Modeling Free Fall

T. Mazzola, M. Fok, J. Okerago and Z. Suresh

Mentor: Dr. Samuel Brzezicki

October 2025 - March 2026

1
 Abstract
Free fall can be described as a body accelerating in a downward move-
ment due only to Earth’s gravitational pull, meaning true free fall can
only happen in a vacuum. The aim of this report is to use primary ex-
perimental data to compare predictions made by mathematical models of
free fall with real observations. Variables affecting an object in free fall
will be identified, primarily the force of gravity and drag, and we will ex-
plore the best way of modeling these, beginning with the use of Newton’s
Second Law of Motion to derive various models. It will be assumed that
acceleration due to gravity, g, is approximately 9.81 metres per second
squared [9]. Using computer software, it is possible to simulate a falling
object using a model, allowing for a direct comparison between what a
model predicts and actuality. At a later stage, we aim to explore how
initial conditions, such as rotational velocity, cause spin and affect the
way in which an object falls while still in the air. Finally, we shall discuss
the limitations and strengths of such a model, assessing its accuracy and
the parameters within which it works.

2
Contents
1 Introduction 5
1.1 Aim of the report . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2 What constitutes an accurate model? . . . . . . . . . . . . . . . . 5

2 Collecting Experimental Data Regarding

Motion in Free Fall 6
2.1 Aim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.3 Equipment List . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Figure of results . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.6 A Simple View of Free fall: Modeling Constant Acceleration . . . 9
2.7 Findings of the practical . . . . . . . . . . . . . . . . . . . . . . . 10

3 Improving the Free fall Model - Accounting for air resistance 11

3.1 Evaluating air resistance as proportional to vn . . . . . . . . . . 11
3.2 Modeling Linear drag: FD = kv . . . . . . . . . . . . . . . . . . . 12
3.2.1 Solving for v(t) . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.2 Integrating v(t) with s(0) = 0 . . . . . . . . . . . . . . . . 12
3.2.3 A Short Explanation . . . . . . . . . . . . . . . . . . . . . 13
3.3 Modeling Quadratic drag: FD = kv 2 . . . . . . . . . . . . . . . . 13
3.3.1 Integrating v(t) with s(0) = 0 (again) . . . . . . . . . . . 13
3.3.2 A Short Explanation for the Quadratic Drag Model . . . 14
3.4 Summary of Models explored so far . . . . . . . . . . . . . . . . . 14

4 Using Numerical Methods to Approximate Models for drag

force 15
4.1 Modeling Cubic and Quartic Drag . . . . . . . . . . . . . . . . . 15
4.1.1 Euler’s Method . . . . . . . . . . . . . . . . . . . . . . . . 16
4.1.2 Modeling Drag using fractional powers of velocity . . . . 16
4.2 Comparing theoretical models to experimental data . . . . . . . . 18
4.3 Checking accuracy of k-values . . . . . . . . . . . . . . . . . . . . 20
4.3.1 Determining uncertainty in velocity measurement . . . . . 21

5 Comparing the suitability of models 22

5.1 Testing Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.1.1 Comments on the accuracy of the models . . . . . . . . . 23

6 Using Least Squares Estimates (LSE) 25

6.1 Comparing accuracy of velocity predictions given by models . . . 26
6.1.1 Terminal Velocity Predictions . . . . . . . . . . . . . . . . 26
6.1.2 Predictions for Medium t- values . . . . . . . . . . . . . . 28
6.2 Using Least Squares Estimates to optimise k . . . . . . . . . . . 29

7 Findings and Discussion 31

3
 7.1 Aims of Further Research . . . . . . . . . . . . . . . . . . . . . . 32

8 Further Factors affecting free fall 33

8.0.1 Coriolis Force . . . . . . . . . . . . . . . . . . . . . . . . . 33
8.0.2 Modeling the Coriolis Force . . . . . . . . . . . . . . . . . 33
8.1 Considering Upthrust . . . . . . . . . . . . . . . . . . . . . . . . 33

9 free fall in Two Dimensions 35

10 Conclusion 37

11 Bibliography 38

12 Appendix 40
12.1 Raw Data Values . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
12.1.1 Data for values of k . . . . . . . . . . . . . . . . . . . . . 40
12.2 Deriving ’k’ through solving differential equations . . . . . . . . 41
12.2.1 kv and kv2 models . . . . . . . . . . . . . . . . . . . . . . 41
12.2.2 kv3 andkv 4 models . . . . . . . . . . . . . . . . . . . . . . 42
12.2.3 Modeling Drag using fractional powers of velocity . . . . 42
12.2.4 Using the Taylor Series to generate terms . . . . . . . . . 43
12.3 Breakdown of the Statistical approach used . . . . . . . . . . . . 43
12.3.1 Checking k values . . . . . . . . . . . . . . . . . . . . . . 44
12.4 Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4
1 Introduction
1.1 Aim of the report
The study of free fall dates back to the sixteenth century, and since its discovery
in 1551, it has remained vital in areas such as astrophysics, aerospace engineer-
ing, and space science. These applications make it necessary to build a clear
and reliable mathematical model for free-fall motion. An object undergoing free
fall on Earth will be subject to air resistance, and exploring this force will form
a significant part of the report. As this force is based on multiple parameters,
such as fluid density, area, and temperature, it is evident that a general equa-
tion to model free fall, although convenient, might not be the most accurate for
a variety of different objects. For this reason, we will consider which models
for air resistance are most appropriate for varying scenarios, as well as explore
other factors that may act on falling objects, using exact or numerical methods
when needed.

1.2 What constitutes an accurate model?

A model is accurate if it is consistent with empirical data, meaning it can reliably
make predictions by defining the relationship between variables. Therefore, a
mathematical model relies on certain assumptions being fulfilled and works best
within specific parameters. In this report, the focus will be on the mass of the
falling object, as this greatly impacts its terminal velocity.

5
2 Collecting Experimental Data Regarding
Motion in Free Fall
2.1 Aim
To obtain a set of accurate empirical data describing an object in free fall to
verify predictions made by mathematical models that will later be derived.

