Complex 1
Complex 1
Complex 1
Contents
1 Introduction to the course 3
1.1 Complex numbers . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Exponential representation of complex numbers . . . . 5
1.1.2 Algebra of complex numbers . . . . . . . . . . . . . . . 6
1.1.3 Complex conjugates and inequalities . . . . . . . . . . 6
1.2 Solving simple complex equations . . . . . . . . . . . . . . . . 9
1
3.5.2 Logarithms . . . . . . . . . . . . . . . . . . . . . . . . 41
3.5.3 Contours and multivaluedness . . . . . . . . . . . . . . 42
3.5.4 Branch cuts to ensure single-valuedness . . . . . . . . . 44
3.6 “Composite” multifunctions . . . . . . . . . . . . . . . . . . . 45
3.6.1 Example: Product of square-roots . . . . . . . . . . . . 46
3.6.2 Example: Logarithm of a product . . . . . . . . . . . . 48
3.7 Summary and concluding remarks on multifunctions . . . . . . 48
4 Complex Integration 50
4.1 Basic methods of complex integration . . . . . . . . . . . . . . 50
4.2 General results for complex integration . . . . . . . . . . . . . 53
4.3 Harmonic functions to construct analytic functions . . . . . . 59
5 Cauchy’s theorem 60
5.1 Cauchy’s theorem I . . . . . . . . . . . . . . . . . . . . . . . . 60
5.2 Cauchy’s theorem II . . . . . . . . . . . . . . . . . . . . . . . 62
2
9.5.2 Integrals around ”sector” contours . . . . . . . . . . . 132
9.5.3 Integrals of functions with branch-points . . . . . . . . 134
9.5.4 Integrals around rectangular contours . . . . . . . . . . 139
9.6 Principle of the argument, and Rouché’s theorem . . . . . . . 143
Book: M.J. Ablowitz & A.S. Fokas. Complex variables: Introduction and
Applications (2nd edition). Cambridge University Press (2003).
3
Two complex numbers are equal if and only if both their real and imaginary
parts are equal. The absolute value or magnitude of the complex number
z is defined by the length of the vector (x, y) associated with z:
�
|z| = x2 + y 2 , (2)
Here again r is the absolute value of z, r = |z|, and the angle θ is known as
the argument of z, written θ = arg(z). Note that θ in this representation of
z is not single-valued: if z0 = r0 (cos θ0 + i sin θ0 ) is one representation, then
replacing θ0 by θ0 + 2nπ also gives the same complex number z. Any such
value of θ is an argument of the complex number, and arg(z) is therefore really
an infinite set of θ-values. It is often useful to define a unique principal
argument, which is usually taken to be the argument lying in the range
0 ≤ θ < 2π (but sometimes the range −π < θ ≤ π; convenience usually
dictates the choice in a particular application). The principal argument is
sometimes written θ = Arg(z) to distinguish it from the set arg(z). We
will see further consequences of this nonuniqueness later on, when we discuss
multi-valued functions.
Example 1.2 Find the √ polar form of the complex numbers (i) z1 = i, (ii)
z2 = 1 − i, (iii) z3 = 3 − i.
4
Solution: (i) For z1 we have |z1 | = 1 and θ = cos−1 (0) = sin−1 (1) =
π/2 + 2nπ, for any integer n. The principal argument would be θ = π/2. In
any case it is clear that z1 = cos(π/2)
√ + i sin(π/2). √ √
(ii) For z2 = 1 − i we have |z2 | = 2, and
√ thus 1 = 2 cos θ, −1 = 2 sin θ.
Thus θ = −π/4, modulo
√ 2π, and z2 = 2(cos(−π/4) + i sin(−π/4)).
(iii) |z3 | = 2, and 3 = 2 cos θ, −1 = 2 sin θ. Thus θ = −π/6 (modulo 2π),
and z3 = 2(cos(−π/6) + i sin(−π/6)).
5
Example 1.3 For the numbers in example 1.2 we have (i) z1 = eiπ/2 , (ii)
√ −iπ/4
z2 = 2e , and (iii) z3 = 2e−iπ/6 .
z + w = x + u + i(y + v),
zw = (x + iy)(u + iv) = xu − yv + i(xv + yu).
From these laws and the properties already noted, it is easy to show that
the complex numbers C form a field. The inverse of the complex number z,
given by 1/z, satisfies
1 1 x − iy x − iy
z −1 = = = = 2 . (5)
z x + iy (x + iy)(x − iy) x + y2
This identity may be alternatively obtained by the complex polar represen-
tation (4), which gives
−1 e−iθ re−iθ
z = = . (6)
r |z|2
The polar representation also gives the neatest interpretation of complex
multiplication (and division which, with the notion of an inverse given above
is really just multiplication also), since if z = reiθ and w = Reiφ then
zw = rRei(θ+φ) .
√ iπ/4
Example 1.4 The product of z1 and z2 in example 1.2 √ is z1 z2 = 2e =
iπ/3
1 + i; the product of z√1 and z3 is z1 z3 = 2e = 1 + i 3; and the product of
−5iπ/12
z2 and z3 is z2 z3 = 2 2e .
The first equality above shows that geometrically, the complex conjugate of a
point z ∈ C is obtained by reflection of z in the real (x) axis. The definition
6
(7) also shows that the inverse of z, given by (5) or (6) above, is related to
its complex conjugate by
z
z −1 = .
|z|2
It is evident that
z + w = z + w, zw = z w.
and then that, for any complex number, the absolute value of its real part
must be less than or equal to the absolute value of the complex number itself,
which gives
Thus,
Noting that |z +w| ≥ 0 and |z|+|w| ≥ 0, the second inequality of (8) follows.
For the first inequality, we use the result just proved, and set Z = z + w,
W = −w (arbitrary complex numbers, since z and w are arbitrary). Then
z = Z + W , w = −W , and
|Z| ≤ |Z + W | + |W | ⇒ |Z| − |W | ≤ |Z + W |.
7
If |Z| ≥ |W | then the result follows; if not then repeat the above argument
with W = z + w and Z = −w to obtain
|W | − |Z| ≤ |Z + W |,
proving the result for |W | ≥ |Z| also.
The right-hand inequality in (8) generalizes easily to an arbitrary sum of
complex numbers:
� �
�� N � � N
� �
� zn � ≤ |zn |.
� n=1 � n=1
When the complex numbers z, w are plotted in the complex plane |z| is just
the length of the straight line joining z to the origin. The complex number
z+w is found from the usual parallelogram law for vector addition. Therefore
(8) has a geometrical interpretation in terms of lengths of sides of a triangle.
8
Remark Note that inequalities pertain only to the absolute value of com-
plex numbers. Unlike the real numbers, there is no ordering of the field of
complex numbers, so no notion of one complex number being greater than
or less than another.
az 2 + bz + c = 0 a, b, c ∈ C
� �2
b b2
⇒ a z+ +c− =0
2a 4a
� 2 �1/2
b b c
⇒ z=− + − .
2a 4a2 a
We will see later how to take the square-root of the complex number b2 /(4a2 )−
c/a and obtain two distinct roots, just as in the real case. [Note: for
quadratic equations with complex coefficients, the roots do not occur in
complex conjugate pairs.]
Higher order polynomials are more difficult to solve analytically, but the
equation z n = a, for a ∈ C, has a general solution, which we find by writing a
using the complex polar representation (4) and using the multi-valued nature
of the argument:
z n = a = |a|ei(φ+2mπ) m ∈ Z. (9)
Obviously, for any m ∈ Z the right-hand side of (9) is the same, but when
we take the nth root to solve for z, different m-values give different roots for
z. Thus
9
and so on). In the special case a = 1 = 1.ei.0 we have
zm = e2mπi/n = ωnm , m = 0, 1, . . . n − 1;
these are the n roots � of unity (ωn = e2πi/n ). More generally, any nth order
polynomial equation nj=1 aj z j = 0 has exactly n complex solutions – we
will prove this result later.
Example 1.8 If n = 2 and a = 1 in (9), a familiar real case, (10) gives two
solutions zm = e2mπi/2 , m = 0, 1, that is, z0 = 1 and z1 = eiπ = −1.
If n = 3 and a = 1 then we have the three cube roots of unity, z0 = 1, z1 =
e2πi/3 , z2 = e4πi/3 . Note that z2 can be written equivalently as z2 = e−2πi/3 .
Returning to the equation z n = 1, we see that the factor z − 1 is easily
isolated:
z n − 1 = (z − 1)(z − ωn ) . . . (z − ωnn−1 ) = 0,
though it is far from obvious that the product of these n factors gives z n − 1.
Try showing that this is the case for small values of n, e.g., n = 3, n = 4.
10
parametrizing subsets of C, which will be useful later on when we consider
complex integration.
Perhaps the most basic subsets of C are lines, or line segments. For
example, the real axis (the set of complex numbers with zero imaginary
parts) may be described by any of the following equations:
�(z) = 0; z = z; |z − i| = |z + i|.
More generally, noting that for any complex number a ∈ C, a is its reflection
in the real axis, we can state that the real axis is the set of all points that
are equidistant from a and a. This geometrical observation leads to the
description of the real axis by the equation
LR = {z ∈ C : z = x, x ∈ R}.
If we wish to restrict to one or more line segments then we must restrict the
value of the parameter x appropriately. In general, a line L of finite length
between two points a, b ∈ C is defined by
L = {z ∈ C : z = (1 − t)a + tb, 0 ≤ t ≤ 1}
11
Exercise: What subset is described by the parametrization
L = {z ∈ C : z = (1 − t2 )a + t2 b, 0 ≤ t ≤ 1}?
What about
L = {z ∈ C : z = (1 − t)a + t2 b, 0 ≤ t ≤ 1}?
S = {z ∈ C : |z − a| ≤ r}.
12
Figure 1: Two possible circular arcs passing through given points a, b ∈ C
arg(z − a) − arg(z − b) = µ
(see figure 1) or, using the fact that for any two complex numbers z1 and z2 ,
arg(z1 /z2 ) = arg(z1 ) − arg(z2 ),
� �
z−a
arg = µ mod 2π.
z−b
The case µ = π is the degenerate case when the arc collapses to a line segment
joining a and b.
To find the equation of the other arc joining a and b that completes the
circle, note that for a point w on that arc the angle awb must equal the
supplement of azb, so that
� � � �
b−w w−b
arg(b − w) − arg(a − w) = arg = arg = π − µ mod 2π.
a−w w−a
13
We now consider more general subsets of the complex plane, introducing
several formal definitions.
Since the inequality in the definition is strict, the boundary of the disk is
excluded from the neighborhood (this point seems trivial, but is not, as
we shall see). Related to neighborhoods is the concept of a ball or disk
surrounding a given point z0 .
Definition 2.2 The ball, or disk, of radius r about the point z0 is denoted
by B(z0 ; r) (or sometimes D(z0 ; r)), and is defined by
If we remove the point z0 from the ball (or disk) the resulting set is called
the punctured ball or disk, and is denoted by B � (z0 ; r) (or D� (z0 ; r)).
Definition 2.4 (Open set) The set S is open if all its points are interior.
S1 in the example above is open, but S2 is not. To see that S1 is open, we
need to show that for any point z0 ∈ S1 , we can find an �-neighborhood of z0
that is contained within S1 . Let z0 ∈ S1 , then, by definition of S1 , |z0 | < 1.
14
Let � = 1 − |z0 | > 0. The �-neighborhood Nz0 ,� of z0 lies within S1 because
if z1 ∈ Nz0 ,� then, by (12), z1 can be written
where we used (8) in the above. Since |z1 | < 1, z1 ∈ S1 , and since z1 ∈ Nz0 ,�
is arbitrary, we have shown that a neighborhood of z0 lies within S1 for any
z0 ∈ S1 . Hence all points of S1 are interior points, and so S1 is an open set.
To see that S2 is not open we just need to find a point z0 ∈ S2 for which
no neighborhood lies within S2 . In fact any point z0 ∈ S2 will do for this
purpose. Using the description (12), a neighborhood of any z0 ∈ S2 contains
all the points
15
Lemma 2.8 A set S is closed if and only if it contains all its limit points.
Proof
S is closed ⇐⇒ C \ S is open
⇐⇒ given z �∈ S, ∃ � > 0 such that B(z, �) ⊂ C \ S
⇐⇒ given z �∈ S, ∃ � > 0 such that B � (z, �) ∩ S = ∅
⇐⇒ no point of C \ S is a limit point of S.
Since we know the final statement is true if and only if the set contains all
its limit points, the equivalence of this with the set being closed is proved.
16
2.3 Stereographic projection
A one-to-one correspondence may be made between the ”flat” complex plane
and the surface Σ of a sphere of unit diameter in R3 (which we take as
(x, y, Z)-space for the purposes of this discussion). Take the sphere to rest
with its base (its South pole, (0, 0, 0)) at the origin of C, z = 0. Then, for any
z = x+iy ∈ C, draw a straight line joining z to the North pole, N = (0, 0, 1),
of the sphere. This line intersects the surface of the sphere at a unique point
P (x, y), which depends only on the coordinates (x, y) associated with z. As
the magnitude of the complex number |z| → ∞, the point P approaches
the North pole (0, 0, 1), so that if we consider ”∞” to be a point in C then
this maps uniquely to the North pole of the sphere. This representation is
sometimes referred to as a compactification of the complex plane; and the
complex plane with this point ”∞” appended is called the extended complex
plane. A detailed description of such a correspondence may be found in
Ablowitz & Fokas [2].
The explicit correspondence may be worked out, using elementary geom-
etry, as
� �
x y r2
z = x + iy = re ∈ C ↔
iθ
, , ∈Σ
1 + r2 1 + r2 1 + r2
17
Figure 2: The stereographic projection of the extended complex plane onto
the unit sphere.
18
through the complex plane. Paths are often defined in terms of a real pa-
rameter. We have already seen a couple of examples of simple paths that
can be described in this way in §2.1. The circle, centered on z0 with radius
r, can be thought of as a path, whose end-point coincides with its beginning
point (an example of a closed path). Paths are denoted variously by γ, Γ, or
C, possibly with parametric dependence denoted too, e.g. γ(t). If the path
in question is the boundary of some domain D then the alternative notation
∂D may also be used. Thus, we might describe a circular path by
γ = {z = z0 + reiθ : 0 ≤ θ ≤ 2π}. (14)
Here θ is the real parameter, and as θ increases from 0 to 2π we traverse the
circle in the anticlockwise direction, starting at the point z = z0 + r. The
quantity γ(θ0 ) denotes the point on the path corresponding to parameter
value θ0 .
We also saw in §2.1 how a straight line between two points z1 and z2 may
be represented parametrically, as
γ1 = {z = (1 − t)z1 + tz2 : t ∈ [0, 1]} = [z1 , z2 ]. (15)
A path along a second line segment from z2 to z3 may obviously be defined
in the same way:
γ2 = {z = (1 − t)z2 + tz3 : t ∈ [0, 1]} = [z2 , z3 ].
Continuing in this way, any sequence of points zn ∈ C may be connected
by a sequence of such line segments, forming a piecewise linear or polygonal
path within C.
A general path γ through complex space defined in terms of a real param-
eter t may be described by z(t) = x(t) + iy(t), with t lying in some interval
[a, b]. As for the specific examples above we write
γ = {z(t) : t ∈ [a, b]}. (16)
Definition 2.11 (Closed curve) A curve γ defined as in (16) is closed if
z(a) = z(b).
19
Homework: 1. Sketch the contours γ defined by: (i) γ = {z : z = 1 +
ieit , t ∈ [0, π]}. (ii) γ = [−1, 1] ∪ [1, 1 + i] ∪ [1 + i, −1 − i] ∪ [−1 − i, −1] (the
join of these line segments). (iii) γ = {z : z = cos teit , t ∈ [0, 2π]}.
2. Define parametrically the following paths: (i) the square with vertices at
±1 ± i; (ii) the closed semicircle in the right half-plane with [−iR, iR] as
diameter.
The connectedness of S will then imply that one of these two sets is empty.
This cannot be S0 , since a ∈ Sa . Thus Sa = S, and S is also polygonally-
connected.
20
Proof of claim: Let z ∈ S be arbitrary. Since S is open and z ∈ S,
∃ � > 0 such that B(z; �) ⊂ S. For each w ∈ B(z; �) the line segment
[z, w] ⊂ B(z; �) ⊂ S. It follows that z can be joined to a by a polygonal path
in S if and only if w can be. Since S = Sa ∪ S0 and Sa , S0 are disjoint, it
follows that, for k = a, 0, z ∈ Sk ⇒ B(z; �) ⊂ Sk , and thus both Sa and S0
are open, as claimed.
This proves the theorem one way, that a connected set is polygonally-connected.
To prove the converse, suppose that S is non-empty, open, and polygonally-
connected. We will use contradiction, and hence we suppose that S = S1 ∪S2 ,
where both S1 and S2 are open, non-empty, and disjoint. Let a ∈ S1 and
b ∈ S2 , then there exists a polygonal path γ lying in S and joining a to b.
At least one line segment of γ has one endpoint p in S1 , and the other, q, in
S2 , and of course the entirety of γ is contained in S1 ∪ S2 . Set
21
Remark Implicit in the proof above was the notion of convexity, when
we stated that for each w ∈ B(z; �) the line segment [z, w] ⊂ B(z; �). More
generally we have the definition:
22
This definition allows the orientation of a contour to be specified. We
follow the standard convention in complex (and real) analysis, which is the
following:
23
Many of the results and theorems on convergence of real sequences and series
follow through for the complex case (we must replace any modulus notation
in the real case by the absolute value of the appropriate complex quantity).
Often results can be proved simply by considering the sequence or series as
a sum of real and imaginary parts, and applying the real result to each part.
Particular examples of results for complex series include:
�
1. If n an converges then (i) |an | → 0 as n → ∞; and (ii) ∃ M > 0 such
that |an | < M ∀ n. (i) may be proved by noting that since the series
converges (to s say) then given � > 0 ∃N such that, for all n ≥ N ,
|sn − s| < � (where sn denotes the partial sum of the series to n terms.
Thus, choosing m, n ≥ N , we have
Recalling that � > 0 was arbitrary, this proves the convergence of |an |
to zero.
24
4. (Comparison�test.) If we have a convergent series with non-negative
� n bn , and a positive constant k such that |an | ≤ kbn
real terms,
∀n, then n an converges. Armed with the result 3 above, this is a
straightforward extension of the comparison
� test for real functions -
the real result gives convergence of |a
n n |, and then this
� absolute
convergence implies the convergence of the original series n an .
