MATH2019 ENGINEERING MATHEMATICS 2E

ADDITIONAL LECTURE NOTES

These notes are intended to give a brief outline of the course to be used as an
aid in learning. They are not intended to be a replacement for attendance at
lectures, problem classes or tutorials. In particular, they contain few exam-
ples. Since examination questions in this course consist mainly of examples,
you will seriously compromise your chances of passing by not attending lec-
tures, problem classes and tutorials where many examples will be worked out
in detail.

2015
c School of Mathematics and Statistics, UNSW

1
 TOPIC 1 – PARTIAL DIFFERENTIATION
Partial derivatives are the derivatives we obtain when we hold constant all but
one of the independent variables in a function and differentiate with respect
to that variable.
Functions of Two Variables
Suppose z = f (x, y). Define
∂f
= lim f (x+∆x,y)−f
∆x
(x,y)
∂x ∆x−→0
∂f
= lim f (x,y+∆y)−f
∆y
(x,y)
∂y ∆y−→0

These are both functions of x and y and the usual differentiation rules (prod-
uct, quotient etc) apply.
Notation
∂f ∂f
= fx = zx , = fy = zy
∂x ∂y
i.e. subscripts are used to denote differentiation with respect to the indicated
variable. Further
∂f ∂f
(x0 , y0 ) = fx (x0 , y0 ) means evaluated at the point (x0 , y0 ).
∂x ∂x

Higher-Order Derivatives

∂ 2f
!
∂ ∂f
= = (fx )x = fxx
∂x2 ∂x ∂x
∂ 2f
!
∂ ∂f
= = (fy )x = fyx
∂x∂y ∂x ∂y
∂ 2f
!
∂ ∂f
= = (fx )y = fxy
∂y∂x ∂y ∂x
∂ 2f
!
∂ ∂f
= = (fy )y = fyy
∂y 2 ∂y ∂y

Mixed Derivatives Theorem (M.D.T.)

The functions fxy (the y derivative of fx ) and fyx (the x derivative of fy )
are obtained by different procedures and so would appear to be different
functions. In fact, for almost all functions we meet in practical applications,
they are identical because of the M.D.T. which says
If f (x, y) and its partial derivatives fx , fy , fxy and fyx are all defined and
continuous at all points in a region surrounding the point (a, b) then

fxy (a, b) = fyx (a, b).

2
 This readily extends to higher order derivatives. In particular, if all deriva-
∂ n+m f
tives are continuous then ∂x n ∂y m can be used to denote the partial derivative

of f n times with respect to x and m times with respect to y in any order

whatsoever.
Chain Rule
Recall that if u = f (x) and x = g(t) then
du du dx df dg
= = = f 0 (x)g 0 (t).
dt dx dt dx dt
This readily generalises. If w = f (x, y) and x and y are themselves differen-
tiable functions of t (e.g. x and y are the coordinates of a moving point and
t is time), then
dw ∂f dx ∂f dy
= + .
dt ∂x dt ∂y dt
Functions of Three or more Variables
If z = f (x1 , x2 , x3 , . . .) then the partial derivative of z with respect to any
one variable (call it xi ) is obtained by holding all the other variables constant
and then differentiating with respect to xi . The mixed derivatives theorem
extends to these cases.
Chain Rule
This readily extends to functions of three or more variables. For example, if
w = f (x, y, z) and x, y, z are themselves functions of t, then
dw ∂f dx ∂f dy ∂f dz
= + + .
dt ∂x dt ∂y dt ∂z dt
Chain Rules for Functions defined on Surfaces
Suppose w = f (x, y, z) and x = x(r, s), y = y(r, s), z = z(r, s) (the last
three define a surface in 3D space) then
∂f ∂f ∂x ∂f ∂y ∂f ∂z
= + +
∂r ∂x ∂r ∂y ∂r ∂z ∂r
and
∂f ∂f ∂x ∂f ∂y ∂f ∂z
= + +
∂s ∂x ∂s ∂y ∂s ∂z ∂s
where ∂x
∂r
etc are taken holding s constant and ∂x
∂s
etc are taken holding r
constant.
Multivariable Taylor Series
From first year, we know that if f is a function of a single variable x, then
1
f (x) = f (a) + (x − a)f 0 (a) + (x − a)2 f 00 (a) + · · ·
2!
∞
1 (n)
f (a)(x − a)n
X
=
n=0 n!

3
This extends to functions of 2 or more variables. We consider only f (x, y).
The Taylor Series of f (x, y) about the point (a, b) is

∂f ∂f
f (x, y) = f (a, b) + (x − a) (a, b) + (y − b) (a, b)
∂x ∂y

∂ 2f ∂ 2f
(
1
+ (x − a)2 2 (a, b) + 2(x − a)(y − b) (a, b)
2! ∂x ∂x∂y
∂ 2f
)
2
+(y − b) (a, b) + higher-order terms.
∂y 2
Standard Linear Approximation
If y = f (x) then a reasonable approximation when x is close to x0 is

f (x) ' f (x0 ) + (x − x0 )f 0 (x0 )

obtained by truncating the Taylor series after the linear term. Geometrically,
we are approximating the curve y = f (x) for x near x0 by the tangent to the
curve at (x0 , f (x0 )).
This idea readily extends to functions of two or more variables. All we
do is truncate the Taylor Series after the linear terms. The standard linear
approximation of f (x, y) near (x0 , y0 ) is therefore f (x, y) ' L(x, y) where

L(x, y) = f (x0 , y0 ) + (x − x0 )fx (x0 , y0 ) + (y − y0 )fy (x0 , y0 ).

Geometrically, we are approximating the curved surface z = f (x, y) near

(x0 , y0 ) by the tangent plane at (x0 , y0 , f (x0 , y0 )).
Differentials
The expression
∂f ∂f
df = (x0 , y0 )dx + (x0 , y0 )dy
∂x ∂y
is called the differential. You can think of it as the “infinitesimal” change
df produced in f by “infinitesimal” changes dx in x and dy in y. It is obtained
from L(x, y) by replacing ∆x = (x − x0 ) by dx and ∆y = (y − y0 ) by dy.
Error Estimation
The differential can be used to estimate changes in f due to small changes
in its arguments. If ∆f = f (x0 + ∆x, y0 + ∆y) − f (x0 , y0 ) then

∂f ∂f
∆f ' ∆x + ∆y.
∂x ∂y
∂f ∂f
where ,
∂x ∂y
are evaluated at (x0 , y0 ).

4
If ∆x and ∆y are known, we just substitute them in. Usually, however, all
we know are bounds on ∆x and ∆y. For example, we may only be able to
measure temperature to ±0.01◦ C. In that case we have, approximately,
   
 ∂f   ∂f 
|∆f | ≤  |∆x| +   |∆y|.
   

 ∂x   ∂y 

For functions of several variables f (x1 , x2 , · · · , xn )

n
X ∂f
∆f ' ∆xk
k=1 ∂xk

and n 
 
 ∂f 
X 
|∆f | ≤   |∆xk |.
 ∂xk 
k=1

Leibniz Rule

d Z v(x) Z v(x)
∂f dv du
f (x, t)dt = dt + f (x, v(x)) − f (x, u(x)) .
dx u(x) u(x) ∂x dx dx

TOPIC 2 – EXTREME VALUES

Extrema for functions of two variables
Suppose we have f (x, y) continuous on some region R. What are the
extreme values of f (x, y) (i.e. the maxima and minima) and how do we find
them?
Definition: The function f (x, y) has a global maximum or absolute max-
imum at (x0 , y0 ) if f (x, y) ≤ f (x0 , y0 ) for all (x, y) ε R.
Definition: The function f (x, y) has a global minimum or absolute min-
imum at (x0 , y0 ) if f (x, y) ≥ f (x0 , y0 ) for all (x, y) ε R.
Definition: The function f (x, y) has a local maximum or relative maxi-
mum at (x0 , y0 ) if f (x, y) ≤ f (x0 , y0 ) for all (x, y) in some neighbourhood of
(x0 , y0 ).
Definition: The function f (x, y) has a local minimum or relative mini-
mum at (x0 , y0 ) if f (x, y) ≥ f (x0 , y0 ) for all (x, y) in some neighbourhood of
(x0 , y0 ).
Definition: A point (x0 , y0 ) ε R is called a critical point of f if

fx (x0 , y0 ) = fy (x0 , y0 ) = 0,

or if f is not differentiable at (x0 , y0 ).

Definition: A local maximum or minimum is called an extreme point of f .
These can only occur at
(i) boundary points of R

5
(ii) critical points of f

Second Derivative Test

If f and all its first and second partial derivatives are continuous in the
neighbourhood of (a, b) and fx (a, b) = fy (a, b) = 0 then
2
(i) f has a local maximum at (a, b) if fxx < 0 and D = fxx fyy − fxy > 0 at
(a, b).
2
(ii) f has a local minimum at (a, b) if fxx > 0 and D = fxx fyy − fxy >0
at (a, b).
2
(iii) f has a saddle point at (a, b) if D = fxx fyy − fxy < 0 at (a, b).
2
(iv) If D = fxx fyy − fxy = 0 at (a, b) the second derivative test is inconclu-
sive.

The application of these ideas to practical problems will be illustrated in

the lectures. The candidates for maxima and minima are found by looking at
i) boundary points, ii) points where one or more of the first partial derivatives
fail to exist and iii) points where all the first partial derivatives vanish.
The ideas readily generalise to functions of 3 or more variables although
the second derivatives test becomes quite messy.
Extreme values for parameterised curves
To find the extreme values of a function f (x, y) on a curve x = x(t), y =
y(t) we find where
df ∂f dx ∂f dy
= +
dt ∂x dt ∂y dt
is zero. The extreme values are found at

(i) Critical points (where f 0 = 0 or f 0 does not exist).

(ii) The endpoints of the parameter domain.

Constrained extrema and Lagrange multipliers

Motivation: Suppose we are asked to find the minimum (or maximum) of a
function subject to a constraint.
Example: Find the point P (x, y, z) on the plane 2x + y − z − 5 = 0 that lies
closest to the origin.
This involves finding the minimum of the function
q
f (x, y, z) = x2 + y 2 + z 2

subject to the constraint that x, y and z satisfy

g(x, y, z) = 2x + y − z − 5 = 0

6
In this simple case, it is easy to use the constraint equation to find an explicit
expression for one of the variables (say z) in terms of the other two and to
then substitute this into f which thus becomes a function of two variables
only and then to find the extrema of f as a function of x and y. For a more
complicated constraint, it may not be possible to use the constraint equation
to obtain an explicit expression for one of the variables in terms of the others
so a more general procedure is required.
The method of Lagrange multipliers
To start off, suppose that f (x, y) and g(x, y) and their first partial deriva-
tives are continuous. To find the local minima and maxima of f subject to
the constraint g(x, y) = 0 we find the values of x, y and λ that simultaneously
satisfy the equations
∂f ∂g ∂f ∂g
−λ = 0, −λ = 0, together with g(x, y) = 0 (1)
∂x ∂x ∂y ∂y
Justification: We can, in principle, use the equation g(x, y) = 0 to write y
as a function of x although, as indicated above, this may not be possible in
practice. Hence, we may consider f to be a function of a single variable x and
look for points where df /dx = 0. Let (x, y) = (a, b) be such a point. But, by
the chain rule
df ∂f d x ∂f d y ∂f ∂f d y
= + = +
dx ∂x d x ∂y d x ∂x ∂y d x
Thus
∂f ∂f d y
+ =0 at (x, y) = (a, b) (2)
∂x ∂y d x
However, since g(x, y) = 0, dg/dx = 0 everywhere (including (a, b)). Thus
∂g ∂g d y
+ =0 at (x, y) = (a, b) (3)
∂x ∂y d x
Thus, eliminating dy/dx from (2) and (3) we obtain
∂f ∂g ∂f ∂g
− =0 at (x, y) = (a, b)
∂x ∂y ∂y ∂x
which can also be written
 ∂f ∂f 
 
 ∂x ∂y 
 ∂g ∂g  =0 at (x, y) = (a, b)
 
∂x ∂y

Hence, the rows of this determinant must be linearly dependent, Thus there
exists a real number λ such that
! !
∂f ∂f ∂g ∂g
, =λ ,
∂x ∂y ∂x ∂y

These equations, together with g(x, y) = 0, are just (1).