2.2 Variables
Independent Variable:
• Mass of the cupcake cases
• Height of the drop
Dependent Variable:
• Time taken for the cupcake cases to fall
Control Variables:
• Room temperature
• Room pressure
• Initial velocity
• Release method

2.3 Equipment List

• Metre rule or measuring tape
• Stopwatch
• Video camera
• Cupcake cases
• Marker or tape

6
2.4 Method
1. Measure the mass of 1 cupcake case and record it.
2. Measure the vertical distance of the two heights with a meter rule 0.85 m−
h1 and 2.05 m − h2 and mark the positions using a marker or tape.
3. Drop the cupcake case from rest at fixed height, h1 , while filming with the
video camera and start the stopwatch.
4. Ensure the stopwatch is visible in the frame of the video camera.
5. Stop the stopwatch when the cupcake cake touches the ground and then
stop recording.
6. Using the video and the time on the stopwatch, calculate the time taken
for the cupcake case to reach the ground.
7. Repeat steps 3 - 6 at least twice more to obtain a set of data for one
cupcake case at h1 .
8. Calculate the average time taken for the cupcake case to fall.
9. Repeat steps 1 - 8 for different masses of cupcake cases by increasing the
number of cupcake cases falling.
10. Repeat steps 1 - 9 for all masses of the cupcake cases at h2
11. Record the data collected in a table.
12. Use the constant acceleration suvat equations to calculate values for final
velocity of the cupcake cases.
NB: h1 and h2 are not constant values. They can be changed according to
available height. Due to the environment in which the practical was conducted,
0.85 m and 2.05 m were the most convenient values of h1 and h2 to use.

7
2.5 Figure of results

Figure 1: Results from the experiment 0.85m drop

Figure 2: Results from the experiment 2.05m drop

the values for s, u, t, and m (from left to right) are experimental, v and a have
been derived through use of the constant acceleration equations. Discrepancies
between results show that this is not at all a good model for practical application.

8
2.6 A Simple View of Free fall: Modeling Constant Ac-
celeration
The motion of an object in free fall can be treated as uniformly accelerating
motion when air resistance is ignored. Using this assumption, the only other
force acting on the object in controlled conditions is gravity, which produces a
constant acceleration, g. Constant acceleration allows the use of the kinematic
relations, referred to as the constant acceleration (suvat) equations. The term
suvat comes from the five quantities involved:
s: displacement
u: initial velocity
v : final velocity
a: acceleration
t: time
The equations connecting these variables are:

v = u + at,

s = ut + 21 at2 ,

v 2 = u2 + 2as,

s = 12 (u + v)t,

s = vt − 12 at2 .

These equations are valid only when the acceleration remains constant. In a
free-fall model where air resistance is neglected, this requirement is met, so
the equations allow a direct way to calculate velocity, displacement, and time
throughout the motion. (13)
Following on from the practical, the raw data values were used to obtain a value
of k for each model we will explore later in the report.

9
2.7 Findings of the practical
Conclusion
The tables of results show there is a difference between the theoretical values ob-
tained using the constant acceleration (suvat) equations and the values obtained
from the practical. By comparing experimental values to the theoretical values
given by a model, it is possible to determine the accuracy of a model within
some uncertainty due to error. Through proof by contradiction it is shown that
the suvat equations did not successfully model the different masses in free fall.
The equations do not factor in variables such as air resistance as resistive forces
cause acceleration to vary during motion and therefore cannot accurately be
used to estimate final velocity. This means that the object’s mass affects the
rate at which different numbers of cupcake cases falls (in the theoretical model,
mass should not affect the values as it is not included in constant acceleration
calculations). Therefore, there must be other variables that cannot be taken
into account, such as lift.

Evaluation
In terms of practicality, multiple factors other than air resistance may have
caused a decrease in the experimental accuracy of our results. Human errors,
such as whether the cupcake cases are dropped exactly at the measured height
caused our calculated values to vary from the ideal theoretical values, decreasing
the accuracy of the practical. Additionally, the assumption was made that all
the cupcake cases would reach terminal velocity during their descent, allowing
the use of the constant acceleration (suvat) equations. Furthermore, increasing
the number of cupcake cases may have affected the weight distribution of the
object falling, meaning that different objects of the same mass may have followed
different flight paths. During the practical, we observed the object ”wobbling”
in the air as it fell due to its rotational velocity; however, not every object
exhibited the same deviation in its flight path. This variation would affect the
time of flight and reduce experimental accuracy.
To improve this investigation, some aspects of the practical could be changed.
An important change would be to use a greater range of drop heights and release
them from above or at the measured height. This would reduce the likelihood
of human error and allow the person to ensure that the object is dropped from
the correct height. As it is difficult to alter the weight distribution of cupcake
cases, different objects could be used, such as a cube, which would also simplify
calculations.

10
3 Improving the Free fall Model - Accounting
for air resistance
Following the practical, it is possible to consider further forces acting on a falling
mass, beginning with air resistance or drag force. The currently accepted model
for drag force is:
1
FD = CD ρAv 2
2
where A is the surface area of the object, CD is the drag coefficient and ρ is
air density. [4] However, it is not certain that this is the most accurate model
for the parameters of a free-falling cupcake holder, because of its light mass
and large surface area: volume ratio [8], so it is necessary to compare different
models for air resistance in order to determine the most accurate one.

3.1 Evaluating air resistance as proportional to vn

Considering air resistance, a free body force diagram of the falling mass can be
depicted using this diagram:

Figure 3: Force Diagram considering drag force (bv) and Weight (mg) [5]

We consider a mass, m, falling under gravity with acceleration, g. The air

resistance or drag force acting on the object opposes its downward motion,
which is defined as the positive direction. Initially, it can be considered to
model the drag force as:
• 1) Linear drag: FD = kv, or
• 2) Quadratic drag: FD = kv 2
For both models, it is assumed that the object starts from rest meaning at s = 0
and t = 0, u = 0 , which aligns with the experimental conditions to increase

11
the validity of the model. Note that the k terms in these two models do not
represent the same value, but rather an arbitrary constant that depends on
which exponent of v is used to model the drag force. It is necessary to set up
an equation in terms of v and k to find by calculation and modeling the value
of k.

The equation of motion for case 1, as is made evident from Figure 3 (by Newton’s
Second Law) is:
dv
m = mg − FD
dt

3.2 Modeling Linear drag: FD = kv

We can substitute kv for the drag force. This gives:
dv
m = mg − kv
dt
(1)
This is a first-order linear differential equation (ODE). We can solve this to
eliminate the acceleration, a, term and derive an equation for v(t) .

3.2.1 Solving for v(t)

Eq. 1 can be re-written as:
−m
Z
1
dv = t + c
k gm/k − v
so
−m gm
ln| − v| = t + c
k k
It is verifiable that c = 0 because of the nature of the initial conditions of the
experiment - initially, the speed of the object is 0. Exponentiating and solving
for v gives:
gm
v(t) = (1 − e−tk/m )
k
(2)
In this form, it is impossible to calculate the coefficient of air resistance, k , as
v(t) and k are both unknown. Therefore, it is necessary to integrate both sides
of the equation again in order to solve for s(t), which is a known value.

