MUSIC algorithm for Young Mathematicians and

Engineers
Raveen De Silvaa, Amir Antonira, Aleksandar Ignjatovica, and Chamith Wijenayakeb
School of Computer Science, University of New South Wales, Sydney, Australiaa
School of Information Technology and Electrical Engineering, University of Queensland, Brisbane, Australiab
Corresponding Author Email: [email protected]

ABSTRACT

CONTEXT
Training of young mathematicians most often focuses only on mathematical and computational
methods, without giving students a deeper insight into foundations and context of engineering,
which is the most common source of the problems they will be tackling. On the other hand,
engineering students are often taught how to manipulate mathematical formulas without having a
deeper understanding of the underlying first principles. We believe that there are two reasons for
such an unhappy state of modern education. (1) In the modern world, we have become super-
specialised, and for one individual to frequently cross the boundaries between fields is not
generally perceived as a good career move. Education is often reduced to professional training,
where the aim is not to nourish deeper understanding, but just to acquire sufficient skills to perform
certain tasks. (2) Specialist textbooks are often replete with discipline specific technical jargon.

PURPOSE OR GOAL
While one might see this situation as the most cost-effective way to provide industry with workforce
in the short term, we believe that this approach will eventually erode our ability to make paradigm
shifts. Perhaps it is time to take a bit of the attitude of the Renaissance and of Enlightenment and
give our education just a broader, more humanistic focus than just providing a skill set.

APPROACH OR METHODOLOGY/METHODS
As an example for how this can be accomplished, we offer our presentation of the MUSIC (Multiple
Signal Classification) frequency estimation algorithm, free of any signal processing jargon, not
requiring absolutely any knowledge of signal processing but only knowledge of basic linear
algebra. We believe that such an approach can give a small hint how to bridge the chasm between
education of mathematicians and education of signal processing engineers.

ACTUAL OR ANTICIPATED OUTCOMES

We hope it might help young mathematicians appreciate how spectacularly mathematics is applied
in signal processing and we hope that we might surprise young signal processing engineers that
one can understand functioning of a key signal processing algorithm relying only on linear algebra.

CONCLUSIONS/RECOMMENDATIONS/SUMMARY
Engineering education should be more integrated, unifying teaching a good deal of science with
teaching practical engineering. We must resist the pressures to be “the industry of higher
education” and be what we used to be – academia, educating not only practitioners of
sophisticated skills but also independent and creative thinkers and innovators.

KEYWORDS
Science in engineering, wholistic education, thinkers, innovators
Introduction
Education for Mathematics and Engineering students often have different focus, thus creating two
different worlds for mathematicians and engineers. Students in Engineering often taught to use and
manipulate various mathematical formulas with omission of proofs and first principles. Students in
mathematics often learn underlying first principles, proofs, mathematical and computational
methods with limited exposure in the context of engineering problems. Engineering textbooks, as a
rule, try to avoid excursions into mathematical theory as much as possible, while mathematical
textbooks usually state the relevant equations without placing them into a proper engineering
context. In fact, we feel that engineers and mathematicians mostly live in two separate worlds,
often unwilling to cross the boundaries between fields or at least trying to minimise such
excursions. Examples are abundant: young mathematicians most often learn harmonic analysis
without even a mention of one of the most spectacular applications of harmonic analysis, namely
signal processing. On the other hand, young electrical engineers are most often taught the
Discrete Fourier Transform (DFT) by simply giving them the relevant formulas, without ever telling
them that the DFT amounts to a change of basis of the underlying signal space.
As an approach to reduce this disconnect, we explore the possibility of redesigning learning
resources (such as self-paced asynchronous tutorials), which take a commonly used textbook
application from Engineering (e.g., a signal processing algorithm from Electrical Engineering) and
explain the underlying theory and operation, purely based on mathematical details (such as linear
algebra) without any prior knowledge in signal processing. As an example, we present the MUSIC
(Multiple Signal Classification) and the root-MUSIC frequency estimation algorithms used in
standard signal processing, essentially without any reference to traditional signal processing
concepts and without using any signal processing terminology, thus making these algorithms
accessible to all students who have studied basic linear algebra, regardless of their field of study.
Students of signal processing can also benefit from reading this tutorial which presents very clearly
the underlying mathematical foundations of these algorithms.
While the rest of this paper takes the form a mathematical tutorial, the main objective of this paper
is to demonstrate how sophisticated technical algorithms in Engineering can be presented in a
manner making them accessible to a wide audience, bypassing technical jargon and often
requiring minimal background preparation.