7
N. B. The quantity λ is called a Lagrange multiplier and the method also
works if f and g are also functions of z. In that case we have the additional
equation ∂f /∂z = λ∂g/∂z to solve. It is also possible to introduce the so-
called Lagrangian function

L(x, y, λ) = f (x, y) − λg(x, y)

The equations (1) and the constraint g(x, y) = 0 are obtained by setting to
zero the first partial derivatives of L(x, y, λ) with respect to x, y and λ.
Lagrange multipliers with two constraints
Suppose we now want to find the maxima and minima of f (x, y, z) subject
to
g1 (x, y, z) = 0 and g2 (x, y, z) = 0.
To do this, we introduce two Lagrange multipliers (one for each constraint)
and the Lagrangian function for this situation

L(x, y, z, λ, µ) = f (x, y, z) − λg1 (x, y, z) − µg2 (x, y, z)

We now need to find the values of x, y, z, λ and µ which simultaneously satisfy

the five equations obtained by setting to zero the partial derivatives of L with
respect to x, y, z, λ and µ.

TOPIC 3 - VECTOR FIELD THEORY

Quick Revision of Vector Algebra
Scalars are quantities which have only a magnitude (and sign in some cases)
such as temperature, mass, time and speed. Vectors have a magnitude and
a direction. We will work only in 3D physical space and use the usual right-
handed xyz coordinate system. We denote vector quantities by using bold
symbols, e.g. a . We let i, j, k be the three unit vectors parallel to the x, y
and z axes respectively. If a point P has coordinates (p1 , p2 , p3 ) and Q has
−→
coordinates (q1 , q2 , q3 ) then the vector P Q from P to Q has components

a1 = q 1 − p 1 , a 2 = q 2 − p 2 , a 3 = q 3 − p 3
−→ −→ −→
and
q a = P Q= a 1 i + a 2 j + a 3 k = OQ − OP . The length of a is |a| =
2 2 2
a1 + a2 + a3 . The position vector of a typical point with coordinates (x, y, z)
is usually written r = xi + yj + zk.
Addition etc.
Define 0 = 0i + 0j + 0k. This is the vector all of whose components are zero,
and is not to be confused with the scalar 0. All the usual rules apply, for
example

a + b = (a1 + b1 )i + (a2 + b2 )j + (a3 + b3 )k

8
 a+0 = a
ca = ca1 i + ca2 j + ca3 k
−a = (−1)a = −a1 i − a2 j − a3 k
a+b = b+a
(a + b) + c = a + (b + c) = a + b + c
a + (−a) = 0.
c(a + b) = ca + cb
Inner or Dot or Scalar Product of Vectors
a · b = a1 b1 + a2 b2 + a3 b3 = |a||b| cos γ
where γ (0 ≤ γ ≤ π) is the angle between a and b.
Then a · a = |a|2 and the dot product of two (non-zero) vectors is 0 if and
only if they are orthogonal (γ = π2 ).
Observe that i · i = j · j = k · k = 1 and
a·b a1 b 1 + a2 b 2 + a3 b 3
cos γ = =q q
|a||b| a22 + a22 + a23 b21 + b22 + b23
The component of a vector a in the direction of b (otherwise known as the
projection of a onto b) is
|a|a · b a·b
p = |a| cos γ = = .
|a||b| |b|
Vector or Cross Product of Vectors
v = a × b is a vector whose magnitude is |v| = |a||b| sin γ (where γ is the
angle (0 ≤ γ ≤ π)) between a and b. The vector v is perpendicular to the
plane defined by a and b, in such a way that a right-handed screw turn in
the direction of v turns a into b through an angle of less than π.
Properties

a × b = −b × a
a × a = 0 
 i j k 
 
a × b =  a1 a2 a3 
 
 b 1 b2 b3 

Triple Scalar Product

 
a a2 a3 
 1
a · (b × c) =  b1

b2 b3 
 
 c1 c2 c3 
def
= b · (c × a) = c · (a × b) = [a b c]

9
Also, |a · (b × c)| is the volume of the parallelepiped defined by a, b and c.
Scalar and Vector Fields
Consider some region Ω of 3-dimensional space. Let a typical point in Ω have
coordinates (x, y, z). A scalar field is a scalar quantity f (x, y, z) defined on
Ω. It often depends on time t as well. The temperature or density in the
atmosphere are examples of scalar fields.
A vector field is a vector each of whose components is a scalar field. Thus

v = v1 i + v2 j + v3 k

where v1 , v2 and v3 all depend on x, y, z (and t usually) is a vector field.

Velocity and acceleration in a fluid are good examples. If r = xi + yj + zk we
sometimes write v = v(r, t) to indicate that v depends on position and time.
Differentiation of Vectors
Suppose v is a vector field which depends on a single quantity ξ (e.g. ξ =
time t). Define
dv v(ξ + ∆ξ) − v(ξ)
= lim
dξ ∆ξ−→0 ∆ξ
Thus
dv dv1 dv2 dv3
=i +j +k .
dξ dξ dξ dξ
By applying the product rule to each component, we readily derive:
d dρ dv
(ρv) = v + ρ
dξ dξ dξ
d du dv
(u · v) = ·v + u·
dξ dξ dξ
d du dv
(u × v) = ×v + u× .
dξ dξ dξ
where ρ is a scalar.
Partial Derivatives
If v depends on several independent variables, the partial derivative of v
with respect to any one of these is obtained by holding all other independent
variables constant and differentiating with respect to the nominated variable.
Velocity and Acceleration
Consider a point moving through space. Let its coordinates at time t be
(x(t), y(t), z(t)). Then its position vector is

r(t) = x(t)i + y(t)j + z(t)k.

The velocity of the point is

dr
v= .
dt

10
The speed is |v| = (v · v)1/2 and the acceleration is

dv d2 r
a= = 2.
dt dt
Gradient of a Scalar Field
∂φ ∂φ ∂φ
∇φ = grad φ = i+ j+ k.
∂x ∂y ∂z
Directional Derivative
Consider a scalar field φ. What is the change in φ as we move from P (x, y, z)
to Q(x + ∆x, y + ∆y, z + ∆z) keeping t constant?
If ∆s is the distance from P to Q then

(∆s)2 = (∆x)2 + (∆y)2 + (∆z)2 .

So, letting ∆s −→ 0, the vector

dx dy dz
u
b = i+ j+ k
ds ds ds
is seen to be a unit vector in the direction from P to Q.
Now, by our earlier work on increment estimation.
∂φ ∂φ ∂φ
∆φ = φQ − φP = ∆x + ∆y + ∆z + smaller terms
∂x ∂y ∂z
!
∂φ ∆x ∂φ ∆y ∂φ ∆z
= + + ∆s + smaller terms
∂x ∆s ∂y ∆s ∂z ∆s

Hence, letting ∆s −→ 0,
dφ ∂φ dx ∂φ dy ∂φ dz
= + +
ds ∂x ds ∂y ds ∂z ds
= ∇φ · u
b

Now, the rate of change with respect to distance in the direction specified
by the unit vector u
b is called the directional derivative and is denoted by
Dub φ. We have shown that

b φ = ∇φ · u
Du b.

(N.B. ub is a vector of unit length).

b , dφ = |∇φ| cos θ since
Now if θ(0 ≤ θ ≤ π) is the angle between ∇φ and u ds
b | = 1. Thus, dφ has the maximum value |∇φ| when θ = 0 (i.e. u
|u b is in the
ds
direction of ∇φ ) and the minimum value −|∇φ| (when u b is in the direction
of −∇φ.)

11
Normal to a Surface
Next, consider a level surface φ = C. This defines a surface S in space. For
example, meteorologists talk about surfaces of constant pressure such as the
500 millibar surface. Let P and Q be any two nearby points on S. Then
φP = φQ = C, i.e. dφ ds
= 0 at P in any direction tangential to S at P. Thus
−→
∇φ at P is orthogonal to P Q.
Since this holds for any point Q close to P, i.e. is independent of the direction
from P to Q, it follows that ∇φ at P must be orthogonal to the level surface
∇φ
φ = C. A unit normal is |∇φ| .
Equation of Tangent Plane
If P has coordinates (x0 , y0 , z0 ) and (x, y, z) is any point in the plane tan-
gent to S at P then ∇φ is normal to this tangent plane which therefore has
equation
∇φ · [(x − x0 )i + (y − y0 )j + (z − z0 )k] = 0
where ∇φ is evaluated at P.
Divergence of a Vector Field
If F = F1 i+F2 j+F3 k then ∇·F = div F = ∂F ∂x
1
+ ∂F
∂y
2
+ ∂F
∂z
3
. It may be regarded
∂ ∂ ∂
at the dot product of the vector differential operator ∇ = i ∂x + j ∂y + k ∂z and
the vector F. It is just a scalar.
N.B. ∇ · F 6= F · ∇. The latter is the differential operator
∂ ∂ ∂
F · ∇ = F1 + F2 + F3
∂x ∂y ∂z

Theorem ∇ · (φv) = φ(∇ · v) + v · (∇φ).

Proof
∂ ∂ ∂
LHS = (φv1 ) + (φv2 ) + (φv3 )
∂x ∂y ∂z
∂v1 ∂φ ∂v2 ∂φ ∂v3 ∂φ
=φ + v1 + φ + v2 + φ + v3
∂x ∂x ∂y ∂y ∂z ∂z
!
∂v1 ∂v2 ∂v3 ∂φ ∂φ ∂φ
=φ + + + v1 + v2 + v3
∂x ∂y ∂z ∂x ∂y ∂z
= φ∇ · v + v · ∇φ = RHS
Q.E.D.
Laplacian
∂ 2 φ ∂ 2 φ ∂ 2 φ def 2
∇ · (∇φ) = + + 2 = ∇ φ.
∂x2 ∂y 2 ∂z
Curl of a Vector Field
 
 i j k 
 
 ∂ ∂ ∂ 
∇ × F = curl F =  ∂x ∂y ∂z 
 
 F1 F2 F3 

12
 ! ! !
∂F3 ∂F2 ∂F1 ∂F3 ∂F2 ∂F1
=i − +j − +k −
∂y ∂z ∂z ∂x ∂x ∂y
Theorem ∇ × (∇φ) = 0.
Proof  

 i j k 
 ∂ ∂ ∂ 
L.H.S = 
 ∂x ∂y ∂z


 ∂φ ∂φ ∂φ 
 ∂x ∂y ∂z 

∂ 2φ ∂ 2φ ∂ 2φ ∂ 2φ ∂ 2φ ∂ 2φ
! ! !
=i − +j − +k −
∂y∂z ∂z∂y ∂z∂x ∂x∂z ∂x∂y ∂y∂x
= 0i + 0j + 0k = 0 = R.H.S.
Vector fields F for which ∇ × F = 0 are called irrotational or conservative
.
Line Integrals
These are used for calculating, for example, the work done in moving a particle
in a force field.
Consider a vector field F(r) and a curve C from point A to point B. Let the
equation of C be r = r(t) where t is parameter. Let t = a at point A and
t = b at point B. We define
Z b
Z
dr
F(r) · dr = F(r(t)) · dt
C a dt
In terms of components, this can be written
Z b !
Z
dx dy dz
(F1 dx + F2 dy + F3 dz) = F1 + F2 + F3 dt
C a dt dt dt
where dx, dy, dz are displacements measured along C and F is evaluated on C.
In general, this integral depends not only on F but also on the path we take
between A and B. If A and B coincide, we are integrating around a closed
curve. This is denoted by I
F · dr.
C
F · dr is the work done in moving from A to B
R
Work If F is a force field, C
along C.
Simple properties
· dr = k F · dr
R R
(i) If k is a constant C (kF) C

+ G) · dr = F · dr + G · dr.
R R R
(ii) C (F C C

F · dr = − F · dr.
R R
(iii) C1 C
where C1 is the same curve as C except that we start at B and finish
at A, i.e. reversing the order of integration changes the sign of a line
integral.