3.2.2 Integrating v(t) with s(0) = 0

Integrating both sides of Eq. 2 by separating variables [11] gives:
Z t
m
v(t′ ) dt′ = vt t − vt 1 − e−tk/m .


s(t) =
0 k

12
and writing out within parameters gives:

mg m2 g  k

s(t) = t − 2 1 − e− m t .
k k

3.2.3 A Short Explanation

At t = 0, the drag force acting on the mass is zero since initially the object is at
rest. Therefore the object initially accelerates at a rate g. As velocity increases
the drag force kv also increases in proportion, reducing net acceleration, as seen
in Eq. 1. Eventually the resultant (net) force acting on the mass is zero, and
at this point the mass is at terminal velocity vt .

3.3 Modeling Quadratic drag: FD = kv 2

Now that the drag force is proportional to the square of the velocity v of the
object, Eq 1. can be rewritten as:
dv
m = mg − kv 2 .
dt
Similarly to Section 4.2.1, we can solve this ODE to be certain that
r r
mg kg
v(t) = ( )tanh( t)(3)
k m

3.3.1 Integrating v(t) with s(0) = 0 (again)

Similarly to the previous model, we must integrate v(t) so that an equation for
k can be set up including only known values. This can be done through use of
the ’u’-substitution method:
Z t Z t  gt′ 
s(t) = v(t′ ) dt′ = vt tanh dt′ .
0 0 v t

gt′ vt
Change variable u = so dt′ = du:
vt g
gt/vt
vt2 vt2 
Z
gt 
s(t) = tanh(u) du = ln cosh .
g 0 g vt

Hence
vt2  gt  m  gt 
s(t) = ln cosh = ln cosh p (4)
g vt k mg/k

13
3.3.2 A Short Explanation for the Quadratic Drag Model
Initially, (for very small t) drag is negligible and v ≈ gt. As v grows, the
quadratic drag kv 2 grows rapidly andpquickly balances gravity; the velocity
approaches the terminal speed vt = mg/k. The approach to vt follows a
hyperbolic tangent law: v(t) = vt tanh(gt/vt ). Positions grow like s ∼ 12 gt2 for
very small t and transition to linear growth with slope vt for large t; the exact
expression is the logarithm of a hyperbolic cosine shown above.

3.4 Summary of Models explored so far

Linear drag F = kv

mg mg  k
 mg m2 g  k

vt = , v(t) = 1 − e− m t , s(t) = t − 2 1 − e− m t (5)
k k k k

Quadratic drag F = kv 2

vt2 
r
mg  gt  gt 
vt = , v(t) = vt tanh , s(t) = ln cosh (6)
k vt g vt

Eq. 6 includes terminal velocity. As shown, this can be expressed in terms of

other constants. Alternatively, footage from the practical experiment in Section
3 can be reviewed to determine the terminal velocity, as well as ascertaining
that the falling object is moving at a constant speed.

14
4 Using Numerical Methods to Approximate Mod-
els for drag force
4.1 Modeling Cubic and Quartic Drag
For cubic and quartic drag, similar to previous models, the first step is to set
up an equation based on Newton’s Second Law. Upon attempting to solve the
resulting ODEs (first order differential equations), it becomes clear that they
cannot be solved analytically. For example, solving for the velocity function of
v cubed gives
dv 1
. k 3
=1
dt g − m v
which cannot be solved. This is also true for further powers of v that will be
considered. Therefore, it is necessary to approximate the solutions and the two
respective resulting functions in the form of a polynomial in terms of t that is
much easier to manipulate and use a numerical method to solve. This can be
achieved by using a truncated form of the Taylor Series to generate terms of the
solution v(t) , in this case the first four terms. A full method of this expansion
is shown in Section 10.2.3 of the Appendix.
Generating the first four terms of the Cubic and Quartic Drag functions by
taking t = 0 as the centre of the expansion gives

kg 3 4 3k 2 g 5 7 57k 3 g 7 9
v1 (t) = gt − t + 2
t − t (7)
4m 28m 1120m3

for Cubic Drag, and

kg 4 5 4k 2 g 7 9 134k 3 g 1 0 13
v2 (t) = gt − t + t − t (8)
5m 45m2 2925m3

for Quartic Drag

It is now possible to integrate these approximations using the power rule for
integration, as the two functions are simply polynomials.
For Cubic Drag:

g 2 kg 3 5 3k 2 g 5 8 57k 3 g 7 10
s1 (t) = t − t + 2
t − t (9)
2 20m 224m 11200m3

For Quartic Drag:

15
 g 2 kg 4 6 2k 2 g 7 10 67k 3 g 10 14
s2 (t) = t − t + 2
t − t (10)
2 30m 225m 20475m3

For now it is acceptable to only generate the first four terms, as further terms
will have larger and larger powers of t. This means that since the object’s
journey time is under 1 second ( as shown in Figures 1 and 2), further terms
become so small that they are negligible.

4.1.1 Euler’s Method

Alternatively, it is also possible to use Euler’s method to approximate solutions
to the ODEs. This is given by

yn+1 = yn + ∆t(xn , yn )

considering x or y depending on which is required. Although this may also have

produced an accurate result, it would not have generated a function v(t) further
a known value would have had to be used as a starting point however, using v
= 0 at time 0 is impractical as this does not iterate.

4.1.2 Modeling Drag using fractional powers of velocity

It is reasonable to assume that the most accurate model for air resistance will be
using a non-integer power of v. For example, say the two most accurate models
of the ones already derived are the linear and quadratic ones. Naturally models
with fractional powers between 1 and 2 should also be considered. Now v to the
power of 3/2 will be considered. Applying Newton’s Second Law gives:

dv 3
= a(r − v 2 )
dt
this is a simplified version of the original integrand. We can use a Puiseux Series
to generate the first three terms of the v(t) function to approximate a solution.
When t is small, v(t) is also close to 0. Therefore the v component is negligible.
For the first term, Z
v(t) = ardt = art

for term two, we assume it has a general form

art + Ctn

using the Puiseux Series:

d 3
(art + Ctn ) = a(r − (art + Ctn ) 2 )
dt

16
using the first two terms of the trinomial expansion formula
3
ar + Cntn−1 = a(r − (art) 2 )

