Advanced Mechanical Vibrations - Physics, Mathematics and Applications

Advanced Mechanical
Vibrations
Advanced Mechanical
Vibrations
Physics, Mathematics and
Applications
Paolo Luciano Gatti

First edition published 2021
by CRC Press
2 Park Square, Milton Park, Abingdon, Oxon, OX14 4RN
and by CRC Press

6000 Broken Sound Parkway NW, Suite 300, Boca Raton, FL 33487-2742
© 2021 Paolo Luciano Gatti
CRC Press is an imprint of Informa UK Limited
The right of Paolo Luciano Gatti to be identified as author of this work has been asserted by him
in accordance with sections 77 and 78 of the Copyright, Designs and Patents Act 1988.
All rights reserved. No part of this book may be reprinted or reproduced or utilised in any form
or by any electronic, mechanical, or other means, now known or hereafter invented, including
photocopying and recording, or in any information storage or retrieval system, without
permission in writing from the publishers.
For permission to photocopy or use material electronically from this work, access www.
copyright.com or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive,
Danvers, MA 01923, 978-750-8400. For works that are not available on CCC please contact
[email protected]
Trademark notice: Product or corporate names may be trademarks or registered trademarks, and
are used only for identification and explanation without intent to infringe.
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library
Library of Congress Cataloging‑in‑Publication Data

Names: Gatti, Paolo L., 1959- author.
Title: Advanced mechanical vibrations : physics, mathematics and
applications / Paolo Luciano Gatti.
Description: First edition. | Boca Raton : CRC Press, 2021. |
Includes index.
Identifiers: LCCN 2020031525 (print) | LCCN 2020031526 (ebook) |
ISBN 9781138542280 (hardback) | ISBN 9781351008600 (ebook) |
ISBN 9781351008587 (epub) | ISBN 9781351008570 (mobi) |
ISBN 9781351008594 (adobe pdf)
Subjects: LCSH: Vibration.
Classification: LCC TA355 .G375 2021 (print) | LCC TA355 (ebook) |
DDC 620.3–dc23
LC record available at https://lccn.loc.gov/2020031525
LC ebook record available at https://lccn.loc.gov/2020031526
ISBN: 978-1-138-54228-0 (hbk)

ISBN: 978-1-351-00860-0 (ebk)
Typeset in Sabon
by codeMantra
To my wife Simonetta and my daughter
Greta J., for all the future ahead.
And in loving memory of my parents Paolina and Remo
and my grandmother Maria Margherita, a person
certainly endowed with the ‘wisdom of life’.
Contents
Preface xi
Acknowledgements xiii
Frequently used acronyms xv
1 A few preliminary fundamentals 1

1.1 Introduction 1
1.2 Modelling vibrations and vibrating systems 1
1.3 Some basic concepts 3
1.3.1 The phenomenon of beats 5
1.3.2 Displacement, velocity, acceleration and decibels 6
1.4 Springs, dampers and masses 8
2 Formulating the equations of motion 13

2.1 Introduction 13
2.2 Systems of material particles 14
2.2.1 Generalised co-ordinates, constraints
and degrees of freedom 15
2.3 Virtual work and d’Alembert’s principles –
Lagrange and Hamilton equations 16
2.3.1 Hamilton’s equations (HEs) 20
2.4 On the properties and structure of Lagrange’s equations 24
2.4.1 Invariance in the form of LEs
and monogenic forces 24
2.4.2 The structure of the kinetic energy
and of Lagrange equations 24
2.4.3 The energy function and the conservation of energy 28
2.4.4 Elastic forces, viscous forces and
Rayleigh dissipation function 29
vii
viii Contents
2.4.5 More co-ordinates than DOFs:

Lagrange’s multipliers 32
2.5 Hamilton’s principle 34
2.5.1 More than one independent variable:
continuous systems and boundary conditions 38
2.6 Small-amplitude oscillations 44
2.7 A few complements 48
2.7.1 Motion in a non-inertial frame of reference 48
2.7.2 Uniformly rotating frame 51
2.7.3 Ignorable co-ordinates and the Routh function 53
2.7.4 The Simple pendulum again: a
note on non-small oscillations 56
3 Finite DOFs systems: Free vibration 59

3.1 Introduction 59
3.2 Free vibration of 1-DOF systems 59
3.2.1 Logarithmic decrement 65
3.3 Free vibration of MDOF systems: the undamped case 67
3.3.1 Orthogonality of eigenvectors and normalisation 68
3.3.2 The general solution of the undamped
free-vibration problem 70
3.3.3 Normal co-ordinates 72
3.3.4 Eigenvalues and eigenvectors sensitivities 78
3.3.5 Light damping as a perturbation
of an undamped system 80
3.3.6 More orthogonality conditions 82
3.3.7 Eigenvalue degeneracy 83
3.3.8 Unrestrained systems: rigid-body modes 84
3.4 Damped systems: classical and non-classical damping 87
3.4.1 Rayleigh damping 88
3.4.2 Non-classical damping 90
3.5 GEPs and QEPs: reduction to standard form 92
3.5.1 Undamped Systems 93
3.5.2 Viscously damped systems 94
3.6 Eigenvalues sensitivity of viscously damped systems 96
4 Finite-DOFs systems: Response to external excitation 99

4.1 Introduction 99
4.2 Response in the time-, frequency- and s-domains:
IRF, Duhamel’s integral, FRF and TF 100
Contents ix
4.2.1 Excitation due to base displacement,

velocity or acceleration 105
4.3 Harmonic and periodic excitation 107
4.3.1 A few notes on vibration isolation 110
4.3.2 Eccentric excitation 112
4.3.3 Other forms of FRFs 114
4.3.4 Damping evaluation 116
4.3.5 Response spectrum 117
4.4 MDOF systems: classical damping 120
4.4.1 Mode ‘truncation’ and the
mode-acceleration solution 122
4.4.2 The presence of rigid-body modes 125
4.5 MDOF systems: non-classical viscous
damping, a state-space approach 126
4.5.1 Another state-space formulation 129
4.6 Frequency response functions of a 2-DOF system 133
4.7 A few further remarks on FRFs 137
5 Vibrations of continuous systems 139

5.1 Introduction 139
5.2 The Flexible String 140
5.2.1 Sinusoidal waveforms and standing waves 142
5.2.2 Finite strings: the presence of
boundaries and the free vibration 143
5.3 Free longitudinal and torsional vibration of bars 148
5.4 A short mathematical interlude: Sturm–Liouville problems 150
5.5 A two-dimensional system: free
vibration of a flexible membrane 156
5.5.1 Circular membrane with fixed edge 158
5.6 Flexural (bending) vibrations of beams 162
5.7 Finite beams with classical BCs 163
5.7.1 On the orthogonality of beam eigenfunctions 167
5.7.2 Axial force effects 168
5.7.3 Shear deformation and rotary
inertia (Timoshenko beam) 170
5.8 Bending vibrations of thin plates 174
5.8.1 Rectangular plates 176
5.8.2 Circular plates 180
5.8.3 On the orthogonality of plate eigenfunctions 181
5.9 A few additional remarks 182
x Contents
5.9.1 Self-adjointness and positive-definiteness

of the beam and plate operators 182
5.9.2 Analogy with finite-DOFs systems 185
5.9.3 The free vibration solution 188
5.10 Forced vibrations: the modal approach 190
5.10.1 Alternative closed-form for FRFs 199
5.10.2 A note on Green’s functions 201
6 Random vibrations 207

6.1 Introduction 207
6.2 The concept of random process, correlation
and covariance functions 207
6.2.1 Stationary processes 212
6.2.2 Main properties of correlation
and covariance functions 214
6.2.3 Ergodic processes 216
6.3 Some calculus for random processes 219
6.4 Spectral representation of stationary random processes 223
6.4.1 Main properties of spectral densities 227
6.4.2 Narrowband and broadband processes 229
6.5 Response of linear systems to
stationary random excitation 232
6.5.1 SISO (single input–single output) systems 233
6.5.2 SDOF-system response to broadband excitation 236
6.5.3 SDOF systems: transient response 237
6.5.4 A note on Gaussian (normal) processes 239
6.5.5 MIMO (multiple inputs–multiple outputs) systems 241
6.5.6 Response of MDOF systems 243
6.5.7 Response of a continuous system to distributed
random excitation: a modal approach 245
6.6 Threshold crossing rates and peaks distribution
of stationary narrowband processes 249
Appendix A: On matrices and linear spaces 255
Appendix B: Fourier series, Fourier and
Laplace transforms 289
References and further reading 311
Index 317
Preface
In writing this book, the author’s main intention was to write a concise
exposition of the fundamental concepts and ideas that, directly or indi-
rectly, underlie and pervade most of the many specialised disciplines where
linear engineering vibrations play a part.
The style of presentation and approach to the subject matter places
emphasis on the inextricable – and at times subtle – interrelations and inter-
play between physics and mathematics, on the one hand, and between the-
ory and applications, on the other hand. In this light, the reader is somehow
guided on a tour of the main aspects of the subject matter, the starting point
being (in Chapter 2, Chapter 1 is for the most part an introductory chapter
on some basics) the formulation of the equations of motion by means of
analytical methods such as Lagrange’s equations and Hamilton’s principle.
Having formulated the equations of motion, the next step consists in deter-
mining their solution, either in the free vibration or in the forced vibration
conditions. This is done by considering both the time- and frequency-
domain solutions – and their strict relation – for different types of systems
in order of increasing complexity, from discrete finite degrees-of-freedom
systems (Chapters 3 and 4) to continuous systems with an infinite number
of degrees-of-freedom (Chapter 5).
Having obtained the response of these systems to deterministic excita-
tions, a further step is taken in Chapter 6 by considering their response to
random excitations – a subject in which, necessarily, notions of probability
theory and statistics play an important role.
This book is aimed at intermediate-advanced students of engineering,
physics and mathematics and to professionals working in – or simply inter-
ested in – the field of mechanical and structural vibrations. On his/her part,
the reader is assumed to have had some previous exposure to the subject
and to have some familiarity with matrix analysis, differential equations,
Fourier and Laplace transforms, and with basic notions of probability and
statistics. For easy reference, however, a number of important points on
xi
xii Preface
some of these mathematical topics are the subject of two detailed appen-
dixes or, in the case of short digressions that do not interrupt the main flow
of ideas, directly included in the main text.
Milan (Italy) – May 2020
Paolo Luciano Gatti
Acknowledgements
The author wishes to thank the staff at Taylor & Francis, and in particular
Mr. Tony Moore, for their help, competence and highly professional work.
A special thanks goes to my wife and daughter for their patience and
understanding, but most of all for their support and encouragement
in the course of a writing process that, at times, must have seemed like
never-ending.
Last but not the least, a professional thanks goes to many engineers with
whom I had the privilege to collaborate during my years of consulting work.
I surely learned a lot from them.
xiii
Frequently used acronyms
BC boundary condition
BVP boundary value problem
C clamped (type of boundary condition)
F free (type of boundary condition)
FRF frequency response function
GEP generalised eigenvalue problem
HE Hamilton equation
IRF impulse response function
LE Lagrange equation
MDOF multiple degrees of freedom
MIMO multiple inputs–multiple outputs
n-DOF n-degrees of freedom
pdf probability density function
PDF probability distribution function
PSD power spectral density
QEP quadratic eigenvalue problem
r.v. random variable
SEP standard eigenvalue problem
SDOF (also 1-DOF) single degree of freedom
SISO single input–single output
SL Sturm–Liouville
SS simply supported (type of boundary condition)
TF transfer function
WS weakly stationary (of random process)
xv
Chapter 1
A few preliminary fundamentals
1.1 INTRODUCTION
In the framework of classical physics – that is, the physics before the two
‘revolutions’ of relativity and quantum mechanics in the first 20–30 years
of the twentieth century – a major role is played by Newton’s laws. In par-
ticular, the fact that force and motion are strictly related is expressed by
Newton’s second law F = dp dt , where p = mv and we get the familiar
F = ma if the mass is constant. This equation is definitely a pillar of (clas-
sical) dynamics, and one of the branches of dynamics consists in the study,
analysis and prediction of vibratory motion, where by this term one typi-
cally refers to the oscillation of a physical system about a stable equilibrium
position as a consequence of some initial disturbance that sets it in motion
or some external excitation that makes it vibrate.
1.2 MODELLING VIBRATIONS AND

VIBRATING SYSTEMS
In order to make sense of the multifarious complexities of real-life physical

systems and achieve useful results, one must resort to models, that is, ide-
alisations of the actual system/phenomenon under study based on some set
of (simplifying) initial assumptions. Models can be mathematical or non-
mathematical but, by their very nature, have limits of validity and entail
some kind of division into classes or categories that, although often conve-
nient, are in almost all cases neither absolute nor sharply defined. Needless
to say, the field of vibrations is no exception.
First, according to their response behaviour to excitations, systems can be
classified as linear or nonlinear, where, formally, linear systems obey linear
differential equations. The fundamental fact is that for a linear system, the
principle of superposition applies, this meaning that (a) its response/output is
proportional to the excitation/input and (b) its response to the simultaneous
1
2 Advanced Mechanical Vibrations
application of the excitations f1 , f2 is x1 + x2, where x1 , x2 are the system’s

responses to the individual application of f1 and f2. Linearity, however, is not
an intrinsic property of a system but depends on the operating conditions,
and it generally applies only for small amplitudes of vibration. In this book,
our attention will be focused on linear systems, whereas for non-linear vibra-
tions – where things are definitely more complicated, and so far, there is no
comprehensive theory – we refer the interested reader to the specific litera-
tures (e.g. Schmidt and Tondl 1986, Thomsen 2003 or Moon 2004).
Second, according to their physical characteristics – typically mass, e lasticity
and energy dissipation mechanisms, the so-called system’s parameters –
vibrating systems can be continuous or discrete, where discrete-parameters
systems are characterised by a finite number of degrees of freedom (DOFs),
while an infinite number of them is needed for continuous ones. In this regard,
it should be noted that the distinction is not limited to a mere enumeration of
the DOFs, but it also involves a different mathematical formalism – a set of
simultaneous ordinary differential equations for discrete systems and one or
more partial differential equations for continuous ones – with the consequence
that a rigorous treatment of continuous systems is generally much more diffi-
cult, if not even impracticable in many cases. The result of this state of affairs,
therefore, is that continuous systems are in most practical cases modelled by
means of discrete finite-DOFs approximations, which have the advantage of
being very well-suited for computer implementation. A key point in this respect
is the mathematical fact that continuous systems can be seen as limiting cases
of discrete ones as the number of DOFs goes to infinity, thus implying that the
accuracy of the analysis can be improved by increasing the number of DOFs.
Remark 1.1
i. The well-known finite element method – probably the most widely

used technique for engineering design and analysis – is in essence a
highly refined discretisation procedure.
ii. In terms of understanding (as opposed to merely obtaining numerical
results), however, it may be worth observing that the study and analy-
sis of continuous systems provide physical insights that are not at all
evident in discrete cases.
Third, in studying the response of a system to an external excitation,

sometimes it is the type of excitation rather than the system itself that
dictates the strategy of analysis. From this point of view, in fact, the
distinction is between deterministic and random vibrations, where,
broadly speaking, deterministic vibrations are those that can be
described and predicted by means of an explicit mathematical relation-
ship, while there is no way to predict an exact value at a future instant
A few preliminary fundamentals 3
of time for random ones. Then, in light of the fact that with random
data each observation/record is in some sense ‘unique’, their description
can only be given in statistical terms.
1.3 SOME BASIC CONCEPTS
Since the motion of a point particle necessarily occurs in time, it is natural

to describe it by means of an appropriate function of time, say x(t), whose
physical meaning and units depend on the scope of the analysis and, in
applications, also on the available measuring instrumentation. Having
already pointed out above that our main interest lies in oscillations about
an equilibrium position, the simplest case of this type of motion is called
harmonic and is mathematically represented by a sine or cosine function
of the form
x(t) = X cos(ω t − θ ) (1.1)
where
X is the maximum or peak amplitude (in the appropriate units)

(ω t − θ ) is the phase angle (in radians)
ω is the angular frequency (in rad/s)
θ is the initial phase angle (in radians), which depends on the choice of
the time origin and can be assumed to be zero if there is no relative
reference to other sinusoidal functions.
The time interval between two identical conditions of motion is the period T
and is the inverse of the (ordinary) frequency ν = ω 2π (expressed in Hertz;
symbol Hz, with dimensions of s −1), which, in turn, represents the number
of cycles per unit time. The basic relations between these quantities are
ω = 2πν , T = 1 ν = 2π / ω (1.2)
Harmonic motion can be conveniently represented by a vector x of mag-

nitude X that rotates with an angular velocity ω in the xy-plane. In this
representation, x(t) is the instantaneous projection of x on the x-axis (or on
the y-axis if we prefer to use a sine function). This idea can be further gen-
eralised by using complex numbers. If, in fact, we recall the Euler relations
ei z + e− i z ei z − e− i z
e ± i z = cos z ± i sin z, cos z = , sin z = (1.3)
2 2i
where i is the imaginary unit ( i ≡ −1 ) and e = 2.71828 is the basis of

Napierian (or natural) logarithms, then our harmonically varying quantity
can be written as
x(t) = C e i ω t = Xe i (ωt −θ ) (1.4)
where C = Xe − iθ is the complex amplitude, which – as well-known

from basic mathematics – can also be expressed in terms of its real
and imaginary parts (here we call them a, b, respectively) as C = a + ib.
Then, taking the square root of the product of C with its complex con-
jugate C ∗ = a − ib = Xe iθ gives the magnitude C of C, and we have
C = CC ∗ = X = a2 + b2 .
The idea of Equation 1.4 – sometimes called the phasor representation of
x(t) – is the temporary replacement of a real physical quantity by a complex
quantity for purposes of calculation, with the understanding that only the
real part of the phasor has physical significance. With these considerations
in mind, x(t) can be expressed in any one of the four ways
x(t) = a cos(ω t) + b sin(ω t) = X cos(ω t − θ ) = Ce − iω t = Xe − i (ω t −θ ) (1.5)
where only the real part of the complex expressions is assigned a physical
meaning.
Phasors are often very convenient, but some care must be exercised when
considering the energy associated with the oscillatory motion because the
various forms of energy (energy, energy density, power, etc.) depend on the
( )
square of vibration amplitudes. And since Re x2 ≠ ( Re(x)) , we need to take
2
the real part first and then square to find the energy. Complex quantities,
moreover, are also very convenient in calculations. For example, suppose
that we have two physical quantities of the same frequency but different
phases expressed as x1(t) = X1 cos (ω t − θ1 ) and x2 (t) = X2 cos (ω t − θ 2 ). Then,
the average value x1 x2 of the product x1 x2 over one cycle is
T
1 1
x1 x2 ≡
T ∫ x (t)x (t) dt = 2 X X cos (θ − θ ) (1.6)
0
1 2 1 2 1 2
where the calculation of the integral, although not difficult, is tedious. With
complex notation the calculation is simpler, we express the two harmonic
quantities as x1(t) = X1 e i (ω t −θ1 ) and x2 (t) = X2 e i (ω t −θ2 ) and obtain the result by
simply determining the quantity Re x1*x2 2. ( )
Remark 1.2
With complex notation, some authors use j instead of i and write e j ω t, while
some other authors use the negative exponential notation and write e − i ω t or
e − j ω t. However, since we mean to take the real part of the result, the choice
is but a convention and any expression is fine as long as we are consistent. In
any case, it should be observed that in the complex plane, the positive expo-
nential represents a counter-clockwise-rotating phasor, while the negative
exponential represents a clockwise-rotating phasor.
1.3.1 The phenomenon of beats

Consider two sinusoidal functions of slightly different frequencies ω 1 and
ω 2 = ω 1 + ε , where ε is small compared to ω 1 and ω 2. Assuming for simplic-
ity, equal magnitude and zero initial phase for both oscillations x1 , x2 in
complex notation, we get
x1(t) + x2 (t) = Xe i ω1 t + Xe i ω 2 t = Xe i (ω1 +ω 2 ) t 2 e i (ω 2 −ω1 ) t 2 + e − i (ω 2 −ω1 ) t 2 
whose real part – defining ω avg = (ω 1 + ω 2 ) 2 and ω = (ω 2 − ω 1 ) 2 – is
2X cos(ω t) cos (ω avg t ) (1.7)
which can be seen as an oscillation of frequency ω avg and a time-dependent

amplitude 2X cos (ω t ). A graph of this quantity is shown in Figure 1.1 with
ω 1 = 8 rad/s and ω 2 − ω 1 = 0.6
Physically, what happens is that the two original waves remain nearly in
phase for a certain time and reinforce each other; after a while, however,
2,4
Amplitude (arbitrary units)
1,6
0,8
0,0
-0,8
-1,6
-2,4
0 5 10 15 20 25 30
seconds
eats (ω 2 − ω1 = 0.6 ).
Figure 1.1 B
the crests of the first wave correspond to the troughs of the other and they
practically cancel out. This pattern repeats on and on, and the result is
the so-called phenomenon of beats shown in the figure. The maximum
amplitude occurs when ω t = nπ ( n = 0,1, 2, ), that is, every π ω seconds,
and consequently, the frequency of the beats is ω π = ν 2 − ν1, equal to the
(ordinary) frequency difference between the two original signals. For sig-
nals with unequal amplitudes (say, A and B), the total amplitude does not
become zero and varies between A + B and A − B , but in general, the typi-
cal pattern can still be easily identified.
1.3.2 Displacement, velocity,
acceleration and decibels
If the oscillating quantity x(t) of Equation 1.4 is a displacement, we can
recall the familiar definitions of velocity and acceleration (frequently
denoted with overdots as x (t) and x
(t), respectively)
dx(t) d 2 x(t)
v(t) ≡ x (t) = , a(t) ≡ x
(t) =
dt dt 2
and calculate the derivatives to get
v(t) = iω Ce iω t = ω Ce i (ω t + π 2) = V e i (ω t + π 2)
(1.8)
a(t) = −ω 2 Ce iω t = ω 2Ce i (ω t + π) = A e i (ω t + π)
where the second equality in Equation 1.81 follows from Euler’s relation
(Equation 1.3) by observing that e i π 2 = cos ( π 2) + i sin ( π 2) = i . Similarly,
for Equation 1.82 , the same argument shows that e i π = −1.
Physically, Equations 1.8 tell us that velocity leads displacement by 90°
and that acceleration leads displacement by 180° (hence, acceleration leads
velocity by 90°). In regard to amplitudes, moreover, they show that the
maximum velocity amplitude V and maximum acceleration amplitude A are
V = ω C and A = ω 2C = ω V . Clearly, these conclusions of physical nature
must not – and in fact do not – depend on whether we choose to represent
the quantities involved by means of a negative or positive exponential term.
In principle, therefore, it should not matter which one of these quantities –
displacement, velocity or acceleration – is considered, because all three pro-
vide the necessary information on amplitude and frequency content of the
vibration signal. In practice, however, it is generally not so, and some physi-
cal considerations on the nature of the vibrations to be measured and/or on
the available measuring instrumentation often make one parameter prefer-
able with respect to the others.
The amplitudes relations above, in fact, show that the displacement tends to
give more weight to low frequency components while, conversely, acceleration
tends to give more weight to high-frequency components. This, in turn, means

that the frequency range of the expected signals is a first aspect to take into
account – while at the same time indirectly suggesting that for wide-band sig-
nals, velocity may be the more appropriate quantity because it tends to give
equal weight to low- and high-frequency components. By contrast, it is also
a fact that accelerometers are often preferred in applications because of their
versatility – where by this term, we mean a number of desirable properties such
as small physical dimensions, wide frequency and dynamic ranges, easy com-
mercial availability, and so on.
In any case, the primary factors to be considered are, not surprisingly,
the nature of the problem and the final scope of the investigation. Let us
make a few simple heuristic considerations from a practical point of view.
Suppose, for example, that we expect a vibration whose main frequency
component is ν ≅ 1 Hz with an expected amplitude C = ±1 mm. Then, if
possible, displacement would be the quantity to be preferred in this case
because a relatively inexpensive displacement transducer with, say, a total
range of 10 mm and a sensitivity of 0.5 V/mm would produce a peak-to-
peak output signal of 1 V, which means a very good signal-to-noise ratio
in most practical situations. For the same problem, on the other hand, the
peak-to-peak acceleration would be ( 2πν ) C = 7.9 × 10−2 m s2 ≅ 8 × 10−3 g,
2
thus implying that a typical general-purpose accelerometer with a sen-

sitivity of, say, 100 mV/g would produce an output of 0.81 mV. This is
definitely much less satisfactory in terms of the signal-to-noise ratio.
By contrast, if the expected vibration occurs mainly at a frequency of,
say, 100 Hz with an amplitude of ±0.05 mm, the most convenient solution
in this case would be an acceleration measurement. In fact, since the peak-
to-peak acceleration amplitude is now 39.5 m s2 ≅ 4g, a general-purpose
accelerometer with a sensitivity of 100 mV/g would produce a satisfactory
peak-to-peak signal of about 400 mV. In order to measure such small dis-
placements at those values of frequency, we would probably have to resort
to more expensive optical sensors.
Another aspect worthy of mention is the use of the decibel scale, due to
the fact that in many applications, vibration amplitudes may vary over wide
ranges. And since the graphical presentation of signals with wide dynamic
ranges can be impractical on a linear scale, the logarithmic decibel (dB)
scale (which is a standard in the field of acoustics) is sometimes used. By
definition, the dB level L of a quantity of amplitude y is given with respect
to a specified reference value y0, and we have
L (dB) = 20 log10 ( y /y0 ) (1.9)
where, in vibrations, y can be displacement, velocity or acceleration.

Decibels, like radians, are dimensionless; they are not ‘units’ in the usual
sense and consistency dictates that the reference value y0 must be – as it is in
acoustics – universally accepted. It is not always so in the field of vibrations,

and the consequence is that the reference value should always be specified.
In this respect, some typical reference values are d0 = 10−11 m, v0 = 10−9 m s
and a0 = 10−6 m s2 for displacement, velocity and acceleration, respectively.
Note however that, as mentioned earlier, they are not universally accepted.
From the definition itself, it is evident that decibel levels are not added and
subtracted in the usual way. In order to perform these operations, in fact,
we must first go back to the original quantities by taking anti-logarithms –
and, in this regard, we note that Equation 1.9 gives y y0 = 10L 20 – add or
subtract the original quantities and then reconvert the result into a decibel
level by taking the logarithm.
1.4 SPRINGS, DAMPERS AND MASSES
Most physical system possessing elasticity and mass can vibrate. The sim-
plest models of such systems are set up by considering three types of basic
(discrete) elements: springs, viscous dampers and masses, which relate
applied forces to displacement, velocity and acceleration, respectively. Let
us consider them briefly in this order.
The restoring force that acts when a system is slightly displaced from
equilibrium is due to internal elastic forces that tend to bring the system
back to the original position. Although these forces are the manifestation
of short-range microscopic forces at the atomic/molecular level, the simplest
way to macroscopically model this behaviour is by means of a linear mass-
less spring (Figure 1.2). The assumption of zero mass assures that a force
F acting on one end is balanced by a force − F on the other end, so that the
spring undergoes an elongation equal to the difference between the dis-
placements x2 and x1 of its endpoints. For small elongations, it is generally
correct to assume a linear relation of the form
F = k ( x2 − x1 ) (1.10)
where k is a constant (the spring stiffness, with units N/m) that represents
the force required to produce a unit displacement in the specified direction.
If, as it sometimes happens, one end of the spring is fixed, the displace-
ment of the other end is simply labelled x and Equation 1.10 becomes
Figure 1.2 Ideal massless spring.

Figure 1.3 Ideal massless dashpot.
F = −kx , where the minus sign indicates that the force is a restoring force
opposing displacement. The reciprocal of stiffness 1 k is also used, and it is
called flexibility or compliance.
In real-world systems, energy – where, typically, the energies of inter-
est in vibrations are the kinetic energy of motion and the potential strain
energy due to elasticity – is always dissipated (ultimately into heat) by some
means. So, although this ‘damping effect’ can, at least on a first approach,
often be neglected without sacrificing much in terms of physical insight into
the problem at hand, the simplest model of damping mechanism is provided
by the massless viscous damper. This is a device that relates force to veloc-
ity, of which a practical example can be a piston fitting loosely in a cylinder
filled with oil so that the oil can flow around the piston as it moves inside
the cylinder. The graphical symbol usually adopted is the dashpot shown in
Figure 1.3 for which we have a linear relation of the form
F = c ( x 2 − x1 ) (1.11)
where c is the coefficient of viscous damping, with units Ns/m.

If one end is fixed and the velocity of the other end is labelled x , then
Equation 1.11 becomes F = −cx , with the minus sign indicating that the
damping force resists an increase in velocity.
Finally, the quantity that relates forces to accelerations is the mass and
the fundamental relation now is Newton’s second law, which (with respect
to an inertial frame of reference) can be written as
F = mx
 (1.12)
As well-known from basic physics, in the SI system of units, the mass is

measured in kilograms and represents the inertia properties of physical
bodies that, under the action of an applied force F, are set in motion with
an acceleration inversely proportional to their mass.
Now, going back to springs, in practical cases, it is often convenient to
introduce the notion of equivalent spring keq , meaning by this term the
replacement of one or more combination of stiff elements with a single
spring of stiffness keq that, for the problem at hand, represents the stiffness
of such combination. For example, two springs can be connected in series
or in parallel, as shown in Figures 1.4a and b, and in the two cases, respec-
tively, the equivalent stiffness is
(a)
(b)
Figure 1.4 (a) Springs connected in series. (b) Springs connected in parallel.
(a)
(b)
(c)
Figure 1.5 A
few examples of local stiffness for continuous elastic elements.
1 1 1 k1 + k2
= + = , keq = k1 + k2(1.13)
keq k1 k2 k1k2
In the first case, in fact, we have F = k1x1 , F = k2 x2 because the force is

the same in both springs, consequently x1 = F k1 , x2 = F k2 . But since the
total elongation is x = x1 + x2 and the equivalent spring must be such that

F = keq x ⇒ x = F keq , then Equation 1.131 readily follows. For the con-
nection in parallel, on the other hand, both springs undergo the same
displacement x, while the forces F1 = k1 x, F2 = k2 x satisfy the condition
F = F1 + F2 = ( k1 + k2 ) x. Comparing this with the equivalent spring relation
F = keq x leads to Equation 1.132 . Clearly, the two relations 1.13 can be
∑ (1 ki ) and keq = ∑ i=1 ki ,

n n
extended to the case of n springs as 1 keq =
i =1
respectively, and it is not difficult to show that, by appropriately replacing
the k’s with the c’s, similar relations hold for viscous dampers connected in
series and in parallel.
Finally, having pointed out that the stiffness of an elastic element can
be obtained as the applied force divided by the displacement of its point
of application, another example is the cantilever with fixed-free boundary
conditions and a force F applied at the free end (Figure 1.5a). Recalling from
strength of materials analysis that in this case, the vertical d
isplacement x
of the cantilever free end is x = FL3 3EI (where E is Young’s modulus of the
material in N m 2 and I is the cross-sectional moment of inertia in m 4), then
the stiffness at free end is k = F x = 3EI L3
By similar considerations, two more examples are shown in the following
figure:
• Figure 1.5b: fixed-fixed bar of length L with transverse localised force

F at L 2; k is the local stiffness at the point of application of the force;
• Figure 1.5c: bar simply supported at both ends with force F at L 2; k
is the local stiffness at the point of application of the force.
Chapter 2
Formulating the
equations of motion
2.1 INTRODUCTION
In order to describe the motion of a physical object or system, we must spec-

ify its position in space and time with respect to some observer or frame of
reference. In this regard, physics teaches us that not all observers are on an
equal footing because for a special class of them, called inertial observers,
the laws of motion have a particularly simple form. More specifically, if
one introduces the convenient concept of material particle – that is, a body
whose physical dimension can be neglected in the description of its motion
and whose position in space at time t is given by the vector r(t), then for all
inertial observers, the particle’s equation of motion is given by Newton’s
second law F = m a, where F is the vector sum of all the forces applied to
the particle, a(t) = d 2r dt 2 (often also denoted by r(t)) is the particle accel-
eration and m, which here we assume to be a constant, is its (inertial) mass.
This equation, together with the first law: ‘a body at rest or in uniform
rectilinear motion remains in that state unless acted upon by a force’ and
the third law: ‘for every action there is an equal and opposite reaction’ –
both of which, like the second law, hold for inertial observers – is the core
of Newtonian mechanics, which, as we mentioned at the beginning of the
book, is one of the pillars of classical physics.
Also, note that we spoke of all inertial observers because the class of
inertial observers is potentially unlimited in number. In fact, any observer
at rest or in uniform rectilinear motion with respect to an inertial observer
is an inertial observer himself.
Remark 2.1
i. In accordance with the knowledge of his time, Newton regarded the

concepts of space and time intervals as absolute, which is to say that
they are the same in all frames of reference. At the beginning of the
20th century, Einstein showed that it is not so and that Newton’s
13
assumption is only an approximation. It is, however, an excellent

approximation for all phenomena in which the velocities involved are
much smaller than the speed of light c = 2.998 × 108 m/s.
ii. In the most general case, the force F is a function of position, veloc-
ity and time, i.e. F = F(r , r , t). In mathematical terms, Newton’s law
m r(t) = F(r , r , t) is a (Cauchy) initial value problem and admits a
unique solution whenever the initial position r(t = 0) = r0 and initial
velocity r(t = 0) = v0 are given.
2.2 SYSTEMS OF MATERIAL PARTICLES
In most problems, it is convenient to ideally separate a ‘system’ of N mutu-

ally interacting particles from its surrounding environment and classify as
external any interaction between the system and the environment. By so
doing, we can distinguish between external and internal forces and write
the force Fk acting on the kth particle ( k = 1, , N ) as Fk(ext) + Fk(int), with, in
addition, Fk(int) = ∑ j ( j ≠k) Fkj, where Fkj is the (internal) force exerted upon
the kth particle by the jth ( j ≠ k) particle of the system. Then, it is shown
(Goldstein 1980) that by invoking Newton’s third law in its weak and
strong form (see Remark 2.2 below), we are led to two equations in which
the internal forces do not appear; these are
N N N N
∑ p = ∑ F
k =1
k
k =1
(ext)
k , ∑ ( r × p ) = ∑ ( r × F ) (2.1)
k =1
k k
k =1
k
(ext)
k
where × is the well-known symbol of vector (or cross) product. By introduc-

ing the total linear and angular momentum P, L of the system together with
the total external force F(ext) and its moment (or torque) N (ext), Equations
2.1 can be rewritten as P = F(ext) and L = N (ext) , respectively. In particular, if
the system is not acted upon by any external force then P = 0 ⇒ P = const
and L = 0 ⇒ L = const, that is, we obtain, respectively, the conservation
theorem of total linear momentum and the conservation theorem of total
angular momentum.
Remark 2.2
The weak form of the third law states that the mutual forces of the two par-
ticles are equal and opposite. In addition to this, the strong form – which is
necessary to arrive at Equation 2.12 – states that the internal forces between
the two particles lie along the line joining them.
Formulating the equations of motion 15
2.2.1 Generalised co-ordinates, constraints

and degrees of freedom
In the considerations above, the particles of the system are identified by
their Cartesian co-ordinates r1 = ( x1 , y1 , z1 ) , , rN = ( xN , yN , zN ). This is by
no means necessary and we can use some other 3N – and often, as we will
see shortly, even less than that – independent quantities q1 , , q3N for which
there exists a continuous one-to-one correspondence with the original co-
ordinates x1 , , zN . These quantities are called generalised co-ordinates
and are not necessarily ‘co-ordinates’ in the traditional sense; they can be
angles, linear or angular momenta, or whatever may turn out to be conve-
nient in order to (possibly) simplify the problem at hand.
The use of generalised co-ordinates is particularly convenient if between
the co-ordinates of the particles there exist some, say m ( m ≤ 3N ), indepen-
dent relations, called constraints equations, of the form
fi ( r1 , , rN , t ) = 0 ( i = 1, , m) (2.2)
which mathematically represent kinematical conditions that limit the parti-
cles motion. Examples are not hard to find: any two particles of a rigid body
must satisfy the condition ( rk − rj ) − dkj2 = 0 for all t because their mutual
2
distance dkj is fixed; a particle constrained to move on a circle of radius R

in the xy-plane must satisfy x2 + y 2 − R2 = 0 , and so on. For systems with
constraints of the form (2.2), the number of independent co-ordinates that
unambiguously specify the system’s configuration is n = 3N − m, because
these constraint equations can be used to eliminate m co-ordinates. Then,
passing to a set of generalised co-ordinates, we will have n of them, related
to the old co-ordinates by a transformation of the form
r1 = r1 ( q1 , , qn , t ) , , rN = rN ( q1 , , qn , t ) (2.3)
which implicitly contains the information on the constraints. As above, we

assume the transformation (2.3) to be invertible.
Constraints of the type 2.2 are classified as holonomic, and in particular,
they are further subdivided into rheonomic if time t appears explicitly (as
in Equation 2.2) or scleronomic if t does not appear explicitly and we have
fi ( r1 , , rN ) = 0. Needless to say, in all problems involving motion, the vari-
able t always appears implicitly because each ri or qi is a function of time.
Speaking of holonomic constraints implies that there exist non-holonomic
ones. Typically, these constraints have the form of non-integrable rela-
tions between the differentials dri or dqi of the co-ordinates (most books
of mechanics show the classical example of a disk that rolls without slip-
ping on a horizontal plane). Non-holonomic constraints – unlike holonomic
ones – cannot be used to eliminate some of the variables. Moreover, since,
as a general rule, non-holonomic systems must be tackled individually; in

the following developments (unless otherwise stated), we will confine our
attention to holonomic constraints.
Remark 2.3
i. For a system of N particles with m constraints, one defines the num-

ber of degrees of freedom (DOFs for short) as n = 3N − m. Then, as
mentioned earlier, a holonomic system is such that n is the exact num-
ber of generalised co-ordinates necessary to completely describe it;
less than n are not enough while more that n could not be assigned
without satisfying certain conditions. Not so for non-holonomic sys-
tems; here we must operate with more than n co-ordinates and retain
the constraint equations as auxiliary conditions of the problem.
ii. The rheonomic-scleronomic distinction applies also to non-holonomic
constraints, which – just like holonomic ones – may or may not con-
tain the variable t explicitly. In this respect, moreover, it should be
noted that t may appear explicitly in the transformation (2.3) because
(a) the constraints are time-dependent, or (b) our frame of reference is
in relative motion with respect to the system under study.
iii. The constraints discussed so far are bilateral; other types of con-
straints, called unilateral, may involve inequalities and be, for exam-
ple, of the form f ( r1 , , rN , t ) ≤ 0. Although it can be correctly argued
(Greenwood 1997) that they are holonomic in nature, we follow
Goldstein 1980 and classify them as non-holonomic.
A fundamental difficulty due to the presence of constraints is that they

imply the existence of constraint forces, which, a priori, are completely
undetermined in both magnitude and direction. In fact, the information
provided by Equations 2.2 is on the effect the constraint forces on the sys-
tem’s motion, and not on the forces themselves (if we knew the forces, we
would not need the constraint equations). The difficulty lies in the fact that
in order to solve the problem of the system’s motion, we must consider these
constraint forces together with the applied forces. So, unless we are spe-
cifically interested in the determination of the constraint forces, it is highly
desirable to obtain a set of equations of motion in which the constraint
forces do not appear.
2.3 VIRTUAL WORK AND D’ALEMBERT’S PRINCIPLES –

LAGRANGE AND HAMILTON EQUATIONS
Let rk be the position vector of the kth particle of a system; we call virtual
displacement δ rk an imaginary infinitesimal displacement consistent with
the forces and constraints imposed on the system at the time t; this meaning
that we assume any time-dependent force or moving constraint to be ‘fro-
zen’ at time t (this justifies the term ‘virtual’ because in general δ rk does not
coincide with a real displacement drk occurring in the time dt). If the system
is in static equilibrium and Fk is the total force acting on the particle, equi-
librium implies Fk = 0, and consequently, ∑k Fk ⋅ δ rk = 0 – where we recog-
nise the l.h.s. as the system’s total virtual work. Since, however, each force
can be written as the sum of the applied force Fk(a) and the constraint force
( )
fk, we have ∑k Fk(a) + fk ⋅ δ rk = 0. If now, at this point, we limit ourselves to
workless constraints, that is, constraints such that ∑k fk ⋅ δ rk = 0, we arrive
at the principle of virtual work
N
δ W (a) ≡ ∑F
k =1
( a)
k ⋅ δ rk = 0 (2.4)
stating that the condition for static equilibrium of a system with workless
constraints is that the total virtual work of the applied forces be zero. Note,
however, that Equation 2.4 does not imply Fk(a) = 0 because the δ rk , owing
to the presence of constraints, are not all independent.
Remark 2.4
The assumption of workless constraints may seem overly restrictive, but it is

not so. Many common constraints – holonomic and non-holonomic, rheonomic
and scleronomic – are, in fact, workless. Moreover, if a c onstraint is not
frictionless, Equation 2.4 is still valid if we count the tangential components
of friction forces as applied forces. This aspect, however, is of minor impor-
tance; since friction hampers motion and since the principle implies equilib-
rium with frictionless constraints, it is even more so if friction is present.
The principle of virtual work applies to static equilibrium. But if we

note – as d’Alembert ingeniously did – that the equation of motion of the
kth particle rewritten as Fk − mkrk = 0 (we assume the masses to be constant)
expresses a condition of dynamic equilibrium, then the principle of virtual
work leads to d’Alembert’s principle, that is
N N N
∑ (F
k =1
( a)
k )
− mkrk ⋅ δ rk = ∑F
k =1
( a)
k ⋅ δ rk − ∑ m r ⋅ δ r
k =1
k k k = 0 (2.5)
which, when rewritten in terms of generalised co-ordinates, will give us the

possibility to exploit the independence of the δ qi . The procedure is a stan-
dard derivation of analytical mechanics and here we only mention it briefly.
In essence, the four basic relations used to accomplish this task are
n n
∂ rk ∂r d  ∂ rk  ∂ rk
rk = ∑ i =1
∂ qi
q i + k , δ rk =
∂t ∑ ∂∂ qr δ q ,
i =1
k
i
i
∂ rk ∂ rk
=
∂ qi ∂ q i
, =
dt  ∂ qi  ∂ qi
(2.6)
and it can be shown (Goldstein 1980, Greenwood 1977 or Gatti 2014)

that, in the end, the first term on the l.h.s. (left-hand side) of Equation 2.5
becomes
N n
 N
∂ rk 
n
∑k=1
Fk(a) ⋅ δ rk = ∑ ∑
i =1

 k=1
Fk(a) ⋅  δ qi =
∂ qi  ∑ Q δ q (2.7)
i =1
i i
where Qi ( i = 1, , n ) is called the ith generalised force and is defined by the
term in parenthesis, while, with some lengthier manipulations, the second
term on the l.h.s. of 2.5 becomes
N n
   
∑k=1
mk rk ⋅ δ rk = ∑  dtd  ∂∂qT  − ∂∂qT  δ q (2.8)
i =1
i i
i
where T is the system’s kinetic energy (well-known from basic physics)

N N
T=
1
2 ∑
k =1
mkvk2 =
1
2 ∑ m r ⋅ r (2.9)
k =1
k k k
Together, Equations 2.7 and 2.8 give d’Alembert’s principle in the form
n
   
∑  dtd  ∂∂qT  − ∂∂qT − Q  δ q = 0 (2.10)
i =1
i i
i i
where now the virtual displacements δ qi – unlike their Cartesian counter-

parts – are independent. Since this means that Equation 2.10 holds only if
the individual coefficients within brackets are zero, we obtain the system of
n second-order differential equations
d  ∂T  ∂T
− = Qi ( i = 1,… , n ) (2.11)
dt  ∂ q i  ∂ qi
called Lagrange’s equations (LEs for short).

If, moreover, all the applied forces Fk(a) are conservative (see Remark
2.5(iii) below), then there exists a scalar function V ( r1 , , rN , t ) such that
Fk(a) = −∇k V , the generalised force Qi becomes
Qi = − ∑ ∇ V ⋅ ∂∂ qr
k=1
k
k
i
=−
∂V
∂ qi
(2.12)
and we can move the term − ∂ V ∂ qi to the l.h.s. of Equation 2.11. Then,
observing that ∂ V ∂ q i = 0 because in terms of generalised co-ordinates V
is a function of the form V = V ( q1 , , qn , t ), we can define the Lagrangian
function (or simply Lagrangian) L of the system as
L = T − V (2.13)
and write LEs 2.11 in what we can call their standard form for holonomic
systems, that is
d  ∂L  ∂L
− =0 ( i = 1,… , n ) (2.14)
dt  ∂ q i  ∂ qi
Remark 2.5
i. In writing the equation of motion of the kth particle as Fk − mkrk = 0,

the term I k = −mkrk is called inertia force. In this light, d’Alembert’s
principle states that the moving particle is in equilibrium if we add the
inertia force to the impressed forces Fk(a) and fk. Then, if we turn our
attention to the system, we can interpret also LEs 2.11 as an equilib-
rium condition: Qi plus the ith generalised inertia force (the negative
of the l.h.s. term in Equation 2.11) equal zero.
ii. In general, the Qi do not have the dimensions of force, but in any case,
the product Qi δ qi has the dimensions of work.
iii. Summarising, we obtain LEs in the form 2.11 under the assumptions
of workless and holonomic constraints. We obtain the form 2.14 if,
in addition, all the applied forces are conservative. In this respect, we
recall from basic physics that V is called the potential energy of the
system and that the expression ∇k V (the gradient with respect to the
Cartesian co-ordinates xk , yk , zk of the kth particle) is the vector of
components ∂V ∂ xk , ∂V ∂ yk , ∂V ∂zk .
iv. In the general case, the Lagrangian is a function of the form
L = L ( q1 ,… , qn , q1 ,… , q n , t ). For brevity, this functional dependence is
often denoted by L = L ( q, q , t ).
Finally, if some of the applied forces Fk(a) are conservative while some oth-
ers are not, then the generalised forces Qi are written as Qi = Q  i − ∂V ∂qi ,

where the Qi are those generalised forces not derivable from a potential
function. In this case, LEs take the form
d  ∂L  ∂L
− i
=Q ( i = 1,… , n ) (2.15)
dt  ∂ q i  ∂ qi
2.3.1 Hamilton’s equations (HEs)

In some cases, instead of the system of n Lagrange’s second-order differ-
ential equations, it may be more convenient to express the equations of
motion as an equivalent system of 2n first-order differential equations. One
way of doing this is as follows. If, starting from the Lagrangian L ( q, q , t ),
we define the n new variables called generalised momenta as
∂L
pi = (i = 1,..., n) (2.16)
∂ q i
then Lagrange’s equation (2.14) can be written as
∂L
p i = ( i = 1,… , n ) (2.17)
∂ qi
and we can introduce the so-called Hamilton function (or Hamiltonian)
H ( q, p, t ) ≡ ∑ p q − L (2.18)
i
i i
with the understanding that all the q i on the r.h.s. are expressed as func-
tions of the variables q, p, t (see Remark 2.6 below). Then, since the func-
tional form of the Hamiltonian is H = H ( q, p, t ), its differential is
 
dH = ∑  ∂∂ Hq dq + ∂∂ Hp dp  + ∂∂Ht dt (2.19a)
i
i
i
i
i
which can be compared with the differential of the r.h.s. of Equation 2.18,
that is, with
   
d

∑
i
pi q i − L =

∑ q dp + ∑ p dq −∑  ∂∂qL dq + ∂∂qL dq  − ∂∂Lt dt
i
i i
i
i i
i
i
i
i
i

= ∑ q dp − ∑ ∂∂qL dq − ∂∂Lt dt
i
i i
i
i
i
(2.19b)
where the second equality is due to the fact that, owing to Equation 2.16,
the term ∑ pi dq i cancels out with the term ∑ (∂ L / ∂ q i ) dq i . The comparison
of the coefficients of the various differential terms of Equations 2.19a and

b leads to (besides the evident relation ∂ L ∂ t = − ∂ H ∂t)
∂H ∂H
p i = − , q i = (i = 1,… , n) (2.20a)
∂ qi ∂ pi
where in writing the first of Equation 2.20a, we took Lagrange’s equations

(2.17) into account. Equations 2.20a form a system of 2n first-order differ-
ential equations known as Hamilton’s canonical equations.
Remark 2.6
In Equation 2.18, all the q i must be expressed as functions of the variables

q, p; clearly, this includes also the q is that appear in the functional depen-
dence L ( q, q , t ) of the Lagrangian. The functions q i = q i (q, p, t) needed to
accomplish this task are obtained by means of the inverses of Equations
2.16, and we recall that a necessary and sufficient condition for these equa-
( )
tions to be invertible is det ∂ 2 L ∂ q i ∂ q j ≠ 0 – which, in general, is satis-
fied in most cases of interest. Also, we anticipate here that since Equation
2.30 in the following Section 2.4.2 can be written in the matrix form as
p = Mq + b, where the mass matrix M =  Mij  is generally positive definite
and therefore invertible, then the q i are expressed in terms of the momenta
pi by means of the matrix relation q = M −1 ( p − b ).
From the derivation above, it is clear that Hamilton’s canonical equations

2.20a apply to the same systems as the standard form of Lagrange’s equa-
tions 2.14. If, however, there are generalised forces not derivable from a
potential function, LEs take the form (2.15) and it is reasonable to expect
that also HEs may have a different form. This is, in fact, the case because
now the definition of generalised momentum (2.16) modifies Equation 2.17
i + ∂ L ∂ qi . Then, the same procedure as above leads to HEs in
into p i = Q
the form
∂H ∂H
p i = − + Qi , q i = (i = 1,… , n) (2.20b)
∂ qi ∂ pi
Example 2.1
A paradigmatic example of oscillating system is the simple pendulum
of fixed length l (Figure 2.1).
The position of the mass m is identified by the two Cartesian co-
ordinates x, y , but the (scleronomic) constraint x2 + y 2 − l 2 = 0 tells us
that this is a 1-DOF system. Then, since a convenient choice for the
generalised co-ordinate is the angle θ , we have
Figure 2.1 Simple pendulum.

x = l sin θ x = lθ cos θ x = lθ 2 sin θ + lθ cos θ
⇒ ⇒
y = −l cos θ y = lθ sin θ y = lθ 2 cos θ + lθ sin θ
(2.21)
The system’s kinetic energy, potential energy and Lagrangian function

are therefore
( )
T = m x 2 + y 2 2 = ml 2θ 2 2
( )
⇒ L θ , θ ≡ T − V =
ml 2θ 2
+ mgl cos θ
V = mgy = − mgl cos θ 2
(2.22)
from which it follows that the prescribed derivatives of Lagrange equa-

tions in the form 2.14 are
d  ∂L 2  ∂L
  = ml θ , = − mgl sin θ
dt  ∂θ  ∂θ
thus leading to the well-known equation of motion
θ + ( g l ) sin θ = 0 (2.23)
which is clearly non-linear because of the sine term. Not surprisingly,

the same equation of motion can be obtained by using d’Alembert’s
principle in its original form (Equation 2.5) and calculating
(F ( a)
)
− mr ⋅ δ r = ( Fx − mx
) δ x + ( Fy − my) δ y = 0 (2.24)
where in this case Fx = 0, Fy = − mg, the accelerations x , y are given
by Equations 2.213 and the virtual displacements, using Equations
2.211, are δ x = l cos θ δθ , δ y = l sin θ δθ . Also, we can now determine
the generalised force Q; since F ⋅ δ r = Fx δ x + Fy δ y = − mgl sin θ δθ , then
Q = − mgl sin θ , which, being associated with the angular virtual dis-
placement δθ , is a torque.
Finally, since in this example we assumed no energy dissipation, we

can check that the total energy is conserved. This can be done imme-
diately by first determining the total energy E = T + V and then seeing
that, because of the equation of motion 2.23, dE dt = 0.
It is left to the reader to show that the Hamilton function is
p2
H ( p, q ) = − mgl cos θ
2ml 2
where the generalised momentum is p ≡ ∂ L ∂θ = ml 2θ and Hamilton
canonical equations are p = − mgl sin θ and θ = p ml 2 .
Example 2.2: Double pendulum

The double pendulum of Figure 2.2 is a 2-DOF system, and the two
convenient generalised co-ordinates for this case are the angles θ , φ .
As it should be expected, the formulas become now more involved,
but since the method is a rather straightforward extension of the pre-
ceding example, we only outline the main results, leaving the details
to the reader as an exercise. The kinetic and potential energies for this
system are
 m + m2  2  2 m2 l22φ2   cos (θ − φ )
T = 1  l1 θ + + m2 l1l2 θφ
 2 2 (2.25)
V = − ( m1 + m2 ) gl1 cos θ − m2 gl2 cos φ
from which we readily obtain the Lagrangian L = T − V . Then, calcu-

lating the prescribed derivatives and defining α = θ − φ , we get the two
(coupled and non-linear) equations of motion
( m1 + m2 ) l1θ + m2 l2 (φ cos α + φ2 sin α ) + ( m1 + m2 ) g sin θ = 0
(2.26)
l2 φ + l1 θ cos α − l1 θ 2 sin α + g sin φ = 0
Figure 2.2 Double pendulum.

2.4 ON THE PROPERTIES AND STRUCTURE

OF LAGRANGE’S EQUATIONS
2.4.1 Invariance in the form of LEs

and monogenic forces
Unlike Newton’s equations, Lagrange equations hold in an arbitrary frame
of reference. In other words, with some rather lengthy but not difficult cal-
culations, it can be shown that if LEs 2.14 hold for the q-co-ordinates and
u1 , , un is a new set of generalised co-ordinates such that the transforma-
tion qi = qi ( u1 , , un , t ) is invertible, then the system’s equations of motion
are given by
d  ∂ L′  ∂ L′
− =0 ( i = 1,… , n ) (2.27)
dt  ∂ u i  ∂ ui
with L′(u, u , t) = L ( q(u, t), q (u, u , t), t ), where by this equality, we mean that
the ‘new’ Lagrangian L′ is obtained from the ‘old’ by substituting for qi , q i
the functions which express them in terms of the new variables ui , u i .
In light of the fact that the ‘invariance property’ of LEs is the reason
why the form 2.14 is particularly desirable, it could be asked if this form
also applies to cases that are more general than the one in which the forces
are conservative – such as, for instance, forces that may also depend on
time and/or on the velocities of the particles. As it turns out, the answer is
affirmative and LEs have the standard form whenever there exists a scalar
function V = V ( q, q , t ), such that the generalised forces are given by
∂V d  ∂V 
Qi = − + (2.28)
∂ qi dt  ∂ q i 
and the Lagrangian is defined as L = T − V . These forces are sometimes

called monogenic while the function V from which they derive is called gen-
eralised potential (but the name work function for the negative of V is also
common). In this respect, however, it should be noted that the term ‘mono-
genic forces’ refers to the fact that they are derivable from a scalar function,
irrespective of whether they are conservative or not. Clearly, conservative
forces and the familiar potential energy V are special cases of monogenic
forces and of generalised potential.
2.4.2 The structure of the kinetic energy

and of Lagrange equations
The system’s kinetic energy in Cartesian co-ordinates is given by Equation
2.9. Then, substitution of Equations 2.61 into 2.9 leads to an expression of
T which is a sum of three terms. More precisely, we have
 ∂r   ∂r 
N n n
∂ rk ∂ rk
T=
1
2 ∑
k=1
mk 

∑
i =1
∂ qi
q i + k  ⋅ 
∂ t   ∑ j =1
∂ qj
q j + k 
∂ t 
n
 N
∂ rk ∂ rk 
n
 N
∂ rk ∂ rk 
N
 ∂r 
2
=
1
∑∑ 
2 i , j =1  k=1
mk ⋅  q i q j +
∂ qi ∂ q j  ∑∑
i =1

 k=1
mk ⋅
∂ qi ∂ t 
 q i +
1
2 ∑
k=1
mk  k 
 ∂t 
n n
=
1
∑
2 i , j =1
Mij ( q, t ) q i q j + ∑ b (q, t ) q + T
i =1
i i 0 = T2 + T1 + T0
(2.29)
where, by definition, we called T0 the term with ( ∂ rk ∂t ) while the coeffi-

2
cients Mij ( i, j = 1, , n ) and bi ( i = 1, , n ) are defined by the corresponding

terms within parenthesis in the second line of Equation 2.29. Three things
should be noted in this equation:
1. T2 = T2 ( q, q , t ) is a homogeneous quadratic (i.e. of order two) function

of the generalised velocities q i, T1 = T1 ( q, q , t ) is a homogeneous linear
(i.e., of order one) function of the q i and T0 = T0 (q, t) is independent of
the q i. This is the reason for the subscripts 2, 1 and 0.
2. The coefficients Mij are symmetric (i.e. Mij = M j i ),
3. When the co-ordinate transformation (2.3) does not depend explicitly
on time, then T1 = T0 = 0 and T = T2. Systems of this kind are often
referred to as natural, while the term non-natural is used when T1 or
T0 , or both, are non-zero.
Passing now to LEs, we focus our attention on the lth equation and observe
that in order to determine its structure more explicitly, we need to manipu-
late the two terms on the l.h.s., that is, the terms
d  ∂ L  dpi d  ∂ T  d  ∂ T2 ∂ T1 
  = = = +
dt  ∂ q l  dt dt  ∂ q l  dt  ∂ q l ∂ q l 

∂ L ∂ (T − V ) ∂ T2 ∂ T1 ∂ T0 ∂ V
= = + + −
∂ ql ∂ ql ∂ ql ∂ ql ∂ ql ∂ ql
where in writing these two relations, we took into account (a) the defini-
tion 2.16 of pl , (b) the fact that, in most cases, V does not depend on the q i
(and therefore ∂ V ∂ q l = 0), and (c) that, in the general (non-natural) case,
T = T2 + T1 + T0.
The first step consists in using the explicit forms of T2 , T1 to obtain

∂ T2 ∂ T1

∂ q l
= ∑ M q ,
j
lj j
∂ q l
= bl ⇒ pl = ∑ M q + b (2.30)
j
lj j l
from which it follows
∂ pj ∂ pj ∂ pl
dpl
dt
= ∑ ∂qr
r
q r + ∑ ∂ q
r
r
qr +
∂t

 ∂ Mlj ∂ bl     ∂ Mlj ∂ bl 
= ∑ ∂ qr
q j q r + ∑ q r  + 
∂ qr   ∑ Mlj qj  +  ∑ ∂t
q j +
∂t 

 r,j r   j   j 
(2.31a)
which, in turn, can be further rearranged as
 ∂ Mlj ∂ Ml r  ∂ Mlj ∂ bl
dpl 1
dt
=
2 ∑  ∂ q
j ,r
r
+
∂ q j 
q j q r + ∑ M q + ∑
j
lj j
j
∂t
q j + ∑ ∂∂qb q + ∂ t
j
l
j
j
(2.31b)
In the second step, we consider the term ∂ L ∂ ql = ∂ T ∂ ql − ∂ V ∂ ql and

obtain
∂ T2 ∂ T1 ∂ T0 ∂ V  1 ∂ Mij   
+ + − =
∂ ql ∂ ql ∂ ql ∂ ql  2 ∑ ∂q l
q i q j  + 
 
∑ ∂∂ qb q  + ∂∂Tq
i
l
i
0
l
−
∂V
∂ ql
 i, j i
(2.32)
Then, putting Equations 2.31 and 2.32 together and renaming dummy
indexes as appropriate, we get the desired result, that is, the explicit struc-
ture of the lth Lagrange equation. This is
 ∂ Mlj ∂ Ml r ∂ Mrj 
∑ M q + 12 ∑  ∂ q
j
lj j
j ,r
r
+
∂ qj
−
∂ ql 
q j q r
 ∂ Mlj ∂ bl ∂ bj  ∂ bl ∂ V ∂ T0
+ ∑  j
∂t
+ − 
∂ q j ∂ ql 
q j + + −
∂ t ∂ ql ∂ ql
=0 (l = 1,… , n )
(2.33a)
which can be rewritten as
∂ Mlj ∂ bl ∂ (V − T0 )
∑M j
lj qj + ∑[ jr, l ]q q + ∑
j ,r
j r
j
∂t
q j + ∑ g q + + ∂ t +
j
lj j
∂ ql
=0
(2.33b)
when one introduces the Christoffel symbol of the first kind [ jr , l ] and the
skew-symmetric coefficients glj (i.e. such that glj = − g j l ) defined, respec-
tively, as
1  ∂ Mlj ∂ Ml r ∂ Mrj  ∂ bl ∂ bj
[ jr, l ] = 2  + −
∂ ql 
, glj = − (2.34)
 ∂ qr ∂ qj ∂ q j ∂ ql
Clearly, if – as it is often the case – the M and b coefficients do not depend

on t, the third and fifth terms on the l.h.s. of Equation 2.33b are zero. Also,
in respect to the various terms of Equation 2.33b, note that the first three
originate from the T2 -part of the kinetic energy, while the fourth and fifth
terms originate from the T1-part. For these two terms, in fact, the relation
(2.30)2 shows that we can write them as
∂ 2 T1 ∂ 2 T1 ∂ bl ∂ 2 T1
glj = − , = (2.35)
∂ q j ∂ q l ∂ ql ∂ q j ∂ t ∂t ∂ q l
In particular, the terms glj q j in Equations 2.33b are called gyroscopic

and, when shifted to the r.h.s. of the equation, represent the gyroscopic
forces Ql = − ∑ j glj q j ( l = 1,… , n ). In this respect, the skew-symmetry of the
g-coefficients (which implies gll = 0) leads to two important characteris-
tics of the gyroscopic forces: (a) they always involve coupling between the
motion of two or more co-ordinates, and (b) their rate of doing work is zero
because ∑l Ql q l = − ∑l , j glj q l q j = 0. A further point worthy of mention is
that in the last term on the l.h.s., the T0 -part of the kinetic energy is consid-
ered together with V to form the so-called dynamic potential or modified
potential energy, defined as U = V − T0 (not to be confused with the gener-
alised potential of Section 2.4.1).
Equations 2.33 are, in general, non-linear. Since they are, however, linear
in the generalised accelerations q and since the matrix of the coefficients
Mlj is non-singular, they can be solved for the qs in terms of the q, q , t and
expressed in the form
ql + fl ( q, q , t ) = 0 (l = 1,… , n ) (2.36)

2.4.3 The energy function and the

conservation of energy
Let us now consider the case L = L ( q, q ) by also assuming that some of the
applied forces are not derivable from a potential. Then, LEs have the form
2.15 and we can write the total derivative of L as (all sums are from 1 to n)
 
dL
dt
= ∑  ∂∂qL q + ∂∂qL q 
i
i
i
i
i

 d  ∂L  i q i + ∂ L qi  =  d  ∂L  
= ∑i
  
 dt  ∂ q i 
q i − Q
∂ q i  ∑
i
  q i  − Qi q i 
 dt  ∂ q i  
(2.37)
where the second equality follows from LEs 2.15 by isolating ∂ L ∂ qi on

the l.h.s., while the third is due to the familiar rule for the derivative of a
product. Observing that Equation 2.37 can be rewritten as
d  

dt  ∑ ∂∂qL q − L = ∑ Q q (2.38)
i
i
i
i
i i
we are led to the following conclusion: if, by definition, we denote by

h = h( q, q ) and call energy function, or simply energy (of the system), the
function within parenthesis on the l.h.s., its time-derivative equals the total
power of the non-potential forces. In particular, this means that if all the
applied forces originate from a potential, then the r.h.s. of Equation 2.38
is zero and h is a constant, or integral of the motion (which is sometimes
called Jacobi’s integral). We have, in other words, an energy conservation
theorem.
If now, in addition, we recall from Section 2.4.2 that, in general, the
Lagrangian has the form L = T2 + T1 + T0 − V and that T0 , V do not depend
on the velocities q, we can use Euler’s theorem for homogeneous functions
(see Remark 2.7 (i) below) to obtain ∑ i (∂ L ∂ q i ) q i = 2T2 + T1. From the defi-
nition of h, it then follows that
h = T2 − T0 + V = T2 + U (2.39)
where U = V − T0 is the dynamic potential introduced at the end of the pre-

ceding section. In general, therefore, we have h ≠ T + V and the ‘familiar’
energy conservation theorem in the form d (T + V ) dt = 0 is obtained in the
special case when T = T2, that is, when the system is – following a common
terminology – natural and its kinetic energy is a homogeneous quadratic
function of the q s.
Finally, note that the term T1 does not appear on the r.h.s. of Equation 2.39
because – as pointed out at the end of the preceding subsection 2.4.2 – the
gyroscopic forces do no work on the system.
Remark 2.7
Euler’s theorem for homogeneous functions. A function f ( x1 , , xn )

i.
is homogeneous of order m if f (α x1 , , α xn ) = α m f ( x1 , , xn ). In this
case, Euler’s theorem states that mf = ∑ i xi (∂ f / ∂ xi ).
ii. Clearly, the definition of h applies even in the case in which L depends
explicitly on t . In this more general case, we have h = h(q, q , t) and
Equation 2.38 reads

dh(q, q , t)
dt
= ∑ Q q − ∂∂Lt (2.40)
i
i i
or dh dt = − ∂ L ∂ t if all applied forces are derivable from a potential.
iii. Recalling the definitions of the momenta pi and of the function

H (Equations 2.16 and 2.19, respectively), we see that h is just the
Hamiltonian expressed in terms of the variables q, q , t instead of
q, p, t. This suggests that we expect H to be a constant of the motion
whenever it does not depend explicitly on t . This is, indeed, the case;
if H = H (q, p), then
 

dH
dt
= ∑  ∂∂ Hq q + ∂∂ Hp p  = ∑ (− p q + q p ) = 0 (2.41)
i
i
i
i
i
i
i i i i
where the second equality is a consequence of Hamilton’s

Equations 2.20.
2.4.4 Elastic forces, viscous forces and

Rayleigh dissipation function
In LEs, the inertia forces are accounted for by the kinetic energy term while
the potential energy V, or the generalised potential V, accounts for conser-
vative or monogenic forces. Here we consider two types of forces that are
frequently encountered in the field of vibrations: elastic forces and damping
forces of viscous nature. As we will see shortly, elastic forces are conserva-
tive (hence monogenic) while viscous forces are non-monogenic.
If, as usual, we let rk ( k = 1, , N ) be the position vector of the kth particle
of the system and we remain within the linear range, the elastic and viscous
forces acting on the particle can be written as (Chapter 1) Fk(E) = −kk rk and
Fk(V) = −ck rk, where kk , ck are two non-negative constants called the (kth)
stiffness and viscous coefficients, respectively. Let us now consider elastic
forces first. If we introduce the scalar function
N
V (E) ( r1 , , rN ) =
1
2 ∑ k r ⋅ r (2.42)
k =1
k k k
it is almost immediate to see that V (E) is the potential energy because

Fk(E) = −∇k V (E) and also (Equation 2.12) Qi(E) = − ∂ V (E) ∂ qi . This means that
LEs hold in the form 2.14 and that elastic forces are accounted for by a
potential energy term in the Lagrangian.
Example 2.3
The elastic potential energy is also called strain energy. In the sim-
plest example of a spring that is stretched or compressed (within its
linear range), the force–displacement relation is linear and we have
F = −kx, where we assume x = 0 to be the undeformed position.
Dispensing with the minus sign, which is inessential for our present
purposes because x may be a compression or an elongation, the work
done by this force from x = 0 to x equals the strain energy and we have
x
∫
(E)
V (E) = k r dr = kx2 2. Consequently, we can write V = F x 2 , a for-
0
mula known as Clapeyron’s law, which states that the strain energy
is one-half the product Fx. In this light, the example we consider here
is the calculation of the strain energy of a rod of length L and cross-
sectional area A under the action of a longitudinal force F(x, t). Calling
u(x, t) and ε (x, t), respectively, the displacement and strain at point x
and time t, the infinitesimal element dx of the rod undergoes a deforma-
tion (∂ u ∂ x)dx = ε (x, t) dx and the strain energy of the volume element
A dx, by Clapeyron’s law, is dV (E) = ε F dx 2. Then, from the definition
σ (x, t) = F A of axial stress and the assumption to remain within the
elastic range (so that σ = Eε , where E is Young's modulus), we are led
to F = EA ε , and consequently, dV (E) = ε 2 EA dx 2. Integrating over the
rod length, we obtain the rod strain energy
L L 2
1 1  ∂u 
V (E) =
2 ∫
0
EAε 2 dx =
2 ∫
0
EA   dx (2.43)
 ∂x 
The considerations above do not apply to frictional forces of viscous nature.