1 − z n+1
Sn = z �= 1,
1−z
from which it is clear that for |z| < 1 Sn → 1/(1 − z), while for |z| ≥ 1 the
series diverges.
(The partial sum Sn in this example may be obtained by noting that Sn+1 =
zSn + 1, and also Sn+1 = Sn + z n+1 .)
The notion of series leads naturally to the definition of complex power se-
ries, of the general form
∞
�
cn z n , cn ∈ C. (17)
n=0
25
(R may be infinite). The results listed above guarantee that the power series
converges within the circle of convergence |z| < R. To see this, let |z| < R.
By the definition of R as supremum of the convergence � set, there exists
some w ∈ C, with |z| < |w| ≤ R, such that n
|cn w | �
converges. Since
|cn z n | < |cn wn |, result 4 above guarantees convergence of n cn z n .
Proof
� Since the radius of convergence is defined in terms of the real series
n
|cn z�|, we can use results from real analysis. The ratio test for the real
series ∞ n=0 an with positive terms states that if
� �
� an+1 �
L = lim �� �
n→∞ an �
exists then for L < 1 the series � converges, while if L > 1 then the series
diverges. Applying this result to |cn z n |, we have convergence for
� �
� cn+1 z n+1 �
lim � � < 1,
n→∞ � cn z n �
and divergence for
� �
� cn+1 z n+1 �
lim � � > 1.
n→∞ � cn z n �
Thus, we have convergence for
� �
� cn �
�
|z| < lim � �,
n→∞ cn+1 �
and divergence for |z| > limn→∞ |cn /cn+1 |. The lemma is proved.
�
Example 3.6 The radius of convergence of the power series ∞ n
n=0 z is 1.
26
�∞
Example 3.7 The radius of convergence of the power series n=0 an z n ,
where a ∈ C is nonzero, is R = 1/|a|.
This is just a straightforward extension of the example above, and again may
be obtained by either applying lemma 3.5 or by the direct method.
lim sup |cn z n |1/n = |z| lim sup |cn |1/n < 1,
n→∞ n→∞
Hence R = 0.
Exercise: (1) Try showing this � byn other means. (2) Find the radius of
convergence of the power series n z /n!.
Within the circle of convergence, power series such as (17) can be thought
of as defining a function of the complex variable z. More generally, a func-
tion of the complex variable z on a domain D is a mapping or rule f that
27
assigns a unique complex number w to each z ∈ D. We write w = f (z) to
represent this mapping, and use the notation f : D → C to represent the
fact that f takes values in D to other complex values. In general of course if
the domain D is only a subset of C, the image f (D) of D under f will only
be a subset of C, so you may also see notation f : D → f (D), or similar.
While some real functions require care in their generalization to the complex
plane (fractional powers being one example that we consider later), some are
straightforward to generalize. For example, integer powers of z are easily
generated using the rules of algebra we have already introduced, so that
z 2 = (x + iy)(x + iy) = x2 − y 2 + 2ixy,
z 3 = (x + iy)(x2 − y 2 + 2ixy) = x3 + 3ix2 y − 3xy 2 − iy 3 ,
and so on. From this it is a simple matter to generate complex polynomial
functions,
N
�
P (z) = an z n , an ∈ C;
n=0
the power series given above are clearly a limiting case of such functions.
From polynomials complex rational functions are then ratios of two poly-
nomials:
�N
P (z) an z n
R(z) = = �Mn=0 ,
Q(z) m=0 b m z m
Exercise: Find the image of the first quadrant, �(z) ≥ 0, �(z) ≥ 0, under
the functions f1 (z) = z 2 , f2 (z) = z 4 .
In general, since we know z = x + iy, and since the function f (z) takes values
in C, we can write
f (z) = f (x + iy) = u(x, y) + iv(x, y), (18)
for real-valued functions u, v, of the two real variables (x, y). We say that u
is the real part of f , and v is the imaginary part.
The exponential function can be generalized using the relation (3) that
we have already seen. Thus,
ez = ex+iy = ex (cos y + i sin y).
28
We note, for future reference, that the absolute value of this complex expo-
nential depends only on the real part of z:
|ez | = |ex eiy | = ex .
We also have
eiz = e−y+ix = e−y (cos x + i sin x), e−iz = ey−ix = ey (cos x − i sin x). (19)
There is a “natural” way to motivate the generalization of the trigonometric
functions to the case of complex argument. We define
eiz − e−iz eiz + e−iz
sin z = , cos z = . (20)
2i 2
This makes sense, because of the relation (3) for real θ noted earlier, which
gives
eiθ − e−iθ eiθ + e−iθ
sin θ = , cos θ = ,
2i 2
so that (20) is the natural generalization when θ is replaced with a complex
number z. The definitions (20) satisfy the usual trigonometric addition for-
mulae for sin(z + w), etc., and the identity cos2 z + sin2 z = 1 is also easily
verified to hold. Other trigonometric functions (tan z, cot z, sec z, cosecz)
can then be defined as in the real case, in terms of sine and cosine.
Complex hyperbolic functions can be defined similarly from their real
counterparts, as
ez − e−z ez + e−z
sinh z = , cosh z = (21)
2 2
(from which the other hyperbolic functions can be defined); and again, the
addition formulae generalize from the real case as we would expect.
From the definitions (20) and (21) we note that sin and sinh, cos and
cosh, are simply related by
sinh iz = i sin z, sin iz = i sinh z, cosh iz = cos z, cos iz = cosh z.
We can also check the consistency of the definitions using the x + iy repre-
sentation and (19), which gives
eiz − e−iz 1 −y
= (e (cos x + i sin x) − ey (cos x − i sin x))
2i 2i
29
1
= (− cos x(ey − e−y ) + i sin x(ey + e−y )
2i
= i cos x sinh y + sin x cosh y
= cos x sin iy + sin x cos iy
= sin(x + iy)
= sin z.
Remark Note that intuition from real-valued functions does not usually help
us with their complex analogs. For example, though the real sine and cosine
functions are bounded in absolute value by 1, the complex trigonometric
functions are unbounded. If z = iy for y ∈ R then sin z = i sinh y, cos z =
cosh y, so that both sin z and cos z are unbounded as y → ∞. Likewise,
though the real hyperbolic cosine is always real and positive (always greater
than or equal to 1, in fact), the complex hyperbolic cosine has infinitely many
zeros in the complex plane – note that for z = iy, cosh z = cos y, and thus
the complex hyperbolic cosine has zeros along the imaginary axis at points
z = (n + 1/2)π, n ∈ Z.
lim f (z) = w0
z→z0
(that is, the limit of f (z) as z → z0 exists and equals w0 ) if, ∀ � > 0
(sufficiently small) ∃ δ > 0 such that
30
Pictorially, we require the value of the function f to approach w0 when z
approaches z0 from an arbitrary direction in D. The definition 3.10 may
be extended to the case z0 = ∞ simply by considering |z| arbitrarily large;
thus
lim f (z) = w0 (|w0 | < ∞)
z→∞
Homework: (1) Evaluate the following limits, if they exist: (i) limz→0 sin z/z;
(ii) limz→∞ sin z/z; (iii) limz→∞ z 2 /(3z + 1)2 ; (iv) limz→∞ z/(z 2 + 1).
(2) If |g(z)| < M for all z in a neighborhood of z0 , and if limz→z0 f (z) = 0,
show that limz→z0 f (z)g(z) = 0.
Example 3.12 Show that the functions f (z) = z and f (z) = z are con-
tinuous everywhere in C, and that g(z) = z 2 is continuous on any bounded
domain.
The choice δ = � works for both cases f (z) = z, z here, since, if |z − z0 | < �
then
|f (z) − f (z0 )| = |z − z0 | < �.
For g(z) we note that points on a bounded domain satisfy |z| < M for some
fixed M . Then, if |z − z0 | < δ we have
|g(z) − g(z0 )| = |z 2 − z02 | = |z − z0 ||z + z0 | < δ|z + z0 | ≤ δ(|z| + |z0 |) < 2M δ,
31
where we used (8). Thus, choosing δ = �/(2M ), the continuity is proved.
We record one more important theorem on continuous complex functions,
that will be needed later:
Since the theorem pertains only to the magnitude of f , a real quantity, the
proof can proceed as in the real case.
32
in the region of analyticity, and all these derivatives are themselves analytic.
So, analytic ⇒ infinitely differentiable. Differentiability of f at a point z0
evidently guarantees continuity of f at z0 :
f (z0 + h) − f (z0 )
�h := f � (z0 ) − → 0 as h → 0.
h
Then
z+h−z h
f � (z) = lim = lim .
h→0 h h→0 h
33
This limit must exist independently of how the limit is taken, for all possible
paths h → 0 in C. Writing h = |h|eiθ , where θ is an arbitrary argument of
h, we find
This is clearly not a well-defined limit, since it depends explicitly on the way
h → 0 (θ can take any real value here). So, the complex conjugate function
is not differentiable anywhere – though it is continuous everywhere.
Example 3.18 Likewise, the function f (z) = |z|2 is not differentiable any-
where in C except the origin.
This follows since for z �= 0,
which depends on the way we take the limit h → 0. Note however that
|z|2 = x2 + y 2 is very well-behaved – infinitely differentiable, in fact – as a
real valued function in R2 .
These examples illustrate the remark above, since in both cases f is a
function of both z and z̄, and cannot be expected to be differentiable. For
each of these functions, all points in C are singular points. (Even though
f (z) = |z|2 is differentiable at z = 0, it is not differentiable in any neighbor-
hood of zero, and therefore not analytic at z = 0.)
34
Example 3.20 The function f (z) = z n is differentiable, with f � (z) = nz n−1 .
If n is a positive integer (or zero) then this result holds ∀ z ∈ C, but if n is
a negative integer then f is not differentiable at the point z = 0.
In fact the result holds also for arbitrary powers of z, but since we have not
yet learned how to deal with non-integer powers we restrict to integers for
now. To prove the result for n ∈ N we use the standard binomial expansion
for (z + h)n (knowing how to evaluate integer powers of complex numbers,
this may be proved by induction as in the real case), giving
(z + h)n − z n
f � (z) = lim
h→0 h
z n + nhz n−1 + n(n − 1)h2 z n−2 /2! + · · · + hn − z n
= lim
h→0 h
n−1
→ nz as h → 0.
Almost the same proof works for the case of negative integers n if we assume
that the binomial expansion for this case is valid for complex numbers (it
is). For fixed z �= 0 we simply extract the factor z n from (z + h)n and then
expand (1 + h/z)n for |h/z| < 1. We obtain the same result as h → 0.
Homework: (1) Show from first principles that the function f (z) = 1/z is
(a) continuous; (b) analytic in C \ B(0; r) for any r > 0. (The point z = 0 is
the only singular point of this function.)
(2) Show from first principles that, if f is continuous, nonzero and differen-
tiable at z0 ∈ C, then 1/f is differentiable at z0 , and its derivative at this
point is given by −f � (z0 )/f (z0 )2 . [NOTE: You are being asked to prove this
form of the quotient rule here, not apply it!]
(3) Find where the following functions are differentiable (and hence analytic):
(i) sin z; (ii) tan z; (iii) (z − 1)/(z 2 + 1); (iv) �(z).
You may assume that the sum and product of two functions is differentiable
on any region where the two functions are themselves differentiable. Any
results shown in lectures may be assumed.
(4) Let f (z) be continuous everywhere. Show that, if f (z0 ) �= 0, then there
exists a neighborhood of z0 in which f (z) �= 0.
35
3.4 The Cauchy-Riemann theorem
Differentiability (or non-differentiability) of a complex function f (z) may also
be established by appealing to the Cauchy-Riemann theorem.
ux = v y , uy = −vx (24)
in a neighborhood of z.
f (z + i|h|) − f (z)
f � (z) = lim
h→0 i|h|
u(x, y + |h|) − u(x, y) + i(v(x, y + |h|) − v(x, y)
= lim
|h|→0 i|h|
= −iuy (x, y) + vy (x, y). (26)
These two results must be identical for a well-defined limit and thus
ux = vy , uy = −vx .
36
consider taking the limit of a small variation from the point z = x+iy, which
we can represent variously as
z → z + δz = x + iy + δx + iδy.
This variation induces changes δu = u(x + δx, y + δy) − u(x, y), δv = v(x +
δx, y + δy) − v(x, y), δf = f (z + δz) − f (z), in u, v and f , which, by results
from real analysis (Taylor’s theorem), satisfy
where we used the Cauchy-Riemann equations (24) and the limiting behavior
of �u and �v in the above. Thus, the derivative of f at the point z exists, and
the proof is complete.
Exercise: It is clear that there are several possible expressions for f �� (z).
Find them all.
37
Note 2 The Cauchy-Riemann equations also confirm the remark made
above, that only if f turns out to be purely a function of z can it be analytic.
To see this we note that
� �
1 i ∂ 1 ∂ ∂
x = (z + z̄), y = − (z − z̄) ⇒ = +i
2 2 ∂ z̄ 2 ∂x ∂y
(using the chain rule for partial differentiation). Thus, if u and v satisfy the
Cauchy-Riemann equations then
∂f 1 1
= (ux + iuy + i(vx + ivy )) = (ux − vy + i(uy + vx )) = 0,
∂ z̄ 2 2
so that f is independent of z̄.
It is becoming clear that the real and imaginary parts of an analytic
function are closely related, and that one cannot hope to choose real-valued
functions u(x, y) and v(x, y) independently, stick them together as real and
imaginary parts, and obtain an analytic function. The choice of real part
u(x, y) almost uniquely determines the imaginary part v (we will make this
statement precise soon).
Returning to example 3.18 above, with f (z) = |z|2 we have u(x, y) =
x2 + y 2 , v(x, y) = 0. The Cauchy-Riemann equations are not satisfied in the
neighborhood of any z ∈ C, confirming the above finding that this function
is nowhere analytic.
Homework: Ablowitz & Fokas, problems for section 2.1, questions 1,2.
∂2u ∂2u
∇2 u = + =0 (27)
∂x2 ∂y 2
38
Lemma 3.23 The real and imaginary parts of an analytic function f (z) =
u(x, y) + iv(x, y) are harmonic.
We are anticipating a result that will be established later, that analyticity
of f guarantees that the second derivatives (in fact, derivatives of all orders)
of its real and imaginary parts exist and are continuous on the domain D on
which f is defined and analytic.
Proof Since f is analytic its derivatives of all orders exist, and thus the
partial derivatives of u and v exist and are continuous at all orders in the
domain of analyticity. By the Cauchy-Riemann equations we then have
∂ ∂ ∂ ∂
uxx + uyy = (ux ) + (uy ) = (vy ) + (−vx ) = vxy − vxy = 0.
∂x ∂y ∂x ∂y
Similarly we can show that vxx + vyy = 0, so that both u and v are harmonic.
We will later prove a converse (lemma 3.24 below) to this theorem: that
harmonic functions can be used to construct analytic functions.
Example 3.25 The analytic function f (z) = z 4 has real and imaginary
parts u(x, y) = x4 − 6x2 y 2 + y 4 , v(x, y) = 4x3 y = 4xy 3 , both of which are
harmonic everywhere. We know this without doing any differentiation, by
lemma 3.23. It’s also easily checked that u and v satisfy the Cauchy-Riemann
equations.
39
a positive number
√ gives two values: the equation y 2 = x > 0 has two solu-
tions, y = ± x, for any real positive number x. Likewise, real-valued inverse
trigonometric functions are also multivalued. In the complex case also we
would expect to be able to obtain two solutions to the equation
w2 = z, (28)
√
for any given z ∈ C, but how do we interpret z? Likewise (though this
does not arise for the real case), we saw that the exponential function ew
can return the same value for many different choices of w (for example, all
values wn = i(θ + 2nπ) for any integer n give the same result for ew ). Thus
a sensible inverse for the complex exponential function (which is a complex
version of the logarithm) should be a multi-valued function.
We begin by considering how one can define a complex square-root (solv-
ing equation (28) for w(z)). If we try to follow the route of taking real and
imaginary parts things get quite cumbersome. Writing z = x+iy, w = u+iv,
we find
u2 − v 2 = x, 2uv = y;
substituting for v from the second of these into the first gives a real quartic
equation for u – four possible solutions (but hard to find), and some may
(will!) be spurious when substituted back into the original equation w2 = z.
There is a much better way to find solutions to equations such as (28), and
it involves writing the complex number z, whose square-root (or any other
fractional power for that matter) we wish to take, in the complex exponential
form (3). Equation (28) is then
w2 = |z|eiθ = |z|ei(θ0 +2nπ) ,
where θ is any argument of the complex number z and θ0 is a specific choice
(which will usually be taken as the principal argument in applications). Then,
as in the real case, we set
w = f (z) = |z|1/2 ei(θ0 /2+nπ) (29)
(where |z|1/2 is interpreted as in the real case, since |z| ∈ R). Again, this
relation holds for any integer n, but in fact w can only take two distinct
values:
n=0: w = |z|eiθ0 /2 , (30)
n = 1 : w = |z|1/2 ei(θ0 /2+π) = −|z|1/2 eiθ0 /2 . (31)
40
Other integer values of n give these same values (even values of n give the
same as n = 0, and odd n the same as n = 1). These two distinct definitions
of z 1/2 are known as branches of the square-root function f (z).
The simplest case is the mth root function (α = m ∈ N), which has m
distinct branches, defined by
3.5.2 Logarithms
The second major class of multifunctions is given by the complex logarithm.
As mentioned, this function is obtained by inverting the complex exponential,
z = ew . Writing z in its complex exponential form again, and w = u + iv,
we see that
where θ0 is any choice from the set of values arg(z) (again, the principal ar-
gument of z is usually the most sensible definition of the complex logarithm).
As above, the different forms of the function that arise from taking different
values of n in (34) are known as the branches of the complex logarithm.
There are always infinitely many of them.
It is easily verified that the usual algebraic rules for logarithms, such as
log(z1 z2 ) = log z1 + log z2 , etc., are satisfied in the complex case also (where
41
distinct branches of the logarithms are understood; if we have the n1 branch
of log z1 and the n2 branch of log z2 , then in general we obtain the (n1 + n2 )-
branch of log(z1 z2 )). To see this we note that z1 z2 = |z1 z2 |ei(θ1 +2n1 π+θ2 +2n2 π)
(with obvious notation), so that
using the real result for ln(|z1 ||z2 |) in the second line. Also, we note that for
z = |z|ei(θ0 +2nπ) we can define a multifunction
Thus, by traversing the contour and returning to the original point, we have
moved from one branch of the function (30) to the other (31) (clearly, this
branch as defined is discontinuous across the positive real axis). Likewise,
if we start with the n = 1 branch (31) of the square-root and allow z to
traverse the same circular contour then we see that this branch has
�
√ 1/2 iθ/2 −|z|1/2 θ = 0,
z = −|z| e =
|z|1/2 θ = 2π − ,
42
and we move from the n = 1 branch back to the n = 0 branch.