Simplifying and comparing power son both sides gives

3
tn−1 = t 2
5 3
2.5C = −a 2 r 2
To find the final term begin by assuming

v(t) = C + Dtm

this gives
d 2 5 3 5 2 5 3 5
(art − a 2 r 2 t 2 + Dtm ) = ar − a(art − a 2 r 2 t 2 )
dt 5 5
simplifying and compare powers as before gives
3 4 2
m = 4, 4D = a r
5
this means the full three term expansion is
2 5 3 5 3
v(t) = art − a 2 r 2 t 2 + a4 r2 t4
5 20
where
k gm
a= ,r = , ar = g
m k
substituting these values for a and r gives

3
2g 2 k 5 3g 2 k 2 4
v(t) = gt − t2 + t (11)
5m 20m2

Finally we integrate this to get the position function

3
1 4g 2 k 7 3g 2 k 2 5
s(t) = gt2 − t2 + t (12)
2 35m 100m2

Again, this is only an approximate solution. This is also based on the assumption
that the initial velocity is zero, otherwise the a and r values would change. Now
that all the velocity and position equations have been compared the values of k
for each model must be found so that they can then be used. This model will
be referred to as the ’fractional power’ model for convenience.

17
4.2 Comparing theoretical models to experimental data
Once all the position equations s(t) have been derived it is possible to use values
from the Raw Data Table( see 10.1) to obtain a value for k, the coefficient of
air resistance, for each model. For example in the Linear drag model you can
substitute the raw data into position Equation 5. This gives six different curves
from the 6 different rows of values for each of the models. Beginning with the
kv model this gives:

Figure 5: A range of values for k in the linear model

To calculate an accurate value of k ( given by the x-intercepts) it is necessary to

determine an average value that is close to the ’true value’ of k . With a small
dataset such as this, it may be difficult to do so, as a single anomaly – like the
black line in the graph of Figure 5 can skew the average. Therefore it is sensible
to exclude these anomalies when attempting to determine the true value of k
. To achieve this, the standard deviation of the data can be used as a metric
to identify outliers; however, with only 6 data points, it is more reasonable to
use the median absolute deviation (MAD) which measures the average distance
from the median rather than from the mean of the data. In addition to being
less sensitive to outliers than standard deviation, the MAD does not require
the data to be normally distributed. For the raw data in this report, using the
standard deviation to measure variability would itself be an assumption.
To identify anomalies we can compute a modified standardised score for each
data point:

0.6745(xi − m)
zi =
M AD
where the MAD is given by

M AD(X) = m(|xi − m|)

18
Commonly, if a data point has a standardised score greater than 3.5 it is con-
sidered an outlier [14]. 3 is also commonly used however for a small dataset
we decided it would be more sensible to use a less restrictive value in order to
include more data into the mean so that the final value of k for each model
would be more representative of its data points.
Applying this to the data from Figure 5 we find,
m = 1.7125 × 10−3
To shortcut calculation by hand we can use this code in R to flag outliers:

As seen from the data frame, x = 0.002420 is an outlier so must not be considered
when finding an average. As the remaining data does not appear to be skewed
and is instead normally distributed, it is sensible to find a mean for the remaining
values. This gives:
klm = 1.6978 × 10−3

This same process is repeated for the quadratic, cubic and quartic drag models.

Figure 6: A range of values for k in the quadratic model

Using similar lines of code from the linear drag model it can be verified that the
larget value for k, 0.001425, is also an outlier.

19
Therefore,
kquadratic = 8.786 × 10−4

For Cubic Drag:

kcubic = 9.71 × 10−5

to three significant figures and there are no anomalous values.

For Quartic Drag:
the largest value of k was anomalous and so
kquartic = 1.57 × 10−5

repeating this process for the fractional power model gave

kf rac = 3.788 × 10−4

Please see Section 10.1 to view the graphs for the Cubic and Quartic Models.
In Section 10.3 of the appendix we will discuss the limitations of our statistical
approach, as well as some inaccuracies in the method used.

4.3 Checking accuracy of k-values

Now that we have estimates for the coefficient of air resistance, k, for each
model, we must be certain of their accuracy. As (0,0) is a known value for the
velocity and position functions of each model, we could input the estimated
value of k into either function at t=0. However, this would give v =0 regardless
of the value of k. So instead, we can use a very small value of t, where v
approximately 0 to check for accuracy. Ideally this should result in v = 0.
The velocity equations for each model were used for simplicity, compared the
the more complex position equations. Note: to check accuracy the mass of 1
cupcake falling over 2.05m was used. As this was the lightest mass it showed
the smallest range in fall times so is best to use to test if the k value is truly
accurate, as this reading would have the lowest uncertainty.
using t = 0.0001 For the linear model,this gives:

9.81 × 2.738 × 10−4

v0 = (1 − e0.0001 )
1.6978 × 10−3
so v (-) 0.0016 metres per second. This shows that our k value is a slight
overestimate; however, ±0.0016 is definitely within the uncertainty range of the
experiment so this value for k is considered valid. Repeating this process for all
other models shows that valid values for k were obtained, accurate to 3 decimal
places. The full accuracy breakdown can be found in the Appendix.

20
4.3.1 Determining uncertainty in velocity measurement
To determine the absolute uncertainty of the velocity reading we use
∆s ∆t
∆v = v( + )
s t
where v is the average velocity throughout the fall. The values of t and s are
shown in the raw data table. The uncertainty of s is half a millimetre and the
uncertainty in t is half the range of the fall times. Using these value gives

∆v = ±0.03ms−1

All predictions given by the k-values were within this range from the measured
value of 0.

21
5 Comparing the suitability of models
Once all the models have been derived, their relative strengths must be com-
pared. This includes the efficacy, accuracy, range and consistency of each model.
Later, we will consider whether the ’best’ model may change depending on the
parameters of a given falling object.

5.1 Testing Accuracy

To begin testing the accuracy of the models, we input the heights from which
the cupcakes were dropped - these can be found in the Raw Data Table and
record the corresponding predictions that each model gave for the cupcake’s
fall. Collating these results in a table gives:

Figure 7: Table showing fall time estimates

In Figure 7 each model is used to estimate the fall time for a drop of 2.05m and
0.85m. The last column shows the residual, defined as the difference between t
estimate and ta vg.T hepredictedtimesareobtainedgraphically, inamannersimilartohowthevaluesof kweredeter
−f orexample, Equation5f orthelineardragmodel.T hisgeneratesarootatx = [Link], itisn

Figure 8: Graph showing root where x=t

This process is repeated for the quadratic drag model:

22
 Figure 9: Quadratic model prediction for fall time
Again, we consider the positive root and repeat for the remaining models to fill
the table.