In this case, however, we can introduce a scalar function D, called Rayleigh
dissipation function, defined as
N
D ( r1 ,… , rN ) =
1
2 ∑ c r ⋅ r (2.44)
k =1
k k k
which, denoting by ∇k the gradient with respect to the kth velocity vari-
ables x k , y k , zk, is such that Fk(V) = −∇k D. The dissipative nature of these
forces is rather evident; since D is non-negative and Equation 2.44 gives
2D = − ∑k Fk(V) ⋅ rk, then 2D is the rate of energy dissipation due to these
forces.
With respect to LEs, on the other hand, we can recall the relation
∂ rk ∂ qi = ∂ rk ∂ q i given in Section 2.3 and determine that the ith gener-
alised viscous force is
N N
∂ rk
∑ ∑ ∇ D ⋅ ∂∂ qr ∂D
k
Qi(V) = − ∇k D ⋅ = k =− (2.45)
i =1
∂ qi i =1
i ∂ q i
which, in turn, means that in this case LEs are written as
d  ∂L  ∂L ∂D
− + = 0 (2.46)
dt  ∂ q i  ∂ qi ∂ q i
A final point worthy of mention is that whenever the transformation 2.3
does not involve time explicitly, the function D ( q, q ) has the form
n N
∂ rk ∂ rk
D(q, q) =
1
2 ∑
i , j =1
Cij q i q j , Cij (q) = ∑c
k=1
k ⋅
∂ qi ∂ q j
(2.47)
(with Cij = C ji) and, just like the kinetic energy, is a homogeneous function
of order two in the generalised velocities to which Euler’s theorem (Remark
2.8(i)) applies. Then, using this theorem together with Equation 2.38, we
are led to

dh
dt
= ∑Q
i
(V)
i q = −
i ∑ ∂∂ qD q = −2D (2.48)
i
i
i
which, as shown above, confirms the physical interpretation of the

‘velocity-dependent potential’ D and the dissipative nature of these forces.
Remark 2.8
i. In the more general case in which the system is acted upon by mono-
genic, viscous and non-monogenic forces, LEs are written as
d  ∂L  ∂L ∂D
− + = Qi (2.49a)
dt  ∂ q i  ∂ qi ∂ q i
where, since monogenic forces are part of the Lagrangian and

D accounts for viscous forces, the r.h.s. term accounts for the
non-monogenic forces. For this case, moreover, it is left to the

reader to show that Hamilton’s equations have the form
∂H ∂D ∂H
p i = − − + Qi , q i = (i = 1,… , n) (2.49b)
∂ qi ∂ q i ∂ pi
ii. In some cases, the function D is homogeneous of some order m other

than two with respect to the generalised velocities. Then, the r.h.s. of
Equation 2.48 is −mD and the terms dry friction and aerodynamic
drag are frequently used for m = 1 and m = 3, respectively. Obviously,
m = 2 is the viscous damping case considered earlier.
2.4.5 More co-ordinates than DOFs:

Lagrange’s multipliers
Sometimes it may be convenient to work with more co-ordinates than there
are DOFs. Since – recalling Remark 2.3(i) – we know that this is a possibil-
ity for holonomic systems but it is what we must do if the constraints are
non-holonomic, we examine here the non-holonomic case.
Consider a system with generalised co-ordinates q1 , , qn and m non-
holonomic constraint equations. Then, we have
n n
∑j =1
alj dq j + al dt = 0 ⇒ ∑ a δq
j =1
lj j =0 (l = 1, , m) (2.50)
where the first relation is the differential form of the constraints while the
second is the relation that, on account of 2.501, must hold for the virtual
displacements δ q j . Now, since the constraints 2.501 are non-integrable and
cannot be used to eliminate m co-ordinates/variables in favour of a remain-
ing set of n − m co-ordinates, we must tackle the problem by retaining all
the variables. By proceeding as in Section 2.3, we arrive at Equation 2.10,
but now the δ q j s are not independent. If, however, we multiply each one of
Equations 2.502 by an arbitrary factor λl (called Lagrange multiplier) and
form the sum
m n n
 m

∑ ∑
l =1
λl
j =1
alj δ q j = ∑ ∑ λ a  δ q
j =1

 l =1
l lj j =0
and subtract it from Equation 2.10, we get
n
 d  ∂T  ∂T m

∑j =1
  −
 dt  ∂ q j  ∂ q j
− Qj − ∑l =1
λl alj  δ q j = 0 (2.51)

where nothing has changed because we simply subtracted zero. The δ q j ,

however, are still not independent, being connected by the m relations 2.502 .
But since the multipliers λ1 ,..., λm are arbitrary and at our disposal, we can
choose them in such a way that m terms of the sum 2.51 – say, the last m
terms – are zero. Then
m
d  ∂T  ∂T
−
dt  ∂ q j  ∂ q j
− Qj − ∑λ a
l =1
l lj =0 ( j = n − m + 1,… , n ) (2.52)
and Equation 2.51 reduces to the sum of the first n − m terms, where now
the remaining δ q j are independent. Consequently, we obtain
m
d  ∂T  ∂T
−
dt  ∂ q j  ∂ q j
= Qj + ∑λ al =1
l lj ( j = 1,… , n − m) (2.53)
Together, Equations 2.52 and 2.53 form the complete set of n LEs for non-
holonomic systems. But this is not all; these equations and the m equa-
tions of constraints (expressed in the form of the first-order differential
equations ∑ j alj q j + al = 0) provide n + m equations to be solved for the n + m
unknowns q1 , , qn , λ1 , , λm.
In particular, if the generalised forces Qj – which, we recall, correspond
to applied forces – are conservative (or monogenic), then the potential V
(or V) is part of the Lagrangian and we obtain the standard non-holonomic
form of Lagrange equations
m
d  ∂L  ∂L
−
dt  ∂ q j  ∂ q j
= ∑λ a
l =1
l lj ( j = 1,… , n ) (2.54)
Equations 2.53 or 2.54 also suggest the physical meaning of the λ -multipli-
ers: the r.h.s. terms (of LEs) in which they appear are the generalised forces
of constraints. The method, therefore, does not eliminate the unknown con-
straint forces from the problem but provides them as part of the solution.
Remark 2.9
i. As mentioned earlier, the method of Lagrange multipliers can also be

used in the case of m holonomic constraints fl ( q1 , , q3N , t ) = 0, with
l = 1, , m. Since these constraints’ equations imply
∑ ∂∂qf q + ∂∂ft = 0 (2.55)

j =1
l
j
j
l
a comparison with Equations 2.50 shows that this is the special

case that we get when the alj , al have the form alj = ∂ fl ∂ q j and
al = ∂ fl ∂ t, respectively. For holonomic constraints, however, we
do not need to use the multipliers method unless we are specifically
interested in the constraint forces.
ii. Following the developments of section 2.3.1, it is now not difficult
to determine that Lagrange’s equations in the non-holonomic form
(2.54) lead to Hamilton’s equations in the form
m
∂H ∂H
p j = −
∂ qj
+ ∑λ a
l =1
l lj
j,
+Q q j =
∂ pj
( j = 1,… , n )
2.5 HAMILTON'S PRINCIPLE
For holonomic systems, Hamilton’s principle belongs rightfully to the

branch of mathematics known as calculus of variations. In its ‘classical’
form, in fact, the principle is expressed as
t1
δS =δ
∫ L (q, q , t) dt = 0 (2.56)
t0
t1
where S ≡
∫t0
L dt is called the action – or action integral or action func-
tional – L is the system’s Lagrangian L = T − V and t0 , t1 are two fixed
instants of time. In words, the principle may be stated by saying that the
actual motion of the system from time t0 to t1 is such as to render the
action S stationary – in general, a minimum – with respect to the func-
tions qi (t) ( i = 1, , n ) for which the initial and final configurations qi (t0 )
and qi (t1) are prescribed.
Remark 2.10
For a given Lagrangian L, the action S is, in mathematical terminology, a

t1
functional because it assigns the real number
∫ t0
L dt to each path qi (t) in
the configuration space. Then, broadly speaking, the condition δ S = 0 –
where, in the context of functionals, δ S is called the first variation of S – is
the ‘variational counterpart’ of the familiar condition df dx = 0 of basic
calculus, which identifies the stationary points of the function f (x). Also,
we recall from basic calculus that a stationary point, say x = x0 , can cor-
respond to a minimum, a maximum (i.e. the ‘true extremals’) of f (x) or
to an inflection point with a horizontal tangent. The general situation is
quite similar for functionals, and the analogy extends also to the fact that
the character of the extremum can be judged on the basis of the sign of the
second variation δ 2S (although in most applications of mechanics, the sta-
tionary path turns out to be a minimum of S). As for terminology, it is quite
common (with a slight abuse of language) to refer to δ S = 0 as a principle
of least action.
In order to see that Hamilton’s principle is a fundamental postulate from

which we can obtain the equations of motion, let us consider briefly the
type of manipulations involved in the calculus of variations (for a detailed
account, the reader can refer to Gelfand and Fomin (2000) or to Vujanovic
and Atanackovic (2004)) and show that LEs follow directly from 2.56. We
only do it for a 1-DOF system for two reasons: first, because it does not
affect the essence of the argument, and second, because the generalisation
to an n-DOFs holonomic system is rather straightforward.
Let q(t) be a path (in configuration space) satisfying the end condi-
tions q ( t0 ) = q(0) , q ( t1 ) = q(1) and let us consider a neighbouring varied
path q (t) = q(t) + ε r(t), where ε is an infinitesimal parameter and r(t) is an
arbitrary continuously differentiable function such that r ( t0 ) = r ( t1 ) = 0
(so that also the varied path q (t) satisfies the end conditions at t0 and t1).
Then, since in variational notation, it is customary to denote the increment
ε r(t) = q (t) − q(t) by δ q(t) – thus implying that the end conditions are written
δ q ( t0 ) = δ q ( t1 ) = 0 – and call it as variation of q, we obtain the variation δ S
corresponding to δ q(t) as
t1 t1
 ∂L ∂L 
δS =
∫ L (q + δ q, q + δ q , t ) − L (q, q , t ) dt = ∫  ∂ q δ q + ∂ q δ q  dt
t0 t0

t1 t1
∂L   ∂L d  ∂L
=
 ∂ q
δq + 
∫ −    δ qdt
 t0 t0  ∂ q dt  ∂ q  
(2.57)
where the last expression is obtained by integrating by parts the term

(∂ L ∂ q ) δ q under the assumption that δ q = d (δ q )
dt holds, that is, that the
variation operator δ commutes with the time derivative (see Remark 2.11(i)
below). Owing to the zero end conditions on δ q, the boundary term within
square parentheses on the r.h.s. vanishes and only the other term survives.
But then in order to have δ S = 0 for an arbitrary variation δ q, the term
within parenthesis must be zero. Therefore, it follows that the Lagrange
equation
∂L d  ∂L
− = 0 (2.58)
∂ q dt  ∂ q 
is a necessary condition for the path q(t) to be an extremal of the action

S. Then, on account of the fact that for an n-DOFs holonomic system,
the total variation δ S is the sum of the variations δ S1 , , δ Sn correspond-
ing to the independent variations δ q1 , , δ qn satisfying the end conditions
δ qi ( t0 ) = δ qi ( t1 ) = 0 ( i = 1, , n ); the generalisation to n-DOFs systems is
straightforward. The conclusion is that the n LEs 2.14 provide a necessary
condition for the path qi (t) in the n-dimensional configuration space to be
t1
an extremal of the action integral S =
∫t0
L ( q1 ,… , qn , q1 ,… , q n , t ) dt .
In addition to the above considerations, it is now instructive to follow
a different route and start from d’Alembert’s principle in the form 2.5.
Integrating between two instants of time t0 , t1, the term corresponding to
t1
the applied forces Fk(a) becomes
∫ t0
δ W (a) dt – where δ W (a) is the virtual
work of these forces – while for the term corresponding to inertia forces, we
can write the chain of relations
t1 t1 t
  1
∑∫
k
d
dt
( mkrk ) ⋅ δ rk dt = 

∑
k
mkrk ⋅ δ rk  −
 t t0 ∫ ∑ m r ⋅ δ r dt
k
k k k
t0 0

t1
  k 
∫ ∑ 2  dt = [...] − ∫
m r ⋅r k k
t1
= [...] tt10 − δ t1
t0 δ T dt
t0
 k
t0
(2.59)
where we first integrated by parts, took into account that the fact that the
δ -operator commutes with the time derivative, and then used the relation
δ ( mk rk ⋅ rk 2) = mk rk ⋅ δ rk by observing that δ ( mk rk ⋅ rk 2) = δTk (thereby
implying that ∑k δTk is the variation δT of the system’s total kinetic energy).
If now we further assume that at the instants t0 , t1, the position of the sys-
tem is given, then δ rk ( t0 ) = δ rk ( t1 ) = 0 and the boundary term in square
brackets vanishes. Finally, putting the pieces together, we arrive at the
expressions
t1 t1

∫(
t0
)
δ W (a) + δ T dt = 0,
∫ δ L dt = 0 (2.60)
t0
where the first expression is often called the extended Hamilton’s principle,
while the second follows from the first when all the applied forces are deriv-
able from a potential function V (q, t), and we have δ W (a) = −δ V . Also, we
notice that since δ W (a) and δ T (and δ L) are independent on the choice of
co-ordinates, we can just as well see the principles in terms of generalised
co-ordinates (with, clearly, the end conditions δ q j ( t0 ) = δ q j ( t1 ) = 0 for

j = 1, , n). If, moreover, the system is holonomic, then the δ -operator com-
mutes with the definite integral and Equation 2.602 becomes
t1
δ
∫ L dt = 0 (2.61)
t0
which is, as mentioned at the beginning of this section, the classical form
of Hamilton’s principle. In this respect, however, it is important to point
out that the classical form 2.61 does not apply to non-holonomic systems
because in this case the shift from 2.602 to 2.61 cannot be made. For non-
holonomic systems, in fact, it can be shown that the varied path is not in
general a geometrically possible path; this meaning that the system cannot
travel along the varied path without violating the constraints. Consequently,
the correct equations of motion for non-holonomic systems are obtained by
means of Equations 2.60, which, it should be noticed, are not variational
principles in the strict sense of the calculus of variations, but merely inte-
grated forms of d’Alembert’s principle. Detailed discussions of these aspects
can be found in Greenwood (2003), Lurie (2002), Rosenberg (1977) and
in the classical books by Lanczos (1970) and Pars (1965). In any case, the
great advantage of Hamilton’s principle (in the appropriate form, depend-
ing on the system) is that it can be used to derive the equations of motion
of a very large class of systems, either discrete or continuous. In this latter
case, moreover, we will see in the next section that the principle automati-
cally provides also the appropriate spatial boundary conditions.
Remark 2.11
i. Since the δ -operator corresponds to variations (between the actual

and varied path) in which time is held fixed – i.e. the so-called con-
temporaneous variation – it is eminently reasonable to expect it to
commute with the time derivative. However, it should be noticed that
this ‘commutative property’ may not hold for variations that are dif-
ferent from the (contemporaneous) δ -variations considered here.
ii. In the case in which some forces are derivable from a potential func-
tion and some others are not, the extended Hamilton’s principle of
Equation 2.601 expressed in terms of generalised co-ordinates has the
form
t
1
 n

 δ L +
 ∫ ∑ Q δ q  dt = 0 (2.62)
i =1
i i
t0
 i are the ‘non-potential’ generalised

where, as in Equation 2.15, Q
forces.
2.5.1 More than one independent variable:

continuous systems and boundary conditions
So far we have considered Hamilton’s principle in the case of one indepen-
dent variable (time t ) and one or more dependent variables (the functions
q1(t), , qn (t)). However, a different case important in many applications
is when we have more than one independent variables and one function
u of these variables. In particular, for our purposes, we restrict our atten-
tion to cases in which the independent variables are time t together with at
most two spatial variables x1 , x2 and the dependent variable is a function
u = u ( x1 , x2 , t ). Physically, these cases represent continuous systems with an
infinite number of DOFs extending over a finite 1- or 2-dimensional spatial
region/domain R with boundary S. Typical examples, as we will see shortly
(and in Chapter 5), are strings, bars, membranes and plates.
In a first ‘subclass’ of these cases, the action integral has the form
t1 t1
S=
∫ L ( x , x , t, u, ∂ u, ∂ u, ∂ u) dt = ∫ ∫ Λ ( x , x , t, u, ∂ u, ∂ u, ∂ u) dx dt
t0
1 2 t 1 2
t0 R
1 2 t 1 2
(2.63)
where dx = dx1dx2 and we write ∂t u, ∂ j u for the partial derivatives ∂u ∂t

and ∂u ∂ x j ( j = 1, 2). Also, note that in Equation 2.63 we introduced the
Lagrangian density Λ, which is such that L =
∫
R
Λ dx, that is, its integral
over the region R gives the Lagrangian L of the system.
In order to illustrate the main ideas without getting too much involved in
the cumbersome calculations, consider the case of only one spatial variable x
(so that u = u(x, t)), a 1-dimensional spatial domain that extends from x = 0
to x = l and a system Lagrangian density of the form Λ = Λ( x, t , u, u , u′ ),
where here for brevity of notation, we write u , u′ for the time derivative ∂ t u
and the spatial derivative ∂ x u, respectively.
Hamilton’s principle δ S = 0 for this case is
t1 l
 ∂Λ ∂Λ ∂Λ 

∫ ∫  ∂ u δ u + ∂ u δ u + ∂ u′ δ u′ dxdt = 0 (2.64a)
t0 0
where the variation δ u(x, t) is required to be zero at the initial and final
times t0 , t1, i.e.
δ u ( x, t0 ) = δ u ( x, t1 ) = 0 (2.64b)
At this point, in order to have all variations expressed in terms of δ u, we can

integrate by parts the second and third term on the l.h.s. of Equation 2.64a
under the assumption that the δ -operator commutes with both the time-
and spatial derivatives. For the second term, the integration by parts is with
respect to time, and we get
l  t1  l  t1 t1 
 ∂Λ   ∂ Λ  ∂  ∂Λ 
∫ ∫
0

 t0 ∂ u
δ u dt  dx =

∫
0

 ∂ u
δ u
 t0
−
∫
t0
  
∂t ∂u
δ u dt  dx

(2.65a)
l t1
∂  ∂Λ
=−
∫ ∫ ∂ t  ∂ u  δ u dt dx
0 t0
where in writing the last equality, we took the conditions 2.64b into
account. For the third term of 2.64a, on the other hand, the integration by
parts is with respect to x and we get
t1
 l ∂Λ  t1
 ∂Λ l l
∂  ∂Λ 
    

∫ ∫
t0

 0 ∂ u ′
δ u′ dx  dt =

∫
t0
 δ u −
 ∂ u′  0
∫ 0

∂ x ∂ u′
 δ u dx  dt (2.65b)

Using Equations 2.65a and 2.65b in 2.64a, the result is that we have trans-
formed the l.h.s. of Equation 2.64a into
t1 l t1 l
∂ Λ ∂  ∂ Λ  ∂  ∂ Λ    ∂Λ 

∫∫
t0 0

 ∂ u
− 
∂ t  ∂ 
u
 −
∂ x

∂ u ′
  δ u dx dt + 
 t0
∂ u′  ∫
δ u dt (2.66)
0
which, according to Hamilton’s principle, must vanish for the function

u(x, t) that corresponds to the actual motion of the system. Now if we first
suppose that δ u(x, t) is zero at the extreme points x = 0 and x = l, i.e.
δ u (0, t ) = δ u ( l , t ) = 0 (2.67)
(note that if δ S vanishes for all admissible δ u(x, t), it certainly vanishes for
all admissible δ u(x, t) satisfying the extra condition 2.67), then the second
integral in 2.66 is zero and only the double integral remains. But then,
owing to the arbitrariness of δ u, the double integral is zero only if
∂Λ ∂  ∂Λ ∂  ∂Λ 
−  −   = 0 (2.68a)
∂ u ∂ t  ∂ u  ∂ x  ∂ u′ 
which must hold for 0 ≤ x ≤ l and all t. This is the Lagrange equation of
motion of the system.
Now we relax the condition 2.67; since the actual u( x, t ) must satisfy
Equation 2.68a, the double integral in 2.66 vanishes and the Hamilton’s
principle tells us that we must have
 ∂Λ   ∂Λ 
 δ u = 0,  δ u = 0 (2.68b)
∂ u′  x=0
∂ u′  x=l
which provide the possible boundary conditions of the problem. So, for
example, if u is given at x = 0 then δ u(0, t) = 0, and the first boundary condi-
tion of 2.68b is automatically satisfied. If, however, u is not pre-assigned at
x = 0, then Equation 2.68b1 tells us that we must have ∂ Λ ∂ u′ x = 0 = 0. Clearly,
the same applies to the other end point x = l. As for terminology, we note the
following: since the boundary condition δ u(0, t) = 0 is imposed by the geom-
etry of the problem, it is common to call it a geometric (or imposed) bound-
ary condition. On the other hand, the boundary condition ∂ Λ ∂ u′ x = 0 = 0
depends on Λ – that is, on the nature of the system’s kinetic and potential
energies, and consequently on inertial effects and internal forces – and for
this reason, it is referred to as a natural (or force) boundary condition.
Remark 2.12
i. Note that in the calculations of the derivatives of Equation 2.68a, the

quantities u, u and u′ are treated as independent variables (see the fol-
lowing Examples 2.4 and 2.5).
ii. If we have two spatial variables – that is, u = u ( x1 , x2 , t ), like, for
example, a membrane – then with a Lagrangian density of functional
form Λ = Λ ( x1 , x2 , t , u, ∂ t u, ∂1 u, ∂ 2 u ), Equation 2.68a becomes
2
∂Λ ∂  ∂Λ   
−
∂ u ∂ t  ∂ (∂ t u) 
− ∑ ∂∂x  ∂ (∂∂Λu)  = 0
j =1
j j
(2.69)
where here we reverted to the derivatives notation used at the

beginning of this section.
iii. For a class of more complex systems (for example, beams and plates),
the arguments of the Lagrangian density include also second-order
derivatives. In these cases, it can be shown that in the l.h.s. of Equation
2.69, we have the additional term
2
∂2  ∂Λ 
∑
j , k=1
∂ x j ∂ xk  ∂ (∂ jk2 u) 
(2.70a)
with also additional boundary conditions (see the following

Example 2.5).
iv. Using the standard notation x1 = x, x2 = y, it should be noticed that

for a 1-dimensional system (e.g. beams), Equation 2.70a gives only
one term, which is
∂2  ∂Λ 
(2.70b)
∂ x2  ∂ (∂ xx
2
u) 
On the other hand, for a 2-dimensional system (e.g. plates),
Equation 2.70a gives the four terms
∂2  ∂Λ  ∂2  ∂Λ  ∂2  ∂Λ  ∂2  ∂Λ 
+ 2 + + (2.70c)
∂ x  ∂ (∂ xx u)  ∂ y  ∂ (∂ yy u)  ∂ x ∂ y  ∂ (∂ xy u)  ∂ y ∂ x  ∂ (∂ yx
2  2  2   2  2
u) 
Example 2.4
As an application of the considerations above, we obtain here the equa-
tion of motion for the longitudinal (or axial) vibrations of a bar with
length l, mass per unit length m ˆ (x), cross-sectional area A(x) and axial
stiffness EA(x) – a typical continuous 1-dimensional system. If we let
the function u(x, t) represent the bar’s axial displacement at point
x (0 < x < l ) at time t, it is not difficult to show (see, for example, Petyt
(1990)) that the Lagrangian density for this system is
2 2
1  ∂u 1  ∂ u  (2.71)
Λ ( x, ∂ t u, ∂ x u ) = ˆ
m − EA 
2  ∂ t  2  ∂ x 
where, respectively, the two terms on the r.h.s. are the kinetic and
potential energy densities. Then, noting that the various terms of
Equation 2.68a are in this case explicitly given by
∂Λ ∂  ∂Λ  ∂2u ∂  ∂Λ  ∂  ∂u 
= 0, = ˆ
m , =− EA
∂u ∂ t  ∂ (∂ t u)  ∂ t2 ∂ x  ∂ (∂ x u)  ∂ x  ∂ x 
(2.72)
we obtain the second-order differential equation of motion
∂  ∂u  ∂2u
EA − ˆ
m = 0 (2.73a)
∂ x  ∂ x  ∂ t2
Also, from Equation 2.68b, we get the boundary conditions
∂u ∂u
EA δu = 0, EA δu = 0 (2.73b)
∂x x =0 ∂x x=l
Example 2.5
Consider now the transverse (or flexural) vibration of the same bar
of Example 2.4, where here u(x, t) represents the transverse displace-
ment of the bar whose flexural stiffness is EI (x), where E is the Young’s
modulus and I (x) is the cross-sectional moment of inertia. Since the
Lagrangian density is now
2 2
1  ∂u 1  ∂2 u
Λ ( x, ∂ t u, ∂ xx u ) = ˆ
m − EI  ∂ x2  (2.74)
2  ∂ t  2
we have a second-order derivative in the potential energy, and conse-

quently, we must take into account the additional term of Equation
2.70b. Explicitly, this term is
∂2  ∂Λ  ∂2  ∂2u
 =− 2 EI
∂ x2 (
 ∂ ∂ x2 x u )  ∂x  ∂ x2 
and it is now left to the reader to calculate the other terms, put the
pieces together and show that the result is the fourth-order differential
equation of motion
∂2  ∂2u  ∂2u
 EI ∂ x2  + m
ˆ = 0 (2.75)
∂ x2 ∂ t2
Given this result, it may be nonetheless instructive to obtain the equa-

tion of motion 2.75 and the corresponding boundary conditions
directly from the Hamilton’s principle (with the assumption that the
δ -operator commutes with both the time and spatial derivatives).
Denoting indifferently the time derivative by an overdot or by ∂t , the
t1
term
∫ t0
δ T dt is given by
l t1 l t1
1
2 ∫∫ ( )
ˆ δ u 2 dt dx =
m
∫ ∫ mu
ˆ  ∂ (δ u ) dt dx
t
0 t0 0 t0
l l t1 t1 l
=
∫
0
ˆ u δ u  tt10 dx −
 m
∫∫
0 t0
 δ u dxdt = −
ˆ u
m
∫ ∫ mˆ u δ u dxdt
t0 0
(2.76)
where the first equality is a consequence of the relations

δ ( u u ) = 2u δ ( ∂t u ) = 2u ∂t (δ u ), the second is an integration by parts in
the time integral and the third is due to the assumption of vanishing δ u
at the instants t0 , t1. A similar line of reasoning – in which the slightly
lengthier calculation now requires two integrations by parts in the dx
t1
integral – gives the term
∫ t0
δ V dt as
t1 l t1 l
 ′ 

∫
t0
 EI u ′′ δ ( u ′ ) − ( EI u ′′ ) δ u  dt +
0
∫ ∫ (EI u′′)′′ δ u dx dt (2.77)
t0 0
where the primes denote the spatial derivatives. Putting Equations 2.76
and 2.77 together and reverting to the standard notation for deriva-
tives, we arrive at Hamilton’s principle in the form
t1  l
  ∂ 2 u   ∂ u   
l
 ∂  ∂2u 
∫
t0
 
∂x 
 
EI δ u
∂ x2  

0
−  EI  2 δ   dt

  ∂ x   ∂ x   0 


t1 l
 ∂ 2  ∂2u  ∂ 2 u  
−
∫ ∫  ∂ x
t0 0
2 

EI 2
∂x 
ˆ  2   δ u dx dt = 0
+m
 ∂ t  
which gives the equation of motion 2.75 and the four boundary
conditions
x=l x=l
 ∂  ∂2u    ∂2u  ∂u 
  EI δ u = 0,  EI  2  δ   = 0 (2.78)
 ∂ x  ∂ x2     ∂ x   ∂ x   x=0
x =0
where it should be noticed that four is now the correct number because
Equation 2.75 is a fourth-order differential equation.
So, for example, if the bar is clamped at both ends, then the geom-
etry of the system imposes the four geometric boundary conditions
∂u ∂u
u (0, t ) = u ( l , t ) = 0, = = 0 (2.79)
∂x x =0 ∂x x=l
and there are no natural boundary conditions.

If, on the other hand, the bar is simply supported at both ends, the
geometric boundary conditions are two, that is, u (0, t ) = u ( l , t ) = 0. But
then, since the slope ∂ u ∂ x is not preassigned at either end, Equation
2.782 provides the two missing (natural) boundary conditions
 ∂2u  ∂2u
EI  2  = 0, EI  2  = 0 (2.80)
 ∂x  x =0
 ∂x  x=l
which physically mean that the bending moment must be zero at both
ends. This is also the case at a free end, where together with zero bend-
ing moment, we must have the (natural) boundary condition of zero
transverse shear force. Denoting by primes derivatives with respect to
x , this condition on shear force is expressed by ( EIu ′′ )′ = 0.
2.6 SMALL-AMPLITUDE OSCILLATIONS
As shown in Example 2.1, the equation of motion of a simple pendulum is

θ + ( g l ) sin θ = 0, which is non-linear because of the sine term. However, if
θ << 1, the approximation sin θ ≈ θ for small oscillation angles (‘small’ with
respect to the rest position θ = 0) gives the linear equation θ + ( g l ) θ = 0
and leads – as shown in every physics textbook – to the period of oscillation
T = 2π l g and to the corresponding frequency ω = g l .
Remark 2.13
We also recall here from basic physics that the compound (or physical)
pendulum is a rigid body of mass m pivoted at a point O distant d from its
centre of mass G. Since the body is free to rotate under the action of grav-
ity, the equation of motion is θ + (Wd JO ) sin θ = 0, where W = mg is the
weight of the body and JO is its moment of inertia about a horizontal axis
passing through the centre of rotation O. Then, for small oscillations, we
get TCP = 2π JO Wd and ω CP = Wd JO , thus implying that in terms of
period and frequency of oscillation, our compound pendulum is equivalent
to a simple pendulum with length L = JO md (which is sometimes called the
reduced length of the compound pendulum).
Similarly, in the approximation of small angles (relative to the rest posi-

tion θ = φ = 0), the equations of motion 2.26 of the double pendulum of
Example 2.2 become
( m1 + m2 ) l1θ + m2 l2 φ + ( m1 + m2 ) gθ = 0, l2 φ + l1 θ + gφ = 0 (2.81)
which, it should be noticed, can be obtained from the small-amplitude

Lagrangian
m1 + m2 2  2 m2 2 2   − m1 + m2 gl1θ 2 − m2 gl2φ 2 (2.82)

L= l1 θ + l2φ + m2l1l2 θφ
2 2 2 2
(where we ignored the constant terms because they have no effect upon
differentiation).
These introductory considerations are the starting point for the discus-
sion of this section, which concerns the linearisation of the equations of
motion and is fundamental for the subject of linear vibrations.
A first observation in this respect is that ‘linearisation’ means lineari-
sation about some reference state, that is, some reference point, with the
understanding that the relevant variables undergo small deviations relative
to the reference point and that the system’s motion remains within a small
neighbourhood of that point. If, as it is often the case, the reference point
is the equilibrium position, this circumstance carries with it three implica-
tions: (a) that the equilibrium point must be a solution of the (non-linear)
equations of motion such that qi = qi 0 = const ( i = 1, , n ), and consequently
q i 0 = 0, (b) that the reference state must be a stable equilibrium point (oth-
erwise a small departure from an unstable equilibrium would lead to a
growing state of motion that does not remain in the neighbourhood of the
equilibrium point), and (c) that non-linear terms can be approximated by
the linear terms in their Taylor series expansion.
Given these preliminary considerations, we can now return to the pendu-
lum examples above and note that we first obtained the exact expressions
for T and V, derived the equations of motion and then linearised them. This
is perfectly legitimate, but since it is generally more convenient to approxi-
mate the expressions of T and V in the first place and then use the resulting
Lagrangian (as, for example, the Lagrangian of Equation 2.82) to obtain
the linearised equations of motion, this is what we do now by first consider-
ing an n-DOF natural system – that is, we recall, a holonomic system for
which the T1 and T0 parts of the kinetic energy are zero.
In this case, the equilibrium positions are the stationary points of the
potential energy – i.e. the points such that ∂ V ∂ qi = 0 ( i = 1, , n ) – with
the additional requirement that a stable equilibrium point corresponds to a
relative minimum of V. Observing that in many practical cases, the stable
equilibrium point can be identified by inspection, let us denote by q10 , , qn0
(q0 for brevity) the generalised co-ordinates of this equilibrium position and
let ui = qi − qi 0 be small variations from this position. We can then expand
the potential energy as (all sums are from 1 to n)
 ∂V   ∂ 2V 
V = V ( q0 ) + ∑ i
 ∂ q 
i q =q
0
ui +
1
2 ∑
i,j
 ∂q ∂q 
 i j  q=q0
ui u j +  (2.83)
and note that the term V ( q0 ) can be ignored (because an additive constant
is irrelevant in the potential energy) and that the first sum is zero because
of the equilibrium condition. The expansion, therefore, starts with second-
order terms, thus implying that if, under the small-amplitude assumption,
we neglect higher-order terms, we are left with the homogeneous quadratic
function of the ui
 ∂ 2V 
V≅
1
2 ∑
i,j
 ∂q ∂q 
 i j  q=q0
ui u j =
1
2 ∑ k u u (2.84)
i,j
ij i j
where the second expression defines the constants kij , called stiffness coef-
ficients, that are symmetric in the two subscripts (i.e. kij = kj i ).
For the kinetic energy, on the other hand, we already have a homoge-
neous quadratic function of the generalised velocities, because T = T2 for
natural systems. Therefore, we can write
T=
1
2 ∑ M (q) q q
i, j
ij i j =
1
2 ∑ M (q) u u ≅ 12 ∑ M (q ) u u (2.85)
i, j
ij i j
i, j
ij 0 i j
where in the last expression, we directly assigned to the functions Mij (q)
their equilibrium values Mij ( q0 ). Then, denoting by mij the mass-coefficients
Mij ( q0 ) in Equation 2.85 and observing that mij = m j i because they inherit
this property from the original coefficients Mij (q), the ‘small-amplitude’
Lagrangian of the system is
L ( u, u ) =
1
2 ∑ (m
i, j
ij u i u j − kij ui u j ) (2.86)
Finally, since for r = 1,..., n, we get

d  ∂L 
dt  ∂u r 
= ∑m j
rj j ,
u
∂L
∂ur
=− ∑k u
j
rj j
Lagrange’s equations 2.14 lead to the equations of motion
∑ (m
j
rj j + krj u j ) = 0,
u  + Ku = 0 (2.87)
Mu
where the first relation is a set of n equations ( r = 1, , n ), while the second
is the more compact matrix form obtained by introducing the n × n matrices
M, K of the m- and k-coefficients, respectively, and the n × 1 column vectors
, u. Clearly, the matrix versions of the properties mij = m j i and kij = kj i are
u
M = MT and K = KT , meaning that both the mass and stiffness matrices are
symmetric.
Remark 2.14
i. In matrix notation, the Lagrangian 2.86 is written as
L=
1 T
2
(
u Mu − uT Ku (2.88) )
ii. If, as a specific example, we go back to the double pendulum at the
beginning of this section, the matrix form of the Lagrangian 2.82 is
 ( m1 + m2 ) l12 m2 l1 l2   θ
( )

2L = θ φ 
 m2 l1 l2

m2l22   φ
−
 ( θ φ )
 ( m1 + m2 ) g l1 0  θ 
×  
 0 m2 g l2   φ 
from which it is evident that the mass- and stiffness matrix are both
symmetric. And since – as shown in Appendix A and as we will see
in future chapters – symmetric matrices have a number of desirable
properties, the fact of producing symmetric matrices is an impor-
tant feature of the Lagrangian method.
iii. In the special case of a simple 1-DOF system, Equations 2.87 reduce
to the single equation mu  + ω 2u = 0 if one defines
 + ku = 0 , or u
ω = k m . If u is a Cartesian co-ordinate (say, the familiar x), then m
2
is the actual mass of the system.
If our system is also acted upon by dissipative viscous forces, we can recall
the developments of Section 2.4.4 and notice the formal analogy between
T2 and the Rayleigh dissipation function D. This means that – as we did for
the Mij (q) – we can expand the coefficients Cij (q) and retain only the first
term Cij ( q0 ) ≡ cij to obtain
D≅
1
2 ∑c
i, j
ij u i u j (2.89)
Then, owing to Lagrange’s equations in the form 2.46, we are led to the
equations of motion
∑ ( m u + c
j
rj j rj u j + krj u j ) = 0,  + Cu + Ku = 0 (2.90)
Mu
where, as for Equations 2.87, the first relation holds for r = 1, , n while
the second is its matrix form (and the damping matrix is symmetric, i.e.
C = CT , because cij = c j i).
For non-natural systems, on the other hand, we may have two cases:
(i) T1 = 0 and T0 ≠ 0 or (ii) both T1 , T0 non-zero. In the first case, the sys-
tem can be treated as an otherwise natural system with kinetic energy
T2 and potential energy U = V − T0 (i.e. the dynamic potential of Section
2.4.2), thus implying that the stiffness coefficients are given by the sec-
ond derivatives of U instead of the second derivatives of V. In the second
case, we can recall from Section 2.4.2 that T1 = ∑ i bi (q) q i. Expanding in
Taylor series the coefficients bi , the small amplitude approximation gives
T1 ≅ ∑ i bi (q0 ) u i + ∑ ij bij u i u j , where bij = (∂ bi / ∂ q j )q = q . When this approxi-

0
mate T1-term is inserted in the Lagrangian, the calculation of the prescribed

derivatives leads to the term ∑ (b j
rj − bjr ) u j in the rth equation of motion.
Then, by defining the gyroscopic coefficients grj = brj − bjr and passing to the
matrix notation, we have G = B − BT , where it is immediate to see that the
gyroscopic matrix G is skew-symmetric (i.e. G = − GT ) because grj = − g jr. In
the end, all this means that in the matrix form of the equations of motion,
we have an additional term Gu. 
Finally, generalised external forces that are non-conservative are
accounted for by a term Qr on the r.h.s. of Equations 2.871 or 2.901. In the
matrix version of these equations, this means that we have a n × 1 column
vector f on the r.h.s., so that, for example, we have
 + Cu + Ku = f (2.91)
Mu
for a system with dissipative viscous forces. The components of f are the
generalised forces Q1 , , Qn, but we do not denote this vector by q to avoid
possible confusion with a vector of generalised co-ordinates.
As for terminology, the equations with a zero r.h.s. define the so-called
free-vibration problem, while the equations with a non-zero r.h.s. define
the forced vibration problem; we will consider the solutions of both prob-
lems in the following chapters.
Remark 2.15
The linearised equations of motion in the form 2.90 or 2.91 are sufficiently
general for many cases of interest. However, it is worth observing that in
the most general case, the u and u terms, respectively, are (C + G)u and
(K + H) u, where G is the gyroscopic matrix mentioned above while the
skew-symmetric matrix H (denoted by N by some authors) is called circula-
tory. For these aspects, we refer the interested reader to Meirovitch (1997),
Pfeiffer and Schindler (2015) or Ziegler (1977).
2.7 A FEW COMPLEMENTS
2.7.1 Motion in a non-inertial frame of reference

In an inertial frame of reference – we call it K, with origin O – the Lagrangian
of a moving particle P of mass m in a potential field is L = mv 2 2 − U(r),
where r, v are the particle position and velocity vectors (with v 2 = v ⋅ v),
respectively, and in this whole section, we denote the potential by U in
order not to generate confusion with the velocity V (with modulus V) to

be defined shortly. From the Lagrangian L, we readily obtain the familiar
equation of motion ma = − ∂ U ∂ r.
If, however, we now consider (as it may be convenient in some types of
problems) another frame of reference K ′ that – relative to K – rotates with
angular velocity w(t) and whose origin O′ moves with translational velocity
V(t), then (as pointed out in Section 2.4.1) the form of Lagrange’s equa-
tions does not change but the Lagrangian L′ of this non-inertial observer
will differ from L because it will be expressed in terms of r ′, v′ , that is, the
particle position and velocity relative to K ′. In order to determine this ‘new’
Lagrangian, we start from the fact that (as shown in most physics textbooks)
the relation between the velocities in the two frames is v = V + v′ + w × r ′.
Using this in L gives
mV 2 m v′ 2 m dr ′
+ ( w × r ′ ) + mV ⋅ + mv′ ⋅ ( w × r ′ ) − U (2.92)
2
L= +
2 2 2 dt
where dr ′ dt = v′ + ( w × r ′ ) in the fourth term on the r.h.s. is the varia-

tion of r ′ relative to K (while, it should be noticed, v′ is the variation of
r ′ relative to K ′, so that in order to distinguish it from dr ′ dt , one often
writes v′ = d ′ r ′ dt ). Then, by further observing that this fourth term can
be expressed as
dr ′ d dV
mV ⋅ = ( mV ⋅ r ′ ) − m ⋅ r ′ (2.93)
dt dt dt
where dV dt = A is the translational acceleration of O′ relative to K, we can

now introduce Equation 2.93 into 2.92. This gives the Lagrangian in terms
of the variables r ′, v′ , and we get
m v′ 2 m
+ ( w × r ′ ) − m A ⋅ r ′ + m v′ ⋅ ( w × r ′ ) − U ( r ′ ) (2.94)
2
L′ =
2 2
where, in writing Equation 2.94, we
i. ignored the terms mV 2 2 and d ( mV ⋅ r ′ ) dt because they are both

total time derivatives (see Remark 2.16 (i) below),
ii. expressed the potential U in terms of r ′ by considering that r = R + r ′,
where R is the position vector of O′ relative to K.
Passing to the equation of motion in the frame K ′, the first relation we get
from Lagrangian 2.94 is
∂ L′ d  ∂ L′ 
= mv′ + m ( w × r ′ ) ⇒   = ma′ + m ( w
 × r ′ ) + m ( w × v′ ) (2.95)
∂ v′ dt  ∂ v′ 
while for the second relation, we need the derivative ∂ L′ ∂ r ′. In order

to obtain this term, it is convenient to take into account the two rela-
tions (see Remark 2.16(ii) below): (a) v′ ⋅ ( w × r ′ ) = r ′ ⋅ ( v′ × w ) and (b)
( w × r ′ )2 = w 2r ′2 − ( w ⋅ r ′ )2.
Remark 2.16
i. It can be shown (see, for example, Goldstein (1980) or Landau and

Lifshitz (1982)) that two Lagrangians that differ by an additive term
which is the total time derivative of an arbitrary function of the co-
ordinates and time are physically equivalent in order to determine the
equations of motion.
ii. Relation (a) is due to the property of the scalar triple product that
for any three vectors a, b, c, we have a ⋅ (b × c) = b ⋅ (c × a) = c ⋅ (a × b).
On the other hand, relation (b) follows from the chain of equalities
( )
(a × b )2 = a2b2 sin2 α = a2b2 1 − cos2 α = a2b2 − (a ⋅ b )2,
where α is the angle between a and b.
iii. Also, we anticipate here that in writing the second line of the follow-
ing Equation 2.96, we use the property of the triple vector product
a × (b × c) = b (a ⋅ c) − c (a ⋅ b).
Using the relations (a) and (b) above, we get
∂ L′ ∂U
∂ r′
{ }
= m w 2 r ′ − ( w ⋅ r ′ ) w − m A + m ( v′ × w ) −
∂ r′

∂U
= m w × ( r ′ × w ) − m A + m ( v′ × w ) − (2.96)
∂ r′
where, owing to the property of Remark 2.16(iii), the term within curly
brackets in the first line of Equation 2.96 is equal to w × (r ′ × w). Finally,
by further recalling the property a × b = − b × a, Equations 2.95 and 2.96
together lead to the equation of motion in the frame K ′; this is
∂U
m a′ = − − mA − m ( w
 × r ′ ) − 2m ( w × v′ ) − m w × ( w × r ′ ) (2.97)
∂ r′
which shows that the particle acceleration in the frame K ′ is determined,

in addition to the force field − ∂ U ∂ r ′ , by additional forces – the so-called
fictitious or inertial forces – due to its non-inertial state of motion. In par-
ticular, the last three terms on the r.h.s. are due to the rotation: the first is
associated to the non-uniformity of the rotation (and is clearly zero for con-
stant angular velocity), while the second and the third are well-known from
basic physics, being, respectively, the Coriolis and the centrifugal force.
Remark 2.17
i. Since w × v′ = wv′ sin α where α is the angle between the two vectors,
the Coriolis force – which is linear in the velocity and is a typical
example of gyroscopic term mentioned at the end of Section 2.4.2 – is
zero if v′ is parallel to w. Also, the Coriolis force is always perpen-
dicular to the velocity and therefore it does no work.
ii. The centrifugal force lies in the plane through r ′ and w, is perpen-
dicular to the axis of rotation (i.e. to w) and is directed away from the
axis. Its magnitude is mw 2d, where d is the distance of the particle
from the axis of rotation.
2.7.2 Uniformly rotating frame

If now, as a special case, we assume that the frame K ′ is not translating
but only rotating with constant angular velocity w, the Lagrangian 2.94
becomes
mv′ 2 m
+ m v′ ⋅ ( w × r ′ ) + ( w × r ′ ) − U ( r ′ ) (2.98)
2
L′ =
2 2
where we readily recognise that the first, second and third term on the r.h.s.
are, respectively, the T2 , T1 and T0 parts of the kinetic energy. Then, from L′,
we obtain the equation of motion
m a′ = − ∂ U ∂ r ′ − 2m ( w × v′ ) − m w × ( w × r ′ ) (2.99)
where, as above, the last two terms on the r.h.s. are the Coriolis and the
centrifugal force. In this respect, note that the Coriolis term comes from
the T1-part of the kinetic energy while the centrifugal term comes form the
T0 -part.
In this special case, the particle generalised momentum in frame K ′ is
p′ = ∂ L′ ∂ v′ = m ( v′ + w × r ′ ), but since the term within parenthesis is the
particle velocity v relative to K, then p′ coincides with the momentum p in
frame K. If, in addition, the origins of the two reference systems coincide,
then r = r ′ and it also follows that the angular momentum M′ = r ′ × p′ in
frame K ′ is the same as the angular momentum M = r × p in frame K. As
for the energy function, in frame K ′, we have
h′ = ∑ ∂∂ Lx ′′ x ′ − L′ = p′ ⋅ v′ − L′ (2.100)
i
i
i
from which, using the above expression of p′ and L′, it follows:
mv′ 2 m mv′ 2
+ U − ( w × r′ ) =
2
h′ = + U + UCtf (2.101)
2 2 2
where in the rightmost expression, the subscript ‘Ctf’ stands for ‘centrifu-
gal’, thus showing that the rotation of K ′ manifests itself in the appear-
ance of the additional potential energy term UCtf = − m ( w × r ′ ) 2 , which is
2
independent on the particle velocity v′ . In this regard, note that the energy
contains no term linear in the velocity.
At this point, recalling that with no translational motion we have
v′ = v − w × r ′ , we can use this in Equation 2.101 to get
mv 2
h′ = + U − mv ⋅ ( w × r ′ ) = h − w ⋅ M (2.102)
2
where in writing the rightmost expression, we took into account that
i. mv 2 2 + U is the energy h in frame K,

ii. r = r ′ implies mv ⋅ ( w × r ′ ) = mv ⋅ (w × r) = w ⋅ (r × mv) = w ⋅ M,
where the last equality is due to the first property mentioned in
Remark 2.16 (ii).
Equation 2.102 is the relation between the particle energies in the two
frames under the assumptions that K ′ is uniformly rotating relative to K and
that the origins of the two frames coincide. Also, recalling that under these
assumptions we have M′ = M, we can equivalently write h′ = h − w ⋅ M′ ,
thus showing that the particle energy in frame K ′ is less than the energy in
frame K by the amount w ⋅ M′ = w ⋅ M.
Example 2.6
Relative to an inertial frame of reference K with axes x, y, z , a particle
of mass m moves in the horizontal xy-plane under the action of a force
field with potential U(x, y). We wish to write the particle equation of
motion from the point of view of a non-inertial frame K′ that is rotat-
ing with angular velocity w(t) with respect to K. Assuming that the
axis of rotation is directed along the vertical direction z (i.e. w = w k,
where k is the vertical unit vector), let us call q1 , q2 , q3 the co-ordinates
of frame K′ by also assuming that q3 is parallel to z and that the origins
of the two frames coincide. Then, the relation between the particle co-
ordinates in the two systems is
x = q1 cos θ − q2 sin θ , y = q1 sin θ + q2 cos θ (2.103)
where θ = w. Since for observer K the particle Lagrangian is

( )
L = 2−1 m x 2 + y 2 − U , the desired result can be obtained by expressing
it in terms of the K′-co-ordinates q1 , q2. From the co-ordinate transfor-
mation 2.103, it follows
x = q1 cos θ − q1w sin θ − q 2 sin θ − q2 w cos θ

(2.104)
y = q1 sin θ + q1w cos θ + q 2 cos θ − q2 w sin θ
and substitution of Equation 2.104 into L leads to the K′-Lagrangian
m w2 2
L′ =
2
(
m 2
)
q1 + q 22 + mw ( q1q 2 − q2 q1 ) +
2
( )
q1 + q22 − U (2.105)
Then, the calculation of the prescribed derivatives d ( ∂L′ / ∂q i ) dt and

∂L′ ∂qi for i = 1, 2 yields the two equations of motion
 2 − 2mwq 2 − mw 2 q1 + ∂U ∂q1 = 0
m q1 − mwq
(2.106)
 1 + 2mwq1 − mw 2 q2 + ∂U ∂q2 = 0
m q2 + mwq
which, in addition to the expected qi -terms and to the ‘real’ field forces
− ∂U ∂qi , show – as mentioned earlier – the appearance of the fictitious
forces due to the non-inertial state of motion of the frame K′ . Clearly,
the terms with w do not appear if the rotation is uniform. Also, in the
Lagrangian 2.105, it is immediate to identify the three parts T2 , T1 , T0
of the kinetic energy – respectively, quadratic, linear and independent
on the velocities q .
2.7.3 Ignorable co-ordinates and the Routh function

We have seen that Hamilton’s canonical equations for a holonomic sys-
tem are Equations 2.20. Suppose now that some co-ordinates, say the
first k, do not appear in the Hamiltonian function and that we have
H = H ( qk +1 , , qn , p1 , , pn , t ). These missing co-ordinates q1 ,..., qk are
then called ignorable (or cyclic), and owing to Hamilton’s equations (since
p i = − ∂H ∂qi ), we have
p i = 0 ⇒ pi = const ≡ β i ( i = 1,… , k) (2.107)

which tell us that the generalised momenta p1 , , pk are constants (or
integrals) of the motion, with the constants β i being determined from the
initial conditions. But then the functional form of the Hamiltonian is
H = H ( qk +1 , , qn , pk +1 , , pn , β1 , , β k , t ) and involves only the n − k non-
ignorable co-ordinates (for which Hamilton’s equations hold unchanged).
This, in other words, means that k DOFs have been eliminated from the
equations of motion and that we are left with a ‘reduced’ system with n − k
DOFs. As for the behaviour of the ignorable co-ordinates with time, it can
be recovered by integrating the equations q i = ∂ H ∂ β i ( i = 1,… , k).
A similar reduction is possible also in the Lagrangian formulation, but

things are a bit less straightforward. A first observation is that if some
qi do not appear in the Hamiltonian, then they do not appear in the
Lagrangian because, we recall from Section 2.3.1, ∂ L ∂ qi = − ∂ H ∂ qi . So,
if we let q1 , , qk be ignorable co-ordinates, then the generalised momenta
pi = ∂ L ∂ q i are constants of the motion; owing to LEs, in fact, ∂L ∂qi = 0
implies
d  ∂L  ∂L
=0 ⇒ = const ≡ β i (i = 1,… , k) (2.108)
dt  ∂ q i  ∂ q i
with the constants β i being determined from the initial conditions. However,
although the ignorable co-ordinates do not appear in L, all the velocities q i
do, and one would like to eliminate the qs corresponding to the ignorable
qs in order to reduce – as in the Hamiltonian formulation – the number of
DOFs. This can be done by solving the k equations (2.108) for the q i (as
functions of qk +1 ,… , qn , q k +1 ,… , q n , c1 ,… , ck , t ) and then using them in the
so-called Routh function (or Routhian), defined as
k
R = L− ∑ β q (2.109)
i =1
i i
By so doing, the Routhian functional form is R ( qk+1 ,… , qn , q k+1 ,… , q n ,

β1 ,  , βk , t ) , which implies
n n k
∂R ∂R
dR = ∑
i = k+1
∂ q i
dqi + ∑
i = k+1
∂ q i
dq i + ∑ ∂∂βR dβ + ∂∂Rt (2.110)
i =1
i
i
But we can also use the definition on the r.h.s. of Equation 2.109 to write
dR as
 k
 n
∂L
n
∂L ∂L
k k
dL−

∑i =1
∑
β i q i  =
∂
 i =k+1 qi
dqi + ∑ i =1
∂ q i
dq i +
∂t
− ∑
i =1
q i dβ i − ∑ β dq
i =1
i i
(2.111)
and then notice that the first sum on the r.h.s. with the dq i can be split into
two parts as ∑ki =1 () + ∑ ni =k+1 () and that – since β i = ∂L ∂q i – the first part
cancels out with the rightmost sum. At this point, we can compare the vari-
ous differential terms of Equations 2.110 and 2.111 to obtain (besides the
evident relation ∂ R ∂ t = ∂ L ∂ t )
∂ R ∂ qi = ∂ L ∂ qi , ∂ R ∂ q i = ∂ L ∂q i (i = k + 1,… , n)
(2.112)
− ∂ R ∂β i = q i (i = 1,… , k)
where Equations 2.1121 can now be substituted in LEs to give the ‘reduced’
system of n − k equations of motion
d  ∂R  ∂R
−
dt  ∂ q i  ∂ qi
=0 ( i = k + 1,… , n ) (2.113)
with exactly the same form as Lagrange’s equations, but with R in place
of L. As for the ignorable co-ordinates, in most cases, there is no need to
solve for them but, if necessary, they can be obtained by direct integration
of Equations 2.1123.
Example 2.7
In order to illustrate the procedure, a typical example is the so-called
Kepler problem, where here we consider a particle of unit mass
attracted by a gravitational force (i.e. with a r −2 force law) to a fixed
point. Using the usual polar co-ordinates, we obtain the Lagrangian
( )
L = 2−1 r2 + r 2θ 2 + µ r, where µ is a positive (gravitational) constant.
( )
Since L = L r , r, θ , the ignorable co-ordinate is θ and Equation 2.108
for this case is r 2θ = β , from which it follows θ = β r 2 . Using this last
relation to form the Routhian, we get
r2 β2 µ
R = L − β θ = − 2 + (2.114)
2 2r r
which, as expected, is of the form R = R ( r , r, β ) and leads to the 1-DOF

equation of motion  ( ) ( )
r − β 2 r 3 + µ r 2 = 0. Needless to say, this is
the same as the r-equation r − rθ 2 + µ r 2 = 0 that we obtain from the
Lagrangian approach when one considers that r 2θ = β .
A final point worthy of mention is that the existence of ignorable co-
ordinates depends on the choice of co-ordinates; therefore, an inappro-
priate choice may lead to a Lagrangian with no ignorable co-ordinates.
Remark 2.18
i. In the example above, the careful reader has probably noticed that the
Routhian procedure has led to the appearance of an additional poten-
tial energy term (the term proportional to r −2 in the example).
ii. In case of systems with more DOFs, the procedure may also lead to the
appearance of gyroscopic terms (linear in the qs), even if there were no
such terms in the original Lagrangian (see, for example, Greenwood
(1977) or Lanczos (1970)).
2.7.4 The Simple pendulum again: a note

on non-small oscillations
Returning to the paradigmatic example of the simple pendulum, we outline
here how one can proceed when it is not legitimate to make the approximation
sin θ ≈ θ . First, by multiplying both sides of the equation of motion 2.23 by θ,
we get θ θ = −θ ( g l ) sin θ , from which it follows θ dθ = − ( g l ) sin θ dθ . This
can be readily integrated to give θ 2 = ( 2 g l ) cos θ + C . If now we call α be the
maximum swing amplitude, then θ = 0 when θ = α and we obtain the con-
stant of integration C = −(2 g l) cos α . Therefore, θ 2 = ( 2 g l ) ( cos θ − cos α ),
and we get
α S α
dθ 2g dθ 2g
∫
0
cos θ − cos α
=
l ∫
0
dt ⇒
∫0
cos θ − cos α
=S
l

(2.115)
where S = Tα 4 and Tα is the period for swings from −α to α and back.

Now, using the trigonometric relation cos γ = 1 − 2 sin2 (γ 2), we can write
the integral on the l.h.s. of Equation 2.115 as
α
1 dθ

2 ∫ sin (α 2)
0
1 − sin2 (θ 2) sin2 (α 2) 
(2.116)
At this point, defining the new variable x = sin(θ 2) sin(α 2), differentiation
gives
2 sin (α 2) dx 2 sin (α 2) dx 2 sin (α / 2) dx

dθ = = = (2.117)
cos (θ / 2) 1 − sin (θ / 2)
2
1 − x2 sin2 (α / 2)
where the denominator in the last expression is obtained by using the rela-
tion sin2 (θ 2) = x2 sin2 (α 2), which follows from the definition of x. Finally,
substituting Equation 2.117 in 2.116 and using the result in Equation 2.115,
we are led to
1
dx Tα g l

∫ (1 − x )(1 − k x ) =
0
2 2 2 4 l
⇒ Tα = 4
g
K (2.118)
where we defined k = sin(α 2) and K is the so-called elliptic integral (in the
Jacobi form). Elliptic integrals are extensively tabulated, but for our present
purposes, it suffices to consider the expansion
π   1 2  1 3  4  1 3 5 6 
2 2 2
K= 1 +   k +   k +  k + 
2   2   2 4  2 4 6 

π
= 1 + 0.25sin2 (α 2) + 0.1406 sin4 (α 2) + 0.0977 sin6 (α 2) + 
2
(2.119)
which shows that the period (and, clearly, the frequency) of oscillation
depends on the amplitude. This is one of the typical phenomena of non-
linear vibrations.
Finally, note that for small angles (sin(α 2) ≈ α 2 << 1), we can write
l  α2 
Tα = 2π  1 + 16 +  (2.120)
g
where, as expected, the first term corresponds to the familiar formula.

Chapter 3
Finite DOFs systems

Free vibration
3.1 INTRODUCTION
As mentioned at the end of Section 2.6 on small-amplitude oscillations,

the typical equations of motion in which one speaks of free-vibration
are Equations 2.85 and 2.88, where the zero on the r.h.s. means that the
system – undamped in the former case and viscously damped in the latter –
is not subjected to any external force but is set into motion by an initial
disturbance. In this respect, the simplest possible vibrating systems have
one degree of freedom (1-DOF for brevity) and their equations of motion –
undamped and viscously damped, respectively – can be written in the form
 + ku = 0,
mu  + cu + ku = 0 (3.1)
mu
where the system’s physical parameters on the l.h.s., that is, mass, stiffness
and damping characteristics, are not represented by matrices but by simple
scalar quantities.
3.2 FREE VIBRATION OF 1-DOF SYSTEMS
Starting with the undamped case, it is convenient to rewrite Equation 3.11

as
 + ω n2 u = 0,
u ω n2 ≡ k m (3.2)
where the second relation defines the quantity ω n (whose physical meaning
will be clear shortly). Assuming a solution of the form u = eα t , we obtain
the characteristic equation α 2 + ω n2 = 0; then α = ± iω n, and the solution of
Equation 3.21 can be written as u(t) = C1 e iω n t + C2 e − iω n t , where the complex
constants C1 , C2 are determined from the initial conditions and must be
such that C1 = C2∗ because the displacement u(t) is a real quantity. So, if the
initial conditions of displacement and velocity at t = 0 are given by
59
u(0) = u0 , u (0) = v0 (3.3)
some easy calculations lead to the explicit solution
1 v  1 v 
u(t) = u0 − i 0  e iω n t +  u0 + i 0  e − iω n t (3.4)
2  ωn  2 ωn 
which represents an oscillation at the (angular) frequency ω n . For this rea-

son, ω n is called the system’s undamped natural frequency. Clearly, the
solution (3.4) can also be expressed in sinusoidal form as
u(t) = A cos ω n t + B sin ω n t , u(t) = C cos (ω n t − θ ) (3.5)
where the two constants (A, B in the first expression and C, θ in the second)
are determined from the initial conditions (3.3). We leave to the reader the
easy task of checking the relations
A = C1 + C2 = C cos θ = u0 , B = i (C1 − C2 ) = C sin θ = v0 ω n

(3.6)
C = A2 + B2 = u02 + (v0 ω n )2 , tan θ = B A = v0 ω n u0
Remark 3.1
i. The most frequently used example of undamped 1-DOF system is a

mass m attached to a fixed point by means of a spring of stiffness k
and constrained to move in one direction only. If the direction is hori-
zontal, it is assumed that the mass can slide without friction on the
surface that supports it and that the static equilibrium position corre-
sponds to the unstretched spring. If the direction is vertical, the static
equilibrium position does not correspond to the unstretched spring,
but to an elongation δ st = mg k (where g = 9.81 m/s2 is the accelera-
tion of gravity). Since, however, by measuring the displacement from
the static equilibrium position, the weight mg and the equilibrium
spring force kδ st cancel out in Newton’s second law, Equation 3.11
holds unchanged. This, it should be noted, is a general fact and is the
reason why, in order to eliminate gravity forces from the equation of
motion, the static equilibrium position is chosen as the reference posi-
tion. Then, the displacements, deflections, stresses, etc., thus deter-
mined give the dynamic response; the total deflections, stresses, etc.,
are obtained by adding the relevant static quantities to the result of
the dynamic analysis.
ii. Note that for the vertically suspended mass, the undamped natural
frequency can be obtained from the static deflection only. In fact, since
mg = kδ st at equilibrium, Equation 3.22 gives ω n = g δ st . This rela-
tion applies in general and is often used in vibration isolation p
roblems
Finite DOFs systems 61
in order to estimate the fundamental vertical natural frequency of

machines mounted on springs.
iii. In the case of a viscously damped 1-DOF system – which we shall
consider shortly – the typical graphical example found in every book
is a mass attached to a fixed point by means of a spring (of stiffness k)
and a dashpot (with viscous constant c) in parallel.
In terms of energy of vibration, an undamped system in free vibration is

clearly conservative because no energy is fed to the system by an external
excitation and no energy is lost by friction or any other damping mecha-
nisms. Therefore, the system’s total energy ET = Ek + Ep – that is, the sum
of its kinetic and potential energies – is constant and we have dET dt = 0.
More explicitly, since Ek = mu 2 2 and Ep = ku2 2, we can use the solution
in the form (3.5)2 to obtain
kC 2 mω n2C 2 mV 2
ET = = = (3.7)
2 2 2
which shows that the energy is proportional to the amplitude squared (and
where, in writing the last relation, we recalled that the velocity amplitude is
V = ω nC ). Also, observing that the total energy equals the potential energy
at maximum displacement and the kinetic energy at maximum velocity, the
energy equality Ek(max) = Ep(max) leads immediately to the relation ω n2 = k m.
Lastly, we leave to the reader to determine that the average kinetic and
potential energies over a period T = 2π ω n are equal and that their value
is ET 2.
If now we turn our attention to a viscously damped 1-DOF system, it is
convenient to introduce the two quantities ccr , ζ called the critical damping
and the damping ratio, respectively, and defined as
ccr ≡ 2 km = 2mω n , ζ ≡ c ccr = cω n 2k (3.8)
By so doing, we can rewrite the free-vibration Equation 3.12 as
 + 2ω n ζ u + ω 2n u = 0 (3.9)
u
so that, by again assuming a solution of the form u = eα t , the characteristic

equation α 2 + 2ω nζ α + ω n2 = 0 gives the roots
( )
α 1,2 = −ζ ± ζ 2 − 1 ω n (3.10)
which in turn implies that the displacement (general) solution

u(t) = C1 eα1t + C2 eα 2t shows different behaviours depending on whether we
have ζ > 1, ζ = 1 or ζ < 1. These three cases, respectively, correspond to
c > ccr , c = ccr and c < ccr , and one speaks of over-damped, critically damped
and under-damped system – where this last case is the most important in
vibration study because only here the system does actually ‘vibrate’.
Critically damped case: If ζ = 1, then α 1 = α 2 = −ω n and the solution –

taking the initial conditions (3.3) into account – becomes
u(t) = e −ω n t [ u0 + (v0 + ω n u0 ) t ] (3.11)
which is a non-oscillatory function, telling us that the mass simply

returns to its rest position without vibrating.
Over-damped case: A qualitatively similar behaviour to the criti-

cally damped case is obtained if the system is over-damped, that is,
when ζ > 1. Now the two roots are separate and real, and the general
solution takes the form
u(t) = C1 e
(ζ )
− + ζ 2 −1 ω n t
+ C2 e
(ζ )
− − ζ 2 −1 ω n t
(3.12a)
where the constants determined from the initial conditions are
C1 =
(
v0 + ω n u0 ζ + ζ 2 − 1 ), C2 =
(
−v0 − ω n u0 ζ − ζ 2 − 1 ) (3.12b)
2ω n ζ 2 − 1 2ω n ζ 2 − 1
Also, in this case, therefore, the system does not vibrate but returns
to its rest position, although now it takes longer with respect to the
critically damped case. Moreover, it is worth observing that if, for
brevity, we define ω = ω n ζ 2 − 1 , the solution of Equations 3.12a
and b can equivalently be expressed as
 v + ζ ω n u0 
u(t) = e −ζ ω n t  u0 cosh ω t + 0 sinh ω t  (3.12c)
 ω 
Under-damped case: As mentioned above, in this case (ζ < 1), the sys-
tem actually vibrates. In fact, by substituting the two roots (3.10) –
which are now complex conjugates with negative real part – in the
general solution, we get
(
u(t) = e −ζ ω n t C1 e i ω n t 1−ζ 2
+ C2 e − i ω n t 1−ζ 2
) (3.13a)
which represents an oscillation at frequency ω d ≡ ω n 1 − ζ 2 – the

so-called frequency of damped oscillation or damped natural
frequency – with an exponentially decaying amplitude proportional
to the curves ± e −ζω n t that ‘envelope’ the displacement time-history.

From the initial conditions, we obtain the two (complex conjugate,
since u(t) is real) constants
u0 ω d − i ( v0 + ζ u0 ω n ) u0 ω d + i ( v0 + ζ u0 ω n )
C1 = , C2 = (3.13b)
2ω d 2ω d
Other equivalent forms of the solution are
u(t) = e −ζ ω n t ( A cos ω d t + B sin ω d t ) , u(t) = C e −ζ ω n t cos (ω d t − θ ) (3.14)
where the relations among the constants and the initial conditions are
2
 v + ζ u0ω n  B v0 + ζ u0ω n
C = A2 + B2 = u02 +  0  , tan θ = = (3.15)
 ωd A u0ω d
Remark 3.2
i. The property of a critically damped system to return to rest in the

shortest time possible is often used in applications – for example,
in moving parts of electrical and/or measuring instrumentation – in
order to avoid unwanted overshoot and oscillations.
ii. In the under-damped case, the time it takes for the amplitude to
decrease to 1 e of its initial value is called decay (or relaxation) time τ
and is τ = 1 ζω n .
iii. The decaying motion of an under-damped system is not periodic;
strictly speaking, therefore, the term ‘frequency’ is slightly improper.
For most vibrating systems, however, ω d ≅ ω n (i.e. ζ is quite small)
and the motion is very nearly periodic. A better approximation is
( )
ω d ≅ ω n 1 − ζ 2 2 , which is obtained by retaining the first two terms
of the expansion of the square root in ω d = ω n 1 − ζ 2 . Also, note that
the time between two zero crossings in the same direction is con-
stant and gives the damped period Td = 2π ω d . This is also the time
between successive maxima, but the maxima (and minima) are not
exactly halfway between the zeros.
iv. If one wishes to use complex notation, it can be noticed that Equation
3.142 is the real part of Xe − i (ω d − i ζ ω n ) t = Xe − i ω d t , where X = Ce i θ is

the complex amplitude and ω d = ω d − i ζ ω n is the complex damped

frequency, with information on both the damped frequency and the
decay time.
Turning now to some energy considerations, it is evident that a viscously

damped oscillator is non-conservative because its energy dies out with
time. The rate of energy loss can be determined by taking the derivative of
the total energy: by so doing, we get
dET d  ku2 mu 2 

= + = kuu + muu ) = −cu 2 (3.16)
  = u (ku + mu
dt dt  2 2 
where in the last equality, we used the free-vibration Equation 3.12 . Since
the Rayleigh dissipation function is D = cu 2 2 in this case, Equation 3.16
confirms the result of Section 2.4.4, that is, that the rate of energy loss
is −2D.
Remark 3.3
i. So far we have tacitly assumed the conditions k > 0 and c ≥ 0. Since,

however, in some cases it may not be necessarily so, the system’s
motion may be unstable. In fact, if, for instance, c is negative, it is
easy to see that the solution (3.14) is a vibration with exponentially
increasing amplitude. More specifically, if we go back to the charac-
teristic equation and express its two roots (which, we recall, can be
real, purely imaginary or complex; in these last two cases, they must
be complex conjugates because u (t) is real) in the generic complex
form a + ib, the system’s motion can be classified as asymptotically
stable, stable or unstable. It is asymptotically stable if both roots have
negative real parts (i.e. a1 < 0 and a2 < 0), stable if they are purely
imaginary (i.e. a1 = a2 = 0) and unstable if either of the two roots has
a positive real part (i.e. a1 > 0 or a2 > 0, or both). From the discussion
of this section, it is evident that the motion of under-damped, criti-
cally damped and over-damped systems falls in the asymptotically
stable class, while the harmonic motion of an undamped system is
stable.
ii. Just like the decreasing motion of the (asymptotically stable) over-
damped and under-damped cases may decrease with or without
oscillating, the motion of an unstable system may increase with or
without oscillating: in the former case, one speaks of flutter insta-
bility (or simply flutter), while in the latter, the term divergent
instability is used.
Example 3.1
At this point, it could be asked how a negative stiffness or a nega-
tive damping can arise in practice. Two examples are given
Figure 3.1 Inverted pendulum.
in Inman (1994). A first example is the inverted pendulum of

Figure 3.1. This is clearly a 1-DOF system, with the angle θ being a con-
venient choice of generalised co-ordinate. Then, from the Lagrangian
( )
L = ml 2θ 2 2 − mgl cos θ − kl 2 4 sin2 θ , the approximation θ << 1 gives
(
the small-amplitude Lagrangian L = ml 2θ 2 2 − mgl 1 − θ 2 2 − kl 2θ 2 4 . )
At this point, the application of the prescribed derivatives of Lagrange’s
equation leads to
 kl 
ml θ +  − mg  θ = 0 (3.17)
 2 
which is the equation of motion of a 1-DOF undamped system whose

‘effective stiffness’ is given by the term in parenthesis. And since this
term is negative when kl 2 < mg , in this case, the system is unstable
because the weight of the mass acts as a destabilizing force.
The second example is the equation of motion mu  + cu + ku = γ u ,
which is a very simplified 1-DOF model for the vibration of an aircraft
wing and where γ u (γ > 0) approximates the aerodynamic forces on the
wing. Re-arranging the equation, we get mu  + (c − γ )u + ku = 0, and we
have asymptotic stability when c − γ > 0 and a case of flutter instability
when c − γ < 0.
3.2.1 Logarithmic decrement
Since, in general, the damping characteristic of a vibrating system is the
most difficult parameter to estimate satisfactorily, a record of the actual
system’s free oscillation response can be used to obtain such an estimate.
Assuming that our system behaves as an under-damped 1-DOF system, let
t1 , t2 be the times at which we have two successive peaks with amplitudes
u1 , u2 , respectively. Then t2 − t1 = Td = 2π ω d and we can use the complex
form of the solution u(t) given in Remark 3.2-iv to write
u1 Xe −ζ ω n t1 e − iω d t1
= = e i ω d Td eζ ω n Td = e i 2π e2π ζ ω n ωd
= e2π ζ (ω n / ω d ) (3.18)
u2 Xe −ζ ω n t2 e − iω d t2
so that defining the logarithmic decrement as the natural logarithm of the

amplitude ratio, we obtain
u1 ω 2πζ δ
δ ≡ ln = 2π ζ n = ⇒ζ = (3.19)
u2 ωd 1−ζ 2
4π 2 + δ 2
thus implying that if we determine the ratio u1 u2 from the measure-

ment, we can immediately obtain δ and then use it in Equation 3.192 to
get the damping ratio. Even simpler than this, we can expand the expo-
nential in Equation 3.18 and retain only the first two terms, so that
u1 u2 = 1 + 2πζ ω n ω d = 1 + δ ≅ 1 + 2π ζ (where the last relation, i.e. the
approximation δ ≅ 2π ζ or equivalently ω n ω d ≅ 1, is due to the fact that ζ
is in most cases quite small). Then, from u1 u2 ≅ 1 + 2π ζ , we obtain
u1 − u2
ζ ≅ (3.20)
2π u2
which can be readily calculated once u1 , u2 are known.

If the system is very lightly damped, the amplitude of two successive peaks
will be nearly equal, and it will be difficult to perceive any appreciable dif-
ference. In this case, it is better – and more accurate in terms of result – to
consider two peaks that are a few, say m, cycles apart. If the two amplitudes
are ui , ui + m , the same argument as above leads to ln ( ui ui + m ) = 2mπζ ω n ω d
and to the approximate relation
ui − ui + m
ζ ≅ (3.21)
2mπ ui + m
Finally, it may be worth observing that δ can also be expressed in terms

of energy considerations. In fact, from its definition, we get u2 u1 = e −δ ,
and since at the peak amplitudes we have Ep(1) = ku12 2 , Ep(2) = ku22 2 , then
∆Ep Ep(1) = 1 − ( u2 u1 ) = 1 − e −2δ , where ∆Ep = Ep(1) − Ep(2). Expanding the
2
exponential as e −2δ ≅ 1 − 2δ gives the approximate relation
δ ≅ ∆Ep 2Ep(1) (3.22)
Remark 3.4
It is left as an exercise to the reader to determine, as a function of ζ , the

number of cycles required to reduce the amplitude by 50%. As a quick rule
of thumb, it may be useful to remember that for ζ = 0.1 (10% of critical
damping), the reduction occurs in one cycle, while for ζ = 0.05 (5% of criti-
cal damping), it takes about 2 cycles (precisely 2.2 cycles).
3.3 FREE VIBRATION OF MDOF SYSTEMS:

THE UNDAMPED CASE
It was shown in Section 2.6 that the (small-amplitude) equation of motion

for the free vibration of a multiple, say n, degrees of freedom system can be
written in matrix form as
 + Ku = 0,
Mu  + Cu + Ku = 0 (3.23)
Mu
for the undamped and viscously damped case, respectively. Also, it was
shown that M, K, C are n × n symmetric matrices (but in general they are not
diagonal, and the non-diagonal elements provide the coupling between the
n equations) while u is a n × 1 time-dependent displacement vector. Starting
with the undamped case and assuming a solution of the form u = z e iωt in
which all the co-ordinates execute a synchronous motion and where z is a
time-independent ‘shape’ vector, we are led to
( K − ω M ) z = 0 ⇒ Kz = ω
2 2
Mz (3.24)
( )
which has a non-zero solution if and only if det K − ω 2 M = 0. This, in
turn, is an algebraic equation of order n in ω 2 known as the frequency
(or characteristic) equation. Its roots ω 12 , ω 22 , , ω n2 are called the eigen-
values of the undamped free vibration problem, where, physically, the
positive square roots ω 1 , , ω n represent the system’s (undamped) natural
frequencies (with the usual understanding that ω 1 is the lowest value, the
so-called fundamental frequency, and that the subscript increases as the
value of frequency increases). When the natural frequencies ω j ( j = 1, , n )
have been determined, we can go back to Equation 3.241 and solve it
for z for each eigenvalue. This gives a set of vectors z1 , , z n , which are
known by various names: eigenvectors in mathematical terminology,
natural modes of vibration, mode shapes or modal vectors in engineering
terminology. Whatever the name, the homogeneous nature of the math-
ematical problem implies that the amplitude of these vectors can only
be determined to within an arbitrary multiplicative constant – that is, a
scaling factor. In other words, if, for some fixed index k, z k is a solution,
then a z k (a constant) is also a solution, so, in order to completely deter-
mine the eigenvector, we must fix the value of a by some convention. This
process, known as normalisation, can be achieved in various ways and
one possibility (but we will see others shortly) is to enforce the condition
of unit length zTk z k = 1.
Then, assuming the eigenvectors to have been normalised by some
method, the fact that our problem is linear tells us that the general solution
is given by a linear superposition of the oscillations at the various natural
frequencies and can be expressed in sinusoidal forms as
n n
u= ∑ C z cos(ω t − θ ) = ∑ ( D cos ω t + E sin ω t ) z (3.25)

j =1
j j j j
j =1
j j j j j
where the 2n constants – C j , θ j in the first case and Dj , Ej in the second case –
are obtained from the initial conditions (more on this in Section 3.3.2)
u(t = 0) = u0 , u (t = 0) = u 0 (3.26)
Remark 3.5
i. The mathematical form of Equation 3.242 – which is also often writ-

ten as Kz = λ Mz , with λ = ω 2 – is particularly important and is called
generalised eigenvalue problem (or generalised eigenproblem, GEP
for short) – where the term ‘generalised’ is used because it involves two
matrices, while the SEP, standard eigenvalue problem (see Section A.4
of Appendix A), involves only one matrix. The two eigenproblems,
however, are not unrelated because a generalised eigenproblem can
always be transformed to standard form.
ii. On physical grounds, we can infer that the roots of the frequency
equation are real. In fact, if ω had an imaginary part, we would have
an increasing or decreasing exponential factor in the solution (which,
we recall, is of the form u = z e iωt). But since this implies a variation of
the total energy with time, it contradicts the fact that our undamped
system is conservative.
iii. The n roots of the frequency equation are not necessarily all distinct,
and in this case, one must consider their algebraic multiplicities, that
is, the number of times a repeated root occurs as a solution of the fre-
quency equation. We leave this slight complication, sometimes called
eigenvalue degeneracy, for later (Section 3.3.7).
iv. We also leave another complication for later (Section 3.3.8): the fact
that one of the two matrices entering the eigenproblem – typically, the
stiffness matrix K in vibration problems – may not be positive-definite.
v. Just as we do in Appendix A, it is usual to call λ j , z j – that is, an eigen-
value and its associated eigenvector – an eigenpair.
3.3.1 Orthogonality of eigenvectors
and normalisation
Starting with the eigenvalue problem Kz = λ Mz , let z i , z j be two eigen-
vectors corresponding to the eigenvalues λi , λ j , with λi ≠ λ j . Then, the two
relations
Kz i = λi Mz i , Kz j = λ j Mz j (3.27)
are identically satisfied. If we
a. pre-multiply Equation 3.271 by zTj and then transpose both sides, we

get (since both matrices M, K are symmetric) zTi Kz j = λi zTi Mz j, and
b. pre-multiply Equation 3.272 by zTi to give zTi Kz j = λ j zTi Mz j ,
we can now subtract one equation from the other to obtain ( λi − λ j ) zTi Mz j = 0 .
But since we assumed λi ≠ λ j , this implies
zTi M z j = 0 (i ≠ j) (3.28)
which is a generalisation of the usual concept of orthogonality (we recall

that two vectors x, y are orthogonal in the ‘ordinary’ sense if x T y = 0).
Owing to the presence of the mass matrix in Equation 3.28, one often
speaks of mass-orthogonality or M-orthogonality.
The two vectors, moreover, are also stiffness-orthogonal (or
K-orthogonal) because the result zTi Kz j = λi zTi Mz j of point (a) above
together with Equation 3.28 leads immediately to
zTi K z j = 0 (i ≠ j) (3.29)
On the other hand, when i = j , we will have
zTi M z i = Mi , zTi K z i = Ki (i = 1, , n) (3.30)
where the value of the scalars Mi , Ki – called the modal mass and modal
stiffness of the ith mode, respectively – will depend on the normalisation of
the eigenvectors. However, note that no such indetermination occurs in the
ratio of the two quantities; in fact, we have
Ki zT Kz i zT Mz i
= Ti = λi Ti = λi = ω 2i (i = 1, , n) (3.31)
Mi z i Mz i z i Mz i
in evident analogy with the 1-DOF result ω 2 = k m.

