3 Systems

In its simplest form, a system is a mathematical transformation that acts on an input signal to
produce an output signal. We depict a system in block diagram form as follows:

x(t) T y(t)

y(t) = T {x(t)}

Associated with every system is an input-output relation, or recipe, expressed above in the form
y(t) = T {x(t)}. For any given input x(t), this recipe lets you find the corresponding output signal
y(t). Different systems obviously have different input-output relations, and in general two different
systems will produce a different output for the same input.
We’ve already encountered two examples of systems, without directly thinking about them in this
formal context. At the end of the previous section we considered two signals x(t) and y(t) linked
d
by the differentiation property: y(t) = dt x(t). This has a system-level interpretation: consider a
system called a differentiator, which takes a signal x(t) at the input, produces a signal y(t) at the
d
output, and with the input-output recipe y(t) = T {x(t)} = dt x(t):

x(t) Differentiator y(t)

d
y(t) = dt x(t)

If I tell you that the input signal to this system is x(t) = sin(ω0 t), then you can use the input-
output relation to find the corresponding output:
d d
y(t) = T {x(t)} = {x(t)} = sin(ω0 t) = ω0 cos(ω0 t).
dt dt
If the input signal is x(t) = u(t), then the output is

d d
y(t) = T {x(t)} = {x(t)} = u(t) = δ(t).
dt dt
The recipe that governs the action of the system therefore lets you calculate the output for any
given input.
An integrator system can also be defined by means of an input-output relation:

x(t) Integrator y(t)

Rt
y(t) = −∞
x(τ )dτ

Again, for any input x(t) you can use the input-output recipe for the system to find the corre-
sponding output. For example, if the input is x(t) = δ(t) then the output must be
Z t Z t
y(t) = T {x(t)} = x(τ )dτ = δ(τ )dτ = u(t).
−∞ −∞
Alternatively, if the input is x(t) = p1 (t) then the output must satisfy
Z t Z t
y(t) = T {x(t)} = x(τ )dτ = p1 (τ )dτ,
−∞ −∞

which can be evaluated and plotted as follows:

1 y(t)

t
− 21 0
1
2

It is really important to understand that an integrator takes a signal at the input and produces
a signal at the output. This is due to the presence of the time variable t as the upper limit
of the integral: for every value of t the limits of the integral changes, and so does the value of
the integral. Most students, when asked ”What is the inverse operation of differentiation?” will
answer ”Integration”. This is not correct. The inverse of differentiation is indefinite integration.
Indefinite integration, when applied to a function, produces another function, while (in the usual
sense) when you integrate a function you get a value. The difference is critical.
We can also string systems together. The system

Combined system

z(t)
x(t) Integrator Differentiator y(t)

Rt d
as two components, and we can trace the signals through: z(t) = −∞ x(τ )dτ and y(t) = dt z(t).
Since integration and indefinite integration are inverses of one another, this means that y(t) = x(t).
Considered as a whole, we can therefore think of this as a combined system with input x(t) and
output y(t). The input-output relation for this combined system is y(t) = T {x(t)} = x(t), which
produces at the output exactly the same signal as appeared at the input.

3.1 Differential equations as input-output relations

We can use the notion of a system to represent the relationship between signals in a physical
setting. For example, the circuit below contains a current source driving a parallel combination
of a resistor and a capacitor:

iC (t) = y(t)
iR (t)

x(t) = i(t)
C
R

Suppose we are interested in the signal corresponding to the current through the capacitor as we
change the current produced by the source. These are functions of time, and can be thought of as
signals. In this case we can think of the circuit as a system: the input x(t) is the input current
i(t), and the output y(t) is the capacitor current iC (t). The laws of physics (expressed as the
theory of electrical circuits) let us find the output signal for any given input signal.

3-2
The voltage-current relationship for a resistor can be expressed in terms of signals as v(t) = RiR (t),
d
and the relationship for a capacitor is iC (t) = C dt v(t). Note that the voltage signal across the
two circuit elements is equal, and is given by v(t). We can eliminate v(t) from these expressions,
d d
yielding iC (t) = RC dt iR (t). However, we know that iR (t) = i(t) − iC (t), so iC (t) = RC dt (i(t) −
iC (t)). This provides a relationship between i(t) and iC (t):

d d
iC (t) + RC iC (t) = RC i(t).
dt dt
Using our definitions of input and output signals, the input and output of the system must obey
the relation
d d
y(t) + RC y(t) = RC x(t).
dt dt
This is just a recipe for finding the output signal for any given input signal, which is all that is
required to specify a system. If I tell you that the input is x(t) = u(t), then you know that the
output must satisfy
d d
y(t) + RC y(t) = RC u(t),
dt dt
which is just a differential equation with one unknown y(t). Assuming you know how to solve
1
differential equations, you could solve for the output as y(t) = e− RC t u(t), which for RC = 1 is
shown below:

1
Output y(t)

0.5

0
-2 -1 0 1 2 3 4
t

After the step the current initially all flows through the capacitor, which begins charging. Once
charged, all the current flows through the resistor so the output signal goes to zero.
There is one subtlety in this example. The input-output relation for the system just described is
a first-order differential equation in the output y(t), so the solution is only defined up to a single
unknown integration constant. An auxiliary condition is required in order to specify the output
y(t) uniquely for a given input x(t). In the description given above the initial condition was that
prior to t = 0 the capacitor held no charge — this is often called the initial rest condition. At this
stage we do not concern ourselves with the technical details.
The power of thinking at a system level comes from abstraction. Once you (or somebody else)
have done the physical modelling it no longer matters that the system represents an electrical
circuit. You have a mathematical model, expressed in terms of an input-output relation, which
for all practical purposes replaces the physical system with a mathematical representation. Thus
you can work with signals and systems without having to think about details that are irrelevant
to solving the problem.

3.2 System properties

The most general form for an input-output relation for a system is y(t) = T {x(t)}. This expression
indicates that the system is characterised by a transformation T {·}, which acts on the input to
produce the output.

3-3
It usually makes things much clearer if you think about a signal as an object, rather than as a
set of values of a function for different instants in time. The expression y(t) = T {x(t)} can then
be interpreted as follows: the system takes the object x(t), applies the transformation T to it, to
produce the object y(t).
You’ve seen this way of thinking with vectors and matrices. If I were to say that x and y were
vectors (say in Rn ), and that they are linked by the matrix A in the relation y = T {x} = Ax, this
would be the natural interpretation. The transformation T is now represented by the matrix A,
which acts on the object x to produce the object y by matrix-vector multiplication. In reality the
objects x and y are (ordered) collections of numbers, but as far as the structure of transformations
is concerned this fact is not particularly useful. Exactly the same is true for signals1 .
In practice the general transformation y(t) = T {x(t)} is not particularly useful: the set of all
possible transformations that can be applied to x(t) to produce y(t) is too big to permit a useful
mathematical theory. We therefore have to restrict the set of possible transformations that we
consider, and this corresponds to making assumptions about the properties of the systems that
they represent.

3.2.1 Causal systems

In the real world, physical systems always exhibit a property called causality: the state of an
object can only depend on things that happened to it in the past. A soccer ball cannot know that
somebody is going to kick it tomorrow, so its position today can’t in any way depend on what is
going to happen to it in the future. Put another way, the cause of something always precedes the
effect.
Suppose I think of the soccer ball as a system, and assume that it is only free to move in a straight
line. I can put a coordinate system down along this straight line (with an arbitrary origin), and at
any instant in time I can express the position of the ball with respect to this coordinate system.
The position of the ball at any time can be expressed by the value of a function y(t) at that
instant. The signal y(t) then represents the position of the ball at every possible time instant,
and I consider this to be the output of the system. The input x(t) is the force that I apply to
the ball, in the direction that it is free to move, expressed as a function of time. Clearly we now
have a physically-realisable system, where the input signal is the force applied to the ball and the
output is its position. Using Newton’s laws of motion I could derive a relationship linking input
to output.
I could certainly drive this system with the input signal x(t) = u(t), which contains a step change
in the force applied at time t = 0. But what do we mean by t = 0? And what do we mean by
negative time? The answer is that, as was the case with the position axis before, we can put the
origin anywhere we like — but once we’ve fixed what we consider to be t = 0 then this is the
time origin for all our signals. I could, for example, define time t = 0 to be ”now!”, in which case
yesterday would be on the negative time axis and tomorrow would be on the positive axis (as long
as I put the positive time direction going into the future, which is the convention).
Returning to the soccer ball example with x(t) = u(t), I denote by t = 0 the instant at which I
start applying the force. For all time prior to t = 0, no force has ever been applied to the ball. It
is reasonable, therefore, to assume initial rest conditions for the ball — it carries no energy and
isn’t moving. For simplicity we may as well place the (arbitrary) origin of the position axis at the
position of the ball for negative time — in this case we would have y(t) = 0 for t < 0. However,
as soon as the force is applied the ball starts moving, so we would be quite satisfied if someone
told us that an input-output pair for the system looked as follows:
1 This is more true than it might appear. The abstract mathematical construct of a vector space is entirely

appropriate for signals and systems: the structure of the mathematics on ”ordinary” vectors is exactly the same as
the structure of the mathematics for signals. In this context a function (or a signal) x(t) is a vector, and it really
does no harm thinking of it as such.

