 6= 1,
Z n
−p x 1
=
In = x dx = −p + 1 −p + 1
1 log n if p = 1.
P −p
Hence (In ) converges (to a finite limit) iff p > 1. Therefore, as claimed in 11.12, k
converges
P −p if p > 1 and diverges if 0 < p 6 1. [If p < 0 the Integral Test does not apply
but k diverges because k −p 9 0.]
(b) The Integral Test applied with f (x) = 1/(x log x) (note that the conditions for the test
are met, except that we need to start from x = 2 rather than x = 1) implies that
P
k>2 1/(k log k) diverges. This is, perhaps, a little surprising.

(c) Euler’s constant, γ. Apply the Integral Test Theorem in the special case that f (x) =
1/x; certainly f is non-negative and monotonic decreasing, and its integral exists over
any interval [k, k + 1]. We have that
1 1 1
γn := 1 + + + · · · + − log n → γ,
2 3 n
where γ is a constant between 0 and 1. This shows that the partial sums of the divergent
harmonic series tend to infinity slowly—about as fast as log n. The constant γ, known
as Euler’s constant, is rather mysterious: it remains unknown whether γ is rational or
irrational.
1 1 1
(d) A series for log 2. Let sn = 1 − + + · · · + (−1)n−1 . Then
2 3 n
1 1 1
s2n = 1 − + + · · · −
 2 3 2n   
1 1 1 1 1 1
= 1 + + + ··· + −2 + + ··· +
2 3 2n 2 4 2n
= (γ2n + log(2n)) − (γn + log n)
= log 2 + γ2n − γn
→ log 2.

12.14. Example: exploiting the existence of Euler’s constant. [details omitted from
lectures] Consider
1 1 1 1
1 − + − + − ... and
2 3 4 5
1 1 1 1 1 1 1
1+ + − + + + − + ...,
3 5 2 7 9 11 4
so the second series contains the same terms as the first, but in a different order. We analyse
the limiting behaviour of the second series. The terms come in groups of four, three positive
terms followed by one negative one, so we first look at the sum of 4n terms, in n groups:
     
1 1 1 1 1 1 1 1 1 1 1
s4n = 1 + + − + + + − + ··· + + + −
3 5 2 7 9 11 4 6n − 5 6n − 3 6n − 1 2n
     
1 1 1 1 1 1 1 1 1
= 1 + + + ··· + − + + ··· + − + + ··· +
2 3 6n 2 4 6n 2 4 2n
1 1
= (γ6n + log(6n)) − (γ3n + log(3n)) − (γn + log n)
2 2
 49

36n2
 
1 1 1
= log + γ6n − γ3n − γ2n
2 3n2 2 2
1 1
→ s := log 12 = log 2 + log 3.
2 2
 So far we have only looked at a particular subsequence of (sn ). Not good enough! But it is
easy to remedy the omission:
1
s4n+1 = s4n + → s,
6n + 7
1
s4n+2 = s4n+1 + → s,
6n + 9
1
s4n+3 = s4n+2 + → s.
6n + 11
We deduce that sn → s = log 2 + 21 log 3 (using an extension of the result of Problem sheet
4, Q.2(a)).
 So, by changing the order P
of the terms we have changed the value of the sum! We showed
that the alternating series (−1)k−1 1/k has sum log 2, whereas the ‘3 pluses, 1 minus’
rearrangement has sum log 2 + 21 log 3.

: N>0 → N>0 be a
P
12.15. Definition: rearrangement.P Take any series ak and let gP
bijection. Let bk = ag(k) . Then bk is said to be a rearrangement of ak .

Facts about rearrangement

P P
(1) Theorem: If ak is absolutely convergent then all rearrangements of ak converge,
and the value of the sum is not changed by rearrangement.
P
(2) If ak is convergent but not absolutely convergent, then rearrangement may result in
a series converging to a different sum, or to one which diverges. (See, for example, Scott
& Tims, Mathematical Analysis, for an example; the idea is quite simple—we aim to
rearrange so that the positive terms predominate strongly in the partial sums of the
rearranged series.