As for normalisation – which, as mentioned above, removes the inde-
termination on the length of the eigenvectors – it can be observed that, in
principle, any scaling factor will do as long as consistency is maintained
throughout the analysis. Some conventions, however, are more frequently
used and two of the most common are as follows:
1. assign a unit length to each eigenvector by enforcing the condition

zTi z i = 1( i = 1, , n ) ,
2. scale each eigenvector so that zTi Mz i = 1( i = 1, , n ). This is called
mass normalisation and is often the preferred choice in the engineer-
ing field called modal analysis.
Once the eigenvectors have been normalised, it is often convenient to denote

them by a special symbol in order to distinguish them from their ‘non-normalised
counterparts’ z i; here, we will use the symbol p for mass-normalised eigenvec-
tors and, if needed, the symbol b for unit-length eigenvectors.
In particular, since for mass-normalised eigenvectors we have pTi Mpi = 1,
Equation 3.31 shows that pTi Kpi = λi = ω i2, thus implying that we can write
pTi Mp j = δ ij , pTi Kp j = λi δ ij = ω i2 δ ij (i, j = 1, , n) (3.32)
where δ ij is the well-known Kronecker delta (equal to 1 for i = j and

zero for i ≠ j ) and where the relations for i ≠ j express the mass- and
stiffness-orthogonality properties of the eigenvectors. Also, if we form the
so-called modal matrix by assembling the column vectors pi side by side as
P = [ p1 p2  pn ] – i.e. so that the generic element pij of the matrix is the
ith element of the jth (mass-normalised) eigenvector –Equations 3.32 can
equivalently be expressed as
P T MP = I, P T KP = diag ( λ1 , , λn ) ≡ L (3.33)
where I is the n × n unit matrix and we denoted by L the n × n diagonal

matrix of eigenvalues, which is generally called the spectral matrix.
Remark 3.6
i. Two other normalisation conventions are as follows: (a) set the largest
component of each eigenvector equal to 1 and determine the remain-
ing n − 1 components accordingly, and (b) scale the eigenvectors so
that all the modal masses have the same value M, where M is some
convenient parameter (for example, the total mass of the system).
ii. It is not difficult to see that the relationship between pi and its
non-normalised counterpart is pi = z i Mi . Also, it is easy to see that the
non-normalised versions of Equations 3.33 are Z T MZ = diag ( M1 , , Mn )
and Z T KZ = diag ( K1 , , Kn ), where Z is the modal matrix
Z =  z1 z 2  z n  of the non-normalised eigenvectors.
iii. If we need to calculate the inverse of the modal matrix, note that from
Equation 3.331, it follows P −1 = PT M .
3.3.2 The general solution of the undamped

free-vibration problem
If, for present convenience but without the loss of generality, we use the
mass-normalised vectors p j (instead of z j) in the general solution of Equation
3.25 – let us say, in the second expression – it is readily seen that at time
t = 0, we have u0 = ∑ j Dj p j and u 0 = ∑ j ω j Ej p j . Then, pre-multiplying both
expressions by pTi M and taking Equations 3.321 into account, we get
pTi Mu0 = ∑ D p Mp
j
j
T
i j = Di , pTi Mu 0 = ∑ ω E p Mp
j
j j
T
i j = ω i Ei (3.34)
thus implying that we can write the general solution of the undamped
problem as
n
 
u= ∑  p Mu
i =1
T
i 0 cos ω i t +
pTi Mu 0
ωi
sin ω i t  pi (3.35a)

or, equivalently, as
n
u= ∑ p p M  u
i =1
i
T
i 0 cos ω i t +
u 0
ωi

sin ω i t  (3.35b)

On the other hand, if we wish to express the solution in the first

form of Equation 3.25, it is easy to show that the relations between
the constants Di , Ei of Equations 3.34 and the constants Ci , θ i are
Di = Ci cos θ i , Ei = Ci sin θ i ( i = 1, , n ).
Remark 3.7
i. If we choose to use the vectors z j instead of the mass-normalised pi ,

the cosine coefficient in Equation 3.35a is replaced by zTi Mu0 Mi
while the sine coefficient is replaced by zTi Mu 0 ω i Mi .
ii. It is interesting to note from Equations 3.35 that if the initial condi-
tions are such that u 0 = 0 and u0 ≅ b pr (b constant) – i.e. the initial
displacement is similar to the rth eigenvector – then u ≅ b pr cos ω r t ,
meaning that the system vibrates harmonically at the frequency ω r
with a spatial configuration of motion that resembles the rth mode at
all times. This fact is useful in applications because it shows that any
one particular mode can be excited independently of the others by an
appropriate choice of the initial conditions.
The fact that in Equation 3.35a the vector u is a linear combination of the
mode shapes pi tells us that the vectors pi form a basis of the n-dimensional
(linear) space of the system’s vibration shapes. This fact – which can be
mathematically proved on account of the fact that the matrices K, M are
symmetric and that in most cases at least M is positive-definite (see, for
example, Hildebrand (1992), Laub (2005) and Appendix A) – means that
a generic vector w of the space can be expressed as a linear combination of

the system’s modal vectors; that is, that we have w = ∑ j α j p j for some set of
constants α 1 , , α n . These constants, in turn, are easily determined by pre-
multiplying both sides by pTi M because, owing to Equation 3.321, we get
pTi Mw = ∑ α p Mp = ∑ α δ
j
j
T
i j
j
j ij ⇒ α i = pTi Mw (3.36)
thus implying the two relations
w= ∑ ( p Mw ) p = ∑ ( p p M) w,
j
T
j j
j
j
T
j I= ∑p p
j
j
T
j M (3.37)
where Equation 3.371 is the modal expansion of w, while Equation 3.372 –

which follows directly from the second expression in Equation 3.371 – is the
modal (or spectral) expansion of the unit matrix I (and can also be used as
a calculation check for the matrices on the r.h.s.).
Remark 3.8
Observing that M = M I , we can use Equation 3.372 to obtain the spectral

expansion of the mass matrix as M = ∑ j M p j pTj M. Similarly, K = K I gives
K = ∑ j K p j pTj M, but since Kp j = λ j Mp j, we can write the spectral expan-
sion of the stiffness matrix as K = ∑ j λ j M p j pTj M. On the other hand, by
starting from the expressions M −1 = I M −1 and K −1 = I K −1, respectively (obvi-
ously, if the matrices are non-singular) and using again Equation 3.372 , it
is now left to the reader to show that we arrive at the modal expansions
M −1 = ∑ j p j pTj and K −1 = ∑ j λ j−1 p j pTj .
3.3.3 Normal co-ordinates
Having introduced in Section 3.3.1 the modal matrix, we can now consider
the new set of co-ordinates y related to the original ones by the transforma-
tion u = Py = ∑ i pi yi. Then, Equation 3.231 becomes MPy  + KPy = 0 and
we can pre-multiply it by PT to obtain PT MPy  + PT KPy = 0 , or, owing to
Equations 3.33,
y + Ly = 0
⇒ yi + ω i2 yi = 0 (i = 1,… , n) (3.38)
where the second expression is just the first written in terms of components.
The point of the co-ordinate transformation is now evident: the modal
matrix has uncoupled the equations of motion, which are here expressed
in the form of n independent 1-DOF equations – one for each mode, so

that each yi represents an oscillation at the frequency ω i. Clearly, the initial
conditions for the ‘new’ problem (3.38)1 are
y0 = PT Mu0 , y 0 = PT Mu 0 (3.39)
which are obtained by pre-multiplying the relations u0 = Py0 , u 0 = Py 0 by

PT M and taking Equation 3.331 into account. Given the noteworthy sim-
plification of the problem, the co-ordinates y have a special name and are
called normal (or modal) co-ordinates. Then, because of linearity, the com-
plete solution of the free-vibration problem is given by the sum of these
1-DOF normal oscillations.
Along the same line of reasoning, we can now consider the system’s
kinetic and potential energies and recall from Chapter 2 that for small oscil-
lations, they are T = u T Mu 2 and V = uT Ku 2 , respectively. Then, since
u = Py implies uT = yT PT , we can use these relations in the expressions of
the energies to obtain, by virtue of Equations 3.33,
n n n
ω i2 yi2
T=
y T I y
2
= ∑
i =1
y i2
2
, V=
yT L y
2
= ∑
i =1
2
⇒L=
1
2 ∑ ( y
i
2
i − ω i2 yi2 )
(3.40)
where the system’s Lagrangian is simply a sum of n 1-DOFs Lagrangians.

For this reason, it can be concluded that there is no energy interchange
between any two normal co-ordinates, thus implying that the energy con-
tribution (to the total energy, which is obviously constant because we are
dealing with a conservative system) of each normal co-ordinate remains
constant. Note that this is not true for the original co-ordinates ui because,
in general, the energy expressions in terms of the ui contain cross-product
terms that provide the mechanism for the energy interchange between dif-
ferent co-ordinates.
Example 3.2
In order to illustrate the developments above, consider the simple
2-DOF system of Figure 3.2. Taking the position of static equilibrium
as a reference and calling u1 , u2 the vertical displacements of the two
masses, it is not difficult to obtain the (coupled) equations of motion
M u1 = − k1 u1 + k 2 ( u2 − u1 ) 3mu1 + 5ku1 − ku2 = 0 (3.41)

⇒
mu2 = − k 2 ( u2 − u1 ) mu2 − ku1 + ku2 = 0
Figure 3.2 Simple undamped 2-DOF system.
where, for the sake of the example, in the rightmost expressions, we

have considered the values M = 3m, k 2 = k and k1 = 4 k. So, for our sys-
 + Ku = 0 is
tem, the explicit form of the matrix relation Mu

 3m 0   u1   5k − k   u1   0  (3.42)
 + =
 0
 m   u2   − k k   u2   0 
and the explicit form of the eigenvalue problem Kz = ω 2 Mz is

 5k − k   z1   3m 0   z1  (3.43)
  =ω2 
 −k


k   z2   0 m   z 2 
Then, the condition det (K − ω 2 M) = 0 leads to the characteristic

equation 3m 2ω 4 − 8 kmω 2 + 4 k 2 = 0 , whose roots are ω 12 = 2 k 3m and
ω 22 = 2 k m. Consequently, the two natural frequencies of the system are
ω 1 = 2 k 3m , ω 2 = 2 k m (3.44)
and we can use these values in Equation 3.43 to determine that for
the first eigenvector z1 = [ z11 z 21 ]T we get the amplitude ratio
T
z11 z 21 = 1 3 , while for the second eigenvector z 2 =  z12 z 22  we

get z12 z 22 = −1. This means that in the first mode at frequency ω 1 ,
both masses are, at every instant of time, below or above their equi-
librium position (i.e. they move in phase), with the displacement of the
smaller mass m that is three times the displacement of the mass M. In
the second mode at frequency ω 2 , on the other hand, at every instant
of time, the masses have the same absolute displacement with respect
to their equilibrium position but on opposite sides (i.e. they move in
opposition of phase).
Given the amplitude ratios above, we can now normalise the eigen-
vectors: if we choose mass-normalisation, we get
 1 1 
 
1  1  1  1  12m 2 m
p1 = , p2 = ⇒ P= 
12m  3  2 m  −1   3 1 
 − 
 12m 2 m 
(3.45)
where P is our mass-normalised modal matrix. At this point, it is left

to the reader to check the orthogonality conditions p1T Mp 2 = 0 and
p1T K p 2 = 0 together with the relations p1T K p1 = ω 12 and pT2 K p 2 = ω 22.
As for the normal co-ordinates y, the relation u = Py gives explicitly
1 1 3 1
u1 = y1 + y2 , u2 = y1 − y2 (3.46a)
12m 2 m 12m 2 m
and if we want to express the normal co-ordinates in terms of the

original ones, the most convenient way is to pre-multiply both sides of
u = Py by P T M ; owing to Equation 3.331, this gives y = P T Mu, and the
matrix multiplication yields
3m m m 3 ( u1 + u2 ) 
y1 = (u1 + u2 ) , y2 = (3u1 − u2 ) ⇒ y =  
2 2 2  u1 − u2 
(3.46b)
so that using Equations 3.46b, it can be shown that the energy
(
expressions in the original co-ordinates – that is, T = 3mu12 + mu22 2 )
( 1
2
2 )
and V = 5ku + ku − 2 ku1u2 2 – become T = y + y 2 and
2
( 2
1
2
2 )
( ) ( )
V = k 3m y12 + k m y22 , with no cross-product term in the potential
energy.
Finally, if now we assume that the system is started into motion with
the initial conditions u 0 = [ 1 1]T , u 0 = [ 0 0]T , we can write the
general solution in the form given by of Equations 3.35 to get
 u1  1  1  1 1 
 u  =  3  cos ω 1t +  −1  cos ω 2t (3.47)
 2  2   2 
Example 3.3
For this second example, which we leave for the most part to the reader,
we consider the coupled pendulum of Figure 3.3. The small-amplitude
Lagrangian is
L=
2
(
1  2 2 2
) ( )
ml θ1 + θ 2 − mgl θ12 + θ 22 − kl 2 (θ1 − θ 2 )  (3.48)
2

which leads to the equations of motion mlθ1 + mgθ1 + kl (θ1 − θ 2 ) = 0 and

mlθ2 + mgθ 2 − kl (θ1 − θ 2 ) = 0, or in matrix form
   kl + mg   θ1   0 
 ml 0  θ1 − kl
(3.49)
 0    +  =
 ml   θ2   − kl kl + mg   θ 2   0 

( )
Then, the condition det K − ω 2 M = 0 gives the frequency equation
 g k  g  g 2k  g  g 2k 
ω 4 − 2ω 2  +  +  +  = 0 ⇒  ω 2 −   ω 2 − −  = 0
 l m l  l m   l  l m
(3.50)
from which we obtain the natural frequencies
ω1 = g l , ω 2 = g l + 2 k m (3.51)
which in turn lead to the amplitude ratios z11 z 21 = 1 for the first mode
at frequency ω 1 and z12 z 22 = −1 for the second mode at frequency ω 2 .
In the first mode, therefore, the spring remains unstretched and each
mass separately acts as a simple pendulum of length l ; on the other hand,
in the second mode, the two masses move in opposition of phase. Note
that if the term 2k m in the second frequency is small compared with
g l, the two frequencies are nearly equal and this system provides a nice
example of beats (recall Section 1.3.1). In fact, if one mass is displaced
Figure 3.3 C
oupled pendulum.
a small distance while the other is kept in its equilibrium position and
then both masses are released from rest, the disturbed mass vibrates for
a number of cycles without apparently disturbing the other, but then the
motion of this second mass slowly builds up while that of the first one
slowly dies away (and the pattern repeats on and on).
The mass-normalised eigenvectors are now obtained from the condi-
tion zTi Mzi = 1( i = 1,2 ), and we get
1  1  1  1  1  1 1  (3.52)
p1 = , p2 = ⇒ P=
2ml  1  2ml  −1  2ml  1 −1 
T
while, on the other hand, we get b1 =  1 2 1 2  and
 
T
b2 =  1 2 −1 2  if we prefer to work with unit-length eigen-
 
vectors. At this point, it is now easy to determine that the relation
y = P T Mu gives the normal co-ordinates
ml ml ml  u1 + u2  (3.53)
y1 = (u1 + u2 ) , y2 = (u1 − u2 ) ⇒ y=  
2 2 2  u1 − u2 
One point worthy of mention is that in both examples above, the equa-
tions of motions (expressed in terms of the original co-ordinates) are cou-
pled because the stiffness matrix K is non-diagonal – a case which is often
referred to as static or stiffness coupling. When, on the other hand, the
mass matrix is non-diagonal one speaks of dynamic coupling, and a case
in point in this respect is the double pendulum encountered in Chapter 2.
From the small-amplitude Lagrangian of Equation 2.82, in fact, we can
readily obtain the linearised equations of motion; in matrix form, we have
 ( m1 + m2 ) l12 m2l1l2   θ   ( m1 + m2 ) gl1 0 

   + 
 m2l1l2 m2l22   φ   0 m2 gl2 
 θ   0 
×  = (3.54)

 φ   0 
Then, defining the two quantities M = m1 + m2 and L = l1 + l2, the charac-

teristic equation is m1 l12 l22 ω 4 − ML l1 l2 gω 2 + Ml1 l2 g 2 = 0, and we obtain the
natural frequencies
ω 1,2
2
=
g
2m1l1l2 {
ML ± M 2 L2 − 4Mm1 l1l2 (3.55) }
3.3.4 Eigenvalues and eigenvectors sensitivities

Suppose that the mass and stiffness characteristics of a vibrating system
depend on some variable g that varies (or can be varied) with respect to the
unperturbed situation in which M, K are the (known) system’s matrices.
( )
By also assuming that we already know the eigenpairs λi , pi λi = ω 2j of the
unperturbed system, we can then consider the effect that some small struc-
tural modification/perturbation of its mass and/or stiffness characteristics
may have on its eigenvalues and eigenvectors.
In order to do so, let us start with the relation Kpi = λi Mpi – which we
know to be identically satisfied – and let the variable g undergo a small
change; then, denoting for brevity by ∂ the partial derivative ∂ ∂ g , the
calculation of the derivatives in ∂ ( Kpi ) = ∂ ( λi Mpi ) leads to
{K − λi M} ∂ pi = {M ∂ λi + λi ∂ M − ∂ K} pi (3.56)
which we now pre-multiply on both sides by pTi . By so doing, and by taking
into account that (a) pTi Mpi = 1 and (b) pTi ( K − λi M ) = 0 (which is just the
transpose equation of the eigenproblem rewritten as ( K − λi M ) pi = 0 ), we
obtain the first-order correction for the ith eigenvalue, that is
∂ λi = pTi (∂ K − λi ∂ M ) pi (3.57)
Remark 3.9
A basic assumption in the calculations is that the system’s parameters are

smooth and differentiable functions of g. The variable g, in turn, can be
either a ‘design’ variable – e.g. mass density, Young’s modulus, length, thick-
ness, etc. – that is intentionally (slightly) changed in order to re-analyse a
known system for design purposes or it can be an ‘environmental’ variable
that may vary with time because of aging and deterioration of the system/
structure. In this latter case, the sensitivity analysis may be useful in the
(important and very active) fields of fault detection, predictive maintenance
and structural health monitoring.
Turning to eigenvectors, one way to obtain the first-order correction

∂ pi is to start by expanding it on the basis of the unperturbed eigenvec-
tors pi and write ∂ pi = ∑ r ci r pr. Substituting this into Equation 3.56 and
then pre-multiplying the result by pTk (k ≠ i , so that pTk Mpi = 0) leads to
ci k ( λk − λi ) = pTk ( λi ∂ M − ∂ K ) pi and, consequently,
pTk (∂ K − λi ∂ M ) pi
ci k = (i ≠ k) (3.58)
λ i − λk
At this point, the only missing piece is the c coefficient for i = k; enforcing
the normalisation condition ( pk + ∂ pk ) ( M + ∂ M ) ( pk + ∂ pk ) = 1 for the
T
perturbed eigenmodes and retaining only the first-order terms gives
pTk (∂ M ) pk
ckk = − (3.59)
2
Finally, putting the pieces back together and denoting by λˆ i , p̂i the per-
turbed eigenpair, we have
λˆ i = λi + pTi (∂ K − λi ∂ M ) pi
pTi (∂ M ) pi  pTr (∂ K − λi ∂ M ) pi  (3.60)
pˆ i = pi −
2
pi + ∑ 
r (r ≠ i ) 
λi − λr  pr

where these expressions show that only the ith unperturbed eigenpair is
needed for the calculation of λ̂i , while the complete unperturbed set is
required to obtain p̂i . Also, note that the greater contributions to ∂ pi come
from the closer modes, for which the denominator λi − λr is smaller.
Example 3.4
As a simple example, let us go back to the system of Example 3.2 in
Section 3.3.3. If now we consider the following modifications: (a)
increase the first mass of 0.25m, (b) decrease the second mass of 0.1m
and (c) increase the stiffness of the first spring of 0.1k , the ‘perturbing’
mass and stiffness terms ∂ M, ∂ K are
 0.25m 0   0.1k 0 
∂M =  , ∂K = 
 0 −0.1m   0 0 
Then, recalling that the unperturbed eigenvalues and eigenvectors are,

respectively, λ 1 = 2 k 3m , λ 2 = 2 k m and
1
( ) 1
( )
T T
p1 = 1 3 , p2 = 1 −1
12m 2 m
the first-order calculations of Equation 3.57 give
∂λ1 = p1T (∂ K − λ1 ∂ M ) p1 = 0.0444 ( k m ) ,
∂λ 2 = pT2 (∂ K − λ2 ∂ M ) p 2 = −0.0500 ( k m )
and therefore
λˆ1 = λ1 + ∂λ1 = 0.711( k m ) , λˆ2 = λ2 + ∂λ 2 = 1.950 ( k m ) (3.61)

For the first eigenvector, on the other hand, we obtain the expansion
coefficients (Equations 3.58 and 3.59)
c12 =
(
pT2 ∂ K − λ1 ∂ M p1 ) = 0.0289, c11 = −
p1T (∂ M ) p1
= 0.0271
λ1 − λ2 2
1
( ) (3.62a)
T
pˆ 1 = p1 + c11 p1 + c12 p 2 = 0.311 0.875
m
Similarly, for the second eigenvector, we get
1
( ) (3.62b)
T
pˆ 2 = 0.459 −0.584
m
Owing to the simplicity of this example, these results can be compared

with the exact calculation for the modified system, which can be carried
out with small effort. Since the exact eigenvalues are λ1(exact) = 0.712( k m)
and λ (exact)
2 = 1.968( k m) , we have a relative error of 0.07% on the first
frequency and 0.46% on the second. The exact eigenvectors are
1
( ) 1
( )
T T
p1(exact ) = 0.313 0.871 , p(exact)
2 = 0.458 −0.594
m m
3.3.5 Light damping as a perturbation

of an undamped system
Along a similar line of reasoning as above, let us now examine the case in
which we consider a small amount of viscous damping as a perturbation
of an undamped system. The relevant equation of motion is now Equation
3.232 , and we can assume a solution of the form u = z e λ t to obtain the
so-called quadratic eigenvalue problem (often QEP for short, just like the
acronyms SEP and GEP are frequently used for the standard and gener-
alised eigenvalue problems, respectively)
(λ 2
j M + λj C + K z j = 0 ) ( j = 1, , n ) (3.63)
where we call λ j , z j the eigenpairs of the perturbed – that is, lightly damped –
system. By assuming that they differ only slightly from the undamped eigen-
pairs ω j , p j, we can write the first-order approximations
λ j = iω j + ∂λ j , z j = p j + ∂ p j (3.64)
and substitute them in Equation 3.63. Then, if in the (rather lengthy) resulting
expression, we take into account the undamped relation Kp j − ω 2j Mp j = 0
and neglect second-order terms (including terms containing (∂λ j ) C and
C ∂ p j because we already consider C a first-order perturbation), we are left

with
{K − ω M} ∂ p
2
j j { }
+ iω j C + 2 (∂λ j ) M p j = 0 (3.65)
(
At this point, by pre-multiplying by pTj and observing that pTj K − ω 2j M = 0, )
{ }
we obtain pTj C + 2 (∂λ j ) M p j = 0, which, recalling that pTj M p j = 1, gives
pTj C p j
∂λ j = − (3.66)
2
A first comment on this first-order approximation is that ∂λ j – since C is

generally positive-definite – is a real and negative quantity, in agreement
with what we usually expect from a damped system – that is, the physical
fact that its free vibration dies out with time because of energy dissipa-
tion. A second comment is that the correction involves only the diagonal
elements of the damping matrix Ĉ = PT CP , thereby supporting the general
(and often correct, but not always) idea that for small damping the non-
diagonal terms of Ĉ can be neglected.
Passing to the eigenvectors, we can expand ∂ p j on the basis of
the unperturbed eigenvectors as ∂ p j = ∑ r a j r pr , substitute this
in Equation 3.65 and then pre-multiply the result by pTk to get
( )
a j k ω k2 − ω 2j + iω j pTk Cp j + 2iω j (∂λ j ) pTk Mp j = 0. For k ≠ j , we obtain the
expansion coefficient
pTk Cp j
aj k = − i ω j (3.67)
ω k2 − ω 2j
so that, by Equation 3.642 , we have
 i ω j pTk Cp j 
z j = pj + ∑
k (k≠ j)
 ω 2 − ω 2  pk (3.68)
 j k 
which in turn shows that – unless pTk Cp j = 0 ( k ≠ j ) – the perturbed mode

is now complex, that is, it has a non-zero imaginary part. Physically, this
means that there is a phase relation between any two system’s DOFs that is
not – as is the case for real eigenmodes – simply zero (in phase) or π (oppo-
sition of phase), thus implying that the different DOFs do not reach their
peaks and/or troughs simultaneously.
3.3.6 More orthogonality conditions

The mass- and stiffness-orthogonality relations of Equations 3.32 are
only special cases of a broader class of orthogonality conditions involving
the eigenvectors pi . In fact, let λi , pi be an eigenpair (so that Kpi = λi Mpi
holds); pre-multiplying both sides by pTj KM −1 and taking Equations 3.32
into account, we are led to
pTj KM −1 Kpi = λi pTj KM −1 Mpi = λi pTj K p i = λ 2i δ ij (3.69)
Also, we can pre-multiply both sides of the eigenproblem by pTj KM −1 KM −1

to get, owing to Equation 3.69, pTj KM −1 KM −1 Kp i = λi pTj KM −1 Kp i = λi3 δ ij.
The process can then be repeated to give
( )
a
pTj KM −1 Kpi = λia+1δ ij ( a = 1, 2,) (3.70)
which, by inserting before the parenthesis in the l.h.s. the term MM −1, can
be expressed in the equivalent form
( ) (b = 0,1, 2,) (3.71)

b
pTj M M −1K pi = λ bi δ ji
where the cases b = 0 and b = 1 correspond to the original orthogonality

conditions of Equations 3.32.
A similar procedure can be started by pre-multiplying both sides of the
eigenproblem by pTj MK −1 to give δ ij = λi pTj MK −1 Mpi, which, provided
that λi ≠ 0, implies
( )
−1
pTj MK −1 Mpi = pTj M M −1 K pi = λi−1 δ ij (3.72)
where in the first equality we took the matrix relation

( )
−1 −1 −1
K M= M K into account. Now, pre-multiplying both sides of
the eigenproblem by pTj MK −1 MK −1 and using Equation 3.72 gives
( )
−2
pTj MK −1 MK −1 Mpi = pTj M M −1 K pi = λ i−2 δ ij, so that repeated applica-
tion of this procedure leads to
( ) (b = 0, −1, −2,) (3.73)

b
pTj M M −1K pi = λ bi δ ij
Finally, the orthogonality conditions of Equations 3.71 and 3.73 can be

combined into the single equation
( ) (b = 0, ±1, ±2,) (3.74)

b
pTj M M −1 K pi = λ bi δ ji
3.3.7 Eigenvalue degeneracy
So far, we have assumed that the system’s n eigenvalues are all distinct and have
postponed (recall Remark 3.5 (iii)) the complication of degenerate eigenvalues –
that is, the case in which one or more roots of the frequency equation has an
algebraic multiplicity greater than one, or in more physical terms, when two or
more modes of vibration occur at the same natural frequency. We do it here.
However, observing that in most cases of interest the system’s matrices
are symmetric, the developments of Appendix A (Sections A.4.1 and A.4.2)
show that degenerate eigenvalues are not a complication at all but only a
minor inconvenience. In this respect, in fact, the main result is Proposition
A.7, which tells us that for symmetric matrices we can always find an ortho-
normal set of n eigenvectors. This is because symmetric matrices (which are
special cases of normal matrices) are non-defective, meaning that the alge-
braic multiplicity of an eigenvalue λi always coincides with the dimension of
the eigenspace e ( λi ) associated with λi , i.e. with its geometric multiplicity.
So, if the algebraic multiplicity of λi is, say, m (1 < m ≤ n ), this implies
that we always have the possibility to find m linearly independent vectors in
e ( λi ) and – if they are not already so – make them mutually orthonormal by
means of the Gram–Schmidt procedure described in Section A.2.2. Then,
since the eigenspace e ( λi ) is orthogonal to the eigenspaces e ( λk ) for k ≠ i ,
these resulting m eigenvectors will automatically be orthogonal to the other
eigenvectors associated with λk for all k ≠ i .
Remark 3.10
i. Although the developments of Appendix A refer to a SEP while here

we are dealing with a GEP, the considerations above remain valid
because a symmetric generalised problem can always be transformed
into a symmetric problem in standard form. One way of doing
this, for example, is to exploit a result of matrix analysis known as
Cholesky factorisation and express the mass matrix – which is in
most cases symmetric and positive-definite – as M = S ST , where S is a
non-singular lower-triangular matrix with positive diagonal entries.
Then the eigenvalue problem becomes Kz = λ SST z , and we can pre-
multiply this by S−1 to get S−1 Kz = λ ST z. Finally, defining the new set
of co-ordinates ẑ = ST z (so that z = S−T zˆ ), we arrive at the standard
eigenproblem Azˆ = λ zˆ , where A is the symmetric matrix S−1 KS−T . The
eigenvalues of A are the same as those of the original problem, while
the eigenvectors are related to the original ones by z = S−T zˆ .
ii. The Cholesky factorisation of a symmetric and positive-definite matrix
is often written as L LT . Here, we avoided the symbol L because it can
be confused with the diagonal matrix of eigenvalues of the preceding
(and following) sections.
3.3.8 Unrestrained systems: rigid-body modes

If, for a given system, there is some non-zero vector r such that Kr = 0 then
the stiffness matrix K is singular and the potential energy is a positive-
semidefinite quadratic form. Moreover, substituting this vector into the
eigenvalue problem 3.24 gives λ Mr = 0 , which, observing that M is gener-
ally positive-definite, means that r can be looked upon as an eigenvector
corresponding to the eigenvalue λ = ω 2 = 0. At first sight, an oscillation at
zero frequency may seem strange, but the point is that this solution does
not correspond to an oscillatory motion at all and one speaks of ‘oscil-
latory motion’ because Equation 3.24 was obtained by assuming a time-
dependence of the form e i ωt . If, on the other hand, we assume a more
general solution of the form u = rf (t) and substitute it in Equation 3.231 we
get f(t) = 0, which corresponds to a uniform motion f (t) = at + b, where a, b
are two constants. In practice, the system moves as a whole and no strain
energy is associated with this motion because such rigid displacements do
not produce any elastic restoring forces.
The eigenvectors corresponding to the eigenvalue λ = 0 are called rigid-
body modes (this is the reason for the letter r) and their maximum num-
ber is six because a 3-dimensional body has a maximum of six rigid-body
degrees of freedom: three translational and three rotational. In most cases,
moreover, these modes can be identified by simple inspection.
The presence of rigid-body modes leaves the essence of the arguments
used to obtain the orthogonality conditions practically unchanged because
we have
a. the rigid-body modes are both mass- and stiffness-orthogonal to the

‘truly elastic’ modes since they correspond to different eigenvalues,
b. being a special case of eigenvalue degeneracy, the rigid-body modes
can always be assumed to be mutually mass-orthogonal (for stiffness-
orthogonality, on the other hand Kr = 0 now implies riT Krj = 0 for all
indexes i, j ).
Similarly, the presence of rigid-body modes does not change much also the
arguments leading to the solution of the free-vibration problem. In fact, if
we consider an n-DOF system with m rigid-body modes r1 , , rm, we can
express the transformation to normal co-ordinates as
m n− m
u= ∑
j =1
rj w j (t) + ∑ p y (t) = Rw + Py (3.75)
k=1
k k
instead of the relation u = Py of Section 3.3.3, where here R =  r1  rm 

 
T

is the n × m matrix of rigid-body modes and w = w1  wm  is the
 
m × 1 column matrix (vector) of the normal co-ordinates associated with

these modes. The matrices P, y – now, however, with dimensions n × (n − m)
and (n − m) × 1, respectively – are associated with the elastic modes and retain
their original meaning. Substitution of Equation 3.75 into Mu  + Ku = 0
(since KRw = 0) gives MRw  + MPy  + KPy = 0 , which in turn can be pre-
multiplied (separately) by RT and PT to obtain, in the two cases (owing to
the orthogonality conditions)
 = 0,
w y + Ly = 0 (3.76)

where L is the (n − m) × (n − m) diagonal matrix of non-zero eigenvalues.

Equations 3.76 show that the normal-co-ordinate equations for the
elastic modes remain unchanged while the normal-co-ordinate equations
for rigid-body modes have solutions of the form w j = a j t + bj ( j = 1, , m ).
Then, paralleling the form of Equation 3.25, the general solution can be
written as the eigenvector expansion
m n− m
u= ∑ ( aj t + bj ) rj + ∑ ( Dk cos ω kt + Ek sin ω kt ) pk (3.77)

j =1 k=1
where the 2n constants are determined by the initial conditions

m n− m m n− m
u0 = ∑j =1
bj rj + ∑
k=1
Dk pk , u 0 = ∑ j =1
a j rj + ∑ω E p
k=1
k k k
By using once again the orthogonality conditions, it is then left to the reader
the easy task to show that we arrive at the expression
m n− m
 T 
∑( ∑
pTk Mu 0
u=  )
 rjT Mu 0 t + rjT Mu0  rj +
 p
 k Mu 0 cos ω k t +
ωk
sin ω k t  pk
j =1 k=1  
(3.78a)
or, equivalently
m n− m
∑ ∑ p p M  u
 u 0 
u= rj rjT M ( u0 + u 0 t ) + k
T
k 0 cos ω k t + sin ω k t  (3.78b)
j =1 k=1
ωk 
which are the counterparts of Equations 3.35a and 3.35b for a system with
m rigid-body modes. Using Equation 3.75, it is also left to the reader to
show that the kinetic and potential energies become now
m n− m n− m
T=
1
2 ∑ j =1
2
w +
j
1
2 ∑ y ,
k=1
2
k V=
yT Ly 1
2
=
2 ∑ω
k=1
2
k yk2 (3.79)
where L is the diagonal matrix of Equation 3.762 . These expressions show

that the elastic and the rigid-body motion are uncoupled and that, as it was
to be expected, the rigid-body modes give no contribution to the potential
(strain) energy.
Remark 3.11
i. A possibility to overcome the singularity of the stiffness matrix is

called shifting. The method consists in calculating the shifted matrix
K̂ = K − ρ M (where ρ is any shift that makes K̂ non-singular) and
solving the eigenproblem K̂z = µ Mz . By so doing, the result is that the
original eigenvectors are unaffected by the shifting, while the original
eigenvalues λi are related to the µi by λi = ρ + µi . As an easy illustrative
 5 −5   2 0 
example, consider the eigenproblem  z = λ  z
 −5 5   0 1 
whose eigenvalues are λ1 = 0 and λ2 = 15 2, with corresponding mass-
T
normalised eigenvectors p1 =  1 3 1 3  (the rigid-body
 
T
mode) and p2 =  1 6 −2 6  . Then, a shift of, say, ρ = −3
 
 11 −5   2 0 
gives the shifted eigenproblem  z = µ   z whose
 −5 8   0 1 
characteristic equation 2µ 2 − 27 µ + 63 = 0 has the roots µ1 = 3 and
µ2 = 21 2. It is then left to the reader to check that the eigenvectors are
the same as before.
ii. Another possibility consists in the addition of fictitious elastic
elements along an adequate number of DOFs of the unrestrained
system. This is generally done by adding spring elements, which
prevent rigid-body motions and make the stiffness matrix non-
singular. If these additional springs are very ‘soft’ (i.e. have a very
low stiffness), the system is only slightly modified and its frequen-
cies and mode shapes will be very close to those of the original
(unrestrained) system.
iii. Since a rigid-body mode involves the motion of the system as a whole,
it can be ‘eliminated’ if one considers the relative motion of the sys-
tem’s DOFs. As a simple example, consider an undamped 2-DOFs sys-
tem of two masses m1 , m2 connected by a spring k, with the masses that
can move along the horizontal x-direction on a frictionless surface.
Calling x1 , x2 the displacements of the two masses, it is not difficult to
determine that the equations of absolute motion m1 x 1 + k ( x1 − x2 ) = 0
and m2 x2 + k ( x2 − x1 ) = 0 lead to the two eigen frequencies
k ( m1 + m2 )
ω 1 = 0, ω2 =
m1 m2
If now, on the other hand, we consider the relative displacement z = x1 − x2
of m1 with respect to m2, the equation of motion is m1 m2 z + k ( m1 + m2 ) z = 0,
from which we obtain only the non-zero natural frequency ω 2 above.
3.4 DAMPED SYSTEMS: CLASSICAL AND

NON-CLASSICAL DAMPING
The equation of motion for the free vibration of a viscously damped system
is Mu + Cu + Ku = 0 . As anticipated in Section 3.3.5, assuming a solu-
tion of the form u = z e λ t gives the quadratic eigenvalue problem (QEP for
short; but it is also called complex eigenvalue problem by some authors) of
( )
Equation 3.63, i.e. λ 2 + λ C + K z = 0. If now we determine the solution of
the undamped problem, form the modal matrix P and – just like we did in
Section 3.3.3 – pass to the normal co-ordinates y by means of the transfor-
mation u = Py, we are led to
 + PT CPy + Ly = 0 (3.80)
Iy
which is not a set of n uncoupled equations unless the so-called modal

damping matrix Ĉ = PT CP (which, we note in passing, is symmetric
because so is C) is diagonal. When this is the case, one speaks of classically
damped system, and an important result is that the (real) mode shapes of
the damped system are the same as those of the undamped one. In this
respect, moreover, it was shown in 1965 by Caughey and O’Kelly that a
necessary and sufficient condition for a system to be classically damped is
CM −1 K = KM −1C (3.81)
When Ĉ is diagonal, the n uncoupled equations of motion can be written as
yj + 2ω j ζ j y j + ω 2j y = 0 ( j = 1, 2,… , n ) (3.82)

so that, in analogy with the 1-DOF equation 3.9, ζ j is called the jth modal
damping ratio and is now defined in terms of the diagonal elements cˆ jj of
Ĉ by the relation cˆ jj = 2ω j ζ j, where ω j is the jth undamped natural fre-
quency. The solution of each individual Equation 3.82 was considered in
Section 3.2, and for 0 < ζ j < 1 – which, as already pointed out, is the (under-
damped) case of most interest in vibrations – we are already familiar with
its oscillatory character at the damped frequency ω j 1 − ζ j2 and its expo-
nentially decaying amplitude.
3.4.1 Rayleigh damping
In light of the facts that the decoupling of the equations of motion is a note-
worthy simplification and that the various energy dissipation mechanisms of a
physical system are in most cases poorly known, a frequently adopted model-
ling assumption called proportional or Rayleigh damping consists in express-
ing C as a linear combination of the mass and stiffness matrices, that is
C = a M + b K (3.83a)
where a, b are two scalars. A damping matrix of this form, in fact, leads to
a diagonal modal matrix Ĉ in which the damping ratios are given by
1 a 
2ζ j ω j = pTj ( aM + bK ) p j = a + bω 2j ⇒ ζj =  + bω j  (3.83b)
2  ωj 
where the two scalars a, b can be determined by specifying the damp-

ing ratios for two modes (say, the first and the second). Then, once a, b
have been so determined, the other damping ratios can be obtained from
Equation 3.83b.
Remark 3.12
i. It should be observed that the condition 3.83a of Rayleigh damping is

sufficient but not necessary for a system to be classically damped.
ii. In condition 3.83a, one of the two scalars can be chosen to be zero,
and one sometimes speaks of stiffness- or mass-proportional damp-
ing, respectively, if a = 0 or b = 0. Note that stiffness-proportional
damping assigns higher damping to high-frequency modes, while
mass-proportional damping assigns a higher damping to low-
frequency modes.
iii. With a damping matrix of the Rayleigh form, we have pTk Cp j = 0 for
any two modes with k ≠ j . Then, going back to Equations 3.67, we get
a j k = 0 and, consequently, by Equation 3.68, z j = p j. This confirms the
fact mentioned earlier that the real mode shapes of the damped system
are the same as those of the undamped one.
Example 3.5
By considering the simple 2-DOF system with matrices
 2 0   32 −1 2   3 −1 
M= , C= , K=
 0 1   −1 2 1 2   −1 1 
(which is clearly proportionally damped because C = 0.5 K), the reader

is invited to check that condition 3.81 is satisfied and determine that
T
the two undamped eigenpairs are ω 1 = 1 2 , p1 =  1 6 2 6 
 
T
and ω 2 = 2 , p 2 =  1 3 −1 3  , where p1 , p 2 are the mass-
 
normalised mode shapes.
Then, after having obtained the modal damping matrix
 0 
ˆ = 14
C  
 0 1 
the reader is also invited to show that the damping ratios and damped
frequencies are (to three decimal places) ζ 1 = 0.177, ζ 2 = 0.354 and
ω d 1 = 0.696, ω d 2 = 1.323, respectively. As a final check, one can show
( )
that the calculation of det λ 2 M + λ C + K leads to the characteristic
equation 2λ 4 + 2.5λ 3 + 5.5λ 2 + 2λ + 2 = 0, whose solutions are the two
complex conjugate pairs λ1,2 = −0.125 ± 0.696 i and λ3,4 = −0.500 ± 1.323 i.
Then, recalling from Section 3.2 that these solutions are represented
in the form −ζ jω j ± iω j 1 − ζ 2j (where ω j , for j = 1,2 , are the undamped
natural frequencies), we readily see that the imaginary parts are the
damped frequencies given above, while the decay rates of the real
parts correspond to the damping ratios ζ 1 = 0.125 ω 1 = 0.177 and
ζ 2 = 0.5 ω 2 = 0.354.
A greater degree of flexibility on the damping ratios can be obtained if one

takes into account the additional orthogonality conditions of Section 3.3.6.
More generally, in fact, the damping matrix can be of the form known as
Caughey series; that is
r −1
∑ a ( M K ) (3.84a)
−1 k
C=M k
k=0
where the r coefficients ak can be determined from the damping ratios speci-
fied for any r modes – say, the first r – by solving the r algebraic equations
r −1
2ζ j = ∑a ω
k=0
k
2k−1
j (3.84b)
Then, once the coefficients ak have been determined, the same Equation
3.84b can be used to obtain the damping ratios for modes r + 1, , n . By
so doing, however, attention should be paid to the possibility that some of
these modes do not end up with an unreasonable (for example, negative)

value of damping ratio. A further point worthy of notice is that for r = 2, the
Caughey series becomes the Rayleigh damping of Equation 3.83a.
Finally, a different strategy can be adopted if we specify the damping
ratios for all the n modes and wish to construct a damping matrix C for
further calculations – for example, the direct integration of the equations
of motion instead of a mode superposition analysis. In this case, in fact,
we have Cˆ = diag ( 2ζ 1 ω 1 , ,2ζ nω n ), and from the relation Ĉ = PT CP , we
readily get C = P −T Cˆ P −1. Then, observing that the orthogonality relation
PT MP = I implies P −T = MP and P −1 = PT M , we obtain the desired damp-
ing matrix as
 n 
C = MP Cˆ PT M = M 
 ∑ 2ζ ω
j =1
j j p j pTj  M (3.85)

where the second expression shows clearly that the contribution of each
mode to the damping matrix is proportional to its damping ratio. Obviously,
setting ζ = 0 for some modes means that we consider these modes as
undamped.
3.4.2 Non-classical damping
Although certainly convenient, the assumption of proportional or classical
damping is not always justified and one must also consider the more general
case in which the equations of motion 3.232 cannot be uncoupled (at least
by the ‘standard’ method of passing to normal co-ordinates). Assuming a
solution of the form u = z e λ t , substitution in the equations of motion leads
( )
to the QEP λ 2 M + λ C + K z = 0 , which has non-trivial solutions when
(
det λ 2 M + λ C + K = 0 (3.86) )
holds. This is a characteristic equation of order 2n with real coefficients,
thus implying that the 2n roots are either real or occur in complex conju-
gate pairs. The case of most interest in vibrations is when all the roots are
in complex conjugate pairs, a case in which the corresponding eigenvectors
are also in complex conjugates pairs. In addition, since the free-vibration
of stable systems dies out with time because of inevitable energy loss, these
complex eigenvalues must have a negative real part.
Given these considerations, linearity implies that the general solution to
the free-vibration problem is given by the superposition
2n
u= ∑c z e
j =1
j j
λj t
(3.87)
where the 2n constants c j are determined from the initial conditions. Also,
since u is real, we must have c j +1 = c∗j if, without the loss of generality, we
assign the index ( j + 1) to the complex conjugate of the jth eigenpair.
Alternatively, we can assign the same index j to both eigensolutions,
write the eigenvalue as λ j = µ j + iω j (with the corresponding eigenvector z j)
and put together this eigenpair with its complex conjugate λ j∗ , z ∗j to form a
damped mode s j defined as
s j = c j z j e(
µ j +i ω j ) t
+ c∗j z ∗j e(
µ j − i ω j )t
(3.88a)
{ }
= 2 e µ j t Re c j z j e i ω jt = C j e µ j t Re z j e ( { i ω j t −θ j )
}
where Re{•} denotes the real part of the term within parenthesis, and in
the last relation, we expressed c j in polar form as 2c j = C j e − iθ j . Clearly,
by so doing, Equation 3.87 becomes a superposition of the n damped
modes of vibration s j. Moreover, writing the complex vector z j as
T
z j =  r1 j e − i φ1 j  rnj e − i φnj  , we have
 
 1j ( j j 1j )
 r cos ω t − θ − φ 

s j = Cj eµj t    (3.88b)
 
 rnj cos (ω j t − θ j − φnj ) 
 
The 2n eigensolutions z j e λ j t are called complex modes. On the other hand, the

n damped modes s j – which coincide with the classical natural modes if the
system is undamped or classically damped – are real, physically excitable
and are essentially the real parts of the conjugate complex modes. Equation
3.88b shows clearly that in each damped mode, all system DOFs vibrate at
the same frequency ω j and with the same rate of decay µ j but, in general, with
different phase angles φkj . This confirms what was anticipated at the end of
Section 3.3.5, that is, that the different DOFs do not reach their maximum
excursions or their positions of zero displacement simultaneously.
Remark 3.13
i. For a classically damped mode, we have z j = z ∗j and φ1 j = φ2 j =  = φnj = 0.

If, moreover, the system is undamped, then µ j = 0.
ii. Although in a non-classically damped mode s j not all phase angles
φ1 j ,..., φnj are zero, the phase difference between any two elements/
DOFs of s j is constant. Physically, this means that the order in which
the system’s DOFs pass through their equilibrium positions remains
unchanged and that after one cycle the DOFs return to positions
separated by the same phase angles as in the beginning of the cycle.

Therefore, except for the fact that the motion decays exponentially,
there is an unchanging pattern from cycle to cycle. In this respect, it is
interesting to point out that this characteristic of damped modes can be
exploited to transform a non-classically damped system into one with
classical damping. This is done by introducing suitable phase shifts
into each non-classically damped mode so that all DOFs are either
in-phase or out-of-phase. The method is called phase synchronisation
and is described in the article by Ma, Imam and Morzfeld (2009).
In regard to complex modes, an important difference with the classically

damped modes concerns the orthogonality conditions, which turn out
to be less simple than their ‘classical’ counterparts. Suppose that z j is an
eigenvector of the damped problem with eigenvalue λ j. Then, the QEP
( )
λ j2 M + λ j C + K z j = 0 is identically satisfied and pre-multiplication by
( )
zTk gives zTk λ j2 M + λ j C + K z j = 0 . On the other hand, we can start from
the QEP for the kth eigenpair, transpose it and postmultiply by z j to get
( )
zTk λk2 M + λk C + K z j = 0. If λk ≠ λ j, subtraction of one relation from the
other gives the orthogonality condition
( λk + λ j ) zTk Mz j + zTk Cz j = 0 ( λ j ≠ λk ) (3.89)

The same two relations above lead to another orthogonality condition. In
fact, multiplying the first relation by λk, the second by λ j, subtracting the
two results and observing that λk λ j2 − λk2 λ j = λk λ j ( λ j − λk ) leads to
λk λ j zTk Mz j − zTk Kz j = 0 ( λ j ≠ λk ) (3.90)

Then, if we pursue the analogy with the SDOF case and represent a com-
plex eigenvalue in the form λ j = −ζ j ω j + iω j 1 − ζ j2 , we can set λk = λ j∗ in
the two orthogonality conditions to get
z Hj Cz j z Hj Kz j
2ζ jω j = , ω 2j = (3.91)
z Hj Mz j z Hj Mz j
( )
T
where, following a common notation (also used in Appendix A), z Hj = z ∗j .
3.5 GEPs AND QEPs: REDUCTION

TO STANDARD FORM
The preceding sections have shown that for both conservative (undamped)
and non-conservative (damped) systems, we must solve an eigenproblem: a
GEP in the first case and a QEP in the second case. Since, however, there
exists a large number of computationally effective and efficient algorithms

for the solution of the SEP, a widely adopted strategy is to organise the
equations of motions so that the eigenvalue problem can be expressed in the
standard form Av = λ v discussed in Appendix A, where the matrix A and
the vector v depend on the particular transformation procedure.
In the following, we merely outline a few possibilities, and for a more
detailed account of these aspects – including the advantages and drawbacks
of the various transformation methods – we refer the interested reader to
the wide body of specialised literature.
3.5.1 Undamped Systems
Let us consider the conservative case first. Provided that M is non-singular,
both sides of the GEP Kz = λ Mz can be pre-multiplied by M −1 to give
M −1 Kz = λ z , which is an SEP for the matrix A = M −1 K (often referred to as
the dynamic matrix). Similarly, if K is non-singular, we can pre-multiply the
generalised problem by K −1 to get Az = γ z, where now the dynamic matrix
is A = K −1 M and γ = 1 λ . The main drawback of these methods is that the
dynamic matrix, in general, is not symmetric.
In order to obtain a more convenient symmetric problem, we recall that
one possibility (using the Cholesky factorisation) was considered in Remark
3.10 of Section 3.3.7. A second possibility consists in solving the SEP for the
(symmetric and positive-definite) matrix M and consider its spectral decom-
position M = RD2 RT , where R is orthogonal (i.e. RRT = I ) and D2 is the
diagonal matrix of the positive – this is why we write D2 – eigenvalues of M.
Then, substitution of the spectral decomposition into the original GEP gives
Kz = λ RD2 RT z and consequently, since DRT = (RD)T , Kz = λ NNT z where
N = RD. If now we pre-multiply both sides of the eigenproblem by N −1, insert
the identity matrix N −T NT between K and z on the l.h.s. and define the vec-
tor x = NT z, we obtain an SEP for the symmetric matrix S = N −1 KN −T .
A different strategy consists in converting the set of n second-order ordi-
nary differential equations into an equivalent set of 2n first-order ordinary
differential equations by introducing velocities as an auxiliary set of vari-
ables. This is called a state-space formulation of the equations of motions
because, in mathematical terms, the set of 2n variables u1 ,… , un , u1 ,… , u n
defines the so-called state space of the system. In essence, this approach
is similar to what we did in Section 2.3.1, where we discussed Hamilton’s
canonical equations. In that case, we recall, the generalised momenta p j
played the role of auxiliary variables (and the set of 2n variables q j , p j defines
the so-called phase space of the system. By contrast, the n-dimensional space
defined by the variables u1 , , un is known as configuration space).
Starting from the set of n second-order ordinary differential equa-
tions Mu  + Ku = 0 , pre-multiplication by M −1 leads to u  = − M −1 Ku.
T
Now, introducing the 2n × 1 vector x =  u u  (which clearly implies
T
x =  u   ), we can put together the equation u
u  = − M −1 Ku with the
 
trivial identity u = u in the single matrix equation
 0 I 
x = Ax , A= −1
 (3.92)
 − M K 0 

where A, defined in 3.922 , is a 2n × 2n matrix. At this point, assuming a

solution of Equation 3.92 in the form x = v e λ t leads to the SEP Av = λ v
of order 2n. The 2n eigenvalues of this problem are related to the system’s
natural frequencies by λ1,2 = ± i ω 1 ,  , λ2n−1, 2n = ± i ω n, and only the posi-
tive values of frequency have physical meaning. Using these values, we can
then obtain the associated eigenvectors; these, in turn, are 2n-dimensional
T
vectors and have the form v j =  z j λ j z j  , where z j are the eigenvectors
of the original problem Kz = λ Mz.
3.5.2 Viscously damped systems

A state-space formulation of the eigenproblem is more common for the case
of a viscously damped n-DOF system when the transformation to normal
co-ordinates fails to uncouple the equations of motion. One way of doing
this is to put together the equation of motion Mu  + Cu + Ku = 0 and the
trivial identity Mu − Mu = 0 in the single matrix equation
 C M   u   K 0  u   0 
  +  =  (3.93a)
 M 0   u
   0 − M   u   0 
T
which, defining the 2n × 1 state vector x =  u u  and the 2n × 2n real
 
symmetric (but not, in general, positive-definite) matrices,
ˆ =  C M   K 0 
M , Kˆ =   (3.93b)
 M 0   0 − M 
can be expressed as Mxˆ  + Kx

ˆ = 0. Then, assuming a solution of the form
λt
x = v e leads to the symmetric GEP
ˆ + λ Mv
Kv ˆ = 0 (3.94)
whose characteristic equation and eigenvalues are the same as the ones of
the original QEP. The 2n-dimensional eigenvectors of Equation 3.94 have
T
the form v j =  z j λ j z j  , where z j are the n-dimensional eigenvectors of
the quadratic eigenproblem. Besides the increased computational effort, the

solution of the GEP 3.94 develops along the same line of reasoning of the
undamped case shown in the preceding section. As compared with the QEP,
however, the advantage of the present formulation is that the eigenvectors
v j are now orthogonal with respect to the matrices M̂ and K̂, and we have
the relations
ˆ j =M
ˆ j δ ij , Kˆ j vTj Kˆ v j
vTi Mv vTi Kˆ v j = Kˆ j δ ij , λj = − =− T (3.95)
Mˆj vj M ˆ vj
where the constants M ˆ j , Kˆ j depend on the normalisation. In this respect,

note that the r.h.s. of Equations 3.95 are zero even when the two eigenvec-
tors on the l.h.s. are a complex conjugate pair; this is because, as a matter
of fact, they correspond to different eigenvalues.
Remark 3.14
i. The fact that the symmetric GEP 3.94 leads, in the under-damped case,
to complex eigenpairs may seem to contradict the ‘educated guess’
(based on the developments of preceding sections and of Appendix A)
that a symmetric eigenvalue problem should produce real eigenvalues.
However, there is no contradiction because, as pointed out above, the
matrices Mˆ , Kˆ are not, in general, positive-definite.
ii. Note that some authors denote the matrices M ˆ , Kˆ by the symbols A, B,
respectively.
Also, the GEP 3.94 can be converted into the standard form by pre-multiplying
both sides by Mˆ −1 or Kˆ −1 (when they exist), in analogy with the procedure con-
sidered at the beginning of the preceding section. In this case, it is not difficult
to show that, for example, the matrix M ˆ −1 can be obtained in terms of the
original mass and damping matrices as
 M −1 
ˆ −1 =  0
M 
 M −1 − M −1 CM −1 
 
The last possibility we consider here parallels the method described at the
end of the preceding Section 3.5.1 for the undamped case. Starting from
Mu + Cu + Ku = 0 , pre-multiplication by M −1 gives u
 = − M −1 K u − M −1C u.
This equation together with the trivial identity u = u can be combined into
T
the single matrix equation x = Ax , where x =  u u  and A is now the
2n × 2n matrix
 0 I 
A= −1 −1
 (3.96)
 − M K −M C 

Then, assuming a solution of the form x = v e λ t leads to the (non-symmetric)

SEP Av = λ v of order 2n.
3.6 EIGENVALUES SENSITIVITY OF
VISCOUSLY DAMPED SYSTEMS
Proceeding along the lines of Sections 3.3.4 and 3.3.5, we consider now
the sensitivity of the eigenvalues of a non-proportionally damped system.
Assuming that the system does not possess repeated eigenvalues, our start-
ing point here is the fact that the eigenpairs of a damped system satisfy
( )
the QEP λ j2 M + λ j C + K z j = 0, which, defining for present convenience
the matrix Fj = λ M + λ j C + K , can be rewritten as Fj z j = 0. Then, pre-
2
j
multiplication by zTj gives zTj Fj z j = 0 and, consequently, by differentiating

this last relation,
(∂ z )F z
T
j j j + zTj (∂ Fj ) z j + zTj Fj (∂ z j ) = 0 (3.97)
Now, observing that Fj z j = 0 (and that this, by transposing, implies zTj Fj = 0

because Fj is symmetric), it turns out that the first and last terms on the
l.h.s. of 3.97 are zero and that we are left with zTj (∂ Fj ) z j = 0 . Taking into
account the explicit form of Fj, this equation reads
zTj ∂λ j ( 2λ j M + C ) + λ j2 ∂ M + λ j ∂ C + ∂ K  z j = 0

⇒ ( )
∂λ j zTj ( 2λ j M + C ) z j = − zTj λ j2 ∂ M + λ j ∂ C + ∂ K z j
from which we get the first-order perturbation ∂λ j of the complex

eigenvalue λ j as
∂λ j = −
(
zTj λ j2 ∂ M + λ j ∂ C + ∂ K z j ) (3.98)
z T
j ( 2λ j M + C ) z j
This is, as a matter of fact, a generalisation to the damped case of the
‘undamped equation’ 3.57 because it is not difficult to show that Equation
3.98 reduces to Equation 3.57 if the system is undamped. When this is the
case, in fact, C = 0, the eigenvalues are λ j = iω j and the complex eigenvec-
tors z j (with the appropriate normalisation, see Remark 3.15 below) become
the (real) mass-normalised eigenvectors p j. Consequently, Equation 3.98

becomes
( )
pTj −ω 2j ∂ M + ∂ K p j
i (∂ω j ) = −
2iω j
⇒ ( ) ( )
∂ ω 2j = pTj ∂ K − ω 2j ∂ M p j
which is exactly the undamped equation when one recalls that in Equation
3.57, we have λ j = ω 2j .
The sensitivity of the eigenvectors of a damped system is definitely more
involved, and for a detailed account, we refer the interested reader to
Chapter 1 of Adhikari (2014b).
Remark 3.15
Going back to Equation 3.89, for j = k, the complex eigenvectors can be

normalised according to 2λ j zTj Mz j + zTj Cz j = 2i ω j , which reduces to the
familiar mass-normalisation pTj Mp j = 1 if the system is undamped.
Chapter 4
Finite-DOFs systems
Response to external excitation
4.1 INTRODUCTION
When subjected to an external excitation – typically, but not necessar-

ily, a t ime-dependent force f (t) – the equation of motion of a viscously
damped 1-DOF system is the scalar version of Equation 2.91, i.e.
mu  + cu + ku = f . Then, mathematics teaches us that the solution u(t) of
this ordinary differential equation is the sum of two parts, that is
u(t) = uc (t) + up (t) (4.1)
where the complementary function uc (t) is the solution of the homogeneous

equation (i.e. the equation of motion with zero on the r.h.s.) and up (t) is
called the particular integral or, in engineering terminology, the steady-
state solution. The first part has been considered in Section 3.2, and we
saw that (a) it involves two arbitrary constants in order to satisfy the initial
conditions, and (b) it is transient in nature because damping makes it decay
to zero in a relatively short interval of time (this is why it is often called
transient solution in engineering terminology).
On the other hand, the steady-state solution up (t) persists as long as
the external excitation does, does not involve any arbitrary constants
and, by itself, does not satisfy the initial conditions. But since this is the
part that remains after the transients have disappeared, in most cases
of external excitations of relatively long duration – say, longer than a
few times the system’s natural period – the complementary solution is
often ignored and emphasis is placed only on up (t). As for terminology,
the condition in which a system vibrates under the action of a non-zero
r.h.s.-term in the equations of motion is commonly referred to as forced
vibration.
99
4.2 RESPONSE IN THE TIME-, FREQUENCY-

AND s-DOMAINS: IRF, DUHAMEL’S
INTEGRAL, FRF AND TF
Suppose now that our 1-DOF system, initially at rest, is subjected at t = 0 to

an impulse in the form of a relatively large force that lasts for a very short
time – for example, a hammer blow. Then, since a convenient mathematical
representation of this type of force is f (t) = fˆ δ (t), where fˆ is the magnitude
of the impulse and δ (t) is the Dirac delta ‘function’ of Appendix B (Section
B3), the impulse-momentum relation of basic physics tells us that in the
very short time of the ‘hammer blow’, we have a sudden change of the
system’s velocity from zero to fˆ m without an appreciable change of its dis-
placement. Physically, this corresponds to having a force-free system with
( ) ( )
the initial conditions u 0+ = 0 and u 0+ = fˆ m, which can now be substi-
tuted in Equations 3.5 (undamped system) or 3.14 (underdamped system)
(
to give, respectively, the displacement solutions u(t) = fˆ mω n sin ω n t and )
( )
u(t) = fˆ mω d e −ζ ω n t sin ω d t .
Despite its apparent simplicity, this is an important result for linear sys-
tems, and the response to a unit impulse is given a special name and a spe-
cial symbol; the name is impulse response function (IRF) and the symbol is
h(t). Consequently, it follows that
1 e −ζ ω n t
h(t) = sin ω n t , h(t) = sin ω d t (4.2a)
mω n mω d
are the IRFs of an undamped and a viscously damped system, respectively.

Clearly, both functions are zero for t < 0 because of the causal requirement
that the response must occur after the input (applied at t = 0 in this case). If
the impulse is applied at some time τ other than zero, then the appropriate
IRFs are
1 e −ζ ω n (t −τ )
h(t − τ ) = sin ω n (t − τ ), h(t − τ ) = sin ω d (t − τ ) (4.2b)
mω n mω d
which now are zero for t < τ .

Using these definitions, the system’s response to the impulsive force at
t = 0 can be written as u(t) = fˆ h(t). Proceeding along this line of reasoning,
it then makes sense to consider a generic loading with time-history f (t) as
a series of δ -impulses of the form f (τ ) dτ δ (t − τ ), where f (τ ) dτ is the mag-
nitude of the impulse and τ varies along the time axis. But then, since the
system’s response to f (τ ) dτ δ (t − τ ) is du = f (τ ) h(t − τ ) dτ , linearity implies
that the response u(t) at time t is the superposition of the responses to the
Response to external excitation 101
impulses that have occurred between the onset of the input at t = 0 and
time t. Therefore, we can write
t
u(t) =
∫ f (τ ) h(t − τ ) dτ (4.3)
0
which, recalling Equation B.29 of Appendix B, we recognise as the con-

volution ( f ∗ h) (t) of the forcing function f (t) with the system’s IRF (the
different integration limits with respect to definition B.29 are considered
in Remark 4.1(ii) below). As for terminology, it is common to refer to the
expression 4.3 as Duhamel’s integral.
Equation 4.3, however, holds when the system is at rest at time t = 0. If it
is not so, the non-zero initial conditions must be taken into account, so that,
for example, for a damped 1-DOF system with initial conditions u(0) = u0
and u (0) = v0 , we have the response
 v + ζ ω n u0 
u(t) = e −ζ ω n t  u0 cos ω d t + 0 sin ω d t 
 ωd 
t
1
+
mω d ∫ f (τ ) e
0
−ζ ω n (t −τ )
sin [ω d (t − τ )] dτ (4.4)
Clearly, setting ζ = 0 (so that ω d = ω n) gives the response of an undamped

system as
t
v 1
u(t) = u0 cos ω n t + 0 sin ω n t +
ωn mω n ∫ f (τ )sin[ω (t − τ )]dτ (4.5)
0
n
which, with the appropriate modifications, is Equation B.59 obtained in

Appendix B in the context of Laplace transforms.
Remark 4.1
( )
i. Substitution of the initial conditions u 0+ = 0 and u 0+ = fˆ m in ( )
Equations 3.11 and 3.12c gives the (non-oscillatory) IRFs of a criti-
cally damped and over-damped system, respectively. However, as
already mentioned before, the undamped and under-damped cases
considered earlier are the ones of most interest in vibrations.
ii. Considering the integration limits of Equation 4.3, it can be observed
that the lower limit is zero because we assumed the input to start at
t = 0. But since this is not necessarily the case, we can just as well
extend this limit to −∞. For the upper limit, on the other hand, we
know that h(t − τ ) is zero for t < τ . But this is the same as τ > t , and we
can therefore extend the limit to +∞ without affecting the result. The
conclusion is that the Duhamel integral of Equation 4.3 is, as a matter
of fact, a convolution product in agreement with definition B.29.
iii. If now, with the lower and upper limits at infinity, we make the change
of variable α = t − τ ; it is almost immediate to obtain
∞
u(t) =
∫ f (t − α ) h(α ) dα (4.6)
−∞
thus showing that it does not matter which one of the two functions,
f (t) or h(t), is shifted (note that, with a different notation, this is prop-
erty (a) of Remark B.5(ii) in Appendix B). In calculations, therefore,
convenience suggests the most appropriate shifting choice.
iv. A system is stable if a bounded input produces a bounded output. By a
∞
well-known property of integrals, we have u(t) ≤
∫ −∞
f ( t − α ) h(α ) dα ,
but since a bounded input means that there exists a finite constant K
such that f (t) ≤ K , then it follows that a system is stable whenever its
∞
IRF is absolutely integrable, that is, whenever
∫ −∞
h(α ) dα < ∞.
The preceding considerations show that, in essence, the IRF of a linear system
characterises its input–output relationships in the time domain and that for this
reason it is an intrinsic property of the system. Consequently, if we Fourier and
Laplace transform the functions f (t), h(t), call F(ω ), F(s) the two transforms of
f (t) and observe that Equations B.30 and B.52 of Appendix B show that
F ( f ∗ h) (t)  = 2π F(ω ) F [ h(t)], L ( f ∗ h) (t)  = F(s)L [ h(t)] (4.7)
it is now eminently reasonable to expect that H (ω ) ≡ 2π F [ h(t)] and

H (s) ≡ L [ h(t)] should also be intrinsic properties of the system that spec-
ify its input–output characteristics in the frequency- and Laplace-domain,
respectively. This is indeed the case, and the two functions are called fre-
quency response function (FRF) and transfer function (TF). Taking the
Fourier and Laplace transforms of the damped IRF of Equation 4.2a2 , these
functions are
1 1
H (ω ) = , H (s) = (4.8)
(
k 1 − β + 2i ζ β
2
) (
m s + 2 ζ ω n s + ω n2
2
)
where, in the first relation, we introduced the frequency ratio β = ω ω n . So,
Equations 4.7 and 4.8 show that the system’s response in the frequency- and
in the s-domain – that is, the functions U(ω ) = F [ u(t)] and U(s) = L [ u(t)],
respectively – is given by
U(ω ) = H (ω ) F(ω ), U(s) = H (s)F(s) (4.9)
Remark 4.2
i. While the TF H (s) is the Laplace transform of h(t), H (ω ) is not exactly

the Fourier transform of the IRF but 2π times this transform. This
minor inconvenience is due to our definition of Fourier transform
(Equation B.16 in Appendix B; see also Remarks B.4(i) and B.5(i))
and has no effect whatsoever as long as consistency is maintained
throughout. In this respect, however, attention should be paid to the
definition used by the author when consulting a table of Fourier trans-
forms. Also, note – and this is a general fact that applies not only to
1-DOF systems – that H (ω ) can be obtained from H (s) by making the
substitution s = iω .
ii. If the reader wishes to do so, the transforms of Equations 4.8 can
be calculated with the aid of the tabulated integral of Equation 4.11
below.
iii. Not surprisingly, Equations 4.9 can be directly obtained by Fourier
or Laplace transforming the equations of motion. Consider, for exam-
ple, Equation 4.92; for zero initial conditions, the Laplace transform
of mu ( )
 + cu + ku = f gives s2m + sc + k U(s) = F(s), from which it fol-
( )
2 −1
lows U(s) = s m + sc + k F(s) and, consequently, observing that
c m = 2ζ ω n and k m = ω n2 ,
F(s)
U(s) = = H (s)F(s) (4.10a)
m(s2 + 2ζ ω n s + ω n2 )
Clearly, a similar procedure with the Fourier transform leads to

Equation 4.91.
iv. Equation 4.10a holds if the system is initially at rest. But since it is
shown in Appendix B (Equations B50a) that the Laplace transform
‘automatically’ takes into account the initial conditions at t = 0, it is
now left to the reader to determine that with the non-zero initial con-
ditions u(0) = u0 and u (0) = v0 , we have
F(s) m u0 ( s + 2ζ ω n ) v0
U(s) = 2 + 2 2 + 2 (4.10b)
s + 2ζ ω n s + ω n s + 2ζω n s + ω n s + 2ζω n s + ω n2
2

whose inverse Laplace transform gives, as it should be expected,
Equation 4.4.
Example 4.1: The indicial response

In addition to the IRF, another useful function in applications is the
response to the Heaviside unit step function θ (t) defined in Equation B.40
of Appendix B. This response is sometimes called indicial response and
here we denote it by the symbol r(t). Assuming the system to start from
t
rest, Equation 4.3 gives r(t) = ( mω d )
∫ e −ζ ω n (t − τ ) sin ω d ( t − τ )  dτ .
−1
0
Then, making use of the tabulated integral
e at
∫ e at sin ( b t ) dt =
a + b2
2 ( a sin bt − b cos bt ) (4.11)
some easy manipulations lead to the explicit expression
1 e −ζ ω n t  ζ ω n 
r(t) = − sin ω d t + cos ω d t  (t > 0) (4.12)
k k  ω d 
which reduces to r(t) = k−1 (1 − cos ω n t ) if the system is undamped and

shows that the maximum response is twice the static deflection (1 k for
a force of unit amplitude). With some non-zero damping, the maxi-
mum deflection is smaller, but for lightly damped systems, an amplifi-
cation factor of two provides in any case a conservative estimate of the
severity of the response.
Example 4.2: Rectangular pulse

If the excitation to an undamped system is a rectangular pulse of ampli-
tude f0 that starts at t = 0 and ends at t = t1, the system’s response can
be obtained by first considering the ‘forced-vibration era’ 0 < t < t1 (also
called primary region) and then the ‘free-vibration era’ t > t1 (also called
residual region). Since, from the preceding example, we know that the
response in the ‘forced-vibration era’ is u1 (t) = f0 k−1 (1 − cos ω n t ) , the
response u2 (t) in the ‘free-vibration era’ will be given by the free-vibra-
tion solution of Equation 3.51 with initial conditions determined by the
state of the system at t = t1, i.e. u1 ( t1 ) and u1 ( t1 ) . This gives
f  f 
u2 (t) =  0 (1 − cos ω n t1 )  cos ω n ( t − t1 ) +  0 sin ω n t1  sin ω n ( t − t1 )
 k   k 

f0
=  cos ω n ( t − t1 ) − cos ω n t  (4.13)
k
where, for clarity, in the first expression we put the initial conditions
within curly brackets (the second expression then follows from a well-
known trigonometric relation).
It is instructive to note that this same problem can be tackled by
using Laplace transforms. In fact, by first observing that the rectangu-
lar pulse of duration t1 (of unit amplitude for simplicity) can be written
in terms of the Heaviside function as f (t) = θ (t) − θ ( t − t1 ), it follows

( )
that its Laplace transform is F(s) = s −1 1 − e − st1 . Then, the system’s
response in the s-domain is
1m e − s t1 m
U(s) = −
(
s s + ωn
2 2
) (
s s2 + ω n2 )
whose inverse Laplace transform L−1 [U(s)] leads to the same
result as earlier, i.e. u1(t) = k−1 (1 − cos ω nt ) for 0 < t < t1 and
u2 (t)u = k−1  cos ω n ( t − t1 ) − cos ω n t  for t > t1 (by using a table of Laplace
transforms, the reader is invited to check this result).
4.2.1 Excitation due to base displacement,

velocity or acceleration
In many cases of practical interest, the excitation is not given in terms of a
force applied to the mass but in terms of the motion of the base that supports
the system. This motion, in turn, may be known – or, often, m easured – in
terms of displacement, velocity or acceleration. Typical examples, just to
name a few, are earthquakes loadings (if the response is assumed to remain
within the linear range), a vehicle suspension system or the vibration of
a piece of equipment due to the motion – caused by nearby heavy traffic,
trains, subways, etc. – of the building in which it is housed.
In this respect – and with the implicit understanding that damping in
most cases limits motion – the undamped 1-DOF system is often considered
as a conservative ‘standard reference’ in order to approximately determine
the response to be expected under these types of loadings.
Our starting point here is the fact that for an undamped 1-DOF, the fore-
going considerations have shown that the response to an excitation due to a
time-dependent force f (t) starting at t = 0 is
t
1  f (τ ) 
u(t) =
ωn ∫ 
0
m 
sin ω n (t − τ ) dτ (4.14)
where, for present convenience, we put into evidence the term f m because
this is the r.h.s. of the equation of motion when it is rewritten in the form
 + ω n2 x = f m (4.15)
u
So, if the excitation is a displacement x(t) of the base relative to a

fixed frame of reference and u(t) is the displacement of the system’s
mass relative to the same frame of reference, the equation of motion
is mu + k ( u − x ) = 0. Rewriting it as u
 + ω n2 u = ω n2 x, we have a differential
equation of the form 4.15 with ω n2 x(t) on the r.h.s. in place of f m. By

Equation 4.14, therefore, the system’s response is
t
u(t) = ω n
∫ x(τ ) sinω (t − τ ) dτ (4.16)
0
n
If, on the other hand, the excitation is in the form of base velocity x(t), we
 + ω 2n u = ω 2n x to get
can differentiate the equation u
d 2u
+ ω 2n u = ω n2 x (4.17)
dt 2
which is a differential equation in u formally similar to Equation 4.15.
Then, paralleling the result of Equation 4.14, we get the velocity response
t
u (t) = ω n
∫ x(τ )sinω (t − τ ) dτ (4.18)
0
n
Finally, if the base motion is given in terms of acceleration x (t), we can con-
sider the relative coordinate z = u − x of the mass with respect to the base.
Since z = u − x
 (and, consequently, u  = z + x
), we can write the equation of
motion u  + ω 2n (u − x) = 0 as z + ω 2n z = − x  and obtain the response in terms
of relative displacement. We get
t
1
z(t) = −
ωn ∫ x(τ ) sinω (t − τ ) dτ (4.19)
0
n
Moreover, since the relevant equation of motion can be written as

 + ω n2 z = 0, then u
u  = −ω 2n z, so that, owing to Equation 4.19, we can obtain
the response in terms of absolute acceleration as
t
(t) = ω n
u
∫ x(τ ) sinω (t − τ ) dτ (4.20)
0
n
Remark 4.3
i. Equations 4.19 and 4.20 are frequently used in practice because in most
applications the motion of the base is measured with accelerometers.
ii. The relative motion equation (4.19) is important for the evaluation of
stress. For example, failure of a simple 1-DOF system generally cor-
responds to excessive dynamic load on the spring. This occurs when
z(t) max exceeds the maximum permissible deformation (hence stress)
of the spring.
4.3 HARMONIC AND PERIODIC EXCITATION
An important point for both theory and practice is that the FRF H (ω ) of
Equation 4.81 provides the system’s steady-state response to a sinusoidal forc-
ing function of unit amplitude at the frequency ω . Using complex notation,
in fact, the equation of motion in this case is mu + cu + ku = e iω t, which, by
iω t
(
assuming a solution of the form u(t) = Ue , leads to −ω 2m + iω c + k U = 1 )
and consequently
1 1
U= = = H (ω ) (4.21)
(
k − mω 2 + icω k 1 − β 2 + 2iζ β )
where, in writing the second expression, we used the relations
m k = 1 ω n2 , c k = 2ζ ω n and the definition of the frequency ratio β = ω ω n .
Then, the fact that H (ω ) is a complex function means that it has a magni-
tude and phase and therefore that the response may not be in phase with the
excitation. More explicitly, we can write the polar form H (ω ) = H (ω ) e − iφ
and, after a few simple calculations, determine the magnitude and phase
angle as
1 2ζβ
H (ω ) = , tan φ = (4.22a)
(
k 1− β ) + (2ζβ )
2 2 2 1− β2
where φ is the angle of lag of the displacement response relative to the har-
monic exciting force, and we have φ ≅ 0 for β << 1, φ = π 2 radians for β = 1
and φ that tends asymptotically to π for β >> 1. Also, it is not difficult to
show that the real and imaginary parts of H (ω ) are
 1 1− β 2
 1 2ζ β
Re [ H (ω )] =   , Im [ H (ω )] = −  
( )
 k  1 − β 2 + ( 2ζβ )2
2
( )
 k  1 − β 2 + ( 2ζβ )2
2
(4.22b)
Remark 4.4
i. For a unit amplitude force applied statically, the system displace-

ment is 1 k , i.e. H (ω = 0) = 1 k. For this reason, it is often convenient
to consider the non-dimensional quantity D(ω ) = k H (ω ) called as
dynamic magnification factor, which is basically the magnitude of
the system’s response ‘normalised’ to the static response (so that
D(ω = 0) = 1).
ii. In regard to phase angles, we recall from Chapter 1 that veloc-
ity leads displacement by π 2 radians and that acceleration leads
velocity by π 2 radians (hence, acceleration leads displacement by π

radians). At resonance, therefore, velocity is in phase with the exter-
nal force.
iii. In graphic format, the FRF is usually represented by plotting two
graphs: the magnitude and phase versus frequency (sometimes called
the Bode representation or diagram) or the real and imaginary parts
versus frequency. Moreover, the magnitude plot is often drawn using
logarithmic scales (or, at least, a logarithmic scale for the vertical
axis) in order to cover a greater range of frequencies and amplitudes
in a readable graph of reasonable size. Also, in the field of modal
analysis, the so-called Nyquist plot is frequently used; this is a single
plot which displays the imaginary part of the FRF as a function of its
real part. The disadvantage of this type of plot is that it does not show
the frequency information explicitly and captions must be added to
indicate the values of frequency.
As usual, the undamped case is obtained by setting ζ = 0 in the equations

mentioned earlier. By so doing, it is almost immediate to notice that for
β = 1 (that is, ω = ω n ), the undamped response diverges to infinity. This is
the well-known condition of resonance, where in real-world systems, the
divergence is prevented by the unavoidable presence of damping. The result
is that for systems with a damping ratio ζ < 1 2 ≅ 0.707, the magnitude
D(ω ) shows a peak at β ≅ 1, where some simple calculations show that the
( )
−1
maximum occurs at β = 1 − 2ζ 2 and that Dmax = 2ζ 1 − ζ 2 , which is
often approximated by Dmax ≅ 1 2ζ for small damping. This, in turn, implies
that for values of, say, ζ = 0.05 or ζ = 0.1 – which are not at all uncommon
in applications – the amplitude of the displacement response at resonance
is, respectively, ten times or five times the static response. For small damp-
ing, such high values of the response are due to the fact that at resonance
the inertia force is balanced by the spring force and that, consequently, the
external force overcomes the (relatively small) damping force. Also, note
that damping plays an important role only in the resonance region; away
from resonance – that is, in the regions β << 1 and β >> 1 – damping is defi-
nitely of minor importance. In this respect, it can be observed that in the
resonance region, the system’s steady-state response can be approximated
by u(t) ≅ ( f0 cω n ) e i(ω t −π 2), thus showing that c is the ‘controlling parameter’
for ω close to ω n . By contrast, when β << 1, we are not far from the condi-
tion of static excitation and we expect the stiffness k to be the ‘controlling
parameter’. This is confirmed by the approximation u(t) ≅ ( f0 k) e iω t , which
in fact holds for ω << ω n. At the other extreme – that is, when β >> 1 – the
( )
approximation u(t) ≅ f0 mω 2 ) e i (ω t −π) indicates that mass is the ‘controlling
parameter’ in this region.
Example 4.3: Resonant response

As already pointed out earlier, the FRF gives the steady-state response.
For a relatively short interval of time after the harmonic driving force
(which, in this example, we assume of amplitude f0 instead of unity)
has been ‘turned on’, we must take into account also the transient part
of the solution and write, for instance, in sinusoidal notation
u(t) = e −ζ ω n t ( A cos ω d t + B sin ω d t ) + f0 H (ω ) cos (ω t − φ ) (4.23)
With this solution, the initial conditions u(0) = u0 and u (0) = v0 give the
constants
1
A = u0 − f0 H (ω ) cos φ , B=
ωd
(v0 − ω f0 H (ω ) sin φ + ζω n A) (4.24a)
Now, assuming the system to start from rest and considering that at
resonance (ω = ω n ) we have φ = π 2 and f0 H (ω n ) = f0 2ζ k, the two
constants reduce to
ω n f0 f
A = 0, B=− ≅ − 0 (4.24b)
2ζ kω d 2ζ k
where in writing the rightmost expression for B, we further assumed

small damping, so that ω n ω d ≅ 1. Then, the solution 4.23 becomes
u(t) ≅
(
f0 1 − e −ζ ω n t ) sin ω t (4.25)
n
2ζ k
which shows that the response builds up asymptotically to its max-

imum value f0 2ζ k. It is now left to the reader to draw a graph of
Equation 4.25 for a few different values of ζ .
Having determined that the system’s steady-state response to a harmonic

excitation of frequency ω and amplitude f0 is u(t) = f0 H (ω ) e i (ω t −φ ) , it is
now almost immediate to obtain the response to a driving force such that
f (t) = f (t + T ), that is, an excitation of period T. In Appendix B, in fact, we
have seen that an excitation of this type can be represented in the form of
a Fourier series as
∞
f (t) = ∑C e
r =−∞
r
iω r t
(4.26)
where ω r = rω 1 ( r = 1,2,) and ω 1 = 2π T , and the Cr coefficients are

obtained as shown in Equations B.5. Then, since linearity implies that the
superposition principle holds, the response is itself a Fourier series in which

each term is the response to the corresponding individual term of the input
series. Consequently, the response to the forcing input 4.26 is
∞
u(t) = ∑C
r =−∞
r H ( rω 1 ) e i ( rω1 t −φr ) (4.27)
and we can have a resonance condition whenever one of the exciting frequen-
cies ω 1 ,2ω 1 ,3ω 1 , is close or equal to the system’s natural frequency ω n .
4.3.1 A few notes on vibration isolation

If the external excitation to a damped 1-DOF system is a base motion with
displacement time-history x(t) – and where, as earlier, we call u(t) the dis-
placement of the system’s mass – we may be interested in limiting as much
as possible the motion of the mass (which can be, for instance, a sensitive
machinery unit). For this case, we have already seen in Section 4.2.1 that
the equation of motion of an undamped system is mu  + k(u − x) = 0. With
some non-zero damping, this becomes
 + c ( u − x ) + k ( u − x ) = 0 ⇒
mu  + cu + ku = cx + kx (4.28)
mu
and if the support/base displacement is harmonic at frequency ω , we can write

x(t) = Xe iω t. Then, assuming that the mass displacement is also harmonic of
the form u(t) = Ue iω t = U e i (ωt −φ ), we can calculate the prescribed derivatives
and substitute them in Equation 4.28 to obtain the output–input ratio
U k + icω 1 + 2 iζ β
= = (4.29)
X k − mω + icω 1 − β 2 + 2 iζ β
2
which characterises the motion transmissibility between the base and the
mass. At this point, some easy calculations show that the magnitude and
phase of the complex function U X are
U k2 + (cω )2 1 + (2ζβ )2
= =

X (k − mω ) 2 2
+ (cω )2 (1 − β ) + (2ζβ )
2 2 2
(4.30)
2ζβ 3
tan φ =
1 − β 2 + 4ζ 2β 2
where now φ is the angle of lag of the mass displacement with respect to the
base motion.
A different problem is when the system’s mass is subjected to a har-
monic force f (t) = F0 e iω t, and we are interested in limiting the force fT (t)
transmitted to the supporting base. Observing that under the action of

f (t) the mass displacement is harmonic of the form u(t) = Ue iω t , the sys-
( )
tem’s equation of motion gives F0 = − mω 2 + icω + k U . On the other
hand, since fT (t) = cu + ku because it is transmitted to the base through the
damper and the spring, we can assume a harmonic form fT (t) = FT e iω t to get
FT = ( icω + k) U .
Then, putting together this equation with the F0 equation in order to
determine the force transmissibility FT F0 , we obtain
FT k + icω
= (4.31)
F0 k − mω 2 + icω
which is exactly the same function of Equation 4.29 – although with a dif-
ferent physical meaning – and implies that the magnitude FT F0 is given by
the r.h.s. of Equation 4.301 (clearly, the phase of FT relative to F0 is the same
as the r.h.s. of Equation 4.302 , but in these types of problems, the phase is
generally of minor importance).
A first conclusion, therefore, is that the problem of isolating a mass from
the motion of the base is the same as the problem of limiting the force trans-
mitted to the base of a vibrating system. And since the magnitude of U X
or FT F0 is smaller than one only in the region β > 2 , it turns out that we
must have ω > 2 ω n (or ω n < ω 2 ) in order to achieve the desired result. It
should also be noticed that for β = 2 , all magnitude curves have the same
value of unity irrespective of the level of damping.
In addition to this, a rather counterintuitive result is that in the isolation
region β > 2 , a higher damping corresponds to a lower isolation effect
(it is not so, however, in the region β < 2 ), so that, provided that we stay
in the ‘safe’ isolation region, it is advisable to have low values of damping.
When this is the case, it is common to approximate the transmissibility – we
( )
denote it here by T – by T ≅ 1 β 2 − 1 and refer to the quantity 1 − T as the
isolation effectiveness.
Remark 4.5
It is important to point out that the force transmissibility 4.31 is the same
as the motion transmissibility 4.29 only if the base is ‘infinitely large’, that
is, if its mass is much larger than the system’s mass. If it is not so, it can be
shown (Rivin, 2003) that we have
FT mb k2 + (cω )2
= 2 (4.32)
F0 m + mb  mmb 2
 k − m + m ω  + (cω )
2
b
where mb is the mass of the base. As expected, in the limit of mb → ∞, this

becomes the magnitude on the r.h.s. of Equation 4.301
A further result can be obtained if we go back to the first of Equations

4.28 and observe that z = u − x is the relative displacement of the mass with
respect to the base. Then, by subtracting the inertial force mx  from both
sides, we get the equation of relative motion mz + cz + kz = − mx , which,
assuming x(t) = Xe iω t and z(t) = Ze iω t , leads to
Z mω 2 β2 Z β2
= = ⇒ =
X k − mω + icω 1 − β + 2iζ β
(1 − β )
2 2
X 2 2
+ (2ζ β )2
(4.33)
This result is often useful in practice for the two reasons already mentioned
in Remark 4.3: first because in some applications relative motion is more
important than absolute motion and, second, because the base acceleration
is in general easy to measure.
4.3.2 Eccentric excitation
Eccentric excitation is generally due to an unbalanced mass me with eccen-
tricity r that rotates with angular velocity ω (see Figure 4.1). A typical exam-
ple is an engine or any rotary motor mounted on a fixed base by means of
a flexible suspension.
The equation of motion for the system of Figure 4.1 is mu  + cu + kx = fecc (t),
where the magnitude of fecc is Fecc = me rω 2 . Along the lines of the preceding
section, assuming harmonic motions for the driving force and the displace-
ment response of the form fecc (t) = Fecc e iω t and u(t) = Ue iω t , we get
Figure 4.1 E ccentric excitation.