3-4
 y(t)

1 x(t)

t t
0 0

But what if someone told you that the input-output pair was

y(t)

1 x(t)

t t
0 0

This would not be plausible — the ball started moving (the effect) before the force was applied
(the cause). The response in the second case couldn’t be the output of a causal system.
Since the behaviour of a system is determined by its input-output relation, it must be possible
to examine this relation in order to determine if the system is causal. The easiest way of doing
this is to fix a value of t, say t = t0 , and think about calculating the output value y(t0 ) using the
input-output relation. Ask yourself: ”In order to find the value of y(t0 ), what values of the input
x(t) do I need to know?” This answer defines a set of values of t for which you need to know x(t).
If you only need to know x(t) at instants in time that occur earlier than t = t0 , then causality is
indicated. In practice you need to ask this question for every possible value of t0 in order to prove
causality.
Here’s an example. Suppose the input-output relation for a system is given by
Z 1
y(t) = x(t − τ )dτ.
−1

Is it causal? To get some insight, consider finding y(0) = y(t)|t=0 . According to the input-output
R1 R1 R1
relation, y(0) = −1 x(0 − τ )dτ = −1 x(−τ )dτ = −1 x(p)dp. To calculate this integral we need
to know x(t) over the entire integration interval [−1, 1]. This interval includes instants later than
t = 0. Thus the system is not causal.
Is the system represented by the input-output relation
Z 1
y(t) = x(t − τ )dτ
0
R1
causal? In this case y(0) = 0 x(0 − τ )dτ , so we need to know x(−τ ) over the interval [0, 1].
Equivalently, we need to know x(t) over the interval [−1, 0], all of which are in the past. For
t = 0, at least, the system is exhibiting the required condition for causality. To prove that it for
all values of t, do the change of variables p = t − τ , where t is a constant in the integration. The
input-output relation is then Z t
y(t) = x(p)dp.
t−1

Clearly, to find the value of the output at time t we need to know the values of the input over the
interval [t − 1, t], and none of these inputs are in the future. The system is therefore causal.
Causality is a property of physical systems, and if we construct models of such systems then it is
sensible to require that the models are causal. However, the mathematical theory of signals and
systems doesn’t really differentiate between models that are causal and those that are not: working
with a causal system isn’t really any easier than working with a non-causal system. There are other

3-5
properties that a system can have for which this is absolutely not the case: certain systems have
properties that make it much easier to work with them. The remainder of this section discusses
some of these properties.
Finally, it is worth noting that while causality is important in modelling a physical system, non-
causal systems are both important and useful. Suppose someone gives you a DVD full of data
obtained by sampling the swell size of the ocean in False Bay over a period of 100 years, and you
want to process this data to highlight important events. You could design a system (or a filter)
that takes this data as an input x(t), and transforms it to an output y(t) that is more informative
for your purposes. There’s no good reason to require that the system you design be causal — all
the data is available, and it could do a better filtering job if the output at any point depended on
input values that are both in the future and the past relative to that point. Another example is in
image processing: in an image there is no time axis, so causality has no meaning and no relevance.

3.2.2 Linear systems

Let f (x) be a real-valued function, and let x and y be real-valued variables: f is homogeneous if
for any constant a we have f (ax) = af (x) for all x, and f is additive if f (x + y) = f (x) + f (y)
for all x and y. If these two properties hold then the function f is said to be linear, and for any
constants a and b it must be true that f (ax + by) = af (x) + bf (y) for all x and y. Clearly a linear
function is a very restricted form.
Systems can also be linear, and the requirements are similar to those just outlined for functions. A
linear system has a very particular structure, and behaves nicely with regard to sums and scalings
of signals. Systems that are linear have very tractable mathematics.
Consider again the canonical system

x(t) T y(t)

y(t) = T {x(t)}

Homogeneity and additivity (and therefore linearity) of the system are related to the behaviour
of the transformation T . If T {ax(t)} = aT {x(t)} for all signals x(t) and any constant a, then the
system is homogeneous. If T {x(t) + y(t)} = T {x(t)} + T {y(t)} for all signals x(t) and y(t), then
the system is additive. The system is linear if it is both additive and homogeneous, in which case
for any constants a and b we must have T {ax(t) + by(t)} = aT {x(t)} + bT {y(t)} for all signals
x(t) and y(t).
These statements are cryptic — what do they mean? Let’s consider a system that is homogeneous,
and consider driving it with the input signal x(t) = x1 (t). The output will be y(t) = y1 (t) =
T {x1 (t)}, so x1 (t) and y1 (t) are a valid input-output pair for the system:

x1 (t) −→ y1 (t).

Suppose now that we drive it with the input x(t) = ax1 (t), which is just the signal x1 (t) scaled by
a constant a. If the system is homogeneous then the output will be y(t) = T {x(t)} = T {ax1 (t)} =
aT {x1 (t)} = ay1 (t), so the following will also be a valid input-output pair for the system:

ax1 (t) −→ ay1 (t).