12.16. Which test? A miscellany of examples in outline. In all cases, restriction to

tails is permissible or necessary.
X log(k 2 + 1)
(a) (−1)k−1 uk , where uk = √ .
k+5
√
Apply AST. Consider the derivative of f (x) = log(x2 + 1)/ x + 5 to show (uk ) is
monotonic decreasing and sandwiching techniques to show uk → 0.
X 1
(b) √ .
k2 + k
Comparison Test, limit form, works easily. (Ratio Test gives no result here.)
X 1
(c) .
k log k(log log k)2
A classic case for the Integral Test.
X (4k)!
(d) .
(k!)2
Slog it out with the Ratio Test.
50

12.17. Tips for finding counterexamples.

(1) Sometimes you are given a statement that looks rather like one you’ve seen before. Ask
yourself if all the conditions needed to make the statement true have been included. If
not, you’re looking for a counterexample which fails to satisfy some missing condition
(such as non-negative terms).
(2) Series which serve as counterexamples to plausible assertions often ‘only just’ diverge
or converge. The following can be useful: P
• series which
P diverge even more slowly than 1/k: an example is provided by (a
tail of) 1/(k log k); √
• alternating series such as (−1)k−1 / k—the terms are large in modulus, and
P
(non-absolute) convergence occurs because of the alternating signs.

12.18. A ‘true-or-false?’ worked example. There are many conjectures one might make
about convergence of series in general. A few, some true, some false, can be decided even on
our limited treatment of series, and limited examples, thus far.
P P
(a) For ak , bk real, ak 6 bk and bk convergent implies ak convergent.
FALSE: Consider for example ak = −1, bk = 0. Comparison Test needs non-negative 
terms.
P
(b) ak > 0 and kak → 0 implies ak converges.
FALSE: Take for example ak = 1/((k + 1) log(k + 1)).
P
(c) If ak is a convergent series of positive terms then 0 < ak < 1/k for k sufficiently
large.
FALSE: Take (
1/m2 if k = 2m for some m,
ak =
0 otherwise.
1/k 2 ). But, for
P P
Then ak converges (its partial sum sequence behaves like that for
k = 2m ,
2m
|kak | = 2 ,
m
and this is not bounded by 1 for large k—in fact it tends to infinity, as 8.9(b) shows.
Morals from (c):
(1) Series with gaps can be useful for counterexamples.
1/k p does not provide a counterexample here. It is naive to think that
P
(2) A series
behaviour of series of this special form is typical.

12.19. Postscript: other tests for convergence. Not every series you may meet can be
handled by applying directly one of the tests presented above. In particular you should be
aware that the Ratio Test is rather crude and quite often fails to give a result.
The list of tests for convergence, or for absolute convergence, that we have given is far
from exhaustive, and many other convergence tests exist. We note in particular
• Cauchy’s nth Root Test and Raabe’s Test—both useful alternatives to the Ratio
Test, the first because it works neatly on many power series (see Section 13) and the
second because it provides a backstop to the Ratio Test, giving a result in a number of
cases where the Ratio Test does not;
• Cauchy’s CondensationPTest Under the same conditions P mon f mas in the Integral
Test, the test asserts that f (k) converges if and only if 2 f (2 ) converges.
 51

• Abel’s Test and Dirichlet’s Test, for conditional convergence.

Statements and proofs can be found in various analysis textbooks and no doubt on the
internet too.
You will not encounter series in Prelims for which any one of these additional tests is
needed, but they could nevertheless provide you with techniques you might sometimes find
useful, now or later.

13. Power Series

13.1. Power series. A series of the form

X
ck xk , where ck (k > 0) are real constants and x ∈ R,
is a real power series. Likewise a series
X
ck z k , where ck (k > 0) are complex constants and z ∈ C,
is a complex power series. Where subsequently we are able to handle the complex case
we shall do so; this subsumes the real case.
Mostly we shall be concerned with the real case, but the theory of convergence works in
the same way for complex power series, and even for elementary applications it is sometimes
convenient to venture into the complex plane.
We think of x (or z) as a variable, and want to investigate the functions we can define by
power series.
Note that a real or complex power series always converges at 0. So we can exclude
consideration of x = 0 (or z = 0) when testing for convergence.