13
 F · dr = F · dr + F · dr
R R R
(iv) C C1 C2

where C1 is the curve from A to B, C2 the curve from B to C and C the

curve from A to C following the same path.

TOPIC 4 – DOUBLE INTEGRALS

Reminders on Coordinate Systems
1. Cartesian Coordinates: (x, y, z)
We often use 3D Cartesian coordinates xyz. When we do so, the system is
always taken as being right-handed. By this we mean that a right-handed
turn through 90◦ along Ox turns Oy into Oz.
2. Cylindrical Polar Coordinates: (r, θ, z)

x = r cos θ 
 r≥0
y = r sin θ
 0 ≤ θ < 2π

z=z

Definition of a Double Integral

Let Ω be some sub-region of the xy plane and f (x, y) be a function de-
fined at all points of Ω. We divide Ω up into N non-overlapping sub-regions
N
∆Ω1 , ∆Ω2 , · · · , ∆ΩN (so Ω =
S
∆Ωi ). Let ∆Ωj have area ∆Aj and let (ξj , ηj )
i=1
be a typical point in ∆Ωj . Form the sum
N
X
IN = f (ξj , ηj )∆Aj .
j=1

Now let N −→ ∞ in such a way that the largest linear dimension of each
∆Ωj −→ 0 as N −→ ∞. Then if IN tends to some limit as N −→ ∞ we
define Z
I = f (x, y)dA = lim IN .
N −→∞
Ω

If f is continuous on Ω, this limit will exist.

Interpretations
R
1. If f (x, y) = 1 then dA is the area of Ω.
Ω

2. If f ≥ 0 on Ω, then I is the volume of the solid whose base is Ω in the

x − y plane and whose top has equation z = f (x, y)

14
Simple Properties

1. If k is a constant
Z Z
kf (x, y)dA = k f (x, y)dA.
Ω Ω

R R R
2. (f (x, y) + g(x, y))dA = f (x, y)dA + g(x, y)dA.
Ω Ω Ω

3. If f (x, y) ≥ 0 on Ω then
Z
f (x, y)dA ≥ 0
Ω

4. If Ω1 and Ω2 are non-overlapping regions and Ω = Ω1 ∪ Ω2 then

Z Z Z
f (x, y)dA = f (x, y)dA + f (x, y)dA.
Ω Ω1 Ω2

Evaluation as Repeated Single Integrals

The above definition is not useful for practical evaluation. Instead we almost
always evaluate double integrals as repeated single integrals. In xy cartesian
coordinates
dA = dx dy = dy dx.
The simplest case is when Ω is the rectangle defined by a ≤ x ≤ b, c ≤ y ≤ d.
Then Z Z dZ b Z bZ d
f (x, y)dA = f (x, y)dx dy = f (x, y)dy dx.
c a a c
Ω

The integral can be evaluated in either way.

In the first way, we fix y betweenR c and d and integrate with respect to
x from a to b. In this inner integral, ab f (x, y)dx, y is held constant and the
result is a function of y only. We then integrate this from y = c to y = d. In the
second form, we do analogous things with the rôles of x and y interchanged.
The two results must give the same answer, so it is only necessary to do it
one way.
Non Rectangular Regions

1. If Ω is defined by a ≤ x ≤ b, f1 (x) ≤ y ≤ f2 (x) then

Z Z b Z f2 (x)
f (x, y)dA = f (x, y)dy dx
a f1 (x)
Ω

15
 2. If Ω is defined by c ≤ y ≤ d, g1 (y) ≤ x ≤ g2 (y) then
Z Z d Z g2 (y)
f (x, y)dA = f (x, y)dx dy
c g1 (y)
Ω

The evaluation of such integrals is best understood by examples as will be

given in lectures. It is important that you always draw a sketch of the region
of integration in order to get the limits correct. Regions of more complicated
shape may need to be partitioned into a collection of subregions of type 1 or
2 in order to evaluate the integral.
Reversal of Order
Since the value of a double integral is independent of the order in which we
do the integration, it is sometimes easier to reverse the originally specified
order of integration. However, we must then be careful to choose the new
limits of integration so as to cover the same region Ω as the original integral.
Again, this technique is best illustrated by the examples which will be given
in lectures.
Density, Mass, Centre of Mass
Consider a lamina in the x − y plane (e.g. a piece of sheet metal). If a small
element of area ∆A has mass ∆m, we define the surface density (areal density
or superficial density are terms also used) as

∆m
δ(x, y) = lim .
∆A−→0 ∆A

Due to varying composition, this may be a function of x and y. A small

element dΩ with area dA then has mass dm = δ(x, y)dA.
If the lamina occupies the region Ω then its total mass will be
Z Z
M= dm = δ(x, y)dA.
Ω Ω

Consider a small element dΩ with area dA located at (x, y). Its distance from
the y axis is x and its distance from the x axis is y. The first moment of the
lamina about the y axis is
Z
My = xδ(x, y)dA
Ω

and the first moment about the x axis is

Z
Mx = yδ(x, y)dA.
Ω

16
The centre of mass has coordinates (xm , ym ) defined by
My Mx
xm = ym =
M M
If δ is constant, it will cancel and the centre of mass then coincides with the
centroid or centre of area of Ω which has coordinates (x, y) defined by
R R
xdA ydA
x = ΩR y = ΩR
dA dA
Ω Ω

Moments of Inertia
The moments of inertia of the above lamina about the x and y axes are

Ix = y 2 δ(x, y)dA
R
Ω
Iy = x2 δ(x, y)dA.
R
Ω

The polar moment of inertia about the origin is defined by

Z
I0 = Ix + Iy = (x2 + y 2 )δ(x, y)dA.
Ω

Polar Coordinates
If the region Ω is easily described using polar coordinates (x = r cos θ, y =
r sin θ) it is often better to evaluate the double integral in polar coordinates.
The main task is to express dA in polar coordinates.
If r increases by dr and θ by dθ, the little element of area so generated is

dA = (rdθ) × dr = r dr dθ

so
Z Z Z
f (x, y)dA = f (r cos θ, r sin θ)r dr dθ
Ω Ω
Z Z
= f (r cos θ, r sin θ)rd θ dr
Ω

In either form the r and θ limits are chosen to cover the region Ω. This will
be illustrated by examples in lectures. It is essential to draw a diagram of Ω.
Jacobian Transformation
Rb
The evaluation of a f (x)dx is often facilitated by the substitution x = x(u)
to give Z b Z β
dx
f (x)dx = f (x(u)) du
a α du

17
where x(α) = a, x(β) = b. For double integrals, if x = x(u, v), y = y(u, v)
then Z Z Z Z
f (x, y)dx dy = f (x(u, v), y(u, v)) |J|du dv
Ω Ω∗

where Ω∗ is the region in the (u, v) plane corresponding to Ω in the (x, y)

plane and J is the Jacobian Determinant
 
 ∂x ∂x  ∂x ∂y ∂x ∂y
−
 ∂u ∂v 
J=  ∂y ∂y  = .

∂u ∂v
 ∂u ∂v ∂v ∂u
(Thus dA = |J|du dv).
(N.B We take the absolute value of J)
This is valid provided J does not change sign on Ω∗ . Further x(u, v) and
y(u, v) must be continuous functions of u and v with continuous first partial
derivatives. The point (x, y) corresponding to any (u, v) in Ω∗ lies in Ω and
to every point (x, y) in Ω there corresponds one and only one (u, v) in Ω∗ .

TOPIC 5 – ORDINARY DIFFERENTIAL EQUATIONS

Differential equations arise naturally in science, engineering, biology and eco-
nomics. Typically, they relate the rate at which a quantity changes (with
respect to time or distance) to the quantity itself as well as to time or dis-
tance.
First Order Equations
These are of the form
dy
= F (x, y).
dx
There is no general method of solution but certain equations fall into classes
which can be solved. You must learn to recognise these.
1. Separable
These are of the form
dy
= f (x)g(y).
dx
dy
This may be written g(y)
= f (x)dx.
R dy R
So g(y)
= f (x)dx + c.

This is an implicit relation between y and x and involves only one integration
constant. If we were given some initial condition such as y(0) = 1, we would
impose it now to find c.
2. Linear
The general first-order linear o.d.e. is
dy
+ P (x)y = Q(x).
dx

18
 R
Multiplication by the integrating factor I(x) = exp{ P (x)dx} reduces this
to
d
(Iy) = IQ
dx
which can immediately be solved by integration to give
R
( IQdx + c)
y= .
I

Second-Order Linear Homogeneous O.D.E. with Constant Coeffi-

cients
These are of the form
y 00 + ay 0 + by = 0 (4)
where a and b are real constants. They arise in many engineering applications.
The general solution will contain two arbitrary constants α1 and α2

y = α1 y1 (x) + α2 y2 (x)

where y1 and y2 are linearly independent functions.

How to solve
We look for a trial solution y = αeλx . Substituting into (4) gives the charac-
teristic equation
λ2 + aλ + b = 0 (5)
which has the roots
1
h√ i
λ1 = 2
a2 − 4b
−a +
h √ i
λ2 = 12 −a − a2 − 4b

Three Cases Arise

1. a2 > 4b
Then λ1 and λ2 are both real and the general solution is

y = α1 eλ1 x + α2 eλ2 x .

2. a2 = 4b. Then λ1 = λ2 = − 21 a and the above procedure produces only

one solution. In this case only, it can be shown that another solution is
y = xeλ1 x so the general solution is

y = α1 eλ1 x + α2 xeλ1 x
= (α1 + α2 x)e−ax/2

19
 3. a2 < 4b. Then the roots are complex conjugates
1 1
λ1 = − a + iw, λ2 = − a − iw
2 2
where
1√ q
w= 4b − a2 = b − a2 /4 .
2
Then
1 1
y = α1 e(− 2 a+iw)x + α2 e(− 2 a−iw)x
= e−ax/2 {α1 eiwx + α2 e−iwx }
= e−ax/2 {β1 cos wx + β2 sin wx}

Free Oscillations
Several systems (e.g. a mass oscillating at the end of a spring or an electrical
circuit) may be described by the d.e.

my 00 + cy 0 + ky = 0

where m > 0, c > 0, k > 0. In a mechanical system, m is typically mass,

c a friction coefficient and k a restoring-force coefficient. The independent
variable is t and y is a displacement. Such a system is unforced since the
term on the right is 0.
Seeking solutions y = Aeλt gives the characteristic equation

mλ2 + cλ + k = 0,

which has the solutions

1 h √ i
λ1 = −c + c2 − 4mk
2m
1 h √ i
λ2 = −c − c2 − 4mk .
2m
Three cases arise
1. c2 > 4mk. This is called “overdamping”√ since the damping or frictional
coefficient c is large compared with 2 mk. In this case λ1 and λ2 are
both real and negative. The solution is y = Aeλ1 t + Beλ2 t which decays
to zero as t −→ ∞.

2. c2 = 4mk. This is “critical damping” and λ1 = λ2 so the solution is

y = (A + Bt)e−ct/2m .

The solution also decays to 0 as t −→ ∞.

20
 √
3. c2 < 4mk. This is called “underdamping” as c is smaller than 2 mk.
Then
λ1 = −α + iΩ, λ2 = −α − iΩ
q
c k c2
where α= 2m
, Ω= m
− 4m2
Thus
y = (A cos Ωt + B sin Ωt)e−αt
= Re−αt cos(Ωt − δ)
√
where R = A2 + B 2 and tan δ = B/A. This represents decaying
qthe idealised case c = 0 (no friction), y = R cos(Ω0 t − δ)
oscillations. In
where Ω0 = k/m. This has period Ω2π0 . In reality, c > 0 and these
oscillations are killed off by friction.