Background
In the modern university, students often specialise very early in their studies, and have little
substantial engagement with other fields. Henderson and Broadbridge (Henderson & Broadbridge,
2009) describe the state of engineering mathematics education in Australia, including the
challenges presented by the diversity of mathematical experience and proficiency in the student
body, the need to incorporate computing and statistics, and administrative issues such as class
sizes, assessment and budgets. The content and skills taught in these courses are fundamental to
the quantitative methods used in various engineering disciplines, as discussed by Maass et al.
(Maass, Geiger, Romero Ariza, & Goos, 2019) who write that “STEM education in general, and
mathematics education in particular, can contribute to preparing individuals better for twenty-first
century challenges”. However, it is well documented that many engineering students struggle in
these mathematics courses, and they often perceive the content as unnecessarily abstract and of
little relevance to their engineering studies. This issue has received much attention in the literature,
such as case studies of the problem-based learning approach by Flegg et al. (Flegg, Mallet, &
Lupton, 2012) and Bischof et al. (Bischof, Bratschitsch, Casey, & Rubesa, 2007). In addition,
authors such as Ooi (Ooi, 2007) and Klingbeil et al. (Klingbeil, Mercer, Rattan, Raymer, &
Reynolds, 2004) write on the matters of what mathematics should be presented to engineering
students and at what stage of their studies, with the latter advocating for the traditional calculus
sequence to be delayed in favour of teaching freshmen “only the math they really needed to know
in order to progress into their sophomore and junior years”.

Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.
In this paper, we primarily examine the converse problem: that of demonstrating engineering
applications to mathematics students who may be otherwise more interested in mathematics for its
own sake. This is much less widely studied, but we believe it to be important in order to both
diversify the interests and expertise of mathematics students and foster effective interdisciplinary
collaboration. Many mathematics students are not aware of the power of the material they study in
solving real-world problems, and here, we demonstrate how they might be introduced to the
application of linear algebra in signal processing, without any of the jargon from signal processing,
that might be off-putting to a student without any formal training in electrical engineering.
The rest of the paper presents our example tutorial of MUSIC algorithm based on linear algebra.

Example Tutorial on MUSIC Algorithm

The roots of the MUSIC (Multiple Signal Classification) algorithm for frequency estimation of real
sinusoids or complex exponentials lie in the early work of Prony (Prony, 1795) and Pisarenko
(Pisarenko, 1973) which we present here very briefly and in a very simplified manner.

The methods of Prony and Pisarenko

Let a signal 𝑠(𝑡) be a linear combination of 𝑛 complex exponentials, i.e., of the form
)

𝑠(𝑡) = ' 𝐴! 𝑒 "($! %&'! ) .

!*+
(1)

Here, 𝑗 is the imaginary unit and real numbers 𝐴! , 𝜔! and 𝜑! are the amplitude, the frequency,
and the phase of the 𝑘 "# component, respectively; we assume that 𝐴! > 0 and −𝜋 < 𝜔! < 𝜋,
0 < 𝜑! < 2𝜋 for all 1 ≤ 𝑘 ≤ 𝑛.
We also assume that there exists a sequence of samples 𝑠(𝑡 + 𝑚) of such a signal, taken at
consecutive instants a unit distance apart, starting with an instant t. Let us now form a linear
combination of n + 1 such consecutive samples with coefficients 𝑐, , … , 𝑐) to be specified below.
Using (1)Error! Reference source not found., after some simplification we obtain
) ) ) ) )
-
' 𝑐- 𝑠(𝑡 + 𝑚) = ' 𝑐- ' 𝐴! 𝑒 "($! (%&-)&'! ) = ' 2 ' 𝑐- 3𝑒 "$! 4 5 𝐴! 𝑒 "($! %&'! )
-*, -*, !*+ !*+ -*,

