Mathematical Modelling

What is a Mathematical Model?

Well, ask a dozen different people and you will get a dozen different answers. Here are a couple
of definitions that I have come across:

• A mathematical model is a representation of reality (a rather loose definition).

• A mathematical model is an abstract, simplified, mathematical construct related to a

part of reality and created for a particular purpose (E.A. Bender, An introduction to
mathematical modelling, Dover, USA (2000)) (a more formal definition).

Perhaps you could type the phrase “mathematical model” into Google and see what other
people think a mathematical model is. Try consulting Wikipedia for “mathematical model”.

What makes mathematical models useful?

If we use the language of mathematics to formulate our mathematical models then

1. We must formulate our ideas precisely and so are less likely to let implicit assumptions
slip by.

2. We have a concise language which encourages manipulation.

3. We have a large body of theory available to us.

4. We have high speed computers available for carrying out calculations.

There is a trade-off between points 3 and 4 because theory is good for drawing general conclu-
sions from simple models, whereas computers are useful for drawing specific conclusions from
complicated models.

Properties of Models
When we begin to formulate a model, we need to divide the universe into three distinct parts:

1. Things whose effects are neglected

2. Things that affect the model but whose behaviour the model is not designed to study

3. Things the model is designed to study the behaviour of

In other words, the model completely ignores item 1. Item 2 consists of constants, functions
and so on that are external to the model and these are called parameters, input or independent
variables. More formally, they are called exogenous variables. Item 3 consists of the things that

1
the model is trying to explain and these are called endogenous variables or output or dependent
variables.
So we have three types: neglected, input and output and all of these are important in
modelling. For example, if the wrong things are neglected, the model will be no good. On the
other hand if too many things are included, the model can be very complex, require a large
amount of data input and can be incredibly time consuming to solve. Also, a proper choice of
output is essential, we must seek to explain the things we can explain.
Incorporating all these aspects can make modelling a reasonably difficult procedure. Differ-
ent types of simplifying assumptions (ie the things we choose to ignore) usually lead to different
models and so there is no single best model for describing a situation. In some situations one
is often content with a statement that something will increase or decrease for a given change
in the input. Here we have sacrificed precision for realism and generality. On the other hand,
simulation models usually try for precision and realism, but are not very general. Hopefully,
you will gain some idea of what these trade-offs mean after we have studied a few models.
In the modelling process we need to make assumptions and these are reflected in the variables
and their interrelationship. We can then use the model to make predictions. Modelling is based
on the following idea: If the assumptions of the model are true, the conclusions must also be
true. A more colourful way of saying this is: garbage in — garbage out. So, if we get a false
prediction from the model, then there must be something wrong with the model. We must
realise that the model is only an approximation and so we cannot expect perfect predictions.
The question is then: How can we judge the quality of a model?
To start with, we can have crude models (very few, or just bad, assumptions). This may
lead to predictions which are not believable, especially if another model makes contradictory
predictions. We say a result is robust if it can be derived from a variety of different models of
the same situation, or from a rather general model. Alternatively, a prediction that depends
on very special assumptions for its validity is fragile. The cruder the model, the less believable
its fragile predictions. Generally, a decision about the validity of a model is usually based on
the accuracy of its predictions.
Mathematical models can, in some sense, provide an explanation for the physical situation
under consideration. However, two different models may make the same predictions, but offer
different explanations. To see this, consider a model of a perfect LC circuit and a perfect
spring. The differential equations governing these two situations are identical, but Hooke’s law
for springs cannot explain the way in which an electric circuit behaves.

How to build a mathematical model

What are the rules for building mathematical models? Well, there are no rules!!! Model
building involves imagination and skill. Since imagination cannot be taught, only enhanced
through practice, there are only guidelines as to how to build a model.

2
 The following is an outline of the modelling process (again from E.A. Bender) (you may
like to research other processes):

1. Formulate the Problem. What is it that you wish to know? The nature of the model
you choose depends very much on what you want it to do.

2. Outline the Model. At this stage you must separate the various parts of the universe
into unimportant, exogenous (input) and endogenous (output) variables. The interrela-
tions among the variables must also be specified.

3. Is it useful? Now stand back and look at what you have. Can you obtain the needed
data and then use it in the model to make the predictions you want? If the answer is no,
then you must reformulate the model (step 2) and perhaps even the problem (step 1).
Note that “useful” does not mean reasonable or accurate; they come in step 4. It means:
If the model fits the situation, will we be able to use it?

4. Test the Model. Use the model to make predictions that can be checked against data or
common sense. It is not advisable to rely entirely on common sense, because it may well be
wrong. Start out with easy predictions — don’t waste time on involved calculations with
a model that may be no good. If these predictions are bad and there are no mathematical
errors, return to step 2 or step 1. If these predictions are acceptable, they should give
you some feeling for the accuracy and range of applicability of the model. If they are less
accurate than you anticipated, it is a good idea to try to understand why, since this may
uncover implicit or false assumptions.