Similar results hold for the other multifunctions. If we start with the
n-branch of the mth root function (33) then on the circular contour |z| = r
we have
43
Definition 3.27 (Branch-point at infinity) We say z = ∞ is a branch point
of the multifunction f (z) if the point 0 is a branch point of the multifunction
fˆ(z) = f (1/z).
It is clear from these definitions that both 0 and infinity are branch points of
the multifunctions (29), (32), (33) and (34) (to see that ∞ is a branch point
of (34) use (35)).
Clearly the branch is not continuous on the contour (we have moved from
the n to the (n + 1)-branch), and so z = 0 is a branch point of fˆ(z), thus
z = ∞ is a branch point of f .
44
along (−∞, 0]. Such cuts in the complex plane are known as branch cuts,
since they effectively make individual branches of the multifunctions single-
valued. Normally a single branch of the multifunction suffices in applications;
a convenient choice is often the principal branch (defined using the principal
argument). Most of the multifunctions we shall consider are not only single-
valued in the cut plane, but are also analytic there.
Example: The square-root function. We cut the plane along the real
positive axis, and work with the principal argument, which is constrained to
lie in the interval 0 ≤ θ < 2π. This argument function is continuous in the
cut plane, since points of the cut are excluded. The principal branch of the
square-root function is then just the n = 0 branch (30):
taking values |z|1/2 on the top side of the cut, θ = 0+ , and −|z|1/2 on the
bottom side of the cut, θ = 2π − . [The n = 1 branch (31) is also single-valued
in this cut plane, the values taken on the upper and lower sides of the cut
being reversed from the n = 0 case.]
The discussion above suggests that branch cuts will always be taken between
branch points, so as to outlaw contours that encircle a branch point (by def-
inition 3.26, any such contour would lead to a discontinuity in f ). There
is usually a certain amount of freedom in the choice of branch cuts, and
convenience dictates how we choose them in a given situation. It is fairly
straightforward to define suitable branch cuts in the simple examples of mul-
tifunctions discussed so far, but we shall now consider more complicated
multifunctions for which the choice is not so obvious.
45
3.6.1 Example: Product of square-roots
We consider first a generalization of the square-root function,
As with (29) only the values n = 0, 1 give distinct results (other values of
n repeat these), so there are just two distinct branches. It is clear that
if we let z vary along any simple closed contour γa that surrounds just a
and not b then, over one complete circuit of γa , θa changes by 2π while θb
returns to its original value. Thus, the quantity i(θa + θb )/2 appearing in the
exponent in (37) has a net change of iπ, and so f (z) moves from one branch
to the other as we traverse γa once. In the same way, if z traverses a simple
closed contour γb that encloses b and not a, then θb changes by 2π over one
complete circuit, while θa returns to its original value. However, if z makes
one complete circuit of a simple closed contour γab that encloses both a and
b then, since both θa and θb change by an integer multiple of 2π, the quantity
(θa + θb )/2 appearing in the definition (37) changes by an integer multiple of
2π also. Thus in this case f (z) returns to its original value on completion of
the circuit.
Looking at the definition of branch-points again, it is clear that both a
and b are branch points for the function (37). To check the point at infinity,
note that
1
f˜(z) = f (1/z) = [(1/z − a)(1/z − b)]1/2 = [(1 − az)(1 − bz)]1/2 ,
z
which, although it is singular at z = 0, does not have a branch point there.
So z = ∞ is not a branch point of (37).
46
The above discussion about contours suggests that we should take a
branch-cut between a and b in the z-plane, since a cut here prevents z varying
continuously along contours such as γa and γb , but allows z to make excur-
sions that encircle both a and b. Again, a cut may in principle be made along
any curve joining the two points, but the simplest choice is to cut along the
straight line [a, b].
Looking at the definition (37), we can now check that such a branch-cut
leads to a single-valued function. For simplicity in this checking we shall look
at the case a = −1, b = 1, where the branch cut runs along [−1, 1] in the
z-plane. We take the n = 0 branch and restrict θa and θb to be the principal
arguments:
1/2 1/2
f0 (z) = [(z + 1)(z − 1)]1/2 = r−1 r1 ei(θ−1 +θ1 )/2 , 0 ≤ θ−1 , θ1 < 2π.
For most of the cut plane it is evident that f0 (z) is continuous and single-
valued, because θ−1 and θ1 are. The definition is only problematic as z
crosses the (allowed part of) the positive real axis, z ∈ R+ , |z| > 1, where
θ−1 and θ1 have jump discontinuities. Just above the positive real axis, where
arg(z) = 0+ (and |z| > 1) we have θ−1 = θ1 = 0; so here
1/2 1/2
f0 (z) = r−1 r1 .
Just below this part of the positive real axis, where arg(z) = 2π − , we have
θ−1 = θ1 = 2π; so here
1/2 1/2 1/2 1/2
f0 (z) = r−1 r1 ei(2π+2π)/2 = r−1 r1 .
so this branch also will be continuous along any contour that encloses both
1 and −1.
47
3.6.2 Example: Logarithm of a product
We next consider a logarithmic function with a more complicated argument:
In this case, any (simple closed) contour for which a complete circuit gives a
discontinuity in either θa or θb , will also give a discontinuity in f (z). Thus
we must cut the plane so as to prevent z making a complete circuit around
either or both of the points a or b. The way to do this is to make two cuts,
one from a to ∞, and the other from b to ∞. The fact that our cuts must
extend to infinity suggests that this is another branch point, and we can show
this explicitly by considering
Because of the log z term, a small circuit around z = 0 (that does not also
enclose z = 1/a or z = 1/b) will lead to a discontinuity in arg(z) when a
complete circuit is made, and hence a discontinuity in f (1/z), showing that
z = ∞ is indeed also a branch point of this function.
For the specific case a = −1, b = 1, suitable branch-cuts are from the
point z = −1 to infinity, along the negative real axis, and from the point
z = 1 to infinity, along the positive real axis.
48
to definitions of functions with several (possibly infinitely many) branches,
characterized by an integer n. Each branch corresponds to a different se-
lection of argument of the complex variable, each choice of argument being
restricted to an interval of length 2π.
This led us to think about how to define a single-valued (analytic)
branch of the function. In each example considered we chose the simplest
branch (usually the principal n = 0 branch defined via the principal argu-
ment(s)), though we could, of course, use any of the branches to construct
a single-valued function. Single-valuedness is ensured by introducing appro-
priate branch cuts in the complex z-plane, which prevent z from moving
in the plane in any manner that would lead to a discontinuity in f . Only
circuits around certain points in the z-plane can induce such discontinuities;
such points are known as branch-points of the function (definitions 3.26
and 3.27).
We saw that if we do not restrict the plane by branch-cuts, then f is
discontinuous along certain lines in the z-plane. Mathematically, as z crosses
such lines, we move from one branch of the function to another. An alterna-
tive way of rationalizing the structure of multifunctions is to suppose that in
fact a multifunction is defined on several ”copies” of the complex plane, one
for each branch. As z encircles a branch point, we move from one copy of
the plane to another (for the function log z, for example, this complex plane
would be something like a helix: see figure 2.3.6 in [1]). Such a multi-layered
complex plane is known as a Riemann surface, with each complete copy of
C being a single sheet of this surface. For the square-root function the Rie-
mann surface is more analogous to a Mobius strip; two complete circuits of
the Riemann surface bring us back to our starting point.
49
4 Complex Integration
With the ideas of paths through complex space introduced in §§2.4 and 2.4.1,
and our notions of complex functions, we are now in a position to integrate
complex-valued functions along paths in the complex plane.
50
cumbersome, and parametrization is clearly the best option. We now consider
an example where it is best to stick with the cartesian representation.
�
Example 4.2 (Splitting into real and imaginary parts) Evaluate (a) γ z dz,
�
(b) γ z dz, where γ is (i) the straight line joining z0 = 0 to z1 = 1+i; (ii) the
parabolic contour �(z) = �(z)2 joining the same two points in the complex
plane.
(a(i)) We write the integral as real and imaginary parts,
� � �
I = z dz = (x − iy)(dx + idy) = xdx + ydy + i(xdy − ydx),
γ γ γ
(a(ii)) On the second curve y = x2 and so the above result gives the integral
as
� � 1 � 1
3
I = xdx + ydy + i(xdy − ydx) = (x + 2x ) dx + i (2x2 − x2 ) dx
γ 0 0
= 1 + i/3.
51
Observe that in case (a) we get a different result for the two contours, hence
a nonzero result when we integrate z around the simple closed contour made
by joining the two curves together. In case (b) however the two integration
curves give the same result, and hence integrating z around the simple closed
contour made by joining the curves will give zero – again, this is due to the
fact that z is a complex analytic function while z is nowhere analytic.
Example 4.3 (Example 4.1 revisited) Replace the integrand in example 4.1
by a/z + bz:
� �a �
I= + bz dz.
C(0;r) z
we have
� � b
f (z) dz = f (z(t))z � (t) dt. (40)
γ a
52
Alternatively, for a general function f (z) = u(x, y) + iv(x, y) we have
� �
f (z) dz = (u(x, y)dx − v(x, y)dy + i(u(x, y)dy + v(x, y)dx), (41)
γ γ
Homework: Ablowitz & Fokas, problems for section 2.4, questions 1,2,3,4.
53
� b
d
= (F (z(t))) dt
a dt
= F (z(b)) − F (z(a))
= F (z2 ) − F (z1 )
(we go from the third line to the fourth line by applying the real Fundamental
theorem of calculus to the real functions �(F (z(t))), �(F (z(t)))).
If the contour γ is only piecewise smooth the result still follows by using
(42) and following through the above steps for each sub-curve γk . Contribu-
tions from the ”internal” endpoints cancel.
Corollary 4.5 For any closed contour γ lying within the domain of ana-
lyticity D of the function F , if f = F � on D, we have
�
f (z) dz = 0.
γ
Remark Note that it is certainly not the case in general for real-valued
functions u(x, y), even when perfectly smooth, that the real line integral
around a simple closed contour in R2 vanishes. For example, integrating
the function u(x, y) = x2 + y 2 around the contour x2 + y 2 = 1, we have
(transforming to polar coordinates in which ds = dθ on γ):
� � 2π
2 2
(x + y ) ds = 1 dθ = 2π.
γ 0
54
Corollary 4.6 Let f (z) = F � (z) for F (z) analytic on a domain D ⊂ C.
If γ1 and γ2 are two (piecewise smooth) curves lying in D, with the same
endpoints z1 and z2 , then
� �
f (z) dz = f (z) dz,
γ1 γ2
that is, the result of integrating between two given points in D is path-
independent.
We note that this last result does not hold for the example 4.2(a), since the
function f (z) = z is nowhere analytic. However it does hold for 4.2(b).
Proof (Theorem 4.7) We know that for a real-valued function g(t) integrable
on [a, b]
�� b � � b
� �
� g(t) dt�� ≤ |g(t)| dt. (44)
�
a a
We have
�� � �� b � � b
� � � �
� f (z) dz � = � � �
f (z(t))z (t) dt� = eiφ
f (z(t))z � (t) dt
� � �
γ a a
55
for some φ ∈ R. Then
�� � �� b �
� � � �
� f (z) dz � = � �[e f (z(t))z (t)] dt�� .
iφ �
� � �
γ a
Applying the real inequality (44) with g(t) = �[eiφ f (z(t))z � (t)], we find
�� � � b
� �
� f (z) dz � ≤ |�[eiφ f (z(t))z � (t)]| dt
� �
γ a
� b
≤ |eiφ f (z(t))z � (t)| dt
a
� b
≤ |f (z(t))z � (t)| dt.
a
Corollary 4.8 Suppose γ is a path with parametrization (39), and that the
function f (z) is continuous on γ, with |f (z)| ≤ M ∀z ∈ γ. Then
�� � � b
� �
� f (z) dz � ≤ M |z � (t)| dt ≡ M L(γ).
� �
γ a
�b
Here, L(γ) = a |z � (t)| dt is the length of the contour. It is easily checked
that this definition of contour length agrees with that for the length of a real
curve in R2 . Simply switch from the given parametrization t to arclength
parametrization s, with respect to which γ is specified by (x(s), y(s)), with
endpoints at s = sa and s = sb . Then note that (dz/dt)dt = (dz/ds)ds, so
that
� sb � � sb � b
�
L(γ) = 2 2
xs + ys ds = |z (s)| ds = |z � (t)| dt.
sa sa a
The proof of corollary 4.8 is trivial. The inequality is also often written as
�� � � b
� �
� f (z) dz � ≤ M |dz|,
� �
γ a
56
Note: This circular contour arises frequently, and we will usually denote it
by C(0; R). More generally, the circle with center a and radius r is denoted
C(a; r).
Points z(t) on the contour are given by z(t) = Reit , 0 ≤ t < 2π, so that the
estimation theorem 4.7 gives
�� � � 2π � � � 2π
� dz � � iReit � R 2πR
� � � �
� z2 + 1 � ≤ � R2 e2it + 1 � dt ≤ R2 − 1
dt =
R2 − 1
→ 0 as R → ∞.
γ 0 0
[Results such as this can be useful in calculating infinite sums, as we will see
later.]
We split the contour into its four straight-line segments: γ1 : �(z) = N ,
−N < �(z) < N ; γ2 : �(z) = N , −N < �(z) < N ; γ3 : �(z) = −N ,
−N < �(z) < N ; γ4 : �(z) = −N , −N < �(z) < N , and consider each
segment separately.
Our strategy here is to bound tan πz on each part of the contour, which
is not trivial since, as we now know, the sine and cosine function are not
bounded in the complex plane. On γ1 z = N + iy, and
� � � �
� sin πz � � sin π(N + iy) �
| tan πz| = �� �=� �
cos πz � � cos π(N + iy) �
� �
� sin N π cos iπy + cos N π sin iπy �
= �� �
cos N π cos iπy − sin N π sin iπy �
� � � �
� sin iπy � � sinh πy �
= � � � = � � = | tanh πy| ≤ 1.
cos iπy � � cosh πy �
57
� �
� sin πx cos iπN + cos πx sin iπN �
= � �
� cos πx cos iπN − sin πx sin iπN �
� �
� sin πx cosh πN + i cos πx sinh πN �
= � �
� cos πx cosh πN − i sin πx sinh πN �
sin2 πx cosh2 πN + cos2 πx sinh2 πN
⇒ | tan πz|2 =
cos2 πx cosh2 πN + sin2 πx sinh2 πN
cosh2 πN − cos2 πx
=
cosh2 πN − sin2 πx
cosh2 πN
≤ .
cosh2 πN − 1
Clearly this is bounded for large values of N (it becomes arbitrarily close to
1 as N increases) – in fact, for N ≥ 1 it is always less than 1.01! Similarly,
tan πz is bounded on γ4 . So, we can say that
| tan πz| ≤ M on γN ,
for some fixed M > 0. The condition on the function φ(z) gives
|zφ(z)| ≤ �N on γN ,
58
4.3 Harmonic functions to construct analytic functions
We now recall Lemma 3.24 stated, but not proved, earlier (and also our
comments when considering the Cauchy-Riemann theorem, that the real and
imaginary parts of a complex analytic function are closely related). With our
knowledge of complex integration we are now in a position to prove the result,
restated below:
Theorem 4.12 (Green’s theorem in the plane) Let p(x, y), q(x, y) be real-
valued functions, with continuous partial derivatives in a simply-connected
region S made up of the interior of a simple closed contour γ (described in
the usual anticlockwise sense), and the contour itself. Then
� � � � �
∂q ∂p
(pdx + qdy) = − dxdy.
γ S ∂x ∂y
The proof of this theorem can be found in any textbook on vector calculus.
59
where S is a spanning surface within D for the closed curve C (this is where
the simple-connectedness is required), by Green’s theorem in the plane; and
the right-hand side of (45) vanishes because u is harmonic on D. Thus a
conjugate function v can always be defined in this way; and by construction
u and v satisfy the Cauchy-Riemann equations
vy = ux , vx = −uy ,
5 Cauchy’s theorem
5.1 Cauchy’s theorem I
We shall now prove the main result concerning complex integration, from
which many other useful theorems follow. There are many different ways
of proving this theorem; in general the fewer restrictions we impose in the
theorem statement, the more technical the proof required. We shall give
the most general theorem statement later, and for now restrict ourselves to
proving a more basic version.
60
analytic function – but we have not proved this yet! (This proof of the exis-
tence and continuity of (all) higher derivatives of f relies itself on Cauchy’s
theorem.) It is possible to prove Cauchy’s theorem without the assumption
of continuity on f � , but the proof is more technical.
With f � (z) continuous, u and v also have continuous partial derivatives (see
(25), (26)), thus Green’s theorem 4.12 is applicable, and
� � � � � � � � �
∂v ∂u ∂u ∂v
f (z) dz = − + dxdy + i − dxdy.
γ R ∂x ∂y R ∂x ∂y
Example 5.2 Evaluate the integral of f (z) = z n around the unit circle |z| =
1.
By Cauchy’s theorem we know that the result must be zero for integers n ≥ 0,
because f is analytic then. We can show the result directly by parametrizing
the contour by z = eiθ , dz = ieiθ , then
� � 2π � 2π
inθ iθ
f (z) dz = e ie dθ = i (cos(n + 1)θ + i sin(n + 1)θ)dθ
γ 0 0
� �2π � �2π
sin(n + 1)θ cos(n + 1)θ
= i + = 0, n �= −1.
n+1 0 n+1 0
So in fact the result is zero for a wider range of n than predicted by Cauchy’s
theorem. This was easy to demonstrate for the chosen circular contour, but
would be more difficult if we chose (e.g.) the square contour in the next
example. We will soon prove another theorem (again due to Cauchy) that
can predict the above result for all values of n and for all contours (Cauchy’s
Residue theorem).
Example 5.3 The integral of f (z) = sin z around the square contour γ with
sides �(z) = 0, π, �(z) = 0, π, is zero by Cauchy’s theorem.
61
We could of course have chosen a more complicated contour, but then we
would not be able (easily) to verify the result directly. For a direct verification
we parametrize each part of the contour separately, writing
Homework: Ablowitz & Fokas, problems for section 2.6, questions 1,3,4.