Non-Homogeneous Second-Order Linear O.D.E with Constant Co-

efficients
These are of the form
y 00 + ay 0 + by = r(x) (6)
From First-Year Calculus, we know that the general solution of (6) is
y = yH + yP
where yP is any solution of (6) and yH is the most general solution of the
homogeneous equation
y 00 + ay 0 + by = 0. (7)
Thus, we solve (6) in the following way:
1. Find the general solution yH of (7) by means already discussed. This
will contain two arbitrary constants and is sometimes called the com-
plementary function or complementary solution.
2. Produce any solution yP of (6), no matter how special. If r(x) is sim-
ple, the method of undetermined coefficients may be used (see later).
This solution is sometimes called the particular solution or particular
integral.
3. Add the above two solutions to give the general solution of (6).
4. If initial or boundary conditions are specified, impose them now to
determine the constants in yH .
Clearly, the crucial step is 2.

Method of undetermined coefficients

If r(x) is sufficiently simple, we can guess what yP must be as in the following
table.

21
 Term in r(x) choice of yP
keσx Ceσx
Pn (x) Qn (x)
k cos θx
or k sin θx K cos θx + M sin θx
keσx cos θx or keσx sin θx σx
e (K cos θx + M sin θx)
Pn (x)eσx cos θx
or Pn (x)eσx sin θx eσx (Qn (x) cos θx + Rn (x) sin θx)

Rules for Method of Undermined Coefficients

1. If a term in r(x) appears in the first column, choose the yP from the
corresponding second column and determine the unknown coefficients
by substituting into (6). In the table, Pn , Qn and Rn are polynomials
of degree n. Even if Pn (x) has only one term (xn ), Qn and Rn will in
general have all (n + 1) terms, i.e. Qn (x) = nj=0 qj xj .
P

2. If a term in our choice of yP happens to be a solution of the homogeneous

equation (7), we multiply that term by x. If our new guess is still a
solution of (7) we multiply by a further x. (This last case only happens
when the characteristic equation has equal roots).

3. If r(x) is a sum of functions, choose for yP (x) the sum of the appropriate
terms in the table.

Forced Oscillations
When simple periodic forcing is added to the mechanical or electrical
system studied earlier, we have to solve an equation like

my 00 + cy 0 + ky = F0 cos wt (8)

where, as before, m > 0, c > 0, k > 0.

This models a mechanical system driven periodic forces or an electrical system
forced by a periodic voltage. The solution of the homogeneous equation has
already been discussed. For yP we try

yP = a cos wt + b sin wt

and find a = m(w02 − w2 ) F∆0 , b = wc F∆0 where

k
w02 = and ∆ = m2 (w02 − w2 )2 + w2 c2 .
m
Thus
y = Θ cos(wt − δ)

22
where
F02
Θ2 = a2 + b2 =
∆
and
b wc
tan δ = = 2
.
a (m(w0 − w2 ))
Undamped Oscillations (c = 0)
In the ideal case of no mechanical friction or electrical resistance, the complete
solution of (8) is
F
y = A cos w0 t + B sin w0 t + cos wt
m(w02 − w2 )
q
k
for w 6= w0 . This is the sum of free oscillations with frequency w0 = m and
a forced oscillation with frequency w.
As w −→ w0 , the amplitude of the latter gets larger and larger. For
w = w0 , the above is not valid and the appropriate solution (using rule 2
above) is
F0
y = A cos w0 t + B sin w0 t + t sin w0 t
2mw0
The forced response thus consists of a sinusoid with linearly increasing ampli-
tude. This is the phenomenon of resonance and occurs when the frequency
w of the forcing is exactly equal to the natural frequency w0 with which the
system likes to oscillate.
Effects of Friction
In reality, c > 0 so the above is modified by friction or electrical resistance.
As we showed earlier, all solutions of the homogeneous equation decay to zero
as t −→ ∞ when c > 0. These are called the transients of the system. Thus
we are ultimately left with the directly forced response

yP = Θ cos(wt − δ)
q
where Θ
F0
=M = √ 1
and tan δ = wc
m(w02 −w2 )
and w0 = k
m
is
m2 (w02 −w2 )2 +w2 c2
the natural frequency of the undamped (c = 0) oscillations. The quantity M
is the magnification ratio for the amplitude. In engineering design we might
want this to be small so as to avoid damaging resonances or to be large to
magnify weak signals (e.g. tuning an AM radio). This will be discussed in
lectures.
The Method of Variation of Parameters
This is a general method for finding yP .
Consider
y 00 + p(x)y 0 + q(x)y = f (x)

23
The general solution is
yg = yh + yp
Complementary solution assumed known

yh (x) = Ay1 (x) + By2 (x)

For particular solution assume

yp (x) = A(x)y1 (x) + B(x)y2 (x)

Then yp (x) is a solution if

y1 A0 + y2 B 0 = 0
y10 A0 + y20 B 0 = f (x)

The variable parameters A(x), B(x) are found by solving these two equations
first for A0 (x), B 0 (x).
PROOF:

yp = Ay1 + By2
yp0 = A0 y1 + Ay10 + B 0 y2 + By20

But if
A 0 y1 + B 0 y2 = 0
N

Then
y 0 = Ay10 + By20
Now
y 00 = A0 y10 + Ay100 + B 0 y20 + By200
Substitute trial yp (x) into the ODE

A[y100 + py1 + qy1 ] +B[y200 + py20 + qy2 ]

+A0 y10 + B 0 y20 = f (x)

But y1 , y2 are solutions of the homogeneous part and so

A0 y10 + B 0 y20 = f (x)

L

N L
Thus yp (x) is a solution of the ODE provided that and are satisfied.
N L
Now solve and to find A(x) and B(x) First write in matrix form

! ! !
y1 y2 A0 0
=
y10 y20 B0 f

24
We can find a unique solution if the determinant is non-zero. However the

determinant is simply the Wronskian

!
y1 y2
W = det
y10 y20

Clearly this is non-zero because y1 , y2 are linearly independent

EXERCISE Solve the matrix equations and show that
Z
y2 (x)f (x) Z
y1 (x)f (x)
yp (x) = −y1 (x) dx + y2 (x) dx
W (x) W (x)

NOTE This method also works if p and q are constants.

TOPIC 6 – MATRICES
Brief revision (including special matrices)
A matrix is a rectangular array of numbers (real or complex) in the form

a11 a12 . . . a1n

 
 a21 a22 . . . a2n 
A=
 .. .. .. .. 

 . . . . 
am1 am2 . . . amn

Here, A is an m × n matrix which has m rows and n columns. We write

A = (ajk )

in which j is the row suffix and k the column suffix, e.g., a32 is the entry in
the 3rd row, 2nd column.
If all entries of A are real, we call A a real matrix; otherwise it is a complex
matrix.
Row vector : (matrix with one row)

a = (a1 , a2 , . . . , an )

Column vector : (matrix with one column)

b1
 
 .. 
b =  . .
bm

Square matrix : If m = n (i.e. A has the same number of rows and

columns) we have a square matrix A and we call n its order.

Addition of matrices, multiplication by scalars

25
Definition: Two matrices A = (ajk ) and B = (bjk ) are equal if and only if A
and B have the same number of rows and columns and

ajk = bjk for all j, k.

We write A = B.
Definition: Addition of matrices is defined only for matrices with the same
number of rows and columns. The sum of two m × n matrices A and B is
an m × n matrix A + B with entries

ajk + bjk (j = 1, . . . , m; k = 1, . . . n).

We define the zero matrix 0 to be the m × n matrix with all entries zero.
Properties of Addition:

a) A + B = B + A

b) (U + V ) + W = U + (V + W )

c) A + 0 = A

d) A + (−A) = 0, −A = (−ajk ).

Definition: Multiplication by a scalar

ca11 ca12 . . . ca1n

 
 ca21 ca22 . . . ca2n 
cA = Ac =  .. .. .. 
 
 . ...
. . 
cam1 cam2 . . . camn

or cA = (cajk ).
For any m × n matrices A and B and any scalars c and k

a) c(A + B) = cA + cB

b) (c + k)A = cA + kA

c) c(kA) = (ck)A

d) 1A = A

Matrix multiplication
Let A = (ajk ) be an m × n matrix and B = (bjk ) an r × p matrix. Then
the product AB (in this order) is defined only when r = n. (i.e. the number
of rows of B = the number of columns of A). Then AB is an m × p matrix
C = (cjk ) where

cjk = (j th row vector of A) · (k th column vector of B)

26
or
n
X
cjk = ajl blk = aj1 b1k + aj2 b2k + · · · + ajn bnk (j = 1, · · · m, k = 1, · · · p).
l=1

Properties of Matrix Multiplication: Assuming matrix multiplication is de-

fined

a) (kA)B = k(AB) = A(kB)

b) A(BC) = (AB)C

c) (A + B)C = AC + BC

d) C(A + B) = CA + CB

Important notes:
(1) If A and B are matrices such that AB and BA are defined then, in general,

AB 6= BA

(2) AB = 0 does not necessarily imply A = 0 or B = 0.

Special Matrices:
(1) Triangular Matrices: A square n × n matrix whose entries above the
main diagonal are all zero is called a lower triangular matrix. Similarly, a
square matrix whose entries below the diagonal are all zero is called a upper
triangular matrix.
(2) Diagonal Matrices: A square matrix A = ajk whose entries above and
below the main diagonal are all zero (i.e. ajk = 0 for j 6= k) is called a
diagonal matrix.
A scalar matrix is a diagonal matrix whose entries on the main diagonal
are all equal.
c 0 ... 0
 
0 c ... 0
S =  .. .. . . .
 
. . . .. 
0 0 ... c
We then have AS = SA = cA (for any n × n matrix A).
The unit matrix
1 0 ... 0
 
0 1 ... 0
I=  .. .. . . . .. 

. . .
0 0 ... 1
We then have AI = IA = A (for any n × n matrix A).

27
Transpose of a matrix: The transpose AT of an m × n matrix A = (ajk ) is
the n × m matrix in which the k th column of A becomes the k th row of AT
(and the j th row of A becomes the j th column of AT ).
a11 a21 . . . am1
 
 a12 a22 . . . am2 
AT = (akj ) =  .. .. .. 
 
..
 . . . . 
a1n a2n . . . amn
The transpose of a product of matrices is given by

(AB)T = B T AT .

Definition: A real square matrix A = (ajk ) is symmetric if AT = A.

Definition: A real square matrix A = (ajk ) is skew–symmetric if AT = −A.
Note: We can write any real square matrix as A = S + R (where S is
skew–symmetric and R is symmetric) where
1 1
R = (AT + A), S = (A − AT ).
2 2
Systems of linear equations, Gaussian Elimination

Consider the system of linear equations


a11 x1 + · · · + a1n xn = b1 

a21 x1 + · · · + a2n xn = b2



.. (9)
. 



am1 x1 + · · · + amn xn = bm


where the ajk are the coefficients. The system of equations (9) has m equa-
tions with n unknowns. If all bi = 0 the system is homogeneous. If at least
one of the bi 6= 0 the system is nonhomogeneous. We can write the system
(9) as
Ax = b
where
a11 a12 . . . a1n
 
x1 b1
   
 a21 a22 . . . a2n 
..  .. 
x= . , b= . , A =  .. .. .. 
   
 
 . ...
. . 
xn bm
am1 am2 . . . amn

We define the augmented matrix à or [A|b] as

a11 a12 . . . a1n b1
 
 a21 a22 . . . a2n b2 
à = [A|b] =  .. .. ..
 
