MATH 18500 - Autumn 2022

Math Methods in the Physical Sciences

Definitions, Theorems, Algorithms and Techniques

Properties of differential equations

A differential equation is an equation that describes a function y(x) in terms of itself and its derivatives
0 = f (x, y(x), y ′ (x), . . . , y (n) (x)),
for some given function f . The order of a differential equation is the highest derivative of the function y used
in the equation. The above equation would be called an nth-order differential equation.

If the unknown function has only a single dependent variable, usually y(x) (for spatial problems) or y(t)
(for temporal problems), it is called an ordinary differential equation (ODE), Alternatively, if the unknown
function has multiple dependent variables, usually u(x, t) (depending on both time and space), it is called a
partial differential equation (PDE).

An ODE is called linear if the differential equation can be written

an (x)y (n) (x) + an−1 (x)y (n−1) (x) + . . . + a1 (x)y ′ (x) + a0 (x)y(x) = q(x),
where q(x) and ai (x) for i = 0, 1, 2, . . . , n are known functions of x. If the functions ai (x) are all constants
(independent of x) then the equation is a constant-coefficient linear ODE. If the functions ai (x) are all
polynomial (and not all constant) then the equation is a polynomial-coefficient linear ODE. If q(x) = 0
then the equation is a homogeneous linear ODE, otherwise it is an non-homogeneous linear ODE.
• While we will see other classifications, these are the common and important ones for the course.
• The classification of differential equations is most often used to describe collections that have useful
properties or can be solved using a common method.
• Homogeneous ODEs have a broader description beyond linear ODEs, however this will not be relevant
in the context of the course.

Separable first-order ODEs

A first-order ODE for y(x) is called separable if it can be rewritten in the form
y ′ = f (y)g(x),
where f and g are known functions of y and x respectively. From this form, the problem can be solved via
ˆ ˆ
1
dy = g(x) dx.
f (y)

• A relationship between y and x without y ′ (x) can only be obtained if both of the above integrals have
analytic solutions.
• Further, solving both integrals does not ensure y(x) can be isolated to obtain an explicit expression for
the solution in term of x.
• The left integral is obtained by integrating both sides with respect to x and then applying the chainrule.
Integrating factor for linear first-order ODEs

Given any linear first-order ODE for y(x) that can be rewritten in the form
y ′ + p(x)y = q(x),
for known arbitrary functions p, q. The solution to this problem is given by
ˆ  ´
1
y(x) = I(x)q(x) dx , where I(x) = e p(x)dx ,
I(x)
and the function I(x) is known as the integrating factor.
• The integrating factor is defined so that when multiplied to the ODE we obtain
d 
I(x)y ′ + p(x)I(x)y = I(x)q(x) =⇒ I(x)y(x) = I(x)q(x),
dx
by undoing the product rule. This occurs when I ′ (x) = p(x)I(x), a separable first-order ODE, which
yields the exponential solution given for I(x) above.
´
• Notably, both solving I ′ (x) = p(x)I(x) and solving p(x) dx give constants of integration. We can
ignore these constants because we only need one of the infinitely many I(x) that allow the product rule
trick described above. By “ignoring” the constants, we are essentially setting both to zero, which results
in the simplest expression.

General solutions, initial conditions and initial value problems

The general solution is a description of all solution curves that satisfy a given ODE, indicated by the
presence of arbitrary constants. Typically, we expect the general solution to an nth order ODE to have exactly
n arbitary constants. A unique solution to a differential equation can then be obtained through the use of
appropriate initial conditions.

A differential equation with an initial condition is referred to as an initial value problem (IVP). We typically
write the IVP for a first-order ODE in the form
y ′ = f (t, y(t)), y(t0 ) = y0 ,
where t0 , y0 ∈ R.

For an nth-order differential equation, we typically require n initial conditions in order to find the unique
solution to the differential equation. The IVP is written in the form
(n−1)
y (n) = f (t, y, y ′ , . . . , y (n−1) ), y(t0 ) = y0 , y ′ (t0 ) = y0′ , . . . , y (n−1) (t0 ) = y0 ,
(n)
where t0 , y0 , y0′ , . . . , y0 ∈ R.
• Given the nth-order IVP above, the problem will give exactly one solution so long as the function f and
all its partial derivatives are continuous around the initial conditions.
• Initial conditions and initial value problems correspond to physical contexts in which time t is the
independent variable of the unknown function. We may also consider boundary conditions and boundary
value problem, in which the position x is the independent variable.
Visualization of solutions to first-order differential equation

Consider a first-order differential equation of the form

y ′ = f (t, y(t)).
Since y ′ (t) is the slope of the tangent line to the curve y at time t, we can draw small line segments to indicate
the slope at any given point.

A slope field is the visualization of a regular grid of the small slope segments on a ty-plane. Starting
from some initial condition (t, y) = (t0 , y0 ) and following the direction locally indicated by the slope field, an
approximate solution curve can be constructed.
• While the construction of slope fields is visually beneficial and gives an observer some intuition towards
the set of solutions to an ODE, it is not far from the ideas surrounding numerical computation of ODEs.
• Autonomous ODEs are those that have no explicit time dependence. For first-order ODEs, these take
the form y ′ = f (y(t)). The slope field for a first-order autonomous ODE can be used to easily predict
the long-term behaviour of the system, as for a given y-value, the slope will be the same for all time.

Power series solutions

Consider an nth-order homogeneous linear ODE with polynomial coefficients pi (x) for i = 0, 1, . . . , n
pn (x)y (n) + pn−1 (x)y (n−1) + . . . + p1 x + p0 = 0.
We may be able to find a solution to this ODE by guessing a power series solution, that is
∞
X
y(x) = cm xm = c0 + c1 x + c2 x2 + . . . ,
m=0

where each ci is an unknown coefficient. To find the value of the coefficients, we first calculate the required
derivatives
∞
X ∞
X
y ′ (x) = cm mxm−1 , y ′′ (x) = cm m(m − 1)xm−2 , . . . ,
m=0 m=0

and then substitute them into the differential equation. By then finding expressions for the coefficient of each
exponent of x, we can construct a recurrence relation, determining a relationship between the coefficients.
• The power series presented above is centered at x = 0 and hence will fail if the function is not defined
at x = 0. Similarly, if the function is discontinuous at some non-zero value, the power series will only
converge to solution on the continuous interval.
∞
• We can also center the power series at any value x = x0 by guessing the solution y = cm (x − x0 )m .
P
m=0

• Finding the recurrence relation can be difficult as isolating each exponent of x formally requires manipu-
lations of summation terms. This can be informally bypassed by constructing a table of each term versus
their x exponent and observing the general relationship.
• The recurrence relation allows you to write some coefficients in terms of others. For a first-order ODE,
we should expect to write all non-zero coefficients in terms of c0 , for a second-order ODE, we should
expect to write all non-zero coefficients in terms of c0 and c1 , etc.
Taylor series solutions

Consider a general IVP for some first-order ODE

y ′ = f (t, y(t)), y(t0 ) = y0 ,
where f and its partial derivatives are continuous around the initial condition. The initial value of y ′ (t) can
be determined by substituting t = t0 into the equation
y ′ (t0 ) = f (t0 , y0 ).
Further, we can calculate the value of y ′′ (t0 ) by first taking the derivative of the ODE with respect to t and
then substituting t = t0
∂f ∂f dy
y ′′ = + =⇒ y ′′ (t0 ) = ft (t0 , y0 ) + fy (t0 , y0 )y ′ (t0 ) = ft (t0 , y0 ) + fy (t0 , y0 )f (t0 , y0 ).
∂t ∂y dt
We can repeat this process indefinitely (for sufficiently well-behaved f ), obtaining the value of all derivatives
of y at t = t0 . With this information, we can obtain the Taylor series solution for y(t) centered at t = t0
∞
X y (n) (t0 )
y(t) = (t − t0 )n .
n!
n=0
When considering the IVP for some nth-order ODE we will obtain a similar result with
(n−1)
y (n) = f (t, y, y ′ , . . . , y (n−1) ), y(t0 ) = y0 , y ′ (t0 ) = y0′ , . . . , y (n−1) (t0 ) = y0 ,
allowing for the calculation of y (n) (t0 ) via substitution into the ODE and the calculation of y (n+1) (t0 ) being
possible if we can calculate the partial derivatives of f (t, y, y ′ , . . . , y (n−1) ).
• Notably, the above results demonstrate that, for a sufficiently well behaved f , the IVP corresponding to
an nth-order ODE requires exactly n initial conditions.
• In general, it is uncommon to find a general formula for y (n) (t0 ) using the above methods, particularly
with higher order ODEs. In practice, we can instead calculate the first m derivatives and construct an
mth-order Taylor approximation, finding an approximate solution to the ODE close to t = t0 .

Existence and uniqueness of solutions to IVPs

The Taylor series solutions allow us to make some claims about the existence and uniqueness of solutions
to IVPs, specifically that for an nth order IVP
(n−1)
y (n) = f (t, y, y ′ , . . . , y (n−1) ), y(t0 ) = y0 , y ′ (t0 ) = y0′ , . . . , y (n−1) (t0 ) = y0 ,
we can guarantee a unique solution so long as the function f and all of its partial derivatives are continuous.

Existence is demonstrated as under these conditions we are ensured to be able to create the Taylor series
(at least locally to the initial conditions). Uniqueness is guaranteed because the Taylor series of a function is
unique (that is, if two functions have the same Taylor series, they must be the same function).
• The demonstration of existence and uniqueness implies that, for a well-behaved function f , an nth-order
ODE should have a general solution with exactly n arbitrary constants. So long as we apply the n initial
conditions, we will always obtain the same unique solution to the IVP.
• Notably, the above conditions on f ensure the uniqueness and existence of the solution to the IVP,
however they are not necessarily required.
Linear constant-coefficient homogeneous ODEs and the auxiliary equation

The following differential equation is a nth-order linear constant-coefficient homogeneous ODE

an y (n) (t) + an−1 y (n−1) (t) + . . . + a2 y ′′ (t) + a1 y ′ (t) + a0 y(t) = 0,
where each ai is a known real constants (i = 0, 1, . . . , n) and an ̸= 0. The auxiliary equation corresponding to
this ODE is an nth-order polynomial, in which the ith derivative of y in the ODE is replaced with λi . That is
an λn + an−1 λn−1 + . . . + a2 λ2 + a1 λ + a0 = 0.
Finding the roots of the auxiliary equation will allow us to find the general solution to the ODE.
• The auxiliary equation is derived by assuming solutions of the form y(t) = eλt . This assumption is called
an exponential ansatz, and while not all solutions will take exactly this form, we will be able to use the
information from this guess to deduce all solutions.
• The fundamental theorem of algebra tells us that the auxiliary equation, an nth-order polynomial, will
always be able to be factored into exactly n terms
(λ − λ1 )(λ − λ2 ) . . . (λ − λn ) = 0,
where λi is some complex number. Although we cannot ensure that each root will be unique, we will
see that each of the n roots will be associated with the n independent solutions used to form the general
solution.

Superposition of solutions for linear homogeneous ODEs

Consider the following nth-order linear homogeneous ODE

an (t)y (n) (t) + an−1 (t)y (n−1) (t) + . . . + a2 (t)y ′′ (t) + a1 (t)y ′ (t) + a0 (t)y(t) = 0,
where each ai (t) is a known function of t (i = 0, 1, . . . , n). Then, for any two known solution to the ODE, say
y(t) = y1 (t) and y(t) = y2 (t), we can guarantee that any superposition of multiples of these solutions is also a
solution, namely that
y(t) = C1 y1 (t) + C2 y2 (t),
is a solution to the ODE for any constants C1 , C2 .

Using the superposition property, along with the uniqueness and existence theorem, we can say that, so
long as an (t) ̸= 0, the general solution to the nth-order linear homogeneous ODE takes the form
y(t) = C1 y1 (t) + C2 y2 (t) + . . . + Cn yn (t).

• The full superposition is a combination of the additive and multiplicative linear property. The multiplica-
tive property says that for any known solution y(t) = y1 (t), we know that y(t) = C1 y1 (t) is a solution.
The additive property says that for any known solutions y(t) = y1 (t) and y(t) = y2 (t), we know that
y(t) = y1 (t) + y2 (t) is a solution.
• The linear property of these differential equations is analogous to the linear property of solutions to
homogeneous systems of equations, A⃗x = ⃗0. In Math 183, we showed that if ⃗x = ⃗x1 and ⃗x = ⃗x2 are
solutions to A⃗x = ⃗0, then ⃗x = a1 ⃗x1 + a2 ⃗x2 is also a solution to the system. This idea eventually leads to
the construction of a one-to-one correspondence between solutions to systems of equations and solutions
to linear differential equations.
Real solutions to the auxiliary equation

Consider a second-order linear constant-coefficient homogeneous ODE

a2 y ′′ (t) + a1 y ′ (t) + a0 y(t) = 0

where a2 , a1 , a0 are real constants with a2 ̸= 0. If the auxiliary equation has unique real roots λ = λ1 , λ2 then
the general solution to the ODE is

y(t) = C1 eλ1 t + C2 eλ2 t .