U 1 U mω 2
= ⇒ = (4.34)
Fecc k − mω 2 + icω r µ k − mω 2 + icω
where in writing the second expression, we took the relation Fecc = me rω 2

into account and also defined the dimensionless parameter µ = me m. Then,
some easy calculations show that the magnitude and phase of the complex
function U rµ are
U β2 2ζβ
= , tan φecc = (4.35)
rµ
(1 − β ) + (2ζβ )
2 2 2 1− β2
where β ,ζ are the usual frequency and damping ratios.

A different case arises when the system is supported on a base of mass mb ,
where there is a rotating eccentric mass me with eccentricity r (Figure 4.2).
This eccentric-base-excited system can model, for example, a piece of
equipment of mass m mounted on a rotary motor.
Calling u(t), x(t) the absolute motion of the mass and of the base, respec-
tively – thus implying that z = u − x is the relative motion of the mass with
respect to the base – we have the equations of motion
 − cz − kx = − fecc (t),

mb x  + cz + kx = 0 (4.36a)
mu
where, as above, the magnitude of fecc is Fecc = me rω 2 . Then, multiplying the

first of Equations 4.36a by m, the second by mb and then subtracting one
result from the other, we get the equation of relative motion
mEz + cz + kz = ( mE mb ) fecc (4.36b)
Figure 4.2 Eccentric excitation of the base.

where we defined mE = mmb ( m + mb ). If now, as usual, we assume har-

monic motions of the form fecc (t) = Fecc e iω t and z(t) = Ze iω t , substitution in
Equation 4.36b leads to
 mb  Z 1 Z mEω 2
 m  F = k − m ω 2 + icω ⇒ =
r µ k − mEω 2 + icω
(4.37)
E ecc E
where now the dimensionless parameter µ is defined as µ = me mb . Finally,

by further introducing the definitions
k c ω
ωn = , ζ= , β= (4.38)
mE 2mE ω n ωn
the magnitude and phase of the complex function 4.37 are
Z β2 2ζ β
= , tan φecc = (4.39)
rµ
(1 − β ) + (2ζβ )
2 2 2 1− β2
which, although formally identical to the results of Equation 4.35, have a

different physical meaning (first of all, it may be worth reiterating, the fact
that Equation 4.39 refers to the relative motion of the mass with respect to
the base while Equation 4.35 refers to the absolute motion of the mass on
a fixed base, and second, the fact that in Equations 4.39 the parameters
ω n , ζ , β do not have their usual meaning).
4.3.3 Other forms of FRFs

Going back to Equation 4.91 and rewriting it as H (ω ) = U(ω ) F(ω ) we
see that, in the frequency domain, the FRF is the ratio of the system’s
response – or, with a different term, the output – divided by the excitation-
input (clearly, the same is true for the TF in the Laplace domain, but here
we focus our attention on FRFs because these are the functions used in
experimental methods, measurements and analyses).
Then, the fact that the system’s response can be expressed, or measured,
in terms of displacement, velocity or acceleration suggests that the ratios
velocity/force or acceleration/force are just different forms of FRFs other
than the displacement/force given above. This is indeed the case, and the
various FRFs of the type motion/force have different names depending on
the response quantity. The nomenclature may vary depending on the author
but the most common and widely accepted terms are as follows:
Displacement/Force = Receptance
Velocity/Force = Mobility
Acceleration/Force = Accelerance
As for notation, H (ω ) is often used as a generic symbol for FRFs – be

it receptance, mobility or accelerance – because the meaning is generally
clear from the context. Sometimes this ‘ambiguity’ is not a problem, but
if it is important to know which is which, it is advisable to use different
symbols; here, whenever needed, we will use R(ω ) for receptance, M(ω ) for
mobility and A(ω ) for accelerance (note, however, that these symbols are
mnemonically simple but they are not standard; for example, one often
finds the symbols α (ω ) for receptance and Y (ω ) for mobility). More specifi-
cally, therefore, the FRF of Equations 4.81 and 4.21 is a receptance, but
since the various output quantities are mathematically related (Chapter 1,
Section 1.3.2), it is now not difficult to determine that the mobility and
accelerance are
iω −ω 2
M(ω ) = , A(ω ) = (4.40a)
k − mω 2 + icω k − mω 2 + icω
with respective magnitudes
ω ω2
M(ω ) = , A(ω ) = (4.40b)
(k − mω )
2 2
+ (cω )2 (k − mω )
2 2
+ (cω )2
In regard to the polar forms of the various FRFs, we adopt the conven-
tion of writing R(ω ) = R(ω ) e − i φ D (see Section 4.3, where here we added
the subscript ‘D’ for ‘displacement’ to the phase angle of receptance),
M(ω ) = M(ω ) e − i φ V and A(ω ) = A(ω ) e − i φA . By so doing, φD ,φ V ,φA are under-
stood as angles of lag behind the driving excitation, this meaning that a
positive value of φ corresponds to an angle of lag while a negative value
corresponds to an angle of lead. Then, in order for our convention to be in
agreement with the physical fact that the velocity leads displacement by π 2
and acceleration leads displacement by π, we have the relations φV = φD − π 2
and φA = φD − π = φV − π 2.
Remark 4.6
i. With our convention, the graphs of φD ,φ V ,φA are all monotonically

increasing functions of β , with φD that ranges from zero (for β = 0)
to π (for β >> 1), φV that ranges from − π 2 to π 2 and φA that ranges
from −π to zero. Also, at resonance, we have φD = π 2, φ V = 0 and
φA = − π 2 , meaning that when β = 1 displacement lags behind the
excitation by π 2, velocity in phase with the excitation and accelera-
tion leads the excitation by π 2.
ii. Some authors use different conventions of writing the phase exponen-
tial as e i φ , so that a positive φ corresponds to an angle of lead while a
negative value to an angle of lag. With this convention, the graphs of

φD ,φ V ,φA are all monotonically decreasing functions of β , the phase
angle φD ranges from zero (for β = 0) to −π (for β >> 1) and we have the
relations φV = φD + π 2 and φA = φD + π = φV + π 2.
iii. As a useful exercise, the reader is invited to draw a graph of the mag-
nitudes 4.40b and note the different behaviours of receptance, mobil-
ity and accelerance in the regions β << 1 and β >> 1.
Finally, it is worth pointing out that in some applications, one can find
the inverse relations of the FRFs mentioned above; these are the force/
motion ratios
Force/Displacement = Dynamic stiffness
Force/Velocity = Mechanical impedance
Force/acceleration = Apparent mass
with the frequently used symbols K(ω ) and Z(ω ) for dynamic stiffness and
mechanical impedance (while, to the author’s knowledge, there is no stan-
dard symbol for apparent mass, so here we will use J(ω )). Then, from the
developments above, it is not difficult to see that mathematically we have
K = R−1 , Z = M −1 and J = A−1.
4.3.4 Damping evaluation
In Section 3.2.1, it was shown that damping can be obtained from a
graph of the decaying time-history of the system free-response. With the
FRFs at our disposal, we now have other methods to evaluate damping
and here we consider three of the most common. The first consists in
simply determining the maximum value of the receptance, which occurs
at resonance and is f0 2ζ k if the driving force has magnitude f0 . Then,
we have
f0
ζ≅ (4.41)
2k R(ω ) max
In the second method called ‘half-power bandwidth’, we must identify the

two points where the response is reduced to 1 2 = 0.707 of its peak value.
These points (also called -3dB points because with a logarithmic scale
( )
20log10 1 2 = −3) can be mathematically identified by the condition
1 1
= (4.42)
2ζ 2
(1 − β ) + (2ζβ )
2 2 2
( )
which, in turn, leads to the equation β 4 − 2β 2 1 − 2ζ 2 + 1 − 8ζ 2 = 0 whose
roots are β = 1 − 2ζ ± 2ζ 1 + ζ . For small damping, we can write
2
1,2
2 2
β1,2
2
≅ 1 ± 2ζ and consequently β1,2 = 1 ± 2ζ ≅ 1 ± ζ , from which it follows:
β1 − β2 ω 1 − ω 2
ζ= = (4.43)
2 2ω n
The third method is based on the Nyquist plot of mobility. Recalling that
we mentioned Nyquist plots in Remark 4.4(iii), we add here that one note-
worthy feature of these graphs that they enhance the resonance region
with an almost circular shape. For viscously damped systems, however, the
graph of mobility traces out an exact circle, and the radius of the circle can
be used to evaluate the damping constant c. More specifically, since the real
and imaginary parts of mobility are
Re [ M(ω )] =
cω 2
, Im [ M(ω )] =
(
ω k − mω 2 )
(k − mω ) 2 2
+ (cω )
2
(k − mω ) 2 2
+ (cω )2
(4.44)
we can define
1
U = Re [ M(ω )] − , V = Im [ M(ω )] (4.45)
2c
and determine after some easy mathematical manipulations that we have

U 2 + V 2 = 1 4c2 , which, in the U ,V -plane, is the equation of a circle of
radius 1 2c with centre at the origin. Then, using Equations 4.45, it is
immediate to see that in the original Nyquist plane, the centre is at the
point (1 2c ,0). Also, it is worth observing that the resonant frequency ω n
is obtained from the point where the circle intersects the real axis, that
is, when Im [ M(ω )] = 0.
4.3.5 Response spectrum
In Example 4.2, we have determined the response of an undamped system
to a rectangular impulse of amplitude f0 and duration t1. Here, however,
we are not so much interested in the entire time-history of the response
but we ask a different question: what is the maximum value umax of the
response and when does it occur? As we shall see shortly, the answer to this
question – which is often important for design purposes – leads to the very
useful concept of response spectrum.
Going back to the specific case of Example 4.2, the system’s response
‘normalised’ to the static response f0 k is
ku ( t < t1 ) ku ( t > t1 )
= 1 − cos ω nt , = cos ω n ( t − t1 ) − cos ω nt (4.46)
f0 f0
in the primary and residual regions, respectively. And since, depending on

the value of t1, the maximum response can occur either in the first or in the
second region, we must consider them separately.
Primary region. For t < t1, equating the time derivative of 4.461 to zero leads
to the condition sin ω nt = 0, thus implying that the first maximum occurs
at time t M = π ω n = T 2, which, introducing the dimensionless parameter
η = t1 T and observing that we must have t M < t1 (otherwise the maximum
would not be in the primary region), gives η > 1 2. The maximum response
is now obtained by substituting t M in Equation 4.461. By so doing, we get
kumax f0 = 2.
Residual region. In the residual region t > t1, the condition for a maxi-
mum is sin ω n ( t − t1 ) = sin ω nt , which – considering again the first maxi-
mum – is satisfied at the time t M given by ω n ( t M − t1 ) = π − ω nt M, that is,
t M = (T 4) + ( t1 2). This, in turn, implies η ≤ 1 2 when we observe that now
we must have t M > t1 (otherwise, the maximum would not be in the residual
region). Then, substitution of t M in Equation 4.462 gives, after a few easy
manipulations, kumax f0 = 2sin(πη).
Finally, combining the two results, we get
kumax  2sin(πη) (η ≤ 1 2)
= (4.47)
f0 2 (η > 1 2)

so that a graph of kumax f0 as a function of η gives the so-called response
spectrum to a finite-duration rectangular input. In just a few words, there-
fore, one can say that the response spectrum provides a ‘summary’ of the
largest response values (of a linear 1-DOF system) to a particular input
loading – a rectangular pulse in our example – as a function of the natural
period of the system. Also, note that by ignoring damping, we obtain a
conservative value for the maximum response.
In order to give another example, it can be shown (e.g. Gatti, 2014, ch. 5)
that the response of an undamped 1-DOF system to a half-sine excitation
force of the form
 f0 sin ω t 0 ≤ t ≤ t1
f (t) =  (4.48)
 0 t > t1
(with ω = π t1 , so that in the ‘forced vibration era’, we have exactly half a

cycle of the sine) is, in the primary and residual regions, respectively,
f0
u( t ; t ≤ t1 ) = (sin ω t − β sin ω nt )
(
k 1− β2 )
(4.49)
f0 β
u( t ; t > t1 ) = sin ω n t + sin ω n ( t − t1 ) 
(
k β2 −1  )
which in turn – and here the reader is invited to fill in the missing details
(see Remark 4.7(i) below for a hint) – lead to the maximum ‘normalised’
displacements
k 1  2πβ  2η  2π 
umax ( t ≤ t1 ) = sin  = sin
f0 1− β  1 + β  2η − 1  1 + 2η 
(4.50)
k 2β  π  4η
umax ( t > t1 ) = 2 cos  = cos(πη)
f0 β −1  2β  1 − 4η 2
where the first expressions on the r.h.s. are given in terms of the frequency
ratio β = ω ω n , while the second are in terms of η = t1 T . In this respect,
moreover, two points worthy of notice are as follows:
a. Equation 4.501 holds for β < 1 (or equivalently η > 1 2), while Equation
4.502 holds for β > 1 (or equivalently η < 1 2),
b. At resonance (β = 1 or η = 1 2), both expressions (4.50) become inde-
terminate. Then, the well-known l’Hôpital rule from basic calculus
gives (using, for example, the first expression of 4.501)
k π
umax ( β = 1) = ≅ 1.571
f0 2
Remark 4.7
i. As a hint to the reader, we observe that in the primary region, the

maximum occurs at the time t M such that ω t M = 2π − ω nt M (pro-
vided that t M ≤ t1); on the other hand, provided that t M > t1, the
maximum in the residual region is attained at the time t M such that
ω nt M = π − ω n ( t M − t1 ).
ii. In the examples above, we considered two inputs with a finite dura-
tion t1. For step-type inputs whose final value is not zero – like, for
instance, the ‘constant-slope excitation’ given by f (t) = f0 t t1 for
0 ≤ t < t1 and f (t) = f0 for t ≥ t1 – the parameter t1 of interest is not the
duration of the input but its rise time, that is, the time it takes to reach
its full value.
4.4 MDOF SYSTEMS: CLASSICAL DAMPING
From the developments of Chapter 3, we know that the equations of motion

of a classically damped system can be uncoupled by passing to the set of
normal (or modal) co-ordinates y. This is accomplished by means of the
transformation u = Py, where P is the matrix of eigenvectors (which here we
assume to be known and mass-normalised). In the case of forced vibrations
with a non-zero term f(t) on the r.h.s. of the equation of motion, the same
procedure of Section 3.4 leads to
 + PT CP y + Ly = PT f (4.51)
Iy
where L = diag ( λ1 ,, λn ) is the diagonal matrix of eigenvalues (with λ j = ω 2j ),

the matrix PT CP is diagonal and the forcing vector PT f on the r.h.s. is called
modal force vector. Then, Equation 4.51 is a set of n 1-DOF equations of
motions, and explicitly, we have
yj + 2ω jζ j y j + ω 2j y j = ϕ j ( j = 1,2,…, n ) (4.52)

where ϕ j (t) = pTj f (t) is the jth element of the modal force vector and is
called jth modal participation factor. Also, since the transformation to
normal coordinates gives the initial conditions y0 , y 0 of Equations 3.39, it
is evident that, in agreement with Equation 4.4, the solutions of Equation
4.52 are
 y0 j + ζ jω j y0 j 
y j (t) = e −ζ jω j t  y0 j cos ω dj t + sin ω dj t 
 ω dj 
t
1
∫ ϕ (τ ) e sin ω dj ( t − τ )  dτ
−ζ jω j (t −τ )
+ j (4.53)
ω dj
0
where y0 j , y0 j are the initial displacement and velocity of the jth modal coor-
dinate and ω dj = ω j 1 − ζ j2 is the jth damped frequency. If, as it is often the
case, the initial conditions are zero, we can compactly write in matrix form
t
y(t) =
∫ diag hˆ (t − τ ),, hˆ (t − τ ) P f(τ ) dτ (4.54)
0
1 n
T
( )
where hˆ j (t) = e −ζ jω j t ω dj sin ω djt is the jth modal IRF. Then, with the trans-
formation u = Py, we can go back to the original physical co-ordinates, and
Equation 4.54 gives
t t
u(t) =
∫
0
P diag  hˆ1(t − τ ),, hˆ n (t − τ )  PT f (τ ) dτ =
  ∫ h(t − τ ) f(τ ) dτ (4.55)
0
where, in the last expression, we introduced the IRF matrix in physical

coordinates, that is, the matrix
n n
h(t) = P diag  hˆ j (t)  PT =

  ∑m=1
hˆ m (t) pm pTm , hjk (t) = ∑ hˆ (t) p
m=1
m p
j m km (4.56)
where, for brevity, in Equation 4.561, we write diag  hˆ j (t)  for

 
diag  hˆ1(t),, hˆ n (t)  and where Equation 4.562 gives the explicit form of the
 
j, kth element of h(t). Also, note that Equations 4.56 clearly show that h(t)
is symmetric.
Remark 4.8
If, with zero initial conditions, the external loading f is orthogonal to one of
the system’s modes, say the kth mode pk, then ϕ k = pTk f = 0 and consequently
y j = 0, meaning that this mode will not contribute to the response. By the
same token, we can say that if we do not want the kth mode to contribute
to the response, we must choose ϕ k = 0.
If now, in the 1-DOF Equations 4.52, we assume a harmonic excitation of

the form ϕ j (t) = pTj F e iω t , the steady-state response will also be harmonic
with the form y j (t) = Yj e iω t . Consequently, we are led to
(ω 2
j )
− ω 2 + 2iζ jω jω Yj = pTj F ( j = 1,2,, n )
and these n equations can be compactly expressed in the matrix form as
 1  T
Y = diag  2 P F = diag  Hˆ j (ω )  PT F (4.57)
 ω j − ω + 2iζ jω jω   
2
( )
−1
where Hˆ j (ω ) = ω 2j − ω 2 + 2 iζ jω jω is the jth modal FRF ( j = 1,2,, n ).
Then, the response in terms of physical coordinates is given by
U = P diag  Hˆ j (ω )  PT F (4.58)
 
so that, in agreement with the developments and nomenclature of the pre-
ceding sections, the ratio of the displacement response U and the forcing
excitation F is the system’s receptance. This is now an n × n matrix, and
Equation 4.58 rewritten as U = R(ω ) F shows that we have
R(ω ) = P diag  Hˆ j (ω )  PT =
  ∑ Hˆ
m=1
m (ω ) pm pTm (4.59a)
from which it is easy to see that R(ω ) is a symmetric matrix. Its j,kth
element is
n n
∑ ∑ω
p j m pkm
Rj k (ω ) = Hˆ m (ω )p j m pkm = (4.59b)
m=1 m=1
2
m − ω 2 + 2 iζ mω mω
and, in physical terms, expresses the (complex) amplitude of the displacement

response of the jth DOF when a unit harmonic force is applied at the kth DOF.
Remark 4.9
i. In light of the developments of Section 4.2, the fact that the modal
IRF hˆ j (t) and the modal (receptance) FRF Hˆ j (ω ) form a Fourier trans-
form pair is no surprise. More precisely, owing to our definition of
Fourier transform (see Appendix B and Remark 4.2(i)), we have
Hˆ j (ω ) = 2π F  hˆ j (t)  ,
hˆ j (t) = (1 2π ) F −1  Hˆ j (ω )  ( j = 1, , n )
   
(4.60a)

R(ω ) = 2π F[ h(t)], h(t) = (1 2π) F−1[ R(ω )] (4.60b)
ii. The fact that the IRF and FRF matrices h(t), R(ω ) are symmetric is a
consequence of the reciprocity theorem (or reciprocity law) for linear
systems, stating that the response – displacement in this case – of the
jth DOF due to an excitation applied at the kth DOF is equal to the dis-
placement response of the kth DOF when the same excitation is applied
at the jth DOF. Clearly, reciprocity holds even if the system’s response
is expressed in terms of velocity or acceleration, thus implying that the
mobility and accelerance FRF matrices M(ω ), A(ω ) are also symmetric.
4.4.1 Mode ‘truncation’ and the mode-

acceleration solution
From the developments of the preceding section, we see that for undamped
and classically damped systems, the response u(t) is given in the form of
∑
t
a mode superposition u(t) = p y (t), where y (t) =
j
ϕ (τ )hˆ (t − τ ) dτ is
j j j
∫0
j j
the solution for the jth modal coordinate and hˆ j (t) is the jth modal IRF.
However, observing that in many cases of interest, only a limited num-
ber of lower-order modes (say the first r, with r < n and sometimes even
r << n ) gives a significant contribution to u(t); it is reasonable to expect
that perhaps we can obtain a satisfactory approximate solution by simply
‘truncating’ the modes higher than the rth and by considering only the
sum
u (r ) = ∑ p y (4.61)
j =1
j j
This is certainly possible, but since the response depends on the system
under investigation and on the type of loading (its frequency content,
spatial distribution, etc.), the question arises of how many modes should
be retained; too many implies a larger and unnecessary computational
effort and too few may mean a poor and inaccurate solution – and espe-
cially so if the excitation has some frequency components that are close
to one of the truncated modes (in this respect, it is eminently reasonable
that, at a minimum, one should retain all the modes that fall within
the frequency band of the excitation). So, provided that in any case, the
truncation must be made with some ‘educated engineering judgment’; the
idea of the so-called mode-acceleration method is to add a complement-
ing term that takes into account the contribution of the truncated n − r
modes. With respect to simple truncation, experience has shown that
this method often provides a significant improvement on the quality of
the solution.
Considering an undamped system for simplicity, the equations of motion
can be rewritten as Ku = f − Mu,  and consequently, if K is non-singular,
u = K −1f − K −1Mu
 (4.62)
Then, observing that Equation 4.61 implies u (r) = ∑ rj =1 p j yj , we can use this
in the r.h.s. of Equation 4.62 to obtain a truncated mode acceleration solu-
tion u(r) as
r r
∑ ∑ω
yj
u(r) = K −1f − K −1 yj M p j = K −1f − 2 p j (4.63)
j
j =1 j =1
where in writing the last expression we took the relation MPj = ω j−2 KP into
account. As for terminology, the term K −1 f on the r.h.s. is called the pseudo-
static response, while the name of the method is due to the accelerations yj
in the other term.
t
Now, since each y j is given by y j = (1 ω j )
∫ p f(τ )sin ω (t − τ ) dτ , we
0
T
j j
can insert this expression into the jth modal equation of motion rewritten
as yj = pTj f − ω 2j y j and substitute the result in Equation 4.63 to obtain
r  r
p j pTj 
∑ ∑
p j pTj t
u (r ) =
j =1
ωj ∫ 0
f (τ )sin ω j (t − τ )  dτ +  K −1 −
 j =1
ω 2j 
 f (4.64a)
which is an expression in terms of the lower-order modes only. If now we

recall from Remark 3.8 the spectral expansion of the matrix K −1, that is,
K −1 = ∑ nj =1 ω j−2 p j pTj , we can write u r in the form
r n
∑ ∑
p j pTj t
p j pTj
u (r ) =
j =1
ωj ∫ 0
f (τ )sin ω j (t − τ )  dτ +
j = r +1
ω 2 f (4.64b)
j
where the last term represents the contribution of the n − r truncated modes.
Remark 4.10
If now, in the frequency domain, we consider the receptance of the same

undamped system, define β j = ω ω j and assume (a) that the excitation lies
within the limited bandwidth ω < ω B and (b) that ω B is significantly smaller than
the modes with frequencies ω r +1 ,, ω n , then we can write the approximation
n r n
p j pTj  1  p j pTj  1 
∑ ∑ ∑
p j pTj
R(ω ) = ≅ + (4.65a)
j =1
ω 2j  1 − β j2  j =1
ω 2j  1 − β j2  j = r +1 ω 2j
where in the last expression the first term represents the dynamic response
of the lower-order modes in the frequency bandwidth of the excitation,
while the second term is a pseudo-static correction due to the fact that for
the modes with indexes j = r + 1,, n it is legitimate to make the ‘static’
( )
−1
approximation 1 − β j2 ≅ 1. If now, as we did mention earlier, we take
into account the modal expansion of the matrix K −1 and observe that
( )
−1
Hˆ j (ω ) = ω 2j − ω 2 is the jth modal FRF (in the form of receptance) of our
undamped system, we can write
n r
∑ ∑ω
p j pTj
R(ω ) ≅ Hˆ j (ω ) p j pTj + K −1 − 2 (4.65b)
j
j =1 j =1
which is an expression in terms of the lower-order modes only.

4.4.2 The presence of rigid-body modes

Assuming a harmonic excitation and a harmonic response of the forms
f (t) = F e iω t and u(t) = U e iω t, respectively, the equations of motion
Mu  + Cu + Ku = f lead to
( −ω 2
M + iω C + K U = F (4.66) )
where the vector U can be expanded in terms of the system’s modes and
∑
n
expressed as U = bj p j . For a classically damped system with no rigid-
j =1
body modes, in fact, Equation 4.58 is such an expansion because, in light
of Equation 4.59a, it is not difficult to show that the expansion coefficients
are bj = Hˆ j (ω ) pTj F.
On the other hand, if the system has m rigid-body modes rj , the expan-
sion becomes
m n− m
U= ∑ j =1
a j rj + ∑ b p (4.67)
j =1
j j
which can be substituted in Equation 4.66. By so doing, we get a somewhat

lengthy expression, which in turn – taking into account the usual orthogo-
nality conditions for the elastic modes and those for the rigid-body modes
(Section 3.3.8, with also riT C rj = 0 for all i, j because we are dealing with
rigid, non-vibrational modes) – can be pre-multiplied by rkT to obtain the
a-coefficients and by pTk to obtain the b-coefficients. The results are
rkT F pTk F
ak = − , bk = = Hˆ k (ω ) pTk F (4.68)
ω2 ω − ω + 2i ζ k ω kω
2
k
2
and show that the coefficients of the elastic modes are the same as in the
absence of rigid-body modes. Then, Equation 4.67 becomes
m n− m  m n− m 
∑( )
 rjT F 
∑ ∑ ∑
rj rjT
U=−  ω 2  rj + Hˆ j (ω ) pTj F p j =  − + Hˆ j (ω )p j pTj  F
j =1 j =1
 j =1
ω2 j =1

(4.69)
which in turn tells us that the receptance matrix R = U F is in this case

m n− m
∑ ∑ Hˆ (ω ) p p (4.70)
rj rjT
R(ω ) = − + j j
T
j
j =1
ω2 j =1
4.5 MDOF SYSTEMS: NON-CLASSICAL VISCOUS

DAMPING, A STATE-SPACE APPROACH
For non-classically damped systems, the matrix PT CP is not diagonal, thus

implying that the equations of motion are coupled through the PT CP term
and cannot be reduced to a set of n uncoupled 1-DOF equations.
Having considered the homogeneous case in Chapter 3, we have seen
(
there that now we are confronted with the QEP λ 2 M + λ C + K z = 0, whose )
eigenpairs are no longer real and whose eigenvectors satisfy the orthogonal-
ity conditions of Equations 3.89 and 3.90 instead of the definitely simpler
conditions of Equations 3.28 and 3.29.
One possibility to overcome these complications is to adopt a state-space
formulation and transform the n coupled second-order equations into a set
of 2n first-order equations, where, for instance – we recall from Chapter 3,
Section 3.5.2 – one such formulation is given by Equation 3.93a, to which it
corresponds to the symmetric GEP of Equation 3.94. Then, the 2n-dimen-
sional eigenvectors v j of this GEP are related to the original eigenvectors z j
T
of the second-order system by v j =  z j λ j z j  and satisfy the orthogo-
nality conditions of Equations 3.95.
Adopting here the same type of formulation for the non-homogeneous
problem, we can put together the equations of motion Mu  + Cu + Ku = f
with the identity Mu − Mu = 0 and write the matrix equation
 C M   u   K 0  u   f 
  +  =  ⇒ ˆ  + Kx
Mx ˆ =q
 M 0   u
   0 − M   u   0 
(4.71)
where now on the r.h.s. we have the 2n-dimensional forcing vector

T
q(t) =  f (t) 0  , while the 2n × 2n matrices Mˆ , Kˆ are the same as in
T
Equation 3.93b and, as in Section 3.5.2, x(t) =  u(t) u (t)  .
Remark 4.11
Recall that since the coefficients of the matrices involved are real, for
underdamped systems, the eigensolutions occur in complex conjugate
pairs. For calculations, therefore, it is useful to arrange them in some con-
( )
venient way; for example, for j = 1,, n , as ( λ2 j , z 2 j ) = λ2∗ j −1 , z ∗2 j −1 or as
( )
( λ j+n , z j+n ) = λ j∗ , z∗j .
Under the assumption that the eigenvectors form a complete set – which
is true if the eigenvalues are all distinct and, more generally, if the matrices
involved are non-defective – we can express the solution of Equation 4.71

as the superposition of eigenvectors
2n
x(t) = ∑ yˆ (t) v (4.72)

i =1
i i
so that substituting Equation 4.72 into 4.71, pre-multiplying by vTj and tak-
ing the orthogonality conditions 3.95 into account (and here, without the
ˆ j = 1 for all j) give the
loss of generality, we also adopt the normalisation M
2n independent first-order equations
dyˆ j d
− λ j yˆ j = φ j (t) ⇒  yˆ j (t)e − λ jt  = φ j (t)e − λ jt (4.73)
dt dt  
where φ j = vTj q and where the second expression follows from the first by
multiplying both sides by e − λ jt and rearranging the result. From Equation
4.732 , we readily obtain the solution yˆ j (t); assuming for simplicity zero ini-
tial conditions (i.e. yˆ j (0) = 0 for all j), we have
t

∫
yˆ j (t) = φ j (τ ) e λ j (t −τ ) dτ
0
( j = 1,,2n ) (4.74)
Then, introducing the state-space modal matrix of eigenvectors

V =  v1  v 2n , we can write the 2n Equations 4.74 more compactly
as
t
∫ diag e
λ j (t −τ )
yˆ(t) =  V T q(τ ) dτ (4.75)

0
from which we get, observing that the matrix version of Equation 4.72 is
x(t) = Vyˆ(t),
∫ V diag e
λ j (t −τ )
x(t) =  V T q(τ ) dτ (4.76)

0
T
Finally, recalling that v j =  z j λ j z j  , the 2n × 2n matrices V , V T can be
partitioned as
 Z 
V =  , VT =  ZT diag ( λ j ) ZT  (4.77a)


Z diag (λj ) 

 
where the sizes of Z, ZT and diag ( λ j ) are n × 2n, 2n × n and 2n × 2n, respec-
tively. Also, by further observing that
 f 
V T q =  ZT diag ( λ j ) ZT   T
 = Z f (4.77b)
   0 
T
we can use these last three relations together with x =  u u  in
 
Equation 4.76 to obtain the solution in terms of the original n-dimensional
displacement vector u(t) as
t
∫ Z diag e
λ j (t −τ )
u(t) =  ZT f (τ ) dτ (4.78)

0
Harmonic excitation (and the receptance FRF matrix): For a harmonic

excitation of the form f (t) = F e i ω t , the r.h.s. of Equation 4.712 is q(t) = Q e iω t ,
with Q = [ F 0 ] . Also, by assuming – as mentioned earlier – M ˆ j = 1 for the
T
normalisation of the eigenvectors, on the r.h.s. of Equation 4.73, we have

φ j (t) = vTj Q e i ω t and Equation 4.74 gives (assuming zero initial conditions)
t
iω t
( )
T T
e( ) dτ = v j Q e i ω t − e λ j t = v j Q e (4.79)
∫
T λ jt i ω −λ j τ
yˆ j (t) = v Q e
j
iω − λ j iω − λ j
0
where in the last equality we dropped the e λ jt term because it is transient in

nature and our main interest here is in the steady-state solution. Passing to
matrix notation and recalling that x = Vŷ, we arrive at
 1  T iω t  Z   1  T iω t
x = V diag  V Qe =   diag  Z Fe

 iω − λ j   Zdiag ( λ j )   iω − λ j 
 
(4.80)
where in writing the last expression we took Equations 4.77a and b into
account. Then, the steady-state solution for the original displacement
vector is
2n
 1  T iω t  z m zTm  i ω t
u(t) = Z diag 
 iω − λ j 
Z Fe = ∑
m=1
 iω − λ  F e (4.81)
m
thus implying that the system’s receptance matrix is given by

2n
 1  T
R(ω ) = Z diag 
 iω − λ j 
Z = ∑
m=1
z m zTm
iω − λm
(4.82a)
and its j, kth element is

2n
∑ iω − λ
z j m zkm
Rj k (ω ) = (4.82b)
m
m=1
Finally, by further recalling that for underdamped systems the eigensolutions

occur in complex conjugate pairs, we can rewrite the receptance 4.82b as
n
 z j mzkm z ∗j mzkm
∗

Rj k (ω ) = ∑
m=1
 iω − λ
m
+
iω − λm  ∗ 

n  
z ∗j m zkm
∗
∑  z j m zkm 
= +

(
m=1  ζ mω m + i ω − ω m 1 − ζ m
2
) (
ζ mω m + i ω + ω m 1 − ζ m2 ) 

(4.82c)
where in the last expression we took the relation λm = −ζ mω m + iω m 1 − ζ m2

into account.
Remark 4.12
Especially in the discipline of control theory, many authors use the term
poles for the eigenvalues λm and call as residue for mode m the term z j m zkm .
Also, in modal analysis literature, this residue is often given a symbol of its
own such as, for example, m Aj k or rj k,m.
4.5.1 Another state-space formulation

The state-space formulation given earlier leads to a GEP (of order 2n) in
which the two matrices involved are symmetric. At the end of Section 3.5.2,
on the other hand, we mentioned a different state-space formulation that
leads to the SEP (of order 2n) Av = λ v, where the non-symmetric matrix A
is as shown in Equation 3.96. In this formulation, we recall, one assumes
that the mass matrix M is non-singular, rewrites the equation of motion
as u = − M −1Ku − M −1Cu and puts together this equation with the iden-
T
 Then, defining the 2n-dimensional vector x(t) =  u(t) u (t)  , the
tity u = u.
two equations together are expressed in matrix form as x = Ax, which
in turn gives the SEP Av = λ v when one assumes a solution of the form
x(t) = ve λ t . The solution of this SEP is a set of 2n eigenpairs λ j , v j , which
(for the underdamped case) occur in complex conjugate pairs and where the
eigenvectors are related to the n-dimensional eigenvectors z j of the original
( ) T
QEP λ 2 M + λ C + K z = 0 by v j =  z j λ j z j  . Together, the eigenvectors
v j form the state-space modal matrix V and we have (see Equation A.41 of
Appendix A and Remark 4.13(ii) below) V −1AV = diag ( λ j ).

With this formulation, it is not difficult to show that by introducing the
2n-dimensional forcing vector q =  0 M −1f  , the non-homogeneous
 + Cu + Ku = f(t) gives the state-space equation
problem Mu
x (t) = Ax(t) + q(t) (4.83)
which, defining the set of normal coordinates ŷ as x(t) = Vyˆ(t), becomes

Vyˆ(t) = AVyˆ + q(t) and, consequently, pre-multiplying both sides by V −1,
dyˆ(t)
= V −1AVyˆ + V −1q(t) = diag ( λ j ) yˆ + V −1q(t) (4.84)
dt
which is a set of uncoupled equations formally similar to Equations 4.731

but with the difference that now we have φ j (t) = wTj q, where w j is the jth left
eigenvector of A (see the following Remark 4.13(ii)).
Remark 4.13
i. Since in matrix form, Equations 4.731 read y̂ (t) = diag ( λ j ) yˆ + V T q(t);
the difference with Equation 4.84 is that now in the last term on the
r.h.s., we have V −1 instead of V T . In addition to this, it should be
noticed that here the modal matrix V is the matrix of eigenvectors of
the non-symmetric SEP Av = λ v, while in the preceding Section 4.5,
V is the matrix of eigenvectors of the symmetric GEP 3.94. So, even
if we are using the same symbol V , the fact itself that here and in
Section 4.5, we are considering two different state-space formulations
should make it sufficiently clear that we are not dealing with the same
matrix.
ii. From Appendix A, Section A.4, we recall that for non-symmetric
matrices, we have right and left eigenvectors which satisfy the bi-
orthogonality conditions of Equations A.41. If, with the symbols of
the present section, we denote by V, W the matrices of right and left
eigenvectors of A, respectively, Equations A.41 are written as WT V = I
and WT AV = diag ( λ j ). Moreover, since the first equation implies
WT = V −1, the second equation becomes the relation V −1AV = diag ( λ j )
mentioned earlier and used in Equation 4.84.
In light of Remark 4.13(i), we can now parallel the developments of the

preceding section (see Equations 4.75 and 4.76) and – assuming zero initial
conditions – write the solutions ŷ(t) and x(t) as
t t
∫ diag e λ j (t −τ )  V −1q(τ ) dτ ,

∫ V diag e
λ j (t −τ )
yˆ(t) = x(t) =  V −1q(τ ) dτ

0 0
(4.85)
At this point, we partition the two matrices V, V −1 as
 Vupper   Vupper 
V= = , V −1 =  Vleft
−1 −1
Vright 
 Vlower   Vupper diag ( λ j )   
 
(4.86a)
and observing that
 0 
V −1q =  Vleft
−1 −1
Vright  −1
 = Vright M −1 f (4.86b)
   M −1 f 

(
we can use these relations together with x =  u u  in Equation 4.852
T
)
to obtain the system’s displacement response u(t) to an arbitrary excitation
f(t); that is,
u(t) =
∫V
0
upper diag e λ j (t −τ )  Vright
−1
M −1 f (τ ) dτ (4.87)
If now we consider the harmonic excitation f (t) = F e i ω t ; in the present for-

T
mulation, we have q(t) = Q e iω t with Q =  0 M −1F  . Then, paralleling
again the developments of the preceding section, we use Equations 4.86a
and b to obtain the steady-state displacement response as
 1  −1
u(t) = Vupper diag  Vright M −1 F e i ω t (4.88)
 iω − λ j 
which, for this formulation, is clearly the counterpart of Equation 4.81.

Example 4.4
Applying the state-space formulation of this section to a damped
1-DOF system, Equation 4.83 reads
 u   0 1  u   0 
 =
x = Ax + q ⇒
 + 
  
 u − k m − c m   u   f (t ) m 

(4.89)
and it is not difficult to determine that the homogeneous problem leads

to the eigenvalues and eigevectors
 1 1  −1 1  λ2 −1 
λ1,2 = −ζω n  ω n ζ 2 − 1,
V= ⇒V =  
 λ1 λ2  λ2 − λ1  − λ1 1 
(4.90)
where the symbols ω n ,ζ have their usual meaning and where in

Equation 4.902 the two eigenvectors have already been arranged in the
modal matrix V (from which we obtain the inverse matrix of Equation
4.903; we leave to the reader to check that V −1AV = diag ( λ1 , λ2 )).
Then, since Equations 4.86a and b for this case give
1  −1  1  −f m 
Vupper = [1 1] , −1
Vright =  ,
−1
Vright M −1f =  
λ2 − λ1  1  λ2 − λ1  f m 

it follows that the 1-DOF version of Equation 4.87 is

t
∫ f (τ ){e }
1 λ2 (t − τ )
u (t) = − e λ1(t − τ ) dτ (4.91a)
m ( λ2 − λ1 )
0
At this point, observing that for an under-damped system, the eigenval-

ues are λ1,2 = −ζω n  iω n 1 − ζ 2 = −ζω n  iω d and that, consequently,
λ2 − λ1 = 2 iω d (where, as usual, ω d = ω n 1 − ζ 2 is the damped fre-
quency), Equation 4.91a becomes
u (t) =
1
2 imω d ∫ f (τ ) e −ζω n (t − τ )
{e iω d (t − τ )
}
− e − iω d (t − τ ) dτ
0
t
1
=
mω d ∫ f (τ ) e
0
−ζω n (t − τ )
sin [ω d (t − τ )] dτ (4.91b)
where in writing the last expression we used the Euler formula of

Equation 1.33. Note that except for the initial conditions (that here are
assumed to be zero), Equation 4.91b is, as it must be, Equation 4.4 of
Section 4.2.
On the other hand, for a harmonic excitation of the form f (t) = Fe iω t , the
1-DOF version of Equation 4.88 is
Fe iω t  1 1  Fe iω t  1 
u(t) = − =
m ( λ2 − λ1 )  iω − λ2 iω − λ1 
 m  ( iω − λ2 )( iω − λ1 ) 
(4.92a)
so that some easy calculations (in which we take into account the rela-
tions λ1λ2 = ω n2 and λ1 + λ2 = −2ζω n ) lead to
Fe iω t  1   1 
u(t) = =   Fe iω t (4.92b)
(
m  ω n − ω + 2 iζωω n   k 1 − β 2 + 2iζ β 
2 2
)
where the term within parenthesis in the last expression is, as expected,
the receptance FRF H (ω ) of Equation 4.81. It is now left to the reader as
an exercise to work out the state-space formulation of Section 4.5 with
the two symmetric matrices
 k 
ˆ =  c
M
m 
, Kˆ = 
0

 m 0   0 −m 
4.6 FREQUENCY RESPONSE FUNCTIONS

OF A 2-DOF SYSTEM
As an illustrative example of the preceding discussions, consider a damped

2-DOF (Figure 4.3) system with physical characteristics of mass and stiff-
ness given by
 m1 0   1000 0 
M= = ,
 0 m2   0 500 
 k1 + k2 −k2   10 −5  5
K= =  × 10
 −k2 k2   −5 5 
and viscous damping matrix
 c1 + c2 −c2   1000 −500 

C= = 
 −c2 c2   −500 500 
Since one immediately notices that the damping matrix is proportional to

the stiffness matrix, we have a classically damped system and we know
Figure 4.3 S chematic 2-DOF system.
that the undamped modes uncouple the equations of motion. Solving the
undamped free-vibration problem, we are then led to the following eigenval-
ues and mass-orthonormal eigenvectors (already arranged in matrix form)
 λ1 0   292.9 0 
diag ( λ j ) =  = ,
 0 λ2   0 1707.1 
 0.0224 −0.0224 
P= 
 0.0316 0.0316 
thus implying that the system’s natural frequencies are ω 1 = 17.11 and
ω 1 = 41.32 rad/s. For the modal damping ratios, on the other hand, we use
the relations pTj Cp j = 2ζ j ω j ( j = 1,2) to obtain ζ 1 = 0.0086 and ζ 2 = 0.0207.
With these data of frequency and damping, we can now readily write the
two modal FRFs receptances as
1 1
Hˆ 1(ω ) = , Hˆ 2 (ω ) = (4.93)
292.9 − ω + 0.293 iω
2
1707.1 − ω 2 + 1.707 iω
and use Equation 4.59a to obtain the receptance matrix in physical coordi-
nates. This gives
R(ω ) = Hˆ 1 (ω ) p1 p1T + Hˆ 2 (ω ) p2 pT2
 5.00 × 10−4 7.07 × 10−4 

= Hˆ 1 (ω )  
 7.07 × 10−4 1.00 × 10−3 
 5.00 × 10−4  (4.94)

−7.07 × 10−4
+ Hˆ 2 (ω )  
 −7.07 × 10−4 1.00 × 10−3 
Figures 4.4 and 4.5, respectively, show in graphic form the magnitude of
the receptances R11(ω ) and R12 (ω ) (where R12 (ω ) = R21(ω ) because the matrix
R(ω ) is symmetric).
At this point, although we know that a state-space formulation is not neces-
sary for a proportionally damped system, for the sake of the example, it may be
nonetheless instructive to adopt this type of approach for the system at hand.
Using, for instance, the formulation of Section 4.5.1, we now have the matrix
 0 0 1 0 
   
0 I 0 0 0 1
A= = 
 − M −1
K − M −1C   −1000 500 −1 0.5 
 1000 −1000 −1 −1 
 
-70
R11 - Magnitude (m/N, dB values)
-90
-110
-130
-150
0 10 20 30 40 50 60
frequency (rad/s)
Figure 4.4 R
eceptance R11 – magnitude.
-70
R12 =R 21 - Magnitude (m/N, dB values)
-90
-110
-130
-150
0 10 20 30 40 50 60
frequency (rad/s)
eceptance R12=R 21 – magnitude.

Figure 4.5 R
and we obtain the following eigenvalues and eigenvectors (already in matrix form)
 −0.854 + 41.308 i 0 0 0 
 
 0 −0.854 − 41.308 i 0 0 
diag λ j = 
( ) 
 0 0 −0.146 + 17.114 i 0 
 0 0 0 −0.146 − 17.114 i 
 
 0.0138 + 0.0020i 0.0138 − 0.0020i 0.0232 − 0.0244i 0.0232 + 0.0244i 

 
 −0.0196 − 0.0028i −0.0196 + 0.0028i 0.0328 − 0.0345i 0.0328 + 0.0345i 

V=
 −0.0948 + 0.5693i −0.0948 − 0.5693i 0.4140 + 0.4010i 0.4140 − 0.4010i 
 0.1340 − 0.8052i 0.1340 + 0.8052i 0.5855 + 0.5671i 0.5855 − 0.5671i 
Then, following the developments of Section 4.5.1, we (a) form the matrix Vupper with the first two rows of V, (b) invert V,
−1 −1
(c) form the matrix Vright with the last two columns of V −1 and (d) calculate the product Vright M −1, which results
 −0.0623 − 0.4287 i 0.0881 + 0.6163i 

 
−1 −0.0623 − 0.4287 i 0.0881 − 0.6063i 
Vright M −1 = 
 0.3141 − 0.2991i 0.4443 − 0.4230i 
 0.3141 + 0.2991i 0.4443 + 0.4230i 
 
At this point, we can finally obtain the receptance matrix as

 1  −1
R(ω ) = Vupper diag  Vright M −1 (4.95)
 iω − λ j 
where, just to give an example, the calculations show that the element R11(ω )
of R(ω ) thus obtained is
6.052 × 10−6 i 6.052 × 10−6 i 1.4608 × 10−5 i 1.4608 × 10−5 i

R11(ω ) = − + − +
iω − λ1 iω − λ2 iω − λ3 iω − λ4
(4.96)
with the eigenvalues ordered as in the matrix diag ( λ j ) above. Then, since
λ2 = λ1∗ and λ4 = λ3∗ , it is left to the reader to show that the substitution of their
numerical values into Equation 4.96 gives the (1,1)-element of the matrix
in Equation 4.94. Clearly, the same applies to the FRFs R12 (ω ) = R21(ω ) and
R22 (ω ).
4.7 A FEW FURTHER REMARKS ON FRFS
Having already made some comments on FRFs of 1-DOFs systems in

Section 4.3.3, we can now consider a few facts worthy of mention about
FRFs of MDOF systems.
A first point we want to make is that whether a system is classically
damped or not, the receptance matrix is
( )
−1
R(ω ) = K − ω 2 M + iω C (4.97)
which follows immediately from the equations of motion Mu  + Cu + Ku = f

when one assumes a harmonic excitation and a steady-state harmonic
response of the forms f (t) = F e iω t and u(t) = U e iω t and uses them in the equa-
( )
tions of motion to get −ω 2 M + iω C + K U = F. Then, Equation 4.97 follows
because, by definition, the receptance matrix is the ratio U F. However, the
apparent simplicity of Equation 4.97 can be deceiving because this approach
of direct matrix inversion, although possible in principle, is in general com-
putationally very expensive. In fact, we must invert a n × n matrix – and n
can be quite large in most practical cases – at every frequency of interest.
Moreover, on more physical grounds, it can be argued that this method
definitely provides little understanding and insight on the nature of the
response, on the system’s natural frequencies and on its modal properties
in general.
A second point to make concerns the physical interpretation of the ele-
ments of the receptance matrix. By first noticing that the relation U = R(ω ) F
implies that the response amplitude of the jth DOF is U j = ∑ r Rj r (ω )Fr , it
follows that if all excitation forces but the kth are zero, we have U j = Rj k (ω )Fk
and, consequently,
U 
Rj k (ω ) =  j  (4.98)
 Fk  Fr =0; r ≠k
which in turn shows that Rj k (ω ) gives the displacement response of the jth
DOF when the excitation force is applied at the kth DOF only. From an
experimental point of view, this condition is easy to achieve because we
must only apply the excitation force at the kth DOF and measure the sys-
tem’s response at the jth DOF (clearly, the same applies to the mobility
M j k (ω ) or the accelerance Aj k (ω ) when the motion is expressed or measured
in terms of velocity or acceleration, respectively).
If now, on the other hand, we consider the FRFs in the form of force/motion
ratios – say, for example, the dynamic stiffness matrix K(ω ) = F U – then it
(
is easy to see that the equations of motion lead to K(ω ) = −ω 2 M + iω C + K , )
which implies K = R −1. However, besides the general fact that given a non-
singular matrix A =  a jk , we cannot expect the j, kth element of A −1 to be
given simply by 1 a jk (unless A is diagonal), a more important reason why
in general we have K jk (ω ) ≠ {Rjk (ω )} lies in the physical interpretation of
−1
the elements of the dynamic stiffness matrix. Following the same line of
reasoning as earlier, in fact, we have
 F 
K j k (ω ) =  j  (4.99)
 Uk  Ur =0; r ≠k
which tells us that we have to measure the displacement at the kth DOF
with all the other DOFs fixed at zero displacement. In just a few words, this
means that we are basically dealing with a different system; no longer – as
for Rj k (ω ) – a force-free system except for a single input force, but with a
system clamped at all its DOFs except the one where we measure the dis-
placement. Needless to say, this condition is experimentally very difficult
(if not even impossible in most cases) to achieve.
Finally, another point worthy of notice is that for systems with reason-
ably well-separated modes, only one row or one column of the FRF matrix
(typically accelerance because acceleometers are probably the most used
motion transducers in experiments) is necessary to extract the system’s
natural frequencies, damping ratios and mode shapes. This is particularly
important in the applied field of Experimental Modal Analysis, because
out of the n2 elements of the FRF matrix – of which only n(n + 1) 2 are
independent because the matrix is symmetric for linear systems – only n are
necessary to determine the system’s modal parameters. For more details on
this aspect, we refer the interested reader to specialised texts such as Ewins
(2000), Maia and Silva (1997) or Brandt (2011).
Chapter 5
Vibrations of continuous
systems
5.1 INTRODUCTION
Since at the macroscopic level of our senses matter appears to be ‘continu-

ous’, we are naturally led to consider models in which the physical proper-
ties of (macroscopic) vibrating systems are continuously distributed in a
more or less extended region of space. Also, we expect these models to pro-
vide more accurate solutions with respect to discrete (i.e. finite-DOFs) ones
in light of the fact that we can get better and better approximations of the
‘exact’ solution by making the model finer and finer – that is, by increasing
the number n of DOFs – up to the point that, in the limit of n → ∞, the con-
tinuous model thus obtained will, in principle, give us the ‘exact’ solution.
The passage to the limit, however, comes with a price and has a num-
ber of important consequences. Most of them are of mathematical nature
and bring into play concepts of functional analysis, such as, for example,
infinite-dimensional linear spaces (in particular, Hilbert spaces), Lebesgue
integration, boundary-value problems (BVPs) and linear operators on
Hilbert spaces. For the most part, we will not be concerned with these
aspects (for which we refer the reader to the specialised literature; for
instance, the last three chapters of Davis and Thomson (2000), Daya Reddy
(1998), Debnath and Mikusinski (1999) or Naylor and Sell (1982)) and
will only touch upon them when needed in the course of the discussion.
Instead, our intention is to give more emphasis to the similarities that exist
with finite-DOFs systems despite the more delicate and subtle mathematical
treatment needed for continuous ones.
On a more practical side, it is fair to say that problems involving con-
tinuous systems are in general much more difficult than finite-DOFs ones
and that closed-form solutions are possible only in relatively few cases.
However, although this state of affairs implies, on the one hand, that for
most problems, we must adopt some kind of discretisation (typically, finite
element methods; for this see Bathe (1996) or Petyt (1990)) so that, in the
end, we are back to finite-DOFs models; on the other hand, it does not at all
mean that the study of continuous systems is unnecessary. In fact, it is often
the case that the knowledge and physical insight that this study provides
139
turn out to be of great value in the understanding of vibrating systems’

behaviour, of the various approximation methods and, last but not least, of
the travelling-waves/standing-waves ‘duality’.
5.2 THE FLEXIBLE STRING
A taut flexible string with a uniform mass per unit length µ and under the
stretching action of a uniform tension T0 is the ‘classical’ starting point in
the study of continuous systems. If the string undisturbed position coin-
cides with the x-axis and its (small, we will see shortly what this means)
transverse motion occurs in the y-direction, then its deflected shape at point
x and time t is mathematically described by a field function y(x, t), so that
at a given instant of time, say t = t0, y ( x, t0 ) is a snapshot of the string shape
at that instant, while at a fixed point x = x0, the function y ( x0 , t ) represents
the time-history of the string particle located at x0 .
In order to obtain the equation of motion, we observe that the small-
amplitude Lagrangian density (given by the difference between the kinetic
and potential energy densities) for the string has the functional form
Λ(∂ t y, ∂ x y ), and we have
2 2
µ  ∂y T0  ∂ y 
Λ ( ∂ t y, ∂ x y ) =   −   (5.1a)
2 ∂t 2 ∂x
where the expressions of the energy densities on the r.h.s. can be found in
every book on waves (for example, Billingham and King (2000) or Elmore
and Heald (1985)). Then, by calculating the prescribed derivatives of
Equation 2.68a, we get
∂Λ ∂  ∂Λ  ∂2y ∂  ∂Λ  ∂2y
= 0, = µ , = −T (5.1b)
∂ t  ∂ (∂ t y )  ∂ x  ∂ (∂ x y ) 
0
∂y ∂ t2 ∂ x2
thus leading to the equation of motion
∂2y ∂2y ∂2y 1 ∂2y

−µ + T0 2 = 0 ⇒ = (5.2)
∂t 2
∂x ∂ x2 c 2 ∂ t 2
where, in the second expression, we defined c = T0 µ . As it is probably

well-known to the reader, Equation 5.2 is called the one-dimensional wave
equation, while c is the velocity of small-amplitude transverse waves on the
string (which is different from the velocity ∂ t y(x, t) with which the string
particles move transversely to the string).
If now, for the moment, we pose no restrictions on the length of the
string, Example B.6 of Appendix B shows that the solution of the initial
Vibrations of continuous systems 141
value problem (IVP) given by Equation 5.2 supplemented with the initial
(i.e. at t = 0) conditions y(x,0) = u(x) and ∂ t y(x,0) = w(x) is the so-called
d’Alembert solution of Equation B.67. Also, it can be shown that this same
solution can be obtained without the aid of Laplace and Fourier transforms
(which we used in Example B.6) by imposing the initial conditions (ICs) at
t = 0 to a general solution of the form
y ( x, t ) = f ( x − ct ) + g ( x + ct ) (5.3)
where f and g are two unrelated arbitrary functions that represent, respec-
tively, a waveform (or wave profile or progressive wave) propagating without
the change of shape in the positive x-direction and a waveform propagat-
ing without the change of shape in the negative x-direction. Moreover, if
the two wave profiles have a finite spatial extension, the fact that the wave
Equation 5.2 is linear also tells us that when the two waves come together,
they simply ‘pass through’ one another and reappear without distortion.
Remark 5.1
i. More precisely, the two functions f , g are not completely arbitrary

because they have to be twice-differentiable with respect to their
arguments in order to satisfy the wave Equation 5.2. Under this con-
dition, it can be easily checked that they are solutions of the wave
equation. In fact if, for example, y = f (u) and u = x − ct, then we have
∂ x2 y = f ′′(u) and ∂ t2 y = c2 f ′′(u), from which it follows that f (x − ct) sat-
isfies Equation 5.2 identically. Clearly, the same applies to the func-
tion g(x + ct).
ii. The small-amplitude approximation that leads to the linear equation
of motion 5.2 implies the requirement ∂ x y << 1, which, in physical
terms, means that the slope of the string must be small. The approxi-
mation also implies ∂ t y << c, meaning that the transverse particle
velocity must be small compared with the wave propagation velocity.
In fact, if for example we again consider the rightward-propagating
waveform y = f (u), we have ∂ x y = f ′(u) and the requirement ∂ x y << 1
becomes f ′(u) << 1. But then the condition ∂ t y << c is a direct conse-
quence of the relation ∂ t y = −cf ′(u).
Turning to the energy carried by a progressive wave it is easy to see that

if, for instance, y = f (u) with u = x − ct (and assuming that f is a local-
ised pulse or decays sufficiently rapidly as x → ±∞), the expressions of the
∞
kinetic and potential energies given above lead to Ek = µ c2 2 ( ) ∫ (f ′)
−∞
2
dx
∞
and Ep = (T0 2)
∫ −∞
( f ′ )2 dx, which in turn imply Ek = Ep when one observes
that c2 = T0 µ. The total energy E = Ek + Ep, therefore, is equally divided

into kinetic and potential forms. Note that the same applies to the leftward-
propagating waveform g, but it does not, in general, apply to the solution
y = f + g.
5.2.1 Sinusoidal waveforms and standing waves

Out of the potentially infinite variety of functions permitted as solutions of
the wave equation, the fact that sinusoidal waves play a fundamental role
comes as no surprise. In mathematical form, one such wave travelling in the
positive x-direction can be written as
 2π 
y(x, t) = A sin  (x − ct)  = A sin (kx − ω t) (5.4)
 λ 
where A is the amplitude, k = 2π λ is called the wavenumber, ω = 2π c λ = kc

is the angular frequency and λ is the wavelength. The function 5.4 is peri-
odic in both space and time; in space (for a fixed instant of time t = t0), it
repeats itself at any two points x1 , x2 such that x2 − x1 = λ, while in time (for
a fixed position x = x0 ), it repeats itself at any two instants t1 , t2 such that
T ≡ t2 − t1 = λ c, thus implying that ν = 1 T is the frequency of the wave.
Other useful relations are easily obtained, and we have
2π λ ω
ω = 2πν = , c = λν = = (5.5)
T T k
where it should also be noticed that the relation k = ω c is necessary if we

want the sinusoidal waveform 5.4 to satisfy the wave equation.
Remark 5.2
i. In light of the notation used in preceding chapters, the symbols k and

λ used here should not be confused with a stiffness coefficient and an
eigenvalue, respectively. These symbols for the wavenumber and the
wavelength are so widely used that it would be pointless to adopt a
different notation.
ii. The term (kx − ω t) is often referred to as the phase of the wave. The
fact that the wave 5.4 moves to the right can be deduced by noticing
that increasing values of x are required to maintain the phase constant
as time passes.
iii. For a sinusoidal wave of the form 5.4, it is not difficult to show that
the small-slopes approximation ∂ x y << 1 of Remark 5.1(ii) translates
into the restriction A << 1 k = λ 2π, meaning that the wave amplitude
must be quite small in comparison to its wavelength.
iv. Even with progressive sinusoidal waves, the exponential form is

often very convenient. So, for example, the wave of Equation 5.4
can be regarded as the imaginary part of the complex quantity
y(x, t) = A e i (k x −ω t ), although, as stated in previous chapters, the real-
part convention works just as well as long as consistency is main-
tained throughout the calculations.
v. It is left to the reader as an exercise to determine that for a sinusoi-
dal wave, the kinetic and potential energy densities averaged over one
period are equal and that they are given by T0k2 A2 4 or µ ω 2 A2 4
(which is the same because T0k2 = µ ω 2).
If now we consider the special case of two sinusoidal waves with equal
amplitude and velocity travelling on the string in opposite directions,
we have the solution of the wave equation given by the superposition
y(x, t) = A sin ( kx − ω t ) + A sin ( kx + ω t ). Using well-known trigonometric
relations, however, this can be transformed into
y(x, t) = 2A sin(kx) cos(ω t) (5.6)
which is a solution in its own right but no longer with the space and time
variables associated in the ‘propagating form’ x ± ct . Equation 5.6, in fact,
represents a stationary or standing wave, that is, a wave profile that does
not advance along the string but in which all the elements of the string
oscillate in phase (with frequency ω ) and where the particle at x = x0 sim-
ply vibrates back and forth with amplitude 2A sin kx0. Moreover, the par-
ticles at the positions such that sin kx = 0 do not move at all and are called
nodes; a few easy calculations show that they are spaced half-wavelengths
apart, with antinodes – i.e. particles of maximum oscillation amplitude –
spaced halfway between the nodes. Physically, therefore, it can be said that
standing waves are the result of constructive and destructive interference
between rightward- and leftward-propagating waves.
5.2.2 Finite strings: the presence of

boundaries and the free vibration
The fact that a real string has finite length and must end somewhere affects
its motion by imposing appropriate boundary conditions (BCs), which – as
opposed to initial conditions (ICs) that must be satisfied at a given time –
must hold for all time. So, for instance, if the string is rigidly fixed at x = 0,
then we must have y(0, t) = 0 for all t ; since this, on the one hand, implies
that a single waveform of the type f (x − ct) or g(x + ct) cannot possibly sat-
isfy this requirement; on the other, it shows that a solution of the form 5.3
can do it if f (x − ct) = − g(x + ct). Physically, this means that an incoming
(i.e. leftward-propagating, since we assume the string to extend in the region
x > 0) localised disturbance is reflected back into the string by the fixed
boundary in the form of an outgoing (i.e. rightward-propagating) distur-
bance which is the exact replica of the original wave except for being upside
down. This ‘wave-reversing’ effect without change of shape applies to (and
is characteristic of) the fixed boundary, but clearly other BCs – such as, for
example, ky(0, t) = T0 {∂ x y(0, t)}, which corresponds to an elastic boundary
with stiffness k – will give different results, with, in some cases, even con-
siderable distortion of the reflected pulse with respect to the incoming one.
Given these preliminary considerations, our interest now lies in the
free vibration of a uniform string of length L fixed at both ends (x = 0
and x = L). Mathematically, the problem is expressed by the homogeneous
wave equation
∂2y 1 ∂2y
− = 0 (5.7a)
∂ x2 c 2 ∂ t 2
supplemented by two BCs and two ICs, that is,
y(0, t) = 0, y(L, t) = 0
(5.7b)
y(x,0) = y0 (x), ∂ t y(x,0) = y0 (x)
respectively. In order to tackle the problem, we can use the standard and
time-honoured method of separation of variables, which consists in looking
for a solution of the form y(x, t) = u(x)g(t) and substitute it in Equation 5.7a
to obtain c2u′′ u = g g. But since a function of x alone (the l.h.s.) can be
equal to a function of t alone (the r.h.s) only if both functions are equal to
a constant, we denote this constant by −ω 2 and obtain the two ordinary
differential equations
d 2u d2g
+ k2 u = 0, + ω 2 g = 0 (5.8)
dx2 dt 2
(with, we recall, k= ω c) where the BCs are ‘transferred’ to the spatial part
u(x) and become u(0) = u(L) = 0. Then, recalling that the solution of the
spatial Equation 5.81 is u(x) = C sin kx + D cos kx, enforcing the BCs leads
to D = 0 and to the frequency (or characteristic) equation sin kL = 0, which
in turn implies kL = nπ ( n = 1, 2,). So, it turns out that the only allowed
wavenumbers are kn = nπ L, meaning that the string can only vibrate har-
monically at the frequencies
nπ c n π T0
ωn = = ( n = 1, 2,) (5.9a)
L L µ
where to each frequency there corresponds a vibrational mode shape given by
un (x) = Cn sin kn x = Cn sin ( nπ x L ) ( n = 1, 2,) (5.9b)

and it is not difficult to see that (a) the odd modes (n = 1,3, 5,) are sym-
metric with respect to the midpoint of the string, (b) the even modes
(n = 2, 4,6,) are antisymmetric, and (c) the nth mode has n − 1 nodes (end
points excluded).
Combining these results with the solution g(t) = A cos ω t + B sin ω t of the
time-Equation 5.82 , we can now write yn (x, t) = ( An cos ω nt + Bn sin ω nt ) sin kn x
for each value of n (the constant Cn has been absorbed in the constants
An , Bn ) and – owing to the linearity of the wave equation – express the
general solution in the form of the series
∞
y(x, t) = ∑ ( A cos ω t + B sin ω t ) sin k x (5.10)

n =1
n n n n n
where the constants An , Bn are determined by the ICs. Since in light of

Equation 5.10, the ICs at t = 0 are given by the Fourier series
∞ ∞
y0 (x) = ∑n =1
An sin kn x, y0 (x) = ∑ ω B sin k x (5.11)
n =1
n n n
the constants An , Bn can be determined by first recalling the ‘orthogonality’

property of the sine functions (over the interval (0, L))
L
L

∫ sin k x sin k x dx = 2 δ
0
n m nm (5.12)
and then by multiplying both sides of the ICs by sin km x and integrating over
the string length. By so doing, the use of Equation 5.12 in Equations 5.111
and 5.112 , respectively, leads to (renaming the index as appropriate)
L L
2 2
An =
L ∫ y (x)sin k x dx,
0
0 n Bn =
Lω n ∫ y (x)sin k x dx (5.13)
0
0 n
Together, Equations 5.10 and 5.13 provide the desired solution, from which
we can clearly see the role played by BSs and ICs; the BCs determine the
natural frequencies and the mode shapes – that is, the eigenvalues and eigen-
functions in more mathematical terminology – while the ICs determine the
contribution of each mode to the string vibration. The reason for the terms
‘eigenvalues’ and ‘eigenfunctions’ becomes clearer if we define k2 = λ (here
λ is not a wavelength) and rewrite the spatial part of the problem 5.7 – i.e.
Equation 5.81 and the corresponding BCs – as
− u′′ = λ u
(5.14)
u(0) = 0, u(L) = 0
where it is now evident that Equation 5.141 is an eigenvalue problem for the
differential operator − d 2 dx2.
Together, Equations 5.14 define the so-called BVP belonging to a par-
ticular class that mathematicians call Sturm-Liouville problems (SLps).
We will have more to say about this in Section 5.4, but for time being, a
first observation is that the class of SLps is quite broad and covers many
important problems of mathematical physics. A second observation con-
cerns the term ‘orthogonality’ used in connection with Equation 5.12. In
fact, since it can be shown that for any two functions f , g belonging to the
(real) linear space of twice-differentiable functions defined on the interval
L
0 ≤ x ≤ L, the expression
∫
0
f (x)g(x) dx defines a legitimate inner product
(where ‘legitimate’ means that it satisfies all the properties given in Section
A.2.2 of Appendix A), we can write
L
f g ≡
∫ f (x)g(x) dx (5.15)
0
and say that the two functions are orthogonal if f g = 0. This is indeed
the case for the eigenfunctions (of the BVP 5.14) sin kn x and sin km x (n ≠ m).
In this light, moreover, note that Equations 5.13 can be expressed as the
inner products
2 2
An = sin kn x y0 (x) , Bn = sin kn x y0 (x) (5.16)
L Lω n
Remark 5.3
i. For each term of the solution 5.10, we have
2An cos ω nt sin kn x = An sin ( kn x + ω nt ) + An sin ( kn x − ω nt )

2Bn sin ω nt sin kn x = Bn cos ( kn x − ω nt ) − Bn cos ( kn x + ω nt )
which, once again, shows the ‘duality’ between travelling and

standing waves.
ii. The reader is invited to show that the total energy ‘stored’ in each
mode is
µ ω 2L 2
En =
4
(
An + Bn2 (5.17) )
and that, in addition, by virtue of the modes orthogonality
∑
∞
(Equation 5.12), the total energy of the string is given by E = En
n =1
meaning that each mode vibrates with its own amount of energy
and there is no energy exchange between modes.
iii. The homogeneous nature of the BVP 5.14 determines the eigenfunc-
tions un (x) to within an arbitrary scaling (or normalisation) factor.
Fixing this factor by some convention gives the normalised eigenfunc-
tions (which we will denote by φn (x) in order to distinguish them
from their ‘unnormalised’ counterparts un (x)). So, for example, in the
case of the string, one possibility is to normalise the eigenfunctions
so that they satisfy the ortonormality relations φn φm = δ n m. With
this convention, it is then not difficult to see that, in light of Equation
5.12, the normalised eigenfunctions are φn (x) = ( 2 L )sin kn x .
Example 5.1
If the string is set into motion by pulling it aside by an amount a (a << L)
at the midpoint and then releasing it at t = 0, the initial velocity is zero
(thus implying that all the Bn coefficients of Equations 5.132 are zero),
while the initial shape is
 2ax L 0≤x≤L 2
y0 (x) =  (5.18)
2a(L − x) L L 2≤x≤L

Then, using Equation 5.131, the coefficients An are given by
L 2 L 
4a 
An = 2
L 
0
∫ L2
∫
x sin kn x dx + ( L − x ) sin kn x dx  (5.19)


and we leave it to the reader to determine (a) that An = 0 when n is even

and (b) that the final result is the superposition of odd harmonics
∑ (2(n−1)+ 1)
n
8a
y(x, t) = sin ( k2n+1 x ) cos (ω 2n+1t )
π2 2
n =0
8a  πx 1 3π x  (5.20)
=
π2 sin L cos ω 1t − 32 sin L x cos 3ω 1t + 
where in writing the second expression we took into account the rela-
tions kn = nπ L and ω n = nω 1. Note that the absence of even-order
modes is not at all unexpected; these modes, in fact, have a node at
x = L 2, which is precisely the point where we displaced the string
when we applied the IC 5.18.
5.3 FREE LONGITUDINAL AND

TORSIONAL VIBRATION OF BARS
Consider a slender bar of length L with uniform cross-sectional area A and

uniform density ρ = µ A, where µ is the mass per unit length. If the bar
vibration is in the longitudinal direction – in which case the term rod is
also frequently used – the axial displacement at point x and time t is given
by the field function y(x, t). If, on the other hand, the vibration is rotational
around its longitudinal axis, the relevant variable is the angle of twist θ (x, t)
and it is quite common to refer to this system in torsional motion – typically
with circular cross-section, as we are assuming here – as a shaft. In both
cases, the governing theory is based on strength-of-materials considerations
and on certain assumptions on the kinematics of deformation (for exam-
ple, ‘plane cross-sections remain plane during the motion’ for longitudinal
vibrations, or ‘each transverse section remains in its own plane and rotates
about its centre’ for torsional vibrations). We have, in just a few words, two
different physical phenomena, which in turn also seem quite different from
the transverse vibrations of a string.
In spite of this, however, the point of interest for us here and in the fol-
lowing section is that the mathematical treatment is quite similar because
the rod and shaft equations of motion turn out to be one-dimensional wave
equations. In fact, we have, respectively,
∂2y ρ ∂2y ∂ 2θ ρ ∂ 2θ
= , = (5.21)
∂ x2 E ∂ t 2 ∂ x2 G ∂ t 2
where E is Young’s modulus and G is the shear modulus for the material,
and a comparison of Equations 5.21 with the string Equation 5.2 shows
that cl = E ρ and cs = G ρ are the propagation velocities of longitudinal/
shear waves along the rod/shaft. In light of this close mathematical analogy,
our discussion here will be limited to the rod longitudinal vibration because
it is evident that by simply using the appropriate symbols and physical
parameters for the case at hand, all the considerations and results obtained
for the string apply equally well to the rod and the shaft.
So, if now we proceed by separating the variables and looking for a solu-
tion of the form y(x, t) = u(x)g(t), for the spatial part u(x) of Equation 5.211,
we obtain the ordinary differential equation
d 2u(x)
+ γ 2u(x) = 0 (5.22)
dx2
where γ 2 = ρω 2 E and where we called −ω 2 the separation constant.