13.2. Elementary functions: exponential, trigonometric and hyperbolic functions.

P zk
• Exponential: Consider . By convention, here and elsewhere, 0! = 1. When z 6= 0
k!
we can test for absolute convergence using (the corollary to) the Ratio Test:
 k+1 
 = |z| → 0 < 1 as k → ∞.
 z /(k + 1)! 

 k
z /k!  k+1
Hence the exponential series converges absolutely, and so converges, for all z ∈ C and
we define
∞
z
X zk
e = .
k=0
k!
This is also written exp(z), when convenient.

z 2k+1 z 2k
(−1)k and (−1)k
P P
• Sine and cosine: Consider . Applying the Ratio
(2k + 1)! (2k)!
Test (for z 6= 0) we can prove (do it for yourself!) that each of these series converges
absolutely for all z ∈ C and we define
∞ 2k+1 ∞
X
k z
X z 2k
sin z = (−1) and cos z = (−1)k .
k=0
(2k + 1)! k=0
(2k)!
52

• Sinh and cosh: We define

∞ ∞
X z 2k+1 X z 2k
sinh z = and cosh z =
k=0
(2k + 1)! k=0
(2k)!

—noting that we make the definitions legitimate by using the Ratio Test to prove that
the associated power series converge absolutely for all z ∈ C.
Note: We would not expect to be able to define tan, cot, sec and cosec, or their hyperbolic
analogues, as power series converging absolutely for all z ∈ C.
We’d like to work our way towards showing that, for a real variable anyway, these series
definitions do capture the properties we expect of the functions familiar from elementary
mathematics. We begin with some elementary facts.

13.3. AOL
P for series in general and power series in particular. Recall the key fact a
series
P a kPconverges if and only if its partial sum sequence (sn ) converges. Given two series
ak and bk with partial sum sequences (sn ) and (tn ), we have
(a1 + b1 ) + · · · + (an + bn ) = (a1 + · · · + an ) + (b1 + · · · + bn ) = sn + tn
P
(by properties of arithmetic—only
P a FINITE P number of terms involved!). Hence, if ak
convergesP
to s and bk converges
P to t, then (a k +b k ) converges, to s+t. Similar arguments
apply to (ak − bk ) and to cak , where c is a constant.
It follows in particular that we can add/subtract and scalar-multiply convergent power
series in the expected ‘term-by-term’ way.

Examples of elementary connections

1 1
cos z = (eiz + e−iz ), cosh z = (ez + e−z ),
2 2
1 1 z
sin z = (eiz − e−iz ), sinh z = (e − e−z ),
2i 2

eiz = cos z + i sin z.

One can also derive, straight from the definitions, the familiar Osborn’s Rules linking cos
and cosh and linking sin and sinh. For example, cos iz = cosh z, and this is valid for complex
z and not just when z is a real number.

13.4. The radius of convergence of a power series.

A power series which fails to converge is useless for defining a function. So it’s critically
important to know when power series do converge.
ck z k to be R, where
P
We define the radius of convergence of
(
sup{ |z| ∈ R | |ck z k | converges } if the sup exists,
P
R=
∞ otherwise.

(Remember that sup E (for E ⊆ R) exists if and only if E is non-empty and bounded above.
|ck 0k | converges, non-emptyness is not at issue here.)
P
Since
 53

ck z k be
P
13.5. Proposition on radius of convergence (technical but important). Let
a power series with radius of convergence R (> 0). Then
|ck z k | converges for |z| < R, and hence ck z k converges for |z| < R.
P P
(i)
ck z k diverges if |z| > R.
P
(ii)

Proof. Assume R 6= ∞ (minor adaptation needed for the case R = ∞).