In general, if an nth-order linear constant-coefficient homogeneous ODE has a real root λ, then we can say
that the ODE has a solution

y1 (t) = eλt .

Using the superposition property, we can include this in our general solution

y(t) = C1 eλt + C2 y2 (t) + . . . + Cn yn (t).

• As described above, each unique, real λ found as a solution to the auxiliary equation can be included
in the general solution via superposition. In the second-order case, we only find two solution via the
auxiliary equation, namely y1 (t) = eλ1 t and y2 (t) = eλ2 t , but can immediately form the general solution
via the superposition of these solutions.

Complex solutions to the auxiliary equation

Consider a second-order linear constant-coefficient homogeneous ODE

a2 y ′′ (t) + a1 y ′ (t) + a0 y(t) = 0

where a2 , a1 , a0 are real constants with a2 ̸= 0. If the auxiliary equation has complex roots λ = α ± βi then
the general solution to the ODE is
 
y(t) = eαt C1 cos(βt) + C2 sin(βt) .

In general, if an nth-order linear constant-coefficient homogeneous ODE has a set of complex conjugate roots
λ = α + βi, then we can say that the ODE has solutions

y1 (t) = eαt cos(βt), and y2 (t) = eαt sin(βt).

Using the superposition property, we can include these in our general solution

y(t) = C1 eαt cos(βt) + C2 eαt sin(βt) + . . . + Cn yn (t).

• To demonstrate this solution form from our guess of y(t) = eλt , we apply Euler’s identity to e(α+βi)t , in
which the complex components become sinusoidal functions.

• As the ODE has real constant-coefficients, we know the auxiliary equation will also have real coefficients.
This ensures that complex roots always come in complex conjugate pairs, allowing our simplified result.
Reduction of order

Consider the nth-order linear homogeneous ODE

an (t)y (n) (t) + an−1 (t)y (n−1) (t) + . . . + a2 (t)y ′′ (t) + a1 (t)y ′ (t) + a0 (t)y(t) = 0,
where each ai (t) is a known function of t (i = 0, 1, . . . , n). Then if y(t) = y1 (t) is some known solution to the
ODE, we can apply the reduction of order technique by guessing a solution of the form
y(t) = v(t)y1 (t),
allowing us to write a differential equation in terms of v(t). After substituting into the ODE and simplifying,
this guess will always remove all of the v(t) terms from the original ODE. The reduction of order step is then
achieved by defining u(t) = v ′ (t), allowing us to obtain an (n − 1)th-order linear homogeneous ODE
bn−1 (t)u(n−1) (t) + bn−2 (t)u(n−2) + . . . + b2 (t)u′′ (t) + b1 (t)u′ (t) + b0 (t)u(t) = 0,
where each bi (t) (i = 0, 1, . . . , n − 1) is a function of t based on the original coefficient functions aj (t) and the
known solution y1 (t).
• When we guess the form y(t) = v(t)y1 (t), we know that the mth derivative will take the form y (m) (t) =
v (m) (t)y1 (t) + . . . + v(t)y (m) (t). The final term on each expression will have a common factor of v(t), and
substituting into the ODE, removing v(t) from all terms in which it is a factor will reveal the differential
equation in terms of y1 (t). Since y1 (t) is a known solution, these terms will be zero, removing the v(t).
• While reduction of order does not ensure the new ODE will be solvable, it does ensure that it will be a
lower order than your original ODE, which is typically desirable.
• There is the potential to reduce the order of the ODE by more than one. The total possible reduction
depends on the exact structure of the known solution y1 (t) compared to the general solution. This can
be seen in the case where a linear constant-coefficient ODE has repeated roots in the auxiliary equation.

Repeated roots in the auxiliary equation

Consider a second-order linear constant-coefficient homogeneous ODE

a2 y ′′ (t) + a1 y ′ (t) + a0 y(t) = 0
where a2 , a1 , a0 are real constants with a2 ̸= 0. If the auxiliary equation takes the form (λ − λ1 )2 = 0, we say
that λ = λ1 is a repeated root and the general solution to the ODE is
y(t) = C1 eλ1 t + C2 teλ1 t .
In general, if an nth-order linear constant-coefficient homogeneous ODE has a root repeated k times, then it
will correspond to k terms in the solution, namely
y1 (t) = eλt , y2 (t) = teλt , . . . , yk = tk−1 eλt .
Using the superposition property, we can include these in our general solution
y(t) = C1 eλt + C2 teλt + . . . + Ck tk−1 eλt + Ck+1 yk+1 (t) + . . . + Cn yn (t).

• For higher order ODEs, we can also have repeated complex conjugate roots. These cases will take the
same form as complex solutions, with each repeated root giving a solution that is multiplied by t.
• Using the exponential ansatz, a repeated root technically only gives one solution, however we can demon-
strate the appearance of the extra solutions that include the polynomial terms using reduction of order.
Operators and linear operators

An operator O maps an element of a set onto another set (often the same set). A linear operator is an
operator that also satisfies the linear property, that is, if an operator O can be applied to elements of a set X
and f, g ∈ X then
O(k1 f + k2 g) = k1 O(f ) + k2 O(g),
where k1 , k2 are constants.
• We have already used many examples of linear operators, for instance, matrices when applied to vectors
or derivatives when applied to functions.

Addition and multiplication of linear operators

Consider a the linear operators O1 , O2 , that act on the same set. We can construct a new operator by
adding the two operators together, which results in the following when applied to a function y
(O1 + O2 )y = O1 y + O2 y.
Similarly, we can multiply operators together. Notably, the product of operators is resolved from right to left,
that is
(O2 O1 )y = O2 (O1 y).
In general operators do not commute multiplicatively, that is
(O1 O2 )y ̸= (O2 O1 )y.
When an operator is applied twice in a row, we can also use the notation
O1 (O1 y) = O12 y.

Repeated integration for second-order linear constant-coefficient homogeneous ODEs

Consider a second-order linear constant-coefficient homogeneous ODE

y ′′ (t) + a1 y ′ (t) + a0 y(t) = 0
where a1 , a0 are real constants. Using operators D, the differential operator (Dy = y ′ ), and I, the identity
operator (Iy = y), and using the multiplication of linear operators to factor, we can then rewrite the ODE as
(a2 D2 + a1 D + a0 I)y = (D − λ1 I)(D − λ2 I)y = 0,
where λ1 + λ2 = −a1 and λ1 λ2 = a0 . The solution to the ODE can then be found by solving the following pair
of first-order linear ODEs
(D − λ2 I)y(t) = u(t), and (D − λ1 I)u(t) = 0.

• This approach is possible because the operators corresponding to constant coefficient problems have
easily factorable operators. Attempting to factor ODEs with non-constant coefficients is more difficult.
• This repeated integration technique can be implemented without the need to guess the solution as we
did with the exponential ansatz. Despite this, using the exponential ansatz is preferred in practice as it
arrives at the solution far more efficiently.
Euler-Cauchy equations

The following nth-order linear homogeneous ODE is known as an Euler-Cauchy equation

an xn y (n) (x) + an−1 xn−1 y (n−1) (x) + . . . + a2 x2 y ′′ (x) + a1 xy ′ (x) + a0 y(x) = 0,

where each ai (i = 0, 1, 2, . . .) is a constant. The polynomial ansatz y(x) = xm yields the “auxiliary equation”
of the ODE, taking the form

xm an m(m − 1) . . . (m − (n + 1)) + an−1 m(m − 1) . . . (m − n) + . . . + a2 m(m − 1) + a1 m + a0 = 0,

 

in which the xm ̸= 0 implies the remaining polynomial must be zero for some choice(s) of m.

For each real unique m = m1 , m2 , . . . , mr we obtain a solution term of the form

y1 (x) = xm1 , y2 (x) = xm2 , . . . , yr (x) = xmr ,

for complex values m = α + βi we obtain a pair of solution terms of the form

y1 (x) = xα cos(β ln(x)), y2 (x) = xα sin(β ln(x)),

and for any repeated roots of m, we simply take our known solution term and multiply by ln(x). That is, if m
is repeated k times we have the k solutions

y1 (t) = xm , y2 (x) = ln(x)xm , . . . , yk (x) = (ln(x))k xm .

• The ansatz of y(x) = xm can be predicted based on the structure of the differential equation. Notably,
every term is of the form xi y (i) . If we guess xm , an mth order polynomial term, then taking the derivative
i times reduces the order to m − i, but then multiplying by the ith order coefficient returns the term to
order m. This means all terms will have a common factor xm after substitution. As seen in the above
equation, this leaves us with an mth order polynomial.

• The repeated solutions can be demonstrated using reduction of order on our known solution y1 (x) = xm .

• Observing these solutions, you might notice that the solution form is similar to the solutions just found for
constant-coefficient ODEs. In fact, if we were to define t = ln(x), you can see that we obtain exactly the
same solutions upon substitution. This is because the application of this definition t = ln(x), along with
dn y 1 dn y
the relevant derivatives dxn = xn dtn , allows us to transform between these two differential equations.

Proving the relevant derivative rule for this transformation can be done using proof by induction.
Systems of first-order ODEs

The following is a system of first-order ODEs

 ′   
x1 (t) f1 (t, ⃗x(t))
 x′ (t)   f2 (t, ⃗x(t)) 
 2
⃗x ′ (t) =  .  = 
 ⃗
 = f (t, ⃗x(t)),
 
..
 ..   . 
x′n (t) fn (t, ⃗x(t))

where the functions fi (t, ⃗x(t)) describe the change in each function xi (t) in terms of time and the other
components of ⃗x(t).

• Systems of ODEs describe the dynamics of coexisting populations, in particular, how each population
changes based on the quantity of each other population.

Autonomous systems of ODEs and equilibrium solutions

A system of first-order ODEs is called autonomous it takes the form

⃗x ′ (t) = f⃗(⃗x(t)),

that is, if the function f⃗ has no explicit time dependence. This means that the change in ⃗x(t) depends only
on the current state of the system.

An autonomous system is said to be at an equilibrium ⃗x(t) = ⃗x∗ if ⃗x ′ (t) = ⃗0, that is, the system is no
longer undergoing any change. The equilibria of a system can be determined by solving the system f⃗(⃗x∗ ) = ⃗0.

Systems of linear constant-coefficient homogeneous ODEs

A system of linear constant-coefficient homogeneous ODEs takes the form

⃗x ′ (t) = A⃗x(t),

where A is a constant matrix, and the entry aij in row i column j describes the rate that variable xi changes
with respect to the size of the population xj .

The two-variable system can be written explicitly as

x′1 (t) = a11 x1 (t) + a12 x2 (t)

x′2 (t) = a21 x1 (t) + a22 x2 (t),

where aij are constants for i, j = 1, 2.

• Notably, due to the system being constant-coefficient and homogeneous, it is also autonomous.

• The solution vector ⃗x(t) = ⃗0 is always an equilibrium solution of the system ⃗x ′ (t) = A⃗x(t). If the matrix
A is invertible, then ⃗x(t) = ⃗0 will be the only equilibrium solution of this system.
Converting between systems of ODEs and nth-order ODEs

We can convert between an arbitrary two-variable system of ODEs and a second-order linear constant-
coefficient ODE with the following relationship

x′1 (t) = a11 x1 (t)+ a12 x2 (t)

y ′′ (t) − (a11 + a22 )y ′ (t) + (a11 a22 − a12 a21 )y(t) = 0 ⇐⇒ ,
x′2 (t) = a21 x1 (t)+ a22 x2 (t)

where y(t) can be defined as either variable x1 (t) or x2 (t). Systems with more variables can also be converted
to higher-order ODEs, however the relationship quickly becomes more complicated.

Given an nth-order ODE in terms of y(t), we can define variables xi (t) = y (i−1) (t) for i = 1, 2, . . . , n. This
allows us to write the system

x′1 (t) = x2 (t)

x′2 (t) = x3 (t)
y (n) (t) + an−1 y (n−1) + . . . + a1 y ′ (t) + a0 y(t) = 0 ⇐⇒ .. ,
.
x′n−2 (t) = xn−1 (t)
x′n−1 (t) = y (n) (t)

where the last equation can be found by rearranging the ODE: x′n−1 (t) = −an−1 y (n−1) − . . . − a1 y ′ (t) − a0 y(t).

• The above relationships allow us to see that there is a relationship between the system of ODEs and the
nth-order ODE. This motivates our use of an exponential ansatz for solutions to the systems, guessing
the form ⃗x(t) = ⃗v eλt for some unknown λ and constant vector ⃗v .

• This relationship also indicates that for an n-variable system we should expect the solution to take the
form of a superposition of n independent solution terms.

Exponential ansatz yields eigenvalues and eigenvectors for system solutions

The system of ODEs of the form ⃗x ′ (t) = A⃗x(t) can be solved using an exponential ansatz. If we assume
the solution form

⃗x(t) = ⃗v eλt ,

where ⃗v is a constant vector then substitution into the ODE yields

⃗x′ (t) = A⃗x(t) =⇒ λ⃗v eλt = A⃗v eλt =⇒ A⃗v = λ⃗v ,

revealing that if λ is an eigenvalue of A with corresponding eigenvector ⃗v then the above ansatz is a solution
to the system of ODEs.