62
can actually be deduced from Cauchy’s theorem. However, as indicated
above, such a proof, though possible, is much more technical, requiring the
development of several topological ideas, and the proof of several preliminary
results. Since many of these concepts are interesting and useful in their own
right however; and since results such as infinite smoothness of f may be
deduced from Cauchy’s theorem (so it would be nice not to have to assume
any smoothness of f � in the proof); we shall investigate how Cauchy’s theorem
may be proved in more generality.
Our first step is to prove Cauchy’s theorem for an arbitrary triangular
contour.
Proof We use the notation ABC to denote the triangle formed by joining
points z = a, z = b, z = c, and we suppose γ is the triangle A0 B0 C0 .
Take points A1 , B1 , C1 to be the midpoints of edges B0 C0 , A0 C0 , A0 B0 ,
respectively, from which we can construct 4 more “ABC” triangles (A1 B1 C1 ,
A0 B1 C1 , A1 B0 C1 and A1 B1 C0 ). We denote these 4 triangles by γ1k , k =
1, . . . , 4 (the ”1” to denote that this is the first step in an iterative triangu-
lation process, and the k just as a convenient label for each sub-triangle).
Then we have
� �4 �
I := f (z)dz = f (z)dz,
γ k=1 γ1k
using (42) (line integrals along interior segments of the sub-triangles cancel).
The inequality (8) applied to this result gives
� � �� �
�4 �� � � �
� � � �
|I| ≤ � f (z)dz � ≤ 4 � f (z)dz � ,
� γ1k � � γ1k∗ �
k=1
where k∗ is the choice of k giving the maximal absolute value to the integral.
Label this triangle γ1k∗ as γ1 , and repeat the above argument, with sub-
triangles (γ2k ) of γ1 . This procedure, illustrated schematically in figure 3,
63
Figure 3: Iterative triangulation
� of the original contour γ (the outermost
triangle), so as to maximize | γk f (z) dz| at each step.
64
generates a sequence of triangles (γj ) (with γ0 = γ), such that, for all j ≥ 1,
Claim: The set ∩∞ j=0 I(γj ) contains a point Z common to all I(γj ). [Recall
that I(γ) is the interior of the contour γ, a bounded open set.]
Proof of claim: For each j choose zj ∈ I(γj ). The resulting sequence
(zj ) is bounded, since all points zj ∈ I(γ). By result 2 of §3.1, we have a
convergent subsequence with limit Z. For any j, Z is a limit point of the se-
quence {zk : k ≥ j} ⊂ I(γj ), and so (since I(γj ) is closed), Z belongs to I(γj ).
Now let � > 0 (arbitrarily small). f is differentiable at Z so, for some δ > 0,
and
�
(f (Z) + (z − Z)f � (Z))dz = 0, (48)
γN
by application of theorem 4.4 (recall that Z, f (Z), f � (Z) are all fixed con-
stants here). Hence, using (46)–(48), and the estimation theorem 4.7, we
have
�� � �� � �
� � � �
� f (z) dz � = � f (z) dz − (f (Z) + (z − Z)f �
(Z))dz � using (48)
� � � �
γN
� γN γN
65
Finally we use result (iii) listed above to deduce that
�� �
�
N �
�
|I| ≤ 4 � f (z) dz �� < �L(γ)2 .
γN
is analytic in D, with F � = f .
Here the integral is taken along the line segment [a, z]. Note that this the-
orem is immediately applicable to any f analytic on D, since we have just
proved that for such f , the integral along arbitrary triangular contours is
zero (theorem 5.4).
Proof Let z ∈ D. Since D is open, ∃δ > 0 such that B(z; δ) ⊂ D. Then for
all |h| < δ, z + h ∈ D. For such h, the three line segments [a, z], [z, z + h],
[a, z + h], all lie in D (convexity), and form a triangle within D. The integral
of f around this triangle must be zero, by the theorem hypothesis. Hence we
66
have
� �
F (z + h) − F (z) = f (w)dw − f (w)dw
[a,z+h] [a,z]
� �
= − f (w)dw − f (w)dw
[z+h,a] [a,z]
�
= f (w)dw.
[z,z+h]
We then have
� � �� �
� F (z + h) − F (z) � 1 �� �
� − f (z)� = (f (w) − f (z))dw �
� h � |h| � [z,z+h] �
1
≤ |h| sup |f (w) − f (z)|
|h| w∈[z,z+h]
by the estimation theorem result 4.8. The right-hand side here goes to zero
as h → 0 by continuity; and the theorem is proved.
For convenience, we restate this result specifically for analytic functions (for
which it holds automatically as noted above):
67
An immediate consequence of theorem 5.7 is that for any function that is
complex analytic on a convex domain D, its integral along a path between
two points a, b ∈ D is path independent – the result is the same no matter
which path within D we choose.
The next version of Cauchy’s theorem that we give (theorem 5.10 below)
relates to the integral of an analytic function around an arbitrary simple
closed contour, and is sufficient for most purposes. Since we cannot assume
that we are dealing with contours composed of simple line segments as in the
results derived so far, we must first prove an important preliminary result,
which we now state:
Informally, this theorem guarantees that when a contour lies inside an open
subset D ⊂ C, it may be entirely covered by a finite sequence of (N + 1)
overlapping balls, of fixed radius, such that all the balls lie within D. The
construction also partitions the contour into N individual segments, each of
which is entirely contained within one of the balls.
ρ(z) ≤ |z − w| ≤ |z − z � | + |z � − w|
⇒ |z � − w| ≥ ρ(z) − |z � − z|.
68
Taking the infimum over w �∈ D we find
ρ(z � ) ≥ ρ(z) − |z − z � |.
ρ(z) ≥ ρ(z � ) − |z − z � |;
Thus ρ(z) is a continuous function; and on the compact set represented by the
contour γ, ρ is bounded and attains its bounds (theorem 3.13). In particular
it attains its infimum at some point z ∗ ∈ γ. Then, because D is open and
z ∗ ∈ D, ∃δ > 0 such that B(z ∗ ; δ) ⊂ D. By definition of the infimum on the
set γ, and of the distance function ρ
ρ(z) ≥ ρ(z ∗ ) ≥ δ ∀z ∈ γ.
with a similar result for �(z(t)). The continuous real functions �(z � (t)),
�(z � (t)) are defined on the compact set [a, b], so are bounded (and attain
their bounds) on [a, b]. Let M be a common bound for these functions. We
therefore have
69
which together imply (using (8)), that
Proof Note that if γ is a polygonal contour then the result is immediate from
Cauchy’s theorem for triangular contours (theorem 5.4). As noted earlier,
we can simply triangulate the polygon and write
� ��
f (z) dz = f (z) dz,
γ j γj
where the γj are the triangulating contours (the integrals along the interior
line segments are each covered twice, in the opposite sense, and thus cancel).
Since the integral around each γj vanishes (theorem 5.4), the result follows.
For an arbitrary contour γ, our approach is to approximate γ by a
polygonal contour. We use theorem 5.8 to cover γ by overlapping discs
Bk = B(z(tk ); δ), k = 0, . . . , N satisfying all conditions listed there and,
additionally, with z(tN ) = z(t0 ) (a closed contour with z(a) = z(b)). Within
each disk, which is a convex domain, Cauchy’s theorem 5.7 applies, giving
� �
f (z) dz = f (z) dz,
γk [zk ,zk+1 ]
where γk is the restriction of the contour γ to the parameter interval [tk .tk+1 ],
and [zk , zk+1 ] is the line segment from zk = z(tk ) to zk+1 = z(tk+1 ). Hence,
the total integral along the contour γ is equal to the integral along the closed
70
polygonal contour γP made by joining the line segments [z0 , z1 ], [z1 , z2 ], . . . ,
[zN −1 , zN ],
� N
� −1 � �
f (z) dz = f (z) dz = f (z) dz.
γ k=0 [zk ,zk+1 ] γP
The result then follows, by the noted result for polygonal contours.
Remark We note that we require the concept of the interior of the con-
tour (introduced in definition 2.20), since it is not possible to triangulate an
unbounded domain. Hence the theorem is applicable only to the contour’s
interior. This appeal to triangulation in the final step is also the reason
why the theorem works only for a simply-connected domain – we cannot
triangulate a closed polygonal curve on a non-simply-connected domain, in
general.
The analyticity of f in the region between the two contours means that one
contour can be ”deformed” smoothly into the other, without changing the
value of the integral. The proof is almost immediate from Cauchy’s theorem
5.10:
Proof (of theorem 5.11). Introduce two adjacent straight-line contours: γ12 ,
joining γ1 to γ2 , and γ21 = −γ12 , joining γ2 to γ1 . Then the composite contour
γ1 ∪ γ12 ∪ (−γ2 ) ∪ γ21 is a simple closed contour, within which f is analytic.
By Cauchy’s theorem 5.10 then,
� � � �
f (z) dz + f (z) dz + f (z) dz + f (z) dz = 0.
γ1 γ12 −γ2 γ21
� � � � �
Since γ21
= −γ12
=− γ12
and −γ2
=− γ2
the result follows.
71
We also have the contour deformation theorem for two open contours with
the same endpoints:
For completeness, we note the other forms of Cauchy’s theorem that may
be proved:
Clearly this definition always gives zero for points w0 that lie exterior to
the closed contour, by Cauchy’s theorem 5.10 (because the function f (z) =
1/(z − w0 ) is analytic inside and on the contour). However, if w0 lies inside
72
the contour the integrand has a singularity there, and we expect a nonzero
result in general. Consider the simplest case in which γ = C(w0 , r), traversed
N times. We parametrize γ by γ = {w(t) : w(t) = w0 + reiθ , 0 ≤ θ ≤ 2N π}
then the formula (49) gives
� 2N π
1 ireiθ
n(γ; w0 ) = dθ = N,
2πi 0 reiθ
so the definition holds true for this case at least. More generally then the
result should hold true by adapting the contour deformation results we have
already proved – an arbitrary contour wrapping N times round w0 being
equivalent to the circle wrapping N times around it.
The definition 5.14 is given meaning by the following theorems. We state
and prove the theorems for w0 = 0, but the results generalize easily to nonzero
w0 . Not surprisingly, given its definition and our knowledge of real calculus,
the winding number is closely related to the complex logarithm that we came
across in §3.5.2. We first prove a result on conditions under which an analytic
branch of the logarithm may be found on a domain D, and then use this result
to deduce properties of the winding number.
Proof As in the proof of theorem 5.5, it is easily checked that with the
definition (50) we have f � (z) = 1/z and, since 0 �∈ D (so that 1/z is analytic
on D), f is analytic on D. (Note that, without loss of generality, we may in
fact take the contour γ joining a and z to be a straight line, since the definition
is path-independent – this is easily shown by the deformation theorem 5.12
just proved.)
73
We also have
d
(ze−f (z) ) = e−f (z) − zf � (z)e−f (z) = 0.
dz
Hence, z = Cef (z) (C �= 0). Since we seek f such that ef (z) = z, C = 1.
To check uniqueness, suppose that there exists another function g(z) an-
alytic on D such that ef (z) = eg(z) = z for all z ∈ D. Differentiation of both
sides of this relation gives zf � (z) = zg � (z) ⇒ f � (z) = g � (z) for all z ∈ D, and
thus f (z) − g(z) = K, constant, where eK = 1. This gives K = 2nπi; thus
all such representations f differ by additive multiples of 2πi.
The final part of the theorem follows as in our definition of the complex
logarithm earlier, using the relation z = ef (z) with z = |z|eiθ and f = u + iv,
which gives |z| = eu and v = θ (modulo 2π).
74
and
The second part of the theorem relates to the fact that the sequence of loga-
rithms fk defined above do not necessarily give a single continuous function
when “patched” together; but if one defines an argument function Θ suit-
ably, then one can ensure continuous variation of argument as the contour γ
is followed.
75
To do this, we use the parametrization of the contour, and first define
argument functions on each section of γ: Θk (t) = θk (z(t)), t ∈ [tk , tk+1 ].
Each Θk (t) is continuous (since z(t) and θk (z) are), but in general there is
no continuity between Θk and Θk+1 at t = tk+1 (they differ by an integer
multiple of 2π: Θk+1 (tk+1 ) − Θk (tk+1 ) = 2πnk ). If we define Θ(t) recursively
by
�
Θ0 (t) t ∈ [t0 , t1 ]
Θ(t) = (52)
Θk (t) − (Θk (tk ) − Θ(tk )) t ∈ (tk , tk+1 ], 1 ≤ k ≤ N − 1,
Θ(t+
1 ) = Θ1 (t1 ) − (Θ1 (t1 ) − Θ(t1 )) = Θ(t1 ),
Θ(t+
k ) = Θk (tk ) − (Θk (tk ) − Θ(tk )) = Θ(tk ).
showing explicitly that Θ(t), since it only ever differs from one of the argu-
ment functions by an integer multiple of 2π, is indeed an argument of the
complex number z on γ. This proves part 2(i).
76
Finally, using (53) with t = tN = b and k = N − 1 we have
N
� −1 N
�
Θ(b) = ΘN −1 (b) − 2π nj−1 = ΘN (b) − 2π nj−1 .
j=1 j=1
Recalling that ΘN (b) = θN (zN ) = θ0 (z0 ) = Θ0 (a) = Θ(a) (using the equality
of the complex logarithms fN and f0 , and (52)), this leads to
N
�
Θ(b) − Θ(a) = −2π nj−1 ≡ 2πn(γ; 0),
j=1
proving 2(ii).
77
Theorem 6.1 (Cauchy’s Integral Formula) Let f (z) be analytic inside and
on a simple closed contour γ. Then for any z inside γ (z ∈ I(γ)),
�
1 f (w)
f (z) = dw. (54)
2πi γ w − z
This theorem is remarkable – it says that, if a function is known to be ana-
lytic inside and on a given curve, then its values on that curve are sufficient
to determine uniquely its values everywhere inside the curve. Recall the
connection that we know exists between harmonic functions and analytic
functions – we know from PDE theory (or we should know!) that, if we wish
to solve Laplace’s equation in a given domain, then the solution u is uniquely
determined by the values of u at all points of the boundary. We also know
(lemma 3.24) that a harmonic conjugate v of u may be defined, unique up to
an additive constant, such that u + iv is a complex analytic function. Hence
theorem 6.1 does fit with several facts we already knew. What we cannot
do, however, is arbitrarily prescribe the values of a function f on a curve
γ and expect (54) to yield an analytic function at all points of the interior of
γ. As the results for Laplace’s equation suggest, in fact we can really only
arbitrarily prescribe either the real or imaginary part of an analytic function
on a simple closed contour γ – and this is sufficient to uniquely determine
the values of the corresponding analytic function at all points within γ (at
least, up to an arbitrary constant). The real and imaginary parts of f are
closely related, and cannot be prescribed independently, even on a contour
in C.
Proof Let z ∈ I(γ), then since I(γ) is open ∃� > 0 such that B(z; �) ⊂ I(γ).
For any 0 < δ < �,
� � � �
f (z) f (z) f (w) f (w)
dw = dw, and dw = dw,
γ w−z C(z;δ) w − z γ w−z C(z;δ) w − z
(where C(z; δ) denotes the circle with center z and radius δ), by the defor-
mation theorem 5.11. Thus, using the result
� � 2π
dw iδeiθ dθ
= = 2πi,
C(z;δ) w − z 0 δeiθ
in which we parametrized C(z; δ) by w = z + δeiθ , 0 ≤ θ < 2π, we have
� � � � � �
� 1 f (w) � � 1 f (w) − f (z) ��
� dw − f (z)� = � dw�
� 2πi w − z � � 2πi w−z
γ C(z;δ)
78
� � �
� 1 2π
f (z + δeiθ ) − f (z) iθ ��
= �� iδe dθ�
2πi 0 δeiθ
1
≤ sup |f (z + δeiθ ) − f (z)|.
2π 0≤θ<2π
Proof For k = 0 this is just theorem 6.1 proved above. So, we suppose (55)
holds for k = n, and show that it must then hold for k = n + 1 (the general
result will then hold, by induction). As above we may use theorem 5.11 to
deform the contour of integration to a (sufficiently small) circle, C = C(z; 2r)
for some fixed r > 0. We use the definition of derivative, (23), to try to
evaluate f (n+1) (z) from f (n) (z) which, by our induction hypothesis, is given
by (55).
We now let h ∈ C with |h| < r. Then
� � �
(n) (n) n! 1 1
f (z + h) − f (z) = f (w) − dw
2πi C (w − z − h)n+1 (w − z)n+1
� �� �
(n + 1)! 1
= f (w) n+2
dζ dw,
2πi C [z,z+h] (w − ζ)
79
again applying theorem 4.4. Since f is analytic, it is continuous, and therefore
bounded (by M , say) on the compact set C. Note that for any w ∈ C =
C(z; 2r), we have |w − z| = 2r, while |h| < r; and in the above integrals,
80
For (c) we use (55) with k = 2, f (w) = sin w/(w +4)2 and z = 1 to obtain
�
sin w 2πi ��
2 3
dw = f (1).
γ (w + 4) (w − 1) 2!
The integrals in this example could all be worked out by a similar, but much
more general method, using Cauchy’s residue theorem, which we shall soon
derive. (Both Cauchy’s integral formula and Cauchy’s formula for derivatives
may be thought of as special cases of the Residue theorem).
These representations (56) and (57) state that the value of f at the point
z is equal to its mean value integrated around the circle of radius R, and
is also equal to its mean value integrated over the area of the same circle,
respectively. [Of course, Cauchy’s integral formula (54) itself may be con-
sidered as a generalized mean value result relating the value of f at any
point z ∈ γ to a weighted average of its values along γ.] Proofs of (56)
and (57) are easy – for (56) just take the contour γ in theorem 6.1 to be
γ = {z ∈ C : z = Reiθ , 0 ≤ θ < 2π}. Then (57) follows by multiplying both
sides of (56) by r dr and integrating from r = 0 to r = R.
81
B(z; R) ⊂ D. Writing ζ = z + reiθ for any point ζ in this ball (0 ≤ r < R)
(57) then gives
� R � 2π
1
f (z) = f (ζ)rdrdθ
πR2 0 0
�� �
1 �� �
�
⇒ |f (z)| = f (ζ)dS
πR2 � B(z;R) �
�
1
≤ |f (ζ)|dS
πR2 B(z;R)
�
|f (z)|
≤ dS
πR2 B(z;R)
= |f (z)|,
where we used a variant of the estimation theorem 4.7 (for double integrals;
the proof is exactly similar) in the first inequality, and the contradiction
assumption in the second. Since we end up again with |f (z)| at the end of
this chain of inequalities we must have equality throughout; and in particular,
�
1
|f (z)| = |f (ζ)|dS
πR2 B(z;R)
�
1
⇒ (|f (z)| − |f (ζ)|)dS = 0.