(2)
Consider now a polynomial 𝑃(𝑧) with the leading coefficient 1, given by the product 𝑃(𝑧) =
3𝑧 − 𝑒 "$" 4 … 3𝑧 − 𝑒 "$# 4. Let 𝑐, , … , 𝑐) be the coefficients of this polynomial, such that
) )
-
' 𝑐- 𝑧 = 93𝑧 − 𝑒 "$! 4.
-*, !*+

Then, since the right-hand side of (2) is of the form ∑)!*+ 𝑃3𝑒 "$! 4 𝐴! 𝑒 "($! %&'! )
and since 𝑒 "$" , … , 𝑒 "$# are the roots of 𝑃(𝑧), the right-hand side (2) of will be equal to zero for all
real t, and vice versa: since complex exponentials with distinct frequencies are linearly
independent functions, if the right side of (2) is equal to zero for all real t, then 𝑒 "$! must be the
roots of the polynomial 𝑃(𝑧). Consequently, in order to find the unknown frequencies 𝜔! it is
enough to find then coefficients 𝑐, , … , 𝑐) such that for all 𝑡,
)

' 𝑐- 𝑠(𝑡 + 𝑚) = 0
-*,

(3)

Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.
and then find the roots of the associated algebraic equation ∑)-*, 𝑐- 𝑧 - = 0; such roots lie on the
unit circle and their arguments are the frequencies sought.
To find a non-zero vector 𝒄 = 𝑐, , … 𝑐) such that
(3) holds for all t, we instantiate
(3) from 𝑡 = 1 to = 𝑛, and if 2𝑛 samples 𝑠(1), 𝑠(2), … 𝑠(2𝑛) of the signal 𝑠(𝑡) are available, we
obtain a system of linear equations in unknown coefficients 𝑐, , … 𝑐) of the form
𝑠(1) 𝑠(2) ⋯ 𝑠(𝑛 + 1) 𝑐,
⎛ 𝑠(2) 𝑠(3) ⋯ 𝑠(𝑛 + 2) ⎞ ⎛ 𝑐+ ⎞
⋱
⎜ ⋮ ⋮ ⋮ ⎟⎜ ⋮ ⎟ = 𝟎
𝑠(𝑛 − 1) 𝑠(𝑛) ⋯ 𝑠(2𝑛 − 1) 𝑐).+
⎝ 𝑠(𝑛) 𝑠(𝑛 + 1) ⋯ 𝑠(2𝑛) ⎠ ⎝ 𝑐) ⎠
(4)
which, in absence of any noise, we can solve exactly. This is the Prony method, dating back to
year 1795 [Error! Reference source not found.], which is summarised in Fig. 1.

Figure 1: a summary of the Prony method.