Note that a can be any constant value. This property is saying that, for an homogeneous system,
scaling the input by some fixed constant just changes the scaling of the output by the same
constant — increasing the amplitude of the signal at the input causes the amplitude of the signal
at the output to change by the same amount. An homogeneous system is easy to characterise

3-6
in terms of multiplicative scaling. The overall property can be summarised as below, in terms of
input-output pairs:

Homogeneous:

If x1 (t) −→ y1 (t)
then ax1 (t) −→ ay1 (t) for all a.

In a similar manner, consider any two sets of valid input-output pairs: y1 (t) = T {x1 (t)} and
y2 (t) = T {x2 (t)}. If a system exhibits the additivity property then y1 (t)+y2 (t) = T {x1 (t)+x2 (t)}
must also be a valid input-output pair. It doesn’t matter whether we add to signals and then put
the result through the system, or if we put the two signals through the system and then add
the result. Perhaps more importantly, additivity allows us to decompose an input signal into
components, and think about what happens when each of these components passes through the
system. Summarising this property, if we have two input-output pairs for the system, and we
know that it is additive, then we can deduce a third:

Additive:

If x1 (t) −→ y1 (t) and x2 (t) −→ y2 (t)

then x1 (t) + x2 (t) −→ y1 (t) + y2 (t).

If a system is both homogeneous and additive then it is linear. The building block of the
mathematics of linear systems is linear combinations (or weighted sums) of signals. The sig-
nal x(t) = ax1 (t) + bx2 (t) is called a linear combination of the two signals x1 (t) and x2 (t): the
two components are combined using the addition operation, with weights that determine the pro-
portion of each. Different values of a and b lead to a different combined signal, but all the signals
that can be generated in this way are basically built up from the two component signals x1 (t) and
x2 (t).
Given two valid input-output pairs y1 (t) = T {x1 (t)} and y2 (t) = T {x2 (t)}, a linear combination
of these input signals will yield an equivalent linear combination of the output signals if the system
is linear:

T {ax1 (t) + bx2 (t)} = T {ax1 (t)} + T {bx2 (t)} (using additivity)
= aT {x1 (t)} + bT {x2 (t)} (using homogeneity)
= ay1 (t) + by2 (t).

In general, given N valid input-output pairs yi (t) = T {xi (t)} and N scalar values ai for i =
PN PN
1, · · · , N , for a linear system it will be the case that T { i=1 ai xi (t)} = i=1 ai yi (t). This is
often referred to as the principle of superposition for linear systems. A linear system is easy to
characterise in terms of linear combinations (or superpositions) of signals.
Given two valid input-output pairs for a linear system, we can deduce a whole class of valid
input-output pairs by using different linear combinations of them:

Linear:

If x1 (t) −→ y1 (t) and x2 (t) −→ y2 (t)

then ax1 (t) + bx2 (t) −→ ay1 (t) + by2 (t) for all a and b.

The simplest linear system is an ideal amplifier, which obeys the input-output relation y(t) =
Kx(t) for some constant K usually greater than one. To prove that a system is linear we need

3-7
to show that homogeneity and additivity hold for all possible input signals — this is harder than
proving nonlinearity, where a single counterexample would suffice.
Assume that we have two valid input-output pairs for the system, namely y1 (t) = T {x1 (t)} =
Kx1 (t) and y2 (t) = T {x2 (t)} = Kx2 (t). We haven’t specified x1 (t) or x2 (t), so these could be any
possible signals: all we know is that once they are specified, the corresponding outputs y1 (t) and
y2 (t) are completely determined. Consider now driving the system with the input x(t) = ax1 (t)
for some value of a: the output will be y(t) = Kx(t) = Kax1 (t) = aKx1 (t) = ay1 (t). Since this
holds for all a and any x1 (t), the ideal amplifier is homogeneous.
Now consider the input x(t) = x1 (t) + x2 (t). The output will be

y(t) = Kx(t) = K(x1 (t) + x2 (t)) = Kx1 (t) + Kx2 (t) = y1 (t) + y2 (t).

Since x1 (t) and x2 (t) can be anything, this proves additivity. The system is therefore additive and
homogeneous, and is therefore linear.
One could also prove linearity directly by considering the input signal x(t) = ax1 (t) + bx2 (t) for
some a and b. According to the input-output recipe for the system the output will be

y(t) = Kx(t) = K(ax1 (t) + bx2 (t)) = aKx1 (t) + bKx2 (t) = ay1 (t) + by2 (t).