 . ... 
. . 
am1 am2 . . . amn bm

28
This completely determines the system (9).
The system (9) is overdetermined if it has more equations than un-
knowns (m > n). It is determined if m = n and is undetermined if
m < n (i.e. the system has fewer equations than unknowns).
Theorem The system (9) has

1. No solution if the rank of A is not equal to the rank of the augmented

matrix Ã

2. A unique solution if the ranks of A and à both equal n.

3. Infinitely many solutions if the ranks of A and à are equal and < n.

Recall that the rank of a matrix is the maximum number of linearly inde-
pendent rows of the matrix. Somewhat surprisingly, this is also the maximum
number of linearly independent columns of the matrix or, in other words, row-
rank equals column-rank.
The solutions of (9) may be found by Gaussian elimination which is a
systematic process of elimination to reduce the matrix to its echelon form,
followed by back-substitution. Gaussian elimination is done using elementary
row operations:

1. interchange two rows

2. multiplication of a row by a nonzero constant

3. addition of a constant multiple of one row to another row.

{See first-year algebra notes for further details of Gaussian elimination.}

Inverse of a Matrix: The inverse of an n × n matrix A = (ajk ), denoted by

A−1 , is an n × n matrix such that

AA−1 = A−1 A = I

where I is the n × n identity (unit) matrix. If A has an inverse it is called

nonsingular. If A has no inverse it is called singular.
Existence of an inverse: For an n×n matrix A the inverse exists if and only if
the rank of A is equal to n. This is equivalent to saying that the determinant
of A (written det A or |A|) is non-zero. If the inverse exists it is unique and
the solution of (9) when A is a non-singular n×n matrix is given by x = A−1 b
. The inverse is of great theoretical importance. However, in practical prob-
lems, we solve the system by Gaussian elimination and back-substitution and
not by calculation of the inverse followed by A−1 b since the latter approach
involves more work than the former.

29
 Useful formula: For a nonsingular 2 × 2 matrix

a11 a12 1 a22 −a12

   
−1
A= A =
a21 a22 det A −a21 a11

where det A = a11 a22 − a12 a21 is the determinant of A. Note that a matrix
is nonsingular if det A 6= 0 (this holds for any general n × n matrix.)
Inverse of a diagonal matrix:
 
a11

0 ... 0

1/a11 0 ... 0
 0 a22 ... 0   0 1/a22 ... 0
 
−1

A=
 .. .. .. .. 

A =
 .. .. .. .. 
 . . . .   . . . .


0 0 . . . ann 0 0 . . . 1/ann

provided ajj 6= 0 for all j = 1, . . . n.

Inverse of a product of matrices:

(AC)−1 = C −1 A−1 .

Eigenvalues and eigenvectors:

Let A = (ajk ) be an n × n matrix. Consider the vector equation

Ax = λx, (10)

where λ is some scalar. The vector x = 0 is a solution of (10) for all λ. A

value of λ for which (10) has a nonzero solution is called an eigenvalue of
A; the corresponding vector x is called an eigenvector of A. Observe that
cx is also an eigenvector for any scalar c 6= 0.
Determination of eigenvectors: Any n × n matrix has at least one and at
most n distinct (real or complex) eigenvalues. Rewrite (10) as

(A − λI)x = 0.

This is a set of homogeneous linear equations and has a nontrivial (x 6= 0)

solution if and only if
 a11 − λ a12 ... a1n 
 

 a21 a22 − λ . . . a2n 
det(A − λI) = 
 .. .. .. .. =0 (11)

 . . . . 


an1 an2 . . . ann − λ 

Equation (11) then gives the characteristic equation (or polynomial) of

A.
To get the eigenvalues we must evaluate the determinant. This can be
done for any order matrix but in practice if n ≥ 4 (say) it is usually simpler
to determine the eigenvalues numerically.

30
 For a 2 × 2 matrix we get a quadratic in λ:
 a11 − λ a12 
 
det(A − λI) =  = (a11 − λ)(a22 − λ) − a12 a21
 a
21 a22 − λ 
For a 3 × 3 matrix we get a cubic in λ. For a general n × n matrix A the
characteristic equation is a polynomial of order n in λ. Even if A is real, some
of the eigenvalues may be complex and hence so will be the corresponding
eigenvectors.

Definition: If λ is an eigenvalue of order Mλ (i.e. a root of the characteristic

equation of order Mλ ) then Mλ is called the algebraic multiplicity of λ.
The number, mλ , of linearly independent eigenvectors corresponding to λ is
called the geometric multiplicity. In general mλ ≤ Mλ (i.e. geometric
multiplicity ≤ algebraic multiplicity).
Some Properties of Eigenvalues & Eigenvectors
In the following list of properties, λ is an eigenvalue of A with x being the
corresponding eigenvector.
1. Zero is an eigenvalue of A if and only if A is singular.
2. If k is a constant, the matrix kA has eigenvectors identical with those of
A and eigenvalues kλ
3. If m is a positive integer, the matrix Am has eigenvectors identical with
those of A and eigenvalues λm
4. A−1 has eigenvectors identical with those of A and eigenvalues λ−1
5. The transposed matrix AT has the same eigenvalues as A but, in general,
different eigenvectors
6. If A is a real symmetric matrix, (AT = A ), the eigenvalues of A are all
real and hence the eigenvectors may be taken to be real also.
7. The eigenvectors associated with different eigenvalues are linearly inde-
pendent
8. If A is a real n × n symmetric matrix, we can always find n linearly inde-
pendent eigenvectors of A, even if some of the eigenvalues are repeated.
9. If A is a real symmetric matrix, the eigenvectors associated with different
eigenvalues of A are orthogonal to one another.

Reminder : Orthogonality of Vectors

Two n × 1 column vectors a and b are said to be orthogonal if aT b = 0.
Since aT b = ni=1 ai bi , this is the n-dimensional version of the dot product
P

of two vectors in 3-dimensional space and so is sometimes written a · b.

Quadric surfaces
Quadric surfaces are surfaces in space whose equations combine quadratic
terms with linear terms and constants.
Examples:
x2 y 2 z 2
ellipsoid : + 2 + 2 = 1,
a2 b c

31
 x2 y2 z
elliptic paraboloid : + = ,
a2 b2 c
x2 y2 z2
elliptic cone : + 2 = 2,
a2 b c
x2 y2 z2
hyperboloid (one sheet) : + 2 − 2 = 1,
a2 b c
z2 x2 y2
hyperboloid (two sheet) : − 2 − 2 = 1.
c2 a b
Quadratic forms: Quadric surfaces can be written in terms of a quadratic
form. Consider a real n × n matrix A and real vector x. Then
n X
n
xT Ax = ajk xj xk = a11 x21 + a12 x1 x2 + · · · + a1n x1 xn
X

j=1 k=1

+ a21 x2 x1 + a22 x22 + · · · + a2n x2 xn

..
.
+ an1 xn x1 + an2 xn x2 + · · · + ann x2n ,

and the matrix A may be assumed to symmetric by re-defining the matrix

elements as (aij )new = (aji )new = 12 (aij + aji )old . We can then characterise
any quadric surface by a corresponding symmetric matrix A = AT .

Orthogonal matrices and diagonalisation

An orthogonal matrix is one for which

AT = A−1

An n × n matrix  is said to be similar to an n × n matrix A if

 = T −1 AT

for some nonsingular matrix T . The transformation, which gives  from A,

is called a similarity transformation.
Theorem: If  is similar to A, then  has the same eigenvalues as A. Also, if
x is an eigenvector of A then y = T −1 x is an eigenvector of  corresponding
to the same eigenvalue.
Proof:

Ax = λx
−1
⇒ T Ax = λT −1 x
⇒ T AT T −1 x
−1
= λT −1 x
⇒ Â(T −1 x) = λ(T −1 x)

Therefore λ is an eigenvalue of  with eigenvector T −1 x.

32
Theorem: Let λ1 , λ2 , . . . λk be distinct eigenvalues of an n × n matrix. Then
the corresponding eigenvectors x1 , x2 , . . . , xk form a linearly independent set.
This is basically just Property 7 above.
Theorem: If an n × n matrix A has n distinct eigenvalues, then A has n
linearly independent eigenvectors which therefore constitute a basis for Rn
(or C n ). This follows from Property 7 above.
Orthonormal Sets of Vectors: Consider the set of vectors v1 , v2 , . . . , vn
This set is said to be orthonormal if vjT vj = 1 for j = 1, . . . , n and vjT vk = 0
for j 6= k. The matrix with these vectors as column vectors is orthogonal i.e..
if P = (v1 , v2 , . . . , vn ) then P −1 = P T .

Theorem: The eigenvalues of a real, symmetric matrix are all real. Hence, we
may take the eigenvectors to be real also.
Theorem: A real, symmetric matrix has an orthonormal basis of eigenvec-
tors in Rn .
Hence, even though some of the eigenvalues of a real, symmetric matrix
may be equal, we can always find n orthogonal eigenvectors. By normalising
each of these to have length 1, we can construct an orthonormal basis for Rn .
Diagonalisation of a Matrix
Theorem: If an n × n matrix A has a basis of eigenvectors (i.e. n linearly
independent eigenvectors) then
D = P −1 AP
is diagonal, with the eigenvalues of A as the entries on main diagonal. P is
the matrix with the eigenvectors of A as column vectors. Also
Dm = P −1 Am P.
Diagonalisation of a Symmetric Matrix
Not all matrices can be diagonalised - it relies upon us being able to find n
linearly independent eigenvectors. Fortunately, a real symmetric matrix can
always be diagonalised (because of Property 8 above). Even better, because
of Property 9, the columns of P are orthogonal to one another. If we further
normalise each eigenvector by dividing it by its length (the square root of the
sum of the squares of its n components), the matrix P is then an orthogonal
matrix and so its inverse is just its transpose. This saves a lot of work.
Transformation of a quadratic form to principal axes
Suppose we have a quadratic form
Q = xT Ax (12)
where A is real–symmetric. Then by a previous theorem, A has an orthonor-
mal basis of n eigenvectors. Hence the P matrix with eigenvectors of A as
column vectors is orthogonal. Now
D = P −1 AP ⇒ A = P DP −1 ⇒ A = P DP T

33
Substitution of this into (12) gives

Q = xT P DP T x

Set P T x = y ⇒ x = P y (using P −1 = P T ). Then

Q = yT Dy = λ1 y12 + λ2 y22 + · · · + λn yn2 (13)

where λj are the eigenvalues of A. This proves the following

Theorem: The substitution x = P y transforms the quadratic form
n X
n
Q = xT Ax =
X
ajk xj xk
j=1 k=1

to the principal axis form (13) where λ1 , . . . , λn are the eigenvalues (not
necessarily distinct) of the symmetric matrix A and P is the orthogonal ma-
trix with the corresponding eigenvectors p1 , . . . , pn as column vectors.

Systems of linear ordinary differential equations

Consider the first order system of equations
n
dyj X
= ajk yk + hj (t), j = 1, . . . n. (14)
dt k=1

(for simplicity we will assume that the coefficients ajk are constant).
Note: any higher (nth) order o.d.e can be reduced to a system of n first order
o.d.e’s.
Example:
d3 x d2 x
+ 2 2 + x = 0.
dt3 dt
Define

x = y1 ,
dx dy1
= = y2 ,
dt dt
d2 x dy2
= = y3 ,
dt2 dt
d3 x dy3 d2 x
= = −2 2 − x = −2y3 − y1 .
dt3 dt dt
Then
dy1 dy2 dy3
= y2 , = y3 , = −2y3 − y1 .
dt dt dt
Returning to (14): Rewrite this in matrix form

dy
= Ay + h, A = (ajk ). (15)
dt

34
Assume A has a basis of eigenvectors p1 , . . . pn . Then P = (p1 , . . . , pn ) is
nonsingular. To diagonalise (15) we write

y = Pz

Then assuming that all the ajk are constant (⇒ P is constant) we have
y0 = P z0 and (15) becomes

P z0 = AP z + h
⇒ z0 = P −1 AP z + P −1 h,
⇒ z0 = Dz + P −1 h,

where D is the diagonal matrix with the eigenvalues of A along the diagonal.
Writing the last expression in component form gives

zj0 − λj zj = rj (t),

where rj is the jth component of P −1 h. This is then a first order, constant

coefficient o.d.e whose general solution is
Z 
zj (t) = eλj t e−λj t rj (t) dt + Cj

with Cj arbitrary constants. These can be determined by supplying initial

conditions
y(t0 ) = y0 ⇒ z(t0 ) = z0 = P −1 y0 . (16)
Existence and Uniqueness: Let h(t) in (15) be continuous in the interval
α < t < β and let t0 be any given point in that interval. Assume that A has
a set of n linearly independent eigenvectors. Then the initial value problem
(15), (16) has a unique solution in the interval.
Stability: All solutions of (15) with h = 0 approach zero as t −→ ∞ if and
only if all the eigenvalues of A have negative real parts.