• The characteristic polynomial for the matrix A, namely det(A−λI) = 0, generally takes the same form as
the corresponding auxiliary equation for the nth-order linear constant-coefficient ODE. This is necessary
because we expect consistent solutions forms due to their relationship.
Solutions to linear systems of ODEs with real eigenvalues

Consider the two-variable system of linear ODEs of the form ⃗x ′ (t) = A⃗x(t) for some constant matrix A. If
the matrix A has two real unique eigenvalues λ1 and λ2 with corresponding eigenvectors ⃗v1 and ⃗v2 then the
general solution to the system of ODEs is

⃗x(t) = c1⃗v1 eλ1 t + c2⃗v2 eλ2 t .

In general, for an n-variable system, if A has an real eigenvalue λ with algebraic multiplicity one, then λ will
have a corresponding eigenvector ⃗v and hence a solution

⃗x(t) = ⃗v eλt .

Using superposition, this terms can be written as part of our general solution

⃗x(t) = c1⃗v eλt + c2 ⃗x2 (t) + . . . + cn ⃗xn (t).

• If A has an eigenvalue λ with equal algebraic and geometric multiplicities m > 1, then any set of linearly
independent eigenvectors ⃗v1 , ⃗v2 , . . . , ⃗vm that span the eigenspace form m independent solution terms

⃗x1 (t) = v⃗1 eλt , ⃗x2 (t) = ⃗v2 eλt , . . . , ⃗xm (t) = ⃗vm eλt .

Solutions to linear systems of ODEs with complex eigenvalues

Consider the two-variable system of linear ODEs of the form ⃗x ′ (t) = A⃗x(t) for some constant matrix
A. If the matrix A has complex eigenvalues λ = α ± βi (where α, β ∈ R) with corresponding eigenvectors
⃗1 ± w
⃗v = w ⃗ 2 i (where w ⃗ 2 ∈ R2 ) then the general solution to the system of ODEs is
⃗ 1, w
h
i
⃗x(t) = eαt c1 w


⃗ 1 cos(βt) − w
⃗ 2 sin(βt) + c2 w ⃗ 1 sin(βt) .
⃗ 2 cos(βt) + w

In general, for an n-variable system, if A has a complex conjugate pair of eigenvalues λ = α ± βi with algebraic
⃗1 ± w
multiplicity one, then λ will have a corresponding eigenvector ⃗v = w ⃗ 2 i and hence a solution of the form
given above.
• For a real matrix A, the complex eigenvalues will always come in complex conjugate form and result in
complex conjugate eigenvectors. As such, we only need to calculate one eigenvector to find w
⃗ 1 and w
⃗ 2,
and hence the solution terms corresponding to these eigenvalues.

• When using the above solution form, make sure that λ = α + βi corresponds to ⃗v = w ⃗1 + w⃗ 2 i when
defining w
⃗ 1 and w
⃗ 2 . The solution form will not work if you pair the positive eigenvalue with the negative
eigenvector.

• Writing the solution corresponding to a complex eigenvalue/eigenvector and applying Euler’s identity
h i
⃗x(t) = (w ⃗ 2 i)e(α+βi)t = eαt w
⃗1 + w ⃗ 1 cos(βt) + iw ⃗ 2 cos(βt) − w
⃗ 1 sin(βt) + iw ⃗ 2 sin(βt) ,

we can obtain two independent solutions to the system via Re[⃗x(t)] and Im[⃗x(t)]. The complex conjugate
eigenvalue/eigenvector would yield the same solutions.
Solutions to linear systems of ODEs where A is diagonalizable

Consider the n-variable system of linear ODEs of the form ⃗x ′ (t) = A⃗x(t) for some constant n × n matrix
A. If A is diagonalizable with diagonal matrix D, containing the eigenvalues of A (λ1 , λ2 , . . . , λn ), and change
of basis matrix S, containing the corresponding eigenvectors (⃗v1 , ⃗v2 , . . . , ⃗vn ), then we can write
⃗x ′ (t) = SDS −1 ⃗x(t) ⇐⇒ S −1 ⃗x ′ (t) = DS −1 ⃗x(t) ⇐⇒ X
⃗ ′ (t) = DX(t),
⃗
⃗
where X(t) = S −1 ⃗x(t) is a transformation ⃗
to a new set of variables X(t). ⃗ ′ (t) =
The diagonalized system X
⃗
DX(t) can then be written
 ′      
X1 (t) λ1 0 ... 0 X1 (t) λ1 X1 (t)
 X ′ (t)   0 λ2 . . . 0  X2 (t)   λ2 X2 (t) 
 2
 =  .. = ,
     
.. .. . . . .. ..
. ..
 
 .   . .  .   . 
Xn′ (t) 0 0 . . . λn Xn (t) λn Xn (t)
in which the change in each transformed variable depends only on itself. This is called a decoupled system
and allows us to simply solve each ODE individually, resulting in the solution
Xi (t) = ci eλi t ,
for i = 1, 2, . . . n, where each ci is an unknown constant. We can then transform back to our original variables
using the transformation
 c1 eλ1 t
 

| | |  c2 eλ2 t 
⃗
⃗x(t) = S X(t) = ⃗v1 ⃗v2 . . . ⃗vn 
  
..  = c1⃗v1 eλ1 t + c2⃗v2 eλ2 t + . . . + cn⃗vn eλn t ,

| | |
 . 
cn e λ n t

obtaining the general solution to the ODE.

• The above procedure only works if A is diagonalizable. Recall that a matrix A will be diagonalizable if
the algebraic multiplicity and geometric multiplicity are equal for all eigenvalues. A notable case of this
is when all eigenvalues are unique, which immediately ensures the condition.

Solutions to linear systems of ODEs with unequal algebraic and geometric multiplicities

Consider the two-variable system of linear ODEs of the form ⃗x ′ (t) = A⃗x(t) for some constant matrix A. If
the matrix A has a single real eigenvalue λ with algebraic multiplicity two and geometric multiplicity one, then
the eigenvector ⃗v1 can be used to obtain the generalized eigenvector ⃗v2 satisfying the property (A − λI)⃗v2 = ⃗v1 .
The general solution to the system of ODEs is given by
⃗x(t) = c1⃗v1 eλt + c2 t⃗v1 + ⃗v2 eλt .

In general, for an n-variable system, if A has an real eigenvalue λ with algebraic multiplicity m, then we will
be able to find solutions using the ansatz
⃗x(t) = tm−1 ⃗u1 + tm−2 ⃗u2 + . . . + t⃗um−1 + ⃗um eλt ,

where each ⃗ui is a constant vector.

• We will find m different possible choices of vectors ⃗ui that satisfy the above form to obtain m
independent
solutions. For instance, in the two-variable problem, the solution form is ⃗x(t) = t⃗u1 + ⃗u2 eλt . The first
solution is given by ⃗u1 = ⃗0, ⃗u2 = ⃗v1 , and the second solution is given by ⃗u1 = ⃗v1 , ⃗u2 = ⃗v2 .
Visualization of solutions for systems of first-order ODEs

Consider the autonomous system of ODEs, ⃗x ′ (t) = f⃗(⃗x(t)). For any fixed location ⃗x(t) = ⃗c, the output
⃗x = f⃗(⃗c) is a constant vector indicating the direction of travel of a solution curve at that point.
′ (t)

We can construct the vector field corresponding to the ODE by calculating ⃗x ′ (t) everywhere on a grid.
For a two-variable problem (and perhaps a three-variable problem) this allows us to visualize the vector field.

A phase plane for a two- or three-variable autonomous system of ODEs is the visualization of a collection
of solution curves that indicate the dynamics of the system. This means that the set of chosen curves should
be sufficient to predict the solution curves everywhere in the system (or at least local to points/regions of
interest).
• To construct the phase plane, the system of ODEs must be autonomous. For an autonomous system, the
vector field will never change, meaning that if any two initial conditions pass through the point ⃗x(t) = ⃗c
that they will then follow the exact same trajectory.
• For a non-autonomous system, the vector field will change over time. This makes the visualization and
overall dynamics of the system much less predictable.

Equilibrium classification for two-variable linear systems of first-order ODEs

Consider a two-variable linear system of first-order ODEs, ⃗x ′ (t) = A⃗x(t). The system will always have an
equilibrium at ⃗x(t) = ⃗0. Let λ1 , λ2 be the eigenvalues of A:

If the eigenvalues are real and non-zero then the equilibrium ⃗x(t) = ⃗0 is classified as follows:
(1) Source : λ1 > λ2 > 0, (2) Sink : λ1 < λ2 < 0, (3) Saddle : λ1 > 0 > λ2 .

If the eigenvalues are complex, λ = α ± βi then the equilibrium ⃗x(t) = ⃗0 is classified as follows:
(4) Spiral source : α > 0, (5) Spiral sink : α < 0, (6) Center : α = 0.

If there is one eigenvalue λ ̸= 0 with algebraic multiplicity two then the equilibrium is classified as follows:
(7) Degenerate source : λ > 0, (8) Degenerate sink : λ < 0.

If one of the eigenvalues is λ2 = 0 then the line spanned by the eigenvector ⃗v2 are all equilibria, and can be
classified as follows:
(9) Line of sources : λ1 > 0, (10) Line of sinks : λ1 < 0, (11) Uniform motion : α = 0.

• While the exact solutions of a given classification will change based on the value of the eigenvalues and
eigenvectors, the overall dynamics are essentially the same.
• Below we consider the phase planes for cases (1) - (6), as these are the most common behaviours.
• The degenerate cases are boundary cases between the real and complex eigenvalue cases, with the single
eigenvalue acting analogous to the real eigenvalue cases, with the other dynamics being rotational as in
the complex cases.
• Case (11) is the boundary of cases (9) and (10) in which the vector field is neither attracted or repelled
from the line of equilibria. Instead, motion occurs parallel to this line.
Phase planes for two-variable linear systems of first-order ODEs, real unique eigenvalues

For a two-variable linear system of first-order ODEs, ⃗x ′ (t) = A⃗x(t), if A has real, unique and non-zero
eigenvectors then the phase plane of the system will have one of the following structures

Source Sink Saddle

The solution curves can be predicted based on the eigenvalues and eigenvectors of A. Solutions parallel to the
eigenvectors ⃗v1 , ⃗v2 are indicated by the red lines. Eigenvectors corresponding to positive eigenvalues describing
movement away from the equilibrium and negative eigenvalue describing movement towards the equilibrium.
• For the source and sink cases, the eigenvector with the smaller magnitude eigenvalue will have a greater
influence on behaviour close to the origin, while the eigenvector with the larger magnitude eigenvalue
will have a greater influence far from the origin.
• The eigenvectors correspond to directions where the vector field is moving directly towards or away from
the equilibrium. Eigenvectors divide the space into regions of clockwise and counterclockwise rotation.

Phase planes for two-variable linear systems of first-order ODEs, complex eigenvalues

For a two-variable linear system of first-order ODEs, ⃗x ′ (t) = A⃗x(t), if A has complex eigenvectors then the
phase plane of the system will have one of the following structures

w
⃗2 w
⃗2 w
⃗2
w
⃗1 w
⃗1 w
⃗1

Spiral source Spiral sink Center

The solution curves are based on λ = α + βi and the eigenvectors ⃗v = w ⃗1 + w

⃗ 2 i. The elliptic shape of the
trajectory is given by w
⃗ 1 and w
⃗ 2 and the direction of the rotation can be determined by the short angle from
w
⃗ 2 toward w⃗ 1.
• For the spiral source and sink cases, the radius increase after a full rotation increases as α increases or
as β decreases.
• For complex eigenvalues, the rotation will always be in the same direction as there are no directions in
which the rotation can change direction (no real eigenvectors).
Equilibrium classification for autonomous non-linear systems of ODEs

Consider the two-variable system of autonomous non-linear first-order ODEs, ⃗x ′ (t) = f⃗(⃗x(t)). For any
equilibrium of the system, that is, any ⃗x(t) = ⃗x0 such that f⃗(⃗x0 ) = ⃗0, the system can be locally approximated
close to the equilibrium by the linear system of ODEs
" ∂f1 ∂f1 #
∂x1 ∂x2
⃗x ′ (t) ≈ J(⃗x0 ) (⃗x(t) − ⃗x0 ) where J(⃗x) = ∂f2 ∂f2
.
∂x1 ∂x2

The matrix J is the Jacobian matrix for the vector function f⃗(⃗x). By calculating the eigenvalues and eigen-
vectors of J(⃗x0 ), we can classify the equilibrium.

For a two-variable autonomous non-linear first-order system of ODEs, we can apply this classification to
all equilibria of the system. Using the local dynamics of each equilibria, we are able to begin visualizing these
sections of the phase plane. Extrapolating the dynamics between each equilibria based on its classifications
allows us to vaguely predict the overall dynamics of a non-linear system.