To calculate an estimated fall time given by the fractional power model, we

substitute the same values into Eq. (12). This gave an estimated fall time and
residual of

Figure 13: Predictions given by the fractional Model

The residual is smaller for the 0.85m drop height as seen in the figure. This
suggests this model performs better over shorter falls, similiarly to the rest of
the models apart from the linear model

5.1.1 Comments on the accuracy of the models

For both the 2.05 and 0.85m drop, the linear drag model’s fall time was closest
to the measured value with a residual of less than 0.02s in both cases, this
suggests it is most accurate. The uncertainty in the fall time is

∆t = ±0.02s

This means both predictions given by the linear model are valid as they are
within this range from the measured value. As it is valid at initial and terminal
velocity. We can be confident it will give accurate predictions by interpolating
between those two points.
Overall, as the powers of v increased, the residuals tended to grow larger, so
the models became less accurate. This is not true for the fractional model and

23
the quadratic models, however, as the quadratic model is more accurate despite
having a larger power of v. Partly, this may be because s(t) function for the
fractional model is only an approximation, adding more terms could improve its
accuracy. This also means the k-value is an approximation of an approximation,
since it was derived from an approximate function. Both factors contribute to
inaccuracy.

However, this analysis only considers one data point, namely (t, 0.85), and
therefore does not provide any insight into the behaviour of the object while
falling before reaching the ground. For example, the linear model could be the
least accurate at some time (t, s) between the moment it was dropped and when
it landed. This will be considered in more detail in the next section.

Further, it is important to note that all models where more accurate in de-
termining the fall time for the 2.05m drop than the 0.85m drop. This may be
because during the 0.85m drop the object was falling at terminal velocity for a
shorter time, meaning the drag force was at a maximum for a shorter time. This
is not accounted for in any of the models, which explains why all the predictions
given are underestimates.

Finally, we found that the cubic and quartic models were the least accurate.
This may be because in these models, air resistance has a much smaller effect.
For example, Equation 8 for velocity implies that velocity will increase at a
faster rate at a time t than in the quartic and linear models; therefore, its po-
sition equation will predict a shorter fall time. The same reasoning applies to
the quartic model.

24
6 Using Least Squares Estimates (LSE)
We can now compare the accuracy of the models in more detail as we compare
the velocity equation curves to known data points.

First, we must identify data points to compare with the predicted values. We
will use six points; however, for a more robust comparison, ideally as many
points as possible should be used. The velocity of the particle at a given time
was determined by measuring the distance travelled over a time interval of 0.05
seconds, equivalent to one frame. Two points should be (0,0) and the terminal
velocity. Terminal velocity should be ensured by using a light object and drop-
ping it over a sufficiently long distance. Velocity should be compared at two
different points to confirm that the object is no longer accelerating. We can plot
these 6 data points on a velocity-time graph, along with the linear drag curve:

Figure 11: velocity-time graph of data points with linear drag model overlayed
Although there are few data points, we observe that the curve of the linear
drag velocity equation fits the data well, giving only a slightly smaller terminal
velocity. We can now put all the ’fixed’ velocity curves on the same graph:

25
 Figure 14:velocity-time graph with all fixed curves
The black points on the graph represent the 5 known data points obtained by
analysing the footage of the 0.85m drop. This drop height was chosen rather
than the 2.05m drop height, as the 0.85m drop showed the largest range in resid-
uals, so would be more suitable to use to compare the different curves given by
the models. The 5 curves are fixed, and correspond to the five models consid-
ered. They are:

Red: Linear Drag

Purple: Quadratic Drag

Blue: Fractional Drag

Orange: Cubic Drag

Green: Quartic Drag

6.1 Comparing accuracy of velocity predictions given by

models
6.1.1 Terminal Velocity Predictions
The final two data points seem to be at the same height on the y axis, and
therefore same velocity. This suggests that terminal velocity has been reached

26
and the forces acting on the cupcake holder are now balanced, meaning that
mg = FD
plotting a horizonal line between the two points gives
vt = 1.6ms−1
which is in line with the general terminal velocity of a light cucake holder.
Clearly, the Quadratic and Linear models are most accurate, shown by Figure
14. Both graphs reach a gradient of zero, suggesting terminal velocity has been
reached. Graphically, it seems the linear model’s terminal velocity is closer to
1.6 than that of the quadratic model. We will later test this using Least Squares
Estimates.

The Cubic, Quartic and Fractional models do not perform well at terminal
velocity. All of these graphs exhibit a maximum point before terminal velocity
is reached and then tend to negative infinity, which is clearly inconsistent with
the observed data. This behaviour occurs because the v(t) functions of these
models are all polynomial small time approximations where v is near zero. As
discussed in Section 4, the approximations derived using variations of the Taylor
Series (detailed in the Appendix) assume that t is close to zero and therefore
that v is also small. Consequently, these models are only accurate for small
values of t, when the velocity remains low. From Figure 14, it can be seen that
all curves overlap for small values of t. Furthermore, all curves cross a point
near (0,0) indicating they are accurate near that start of the motion.

We can still compare the accuracy of the approximated models by compar-

ing their maximum velocities. Clearly the fractional model is best as it peaks
closest to v = 1.6, in comparison to the cubic and quartic models which reach
over 2 metres per second. The prediction given by the Fractional model is an
underestimate. This may be because the value of k we obtained is inaccurate
as previously disussed. To address this issue, we can adjust k to fit the data
better:

27
 Figure 15:A better curve for the fractional model
Using k = 0.0003 gave a much closer maximum velocity to 1.6, without altering
the shape of the curve for the smaller values of t when the drag force was still
very small.

6.1.2 Predictions for Medium t- values

Clearly, the linear model is most suitable over large fall times. For small times
of under 0.1 seconds, all models perform very similarly as the 5 curves in Figure
14 almost perfectly overlap. To compare accuracy where t is in between these
two phases we can use residuals to find the best fitting model. If we define a
residual as
ri = vi − vmodel,i
The Least Squares Error can be computed for each model using

S = Σni=1 ri2

this calculates a metric of ’agreement’ between the known values and each model.
an S-value of 0 means there is perfect agreement as the squares of all residuals
of the 5 data points are 0.