Just like in the string case, note that Equation 5.22 is basically an eigen-
value problem for the operator − d 2 dx2. As we already know, the solution
of this equation is u(x) = C sin γ x + D cos γ x , and we obtain the system’s
eigenvalues and eigenfunctions by enforcing the BCs. If, for example, the
rod is clamped at both ends then we get D = 0 and the frequency equation
sin γ L = 0, from which it follows γ n = nπ L ( n = 1, 2,) and consequently
the eigenpairs
nπ E  nπ x 
ωn = , un (x) = Cn sin  (n = 1, 2,) (5.23)
L ρ  L 
If, on the other hand, the rod is clamped at the end point x = 0 and free
at x = L, the BCs are u(0) = 0 (a geometric BC if we recall the discussion
in Section 2.5.1 of Chapter 2) and u′(L) = 0 (a natural BC). With these
BCs, the sinusoidal solution for u(x) gives D = 0 and the frequency equa-
tion cos γ L = 0, from which it follows γ n L = (2n − 1)π 2. Consequently, the
eigenpairs of the clamped-free rod are
(2n − 1)π E  (2n − 1)π x 

ω n = , un (x) = Cn sin   (n = 1, 2,) (5.24)
2L ρ  2L
Finally, another typical boundary configuration is the free-free rod. In this

case, we have two natural BCs, namely, u′(0) = u′(L) = 0, and it is easy to see
that they give C = 0 and the frequency equation sin γ L = 0. Therefore, we
have the eigenpairs
nπ E  nπ x 
ωn = , un (x) = Cn cos  (n = 0,1, 2,) (5.25)
L ρ  L 
where in this case, however, the eigenvalue zero is acceptable because its
corresponding eigenfunction is not identically zero. In fact, inserting γ = 0
in Equation 5.22 gives u′′ = 0, and consequently, by integrating twice,
u(x) = c1x + c2 , where c1 , c2 are two constants of integration. Enforcing the
BCs now leads to u0 (x) = c2 and not, as in the previous cases, u0 (x) = 0 (which
means no motion at all, and this is why zero is not an acceptable eigenvalue
in those cases). The constant eigenfunction u0 = c2 of this free-free case is
a rigid-body mode in which the rod moves as a whole, without any inter-
nal stress or strain. Just like for finite-DOFs systems, therefore, rigid-body
modes are characteristic of unrestrained systems.
Remark 5.4
i. In the wave equation 5.211, it is assumed that the physical quantities

µ , A are uniform along the length of the rod. When this is not the case
and they depend on x, we can recall Example 2.4 in Chapter 2 to
determine the equation of motion
∂  ∂y  ∂2 y d  du 
 EA(x)  = ρ A(x) 2 ⇒ − 

EA(x)  = ω 2 µ(x)u (5.26)
∂x ∂x ∂t dx dx 
where Equation 5.262 is relative to the spatial part u(x) of the
function y(x, t) after having expressed it in the separated-variables
form y(x, t) = u(x)w(t) and having called −ω 2 the separation con-
stant (or, equivalently for our purposes, after having assumed a
solution of the form y(x, t) = u(x) e i ωt and having substituted it in
Equation 5.261).
ii. For the shaft, the torsional stiffness GJ(x) (where J is the area polar
moment of inertia) replaces the longitudinal stiffness EA(x) appear-
ing on the l.h.s., while on the r.h.s., we have the mass moment of
inertia per unit length I (x) = ρ J(x) instead of the mass per unit length
µ(x) = ρ A(x). Clearly, in this case, u(x) is understood as the spatial
part of the function θ (x, t).
5.4 A SHORT MATHEMATICAL INTERLUDE:

STURM–LIOUVILLE PROBLEMS
Let [ a, b] be a finite interval of the real line and let p(x), p′(x) ≡ dp dx , q(x)
and w(x) be given real-valued functions that are continuous on [ a, b], with,
in addition, the requirements p(x) > 0 and w(x) > 0 on [ a, b]. Then, a regular
Sturm-Liouville problem (SL problem for short) is a one-dimensional BVP
of the form
−(pu′)′ + qu = λ wu
(5.27)
Ba u ≡ α 1u(a) + β1u′(a) = 0, Bb u ≡ α 2u(b) + β2u′(b) = 0
where Equation 5.271 is a second-order differential equation while Equations

5.272 and 5.273 are homogeneous (and separated, because each involves
only one of the boundary points) BCs at x = a and x = b, respectively, with
the condition that the real coefficients α , β must be such that α 1 + β1 ≠ 0
and α 2 + β2 ≠ 0. Moreover, if one defines the Sturm-Liouville operator as
d  d 
L=−  p(x)  + q(x) (5.28)
dx  dx 
then Equation 5.271 is the eigenvalue equation Lu = λ wu for the operator

L, where in this context the function w(x) – which, in some cases, can be
the constant function w(x) = 1 for all x in [ a, b] – is called weight function.
The values of λ for which the SL problem has a non-trivial solution are the
eigenvalues and the corresponding solutions are the eigenfunctions. Here
we assume without mathematical proof – but strongly supported by physi-
cal evidence on the vibrations of strings, rods and shafts – that such eigen-
pairs exist.
In a previous section, we mentioned without much explanation that the
class of SL problems covers many important problems of mathematical phys-
ics. This is basically due to the fact that any second-order ordinary differential
equation of the general form A(x) u′′(x) + B(x) u′(x) + [C(x) + λ D(x)] u(x) = 0
(with A(x) ≠ 0) can be converted into the SL form 5.271 by introduc-
ing the factor p(x) = exp {∫ [ }
B(x) A(x)] dx , multiplying the equation by
p(x) A(x) and then defining the two functions q(x) = − [ p(x)C(x)] A(x) and
w(x) = [ p(x)D(x)] A(x).
In light of this possibility and of the fact that for flexible strings, rods and
shafts, the separation-of-variables technique leads to second-order ordinary
differential equations (recall Equations 5.81, 5.14, 5.22 and 5.262); the point
of this section – within the limits of a short ‘interlude’ – is to show that the
mathematical theory provides us with a number of noteworthy results on
the eigenvalues and eigenfunctions of SL problems.
We start by establishing the relation known as Lagrange’s identity: for
any twice-differentiable functions u, v defined on [ a, b], we have
d
u (Lv) − (Lu) v =  p ( u′v − uv′ )  (5.29)
dx 
which follows from the easy-to-check chain of relations
u (Lv) − (Lu) v = u [ −(pv ′)′ + qv ] − [ −(pu′)′ + qu ] v = − u(pv ′)′ + (pu′)′ v
= − p′uv ′ − puv ′′ + p′u′v + pu′′v = p′ ( u′v − uv ′ )

+ p ( u′′v − uv ′′ ) = [ p (u′v − uv ′)]′
If now we integrate Equation 5.29 on [ a, b] and recall from Equation 5.15

b
that the integral
∫ a
f g dx defines a legitimate inner product, we get
u Lv − Lu v = [ p (u′v − uv′)] ba (5.30)
which is known as Green’s formula. Equation 5.30 implies that when the
boundary terms on the r.h.s. are zero – as is the case for a regular SL problem
(but, we note in passing, not only in this case) – we have u Lv = Lu v ,

meaning that the SL operator is self-adjoint (see also the following Remark
5.5(ii)) or, depending on the author, hermitian or symmetric.
With the appropriate BCs that make L self-adjoint, the first two impor-
tant results are that the eigenvalues of the SL problem are real and that
any two eigenfunctions φk (x), φ j (x) corresponding to different eigenvalues
are orthogonal with respect to the weight function w(x) (or w-orthogonal),
that is, such that
b
φk φ j w
≡
∫ φ (x)φ (x)w(x) dx = 0 (5.31)
a
k j
In order to show that the eigenvalues are real, let λ , φ (x) be a possibly complex
eigenpair and let us write λ = r + is and φ (x) = u(x) + iv(x). Then, the prop-
erty of self-adjointness φ Lφ = Lφ φ implies φ λ wφ = λ wφ φ because
Lφ = λ wφ . If now we generalise the inner product 5.15 to the complex case
b
as
∫ a
f *g dx and recall that w(x) is real, the equation φ λ wφ = λ wφ φ in
explicit form reads
b b b
λ
∫φ
2
wdx = λ ∗
∫φ
2
wdx ⇒ ( λ − λ ) ∫ u
∗ 2
+ v 2  w dx = 0
a a a
which, since the last integral is certainly positive, gives λ − λ ∗ = 2is = 0 and
tells us that λ is real. In this respect, moreover, it can also be shown that the
eigenfunctions can always be chosen to be real.
Passing to w-orthogonality, let φk , φ j be two eigenfunctions corresponding
to the eigenvalues λk , λ j with λk ≠ λ j, so that Lφk = λkwφk and Lφ j = λ j wφ j.
Multiplying the first relation by φ j , the second by φk , integrating over the
b
interval and subtracting the two results leads to ( λk − λ j )
∫ a
φk φ j wdx = 0,
which in turn implies Equation 5.31 because λk ≠ λ j.
Remark 5.5
i. The fact that – as pointed out above – for any two functions f , g continu-
b b
ous on [ a, b], the expressions f g ≡
∫ a
f *g dx and f g w
≡
∫
a
f *gw dx
(note that here we consider the possibility of complex functions; for
real functions, the asterisk of complex conjugation can obviously be
ignored) define two legitimate inner products implies that we can
define the norms induced by these inner products as
f = f f , f w
= f f w
respectively.
ii. In order to call an operator self-adjoint, one should first introduce its
adjoint and see if the two operators are equal (hence the term self-
adjoint). For our purposes, however, this is not strictly necessary and
the property u Lv = Lu v suffices. In this respect, note the evident
analogy with the relation u Av = Au v that holds in the finite-
dimensional case for a symmetric matrix A (also see the final part of
Section A.3 of Appendix A).
Another property of the eigenvalues is that they are positive. More precisely,
we have the following proposition: if q ≥ 0 and α 1β1 ≤ 0, α 2β2 ≥ 0 then all
the eigenvalues are positive. The only exception is when α 1 = 0, α 2 = 0 and
q = 0, a case in which the regular SL problem becomes
−( pu′ )′ = λ wu
(5.32)
u′(a) = 0, u′(b) = 0
and zero is an eigenvalue corresponding to a constant eigenfunction (all

the other eigenvalues are positive).
The positivity of the eigenvalues can be shown as follows: let λ , φ be an
eingenpair, multiply the equation −(pφ ′)′ + qφ − λ wφ = 0 by φ and then inte-
grate on the interval [ a, b]. By so doing, an integration by parts leads to
−[ pφφ ′ ]ba +
∫ ( pφ′ 2
)
+ qφ 2 dx
λ= b
a
(5.33)
∫ φ w dx
a
2
where, explicitly, the boundary term is
α2 α
− p(b)φ (b)φ ′(b) + p(a)φ (a)φ ′(a) = p(b)φ 2 (b) − 1 p(a)φ 2 (a)
β2 β1
and the second expression – which follows directly from the BCs – shows
that, under our assumptions on the α , β -coefficients, the boundary term is
non-negative (note that if β1 and β2 are zero, the BCs give φ (a) = φ (b) = 0 and
the boundary term is zero). Then, since the two integrals in Equation 5.33
are positive, it follows that the eigenvalue λ is also positive.
And this is not all, because it can be shown that
a. the eigenvalues form a countable set that can be ordered as

λ1 < λ2 <  < λn < ,
b. λn → ∞ as n → ∞ .
c. each eigenvalue of the regular problem 5.27 is simple – or non-degen-
erate, meaning that it has both algebraic and geometric multiplicity
1 – and corresponds to only one eigenfunction. The eigenfunctions, in
turn, are determined to within a multiplicative constant, which is often
conveniently chosen so that they satisfy the w-orthonormality condi-
tion φk φ j w = δ kj . Note that this orthonormality condition implies
φk Lφ j = λ j δ kj because φk Lφ j = λ j φk wφ j = λ j φk φ j w = λ j δ kj .
Another key property of SL eigenfunctions is that they form a ‘basis’ – that

is, a complete orthonormal system – in the linear space of functions defined
on [ a, b] and satisfying certain regularity conditions. More specifically, for
our purposes, two important propositions are as follows:
[Boyce and Di Prima (2005)]: Let f (x) be a function defined on [ a, b] and
let φ1(x), φ2 (x), , φn (x), be the w-normalised eigenfunctions of the regular
SL problem 5.27. If f , f ′ are piecewise continuous on [ a, b] then f can be
expressed as
∞ b
f (x) = ∑
n=1
cn φn (x) where cn = φ n f w
=
∫ φ (x)f (x)w(x) dx (5.34)
n
a
and cn are called generalised Fourier coefficients of the series expansion

5.341, which converges to [ f (x+) + f (x−)] 2 at any point in the open inter-
val (a, b) (meaning that it converges to f (x) where f is continuous and to the
mean of the left- and right-hand limits at each point of discontinuity).
[Guenther and Lee (2019)]: For each continuous function f on [ a, b] the
unique solution u(x) to the regular SL problem Lu = f ; Ba u = 0, Bb u = 0 can
be expressed as
∞
u(x) = ∑φ
n =1
n u w
φn (x) (5.35)
where the series is absolutely and uniformly convergent on [ a, b].

Finally, for the last property we consider in this section, we go back to
Equation 5.33 and note that the ratio on the r.h.s. is known as Rayleigh
quotient. More generally, since this quotient is well-defined also if we use
a generic function u(x) in it – that is, not necessarily an eigenfunction – the
definition of Rayleigh quotient is
− [ puu′ ]a +
∫  pu′
b 2
+ qu2  dx
u Lu
R [ u] ≡ = b
a
(5.36)
uuw
∫ u w dx
a
2
where the second expression follows from the first by writing in explicit
form the two inner products and performing an integration by parts in the
numerator. Also, note that the denominator can equivalently be written as
2
u w.
The property of interest is that the lowest eigenvalue λ1 of the regu-
lar SL problem is the minimum value of R for all continuous functions
u ≠ 0 that satisfy the BCs and are such that (pu′)′ is continuous on [ a, b].
Mathematically, therefore, we can write
u Lu
λ1 = min R [ u] = min (5.37)
uuw
to which it must be added that the minimum value is achieved if and only
if u(x) is the eigenfunction corresponding to λ1.
For a proof of this property, we can follow this line of reasoning: by
preliminarily observing that for any continuous function u(x) satisfying the
BCs, the eigenfunction expansion 5.35 implies the two results:
∑ ∑λ
2 2
u w
= uu w
= φn u w
, Lu = n φn u w
wφn (5.38)
n n
we can write a chain of relations that leads to the conclusion stated by the
theorem, namely, that λ1 ≤ R [ u] and λ1 = R [ u] if and only if u = φ1. The chain
of relations is
u Lu = ∑φ
n
n u w
φn Lu = ∑φ
n
n u w
Lφn u = ∑λ
n, m
m φn u w
φm u w
φm φn w
∑λ ∑λ ∑φ
2 2
= n φn u w
φn u w
= n φn u w
≥ λ1 n u w
= λ1 u u w
n n n
where we used the expansion 5.35 in writing the first equality, the self-
adjointness of L in the second, the expansion 5.382 in the third and the
orthogonality of the eigenfunctions in the fourth. Then, the inequality fol-
lows because we know from previous results that λn ≥ λ1 (and the last equal-
ity follows from Equation 5.381).
Remark 5.6
i. Just like Hamilton’s principle of Section 2.5, the minimisation prop-

erty of the Rayleigh quotient belongs rightfully to the branch of math-
ematics known as calculus of variations. In this regard, in fact, it may
be worth mentioning that the SL problem 5.27 can be formulated as
the variational problem of finding a twice-differentiable extremal of
the functional J subject to the constraint K = 1, where
b
α1 α
J[ u] =
∫ {p(u′)
a
2
}
+ qu2 dx −
β1
p(a)u2 (a) + 2 p(b)u2 (b)
β2
(5.39)
b
K[ u] =
∫ u wdx
a
2
and where the theory tells us (see, for example, Collins (2006))
that the constrained variational problem 5.39 is equivalent to the
unconstrained problem of finding the extremals of the functional
I = J − λ K, with λ playing the role of a Lagrange multiplier.
ii. In light of Equation 5.37, it is natural to ask if also the other eigenvalues
satisfy some kind of minimisation property. The answer is a ffirmative
and it can be shown that the nth eigenvalue λn is the minimum of
R[ u] over all twice-differentiable functions that are orthogonal to the
first n − 1 eigenfunctions. The minimising function in this case is φn.
In this respect, it should also be pointed out that these minimisation
properties play an important role in approximate methods used to
numerically evaluate the lowest-order eigenvalues of vibrating systems
for which an analytical solution is very difficult or even impracticable.
5.5 A TWO-DIMENSIONAL SYSTEM: FREE

VIBRATION OF A FLEXIBLE MEMBRANE
Basically, a stretched flexible membrane is the two-dimensional counter-

part of the flexible string, where flexibility means that restoring forces in
membranes arise from the in-plane tensile forces and that there is no resis-
tance in bending and shear. In order to derive the small-amplitude equation
of transverse (that is, in the z-direction if the undisturbed membrane is
flat in the xy-plane) motion, the simplest assumptions are that the mem-
brane has a uniform density per unit area σ and that it is subjected to a
uniform in-plane tensile force per unit length T. This force is the action
that one part of the membrane exerts on the adjacent part across any line
segment (regardless of its orientation, hence the term ‘uniform’) lying in the
membrane; the magnitude of T, moreover, is assumed to remain practically
constant under the small deflections that displace the membrane from its
equilibrium position.
Denoting by w(x, y, t) the field function that describes the membrane
motion at the point x, y and time t , the small-amplitude Lagrangian density
{ 2

2
of motion is obtained by using Equation 2.69; this gives

 }
(per unit area) is Λ = 2−1 σ (∂ t w ) − T (∂ x w ) + (∂ y w )  , and the equation
2
 ∂2w ∂2w  ∂2w 1 ∂2w

T + − σ =0 ⇒ ∇2w − = 0 (5.40)
 ∂ x2 ∂ y 2  ∂ t2 c2 ∂ t 2
where in the second expression, we defined c = T σ – which, in analogy

with the string case, is the wave velocity for transverse waves on the mem-
brane – and introduced the well-known Laplacian operator ∇2 (or simply
the Laplacian, whose two-dimensional expression in Cartesian coordinates
is, as shown above, ∇2 = ∂ xx2
+ ∂ yy
2
). Then, by assuming a solution of the form
iω t
w(x, y, t) = u(x, y) e , substituting it in Equation 5.402 and defining k = ω c,
we can readily determine that the equation for the spatial function u(x, y) is
∇2u + k2u = 0 (5.41)
which is known as Helmholtz equation. At this point, in order to make

progress, we must consider the fact that the Laplacian operator has differ-
ent forms in different types of coordinates and that, consequently, conve-
nience suggests to choose a coordinate system that matches the shape and
boundary of the membrane. So, for example, we choose Cartesian coordi-
nates for rectangular or square membranes, polar coordinates for a circular
membrane, etc. Unfortunately, since the number of useful coordinate sys-
tems is rather limited, so is the number of membrane problems that can be
solved with relative ease.
Example 5.2
Leaving the details to the reader as an exercise, we consider here
the case of a rectangular membrane of size a along x, b along y and
fixed along all edges. Choosing a system of Cartesian coordinates
and assuming a solution of Equation 5.41 in the separated-variables
form u(x, y) = f (x) g(y), we get ( f ′′ f ) = − ( g ′′ g ) − k2, where the primes
denote derivatives with respect to the appropriate argument. Then,
calling −kx2 the separation constant and defining ky2 = k2 − kx2, we are led
to the two ordinary differential equations f ′′ + kx2 f = 0 and g ′′ + ky2 g = 0
whose solutions are, respectively, f (x) = A1 sin kx x + A2 cos kx x and
g(y) = B1 sin ky y + B2 cos ky y .
At this point, enforcing the fixed BCs f (0) = f (a) = 0 and
g(0) = g(b) = 0, it follows that we must have kx = nπ a ( n = 1, 2,) and
ky = mπ b ( m = 1, 2,), thus implying that the eigenfrequencies are
T  n 2 m2 
ω n m = c kx2 + ky2 = π + ( n, m = 1, 2, ) (5.42a)
σ  a2 b2 
with corresponding eigenfunctions
un m (x, y) = Cn m sin ( nπ x a ) sin ( mπ y b ) (5.42b)
As an incidental remark, note that if the parameters a2 , b2 are incom-

mensurable (meaning that a2 b2 cannot be expressed as the ratio of
two integers), Equation 5.42a tells us that the eigenfrequencies are all
distinct. Otherwise, there can be different modes with the same fre-
quency and we have a case of degeneracy; for example, for a = 2b, the
frequency ( πc a ) 200 is a 3-fold degenerate eigenvalue, because it is
the frequency that corresponds to the (n, m) pairs of the modes (2,7),
(10,5) and (14,1). Clearly, for a square membrane, we have ω n m = ω m n.
With the above results, the general solution can then be written as
∞
w(x, y, t) = ∑  A
n , m=1
nm cos ω n m t + Bn m sin ω n m t  un m (x, y) (5.43)
where the constants are determined from the ICs w(x, y, 0) = w0 (x, y)
and ∂ t w(x, y, 0) = w
 0 (x, y) by a direct extension of the procedure fol-
lowed in Section 5.2.2. Here, we get
a b
4 4
An m =
ab ∫ ∫ w sin (nπx a) sin (mπy b) dx dy = ab
0 0
0 un m w0
a b
4 4
Bn m =
abω n m ∫ ∫ w sin (nπx a) sin (mπy b) dxdy = abω
0 0
0
nm
0
un m w
(5.44)
where in writing the rightmost expressions in inner product notation,

we used Cn m = 1 as the normalisation constants for the eigenfunctions
5.42b (i.e. a normalisation scheme such that urs un m = (ab 4) δ r nδ s m ).
It is now left to the reader to investigate a few of the first modal
shapes 5.42b and determine the position of their nodal lines (which,
in this case, are straight lines parallel to the edges of the membrane).
5.5.1 Circular membrane with fixed edge

For a finite circular membrane of radius R, it is convenient to write Helmholtz
equation in polar coordinates r, θ , where x = r cos θ and y = r sin θ . By so
doing, the wave equation 5.402 and Helmholtz equation 5.41 become,
respectively,
∂2w 1 ∂w 1 ∂2w 1 ∂2w ∂2u 1 ∂u 1 ∂2u

+ + = , + + + k2 u = 0 (5.45)
∂ r 2 r ∂ r r 2 ∂θ 2 c2 ∂ t 2 ∂ r 2 r ∂ r r 2 ∂θ 2
and, as above, we look for a solution of the spatial part u in the separated-
variables form u(r , θ ) = f (r) g(θ ). Substituting this solution in 5.452 , multiply-
ing the result by r 2 f g and then calling α 2 the separation constant lead to
the following two equations
d 2 f 1 df  2 α 2  d2g
+ + k − 2  f = 0, + α 2 g = 0 (5.46)
dr 2 r dr  r  dθ 2
where we already know that the solution of the second equation is

g(θ ) = C cos αθ + D sin αθ (or an equivalent form in terms of complex expo-
nentials). Before turning our attention to the first equation, however, it
should be observed that if our solution is to be a single-valued function of
position – which is what we are assuming here, but it is not the case if the
membrane is shaped, for example, like a sector of a circle – we must have
u(r , θ ) = u ( r , θ + 2nπ ). And since this ‘implicit BC’ along the radial lines of
the membrane implies that α must be an integer, it follows that the solution
of Equation 5.462 becomes g(θ ) = C cos nθ + D sin nθ and that we must set
α = n ( n = 0,1, 2,) in Equation 5.461. This latter equation, in turn, is one
of the ‘famous’ equations of mathematical physics and is known as Bessel’s
equation of order n. Its general solution is f (r) = AJn (kr) + BYn (kr) – that is,
a linear combination of the (extensively studied and tabulated) functions
Jn , Yn known as Bessel functions of order n of the first and second kind,
respectively. However, since the functions Yn become unbounded as kr → 0
and here we are assuming our membrane to extend continuously across the
origin, in order to have a finite displacement at r = 0, we must require B = 0
We are then left with the solution f (r) = AJn (kr), on which we must now
enforce the condition of fixed boundary at r = R. This leads to the count-
ably infinite set of frequency equations Jn (kR) = 0 – one for each value of
n – and amounts to determining the roots of the functions Jn.
Fortunately, the zeroes of the functions Jn can easily be found in math-
ematical tables; for n = 0,1, 2,3, we have, for example,
J0 (x) = 0 at x = 2.405, x = 5.520, x = 8.654, 
J1(x) = 0 at x = 3.832, x = 7.016, x = 10.173, 

J2 (x) = 0 at x = 5.136, x = 8.417, x = 11.620, 
J3 (x) = 0 at x = 6.380, x = 9.761, x = 13.015, 
where the zero at x = 0, which is also a root for all the Jn functions with
n ≥ 1, is excluded because it leads to no motion at all. So, it turns out that for
each value of n, we have a countably infinite number of solutions. Labelling
them with the index m ( m = 1, 2,) and recalling that k = ω c, the natural
frequencies of the membrane are
ω n m = ckn m ( n = 0,1, 2, ; m = 1, 2,) (5.47)

where the lowest frequencies are
2.405 c 3.832 c 5.136 c 5.520 c

ω 01 = , ω 11 = , ω 21 = , ω 02 = (5.48)
R R R R
The mode shapes, in turn, are given by the product of the two spatial
solutions; corresponding to the frequencies ω 0 m ( m = 1, 2,), we have – to
within a multiplying constant C0m – the eigenfunctions u0 m (r) = C0m J0 ( k0 m r )
which do not depend on θ . For every fixed n ≥ 1, on other hand, each one of
the frequencies ω n m is 2-fold degenerate because it corresponds to the two
eigenfunctions
un m (r , θ ) = Cn m Jn ( kn m r ) cos nθ , uˆ n m (r , θ ) = Cˆ n m Jn ( kn m r ) sin nθ (5.49)
which have the same shape but differ from one another by an angular rota-
tion of 90°.
Putting together the eigenfunctions un m , uˆ n m with the time part of the
solution, we can write the general solution as
∞
 ∞
w(r , θ , t) = ∑∑
m=1 

 n =0
un m ( An m cos ω n m t + Bn m sin ω n m t )

∞

+ ∑ uˆ ( Aˆ
n=1
nm nm )
cos ω n m t + Bˆ n m sin ω n m t  (5.50)

Remark 5.7
i. Just by drawing some schematic representations of the first few modes,

it can be seen that the (n, m)th mode has n nodal diameters (hence
no nodal diameters for n = 0) and m nodal circles (fixed boundary
included).
ii. The orthogonality of the eigenfunctions 5.49 is a consequence of
R
the property of Bessel functions
∫ 0
Jn1 ( kn1m1 r ) Jn2 ( kn2 m2 r ) rdr = 0
which holds for (a) n1 ≠ n2 and (b) n1 = n2 , m1 ≠ m2. Since, on

the other hand, when n1 = n2 = n and m1 = m2 = m, we have
R
∫ 0
( )
Jn2 ( kn m r )rdr = R2 2 Jn2+1 ( kn m R), it is left to the reader to determine
the constants Cn m , Cˆ n m needed to satisfy the normalisation condition

un m un m = 1 (and a similar relation for the functions uˆ n m ), where the
2π R
inner product in this case is given by un m un m =
∫ ∫
0 0
un2 m (r , θ )rdrdθ
As a final point for this section, we go back to the SLps discussed in

Section 5.4 because there, we recall, it was observed that any second-order
ordinary differential equation can be put in SL form. And since, in this
light, the fact that Bessel’s equation belongs to the SL class is not at all sur-
prising, it may be nonetheless worthwhile to show it explicitly.
Starting from Equation 5.461 (with α = n for the reason given above),
make the change of variable r = ω x k and denote by y(x) the function
f (ω x k). Then, since
df k dy d 2f k2 d 2 y
= , = 2
dr ω dx dr 2
ω dx2
it follows that under this transformation, Equation 5.461 becomes
d 2 y 1 dy  2 n2  d  dy  n2
+ + ω − 2 y =0 ⇒ − x + y = ω 2 xy (5.51)
dx2 x dx  x  dx  dx  x
where in writing the second equation, we multiplied the first by −x

and then observed that xy′′ + y′ = (xy′)′. At this point, it is immedi-
ate to see that Equation 5.512 is the Sturm–Liouville equation 5.27 with
p(x) = x, q(x) = n2 x, weight function w(x) = x and eigenvalue λ = ω 2. It
should be noted, however, that this is not a regular SL problem as the ones
discussed in Section 5.4, but a so-called singular SL problem because the
functions p(x) = x and w(x) = x are zero at x = 0 and therefore satisfy the
regularity requirements p(x) > 0, w(x) > 0 only on the interval (0, R], and
not on the closed interval [0, R]. Moreover, so does the function q(x) = n2 x
which is unbounded at x = 0.
In order to deal with this different type of SL problem, one must modify
the BC at the singular point – x = 0 in our case – by requiring that the solu-
tion y of Equation 5.51 and its derivative y′ remain bounded as x → 0 (note
that this is what we did above when we ‘eliminated’ the functions of the
second kind Yn). Under this condition, it can be shown that the SL operator is
self-adjoint and many of the ‘nice’ properties of regular SL problems – prop-
erties that, we recall, are essentially consequences of self-adjointness – retain
their validity. In particular – we recall from Section 5.4 – the most important
properties for our purposes are that the eigenvalues are real, that the eigen-
functions are orthogonal with respect to the weight w(x) and that they form
a complete orthonormal system in terms of which we can write the series
expansion of Equation 5.34 for any sufficiently regular function f (x).
Remark 5.8
Although it is not the case for the circular membrane considered here, it is
worth mentioning the fact that a singular SL problem may have a continu-
ous spectrum, where this term means that the problem may have non-trivial
solutions for every value of λ or for every value λ in some interval. This is
the most striking difference between regular problems (whose eigenvalues
are discrete) and singular ones. Some singular problems, moreover, may
have both a discrete and a continuous spectrum. However, when the prob-
lem has no continuous spectrum but a countably infinite number of discrete
eigenvalues, its eigenpairs, as pointed out above, have properties that are
similar to those of a regular SL problem.
5.6 FLEXURAL (BENDING) VIBRATIONS OF BEAMS
Going back to a one-dimensional system, it was shown in Example 2.5

of Chapter 2 that the equation of motion for the (small-amplitude) free
transverse vibrations of a beam with bending stiffness EI (x) and mass per
unit length µ(x) = ρ A(x) is (with a slight change of notation with respect to
Chapter 2; in particular, now we denote by y(x, t) the transverse displace-
ment at point x and time t )
∂2  ∂2y  ∂2y ∂4y 1 ∂2y

 EI (x) + µ (x) = 0, + = 0 (5.52)
∂ x2 ∂ x2  ∂ t2 ∂ x 4 a2 ∂ t 2
where in the second equation – which applies when the beam is uniform with
constant stiffness EI and constant cross-sectional area A – we defined the
parameter a = EI µ = EI ρ A. The first thing we notice is that Equation
5.522 is not a ‘standard’ wave equation; first of all, because there is a fourth-
order derivative with respect to x and, second, because a does not have the
dimensions of velocity. As it can be easily checked, moreover, waves of the
functional form 5.3 do not satisfy Equation 5.522 , thus indicating that a
(flexural) wave of arbitrary shape cannot retain its shape as it travels along
the beam. In fact, if we consider a travelling sinusoidal wave of the form
y(x, t) = A cos(kx − ω t), substitution in Equation 5.522 leads to ω = a k2 and
consequently to the phase velocity cp ≡ ω k = ak = 2πa λ , which in turn
tells us that cp , unlike the case of transverse waves on a string, is not the
same for all wavenumbers (or wavelengths). As known from basic physics,
this phenomenon is called dispersion and the rate at which the energy of a
flexural pulse – that is, a non-sinusoidal wave comprising waves with differ-
ent wavenumbers – propagates along the beam is not given by cp but by the
group velocity cg ≡ dω dk. And since in our case the dispersion relation is
ω = a k2, we get cg = 2ak = 2cp = 4π a λ . In passing it may be worth observ-

ing that in the general case, the relation between the two velocities is, in
terms of wavenumber or wavelength, respectively
dcp dcp
c g = cp + k , c g = cp − λ (5.53)
dk dλ
Remark 5.9
The above result on the beam propagation velocities leads to the definitely
unphysical conclusion that both cp and cg tend to increase without limit
as k → ∞ or, equivalently, as λ → 0. This ‘anomaly’ is due to the fact that
the equation of motion 5.52 is obtained on the basis of the simplest theory
of beams, known as Euler-Bernoulli theory, in which the most important
assumption is that plane cross sections initially perpendicular to the axis
of the beam remain plane and perpendicular to the deformed neutral axis
during bending. The assumption, however, turns out to be satisfactory for
wavelengths that are large compared with the lateral dimensions of the
beam; when it is not so the theory breaks down and one must take into
account – as Rayleigh did in his classic book of 1894 The Theory of Sound –
the effect of rotary (or rotatory) inertia. This effect alone is sufficient to
prevent the divergence of the velocities at very short wavelengths, but is not
in good agreement with the experimental data. Much better results can be
obtained by means of the Timoshenko theory of beams, in which the effect
of shear deformation is also included in deriving the governing equations.
We will consider these aspects in Section 5.7.3.
5.7 FINITE BEAMS WITH CLASSICAL BCs
Consider now a uniform beam of length L and constant cross-section A.

Assuming a solution of the form y(x, t) = u(x)e i ω t , substitution into Equation
5.522 gives the fourth-order ordinary differential equation
d 4 u(x) ω 2 ω 2µ
− γ 4 u(x) = 0 where γ4 = = (5.54)
dx4 a2 EI
whose general solution can be written as
u(x) = C1 cosh γ x + C2 sinh γ x + C3 cos γ x + C4 sin γ x (5.55)
where the constants depend on the type of BCs. Here we will consider a few
typical and ‘classical’ cases.
Case 1: Both ends simply supported (SS-SS configuration)

In this case, the BCs require that the displacement u(x) and
bending moment EI u′′ vanish at both ends, that is
u(0) = u(L) = 0, u′′(0) = u′′(L) = 0 (5.56)
where, recalling the developments of Section 2.5.1, we recognise

5.561 as geometric BCs and 5.562 as natural BCs. Substitution
of these BCs into Equation 5.55 leads to C1 = C2 = C3 = 0 and to
the frequency equation sin γ L = 0, which in turn implies γ L = nπ.
Then, for n = 1, 2,, the allowed frequencies and the correspond-
ing (non-normalised) eigenfunctions are
n2 π 2 EI  nπx 
ωn = , un (x) = C4 sin γ n x = C4 sin  (5.57)
L2 µ  L 
Note that the eigenfunctions are the same as those of the fixed-
fixed string.
Case 2: One end clamped and one end free (C-F or cantilever
configuration).
If the end at x = 0 is rigidly fixed (clamped) and the end at x = L is
free, the geometric BCs require the displacement u and slope u′ to
vanish at the clamped end, whereas at the free end, we have the two
(natural) BCs of zero bending moment and zero shear force EI u′′′ .
Then, using the BCs
u(0) = 0, u′(0) = 0; u′′(L) = 0, u′′′(L) = 0 (5.58)
in Equation 5.55, we get the four relations
C1 + C3 = 0, C2 + C4 = 0
C1 cosh γ L + C2 sinh γ L − C3 cos γ L − C4 sin γ L = 0
C1 sinh γ L + C2 cosh γ L + C3 sin γ L − C4 cos γ L = 0
which can be conveniently arranged in matrix form as
 1 0 1 0  C1 
  
 0 1 0 1  C2 
 cosh γ L sinh γ L − cos γ L − sin γ L  C3  = 0 (5.59)
  
 sinh γ L cosh γ L sin γ L − cos γ L  
 C4 
and tell us that we have a non-trivial solution only if the

matrix determinant is zero. This gives the frequency equation
1 + cosh γ L cos γ L = 0, which must be solved numerically. The first

few roots are
γ 1 L = 1.875, γ 2 L = 4.694, γ 3 L = 7.855, γ 4 L = 10.996
and we obtain the natural frequencies
2
EI  2n − 1  EI
ω n = (γ n L ) ( n = 1, 2,) (5.60)
2
≅ π2 
µ L4  2  µ L4
where the rightmost expression is a good approximation for n ≥ 3.

The eigenfunctions, on the other hand, can be obtained from the
first three of Equations 5.59, and we get
 cosh γ n L + cos γ n L 
C2 = −C1  ≡ −κ n C1
 sinh γ n L + sin γ n L 
so that substitution into 5.55 gives
{ }
un (x) = C1 ( cosh γ n x − cos γ n x ) − κ n ( sinh γ n x − sin γ n x ) (5.61)
and it can be shown that the second mode has a node at x = 0.783L
and the third has two nodes at x = 0.504L and x = 0.868L, etc.
Case 3: Both ends clamped (C-C configuration).
All four BCs are geometric in this case, and we must have
u(0) = u(L) = 0, u′(0) = u′(L) = 0 (5.62)
Following the same procedure as above leads to
 1 0 1 0  C1 
  
 0 1 0 1  C2 
 cosh γ L sinh γ L cos γ L sin γ L  C3 =0
  
 sinh γ L cosh γ L − sin γ L cos γ L  
 C4 
with the frequency equation 1 − cosh γ L cos γ L = 0. The first four

roots are
γ 1 L = 4.730, γ 2 L = 7.853, γ 3 L = 10.996, γ 4 L = 14.137
and the approximation γ n L ≅ (2n + 1) π 2 is very good for all n ≥ 3.

As for the eigenfunctions, it is not difficult to see that we get
{ }
un (x) = C1 ( cosh γ n x − cos γ n x ) − κ n ( sinh γ n x − sin γ n x ) (5.63a)
where now κ n is not as in the previous case but we have
cosh γ n L − cos γ n L
κn = (5.63b)
sinh γ n L − sin γ n L
Finally, we leave to the reader the task of filling in the details for
other two classical configurations: the free-free beam and the beam
clamped at one end and simply supported at the other.
Case 4: Both ends free (F-F configuration).
The BCs are now all natural and require the bending moment
and shear force to vanish at both ends, that is
u′′(0) = u′′(L) = 0, u′′′(0) = u′′′(L) = 0 (5.64)
thus leading to the frequency equation 1 − cosh γ L cos γ L = 0, which

is the same as for the C-C configuration. The natural frequencies,
therefore, are the same as in that configuration (not the eigenfunc-
tions!), but with the difference that now the system is unrestrained
and we expect rigid-body modes at zero frequency. This is, in fact,
the case, whereas the elastic modes are
{ }
un (x) = C1 ( cosh γ n x + cos γ n x ) − κ n ( sinh γ n x + sin γ n x ) (5.65)
where κ n is the same as in the C-C case (i.e. Equation 5.63b); sub-
stitution of γ = 0 in Equation 5.54 gives u0 (x) = Ax3 + Bx2 + C x + D
which with the BCs 5.64 does not lead – as in the other cases – to
the trivial zero solution, but to u0 (x) = C x + D, that is, a linear com-
bination of the two functions u0(1) = 1 and u0(2) = x. And since here we
are considering the lateral motions of the beam, u0(1) , u0(2) are physi-
cally interpreted as a transverse rigid translation of the beam as a
whole and a rigid rotation about its centre of mass, respectively.
Also, note that the first elastic mode has two nodes at x = 0.224L
and x = 0.776L.
Case 5: C-SS configuration.
For a beam clamped at the end x = 0 and simply supported at the
other, the frequency equation turns out to be tan γ L = tanh γ L. The
first four roots are
γ 1 L = 3.927, γ 2 L = 7.069, γ 3 L = 10.210, γ 4 L = 13.352
and can be well approximated by γ n L ≅ (4n + 1)π 4. The eigenfunc-

tions are left to the reader as an exercise (also show that the second
mode has one node at x = 0.558L). A final incidental remark: in all

the configurations above (except Case 1, whose frequencies are all
equally spaced), the lower-order frequencies are irregularly spaced,
but as n increases, the difference (γ n +1 − γ n ) L approaches π.
5.7.1 On the orthogonality of beam eigenfunctions

Denoting by primes the derivatives with respect to x, let un , um be two beam
eigenfunctions, so that the equations un′′′′− γ n4 un = 0 and um
′′′′− γ m4 um = 0 are
identically satisfied. Multiplying the first equation by um , the second by un,
subtracting the two results and integrating over the beam length gives
L L

∫ (
( um un′′′′− un um′′′′) dx = γ n4 − γ m4 )∫u u dx (5.66)
m n
0 0
and we can integrate by parts four times the term um un′′′′ to get
L L

∫0
um un′′′′dx = [ um un′′′− um
′ un′′ + um ′′′un ] 0L +
′′ un′ + um
∫ u′′′′u dx (5.67)
0
m n
Then, using Equation 5.67 in 5.66 gives

L
[ umun′′′− um′ un′′ + um′′ un′ − um′′′un ] 0L= (γ n4 − γ m4 ) ∫ um un dx (5.68a)

0
from which we readily see that any combination of the various ‘classical’
BCs (e.g. SS-SS, C-C, C-F, etc.) cause the left-hand-side to vanish. More
than that, the same occurs for any set of homogeneous BCs of the form
au + bu′ + cu′′ + du′′′ = 0 (5.69)
(with a, b, c, d constants), where this type of condition can arise, for exam-
ple, from a combination of linear and torsional springs. So, having assumed
γ n ≠ γ m from the beginning, we conclude that for BCs of the form 5.69, we
L
have the orthogonality relation
∫ 0
um un dx = 0, or, in inner product nota-
tion, um un = 0.
Remark 5.10
The essence of the argument does not change if the beam is not uniform
and has a flexural stiffness s(x) ≡ EI (x) and mass per unit length µ (x) that
depend on x. In this case, the two eigenfunctions satisfy the equations

(sun′′)′′ − µω n2un = 0 and (sum
′′ )′′ − µω m2 um = 0, and it is now left to the reader as
an exercise to show that Equation 5.68 is replaced by
L L
 0
(
 um ( sun′′ )′ − un ( sum′′ )′ + s ( um′′ un′ − um′ u′′ )n  = ω n2 − ω m2 )∫ µum nu dx (5.68b)
0
which, for the classical BCs, implies the orthogonality relation

L
∫ 0
µ um un dx = 0, or um un µ
= 0 in inner product notation (note that here
µ (x) plays the role of the weight function w(x) introduced in Section 5.4 on
SLps).
5.7.2 Axial force effects

If a uniform beam is subjected to a non-negligible constant tension T0 acting
along its longitudinal axis, the equation of motion must include a ‘string-
like’ term that accounts for this additional stiffening effect. So, we have
∂4 y ∂2 y ∂2 y d 4 u T0 d 2u µω 2
EI − T0 + µ =0 ⇒ − − u = 0 (5.70)
∂ x4 ∂ x2 ∂t 2 dx4 EI dx2 EI
where the second equation follows from the first when we assume a solution
of the form y(x, t) = u(x) e i ωt. Looking for solutions of Equation 5.702 in the
form u(x) = A eα x leads to
α 12  T0
2
 T0  µω2
 = ±   +
α 22  2EI  2EI  EI
with α 12 > 0 and α 22 < 0. Consequently, we have the four roots ±η and ±i ξ
where
12 12
 2
µω 2 T   2
µω 2 T 
 T   T 
η =   0  + + 0  , ξ =  0  + − 0  (5.71)
  2EI  EI 2EI    2EI  EI 2EI 
and the solution of Equation 5.702 can be written as
u(x) = C1 cosh η x + C2 sinh η x + C3 cos ξ x + C4 sin ξ x (5.72)
where the constants depend on the type of BCs (note that Equation 5.72
is only formally similar to Equation 5.55 because here the hyperbolic and
trigonometric functions have different arguments). As it turns out, the
simplest case is the SS-SS configuration, in which the BCs are given by
Equation 5.56. Enforcing these BCs on the solution 5.72 – and the reader is
invited to do the calculations – yields C1 = C3 = 0 and the frequency equa-
tion sinh η L sin ξ L = 0. And since sinh η L ≠ 0 for η L ≠ 0, we are left with
sin ξ L = 0, which in turn implies ξ n L = nπ. Thus, the allowed frequencies
are
n2 π 2 EI T0 L2
ωn = + ( n = 1, 2,) (5.73a)
L2 µ n2 π 2 µ
and can also be written as
nπ T0 n2 π 2
ωn =
L µ
(
1 + n2 π 2 R = 2
L
) EI 
µ 
1 
 1 + 2 2  (5.73b)
nπ R
where these two expressions show more clearly the two extreme cases in
terms of the non-dimensional ratio R = EI T0L2 : for small values of R, the
tension is the most important restoring force and the beam behaves like a
string; conversely, for large values of R, the bending stiffness EI is the most
important restoring force and we recover the case of the beam with no axial
force.
Associated to the frequencies 5.73, we have the eigenfunctions
un (x) = C4 sin ξ n x because enforcing the BCs leads – together with the previ-
ously stated result C1 = C3 = 0 – also to C2 = 0.
Remark 5.11
i. If T0 is a compressive force we must reverse its sign in the formulas

above. Worthy of mention in this regard is that for n = 1, we can write
n2 π 2 EI  L2  n2π 2 EI  T 
ω1 = 1 − T0 2  = 2 1 − 0  (5.74)
L2 µ  π EI  L µ  TE 
and note that ω 1 → 0 as T0 → TE = π 2EI L2 . As is well-known from

basic engineering theory, TE is the so-called Euler’s buckling load.
ii. The fact that, as shown by Equations 5.73 and 5.74, the natural fre-
quencies of a straight beam are increased by a tensile load and lowered
by a compressive load applies in general and is in no way limited to the
case of SS-SS BCs.
For BCs other than the SS-SS configuration, the calculations are in general
more involved. If, for example, we consider the C-C configuration, it is
convenient to place the origin x = 0 halfway between the supports. By so
doing, the eigenfunctions are divided into even (i.e. such that u(− x) = u(x))
and odd (i.e. such that u(− x) = − u(x)), where the even functions come from
the combination C1 cosh η x + C3 cos ξ x while the odd ones come from the
combination C2 sinh η x + C4 sin ξ x . In both cases, if we fit the BCs at x = L 2
they will also fit at x = − L 2. For the even and odd functions, respectively,
the C-C BCs u(L 2) = u′(L 2) = 0 lead to the frequency equations
ξ tan(ξ L 2) = −η tanh(η L 2), η tan(ξ L 2) = ξ tanh(η L 2) (5.75)
which must be solved by numerical methods. The first equation gives the
frequencies ω 1 , ω 3 , ω 5 , associated with the even eigenfunctions, while
the second gives the frequencies ω 2 , ω 4 , ω 6 , associated with the odd
eigenfunctions.
5.7.3 Shear deformation and rotary

inertia (Timoshenko beam)
It was stated in Section 5.6 that the Euler-Bernoulli theory provides sat-
isfactory results for wavelengths that are large compared with the lateral
dimensions of the beam, where this latter quantity – we add now – is typi-
cally measured in terms of rg = I A , the radius of gyration of the beam
cross-section. In other words, this means that the Euler-Bernoulli theory
fails when either (a) the beam is short and deep or (b) the beam is sufficiently
slender (say, L rg ≥ 30), but we are interested in higher-order modes – two
cases in which the kinematics of motion must take into account the effects
of shear deformation and rotary inertia.
Assuming for simplicity a uniform beam with constant physical prop-
erties (i.e. shear modulus G, bending stiffness EI, cross-sectional area A
and mass per unit length µ independent of x) and adopting a Lagrangian
perspective, the main observation for our purposes is that shear deforma-
tion is accounted for by an additional term in the potential energy density,
while rotary inertia by an additional term in the kinetic energy density.
Respectively, these terms are (see also the following Remark 5.12)
κ GA µ rg2
Vshear =
2
(∂ x y − ψ )2 , Trot =
2
(∂ t ψ )2 (5.76)
where ψ = ψ (x, t) is the angle of rotation of the beam (at point x and time
t) due to bending alone, and κ is a numerical factor known as Timoshenko
shear coefficient that depends on the shape of the cross-section (typical val-
ues are, for example, κ = 0.83 for a rectangular cross-section and κ = 0.85
for a circular cross-section).
With the terms 5.76, the Lagrangian density becomes
Λ=
µ
2
{(∂ y) + r (∂ ψ ) } − 12 {EI (∂ ψ ) + κ GA (∂ y − ψ ) } (5.77)
t
2 2
g t
2
x
2
x
2
where here we have two independent fields, namely, the bending rotation ψ
and the transverse displacement y (so that the difference (∂ x y − ψ ) describes
the effect of shear).
Remark 5.12
i. In the Euler-Bernoulli theory, the slope ∂ x y of the deflection curve is

given by ∂ x y = ψ , where ψ is the angle of rotation due to bending. In
the Timoshenko theory, the slope is made up of two contributions –
bending and shear – and one writes ∂ x y = ψ + θ , where θ = ∂ x y − ψ
represents the contribution of shear.
ii. In expressing the shear force S as S = κ GAθ, the Timoshenko fac-
tor κ is, broadly speaking, a kind of ‘cross-section average value’
that accounts for the non-uniform distribution of shear over the
cross-section.
iii. The rotary inertia term accounts for the fact that a beam element
rotates as well as translating laterally and therefore adds a contribu-
tion J (∂ tψ ) 2 to the kinetic energy density, where J = ρ I = µ rg2 is the
2
mass moment of inertia per unit length.
Now, observing that in terms of y, the Lagrangian has the functional

form Λ = Λ ( ∂t y, ∂ x y ) while in terms of ψ has the functional form
Λ = Λ (ψ , ∂t ψ , ∂ xψ ), we use Equation 2.68a to obtain the two equations
of motion
(
− µ ∂ tt2 y + κ GA ∂ xx
2
)
y − ∂ xψ = 0
(5.78)
κ GA (∂ x y − ψ ) − µ rg2 ∂ tt2ψ + EI ∂ xx
2
ψ =0
which govern the free vibration of the uniform Timoshenko beam and show
that physically we have two coupled ‘modes of deformation’. In general, the
two equations cannot be ‘uncoupled’, but for a uniform beam, it is possible
and the final result is a single equation for y. From Equations 5.78, in fact,
we obtain the relations
µ
∂ xψ = ∂ xx
2
y− ∂ tt2 y
κ GA
(5.79)
EI ∂ xxx
3
ψ + κ GA ∂ xx
2
( )
y − ∂ xψ − µ rg2 ∂ xtt
3
ψ =0
where the first follows directly from Equation 5.781 while the second is
obtained by differentiating Equation 5.782 with respect to x. Then, using
Equation 5.791 (and its derivatives, as appropriate) in Equation 5.792 gives
the single equation
∂4 y ∂2 y  µ EI  ∂2 y µ 2rg2 ∂4 y
EI +µ 2 − + µ rg2  + = 0 (5.80)
∂x 4
∂t  κ GA 2
 ∂ x ∂t 2
κ GA ∂t 4
which, as expected, reduces to Equation 5.522 when shear and rotary iner-
tia are neglected. With respect to the Euler-Bernoulli beam, the three addi-
tional terms are
µ EI ∂2 y ∂2 y µ 2 rg2 ∂4 y
− , − µ rg2 , (5.81)
κ GA ∂ x2 ∂t 2 ∂ x2 ∂ t 2 κ GA ∂t 4
where the first is due to shear, the second to rotary inertia and the third is a
‘coupling’ term due to both effects. Note that this last term vanishes when
either of the two effects is negligible. Also, note that, in mathematical par-
lance, the shear effect goes to zero if we let G → ∞ while the rotary inertia
term goes to zero if we let rg → 0.
If now we look for a solution of Equation 5.80 in the usual form
y(x, t) = u(x) e i ω t , we arrive at the ordinary differential equation
d 4 u  µ ω 2 µ ω 2rg2  d 2u  µ 2ω 4 rg2 µω2 

+ +  + − u = 0 (5.82)
dx 4
 κ GA EI  dx  EI κ GA
2
EI 
which is now taken as a starting point in order to investigate the individual

effects of shear deflection and rotary inertia on the natural frequencies of
an SS-SS beam.
Shear deflection alone (shear beam)
If we neglect rotary inertia ( rg → 0), Equation 5.82 gives
d 4 u µ ω 2 d 2u µ ω 2
+ − u = 0 (5.83)
dx4 κ GA dx2 EI
and we can parallel the solution procedure used in Section 5.7.2. As we did
there, we obtain the four roots ±η and ±iξ , where now however we have
12
 2 
 µω 2  µω 2 µω 2 
η=  + −
  2κ GA  EI 2κ GA 

(5.84)
12
 2 
 µω 2  µω 2 µω 2 
ξ=  + +
  2κ GA  EI 2κ GA 

and the allowed frequencies follow from the condition ξ n L = nπ . Denoting

by ω (0)
n the natural frequencies of the SS-SS Euler-Bernoulli beam (Equation
5.571), we get
ω (shear)
n κ GAL2 1
= = (5.85)
1 + n π ( rg L ) ( E κ G )
2
ωn (0)
κ GAL2 + n2π 2EI 2 2
where in the second expression the effect of the ratio rg L is put into evi-
dence. As rg L → 0 (that is, very slender beams), we get ω n(shear) ω n(0) → 1.
Rotary inertia alone (Rayleigh beam)
If now we neglect the effect of shear (G → ∞ ), Equation 5.82 gives
d 4 u µ ω 2rg2 d 2u µ ω 2
+ − u = 0 (5.86)
dx4 EI dx2 EI
and we can proceed exactly as above to arrive at
ω (rot)
n 1
= 2 (5.87)
ω (0)
n 1 + n π ( rg L )
2 2
where, again, we denote by ω (0)

n the frequencies of the SS-SS Euler-Bernoulli
beam.
Equations 5.85 and 5.87 show that both effects tend to decrease the
beam natural frequencies, but that, in general, shear is more important
than rotary inertia because E κ G > 1 (typical values are between 2 and 4).
Also, note that for a given beam, both effects become more pronounced for
increasing values of n (see the following Remark 5.13 (iv)).
Remark 5.13
i. In Remark 5.9, Section 5.6, it was mentioned that the inclusion of

rotary inertia is sufficient to prevent the divergence of the phase and
group velocities at very short wavelengths (or very large wavenum-
bers). Since Equation 5.80 with rotary inertia and no shear become
EI ∂ xxxx
4
y + µ ∂ tt2 y − µ rg2 ∂ xxtt
4
y = 0, it is left to the reader to check that
using a solution of the form y(x, t) = A cos ( kx − ω t ) in this equation
does indeed lead to the result that both velocities cp , cg do not diverge
as λ → 0 but tend to the limit E ρ .
ii. For a Timoshenko beam with SS-SS BCs, the reader is invited to show
that from Equation 5.82, we obtain the following equation for the
natural frequencies
µ 2rg2  µ n2π 2 µ n2π 2 µ rg2  2 n4 π 4

ω n4 −  + + ω n + 4 = 0 (5.88)
kGAEI  EI kGAL
2
EI L2  L
which, as it should be expected, leads to the frequencies of Equation

5.87 for a Rayleigh beam (i.e. no shear), to the frequencies of
Equation 5.85 for a shear beam (i.e. no rotary inertia) and to the
frequencies of the Euler-Bernoulli beam if we neglect both shear
and rotary inertia effects.
iii. Equation 5.88 is quadratic in ω n2, meaning that it gives two values of
ω n2 for each n. The smaller value corresponds to a flexural vibration
mode, while the other to a shear vibration mode.
iv. If we consider the effect of rotary inertia alone, Equation 5.87 can
be used to evaluate the minimum slenderness ratio needed in order
to have, for example, ω n(rot) ω n(0) ≥ 0.9. For n = 1, 2,3 (first, second and
third mode), respectively, we get L rg ≥ 6.5, L rg ≥ 13.0 and L rg ≥ 19.5
Using Equation 5.85 and assuming a typical value of 3 for the ratio
E κ G, the same can be done for the effect of shear alone. In order to
obtain the result ω n(shear) ω n(0) ≥ 0.9, for the first, second and third mode,
respectively, we must have slenderness ratios such that L rg ≥ 11.2,
L rg ≥ 22.5 and L rg ≥ 33.7.
5.8 BENDING VIBRATIONS OF THIN PLATES
Much in the same way in which a membrane is the two-dimensional ana-

logue of a flexible string, a plate is the two-dimensional analogue of a beam.
Plates, in fact, do have bending stiffness and the additional complications
arise not only from the increased complexity of two-dimensional wave
motion but also from the complex sorts of stresses that are set up when a
plate is bent. Without entering in the details of these aspects, the ‘classi-
cal’ theory of thin plates – known also as Kirchhoff theory – is for plates
the counterpart of the Euler-Bernoulli theory of beams. Two of the most
important assumptions of this theory are (a) the plate thickness h is small
compared with its lateral dimensions (say, h a ≤ 1 20, where a is the smallest
in-plane dimension) and (b) normals to the mid-surface of the undeformed
plate remain straight and normal to the mid-surface (and unstretched in
length) during deformation. In particular, the theory neglects the effects of
shear deformation and rotary inertia, which are otherwise included in the
more refined theory known as Mindlin plate theory.
Under the classical assumptions, if we let w(x, y, t) be the transverse dis-
placement at point x, y and time t and consider a homogeneous and isotropic
plate with constant thickness h, constant mass density and elastic proper-
ties, it can be shown (see, for example, Graff (1991), Leissa and Qatu (2011)
and Remarks 5.14(i) and 5.14(ii) below, or, for a variational approach using
Hamilton’s principle, Chakraverty (2009), Géradin and Rixen (2015) or
Meirovitch (1997)) that the equation of motion of free vibration is
 ∂4w ∂4w ∂4w ∂2 w ∂2 w

D + 2 + + ρ h = 0 ⇒ D ∇ 4
w + ρ h = 0 (5.89)
 ∂ x4 ∂ x2 ∂ y 2 ∂ y 4  ∂t 2 ∂t 2
where ∇4 = ∇2∇2 (the Laplacian of the Laplacian, whose expression in rect-

angular coordinates is as shown in Equation 5.891) is called biharmonic
operator, ρ is the plate mass density (so that σ = ρ h is the mass density per
unit area) and
Eh3
D= (5.90)
(
12 1 − ν 2 )
is the plate flexural stiffness, where E is Young’s modulus and ν is Poisson’s
ratio.
Remark 5.14
i. Under the classical Kirchhoff assumptions, it is not difficult to see

that the plate kinetic energy density (per unit area) is σ (∂ t w ) 2. Less
2
immediate is to determine that the bending potential energy density

is, in rectangular coordinates,
D  ∂2 w ∂2 w   ∂2 w ∂2 w  ∂2 w  2  
2
 + − 2(1 − ν )  2 −   (5.91)
2  ∂ x2 ∂ y 2   ∂ x ∂ y
2
 ∂ x ∂ y   
 
where we recognise the first term within curly brackets as the rect-
( )
2
angular coordinates expression of ∇2w .
ii. Forming the (small-amplitude) Lagrangian density Λ with the
kinetic and potential energy densities of point (i) and observing that
(
Λ = Λ ∂ t w, ∂ xx
2
w, ∂ yy
2
w, ∂ xy
2
w, ∂ yx
2
)
w , it is left to the reader to check
that the calculations of the appropriate derivatives prescribed by
Equations 2.69 and 2.70 lead to the equation of motion 5.89.
iii. In light of the aforementioned analogy between plates and beams,
it is quite evident that a plate is a dispersive medium. As for Euler-
Bernoulli beams, the Kirchhoff theory of plates predicts unbounded
wave velocity at very short wavelengths, and this ‘unphysical’ diver-
gence is removed when one takes into account the effects of shear and
rotary inertia. Also, for finite plates, both effects – like in beams –
are found to decrease the natural frequencies (with respect to the
Kirchhoff theory), becoming more significant as the relative thickness
of the plate increases and for higher frequency vibration modes.
Assuming a solution of Equation 5.89 of the form w = ue i ωt – where the func-

tion u depends only on the spatial coordinates – substituting it in Equation
5.89 and defining γ 4 = ρhω 2 D, we readily find that the equation for u is
(∇ 4
)
−γ 4 u = 0 ⇒ (∇ 2
)( )
+ γ 2 ∇2 − γ 2 u = 0 (5.92)
where in the second expression we conveniently factored the operator

( ) ( )( )
∇4 − γ 4 into ∇2 + γ 2 ∇2 − γ 2 because by so doing the complete solution
u can be obtained as u = u1 + u2, where u1 , u2 are such that ∇2 + γ 2 u1 = 0 ( )
( )
and ∇2 − γ 2 u2 = 0 (note that the equation for u1 is the Helmholtz equation
of Section 5.5).
Remark 5.15
Substituting u = u1 + u2 in the l.h.s. of Equation 5.922 and taking into account

( ) ( )
that ∇2 + γ 2 u1 = 0 and ∇2 − γ 2 u2 = 0 lead to the chain of relations
(∇ 2
)( ) ( )(
+ γ 2 ∇2u1 − γ 2u1 + ∇2u2 − γ 2u2 = ∇2 + γ 2 ∇2u1 − γ 2u1 )

( )(
= ∇2 + γ 2 ∇2u1 + γ 2u1 − γ 2u1 − γ u ) = −2γ ( ∇
2
1
2 2
)
+ γ 2 u1 = 0
thus confirming that the complete solution can be expressed as the sum
u = u1 + u2.
In order to investigate the free vibrations of finite plates, the equation of

motion must be supplemented by appropriate BCs at the edges, where – as
for beams – the classical BCs are simply supported (SS), clamped (C) and
free (F). In addition, the fact that we are now considering a two-dimensional
system suggests that, just like we did with membranes, we should choose a
type of coordinates that matches the shape and boundary of the plate.
In the following sections, we will briefly examine a few simple cases. For
more detailed and exhaustive treatments, we refer the interested reader to
the vast literature on the subject (essentially due to the importance of the
plate structure in many engineering applications), such as, for instance, the
classic monograph by Leissa (1969), or the books by Chakraverty (2009)
or Szilard (2004). Needless to say, very good accounts can also be found in
books on vibrations of continuous systems, such as Hagedorn and DasGupta
(2007), Leissa and Qatu (2011) or Rao (2007), just to name a few.
5.8.1 Rectangular plates
Assuming complete support per edge, for rectangular plates, there are many
distinct cases involving all possible combinations of classical BCs. Among
these, it turns out that the more tractable ones are those in which two
opposite edges are simply supported and that the simplest case is when all
edges are simply supported. Here, therefore, we use rectangular coordinates
and consider a uniform rectangular plate of length a along the x direction,
length b along y and simply supported on two opposite edges.
If the simply supported edges are x = 0 and x = a, the BCs at these edges are
u(0, y) = u(a, y) = 0 and ∂2xx u(0, y) = ∂2xx u(a, y) = 0, and they are all satisfied if
we choose a solution of the form u(x, y) = Y (y)sin α x = [Y1(y) + Y2 (y)] sin α x
with α = nπ a ( n = 1, 2,), where in writing the last expression, we took
into account the fact that, as pointed out in the preceding section, u is given
by the sum u1(x, y) + u2 (x, y).
Now, substituting u1 = Y1 sin α x into the rectangular coordinates expres-
(
)
sion of equation ∇2 + γ 2 u1 = 0 and u2 = Y2 sin α x into the rectangular
coordinates expression of ( ∇ 2
)
− γ 2 u2 = 0 leads to the two equations
( )
Y1′′+ γ 2 − α 2 Y1 = 0, ( )
Y2′′− γ 2 + α 2 Y2 = 0 (5.93)
whose solutions are, assuming γ 2 − α 2 > 0,
Y1(y) = A1 cos ( )
γ 2 − α 2 y + A2 sin ( γ 2 −α2y )
(5.94)
Y2 (y) = A3 cosh ( γ + α y A4 sinh
2 2
) ( γ +α y
2 2
)
thus implying that the complete solution u = u1 + u2 is
u(x, y) = { A1 cos ry + A2 sin ry + A3 cosh sy + A4 sinh sy} sin α x (5.95a)
where we defined
r = γ 2 −α2 , s = γ 2 + α 2 (5.95b)
At this point, we must enforce the BCs at the edges y = 0 and y = b on the
solution 5.95. If also these two edges are simply supported, the appropriate
BCs are u(x,0) = u(x, b) = 0 and ∂2yy u(x,0) = ∂2yy u(x, b) = 0. Then, since the
BCs at y = 0 give (as the reader is invited to check) A1 = A3 = 0, the solution
5.95 becomes u = { A2 sin ry + A4 sinh sy} sin α x, while for the other two BCs
at y = b, we must have
 sin rb sinh sb 
det   =0 ⇒ (s2 + α 2 )sin rb sinh sb = 0
 −α sin rb
2 2
s sinh sb 
which in turn implies sin rb = 0 and therefore r = mπ b, with m = 1,2,. So,

recalling Equation 5.95b1 together with the definition γ 2 = ω ρ h D , it fol-
lows that the natural frequencies of the SS-SS-SS-SS rectangular plate are,
for n, m = 1, 2,,
D  n 2 m2  ρh   a 
2
ω nm = π 2  a2 + b2  ⇒ ω n m a2 = π 2  n2 + m2    (5.96)
ρh D   b 

where the second expression is written in the non-dimensional form by put-

ting the plate aspect ratio a b into evidence.
Finally, since using rb = mπ into the equation A2 sin rb + A4 sinh sb = 0 or
into −α 2 A2 sin rb + s2 A4 sinh sb = 0 (which are the two equations we get by
enforcing the SS BCs at y = b) gives A4 = 0, the eigenfunctions associated
with the frequencies 5.96 are
 nπ x   mπy 
unm (x, y) = Anm sin   sin  (5.97)
 a   b 
which are the same mode shapes as those of the rectangular membrane
clamped on all edges (Equation 5.42b).
Remark 5.16
The solution 5.95 has been obtained assuming γ 2 > α 2 . If γ 2 < α 2 , Equation
( )
5.931 is rewritten as Y1′′− α 2 − γ 2 Y1 = 0 and Equation 5.941 becomes
Y1(y) = A1 cosh ry ˆ , with r̂ = α 2 − γ 2 . In solving a free-vibration
ˆ + A2 sinh ry
problem, one should consider both possibilities and obtain sets of eigenval-
ues (i.e. natural frequencies) for each case, by checking which inequality
applies for an eigenvalue to be valid. However, in Leissa (1973), it is shown
that while for the case γ 2 > α 2 proper eigenvalues exist for all six rectan-
gular problems in which the plate has two opposite sides simply supported;
the case γ 2 < α 2 is only valid for the three problems having one or more free
sides. For example, for a SS-C-SS-F plate (with a typical value ν = 0.3 for
Poisson’s ratio), it is found that the case γ 2 < α 2 applies for nb a > 7.353.
For the three cases in which the two opposite non-SS edges have the same
type of BC – that is, the cases SS-C-SS-C and SS-F-SS-F (the case SS-SS-
SS-SS has already been considered above) – the free vibration modes are
either symmetric or antisymmetric with respect not only to the axis x = a 2
but also with respect to the axis y = b 2. In this light, it turns out to be con-
venient to place the origin so that the two non-SS edges are at y = ± b 2 and
use the even part of solution 5.95 to determine the symmetric frequencies
and mode shapes and the odd part for the antisymmetric ones.
So, for example, for the SS-C-SS-C case in which we assume γ 2 > α 2,
the even and odd parts of Y are, respectively, Yev (y) = A1 cos ry + A3 cosh sy
and Yodd (y) = A2 sin ry + A4 sinh sy , where r , s are as in Equation 5.95b.
Then, enforcing the clamped BC Y (b 2) = Y ′(b 2) = 0 leads to the frequency

equations
r tan ( rb 2) = − s tanh ( sb 2) , s tan ( rb 2) = r tanh ( sb 2) (5.98)
for the symmetric and antisymmetric modes, respectively.

If, on the other hand, we maintain the origin at the corner of the plate
(just like we did for the SS-SS-SS-SS case above), enforcing the BCs
Y (0) = Y ′(0) = 0 and Y (b) = Y ′(b) = 0 on the solution 5.95 leads to the fre-
quency equation
rs cos rb cosh sb − rs − α 2 sin rb sinh sb = 0 (5.99)
which is more complicated than either of the two frequency equations 5.98
although it gives exactly the same eigenvalues. Note that with these BCs,
there is no need to consider the case γ 2 < α 2 for the reason explained in
Remark 5.16 above.
Remark 5.17
i. Since a plate with SS-C-SS-C BCs is certainly stiffer than a plate with
all edges simply supported, we expect its natural frequencies to be
higher. It is in fact so, and just to give a general idea, for an aspect ratio
a b = 1 (i.e. a square plate) the non-dimensional frequency parameter
ω a2 ρ h D is 19.74 for the SS-SS-SS-SS case and 28.95 for the SS-C-
SS-C case.
ii. If, on the other hand, at least one side is free, we have a value of
12.69 for a SS-C-SS-F square plate (with ν = 0.3), a value of 11.68 for
a SS-SS-SS-F plate (ν = 0.3) and 9.63 for a SS-F-SS-F plate (ν = 0.3).
Note that for these last three cases – which also refer to the aspect
ratio a b = 1 – we specified within parenthesis the value of Poisson’s
ratio; this is because it turns out that the non-dimensional frequency
parameter ω a2 ρ h D does not directly depend on ν unless at least
one of the edges is free. It should be pointed out, however, that the
frequency does depend on Poisson’s ratio because the flexural stiffness
D contains ν .
iii. In general, the eigenfunctions of plates with BCs other that SS-SS-
SS-SS have more complicated expressions than the simple eigen-
functions of Equation 5.97. For example, the eigenfuntions of the
SS-C-SS-C plate mentioned above are
unm (x, y) = {( cosh sb − cos rb)( r sinh sy − s sin ry )

}
− ( r sinh sb − s sin rb )( cosh sy − cos ry ) sin α n x (5.100)
where α n = nπ a and m = 1, 2, is the index that labels the succes-

sive values of γ (hence ω ) that, for each n, satisfy Equations 5.98.
5.8.2 Circular plates
For circular plates, it is convenient to place the origin at the centre of the
( )
plate and express the two equations ∇2 + γ 2 u1 = 0 and ∇2 − γ 2 u2 = 0 ( )
in terms of polar co-ordinates r, θ , where x = r cos θ , y = r sin θ . Then,
since the equation for u1 is Helmholtz equation of Section 5.5, we can
proceed as we did there, that is, write u1 in the separated-variables form
u1(r , θ ) = f1(r)g1(θ ) and observe that (a) the periodicity of the angular part
of the solution implies g1(θ ) = cos nθ + sin nθ ( n = 1, 2,) and that (b)
f1(r) = A1 Jn (γ r) + B1Yn (γ r) because f1(r) is the solution of Bessel’s equation of
order n (we recall that Jn , Yn are Bessel’s functions of order n of the first and
second kind, respectively).
Passing now to the equation for u2 , we can rewrite it as ∇2 + (iγ )2  u2 = 0
and proceed similarly. By so doing, a solution in the ‘separated form’
u2 (r , θ ) = f2 (r)g2 (θ ) leads, on the one hand, to g2 (θ ) = cos nθ + sin nθ and, on
the other, to the equation for f2 (r)
d 2 f2 1 df2  2 n2 
+ − γ + 2  f2 = 0 (5.101)
dr 2 r dr  r 
which is known as the modified Bessel equation of order n. Its solution

is f2 (r) = A2 I n (γ r) + B2 Kn (γ r), where In (γ r), Kn (γ r) are called modified
(or hyperbolic) Bessel functions of order n of the first and second kind,
respectively.
Then, recalling that u(r , θ ) = u1(r , θ ) + u2 (r , θ ), the complete spatial solu-
tion is
u(r , θ ) = { A1 Jn (γ r) + B1Yn (γ r) + A2In (γ r) + B2Kn (γ r)} ( cos nθ + sin nθ ) (5.102)
where, for a plate that extends continuously across the origin, we must
require B1 = B2 = 0 because the functions of the second kind Yn , Kn become
unbounded at r = 0. So, for a full plate, we have the solution
u(r , θ ) = { A1 Jn (γ r) + + A2In (γ r)} ( cos nθ + sin nθ ) (5.103)
on which we must now enforce the BCs.