Partial differential equations

A partial differential equations is a differential equation describes a multivariable function say u(x, y) in
terms of itself and its derivatives. The most general form of a second-order linear homogeneous PDE is
a2,0 (x, y)uxx + a1,1 (x, y)uxy + a0,2 (x, y)uyy + a1,0 (x, y)ux + a0,1 (x, y)uy + a0,0 (x, y)u = 0,
where each ai,j is a known function of x and y.
• Even a very simple partial differential equations can be relatively difficult to solve. To have a chance of
obtaining a solution with the methods we will learn, we require even more idealized forms of the above,
such as no mixed partial terms (a1,1 = 0), or constant-coefficients.

Intuition of PDE terms based on finite differences

Most PDE problems have the context in which variables are either temporal variables (typically t) or spatial
variables (typically x and y). Variables such as ut and utt refer to velocity and acceleration respectively,
however terms such as ux and uxx are sometimes less clear. Using Taylor series, we can write the following
approximations for first and second derivatives
f (x + ∆x) − f (x) f (x + ∆x) − 2f (x) + f (x − ∆x)
f ′ (x) ≈ , f ′′ (x) ≈ ,
∆x ∆x
where smaller values of ∆x increase the accuracy of the approximation.

Using these approximations, we can generally consider the meaning u, ux and uxx terms. The term u(x, t)
refers only to the current position, while both ux and uxx considers the value of u local to x. Notably, the magni-
tude of ux increases when u is not constant near x, giving positive/negative values if u is increasing/decreasing
(respectively) as x is increasing. Similarly, the magnitude of uxx will increase if u is not linear near x, giving
positive/negative values based on whether u close to x is on average above/below u(x, t). Notably, this means
that uxx is often used when there is no directional bias, while ux appears to indicate directional bias based on
the sign of its coefficient.
Heat equation PDE

The heat equation is a PDE for a function u(x, t), taking the form
ut = k 2 uxx ,
for some real constant k. Without context, the PDE describes that the rate of change in u (that is, ut ) will
be largest when u is locally non-linear.

The heat equation is most often used to describe heat spreading through an object. The above equation is
the one-dimensional heat equation, in which we consider, say a metal rod, and assume that the cross-section
at a given x value will have constant temperature u(x, t) at time t, but the temperature can vary between
cross sections. The coefficient k is the heat transfer coefficient, describing how quickly heat diffuses through
the object. We can imagine that if u(x, t) is locally linear, then the temperature gain from one side essentially
cancels the temperature loss from the other side.

Wave equation PDE

The wave equation is a PDE for a function u(x, t), taking the form
utt = c2 uxx ,
for some real constant c. Without context, the PDE describes that the acceleration of u (that is, utt ) will be
largest when u is locally non-linear.

The wave equation describes the dynamics of a one-dimensional wave with u(x, t) representing the dis-
placement from equilibrium at a location x at time t. The coefficient c is the wave speed. In the context of
longitudinal waves, we can consider the wave as a collection of point masses attached by springs, result in
acceleration if adjacent points become too close or too far. Local linearity indicates that, for a given point
mass, the adjacent point masses will be applying an equal and opposite force, resulting in no net force. In
the context of a transverse wave, u(x, t) is the vertical displacement, and vertical movement is based on the
vertical component of local tension forces. Local linearity in u indicates that the tension forces are equal and
opposite, hence no net force.

Laplace’s equation PDE

Laplace’s equation is a PDE for a function u(x, y), taking the form
uxx + uyy = 0.
The left hand side of this equation is often written ∇2 u, where ∇2 = ∇ · ∇, is known as the Laplace operator.
Without context, the PDE describes that either u is locally linear in both x and y or the non-linear aspects of
u with respect to each term are equal and opposite.

Laplace’s equation can be thought of as the equilibrium state (as t → ∞, ut → 0) of a two dimensional heat
equation. In this context, the heat transfer coefficient k is irrelevant because the equilibrium will be achieved
eventually, regardless of the rate at which the temperature spreads.

Alternatively, Laplace’s equation can be derived as a potential function by assuming that a vector field F⃗
is incompressible div(F⃗ ) = ∇ · F⃗ = 0 and irrotational curl(F⃗ ) = 0. Irrotational implies the existence of a
potential function ϕ such that ∇ϕ = F⃗ and then incompressibility implies ∇ · (∇ϕ) = ∇2 ϕ = 0.
Separation of variables (Part 1) - Obtaining ODEs

Separation of variables is a technique for solving PDEs, in particular, second-order linear PDEs. For a PDE
describing a function of two variables, say u(x, t), we make the assumption
u(x, t) = X(x)T (t),
where X(x) and T (t) are unknown functions. Our goal is then to substitute this form into the PDE equation,
and then rewrite the equation in a form that completely isolates all x terms from all t terms. After initially
taking derivatives and substituting into the PDE, progress towards this form is achieved by dividing the
equation by both X and T . We reach the below form, and write
F (x, X(x), X ′ (x), X ′′ (x)) = G(t, T (t), T ′ (t), T ′′ (t)) = λ,
where λ is an unknown real constant. From here, our goal is to solve the two ODEs
F (x, X(x), X ′ (x), X ′′ (x)) = λ, G(t, T (t), T ′ (t), T ′′ (t)) = λ,
that are linked via the shared value λ.
• The separation of variables technique assumes solutions of the separated form X(x)T (t). While this
assumption seems very restrictive, through the use of Fourier series, we can conclude that we will be able
to obtain a solution for any choice of physically reasonable initial conditions.
• When setting F (x, . . .) = G(t, . . .) = λ, we are claiming that F and G are constant functions. This works
because the equation must be true for all x and t. Choosing a value x = x0 fixes the value of F , but
G can still input any choice of t. This can only be true if G is a constant function, which immediately
results in F also being a constant function (as they are equal).
• After obtaining our two ODE equation, we still have an issue that λ could be any real constant. In
practice, dealing with λ in this state is impractical. Instead, we can restrict the possible λ values by
enforcing a boundary condition.

Intuition of common boundary conditions

There are a few boundary conditions that are commonly implemented for PDE problems, both due to their
physical nature and as we are able to obtain solutions in these forms.

The Dirichlet boundary conditions describe fixed boundaries at x = 0, L, taking the form u(0, t) = u(L, t) =
0. For the heat equation, we might consider that the boundaries are attached to a heating/cooling device that
is keeping the temperature fixed. For the wave equation, we might consider a string with fixed endpoints, like
a guitar, that will undergo oscillation when played.

The Neumann boundary conditions describe no-flux boundaries at x = 0, L, taking the form ux (0, t) =
ux (L, t) = 0. For the heat equation, this corresponds to the boundaries being insulated, meaning that no heat
will travel through the boundaries and the net temperature of the system will remain unchanged. If using
Laplace’s equation as a potential function, this corresponds to the vector field being parallel to the boundary.

Periodic boundary conditions describe boundaries points that represent the same location x = −L, L, taking
the form u(−L, t) = u(L, t) and ut (−L, t) = ut (L, t). For the heat equation, this corresponds to considering
a metal rod from x = −L to x = L, that is then bent into a ring by connecting the edges, meaning that
x = −L, L now correspond to a single point. The problem is solved on the interval [−L, L] and we manually
enforce a continuity and differentiable condition to ensure smoothness at x = ±L.
Separation of variables (Part 2) - Solving X ′′ − λX = 0 when λ is unknown

For PDEs that we will be solving, it is relatively common that as part of the separation of variables
technique, one of the ODEs to solve will be
X ′′
= λ =⇒ X ′′ − λX = 0.
F (x, X(x), X ′ (x), X ′′ (x)) =
X
This ODE appears when solving the heat equation, wave equation and Laplace’s equation.

Since this is a second-order linear constant-coefficient ODE, we can solve this using the auxiliary equation,
guessing the form X(x) = eαx . The resulting auxiliary equation is
α2 = λ,
at which point the solution form is dependent on the value of λ. We split the problem into three cases
(1) λ = µ2 > 0 =⇒ X(x) = Aeµx + Be−µx
(2) λ = 0 =⇒ X(x) = Ax + B
(3) λ = −µ2 < 0 =⇒ X(x) = A cos(µx) + B sin(µx).
At this point, we apply the boundary conditions to each case. For most choices of µ (and hence λ), the
boundary condition will only be able to satisfy the ODE if A = B = 0, resulting in a trivial solution. For
these cases, we can neglect the corresponding value of λ. If there is a choice of µ such that either A or B is
non-zero, then we include the term corresponding to that coefficient in our set of potential solutions.

Solutions to X ′′ − λX = 0 for Dirichlet boundary conditions

Consider the Dirichlet boundary conditions u(0, t) = u(L, t) = 0. Under the separable assumption, we
obtain the following boundary conditions
u(0, t) = X(0)T (t) = 0 =⇒ X(0) = 0,
u(L, t) = X(L)T (t) = 0 =⇒ X(L) = 0
The problem X ′′ − λX = 0 with boundary conditions X(0) = X(L) = 0 has solution
 nπx   nπ 2
Xn (x) = sin where λn = − , for n = 1, 2, . . . .
L L
• In this case, we find that only wave-like solutions will satisfy the boundary conditions. For exponential
solutions, the given form will never pass through the origin, meaning it is impossible for X(x) = 0 for
any x ∈ [0, L]. For linear solutions, the line through (0, 0) and (L, 0) is clearly X = 0, hence we also
cannot have this case occur.
• For the final case, it is possible to have non-trivial sine wave pass through both (0, 0) and (L, 0), while
this is clearly impossible for cosine waves that will always be at a peak at x = 0. When a sine wave
travels through both these points, it will do so regardless of the choice of amplitude, meaning that the
coefficient can be any non-zero number.
• Importantly, there are also infinitely many sine waves that do not satisfy the boundary conditions. The
given selection of λn values are those that give waves whose frequencies align with the conditions, but
choosing a λ ∈ (λn+1 , λn ) will also give trivial solutions. We denote the λ’s that satisfy the boundary
conditions as λn , and pair them with their corresponding non-trivial solutions Xn (x).
Solutions to X ′′ − λX = 0 for Neumann boundary conditions

Consider the Neumann boundary conditions ux (0, t) = ux (L, t) = 0. Under the separable assumption, we
obtain the following boundary conditions
ux (0, t) = X ′ (0)T (t) = 0 =⇒ X ′ (0) = 0,
ux (L, t) = X ′ (L)T (t) = 0 =⇒ X ′ (L) = 0.
The problem X ′′ − λX = 0 with boundary conditions X ′ (0) = X ′ (L) = 0 has solution
 nπx   nπ 2
Xn (x) = cos where λn = − , for n = 0, 1, 2, . . . .
L L
• For this set of solutions, while n = 0 technically fits into the pattern, it actually results in a different
form, with X0 = 1 and λ0 = 0. This corresponds to the linear solution case where any constant line will
have derivative zero at the boundaries. This is often worth separating because in practice these terms
result indifferent behaviours than most cosine terms.
• The wave-like solutions here are the opposite of the Dirichlet boundary conditions. Given that cosine is
at its peak at x = 0 the slope will always be zero on the boundary here, while sine curves will always be
going through zero, hence increasing.

Solutions to X ′′ − λX = 0 for periodic boundary conditions

Consider the periodic boundary conditions u(−L, t) = u(L, t) and ux (−L, t) = ux (L, t). Under the separable
assumption, we obtain the following boundary conditions
u(−L, t) = u(L, t) =⇒ X(−L)T (t) = X(L)T (t) =⇒ X(−L) = X(L),
ux (−L, t) = ux (L, t) =⇒ X ′ (−L)T (t) = X ′ (L)T (t) =⇒ X ′ (−L) = X ′ (L).
The problem X ′′ − λX = 0 with boundary conditions X(−L) = X(L) and X ′ (−L) = X ′ (L) has solution
 nπx   nπx   nπ 2
Xn (x) = An cos + Bn sin , where λn = − , for n = 0, 1, 2, . . . .
L L L
• Periodic boundary conditions appear to have both qualities of Dirichlet and Neumann boundary condi-
tions, taking both sets of terms.
• Similar to Neumann boundary conditions, the n = 0 case is somewhat strange and might be best to
consider separately. The n = 0 solution is technically X0 = A0 and λ0 = 0, where the B0 term does not
really exist as n = 0 will always cause the sine term to disappear.
• For this solution form, we have chosen to write the coefficients, while in previous cases we did not.
Technically, we could also do that here, writing something like
 nπx   nπx 
Xn (x) = cos , sin ,
L L
as in both cases, our only goal is collect the set of all separate solutions. There is the potential for the
above notation to become even more confusing, with each Xn corresponding to solutions for a given
n-value, but also each Xn giving two independent solutions. If written in this form (or the forms for the
Dirichlet or Neumann cases above) the arbitrary coefficients will be reintroduced later when we take a
superposition of solutions.
Separation of variable (Part 3) - Obtain a general solution

Upon applying the boundary condition, you should have the following information:
1. the set of values λn that do not contradict the ODE obtained via F (x, X(x), X ′ (x), X ′′ (X)) = λn or the
boundary conditions associated with the x variable,
2. the solutions Xn (x) that correspond to each non-contradictory λn value.

This will leave you with only a single ODE

G(t, Tn (t), Tn′ (t), Tn′′ (t)) = λn ,
which we only wish to solve for non-contradictory λn . Upon solving this equation, you will obtain a solution
form for Tn (t). Under some circumstances, it may be reasonable to apply initial/boundary conditions to Tn (t)
if the form of the conditions allows this. However, in most cases this is not necessary and can simply be dealt
with after the next step.