TOPIC 7 – THE LAPLACE TRANSFORM

Laplace Transforms

Let f (t) be a given function defined for all t ≥ 0. If the integral

Z∞
F (s) = e−st f (t) dt
0

exists it is called the Laplace transform of f (t). We denote it by L(f ) i.e.

Z∞
L(f ) = F (s) = e−st f (t) dt (17)
0

35
The original function f (t) in (17) is called the inverse transform of F (s); we
denote it by L−1 (F ), so
f (t) = L−1 (F ).
Linearity: The Laplace transformation is a linear operation. For any functions
f (t) and g(t) whose transforms exist and any constants a and b
L {af (t) + bg(t)} = aL {f (t)} + bL {g(t)} . (18)
Tables of Laplace Transforms: A short table of Laplace Transforms is at the
end of these notes.
Existence theorem: Let f (t) be a function that is piecewise continuous (i.e.
a function which is continuous except at a finite number of points) on every
finite interval in the range t ≥ 0 and satisfies
|f (t)| ≤ M eγt for all t ≥ 0, (19)
for some constants γ and M . Then the Laplace transform of f (t) exists for
all s > γ.

Transform of derivatives
Theorem: Suppose that f (t) is continuous for all t ≥ 0, satisfies (19) (for
some γ and M ) and f 0 (t) is piecewise continuous on every finite interval in
the range t ≥ 0. Then the Laplace transform of f 0 (t) exists when s > γ and
L(f 0 ) = sL(f ) − f (0) (for s > γ) (20)
Higher derivatives

L(f 00 ) = sL(f 0 ) − f 0 (0),

= s [sL(f ) − f (0)] − f 0 (0)
= s2 L(f ) − sf (0) − f 0 (0)
Continuing gives us the general result
L(f (n) ) = sn L(f ) − sn−1 f (0) − sn−2 f 0 (0) − · · · − f (n−1) (0).
Transform of an integral of a function
If f (t) is piecewise continuous and satisfies our earlier inequality then
 t 
Z
1
L f (τ )dτ  = L(f (t)) (for s > γ).
s
0

Hence if we write L(f (t)) = F (s) then

 t 
Z
F (s)
L  f (τ )dτ  =
s
0

36
Taking the inverse Laplace transform gives
! Zt
−1 F (s)
L = f (τ )dτ.
s
0

Shifting on the s–axis

Suppose f (t) has the transform F (s) where s > γ then eat f (t) has the
transform F (s − a) where s − a > γ, i.e. if

L(f (t)) = F (s)

then
L(eat f (t)) = F (s − a), (21)
So if we know the transform F (s) of a function f (t), (21) gives us the trans-
form of exp(at)f (t) just by shifting on the s axis i.e. replacing s with s − a
to give F (s − a). Taking the inverse transform in (21) gives

L−1 (F (s − a)) = eat f (t).

An example: Find L(eat cos ωt).

We know that
s
L(cos ωt) =
s2 + ω2
s−a
(21) ⇒ L(eat cos ωt) =
(s − a)2 + ω 2
Shifting on the t–axis
If f (t) has the transform F (s) and a ≥ 0 then the function
(
0 if t < a
f˜(t) = (22)
f (t − a) if t > a

has the transform

e−as F (s).
Thus if we know F (s) is the transform of f (t) then we get the transform of
(22) by multiplying F (s) by e−as .

Unit step function u(t)

Define
 0, if t < 0;

1
u(t) = 2
, if t = 0;

1, if t > 0.
(also called the Heaviside function). Note: our previous function

0, if t < a;

f˜(t) =
f (t − a), if t > a

37
can now be written as

f˜(t) = f (t − a)u(t − a)

(strictly, we should define f˜(a) = f (0)/2) and our shifting on the t–axis result
is
L {f (t − a)u(t − a)} = e−as F (s) for a ≥ 0
and the inverse
n o
L−1 e−as F (s) = f (t − a)u(t − a) for a ≥ 0

Another useful formula is

e−as
L {u(t − a)} = for a ≥ 0
s
Partial Fractions
In most applications we obtain Y (s) = L(y), the transform of the function
y, in the form
F (s)
Y (s) = ,
G(s)
where F (s) and G(s) are polynomials in s. For example in the solution of an
ordinary differential equation we would expect a transform of this form.
Assume F (s) and G(s) have real coefficients and no common factors. The
degree of F is lower than the degree of G. In practice we express F/G in
terms of partial fractions
Example:
F (s) 1 1
= 2 =
G(s) (s − 3s − 4) (s − 4)(s + 1)
A B
= +
(s − 4) (s + 1)
The form of the partial fraction depends on the type of factor in F/G. The
four common ones are

Case 1: unrepeated factor (s − a)

Case 2: repeated factor (s − a)m
Case 3: complex factors (s − a)(s − ā) (ā complex conjugate of a)
Case 4: repeated complex factors [(s − a)(s − ā)]2

Cases 1–4 lead to a general form of partial fraction for Y = F/G which then
have a corresponding inverse Laplace transform.
Case 1: In Y = F/G a fraction
A
(s − a)

38
has inverse transform
Aeat
with
A = F (a)/G0 (a).
Case 2: In Y = F/G a sum of m fractions

Am Am−1 A1
+ + · · · + .
(s − a)m (s − a)m−1 (s − a)

Using tables and s–shifting this has inverse transform

( )
at Am Am−1 m−2 A2 1
e tm−1 + t + ··· + t + A1 ,
(m − 1)! (m − 2)! 1!

where
(s − a)m F (s)
Am = lim
s→a G(s)
dm−k (s − a)m F (s)
" #
1
Ak = lim (k = 1, . . . , m − 1)
(m − k)! s→a dsm−k G(s)

Case 3: unrepeated complex factors (s − a)(s − ā). Define a = α + iβ. Then

(s − a)(s − ā) = (s − α)2 + β 2

This corresponds to the partial fraction

As + B A(s − α) + αA + B
or
(s − α)2 + β 2 (s − α)2 + β 2

Using the table and s–shifting theorem gives the inverse transform
!
αt αA + B
e A cos βt + sin βt
β

Here A is the imaginary part and (αA + B)/β is the real part of

1 [(s − α)2 + β 2 ]F (s)

Qa = lim
β s→a G(s)

Case 4: repeated complex factors [(s − a)(s − ā)]2 . These correspond to

partial fractions in Y = F/G

As + B Cs + D
2 2 2
+
[(s − α) + β ] (s − α)2 + β 2

39
which has the inverse transform
" #
αt A αA + B
e t sin βt + (sin βt − βt cos βt)
2β 2β 3
" #
αt αC + D
+ e C cos βt + sin βt
β
The constants are given by formulas similar to those above.
Solving Ordinary Differential Equations
One of the main uses of Laplace transforms is in solving differential equa-
tions, both ordinary and partial.
Example: Using Laplace transforms solve

y 00 + y = 2 cos t, y(0) = 2, y 0 (0) = 0.

Solution: Take Laplace transform of the differential equation. Define Y (s) =

L(y).
h i
s2 L(y) − sy(0) − y 0 (0) + L(y) = L(2 cos t),
2s
⇒ (s2 + 1)L(y) − 2s = 2
s +1
2s 2s
⇒ L(y) = 2 + 2
s + 1 (s + 1)2
We have a complex and repeated complex factor.
2s
 
−1
L 2
= 2 cos t.
s +1
!
−1 2s
L = t sin t.
(s + 1)2
2

Therefore
y(t) = 2 cos t + t sin t.

TOPIC 8 – FOURIER SERIES

Periodic functions, trigonometric series
Definition: A function f (x) is periodic if it is defined for all real x and if
there is some positive number p such that

f (x + p) = f (x) for all x . (23)

The number p is then called the period. From (23) we then have

f (x + 2p) = f (x + p + p) = f (x + p) = f (x)

40
Thus for any integer n

f (x + np) = f (x) for all x.

If f (x) and g(x) have period p then so does

h(x) = af (x) + bg(x) (a, b constants).

Simple examples:

f (x) = constant is periodic

f (x) = sin x is periodic with period 2π
f (x) = x is not periodic

Observe that a function which is constant has period p for any p > 0. If
f is periodic but not constant, the smallest positive p for which (23) holds is
called the primitive period of f .
We will want to represent functions of period p = 2π in terms of simple
trigonometric functions 1, cos x, sin x, cos 2x, sin 2x, etc. Such a representation
will give rise to a trigonometric series.

a0 + a1 cos x + b1 sin x + a2 cos 2x + b2 sin 2x + · · · (24)

where a0 , a1 , b1 etc are real constants—the an and bn are called the coeffi-
cients of the series. We can rewrite (24) as
∞
X
a0 + (an cos nx + bn sin nx).
n=1

Each term in the series has period 2π. Hence if the series converges, its sum
will be a function with period 2π.

Fourier series, Euler formulæ

Assume f (x) is a periodic function of period 2π that can be represented
by the trigonometric series
∞
X
f (x) = a0 + (an cos nx + bn sin nx) (25)
n=1

i.e. assume the series converges and has f (x) as its sum.
Question: How do we determine the coefficients an and bn in (25) ?
First a0 . Integrate both sides of (25) from −π to π.
Zπ Zπ " ∞
X
#
f (x) dx = a0 + (an cos nx + bn sin nx) dx.
−π −π n=1

41
Assuming that term-by-term integration is valid, we have
Zπ Zπ ∞
X Zπ ∞
X Zπ
f (x) dx = a0 dx + an cos nx dx + bn sin nx dx
−π −π n=1 −π n=1 −π

The last two integrals in this expression are zero. Therefore

π
1 Z
a0 = f (x) dx (26)
2π
−π

To determine am multiply (25) by cos mx, where m is any positive integer,

and integrate from −π to π.
Zπ Zπ " ∞
X
#
f (x) cos mx dx = a0 + (an cos nx + bn sin nx) cos mx dx.
−π −π n=1

Integrating term by term gives

Zπ Zπ
f (x) cos mx dx = a0 cos mx dx +
−π −π
Zπ Zπ
 
∞ 
X 
a cos nx cos mx dx + bn sin nx cos mx dx
 n 
n=1 −π −π

If we now make use of

Zπ π π
1Z 1Z
cos nx cos mx dx = cos(n + m)x dx + cos(n − m)x dx
2 2
−π −π −π

Zπ π π
1Z 1Z
sin nx cos mx dx = sin(n + m)x dx + sin(n − m)x dx
2 2
−π −π −π

and evaluating these integrals shows that all four integrals are zero except for
π
1Z 0, if m 6= n;

cos(n − m)x dx =
2 π, if m = n.
−π

We then have
Zπ
f (x) cos mx dx = πam
−π
or π
1Z
am = f (x) cos mx dx (m = 1, 2, . . .) (27)
π
−π

42
Finally to obtain the coefficients bm we multiply (25) by sin mx (where m is a
positive integer) and integrate from −π to π. Following the above procedure
we obtain π
1Z
bm = f (x) sin mx dx (m = 1, 2, . . .) (28)
π
−π

To summarise
Rπ

1
a0 = f (x) dx


2π


−π 

Rπ


1
am = π
f (x) cos mx dx (m = 1, 2, . . .) (29)
−π 

1 Rπ 

bm = f (x) sin mx dx (m = 1, 2, . . .)