πR2 B(z;R)
The integrand here is real and non-negative; so the only way it can integrate
to zero is if it is identically zero. This implies that |f (ζ)| = |f (z)| on B(z; R),
and hence f is constant on this ball. (The fact that f is constant on a
region where |f | is constant is a simple deduction from the Cauchy-Riemann
equations.)
The above shows that f must in fact be constant on the largest ball that
can fit around an interior maximum point z within D. Since all points of
this ball are now maxima of |f | we can repeat the argument for a different
point near the ball boundary and, proceeding in this way, cover the whole of
D with balls on which f must be constant. This proves part (i).
The proof of (ii) is almost immediate from (i). The continuity of the
real function |f | on the closed bounded region D (a compact set) guarantees
that |f | is bounded and attains its bounds; so we know that |f | does have a
maximum on D. Part (i) tells us that there are then two possibilities: if |f |
has a maximum in D, then f is constant on D and, by continuity therefore,
82
f is constant on D, so that |f | attains its maximum on ∂D trivially. The
second possibility is that |f | is non-constant, and has no maximum in D.
Then the only possibility is that the maximum of |f | on D occurs on ∂D.
Remark In fact a slightly stronger (though not much more useful) version
of Morera’s theorem could be stated above: we only require that f integrate
to zero around any given triangular contour in D (the construction used to
demonstrate that F � = f uses only this fact).
83
Entire functions are quite rare in the space of all complex functions, and
strong results can be proved for them. There is some ambiguity in the lit-
erature regarding the notion of analyticity of a function at infinity. Some
would define a function f (z) to be analytic at infinity only if f˜(z) := f (1/z)
is analytic at the origin. However, it is normally accepted that functions such
as complex polynomials (e.g. z n ) – which do not satisfy the above criterion
for analyticity at infinity, are entire.
Proof The proof uses the expression (55) for f � (z). Since f is entire we may
choose any contour γ we wish in this expression, and we choose the circle of
arbitrary radius R centered on z: |ζ − z| = R, or ζ = z + Reiθ , 0 ≤ θ < 2π.
Then, since we know |f | ≤ M for some M > 0, and dζ = iReiθ dθ, we have
�� � � � 2π
� 1 �� f (ζ) �
� 1 |f (ζ)| M M
|f (z)| = � 2
dζ � ≤ 2
|dζ| ≤ dθ = .
2π C (ζ − z) 2π C |ζ − z| 2πR 0 R
We note without proof the following much stronger result due to Picard:
Liouville’s theorem can be used to give a quick and elegant proof of the
fundamental theorem of algebra:
84
Proof The proof proceeds by contradiction: we suppose that p(z) �= 0 ∀z ∈
C. Since |p(z)| → ∞ as |z| → ∞, ∃ R > 0 such that 1/|p(z)| < 1 for
|z| > R. On |z| ≤ R, which is a compact set, 1/|p(z)| is continuous, and hence
bounded. Hence 1/p(z) is bounded on the whole complex plane. It is also
analytic (since p(z) nowhere vanishes; see homework question (2) in §3.3).
By Liouville’s theorem 6.8 then, p(z) must be constant. This contradiction
proves the theorem.
Theorem 6.11 Suppose f (z) is an entire function such that lim|z|→∞ |f (z)/z N | =
M , 0 < M < ∞. Then f (z) is a polynomial of degree N .
Theorem 7.1 (Taylor series) Suppose f (z) is analytic on B(z0 ; R). Then
∞
� f (n) (z0 )(z − z0 )n
f (z) = . (58)
n=0
n!
This is the Taylor series expansion of f about the point z0 , and it converges
uniformly on |z − z0 | ≤ R.
85
The theorem utilizes Cauchy’s integral formula, and in the proof we will need
to interchange the order of summation and integration. This is always fine
for a finite sum, but can pose problems when the sum is infinite, as in the
theorem statement. We shall therefore first prove the following theorem (and
a corollary):
By uniform convergence on D here we mean that, for any z ∈ D, and any � >
0, ∃ N , dependent on � but independent of z, such that |Gn (z) − G(z)| < �
for all n ≥ N .
so that
|G(z) − G(z0 )| ≤ |Gn (z) − Gn (z0 )| + |Gn (z0 ) − G(z0 )| + |G(z) − Gn (z)|.
|Gn (z0 ) − G(z0 )| < �1 /3 and |Gn (z) − G(z)| < �1 /3, ∀n ≥ N.
86
convergence of the sequence Gn to G we know we can find N such that, for
n ≥ N , |Gn (z) − G(z)| < �/L(γ). Then,
�� � � �
� �
� Gn (z) dz − G(z) dz � ≤ |Gn (z) − G(z)||dz| < � L(γ) = �,
� � L(γ)
γ γ γ
Proof (of Taylor’s theorem 7.1) We use Cauchy’s integral formula. Let
z0 ∈ B(z0 ; R) and choose r > 0 such that |z − z0 | < r < R. Let γ = C(z0 ; r),
then (54) gives
�
1 f (w) dw
f (z) = . (60)
2πi γ (w − z)
We write the factor 1/(w − z) here as
1 1 1
=
w−z (w − z0 ) (1 − [(z − z0 )/(w − z0 )])
and, noting that for w ∈ γ we have |z − z0 | < |w − z0 |, we can expand the
right-hand side here as a convergent series. Thus, (60) becomes
� � ∞
1 (z − z0 )n
f (z) = f (w) dw. (61)
2πi γ n=0 (w − z0 )n+1
87
�
where Mn converges, since |z − z0 | < r. By theorem 7.2 and its corol-
lary 7.3, summation and integration may be interchanged in (61), since the
integrand is uniformly convergent on γ, giving
�∞ � � �
1 f (w)
f (z) = n+1
dw (z − z0 )n .
n=0
2πi γ (w − z0 )
From Cauchy’s formula for derivatives (55), we note that this is exactly
∞
� f (n) (z0 )
f (z) = (z − z0 )n ,
n=0
n!
Strictly speaking, the proof above shows only that the expansion (58) is
one possible power series expansion of a given analytic function f about the
point z0 . Unsurprisingly, however, the power series expansion of f about z0
is unique, as we will shortly prove.
Example 7.4 Find the Taylor series expansion of (a) f (z) = ez about z = a,
(b) f (z) = 1/z about z = 1.
For (a) we note that f (n) (z) = ez , ∀n, and so we have the Taylor series
∞
� ea (z − a)n
f (z) = . (62)
n=0
n!
88
convergent for |z − 1| < 1. Here we note that the function f can be rewritten
as
f (z) = z −1 = (1 + (z − 1))−1
which, for |z − 1| < 1 can be expanded binomially to give exactly the result
(63). We will show soon that the power series expansion for an analytic
function about a given point is unique, so that if, by any means, we are able
to find a power series expansion of the function about the given point, then
that expansion must be its Taylor series. The above example (b) illustrates
this fact. Alternatively, once we know that the Taylor expansion is unique,
we can view example (b) as a proof that the binomial theorem is valid here.
Similarly for example (a) there are alternative ways to obtain the power series
– we could write ζ = z − a and ez = ea+ζ = ea eζ . Then use the definition of
eζ , to obtain the same result as (62).
Example 7.5 (Binomial expansion) Show that, for positive integers n, the
binomial expansion
n
�
n n!
(z + h) = hn−k z k
k=0
(n − k)!k!
This example shows that the “standard” binomial expansion holds good for
complex as well as real variables.
89
The power series expansion in Taylor’s theorem 7.1 contains only positive
powers of (z − z0 ), reflecting the fact that the function f is analytic at the
point z0 (so the power series is well-behaved there). If we know only that f
is analytic in some annulus about z0 , R1 ≤ |z − z0 | ≤ R2 , then we will see
that we can find a more general Laurent series expansion for f (z) on the
annulus, that contains both positive and negative powers of (z − z0 ). Clearly
any function whose power series expansion about z0 contains negative powers
of (z−z0 ) is singular at the point z0 . We discuss singularities in detail below,
but first we note a couple of useful consequences of Taylor’s theorem 7.1.
The proof above demonstrates the uniform convergence of the Taylor
series about z0 for balls on which f is analytic. Therefore we are able to
differentiate the Taylor series term-by-term to obtain the Taylor series of the
derivative:
Taylor’s theorem gives us a recipe for finding convergent power series ex-
pansions of given analytic functions. It is also useful to be able to deduce
properties of a given power series. The following theorem tells us that, where
a given power series in z is uniformly convergent, it necessarily defines an
90
analytic function; and within its region of uniform convergence it may (like a
Taylor series) be differentiated term-by-term to obtain the function’s deriva-
tive.
Proof For the proof we take z0 = 0 for simplicity, noting that the result is
easily extended to general z0 by a shift of origin. Let g(z) be the power series
defined by
∞
� ∞
�
n−1
g(z) = nbn z = gn (z). (66)
n=1 n=1
Then g(z) has the same radius of convergence as f (z), because if we take
z ∈ B(0; R) and ρ ∈ (|z|, R), then
� �n
n−1 n |z|
|gn (z)| = |nbn z | = |bn ρn |.
|z| ρ
�
The series n(|z|/ρ)n converges for |z| < ρ, by the ratio test for real series.
Therefore its terms (which are all positive) are all bounded, n(|z|/ρ)n ≤ M
say, for some M > 0. Then
M
|gn (z)| ≤ |bn ρn |,
|z|
�
and so 0 < |z| < R, by the comparison
gn (z) is absolutely convergent for �
test. The absolute convergence of the series n gn (z) for z = 0 is immediate
from setting z = 0 in the sum; thus we have absolute convergence for�all |z| <
R. We note that the same argument applied �∞ to the power series n gn (z)
n−2
also guarantees absolute convergence of n=2 n(n − 1)bn z on B(0; R), a
result we shall use below.
91
We must now show the analyticity of the general power series (65) within
its radius of convergence: we shall show that f � (z) exists and is given by
g(z). Let z, z + h ∈ B(0; R), then
� � ��� � ���
� f (z + h) − f (z) � �∞ (z + h)n
− z n
�
� − g(z)�� = � bn − nz n−1 �
� h � n=1 h �
(the term in n = 0 is identically zero). By the binomial expansion (example
7.5), the term (z + h)n on the right-hand side here can be written
�n � �
n n
(z + h) = z n−k hk ,
k
k=0
giving
∞
� n � ∞
f (z + h) − f (z) 1� � n! k n−k n
�
− g(z) = bn h z −z − bn nz n−1
h h n=1 k=0
(n − k)!k! n=1
∞
� n
� �∞
n! k−1 n−k
= bn h z − bn nz n−1
n=1 k=1
(n − k)!k! n=1
�∞ �n
n!
= bn hk−1 z n−k
n=1 k=2
(n − k)!k!
∞
� �n � �
n
= h bn z n−k hk−2
k
n=1 k=2
� � �∞ n � � �
� f (z + h) − f (z) � �� � �
� n �
⇒ �� − g(z)�� = |h| � bn z n−k hk−2 �
h � k �
n=1 k=2
� ∞ n−2 �
� �
n(n − 1) n−2
≤ |h| |bn | |z|n−2−m |h|m
2 m
n=1 m=0
∞
� n(n − 1)
≤ |h| |bn |(|z| + |h|)n−2 , (67)
n=1
2
where we used the result
� �
n n(n − 1)(n − 2)! n(n − 1)(n − 2)!
= =
k k(k − 1)(n − k)!(k − 2)! (m + 2)(m + 1)(n − 2 − m)!m!
� �
n(n − 1) n − 2
≤
2 m
92
with m �= k − 2 ≥ 0. Choosing ρ ∈ (|z|, R), we now recall that, as shown
∞ n−2
above, n=2 n(n − 1)|bn |ρ converges (to K, say). It follows that for
|h| < ρ − |z|,
� � ∞
� f (z + h) − f (z) � |h| � K|h|
� �
− g(z)� ≤ n(n − 1)|bn |ρn−2 ≤ .
� h 2 n=1 2
Note The above proof can be easily adapted to the power series (65) for
arbitrary z0 �= 0.
Corollary 7.8 The power series defined by (65), with radius of convergence
R > 0, has derivatives of all orders within B(z0 ; R). In particular, f (n) (z0 ) =
n!bn .
It is immediate from this corollary that the coefficients in a power series
expansion are unique. In particular, the Taylor series expansion of theorem
7.1 is unique.
Example 7.9 Find the Taylor series expansions of (a) f (z) = e1/z and (b)
f (z) = (az − b)/(cz − d) about the point z = ∞.
�∞ n
For (a), note that f (ζ) = eζ = n=0 ζ /n! gives the Taylor series about
ζ = 0 (infinite radius of convergence). Therefore the Taylor series of f about
z = ∞ is
�∞
1
f (z) = e1/z = n
.
n=0
z n!
93
For (b), writing ζ = 1/z gives
�∞ � �n
a − bζ dζ −1 dζ
f (ζ) = = c(a − bζ)(1 − ) = c(a − bζ)
c − dζ c n=0
c
�∞ � � n �∞ � � n
d n d
= ac ζ − bc ζ n+1
n=0
c n=0
c
� ∞ � � n−1
d
= ac + (ad − bc)ζ n .
n=1
c
This power series converges absolutely for |ζ| < |c/d|. Thus the Taylor series
of the function about infinity is
� ∞ � �n−1
d
f (z) = ac + (ad − bc)z −n ,
n=1
c
Example 7.10 Can we find the Taylor expansion of the function ez about
z = ∞?
Again we write ζ = 1/z, then the power series expansion of e1/ζ about ζ = 0
is what we need. However, e1/ζ is not analytic in any neighborhood of ζ = 0!
Taylor’s theorem is not applicable, and there is no Taylor series expansion of
ez at z = ∞ – the function is not analytic there.
94
prove this from first principles) which, by Taylor’s theorem 7.1, has a power
series expansion
∞
�
f (z)g(z) = (f g)n z n ,
n=0
absolutely convergent on |z| < R, where the coefficients (f g)n are given by
the usual formula in theorem 7.1. We then have
∞
�∞ �� ∞ �
� � �
(f g)n z n = fn z n gn z n , |z| < R.
n=0 n=0 n=0
Comparison of terms on left- and right-hand sides here then gives the usual
formula
� n
(f g)n = fr gn−r
r=0
for the coefficient of z n , using the fact that the power series (Taylor series)
expansion of an analytic function is unique.
95
where cn = f (n) (a)/n!, and cm �= 0. An alternative characterization of a zero
of order m that can be useful is given by the following:
Applying the triangle inequality in the form |g(z) − g(a)| ≥ ||g(z)| − |g(a)||
we then have
This proves the result one way. The converse is easier; if f (z) has the assumed
representation in terms of g then the Taylor expansion of g must take the
form
∞
�
g(z) = cn+m (z − a)n , cm �= 0,
n=0
96
uniformly convergent on some B(a; δ), which gives
∞
�
f (z) = (z − a)m g(z) = cn (z − a)n , cm �= 0
n=m
as the unique Taylor expansion of f on B(a; δ). It follows then from the
definition 7.11, and from the definition of the Taylor coefficients in (58), that
f has a zero of order m at a.
Proof of theorem 7.13. By Taylor’s theorem, f has the (unique) power series
expansion
∞
�
f (z) = cn (z − a)n , z ∈ B(a; r). (69)
n=0
97
∞
�
(m)
⇒ f (z) = (n + m)(n + m − 1) . . . (n + 1)cn+m (z − a)n ,
n=0
together with the comparison test (the terms of g(z) have absolute value less
than or equal to those of f (m) (z), which we know converges).
Since g(a) �= 0 and g is continuous at a (theorem 7.2, or the stronger
theorem 7.7), g(z) �= 0 for z ∈ B(a; �), some � > 0. In the punctured ball
B � (a; �), f (z) = (z − a)m g(z) �= 0, so a is not a limit point of Z(f ).
In conclusion, if a is a zero of f , then either f ≡ 0 on B(a; r) (some r > 0),
or a cannot be a limit point of zeros (that is, it is an isolated zero). The
theorem follows.
Proof Apply the identity theorem to the function F (z) = f (z) − g(z). F (z)
is analytic on D, and its set of zeros S contains a limit point. Thus F (z) ≡ 0
by theorem 7.14.
98
We illustrate the results by example.
Example 7.16 Using the identity theorem, show that there exists a unique
analytic function f (z), defined on a domain D containing the origin, such
that
� �
1
f = (n2 − 1)−1 (70)
n
for all but finitely many n ∈ N. Give f (z).
Suppose there are two such functions, f1 and f2 . Then g(z) = f1 (z) − f2 (z)
has zeros at points zn = 1/n for all but finitely many n ∈ N. It follows that 0
is a limit point of zeros of g in D. By the identity theorem 7.14 then, g ≡ 0
in D. The function f is therefore unique, and the only such function is
z2
f (z) =
1 − z2
(replace n by 1/z in (70)).
Example 7.17 Let f (z) = u(x, y) + iv(x, y) be analytic on B(0; 1). Prove
that the function g(z) defined by
is analytic on B(0; 1). Show further that, if f takes real values on the imag-
inary axis, then
99
where we used the Cauchy-Riemann equations for u and v. Thus, the Cauchy-
Riemann equations hold for the real and imaginary parts of f (−z̄), and so
this function is analytic on B(0; 1).
For the second part, observe that z = iy = −z̄ on the imaginary axis,
and thus on this line
since f (iy) ∈ R. The zeros of the analytic function g are not isolated; there-
fore by the identity theorem we have g(z) = 0 on its domain of analyticity,
as required.
The indentity theorem 7.14 (or the uniqueness corollary 7.15) allows certain
functional identities to be extended from subsets of C to larger subsets, or
the whole of C.
Example 7.19 (The binomial theorem for negative integers n) Show that,
for any negative integer n,
∞
�
n n!
(1 + z) = zk , |z| ≤ 1.
k=0
(n − k)!k!
We write
∞
�
n n!
f (z) = (1 + z) , g(z) = zk .
k=0
(n − k)!k!
Clearly f is analytic except at z = −1, while the series defining g has radius of
convergence 1. By theorem 7.7 g then defines an analytic function on B(0; 1).
We know from real analysis that f (z) = g(z) when z ∈ R with z| < 1. The
uniqueness corollary 7.15 then tells us that f ≡ g for all z ∈ B(0; 1).