Importantly, since our two ODEs were linked by λ, this means that each solution Xn (x) should have a
corresponding solution Tn (t). We can hence write individual solutions to the PDE in the form
un (x, t) = Xn (x)Tn (t),
in which these solutions will satisfy both the PDE and the boundary conditions for x. Since the PDE is linear,
we can obtain solutions via superposition. As we have obtained infinitely many solutions (for each n-value)
we can write a solution in the form of an infinite series
X∞ ∞
X
u(x, t) = un (x, t) = Xn (x)Tn (t).
n=n0 n=n0

where n0 is the starting value of n (typically zero or one).

For the three PDEs (heat equation, wave equation, Laplace’s equation), with one of the three boundary
conditions (Dirichlet, Neumann, periodic), the infinite series written by following the above steps will be the
general solution to the PDE.
• As stated above, the full separation of variable approach will work for any combination of the three PDEs
and three boundary conditions we have considered, however the method is relatively versatile and can
potentially be applied to a wide variety of PDEs with a range of boundary conditions. It is exceptionally
difficult to narrow down exactly when the method can be applied, and some steps may vary slightly
based on the exact PDE forms or boundary conditions imposed.
• In both the heat equation and the wave equation, using any of our three boundary conditions results in
wave-like solutions. Further, upon applying the initial condition, we obtain an equation that says our
initial condition is equal to a superposition of waves with unknown amplitudes (arbitrary coefficients).
At this point, we have no technique for solving such an equation unless our initial condition is also
wave-like.
• We also claim that the given solution, at least in our ideal cases, is the general solution to the PDE. By
this, we mean that for any choice of initial condition (or potentially boundary condition in cases such as
Laplace’s equation) we should be able to obtain the unique solution corresponding to it. Beyond the issue
that our initial condition may not be wave-like, we should also note that the premise of the separation
of variables method is to restrict our solutions search to functions of the form u(x, t) = X(x)T (t). In
order to resolve both these issues, we must introduce Fourier series.
Periodic functions

A function f (x) is called periodic with period P if f (x) = f (x + P ) for all x ∈ R.

Every function that is periodic with period P also has period kP for all k ∈ N. The minimum period of
a function is the smallest value of P > 0 such that f (x) = f (x + P ) for all x ∈ R.

The functions
 nπx   nπx 
sin , and cos ,
L L
for n = 1, 2, . . . , are all sine and cosine waves with period P = 2L.
• The set of sine and cosine waves given above arose naturally when solving PDEs with Dirichlet, Neumann
and periodic boundary conditions.
• While sine and cosine waves, along with other trigonometric functions, are the most commonly thought
of periodic functions, we can use piecewise functions and domain restrictions to define a wide range of
different periodic functions.

Orthogonality of functions on a finite interval

Consider the set of all real functions defined on a finite interval [−L, L]. For any two functions f (x) and
g(x) in the set, we define these functions to be orthogonal if
ˆ L
f (x)g(x) dx = 0.
−L
The set containing a constant function, and all sine and cosine waves with period 2L is
n  nπx   nπx o
1, cos , sin for n = 1, 2, . . . .
L L
Any two unique elements of this set are orthogonal to each other, that is, for m1 , m2 ∈ N with m1 ̸= m2 then
ˆ L  m πx  ˆ L  m πx  ˆ L  m πx   m πx 
1 1 1 2
1 · cos dx = 0, 1 · sin dx = 0, cos sin dx = 0,
−L L −L L −L L L
ˆ L  m πx   m πx  ˆ L  m πx   m πx 
1 2 1 2
cos cos dx = 0, sin sin dx = 0.
−L L L −L L L
This choice of set also has the following useful property that for any m ∈ N
ˆ L ˆ L  mπx   mπx  ˆ L  mπx   mπx 
1 · 1 dx = 2L, cos cos dx = L, sin sin dx = L.
−L −L L L −L L L
• The implication of the above integrals is that the set of functions in the above set are linearly independent
within the context of real functions defined on the interval [−L, L]. This means that each function
contributes something that no other function (or combination of functions) in the set could contribute.
• Although it is difficult to demonstrate, the integral of any function f (x) with period 2L that is not a
constant, sine or cosine function will be non-orthogonal to at least one of the elements of the above set.
• The implication of the above two result is that the above set is a basis for all f (x) with period 2L. In
other words, all f (x) with period 2L can be uniquely constructed using a linear combination of constants,
and sine and cosine waves with period 2L.
Fourier series - real form

Consider a function f (x) with period 2L. The function f (x) can be written as a Fourier series of the form
∞
a0 X  nπx   nπx 
f (x) = + an cos + bn sin ,
2 L L
n=1

where the coefficients of each term can be determined by calculating the following integrals
ˆ ˆ ˆ
1 L 1 L  nπx  1 L  nπx 
a0 = f (x) dx, an = f (x) cos dx, bn = f (x) sin dx.
L −L L −L L L −L L

• Fourier series is the statement that all functions with period 2L can be uniquely written as a linear
combination of constants, and sine and cosine waves with period 2L.

• The coefficients can be determined using the integrals presented in the orthogonality section above. By
multiply by Fourier series expression by one of the basis terms (constant, sine or cosine function) and
then integrating over the interval [−L, L], all but one of the infinitely many terms will disappear, leaving
us with only a single coefficient, which can then be written as an integral.

• It is useful to remember that each coefficient is simply a number. Many of the results can looks somewhat
unpleasant, particularly being in terms of n, however each n simply corresponds to a term in the sum,
giving some number as the coefficient to each function in the infinite series.

Fourier series - finite number of terms

When thinking of a Fourier series as trying to write a function as a linear combination of constants, sine
and cosine functions, it should be clear that any sum of a finite collection of sine and cosine waves that share
a common period is already in form of a Fourier series. In these cases, we only require finitely many non-zero
coefficients, resulting in a finite collection of terms.

This can also arise in situations where f (x) is the product of sine and cosine waves,

f (x) = cos(k1 x)m1 sin(k2 x)m2 ,

where k1 , k2 ∈ R and m1 , m2 ∈ N (including zero). The above form can always be reduced to the sum of a
finite collection of constant, sine and cosine functions using trigonometric identities. In these cases, so long as
f (x) is periodic (which might not occur for some values of k1 , k2 ), then f (x) can be written as a Fourier series
with a finite number of terms.
• For most 2L periodic functions, we will not have the nice conditions above and hence obtain infinitely
many terms in the Fourier series, writing an and bn in terms of n.

• In cases with a finite collection of constant, sine and cosine terms, it is often more convenient to simply
equate coefficients if we are looking to state the coefficients explicitly.

• Importantly, taking a 2L periodic function that is the sum and/or product of sine and cosine terms
and simply inputting it into the Fourier coefficient formulas will still result in the correct values for the
coefficients, although will typically be more work.
Fourier series of odd and even functions

Consider an odd function f (x) with period 2L. The Fourier series of f (x) take the form
∞  nπx  ˆ L
X 2  nπx 
f (x) = bn sin , where bn = f (x) sin dx.
L L 0 L
n=1

This can referred to as a Fourier sine series.

Consider a even function f (x) with period 2L. The Fourier series of f (x) take the form
∞ ˆ L ˆ L
a0 X  nπx  2 2  nπx 
f (x) = + an cos , where a0 = f (x) dx, an = f (x) cos dx.
2 L L 0 L 0 L
n=1

This can referred to as a Fourier cosine series.

• The above equations can quickly be derived directly from the standard Fourier series. For the Fourier
sine series, we obtain a0 = 0 and an = 0 from their corresponding integrals. Since odd f (x) multiplied
by an even constant or cosine function gives an odd function, the integral over the interval [−L, L] will
be zero. Since this is always the case for odd f (x), we can simply ignore the terms. A similar argument
is made for bn = 0 when f (x) is even.

• The non-zero coefficients are also simplified. For the Fourier sine series, both f (x) and sine functions
are odd, meaning that their product is even. For an even integral, we can simplify the integral from
[−L, L] by instead integrating from [0, L] and multiplying the result by two, which generally simplifies
the procedure. A similar argument is made for a0 and an when f (x) is even.

• We can also make the logical argument that since f (x) is odd/even, the Fourier series terms must be
made up of exclusively odd/even terms respectively. This is a generally good perspective to be aware
of, as it also appears in other instance, such as Taylor series (for instance, the Taylor series for sin(x)
contains only odd polynomial terms).

Complex forms of sine and cosine

Using Taylor series, we are able to derive Euler’s identity

eix = cos(x) + i sin(x).

Using this identity and complex conjugates, we can also write sine and cosine in terms of complex exponentials

eix + e−ix e − e−ix

 ix 
cos(x) = and sin(x) = −i .
2 2

• Writing sine and cosine explicitly in terms of complex exponentials can be extremely beneficial. In
particular, writing these forms can allow for the derivation of almost any trigonometric identity very
quickly without the need to construct and argue the geometric context.

• Notably, for those that have used hyperbolic sine and hyperbolic cosine, we see the relationship between
their naming conventions, despite the functions having no visible relationship geometrically.
Fourier series - complex form

Consider a function f (x) with period 2L. The complex form of the Fourier series of f (x) is
∞ ˆ L
nπx 1 nπx
cn ei( L ) , where cn = f (x)e−i( L ) dx.
X
f (x) =
n=−∞
2L −L
This is simply a different form of the Fourier series formula and is directly related to the real form of the
Fourier series via the following equations
cn = c−n , 2cn = an − ibn , an = 2Re[cn ], bn = −2Im[cn ],
where, for some z ∈ C, z is the complex conjugate of z, and Re[z] and Im[z] are the real and imaginary
components of z respectively.
• The complex form of Fourier series can be derived through applying the complex exponential forms of
cosine and sine to the real form of the Fourier series. At some point in that derivation, the coefficients
are redefined from an and bn to cn and c−n , resulting in the above relationships.
• The complex form of Fourier series, while representing the same series, is somewhat less intuitive. How-
ever, it has the benefit of typically being easier to compute, only requiring the solution of a single
coefficient cn instead of both an and bn . It is often simpler to find the complex coefficients and then
determine an and bn using real and complex parts if the real form is desired.
• If f (x) is odd or even, then typically the complex form is not as efficient due to the simplified forms of
the Fourier sine and cosine series. Notably, if f (x) is odd or even, cn will have no real or imaginary part
respectively to enforce the condition that an = 0 or bn = 0, as expected.
• We might also realize that the implication of writing the Fourier series in this form is that the set of
nπx nπx
functions ei( L ) for all n ∈ Z form a basis for all 2L periodic functions. This indicates that each ei( L )
function is orthogonal to each other one, with them being the minimum set resulting in the unique
coefficients.

Fourier series convergence theorem

Let f (x) be a function with period 2L. If f (x) has finitely many non-differentiable points on [−L, L] and
for each a ∈ [−L, L] where f (x) is not differentiable, the one sided limits exist and are denoted
f (a+ ) = lim f (x), f (a− ) = lim f (x),
x→a+ t→a−

then at every point x = x0 where f (x) is continuous the Fourier series of f (x) converges to the value f (x0 )
∞ ∞
a0 X nπx0
cn ei( L ) ,
X
f (x0 ) = + an cos(nx0 ) + bn sin(nx0 ) =
2 n=−∞
n=1

and at any other point, it converges to the average of the upper and lower limiting values,
− ∞ ∞
f (x+
0 ) + f (x0 ) a0 X nπx0
cn ei( L ) .
X
= + an cos(nx0 ) + bn sin(nx0 ) =
2 2 n=−∞
n=1

• Importantly, this theorem indicates that even when our function f (x) has a finite collection of finite valued
non-differentiable points, we can still construct a Fourier series of the function, with discontinuous points
being evaluated at the midpoint of their one sided limits.
Periodic extensions

Consider a function f (x) defined on a bounded interval [−L, L]. Since the function is not periodic, it
technically cannot be represented by a Fourier series. However, we can define the following function
fp (x) = f (x) for x ∈ [−L, L], and fp (x) = fp (x + 2L).
This new function, fp (x) is exactly f (x) on the interval [−L, L], while also being periodic. As such, we can
construct the Fourier series for fp (x) and then say
∞ ∞
i( nπx
L )
nπx
cn ei( ) , for x ∈ [−L, L].
X X
fp (x) = cn e =⇒ f (x) = L

n=−∞ n=−∞

• The construction of a periodic extension is essentially a mathematical artifact. By performing this trick,
we are able to develop a corresponding periodic function and then simply ignore it on any interval that
is not relevant to our interests.
• While in practice, the function fp (x) is irrelevant to us, if it is not possible to construct fp (x) then we
cannot use this technique. As presented above, the bounded interval of f (x) allowed us to construct a
corresponding periodic function, hence allowing the construction of a Fourier series. If the function was
unbounded on either side, then we would be unable to simply copy the structure to form a corresponding
periodic function fp (x).