(i) Fix z with |z| < R and pick S such that |z| < S < R. Then ε = R − S > 0 and
by Approximation Property for supremum we can find ρPwith R − ε = S < ρ < R such
k
|ck z k | converges, by the simple
P
that |ck ρ | converges. But since |z| < ρ, this implies
Comparison Test.

(ii) Assume for contradiction that there exists z with |z| > R such that ck z k converges.
P
Then ck z k → 0. Therefore (because a convergent sequence is bounded) there exists a constant
M such that |ck z k | 6 M . Pick ρ with |z| > ρ > R. Then
k ρ
 k  ρ k
k
0 6 |ck ρ | = |ck z |   6 M   .
 
z z
k k
P P
But |ρ/z| is convergent, since |ρ/z| < 1. So |ck ρ | converges by comparison, and this
contradicts the definition of R. 

Notes
• Part (i) of the proposition is not just the definition of R.
• Some sources replace |ck z k | by ck z k in the definition of R. The proposition implies
P P
that the two versions are equivalent. We prefer to phrase the definition with moduli in,
as a reminder that we can find R by using tests for convergence of series of non-negative
terms.
• We refer to { z ∈ C | |z| < R } as the disc of convergence. For real power series we
have an interval of convergence.
• If ck xk is a real power series with radius of convergence R then the series may converge
P
or diverge at R and −R, and similarly for |z| = R in the complex case.

13.6. Examples: radius of convergence.

(a) The exponential, trigonometric and hyperbolic series all have R = ∞. Proof: use Ratio
Test to test for absolute convergence.
P k
(b) The geometric series x has R = 1 (from Example 11.3).
k!
ck xk where ck = k . We apply the Ratio Test to test for absolute conver-
P
(c) Consider
k
gence, for x 6= 0.
−k
 (k + 1)!xk+1 /(k + 1)k+1 
  
1
lim  k k
 = lim 1 + |x| = e−1 |x|.
k!x /k  k
|ck xk | converges for |x| < e and diverges for |x| > e, so R = e.
P
Hence
P
(d) Consider ck where ck = 1 if k is prime and ck = 0 otherwise. We cannot use the
Ratio Test here. Note that ck xk 9 0 if |x| P > 1 (because there are infinitely many
primes). Hence R 6 1. For |x| < 1, consider r>1 xpr , where p1 < Pp2 p< p3 < . . . is an
pr r
enumeration of the P primes. Then r 6 pr , so |x | 6 |x | and so |x | converges by
r

comparison with |xr |. Therefore R = 1.

54

(e) Problem sheet 6, Q. 4 is set as an exercise on convergence tests in the context of real
power series. With the definition of R in place it can be seen as asking you to calculate R
for various power series. Working through this exercise is intended to provide practice
in finding radius of convergence and to serve also to illustrate the properties of power
series captured by Proposition 13.5.

Important notes

ck xk it is not correct to say, for example, 

P
• In determining the radius of convergence of
‘ |ck xk | converges for |x| < 3, and hence R = 3.’ Convergence for |x| < 3 only
P
implies R > 3. You need to say where the series diverges as well as where it converges.
(Note the way Problem sheet 6, Q. 4 is worded—to prevent you making this common
beginners’ mistake.)

• The last example shows that the Ratio Test cannot necessarily be used to find R because
the limit one needs to consider may not exist. More advanced texts give a formula for
R involving lim inf |ck+1 /ck |. Here lim inf is a notion which can sometimes be exploited
in place of lim when the latter fails to exist. This formula is not needed for Prelims, 
and is frequently abused. DO NOT CLAIM that R = 1/ lim |ck+1 /ck | gives R for a
general power series.