The result embedded in theorems 7.13, 7.14 and 7.15 represent an elemen-
tary form of a more general principle known as analytic continuation. The
100
basic idea of analytic continuation is to take limited information about an
analytic function f (e.g. its values on some small subset, such as a line or
curve in C), and use this information to deduce the values of the function
in some larger set (even perhaps the whole complex plane). This concept,
which will be explored in more detail in Math 756, is linked to our ear-
lier observations about the connection between analytic functions and real
“harmonic conjugate” functions (see remarks following the Cauchy Integral
formula, theorem 6.1), though it is more general. We observed there that
the values of a harmonic function u on some boundary curve are sufficient
to determine uniquely its values inside the boundary curve (or even on one
side of a semi-infinite boundary curve, provided we impose suitable bounds
on the growth at infinity). Once we have the harmonic function then we
can determine its harmonic conjugate v, and hence construct the associated
analytic function f = u + iv. Since we need only the real (or imaginary)
part of the analytic function on a curve to construct the whole analytic func-
tion away from the curve, it should be no surprise that knowledge of both
real and imaginary parts are certainly sufficient to “continue” the analytic
function away from the curve (or other subdomain in question). In fact if
anything, one has to worry about the compatibility of the data on the subset
– as we saw earlier, the real and imaginary parts of f cannot be prescribed
arbitrarily on a boundary curve or subset of C.
101
D), the identity theorem 7.13 tells us that a is not a limit point of zeros.
Therefore, ∃δ > 0 such that B(a; 2δ) ⊂ D and f (z) �= 0 on B � (a; 2δ).
We next note that f (z) is analytic (thus continuous) on the compact set
represented by C(a; δ) (C(a; δ) = {z : |z − a| = δ} ⊂ B(a; 2δ)); and so it is
bounded and attains its bounds on C(a; δ). We let 2m > 0 be the minimum
of |f (z)| on C(a; δ).
We shall show that B(f (a) = 0; m) ⊂ f (D), proving that f (D) is open
(since a ∈ D was arbitrary). Let w ∈ B � (0; m), so 0 < |w| < m; and for
a contradiction, assume that w �= f (z) for any z ∈ D. Then in particular,
w �= f (z) for any z ∈ B(a; 2δ), and so the function
m
h(z) =
f (z) − w
Note that this result is not true for infinitely differentiable real-valued
functions f (x). A simple counterexample is given by the function f (x) = x2 ,
which maps the open interval (−1, 1) to [0, 1), which is not open.
Example 7.21 The function f (z) = z z̄ takes the unit ball, B(0; 1), onto the
real interval 0 ≤ �(f (z)) < 1, �(f (z)) = 0. This is not an open subset of C,
since we cannot fit an open ball that lies within the set around any point of
the set. Therefore f cannot be analytic on B(0; 1) (as we already knew).
It is also true that analytic functions map compact sets to compact sets in
C – but note that we cannot prove this simply by taking complements in
the open mapping theorem. While it is true that Dc is closed if and only
if D is open, we have no guarantee that f (D)c = f (Dc ); and indeed this is
not true in general. For example, the simple function f (z) = z(z − 2) takes
both points z = 0 and z = 2 to the origin in the image plane. Therefore
for this function, taking D = B(0; 1), both f (D) and f (Dc ) have a point in
common (the origin), and thus they are not complementary sets. We will not
102
prove the result that compact sets are mapped to compact sets by analytic
functions.
The open mapping theorem allows us to give a very easy proof of the max-
imum modulus theorem 6.5 – which was used in proving the open mapping
theorem! The maximum modulus theorem (and its variants) has implications
for real harmonic functions, as we shall see. We now restate the theorem, in
a slightly different form, and give the simple proof.
Proof The open mapping theorem 7.20 guarantees that the image of a do-
main D under a non-constant analytic function f , f (D), is open. (In fact,
f (D) is a domain if D is a domain.)
(i) Let a ∈ D. Then f (a) ∈ f (D), and f (D), being open, contains a neigh-
borhood of f (a), and therefore a point of larger modulus.
(ii) Since the real function |g| is continuous on the closed bounded set D, |g|
has a maximum at some point zmax ∈ D,
Proof If f does not vanish on D then g(z) = 1/f (z) is analytic on D and
so, by theorem 7.22, attains its maximum modulus on the boundary ∂D (or
is constant, in which case the result follows trivially).
103
Proof Since u is harmonic on D it is the real part of a function f (z) analytic
on D. The function g(z) = exp(f (z)) is also analytic and nonvanishing on
D, and so attains its maximum modulus and its minimum modulus on the
boundary ∂D. Since
it follows that u(x, y) must attain its maximum and minimum values on the
boundary ∂D. (Note that this last equality above makes it clear that g(z) is
nonvanishing on D.)
Note: the desired result can also be deduced from the maximum modu-
lus principle, theorem 7.22, alone, by consideration of the analytic function
exp(−f (z)) to show that u(x, y) attains its minimum value on the boundary.
Homework: Ablowitz & Fokas, Problems for Section 2.1, questions 4,5,7.
8 Laurent expansions
Taylor’s theorem, giving power-series expansions about a given point z0 ,
only applies to functions that are analytic in a disk about z0 . If f fails
to be analytic at z0 , but is analytic in some annular region about z0 , then
a generalization of Taylor’s theorem may be found, in which negative as
well as positive powers of (z − z0 ) appear in the series expansion. Such a
series expansion for a function f is called a Laurent series. It is certainly
clear that such an expansion can be written down sometimes: consider for
2 2
example the function f (z) = ez /z 2 . We know that ez has a power series
�∞ 2n about the origin z = 0 that converges uniformly ∀z ∈ C, given by
expansion
n=0 z /n!. Therefore
∞ ∞
1 � z 2n � z 2n
f (z) = 2 = .
z n=0 n! n=−1
(n + 1)!
104
8.1 Laurent’s theorem
Theorem 8.1 (Laurent series) A function f (z) analytic in an annulus R1 ≤
|z − z0 | ≤ R2 may be written as
∞
�
f (z) = Cn (z − z0 )n (71)
n=−∞
in R1 ≤ |z − z0 | ≤ R2 , where
�
1 f (w)dw
Cn = (72)
2πi γ (w − z0 )n+1
and γ is any simple closed contour lying within the region of analyticity and
enclosing the inner boundary |z − z0 | = R1 . This series representation for f
is unique, and converges uniformly to f (z) for R1 < |z − z0 | < R2 .
Definition 8.2 (Residue) The coefficient of the term (z − z0 )−1 in the Lau-
rent expansion, C−1 , is known as the residue of the function f at z0 .
Definition
�−1 8.3 (Principal part) The singular part of the Laurent expansion,
n
n=−∞ Cn (z − z0 ) , is known as the principal part of the Laurent expan-
sion.
Remark In the case that f is analytic on the whole disc |z| ≤ R2 it is easily
checked, using Cauchy’s formula for derivatives (55) for n ≥ 0, and Cauchy’s
theorem 5.10 for n < 0, that the Laurent expansion (71) for f reduces to the
Taylor expansion (58).
105
In the first integral we have |z − z0 | < |w − z0 | for any w ∈ γ2 , and thus
1 1 1
= = � �
w−z (w − z0 ) − (z − z0 ) (w − z0 ) 1 − z−z0
w−z0
∞ � �
1 � z − z0
j
= , (74)
w − z0 j=0 w − z0
again the sum converges uniformly. Using (74) and (75) in (73) gives
� � ∞
1 (z − z0 )j
f (z) = f (w) dw +
2πi γ2 j=0 (w − z0 )j+1
� � ∞
1 (w − z0 )j
f (w) dw. (76)
2πi γ1 j=0 (z − z0 )j+1
The sums and integrals can be interchanged, by theorem 7.3. We also relabel
in the second sum, writing j = −(m + 1), for m running from −∞ to −1.
Then (76) becomes
�∞ � � �
1 f (w) dw
f (z) = j+1
(z − z0 )j +
j=0
2πi γ2 (w − z 0 )
−1
� � � �
1 f (w) dw
m+1
(z − z0 )m .
m=−∞
2πi γ1 (w − z0 )
Finally, noting that the integrand in all cases is analytic within D, we can
use theorem 5.11 to replace each of γ1 and γ2 by an arbitrary contour lying
between γ1 and γ2 , giving
�∞ � � � �∞
1 f (w) dw j
f (z) = j+1
(z − z0 ) ≡ Cj (z − z0 )j ,
j=−∞
2πi γ (w − z0 ) j=−∞
106
as claimed.
Remark From the form of the Laurent expansion and our knowledge of the
properties of power series, it is clear that the function f is analytic at the
point z0 if and only if all the Laurent coefficients Cj for j < 0 are zero. If
some Cj �= 0 for j < 0 then f is unbounded at z0 , and cannot be analytic
there. For such singular functions, the Laurent expansion can be used to
classify the type of singularity, as we discuss in §8.2 below.
Proof Assume z0 = 0 for simplicity (the proof easily generalizes). Let R1 <
r < R2 , then
� � ∞
�
−(n+1)
2πiCn = f (w)w dw = Bk wk−n−1 dw
C(0;r) C(0;r) k=−∞
� ∞
� � �∞
= Bk wk−n−1 dw + B−m w−m−n−1 dw.
C(0;r) k=0 C(0;r) m=1
By the results on uniform convergence of power series (theorem 7.2 and its
corollary 7.3) summation and integration can be interchanged in each of these
two power series to give
∞
� �
2πiCn = Bk wk−n−1 dw = 2πiBn
k=−∞ C(0;r)
�
(it is easily shown by parametric integration that C(0;r)
wk−n−1 dw = 2πiδkn ).
107
This theorem is useful because in general it is very difficult to evaluate the
Laurent coefficients using the formula (72) derived in the theorem 8.1. How-
ever, the uniqueness result assures us that if by any other means we can find
a valid Laurent expansion for the given function f , then by uniqueness, this
expansion must be the unique Laurent expansion for f .
We now consider a few examples to illustrate.
On A2 we have
1 −1
= = −z −1 (1 − z −1 )−1 ,
1−z z(1 − z −1 )
and the final bracketed term here can again be binomially expanded since
|z|−1 < 1 on A2 . Then
�∞ −2
�
1 1 −1 −1 −n
f (z) = + =z −z z =− zk , z ∈ A2 .
z 1−z n=0 k=−∞
108
If we had not noted the partial fraction decomposition we could still obtain
the Laurent expansion quite easily, noting that on A1
�∞ ∞
�
1
f (z) = = z −1 (1 − z)−1 = z −1 zn = zn,
z(1 − z) n=0 n=−1
Example 8.6 Find a Laurent series expansion for the function tan z, about
z = π/2.
109
Writing u = z − π/2 we have
�∞ (−1)n u2n
sin z sin(π/2 + u) cos u n=0 (2n)!
f (z) = = =− = − �∞ (−1)n u2n+1
cos z cos(π/2 + u) sin u n=0 (2n+1)!
�∞ (−1)n u2n
1 n=0 (2n)!
= − �∞ (−1)n u2n
u n=0
(2n+1)!
� �� �−1
1 u2 u 4 u2 u4
= − 1− + + ··· 1− + + ···
u 2! 4! 3! 5!
� �� � � �
1 u2 u 4 u2 4 1 1
= − 1− + + ··· 1+ +u − + + ···
u 2! 4! 3! 5! (3!)2
� � � �
1 u2 4 1 1 1 1
= − 1− +u − + − + ··· .
u 3 (3!)2 5! 4! 2!3!
Setting u = z − π/2 in here we obtain the desired Laurent expansion.
2
Example 8.7 Find a Laurent series expansion of the function f (z) = e−1/z .
Considered as a real function of a real variable z = x, although (strictly
speaking) it is singular at x = 0, this function can be plotted for all real
x and is actually very smooth, approaching zero exponentially as x → 0.
However in the complex plane it is very badly-behaved as z → 0, because if
2
we consider z = iy, y → 0, we have −1/z 2 = 1/y 2 → +∞, so that e−1/z
blows up exponentially as z → 0 along the imaginary axis. This bad behavior
is reflected in its Laurent series expansion which, for z �= 0, can be obtained
from the usual power series representation of the exponential function:
∞
� �0
−1/z 2 (−1)n z −2n (−1)n z 2n
e = = ,
n=0
n! n=−∞
|n|!
uniformly convergent for any z �= 0. Hence this expansion is valid for |z| > R
for any R > 0.
110
The function e−1/w in the integrand has a Laurent expansion convergent on
any |z| > R as above:
� ∞
−1/w (−1)n w−n
e = ,
n=0
n!
and thus
∞
e−1/w � (−1)n w−(n+2)
=
w2 n=0
n!
also uniformly convergent on |z| > R for any R > 0. By theorem 7.3 then,
we can interchange the order of summation and integration in the definition
of A(z) and integrate term-by-term, to find
�∞ �
(−1)n ∞ −(n+2)
A(z) = w dw
n=0
n! z
∞
� (−1)n+1
= [w−(n+1) ]∞
z
n=0
(n + 1)!
�∞
(−1)n −(n+1)
= z
n=0
(n + 1)!
�−1
(−1)n+1 n
= z .
−∞
|n|!
Looking at the third equality above we see that
�∞ �∞
(−1)n −n (−1)n −n
A(z) = − z =1− z = 1 − e−1/z , (79)
1
n! 0
n!
suggesting that we could have evaluated the integral explicitly; and indeed
we see that differentiation of each of (78) and (79) leads to the same result,
A� (z) = −e−1/z /z 2 .
This procedure does not always work – for example, consider the function
E(z) defined by
� ∞ −w
e
E(z) = dw.
z w
The integral is convergent, but the above procedure leads to divergent inte-
grals.
Homework: Ablowitz & Fokas, problems for §3.3, questions 2,3,5.
111
8.2 Singular points
We begin by discussing the simplest kind of functional singularities: where
the function f (z) is analytic on some punctured disk about z0 (0 < |z − z0 | <
R for some R > 0), but is not analytic at z0 itself. Such a singular point z0 is
known as an isolated singular point of f . (Branch points are not isolated
singular points.) There are several types of such isolated singular points.
(i) If z0 is an isolated singular point of f at which |f | is bounded (i.e. there
is some M > 0 such that |f (z)| ≤ M for all |z| ≤ R) then z0 is a removable
singularity of f . Clearly, all coefficients Cn with n < 0 must be zero
in the Laurent expansion
�∞ (71), thusn f in fact has a regular power series
expansion f (z) = n=0 Cn (z − z0 ) valid for 0 < |z − z0 | < R. Since
this power series converges at z = z0 it follows that if we simply redefine
�∞f is analyticnon the whole disc |z − z0 | < R, with Taylor
f (z0 ) = C0 then
series f (z) = n=0 Cn (z − z0 ) .
Example 8.9 The function f (z) = sin z/z has a removable singularity at
z = 0. Strictly speaking this function is undefined at zero, but it has Laurent
series
∞
� (−1)n z 2n
f (z) = , |z| > 0,
n=0
(2n + 1)!
112
(ii) If f (z) can be written in the form
g(z)
f (z) =
(z − z0 )N
e2z − 1
f (z) =
z2
has a simple pole at z = 0.
Its Laurent expansion is given by
∞
� 2(n+2) z n
f (z) = ,
n=−1
(n + 2)!
so the residue at z = 0 is 1.
113
Example 8.13 The function
�0
1/z 2 z 2n
f (z) = e with Laurent expansion f (z) =
n=−∞
|n|!
Homework: Review your notes on branch points, branch cuts and multi-
functions.
114
Homework: Ablowitz & Fokas, problems for section 2.1, question 3.
Ablowitz & Fokas, problems for section 3.2, question 2(b),(f), 6(c).
Ablowitz & Fokas, problems for section 3.3, question 4.
115
9 Cauchy’s Residue Theorem
9.1 Preliminaries: Motivation
We have seen that Cauchy’s theorem is a very powerful result that applies to
functions analytic on a domain in C. However, we also now know that many
functions are not analytic, and have singularities of various kinds within
the domain of interest. We need to be able to evaluate integrals involving
non-analytic functions too (after all, it would be quite boring if the answer
were always zero). Cauchy’s residue theorem enables us to evaluate many
such integrals around closed contours, sometimes even where functions have
infinitely many singularities within the contour of integration.
There are many reasons why we might want to perform integration of
complex functions around contours in the complex plane. A very common
application of such contour integration arises in the inversion of Laplace
(and Fourier) transforms, which can be used to solve many ordinary and
partial differential equations (and which are considered in detail in Math
756). A more commonplace application of complex contour integration is
in the evaluation of certain real integrals that are hard to do by standard
methods from real calculus. We begin with an example of a real integral that
can be reduced to a contour integral.
�∞
Example 9.1 Evaluate I = 0 (1 + x4 )−1 dx.
Method 1: Try to think of a clever (real) substitution that reduces this
integral to one we can evaluate explicitly.
Method 2: Try to rewrite the (real) integral as a complex contour integral.
Methods from complex analysis can often be used to evaluate such integrals
quickly and easily.
116
and, recalling example 4.9, we would expect the integral of f (z) along a large
circular arc to go to zero as the arc radius R → ∞. We therefore consider
f (z) integrated around a semicircular contour γ composed of the real axis
from −R to R (this part will correspond to the integral we want), and a
semicircle in the upper half plane (this part will go to zero as R → ∞); see
figure 4. Parametrizing the semicircular contour in the usual way, z = Reiθ ,
0 ≤ θ ≤ π, we have
� � ∞ � π
iReiθ dθ
lim f (z) dz = 2 f (x) dx + lim .
R→∞ γ 0 R→∞ 0 1 + R4 e4iθ
so if we can evaluate this contour integral then we have the value of the real
integral. The contour integral is not zero; Cauchy’s theorem does not apply
to this case because the contour contains singularities at z = eiπ/4 , e3πi/4 .
Cauchy’s Residue theorem will tell us how to evaluate such integrals.
Similar tricks can be used to convert real integrals involving trigonometric
functions to complex integrals around closed contours, in many instances.
Again, the procedure is best illustrated by example, but often the trick is to
consider an appropriate function integrated around the unit circle, using the
fact that z = eiθ on such a circle.
� 2π
Example 9.2 Use a complex substitution to evaluate I = 0 (1+8 cos2 θ)−1 .
Note that on the contour, z = eiθ , z −1 = e−iθ , and thus 2 cos θ = z + z −1 .
Also, dz = ieiθ dθ, and so dθ = −idz/z. Then
� �
−i dz z dz
I= 2 2
= −i 4 2
.
C(0;1) z(1 + 2z + 4 + 2/z ) C(0;1) 2z + 5z + 2
If the complex integrand here was analytic on the unit disk (within the con-
tour) then Cauchy’s theorem would give a zero result immediately. However,
117
Figure 4: The semicircular contour can be used to evaluate certain real
integrals from −∞ to ∞, or from 0 to ∞, as the circle radius R → ∞.