Odd and even periodic extensions

Consider a function f (x) defined on a bounded interval [0, L]. If we define the following function

f (x) for x ∈ [0, L]
fo (x) = , and fo (x) = fo (x + 2L).
−f (−x) for x ∈ [−L, 0]
then fo (x) is exactly f (x) on the interval [0, L], while also being an odd periodic function. As such, we can
construct the Fourier series for fo (x) and then say
∞
X  nπx  ∞
X  nπx 
fo (x) = bn sin =⇒ f (x) = bn sin , for x ∈ [0, L].
L L
n=1 n=1

If instead we define the function


f (x) for x ∈ [0, L]
fe (x) = , and fe (x) = fe (x + 2L).
f (−x) for x ∈ [−L, 0]
then fe (x) is exactly f (x) on the interval [0, L], while also being an even periodic function. As such, we can
construct the Fourier series for fe (x) and then say
∞ ∞
a0 X  nπx  a0 X  nπx 
fe (x) = + an cos =⇒ f (x) = + an cos , for x ∈ [0, L].
2 L 2 L
n=1 n=1

• We can use the same process to define the [−L, 0] interval to be any function we want, however these
two choices are convenient for specifically giving odd and even functions.
• The requirement to obtain a Fourier sine or cosine series representation for a bounded function arises
naturally in PDE problems.
Solving PDEs with arbitrary initial conditions

The general solution to a PDE when using the separation of variables technique will take the form
∞
X
u(x, t) = Xn (x)Tn (t),
n=n0

where typically n0 = 0 or n0 = 1. Depending on the exact PDE, boundary conditions and initial conditions,
there are numerous different forms that can be obtained, however we do find some consistencies. We will
consider an initial condition u(x, 0) = f (x) (this could also represent an initial condition ut (x, 0) = f (x)).

For a problem with Dirichlet boundary conditions, f (x) will be defined on an interval [0, L]. When the
initial condition is applied to the general solution, we should obtain a form
∞
X  nπx 
f (x) = An sin ,
L
n=1

where the An terms are unknown constants. In this case, we should construct the odd periodic extension for
f (x), allowing us to determine the coefficients An .

For a problem with Neumann boundary conditions, f (x) will be defined on an interval [0, L]. When the
initial condition is applied to the general solution, we should obtain a form
∞
X  nπx 
f (x) = A0 + An cos ,
L
n=1

where the An terms are unknown constants. In this case, we should construct the even periodic extension for
f (x), allowing us to determine the coefficients An .

For a problem with periodic boundary conditions, f (x) will be defined on an interval [−L, L]. When the
initial condition is applied to the general solution, we should obtain a form
∞
X  nπx   nπx 
f (x) = A0 + An cos + Bn sin ,
L L
n=1

where the An and Bn terms are unknown constants. In this case, we should construct a periodic extension
for f (x) (notably this only requires the periodic step as we are already defined on [−L, L]), allowing us to
determine the coefficients An and Bn .

• In each case above, upon learning the unknown coefficients based on the initial condition, they can then
be substituted into the general solution to the PDE to obtain the unique solution corresponding to both
the initial condition(s) and the boundary conditions.
Linear non-homogeneous ODEs

The following differential equation is an nth-order linear non-homogeneous ODE

an (t)y (n) (t) + an−1 (t)y (n−1) (t) + . . . + a1 (t)y ′ (t) + a0 (t)y(t) = f (t),

where each ai (t) is a known real function of t (i = 0, 1, . . . , n), f (t) is a known function of t, and both an (t)
and f (t) are non-zero functions.

• The function f (t), sometimes referred to as a forcing term, corresponds to an influence of the system
independent of the objects own properties. Non-homogeneous terms arise often, for example, in simple
projectile motion, acceleration due to gravity is treated as independent of the position and velocity of
the object.

• At this point, we have already learned to solve the general first order non-homogeneous ODE, namely
y ′ + p(x)y(x) = q(x), using the integrating factor.

Superposition of homogeneous and non-homogeneous solutions

Consider an nth-order linear non-homogeneous ODE

an (t)y (n) (t) + an−1 (t)y (n−1) (t) + . . . + a1 (t)y ′ (t) + a0 (t)y(t) = f (t).

The general solution to the ODE is given by

y(t) = yh (t) + yp (t),

where yh (t) (known as the homogeneous solution) is the general solution to the corresponding homoge-
neous ODE
(n) (n−1)
an (t)yh (t) + an−1 (t)yh (t) + . . . + a1 (t)yh′ (t) + a0 (t)yh (t) = 0,

and yp (t) (known as the particular solution) is any one solution to the non-homogeneous ODE

an (t)yp(n) (t) + an−1 (t)yp(n−1) (t) + . . . + a1 (t)yp′ (t) + a0 (t)yp (t) = f (t).

• Summing the general homogeneous solution and the particular solution indicates that the result are all
solution to the ODE.

• Using a similar approach, subtracting two known particular solutions to the non-homogeneous ODE
will result in the difference satisfying the homogeneous ODE. This means that any solution to the
homogeneous can shift any particular solution to any other particular solution of the system.

• Using this superposition means that only a single solution to the non-homogeneous ODE is required, so
long as the homogeneous solution to the ODE can be obtained.
Superposition of particular solutions

Consider the following two non-homogeneous ODEs with the same corresponding homogeneous ODE but
different forcing terms
an (t)y (n) (t) + an−1 (t)y (n−1) (t) + . . . + a1 (t)y ′ (t) + a0 (t)y(t) = f1 (t)
an (t)y (n) (t) + an−1 (t)y (n−1) (t) + . . . + a1 (t)y ′ (t) + a0 (t)y(t) = f2 (t).
If these ODEs have particular solutions yp1 and yp2 respectively then the linear non-homogeneous ODE
an (t)y (n) (t) + an−1 (t)y (n−1) (t) + . . . + a1 (t)y ′ (t) + a0 (t)y(t) = c1 f1 (t) + c2 f2 (t).
has a particular solution of the form
yp (t) = c1 yp1 (t) + c2 yp2 (t).

• This superposition technique means that we can individually solve for separate particular solutions before
adding them together.

Method of undetermined coefficients - exponential, sinusoidal, polynomial

Consider a linear constant-coefficient non-homogeneous ODE with forcing term f (t). For specific choices
of f (t) we are simply able to guess the form of the particular solution. In each case, the guess yp and its
derivatives are substituted into the ODE and used to determine the values of the unknown coefficients bi .

If f (t) = cekt , where c, k are real non-zero constants then the particular solution will take the form
yp (t) = bekt ,
for some unknown value b, assuming this guess yp (t) is not a term in the homogeneous solution yh (t).

If f (t) = c1 cos(ωt)+c2 sin(ωt), where c1 , c2 , ω are real constants, ω is non-zeros, and at least one is non-zero
then the particular solution will take the form
yp (t) = b1 cos(ωt) + b2 sin(ωt),
for some unknown values b1 , b2 , assuming neither term in yp (t) is a term in the homogeneous solution yh (t).

If f (t) = cn tn + cn−1 tn−1 + . . . + c1 t + c0 , where n is some natural number (including zero), each ci is a real
constant and cn is non-zero then the particular solution will take the form
yp (t) = bn tn + bn−1 tn−1 + . . . + b1 t + b0 ,
for some unknown values bi , assuming no terms in yp (t) are terms in the homogeneous solution yh (t).
• Each guess replies on the fact that these functions have repeating or contained derivative patterns. This
is particularly important in the sinusoidal case where if c1 or c2 were zero, we still guess both sine and
cosine terms, as guessing one will cause the other to appear upon taking derivatives. This idea will be
addressed further in the next section.
• If our desired guess yp (t) contains terms in the homogeneous solution yh (t), we must amend our guess.
This case is addressed in a section below.
• For constant-coefficient ODEs, the homogeneous solution is easily obtained via the auxiliary equation,
hence seeking a single particular solution via this method is ideal.
Method of undetermined coefficients - products of guesses and finite derivative patterns

Consider a linear constant-coefficient non-homogeneous ODE with forcing term f (t). An appropriate guess
can be determined under any circumstance where all possible appearing derivatives are chosen as part of the
guess. If f (t) is the product of exponential, sinusoidal or polynomial terms, we can achieve this goal.

If f (t) takes the form

f (t) = cn tn + cn−1 tn−1 + . . . + c1 t + c0 ekt cos(ωt) + dn tn + dn−1 tn−1 + . . . + d1 t + d0 ekt sin(ωt),

where n is some natural number (including zero), ci , di , k, ω are real constants, and at least one of cn , dn is
non-zero then the particular solution will take the form
yp (t) = an tn + an−1 tn−1 + . . . + a1 t + a0 ekt cos(ωt) + bn tn + bn−1 tn−1 + . . . + b1 t + b0 ekt sin(ωt),

for some unknown values ai , bi , assuming no terms in yp (t) are terms in the homogeneous solution yh (t).
• The above form will work because indefinitely taking the derivative of the above f (t) will result in a form
equivalent to our guess. The method of undetermined coefficients requires a finite collection of terms in
order to create a resolvable system of equation at the end of the procedure, which does not occur for
most choices of f (t).
• Notably, this form also includes the form for each paired combination. For exponential and sinusoidal,
the highest order polynomial terms will be n = 0, for sinusoidal and polynomial, we set the exponent
k = 0, and for exponential and polynomial, we set ω = 0 removing all sine terms and leaving only terms
attached to cosine. Any two of these conditions would reduce to the original guess form.

Method of undetermined coefficients - desired guess is part of the homogeneous solution

Consider a linear constant-coefficient non-homogeneous ODE with forcing term f (t) of the form
f (t) = cn tn + cn−1 tn−1 + . . . + c1 t + c0 ekt cos(ωt) + dn tn + dn−1 tn−1 + . . . + d1 t + d0 ekt sin(ωt),

for some natural number n (including zero), real constants ci , di , k, ω, and at least one of cn , dn being non-zero.

Based on the above form, we can make an initial guess of particular solution, say y1 (t), to apply the method
of undetermined coefficients. Upon doing this, we then check if any of the terms in y1 (t) are also terms in
the homogeneous solution yh (t). If no terms are shared between y1 (t) and yh (t) then we guess the particular
solution yp (t) = y1 (t).

However, if y1 (t) and yh (t) share any terms, we make a new guess by multiplying our initial guess by t,
giving y2 (t) = ty1 (t). We then repeat this process: if y2 (t) and yh (t) share no terms, set yp (t) = y2 (t), but if
they share any terms, construct the new guess y3 (t) = ty2 (t) = t2 y1 (t). If the ODE is mth order, then yh (t)
has at most m terms, hence by guess ym+1 = tm+1 y1 (t) we will have found a form for the particular solution
that will definitely work.
• Between our guessing structure based off f (t) and our amendments to our guess based on f (t), the above
procedures allow us to obtain a solutions for any f (t) made up of any product of exponential, sinusoidal
and polynomial terms.
• If our f (t) is the sum of terms that will require different guesses, then we can resolve each individually
and use the superposition of particular solutions to obtain the particular solution for our ODE.
Periodic forcing terms

Consider an mth-order linear constant-coefficient non-homogeneous ODE

cm y (m) (t) + cm−1 y (m−1) (t) + . . . + c1 y ′ (t) + c0 y(t) = f (t),

such that f (t) = f (t + 2L), that is f (t) has period 2L. If this is the case then f (t) can be written as a Fourier
series, meaning that we can rewrite the ODE
∞
(m) (m−1) ′ C0 X  nπx   nπx 
cm y (t) + cm−1 y (t) + . . . + c1 y (t) + c0 y(t) = + Cn cos + Dn sin ,
2 L L
n=1

where the Ci , Di are known coefficients unique to the periodic function f (t). As such, we can make a guess of
the Fourier series form
∞
a0 X  nπx   nπx 
yp (t) = + an cos + bn sin ,
2 L L
n=1

and seek the constants ai , bi to obtain the particular solution (assuming no terms in yp appear in yh ).
• While this task seems potentially daunting, note that each term in yp (t) will only interact with sine and
cosine terms of the same frequency. This means that when we equate coefficients, we are simply ensuring
that the nth frequency term is satisfied simultaneously for all natural numbers n.

Method of undetermined coefficients - complex form

Consider a linear constant-coefficient non-homogeneous ODE

an y (n) (t) + an−1 y (n−1) (t) + . . . + a1 y ′ (t) + a0 y(t) = c1 cos(ωt) + c2 sin(ωt).

We can obtain the particular solution to this by instead considering the complex ODE

an z (n) (t) + an−1 z (n−1) (t) + . . . + a1 z ′ (t) + a0 z(t) = eiωt ,

which can be solved using the initial guess zp (t) = beiωt where b ∈ C. If we consider that z = Re[z] + iIm[z]
then we can obtain the following two ODEs by considering the real and complex parts of the ODE

an Re[z (n) (t)] + an−1 Re[z (n−1) (t)] + . . . + a1 Re[z ′ (t)] + a0 Re[z(t)] = cos(ωt),
an Im[z (n) (t)] + an−1 Im[z (n−1) (t)] + . . . + a1 Im[z ′ (t)] + a0 Im[z(t)] = sin(ωt).