π


−π

The expressions (29) are known as the Euler formulæ. In these integrals,
the limits need not be −π and π. Since everything has period 2π, we may
take the limits to be α and α + 2π for any real number α. In particular, it
is often convenient to take them to be 0 and 2π. The a0 , an , bn are called the
Fourier coefficients of f (x) and the series
∞
X
a0 + (an cos nx + bn sin nx). (30)
n=1

is called the Fourier series of f (x).

Functions of arbitrary period

Suppose f (x) = f (x+2L) for all x, i.e., f has period 2L. Then its Fourier
series is of the form
∞
X nπx nπx
f (x) = a0 + (an cos + bn sin )
n=1 L L

with the Fourier coefficients given by the Euler formulæ

L
1 Z
a0 = f (x) dx
2L
−L
L
1 Z mπx
am = f (x) cos dx (m = 1, 2, . . .)
L L
−L
L
1 Z mπx
bm = f (x) sin dx (m = 1, 2, . . .)
L L
−L

The proof follows our earlier derivation of the Euler formulæ for 2π–periodic
functions. We can also prove these results by a change of scale. Set v = πx/L.

43
Then x = ±L ⇒ v = ±π and f (x) = g(v) where g(v) has period of 2π in v.
So g(v) has Fourier series
∞
X
g(v) = a0 + (an cos nv + bn sin nv)
n=1

with coefficients given by Euler formulæ (29). Changing variables back to x

gives the required result.

Convergence of Fourier Series

There are a great many theorems on the convergence of Fourier Series.
One of the most straightforward is:
Theorem: If f has period 2L and is piecewise continuous and has a left-
hand and a right-hand derivative at each point then the Fourier series con-
verges. The sum of the series is f (x0 ) if f is continuous at x = x0 and
(f (x0 +) + f (x0 −))/2 if f is discontinuous at x = x0 .

Sums of functions
(1) The Fourier coefficients of a sum f1 +f2 are the sums of the corresponding
Fourier coefficients of f1 and f2 .
(2) The Fourier coefficients of cf are c times the Fourier coefficients of f .

Even and Odd functions

The above Fourier series simplify somewhat if f (x) is either an even func-
tion or an odd function. Remember that y = g(x) is even if g(−x) = g(x)
for all x and y = h(x) is odd if h(−x) = −h(x) for all x.
If g(x) is an even function then
ZL ZL Z0
g(x) dx = g(x) dx + g(x) dx
−L 0 −L

Put v = −x in the second integral to give

ZL ZL Z0
g(x) dx = g(x) dx + (−g(−v)) dv
−L 0 L
ZL ZL
= g(x) dx + g(v) dv
0 0
ZL
=2 g(x) dx
0

Similarly if h(x) is an odd function

ZL ZL Z0
h(x) dx = h(x) dx + h(x) dx
−L 0 −L

44
 ZL Z0
= h(x) dx + (−h(−v)) dv
0 L
ZL ZL
= h(x) dx − h(v) dv
0 0
=0

Fourier series for even and odd functions

Using the above results, we may readily show that the Fourier series of an
even function f (x) of period 2L is a Fourier cosine series
∞
X nπx
f (x) = a0 + an cos
n=1 L

with coefficients
L L
1Z 2Z mπx
a0 = f (x) dx am = f (x) cos x (m = 1, 2, . . .)
L L L
0 0

The Fourier series of an odd function f (x) of period 2L is a Fourier sine

series ∞
X nπx
f (x) = bn sin
n=1 L
with coefficients
L
2Z mπx
bm = f (x) sin dx (m = 1, 2, . . .)
L L
0

Half–Range expansions

Suppose we have a function f (x) defined on some interval 0 ≤ x ≤ L and we

want to find the Fourier series of this function. By taking period 2L we can get

(i) A Fourier cosine series for f (x) by extending it to an even function on the
interval −L ≤ x ≤ L. The cosine half–range expansion for f (x) is thus
∞
X nπx
f (x) = a0 + an cos
n=1 L

where
L L
1Z 2Z mπx
a0 = f (x) dx am = f (x) cos dx (m = 1, 2, . . .)
L L L
0 0

45
(ii) A Fourier sine series for f (x) by extending it to an odd function on the
interval −L ≤ x ≤ L. The sine half–range expansion for f (x) is thus
∞
X nπx
f (x) = bn sin
n=1 L
where
L
2Z mπx
bm = f (x) sin dx (m = 1, 2, . . .)
L L
0
In (0, L) both these series represent f (x). Outside (0, L) they represent differ-
ent functions – the even and odd periodic extensions of f respectively.
Complex form of the Fourier series
Recall that
eiθ = cos θ + i sin θ (31)
Taking the complex conjugate gives
e−iθ = cos θ − i sin θ (32)
Then (31)+(32) and (31)-(32) give
1  iθ  1  iθ 
cos θ = e + e−iθ sin θ = e − e−iθ .
2 2i
Hence
1  inx  1  inx 
cos nx = e + e−inx sin nx = e − e−inx .
2 2i
Now consider the Fourier series
∞
X
f (x) = a0 + (an cos nx + bn sin nx)
n=1
∞
an  inx  bn  inx 
e + e−inx + e − e−inx
X
= a0 +
n=1 2 2i
∞
1 1
(an − ibn )einx + (an + ibn )e−inx
X
= a0 +
n=1 2 2
∞  
cn einx + kn e−inx
X
⇒ f (x) = c0 +
n=1

where
1
cn = (an − ibn ), kn = c∗n .
2
Remember that
π
1Z
an = f (x) cos nx dx
π
−π
Zπ
1
bn = f (x) sin nx dx
π
−π

46
Hence
π π
 
1 1 Z i Z 
cn =  f (x) cos nx dx − f (x) sin nx dx
2 π π
−π −π
Zπ
1
= f (x)(cos nx − i sin nx) dx
2π
−π
Zπ
1
= f (x)e−inx dx
2π
−π

Similarly
π
1 Z
kn = f (x)einx dx
2π
−π

and note that kn = c−n . Then we have

∞ ∞
cn einx + c−n e−inx
X X
f (x) = c0 +
n=1 n=1
∞ −∞
cn einx + cn einx
X X
= c0 +
n=1 n=−1

Finally noting that ei(0)x = 1 we have

∞
cn einx
X
f (x) =
n=−∞

with π
1 Z
cn = f (x)e−inx dx.
2π
−π

This is then the complex form of the Fourier series for f (x). The cn are the
complex Fourier coefficients.

Forced Oscillations
Earlier we saw that forced oscillations of a body of mass m on a spring
are governed by the equation

my 00 + cy 0 + ky = r(t)

where k is the spring modulus, c the damping constant. If r(t) is a cos or sin
we can solve this equation by the method of undetermined coefficients.
What if r(t) is some other periodic function? We can represent r(t) by a
Fourier series; we should then be able to find a particular solution as a similar
Fourier series. To fix our ideas we consider a simple example

47
Example: Find the general solution to
y 00 + ω 2 y = r(t)
where ∞
X 1
r(t) = 2
sin(2n − 1)t.
n=1 (2n − 1)
Solution: We have
∞
00 2
X 1
y +ω y = 2
sin(2n − 1)t (33)
n=1 (2n − 1)

Consider the equation

1
yn00 + ω 2 yn = sin nt (n = 1, 3, 5, . . .) (34)
n2
Find the general solution of this equation. The general solution of the homo-
geneous form of (33) is
yh = A cos ωt + B sin ωt.
For a particular solution of (34) try
yn = An cos nt + Bn sin nt.
Differentiating and substituting into (34) gives
1
(−n2 + ω 2 )An cos nt + (−n2 + ω 2 )Bn sin nt = sin nt
n2
(assuming ω 6= n for n odd) we have
1
An = 0, Bn = .
n2 (ω 2 − n2 )
Thus the particular solution of (34) is
1
yn = sin nt.
n2 (ω 2− n2 )
Since (33) is linear the general solution is then a superposition
y1 + y3 + y5 + · · · + yh
Therefore
∞
X 1
y = A cos ωt + B sin ωt + sin(2n − 1)t.
n=1 (2n − 1)2 [ω 2− (2n − 1)2 ]

TOPIC 9 – PARTIAL DIFFERENTIAL EQUATIONS

Basic concepts
An equation involving one or more partial derivatives of an (unknown)
function of two or more independent variables is called a partial differential
equation.

48
 • The order of the highest derivative is called the order of the equation.

• A p.d.e (partial differential equation) is linear if it is of first degree in

the dependent variable (unknown function) and its partial derivatives.

• A partial differential equation is homogeneous if each term contains

either the dependent variable or one of its derivatives; otherwise it is
inhomogeneous.

• A solution of a p.d.e.. in some region R of the space of the independent

variables is a function that has all the partial derivatives appearing in
the equation in some domain containing R and satisfies the equation
everywhere in R.

Some examples
∂ 2u 2
2∂ u
(1) = c one dimensional wave equation
∂t2 ∂x2
2
∂u 2∂ u
(2) =c one dimensional heat equation
∂t ∂x2
∂ 2u ∂ 2u
(3) + =0 two dimensional Laplace equation
∂x2 ∂y 2
∂ 2u ∂ 2u
(4) + = f (x, y) two dimensional Poisson equation
∂x2 ∂y 2
(1)-(3) are homogeneous, linear. (4) is inhomogeneous, linear. All are second
order.
We may, depending on the problem, have boundary conditions (the
solution has some given value on the boundary of some domain) or initial
conditions (where the value of the solution will be given at some initial time,
e.g. t = 0).
Superposition Theorem: If u1 and u2 are any solutions of a linear homogeneous
p.d.e in some region, then

u = c1 u1 + c2 u2

is also a solution of the equation in the region.

Vibrating Strings
As shown in Section 2 of Chapter 11 of the book by Kreyszig, the small
transverse oscillations of a tightly stretched string satisfy
∂ 2u 2
2∂ u
= c where c2 = T /ρ,
∂t2 ∂x2
T is the tension in the string, ρ is the constant mass per unit length of the
string and u(x, t) is the displacement of the string.

49
D’Alembert’s solution of the wave equation

The equation
∂ 2u 2
2∂ u
= c (35)
∂t2 ∂x2
is called the one-dimensional wave equation.
Introduce new variables

v = x + ct, z = x − ct

then u in (35) becomes a function of v and z. By the chain rule we have

∂u ∂u ∂v ∂u ∂z
= +
∂x ∂v ∂x ∂z ∂x
Thus
ux = uv + uz
Then
∂ 2u
! !
∂ ∂u ∂u ∂v ∂ ∂u ∂u ∂z
2
= + + +
∂x ∂v ∂v ∂z ∂x ∂z ∂v ∂z ∂x
⇒ uxx = uvv + 2uvz + uzz .
Similarly
utt = c2 (uvv − 2uvz + uzz ).
Combining these in (35) gives

c2 (uvv − 2uvz + uzz ) = c2 (uvv + 2uvz + uzz ).

Thus
uvz = 0. (36)
Integrating (36) gives
∂u
= h(v)
∂v Z
⇒ u = h(v)dv + ψ(z)

i.e.
u = φ(v) + ψ(z)
R
where φ = h(v)dv. Thus

u = φ(x + ct) + ψ(x − ct) (37)

where φ and ψ are arbitrary functions. This is D’Alembert’s solution of the

wave equation and is valid for c constant.

50
 The functions φ and ψ can be determined from the initial conditions.
Suppose we have to solve (35) for −∞ < x < ∞ and t > 0 subject to the
initial conditions
∂u(x, 0)
u(x, 0) = f (x), = g(x) for −∞<x<∞
∂t
From (37) we have
∂u
= cφ0 (x + ct) − cψ 0 (x − ct)
∂t
where the 0 denotes differentiation with respect to the entire argument x + ct
and x − ct respectively. Then

cφ0 (x) − cψ 0 (x) = g(x) (38)

φ(x) + ψ(x) = f (x) (39)

We consider only the case g(x) = 0. From (39) we have

cφ0 (x) − cψ 0 (x) = 0 ⇒ φ(x) = ψ(x) + k (constant)

And from (39)

2ψ(x) = f (x) − k
1 1
⇒ ψ = [f (x) − k] , φ= [f (x) + k]
2 2
Hence
1
u(x, t) = [f (x + ct) + f (x − ct)] .
2
If g(x) 6= 0 then
1 1 Z x+ct
u(x, t) = [f (x + ct) + f (x − ct)] + g(s)ds.
2 2c x−ct
Note: In the solution (37), ψ(x − ct) represents a disturbance travelling
to the right with speed c without changing its shape and φ(x + ct)
represents a disturbance travelling to the left with speed c without
changing its shape.