In the presence of noise

(3) will not hold exactly; to deal with this problem, we make the system of equations (4)
overdetermined. Thus, we will assume that we have 𝑁 > 2𝑛 equidistant samples of the signal
𝑠(𝑡). Let us first consider the noise-free case, i.e., assume that 𝑠(𝑡) is as in (1) and let us form the
following Hankel matrix 𝐌/0 of size (𝑁 – 𝑛) × (𝑛 + 1):
𝑠(1) 𝑠(2) ⋯ 𝑠(𝑛 + 1)
⎛ 𝑠(2) 𝑠(3) ⋯ 𝑠(𝑛 + 2)⎞
⋯ 𝑠(𝑛 + 3) .
𝐌/0 = ⎜ 𝑠(3) 𝑠(4) ⎟
⋮ ⋮ ⋱ ⋮
⎝ 𝑠 (𝑁 − 𝑛 ) 𝑠(𝑁 − 𝑛 + 1) ⋯ 𝑠 (𝑁 ) ⎠
(5)
If no noise were present,
(3) would imply that the rank of this matrix is equal to 𝑛. Consequently, if we consider the singular
value decomposition of this matrix, 𝐌/0 = UΣ𝐕 ∗ , the diagonal of Σ would consist of 𝑛 + 1 singular
values of 𝐌/0 , out of which there would be 𝑛 non-zero singular values, while the last, which is the
smallest of the singular values, would be equal to 0. This is true because singular values on the
diagonal of Σ are always non-negative reals and are usually ordered in a descending order; thus,
the smallest singular value is the rightmost one on the diagonal of Σ. Matrices 𝐔 and 𝐕 are both
unitary, i.e., their columns represent a set of orthonormal vectors, called the left (the right) singular
vectors, respectively. Let 𝐯)&+ be the rightmost singular vector which corresponds to the zero
singular value of 𝐌/0 . Then, since 𝐕 is unitary, 𝐕 ∗ 𝐯)&+ = (0,0, … . . ,0,1)2 and, since the last entry on
the diagonal of 𝛴 is zero, this is easily seen to imply that 𝐌/0 𝐯)&+ = 𝐔𝚺𝐕 ∗ 𝐯)&+ = 0. Thus, the
components of the rightmost singular vector 𝐯)&+ corresponding to the singular value zero are the
required coefficients of a linear combination of the columns of the matrix 𝐌/0 which is equal to the
zero vector.

Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.
Assume now that we have 𝑁 > 2𝑛 equidistant samples 𝑓(1), . . . , 𝑓(𝑁) of a signal 𝑓(𝑡) = 𝑠(𝑡) +
3
𝜈(𝑡) which is a sum of a signal 𝑠(𝑡) as in (1) and noise 𝜈(𝑡). We can still form a Hankel matrix 𝐌/
obtained from matrix 𝐌/0 by replacing samples of 𝑠(𝑡) by the corresponding noisy samples of 𝑓(𝑡).
3
Due to the presence of noise, 𝐌/ will generally have a full rank of 𝑛 + 1, and thus all singular
3
values of 𝐌/ will be non-zero. We now take the rightmost singular vector, which corresponds to
the smallest singular value, as an approximation of the rightmost singular vector if no noise were
present, and its components 𝐯)&+ (𝑚 + 1) as approximations of the values of 𝑐- for which
equation (3) holds.
3
As is well known, the right singular vectors of matrix 𝐌/ are the eigenvectors of the product matrix
3 3 ∗ 3
𝐀/ = 3𝐌/ 4 𝐌/ , where 𝐌∗ denotes the conjugate transpose of 𝐌. Note that, in our particular case,
3 3
(1/𝑁)𝐀/ is just the auto-covariance matrix of the samples of the noisy signal 𝑓(𝑡). Since matrix 𝐀/
3
is of size (𝑛 + 1) × (𝑛 + 1) and since 𝑁 is usually much larger than 𝑛, 𝐀/ is of much smaller size
3 3
than 𝐌/ . Consequently, finding the eigen decomposition of 𝐀/ is a computationally lighter task
3
than finding the singular value decomposition of 𝐌/ . Note that this benefit, however, is often offset
3 3 ∗ 3
by the cost of the computation of the matrix product 𝐀/ = 3𝐌/ 4 𝐌/ . In this way we obtain the
Pisarenko frequency estimation algorithm, which can be summarised as follows and in Fig. 2:
3 3 ∗ 3 3
Compute the matrix 𝐀/ = 3𝐌/ 4 𝐌/ and obtain its eigenvalue decomposition 𝐀/ = 𝐐𝚲𝐐∗ . Take the
3
rightmost eigenvector 𝐪)&+ which corresponds to the smallest eigenvalue of 𝐀/ and solve the
associated algebraic equation ∑)-*, 𝐪)&+ (𝑚 + 1)𝑧 - = 0; the arguments of its 𝑛 roots are taken
as estimates of the unknown frequencies 𝜔! of the complex exponentials which are the 𝑛
components of the signal 𝑠(𝑡).

Figure 2: a summary of the Pisarenko method.