Using superposition of the particular solutions, we obtain yp = c1 Re[zp ] + c2 Im[zp ]

• The complex form used here is typically more effective to use than the standard sinusoidal form, due to
the simplicity of the derivatives obtained for eiωt and the fact that we have fewer overall terms.
• This form can be used under any circumstance where a sinusoidal form appears in the particular solution,
even when multiplied by an exponential or polynomial term. For instance, if f (t) = ekt cos(ωt) we can
initially guess yp (t) = e(k+iω)t .
• This form still represents sinusoidal solutions, hence if yh (t) contains terms cos(ωt) and sin(ωt) then we
should guess yp (t) = bteiωt instead of simply beiωt .
Sinusoidal forms

There are three useful forms for sinusoidal functions. The first is the standard form
C1 cos(ωt) + C2 sin(ωt).
This form demonstrates the linear decomposition of sinusoidal waves of frequency ω into two independent
terms, namely a sine wave and a cosine wave.

The second form is the complex exponential form obtained via Euler’s identity
B1 eiωt + B2 e−iωt = (B1 + B2 ) cos(ωt) + i(B1 − B2 ) sin(ωt).
This form is often mathematically the most valuable, for instance, giving a simplified approach for constructing
Fourier series, and allowing for derivations of trigonometric identities without the need for physical intuition.
It is also the form that naturally appears when solving the auxiliary equation, and often allows for more
efficient calculations taking advantage of complex conjugate properties such as in the simplified approach to
the method of undetermined coefficients.

The third form is phase-shifted form

A cos(ωt − ϕ) = (A cos(ϕ)) cos(ωt) + (A sin(ϕ)) sin(ωt).
In this form, we are able to see all the physically meaningful components of the sinusoid. The coefficient A
represents the waves amplitude, ω the waves frequency, and ϕ the waves phase-shift.
• Each form has different benefits, so the ability to transition between each form and to recognize the ideal
form are important skills to develop.

Fourier transform

The Fourier transform of a function f (x) is given by and denoted

ˆ ∞
F{f (x)} = f (x)e−iωx dx = fˆ(ω),
−∞

where fˆ(ω) is a function of the variable ω. The inverse Fourier transform is given by
ˆ ∞
−1 ˆ 1
F {f (ω)} = fˆ(ω)eiωx dω = f (x).
2π −∞

• The Fourier transform is the limiting case of a Fourier series in which we consider the continuum of
frequencies ω and their corresponding amplitudes fˆ(ω).
• Since more frequencies are considered as the period 2L increases, we arrive at a continuum by having
L → ∞, which essentially means the function is not periodic.
• For a Fourier series to be valid, the function f (x) must be an L2 function
ˆ ∞
|f (x)|2 dx < ∞,
−∞

meaning that it must converge sufficiently quickly to zero as we move away from the origin.
Laplace transform

The Laplace transform of a function f (t) with domain t ≥ 0 is given by and denoted
ˆ ∞
L{f (t)} = f (t)e−st dt = F (s),
0

where F (s) is a function of the variable s.

• The Laplace transform is a more general version of a Fourier transform, in which the transformed variable
s can be thought of as a complex exponent s = σ + iω. Convergence will only occur for some range of
values of σ = Re[s], with this restriction indicating the exponential component of the function. Similarly,
any complex components ω = Im[s] indicate sinusoidal frequencies of the function.

• Typically, when using Laplace transforms, lower case functions are used to represent the function prior
to transformation and upper case functions are used to represent the transformed function.

Linear property of Laplace transforms

The Laplace transform is a linear operator, meaning that

L{c1 f (t) + c2 g(t)} = c1 L{f (t)} + c2 L{g(t)},

where c1 , c2 are constants and f (t), g(t) are functions whose Laplace transforms exist.

• Laplace transforms being a linear operator is directly a consequence of integration being a linear operator.

Uniqueness of Laplace transforms and inverse Laplace transforms

Consider two function f (t), g(t) whose Laplace transforms exist. If we observe that the Laplace transforms
of the functions are equal then this indicates that the functions themselves must be equal, that is

L{f (t)} = L{g(t)} ⇐⇒ f (t) = g(t).

Since Laplace transforms are a linear operator and the Laplace transform of a function is unique, we can
immediately conclude that the inverse Laplace transform exists and that

L{f (t)} = F (s) ⇐⇒ f (t) = L−1 {F (s)}.

• Using the above rules, we are able to determine the inverse Laplace transform for any functions whose
Laplace transforms are known. This is much like how the fundamental theorem of calculus allows us to
more easily calculate integrals as anti-derivatives.

• While there does exist a formula for the inverse Laplace transform, it is impractical to use in the context
of our work.
Laplace transforms - exponentials and sinusoids

The Laplace transform of an exponential term is given by

ˆ ∞
1
kt
L{e } = e−(s−k)t dt = ,
0 s−k

which will converge for s > Re[k], where k ∈ C.

The Laplace transform of sine and cosine terms are given by

e + e−iωt
 iωt 
1 1 s
L{cos(ωt)} = L = + = 2 ,
2 2(s − iω) 2(s + iω) s + ω2
e − e−iωt
 iωt 
1 1 ω
L{sin(ωt)} = L = − = 2 ,
2i 2i(s − iω) 2i(s + iω) s + ω2

which will converge for s > 0, where ω ∈ R.

• Given that Laplace transforms aim to identify the exponential and sinusoidal components of a function,
these functions are a natural starting place.

• The Laplace transform identifies exponential and sinusoidal components of a function via the poles
(division by zero) of the resulting transform. That is, if k ∈ R, then ekt is purely exponential with
exponent rate k, hence a pole at s = k. For sine or cosine with frequency ω, the poles occur at s = ±iω.

• Sinusoids converge for s > 0 since they are constructed complex exponentials with no real part.

Laplace transforms - exponential shift (first shifting theorem)

Consider a function f (t) whose Laplace transform exists. Then if we define L{f (t)} = F (s) which converges
for s > s0 , we find that

L{ekt f (t)} = F (s − k),

which will converge for s > s0 + k, where k ∈ R.

Of specific interest is the product of exponentials and sinusoids, which have the following Laplace transforms
s−k ω
L{ekt cos(ωt)} = , L{ekt sin(ωt)} = ,
(s − k)2 + ω 2 (s − k)2 + ω 2

which will converge for s > k, where k, ω ∈ R.

• For the product of exponential and sinusoidal functions, with exponential rate k and frequency ω, the
poles occur at s = k ± iω, recognizing the non-zero exponential and sinusoidal components.

• If we consider k ∈ C, then this result gives us a means of moving the poles of a function f (t). Notably,
the result indicates that the entire complex plane is also shifted, placing the origin at s = k.
Laplace transforms - derivatives of functions

Consider a function f (t) whose Laplace transform exists. Then if we define L{f (t)} = F (s) which converges
for s > s0 , and say that lim y(t)e−st = 0 for s > s1 , then we find that
t→∞

L{f ′ (t)} = sF (s) − f (0),

which will converge for s > max(s0 , s1 ).

This rule be reapplied to higher-order derivatives, giving that

L{f (n) (t)} = sL{f (n−1) (t)} − f (n−1) (0),

and applied recursively gives

n
X
(n) n
L{f (t)} = s F (s) − sn−i f (i−1) (0),
i=1

which will converge for s > max(s0 , s1 , . . . , sn ), where each si is defined such that lim y (i−1) (t)e−st = 0 only
t→∞
converges when s > si .

• The base step of this rule can be derived by applying integration by parts to the definition of the Laplace
transform of a derivative.

• This Laplace transform rule shows that all derivatives of a function can be rewritten in terms of the
Laplace transform of the original function. This allows Laplace transforms to convert differential equa-
tions into algebraic equations, often making solutions easier to obtain in the transformed space.

Laplace transforms - polynomials

The Laplace transform of a constant is given by

1
L{1} = ,
s
which will converge for s > 0.

If we then define the general polynomial term p(t) = tn , realizing that p(n) (t) = n!, and that p(i) (0) = 0 for
i = 0, 1, . . . , n − 1, we can use the derivative rule for Laplace transforms to obtain

L{p(n) (t)} L{n!} n!

L{tn } = L{p(t)} = n
= n
= n+1 ,
s s s
which will converge for s > 0.

• Since all polynomial growth is slower than exponential growth, the conditions for convergence of the
Laplace transform for polynomial terms is always s > 0 (any negative exponential component will allow
for convergence).
Solving ODEs using Laplace transforms

Consider the second-order linear constant-coefficient non-homogeneous ODE, with general initial conditions
a2 y ′′ + a1 y ′ + a0 y = f (t), y(0) = y0 , y ′ (0) = y0′ ,
where a0 , a1 , a2 , y0 , y0′ are real constants and f (t) is a known function. Taking the Laplace transform of the
ODE turns all y terms and derivatives of y into Y (s) terms. Applying the Laplace transform, simplifying and
isolating Y (s) gives
F (s) + a2 sy0 + a2 y0′ + a1 y0
Y (s) = .
a2 s2 + a1 s + a0
Taking the inverse Laplace transform of Y (s) will give the general solution to the non-homogeneous ODE in
terms of the general initial conditions.
• This technique can be used to obtain the general solution to any nth-order linear constant-coefficient
non-homogeneous ODE, so long as the Laplace transform of f (t) and the inverse Laplace transform of
Y (s) can be calculated.
• Calculating the inverse Laplace transform of Y (s) typically requires us to factor the denominator and to
make use of partial fractions.

Particular solution via Laplace transforms

Consider the second-order linear constant-coefficient non-homogeneous ODE

a2 y ′′ + a1 y ′ + a0 y = f (t),
where a0 , a1 , a2 are real constants and f (t) is a known function. Using the initial condition y(0) = y ′ (0) = 0
means that the problem will reduce to
1
Y (s) = H(s)F (s), where H(s) = .
a2 s2 + a1 s + a0
The function H(s) is known as the transfer function. A particular solution to the ODE, corresponding to
the initial condition y(0) = y ′ (0) = 0, can hence be calculated so long as F (s) can be calculated and we can
calculate the inverse Laplace transform
yp (t) = L−1 {Y (s)} = L−1 {H(s)F (s)}.

• Choosing this initial condition results in a simple and memorable expression for Y (s). As the homo-
geneous solution can easily be obtained using the auxiliary equation, the general solution can also be
obtained by adding this particular solution.
• As with the general Laplace transform approach, we can also apply this to any nth-order linear constant-
coefficient non-homogeneous ODE, using the initial condition in which the function y(t) and its first n−1
derivatives are all zero.
• Upon applying partial fractions to H(s)F (s) we will be able to calculate the inverse Laplace transforms
of individual terms. However, we can potentially ignore some of the coefficients when calculating partial
fractions if we know they correspond to terms in the homogeneous solutions (these values will be absorbed
into the arbitrary coefficients of the homogeneous solution). This trick can be used to speed up the partial
fractions procedure.
Partial fractions - unique poles

Consider a rational function of the form

P (x) P (x)
R(x) = = ,
Q(x) (x − c1 )(x − c2 ) . . . (x − cm )
where each ci ∈ C, P (x) is an nth-order polynomial, and Q(x) is an mth-order polynomial, such that n < m.
Assuming that each pole ci is unique, we can rewrite the rational function as follows
m m
A1 A2 Am 1 X Y P (x)
R(x) = + + ... + =⇒ Ai (x − cj ) = ,
x − c1 x − c2 x − cm Q(x) Q(x)
i=1 j=1
j̸=i

where each Ai is an unknown constant. The numerators of these expressions must be equal for all choices of
x, and since each ci is unique, we can determine the kth coefficient Ak by setting x = ck giving
m m
X Y P (ck )
Ai (ck − cj ) = P (ck ) =⇒ Ak = ,
(ck − c1 )(ck − c2 ) . . . (ck − ck−1 )(ck − ck+1 ) . . . (ck − cm )
i=1 j=1
j̸=i

with this step being repeatable for all k = 1, 2, . . . , m.

• Due to the fundamental theorem of algebra, we know that the mth order polynomial Q(x) can always be
factored into its m roots. As such, this approach will always work so long as Q(x) has m unique roots.
• The above procedure, in particular choosing x = ck to find coefficient Ak , is effective as it multiplies each
other coefficient by zero, allowing Ak to be easily isolated.
• In cases where the order of P (x) is greater than or equal to the order of Q(x), polynomial division can be
used to rewrite the initial form as R(x) = P1 (x) + PQ(x)
2 (x)
, where P1 (x), P2 (x) are polynomials and P2 (x)
is a lower order polynomial than Q(x).