Finding the S-value for each of the models we find that the linear model has the
smallest Error as
Slm = 0.037
using this code:

computing the Root Square Mean Error (RSME) gives:

r
0.037
RSM E = = 0.006
6

This means for a typical velocity of around 1 metre per second, there is a typical
disagreement of around 0.006 metres per second. Considering the uncertainty
in the velocity measurement, this is a very accurate model.
Using the same code but for the Quadratic Drag function gives

Squadratic = 0.250, RSM E = 0.204

28
This model shows less agreement with the data. Computing Least Squares Error
with the other models would give a huge disagreement as the graphs tend to
negative infinity, however this does not mean that they are not accurate for
small values of t.

6.2 Using Least Squares Estimates to optimise k

Finally we will use Least Squares Estimates to optimise the value of k. k is
optimised when the v(t) function for the linear drag model agrees perfectly
with the known data. This is when S = 0. For a linear regression line the Least
Squares Estimate Formula is
Σni=1 (xi − x̄)(yi − ȳ)
m=
Σni=1 (xi − x̄)2

Alternatively, this code gives the optimal value for k for a non-linear regres-
sion line:

Inputting this new value of k in Eq. 5 gives the following curve:

29
 Figure 16: Optimised linear drag curve
This curve has almost perfect agreement with the measured points and has a
terminal velocity of 1.63 metres per second.

30
7 Findings and Discussion
Although all models performed well with small time values, overall the linear
drag model was by far the most accurate for all values of t as it had almost
perfect agreement with the measured data. This makes it most suitable for
modeling a falling cupcake case.

Despite this, it is also important to note that other models may have been
less accurate not only because of the model itself, but also because the k value
used was not optimal. To test this, a least squares estimate must be found for
each model in future.
Further, the models derived by numerical methods ( cubic, quartic and frac-
tional models) were not accurate over large time intervals for reasons already
discussed. However, it is possible to make them more accurate for longer time
intervals by generating more terms of the function. As following terms have
larger powers, as t becomes greater the larger terms will have a bigger impact.
generating more terms was not feasible given the time constraints and resources
available to use. In addition, even if more terms had been generated, these
models are not suitable for modeling fall times of over 1 second, as in this case
velocity increases exponentially using the models. Finally, as the methods used
to generate terms gave polynomials, there will not be an accurate representa-
tion of terminal velocity. Therefore these models are generally not suitable to
model free fall, except for modeling very massive objects with a large accelera-
tion falling for short time periods, at least in the way covered in this report -
excluding for modeling very massive objects with a large acceleration falling for
short time periods,.

In contrast, the linear model is suitable to use for modeling free fall, although
for heavier objects. For the purpose of modeling the fall of a very light object,
such as in the cupcake case, the linear drag model is best, as it has a smaller
gradient than the quadratic model for any time t. This is suitable for light
objects that do not accelerate greatly as there is a small downward resultant
force acting on them, as mg is small. This results in a smaller rate of change
of velocity, so a model in which drag force is proportional to v rather than v
squared is more appropriate as generally there will be smaller values of v.

Our Research is consistent with currently accepted models for Drag Force. We
found that a Linear model works best for drag Force when n object is travelling
at low speeds. This is in line with models used to describe objects in laminar
flow at slow speeds for example stokes law:

FD = 6πηrv

which is in form
FD = kv
applies to small spheres travelling at low speeds.

31
Further, the equation most commonly used to model air resistance is
1
FD = CD ρAv 2
2
, which is in the form
FD = kv 2
We found that although this model did not best describe a falling cupcake case,
quadratic drag will be most suitable for most scenarios involving objects of
medium mass and surface area. This explains why is it almost exclusively used
in most fields - it is widely applicable and simple to use, unlike the cubic and
quartic models.

Finally, we found that using numerical methods to approximate solutions to

the cubic, quartic and fractional models was not very effective for reasons al-
ready stated. However, once an optimal k-value had been found it was accurate
in modeling velocity as a function of time for our cupcake holder. despite this as
only 3 terms were generated for each of these functions, they were not accurate
for any more than a second. In our case, using a linear or Quadratic model is
more suitable and easier to calculate.

7.1 Aims of Further Research

We have been successful in creating a linear Drag model that agrees with the
data, as shown be statistical methods. In addition, we have also explained why
it was successful and where a quadratic model would have worked better.

To continue further research, it would be necessary to generate further terms

of the velocity and position functions of the remaining models in order to test
whether they would be suitable for longer fall times for massive objects with
great terminal velocities.

Also, the experiment and entailing research only considered free fall in one di-
mension. This is not particularly realistic or applicable to most scenarios where
air resistance might be considered. In further research, free fall in 2 dimensions
must be explored, as well as the possibility of using different models for the x
and y directions and how these models might change with respect to the angle
of free fall.

One alternative approach to further research would be on the constant k it-

self. Currently used models for drag force mentioned above split k into multiple
parameters that are dependent on mass, density etc, not just on velocity. Fur-
ther research could explain what effect these parameters have on the value of k
and why they were chosen.

32
8 Further Factors affecting free fall
A model for free fall relies on certain underlying conditions to apply, such as con-
stant temperature, and wind speed as well as factors already discussed, namely
constant mass. In this section, further factors and variables will be discussed
which have too small of an impact to be considered in a feasible model, but are
important nonetheless.

8.0.1 Coriolis Force

The Coriolis force is a conceptual force that describes the ’apparent deflection of
the path of an object that moves within a rotating coordinate system’[2], namely
the anti-clockwise rotation of the earth about its axis. In the context of the
experiment shown in Section 3, the Coriolis Force will not produce a measurable
effect. This is because of the small distance over which the masses fell, and their
short duration of fall. However, the Coriolis Force must be considered when an
object is falling from great heights, for example in rocket or satellite re-entry
[10]

8.0.2 Modeling the Coriolis Force

It is important to note that this force has almost no effect on the vertical
component of the forces acting on the falling object. This is because the force
is purely horizontal - acting parallel to the earth’s surface, however over large
fall time will affect where an object will land on the earth’s surface for example
a decommissioned satellite re-entering orbit. Acceleration caused by Coriolis is

ac = v

where omega is a constant representing the earth’s angular velocity about its
axis. [3]

8.1 Considering Upthrust

Archimedes’ principle states that the upward buoyant force acting on an object
that is submerged in a fluid, whether fully or partially, is equal to the weight
of the fluid displaced by that body. This principle is fundamental in fluid me-
chanics and applies not only to liquids but also to gases such as air. [12] For an
object falling through air, the forces acting on it include its weight, which acts
downward, and the upward forces of air resistance and upthrust. The upthrust
equals to the weight of the displaced air and is given by U = ρair V g, where ρair
is the density of air (approximately 1.2 kg m−3 ), V is the volume of the object,
and g is the gravitational field strength. Although upthrust becomes relatively
more noticeable for very light objects, its magnitude remains small because of
the low density of air. For example, a paper cupcake case of mass 0.3g has a
weight of approximately 2.9 × 10−3 N, while its estimated upthrust is on the
order of 4 × 10−6 N. Thus, the upthrust is less than 0.2% of the object’s weight,

33
making its effect on the motion negligible compared with the much larger force
of air resistance.