This holds for all a and b and for any x1 (t) and x2 (t), so linearity has been shown directly.
Thus the system with the input-output relation y(t) = Kx(t) is linear. What about the system
where y(t) = Kx(t) + 1? This also feels like a ”straight-line” relationship between input and
output, so it might seem likely to be linear. The best way to understand the action of a system
is to put some example signals through it. For definiteness let’s assume K = 2, and consider the
response to x1 (t) = u(t): the output will be y1 (t) = 2x1 (t) + 1 = 2u(t) + 1. Thus

u(t) −→ 2u(t) + 1

is a valid input-output pair for the system. If the system is linear then it must be homogeneous,
so
2u(t) −→ 2(2u(t) + 1) = 4u(t) + 2
would have to be a valid input-output pair. However, it is easy to see that the response of the
system to x2 (t) = 2u(t) is y2 (t) = 2x2 (t) + 1 = 2(2u(t)) + 1 = 4u(t) + 1, so it cannot be valid.
Therefore the system is not homogeneous and consequently not linear. [add some diagrams here,
and explicitly include examples]

3.2.3 Time-invariant systems

The final property we consider that a system can have is that of time invariance. A time-invariant
system is one that behaves the same way today as it did yesterday, or as it will tomorrow. If a
system is time invariant, then the following is true with regard to input-output pairs:

Time invariant:

If x1 (t) −→ y1 (t)
then x1 (t − c) −→ y1 (t − c) for all c.

What we see then is that shifting the input to a time-invariant system just causes an equivalent
shift in the output.
As a practical example, suppose you’re using a simple electrical RC circuit in the lab. You drive
it with a step input, and it so happens that the step in the value happens at exactly 3pm. You
observe that the output starts to change at 3pm, and traces a particular curve. If you were to

3-8
repeat this experiment one hour later, the step in the value at the input would occur at 4pm.
However, the circuit is time invariant, so you would expect to see exactly the same signal as
before, but it will just start at 4pm. Delaying the input by one hour just causes the output to be
delayed by one hour.
A soccer ball with a slow leak could be considered to be a system that is not time invariant: if I
kick it tomorrow it would probably behave very differently from if I kick it today.
Another way of thinking about time invariance is that it corresponds to a system that doesn’t
care where we place the t = 0 origin in our signals. Sure, the input and output signals must share
the same origin (or we wouldn’t know the relative time between a signal going in to a system and
the system’s response), but the position of this origin is arbitrary. Effectively, we could take any
input-output pair for a time-invariant system and move the position of the origin by the same
amount in both the input and the output, and the result will also be a valid input-output pair for
the system.
R1
Example: Is the system with input-output relation y(t) = 0 x(t − τ )dτ time invariant? To get
some insight, it is generally useful to choose some ”easy” input and find the response, and then
to repeat for a shifted version of the input. Consider the input x1 (t) = δ(t): the output will be
R1
y1 (t) = 0 δ(t − τ )dτ . In this integral the variable is τ and t is a fixed constant, so the signal being
integrated corresponds to a delta function at τ = t. The integration interval is [0, 1], which will
include the impulse as long as 0 ≤ t ≤ 1. Thus the output is
(
1 0≤t≤1
y1 (t) =
0 otherwise
Consider now the input x(t) = x1 (t − 1) = δ(t − 1), which is just the previous input delayed by one
R1 R1 R1
unit. The output will be y(t) = 0 x(t−τ )dτ = 0 x1 (t−τ −1)dτ = 0 δ(t−τ −1)dτ . The impulse
is now at position t − 1 on the τ axis, so the output will be nonzero as long as 0 ≤ t − 1 ≤ 1, or
equivalently 1 ≤ t ≤ 2. The output is therefore seen to be the same as the previous output, but
with a delay of one time unit as required.
It therefore seems plausible that the system is time invariant. To prove it, we need to show that
the property holds for all possible signals and all shifts. This is quite easy: consider R 1 an arbitrary
input-output pair y1 (t) = T {x1 (t)}. From the system recipe we can write y1 (t) = 0 x1 (t − τ )dτ .
Now consider the shifted input x2 (t) = x1 (t − c), where c is some fixed value. The output in this
case is given by
Z 1 Z 1 Z 1
y2 (t) = x2 (t − τ )dτ = x1 (t − τ − c)dτ = x1 ((t − c) − τ )dτ = y1 (t − c),
0 0 0
which is seen to be an equally shifted version of the original output. Since this is true for all c
and x1 (t) was arbitrary, the system is time invariant.

3.3 Linear time-invariant systems

Systems that exhibit both linearity and time invariance are called linear time-invariant systems
(often just called LTI systems). All of the theory of signals and systems relates to systems that
are LTI, and if a system is not LTI then we don’t really have the mathematics to deal with it.
Any system that is governed by a differential equation linking output y(t) to input x(t) is linear.
For such a system the input-output relationship can be written as
N M
X di X dj
ai (t) y(t) = b j (t) x(t),
i=1
dti j=1
dtj

di
where ai (t) (for i = 1, . . . , N ) and bj (t) (for j = 1 . . . , M ) are functions of time, and dti denotes
the ith derivative with respect to time.