MUSIC and root-MUSIC Algorithms

It is well known that the Pisarenko method explained above is not very noise robust and that it
often has quite a poor performance. The MUSIC algorithm is a generalization of the Pisarenko
method, which significantly improves its noise robustness; it is essentially an averaging procedure
of multiple estimates of the frequencies obtained by the Pisarenko method.
0
To obtain the MUSIC algorithm, let us again first consider the noise-free case and let 𝐌- be the
Hankel matrix of shifted consecutive samples with a possibly larger number of columns 𝐾 ≥ 𝑛 +
1, thus of size (𝑁 − 𝐾 + 1) × 𝐾,
𝑠(1) 𝑠(2) ⋯ 𝑠(𝐾)
⎛ 𝑠(2) 𝑠(3) ⋯ 𝑠(𝐾 + 1)
0
𝐌- = ⎜ 𝑠(3) 𝑠(4) ⋯ 𝑠(𝐾 + 2)⎞
⋱ ⎟
⋮ ⋮ ⋮
⎝𝑠(𝑁 − 𝐾 + 1) 𝑠(𝑁 − 𝐾 + 2) ⋯ 𝑠(𝑁) ⎠
(6)
0
and again consider its singular value decomposition, 𝐌- = UΣ𝐕 ∗ . In absence of noise equation (3)
again implies that the rank of this matrix would be equal to 𝑛 and thus, only 𝑛 singular values on
the diagonal of 𝚺 would be non-zero. Just as in the case of the Pisarenko method, this would imply
that for 𝐾 − 𝑛 right singular vectors 𝐯4 , (𝑛 < 𝑖 ≤ 𝐾), which correspond to 𝐾 − 𝑛 zero singular

Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.
 0 0
values of 𝐌- , we would have 𝐌- 𝐯4 = UΣ𝐕 ∗ 𝐯𝒊 = 0. Thus, each of the right singular vectors 𝑣4 , 𝑛 <
𝑖 ≤ 𝐾, produces an equation of the form
6.+

' 𝐯4 (𝑚 + 1)𝑓(𝑡 + 𝑚) = 0
-*,

that holds for samples 𝑓(𝑡), . . . , 𝑓(𝑡 + 𝐾 − 1) for all integers 1 ≤ 𝑡 ≤ 𝑁 − 𝐾 + 1. Consequently, in the
absence of any noise, for every 𝑛 < 𝑖 ≤ 𝐾 and all 1 ≤ 𝑡 ≤ 𝑁 − 𝐾 + 1, equation (2) would imply
6.+ ) 6.+
"'!
' 𝐯4 (𝑚 + 1)𝑓(𝑡 + 𝑚) = ' 𝐴! 𝑒 2 ' 𝐯4 (𝑚 + 1)(𝑒 "$! )- 5 𝑒 "$$ % = 0.
-*, !*+ -*,

(7)
%
If 𝑁 − 𝐾 + 1 ≥ 𝑛, since the Vandermonde matrix 𝐕 = f3𝑒 "$! 4 ∶ 1 ≤ 𝑘, 𝑡 ≤ 𝑛i is always
nonsingular if all 𝜔! are distinct, equation (7) would imply that 𝑒 "$! must be among the 𝐾 − 1
many roots of each of the polynomials 𝑃4 (𝑧), where
6.+

𝑃4 (𝑧) = ' 𝐯4 (𝑚 + 1)𝑧 - , 𝑛 < 𝑖 ≤ 𝐾.

-*,
(8)
Thus, polynomials 𝑃4 (𝑧) for all 𝑛 < 𝑖 ≤ 𝐾 share the same 𝑛 roots 𝑒 "$" , … , 𝑒 "$# which belong to
the unit circle, and each of the polynomials 𝑃4 (𝑧) has additional 𝐾 − 1 − 𝑛 roots generally not
belonging to the unit circle and which are different and specific to each polynomial 𝑃4 (𝑧). Thus, if
we consider the real valued function
6 6