If our plate has radius R and is clamped at its boundary, the appropriate
BCs are
u(R, θ ) = 0, ∂ r u(R, θ ) = 0 (5.104)

and we obtain the frequency equation
 Jn (γ R) I n (γ R) 
det   =0 ⇒ Jn (γ R)I n′ (γ R) − I n (γ R) Jn′ (γ R) = 0 (5.105)
 Jn′ (γ R) I n′ (γ R) 
from which it follows that to each value of n there corresponds a countably

infinite number of roots, labelled with an index m ( m = 1, 2,). Equation
5.105 must be solved numerically, and if now we introduce the frequency
parameter λn m = γ n m R (and recall the relation γ 4 = ρhω 2 D), the plate nat-
ural frequencies can be written as
λ 2n m D
ω nm = (5.106)
R2 ρh
where some of the first values of λ 2 are
λ01
2
= 10.22, λ11
2
= 21.26, λ21
2
= 34.88,

λ02
2
= 39.77, λ31
2
= 51.04, λ12
2
= 60.82
Also, since from the first BC we get A2 = − A1 [ Jn (γ R) Jn (γ R)], the eigenfunc-

tions corresponding to the frequencies 5.106 are
 Jn (γ n m R) 
un m (r , θ ) = An m  Jn (γ n m r ) − In (γ n m r ) ( cos nθ + sin nθ ) (5.107)
 I n (γ nm R) 
where the constants An m depend on the choice we make for the normalisa-
tion of the eigenfunctions. From the results above, it is found that just like
for the circular membrane of Section 5.5.1, (a) the n, m th mode has n dia-
metrical nodes and m nodal circles (boundary included), and (b) except for
n = 0, each mode is degenerate.
BCs other than the clamped case lead to more complicated calculations,
and for this, we refer the interested reader to Leissa (1969).
5.8.3 On the orthogonality of plate eigenfunctions

Although in order to determine the orthogonality of plate eigenfunctions,
we could proceed as in Section 5.7.1; this would require some rather lengthy
manipulations. Adopting a different strategy, we can start from the fact
that in the static case, the equation ∇4 u = q D gives the plate deflection
under the load q. Now, if we let un m , ulk be two different eigenfunctions,
then the equations
∇4 unm = γ nm
4
unm , ∇4 ulk = γ lk4 ulk (5.108)
are identically satisfied, and we can therefore interpret the first and sec-
ond of Equations 5.108, respectively, by seeing un m as the plate deflection
under the load q1 = D γ nm 4
unm and ulk as the plate deflection under the load
q2 = D γ lk ulk. At this point, we invoke Betti's reciprocal theorem (or Betti’s
4
law) for linearly elastic structures (see, for example, Bisplinghoff, Mar and
Pian (1965)): The work done by a system of forces Q1 under a distortion
u(2) caused by a system Q2 equals the work done by the system Q2 under
a distortion u(1) caused by the system Q1. In mathematical form, this state-
ment reads Q1 u(2) = Q2 u(1).
So, returning to our problem, we observe that (a) the two loads are q1 , q2
as given above, (b) the deflection ulk is analogous to u(2), while the deflection
un m is analogous to u(1), and (c) we obtain a work expression by integrating
over the area of the plate S. Consequently, the equality of the two work
expressions gives
Dγ nm
4
∫
unm ulk dS = Dγ lk4
∫u ulk nm dS ⇒ (γ 4
nm − γ lk4 )∫u u dS = 0
nm lk
S S S
from which it follows, since γ n m ≠ γ lk,

∫u
S
u dS = 0 (5.109)
nm lk
which establishes the orthogonality of the eigenfunctions and can be com-

pactly expressed in inner product notation as un m ulk = 0.
5.9 A FEW ADDITIONAL REMARKS
5.9.1 Self-adjointness and positive-definiteness

of the beam and plate operators
We have seen in Section 5.4 that the most important and useful properties
of the eigenpairs of SLps – that is, real and non-negative eigenvalues, the
orthogonality of the eigenfunctions and the fact that they form a ‘basis’
of an appropriate (infinite-dimensional) linear space – are consequences
of the self-adjointness and positive-(semi)definiteness of the SL operator
under rather general and common types of BCs. This fact, however – as in
part anticipated by the developments on eigenfunctions orthogonality of
Sections 5.7.1 and 5.8.3 – is not limited to second-order SL operators, and
here we show that it applies also to fourth-order systems such as beams and
plates.
Starting with the beam operator
d2  d2 
Lb =  EI dx2  (5.110)
dx2
in which we assume that the bending stiffness EI may depend on x, let u, v

be two sufficiently regular functions for which the formula of integration
by parts holds. If now we go back to Section 5.7.1 and observe that by using
these functions instead of the eigenfunctions un , um, Equation 5.67 becomes
L L
∫ ∫
v ( EI u′′ )′′ dx = [ v(EI u′′)′ − u(EIv′′)′ + EI (v′′u′ − u′′v′)] 0 + (EIv′′)′′ u dx
L

0 0
(5.111)
then it can be easily verified that for the classical BCs (SS, C and F), the
boundary term within square brackets on the r.h.s. vanishes, thus implying
that Equation 5.111 reduces to the self-adjointness relation v Lb u = Lb v u .
Passing to positive-definiteness – which, we recall, is mathematically
expressed by the inequality u Lb u ≥ 0 for every sufficiently regular func-
tion u – we note that two integrations by parts lead to
L
u Lb u =
∫ u (EI u′′)′′dx
0

L L
= [ u(EI u′′)′ − EI u′u′′ ] 0 +

∫ EI (u′′) dx = ∫ EI (u′′) dx ≥ 0
L 2 2
(5.112)
0 0
where the third equality is due to the fact that the boundary term within
square brackets vanishes for the classical BCs and where the final inequality
holds because EI > 0 and (u′′)2 ≥ 0. Then, imposing the equality u′′ = 0 gives
u = C1x + C2, and since it is immediate to show that for the SS and C BCs, we
get C1 = C2 = 0, it follows that u Lb u = 0 if and only if u = 0, thus implying
that Lb is positive-definite. It is not so, however, for an F-F beam because in
this case the system is unrestrained and we know that there are rigid-body
modes (recall Case 4 in Section 5.7). In this case, therefore, the operator Lb
is only positive-semidefinite.
Things are a bit more involved for the plate operator Lp = ∇4 , where here
for simplicity (but without loss of generality for our purposes here), we con-
sider a uniform plate with constant flexural stiffness. In order to show that
Lp satisfies the relation u Lpv = Lp u v for many physically meaningful
types of BCs, let us start by observing that
( ) (
u ( Lpv ) = u∇4v = u∇2 ∇2v = u∇ ⋅ ∇∇2v (5.113) )
where in writing the last equality we used the fact that ∇2 = ∇ ⋅ ∇ (or
∇2 = div ( grad ) if we use a different notation). Then, recalling the vector
calculus relations
∇ ⋅ (u A) = ∇u ⋅ A + u∇ ⋅ A, ∇ ⋅ ( f ∇u ) = f ∇2 u + ∇f ⋅ ∇u (5.114)
we can set A = ∇∇2v in the first relation to get
( ) (
∇ ⋅ u∇∇2v = ∇u ⋅ ∇∇2v + u∇ ⋅ ∇∇2v ) ( ) ⇒ (
u∇ ⋅ ∇∇2v )
= ∇ ⋅ ( u∇∇ v ) − ∇u ⋅ ( ∇∇ v )
2 2
and use it in Equation 5.113 to obtain
( ) (
u ( Lpv ) = ∇ ⋅ u∇∇2v − ∇u ⋅ ∇∇2v (5.115) )
Now we set f = ∇2v in 5.1142 to obtain
( )
∇ ⋅ ∇2v∇u = ∇2v ∇2u + ∇∇2v ⋅ ∇u ⇒ ( )
∇u ⋅ ∇∇2v = ∇ ⋅ ∇2v ∇u − ∇2v ∇2u ( )
and substitute it in Equation 5.115. By so doing, integration over the two-
dimensional region/domain S occupied by the plate gives
∫ u (L v ) dS = ∫ ∇ ⋅ (u∇∇ v ) dS − ∫ ∇ ⋅ (∇ v∇u) dS + ∫ (∇ v∇ u) dS (5.116)

S
p
S
2
S
2
S
2 2
which, using the divergence theorem (see the following Remark 5.18) for
the first and second term on the r.h.s., becomes
∫ u (L v ) dS = ∫ u ∂ (∇ v ) dC − ∫ ∇ v ( ∂ u) dC + ∫ (∇ v∇ u) dS (5.117)
S
p
C
n
2
C
2
n
S
2 2
where C is the boundary/contour of S.
Remark 5.18
If S is a region/domain of space whose boundary/contour is C and A is a

smooth vector field defined in this region, we recall from vector calculus
that the divergence theorem states that

∫ ∇ ⋅ A dS = ∫ A ⋅ n dC
S C
where n is the outward normal from the boundary C. Explicitly, using this
theorem for the first and second term on the r.h.s. of Equation 5.116 gives,
respectively
∫ ∇ ⋅ (u∇∇ v ) dS = ∫ u (∇∇ v ) ⋅ n dC = ∫ u ∂ (∇ v) dC
S
2
C
2
C
n
2
∫ ∇ ⋅ (∇ v∇u) dS = ∫ ∇ v∇u ⋅ n dC = ∫ ∇ v (∂ u) dC
S
2
C
2
C
2
n
where the rightmost equalities are due to the fact that ∇() ⋅ n = ∂n (),
where ∂n denotes the derivative in the direction of n.
At this point, it is evident that by following the same procedure that we

used to arrive at Equation 5.116, we can obtain an expression similar to
Equation 5.117 with the roles of u and v interchanged. Then, subtracting
this expression from Equation 5.117 yields
∫ {u ( L v) − ( L u) v} dS
S
p p
(5.118)
=
∫ {u ∂ (∇ v) − v ∂ (∇ u) − ∇ v ∂ u + ∇ u ∂ v} dC
C
n
2
n
2 2
n
2
n
showing that the plate operator is self-adjoint if for any two functions u, v
that are four times differentiable the boundary integral on the r.h.s. of
Equation 5.118 vanishes – as it can be shown to be the case for the classical
BCs in which the plate is clamped, simply supported or free.
For positive-definiteness, we can consider Equation 5.117 with v replaced
by u, that is
∫ u(Lu) dS = ∫ {u ∂ (∇ u) − ∇ u ( ∂ u)}dC + ∫ (∇ u) dS
2 2 2 2
u Lp u = n n
S C S
from which it follows that u Lp u ≥ 0 because the contour integral vanishes

for the classical BCs. For clamped and simply-supported BCs, moreover, we
have the strict inequality u Lp u > 0.
5.9.2 Analogy with finite-DOFs systems

For the free vibration of (undamped) finite-DOFs systems, we have seen in
Chapter 3 that the fundamental equation is in the form of the generalised
eigenvalue problem Ku = λ Mu (for present convenience, note that here we
adopt a slight change of notation and denote by u the vector denoted by z
in Chapter 3; see, for example, Equation 3.24 and the equations of Section
3.3.1), where K, M are the system’s stiffness and mass matrices, λ = ω 2 is
the eigenvalue and u is a vector (or mode shape). If now we go back to
Equation 5.262 that governs the free longitudinal vibrations of a non-uni-

form rod – but the same applies to strings and shafts by simply using the
appropriate physical quantities – we can either see it as an SL problem of the
form 5.271 (with p(x) = EA(x), q(x) = 0, w(x) = µ(x) and λ = ω 2) or rewrite
it as
Ku = λ Mu (5.119)
where we introduced the rod stiffness and mass operators
d  d 
K=−  EA(x) dx  , M = µ(x) (5.120)
dx 
and where, as we know from Section 5.4, K is self-adjoint for many physi-
cally meaningful BCs. Equation 5.119 is the so-called differential eigen-
value problem, and the analogy with the finite-DOFs eigenproblem is
evident, although it should be just as evident that now – unlike the finite-
DOFs case in which the BCs are, let us say, ‘incorporated’ in the system’s
matrices – Equation 5.119 must be supplemented by the appropriate BCs for
the problem at hand.
Moreover, if we consider membranes, beams and plates, it is not difficult
to show that the differential eigenvalue form 5.119 applies to these systems
as well. In fact, for a non-uniform (Euler-Bernoulli) beam, we have
d2  d2 
K=  EI (x) dx2  , M = µ(x) (5.121)
dx2
with eigenvalue λ = ω 2, while for a uniform (Kirchhoff) plate, the stiffness

and mass operators are K = D ∇4 and M = σ (where σ = ρ h is the plate mass
density per unit area) and the eigenvalue is λ = ω 2.
Now, having shown in the preceding sections that the stiffness operators
of the systems considered above are self-adjoint, the point of this section is
that the analogy between Equation 5.119 and its finite-DOFs counterpart
turns out to be more than a mere formal similarity because the valuable
property of self-adjointness corresponds to the symmetry of the system’s
matrices of the discrete case (in this regard, also recall Remark 5.5(ii)).
In this light, therefore, if we denote by λn , φn the eigenvalues and mass-
normalised eigenfunctions of a self-adjoint continuous system, the counter-
parts of Equations 3.32 are
φn Mφm = δ n m , φn Kφm = λm δ n m (5.122)
and it is legitimate to expand any reasonably well-behaved function f in

terms of the system’s eigenfunctions as
f = ∑c φ
n =1
n n where cn = φn Mf (5.123)
Remark 5.19
i. In the cases considered above, the mass operator is not a differential

operator but simply a multiplication operator by some mass property
(such as µ(x) or σ ). This type of operator is obviously self-adjoint. In
more general cases in which M may be a true differential operator one
speaks of self-adjoint system if both K and M are self-adjoint.
ii. Strictly speaking, the convergence of the series 5.123 is the conver-
gence ‘in the mean’ of the space L2 (the Hilbert space of square-inte-
grable functions defined on the domain of interest, where, moreover,
the appropriate integral in this setting is the Lebesgue integral and not
the ‘usual’ Riemann integral). However, if f is continuously differen-
tiable a sufficient number of times – obviously, a number that depends
on the operators involved – and satisfies the BCs of the problem, the
series converges uniformly and absolutely.
Example 5.1
Pursuing the analogy with finite-DOFs systems, let us go back, for
example, to Section 3.3.4 and notice that Equation 3.57 rewritten
by using the ‘angular bracket notation’ for the inner product reads
∂λi = p i (∂ K − λi ∂ M ) p i , thereby suggesting the ‘continuous system
counterpart’
∂λi = φi (∂ K − λi ∂ M ) φi (5.124)
where φi ( i = 1, 2,) are the system’s mass-normalised eigenfunctions,

∂ K , ∂ M are small perturbations of the unperturbed stiffness and
mass operators K , M and ∂λi is the first-order perturbation of the ith
eigenvalue.
Using Equation 5.124, we can now consider the effect of shear defor-
mation on an SS-SS (uniform) Euler-Bernoulli beam. In order to do so,
we set λ = ω 2 and start by rewriting Equation 5.83 in the form
d 4 u(x)  µ rg2 E d 2 
EI = λ  µ − u(x) (5.125)
dx4  κ G dx2 
which when compared with EI u ′′′′ = λ µ u – that is, the eigenproblem

Ku = λ Mu for a uniform Euler-Bernoulli beam – shows that the per-
turbation operators in this case are
µ rg2 E d 2
∂M = − , ∂ K = 0 (5.126)
κ G dx2
Then, observing that the mass-normalised eigenfunctions of a SS-SS

( )
beam are φi (x) = 2 µ L sin ( i πx L ) , we can use this and Equation
5.126 in 5.124 to get
µ rg2 E d 2φ i
∂λi = − λi φi (∂ M)φ i = λi φi
κG dx2

2 2 2 L 2
2i π r E  iπ x  2 2
i π E  rg 
∫ sin
g
= − λi 2
  dx = − λi (5.127)
κ GL 3
L  κ G  L 
0
To the first order, therefore, the ith perturbed eigenvalue is λˆ i = λi + ∂λi ,

and we have
λˆ i
2
i 2 π 2 E  rg 
= 1− (5.128)
λi κ G  L 
which must be compared with the exact result of Equation 5.85. And
since this equation, when appropriately modified for our present pur-
poses, reads
−1
λˆ i(shear)  i 2 π 2 E  rg  
2
= 1 +  (5.129)
λi  κ G  L  
we can recall the binomial expansion (1 + x)−1 = 1 − x +  to see that

Equation 5.128 is the first-order expansion of Equation 5.129. This
shows that the perturbative approximation is in good agreement with
the exact calculation whenever the effect of shear is small – which is
precisely the assumption under which we expect Equation 5.124 to
provide satisfactory results.
5.9.3 The free vibration solution

In light of the preceding developments, if we denote by x the set of (one or
more) spatial variables of the problem and by K , M the system’s stiffness and
mass operators – so that, for example, K = − EA ∂ 2 ∂ x2 , M = µ for a uni- ( )
( )
form rod, K = EI ∂ ∂ x , M = µ for a uniform beam, or K = −T ∇2 , M = σ
4 4
for a membrane – the free vibration equation of motion can be written as
∂2 w(x, t)
Kw(x, t) + M = 0 (5.130)
∂t 2
Assuming that the system’s natural frequencies and mass-normalised eigen-

functions ω j , φ j (x) are known and that, as is typically the case in practice,
we can expand w(x, t) in terms of eigenfunctions as
∞
w(x, t) = ∑ y (t)φ (x) (5.131)

i =1
i i
substitute it into Equation 5.130 and then take the inner product of the
resulting equation with φ j (x). This gives
i
y i φ j Kφ i +
i
∑
yi φ j Mφi = 0 ∑
and, consequently, owing to the orthogonality of the eigenfunctions
(Equations 5.122),
yj (t) + ω 2j y j (t) = 0 ( j = 1, 2,…) (5.132)

which is a countably infinite set of uncoupled 1-DOFs equations. Then,
since we know from Chapter 3 that the solution of Equation 5.132 for the
jth modal (or normal) co-ordinate is
y j (0)
y j (t) = y j (0) cos ω j t +
ωj
sin ω j t ( j = 1, 2,…) (5.133)
the only thing left to do at this point is to evaluate the initial (i.e. at t = 0)
 0 = ∂ t w(x,0) for the
conditions y j (0), y j (0) in terms of the ICs w0 = w(x,0), w
function w(x, t). The result is
y j (0) = φ j Mw0 , y j (0) = φ j Mw

0 ( j = 1, 2,…) (5.134)
and it is readily obtained once we observe that from Equation 5.131, it
follows
w0 = ∑ y (0)φ (x)
i
i i
⇒
Mw0 = ∑ y (0)Mφ (x)
i
i i
⇒
φ j Mw0 = y j (0)
w 0 = ∑ y (0)φ (x)
i i Mw 0 = ∑ y (0)Mφ (x)
i i
φ j Mw 0 = y j (0)
i i
where in writing the last expressions we took the orthogonality conditions

into account.
Finally, the solution of Equation 5.130 is obtained by inserting Equation
5.133 (with the ICs of Equation 5.134) into Equation 5.131. By so doing,
we get
∞
 φ j Mw 0 
w(x, t) = ∑  φ
j =1
j Mw0 cos ω j t +
ωj
sin ω j t  φ j (x) (5.135)

in analogy with the finite-DOFs Equation 3.35a. Note that the function
w(x, t) automatically satisfies the BCs because so do the eigenfunctions φ j .
Remark 5.20
Since for a uniform string fixed at both ends we have K = −T0 ∂ xx2
, M = µ and
the mass-normalised eigenfunctions φ j (x) = ( 2 µ L )sin ( jπ x L ), the reader
is invited to re-obtain the solution of Equations 5.10 and 5.13 (Section
5.2.2) by using Equation 5.135. Although it is sufficiently clear from the
context, note that in that section the field variable is denoted by y(x, t) (it is
not a modal coordinate), while here we used the symbol w(x, t).
5.10 FORCED VIBRATIONS: THE MODAL APPROACH
Besides its theoretical interest, the possibility of expressing the response of

a vibrating system to an external excitation as a superposition of its eigen-
modes is important from a practical point of view because many experi-
mental techniques – for example, EMA (acronym for ‘experimental modal
analysis’) and, more recently, OMA (‘operational modal analysis’) – that
are widely used in different fields of engineering rely on this possibility.
Given this preliminary consideration, let us now denote by f (x, t) an exter-
nal exciting load; then the relevant equation of motion can be written as
∂2 w(x, t)
Kw(x, t) + M = f (x, t) (5.136)
∂t 2
where K , M are the appropriate system’s stiffness and mass operators.

Under the assumption that the eigenpairs ω j , φ j (x) are known, we now write
the expansion 5.131 and substitute it in Equation 5.136. Taking the inner
product of the resulting equation with φ j (x) and using the eigenfunctions
orthogonality conditions leads to
yj (t) + ω 2j y j (t) = φ j f ( j = 1, 2,…) (5.137)

which, as in the preceding section on free vibration, is a countably infi-
nite set of uncoupled 1-DOFs equations. And since from previous chap-
ters, we know that that the solution of Equation 5.137 for the jth modal
coordinate is
t
y j (0) 1
y j (t) = y j (0) cos ω j t +
ωj
sin ω j t +
ωj ∫φ
0
j f sin ω j ( t − τ )  dτ (5.138)
the solution of Equation 5.136 is obtained by inserting Equation 5.138

into Equation 5.131, where the jth initial modal displacement and velocity
y j (0), y j (0) are obtained by means of Equation 5.134.
Then, if we further recall from Chapter 4 that the (undamped) modal
IRF hˆ j (t) = ω j−1 sin ω j t is the response of the modal coordinate y j to a Dirac
delta input δ (t) (see Appendix B, Section B.3); for zero ICs, Equation 5.138
t
can be written as y (t) = j
∫
φ f hˆ (t − τ ) dτ , so that, by Equation 5.131, the
0
j j
response in physical coordinates is given by the superposition
∞ t
w(x, t) = ∑j =1
φ j (x)
∫φ j f hˆ j (t − τ ) dτ (5.139)
0
where φ j f =
∫ R
φ j (x)f (x, t)dx and R is the spatial region occupied by the
system. In this regard, note that the jth mode does not contribute to the
response if φ j f = 0.
If now we let the excitation be of the form f (x, τ ) = δ ( x − xk )δ (τ ) – that
is, a unit-amplitude delta impulse applied at point x = xk and at time τ = 0 –
then φ j f = φ j ( xk ) δ (τ ) and it follows from Equation 5.139 that the output
at point x = xm is
∞
w ( xm , t ) = ∑φ ( x
j =1
j m )φ j ( xk ) hˆ j (t) (5.140)
But then since by definition the physical coordinate IRF h ( xm , xk , t ) is the

displacement response at time t and at point xm to an impulse applied at
point xk at time t = 0, then Equation 5.140 shows that
∞
h ( xm , xk , t ) = ∑ φ (x
j =1
j m )φ j ( xk ) hˆ j (t) (5.141)
which can be compared with Equation 4.562 of the finite-DOFs case.

Turning to the frequency domain, for a distributed harmonic excitation
of the form f (x, t) = F(x) e iω t and a corresponding harmonic modal response
(
y j = Yj e iω t , Equation 5.137 gives Yj ω 2j − ω 2 = φ j F and, consequently, )
yj = φj F e iω t
ω −ω ( 2
j
2 −1
) . By Equation 5.131, therefore, the response in
physical coordinates is
∞ ∞
φj F
w(x, t) = ∑ (ω
j =1
2
j −ω2 )
φ j (x) e iω t = ∑φ j =1
j F Hˆ j (ω )φ j (x)e iω t (5.142a)
where φ j F =
∫ F(r)φ (r) dr and R is the spatial region occupied by the
R
j
vibrating system, and in the last relation, we recalled from previous chap-
( )
−1
ters that Hˆ j (ω ) = ω 2j − ω 2 is the (undamped) modal FRF.
In particular, if a unit amplitude excitation is applied at the point x = xk ,
then f (x, t) = δ ( x − xk ) e iω t and φ j F = φ j ( xk ), and consequently
∞
w(x, t) = ∑φ (x)φ ( x ) Hˆ (ω ) e
j =1
j j k j
iω t
(5.142b)
which, when compared with the relation w(x, t) = H ( x, xk , ω ) e iωt that

defines the physical coordinates FRF H ( x, xk , ω ) between the point x and
xk, shows that
φ j (x)φ j ( xk )
∞ ∞
H ( x, xk , ω ) = ∑ j =1
φ j (x)φ j ( xk ) Hˆ j (ω ) = ∑
j =1
ω 2j − ω 2
(5.143)
so that the response at, say, the point x = xm is

∑
∞
H (xm , xk , ω ) = φ j (xm )φ j (xk ) Hˆ j (ω ) , whose (damped) counterpart for
j =1
discrete systems is given by Equation 4.59b. Clearly, the FRFs above are
given in the form of receptances but the same applies to mobilities and
accelerances.
Example 5.2
Consider a vertical clamped-free rod of length L, uniform mass per
unit length µ and uniform stiffness EA subjected to an excitation in
the form of a vertical base displacement g(t). In order to determine
the system’s response, we first write the longitudinal rod displacement
w(x, t) as
w(x, t) = u(x, t) + g(t) (5.144)
where u(x, t) is the rod displacement relative to the base. Substituting

Equation 5.144 in the equation of motion Kw + M ∂2tt w = 0 (where
we recall that for a uniform rod the stiffness and mass operators are
K = − EA ∂ xx
2
and M = µ ) gives
∂2 u ∂2 u d2g
− EA 2
+ µ 2 = − µ 2 (5.145)
∂x ∂t dt
where the r.h.s. is an ‘effective force’ that provides the external excita-
tion to the system, and we can write feff = − µ g.  Assuming the system
to be initially at rest, the jth modal response is given by Equation 5.138
with y j (0) = y j (0) = 0. Calculating the angular bracket term by taking

into account the fact that the system’s mass-normalised eigenfunctions
are given by Equation 5.242 with the normalisation constant 2 µ L ,
we get
L
2 g(t) 2µ L

∫
φ j feff = − µ g(t) φ j (x) dx = −
0
(2 j − 1)π
and consequently, by Equation 5.139, we obtain the response in physi-

cal coordinates as
∞ t
φ j (x)
u(x, t) = −2 2µ L ∑j =1
(2 j − 1)πω j ∫ g(τ ) sin ω (t − τ ) dτ (5.146)
j
0
where the natural frequencies are given by Equation 5.241 and the time
integral must be evaluated numerically if the base motion g(t) and its
corresponding acceleration g(t) are not relatively simple functions of
time. Then, the total response w(x, t) is obtained from Equation 5.144.
As an incidental remark to this example, it should be noticed that
here we used a ‘standard’ method in order to transform a homogeneous
problem with non-homogeneous BCs into a non-homogeneous prob-
lem with homogeneous BCs. In fact, suppose that we are given a linear
operator B (B = − EA ∂2xx + µ ∂tt2 in the example), the homogeneous equa-
tion Bw = 0 to be satisfied in some spatial region R and the non-homo-
geneous BC w = y on the boundary S of R. By introducing the function
u = w − v, where v satisfies the given BC, the original problem becomes
the following problem for the function u: the non-homogeneous equa-
tion Bu = − By in R with the homogeneous BC u = 0 on S.
Example 5.3
Let us now consider a uniform clamped-free rod of length L and mass
per unit length µ excited by a load of the form f (x, t) = p(t) δ (x − L) at
the free end. Then, we have
2  ( 2 j − 1) π  j −1 2
φ j f = p(t) φ j (L) = p(t)
µL
sin 
2  = (−1) p(t) µ L (5.147)
 
where in writing the last expression, we used the relation
sin [(2 j − 1)π 2] = (−1) j −1. If we assume the system to be initially at rest,
Equation 5.138 gives
t
(−1) j −1 2
y j (t) =
ωj µL ∫ p(τ ) sin ω (t − τ ) dτ (5.148)
0
j
and if now we further assume that p(t) = θ (t), where θ (t) is a unit ampli-
tude Heaviside step function (i.e. θ (t) = 0 for t < 0 and θ (t) = 1 for t ≥ 0),
the calculation of the integral in Equation 5.148 gives ω j−1 (1 − cos ω j t )

and consequently
(−1) j −1 2
y j (t) =
ω 2j µL
(1 − cos ω j t )
thus implying that the displacement in physical coordinates is given by
the superposition
∞ j −1
w(x, t) =
2
µL ∑ (−ω1)
j =1
2
j
(1 − cos ω t ) φ (x)
j j

∞ j −1
=
8L
π 2 EA ∑ (2(−j1)− 1)
j =1
2 (1 − cos ω t ) sin  (2 j −2L1)π x 
j (5.149)
where in the second expression, we (a) used Equation 5.241 in order to

write the explicit form of ω 2j at the denominator of the first expression
and (b) recalled that the mass-normalised eigenfunctions φ j (x) are given
by Equation 5.242 with the normalisation constant 2 µ L.
Example 5.4
In the preceding Example 5.3, the solution 5.149 has been obtained by
using the modal approach ‘directly’, that is, by considering the rod as
a ‘forced’ or non-homogeneous system (that is, with the non-zero term
f = p(t) δ (x − L) on the r.h.s. of the equation of motion) subjected to the
‘standard’ BCs of a rod clamped at x = 0 and free at x = L. However,
since the external excitation is applied at the boundary point x = L, the
same rod problem is here analysed with a different strategy that is in
general better suited for a system excited on its boundary. In order to
do so, we start by considering the rod as an excitation-free (i.e. homo-
geneous) system subject to non-homogeneous BCs, so that we have the
equation of motion and BCs
∂2 w ∂2 w
− EA 2
+ µ 2 = 0; w(0, t) = 0, EA ∂ x w(L, t) = p(t) (5.150)
∂x ∂t
where here only the BC at x = L (Equation 5.1503) is not homogeneous.

Now we look for a solution of the form
w(x, t) = u(x, t) + r(x)p(t) = u(x, t) + upst (x, t) (5.151)
where the term upst (x, t) = r(x)p(t), often called ‘pseudo-static’ displace-
ment (hence the subscript ‘pst’), satisfies the equation EA ∂ xx 2
upst = 0
and is chosen with the intention of making the BCs for u(x, t) homo-
geneous. Considering first the BCs, we note that with a solution of the
form 5.151, the BCs of Equations 5.150 imply
u(0, t) = − r(0)p(t), EA ∂ x u(L, t) = p(t) [1 − EA r ′(L)]
which in turn means that if we want the BCs for u(x, t) to be homoge-
neous, we must have r(0) = 0 and r ′(L) = 1 EA. Under these conditions,
in fact, we have
u(0, t) = 0, EA ∂ x u(L, t) = 0 (5.152)
If now, in addition, we take into account that EA ∂ xx 2

upst = 0 implies
r(x) = C1 x + C2 and that, consequently, the conditions on r(x) give
C2 = 0, C1 = 1 EA, then we have r(x) = x EA.
Turning to the equation of motion, substitution of the solution 5.151
into Equation 5.1501 leads to the equation for u(x, t) , which is
∂2 u ∂2 u
− EA 2
+ µ 2 = feff (x, t) where
∂x ∂t
(5.153)
d2p d 2r µ x d2p
feff (x, t) = − µ r(x) 2 + EA p(t) 2 = −
dt dx EA dt 2
and where in writing the last expression we took into account that
r(x) = x EA and that, consequently, d 2 r dx2 = 0.
We have at this point transformed the problem 5.150 into the non-
homogeneous problem 5.153 with the homogeneous BC 5.152, where
the forcing term feff accounts for the boundary non-homogeneity of
Equation 5.1503. Now we apply the modal approach to this new problem
by expanding u(x, t) in terms of eigenfunctions as u(x, t) = ∑ j y j (t)φ j (x)
and calculating the modal response y j as prescribed by Equation 5.138.

Assuming zero ICs, this gives
d2p  µ x 
t t L
y j (t) =
∫
0
φ j feff hˆ j (t − τ ) dτ = −
∫0

dt 2  EA ∫
φ
0
j (x) dx  hˆ j (t − τ ) dτ

(5.154a)
and since the integral within parenthesis yields
L
µ 2  (2 j − 1)π x  µ 2 (−1) j −1 4L2

EA µL ∫ x sin 
0
2L
 dx =
EA µ L (2 j − 1)2 π 2
Equation 5.154a becomes
t
µ 2 (−1) j −1 4L2 d 2 p(τ ) ˆ
y j (t) = −
EA µ L (2 j − 1)2 π 2 ∫
0
dτ 2
hj (t − τ ) dτ (5.154b)
Now, assuming p(t) = θ (t) (as in the preceding Example 5.3) and
recalling the properties of the Dirac delta of Equations B.411 and B.42
in Appendix B, the time integral gives
t t
d 2θ (τ ) ˆ dδ (τ ) ˆ d hˆ j (τ )

∫0
dτ 2
hj (t − τ ) dτ =
∫ 0
dτ
hj (t − τ ) dτ = −
dτ
τ =t
= cos ω j t
and therefore
∞ j −1
µ
∑ ((2−j1)− 1)4Lπ
2
2
u(x, t) = − cos ω j t
EA µL j =1
2 2
from which it follows, owing to Equation 5.151,
∞ j −1
w(x, t) =
x
−
8L
EA π 2 EA ∑ (2(−j1)− 1)
j =1
2
 (2 j − 1)π x 
sin 
 2L
 cos ω j t (5.155)
This is the solution that must be compared with Equation 5.149. In

order to do so, we first observe that the calculation of the product in
Equation 5.149 leads to
∞ j −1
w(x, t) =
8L
π 2 EA ∑ (2(−j1)− 1)
j =1
2
 (2 j − 1)πx 
sin 
 2L


∞ j −1
−
8L
π 2 EA ∑ (2(−j1)− 1)
j =1
2
 (2 j − 1)π x 
sin 
 2L
 cos ω j t (5.156)
and then consider the fact that the Fourier expansion of the function
π 2 x 8L is (we leave the proof to the reader)
∞ j −1

π2 x
8L
= ∑ (2(−j1)− 1)
j =1
2
 (2 j − 1)π x 
sin 
 2L

which, when used in Equation 5.156, shows that the first term of this
equation is exactly the term x EA of Equation 5.155, i.e. the function
r(x) of the pseudo-static displacement. So, the two solutions are indeed
equal, but it is now evident that the inclusion of the pseudo-static term
from the outset makes the series 5.155 much more advantageous from
a computational point of view because of its faster rate of convergence.
In actual calculations, therefore, less terms will be required to achieve
a satisfactory degree of accuracy.
Example 5.5
In modal testing, the experimenter is often interested in the response
of a system to an impulse loading at certain specified points. So, if
the same rod of Example 5.3 is subjected to a unit impulse applied at

the point x = L at time t = 0, we can set p(τ ) = δ (τ ) in Equation 5.148
( )
and obtain y j (t) = (−1) j −1ω j−1 2 µ L sin ω j t . From Equation 5.131, it
follows that the physical coordinate response measured, for example,
at the same point x = L at which the load has been applied is
∞ ∞
(−1) j −1 sin ω j t sin ω j t
w(L, t) =
2
µL ∑ j =1
ωj
φ j (L) =
2
µL ∑
j =1
ωj
(5.157)
where in writing the second expression, we took into account the

relations
( )
φ j (L) = 2 µ L sin [(2 j − 1)π 2] and sin [(2 j − 1)π 2] = (−1) . But
then, since under these conditions w(L, t) is the response at point
j −1
x = L to an impulse load applied at the same point, we must have

w(L, t) = h(L, L, t). The fact that it is so is easily verified by noticing that
∑
∞
Equation 5.141 gives h(L, L, t) = φ 2 (L)hˆ (t), which is in fact the
j j
j =1
same as Equation 5.157 when one considers the explicit expressions of
φ j (L) and hˆ j (t).
If, on the other hand, we are interested in the receptance FRF at
x = L , Equation 5.143 gives
∞ ∞
φ j2 (L)
H (L, L, ω ) = ∑ω
j =1
2
j − ω 2
=
µ
2
L ∑ω j =1
2
j
1
− ω2
(5.158)
which, in light of preceding chapters, we expect to be the Fourier

transform of h(L, L, t) . Since, however, the Fourier transform of the
undamped IRF 5.157 does not exist in the ‘classical’ sense, we use
the trick of first calculating its Laplace transform and then – recalling
that the Laplace variable is s = c + iω – passing to the limit as c → 0.
( ) ( )
−1 −1
And since ω j−1 L sin ω j t  = s2 + ω 2j , in the limit we get ω 2j − ω 2 .
Obviously, the trick would not be necessary if the system has some
amount of positive damping that prevents the divergence of the FRF
at ω = ω j .
Example 5.6
As a simplified model of a vehicle travelling across a bridge deck, con-
sider the response of an SS-SS Euler-Bernoulli beam to a load of con-
stant magnitude P moving along the beam at a constant velocity V.
Under the reasonable assumption that the mass of the vehicle is small
in comparison with the mass of the bridge deck (so that the beam eigen-
values and eigenfunctions are not appreciably altered by the presence of
the vehicle), we write the moving load as
 P δ (x − V t) 0≤t ≤LV
f (x, t) =  (5.159)
 0 otherwise
and obtain the angular bracket term in Equation 5.138 as
L
2  j πV t 

∫
φ j f = P φ j (x) δ ( x − Vt ) dx = P
0
µL
sin 
 L 
where in writing the last expression, we recalled that the mass-

normalised eigenfunctions of an SS-SS Euler-Bernoulli beam are
( )
φ j (x) = 2 µ L sin ( j πx L ).
Then, assuming the beam to be initially at rest, we have
t
P 2  jπVt 
y j (t) =
ωj µL ∫ sin 
0
L 
 sin ω j (t − τ )  dτ (5.160a)
which gives, after two integrations by parts (see the following Remark
5.21 for a hint),
2  L2   jπ V  jπ V t  
y j (t) = P
2 2 
sin ω j t − sin 

µL  j π V − ω j L   ω jL
2 2 2
 L  
(5.160b)
thus implying, by virtue of the expansion of Equation 5.131, that the

displacement response in physical coordinates is
L2 sin ( jπx L )  jπ V
∞
 jπ V t  
w(x, t) =
2P
µL ∑
j =1
2 2 
j π V − ω j L  ω jL
2 2 2
sin ω j t − sin 
 L  
(5.161)
which in turn shows that resonance may occur at the ‘critical’ values
of velocity
Lω j jπ EI
Vj(crit) =
jπ
=
L µ
( j = 1, 2, ) (5.162)
where the last expression follows from the fact that the beam natu-
ral frequencies – we recall from Section 5.7 – are ω j = ( j π L ) EI µ .
2
At these values of speed, the transit times are t j = L Vj(crit)

= j π ω j , so
that t j = T 2 j where T = 2π ω 1 is the fundamental period of the beam
vibration.
Also, note that if our interest is, for instance, the deflection at
mid-span, the fact that φ j (L 2) = 0 for j = 2, 4, 6, shows that the
even-numbered (antisymmetric) modes provide no contribution to the
deflection at x = L 2.
Remark 5.21
If, for brevity, we call A the integral in Equation 5.160a and define a j = jπV L ,
( )
two integrations by parts lead to A = (1 a j ) sin ω j t − ω j a2j sin a j t + ω 2j A a2j ,
from which Equation 5.160b follows easily.
5.10.1 Alternative closed-form for FRFs

An alternative approach for finding a closed-form solution of the FRF of a
continuous system is based on the fact that with a harmonic excitation of
the form f (x, t) = F(x, ω ) e i ω t, the steady-state response of a stable and time-
invariant system is also harmonic and can be written as w(x, t) = W (x, ω ) e i ω t .
On substituting these relations into the equation of motion, the exponen-
tial factor cancels out and we are left with a non-homogeneous differential
equation for the function W (x, ω ) supplemented by the appropriate BCs.
Then, observing that, by definition, the FRF H ( x, xk , ω ) is the multiplying
coefficient of the harmonic solution for the system’s response at point x due
to a local excitation of the form F(x, ω ) = δ ( x − xk ) applied at point x = xk;
with this type of excitation, we have
H ( x, xk , ω ) = W (x, ω ) (5.163)
By this method, the solution is not in the form of a series of eigenfunctions

and one of the advantages is that it can be profitably used in cases in which
a set of orthonormal functions is difficult to obtain or cannot be obtained.
Example 5.7
As an example, let us reconsider the uniform clamped-free rod of
Example 5.3 by assuming a harmonic excitation of unit ampli-
tude at the free end x = L, so that we have the equation of motion
− EA ∂2xx w + µ ∂2tt w = 0 for x ∈ (0, L) plus the BC w(0, t) = 0 at x = 0 and
the force condition EA ∂ x w(L, t) = δ (x − L) e iω t at x = L. Proceeding as
explained above and defining γ 2 = µω 2 EA, we are led to
W ′′ + γ 2W = 0; W (0, ω ) = 0, EAW ′(L, ω ) = 1 (5.164)
where the second BC (Equation 5.1643) follows from the fact that the
force excitation has unit amplitude. At this point, since the solution
of Equation 5.1641 is W (x, ω ) = C1 cos γ x + C2 sin γ x, enforcing the BCs
gives C1 = 0 and C2 = (γ EA cos γ L ) , and consequently
−1
1
W (x, ω ) = H (x, L, ω ) = sin γ x (5.165)
γ EA cos γ L
which – having neglected damping – becomes unbounded when

cos γ L = 0 (not surprisingly, because we recall from Section 5.3 that
cos γ L = 0 is the frequency equation for a clamped-free rod). The FRF
of Equation 5.165 must be compared with the FRF that we obtain
from Equation 5.143, that is, with the expansion in terms of the mass-
normalised eigenfunctions φ j (x)
∞ j −1
H (x, L, ω ) =
2
µL ∑ ω(−1)− ω
j =1
2
j
2
 (2 j − 1)πx  (5.166)
sin 
 2L 
where in writing this expression we recalled that the eigenpairs of a

clamped-free rod are given by Equation 5.24 and also used the relation
sin [(2 j − 1)π 2] = (−1) j −1. At first sight, it is not at all evident that the two
FRFs are equal, but the fact that it is so can be shown by calculating the
inner product φ j W , where
L
2 1  (2 j − 1)πx 
φj W =
µ L γ EA cos γ L ∫ sin  2L
 sin γ x dx
0
L
2 1  µ 
=
µ L γ EA cos γ L ∫ sin  ω x
0
j
EA 
sin γ x dx
and where in writing the second equality, we observed that Equation

5.241 implies (2 j − 1)π 2L = ω j µ EA . Leaving to the reader the details
of the calculation (in which one must take into account the relations
cos [(2 j − 1)π 2] = 0, sin ( 2 j − 1) π 2  = (−1) j −1 and γ = ω µ EA), the
result is
2 (−1) j −1
φj W = (5.167)
(
µ L µ ω 2j − ω 2 )
This in turn implies
∞ ∞
W (x, ω ) = ∑
j =1
µ φ j W φ j (x) = ∑φ
j =1
j W µ
φ j (x) (5.168)
and tells us that the r.h.s. of Equation 5.166 is the series expansion of
the function W (x, ω ) of Equation 5.165.
If now, for example, we are interested in the response at x = L,
Equation 5.165 gives
tan γ L 1  µ 
W (L, ω ) = H (L, L, ω ) = = tan  ω L (5.169a)
γ EA ω µ EA  EA 
which must be compared with the FRF that we obtain from Equation
5.143, that is, with
∞
H (L, L, ω ) =
2
µL ∑ω
j =1
2
j
1
− ω2
(5.169b)
In order to show that the two FRFs of Equations 5.169a and b are the
same, the reader is invited to do so by setting θ = 2γ L π in the standard
result found in mathematical tables
∞
πθ 2θ 2θ 2θ

π
2
tan
2
= + +
1 − θ 2 32 − θ 2 52 − θ 2
+ = ∑ (2 j − 12θ) − θ
j =1
2 2
Example 5.8
As a second example of the method, consider the longitudinal vibra-
tions of a vertical rod subjected to a support harmonic motion of unit
amplitude. For this case, the equation of motion Kw + µ ∂ tt2 w = 0 and
the appropriate BCs are
∂2 w ∂2 w
− EA 2
+ µ 2 = 0; w(0, t) = e iωt , ∂ x w(L, t) = 0 (5.170)
∂x ∂t
so that assuming a solution in the form w(x, t) = W (x, ω ) e i ω t leads to

the equation and BCs for W
W ′′ + γ 2W = 0; W (0, ω ) = 1, W ′(L, ω ) = 0 (5.171)
where γ 2 = µ ω 2 EA and primes denote derivatives with respect to x.

Then, enforcing the BCs on the solution W (x, ω ) = C1 cos γ x + C2 sin γ x
of Equation 5.1711 gives C1 = 1, C2 = tan γ L and therefore
W (x, ω ) = H (x, 0, ω ) = cos γ x + (tan γ L) sin γ x (5.172)
5.10.2 A note on Green’s functions

Again, we start from the fact that the steady-state response of a s table,
time-invariant system to a harmonic forcing function of the form
f (x, t) = F(x) e iω t is also harmonic with form w(x, t) = u(x) e iω t. Substitution
of these relations in the equation of motion Kw + M ∂2tt w = f (x, t) then leads
to the non-homogeneous equation Ku − ω 2Mu = F(x) for the function u(x),
which, setting λ = ω 2 and defining the operator L = K − λ M, becomes
Lu(x) = F(x) (5.173)
where, typically, we have seen that the operator L is self-adjoint for a large
class of systems of our interest.
Now, leaving mathematical rigour aside for the moment, let us make some
formal manipulations and suppose that we can find an operator L−1 such
that L−1 L = LL−1 = I , where I is the identity operator. Then L−1 Lu = L−1 F
and the solution of problem 5.173 is u = L−1 F . Moreover, since L is a dif-
ferential operator, it is eminently reasonable (and correct) to expect that L−1
is an integral operator and that the solution u will be expressed in the form
u = L−1F =
∫ G(x, ξ)F(ξ) dξ (5.174)
where G(x, ξ ) is a function to be determined that depends on the problem
at hand. As for terminology, in the mathematical literature, G(x, ξ ) is called
the Green’s function of the differential operator L and is – using a com-
mon term from the theory of integral equations – the kernel of the integral
operator L−1.
In light of these considerations, let us proceed with our ‘free’ manipula-
tions. If now we write

∫
F(x) = LL−1F = L G(x, ξ )F(ξ ) dx =
∫ {LG(x, ξ)} F(ξ) dξ (5.175)
and recall the defining property of the Dirac delta function (see Section B.3
of Appendix B), then the last expression in Equation 5.175 shows that G
must be such that
LG(x, ξ ) = δ (x − ξ ) ⇒ {K − λ M}G(x, ξ) = δ (x − ξ ) (5.176)

from which we deduce the physical significance of G(x, ξ ); it is the solution
of the problem for a unit amplitude delta excitation applied at the point
x = ξ.
Remark 5.22
We have pointed out in Appendix B that although the Dirac delta is not a
function in the ordinary sense but a so-called distribution (or generalised
function), our interest lies in the many ways in which it is used in applica-
tions and not in the rigorous theory. The rather ‘free’ manipulations above,
therefore, are in this same spirit, and a more mathematically oriented reader
can find their full justification in the theory of distributions (see References
given in Remark B.6 of Appendix B).
The problem at this point is how to determine the Green’s function, and
here we illustrate two methods. For the first method, we can go back to
Section 5.10 and recall that Equation 5.143 gives the system’s response at
the point x to a local delta excitation applied at xk. Therefore, with a simple
change of notation in which we set xk = ξ , it readily follows that that same
equation provides the expression of the Green’s function in the form of a
series expansion in terms of the system’s eigenfunctions, that is,
∞
φ j (x)φ j (ξ )
G(x, ξ ) = ∑
j =1
λj − λ
(5.177)
In this light, in fact, the system’s response of Equation 5.142a can be rewrit-
ten as
∞   φ (x)
w(x, t) = u(x) e i ωt = e i ωt ∑∫
j =1
 φ j (ξ )F(ξ ) d ξ 

j
 λ j − λ
R

 ∞
φ j (x)φ j (ξ ) 
= e i ωt 

R
∫∑j =1
λ j − λ  ∫
 F(ξ ) d ξ = e iωt G(x, ξ )F(ξ ) d ξ
R
in agreement with Equation 5.174, and where the Green’s function is given
by Equation 5.177.
For the second method, we refer to a further development – in addition to
the ones considered in Section 5.4 – on SLps. For the non-homogeneous SLp
Lu ≡ − [ p(x) u′ ]′ + q(x) u = F(x)

(5.178)
α 1u(a) + β1u′(a) = 0, α 2u(b) + β2u′(b) = 0
in fact, we have the following result (see, for example, Debnath and
Mikusinski (1999) or Vladimirov (1987)):
Provided that λ = 0 is not an eigenvalue of Lu = λ u subject to the BCs
5.1782 and 5.1783, the Green’s function of the problem 5.178 is given by
1  u1(x) u2 (ξ ) (a ≤ x < ξ )
G(x, ξ ) = −  (5.179)
p(x)W (x)  u2 (x) u1(ξ ) (ξ < x ≤ b)

where u1 is a solution of the homogeneous problem Lu = 0 satisfying the BC

at x = a, u2 is a solution of the homogeneous problem Lu = 0 satisfying the
BC at x = b and W (x) is the so-called Wronskian of u1 , u2, that is,
 u1(x) u2 (x) 
W (x) ≡ det   = u1(x) u2′ (x) − u2 (x) u1′ (x)
 u1′ (x) u2′ (x) 
Given the Green’s function 5.179, the solution of the problem 5.178 is then
b
u(x) =
∫ G(x, ξ) F(ξ) dξ (5.180)
a
Example 5.9
For a uniform finite string fixed at both ends under the action of a
distributed external load f (x, t), we know that we have the equation of
motion Kw + M ∂2tt w = f (x, t) with K = −T0 ∂2xx and M = µ . Then, assum-
ing an excitation and a corresponding response of the harmonic forms
given above, we get T0 u ′′(x) + µω 2 u(x) = − F(x), from which it follows
u ′′(x) + k2 u(x) = δ (x − ξ ) where k2 = ω 2 c2 = ω 2 µ T0 (5.181)
if the excitation is of the form F(x) T0 = −δ (x − ξ ), i.e. a unit amplitude

impulse applied at the point x = ξ (the minus sign is for present con-
venience in order to have a positive impulse on the r.h.s. of Equation
5.1811), where ξ ∈ (0, L). Now, since δ (x − ξ ) = 0 for x ≠ ξ , we have two
solutions for the homogeneous equation: one valid for x < ξ and one
valid for x > ξ , namely,
u1 (x) = A1 cos kx + B1 sin kx (0 < x < ξ )

u2 (x) = A2 cos kx + B2 sin kx (ξ < x < L)
which become
u1 (x) = B1 sin kx, u2 (x) = B2 ( sin kx − tan kL cos kx ) (5.182)
once we enforce the fixed BCs u1 (0) = 0 and u2 (L) = 0 at the two
end points. In addition to this, two more conditions are required in
order to match the solutions at the point x = ξ . The first, obviously,
is u1 (ξ ) = u2 (ξ ) because the string displacement must be continuous at
x = ξ . For the second condition, we integrate Equation 5.181 across the
point of application of the load to get
ξ +ε ξ +ε ξ +ε

∫
ξ −ε
u ′′ dx + k2
∫
ξ −ε
u dx =
∫ δ (x − ξ) dx = 1
ξ −ε
where, in the limit ε → 0 , the second integral on the l.h.s. vanishes

because of the continuity condition u1 (ξ ) = u2 (ξ ) . So, we are left with
ξ +ε
∫ u′′ dx = u′
ξ +ε
= u2′ (ξ ) − u1′ (ξ ) = 1 (5.183)
ξ −ε
ξ −ε
which establishes the discontinuity jump of the slope of the string at

x = ξ . Using these two last conditions in the solutions 5.182 yields
B1 sin kξ = B2 ( sin kξ − tan kL cos kξ ) ,
kB2 ( cos kξ + tan kL sin kξ ) − kB1 cos kξ = 1
from which we get, after some algebraic manipulations,
sin kξ cos kξ sin kξ

B1 = − , B2 =
k tan kL k k tan kL
At this point, substitution in Equations 5.182 and some standard trigo-
nometric relations lead to the solutions
sin k(L − ξ ) sin kx sin kξ sin k(L − x)

u1 (x) = − , u2 (x) = −
k sin kL k sin kL
in the intervals 0 ≤ x < ξ and ξ < x ≤ L, respectively. Putting the two

solutions together, we get
1  sin k(L − ξ ) sin kx (0 ≤ x < ξ )

Gˆ (x, ξ ) = −  (5.184)
k sin kL  sin kξ sin k(L − x) (ξ < x ≤ L)

which in turn tells us that in order to express the string response at x

due to a distributed load per unit length f (x, t) = F(x) e i ωt as
w(x, t) = e iωt
∫ G(x, ξ)F(ξ) dx (5.185)
0
we must have G(x, ξ ) = − Gˆ (x, ξ ) T0 .
Remark 5.23
i. Example 5.9 is clearly a special case of the SLp 5.178 in which we

have p(x) = T0 , q(x) = − µω 2 in the differential equation, β1 = β2 = 0 in
the BCs and where the two end points are a = 0, b = L. Also, since the
two solutions are u1(x) = sin kx and u2 (x) = sin kx − tan kL cos kx, then
p(x)W (x) = T0 k tan kL, and it is evident that the Green’s function of
Equation 5.185 is the special case of Equation 5.179 for the string of
the example.
ii. Note that both Equations 5.177 and 5.179 show that the Green’s func-
tion is symmetric, that is, G(x, ξ ) = G(ξ , x). This is yet another proof of
reciprocity for linear systems, stating that the deflection at x due to a
unit load at ξ is equal to the deflection at ξ due to a unit load at x.
Chapter 6
Random vibrations
6.1 INTRODUCTION
For some physical phenomena, there seems to be no way to satisfactorily

predict the outcome of a single experiment, observation or measurement,
but it turns out that under repeated experiments, observations and/or mea-
surements they do show long-term patterns and regularities – albeit of a
qualitative different nature with respect to the ones we find in deterministic
phenomena.
As is well known, the key underlying idea here is the concept of prob-
ability and the long-term regularities are generally referred to as ‘laws of
chance’ or, as it is often heard, ‘laws of large numbers’. In this light, we call
these phenomena ‘random’, ‘stochastic’ or ‘non-deterministic’, and the best
we can do – at least at the present state of knowledge – is to adopt a ‘method
of attack’ based on the disciplines of probability theory and statistics. The
question as to whether randomness is ‘built into nature’ (as it seems to be
the case at the quantum level of atomic and subatomic particles – which,
however, are not of interest for us here) or is just a reflection of our igno-
rance remains open.
6.2 THE CONCEPT OF RANDOM PROCESS,

CORRELATION AND COVARIANCE FUNCTIONS
Assuming the reader to have had some previous exposure to the basic defi-
nitions and ideas of probability theory and statistics, we start by observing
that the concepts of ‘event’ and ‘random variable’ (r.v. for short) can be con-
sidered as two levels of a hierarchy. In fact, while a single number P(A) – its
probability – suffices for an event A, the information on a random variable
requires the knowledge of the probability of many, even infinitely many,
events. A step up in this hypothetical hierarchy, we find the concept of ran-
dom (or stochastic) process: a family X(z) of random variables indexed by
a parameter z that varies within an index set Z, where z can be discrete or
continuous.
207
For the most part, in the field of vibrations, the interest is focused on con-
tinuous random processes of the form X(t) with t ∈T , where t is time and
T is some appropriate time interval (the range of t for which X(t) is defined,
observed or measured).
Remark 6.1
i. Although the meaning will always be clear from the context, the time
interval T must not be confused with the period of a periodic function
(also denoted by T in previous chapters);
ii. A process can be random in both time and space. A typical example
can be the vibrations of a tall and slender building during a wind-
storm; here, in fact, the effects of wind and turbulence are random
in time and also with respect to the vertical coordinate y along the
structure.
Since X(t) is a random variable for each value of t , we can use the familiar
definitions of probability distribution function (often abbreviated as PDF)
and probability density function (pdf), and write
FX (x; t) = P (X(t) ≤ x), fX (x; t) dx = P (x < X(t) ≤ x + dx), (6.1a)
where
∂FX (x; t)
fX (x; t) = , (6.1b)
∂x
and where the notations FX (x; t) and fX (x; t) for the PDF and pdf, respec-
tively, show that for a random process, these functions are in general time
dependent. The functions above are said to be of first order because more
detailed information on X(t) can be obtained by considering its behaviour
at two instants of time t1 , t2 – and, in increasing levels of detail, at any finite
number of instants t1 ,..., t n . So, for n = 2, we have the (second-order) joint-
PDF and joint-pdf
 2 
FXX ( x1 , x2 ; t1 , t2 ) = P (
 i =1
)
X ( t i ) ≤ xi 

(6.2a)
 2 
fXX ( x1 , x2 ; t1 , t2 ) dx1dx2 = P  (
 i =1
xi < X ( t i ) ≤ xi + dxi ) 

Random vibrations 209
with
∂2 FX ( x1 , x2 ; t1 , t2 )
fXX ( x1 , x2 ; t1 , t2 ) = , (6.2b)
∂ x1 ∂ x2
and where the extension to n > 2 is straightforward. Note that the second-
order functions contain information on the first-order ones because we have
FX ( x1 ; t1 ) = FXX ( x1 , ∞; t1 , t2 ) , FX ( x2 ; t2 ) = FXX ( ∞, x2 ; t1 , t2 )
∞ ∞
(6.3)
fX ( x1 ; t1 ) =
∫
−∞
fXX ( x1 , x2 ; t1 , t2 ) dx2 , fX ( x2 ; t2 ) =
∫
−∞
fXX ( x1 , x2 ; t1 , t2 ) dx1
(this is a general rule and nth-order functions contain information on all kth-
order for k < n). Clearly, by a similar line of reasoning, we can consider more
than one stochastic process – say, for example, two processes X(t) and Y (t′) –
and introduce their joint-PDFs for various possible sets of instants of time.
As known from basic probability theory (see also the following Remark
6.2(i)), we can use the PDFs or pdfs to calculate expectations or expected val-
ues. So, in particular, we have the m th-order ( m = 1,2,) moment E (X m (t)) –
with the mean µX (t) = E(X(t)) as a special case – and the mth-order cen-
{
tral moment E [ X(t) − µX (t)]
m
}, with the variance σ 2
X (t) as the special case
for m = 2. For any two instants of time t1 , t2 ∈T , we have, respectively, the
autocorrelation and autocovariance functions, defined as
RXX ( t1 , t2 ) ≡ E  X ( t1 ) X ( t2 ) 
(6.4a)
{
KXX ( t1 , t2 ) ≡ E  X ( t1 ) − µX ( t1 )   X ( t2 ) − µX ( t2 )  }
and related by the equation
KXX ( t1 , t2 ) = RXX ( t1 , t2 ) − µX ( t1 ) µX ( t2 ). (6.4b)
In particular, for t1 = t2 ≡ t , Equations 6.4 give
RXX (t , t) = E  X 2 (t)  , KXX (t , t) = σ X2 (t), σ X2 (t) = E  X 2 (t)  − µX2 (t).(6.5)
Also, to every process X(t) with finite mean µX (t), it is often convenient to
associate the centred process Xˆ (t) = X(t) − µX (t), which is a process with
zero mean whose moments are the central moments of X(t). In particular,
note that the autocorrelation and autocovariance functions of Xˆ (t) coincide.
When two processes X , Y are considered simultaneously, the counter-

parts of Equation 6.4a are called cross-correlation and cross-covariance
RXY ( t1 , t2 ) = E  X ( t1 )Y ( t2 ) 
(6.6a)
{
KXY (t1 , t2 ) ≡ E  X ( t1 ) − µX ( t1 )  Y ( t2 ) − µY ( t2 )  }
while the ‘two-processes counterpart’ of Equation 6.4b is
KXY ( t1 , t2 ) = RXY ( t1 , t2 ) − µX ( t1 ) µY ( t2 ). (6.6b)
Remark 6.2 (A refresher of probability theory)
i. If X is a continuous random variable with pdf fX (x), it is well known

from basic probability theory that the expectation E(X) of X is given
∞
by E(X) =
∫ −∞
xfX (x) dx . This is the familiar mean, often also denoted
by the symbol µX. Other familiar quantities are the variance σ X2 of X,
∞
defined as σ X2 ≡ E ( X − µX )  =
∫
( x − µX ) fX (x)dx (and often also
2 2
  −∞
denoted by Var(X)) and the mean square value E X 2 . In this respect, ( )
we recall the relation σ = E X 2
X ( )− µ
2 2
X , which is a special case of the
( )
m
general formula E ( X − µX )  = ∑ (−1)k m !
m
µXk E X m−k (with the con-
  k= 0 k !(m− k)!
( )
vention E X 0 = 1) that gives the central moments of X in terms of
its ordinary (i.e. non-central) moments. Also, the square root of the
variance, denoted by σ X , is called standard deviation, while E X 2 ( )
is called root mean square (rms) value;
ii.
In the light of point (i) of the Remark, it is understood
that the autocorrelation of Equation 6.4a1 is obtained as
∞ ∞
RXX ( t1 , t2 ) =
∫ ∫ −∞ −∞
x1x2 fXX ( x1 , x2 ; t1 , t2 ) dx1 dx2 and that the auto-
covariance, cross-correlation and cross-covariance are determined
accordingly;
iii. A key concept in probability theory is the notion of independent
random variables, where in general terms a r.v. Y is independent of
another r.v. X if knowledge of the value of X gives no information
at all on the possible values of Y , or on the probability that Y will
take on any of its possible values. In this respect, important conse-
quences of independence are the relations FXY (x, y) = FX (x)FY (y) and
fXY (x, y) = fX (x)fY (y), meaning that the joint-PDF and joint-pdf fac-
torise into the product of the individual PDFs and pdfs. Also, inde-
pendence implies E(XY ) = E(X)E(Y ) and Cov(X , Y ) = 0, where the
covariance is defined as Cov(X , Y ) = E ( X − µX )(Y − µY ) . However,
note that the first two relations are ‘if and only if’ statements, whereas
the last two are not; they are implied by independence but, when they
hold, in general do not imply independence;
iv. We recall that two random variables satisfying Cov(X , Y ) = 0 are
called uncorrelated. So, the final part of point (iii) of the remark tells
us that independence implies uncorrelation, but uncorrelation does
not, in general, implies independence.
Turning now to more practical aspects, we observe that what we do in

order to describe and/or obtain information on a random variable X is by
means of statistical methods. And since the first step consists in collecting
a number n (where n is possibly large) of sample data x1 ,, xn, this is what
we do with a random process X(t), with the difference that now our data
will be in the form of a number of sample functions x1(t),, xn (t), where
each sample function xi (t) is a time record (or time history) that extends for
a certain time interval. In principle, we should collect an infinite number of
such time records, but since this is an obvious impossibility, the engineer’s
approximate representation of the process is a finite set of time records
x1(t),, xn (t) called an ensemble (of size n).
Example 6.1
As a simple example, consider the vibrations of a car that travels every
day over a certain rough road at a given speed and takes approxi-
mately 10 minutes from beginning to end. So, with t varying in the
time interval T = [0,600] seconds, we will measure a vibration time
history x1(t) on the first day, x2 (t) on the second day, etc. But since
the (hypothetical) ‘population’ associated with this ensemble is the set
of sample functions that, in principle, could be recorded by repeating
the ‘experiment’ an infinite number of times, our representation of the
process will be a finite set of time records. For purpose of illustration,
Figure 6.1 shows an ensemble of size n = 4.
In the light of the considerations above, it is now important to point out

that the expectations µX (t), RXX ( t1 , t2 ) are the (non-random) theoretical
quantities of the ‘population’, while the averages that we determine from
our finite ensemble are their statistical counterparts (the so-called estimates
of µX (t), RXX ( t1 , t2 ), etc.) and are random quantities themselves whose reli-
ability depends in general on a number of factors, among which – not
surprisingly – the size of the ensemble.
Figure 6.1 Ensemble of four time histories.
To all this, however, it should be added that there is an important differ-

ence between a r.v. X and a random process X(t); while in the former case,
a single observation x1 (i.e. a set of data of size one) provides practically no
useful information on X, it may not be so in the latter case. Under appro-
priate conditions, in fact, we will see that a single (sufficiently long) time
record x1(t) can provide meaningful information on the underlying process
X(t). This fact, in turn, is related to what we can call a two-dimensional
interpretation of the random process if now, for present convenience, we
denote it by the symbol X(i, t). Then, for a fixed value of t , say t = t0, X ( i, t0 )
is a one-dimensional random variable and x1 ( t0 ) ,, xn ( t0 ) are realisations/
observations of the r.v. X ( i, t0 ), while for a fixed value of i, say i = 1, X(1, t)
is a function of time – that is the time record x1(t). But then, in the light of
the fact that the expectations µX (t), RXX ( t1 , t2 ) are understood as ensemble
expectation values – that is expectations calculated across the ensemble –
the question arises if these expectations are somehow related, or even equal,
to the time averages calculated along a single (possibly long) realisation of
the process. We will consider this aspect in a future section where we will
see that such relations do exist in some special but important cases.
6.2.1 Stationary processes
The term ‘stationary’ refers to the fact that some characteristics of a ran-
dom process – moments and/or probability laws – remain unchanged under
an arbitrary shift of the time axis, meaning that the process is, broadly
speaking, in some kind of ‘statistical steady state’.
In general, moments’ stationarity is more frequently used in applica-
tions and we call a process mean-value (or first-moment) stationary if
µX (t + r) = µX (t) for all values of t and r, a condition which clearly implies
that the mean µX cannot depend on t . Similarly, a process is second-moment

stationary and two processes X , Y are called jointly second-moment
stationary if, respectively, the equalities
RXX ( t1 + r , t2 + r ) = RXX ( t1 , t2 ) , RXY ( t1 + r , t2 + r ) = RXY ( t1 , t2 ) (6.7)
hold for any value of r and any two times t1 , t2. When this is the case, it
follows that the autocorrelation (or cross-correlation) cannot depend on
the specific values of t1 , t2 but only on their difference τ = t2 − t1 and we can
write RXX (τ ) or RXY (τ ). A slight variation of second-moment stationarity is
obtained if the equalities 6.7 hold for the covariance or cross-covariance
function. In the two cases, respectively, we will then speak of covariant sta-
tionary process or jointly covariant stationary processes and, as above, we
will have the simpler functional dependence KXX (τ ) or KXY (τ ). Note that if
a process X(t) is both mean-value and second-moment stationary, then (by
virtue of Equation (6.4b)) it is also covariant stationary. In particular, the
two notions coincide for the centred process Xˆ (t) – which, being defined as
Xˆ (t) = X(t) − µX (t), is always mean-value stationary.
By extension, a process is called mth moment stationary if
E  X ( t1 + r ) X ( t m + r )  = E  X ( t1 ) X ( t m )  (6.8)
for all values of the shift r and times t1 ,, t m. So, in particular, if
t1 = t2 =  = t m ≡ t , then E  X m (t + r)  = E  X m (t)  and the mth moment
E(X m ) does not depend on t . At the other extreme, if the times t1 ,, t m are
all different, then the mth moment function will not depend on their specific
values but only on the m − 1 time increments τ 1 = t2 − t1 ,, τ m−1 = t m − t m−1.
Remark 6.3
The developments above show that stationarity reduces the number of neces-
sary time arguments by one. This is a general rule that applies also to the other
types of stationarity – that is first order, second order, etc. – introduced below.
As mentioned at the beginning of this section, the other type of station-

arity refers to probability laws (PDFs or, when they exist, pdfs) instead of
moments. Then, a process is called first-order stationary if
FX (x, t + r) = FX (x, t) (6.9)
for all values of x, t and r. This implies that the processes PDF and pdf do
not change with time and can be written as FX (x) and fX (x). The processes
are second-order stationary if
FXX ( x1 , x2 , t1 + r , t2 + r ) = FXX ( x1 , x2 , t1 , t2 ) (6.10)

for all values of x1 , x2 , t1 , t2 and r, and so on, up to the most restrictive type
of stationarity – called strict – which occurs when X(t) is mth-order station-
ary for all m = 1,2.
Given these definitions, the question arises if the various types of sta-
tionarity are somehow related. The answer is ‘partly so’, and here, we only
limit ourselves to two results: (a) mth-order stationarity implies all sta-
tionarities of lower order, while the same does not apply to mth moment
stationarity (so, for example, a second-moment stationary process may
not be mean-value stationary) and (b) mth-order stationarity implies mth
moment stationarity. In this respect, note that from points (a) and (b), it
follows that a mth-order stationary process is also stationary up to the
mth moment.
Remark 6.4
In general, it is not possible to establish a complete hierarchy among order

and moment stationarities because they refer to different characteristics of
the process. Consequently, some types are simply non-comparable, and an
example in this sense is given by second-moment and first-order stationar-
ity. In fact, first-order stationarity means that FX , fX are invariant under
( )
time shifts, thus implying that the moments E X m (t) – which are calculated
using FX or fX – are also time invariant for all m = 1,2,. This fact, however,
gives us no information on the relation between X ( t1 ) and X ( t2 ) for t1 ≠ t2 –
an information which, on the contrary, is provided by second-moment
stationarity.
Turning to a more practical aspect, it turns out that in general, it is

seldom possible to test for more than second-moment or second-order sta-
tionarity, so that the term ‘weakly stationary’ (WS) or ‘wide-sense sta-
tionary’ (WSS) is frequently found in the literature. The term, however,
is slightly ambiguous; for some authors, in fact, a WS process is a process
that is both first- and second-moment stationary, while for some other
authors, it means a second-order stationary process. Here, we adhere to
the first definition.
6.2.2 Main properties of correlation

and covariance functions
Starting with symmetry under the exchange of t1 and t2 , it is evident that the
relations RXX ( t1 , t2 ) = RXX ( t2 , t1 ) and KXX ( t1 , t2 ) = KXX ( t2 , t1 ) follow directly
from the definitions themselves. For WS processes, this means that both
functions are even and we have
RXX (−τ ) = RXX (τ ), KXX (−τ ) = KXX (τ ), (6.11)

thus implying that in practice one can consider only the positive values of τ .
Under the assumption of weak stationarity, a second property is that the
two functions are bounded by their values at τ = 0 because we have
( )
RXX (τ ) ≤ RXX (0) = E X 2 , KXX (τ ) ≤ KXX (0) = σ X2 , (6.12)
which hold for all τ and follow from using the r.v.s X(t) and X(t + τ ) in the
( ) ( )
well-known result E(XY ) ≤ E X 2 E Y 2 called Schwarz’s inequality.
Remark 6.5
If stationarity is not assumed, Schwarz’ inequality gives RXX ( t1 , t2 )

≤ E  X 2 ( t1 )  E  X 2 ( t2 )  and KXX (t1 , t2 ) ≤ σ X (t1)σ X (t2 ). Equation 6.12 are
the stationary versions of these two more general inequalities.
For WS processes, moreover, we have
KXX (τ ) = RXX (τ ) − µX2 , RXX (0) = σ X2 + µX2 , (6.13)
where the second equation is simply the first for τ = 0 and the first, in turn,
is the stationary counterpart of Equation 6.4b. Also, since this is a fre-
quently encountered case in practice, it is worth noting that for µX = 0,
Equation 6.13 become KXX (τ ) = RXX (τ ) and RXX (0) = KXX (0) = σ X2 .
Another property of autocovariance functions concerns their behaviour
for large values of τ because in most cases of practical interest (if the process
does not contain any periodic component), it is found that KXX (τ ) → 0 as
τ → ∞. Rather than a strict mathematical property, however, this is some-
how a consequence of the randomness of the process and indicates that,
in general, there is an increasing loss of correlation as X(t) and X(t + τ ) get
further and further apart. In other words, this means that the process pro-
gressively loses memory of its past, the loss of memory being quick when
KXX (τ ) drops rapidly to zero (as is the case for extremely irregular time
records) or slower when the time records are relatively smooth.
If two WS processes are cross-covariant stationary, the cross-correlation
functions RXY (τ ), RYX (τ ) are neither odd nor even, and in general, we have
RXY (τ ) ≠ RYX (τ ). However, the property of invariance under a time shift
leads to
RXY (τ ) = RYX (−τ ), RYX (τ ) = RXY (−τ ). (6.14)
Also, since in this case Equation 6.6b becomes
KXY (τ ) = RXY (τ ) − µX µY , (6.15)

it follows that two stationary processes are uncorrelated – meaning that

KXY (τ ) = 0 – whenever RXY (τ ) = µX µY . Then, Equation 6.15 shows that for
two uncorrelated processes, we can have RXY (τ ) = 0 only if at least either of
the two means is zero. Finally, worthy of mention are other two properties
called cross-correlation inequalities
2 2
RXY (τ ) ≤ RXX (0) RYY (0), KXY (τ ) ≤ KXX (0) KYY (0) = σ X2 σ Y2 , (6.16)
which, as Equation 6.12, follow from Schwarz’s inequality.
6.2.3 Ergodic processes
By definition, a process is strictly ergodic if one sufficiently long sample
function x(t) is representative of the whole process. In other words, if the
length T of the time record is large and it can reasonably be assumed that
x(t) passes through all the values accessible to it, then we have good rea-
sons to believe that the process is ergodic. The rationale behind this lies
essentially in two considerations of statistical nature. The first is that in
any sample other than x(t), we can expect to find not only the same values
(although, clearly, in a different time order) taken on by the process in the
sample x(t), but also the same frequency of appearance of these values. The
second is that if we imagine to divide our sample function x(t) in a number,
say p, of sections and we can assume the behaviour in each section to be
independent on the behaviour of the other sections, then, for all practical
purposes, we can consider the p sections as a satisfactory and representative
ensemble of the process.
The consequence is that we can replace ensemble averages by time aver-
ages, that is averages calculated along the sample x(t). So, for example, we
say that a WS process is weakly ergodic if it is both mean-value ergodic and
second-moment ergodic, that is if the two equalities
E(X(t)) = x , RXX (τ ) = CXX (τ ) (6.17)
hold, where the temporal mean value x (but we will also denote it by x)
and the temporal correlation CXX (τ ) are defined as
T
1
x = lim
T →∞ T ∫ x(t) dt
0
(6.18)
T
1
CXX (τ ) = x(t) x(t + τ ) = lim
T →∞ T ∫ x(t) x(t + τ ) dt.
0
The reason why the process must be WS in order to be weakly ergodic is

that, as a general rule, ergodicity implies stationarity (but not conversely).
This is particularly evident if we consider mean-value ergodicity; since x ,
by definition, cannot depend on time, Equation 6.171 can hold only
if E(X(t)) does not depend on time, that is if the process is mean-
value stationary. So, if the process is non-stationary, then it is surely
non-ergodic.
Remark 6.6
For an arbitrary process of which we have an ensemble of size n, the time

averages defined by Equation 6.18 will depend on the particular sample
function used to calculate them and we should write x(k) (or x (k)) and
CXX (τ ; k), where k = 1,, n is the sample index. For a weakly ergodic pro-
cess, however, we can omit this index because Equation 6.17 imply that (as
far as the first two moments are concerned) each sample is representative of
all the others.
For the most part, we will be interested in mean-value ergodic pro-

cesses and weakly ergodic processes but it is not difficult to see that
the ideas above can be extended to define as many types of ergodicity
as there are stationarities, each type of ergodicity implying the corre-
sponding type of stationarity (order ergodicity, however, is slightly more
involved than moment ergodicity, and the interested reader can refer, for
example, to Lutes and Sarkani (1997) for this aspect). The reverse impli-
cation, however, is not true in general, and for example, a mean-value
stationary process may not be mean-value ergodic, a second-moment sta-
tionary process may not be second-moment ergodic, etc. Also, note that
second-moment ergodicity does not imply, and is not implied by, mean-
value ergodicity.
Although there exist theorems giving sufficient conditions for ergodicity,
we refer the interested reader to more specialised literature (e.g. Elishakoff
(1999), Lutes and Sarkani (1997) or Sólnes (1997)) for these aspects, and
here we only limit ourselves to some practical considerations. The first is
that one generally invokes ergodicity because it significantly simplifies both
the data acquisition stage and the subsequent data analysis. Clearly, if a
large ensemble of time records is available, then the issue of ergodicity can
be avoided altogether, but the fact remains that obtaining a large ensemble
is often impractical and costly, if not impossible in some cases. So, a gener-
ally adopted pragmatic approach is as follows: Whenever a process can be
considered stationary, one (more or less implicitly) assumes that it is also
ergodic unless there are obvious reasons not to do so. Admittedly, this is
more an educated guess rather than a solid argument but it often turns out
to be satisfactory if our judgement is based on good engineering common
sense and some insight on the physical mechanism generating the process
under study.
The second consideration refers to the generic term ‘sufficiently long
sample’, used more than once in the preceding discussion. A practical
answer in this respect is that the duration of the time record must be
at least longer than the period of its lowest spectral components (the
meaning of ‘spectral components’ for a random process is considered is
Section 6.4).
Remark 6.7
Strictly speaking, no real-world process is truly stationary or truly ergodic

because all processes must begin and end at some time. Since, however, we
noticed that it is often adequate to assume that a process is stationary and
ergodic for the majority of its lifetime, the question arises of how we can
use the available data to check the goodness of the assumption. In order
to answer this question, we must turn to statistics. We can, for example,
divide our sample(s) in shorter sections, calculate (time and ensemble) aver-
ages for each section and groups of sections, and then use the techniques
of hypothesis testing to examine how these section averages are compared
with each other and with the corresponding average(s) for the original
sample(s). Then, based on these tests and the amount of variation that we
are willing to accept, we will be in a position to decide whether to accept or
reject our initial assumption.
Example 6.2
Consider the random process X(t) = U sin(ω t + V ), where ω is a constant
and the two r.v.s U ,V are such that (a) they are independent, (b) the
amplitude U has mean and variance µ U , σ U2 , respectively, and (c) the
phase angle V is a r.v. uniformly distributed in the interval [0,2π ] (this
meaning that its pdf is fV (v) = 1 2π for v ∈[0,2π ] and zero otherwise).
In order to see if the process is WS and/or ergodic, we must calculate
and compare the appropriate averages (the reader is invited to fill in the
details of the calculations). For ensemble averages, on the one hand, we
take independence into account to obtain the mean as

2π 2π
µU
E(X(t)) = E(U)E(V ) = µU
∫
0
fV (v) sin(ω t + v)dv =
2π ∫ sin(ω t + v) dv = 0
0

(6.19)
and then the autocorrelation at times r , s as

RXX (r , s) = E [ X(r)X(s)] = E U 2 sin(ω r + v) sin(ω s + v) 
( )
E U2
2π
=
2π ∫ sin(ω r + v) sin(ω s + v) dv
0
( )
E U2
2π
=
4π ∫ cos ω (s − r) − cos [ω (r + s) + 2v] dv
0
σ +µ
2
U
2
U σ 2 + µU2 (6.20)
= cos ω (s − r) = U cos ω τ ,
2 2
( )
where τ = s − r and we took the relation E U 2 = σ U2 + µ 2U into account.
Since the mean does not depend on time and the correlation depends
only on τ , the process is WS (also note that RXX (τ ) = KXX (τ ) because
E(X) = 0). On the other hand, the time averages calculated (between
zero and T = 2π ω ) for a specific sample in which the two r.v.s have the
values u, v give
T
1

x T
=
T ∫ u sin(ω t + v)dt = 0
0 (6.21)
u2
(T )
CXX (τ ) = u2 sin(ω t + v)sin [ω (t + τ ) + v ] = cos ωτ
T 2
so that in the limit as T → ∞, we obtain x = 0 and CXX (τ ) = u2 2 cos ωτ . ( )

Since this latter quantity depends (through the term u2 2) on the spe-
cific sample used in the calculations, the process is not second-moment
ergodic and therefore is not weakly ergodic. The process, however, is
mean-value ergodic because E(X(t)) = x .
Consider now the case in which V is as above but the ampli-
tude U is a constant (hence non-random) of value u. All the results
of Equation 6.19–6.21 remain formally unaltered but now we have
σ U2 = 0 and µU2 = u2 , which, when substituted in Equation 6.20, lead
to CXX (τ ) = RXX (τ ), thus showing that the process X(t) = u sin(ω t + V )
is both mean-value and second-moment ergodic (i.e. weakly ergodic).
6.3 SOME CALCULUS FOR RANDOM PROCESSES
Since the main physical quantities involved in the study and analysis of vibra-
tions are displacement, velocity and acceleration, it is important to consider
how these quantities are related in the case of random processes. Just like
ordinary calculus, the calculus of random processes revolves around the
notion of limit and hence convergence. We call stochastic derivative of X(t)
the process X (t) defined as
dX(t) X(t + h) − X(t)

X (t) = ≡ lim (6.22)
dt h→0 h
if the limit exists for each t (see the following Remark 6.8). Similarly, when
it exists, we define the stochastic integral of X(t) on [ a, b] as the limit of a
sequence of Riemann sums, that is
n
∑ X (t ) ∆t
b
I=
∫ a
X(t) dt ≡ lim In = lim
n→∞ n→∞
k=1
( n)
k
( n)
k , (6.23)
{ }
where Pn = a = t0(n) ≤ t1(n) ≤, ≤ t n(n) = b is a partition of the interval [ a, b],
( n) ( n) ( n)
∆t = t − t
k k and the sequence of partitions P1 , P2 , is such that
k−1
( n) ( n)
∆ t = max ∆ t → 0 as n → ∞. Note that for fixed a, b, the quantity I is a
k
random variable.
Remark 6.8
i. Since in probability theory we have different types of convergence for

random variables (typically: convergence ‘in probability’, ‘in distribu-
tion’ and ‘almost-sure’ convergence), the question arises about what
type of limit is involved in the definitions of derivative, integral and,
clearly, continuity of a random process. As it turns out, none of the
ones mentioned above. In this context, in fact, the most convenient
type is the so-called mean-square limit. We say that a sequence Xn of
r.v.s converges in the mean square to X – and we write Xn → X [ ms] –
( )
if E Xn2 < ∞ for all n = 1,2, and E  Xn − X  → 0 as n → ∞.