3-9
 Furthermore, if ai (t) = ai and bj (t) = bj are constant (i.e. not functions of time) then the system
is LTI. That is, if the input-output relation for a system can be written in the form
N M
X di X dj
ai y(t) = bj j x(t),
i=1
dti j=1
dt

then it is an LTI system. The above expression is called a linear constant coefficient differential
equation, or LCCDE. Many physical systems are governed by LCCDEs, and the mathematics
of signals and systems are appropriate for them. For example, any electrical circuit made up
of resistors, capacitors, and inductors (any number of them, connected in any configuration) is
governed by a LCCDE, and can therefore be modelled and analysed using systems theory.

3.4 Impulse response for LTI systems

Using the sifting property of the Dirac delta function, the following identity can be proved:
Z ∞
x(τ )δ(t − τ )dτ = x(t).
−∞

Exercise: Plot x(τ )δ(t − τ ) and x(t)δ(t − τ ) as functions of τ , and conclude that the above identity must be
true for all t.
Using the properties of linearity and time invariance, a general form for the output y(t) of a LTI
system for a given input x(t) can be obtained. The only thing we need to know is the output of
the system when the input is a Dirac delta function: this is called the impulse response of the
system and is often denoted h(t). Specifically, if x(t) = δ(t), then the input-output recipe for the
system can be used to find y(t) = h(t). For a LTI system, if we only know the input-output pair
δ(t) −→ h(t),
then a general expression for the input-output relation follows:
Z ∞ 
y(t) = T {x(t)} = T x(τ )δ(t − τ )dτ
−∞
Z ∞
= T {x(τ )δ(t − τ )}dτ using additivity
−∞
Z ∞
= x(τ )T {δ(t − τ )}dτ using homogeneity
−∞
Z ∞
= x(τ )h(t − τ )dτ using time invariance.
−∞

This expression can be used to find the output y(t) for any given input x(t) as long as the impulse
response h(t) for the system is given or known.

RA∞LTI system therefore also has an input-output relation that can be written in the form y(t) =
−∞ x(τ )h(t − τ )dτ . This is mathematical operation is called convolution, and is denoted y(t) =
x(t) ∗ h(t). We could say that for a LTI system the output signal is the convolution of the input
signal with the impulse response for the system. Knowing h(t) completely determines the behaviour
of such a system.
To highlight the importance of the impulse response of a linear time invariant system, we often
denote the system itself by its impulse response:

x(t) h(t) y(t)

3-10
The implication is that in this case the signals are related by y(t) = x(t) ∗ h(t). Note that the
impulse response only characterises the system completely if it is LTI: a system that is not LTI
does not have an impulse response.

3.5 Convolution
As defined, the mathematical process of convolution takes two signals x(t) and h(t), and from
these produces an output signal y(t) according to the relation
Z ∞
y(t) = x(τ )h(t − τ )dτ.
−∞

Suppose you’re given x(t) and h(t), and want to find the result of the convolution. Essentially
what you want is a plot of y(t) versus time for all possible values of t. Since y(t) is just a function,
we can obtain the value y(t) for say t = 3 by calculating
Z ∞
y(3) = x(τ )h(3 − τ )dτ.
−∞

In this expression τ is an integration variable that is eliminated in the integration, resulting in a

number that we call y(3). This gives us one point on our y(t) plot — the value for t = 3. According
to the formula, to find this value we need to plot x(τ ) and h(3 − τ ) as functions of τ , find the
product x(τ )h(3 − τ ) as a function of τ , and then calculate the total area under this resulting
function. In principle to plot the whole of y(t) we need to repeat this process for every value of t.
Suppose for example that x(t) = p10 (t) (a centered pulse of total width 10) and h(t) = u(t − 1):

1 x(t) h(t)
1

t t
−5 5 1
0 0

The required quantities for calculating y(3) are

1
x(τ )

τ
−5 5
0
1
h(3 − τ )

τ
2 3
0
1
x(τ )h(3 − τ )

τ
−5 2
0

Since the total area under x(τ )h(3 − τ ) is 7, we conclude that y(3) = 7.
To find the output y(t) = x(t) ∗ h(t) we need to repeat the above process for all possible values of
t. Noting that plotting h(t − τ ) as a function of τ involves a flip around the origin, and a shift of
the origin to position t, we can generally denote it as follows:

1
h(t − τ )

τ
t−1 t
0

3-11
In forming the product x(τ )h(t − τ ) there are three different cases to consider:
t − 1 < −5 −5 ≤ t − 1 ≤ 5 t−1>5
1 1 1
x(τ )

τ τ τ
−5 5 −5 5 −5 5
0 0 0

h(t − τ )

τ τ τ
t−1 t−1 t−1
0 0 0
1 1 1
x(τ )h(t − τ )