𝐹(𝑧) = ' |𝑃4 (𝑧)|7 llllll

= ' 𝑃4 (𝑧)𝑃8 (𝑧) ,
4*)&+ 4*)&+
(9)
where 𝑧! denotes the complex conjugation, then this function will have 𝑛 of its zeros lying on the
unit circle, namely 𝑒 9$! for 1 ≤ 𝑘 ≤ 𝑛 and the arguments of these zeros are the unknown
frequencies sought.
Clearly, the above no longer holds in the presence of noise. If 𝑓(𝑡) = 𝑠(𝑡) + 𝜈(𝑡) where 𝑠(𝑡) is of
the form given by the right hand side of equation (1) and 𝜈(𝑡) is noise, then we can form matrix
3 0
𝐌- of the same form as matrix 𝐌- but with the samples the noisy signal 𝑓(𝑡) in place of the
3
corresponding samples of the noise-free signal 𝑠(𝑡). However, such a matrix 𝐌- will generally be
of full rank, but we can take its 𝐾 − 𝑛 right singular vectors 𝐯4 which correspond to the smallest 𝐾 −
3
𝑛 singular values of 𝐌- as an approximation of such singular vectors, which would correspond to
0
the noise-free matrix 𝐌- . We can again form the corresponding polynomials 𝑃4 (𝑧) given by
equation (8) and function 𝐹(𝑧) given by equation (9). Note that 𝐹(𝑧) is not a polynomial, due to the
presence of the modulus (or the complex conjugation) function. Due to the effects of noise, no
roots of polynomials 𝑃4 (𝑧) might belong to the unit circle, and also these polynomials might not
"$ 7
share the exact same 𝑛 roots. Thus, function Φ(𝜔) = 𝐹3𝑒 "$ 4 = ∑6 4*)&+n𝑃4 (𝑒 )n might not have
any real zeros belonging to the interval [−𝜋, 𝜋]. Since Φ(𝜔) > 0, the MUSIC algorithm thus instead
searches for the 𝑛 values of 𝜔 which lie in the interval [−𝜋, 𝜋], where Φ(𝜔) attains 𝑛 smallest local
minima. Fig.4(a) illustrates the equivalent but numerically more convenient search for the 𝑛 largest
local peaks of the reciprocal function
1
𝑅(𝜔) = .
∑6
4*)&+|𝑃4 (𝑒 )|
"$ 7

Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.
 (10)
To avoid such a numerical search, the root-MUSIC algorithm instead uses the fact that complex
numbers 𝑧 which are close to the unit circle satisfy 𝑧̅ ≈ 𝑧 .+ and thus instead explicitly solves the
following equation, conveniently reducible to an algebraic (i.e., polynomial) equation
6
u8 (𝑧 .+ ) = 0
' 𝑃4 (𝑧)𝑃
4*)&+

(11)
! 𝑖 (𝑧) denotes the polynomial obtained from the polynomial 𝑃𝑖 (𝑧) by taking the complex
where 𝑃
conjugates of the coefficients of 𝑃𝑖 (𝑧).
It is clear that if 𝑧4 is a root of equation (11) then so is 𝑧lllll
.+ ; thus, the roots of (11) come in pairs of
.+
the form: v𝑝4 𝑒 "$ , 𝜌4 𝑒 "$ y. The root-MUSIC algorithm picks 𝑛 pairs of such roots that lie closest to
the unit circle; the arguments of these pairs are taken as the estimates of the frequencies 𝜔! (1 ≤
𝑘 ≤ 𝑛) of the 𝑛 components of 𝑠(𝑡), as illustrated in Fig.4(b)).
The interested reader can find more details in (Stoica & Moses, 1979) (Pisarenko, 1973).

Figure 3: a summary of the MUSIC and root-MUSIC algorithms.

(a) (b)

Required roots of
equation (11)

Figure 4: (a) Example peak profile of function 𝑹(𝝎) given in equation (10); (b) illustration of roots of
equation (11) as pairs of complex numbers lying close to unit circle.

Discussion
The tutorial explanations given above were presented to a group of students from which 7 survey
responses were recorded. Although the response rate to the feedback survey is quite low, 70% of
Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.
the responses agreed that “the tutorial provides a clear description about the MUSIC algorithm,”
and 56% agreed that “I obtained a good understanding about the MUSIC algorithm after going
through the tutorial.” Survey participants also thought that providing more visual interpretation of
some of the linear algebra operations will be helpful from a student perspective.