34
9 free fall in Two Dimensions
In this report we have only considered free fall in one dimension, this is known
as linear motion and can only describe an object falling perfectly downward,
perpendicular to the surface of the earth. This is obviously unrealistic: a can-
nonball shot from a cannon or a raindrop in a gust of wind are both common
examples that can be used to describe objects in free fall and yet neither of
these fall purely downward. In other words, their velocity v can be split into
two perpendicular components as shown in Figure 10.

Figure 10: Velocity components of Projectile motion [15]

Adding another dimension gives two differential equations one in the x direction
and the other in the y direction. By Newton’s Second Law we can write this
using a linear drag model:
dvy
m = −(mg + kvy )
dt
dvx
m = −kvx
dt
The signs will vary depending on which direction is taken as positive and which
direction the particle is moving in. For simplicity we will assume the particle
is moving upwards and to the right, which we will take as the positive y and x
directions. There will be a drag force acting on both the horizontal and vertical
components of the motion of the ball and these forces are not the same magni-
tude as they are each dependent on the velocity of their respective component.
This means that different drag models can be used to model each component
dependent on the velocities of each component. For example, for a ball thrown
almost vertically you may use a quadratic model for the x direction and a linear
one for the y direction, giving:
dvy
m = −(mg + k1 vy )
dt

35
 dvx
m = −k2 vx2
dt
To conduct further research on the models already discussed in linear motion, it
would be necessary to conduct another experiment or collect secondary data on
an object in free fall travelling across two dimensions and then solve the x and
y differential equations independently, to then test accuracy as we have already
done.

36
10 Conclusion
In summary, this report has been successful in identifying a model to accurately
describe the motion of a light object in free fall. We have also explained why the
model was successful and in which cases other models would be more suitable,
as well as comparing our research to currently accepted models. Further we
have used various statistical methods to optimise our models.
To better the research, more data should have been collected, something that
was not feasible for us. Alternatively, secondary data could also have been used.
Further, the uncertainty in measurements obtained should also be aimed to be
reduced in order to better optimise any model.

37
11 Bibliography
[1]
Newton, Isaac, 1642-1727. Philosophiæ Naturalis Principia Mathematica.
Londini :Apud G. J. Innys, 1726.
[2]
Encyclopædia Britannica, Inc., Coriolis force — description, Place of
publication: Brittanica, Last updated: 18 Oct. 2025, Cited: 19/11/25
[3]
Dr Ben Yelverton, Coriolis force: deflection of a falling particle [VIDEO],
available from: 04/04/21, URL:
([Link] , Last Accessed:
19/11/25
[4]
J. Bae et al. Introduction of an optimised dynamic installation anchor via
model test and numerical analysis: Free fall in the water column, ELSEVIER
[INTERNET], Published: 2024, Cited: 19/11/25. Available from:
([Link]
schallengef r = RR 1)
[5]
Authored by: OpenStax CNX, 6.4 drag force and Terminal Speed, Published
by Lumen candela, available from: ([Link]
suny-osuniversityphysics/chapter/6-4-drag-force-and-terminal-speed/) Last
accessed: 19/11/25
[6]
Cambridge University Press, free fall — English meaning - Cambridge
Dictionary,Last updated: @2025, last accessed: 19/11/25. (URL:
[Link] fall)
[7]
WikiMedia Ltd. , Free Fall, Wikipedia, Last edited: 28/10/25. Accessed:
19/11/25 (URL: [Link] allaction =
history) 17
[8]
American Institute of Physics, Falling Physics, Last Edited: 2023. Accessed
21/11/25 URL: ([Link]
[9]

38
 [Link], Weight and Mass, Last Edited: N/A [ONLINE]. Accessed
21/11/25 (URL:
[Link]
[10]
R. Ocaya et al., Uncontrolled reentry of Low Earth Orbit decaying objects.
Commun Eng. 2025. Viewed: 21/11/25. (URL:
[Link]
[11]
Dr. Treffor Bazett, Separation of Variables // Differential Equations [VIDEO],
available from: 27/01/21. Last Accesed: 21/11/25. (URL:
[Link]
[12]
Wikipedia Contributors (2019). Archimedes’ principle. [online] Wikipedia.
Available at: ([Link]
[13]
Halliday, D., Resnick, R., Walker, J. (2014). Fundamentals of Physics (10th
ed.)
[14]
Iglewicz B, Hoaglin DC. How to Detect and Handle Outliers. Volume 16,
ASQC Basic References in Quality Control: Statistical Techniques.
Milwaukee, WI: Quality Press; 1993
[15]
Monash Australia, Monash University, Projectile Motion - Student Academic
Success

39
12 Appendix
12.1 Raw Data Values

Figure 5: table for raw data

View Figure 5 above. It contains raw data from the Free fall experiment (Sec-
tion 3) for objects of three different masses dropped from two different heights.
The value for time (t) shown for each mass is a mean value calculated from 3
repeat measurements. Notice that the velocity and acceleration columns have
been left blank - these values are not known.
However, a value for terminal velocity (vt ) could be found by reviewing the
footage of the falling object in slow motion and calculating the time taken for
the object to fall a certain distance. To check if this is at terminal velocity,
repeat this check and determine if the two velocities are the same within an
experimental margin of error.

12.1.1 Data for values of k

Figure 7: Range of values of k for Cubic Drag

40
Figure 8: Range of values for k for Quartic Drag Notice the large range in these
two graphs. This is due both to random error within the data used and the fact
that this error has been ’amplified’ through the large powers used. There is also
no obvious distribution, unlike the linear and quadratic model graphs.