τ τ τ
−5 5 −5 t−1 −5 5
0 0 0

For t − 1 < −5, or t < −4, the total area under the product is zero and y(t) = 0 over this range.
For −5 ≤ t − 1 ≤ 5, or −4 ≤ t ≤ 6 the total area under the product is t − 1 + 5, so y(t) = t + 4
over this range. For t − 1 > 5, ort > 6, the total area is 10 and we have y(t) = 10 in this case.
The required convolution can therefore be written in the form

0
 t < −4
y(t) = t + 4 −4 ≤ t ≤ 6

10 t > 6,


so the output is the signal shown below:

10 y(t)

t
−4 0 6

For more complicated signals it may be necessary to use calculus and algebra to find the resulting
output. For example, suppose we are given the signals x(t) and h(t) below, and want to find
y(t) = h(t) ∗ x(t):
1
e−2(t− 2 )
et 1 e−2t 1

x(t) h(t)

t 1
t
−1 1 2 2
0 0
R∞
We flip and shift the ”easier” signal, so use the form y(t) = −∞ x(τ )h(t − τ )dτ . The functions
that need to be multiplied and integrated are expressed as a function of τ are as follows:

eτ 1 e−2τ

x(τ )

τ
−1 1
0

1
e−2((t−τ )− 2 )

h(t − τ )

1
τ
t−2 t− 2 t
0

3-12
As t varies the function h(t − τ ) changes position along the τ axis. To find y(t) we need to find
the area under the product x(τ )h(t − τ ) for each value of t.
For simplicity, let’s consider just finding the value y(1). The signals in the convolution integral
are roughly as drawn above (with t = 1), so the product is only nonzero over the range −1 to
1 − 21 . The required output is therefore
1 1
Z 2
Z 0 Z 2
y(1) = x(τ )h(t − τ )dτ = x(τ )h(t − τ )dτ + x(τ )h(t − τ )dτ
−1 −1 0
1
Z 0 Z 2
1 1
= eτ e−2((t−τ )− 2 ) dτ + e−2τ e−2((t−τ )− 2 ) dτ,
−1 0

where the integral is split into two because the signal x(τ ) is defined differently over two regions
of the domain.
In general, the convolution will take the same form as above as long as t − 12 ≥ 0 and t − 2 ≤ −1:
in this case the nonzero region of the product will be from −1 to t − 21 , with some portion of the
positive τ domain included in this region. The required convolution values will be
Z t− 21 Z 0 Z t− 21
y(t) = x(τ )h(t − τ )dτ = x(τ )h(t − τ )dτ + x(τ )h(t − τ )dτ
−1 −1 0
Z 0 Z t− 12
1 1
= eτ e−2((t−τ )− 2 ) dτ + e−2τ e−2((t−τ )− 2 ) dτ.
−1 0

1
This gives the required output for any 2 ≤ t ≤ 1.
Finding the remainder of the convolution output is laborious but conceptually identical. Critical
values of t are defined by the following: t − 21 = −1, t − 21 = 0, t − 12 = 1, t − 2 = −1, t − 2 = 0,
and t − 2 = 1. Ordering these critical points yields the following values of t: − 21 , 12 , 1, 32 , 2, 3.
The calculation of the convolution will therefore require separate consideration of values of t in
intervals (−∞, − 12 ], [− 21 , 12 ], [ 12 , 1], [1, 32 ], [ 23 , 2], [2, 3], [3, ∞). In each case the integral will differ
either in terms of the integration limits or in the specification of the integrand.

3.6 Convolution properties

The convolution operation has many properties, which provide insight into the operation itself as
well as into the action of LTI systems on input and output signals. A list of basic convolution
properties follows:

x(t) ∗ v(t) = v(t) ∗ x(t) (commutative)

x(t) ∗ (v(t) ∗ w(t)) = (x(t) ∗ v(t)) ∗ w(t) (associative)
x(t) ∗ (v(t) + w(t)) = x(t) ∗ v(t) + x(t) ∗ w(t) (distributive over addition)
x(t) ∗ δ(t) = x(t) (identity element)
x(t) ∗ δ(t − c) = x(t − c) (shifted identity).