Lemma 9.3 (Function with a single pole within γ) Let f be analytic inside
118
and on a simple closed (positively-oriented) contour γ, except at z0 ∈ γ where
it has an isolated singularity, with Laurent expansion
∞
�
f (z) = Cn (z − z0 )n .
−∞
Then
�
f (z) dz = 2πiC−1 .
γ
�∞ � 2π
= Cn rn einθ ireiθ dθ
n=−∞ 0
�∞
= Cn rn+1 (2πiδn,−1 )
n=−∞
= 2πiC−1 .
119
Proof Let fk be the principal part of the Laurent expansion of f about each
zk . Then
m
�
g(z) = f (z) − fk (z)
k=1
120
and
1
Res(f (z); e3iπ/4 ) =
(e3iπ/4
− − e−3iπ/4 )(e3iπ/4 − e−iπ/4 )
eiπ/4 )(e3iπ/4
√
2 2
=
((−1 + i) − (1 + i))((−1 + i) − (−1 − i))((−1 + i) − (1 − i))
√
2 2 i(1 + i)
= ×
−2.2i.(−2)(1 − i) i(1 + i)
√
i 2(1 + i)
= .
−8
Thus by Cauchy’s residue theorem 9.4 we have
� ∞ � � √ √ � √
dx 1 dz i 2(1 − i) i 2(1 + i) π 2
= = πi + = .
0 1 + x4 2 γ 1 + z4 −8 −8 4
121
Example 9.5 Use complex contour integration and the residue theorem to
evaluate
� ∞
cos x
2
dx.
−∞ x + x + 1
eiz
f (z) = 2 ,
z +z+1
whose real part, when z = x ∈ R, gives the integrand we need, then on ΓR
we have
� iR(cos θ+i sin θ) �
� e �
|f (z)| = �� 2 2iθ �
R e + Re + 1 �
iθ
e−R sin θ
≤
R2 − R − 1
1
≤ 2
on ΓR ,
R −R−1
122
and thus for this choice of f (z),
�� � �
� � R
� f (z) dz �≤ dθ → 0 as R → ∞.
� � 2
R −R−1
ΓR ΓR
2
=� √ e (cos(1/2) − i sin(1/2)) = √ cos(1/2)e− 3/2 .
−∞ x +1+1 3 3
123
Figure 5: Geometric construction used in the proof of Jordan’s inequality.
C(O; 1) with the horizontal diameter, and B is the intersection of C(M, sin θ)
with the horizontal diameter, and Q is the reflection of P in this horizontal,
then the inequality is proved by noting that the length of the straight line
P Q (2 sin θ) is less than or equal to the arc length P AQ = 2θ, and the arc
length P AQ is less than or equal to the arc length P BQ = π sin θ (half the
circumference of circle C(M ; sin θ)).
This inequality can be used to prove a general result about integrals of an-
alytic functions along semicircular contours which, as we are seeing, arise
frequently in applications of the residue theorem.
124
Proof The condition on the limiting behavior of f means that on ΓR (z =
Reiθ ) we have |f (z)| ≤ MR , where MR is independent of θ but depends on
R, and MR → 0 as R → ∞. Defining
�
I= eikz f (z) dz
ΓR
we then have z = Reiθ , dz = iReiθ , and eikz = e−kR sin θ+ikR cos θ , so that
� π � π/2
−kR sin θ
|I| ≤ e MR R dθ = 2MR R e−kR sin θ dθ.
0 0
We now apply Jordan’s inequality, sin θ ≥ 2θ/π, which holds on the range of
integration, to deduce that e−kR sin θ ≤ e−2kRθ/π for 0 ≤ θ ≤ π/2. Thus
� π/2
MR π
|I| ≤ 2MR R e−2kRθ/π dθ = (1 − e−kR ) → 0 as R → ∞,
0 k
because MR → 0 as R → ∞.
Remark Note that if k < 0 in the integral then a similar result holds for
the contour Γ−
R , the semicircle in the lower half plane.
Observe, e.g., that we did not need Jordan’s lemma in example 9.5 (eval-
uation of cos x/(x2 + x + 1) from x = −∞ to ∞), because the function
1/(z 2 + z + 1) decayed sufficiently fast as |z| → ∞. It decays as 1/R2 , which
is enough to compensate for the factor of R that arises in the numerator from
the |dz|. If, however, we wished to evaluate the real integral
� ∞
x sin x
2
dx
−∞ x + 1
125
then we do need the lemma, since a straightforward estimate will not give
the required decay in this case. Here we must consider the complex integral
�
zeiz
2
dz,
γ z +1
around the contour made up of ΓR and the straight-line segment along the
real axis from −R to R. The integrand has poles at z = ±i, and so the
Residue theorem gives the result as
�
zeiz iz 2 zeiz πi
2
dz = 2πiRes(ze /(z + 1); i) = 2πi lim (z − i) 2
= .
γ z +1 z→i z +1 e
Thus,
� R �
xeix zeiz πi
dx + dz = .
−R x2 + 1 ΓR z2 + 1 e
Since the function z/(z 2 + 1) goes to zero uniformly on ΓR as R → ∞,
Jordan’s lemma 9.7 applies, and gives
�� �
� zeiz �
� dz �� → 0 as R → ∞.
� 2
ΓR z + 1
Thus,
� R
xeix πi
lim 2
dx = ,
R→∞ −R x +1 e
and taking the imaginary part gives the result we seek,
� ∞
x sin x π
2
dx = .
−∞ x + 1 e
(Note that the real part gives
� ∞
x cos x
dx = 0,
−∞ x2 + 1
a result we could have deduced by symmetry considerations.)
It is important to remember when the “k < 0” generalization of Jordan’s
lemma is required, because then the contour of integration to be used changes,
and the residues must be taken from within the appropriate contour. The
following example (taken from Ablowitz & Fokas [1]) shows such a case:
126
Example 9.8 Evaluate
� ∞
x sin ax cos bx
dx, c > 0, a, b ∈ R.
−∞ x2 + c2
We first note that the trigonometric identity
1
sin ax cos bx = (sin(a − b)x + sin(a + b)x)
2
suggests considering the complex integrals
� �
zei(a−b)z zei(a+b)z
J1 = 2 2
dz and J 2 = 2 2
dz,
γ1 z + c γ2 z + c
127
since the only pole of the integrand inside γ1 is z = −ic. Putting these results
together then, the quantity we require is
1
I = �(I1 + I2 )
2 � � i(a−b)z � � i(a+b)z ��
2π ze ze
= � −Res ; z = −ic + Res ; z = ic
2 z 2 + c2 z 2 + c2
� c(a−b) �
e e−c(a+b)
= π − +
2 2
−bc
= −πe sinh(ac).
128
is exactly the integrand we require. However, this function has a pole at
z = 0, and so we cannot take a contour of integration that passes through
the origin. The way around this difficulty is to use the basic semicircular
contour, but make a small semicircular indentation (of vanishingly small ra-
dius 0 < � � 1) around the pole at the origin. For such a contour γ we have,
with f (z) = eiz /z,
� � −� � � R �
f (z) dz = f (x) dx − f (z) dz + f (x) dx + f (z) dz = 0,
γ −R Γ� � ΓR
For the integral along the small semicircle we have z = �eiθ , so that
� � π i�(cos θ+i sin θ)
eiz e
dz = iθ
i�eiθ dθ → πi as � → 0.
Γ� z 0 �e
129
The above uses a special case of a general result when we indent a contour
of integration along a (vanishingly) small circular arc to avoid a singularity
in the integrand.
Lemma 9.9 Let f be analytic in some B � (a; r) (r > 0), with a simple pole
of residue b at z = a. Then with Γ� = {z : z = a + �eiθ , θ1 ≤ θ ≤ θ2 }, we
have
�
lim f (z)dz = ib(θ2 − θ1 ).
�→0 Γ�
Note that this result does not generalize to the case of a stronger singularity,
such as a pole of higher order at z = a, as a glance at the proof easily shows.
However, sometimes we may be faced with an integral that appears to require
this generalization, such as the following:
� ∞
x − sin x
dx.
−∞ x3
130
The real integrand has only a removable singularity at the origin (consider
the Taylor expansion of x − sin x about x = 0), but we know that we cannot
consider the integral of (z −sin z)/z 3 around the large semicircle as this grows
exponentially on ΓR as R → ∞. The preceding results suggest that we need
to consider the real part of the complex integrand f1 (z) = (z + ieiz )/z 3
which, along the real axis z = x ∈ R, gives the function we want. However,
the Laurent expansion of this function about z = 0 is
� � ��
1 z 2 iz 3
f1 (z) = 3 z + i 1 + iz − − + ···
z 2! 3!
� �
i z 2 iz 3
= 3 1− − + ··· ,
z 2! 3!
which has a triple pole at z = 0, so a simple indentation of the contour as
in the lemma above will not work. We need to find some other function that
also has (x − sin x)/x3 as its real part, but that has at most a simple pole at
the origin, and such a function is
z + ieiz − i
f (z) = ,
z3
with Laurent expansion (from f1 above)
� �
i z 2 iz 3 i
f (z) = 3 1 − − + ··· − 3
z 2! 3! z
� �
i 1 iz
= − + + ···
z 2! 3!
(thus we have just a simple pole of residue −i/2). We can integrate this
function around a large semicircle in the upper half-plane, with an indenta-
tion at the origin. The contribution along the portion Γ+R will go to zero as
R → ∞ (Jordan’s lemma and elementary bounds), and lemma 9.9 applies to
the small indentation, giving
�
(−i) π
lim f (z) dz → iπ = .
�→0 Γ
�
2 2
Since f (z) has no singularities within the integration contour chosen, as
� → 0, R → ∞ we obtain (taking care to observe the sense of integration on
each portion of the contour)
�� 0 � ∞� � �
π x − sin x i(cos x − 1)
= + + dx.
2 −∞ 0 x3 x3
131
Taking the real part gives the result we want (the imaginary part tells us
that the integral, interpreted in the principal-value sense, of the odd function
(cos x − 1)/x3 over the whole real line, is zero).
which does not decay exponentially over the whole contour for either choice
−
Γ+
R or ΓR . We need to derive a variant of Jordan’s lemma 9.7 appropriate to
this integrand. Note that Jordan’s inequality (81) gives
2 sin 2θ
≤ ≤ 1 for θ ∈ (0, π/4],
π 2θ
suggesting that we restrict the arc to a quarter of the usual,
132
Figure 6: Sector contour of angle π/4 used in example 9.10.
2
Moreover, on the ray z = reiπ/4 we have f (z) = e−r , which we can integrate
from r = 0 to ∞. Thus taking the contour γ defined by
In this case, the fact that the integrand is not an even function of x prevents
us from using the usual semicircular contour, since the integral of f (z) =
1/(z 3 +a3 ) along the negative real axis does not give a multiple of I. However,
if we consider rays in the complex plane again, z = reiα for fixed α and
133
r ∈ R+ , then on such a ray f (z) = 1/(r3 e3iα + a3 ); and we see that if
α = 2π/3 then we get the desired integrand, f (z = 1/(r3 + a3 ). Thus,
if we take the integral of f (z) around a large sector of angle 2π/3 we will
get contributions proportional to I from the two straight sides, a vanishing
contribution from the curved portion as R → ∞; and the total value of the
integral will be given by Cauchy’s residue theorem, since there is a simple
pole of the integrand at z = aeiπ/3 inside the contour. Denoting the sector
contour by γ we have on the one hand, using the residue theorem,
�
dz
3 + a3
= 2πiRes(f (z); aeiπ/3 )
γ z
z − aeiπ/3
= 2πi lim
z→aeiπ/3
� z 3 + a3
2πi ��
=
3z 2 �z=aeiπ/3
2πie−2πi/3
=
3a2
2πe−iπ/6
= ,
3a2
where we used l’Hôpital’s rule between the 2nd and 3rd lines to evaluate the
limit. On the other hand, splitting the contour into its constitutive parts we
have
� � ∞ � ∞
dx 2πi/3 dr
lim f (z) dz = 3 3
− e 3 3
= (1 − e2πi/3 )I.
R→∞ γ 0 x +a 0 r +a
Thus,
2πe−iπ/6 −2πe−iπ/6 2πi 2π
I= 2 2πi/3
= 2 iπ/3 iπ/3 −iπ/3
= 2 = √ .
3a (1 − e ) 3a e (e −e ) 3a .2i sin π/3 3a2 3
134
Figure 7: A semicircular contour with branch-cut along the negative real
axis, used in example 9.11.
135
Along the �-semicircle we need another direct estimate of the integral. With
z = �eiθ
�� � �� �
� � � log � + iθ iθ ��
π
� f (z) dz � = � i�e dθ�
� � � 2 2iθ
Γ� 0 1+� e
�(| log �| + π)π
≤ → 0 as � → 0.
1 − �2
Along the positive real axis we have z = x ∈ R+ , and as � → 0 and
R → ∞ this integral converges to the one we want. Along the negative real
axis portion of γ we have, since we are on the top side of the cut, z = reiπ
(r = |z|); dz = eiπ dr = −dr and log z = log r + iπ, and this portion of the
integral is, as � → 0, R → ∞,
� 0 � ∞ � ∞
log r + iπ log r dr
(−dr) = dr + iπ .
∞ 1 + r2 0 1 + r2 0 1 + r2
Again, the first integral here is equal to the one we seek, and putting all
results together we have, letting � → 0 and R → ∞,
� ∞ � ∞
log x dr
2 2
dx = 2πiRes(f (z); i) − iπ ;
0 1+x 0 1 + r2
so we can see we will get two real integrals for the price of one here. Since f
has only a simple pole at z = i we have
log z log(i) π
Res(f (z); i) = lim(z − i)f (z) = lim(z − i) 2
= = .
z→i z→i z +1 2i 4
Thus we find
� ∞ � ∞
log x dx π
dx = 0, and = .
0 1 + x2 0 1+x 2 2
Example 9.12 Show that
� ∞ a−1
x
dx = πcosecπa (0 < a < 1).
0 1+x
The integrand here is another multifunction when considered in the complex
plane, so we must again cut the plane appropriately and define an analytic
branch of the multifunction in the cut plane. This time we make a cut along
136
Figure 8: “Keyhole” contour used in example 9.12.
the positive real axis, so that we represent z in the cut plane by z = |z|eiθ ,
0 ≤ θ < 2π and define the analytic branch of z a−1 by
The contour used in the previous example does not help us here, since the
integral along the negative real axis won’t reduce to something related to the
integral we want (check this). So, we choose a contour that includes straight-
line integrals along both sides of the cut, with a small circular indentation (a
“keyhole”) to avoid the branch-point itself (figure 8). The contour is closed
off by a large circular arc of radius R � 1 that does not cross the cut. This
contour is well-known, and is often called a keyhole contour. Within the
contour the chosen branch of the function is analytic except for a simple pole
at z = −1; and denoting by γ the entire contour we have
�
f (z) dz = 2πiRes(f (z); z = −1).
γ
137
On the outer large circular portion of γ we have z = Reiθ , 0 ≤ θ < 2π, and
so
� a−1 i(a−1)θ �
�R e � a−1
|f (z)| = �� �≤ R .
1 + Reiθ � R − 1
The integral along the upper side of the cut converges trivially to the
desired integral �as � → 0, R → ∞. Along the bottom side of the cut the
contribution to γ is such that z = re2πi in the integrand, with r going from
∞ to 0 to preserve the same sense of integration. Thus we have dz = dr,
and, if we call this portion of the curve γ − then in the limit � → 0, R → ∞
� � 0 a−1 2πi(a−1) � ∞ a−1
r e 2πi(a−1) r
f (z) dz = dr = −e dr
γ− ∞ 1+r 0 1+r
taking care to use the analytic branch of the multifunction as defined above.
Denoting by I the real integral we seek, we then have
138
Remark When calculating integrals such as this, involving integrating a
multifunction in a cut plane, it is important to ensure that you remain con-
sistent in your choice of analytic branch throughout. In example 9.12 just
solved, for example, we would have obtained an incorrect answer had we
worked with the branch
This does define an analytic branch, but only in the plane cut along the
negative real axis – which is inconsistent with the contour we have chosen.
139
We instead take a rectangle, with one side along the real axis from −R
to R (along y = 0), and the parallel side along y = Y . The two shorter sides
along x = ±R complete the rectangle. We need to make a convenient choice
of Y that allows us to evaluate the real integral we want. We choose complex
integrand f (z) = epz /(1 + ez ), since this gives the correct integral along the
real axis. On y = Y we have
ep(x+iY ) epx eipY
f (z) = = ,
1 + ex+iY 1 + ex eiY
and we see that if we choose Y = 2π then we get a multiple of the integral we
want. On the short sides of the rectangle, Γr and Γl say, we have z = ±R+iy,
0 ≤ y ≤ 2π. On Γr ,
�� � �� Y � � Y
� � � e pR iy
e � epR Y epR
� f (z) dz � = � idy �≤ dy ≤
� � � R iy � R eR − 1
Γr 0 1+e e 0 e −1
→ 0 as R → ∞,
since p < 1. On Γl we have
�� � �� 0 −pR iy � �
� � � e e � Y
e−pR Y e−pR
� f (z) dz � = � idy �� ≤ dy ≤
� � � −R
1+e e iy 1 − e−R 1 − e−R
Γl Y 0
→ 0 as R → ∞,
since p > 0. By the Residue theorem then, as R → ∞ we obtain
� ∞ � −∞ 2πip px
epx e e
x
dx + dx = 2πiRes(f (z), iπ)
−∞ 1 + e ∞ 1 + ex
epz eiπp
⇒ (1 − e2πip )I = 2πi lim (z − iπ) = 2πi = −2πieiπp .
z→iπ 1 + ez eiπ
Thus,
2πieiπp π
I= iπp iπp −iπp
= .
e (e − e ) sin πp
We conclude by considering a more complicated variant.
140
Figure 9: Rectangle with indentations to exclude simple poles from the inte-
rior, used in example 9.14.
141
case, the indentation round the pole is taken in the opposite sense to that in
the lemma statement (so we get a change of sign). For the integrand fp we
thus obtain
� ∞ 2πip px
e e
Ip − x
dx − iπRes(fp (z), 0) − iπRes(fp (z), 2πi) = 0.
−∞ 1 − e
Homework
�∞ (1) What is wrong with the following argument: “Let I =
4 −1
0
(1 + x ) dx. Put x = iy, then
� ∞
I= (1 + y 4 )−1 dy = iI.