SEPARATION OF VARIABLES
This is an important technique which can be used to find the solutions of
many linear partial differential equations and is best illustrated by examples.
Wave equation
The one–dimensional wave equation

∂ 2u 2
2∂ u
= c (40)
∂t2 ∂x2

51
(e.g. a model of the deflection u(x, t) of an elastic string) can be solved by
this technique. Physically, conditions will be imposed on the solution of (40).
Boundary conditions (string fixed at ends x = 0 and x = L). Then

u(0, t) = 0, u(L, t) = 0 for all time t (41)

Initial conditions (the motion of the string will depend upon its initial
deflection and its initial velocity i.e. the deflection and velocity at t = 0).

initial deflection u(x, 0) = f (x) (42)


∂u 
initial velocity  = g(x) (43)
∂t  t=0
The problem is to find the solution of (40) satisfying the conditions (41), (42)
and (43) .
This a 3 step process.

Step 1. Apply the method of separation of variables to obtain two

ordinary differential equations.
Step 2. Determine the solution of these two equations that satisfy the bound-
ary conditions.
Step 3. Combine these solutions so that the result will be a solution of (40)
which also satisfies the initial conditions (42) and (43).
Step 1: Separation of variables.
We look for a solution of (40) in the form

u(x, t) = F (x)G(t).

Differentiating gives
∂ 2u d2 G ∂ 2u d2 F
= F , = G
∂t2 dt2 ∂x2 dx2
Substituting into equation (40) gives
d2 G 2
2 d F
F = c G
dt2 dx2
rearranging (divide by c2 F G)
1 d2 G 1 d2 F
=
c2 G dt2 F dx2
1 d2 G 1 d2 F
⇒ = = k (k constant).
c2 G dt2 F dx2
This then gives us two ordinary differential equations:
d2 F d2 G
− kF = 0, − kc2 G = 0 . (44)
dx2 dt 2

52
Step 2: Satisfy the boundary conditions. With u = F G

u(0, t) = F (0)G(t) = 0 u(L, t) = F (L)G(t) = 0 for all t

There are then two possibilities G(t) = 0 for all t ⇒ u ≡ 0. Or

F (0) = F (L) = 0 (45)

We then need to solve the first of (44) subject to (45).

Cases: (A) k = 0 ⇒ F = ax + b but (45) ⇒ a = b = 0. Hence F ≡ 0 ⇒
u ≡ 0 (trivial solution).
(B) k > 0. Take k = µ2 . Then

F = Aeµx + Be−µx

But (45) ⇒ A = B = 0. Hence F ≡ 0 ⇒ u ≡ 0 (trivial solution).

(C) k < 0. Take k = −p2 where ,without loss of generality, p > 0.
Then
F = C cos px + D sin px
Boundary conditions

F (0) = 0 ⇒ C=0
nπ
F (L) = 0 ⇒ D sin pL = 0 ⇒ sin pL = 0 ⇒ p= (n positive integer)
L
So p = nπ/L and D is arbitrary (take D = 1). We then have an infinite
number of solutions
nπx
Fn (x) = sin for n = 1, 2, 3, · · ·
L
Now with k = −p2 = −(nπ/L)2 we have

d2 G
2
cnπ

2
+ G = 0.
dt L
which has the general solution

Gn (t) = Bn cos λn t + Cn sin λn t for n = 1, 2, 3, · · ·

where λn = cnπ/L. Hence

un (x, t) = Fn (x)Gn (t)

or
nπx
un (x, t) = (Bn cos λn t + Cn sin λn t) sin for n = 1, 2, 3, · · ·
L

53
are solutions of (40) satisfying the boundary conditions (41) . We call the
un ’s the eigenfunctions and the λn ’s the eigenvalues. The set {λ1 , λ2 , · · ·}
is called the spectrum.
Step 3: Satisfy the initial conditions. The initial conditions are given by
(42) and (43) . Since the wave equation is linear and homogeneous, a sum of
solutions is also a solution. Consider
∞ ∞
X X nπx
u(x, t) = un (x, t) = (Bn cos λn t + Cn sin λn t) sin
n=1 n=1 L

with λn = cnπ/L. From (42) we have

∞
X nπx
u(x, 0) = f (x) = Bn sin
n=1 L

This is just the Fourier–sine half–range expansion for f (x). Thus

L
2Z nπx
Bn = f (x) sin dx (for n = 1, 2, . . .) (46)
L L
0

We have ∞
∂u X nπx
= λn (−Bn sin λn t + Cn cos λn t) sin
∂t n=1 L
From (43)
∞

∂u  X nπx
 = g(x) = λn Cn sin
∂t t=0

n=1 L
Again, this is the Fourier–sine half–range expansion for g(x). Thus
L
2Z nπx
λn Cn = g(x) sin dx
L L
0

or
L
2 Z nπx
Cn = g(x) sin dx (for n = 1, 2, . . .) (47)
cnπ L
0

Summary

∂ 2u 2
2∂ u
= c
∂t2 ∂x2

∂u 
u(x, 0) = f (x)  = g(x)
∂t t=0
has solution (see note below)
∞
X nπx
u(x, t) = (Bn cos λn t + Cn sin λn t) sin
n=1 L

54
where
L
2Z nπx
Bn = f (x) sin dx (for n = 1, 2, . . .)
L L
0
L
2 Z nπx
Cn = g(x) sin dx (for n = 1, 2, . . .)
cnπ L
0

Note: it is a solution if the infinite series converges and the series obtained
by differentiating twice with respect to x and t converges with sums uxx and
utt (which are continuous).
For the case g(x) = 0, we have Cn = 0 and
∞
X nπx
u(x, t) = Bn cos λn t sin
n=1 L

We can write
cnπt nπx 1 nπ nπ
    
cos sin = sin (x − ct) + sin (x + ct)
L L 2 L L
Then
∞ ∞
1X nπ 1X nπ
   
u(x, t) = Bn sin (x − ct) + Bn sin (x + ct)
2 n=1 L 2 n=1 L

but ∞
X nπx
Bn sin = f (x)
n=1 L
(the Fourier sine series for f (x)). Thus
1 ∗
u(x, t) = {f (x − ct) + f ∗ (x + ct)}
2
where f ∗ is the odd periodic extension of f with period 2L. Since f (x) is
continuous on 0 ≤ x ≤ L and zero at endpoints ⇒ u(x, t) is continuous for
all x, t. Differentiating ⇒ u(x, t) is a solution of (40) (provided f 00 exists
for 0 < x < L and f has one–sided derivatives at x = 0, L). Under these
conditions u(x, t) is a solution of (40) satisfying the conditions (41), (42) and
(43) .
Note: compare the above solution with D’Alembert’s solution.
Heat equation
The heat flow in a body of homogeneous material is governed by the heat
equation
∂u κ
= c2 ∇2 u c2 = .
∂t σρ

55
where u(x, y, z, t) is the temperature in the body, κ =thermal conductivity,
σ =specific heat and ρ =density. Here

∂ 2u ∂ 2u ∂ 2u
∇2 u = + +
∂x2 ∂y 2 ∂z 2
Consider the temperature in a long thin wire of constant cross section
(homogeneous material) that is perfectly insulated laterally (so that heat
flows only in the x direction). Then u depends only on x and t. The heat
equation becomes
∂u ∂ 2u
= c2 2 (48)
∂t ∂x
Consider the case when the ends of the bar are kept at zero temperature and
suppose the initial temperature is f (x). This then gives us the conditions
boundary conditions:

u(0, t) = u(L, t) = 0 for t > 0 (49)

initial conditions:
u(x, 0) = f (x) for 0 < x < L (50)
We want to determine the solution of (48) subject to (49) and (50). Use
separation of variables.
Step 1: Separation of variables. Look for a solution in the form

u(x, t) = F (x)G(t) (51)

Substituting into (48) gives

1 dG 1 d2 F
= = −p2 = constant.
c2 G dt F dx2
Thus
dG
+ c2 p 2 G = 0 (52)
dt
d2 F
+ p2 F = 0 (53)
dx2
Step 2: Satisfy the boundary conditions (49).

(49), (51) ⇒ F (0) = F (L) = 0. (54)

Solving (53) subject to (54) gives

F (x) = A cos px + B sin px

F (0) = 0 ⇒ A=0
nπ
F (L) = 0 ⇒ B sin pL = 0 ⇒ p=
L

56
(setting B = 1 we have)
nπx
Fn = sin (n = 1, 2, . . .).
L
Now equation (53) becomes

dG c2 n2 π 2
+ G=0
dt L2
which has the general solution
2
Gn = Bn e−λn t , λn = cnπ/L.

So the functions
nπx −λ2n t
un = Bn sin e .
L
are solutions of (48) satisfying (49).
Step 3: Satisfy the initial conditions (50). As the heat equation is linear
∞
X nπx −λ2n t
u(x, t) = Bn sin e , (55)
n=1 L

is also a solution. But

∞
X nπx
u(x, 0) = f (x) = Bn sin
n=1 L

which is just the half–range expansion for f (x) (i.e. the Fourier sine series).
Hence
L
2Z nπx
Bn = f (x) sin dx (56)
L L
0

So (55) with coefficients (56) is the solution to the heat equation (48) satisfy-
ing (49) and (50) (provided the series converges). Note: The presence of the
2
e−λn t terms in (55) means that all terms in (55) approach zero as t −→ ∞
(i.e. the temperature tends to zero as t −→ ∞).

Steady state two–dimensional heat flow

If we consider a steady flow of heat, the 2D heat equation reduces to
Laplace’s equation
∂ 2u ∂ 2u
∇2 u = + =0 (57)
∂x2 ∂y 2
Solution involves solving (57) in some region R of the xy-plane with a given
boundary condition of the boundary curve of R—a boundary value problem.
Such a problem is called
Dirichlet problem if u is prescribed on C (the boundary curve of R).
∂u
Neumann problem if the normal derivative un = ∂n is prescribed on C.

57
Mixed problem if u is prescribed on a part of C and un on the rest of C.
We will consider only the Dirichlet problem on the rectangle 0 ≤ x ≤
a, 0 ≤ y ≤ b with u = 0 along the boundary except for the edge y = b where
u(x, b) = f (x) for 0 ≤ x ≤ a.
To solve (57) we use separation of variables. Substituting u(x, y) = F (x)G(y)
into (57) gives
1 d2 F 1 d2 G
= − = −k = (constant).
F dx2 G dy 2
Our boundary conditions give
F (0) = 0, F (a) = 0, G(0) = 0
Solving for F gives
nπx
Fn (x) = sin k = (nπ/a)2
a
Solving for G subject to G(0) = 0 gives
nπy
Gn = 2An sinh
a
Finally we need to satisfy the boundary condition
u(x, b) = f (x).
Write ∞
X nπx nπy
u(x, y) = 2 An sin sinh
n=1 a a
Then
∞
X nπx nπb
f (x) = 2 An sin sinh
n=1 a a
∞
X nπx
= Cn sin
n=1 a
where
nπb
Cn = 2An sinh .
a
These are the Fourier coefficients of f (x)
a
nπb 2Z nπx
2An sinh = f (x) sin dx
a a a
0
Thus
Za
1 nπx
An = nπb f (x) sin dx
a sinh a a
0
Therefore the solution of the problem is
∞
X nπx nπy
u(x, y) = 2 An sin sinh
n=1 a a
with the An ’s given above.

58