Where traditional methods to teach signal processing algorithms are grounded in terminology from
electrical engineering and aim to solve concrete problems, ours is deliberately abstract. We seek to
first demonstrate to mathematics students how the singular value decomposition can be used to
identify the dominant frequencies in a noisy signal numerically, to be later supplemented by
discussion of the application of these methods in digital signal processing. We believe that this
approach is different to traditional classroom teaching of such signal processing algorithms.

We envision that this material could be presented in courses on linear algebra, which often
introduce the singular value decomposition, but may not contextualise it by demonstrating its
application to practical engineering problems. By first relating the theory to a challenging
mathematical problem, we aim to inspire mathematics students to consider how their work is
related to that undertaken in other disciplines.

This tutorial could also be applied to courses on mathematical computing, as the implementation
and analysis of these algorithms in a software package such as MATLAB is an instructive exercise
and provides a tangible outcome from linear algebra subroutines.

Conclusion and Future Work

This tutorial presented an approach to reduce the gap between engineering and mathematics
education by demonstrating an example design of a mathematics tutorial in the context of electrical
engineering. The specific example involved taking an alternative presentation of the conventional
MUSIC frequency estimation algorithm used in signal processing, albeit with no signal processing
jargon and prior knowledge, but only using the knowledge from linear algebra. Future work
includes using employing further visual illustrations of the associated multidimensional vector
spaces via interactive simulations and plots, as well as implementing the tutorial in a linear algebra
or mathematical computing course.

References
Bischof, G., Bratschitsch, E., Casey, A., & Rubesa, D. (2007). Facilitating Engineering Mathematics
Education by Multidisciplinary Projects. Paper presented at the ASEE Annual Conference & Exposition,
Honolulu, Hawaii.
Flegg, J., Mallet, D., & Lupton, M. (2012). Students' perceptions of the relevance of mathematics in
engineering. International Journal of Mathematical Education in Science and Technology, 717-732.
Henderson, S. & Broadbridge, P. (2009). Engineering Mathematics Education in Australia. MSOR
Connections, 9(1), 12-17.
Klingbeil, N., Mercer, R., Rattan, K., Raymer, M., & Reynolds, D. (2004). Rethinking Engineering
Mathematics Education: A Model for Increased Retention, Motivation and Success in Engineering. Paper
presented at the ASEE Annual Conference & Exposition, Salt Lake City, Utah.
Maass, K., Geiger, V., Romero Ariza, M., & Goos, M. (2019). The Role of Mathematics in Interdisciplinary
STEM education. ZDM, 51, 869-884.
Ooi, A. (2007). An analysis of the Teaching of Mathematics in Undergraduate Engineering Courses. Paper
presented at the 18th Conference of the Australasian Association for Engineering Education, Melbourne,
Victoria.
Pisarenko, V. F. (1973). The retrieval of harmonics from a covariance function. Geophysical Journal
International, 33, 347-366.
Prony, R. (1795). Essai experimentale et analytique. J. Ecole Polytechnique, 2, 24-76.

Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.
Stoica, P. & Moses, R. (1979). Introduction to Spectral Analysis (1st ed.). Upper Saddle River, NJ: Prentice
Hall.

Copyright statement

Copyright © 2022 Raveen De Silva, Amir Antonir, Aleksandar Ignjatovic, Chamith Wijenayake: The authors assign to the Australasian
Association for Engineering Education (AAEE) and educational non-profit institutions a non-exclusive licence to use this document for
personal use and in courses of instruction provided that the article is used in full and this copyright statement is reproduced. The authors
also grant a non-exclusive licence to AAEE to publish this document in full on the World Wide Web (prime sites and mirrors), on
Memory Sticks, and in printed form within the AAEE 2022 proceedings. Any other usage is prohibited without the express permission of
the authors.

Proceedings of AAEE 2022 Western Sydney University, Sydney, Australia, Copyright © Raveen De Silva, Amir Antonir, Aleksandar
Ignjatovic, Chamith Wijenayake, 2022.