0
√
Hence, I = 0.” How do we obtain the correct value of the integral (π/(2 2))?
142
9.6 Principle of the argument, and Rouché’s theorem
Residue calculus can also be applied to deduce results about the number of
zeros and poles of a meromorphic function in a given region.
f � (z) ±ni
= + φi (z)
f (z) (z − zi )
where φi (z) = gi� (z)/gi (z) is analytic in B(zi , �i ), and in the region D� made
up of the interior of the curve C, with each B(zi , �i ) removed, f � /f is analytic.
Application of Cauchy’s theorem 5.10 to the region D� then gives
�
1 f � (z)
0 = dz
2πi ∂D� f (z)
� � �
1 f (z) 1 � f � (z)
= dz − dz
2πi C f (z) 2πi i γ(zi ,�i ) f (z)
� � �
1 f (z) 1 � ±ni
= dz − + φi (z)dz
2πi C f (z) 2πi i γ(zi ,�i ) (z − zi )
� � � �
1 f (z)
⇒ dz = ni − nj = N − P,
2πi C f (z) zeros poles
143
Figure 10: A geometrical interpretation of the Argument Principle.
where γ(zi , �i ) denotes the circle of center zi and radius �i , and we used the
residue theorem 9.4 in the last step. The quantities N , P are as defined in
the theorem statement.
To demonstrate the final equality in the theorem, we parametrize the
curve C so that
C = {z(t) : t ∈ [a, b], z(a) = z(b)}.
Then
� � � b �
1 f (z) 1 f (z(t)) � 1
dz = z (t)dt = [log|f (z(t))| + i arg(f (z(t)))]bt=a
2πi C f (z) 2πi a f (z(t)) 2πi
1
= [arg(f (z))]C
2π
(which branch of the logarithm we choose is immaterial for the theorem
result, but for definiteness you can assume the principal branch).
The theorem has an interesting geometrical interpretation, illustrated in fig-
ure 10. The function w = f (z) represents a mapping from the complex
z-plane to the complex w-plane, under which the curve C maps to some im-
age curve C � in the w-plane (see later notes on conformal mappings; Math
756). We then have dw = f � (z)dz, so that
� � �
1 f (z) 1 dw 1 1
dz = = [log(w)]C � = [arg(w)]C � .
2πi C f (z) 2πi C � w 2πi 2π
144
Remark This quantity (1/2π)[arg(w)]C � is known as the winding number
of the curve C � about the origin in the w-plane – the number of times that
the closed curve C � encircles the origin.
Example 9.16 Find the number of zeros of the function f (z) = z 5 +1 within
the first quadrant.
We use the principle of the argument to determine the change in arg(f ) as
we traverse an appropriate contour. Since f is analytic on the first quadrant
of the z-plane, we may take the contour C to be a large quarter-circle, made
up of the three portions C = C1 ∪ C2 ∪ C3 ,
C1 = {z : z = x, 0 ≤ x ≤ R},
C2 = {z : z = Reiθ , 0 ≤ θ ≤ π/2},
C3 = {z : z = iy, 0 ≤ y ≤ R}.
In the limit that R becomes arbitrarily large the contour will enclose the
whole of the first quadrant. The argument of f satisfies arg(f ) = φ, where
tan φ = �(f )/�(f ). Along C1 f = 1 + x5 ∈ R+ , and arg(f ) = 0. Along
C2 f = 1 + R5 e5iθ ≈ R5 e5iθ , and the argument of f therefore increases from
0 to 5π/2 (in the limit R → ∞) as θ increases from 0 to π/2. Finally,
on C3 f = 1 + iy 5 , and for y = R � 1 we have arg(f ) ≈ 5π/2 (it must
vary continuously from C2 to C3 ), while as y decreases from R � 1 to 0,
tan φ = �(f )/�(f ) decreases (continuously) from +∞ to 0+ and thus arg(f )
decreases (continuously) from 5π/2 to 2π.
Thus, the net change in arg(f ) as we traverse C is 2π. Applying theorem
9.15 then, since f has no singularities within the first quadrant (so P = 0),
the number of zeros it has, N , satisfies
1 2π
N= [arg(f )]C = = 1.
2π 2π
Theorem 9.17 (Rouché) Let f (z) and g(z) be analytic inside and on a sim-
ple closed contour C. If |f | > |g| on C, then f and f + g have the same
number of zeros inside C.
This theorem makes sense intuitively, since we can think of g as being a
perturbation to f , which will in turn perturb its zeros. If the size of the
perturbation is bounded on the contour C, then (we can show) the size of
the perturbation is bounded inside C as well (this is not surprising, given
145
Figure 11: Illustration of the proof of Rouch’e’s theorem.
f (z) + g(z)
h(z) = ,
f (z)
then h is analytic and nonzero on C, and the Argument Principle theorem
9.15 may be applied to deduce that
� �
1 h (z) 1
dz = [arg(w)]C � ,
2πi C h(z) 2π
where w = h(z), and C � is the image of the curve C under this transformation
in the w-plane. However,
g(z)
w = h(z) = 1 +
f (z)
146
as we traverse the closed curve C � there is no net change in arg(w), because
C � does not enclose the origin. It follows from the argument principle that
Nh −Ph = 0, where Nh and Ph are the numbers of zeros and poles, respectively
(counted according to multiplicity) of h(z). Since f and g are analytic inside
and on C, the poles of h coincide in location and multiplicity with the zeros
of f , so that Ph = Nf . Also by analyticity of f , h is zero only where f +g = 0,
and we have Nh = Nf +g . Thus,
Nh − P h = 0 ⇒ Nf = Nf +g ,
Example 9.18 Show that 4z 2 = eiz has a solution on the unit disc |z| ≤ 1.
Take the contour C to be |z| = 1, so that points on C are given by z = eit ,
t ∈ [0, 2π]. Then, with f (z) = 4z 2 and g(z) = −eiz , we have |f | = 4 on C,
while
Thus Rouché’s theorem applies, and f + g has the same number of zeros as
f on the unit disc. Clearly, f (z) = 4z 2 has exactly 2 zeros (both at z = 0
but we count according to multiplicity). It follows that 4z 2 = eiz in fact has
2 solutions on the unit disc.
Homework: Ablowitz & Fokas, problems for section 4.4, questions 3(b),
5(a).
147
10 Asymptotic expansions and Stokes lines
10.1 Asymptotic sequences and asymptotic expansions
Asymptotic expansions are about finding an approximation to some function
that depends on a small (usually real) parameter, 0 < � � 1. The function
may be given explicitly, with �-dependence embedded in the definition, or
(as is more common in applications) the function may be a solution of a
differential equation (ordinary or partial) that contains the small parameter.
Typically the approach is to find some means of expanding the function as
a sum, often in powers of the small parameter, to a certain number of terms,
e.g.
f (�) = f0 + �f1 + · · · + �N fN + RN (�)
= SN (�) + RN (�), (84)
in order to achieve a desired level of accuracy. Here, RN (�) is some remainder
term, which intuitively we would expect to be much smaller than the pre-
ceding term in the series. In most asymptotic expansions in � of the above
form, we in fact have |RN (�)| = O(�N +1 ) as � → 0.1 Loosely speaking, this
means that |RN (�)| is bounded by some constant (positive) multiple of �N +1
as � → 0; more precisely, there exists some M > 0 and some �0 > 0 (small)
such that
|RN (�)| ≤ M �N +1
for � < �0 .
�
Thus, for example, any polynomial in �, P (�) = N n
n=1 an � , is of order � as
� → 0. As well as this “order” notation, we also have the concept of smaller
order”, denoted by “o” instead of “O”. Here we write
f (�) = o(g(�)) as � → 0 (85)
to signify that f (�) is much smaller than g(�) as � → 0 – again, more precisely,
we have
� �
� f (�) �
lim �� � → 0 as � → 0.
�→0 g(�) �
148
Returning to our initial idea of approximating a function (see (84)), we say
that SN is an approximation to f valid, or correct, or accurate, to order �N
as � → 0 if
f − SN RN
lim ≡ lim = 0.
�→0 �N �→0 �N
149
but this last series diverges for all � > 0. However, this last exercise is not
futile, because it turns out that if one truncates the divergent infinite series
in (87) at a finite number of terms, N , then for small values of � the finite
sum gives a good approximation to the integral, and the approximation gets
better and better as � → 0. It can be shown that
� N
∞
e−t dt �
= (−1)n n!�n + O(�N +1 ) as � → 0 for N fixed. (88)
0 1 + �t 0
and thus
(though in fact this result also holds as � → 0 with N fixed). The sum on the
right-hand-side of (88) provides an asymptotic expansion of the integral
on the left-hand side; and we write
∞
�
f (�) ∼ (−1)n n!�n
0
150
to denote the statement (88). (We will sometimes also denote the asymptotic
expansion by the finite sum plus the remainder, or just the finite sum itself.)
Since we know the asymptotic series in (88) diverges as N → ∞, a natural
question to ask is, how to decide where to truncate the series? Obviously
for a given value of �, as one takes more and more terms, the approximation
must eventually get worse, since the terms diverge as N → ∞. A given
asymptotic expansion, for a given value of �, will have an optimal truncation
value N (�). For this particular case, optimal truncation is for N ∼ 1/�.
We now generalize this idea somewhat. The series does not need to be
in powers of � (though this simple form is frequently observed). We define
an asymptotic sequence δn (�) to be a decreasing sequence of functions as
� → 0, such that
� �
� δn+1 (�) �
� �
� δn (�) � → 0 as � → 0.
means
N
�
f (�) = δn (�)an + O(δN +1 (�)) as � → 0 for N fixed.
n=0
151
Writing � = 1/k we have
� ∞ � ∞
e−t/� e−τ
f= dt = � dτ.
0 1+t 0 1 + �τ
The discussion above showed us how to obtain the asymptotic expansion of
this integral as � → 0; we have
∞
�
f ∼� (−1)n n!�n ,
0
and thus
∞
� (−1)n−1 (n − 1)!
f∼ ,
n=1
kn
or equivalently,
N
� (−1)n−1 (n − 1)!
f= + R,
n=1
kn
Thus,
� −k � � −k �
e−k e−k e e−k e−k e
I1 = − 1.I2 = − − 2I3 = − 2 +2 − 3I4
k k k2 k k k3
..
.
e−k e−k 2!e−k N −1 (N − 1)!e
−k
= − 2 + 3 + · · · + (−1) N
+ (−1)N N !IN +1 .
k k k k
152
This generates an asymptotic expansion for the integral as k → ∞ provided
|N !IN +1 | � e−k /k N as k → ∞, for fixed N . But clearly
� ∞
N! −t N !e−k e−k
|N !IN +! | ≤ N +1 e dt = N +1 � N ,
k k k k
as k → ∞, for fixed N . We again have the situation that the infinite series
is divergent; but if we take only a finite number N of terms and let k → ∞
then the remainder in the series goes to zero. Note that in this example the
asymptotic series is not a simple power series.
153
Then the function g(z) = f (z) + e−z has an identical asymptotic power series
expansion in S1 , because on that sector −π/2 < θ < π/2 we have
When we move outside S1 however, the two functions will differ in their
asymptotic expansions. If in fact the asymptotic expansion (90) is valid on
an extended sector, say S2 = {z : π/2 < arg(z) < π}, then on the larger
sector S1 ∪ S2 f will have the same expansion as before, but the expansion
of g on S2 as |z| → ∞ will now be dominated by the e−z , which here is
exponentially large:
∞
�
−z an
g(z) ∼ e + , as |z| → ∞.
n=0
zn
It could be the case that on the remaining portion of the complex plane f
itself has a different asymptotic expansion.
Note that in the above, the (convergent) Taylor series expansion of e−z
does NOT provide an asymptotic expansion, because its terms Tn = (−1)n z n /n!
do not satisfy the criterion of subsequent terms becoming asymptotically
smaller as |z| → ∞, for a fixed point N in the series (z n is not an asymptotic
sequence as |z| → ∞). Therefore, if we were to truncate the series at some
fixed value N , the truncated series would not provide a good approximation
to the function for |z| arbitrarily large.
This example with the function g is a very simple (and vague, at this
stage) example of the Stokes phenomenon, in which the asymptotic ex-
pansion of a given function can change discontinuously across a line in com-
plex space. We see that in the example above, the function e−z possesses an
essential singularity at infinity (consider the singular behavior of the function
e−1/z at the origin), which is what gives rise to this behavior. Another simple
example is given by the function f (z) = sinh(1/z), as z → 0. This function
has an essential singularity at the origin, as is easily seen by writing down its
Laurent expansion about z = 0. The Laurent expansion, though convergent
in any neighborhood of the origin, does not provide an asymptotic expan-
sion for the function, because if we truncate it at a finite number of terms,
we do not get a good approximation to the function – there is no “largest”
term as |z| → 0 that we can start our expansion with and order terms in
decreasing size. Again, it is the exponential representation that gives the
154
asymptotic expansion for the function, and different representations are ap-
propriate in different regions of C. The full representation of the function is
f (z) = (e1/z − e−1/z )/2; and
−1
| exp(1/z)| = | exp(r−1 (cos θ − i sin θ))| = er cos θ
,
−1 −r −1 cos θ
| exp(−1/z)| = | exp(−r (cos θ − i sin θ))| = e ,
so that in the sector −π/2 < θ < π/2 (the right-half plane; θ = arg(z)) we
have cos θ > 0, and
1
f (z) ∼ e1/z as |z| → 0
2
is a valid asymptotic expansion for sinh(1/z), while in the sector π/2 < θ <
3π/2 (the left-half plane) we have cos θ < 0, and
1
f (z) ∼ − e−1/z as |z| → 0
2
is a valid asymptotic expansion for sinh(1/z).
155
Suppose first that t ∈ R+ in the integral, with the path of integration taken
along the positive real axis. Then as t → ∞, the integral converges provided
�(zt) > 0, that is, provided −π/2 < θ < π/2, where θ = arg(z). In this
case, for large |z| the integrand is exponentially small except near t = 0.
The dominant contribution will therefore come from this region; and we may
expand the factor (1 + t4 )−1 in the integrand using the binomial expansion:
(1 + t4 )−1 = 1 − t4 + O(t8 ). Thus, for −π/2 < arg(z) < π/2 we have
� ∞ � ∞ �� ∞ �
−zt 4 −zt 8 −zt
f (z) ∼ e dt − t e dt + O t e dt .
0 0 0
�∞
Writing In = 0 tn e−zt dt, we have
� −zt �∞ �
e n n ∞ −zt n−1 n n! n!
In = t + e t dt = In−1 = · · · = n I0 = n+1 .
−z 0 z 0 z z z
Thus, for −π/2 < arg(z) < π/2 (region 1),
1 4!
f (z) ∼ − + O(z −9 ), as |z| → ∞. (91)
z z5
However, this expansion is certainly not valid for values of arg(z) outside this
region. To find the correct far-field behavior in other regions, we consider a
different integration path in t, allowing t complex.
Noting that the function 1/(1 + t4 ) decays as |t| → ∞, and that Jordan’s
Lemma 9.7 applies to the integrand along (any portion of) a large semicircular
contour in the right-half plane (the integral along any such circular arc will go
to zero as the arc goes to infinity), we may form a closed contour by taking a
large quarter-circle, with radii along the positive real t-axis and the negative
imaginary t-axis. Then, applying the residue theorem to this contour, we
have
� −i∞ −zt
e
f (z) = 4
dt − 2πiRes(e−zt /(1 + t4 ); e−iπ/4 )
0 1 + t
� −i∞ −zt
e iπ −ze−iπ/4 3πi/4
= dt + e e . (92)
0 1 + t4 2
The integral in (92) converges provided �(zt) > 0, which now corresponds
to a different range of arg(z), and a different region of the complex z-plane.
With z = reiθ and t = se−iπ/2 , for r, s ∈ R+ , we have
zt = rsei(θ−π/2) , �(zt) = rs cos(θ − π/2),
156
and thus a convergent integral for 0 < θ < π (region 2). So, we can
now compute the far-field behavior of f (z) for the range −π/2 < θ < π
((region 1) ∪ (region 2)). (Note that there is overlap between region 1 and
region 2, so potentially two different asymptotic expansions for f on this
overlap domain.)
To evaluate the far-field behavior in region 2 due to (92), note that in the
contribution from the residue we have
−iπ/4
e−ze = e−r(cos(θ−π/4)+i sin(θ−π/4)) ,
which is exponentially small in −π/4 < θ < 3π/4, but exponentially large in
3π/4 < θ < π.
The same approach as was used to obtain (91) can be applied to the
integral (call it J) in (92), since again, on its region of convergence in the
z-plane, the dominant contribution comes from small values of |t|. Writing
t = −iη in this integral, where η ∈ R+ , we find
� ∞ izη
e
J = −i dη.
0 1 + η4
Since the dominant contribution (for |z| � 1, 0 < θ < π) comes from
0 < η � 1, we can again expand the factor (1 + η 4 )−1 using the binomial,
and writing w = −iz we obtain (exactly as before)
� �
1 4! −9
J = −i − + O(w ) , as |w| → ∞,
w w5
1 4!
= − + O(z −9 ), as |z| → ∞.
z z5
Thus, on the overlap region, (region 1) ∩ (region 2), 0 < θ < π/2, we have an
exponentially small contribution from the residue, and using the approach
for either region, we obtain the same result,
1 4!
f (z) ∼ − 5 + O(1/z −9 ),
z z
−iπ/4
since the exponentially-small term e−ze is much smaller than any element
of the asymptotic sequence in negative powers of z (so would never appear
in the asymptotic expansion). So there is no discrepancy between the two
approaches.
157
The above shows that, on the whole region −π/2 < θ < 3π/4 (region
1, plus that part of region 2 where the exponential contribution in (92) is
exponentially small), the appropriate asymptotic expansion is
1 4!
f (z) ∼ − + O(1/z −9 ), as |z| → ∞,
z z5
while on the remainder of region 2, 3π/4 < θ < π, the exponential term is
far larger than any of the algebraic terms, so that the asymptotic expansion
there is
iπ −ze−iπ/4 3πi/4
f (z) ∼ e e as |z| → ∞.
2
Hence the form of the asymptotic expansion of f changes discontinuously
across the ray θ = 3π/4 (a Stokes line for this example), as we move from
one region to the next.
It remains, of course, to elucidate the behavior of f in the region 3,
−π < θ < −π/2.
158
References
[1] Ablowitz, M.J., Fokas, A.S. Complex variables: Introduction and Ap-
plications (2nd Edition) Cambridge University Press (2003).
159