2

When compared to the other types of convergence, it can be shown
that Xn → X [ ms] implies convergence in probability and in distribu-
tion, while, without additional assumptions, there is in general no
relation between ms convergence and almost sure convergence;
ii. For our present purposes, the important fact is that ms limits and
expectations can be interchanged; more specifically, we have (Sólnes
1997): If Xn → X [ ms] and Yn → Y [ ms], then (a) lim n→∞ E ( Xn ) = E(X),
( )
(b) lim n,m→∞ E ( Xn Xm ) = E X 2 and (c) lim n→∞ E ( XnYn ) = E(XY ). Using
these properties, in fact, the developments that follow will show that
the stochastic derivative and integral defined above exist whenever
some appropriate ordinary derivatives and Riemann integrals of the
functions µX and RXX exist.
With Remark 6.8(ii) in mind, it is easy to determine the mean of X (t) as
dµ (t)

d
( )
µX (t) ≡ E X (t) = E[ X(t)] = X . (6.24)
dt dt
Things are a bit more involved for the cross-correlations between X (t) and
X(t), but if we let r , s be two instants of time with r ≤ s , it is not difficult to
obtain the relations
∂RXX (r , s) ∂RXX (r , s)
 (r , s) =
RXX , RXX (r , s) = ,
∂r ∂s
(6.25)
∂2 RXX (r , s)
  (r , s) =
RXX .
∂r ∂s
If, in particular, X(t) is a WS process, then its mean value is time inde-
( )
pendent and Equation 6.24 implies µX ≡ E X (t) = 0. Moreover, since RXX
depends only on the difference τ = s − r (so that dτ dr = −1 and dτ ds = 1),
the same functional dependence applies to the correlation functions above
and we have
dRXX (τ ) dRXX (τ ) d 2RXX (τ )

 (τ ) = −
RXX , RXX (τ ) = ,   (τ ) = −
RXX , (6.26)
dτ dτ dτ 2
which, in turn, show that RXX  (τ ) = − RXX (τ ).

(In the following, the first and second τ -derivatives of RXX (τ ) appear-
ing on the right-hand sides of Equation 6.26 will often also be denoted by
′ (τ ) and RXX
RXX ′′ (τ ), respectively).
Remark 6.9
Some consequences of the results above are worthy of notice:
i. ( )
E X (t) = 0 implies, owing to Equation 6.15, RXX (τ ) = KXX (τ ),
RXX (τ ) = KXX  (τ ) and RXX  (τ ) = KXX
  (τ ). From this last relation, it follows
that the variance of X (t) is given by σ X2 = KXX   (0) = RXX

  (0) = − RXX
′′ (0);
ii. The equality RXX  (τ ) = − RXX (τ ) right after Equation 6.26 together with
Equation 6.14 shows that the two cross-correlations RXX , RXX  (and there-
fore, by Equation 6.26, RXX ′ (τ )) are all odd functions of τ . This implies
that they are all zero at τ = 0; that is RXX (0) = − RXX  (0) = RXX ′ (0) = 0,
which in turn means that a stationary process X(t) is such that X(t)
and X ( t ′ ) are uncorrelated for t = t ′ (i.e. for τ = 0);
iii. Equations 6.11 together with 6.263 imply that RXX ′′ (τ ) is an even
function of τ .
By a straightforward extension of the ideas above, we can consider the

 (t). The reader is invited to do so and to show
second-derivative process X
that, among other properties, in the WS case, we have
 (τ ) = RXX
 (τ ) = RXX
′′ (τ ),   (τ ) = − RXX
  (τ ) = RXX (τ )
(3)
RXX RXX
(6.27)
  (τ ) = RXX (τ ), σ X2 = RXX
(4) (4)
RXX (0),
where by RXX (3)

(τ ), RXX
(4)
(τ ) we mean the third- and fourth-order derivative
of RXX (τ ) with respect to τ . So, a point to be noticed about the deriva-
tive processes X (t), X (t),… (when they exist in the sense of definition 6.22
and Remark 6.8(i)) is that they are all weakly stationary whenever X(t) is
weakly stationary.
If now we turn our attention to the integral I of Equation 6.23, the prop-
erties mentioned in Remark 6.8(ii) lead to
b b b
E(I) =
∫ µ (t)dt,
a
X
2
E(I ) =
∫ ∫R
a a
XX (r , s) drds, (6.28)
thus implying that Var(I) is given by the double integral of the covariance
function KXX (r , s). A slight generalisation of Equation 6.23 is given by the
integral
Q(z) =
∫ X(t)k(t, z) dt , (6.29)
a
where k(t , z) – a so-called kernel function – is any (possibly complex) deter-

ministic smooth function of its arguments. Then, Q(z) is a new random
process indexed by the parameter z and we obtain the relations
µQ (z) =
∫µ a
X (t) k(t , z) dt
(6.30)
b b
RQQ ( z1 , z2 ) =
∫∫R
a a
XX (r , s) k ( r , z1 ) k ( s, z2 ) drds,
which follow from the fact that the kernel function k is non-random.
A different process – we call it J(t) – is obtained if X(t) is integrable on
t
[ a, b], t ∈[ a, b], and we consider the integral J(t) =
∫ a
X(r) dr . Then, we have
the relations
t r s
µ J (t) = E( J(t)) =
∫µ
a
X (r)dr , RJJ (r , s) =
∫∫R
a a
XX (u, v) dudv
(6.31)
s
RXJ (r , s) =
∫R
a
XX (r , v) dv,
which lend themselves to two main considerations. The first is that, when
compared with Equations 6.24 and 6.25, they show that the process J(t)
can be considered as the ‘stochastic antiderivative’ of X(t). This is a satis-
factory parallel with ordinary calculus and conforms with our idea that
even with random vibrations we can think of velocity as the derivative of
displacement and of displacement as the integral of velocity. The second
consideration is less satisfactory: When X(t) is stationary, it turns out that,
in general, J(t) is not. Just by looking at the first of Equation 6.31 in fact,
we see that the integral process J(t) of a mean-value stationary process X(t)
is not mean-value stationary unless µX = 0. Similarly, it is not difficult to
find examples of second-moment stationary processes whose integral J(t)
is neither second-moment stationary nor jointly second-moment stationary
with X(t).
6.4 SPECTRAL REPRESENTATION OF
STATIONARY RANDOM PROCESSES
∞
A stationary random process X(t) is such that the integral
∫−∞
X(t) dt does
not converge, and therefore, it does not have a Fourier transform in the
classical sense. Since, however – we recall from Section 6.2.2 – randomness
results in a progressive loss of correlation as τ increases, the covariance
function of the most real-world processes is, as a matter of fact, integra-
ble on the real line R. Then, its Fourier transform exists, and, by defini-
tion, we call it power spectral density (PSD) and denote it by the symbol
SXX (ω ). Moreover, if SXX (ω ) is itself integrable on R, then the two functions
KXX (τ ), SXX (ω ) form a Fourier transform pair, and we have
∞
1
SXX (ω ) = F {KXX (τ )} =
2π ∫K
−∞
XX (τ ) e − iωτ dτ
(6.32)
∞
KXX (τ ) = F−1 {SXX (ω )} =

∫S
−∞
XX (ω ) e iωτ dω ,
where, as in preceding chapters and in Appendix B, the symbol F{•} indi-

cates the Fourier transform of the function within braces and we write F-1 {•}
for the inverse transform. Together, Equation 6.32 are known as Wiener–
Khintchine relations.
Remark 6.10
i. Other common names for SXX (ω ) are autospectral density or, simply,
spectral density;
ii. Note that some authors define SXX (ω ) as F {RXX (τ )}. They generally
assume, however, that either X(t) is a process with zero-mean or its
nonzero-mean value has been removed (otherwise, the PSD has a
Dirac-δ ‘spike’ at ω = 0);
iii. The name ‘PSD’ comes from an analogy with electrical systems. If, in
fact, we think of X(t) as a voltage signal across a unit resistor, then
2
X(t) is the instantaneous rate of energy dissipation.
Example 6.3
−c τ
Given the correlation function RXX (τ ) = R0 e (where c is a positive
constant), the reader is invited to show that the corresponding spectral
density is
cR0
SXX (ω ) =
(
π c2 + ω 2 )
and to draw a graph for at least two different values of c, showing that
increasing values of c imply a faster decrease of RXX to zero (meaning
more irregular time histories of X(t)) and a broader spectrum of fre-
quencies in SXX . Conversely, given the PSD, SXX (ω ) = S0 e − c ω , the reader
is invited to show that the corresponding correlation function is
2cS0
RXX (τ ) = .
c2 + τ 2
Example 6.4
Leaving the details of the calculations to the reader (see the
hint below), the PSD corresponding to the correlation function
RXX (τ ) = e − c τ cos bτ is
1  c c 
SXX (ω ) =  + .
2π  c2 + (ω + b)2 c2 + (ω − b)2 
In this case, the shape of the curve depends on the ratio between the
two parameters c and b . If c < b, the oscillatory part prevails and the
spectrum shows two peaks at the frequencies ω = ±b. On the other
hand, if c > b, the decreasing exponential prevails and the spectrum is
‘quite flat’ over a range of frequencies.
Hint for the calculation: Use Euler’s formula 2 cos bτ = e i bτ + e − i bτ
and calculate the Fourier transform of RXX (τ ) as
1  c τ i bτ 
0 ∞
∫ ( ) ( ) 

4π 
 e e +e
 −∞
− i bτ
e − iωτ dτ +
∫
0
e − c τ e i bτ + e − i bτ e − iωτ dτ .

The cross-spectral density SXY (ω ) between two stationary processes

X(t), Y (t) is defined similarly, and if SXY (ω ) is integrable, we have the Fourier
transform pair
SXY (ω ) = F {KXY (τ )} , KXY (τ ) = F−1 {SXY (ω )} (6.33)
and, clearly, similar relations for SYX (ω ).

Given the above definitions, the first question that comes to mind is
whether, by Fourier transforming the covariance function, we are really
considering the frequency content of the process. The answer is affirma-
tive, and the following argument will provide some insight. Let X(t) be a
zero-mean stationary process, which, for present convenience, we write
as an ensemble xk (t) of theoretically infinitely long time records indexed
by k = 1,2,. By so doing, taking expectations means averaging on k, and
we have, for example, E ( xk (t)) = µX = 0. More important, however, is the
relation
 1 T

E [CXX (τ ; k, T )] ≡ E 
 2T ∫
−T
xk (t) xk (t + τ ) dt 

(6.34)
T T
1 1
=
2T
−T
∫ E[ x (t)x (t + τ )] dt = R
k k XX (τ )
2T ∫ dt = R
−T
XX (τ ),
which, in statistical terms, shows that the time correlation CXX (τ ; k, T )

is an unbiased estimator (see Remark 6.11 below) of RXX (τ ). If now
we define the ‘truncated’ time records xk(T ) (t) as xk(T ) (t) = xk (t) for
−T ≤ t ≤ T and zero otherwise, the time correlation can be rewritten as
∞
CXX (τ ; k, T ) = (2T )−1
∫ −∞
xk(T ) (t)xk(T ) (t + τ ) dt, that is as 1 2T times the convo-
lution of xk(T ) (t) with itself. Then, by the convolution theorem (Appendix
B, Section B.2.1), it follows that the Fourier transform SXX (ω ; k, T ) of
{ }
CXX (τ ; k, T ) is related to the modulus squared of X k(T ) (ω ) ≡ F xk(T ) (t) by the
equation
π  (T ) 2
SXX (ω ; k, T ) ≡ F {CXX (τ ; k, T )} = Xk (ω ) . (6.35)
T
Taking expectations on both sides and calling SXX (ω ; T ) the expectation of

SXX (ω ; k, T ), we get
π 2
SXX (ω ; T ) = F {RXX (τ )} = E  X k(T ) (ω )  , (6.36)
T 
where the first equality is due to Equation 6.34 when one takes into account
that the expectation and Fourier operators commute, that is E ( F{•} ) = F {E(•)}.
Finally, since in the limit as T → ∞ the function SXX (ω ; T ) tends to the PSD
SXX (ω ), Equation 6.36 gives
 π  (T ) 2
SXX (ω ) = lim E  Xk (ω )  , (6.37)
T →∞  T 
which, by showing that the PSD SXX (ω ) does indeed ‘capture’ the frequency
information of the process X(t), provides the ‘yes’ answer to our question.
Remark 6.11
Some clarifying points on the argument leading to Equation 6.37 are as

follows:
i. A r.v. Q is an unbiased estimator of an unknown quantity q if

E(Q) = q. If E(Q) − q = b ≠ 0, then Q is biased and b is called the bias.
Unbiasedness is a desirable property of estimators, but it is not the
only one; among others, we mention efficiency, consistency and suf-
ficiency (see, for example, Gatti (2005) or Ivchenko and Medvedev
(1990) for more details);
ii. The fact that CXX (τ ; k, T ), by Equation 6.34, is an unbiased estimator
of RXX (τ ) does not imply that X(t) is ergodic; it might be or it might
not be, but others are the conditions required to establish the ergodic-
ity of X(t);
iii. Obviously, the functions xk(T ) (t) tend to xk (t) as T → ∞. However, the
‘truncated’ transforms X k(T ) (ω ) exist for any finite value of T even if the
process X(t), and therefore the individual records xk (t), are not Fourier
transformable.
Remark 6.12
This additional remark is made because Equation 6.35 may suggest the
following (incorrect) line of thought: Since the index k is superfluous
if X(t) is ergodic, we can in this case skip the passage of taking expec-
tations (of Equation 6.35) and obtain Equation 6.37 in the simpler form
2
SXX (ω ) = lim ( π T ) X (T ) (ω ) . This, in practice, means that the (unknown)
T →∞
PSD of an ergodic process can be estimated by simply squaring the modu-
{ }
lus of F x(T ) (t) and then multiplying it by π T , where x(T ) (t) is a single and
sufficiently long time record. However tempting, this argument is not cor-
2
rect because it turns out that ( π T ) X (T ) (ω ) , as an estimator of SXX (ω ), is
not consistent, meaning that the variance of this estimator is not small and
does not go to zero as T increases. Consequently, we can indeed write the
2
approximation SXX (ω ) ≅( π T ) X (T ) (ω ) , but we cannot expect it to be a reli-
able and accurate approximation of SXX (ω ), no matter how large is T. The
conclusion is that we need to take the expectation of Equation 6.35 even if
the process is ergodic. In practice, this means that more reliable approxima-
tions are obtained only at the price of a somewhat lower-frequency resolu-
tion (this is because taking expectations provides a ‘smoothed’ estimate
of the PSD. More on this aspect can be found, for example, in Lutes and
Sarkani (1997) or Papoulis (1981).
6.4.1 Main properties of spectral densities

Owing to the symmetry properties (Equations 6.11 and 6.14) of the (real)
functions that are Fourier-transformed, the definitions of PSD and cross-
PSD lead easily to
*
SXX (ω ) = SXX (ω ) = SXX (−ω ), *
SXY (ω ) = SYX (ω ) = SXY (−ω ), (6.38)
where the first equation shows that auto-PSDs are real, even functions of ω .
Since this implies that there is no loss of information in considering only
the range of positive frequencies, one frequently encounters the so-called
one-sided spectral densities GXX (ω ), GXX (ν ), where ν = ω 2π is the ordinary
frequency in hertz and where the relations with SXX (ω ) are
GXX (ω ) = 2 SXX (ω ), GXX (ν ) = 4π SXX (ω ). (6.39)
Another property is obtained by setting τ = 0 in Equation 6.32b; this gives
KXX (0) = σ = 2
X
∫S
−∞
XX (ω ) dω , (6.40)
which can be used to obtain the variance of the (stationary) process X(t) by
calculating the area under its PSD curve. For cross-PSDs, Equation 6.382
shows that, in general, they are complex functions with a real part
Re {SXY (ω )} and an imaginary part Im {SXY (ω )}. In applications, these two
functions are often called the co-spectrum and quad-spectrum, respec-
tively, and sometimes are also denoted by special symbols like, for example,
CoXY (ω ) and QuXY (ω ).
If now we turn our attention to the first two derivatives of KXX (τ ), the
properties of Fourier transforms (Appendix B, Section B.2.1) give
F {KXX
′ (τ )} = iω F {KXX (τ )} , F {KXX
′′ (τ )} = −ω 2 F {KXX (τ )}, (6.41)
where on the r.h.s. of both relations we recognise the PSD SXX (ω ) = F {KXX (τ )},
while (owing to Equation 6.26 and Remark 6.9) the l.h.s, are F KXX (τ ) { }
{ }
  (τ ) , respectively. Since, by definition, these two transforms are
and −F KXX
the cross-PSD SXX (ω ) and the PSD SXX
  (ω ), we conclude that
SXX (ω ) = iω SXX (ω ),   (ω ) = ω SXX (ω ). (6.42)

SXX 2
The same line of reasoning applies to the third- and fourth-order derivatives
of KXX (τ ), and we get
  (ω ) = iω SXX (ω ),   (ω ) = ω SXX (ω ), (6.43)

3 4
SXX SXX
which are special cases of the general formula SX( j)X(k) (ω ) = (−1)k (iω ) j +k SXX (ω ),
where we write X ( j) to denote the process d j X(t) dt j .
Finally, it is worth pointing out that knowledge of SXX (ω ) allows us to
obtain the variances of X (t) and X
 (t) as
∞ ∞
σ =
2
X ∫S
−∞
 
XX (ω ) dω =
∫ω
−∞
2
SXX (ω ) dω
(6.44)
∞ ∞
σ X2 =
∫S
−∞
 
XX (ω ) dω =
∫ω
−∞
4
SXX (ω ) dω ,
which, like Equation 6.40, follow by setting τ = 0 in the inverse Fourier

  (ω ) and SXX
transform expressions of the PSDs SXX   (ω ), respectively, and then
taking the two rightmost relations of Equations 6.42 and 6.43 into account.
6.4.2 Narrowband and broadband processes

As suggested by the names themselves, the distinction between narrowband
and broadband processes concerns the extension of their spectral densities
in the frequency domain. Working, in a sense, backwards, we now inves-
tigate what kind of time histories and autocorrelation functions we expect
from these two types of processes.
The spectral density of a narrowband process is very small, that is zero
or almost zero, except in a small neighbourhood of a certain frequency ω 0.
A typical example is given by the spectral density of Figure 6.2, which has
the constant value S0 only in an interval of width ∆ω = ω 2 − ω 1 centred at ω 0
(so that ω 0 = (ω 1 + ω 2 ) 2) and is zero otherwise.
Then, the autocovariance function is obtained by inverse Fourier trans-
forming SXX (ω ), and we get
∞ −ω1 ω2
KXX (τ ) =
∫S
−∞
XX (ω ) e iωτ
dω = S0
∫e
−ω 2
iωτ
dω + S0
∫e
ω1
iωτ
dω

2S0 τ ∆ω 
(sin ω 2τ − sin ω 1τ ) = 0 sin 
4S
=  cos (ω 0τ ) , (6.45)
τ τ 2 
whose graph is shown in Figure 6.3 for the values ω 0 = 50 rad/s, ∆ω = 4 rad/s
(i.e. ∆ω ω 0 = 0.08) and S0 = 1 (Figure 6.3b is a detail of Figure 6.3a in the
neighbourhood of τ = 0). From Figure 6.2, it is immediate to see that the
area under SXX (ω ) is 2S0 ∆ω ; this is in agreement with Equation 6.40 because
setting τ = 0 in Equation 6.45 and observing that τ −1 sin (τ ∆ω 2) → ∆ω 2 as
τ → 0 gives exactly KXX (0) = σ X2 = 2S0 ∆ω .
So, since a typical narrowband process is such that ∆ω ω 0 << 1, its auto-
covariance function is practically a cosine oscillation at the frequency ω 0
enveloped by the slowly varying term ( 4S0 τ ) sin (τ ∆ω 2) that decays to zero
for increasing values of τ . Moreover, the fact that the frequency interval ∆ω
Figure 6.2 S pectral density of narrowband process.

10
-5
-10
-5,0 -2,5 0,0 2,5 5,0
(s)
Figure 6.3 Autocorrelation of narrowband process.
∞
is small means that we can rewrite Equation 6.441 as σ X2 ≅ ω 02
∫ −∞
SXX (ω )dω
and use it with Equation 6.40 to approximate the ‘characteristic frequency’
ω 0 of the process as the ratio of standard deviations
ω 0 ≅ σ X σ X . (6.46)
In the limit ∆ω → 0, the PSD tends to a ‘function’ that consists of two

Dirac-δ spikes at the points ω = ±ω 0 , and its inverse Fourier transform (see
Appendix B) gives
KXX (τ ) = 2S0 cos ω 0τ , (6.47)
which is a simple harmonic function at the frequency ω 0. Recalling from

Example 6.2 that a covariance function of the form 6.47 is associated with
a process of the form X(t) = U sin(ω t + V ), a conclusion of rather general
nature is that the time histories of a typical narrowband process look ‘quite
sinusoidal’.
At the other extreme, we find the so-called broadband processes, whose
spectral densities are significantly different from zero over a relatively large
range of frequencies. An example can be given by a process with a spectral
density like in Figure 6.2 but where now the two frequencies ω 1 , ω 2 are
much more further apart. For illustrative purposes, let us set ω 0 = 50 rad/s
and ∆ω = 80 rad/s (i.e. ω 1 = 10, ω 2 = 90) and draw a graph of the autocor-
relation function, which is still given by Equation 6.45. This graph is shown
in Figure 6.4 where, again, we chose S0 = 1.
200
150
100
50
-50
-100
-1,0 -0,8 -0,6 -0,4 -0,2 0,0 0,2 0,4 0,6 0,8 1,0
(s)
Figure 6.4 Autocorrelation of broadband process.
The fictitious process whose spectral density is equal to a constant S0

over all values of frequencies represents a mathematical idealisation called
white noise (by analogy with the approximately flat spectrum of white light
in the visible range of electromagnetic radiation) but it often turns out to
be useful in applications. The spectral density of such a process is clearly
non-integrable; however, we can once more use the Dirac delta ‘function’
and note that the Fourier transform of the covariance function
KXX (τ ) = 2π S0 δ (τ ) (6.48)
yields the desired spectral density SXX (ω ) = S0. A more realistic process,
called band-limited white noise, has a constant (with value S0) PSD only up
to a cut-off frequency ω = ω C . In this case, we get
2S0 sin ω C τ
KXX (τ ) = F−1 {SXX (ω )} = , (6.49)
τ
whose graph is shown in Figure 6.5 for the values ω C = 150 rad/s and S0 = 1.
Also, note that the area under the PSD is now σ X2 = 2S0ω C , in agreement
with the fact that the function 6.49 is such that KXX (τ ) → 2S0ω C as τ → 0.
Remark 6.13
If, in Equation 6.49, we define ε = 1 ω C (so that ε → 0 as ω C → ∞), we can

recall from Appendix B, Section B.3, that one of the representations of the
Dirac delta δ (x) is the limit of the function (π x)−1 sin(x ε ) as ε → 0. Then, as
300
200
100
-100
-0,8 -0,6 -0,4 -0,2 0,0 0,2 0,4 0,6 0,8
(s)
Figure 6.5 Autocorrelation of band-limited white noise.
intuitively expected, it follows that the covariance 6.49 tends to the white-
noise covariance of Equation 6.48 as ω C → ∞.
Because of Equation 6.48, a white-noise process is sometimes called

delta-correlated, where this term focuses the attention on the time domain
characteristics rather than on its flat spectrum in the frequency domain.
Since a delta-correlated process is such that its covariance is nonzero only
at the origin (so that the r.v.s X(t) and X(t + τ ) are uncorrelated even for
very small values of τ ), it is not difficult to figure out that the associated
time histories are very irregular, with a high degree of randomness. This
confirms the qualitative statement of Section 6.2.2 that the rapidity with
which KXX (τ ) decays to zero can be interpreted as indication of the ‘degree
of randomness’ of the process.
6.5 RESPONSE OF LINEAR SYSTEMS TO

STATIONARY RANDOM EXCITATION
We consider here the response of a deterministic linear system with

constant and non-random physical characteristics to the action of one or
more random excitations, which, unless otherwise stated, we assume to
be in the form of WS processes. With respect to our treatment, a differ-
ent level of sophistication is represented by dynamical models in which
also the system’s characteristics are considered to be random, thereby
contributing in their own right to the randomness of the output. Clearly,
these random-parameters systems produce a random output even if both
the initial conditions and the exciting loads are deterministic (see, for
example, Köylüoglu 1995).
6.5.1 SISO (single input–single output) systems

As a starting point, consider a system with one input/excitation and one
output (a so-called SISO system). Since from Chapters 4 and 5 we know
that the system’s dynamic behaviour is fully represented by the IRF h(t) in
the time domain or by the FRF H (ω ) = 2πF{h(t)} in the frequency domain,
we can write the system’s response to the random input load F(t) as the
process X(t) given by
X(t) =
∫ F(t − α ) h(α ) dα . (6.50)
−∞
Then, calling µF = E(F(t)) the mean input level, a first quantity of interest
is the mean output E(X(t)). Recalling that expectations and integrals com-
mute, we take expectations on both sides of Equation 6.50 to get
E(X(t)) = µF
∫ h(α ) dα = µ H(0), (6.51)
−∞
F
thus implying that E(X(t)) ≡ µX is a bounded, time-independent quantity if

∞
the system is stable, that is if
∫ −∞
h(α ) dα < ∞ (such systems are also called
bibo, an acronym for bounded input–bounded output). Also, note that the
rightmost equality in Equation 6.51 is obtained by simply setting ω = 0 in
∞
the relation H (ω ) =
∫−∞
h(t)e − iω t dt .
Example 6.5
Leaving the calculation to the reader, consider a damped 1-DOF. Since
its IRF is given by Equation 4.2a2 , the integral in Equation 6.51 leads to
∞
µF µF µ
µX =
mω d ∫e
0
−ζ ω n α
sin (ω dα ) dα =
m ω n2
= F . (6.52)
k
This result coincides with the rightmost term in 6.51 because for a
damped SDOF system, we have H (0) = 1 k (see Equation 4.81).
Assuming now without loss of generality that the input process has zero
mean (so that, by Equation 6.51, µX = 0) from Equation 6.50, we get
∞ ∞
X(t)X(t + τ ) =
∫ F (t − α ) h(α ) dα ∫ F(t + τ − γ ) h(γ ) dγ
−∞ −∞

∞ ∞
=
∫ ∫ h(α )h(γ ) F(t − α ) F(t + τ − γ ) dα dγ .
−∞ −∞
Taking expectations on both sides gives the autocorrelation function
∞ ∞
RXX (τ ) =
∫ ∫ h(α )h(γ ) R
−∞ −∞
FF (τ + α − γ ) dα dγ , (6.53)
which, under the assumption of stationary input, shows that RXX is

independent of absolute time t and depends only on the time interval τ .
Together, therefore, the equations above show that the output is WS sta-
tionary whenever the input is WS stationary. If, in particular, the input
is a white-noise process with RFF (t) = δ (t), then Equation 6.53 becomes
∞
RXX (τ ) =
∫ −∞
h(γ ) h (γ − τ ) dγ and the variance of the response (or E X 2 if ( )
the mean of the input process is not zero) is given by
∞
σ = RXX (0) =
2
X
∫ h (γ ) dγ . (6.54)
−∞
2
In the frequency domain, on the other hand, the double integral above turns
into a simpler relation. In fact, with the understanding that all integrals extend
from −∞ to ∞, we can Fourier transform both sides of Equation 6.53 to get
∫ { ∫∫ h(α )h(γ ) R }
1
SXX (ω ) = FF (τ + α − γ ) dα dγ e − iω τ dτ
2π

 1 
=
∫∫ h(α )h(γ )  2π ∫ R FF (τ + α − γ ) e − iω τ dτ  dα dγ ,

which, by introducing the variable y = τ + α − γ in the integral within curly
brackets, leads to
 1 
SXX (ω ) =
∫∫ h(α ) e iω α
h(γ ) e − iω γ 
 2π ∫R FF (y) e − iω y dy  dα dγ

= dα   dγ  SFF (ω ) = H ∗ (ω )H (ω ) SFF (ω ),
 ∫ h(α ) e iω α
 ∫ h(γ ) e − iω γ

from which we obtain the relation between the response PSD and the exci-
tation PSD
2
SXX (ω ) = H (ω ) SFF (ω ). (6.55)
At this point, observing that RXX (τ ) = F−1 {SXX (ω )}, we can use Equation
6.55 to obtain the response variance as
∞ ∞
∫S ∫ H(ω )
2
σ = RXX (0) =
2
X XX (ω ) dω = SFF (ω ) dω , (6.56)
−∞ −∞
thus implying that if X(t) is a displacement response, Equation 6.44 give the
variances of the velocity X (t) and acceleration X  (t).
Other quantities of interest are the cross-relations between input and output.
Starting once again from Equation 6.50, we now have F(t)X(t + τ ) =
∞
∫ −∞
F(t) F ( t + τ − α ) h(α ) dα , so that taking expectations, we get
RFX (τ ) =
∫ h(α )R
−∞
FF (τ − α ) dα , (6.57)
which, when Fourier-transformed, shows that the cross-PSDs between out-

put and input are given by
SFX (ω ) = H (ω ) SFF (ω ), SXF (ω ) = H ∗ (ω ) SFF (ω ), (6.58)
where the second relation follows from Equation 6.382 and from the fact
that SFF (ω ) is real. Note that Equations 6.58 – unlike Equation 6.55 that is a
real equation with no phase information – are complex equations with both
magnitude and phase information. In this respect, it can be observed that in
applications and measurements, they provide two methods to evaluate the
system’s FRF H (ω ); these are known as the H1 estimate and H 2 estimate of
the FRF and are given by
SFX (ω ) SXX (ω )
H1 (ω ) = , H 2 (ω ) = , (6.59)
SFF (ω ) SXF (ω )
where the first relation follows directly from Equation 6.581, while the sec-
ond is obtained by first rewriting Equation 6.55 as SXX (ω ) = H (ω )H ∗ (ω )SFF (ω )
and then using Equation 6.582 .
Remark 6.14
Ideally, Equation 6.59 should give the same result. Since, however, this is
not generally the case in actual measurements (typically, because of extra-
neous ‘noise’ that contaminates the input or output signals, or both), one
can use the H1 H 2 ratio as an indicator of the quality of the measurement
by defining the so-called coherence function
2
H1(ω ) SFX (ω )
γ 2 (ω ) = = ,
H 2 (ω ) SFF (ω )SXX (ω )
which, in some sense, is the counterpart of the familiar correlation coeffi-

cient used in ordinary regression analysis and is a measure of how well (on
a scale from 0 to 1) the output is linearly related to the input. More on these
aspects can be found in McConnell (1995) or in Shin and Hammond (2008).
6.5.2 SDOF-system response to
broadband excitation
From preceding chapters, we know that the FRF of a SDOF system with
( )
−1
parameters m, k, c is given by H (ω ) = k − mω 2 + iω c ; consequently,
2 1 1
H (ω ) = = , (6.60)
(k − mω ) 2 2
+ (ω c) 2

2
(
m  ω n2 − ω 2 )
2
+ ( 2ζ ω nω ) 
2

where ω n2 = k m is the system’s natural frequency. By Equation 6.55, the

response PSD SXX (ω ) is obtained by simply multiplying Equation 6.60 by
the input PSD SFF (ω ). So, when the excitation is in the form of a broadband
process – that is an ‘almost white noise’ with a flat spectrum over a wide
frequency range that includes ω n – and the system’s damping is low, then (a)
the function 6.60 is sharply peaked the vicinity of ω n and small everywhere
else, and (b) the variation SFF (ω ) in the region of the peak will be practically
negligible.
These two facts justify the approximation SFF (ω ) ≅ SFF (ω n ), and therefore,
SFF (ω n )
SXX (ω ) ≅ H (ω ) SFF (ω n ) =
2
, (6.61)
(
m2  ω n2 − ω 2 ) + (2ζ ω ω ) 
2 2

n
which shows that the system’s response is a narrowband process whose

2
PSD has the same characteristics as H (ω ) , that is a sharp peak in the
vicinity of ω n and small everywhere else. The height of the peak depends on
the amount of damping, and it is evident that for low values of damping,
we have a condition of ‘resonant amplification’ in which SXX (ω n ) can be

much greater than SFF (ω n ).
Then, the output variance can be obtained from Equation 6.56, and we get
∞
π SFF (ω n )
σ X2 ≅ SFF (ω n )
∫ H(ω )
2
dω = , (6.62)
kc
−∞
where the rightmost result can be obtained from tables of integrals (see the
following Remark 6.15).
Remark 6.15
With FRF functions of the form
B0 + (iω )B1 + (iω )2 B2 +  + (iω )k−1 Bk−1

H k (ω ) = ,
A0 + (iω )A1 + (iω )2 A2 +  + (iω )k Ak
∫
2
a list of integrals H k (ω ) dω for k = 1,2,5 can be found in Appendix
−∞
1 of Newland (1993).
( )
−1
If now we observe that c = 2mζ ω n, m = k ω n2 and H (ω n ) = 4ζ 2k2
2
,
Equation 6.62 can be rewritten as
σ X2 ≅ 2πζω n H (ω n ) SFF (ω n ), (6.63)

2

which, when compared with Equation 6.62, shows that we can evaluate the
∞
∫
2 2
area under the curve H (ω ) (the integral H (ω ) dω ) by calculating
−∞
the area of the rectangle whose horizontal and vertical sides, respectively,
are 2πζ ω n and the value H (ω n ) . Note that these are two quantities that
2
can be obtained from a graph of H (ω ) without knowing the values of the

system’s parameters m, k and c.
6.5.3 SDOF systems: transient response

The input–output relationships obtained above refer to the steady-state
condition, that is the situation in which the system has already had enough
time to ‘adjust’ to its state of motion. Since, however, it is evident that it
takes some time for the system to reach its steady state, this means that
for a certain interval of time right after the onset of the input, the system’s
response will be non-stationary.
In order to examine this transient response in a simple case, we assume

that the input is a zero-mean, WS process with correlation RFF (τ ) = R0 δ (τ )
and that it starts acting on our SDOF system at t = 0. Considering the vari-
ance of the response, we can start from Equation 6.54, which in this case
t
reads σ X2 (t) = R0
∫0
h2 (γ ) dγ . Then, using the explicit expression of h(t) for a
damped SDOF system, we have
t ωdt
R R
σ (t) = 2 0 2
2
X
m ωd ∫e
0
−2 ζ ω n γ
sin (ω dγ ) dγ = 2 0 3
2
m ωd ∫e
0
−2 ζ (ω n ω d ) y
sin2 y dy , (6.64)
where the second integral is a slightly modified version of the first obtained by
introducing the variable y = ω d γ . We did this because, with a = − 2ζ ω n ω d ,
the r.h.s. of 6.64 is of the same form as the standard tabulated integral
1  ax

∫e ax
sin2 xdx =
a2 + 4  ∫
e sin x ( a sin x − 2cos x ) + 2 e ax dx 

from which we can obtain the desired result
R0  −2ζ ω n t  ζ ωn 2ζ 2ω n2 2  
σ X2 (t) = 3 1 − e 1+ sin ( 2ω d t ) + sin (ω d t )  , (6.65)
4m ζ ω n 
2
 ωd ωd 2
 
thus implying
R0 R0 π S0 π S0
σ X2 = lim σ X2 (t) = 3 = = 3 = , (6.66)
t →∞ 4m ζ ω n 2 kc 2m ζ ω n
2 2
kc
where in writing the last two relations we observed that the PSD of the (white-
noise) input process is the constant S0 given by S0 = R0 F {δ (τ )} = R0 2π .
Note that – as it must be – Equation 6.66 gives precisely the steady-state
value of Equation 6.62.
If, as an illustrative example, we consider an SDOF system with a natural
frequency ω n = 10 rad/s, Figure 6.6 shows a graph of the ratio W (t) = σ X2 (t) σ X2
(i.e. the term within braces in Equation 6.65) for the two values of damp-
ing ζ = 0.05 and ζ = 0.10. From the figure, the effect of damping is rather
evident: The lower the damping, the slower the convergence of σ X2 (t) to its
steady-state value.
1,0
damp. = 10%
0,8
Variance ratio W(t)
damp. = 5%
0,6
0,4
0,2
0,0
0,0 1,0 2,0 3,0
seconds
Figure 6.6 Variance ratio W(t) – evolution towards steady state.
Remark 6.16
i. Incidentally, we note that the result σ X2 = π S0 kc can be used together

with Equation 6.46 rewritten as σ X2 ≅ ω n2 σ X2 to obtain the (steady-
state) variance of the velocity process as σ X2 = π S0 mc;
ii. In the discussion above, we focused our attention on the output vari-
ance because we assumed the input process to have zero mean. If,
however, the input is such that µF ≠ 0, it is evident that also the output
mean value will vary before reaching its steady-state value of Equation
t
6.52. Starting from the relation µX (t) = µF
∫ 0
h(α ) dα and using the
explicit expression of the IRF of an SDOF system, we leave it to the

reader to show that during the transient phase of motion, we have
µF  −ζ ω n t  ζω n 
µX (t) = 1 − e  cos ω d t + ω sin ω d t   , (6.67)
k  d 
which, as expected, tends to the steady-state value µF k of Equation

6.52 as t → ∞.
6.5.4 A note on Gaussian (normal) processes

In applications, Gaussian processes often play a major role for a number of rea-
sons; some of these that are without doubt worthy of mention are as follows:
• A Gaussian process is completely described by its first- and second-

moment properties, that is its mean and autocorrelation (or autoco-
variance) function;
• For two or more Gaussian processes considered jointly, uncorrelation

is equivalent to independence;
• The result of any linear operation – for example, differentiation and
integration – performed on a Gaussian process is another Gaussian
process. This includes the fact that the output of a linear system to a
Gaussian input is also Gaussian, while, in general, there is no simple
relation between the input and output probability distributions;
• The famous result known as Central Limit Theorem, which p rovides
theoretical support to the idea that a process can reasonably be
expected to be Gaussian (or nearly so) if many independent factors –
none of which is predominant on the others – contribute to its
randomness.
Given these important properties, it may now be useful here to briefly recall
some basic mathematical facts. A random variable X has a Gaussian (or
normal) distribution with mean µX and variance σ X2 if its pdf is
1  (x − µX )2 
fX (x) = exp  − . (6.68)
2π σ X  2σ X2 
The idea is extended to the bivariate case and the pdf of two jointly Gaussian
r.v.s X , Y is
1
fXY (x, y) = e − g (x,y), (6.69a)
σ 1σ 2 1 − ρ 2
where
1  ( x − m1 )2 ( y − m2 )2 2 ρ ( x − m1 ) ( y − m2 ) 
g(x, y) =  + −  (6.69b)
(
2 1 − ρ2 )  σ 1
2
σ 22 σ 1σ 2 
and where m1 , m2 and σ 12 , σ 22 are the means and variances of X , Y , respec-

tively. Also, ρ = Cov(X , Y ) σ 1σ 2 is the normalised covariance (also
called correlation coefficient between X and Y ). If the two variables are
uncorrelated – which, as pointed out above, is the same as independent in
the Gaussian case – then ρ = 0 and Equation 6.69 simplifies accordingly,
becoming fXY (x, y) = fX (x)fY (y), where fX (x), fY (y) are the two individually
Gaussian pdfs of X and Y .
Even more generally, the joint-pdf of n correlated Gaussian r.v.s X1 ,, Xn
is expressed in matrix notation as
1  1 
fX (x) = exp  − (x − m)T K −1(x − m) , (6.70)
( 2π ) n2
det K  2 
where X is the n-dimensional random vector formed with the variables

X1 ,, Xn , K is the covariance matrix with elements Kij = Cov ( Xi , X j ), and
m is the vector of mean values. If the variables are uncorrelated (hence inde-
pendent), then K is diagonal and Kij = 0 for i ≠ j . Clearly, Equation 6.69 is
the special case of Equation 6.70 for n = 2, and in this bivariate case, we have
 σ 12 ρ σ 1σ 2   σ 2 σ1 −ρ 
1
K=  ⇒ K −1 =  .
 ρ σ 1σ 2 σ 22  (
σ 1σ 2 1 − ρ 2 )  −ρ σ1 σ 2 
6.5.5 MIMO (multiple inputs–

multiple outputs) systems
Here, we extend the developments of Section 6.5.1 to the MIMO case by
considering a linear system with p (zero-mean and stationary) inputs/exci-
tations and q outputs. It is now convenient to turn to matrix notation and
denote by F(t) the p × 1 input vector and by X(t) the q × 1 output vector.
By so doing, Equation 6.50 generalises to
∞
X(t) =
−∞
∫ h(α ) F(t − α ) dα , (6.71a)
where h(t) is the q × p IRF matrix whose ijth element, the IRF hij (t),
represents the output/response at point i due to a unit Dirac delta excitation
Fj (t) = δ (t) applied at point j. Then
p ∞
X j (t) = ∑ ∫ h (α )F (t − α ) dα
k=1 −∞
jk k ( j = 1,, q ) (6.71b)
is the jth element of the vector X(t) and gives the response at the
jth point to the p inputs. Now, using Equation 6.71a together with
∞
X(t + τ ) =
∫ −∞
h(γ ) F(t + τ − γ ) dγ , we can form the product X(t)XT (t + τ ) and
take expectations on both sides to obtain the q × q correlation matrix

∞ ∞
R XX (τ ) =
∫ ∫ h(α ) R
−∞ −∞
FF (τ + α − γ ) hT (γ ) dα dγ , (6.72)
where R FF (τ ) = E  F(t) FT (t + τ )  is the p × p input correlation matrix.

Equation 6.72 can be Fourier-transformed and we can follow the same
line of reasoning leading to Equation 6.55 to obtain the q × q output PSD
matrix
SXX (ω ) = H∗ (ω ) SFF (ω ) HT (ω ), (6.73)
where the asterisk denotes complex conjugation and H(ω ) = 2π F {h(t)} is the
q × p FRF matrix whose j, kth element is H j k (ω ) = 2π F {hj k (t)}.
Along the same line, it is now not difficult to obtain the cross-quantities
between input and output; in the time and frequency domain, respectively,
we get
R FX (τ ) =
∫R
−∞
FF (τ − α ) hT (α ) dα , SFX (ω ) = SFF (ω ) HT (ω )
∞
(6.74)
R XF (t) =
∫ h(α )R
−∞
FF (τ + α ) dα , ∗
SXF (ω ) = H (ω )SFF (ω ),
where the ‘FX’ matrices have dimensions p × q , while the ‘XF’ matrices
are q × p. Also, note that by virtue of the properties of Equations 6.11 and
6.14, the matrix R FF (τ ) is such that RTFF (−τ ) = R FF (τ ), and so is R XX . On the
other hand, the matrices SFF (ω ) and SXX (ω ) are Hermitian (i.e. such that
∗
SFF (ω ) = SH
FF (ω ), or in terms of components, SFj Fk (ω ) = SFkFj (ω )), where S FF (ω )
H
is the complex conjugate of STFF (ω ).
Remark 6.17
i. The jkth element of, say, R XX (τ ) is the correlation function RX j Xk (τ ),

thus implying that the diagonal elements of the matrix are the out-
puts autocorrelations, while the off-diagonal elements are the outputs
cross-correlations. The same, clearly, applies to R FF (τ );
ii. Owing to Equation 6.742 , note that the PSD SXX (ω ) of Equation 6.73
can also be written as SXX (ω ) = H∗ (ω ) SFX (ω );
iii. The explicit expression of the j, kth (j, k = 1,, q ) element of the cor-
relation matrix 6.72 and the PSD matrix 6.73 are given by
RX j Xk (τ ) = ∑ ∫∫ h (α ) h
l , m=1
jl km (γ ) RFl Fm (τ + α − γ ) dα dγ
(6.75)
p
SX j Xk (ω ) = ∑H
l , m=1
∗
jl (ω ) H km (ω ) SFl Fm (ω ),
respectively, while the j, k th elements of the PSD matrices of Equation

6.74 are
p p
SFj Xk (ω ) = ∑l =1
SFj Fl (ω )H kl (ω ), SX j Fk (ω ) = ∑H
l =1
∗
jl (ω )SFl Fk (ω ); (6.76)
iv. Note that in the special case of multiple inputs and only one output
(i.e. q = 1), the matrices h(t), H(ω ) are 1 × p row vectors whose elements
are generally labelled by a single (input) index. So, for example, for
two inputs and one output, the matrix H(ω ) is the 1 × 2 row vector
 H1 (ω ) H 2 (ω ) , SFF (ω ) is a 2 × 2 matrix, and we have only one out-
put PSD given by
2 2
SXX (ω ) = H1 (ω ) SF1 F1 (ω ) + H 2 (ω ) SF2 F2 (ω )

+ H1∗ (ω ) H 2 (ω ) SF1 F2 (ω ) + H 2∗ (ω ) H1 (ω ) SF2 F1 (ω ).
On the other hand, for a single input–multiple outputs system, the

matrices h(t), H(ω ) are q × 1 column vectors whose elements are
labelled by a single (output) index. So, for example, in the case
of one input and two outputs, the matrix H(ω ) is the 2 × 1 column
vector [ H1 (ω ) H 2 (ω )]T , SFF (ω ) is a scalar quantity, and we get
SX j Xk (ω ) = H ∗j (ω )H k (ω )SFF (ω ) for j, k = 1,2. These ‘one-index-FRFs’
should not be confused with modal FRFs, which are also labelled
by a single index but are denoted by a caret (e.g. Hˆ j (ω )).
A final result concerns the mean value of the output, on which nothing has
been said because at the beginning of this section, we assumed the inputs to
have zero mean. When this is not the case, the input mean values µF1 ,, µFp
can be arranged in a p × 1 column vector m F and we can use Equation 6.71a
to obtain the output mean as
∞ 

 ∫
m X =  h(α ) dα  m F = H(0) m F, (6.77a)
−∞

where m X is the q × 1 column vector of components µX1 ,, µXq . Explicitly,

we have
p ∞ p
µX j = ∑ ∫
l =1
µFl hjl (α ) dα = ∑µ H
l =1
Fl jl (0) ( j = 1,, q ). (6.77b)
−∞
6.5.6 Response of MDOF systems

For our purposes, a particularly important case is represented by an
n-DOF system subjected to random excitation. If, as in preceding chapters,
we call M, C, K the system’s mass, damping and stiffness matrices, then

the response X(t) to the input F(t) is the solution of the vector equation
of motion MX  (t) + CX
 (t) + KX(t) = F(t) and we can now use the MIMO
relations obtained above together with the results of Chapter 4. There, we
recall, it was determined that in the case of classical damping – which we
are assuming is the case here – the n coupled equations of motion of an
n-DOF system can be uncoupled and expressed in the much simpler form
of n SDOF equations. This possibility, in turn, leads to the concept of nor-
mal or modal coordinates and – in the analysis of the system’s response
to an external excitation – to the concepts of modal IRFs hˆ j (t) and modal
FRFs Hˆ j (ω ). When arranged in the form of the two n × n diagonal matrices
( ) ( )
ĥ(t) ≡ diag hˆ j (t) and Ĥ (ω ) ≡ diag Hˆ j (ω ) , the relations with their physical
coordinates counterparts h(t), H(ω ) are given by (Equations 4.56 and 4.59)
h(t) = P hˆ (t) PT , ˆ (ω ) PT , (6.78)

H(ω ) = P H
where P is the n × n matrix of mass-orthonormal eigenvectors such

that PT M P = I and PT K P = L ≡ diag ( λ j ), and λ1 ,, λn are the system’s
eigenvalues.
At this point, we can use Equation 6.78 to rewrite Equations 6.72, 6.73
and 6.74 in terms of modal quantities; in particular, the frequency-domain
relation 6.73 becomes
ˆ ∗ (ω ) PT SFF (ω ) P H
SXX (ω ) = P H ˆ (ω ) PT , (6.79a)
ˆ T (ω ) = H
where we took into account that H ˆ (ω ) and that P∗ = P. Also, using
Equation 4.59a, the PSD matrix SXX (ω ) can be expressed as
 n
  n

SXX (ω ) = 

∑
l =1
Hˆ l∗ (ω ) pl pTl  SFF (ω ) 
 
∑ Hˆ
m=1
m (ω ) pm pTm 


n
= ∑ Hˆ (ω ) p p S
l ,m=1
∗
l l
T
l FF (ω )Hˆ m (ω ) pm pTm , (6.79b)
and the explicit form of its j, kth element is given by
SX j Xk (ω ) = ∑∑∑∑ p
l m r s
jl pml SFmFr (ω ) prs pks Hˆ l∗ (ω ) Hˆ s (ω ), (6.80)
where all sums are from 1 to n and Hˆ l (ω ), Hˆ s (ω ) are the lth and sth modal
FRFs. In particular, if (a) the components of the input vector are white-
noise processes, and (b) the modes of the system under investigation
are lightly damped and well separated, then the major contribution to
the sum on the r.h.s of Equation 6.80 will come from the square terms
of the form H l (ω ) ( l = 1,, n ) because, in comparison, the cross-terms
2
H l*(ω )H s (ω ) ( l ≠ s ) will generally be small. The consequence is that the PSD

6.80 will show the well-spaced peaks at the natural frequencies of the sys-
tem and that in the calculation of the mean square response E(X 2 ) or of σ X2 ,
the modal cross-terms can be neglected.
Remark 6.18
i. Similar considerations apply if the inputs, in addition to being white-

noise processes, are also uncorrelated. In this case, the PSD matrix
SFF (ω ) is diagonal and its nonzero elements (on the main diagonal) are
constants. In this case, moreover, the r.h.s. of Equation 6.80 simplifies
into a triple sum because SFm Fr is proportional to δ m r ;
ii. In Chapter 4, Section 4.4, we defined the modal force vector PT f , whose
jth element pTj f(t) was denoted by ϕ j (t). Now, in case of random excita-
tion with excitation vector F(t), we can similarly define the modal force
vector Q(t) = PT F(t). Then, we have Q(t)QT (t + τ ) = PT F(t)FT (t + τ ) P ,
so that taking expectations on both sides gives RQQ (τ ) = PT R FF (τ ) P ,
which, when Fourier-transformed, leads to SQQ (ω ) = PT SFF (ω ) P . This
last relation, in turn, shows that the PSD of Equation 6.79 can also
be expressed in terms of the PSD SQQ (ω ) of the modal force vector as
SXX (ω ) = PH ˆ ∗ (ω ) SQQ (ω )H
ˆ (ω )PT .
From this, we can obtain the matrix of mean square values
∞
( )
E XXT = R XX (0) =
∫ PHˆ (ω ) S
∗
QQ
ˆ (ω )PT dω ,
(ω )H
−∞
( )
whose diagonal elements are E X 2j , while the off-diagonal elements
are the cross-values E ( X j Xk ) with j ≠ k .
6.5.7 Response of a continuous system to distributed

random excitation: a modal approach
Consider for simplicity a one-dimensional system – for example a beam of
length L, where w(x, t) is the beam displacement at point x and time t –
under the action of a distributed random load F(x, t). Starting from the dis-
crete MIMO case of Section 6.5.5, we recall that Equation 6.752 gives the
output cross-PSD function between the jth and kth point of the system in
response to the excitation of p localised inputs applied at p different points
of the system. With the extension to continuous systems in mind, for our
present purposes we can rewrite this equation as
Sww ( x1 , x2 , ω ) = ∑ H ( x , r ,ω ) H ( x , r ,ω ) S
l , m=1
∗
1 l 2 m FF ( rl , rm , ω ), (6.81)
which gives the cross-PSD of the responses at two generic points x1 , x2 when
p localised inputs are applied at the points r1 ,, rp (the auto-PSD for the
response at x is then the special case x1 = x2 = x ). Then, for a distributed
excitation, we must pass to the limit p → ∞; by so doing, the sums become
spatial integrals on the beam length and we get
L L
Sww ( x1 , x2 , ω ) =
∫ ∫ H ( x , r ,ω ) H ( x , r ,ω ) S
0 0
∗
1 1 2 2 FF ( r1, r2 , ω )dr1dr2. (6.82)
If now we assume that the systems eigenpairs are known and recall from
Chapter 5 that the physical coordinates FRFs are expressed in terms of the
modal FRFs by means of Equation 5.143, we have
∞
H ( x1 , r1 , ω ) = ∑φ ( x )φ ( r ) Hˆ (ω )
j =1
j 1 j 1 j
(6.83)
∞
H ( x2 , r2 , ω ) = ∑φ ( x )φ ( r ) Hˆ (ω )
k=1
k 2 k 2 k
so that we can substitute these relations into Equation 6.82 to obtain

the response PSD in terms of the excitation PSD and the system’s (mass-
normalised) eigenfunctions as
∞
Sww ( x1 , x2 , ω ) = ∑φ ( x )φ ( x )G(ω ), (6.84a)

j ,k=1
j 1 k 2
where
L L
G(ω ) =
∫ ∫ Hˆ (ω )Hˆ (ω )φ ( r )φ (r )S
0 0
∗
j k j 1 k 2 FF ( r1, r2 , ω ) dr1 dr2 . (6.84b)
In obtaining Equation 6.84, we have considered only the frequency domain,

but at this point it can be instructive to see how we can arrive at this same
result by starting the analysis from the uncoupled equations of motion
(Equation 5.137) in the time domain. If now, for present convenience, we
L
call q j (t) the random modal excitation φ j (x) F(x, t) =
∫ 0
φ j (x)F(x, t) dx and
rewrite Equation 5.139 by shifting the function q j (t) instead of the modal
IRF hˆ j (t), a first result that can be easily obtained is the mean value of the
beam displacement. This is
∞ ∞
µw (x, t) = E [ w(x, t)] = ∑φ (x) ∫ E q (t − τ ) hˆ (τ ) dτ

j =1
j j j
−∞

∞ L ∞

= ∑  ∫ ∫
φ j (x)  φ j (ξ ) µF (ξ , t − τ ) d ξ  hˆ j (τ )dτ , (6.85)

j =1 −∞  0
where in writing the last relation we took into account the explicit expres-
sion of q j (t). If, moreover, the excitation is WS stationary, then µF is time
independent and we get
∞ ∞ L
µw (x) = ∑
j =1
∫ ∫
φ j (x) hˆ j (τ ) dτ µF (ξ )φ j (ξ ) d ξ
−∞ 0

∞ L
= ∑
j =1
∫
φ j (x)Hˆ j (0) µF (ξ )φ j (ξ ) d ξ , (6.86)
0
thus showing that the mean response is also time independent. In p articular,
if µF (x) = 0, then µw (x) = 0. Now, assuming this to be the case and by also
assuming the excitation to be WS stationary in time, the cross-correlation
function Rq jqk (τ ) between q j (t) and qk (t + τ ) is given by
L L
Rq jqk (τ ) = E q j (t)qk (t + τ )  =

∫ ∫ φ ( r )φ ( r ) R
0 0
j 1 k 2 FF ( r1, r2 , τ ) dr1 dr2 , (6.87)
where r1 , r2 are two spatial dummy variables of integration representing two

points along the beam. On the other hand, taking into account the expan-
sion of w(x, t) in terms of eigenfunctions (Equation 5.131), the correlation
of the beam displacement w can be written as
Rww ( x1 , x2 , τ ) = E  w ( x1 , t )w ( x2 , t + τ ) 
∞ ∞
= ∑
j ,k=1
E  y j (t)yk (t + τ )  φ j ( x1 ) φk ( x2 ) = ∑R
j ,k=1
y j yk (τ ) φ j ( x1 ) φk ( x2 ),
(6.88)
∞
where, observing that the modal response is y j (t) =
∫
−∞
hˆ j (α ) q j (t − α )dα ,
we have
∞ ∞
Ry j yk (τ ) = E  y j (t)yk (t + τ )  =
∫ ∫ hˆ (α ) hˆ (γ ) R
−∞ −∞
j k q j qk (τ + α − γ ) dα dγ . (6.89)
This, in turn, implies that the correlation of Equation 6.88 becomes
∞ ∞ ∞
Rww ( x1 , x2 , τ ) = ∑φ ( x )φ ( x ) ∫ ∫ hˆ (α )hˆ (γ ) R
j ,k=1
j 1 k 2 j k q j qk (τ + α − γ ) dα dγ , (6.90)
−∞ −∞
in which we can now substitute the expression of Rq jqk given by Equation

6.87 to obtain the response correlation Rww in terms of the correlation RFF
of the excitation as
Rww ( x1 , x2 , τ ) = ∑φ ( x )φ ( x ) g(τ ), (6.91a)

j ,k=1
j 1 k 2
where
∞ ∞ L L
g(τ ) =
∫ ∫ ∫ ∫ hˆ (α )hˆ (γ )φ ( r )φ ( r ) R
−∞ −∞ 0 0
j k j 1 k 2 FF ( r1, r2 , τ + α − γ ) dr1 dr2 dα dγ , (6.91b)
and it is now not difficult to show that the Fourier transform of the correla-
tion of Equation 6.91 leads exactly to the PSD of Equation 6.84.
Given this result, it follows that if, in a certain problem, we know the
PSD of the excitation, the response correlation is obtained as
∞ ∞ ∞
Rww ( x1 , x2 , τ ) =
∫ Sww ( x1 , x2 , ω ) e i ωτ dω = ∑ ∫
φ j ( x1 ) φk ( x2 ) G(ω ) e i ωτ dω ,
−∞ j ,k=1 −∞
(6.92)
where G(ω ) is given by Equation 6.84b. Also, note that this same correlation
can be obtained in terms of the PSD of the modal excitations as
∞ ∞
Rww ( x1 , x2 , τ ) = ∑
j ,k=1
∫
φ j ( x1 )φk ( x2 ) Hˆ ∗j (ω )Hˆ k (ω ) Sq jqk (ω ) e i ωτ dω , (6.93)
−∞
which follows from the fact that Sww ( x1 , x2 , ω ) is the Fourier transform of
Equation 6.90.
6.6 THRESHOLD CROSSING RATES AND

PEAKS DISTRIBUTION OF STATIONARY
NARROWBAND PROCESSES
Consider an ensemble xk (t) of time histories of duration T of a station-

ary narrowband process X(t) – like, for example, the response of a 1-DOF
system to broadband excitation with a flat spectrum that extends over the
system’s natural frequency. A first question we ask is if we can obtain some
information on the number of times in which the process crosses a given
threshold level x = a in the time T . More specifically, we focus our attention
on the number of upward crossings, that is crossings with a positive slope.
Calling na+ (k, T ) the number of such crossings in the kth sample function,
averaging over the ensemble leads to a number N a+ (T ) ≡ E  na+ (k, T ) , which
can be reasonably expected to be proportional to T in the light of the fact
that the process is stationary. So, by introducing the proportionality con-
stant ν a+, we can then write
N a+ (T ) = ν a+ T (6.94)
and interpret ν a+ as the average frequency of upward crossings of the thresh-

old x = a.
Now, turning to the process X(t) and restricting our attention to an infin-
itesimal time interval dt between the instants t and t + dt , we observe that
an up-crossing is very likely to occur within dt if three conditions are met:
(a) X(t) < a, (b) X (t) > 0, that is the derivative/slope of X(t) must be posi-
tive, and (c) X (t)dt > a − X(t), where this last condition means that the slope
must be steep enough to arrive at the threshold value within the time dt.
Then, rewriting condition (c) as X(t) > a − X (t) dt , the probability of having
a − X (t)dt < X(t) ≤ a and X (t) > 0 is given by
∞ a ∞ ∞

∫ ∫
0 a− xdt

fX X (x, x ) dxdx ≅
∫
0
fX X (a, x ) xdt
∫
 dx = dt fX X (a, x ) xdx
0
  , (6.95)
where fX X (x, x ) represent the joint-pdf of X and X , fX X (a, x ) means fX X (x, x )

evaluated at x = a and the approximation holds because dt is small. At this
point, we can put together Equation 6.95 with the interpretation of ν a+ given
∞
above to write ν a+ dt = dt
∫ 0
fX X (a, x ) xdx
  (see the following Remark 6.19),
and consequently,
+
ν =
a
∫f
0
X X (a, x ) xdx
  . (6.96)
Remark 6.19
The interpretation of ν a+ tells us that ν a+ dt is the average number of up-

crossings in time dt. Since, however, for small dt and relatively smooth
time histories such as those of a narrowband process, we have at most one
crossing in dt (in this respect, see also point (ii) of Remark 6.20), ν a+ dt can
also be interpreted as the probability that any one sample record chosen at
random will show a crossing in the time dt, that is the probability expressed
by Equation 6.95.
Equation 6.96 is a general result that holds for any probability distribu-
tion. In the special case of a Gaussian process with joint-pdf
1  x2 x 2 
fX X (x.x ) = exp  − − , (6.97)
2π σ Xσ X  2σ X2 2σ X2 
substitution in Equation 6.96 gives

∞
1 σ X − a2
∫
2 2
2σ X − x 2 2σ 2 2
2σ X
ν a+ = e−a 
xe X
dx = e (6.98)
2πσ Xσ X 2π σ X
0
because the result of the integral is σ X2 .
Remark 6.20
i. The absence of cross-terms xx in the exponential of Equation 6.97 is

due to the fact that X(t) and X (t ′) are uncorrelated for t = t ′ (see Remark
6.9(ii)). Moreover, since for a Gaussian processes, we recall, uncorrela-
tion is equivalent to independence, we have fX X (x.x ) = fX (x)fX (x ). Also
note that the form 6.97 implies that we assumed X(t) to have zero mean;
ii. The choice of threshold level a = 0 in Equation 6.98 gives the value
ν 0+ = σ X 2πσ X , which, in the light of the relatively smooth time his-
tories of narrowband processes, can be interpreted as an ‘average
frequency’ of the process X(t). With this interpretation in mind, we
can now go back to the ‘small’ quantity dt and consider that, on
more physical grounds, ‘small’ does not necessarily mean infinitesi-
mally small in the sense of calculus, but small in comparison with the
average period T0+ = 1 ν 0+ of upward zero crossings.
Using the results above, we can now go back to Section 6.5.2 and evaluate
ν a+ in the special case in which X(t) is the output of an SDOF system sub-
jected to a Gaussian white-noise excitation with constant PSD SFF (ω ) = S0.
The desired result can be readily obtained if we recall the relations
σ X2 = π S0 kc (Equation 6.66) and σ X2 = π S0 mc (Remark 6.16(i)), which can

be substituted in Equation 6.98 to give
ωn  a2kc 
ν a+ = exp  − , (6.99)
2π  2π S0 
where ω n = k m is the system’s natural frequency. Incidentally, note that

in line with the interpretation of Remark 6.20(ii), the average frequency ν 0+
of the output process is – not surprisingly – ω n 2π .
Considering again a stationary narrowband process X(t), we now ask
about the probability distribution of peaks, starting from the fact that if
we call fP (α ) the peak pdf, the probability of a peak with amplitude greater
than a is given by
P ( peak > a ) =
∫ f (α ) dα . (6.100)
a
P
At this point, recalling once again that the time histories of narrowband
processes are typically rather well behaved, we make the following emi-
nently reasonable assumptions: (a) Each up-crossing of x = a results in a
peak with amplitude greater than a and (b) each up-crossing of x = 0 cor-
responds to one ‘cycle’ of our smoothly varying time history. Under these
assumptions, it follows that the ν a+ ν 0+ ratio represents the favourable
fraction of peaks greater than a and that, consequently, we can write the
∞
equality
∫
a
fP (α ) dα = ν a+ ν 0+ . Then, differentiating both sides with respect
to a gives the first result
1 dν a+
− fP (a) = . (6.101)
ν 0+ da
In particular, if our process X(t) is Gaussian, we can use Equation 6.98 in

Equation 6.101 to obtain
a  a2 
fP (a) = 2 exp  − , (6.102)
σX  2σ X2 
which, in probability theory, is known as Rayleigh distribution (see Remark

6.21 below). Also, we can use Equation 6.98 in the r.h.s. of the relation
∞
∫ a
fP (α ) dα = ν a+ ν 0+ and, owing to Equation 6.100, obtain the probability
of a peak greater than a and the probability of a peak smaller than a as

(
P (peak > a) = exp − a2 2σ X2 )
(6.103)
P (peak ≤ a) = F (a) = 1 − exp ( − a
P
2
)
2σ X2 ,
respectively, where FP (a) is the Rayleigh PDF corresponding to the pdf fP (a).
If the probability distribution of the process X(t) is not Gaussian, it turns
out that the peaks distribution may significantly differ from the Rayleigh
distribution, but that a possible generalisation to non-Gaussian cases can
be obtained as follows: If we call a0 the median of the Rayleigh distribu-
tion, we can use Equation 6.1032 together with the definition of median –
that is the relation FP ( a0 ) = 1 2 – to obtain σ X2 = a02 2ln2. Substitution
of this result into FP (a) leads to the alternative form of the Rayleigh PDF
FP ( a a0 ) = 1 − exp  − ( a a0 ) ln2 , which, in turn, can be seen as a special
2
 
case of the more general one-parameter distribution
FP ( a a0 ) = 1 − exp  − ( a a0 ) ln2 , (6.104a)

k

 
in which, it should be noted, a0 is always the median of the peak height
irrespective of the value of k (in fact, we have FP ( a a0 = 1) = 1 2 for all k).
Then, the pdf fP ( a a0 ) corresponding to the PDF 6.104a is obtained by
differentiating this PDF; by so doing, we get
fP ( a a0 ) = (ln2) k ( a a0 ) exp  − ( a a0 ) ln2 , (6.104b)

k−1 k

 
which is sketched in Figure 6.7 for three different values of k and where the
case k = 2 corresponds to the Rayleigh pdf.
2,0
k=5
1,5
f(a/a0)
1,0
k=1
0,5 k=2
0,0
0,0 0,5 1,0 1,5 2,0 2,5 3,0
a/a0
Figure 6.7 Weibull pdf for different values of k.