If we are to make good use of power series as functions we would like to know they
have good behaviour, in particular that it is legitimate to differentiate them. The following
Differentiation Theorem is very important, but a first-principles proof is technical (you’ll see
this in Analysis II) and a better proof for real power series uses integration theory (Analysis
III). So we present the theorem here without proof.

ck xk be a real power
P
13.7. Differentiation Theorem for (real) power series. Let
series and assume that the series has radius of convergence R where 0 < R 6 ∞. Let
∞
X
f (x) := ck x k (|x| < R).
k=0

Then f (x) is well defined for each x with |x| < R and moreover the derivative f 0 (x) exists
for each such x and is given by
∞ ∞ ∞
d X ∗
Xd
X
f 0 (x) = ck x k = ck x k = kck xk−1 .
dx k=0 k=0
dx k=1

In words, we say that f 0 (x) is obtained by term-by-term differentiation.

Note: The theorem is very powerful and it is far from obvious that it is true. Note that 
differentiation is carried out by a limiting process, that summing an infinite series also
involves a limiting process, and that iterated limits may not commute (recall Problem sheet
4, Points to Ponder B.). Saying we can differentiate term-by-term is exactly saying that we
can interchange the order of two limits: in ∗ above
∞ ∞
d X X d
ck x k ck x k


and
dx k=0 k=0
dx

we have iterated limits taken in different orders.

 55

13.8. Some applications of the (real) Differentiation Theorem. The general idea here
is to use the Differentiation Theorem to derive suitable differential equations whose solution
leads to formulae connecting functions defined by power series.
(a) Because the power series defining the functions are absolutely convergent for all real x,
from above, the Differentiation Theorem implies that the derivatives below exist and
are given by the expected formulae:
d x
e = ex ,
dx
d d
sin x = cos x, sinh x = cosh x,
dx dx
d d
cos x = − sin x, cosh x = sinh x.
dx dx
To illustrate:
∞ ∞ ∞ ∞
d x d X xk ∗ X d xk X kxk−1 X xm
e = = = = = ex .
dx dx k=0 k! k=0
dx k! k=1
k! m=0
m!
Here the equality marked ∗ holds by the Differentiation Theorem.

(b) To prove sin2 x + cos2 x = 1 for all x ∈ R.

Proof. Let h(x) = sin2 x + cos2 x. Then by the usual rules for differentiation (proved
in Analysis II) h0 (x) exists for all x and
h0 (x) = 2 sin x cos x − 2 sin x cos x = 0.
Hence h(x) is a constant, A say (see Analysis II for justification). Putting x = 0, we
get A = 1 and the result follows. 

We now consider examples in which we have formulae involving two variables.

(c) ea+b = ea eb for a, b ∈ R.
Proof. We let c be a real constant and define a function g on R by
g(x) = ex ec−x .
Then, differentiating (using the product rule and chain rule, proved in Analysis II),
g 0 (x) = ex ec−x − ex ec−x = 0.
It follows (see Analysis II for justification) that g(x) = A, where A is a constant, which
will depend on c. But A = g(0) = ec , since e0 = 1. We have shown that ec = ex ec−x for
all x, c ∈ R, Now put x = a, c = a + b to get
ea+b = ea eb . 
As a special case of the addition formula we get ex e−x = e0 = 1. Certainly ex > 0 when
x > 0. We now see that ex > 0 for all x ∈ R.

(d) Addition formulae for the trigonometric and hyperbolic functions are proved by adapt-
ing the strategy used for (c).

Note: You might be tempted to try to prove (b) by taking the power series for sin x and
squaring it, and likewise for cos x. Not recommended! Justifying multiplying power series
in the same way as one would multiply polynomials requires serious work. After all, you’ve
seen that infinite sums do not necessarily behave the same way as regards arithmetic as finite
sums do—beware · · · ! (See Problem sheet 7, Point to Ponder A. for more on this.)
56

13.9. Trigonometric functions: what became of π. You will probably have met the
cosine and sine series before, as Maclaurin expansions, but you will have been introduced
to the functions cos x and sin x geometrically. We have turned things around and used the
series to define these functions. One can capture the number known as π in either of the
following (equivalent) ways:
• π is the smallest number x > 0 for which sin x = 0;
• π/2 is the smallest number x > 0 for which cos x = 0.
[In Analysis II you’ll discover why these smallest zeros exist.] One can then use the addition
formulae to prove the periodicity results
cos(x + 2π) = cos x, sin(x + 2π) = sin x.