We are now in a position to test the model. However, don’t push it too far; it is dangerous
to apply a model blindly to problems that differ greatly from those on which the model was
tested. Treat every application of the model as a test of its validity.
Sometimes it is difficult to know which factors to ignore or how accurately the input variable
need to be determined. We can start with a crude model and rough data estimates to see
which factors are important in the model and to determine the accuracy required for the input
variables.

Why do mathematical modelling?

Why develop and study mathematical models? Why not simply deal with the real world
instead? One clear benefit is that mathematical modelling can avoid or reduce the need for
costly, dangerous, undesirable or impossible experiments, such as:

1. What is the most efficient way to divide the fuel between the stages of a multistage rocket?

2. What would be the effect of a very bad nuclear reactor accident?

3
 3. How large a meteor was needed to produce Meteor Crater in Arizona?

In trying to “explain” the world, modelling is essential. This is an ongoing process, back
and forth between the real world and the model. It is also an iterative process: Improved
models to explain new experimental data and more refined experiments to validate modelling
observations.

An Example Problem
You are asked to predict how a population will grow over a few generations. (Step 1.)
Let the input variables be the net reproduction rate per individual, r, the time, t, and N0
the size of the population at t = 0. (Step 2.) The net reproduction rate is the birth rate minus
the death rate or the fractional rate of change of the population size
1 dN
r= (1)
N dt
Here the output variable is the size of the population at time t, N (t).
We have ignored time lag effects (an assumption). If the fraction of the population that is
of reproductive age varies with t, this can be a poor approximation. Hence (1) gives a rather
crude model. However, the model can be useful if it fits the real world. (Step 3.)
The solution is
N (t) = N0 ert .
Unless r = 0, the population will eventually either die out (r < 0) or fill the universe (r > 0).
To get reasonable predictions for the behaviours of the population size is difficult from this
simple model, casting doubt on the validity of the model (using a constant net reproduction
rate). Important point: Study the long term behaviour of your model in limiting cases (ie,
as t → ∞).

4
 This test indicates we should reject the model for long term growth, but it may be valid for
short term growth. (Step 4.) We were asked for long term behaviour.
Now assume the population growth rate depends on the size of the population due to
overcrowding or shortage of food. As the population increases we expect the death rate to
exceed the birth rate. Hence we can replace r in (1) by r(N ) which is a strictly decreasing
function of N for large N and negative when N is very large:
1 dN
= r(N ) (2)
N dt
(Step 2 redone.) The model is less useful as obtaining r(N ) may be difficult. However, rough
estimates can be obtained. On to step 4. If r(N0 ) = 0 for some N0 it can be shown that
N (t) → N0 as t → ∞. This is a robust prediction since very few assumptions have been made
about r(N ).
The cycle of steps 4, 2 and 3 can be repeated since (2) has many drawbacks. For example,
N (t) can only approach N0 as t increases. In reality, it might overshoot or fluctuate around
the one value due to random events.
We can introduce time lags. For example if the death rate, m, is not age dependent and
the birth rate b changes from zero to a constant at p, then we get
dN
= −mN (t) + bN (t − p)
dt
We could make both m and b functions of N (t), N (t − p) or some weighted average of N on
the interval [t − p, t].
To allow for random fluctuations we must replace our deterministic model by a random
one. If age and sex rates are changing, split the population into subpopulations. Demographic
models are designed by breaking time into discrete units (eg 5 years), men are ignored and
women are divided into age classes separated by a single unit. For each age class there is a
death rate, mi , and birth rate, bi , for female children. The number of newborn girls at time t+1
P
is N0 (t + 1) = bi Ni (t) and the number of women in class i + 1 is Ni+1 (t + 1) = (1 − mi )Ni (t)
(survival). If is often assumed that bi and mi are constant as the time periods are short. This
is no longer a differential equation model.
We could also include seasonality.
As you can see the models are becoming more complicated and can get more difficult to
solve.

Another Example
The manager of a large commercial printing company asks your advice on how many sales
people to employ.
Qualitatively: more sales peoples will mean increased sales overhead, while fewer sales
people may mean losing potential customers.

5
 The problem as stated is unanswerable. We need to understand more, we need to ask more
questions:

• What are the production limitations of the company?

• what are the goals of the management?

• Maximum profit?

• Maximum “empire” with satisfactory profit?

• Something else?

We need clear answers to these questions otherwise our recommendations might be inaccu-
rate.
A better way to proceed is to provide a description of the consequences of sales forces of
various sizes, thus leaving the final decision to management.