Partial fractions - complex poles

For a rational function with a pair of complex conjugate poles, c1 = α + βi and c2 = α − βi, we can write
P (x) P (x) P (x)
R(x) = = = ,
Q(x) (x − c1 )(x − c2 ) . . . (x − cm ) ((x − α)2 + β 2 )(x − c3 ) . . . (x − cm )
where each ci ∈ C, P (x) is an nth-order polynomial, and Q(x) is an mth-order polynomial, such that n < m.
Assuming that each pole ci is unique, we can rewrite the rational function as follows
A1 x + A2 A3 Am
R(x) = + + ... +
(x − α)2 + β 2 x − c3 x − cm
where each Ai is an unknown constant. As in the previous case, achieving a common denominator and setting
numerators equal gives an equation that must be true for all x. The coefficients A1 and A2 can then be sought
simultaneously by setting x = α + βi or x = α − βi, and equating real and complex parts (giving two equations
for the two unknowns).
• This procedure is often more useful in practice as factoring in terms of complex roots is unpleasant.
Further, in the context of Laplace transforms, the partial fractions form achieved here is the inverse
Laplace transform of a shifted sine and cosine function, which is often more desirable than a complex
exponential.
Partial fractions - repeated poles

For a rational function with a pole that is repeated r times, c1 = c2 = . . . = cr , we can write
P (x) P (x) P (x)
R(x) = = = ,
Q(x) (x − c1 )(x − c2 ) . . . (x − cm ) (x − c1 )r (x − cr+1 ) . . . (x − cm )
where each ci ∈ C, P (x) is an nth-order polynomial, and Q(x) is an mth-order polynomial, such that n < m.
We can rewrite the rational function as follows
A1 A2 Ar
R(x) = + 2
+ ... + + ...
x − c1 (x − c1 ) (x − c1 )r
where each Ai is an unknown constant, and the remaining terms will take a form depending on whether they
correspond to unique or repeated poles. As in the previous case, achieving a common denominator and setting
numerators equal gives an equation that must be true for all x.

The coefficient Ar can be determined, by setting x = c1 in the numerator equation. To obtain the simplest
equation to find Ar−1 we can take the derivative of the numerator equation and then again choose x = c1 ,
which will give Ar−1 in terms of only Ar . We can repeat taking the derivative and then choosing x = c1 to
obtain each coefficient Ar−2 , Ar−3 , . . . , A1 .
• The process for a large partial fractions problem with repeated roots can be quite tedious and long,
however this procedure can technically be automated.
• In the context of a smaller problem, it may be more convenient to simply choose a different value of x
instead of taking the derivative of the expression.

Heaviside function

The Heaviside function, or unit step function, is given by


1 if t > 0
σ(t) = ,
0 if t < 0
taking the value one when the input is positive and the value zero when the input is negative.

Heaviside functions can be used to write piecewise functions as a single expression, for instance

f1 (t) if t > a  
f (t) = = f2 (t) + f1 (t) − f2 (t) σ(t − a),
f2 (t) if t < a
in which the Heaviside function can be thought of as “turning on” the function f1 (t) and “turning off” the
function f2 (t) at time t = a.
• We might also choose to write the above expression as
 
f (t) = f2 (t) 1 − σ(t − a) + f1 (t)σ(t − a),
in which individual functions are attached to their corresponding on/off timers.
• The Heaviside function is also useful in physical contexts where something occurs quickly relative to the
timescale of the problem being measured. Using the Heaviside function to approximate a quick change
as instantaneous is often simpler in practice.
• The Heaviside function most commonly uses the notation H(x), however in the context of Laplace
transforms, since capital letters represent transformed functions, we try to avoid this.
Laplace transform - Heaviside function (second shifting theorem)

Consider a function f (t) whose Laplace transform exists. Then we obtain the follow Laplace transform rule

L{f (t − a)σ(t − a)} = e−sa L{f (t)}.

In practice, this means that to Laplace transform a function multiplied to a Heaviside function, we must first
recenter the function about t = a. For exponentials we can do this using simple exponential rules

ekt σ(t − a) = eka · ek(t−a) σ(t − a),

for sinusoidal functions, we can do this using the additive trigonometric identities

cos(ωt)σ(t − a) = cos(ωa) cos(ω(t − a))σ(t − a) − sin(ωa) sin(ω(t − a))σ(t − a),

sin(ωt)σ(t − a) = cos(ωa) sin(ω(t − a))σ(t − a) + sin(ωa) cos(ω(t − a))σ(t − a),

and for polynomial terms, we can use the binomial expansion formula

tσ(t − a) = (t − a)σ(t − a) + aσ(t − a),

t σ(t − a) = (t − a)2 σ(t − a) + 2a(t − a)σ(t − a) + a2 σ(t − a),
2

n  
n

n X n k
t σ(t − a) = (t − a) + a σ(t − a) = σ(t − a) a (t − a)n−k .
k
k=0

• In general, it might not be possible to simply recenter a function f (t)σ(t − a) into the form required to
apply the second shifting theorem. The rearrangements shown above are those of our most commonly
used functions, demonstrating that it is possible for each of them.

• For the case where f (t)σ(t), that is where the Heaviside function turns on at a = 0, we find that
L{f (t)σ(t)} = L{f (t)}. This is because Laplace transforms integrate over the interval t ∈ [0, ∞), so
from the perspective of the Laplace transform, every function is zero prior to t = 0.

Inverse Laplace transform - Heaviside function (second shifting theorem)

Consider a function of the form e−sa F (s). To calculate the inverse Laplace transform, we can apply the
inverse of the second shifting theorem by writing

L−1 {e−sa F (s)} = L−1 {e−sa L{f (t)}} = f (t − a)σ(t − a),

in which we have defined F (s) = L{f (t)}. In order to obtain the solution, we require f (t − a) which can be
obtained by solving

f (t) = L−1 {F (s)}.

• Explicitly defining F (s) = L{f (t)} typically makes this problem easier to resolve, as even after finding
f (t), a final shift still needs to be applied to the solution to get the term f (t − a).
Dirac-delta function

The Dirac-delta function is a generalized function or distribution, denoted δ(t), that satisfies the following
properties
ˆ b 
1 if 0 ∈ [a, b]
δ(t) = 0 for t ̸= 0, and δ(t) dt = ,
a 0 otherwise
that is, the Dirac-delta function is zero everywhere except for when its input is zero, and while it takes no
explicit value at zero, integrating the function over an interval that includes zero will give the value one.
• The Dirac-delta function is technically not a function as the value at t = 0 must be unbounded to result
in the finite integral value one given by the second condition. However, we are able to define such an
object by considering it as the limiting case of a sequence of well defined functions. For instance, the
Dirac-delta function could be described as follows
r 
N −x2 N
δ(t) = lim e .
N →∞ π
There are infinitely many potential sequences whose limit could also be used to describe the Dirac-delta
function.
• Thinking of the Dirac-delta function as the limit of the sequence allows us to better think of it as an
impulse, some action that occurs very quickly relative to the measured timescale, but simultaneously
very forcefully.

Relationship between Dirac-delta and Heaviside functions

Consider the following integral of the Dirac-delta function

ˆ T
δ(t − a) dt,
0
for some constants T, a > 0. The region of integration will only include the spike value a if T > a, hence
ˆ T 
1 if T > a
δ(t − a) dt = = σ(T − a).
0 0 if T < a
This observation allows us to rewrite the integral as
ˆ T h iT
δ(t − a) dt = σ(T − a) − σ(0 − a) = σ(t − a) ,
0 0

which takes the form of the fundamental theorem of calculus, implying the relationship
d
σ(t) = δ(t).
dt
• While this relationship will be particularly useful to us, we might realize it is problematic to claim
that the Heaviside function σ(t) is differentiable, given the discontinuity at t = 0. However, if we are
considering the Dirac-delta function as a generalized function, we can also claim this relationship is only
relevant in the same context. For instance
N e−N t N e−N t
 
d 1 1
= , and σ(t) = lim , δ(t) = lim ,
dt 1 + e−N t (1 + e−N t )2 N →∞ 1 + e−N t N →∞ (1 + e−N t )2

in which we see that the derivative of a collection of functions whose sequence converges to the Heaviside
function gives a collection of functions whose sequence converges to the Dirac-delta function.
Laplace Transform - Dirac-delta function

The definite integral of a function f (t) multiplied by the Dirac-delta function with a spike at a over the
interval [t1 , t2 ] is
ˆ t2 
f (a) if a ∈ [t1 , t2 ]
f (t)δ(t − a) dt = .
t1 0 otherwise

Using this, we can calculate the Laplace transform of the Dirac-delta function
ˆ ∞
L{δ(t − a)} = δ(t − a)e−st dt = e−sa ,
ˆ0 ∞
L{f (t)δ(t − a)} = f (t)δ(t − a)e−st dt = f (a)e−sa ,
0

where a > 0.

• The initial integral requires integration by parts and the fact that σ ′ (t) = δ(t) to calculate.

• We might also notice that L{f (t)δ(t − a)} = f (a)L{δ(t − a)}. This makes sense because all values of the
function f (t) will be made zero by the delta function, except for the value that occurs when t = a. That
is, only a single value of f (t) influences the function, namely the constant value f (a).

Constructing arbitrary functions using Dirac-delta functions

Consider the function f (t) defined on t ≥ 0. We can rewrite the function using the following integral
ˆ ∞
f (t) = f (u)δ(u − t) du.
0

This works because for any given point t = t0 and arbitrarily small ϵ, we can write
ˆ t0 +ϵ ˆ ∞
f (t0 ) = f (u)δ(u − t0 ) du = f (u)δ(u − t0 ) du.
t0 −ϵ 0

Essentially, for each value of t ≥ 0, f (u)δ(u − t) is a spike and the area under the spike is the function value
f (t). Since we are integrating over the interval u ∈ [0, ∞), we are guaranteed to find a spike at our chosen
t-value, rather than specifying a different set of integral bounds for each t. It is also possible for us to integrate
over the entire region for any chosen t-value as all other spikes will be ignored.

• This is the continuous analog of superposition, in which integration is used instead of summation.

• In the same way that the Fourier series terms form a basis for all periodic functions, this demonstrates
that the Dirac-delta function forms a basis for all functions.

• While the above integral indicates that we could simply use a small interval around the chosen t-value,
doing so would be of no benefit and simply cause more work due to the inconsistent interval. Further,
when t-values get close to zero, would would have to be careful to have a well defined upper bound.
Particular solutions to Dirac-delta forcing terms

Consider the second-order linear constant-coefficient non-homogeneous ODE

a2 y ′′ + a1 y ′ + a0 y = δ(t − u),

where a0 , a1 , a2 , u are known real constants. Using the initial condition y(0) = y ′ (0) = 0 means that we can
move immediately to our expression for Y (s) which can then be inverse Laplace transformed immediately using
the second shifting theorem

Y (s) = H(s)e−su =⇒ yp (t) = h(t − u)σ(t − u),

where h(t) = L−1 {H(s)}.

• Since h(t) is the inverse Laplace transform of the transfer function, H(s), we expect the terms to be of
the same form as the homogeneous solution.

• Specifically, h(t − u)σ(t − u) represents a solution of the same form as the homogeneous that only starts
at time t = u. This is unsurprising as the Dirac-delta function represents an instantaneous impulse,
essentially interacting with the system for only a single moment. The result is a shifted form of the
natural behaviour of the system.

Solving ODEs using continuous superposition of particular solutions

Consider the set of second-order linear constant-coefficient non-homogeneous ODEs, each corresponding to
a value u ∈ [0, ∞) and whose particular solutions are all known

a2 y ′′ + a1 y ′ + a0 y = f (u)δ(t − u), =⇒ yu (t) = f (u)h(t − u)σ(t − u),

where a0 , a1 , a2 are real constants, h(t) is the inverse Laplace transform of the transfer function, H(s), and
yu (t) is the particular solution to the ODE with forcing term f (u)δ(t − u).

The following second-order linear constant-coefficient non-homogeneous ODE

a2 y ′′ + a1 y ′ + a0 y = f (t),

can have its forcing term rewritten as

ˆ ∞
f (t) = f (u)δ(t − u) du,
0

and hence has a particular solution that is the superposition of the particular solutions for each forcing term
ˆ ∞ ˆ ∞ ˆ t
yp (t) = yu (t) du = f (u)h(t − u)σ(t − u) du = f (u)h(t − u) du,
0 0 0

allowing us to immediately obtain a particular solution for any function f (t) so long as we can calculate the
above integral.
• Essentially, we have resolved the general case of a non-homogeneous ODE by decomposing the forcing
term into point-wise forcing terms occurring at each individual time.
Convolution

The convolution of two functions f (t) and g(t) is defined to be

ˆ t ˆ t
f (t) ∗ g(t) = f (u)g(t − u) du = f (t − u)g(u) du,
0 0

of which both integrals will yield the same result.

• Convolution is a binary operation, much like addition or multiplication, between two functions and
resulting in a new function.

Laplace transforms - convolution

The Laplace transform of a convolution of two functions is

ˆ ∞ˆ t
L{f (t) ∗ g(t)} = f (u)g(t − u)e−st du dt = F (s)G(s),
0 0

where F (s) = L{f (t)} and G(s) = L{g(t)}.

In terms linear constant-coefficient non-homogeneous ODEs, defining the convolution and its Laplace trans-
form allows us to show that

yp (t) = L−1 {Y (s)} = L−1 {H(s)F (s)} = h(t) ∗ f (t),

the particular solution given the initial conditions, y(0) = y ′ (0) = . . . = y (n−1) (0) = 0.

• The Laplace transform of a convolution reveals why it is useful to define, with the Laplace transform
turning convolution of functions into the product of transformed functions.

• Proving the above statement for the Laplace transform of a convolution requires us to change the order
of integration of the statement.