13.10. Looking ahead: the importance of complex power series. We have worked
with complex power series where the arguments are the same as for the real case. We have
restricted consideration of differentiation to the real case, because you have only learned
about differentiation of functions of a real variable so far. In the second year you will study
complex analysis, ‘complex’ meaning ‘in C’ and not ‘more complicated’.
One fact worth noting, and quite easy to prove from the definitions, is that
ex+iy = ex (cos y + i sin y) (for x, y ∈ R).
As a corollary: the facts about π then give

e2πi = 1.

[The remaining examples are on the boundary between Analysis I and Analysis II and III, and
won’t be covered in lectures. They provide additional illustrations of the use of differentiation
of power series to obtain valuable information about important functions.]

13.11. The binomial expansion. Here we shall freely use arbitrary powers of real num-
bers and formulae for their derivatives, noting that such powers are defined using exponentials
(and logs).
Let α ∈ R \ N. (The case that α ∈ N was covered in Introduction to University Mathe-
matics, and doesn’t involve convergence issues.) Consider the infinite sum
∞
X α(α − 1) . . . (α − k + 1)
Bα (x) = xk .
k=0
k!
We apply the Ratio Test to show the series has radius of convergence 1:
 α(α − 1) . . . (α − k)xk+1 /(k + 1)! 
   
α − k 
lim 
  = lim 
  |x| = |x|.
α(α − 1) . . . (α − k + 1)xk /k!  k +1
The Differentiation Theorem tells us we can differentiate term-by-term with respect to x for
|x| < 1, This gives
∞
X α(α − 1) . . . (α − k + 1)
Bα0 (x) = xk−1
k=1
(k − 1)!
∞
X α(α − 1) . . . (α − r) r
= x (writing k = r + 1).
r=0
r!
 57

Now multiply by (1 + x), treating x as constant, and collect together coefficients associated
with the same power of x. We get
∞  
0
X α(α − 1) . . . (α − s) α(α − 1) . . . (α − s + 1)
(1 + x)Bα (x) = + xs
s=0
s! (s − 1)!
∞
X α(α − 1) . . . (α − s + 1)
= ((α − s) + s)) xs .
s=0
s!
Therefore we have a differential equation for Bα :
(1 + x)Bα0 (x) = αBα (x).
Solving, we get Bα (x) = C(1 + x)α , where C is a constant (see Analysis II for justification).
Putting x = 0 we get C = 1.

13.12. The logarithmic series.

Consider ∞
X xk 1 1
L(x) = = x + x 2 + x3 + . . . .
n=1
k 2 3
P xk
The series converges for |x| < 1 and diverges for |x| < 1 (easy application of Ratio
k
Test). By the Differentiation Theorem, we can differentiate term-by-term for |x| < 1 to get
1
L0 (x) = 1 + x + x2 + · · · = .
1−x
For −1 < x < 1, define
M (x) = (1 − x)exp(L(x)).
0
Computing M (x) by the chain and product rules (proved in Analysis II), we get
M 0 (x) = (−1)exp(L(x)) + (1 − x)L0 (x)exp(L(x)) = 0.
Hence M (x) is constant, equal to M (0) = 1 (see Analysis II for justification). Hence
1
exp(L(x)) = . (−1 < x < 1).
1−x
So, from the addition formula for the exponential function, we have exp(−L(x)) = 1 − x.
Let us define a function log to satisfy
 
1
− log(1 − x) = log = L(x); i.e. log(1 + x) = −L(−x).
1−x
This then gives us a log function acting as the inverse to the exponential function, and a
series expansion for log(1 + x), valid for −1 < x < 1. All this is in line with what we want
and expect the logarithm to do. The full picture will emerge in Analysis II (which considers
inverse functions) and Analysis III (which considers the definition of the log function by an
integral).