12.2 Deriving ’k’ through solving differential equations

12.2.1 kv and kv2 models

Solve for v(t)

Separate variables:
dv
= dt.
k
g − v2
m
Divide the numerator and denominator by g and introduce u = v/v∞ , noting
p k g
v∞ = mg/k so = 2 . Then
m v∞
Z Z Z Z
dv 1 du g

= dt =⇒ 2
= dt.
g 1 − (v/v∞ )2 v∞ 1−u v∞

The anti derivative is arctanh(u). With v(0) = 0 (so u(0) = 0) we obtain

 v  g
arctanh = t.
v∞ v∞
Apply tanh to invert:
 g t  r mg  gt 
v(t) = v∞ tanh = tanh p .
v∞ k mg/k

v∞ m
r
Define the characteristic time τ = = , then v(t) = v∞ tanh(t/τ ).
g gk

41
Asymptotic checks and small-time expansions
Small t (both models)
As t → 0 both models reduce to free-fall with negligible drag:

v(t) ∼ gt + O(t2 ), s(t) ∼ 12 gt2 + O(t3 ).

Large t
Both models tend to the terminal speed:

v(t) → vt as t → ∞,

and s(t) grows approximately linearly as s(t) ∼ vt t for large t.

Units and notes

• In the linear model k has units of kg/s so that kv is in newtons.
• In the quadratic model k has units of kg/m (or more often written as
1 2
2 ρCd A in aerodynamic usage) so that kv has units of newtons.

• If you prefer upward positive, simply flip the signs on the velocity and
replace g by −g in the differential equations; the magnitude formulas for
terminal speed remain the same.

12.2.2 kv3 andkv 4 models

In these cases it was nescessary to use the taylor series (10.2.3) to approximate
the v(t) and resulting s(t) functions. This gave the equations seen in Section 5.

12.2.3 Modeling Drag using fractional powers of velocity

Up until this point, the research has been focused on the use of integer powers of
velocity to model air resistance. In order to determine the most accurate model,
it is necessary to explore fractional powers. Solving ODEs (ordinary differential
equations) from Newtons 2nd Law involving fractional powers of v is possible
using the partial fractions technique, and in fact this was what was used to get
the t(v) function. The resulting integral was very complex so we were not able
to find a way to write a v(t) function as v in terms of t. It is also not possible
to re-write the t(v) function using the Taylor Series. this is why the Puiseux
series was used, a generalisation of the Laurent Series that allows for fractional
exponents,truncated at the 3rd term. This Series is given by:
n
f (x) = Σ∞
N =n an (x − a)
e

and is similar to the Taylor Series used to approximate the cubic and quartic
models.

42
12.2.4 Using the Taylor Series to generate terms
When trying to find v(t) for different models for air resistance, solving the resul-
tant ODE will result in an integral that cannot be solved exactly. In this case it
is necessary to use numerical methods such as the Taylor series to approximate
the function v(t) into polynomial one.
The Taylor Series is given by:

f ′′ (a) f n (a)
f (x) = f (a) + f ′ (a)(x − a) + + ... + (x − a)n
2! n!
where a is the ’centre’ of the series. In our research, we took t =0 to be the centre
of the expansion (this is the most convenient value to use). This is an aptly
chosen starting point from which to build your expansion. n terms of the series
can be generated with more terms leading to a more accurate representation of
your chosen function. Each new term will have a greater power of x as the (x-a)
bracket is expanded n times. This is significant as in the case where the time
of fall is under 1 second, terms with larger and larger powers will tend to 0:

lim tn = 0, t < 1
n→∞

so a highly accurate approximation of your given function can be obtained with

little terms. Referring to the table of results it can be seen that in our case t ¡
1. This is why in the Taylor expansions shown only 4 terms where generated,
but be aware that this will not always be the case.

12.3 Breakdown of the Statistical approach used

Due to the limitations in equipment available at the time of conducting our
experiment, our data is not entirely accurate and there is a reasonably large
amount of uncertainty in our values because of the resolution of the equipment
used as well as the presence of random error. This is not to say that the data
collected is not valid but that there are clearly anomalies present which skew
the data, and so should not be used when finding a value of k .
As discussed in Section 5.2, we opted to use the MAD as opposed to the variance
from the mean to find anomalies as we have a small dataset and are also unsure
of the distribution of our data. However, using the MAD also has limitations.

Firstly, we used Desmos Graphing Software to compute the 6 values of k for

each model. This inherently loses some accuracy as variables are only generated
to some accuracy rather than their true value, which may be irrational. This
effect is more noticeable in the cubic and quartic models which require a higher
degree of accuracy as the different values of k are very small.

Further, the MAD also suffers from the same pitfalls as standard deviation but
to a lesser degree. When calculating a standardised score, the data is warped

43
into a normal distribution. It is important to note that unlike for standard de-
viation, this distribution is not assumed, although it might not be the best way
to describe the spread of the data. Although this seems accurate for the linear
model, as seen in Figure five it is unclear whether this can be said for some of
the other models which have a wider spread. When defining an anomaly, we
chose to use a value of 3.5 as a benchmark. Arguably, we should have used all
the data when considering an approximate value for k as removing even 1/6 of
the data will give a result that is not representative.

Finally, we opted to use the mean rather than the median to calculate an av-
erage, as this would include all the non-anomalous data rather than just one
point. However this decision is reliant on the fact that the median mean, which
may not be true. Upon studying the graphs we find that this is only loosely
true in some cases and so using the mean may not be the best option. To ensure
the different models are comparable, however, we must use the mean for all.
To summarise, although using the MAD ensured a good estimate for the ’true
value’ for the coefficient of air resistance of each model. The inherent qualities
of the empirical data used did cause some possible inaccuracies in the values
obtained. This means that although the results of the report are still valid, they
may not be entirely true. For example, if we find that the quartic model does
not fit our data well, although we can deduce that it is not the best option,
we can not deduce by how much. This is because of confounding; we may be
attributing the model’s poor or inaccurate predictions compared to the data to
faults in the model itself, when in reality the value of k used could also have
increased the inaccuracy of the model, and a different value closer to the ’true
value’ would have been better.

12.3.1 Checking k values

Checking k values is essential as the biggest weakness of the statistical approach
was the fact that the solutions for k were obtained graphically by sight so may
have been vastly incorrect. Checking the final estimate ensured it was accurate,
and therefore acceptable to use when comparing models. This is essential as it
would have otherwise resulted in confounding - where a model could have been
labelled as worse than another when really it was the k value that was not valid.
We also considered using Euler’s method to approximate the roots and compare
these to our estimated values. However, we found it was easier to substitute the
estimated values for k into the velocity equations with t = 0.0001.

12.4 Acknowledgements
We would like to thank our mentor, Dr Sam Brzezicki, for taking the time to
help us to conduct our research and write this report with his expertise. We
would also like to express gratitude to our supervisor, Raj Velusamy, for his
consistent help and support throughout the project.

44