These properties can all be derived directly from the definition of convolution. For example,
commutativity states that x(t) ∗ v(t) = v(t) ∗ x(t) and is quite easy to prove via a change of
variables (letting p = t − τ below):
Z ∞ Z ∞ Z ∞
x(t) ∗ v(t) = x(τ )v(t − τ )dτ = x(t − p)v(p)dp = v(p)x(t − p)dp = v(t) ∗ x(t).
−∞ −∞ −∞

Similarly, the shifted identity property follows from the sifting property: since by definition
Z ∞
x(t) ∗ v(t) = v(τ )x(t − τ )dτ,
−∞

3-13
if we let v(t) = δ(t − c) we get:
Z ∞ Z ∞
x(t) ∗ δ(t − c) = δ(τ − c)x(t − τ )dτ = δ(τ − c)x(t − c)dτ
−∞ −∞
Z ∞
= x(t − c) δ(τ − c)dτ = x(t − c).
−∞

Some of the properties can also be justified using basic theory for linear systems. For example,
the distributive property x(t) ∗ (v(t) + w(t)) = x(t) ∗ v(t) + x(t) ∗ w(t) is obvious if we think of a
system with impulse response x(t) being driven by the signal v(t) + w(t):

v(t) + w(t) x(t) y(t)

The output will be y(t) = x(t) ∗ (v(t) + w(t)). However, we know that the following input-output
pairs are valid:
v(t) −→ x(t) ∗ v(t)
w(t) −→ x(t) ∗ w(t).
Since the the system is LTI we know that additivity holds, so it must also be the case that
y(t) = x(t) ∗ v(t) + x(t) ∗ w(t), and the property is seen to be true. The identity and shifted
identity properties follow similarly: if the input is δ(t) then we know that the output must be the
impulse response x(t), so x(t) ∗ δ(t) = x(t); if the input is the shifted impulse δ(t − c) then the
output must be the shifted impulse response x(t − c), so we have x(t) ∗ δ(t − c) = x(t − c).
Consider a series combination or cascade of two LTI systems, as shown below:

z(t)
x(t) f (t) g(t) y(t)

heff (t)

Since z(t) = f (t) ∗ x(t) and y(t) = g(t) ∗ z(t), the overall input-output relation for the combined
system is y(t) = g(t) ∗ (f (t) ∗ x(t)). The associative property allows us to instead write
y(t) = (g(t) ∗ f (t)) ∗ x(t) = heff (t) ∗ x(t),
where heff (t) = g(t) ∗ f (t) is the overall effective impulse response of the combined system. From
commutativity the effective system is exactly the same as the one below:

x(t) g(t) f (t) y(t)

heff (t)

Thus the order of LTI systems arranged in a series combination can be changed without affecting
the overall system response.
Other properties can also be derived that relate inputs and outputs of a convolution. Assuming
that y(t) = x(t) ∗ v(t), the following are also true:

3-14
 y(t − c) = x(t − c) ∗ v(t) = x(t) ∗ v(t − c) (shifting)
d d d
y(t) = dt x(t) ∗ v(t) = x(t) ∗ dt v(t) (derivative)
Rt R tdt Rt
−∞
y(τ )dτ = −∞
x(τ )dτ ∗ v(t) = x(t) ∗ −∞
v(τ )dτ (integration).

The derivative property can be obtained directly from the definition

Z ∞
y(t) = x(t) ∗ v(t) = x(τ )v(t − τ )dτ.
−∞

Taking the derivative with respect to time on both sides gives

Z ∞  Z ∞   Z ∞
d d d
y(t) = x(τ )v(t − τ )dτ = x(τ ) v(t − τ ) dτ = x(τ )v̇(t − τ )dτ,
dt dt −∞ −∞ dt −∞

so ẏ(t) = x(t) ∗ v̇(t) (where the dot above the signal indicates the generalised time derivative).
One form of the shifting property follows again from the interpretation of y(t) = x(t) ∗ v(t) as
the input-output relation for a system with impulse response x(t) driven by the signal v(t). Since
the system is time invariant (or it wouldn’t have an impulse response and convolution would be
meaningless), the response to the shifted input v(t − c) must be y(t − c) and it must always be
true that y(t − c) = x(t) ∗ v(t − c).
The integration property can be justified using basic properties of LTI systems. Consider the
cascade below:
Integrator
y(t)
x(t) v(t) g(t) w(t)

Rt
The input-output relation for the integrator is w(t) = −∞ y(τ )dτ , which can be shown to satisfy
the requirements for linearity and time invariance. (The system therefore does have an impulse
response g(t), but we’re not making use of this fact here.) Also, y(t) = v(t) ∗ x(t).
Since both component systems are LTI, from an overall input-output perspective the combined
system above is equivalent to

Integrator
z(t)
x(t) g(t) v(t) w(t)

In this case Z t
w(t) = v(t) ∗ z(t) = v(t) ∗ x(τ )dτ.
−∞

For the same input x(t) the output w(t) must be the same for both cases shown, so we must have
Z t Z t
y(τ )dτ = v(t) ∗ x(τ )dτ
−∞ −∞

as long as y(t) = v(t) ∗ x(t).

Exercise: The integrator system (also called an accumulator) has an impulse response g(t) = u(t). Assume
this to be true and use the definition of convolution to show that the system has the required
input-output relation.

3.7 Example: smoothing a rectangular pulse train

3-15