Now, since in probability theory the distribution with PDF and pdf given,
respectively, by Equations 6.104a and 6.104b is known as Weibull distri-
bution (see the following Remark 6.21), we can conclude that the Weibull
distribution provides the desired generalisation to the case of non-Gaussian
narrowband processes.
Remark 6.21
The Weibull4 distribution, in its general form, has two parameters, while
the Rayleigh distribution has only one. Calling A, B the two parameters, a
r.v. Y is said to have a Weibull distribution if its PDF and pdf (both defined
for y ≥ 0 and zero for y < 0) are
( )
FY (y) = 1 − exp − Ay B , ( )
fY (y) = AB y B−1 exp − Ay B , (6.105)
respectively, where it can be shown that A is a ‘scale parameter’ that deter-

mines the spread of the values, while B is a ‘shape parameter’. On the other
hand, the r.v. Y is said to have a Rayleigh distribution with parameter R if
its PDF and pdf (where, again, both functions are defined for y ≥ 0 and zero
for y < 0) are, respectively,
(
FY (y) = 1 − exp − y 2 2R2 ,) ( ) ( )
fY (y) = y R2 exp − y 2 2R2 , (6.106)
where it is not difficult to show that the mean and variance of Y are
( )
E(Y ) = R π 2 ≅ 1.25R and σ Y2 = R2 2 ( 4 − π ) ≅ 0.43R2. Also, from
Equations 6.105 and 6.106, it can be readily seen that the Rayleigh distri-
bution is the special case of the Weibull distribution obtained for the choice
of parameters A = 1 2R2 and B = 2. Moreover, it should be noticed that for
B = 1, the Weibull distribution becomes the so-called exponential distribu-
tion fY (y) = A exp ( − Ay ).
At this point, we can obtain a final important result by observing that the
developments above provide a lower bound – which we will call M – on the
amplitude of the highest peak that may be expected to occur within a time
interval T . If, in fact, we call M the (as yet unknown) threshold level that,
+
on average, is exceeded only once in time T , then ν M T = 1 and therefore
+
ν M = 1 T . For a narrowband process, consequently, the favourable fraction
+
of peaks greater than M is ν M ν 0+ = 1 ν 0+T and this, we recall, is the probabil-
ity P ( peak > M ). Since, however, under the assumption that the peaks have a
Weibull probability distribution we have P ( peak > M ) = exp  − ( M a0 ) ln2 ,
k
 
the equality 1 ν 0+T = exp  − ( M a0 ) ln2  leads to
k
 
1k
M  ln(ν 0+ T ) 
= ,(6.107)
a0  ln2 
which, we reiterate, is valid for narrowband processes (with a Weibull peaks

distribution) when it can be assumed that any up-crossing of the zero level
corresponds to a full ‘cycle’ (and to a peak), thus implying that the average
number of cycles/peaks in time T is ν 0+T .
So, for example, with k = 1, Equation 6.107 shows that, on average, a
peak with an amplitude higher than four times the median (i.e. M a0 = 4)
can be expected every 16 cycles or, in other words, one peak out of 16
will exceed, on average, four times the median. If, on the other hand,
the peaks distribution is a Rayleigh distribution (i.e. k = 2), the average
number of cycles needed to observe one peak higher than four times the
median is 65,536, meaning, in other words, that one peak out of (approxi-
mately) 65,500 peaks will exceed an amplitude of four times the median.
Qualitatively, similar results should be expected just by visual inspection
of Figure 6.7, where we note that higher values of k correspond to pdfs
that are more and more strongly peaked in the vicinity of the median, and
hence to lower and lower probabilities for the occurrence of peak ampli-
tudes s ignificantly different from a0.
It is now left to the reader to sketch a graph of Equation 6.107 and plot
M a0 as a function of ν 0+T for different values of k.
Remark 6.22
In applications, the main parameter to be determined is the exponent k.

One way of doing so is by analysing sample records of the process and plot
the probability P(peak > a) – which here we denote by P for brevity – as a
function of a using appropriate scales. From Equation 6.104a, in fact, we
get ln P = − ( a a0 ) ln2, and consequently,
k
ln ( − ln P ) = k ln a − k ln a0 + ln ( ln2) (6.108)
so that k is the slope of the graph. On the other hand, the median peak
height a0 can be obtained from the zero intercept of the graph, which will
occur at a point a1 such that 0 = ln ( a1 a0 ) ln2 , from which it follows
k
 
1 = ( a1 a0 ) ln2 a0 = a1 ( ln2) . (6.109)
k 1k
⇒
Appendix A
On matrices and linear spaces
Although many – if not most – aspects of matrix analysis and manipu-

lation can be examined (almost) without even mentioning the notion of
finite-dimensional vector/linear space and by simply considering matrices
as ordered arrays of numbers associated with a certain set of specified rules,
this approach tends to conceal the key point that vectors, vector spaces and
linear transformations/operators between these spaces are the fundamen-
tal entities, while matrices are convenient ‘representations’ of these entities
particularly well suited for computational purposes. In this light, therefore,
finite-dimensional linear spaces provide the theoretical background behind
matrices, and this is the standpoint we adopt here – also because it provides
a necessary prerequisite for the study of linear spaces of infinite dimension.
A.1 MATRICES
Basically, a matrix A is a rectangular (and ordered) array of scalars from

a field F, which, for our purposes, will always be either the field of real
numbers R or the field of complex numbers C. The dimension of the array,
say m × n, specifies the number of rows (m) and the number of columns (n).
When m = n, we have a square matrix. The element in the i th row and j th
column of A (with 1 ≤ i ≤ m and 1 ≤ j ≤ n) is denoted by aij , and the symbol
[aij ] is often used in place of A to indicate the matrix itself. Also, the set
of all m × n matrices with elements in the field F is frequently denoted by
Mm× n (F), while the symbol Mn (F) is used for the set of square n × n matrices.
Matrices of the same dimension can be added entrywise – that is the ijth
element of the matrix C = A + B is cij = aij + bij – and the operation is associa-
tive and commutative; the matrix 0 with all zero entries is the identity with
respect to the addition, that is A + 0 = A. Multiplication of a matrix by a
scalar b ∈ F is also defined entrywise, and we have b A = [b aij ].
Two matrices can be multiplied only if the number of columns of the first
matrix equals the number of rows of the second. Then, multiplication of
255
256 Appendix A
a m × n matrix A of elements aij by a n × p matrix B of elements bij gives a

m × p matrix C whose ij th element cij is given by
n
cij = ∑a b
k=1
i k kj (1 ≤ i ≤ m, 1 ≤ j ≤ p ) (A.1)
so that, for example,
 1 −1 
 2 −1 3    14 17 
AB =   3 2  =  23 =C
 0 1 4   5 7  
30 
 
because c11 = 2 − 3 + 15 = 14, c12 = −2 − 2 + 21 = 17, and so forth.

In general, the matrix product is not commutative and AB ≠ BA (when,
clearly, the two expressions make sense). Matrix multiplication is associa-
tive and distributive over matrix addition; when the following products are
defined, therefore, we have
A (BC) = (AB) C, A(B + C) = AB + AC, (A + B) C = AC + BC. (A.2)
Two points worthy of mention are that, in general:
AB = 0 does not imply A = 0 or B = 0;

1.
AB = CB does not imply A = C (the reverse, however, is true and A = C
2.
implies AB = CB).
Given a m × n matrix A =  aij , its transpose is the n × m matrix AT =  a ji 

obtained by interchanging rows and columns of the original matrix. A
square matrix such that A = AT is called symmetric.
The Hermitian-adjoint (or Hermitian transpose) of A – denoted by A H –
is the matrix A H =  a*j i  whose elements are the complex conjugates of AT ;
clearly, if all the elements of A are real, then A H = AT . A square matrix with
complex entries such that A = A H is called Hermitian (or self-adjoint). So,
Hermitian matrices with real entries are symmetric, but symmetric matri-
ces with complex entries are not Hermitian. Both the transpose and the
Hermitian-adjoint of products satisfy a ‘reverse-order law’, which reads
(AB)T = BT AT , (AB)H = BH A H . (A.3)
There is, however, no reversing for complex conjugation and (AB)* = A* B*.
Other important definitions for square matrices are as follows:
1. If A = − AT , the matrix is skew-symmetric;

2. If A = − A H, the matrix is skew-Hermitian;
Appendix A 257
3. If AAT = AT A = I, the matrix is orthogonal;

4. If AA H = A H A = I, the matrix is unitary (so, a unitary matrix with
real entries is orthogonal).
In definitions 3 and 4, I is the unit matrix, that is the matrix whose only
nonzero elements are all ones and lie on the main diagonal; for example, the
2 × 2 and 3 × 3 unit matrices are
 1 0 0 
 1 0   
I2 =  , I3 =  0 1 0 ,
 0 1 
 0 0 1 
 
which, for brevity, can be denoted by the symbols diag(1,1) and diag(1,1,1).
The matrix I is a special case of diagonal matrix, where this name indicates
all (square) matrices whose only nonzero elements are on the main diag-
onal. On the other hand, one speaks of upper-(lower)-triangular matrix
if aij = 0 for j < i ( j > i ) and strictly upper-(lower)-triangular if aij = 0 for
j ≤ i ( j ≥ i ). Clearly, the transpose of an upper-triangular matrix is lower-
triangular and vice versa.
A square matrix is called normal if it commutes with its Hermitian-
adjoint, that is if AA H = A H A, and the reader can easily check the following
properties:
1. A unitary matrix is normal (therefore, a unitary matrix with real

entries, i.e. an orthogonal matrix, is normal);
2. A Hermitian matrix is normal (therefore, a Hermitian matrix with
real entries, i.e. a symmetric matrix, is normal);
3. A skew-Hermitian matrix is normal (therefore, a skew-Hermitian
matrix with real entries, i.e. a skew-symmetric matrix, is normal).
In order to simplify calculations, matrices can be partitioned, where

by this term we mean the subdivision of a matrix into a (convenient)
number of submatrices. The only constraint in doing this is that a line
of partitioning must always run completely across the original matrix.
So, for example, the 3 × 5 matrix below can be partitioned into four
submatrices as
 a11 a12 a13 a14 a15 

   A11 A12 
A =  a21 a22 a23 a24 a25 = ,
 a31 a32 a33 a34 a35   A 21 A 22 
 
258 Appendix A
where
 
 a11 a12   a13 a14 a15 
A11 =  , A12 = ,
 a21 a22   a23 a24 a25 
 

A 21 =  a31 a32  , A 22 =  a33 a34 a35  .
   
Provided that the partitioning is consistent with the operation we need

to carry out, partitioned matrices can be added, subtracted and multi-
plied. The reader is invited to check multiplication by first appropriately
partitioning two matrices A, B and then by multiplying the submatri-
ces according to the standard rule (Equation A.1) of matrix multiplica-
tion (i.e. as if the submatrices were individual elements and not matrices
themselves).
A.1.1 Trace, determinant, inverse

and rank of a matrix
Two important scalar quantities associated with square matrices are the
trace and the determinant. The trace of a matrix trA is the sum of its diago-
∑
n
nal elements (n elements if the matrix is n × n ), that is trA = ai i . The
i =1
determinant, on the other hand, involves all the matrix elements and has a
more complicated expression. It is denoted by det A or A and can be calcu-
lated by the so-called Laplace expansion
n n
det A = ∑
j =1
(−1)i + j aij det A ij = ∑ (−1)
i =1
i+ j
aij det A ij , (A.4)
where A ij is the (n − 1) × (n − 1) matrix obtained by deleting the ith row and

the jth column of A (and det A ij is often called the ijth minor of A). The first
sum in A.4 is the Laplace expansion along the ith row, while the second
sum is the Laplace expansion along the jth column (any row or column can
be chosen for the expansion because all expansions give the same result).
More specifically, the definition of determinant proceeds by induction by
defining the determinant of a 1 × 1 matrix to be the value of its single entry
(i.e. det[ a11 ] = a11), the determinant of a 2 × 2 matrix as
 a11 a12 
det   = a11a22 − a12 a21
 a21 a22 
Appendix A 259
and so on. Note that three consequences of the expansions A.4 are that (a)
the determinant of a diagonal or triangular matrix is given by the product
( ) ( )
of its diagonal elements, (b) det AT = det A and (c) det A H = (det A)∗. A
very useful property of determinants is
det(AB) = det A det B, (A.5)
which is important in applications where – as it frequently happens – a

matrix is factorised into a product of two, three or more matrices.
The inverse A −1 of a square matrix A is the matrix such that AA −1 = A −1A = I .
When A −1 exists – and in this case the matrix is called nonsingular – it is unique.
By contrast, a singular matrix has no inverse. When both A, B are nonsingu-
lar, we have the ‘reverse-order law’
(AB)−1 = B−1A −1. (A.6)
The existence of the inverse A −1 is strictly related to the value of det A –

more specifically, since Equation A.5 implies det A −1 = 1 det A , it depends
on whether det A is zero or not – and to a quantity called the rank of A and
denoted by rank A . This is defined as the dimension of the largest (square)
submatrix of A with a nonzero determinant (thus implying that the rank of
a nonsingular n × n matrix is n).
Remark A.1
Firstly, note that the definition of rank applies to any m × n matrix

and is not restricted to square matrices only. Clearly, if A ∈ Mm × n, then
rank A ≤ min(m, n), and when the equality holds, we say that A is a full-
rank matrix. Secondly, although the notion of linear independence will be
introduced in a later section, we anticipate here that the rank of A can be
alternatively defined as the maximum number of columns of A that form
a linearly independent set. This set of columns is not unique but the rank
( )
is, indeed, unique. Moreover, since rank A = rank AT , the rank can be
equivalently defined in terms of linearly independent rows and it can be
shown (see, for example, Hildebrand (1992)) that the definition in terms
of linear independence of rows or columns coincides with the definition in
terms of submatrices determinants.
Summarising in compact form a number of results on the inverse of a

matrix, we have the following proposition.
260 Appendix A
Proposition A.1
If A is a square n × n matrix, the following statements are equivalent
1. A −1 exists (i.e. A is nonsingular);

2. det A ≠ 0;
3. rank A = n (i.e. the matrix has full-rank);
4. The rows/columns of A are linearly independent;
5. If x is a column (n × 1) vector of unknowns, the only solution of the
system of linear equations Ax = 0 is x = 0;
6. The linear system Ax = b has a unique solution for each given (non-
zero) column vector b of scalars;
7. Zero is not an eigenvalue of A.
Although not strictly relevant to the discussion above, properties 5–7

have been given for completeness; properties 5 and 6 because the reader
is assumed to be familiar with systems of linear equations; and property 7
because the concept of eigenvalue – which plays an important role in the
main text – will be introduced and discussed in Section A.4.
A.2 VECTOR (LINEAR) SPACES
Underlying a vector space is a field F, whose elements are called scalars. As

in Section A.1, we will always assume it to be either the real field R or the
complex field C and, respectively, we will speak of real or complex vector
space. A vector (or linear) space V over the field F is a set of objects – called
vectors – where two operations are defined: addition between elements of V
and multiplication of vectors by scalars. Denoting vectors by boldface let-
ters and scalars by either Latin or Greek letters, the two operations satisfy
the following properties: For all x , y, z ∈V and all a, b ∈ F , we have
A1. x + y = y + x
A2. (x + y) + z = x + (y + z)
A3. There exists a unique vector 0 ∈V such that x + 0 = x
A4. There exists a unique vector − x ∈V such that x + (− x) = 0.
M1. a (x + y) = a x + a y
M2. (a + b) x = a x + b x
M3. (ab) x = a(b x)
M4. 1 x = x ,
where in M4 it is understood that 1 is the unity of F. Also, note that the

space is required to be closed under the two operations, which means that
we must have x + y ∈V for all x , y ∈V and a x ∈V for all a ∈ F and x ∈V .
Moreover, using the properties above, it is left to the reader to show that (a)
Appendix A 261
0 x = 0, (b) (−1) x = − x and (c) a 0 = 0, where 0 is the zero element of F, while

0 is the zero vector.
Example A.1
For any positive integer n, the set of all n-tuples of real (or com-
plex) numbers ( x1 , , xn ) forms a vector space on the real (com-
plex) field when we define addition and scalar multiplication term
by term, that is ( x1 , , xn ) + ( y1 , , yn ) ≡ ( x1 + y1 , , xn + yn ) and
a ( x1 , , xn ) ≡ ( a x1 , , a xn ), respectively. Depending on whether F = R
or F = C, the vector spaces thus obtained are the well-known spaces Rn
or C n , respectively.
Example A.2
The reader is invited to verify that the set Mm × n (F) of all m × n matrices
with elements in F is a vector space when addition and multiplication
by a scalar are defined as in Section A.1.
Other basic definitions are as follows:

A subspace U of V is a subset of V that is itself a vector space over the
same scalar field and with the same operations as V. The intersection (i.e.
the familiar set operation ∩) of linear subspaces of V is a linear subspace
of V.
A set of vectors u1 ,, u k ∈V ( k ≥ 1) is linearly independent if the relation
a1 u1 +  + ak u k = 0 implies a1 = a2 =  = ak = 0. The set is linearly depen-
dent if it is not linearly independent. Thus, the set is linearly dependent if
at least one of the scalars a j is nonzero. The sum a1 u1 +  + ak u k is called a
linear combination of the vectors u1 ,, u k .
A set of vectors u1 ,, u n ∈V is said to span the space V if every vector
x ∈V can be written as a linear combination of the u i , that is if
n
x= ∑ x u (A.7)
i =1
i i
for some set of scalars x1 ,, xn ∈ F . The set of vectors is called a basis of
V if (a) the set spans V and (b) the set is linearly independent. In general, a
vector space V has many bases but the number of elements that form any
one of these bases is defined without ambiguity. In fact, it can be shown
(Halmos (2017) that all bases of V have the same number of elements; this
number is called the dimension of V and denoted by dim V . Equivalently,
but with different wording, V is called n-dimensional – that is dim V = n – if
it is possible to find (in V) n linearly independent elements but any set of
n + 1 vectors of V is linearly dependent. When, on the other hand, we can
find n linearly independent elements for every n = 1,2,, V is said to be
262 Appendix A
infinite-dimensional. Clearly, for every linear subspace U of V, we have

dim U ≤ dim V .
If u1 ,, u n (or, for short, {u i } in=1) is a basis of V, then any x ∈V can be
written as in Equation A.7. The scalars xi are called the components (or
coordinates) of x relative to the basis {u i } ni =1, and we can say that the basis
{u i }ni=1 provides a coordinate system for V. The emphasis on the terms ‘rela-
tive to the basis’ and ‘coordinate system’ means that we expect the same
vector to have different components relative to another basis/coordinate
system { v i }i =1.
n
Example A.3
The standard basis in the familiar (three-dimensional) space R3 is
given by the three vectors i = (1,0,0), j = (0,1,0), k = (0,0,1), and the vec-
tor x = 2 i + 3 k has components (2,0,3) relative to this basis. If now we
consider the set of vectors e1 = i + j, e2 = i + 2k , e3 = i + j + k , it is left to
the reader to show that this set is a basis and that the components of x
relative to it are (1, 2, −1).
Having stated that we expect the components of a vector to change under

a change of basis, let us now consider this point. If {u i }i =1 and { v i }i =1 are
n n
two basis of V, then (a) an arbitrary vector x ∈V can be expressed in terms

of its components relative to the u- and v-basis and (b) each vector of the
second basis can be expressed as a linear combination of the vectors of the
first basis. In this light, we can write the relations
n n n
x= ∑x u ,
i =1
i i x= ∑ xˆ v ,
j =1
j j vj = ∑c
k=1
kj u k , (A.8)
where the scalars x1 ,, xn and xˆ 1 ,, xˆ n are the components of x relative
to the u- and v-basis, respectively, while the n2 scalars ckj in Equation
A.83 (which, it should be noted, is a set of n equation; one for each j with
j = 1,2,, n ) specify the change of basis. Then, it is not difficult to see that
substitution of Equation A.83 into A.82 leads to x = ∑k ( ∑ j ckj xˆ j ) u k , which,
in turn, can be compared with Equation A.81 to give
xk = ∑c
j =1
kj xˆ j , [ x ]u = C [ x ] v ,(A.9)
where the first of Equations A.9 is understood to hold for k = 1,, n (and is
therefore a set of n equations), while the second comprises all these n equa-
tions is a single matrix relation when one introduces the matrices
Appendix A 263
 x1   xˆ 1   c11 c12  c1n 

     
x2 xˆ 2  c21 c22  c2n
[ x ]u =  , [ x ]v =  , C =  ,

        
     
 xn   xˆ n   cn1 cn2  cn n 
(A.10)
where C is the change-of-basis matrix, while [ x ]u , [ x ]v are the coordinate

vectors of x relative to the u- and v-basis, respectively.
Remark A.2
Note that the transformation law of components – the first of Equations

A.9, with the sum on the second index – is different from the transforma-
tion law of basis vectors – Equation A.83, where the sum is on the first
index. Also, note that x is an element of V, while the coordinate vectors
[ x ]u , [ x ]v are elements of Rn (if F = R ) or C n (if F = C ).
An important point is that the change-of-basis matrix C is nonsingular.

In fact, by reversing the roles of the two bases and expressing the u vectors
in terms of the v vectors by means of the n relations u i = ∑ j cˆ j i v j , the same
argument leading to Equations A.9 leads to
xˆ j = ∑ cˆ i
ji xi , [ x ]v = Cˆ [ x ]u, (A.11)
where, as in A.9, the second relation is the matrix version of the first. But
then Equations A.9 and A.11 together give CC ˆ = CCˆ = I, which in turn
ˆ −1
proves that C = C as soon as we recall from Section A.1.1 that the inverse
of a matrix – when it exists – is unique. It is now left to the reader to show
that in terms of components the relations CC ˆ = CCˆ = I read
∑ cˆ
j
ij cj k = ∑cj
kj cˆ j i = δ i k, (A.12)
where δ i k is the well-known Kronecker symbol, equal to 1 for i = k and zero

for i ≠ k.
A.2.1 Linear operators and isomorphisms

If V ,U are two vector spaces of dimensions n, m , respectively, over the same
field F, a mapping A : V → U such that
A (x + y) = A x + A y, A (a x) = a (A x) (A.13)
264 Appendix A
for all x , y ∈V and all a ∈ F is called a linear operator (or linear transfor-
mation) from V to U, and it is not difficult to show that a linear operator is
completely specified by its action on a basis u1 ,, u m of its domain V. The
set of all linear operators from V to U becomes a linear (vector) space itself –
and precisely an nm-dimensional linear space often denoted by L(V ,U) –
if one defines the operations of addition and multiplication by scalars as
(A + B) x = A x + B x and (aA) x = a (A x), respectively.
In particular, if U = V , the linear space of operators from V to itself is
denoted by L(V ). For these operators, one of the most important and far-
reaching aspects of the theory concerns the solution of the so-called eigen-
value problem A x = λ x, where x is a vector of V and λ is a scalar. We have
encountered eigenvalue problems repeatedly in the course of the main text,
and in this appendix, we will discuss the subject in more detail in Section A.4.
Particularly important among the set of linear operators on a vector space
are isomorphisms. A linear operator is an isomorphism (from the Greek
meaning ‘same structure’) if it is both injective (or one-to-one) and surjec-
tive (or onto), and two isomorphic linear spaces – as far as their algebraic
structure is concerned – can be considered as the same space for all practi-
cal purposes, the only ‘formal’ difference being in the name and nature of
their elements. In this regard, a fundamental theorem (see, for example,
Halmos (2017)) states that two finite-dimensional vector spaces over the
same scalar field are isomorphic if and only if they have the same dimen-
sion, and a corollary to this theorem is that any n-dimensional real vector
space is isomorphic to Rn and any n-dimensional complex vector space is
isomorphic to C n .
More specifically, if V is an n-dimensional space over R (or C) and
u1 ,..., u n is a basis of V; the mapping that to each vector x ∈V associates
the coordinate vector [ x ]u , that is the n-tuple of components ( x1 ,, xn ), is
an isomorphism from V to Rn (or C n ). The immediate consequence of this
fact is that the expansion (A.7) in terms of basis vectors is unique and that
choosing a basis of V allows us to multiply vectors by scalars and/or sum
vectors by carrying out all the necessary calculations in terms of compo-
nents, that is by operating in Rn (or C n ) according to the operations defined
in Example A.1.
If A : V → U is an isomorphism, then it is invertible (or nonsingular),
which means that there exists a (unique) linear operator B : U → V such
that B (A x) = x for all x ∈V and A (By) = y for all y ∈U . The operator B is
denoted by A−1 and is an isomorphism itself.
A.2.2 Inner products and orthonormal bases

So far, our vector space is only endowed with an algebraic structure and no
mention has been made of metric properties such as length of a vector and
angle between two vectors. This richer structure is obtained by introducing
Appendix A 265
the notion of inner product, where an inner (or scalar) product is a map-
ping • • from V × V (the set of ordered pairs of elements of V) to the field
F satisfying the defining properties
∗
IP1. x y = y x
IP2. x a y1 + b y 2 = a x y1 + b x y 2
IP3. x x ≥ 0 and x x = 0 if and only if x = 0,
where the asterisk denotes complex conjugation and can be ignored for
vector spaces on the real field. It can be shown (see, for example, Horn
and Johnson 1993) that an inner product satisfies the Cauchy–Schwarz
inequality: For all x , y ∈V
2
xy ≤ x x y y (A.14)
and the equal sign holds if and only if x , y are linearly dependent. Moreover,
using the inner product, we can define the length x of a vector (or norm
in more mathematical terminology) and the angle θ between two nonzero
vectors as
x y
x = x x , cos θ = , (A.15)
x y
where 0 ≤ θ ≤ π 2 and it can be shown that these quantities satisfy the usual
properties of lengths and angles. So, for example, in the case of the vector
spaces Rn or C n of Example A.1, it is well known that the ‘standard’ inner
product of two elements x = ( x1 ,, xn ) , y = ( y1 ,, yn ) is defined as
n n
x y R n ≡ ∑
i =1
xi yi , x y C n ≡ ∑ x y , (A.16a)
i =1
*
i i
which, by Equation A.151, lead to the Euclidean norm (or l2-norm) on C n

(or Rn )
∑
2
x = xi . (A.16b)
i
Remark A.3
i. Properties IP1 and IP2 imply a x1 + b x 2 y = a* x1 y + b* x 2 y ,

which means that the inner product is linear in the second slot and
conjugate-linear in the first slot. This is what we may call the ‘physi-
cists’ convention’ because mathematicians, in general, define linearity
266 Appendix A
in the first slot (so that conjugate linearity turns up in the second slot).
In real spaces, this is irrelevant – in this case, in fact, property IP1
reads x y = y x , and we have linearity in both slots – but it is not so
in complex spaces. With linearity in the first slot, for example, instead
of Equation A.18a below, we have x j = x u j ;
ii. A second point we want to make here is that, from a theoretical stand-
point, the concept of norm is more ‘primitive’ than the concept of
inner product. In fact, a vector norm can be axiomatically defined as
a mapping • : V → R satisfying the properties:
N1. x ≥ 0 for all x ∈V and x = 0 if and only if x = 0

N2. a x = a x
N3. x + y ≤ x + y (triangle inequality),
thus implying that we do not need an inner product in order
to define a norm. If, however, V is an inner product space, the
‘natural’ norm of V – which is a norm because it satisfies proper-
ties N1–N3 – is defined as in Equation A.151 and in this case, one
speaks of norm generated by the inner product. For these norms,
we have the additional property that for all x, y ∈V
2 2
N4. x + y + x − y = 2 x + y
2
( 2
)
(parallelogram law).
On the other hand, if a norm does not satisfy N4, then it is not
generated by an inner product and there is no inner product such
2
that x = x x .
Two vectors are said to be orthogonal if their inner product is zero, that
is x y = 0, and in this case, one often writes x ⊥ y. In particular, a basis
u1 ,, u n is called orthonormal if the basis vectors have unit length and are
mutually orthogonal, that is if
u i u j = δ ij (i, j = 1, , n).(A.17)
The use of orthonormal bases is generally very convenient and some

examples will show why it is so. Firstly, if V is an n-dimensional vector
space and u1 ,, u n an orthonormal basis, we can express the components
x j of a vector x ∈V as an inner product. In fact, taking the inner prod-
uct of both sides of the expansion A.7 with u j and using Equation A.17
gives
xj = u j x ( j = 1, 2, , n). (A.18a)
Then, by a similar argument, it is not difficult to show that the elements of

the change-of-basis matrices C, Cˆ of Section A.2 can be written as
cij = u i v j , cˆ ij = v i u j (i, j = 1, , n). (A.18b)

Appendix A 267
Secondly, the inner product of two vectors x , y ∈V can be expressed in a

number of different forms and we can write
n n
 y1 
   
xy V
= ∑x y = ∑ x u
i =1
∗
i i
i =1
i u i y = [ x ] [ y ]u =  x
H
u

∗
1  x ∗
n

 yn
,

 
(A.19)
where xi , yi ( i = 1,, n ) are the components of x , y relative to the u-basis and

the first equality follows from the chain of relations
n
xy V
= ∑ i
xi u i ∑ j
yju j = ∑i,j
xi∗y j u i u j = ∑ i,j
xi∗y j δ ij = ∑x y ,
i =1
∗
i i
while we used Equation A.18a in writing the second equality (and clearly
[ x ]u ,[ y ]u are the coordinate vectors formed with the components xi , yi).
Note that in writing Equation A.19, we assumed V to be a complex space;
however, since in a real space complex conjugation can be ignored and
[ x ]H T
u is replaced by [ x ]u , it is immediate to see that in this case, we have
T
x y V = ∑ i xi yi = [ x ]u [ y ]u .
Finally, a third point concerns the change-of-basis matrix C, which we
know from Section A.2 to be nonsingular. Since, however, orthonormal
bases are special, we may ask if the change-of-basis matrix between ortho-
normal bases – besides being obviously nonsingular – is also special in some
way. The answer is affirmative, and it turns out that C is unitary in complex
spaces and orthogonal in real spaces. In fact, if u1 ,, u n and v1 ,, v n are
two orthonormal bases of the complex space V, then we have v j v k = δ j k
and we can use Equation A.83 together with the properties of the inner
product and the orthogonality of the u-basis to write
δ j k = v j vk = ∑∑c c
i m
∗
ij mk ui um = ∑∑c c
i m
∗
ij mk δim = ∑c c
i
∗
ij ik ,
whose matrix version reads C H C = I (CT C = I in real spaces) and means

that, as stated above, the change-of-basis matrix is unitary (orthogonal).
A.2.2.1 The Gram–Schmidt orthonormalisation process

As a matter of fact, any set x1 ,, x n of n linearly independent vectors in an
n-dimensional space V is a basis of V. But since, as mentioned above, ortho-
normal bases are particularly convenient, it is interesting to note that it is
always possible to use the set x1 ,, x n as a starting point to construct an
268 Appendix A
orthonormal basis. In principle, such an ‘orthonormalisation process’ can

be accomplished in many different ways, but there is a simple and widely
used algorithm – called the Gram–Schmidt process – which we outline here
briefly.
Starting with our arbitrary basis x1 ,, x n , define y1 = x1 and then the
unit-length vector u1 = y1 y1 . Next, define y 2 = x 2 − u1 x 2 u1 (note that
y 2 ⊥ u1) and u 2 = y 2 y 2 , which makes u 2 a vector of unit length. By iter-
ating the process and assuming that u1 ,, u k−1 ( k − 1 < n ) have been deter-
mined, we now define the vector y k = x k − u k−1 x k u k−1 −  − u1 x k u1,
and then set u k = y k y k . When the process has been repeated n times, we
shall have generated the orthonormal basis u1 ,..., u n. We observe in passing
that at each step the orthonormal vectors u1 ,..., u k are linear combinations
of the original vectors x1 ,, x k only.
A.2.3 An observation on inner products

If V is a complex inner product space and u1 ,, u n an orthonormal basis
of V, we have seen that the inner product of two vectors x , y ∈V can be
expressed in terms of coordinate vectors as the matrix product [ x ]H u [ y ]u ,
which, in turn, can be equivalently written as [ x ]H
u I [ y ]u – where I is the n × n
unit matrix. If, on the other hand, the basis u1 ,..., u n is not orthonormal –
and in this case, we say that u1 ,, u n provides an oblique coordinate system
for V – we can introduce the n2 scalars uij = u i u j and obtain the more
general relation
xy = ∑x y
i,j
*
i j ui u j = ∑x u
i,j
*
i ij yj = [x] H
u U [ y ]u , (A.20)
where in the last expression U is the n × n matrix of the uij. Since, however, the
choice of another nonorthonormal basis v1 ,, v n leads to x y = [ x ] H v V [ y ]v,
it turns out that the result of the inner product seems to depend on the basis.
The fact that it is not so can be shown by using Equation A.83 and writing
vij = v i v j = ∑c
k ,m
*
c
ki m j uk u m = ∑c
k ,m
*
ki ukm cm j (i, j = 1,, n),
which in matrix form reads V = C H U C and implies
xy V
= [x] H H H
v V [ y ] v = [ x ] v C U C [ y ] v . (A.21)
Then, noting that Equation A.112 (since Cˆ = C −1) implies [ x ] H H

v = [x] u C
−H
−1
and [ y ]v = C [ y ] u , we can use these results in Equation A.21 to arrive at
[x] H H
v V [ y ] v = [ x ] u U [ y ]u, (A.22)
Appendix A 269
thus fully justifying the fact that the inner product is used to define metric
quantities such as lengths and angles, that is quantities that cannot depend
on which basis – orthonormal or not – we may decide to choose.
Remark A.4
Clearly, the discussion above covers as a special case the situation in

which either of the two basis is orthonormal and the other is not. The
even more special case in which both the u- and v-basis are orthonormal
corresponds to U = V = I (or equivalently uij = vij = δ ij ), and now the rela-
tion [ x ] H H
v [ y ] v = [ x ] u [ y ] u follows immediately from the fact that the change-
of-basis matrix is unitary (i.e. C H = C −1). The equality [ x ] H H
v [ y ] v = [ x ] u [ y ] u,
moreover, is the reason why, with orthonormal bases, we can omit the sub-
script and express the inner product x y V in terms of components and
write simply x H y (or x T y in real spaces).
A.3 MATRICES AND LINEAR OPERATORS
Given a basis u1 ,, u n of an n-dimensional vector space V on a field F, we

saw in Section A.2 that the mapping that associates a vector x ∈V with its
coordinate vector [ x ]u – where, explicitly, [ x ]u ∈ F n is the set of n scalars
( x1,, xn ) arranged in the form of a column matrix – is an isomorphism.
This, in essence, is the reason why we can manipulate vectors by operating
on their components. The components, however, change under a change
of basis – a fact expressed in mathematical terminology by saying that the
isomorphism is not canonical – and the way in which they change was also
discussed in Section A.2.
In a similar way, given a basis of V, a linear transformation (or opera-
tor) T : V → V is represented by a n × n matrix [ T ]u ∈ Mn (F), and it can be
shown (e.g. Halmos (2017)) that the mapping that associates T with [ T ]u is
an isomorphism from the vector space L(V ) to the vector space Mn (F). The
simplest examples are the null and identity operators, that is the operators
Z : V → V and I : V → V such that Z x = 0 and I x = x for all x ∈V , which
are represented by the null matrix 0n and the identity matrix I n, respectively
(even though these two operators are indeed special because they are repre-
sented by the matrices 0n , I n for any basis of V).
However, since this isomorphism is also noncanonical and different
matrices can represent the same operator, the question arises as to whether
there exists a relation among these matrices. We will see shortly that this
relation exists and is called similarity.
If u1 ,, u n and v1 ,, v n are two bases of V, C is the change-of-basis
matrix and x, y are two vectors of V, we know from Section A.2 that their
coordinate vectors relative to the two bases satisfy the relations
270 Appendix A
[ x ]u = C[ x ]v , [ x ]v = C −1[ x ]u , [ y ]u = C[ y ]v , [ y ]v = C −1[ y ]u . (A.23)
If now we let T : V → V be a linear operator and we let y be the vector such

that y = T x, it is reasonable (and correct) to expect that this last relation can
be written in terms of components in either one of the two forms
[ y ]u = [ T ]u [ x ]u , [ y ]v = [ T ]v [ x ]v , (A.24)
where [ T ]u and [ T ]v are two appropriate n × n matrices that represent T in

the u- and v-basis, respectively. Using the first and third of Equations A.23,
we can rewrite Equation A.241 as C [ y ]v = [ T ]u C [ x ]v. Then, premultiplying
both sides by C −1 gives [ y ]v = C −1[ T ]u C [ x ]v , which, when compared with
Equation A.242 implies
[ T ]v = C −1[ T ]u C, [ T ]u = C [ T ]v C −1, (A.25)
where the second equation follows immediately from the first. But since, by
definition, two square matrices A, B are said to be similar if there exists a
nonsingular matrix S – called the similarity matrix – such that B = S−1AS
(which implies A = SBS−1), Equations A.25 show that (a) the matrices
[ T ]u , [ T ]v are similar and (b) the change-of-basis matrix C plays the role of
the similarity matrix.
At this point, the only piece missing is the explicit determination of the
elements of [ T ]u and [ T ]v , that is the two sets of scalars t ij and tˆij (n2 scalars
for each set). This is a necessary step because we can actually use Equations
A.25 only if we know – together with the elements of C and/or Cˆ = C −1 – at
least either one of the two sets. These scalars are given by the expansions of
the transformed vectors of a basis in terms of the same basis, that is
T uj = ∑t k
kj uk , T vj = ∑ tˆ
k
kj vk ( j = 1, , n ). (A.26)
In particular, if the two bases are orthonormal, it is not difficult to show

that we have the expressions
t ij = u i T u j , tˆij = v i T v j (i, j = 1, , n). (A.27)
Remark A.5
If we recall that the elements of C, Cˆ are given by the expansions of the vec-
tors of one basis in terms of the other basis, that is
vj = ∑c
k
kj uk , uj = ∑ cˆ
k
kj vk ( j = 1, , n), (A.28)
Appendix A 271
we can use Equations A.26 and A.28 to write Equations A.25 in terms of
components; in fact, starting from Equation A.281, we have the chain of
equalities
 
T vj = ∑c k
kj (T uk ) = ∑ ckj  ∑ tiku i 
k  i 
   
= ∑ k ,i
ckjt ik 

∑
m
cˆ m i v m  =

∑∑
m

 i ,k
cˆ m it i kck j  v m

and this can be compared to Equation A.262 to give

n
tˆ m j = ∑ cˆ
i ,k=1
t c
m i ik kj (m, j = 1, , n), (A.29)
which is Equation A.251 written in terms of the elements of the various

matrices. By a similar line of reasoning, the reader is invited to obtain the
‘components version’ of A.252 .
Example A.4
As a simple example in R2 , consider the two bases
 1   0   1   −2 
u1 =  , u2 =  ; v1 =  , v2 =  ,
 0   1   1   0 
where u1 , u 2 is the standard (orthonormal) basis of R2 , while v1 , v 2 is

not orthonormal. Then, it is immediate to see that Equations A.28 lead
to the change-of-basis matrices
 c11 c12   1 −2   cˆ11 cˆ12   0 1 

C= = , Cˆ =  = ,
 c21 c22   1 0   cˆ21 cˆ22   − 1 2 12 

where, as expected, Cˆ C = C Cˆ = I. Now, if we consider the linear trans-

formation T : R2 → R2 whose action on a generic vector x (with com-
ponents x1 , x2 relative to the standard basis) is
 x1   x1 + x2 
T = ,
 x2   x2 
the reader is invited to use Equations A.26 and check the easy calcula-
tions that give
272 Appendix A
 t11   1  tˆ tˆ12   0 
t12 1  1
[ T ]v =  =
11
[ T ]u =  = , 
 t21 t22   0 1   t21 ˆt22   − 1 2 1 

 
and then check Equations A.25.
If, in addition to the linear structure, V has an inner product, we can con-
sider the special case in which the two bases are orthonormal. If V is a real
vector space, we know from Section A.2.2 that the change-of-basis matrix
is orthogonal, that is CT C = I. This implies that two orthonormal bases lead
to the special form of similarity
[ T ]v = CT [ T ]u C, [ T ]u = C[ T ]v CT , (A.30)
which is expressed by saying that the matrices [ T ]u and [ T ]v are orthogo-

nally similar. If, on the other hand, V is a complex space, the change-of-
basis matrix is unitary and we have unitary similarity, that is
[ T ]v = C H [ T ]u C, [ T ]u = C[ T ]v C H. (A.31)
Particularly important in inner product spaces (and in applications) are

operators that can act on vectors in either one of the two ‘slots’ of the inner
product without changing the result, that is operators T : V → V such that
T x y = x T y (A.32)
for all x , y ∈V . In real vector spaces, these operators are called symmetric:
Hermitian or self-adjoint in complex spaces.
Then, if u1 ,, u n is an orthonormal basis of the real space V and we call
xi , yi ( i = 1,, n ), respectively, the components of x , y relative to this basis,
we have
Tx = ∑xt
i ,k
i ki uk , Ty = ∑y t
j, m
j mj u m , (A.33)
where the n2 scalars t ij are the elements of the matrix [ T ]u that represents
T relative to the u-basis. Using Equations A.33, the two sides of Equation
A.32 are expressed in terms of components as
Tx y = ∑x t
i ,k, j
i ki y j uk u j = ∑x t
i ,k, j
i ki y j δ kj = ∑x t
i, j
i ji yj
(A.34)
x Ty = ∑
i , j ,m
y j t m j xi u i u m = ∑i , j ,m
y j t m j xi δ i m = ∑
i,j
xi t i j y j ,
Appendix A 273
respectively. Since these two terms must be equal, the relations A.34
imply t j i = t ij, meaning that the matrix [ T ]u is symmetric. Moreover,
if v1 ,, v n is another orthonormal basis and C is the change-of-
basis matrix from the u-basis to the v-basis, Equation A.301 gives
( )
T
[ T ] Tv = CT [ T ]u C = CT [ T ] Tu C = CT [ T ]u C = [ T ]v , which means that [ T ]v
is also symmetric. A similar line of reasoning applies if V is a complex
space; in this case – and the reader is invited to fill in the details – we get
[ T ]H H
u = [ T ]u and then, from Equation A.311, [ T ] v = [ T ]v . The conclusion is
that, relative to orthonormal bases, symmetric operators are represented by
symmetric matrices and Hermitian operators are represented by Hermitian
matrices.
Remark A.6
i. Although not strictly necessary for our present purposes, we can go

back to Equation A.32 and mention the fact that in order to ‘prop-
erly’ define Hermitian (or self-adjoint) operators, one should first
introduce the concept of adjoint operator (of a linear operator T).
Given a linear operator T : V → V , its adjoint T + is the linear operator
T + : V → V such that T x y = x T + y for all x, y ∈V – or, equiva-
lently, x T y = T + x y for all x, y ∈V . In this regard, therefore, it is
evident that the name self-adjoint refers to those special operators
such that T + = T ;
ii. Note that some authors use the symbol T ∗ instead of T + . However,
since in this text, unless otherwise specified, we use the asterisk for
complex conjugation, the symbol T + is to be preferred;
iii. Given the definition above, at this point, it will not come as a surprise
to know that if, with respect to a given orthonormal basis, t ij  is the
matrix that represents the operator T , then the matrix that represents
T + (in the same basis) is its Hermitian-adjoint t ∗j i . In fact, if we call
sij = u i T + u j the elements of the representative matrix of T + relative
to the orthonormal u-basis, then, owing to the definition of adjoint
and the properties of the inner product, we have the chain of relations
∗
sij = u i T + u j = Tu i u j = u j Tu i = t ∗j i .
A.4 EIGENVALUES AND EIGENVECTORS: THE

STANDARD EIGENVALUE PROBLEM
In the light of the strict relation between linear operators and matrices
examined in the preceding section, one of the most important part of
the theory of operators and matrices is the so-called eigenvalue problem.
Our starting point here is the set Mn (F) of all n × n matrices on the field of
274 Appendix A
scalars F = C . For the most part, it will seldom make a substantial differ-
ence if the following material is interpreted in terms of real numbers instead
of complex numbers, but the main reason for the assumption is that C is
algebraically closed, while R is not. Consequently, it will be convenient to
think of real vectors and matrices as complex vectors and matrices with
‘restricted’ entries (i.e. with zero imaginary part).
Let A, x , λ be an n × n matrix, a column vector and a scalar, respectively.
One calls standard eigenvalue problem (or standard eigenproblem, SEP for
short) the equation
Ax = λ x ⇔ (A − λ I)x = 0, (A.35)
where the second expression is just the first rewritten in a different form.
A scalar λ and a nonzero vector x that satisfy Equation A.35 are called,
respectively, eigenvalue and eigenvector of A. Since an eigenvector is always
associated with a corresponding eigenvalue, λ and x together form a so-
called eigenpair. The set of all eigenvalues of A is called the spectrum of A
and is often denoted by the symbol σ (A), or, for some authors, Λ(A).
Three observations can be made immediately: Firstly, if x is an eigenvec-
tor associated with the eigenvalue λ , then any nonzero scalar multiple of
x is also an eigenvector. This means that eigenvectors are determined to
within a multiplicative constant – that is a scaling factor. Choosing this
scaling factor by some appropriate (or convenient) means, a process called
normalisation, fixes the length/norm of the eigenvectors and removes the
indeterminacy. Secondly, if x , y are two eigenvectors both associated with
λ , then any nonzero linear combination of x , y is an eigenvector associated
with λ . Third, recalling point 7 of Proposition A.1, A is singular if and only
if λ = 0 is one of its eigenvalues, that is if and only if 0 ∈σ (A). If, on the
other hand A, is nonsingular and λ , x is an eigenpair of A, then it is immedi-
ate to show that λ −1 , x is an eigenpair of A −1.
Remark A.7
The name standard used for the eigenproblem A.35 is used to distinguish
it from other types of eigenproblems. For example, the generalised eigen-
problem (GEP) – frequently encountered in the main text – involves two
matrices and has the (slightly more complicated) form Ax = λ Bx. However,
it is shown in the main text that a generalised problem can always be recast
in standard form.
From Equation A.352 , we see that λ is an eigenvalue of A, that is λ ∈σ (A),

if and only if A − λ I is a singular matrix, that is
det ( A − λ I ) = 0. (A.36)
Appendix A 275
From the Laplace expansion of the determinant, it follows that det ( A − λ I )

is a polynomial p(λ) of degree n in λ called the characteristic polynomial of
A, while A.36 – that is p(λ) = 0 – is the characteristic equation (of A). Then,
its n roots (see Remark A.8 below) are the eigenvalues of A, thus implying
that every n × n matrix has, in C, exactly n eigenvalues, counting algebraic
multiplicities.
Remark A.8
The fact that the roots of the characteristic equation are n follows from the fun-
damental theorem of algebra: In the field of complex numbers, a polynomial
of degree n with complex coefficients has exactly n zeroes, counting (algebraic)
multiplicities. The algebraic multiplicity of a root – which is to be distinguished
from the geometric multiplicity, to be defined later on – is the number of times
this root appears as a solution of the characteristic equation.
Again, we point out that the considerations above depend on the fact that
the complex field is algebraically closed; for matrices on R, little can be said
about the number of eigenvalues in that field.
Now, since det(A − λ I) = det(A − λ I)T = det(AT − λ I) and
( )
{det(A − λ I)} = det ( A − λ I )H = det A H − λ *I , it follows that (a) AT has theH
*
same eigenvalues of A, counting multiplicities, and (b) the eigenvalues of A

are the complex conjugates of those of A, counting multiplicities.
The fact that the matrices A, AT have the same eigenvalues does not imply
that the eigenvectors (of A and AT ) associated with the same eigenvalue are
the same; in fact, they are, in general, different. In this respect, however,
we can consider the eigenvalue problem for AT by assuming that y j , λ j is
an eigenpair for this problem, so that AT y j = λ j y j holds. Transposing both
sides gives yTj A = λ j yTj , thus showing that y j is a so-called left-eigenvector
of A associated with λ j. If, for k ≠ j , x k is an ordinary (right-)eigenvector of
A corresponding to the eigenvalue λk, then Ax k = λk x k . Premultiplying both
sides by yTj and subtracting this result from the one that we obtain by post-
multiplying yTj A = λ j yTj by x k leads to
yTj x k = 0, (A.37)
which is a kind of orthogonality – the term ‘biorthogonality’ is sometimes

used – between left- and right-eigenvectors corresponding to different eigen-
values (note, however, that since x k and/or y j may be complex, the l.h.s. of
A.37 differs from the ‘standard’ inner product y Hj x k of C n ). Also, taking
Equation A.37 into account, we can now premultiply Ax k = λk x k by yTj to
obtain the additional A-biorthogonality condition
yTj Ax k = 0. (A.38)
276 Appendix A
For eigenvectors corresponding to the same eigenvalue (i.e. k = j ), the r.h.s.

of Equations A.37 and A.38 is nonzero and we must first adopt some kind
of normalisation in order to fix its value. If, for example, we adopt the
frequently used normalisation such that yTj x j = 1, then yTj Ax j = λ j and the
biorthogonality conditions can be summarised in the form
yTj x k = δ j k , yTj A x k = λ j δ j k. (A.39)
Remark A.9
Instead of AT , we can consider the eigenvalue problem for A H . If A H y j = λ j y j,

then y Hj A = λ *j y Hj and y j is a left-eigenvector of A associated with the eigen-
value λ ∗j . The same line of reasoning as above now leads to
y Hj x k = δ j k , y Hj A x k = λ j δ j k , (A.40)
where the l.h.s. of A.401 is an inner product of two complex vectors in the
usual sense. Note, however, that Equations A.39 and A.40 do not neces-
sarily hold for all j, k = 1,, n because the matrix A may be defective (a
concept to be introduced in the next section). If it is nondefective, then
there are n linearly independent eigenvectors, the two equations hold for all
j, k = 1,, n and we can write the matrix Equations A.41 and A.42 below.
As far as eigenvalues are concerned, it turns out that the theory is simpler
when the n eigenvalues λ1 , λ2 ,, λn are all distinct and λ j ≠ λk for j ≠ k .
When this is the case, A is surely nondefective, each eigenvalue is associated
with a unique (to within a scaling factor) eigenvector and the (right-)eigen-
vectors form a linearly independent set (for this, see, for example, Laub
(2005) or Wilkinson (1996)). The same, clearly, applies to the eigenvectors
of AT , that is the left-eigenvectors of A. If now we arrange the n left-(right-)
eigenvectors of A to form the n × n matrix Y (X) whose jth column is given
by the components of y j ( x j ), Equations A.39 can be compactly written as
YT X = I, YT AX = X −1AX = diag ( λ1 ,, λn ), (A.41)
where in A.412 we took into account that the first equation implies YT = X −1 ,
If, as in Remark A.9, we consider A H and, again, let Y be the matrix of n
left-eigenvectors of A, we have
Y H X = I, Y H AX = X −1AX = diag ( λ1 ,, λn ), (A.42)
which is given as Theorem 9.15 in Chapter 9 of Laub (2005). Also, note that
Equations A.42 imply A = X diag ( λ1 , , λn ) X −1 = ∑ ni =1 λi x i y iH (clearly, for
Equations A.41, the y H
i of the last relation is replaced by y i ).
T
Appendix A 277
A.4.1 Similar matrices and diagonalisation

The rightmost expressions in the last two equations give us the possibility of
returning to the definition of similar matrices (mentioned in Section A.3) in
order to introduce the concept of diagonalisable matrices. Firstly, however,
a few more words on similarity are in order. Two n × n matrices A, B – we
recall – are similar if there exists a nonsingular n × n matrix S , the similarity
matrix, such that B = S−1AS (which implies A = SBS−1). The transformation
A → S−1AS = B is called a similarity transformation, and one often writes
B ≈ A to indicate that A and B are similar. On the set Mn of all square n × n
matrices, similarity is an equivalence relation – that is it is a reflexive (A ≈ A
), symmetric (if A ≈ B then B ≈ A) and transitive (if A ≈ B and B ≈ C, then
A ≈ C) relation. As it is known from basic mathematical theory of sets, this
implies that the similarity relation partitions Mn into disjoint equivalence
classes and that each class can be represented by any one of its members.
Since some matrices – for example, diagonal matrices for our present pur-
poses – have a particularly simple and convenient form, it is worthwhile
considering the equivalence class to which diagonal matrices belong. Before
doing so, however, we give two more results on similar matrices, namely:
Proposition A.2
If A, B ∈ Mn and A ≈ B, then they have the same characteristic polynomial.
Proposition A.3
A, B ∈ Mn and A ≈ B, then they have the same eigenvalues, counting

multiplicity.
Proposition A.3 – which is essentially a corollary of Proposition A.2 –

states that the eigenvalues of a matrix are invariant under a similarity
transformation. It should be noted, however, that having the same eigen-
values is a necessary but not sufficient condition for similarity and two
matrices can have the same eigenvalues without being similar.
Remark A.10
Note that if A, B ∈ Mn are similar via S and if λ , x is a nonzero eigenpair

of B, then λ , Sx is an eigenpair of A (the proof is immediate when one
observes that S−1AS x = λ x implies AS x = λ Sx ).
Returning to the equivalence class of diagonal matrices, a matrix A

belongs to this class, and we say that it is diagonalisable, if it is similar to a
diagonal matrix. Then, we have the following results.
278 Appendix A
Proposition A.4
If A ∈ Mn has n distinct eigenvalues, then it is diagonalisable, which means

that there exists S such that B = S−1AS is diagonal. In other words, we have
A ≈ B with B diagonal.
Some complications arise if A has one or more multiple eigenvalues, that

is eigenvalues with algebraic multiplicity (a. m. for short) strictly greater
than one (or, in different terminology, if one or more eigenvalues are degen-
erate). In this case, diagonalisation may not be possible. It is, however, pos-
sible under the conditions stated by the following proposition – which, it
should be noticed, includes Proposition A.4 as a special case.
Proposition A.5
A ∈ Mn is diagonalisable if and only if it has a set of n linearly independent

eigenvectors.
Proposition A.5, in turn, leads to the question: Under what conditions

does A have a set of n linearly independent eigenvectors? In order to answer,
one must introduce the concept of geometric multiplicity (g.m. for short) of
an eigenvalue λ j. This is defined as the maximum number of linearly inde-
pendent eigenvectors associated with λ j or, equivalently, as the dimension of
the subspace generated by all vectors x j satisfying Ax j = λ j x j (the so-called
eigenspace of λ j). An important result in this regard is that g. m. ≤ a. m. (see,
for example, Halmos (2017) or Horn and Johnson (1993)). So, if λ j is a
single root of the characteristic equation (i.e. a.m. = 1), then its associated
eigenvector x j is unique (to within a scaling factor, which is inessential
here) and g.m. = a.m. = 1. If, however, λ j is a multiple root, then we may
have either g. m. = a. m. or g. m. < a. m., and Proposition A.5 tells us that A is
diagonalisable if and only if g. m. = a. m. for all eigenvalues of A. This impor-
tant case is given a special name and one calls A nondefective. On the other
hand, when g. m. < a. m. for at least one eigenvalue, A is called defective. As
a simple example, the reader is invited to show that the matrix
 a 1 
A= 
 0 a 
is defective because the double eigenvalue λ = a is such that a. m. = 2 and

g. m. = 1. With a different wording, therefore, Proposition A.5 can be
equivalently stated by saying that A is diagonalisable if and only if it is
nondefective.
More generally, we say that two diagonalisable matrices A, B ∈ Mn are
simultaneously diagonalisable if there exists a single similarity matrix S ∈ Mn
such that S−1AS and S−1BS are both diagonal. Then, we have the result.
Appendix A 279
Proposition A.6
Two diagonalisable matrices A, B are simultaneously diagonalisable if and

only if they commute, that is if and only if AB = BA. The theorem can also
be extended to a set of diagonalisable matrices A1 , A 2 ,... (finite or not) that
commute in pairs (Horn and Johnson (1993)).
A.4.2 Hermitian and symmetric matrices

Symmetric matrices with real entries arise very frequently in practical cases
and, as a matter of fact, they have played a major role throughout this
book whenever eigenvalue problems have been considered. From the point
of view of the theory, however, symmetric matrices with complex entries
do not, in general, have many of the desirable properties of real symmetric
matrices and one is therefore led to consider Hermitian matrices (i.e. we
recall matrices such that A H = A). By so doing, moreover, one can formulate
the discussion in terms of Hermitian matrices and then note that real sym-
metric matrices are just Hermitian matrices with real entries.
A first important result is that the eigenvalues of a Hermitian matrix are
real. In fact, if Ax = λ x, then x H Ax = λ x H x, where x H x – being the norm
squared of x – is always real and positive for any nonzero (real or complex)
( )
H
vector x. Moreover, since x H Ax = x H Ax implies that the scalar x H Ax is
( )( )
real, λ = x H Ax x H x is also real. Although this clearly means that a real
symmetric matrix has real eigenvalues, it does not necessarily imply that
these eigenvalues correspond to real eigenvectors; it is so for real symmetric
matrices but not, in general, for (complex) Hermitian matrices. Taking a step
further, it is not difficult to show that the left- and right-eigenvectors of a
Hermitian matrix coincide and that – assuming the eigenvectors to be nor-
malised to unity, that is x Hj x j = 1 – we are led to the orthogonality conditions
x Hj x k = δ jk , x Hj Ax k = λ j δ jk.(A.43)
Thus, if a Hermitian matrix A has n distinct eigenvalues, it also has a set

of n linearly independent and mutually orthogonal eigenvectors. In this
case, we can arrange these eigenvectors as columns of a n × n matrix X and
rewrite Equations A.43 as
X H X = I, X H AX = L ≡ diag ( λ1 ,, λn ),(A.44)
which in turn imply:
X H = X −1, that is X is unitary (orthogonal if A is symmetric with real

1.
entries);
A is unitarily similar – orthogonally similar if it is real symmetric – to
2.
the diagonal matrix of eigenvalues L.
280 Appendix A
Similarity via a unitary (or orthogonal) matrix is clearly simpler than ‘ordi-
nary’ similarity because X H (or XT ) is much easier to evaluate than X −1.
Moreover, when compared to ordinary similarity, unitary similarity is an
equivalence relation that partitions Mn into finer equivalence classes because,
as we saw in Section A.2.2, it corresponds to a change of basis between
orthonormal bases.
Example A.5
By explicitly carrying out the calculations, the reader is invited to check
the following results. Let A be the real symmetric matrix
 1 −4 2 
 
A =  −4 3 5 .
 2 5 −1 
 
The roots of its characteristic equation

det(A − λ I) = − λ 3 + 3λ 2 + 46λ − 104 = 0 are the eigenvalues of A; in
increasing order, we have
λ1 = −6.5135, λ2 = 2.1761, λ3 = 7.3375
and the corresponding eigenvectors (normalised to unity) are
 0.4776   0.7847   0.3952 

     
x1 =  0.5576 , x 2 =  0.0768  , x 3 =  −0.8265 .
 −0.6789   0.6151   −0.4009 
     
Then, the similarity matrix is
 0.4776 0.7847 0.3952 

 
X =  x1 x2 x 3  =  0.5576 0.0768 −0.8265 
 
 −0.6789 0.6151 −0.4009 
 
and we have (a) XT X = XXT = I (i.e. X is orthogonal), and (b) XT AX = L ,

where
 −6.5135 0 0 
 
L = diag ( λ1 , λ2 , λ3 ) =  0 2.1761 0 .
 0 0 7.3375 
 
At this point, we can consider the complication of multiple eigenvalues,

which, we recall, may mean that A is defective. One of the most impor-
tant properties of Hermitian matrices, however, is that they are always
Appendix A 281
nondefective. This, in turn, implies that any Hermitian matrix – whether

with multiple eigenvalues or not – can always be unitarily diagonalised
or, in other words, that there always exists a unitary matrix X such that
X H AX = L.
In more mathematical form, we have the following proposition (Horn
and Johnson 1993).
Proposition A.7
Let A =  aij  ∈ Mn have eigenvalues λ1 , λ2 ,, λn , not necessarily distinct.

Then, the following statements are equivalent:
1. A is normal;
2. A is unitarily diagonalisable;
2 2
3. ∑ ni , j =1 aij = ∑ ni =1 λi ;
4. There is an orthonormal set of n eigenvectors of A.
The equivalence of points 1 and 2 in Proposition A.7 is often called the

spectral theorem for normal matrices. In relation to the foregoing discus-
sion, we recall from Section A.1 that a Hermitian matrix is just a special
case of normal matrix.
Remark A.11
Points 1 and 2 of Proposition A.7 tell us that a matrix A is normal if and

only if X H AX = L (where L = diag ( λ1 ,, λn )) but, in general, this diago-
nal matrix is complex. In the special case in which A is Hermitian, how-
ever, L is real because, as shown above, the eigenvalues of a Hermitian
matrix are real. Also, it is worth mentioning that point 3 of the proposition
is a consequence of the fact that for any two unitarily similar matrices
2 2
A =  aij  , B = bij  , we have ∑ ni , j =1 aij ∑ ni , j =1 bij .
So, summarising the results of the preceding discussion, we can say that
a complex Hermitian (or real symmetric) matrix:
1. Has real eigenvalues;

2. Is always nondefective (meaning that, multiple eigenvalues or not,
there always exists a set of n linearly independent eigenvectors, which,
in addition, can always be chosen to be mutually orthogonal);
3. Is unitarily (orthogonally) similar to the diagonal matrix of eigen-
values diag ( λ1 ,, λn ). Moreover, the unitary (orthogonal) similarity
matrix is the matrix X (often called the modal matrix) whose jth col-
umn is the jth eigenvector.
282 Appendix A
A.4.3 Hermitian and quadratic forms

We close this section with a few more definitions and results on Hermitian
matrices. If A ∈ Mn (C) is Hermitian, the expression x H Ax is called a
Hermitian form (of, or generated by, A). Then, recalling from the beginning
of this section that x H Ax is a real scalar, we call A positive-definite if its
Hermitian form is always strictly positive, that is if x H Ax > 0 for all nonzero
vectors x ∈C n. If the strict inequality is weakened to x H Ax ≥ 0, then A is said
to be positive-semidefinite. By simply reversing the inequalities, we obtain
the definitions of negative-definite and negative-semidefinite matrices.
The real counterparts of Hermitian forms are called quadratic forms and
read x T Ax, where A ∈ Mn (R) is a symmetric matrix. In this case, the appro-
priate definitions of positive-definite and positive-semidefinite matrix are,
respectively, x T Ax > 0 and x T Ax ≥ 0 for all nonzero vectors x ∈ Rn.
Given these definitions, an important result for our purposes is as follows.
Proposition A.8
A Hermitian (symmetric) matrix A ∈ Mn is positive-semidefinite if and only

if all of its eigenvalues are non-negative. It is positive-definite if and only if
all of its eigenvalues are positive.
In particular, it should be noted that the fact that a positive-definite

matrix has positive eigenvalues implies that zero is not one of its eigenval-
ues. By virtue of point 7 of Proposition A.1, therefore, a positive-definite
matrix is nonsingular.
Finally, it is left to the reader to show that the trace and the determinant
of a positive-definite (semidefinite) matrix are both positive (non-negative).
A.5 THE EXPONENTIAL OF A MATRIX
If A ∈ Mn, the exponential e A and, for a finite scalar t , the exponential e At

are also in Mn and are defined by the absolutely convergent power series
∞ ∞
A
e = ∑
k= 0
Ak
k!
= I+A+
A2
2!
+ , e At
= ∑
k= 0
(At)k
k!
= I + tA +
t 2A 2
2!
+ …,
(A.45)
respectively. Some important properties of matrix exponentials, which we

give without proof, are as follows:
a. For any two scalars, e A(t + s) = e At e As . In particular, by choosing

t = 1, s = −1, we get e Ae − A = e0 = I, thus implying that e A is always
invertible and that its inverse is e −A regardless of the matrix A;
Appendix A 283
b. If AB = BA, then e A+ B = e A e B and e(A+ B) t = e At e Bt (if A, B do not com-

mute, these properties are not, in general, true);
c. ( )
d e At dt = Ae At = e At A ;
d. If A = diag ( λ1 ,, λn ), then (
e A = diag e λ1 ,, e λn and)
At
(
e = diag e ,, e ; λ1t λn t
)
e. Denoting by L, L−1, respectively, the Laplace transform and the
inverse Laplace transform (see Appendix B, Section B4), then
L e At  = (s I − A)−1 and L−1 ( s I − A )  = e A t .
−1
 
Next, let us now consider two similar matrices A, B, so that we know that
there exists an invertible matrix such that B = S−1AS and A = SBS−1. Then,
passing to exponentials, we get
A
e =e SBS−1
= I + SBS −1
+
(SBS )
−1 2
+  = I + SBS−1 +
SB2S−1
+
2! 2!

 B2 
= S I+ B+ +  S−1 = S e B S−1 , (A.46)
 2! 
where in the third equality we observed that

(SBS −1 2
) 2 −1
= SB S , SBS
−1
(
−1 3
)
3 −1
= SB S , and so on. So, in particular if A is diag-
onalisable and L = S AS is diagonal, Equation A.46 shows that
e A = S e L S−1 , e A t = S e L t S−1 , (A.47)
where the second relation is an extension of the first and t , as above, is a

scalar.
A.6 SCHUR’S TRIANGULARISATION AND THE

SINGULAR-VALUE DECOMPOSITION
In the preceding section, we saw that every normal matrix A ∈ Mn is uni-

tarily diagonalisable and we have X H AX = L. This is the same as saying
that A can be written in the form A = X L X H , where X is unitary (i.e.
X H = X −1) and L is the diagonal matrix of eigenvalues. This is called the
spectral decomposition of A and is a special case of the so-called Schur’s
decomposition, or triangularisation theorem (Horn and Johnson (1993) or
Junkins and Kim (1993)): Given any square matrix A ∈ Mn (C), there is a
unitary matrix U ∈ Mn such that
U H AU = T, (A.48)
284 Appendix A
where T ∈ Mn is an upper-triangular matrix whose diagonal entries are the

eigenvalues of . If, moreover, A is real and all its eigenvalues are real, then
U may be chosen to be real and orthogonal.
Remark A.12
i. Two points worthy of mention in Schur’s theorem are the following.

Firstly, the diagonal elements t ii of T are, as stated above, the (not nec-
essarily distinct) eigenvalues λ1 ,, λn of A but they can appear in any
order. This means that neither U nor T is unique. Secondly, the term
‘upper-triangular’ may be replaced by ‘lower-triangular’ but, clearly,
the matrix U will differ in the two cases;
ii. If A ∈ Mn (R), one should not expect to reduce a real matrix to upper-
triangular form by a real similarity (let alone a real orthogonal
similarity) because the diagonal entries would then be the eigenval-
ues, some of which could be nonreal (in complex conjugate pairs).
However, the Murnaghan–Wintner theorem states that there exists
an orthogonal matrix U such that UT AU is quasi-upper-triangular,
where this term means a block upper-triangular matrix with 1 × 1
diagonal blocks corresponding to the real eigenvalues and 2 × 2
diagonal blocks corresponding to the complex conjugate pairs of
eigenvalues.
Another important result in the direction of unitary diagonalisation – or,

better, ‘quasi-diagonalisation’ in the general case – applies to any rectan-
gular matrix and is given by the singular-value decomposition (SVD): If
A is an m × n matrix of rank k ( k ≤ min(m, n)), then there exist two unitary
matrices U ∈ Mm and V ∈ Mn such that
A = U SV H ,(A.49)
where the elements of S =  sij  ∈ Mm × n are zero for i ≠ j and, calling p the
minimum between m and n, s11 ≥ s22 ≥  ≥ skk > sk+1,k+1 =  = spp = 0. In
other words, the only nonzero elements of S are real positive numbers on
the main diagonal and their number equals the rank of A. These elements
are denoted by σ 1 ,, σ k and called the singular values of A. Moreover, if A
is real, then both U and V may be taken to be real, U, V are orthogonal and
the upper H is replaced by an upper T .
So, for example, if A ∈ M3 × 4 and rank A = 2, we get a matrix of the form
 σ1 0 0 0 
 
S= 0 σ2 0 0 ,
 0 0 0 0 
 
Appendix A 285
with only two nonzero singular values σ 1 = s11 , σ 2 = s22 and the other diago-
nal entry s33 equal to zero. If, on the other hand, rank A = 3, then we have
three singular values and the third is σ 3 = s33 > 0 .
The reason why the singular values are real and non-negative is because
the scalars σ i2 are the eigenvalues of the Hermitian and positive-semidefinite
matrices AA H , A H A. In fact, using Equation A.49 and its Hermitian conju-
gate A H = V SH U H together with the relations U H U = I m , V H V = I n , we get
AA H = U SSH U H , A H A = VSH SV H (A.50)
so that post-multiplying the first of A.50 by U and the second by V leads to
AA H U = U SSH , A H AV = VSH S , (A.51)
where both products SSH and SH S are square matrices whose only nonzero
2 2
elements σ 1 ,, σ k lie on the main diagonal. This, in turn, implies that we
can call u1 , u 2 ,, u m the first, second, …, mth column of U and v1 , v 2 ,, v n
the first, second, …, nth column of V and rewrite Equation A.51 as
2
AA H u i = σ i u i ( i = 1,, m) ; 2
A H Av i = σ i v i ( i = 1,, n ),
(A.52)
which show that the u i are the eigenvectors of AA H corresponding to the

2
eigenvalues σ i , while the v i are the eigenvectors of A H A corresponding to
the same eigenvalues. Since both these matrices are Hermitian – and, we
2
recall, Hermitian matrices have real eigenvalues – we can replace σ i by σ i2
and conclude that the singular values σ i of A are the positive square roots
of the eigenvalues of AA H (or A H A).
In regard to the eigenvectors u i of AA H and the eigenvectors v i of A H A,
they are called, respectively, the left and right singular vectors of A because
they form the matrices U and V appearing to the left and right of the
singular-values matrix S – or, rewriting Equation A.49 as S = U H AV, to the
left and right of the original matrix A.
As a final point, we note that if A ∈ Mn is normal (in particular, Hermitian
or real symmetric) with eigenvalues λ1 ,, λn, its singular values are given by
σ i = λi for i = 1,, n.
A.7 MATRIX NORMS
In Section A.2.2, we observed that a vector norm is a generalisation of the

familiar notion of ‘length’ or ‘size’ of a vector and that its definition is given
in abstract mathematical terms as a mapping from the vector space V to the
286 Appendix A
set of non-negative real numbers satisfying the axioms N1–N3. The same
idea can be extended to matrices by defining a matrix norm as a mapping
• : Mn → R, which, for all A, B ∈ Mn (F) and all a ∈ F satisfies the axioms
MN1. A ≥ 0 and A = 0 if and only if A = 0

MN2. aA = a A
MN3. A + B ≤ A + B (triangle inequality)
MN4. AB ≤ A B (submultiplicative),
where MN1–MN3 are the same as the defining axioms of a vector norm. A
mapping that satisfies these three axioms but not necessarily the fourth is
sometimes called a generalised matrix norm.
Two examples of matrix norms are the maximum column sum and the
maximum row sum norms, denoted, respectively, by the symbols • 1 , • ∞
and defined as
n n
A 1 = max
1≤ j ≤ n ∑ i =1
aij , A ∞
= max
1≤ i ≤ n ∑a
j =1
ij , (A.53)
where, clearly, A 1 = A H . Other two examples are the matrix Euclidean

∞
(or Frobenius) norm • E
and the spectral norm • 2 , defined, respectively, as
∑a ( )
2
A E
= ij = tr A H A , A 2
= σ max (A), (A.54)
i , j =1
where σ max (A) – that is σ 1 if, as in the preceding section, we arrange the
singular values in non-increasing order – is the maximum singular value
of A or, equivalently, the positive square root of the maximum eigen-
value of A H A. Both norms A.54 have the noteworthy property of being
unitarily invariant, where by this term we mean that UAV E = A E and
UAV 2 = A 2 for any two unitary matrices U, V.
Also, associated with any vector norm, it is possible to define a matrix
norm as
Ax
A = max = max Ax , (A.55)
x ≠0 x x =1
which is called the matrix norm subordinate to (or induced by) the vector
norm. This norm is such that I = 1 and satisfies the inequality Ax ≤ A x
for all x. More generally, a matrix and a vector norm satisfying an inequal-
ity of this form are said to be compatible (which clearly implies that a vector
norm and its subordinate matrix norm are always compatible). Note also
Appendix A 287
that I = 1 is a necessary but not sufficient condition for a matrix norm to

be induced by some vector norm.
The matrix norms introduced above are some of the most frequently
used, but obviously they are not the only ones. And in fact, the following
proposition shows that there is also the possibility of ‘tailoring’ a matrix
norm for specific purposes.
Proposition A.9
Let • be a matrix norm on Mn and S ∈ Mn a nonsingular matrix. Then, for

all A ∈ Mn, A S ≡ S−1AS is a matrix norm.
Given a matrix A ∈ Mn, we define its spectral radius ρ(A) as
ρ(A) = max λ , (A.56)

λ ∈ σ (A )
where we recall that λ ∈σ (A) means that λ is an eigenvalue of A and we

note that ρ(A) = A 2 if A is normal. Also, we define the condition number
κ (A) of A as
κ (A) = A A −1 (A.57)
by assuming κ (A) = ∞ if A is singular. Clearly, the condition number depends

on the norm used to define it, and we will have, for example, κ 1(A),κ ∞ (A), and
so on. In any case, however, it can be shown that κ (A) ≥ 1 for every matrix
norm. In particular, the spectral condition number is given by the ratio
σ max (A)
κ 2 (A) = , (A.58)
σ min (A)
( )
where σ min (A) = 1 σ max A −1 is the minimum singular value of A.
With these definitions, the following two propositions may give a general
idea of the results that can be obtained by making use of matrix norms.
Proposition A.10
For any matrix norm, we have

1k
ρ(A) ≤ A , ρ(A) = lim A k . (A.59)
k→∞
Moreover, for any given ε > 0, there exists a matrix norm such that
ρ(A) ≤ A ≤ ρ(A) + ε .
288 Appendix A
Proposition A.11. (Bauer–Fike theorem)
Let A ∈ Mn be diagonalisable with S−1AS = L and L = diag ( λi ,, λn ). If

E ∈ Mn is a perturbative matrix and µ is an eigenvalue of A + E, then there is
at least one eigenvalue λi of A such that
µ − λi ≤ κ (S) E , (A.60)
where • is any one of the norms • 1 , • 2 , • ∞

and κ (•) is its corresponding
condition number.
Proposition A.11 shows that the overall sensitivity of the eigenvalues of A

depends on the size of the condition number of the matrix S of eigenvectors.
If κ (S) is small (near 1), one speaks of well-conditioned eigenvalue problem
because small perturbations of A will result in eigenvalues variations of the
same order as the perturbation. If, however, κ (S) is large, the problem is ill-
conditioned because we may expect relatively large variations of the eigen-
values as a consequence of even a small perturbation of the input data. For
the interested reader, the proof of the last two propositions can be found in
Horn and Johnson (1993), Junkins and Kim (1993) (Proposition A.11) and
Wilkinson (1996).
Appendix B
Fourier series, Fourier and
Laplace transforms
B.1 FOURIER SERIES
It is well known that the basic trigonometric functions sine and cosine are
periodic with period 2π and that, more generally, a function f (t) is called
periodic of (finite) period T if it repeats itself every T seconds, so that
f (t) = f (t + T ) for every t . If now we consider the fundamental (angular or
circular) frequency ω 1 = 2π T and its harmonics ω n = nω 1 ( n = 1, 2,), one
of the great achievements of J.B Fourier (1768–1830) is to have shown that
almost any reasonably well-behaved periodic function of period T can be
expressed as the sum of a trigonometric series. In mathematical terms, this
means that we can write
∞ ∞

A
f (t) = 0 +
2 ∑ (A cosω t +B sinω t),
n=1
n n n n f (t) = ∑ C exp(iω t), (B.1)
n=−∞
n n
where the second expression is the complex form of the Fourier series and
is obtained from the first by using Euler’s formula e ± i x = cos x ± i sin x. The
constants An , Bn or Cn are called Fourier coefficients, and it is not difficult
to show that we have
2Cn = An − iBn , 2C0 = A0 , 2C− n = An + iBn, (B.2)
thus implying that the C-coefficients are generally complex. However, if f (t)
is a real function, then Cn = C−∗n .
Remark B.1
i. In Equation B.11, the ‘static term’ A0 2 has been introduced for future
convenience in order to include the case in which f (t) oscillates about
some nonzero value;
289
290 Appendix B
ii. In the expressions above, f (t) is a time-varying quantity because we are

concerned with vibrations. However, this is not necessary, and in some
applications, time t could be replaced by a spatial variable, say z, so that
the frequency ω would then be replaced by a ‘spatial frequency’ (with
units of rad/m), meaning that f (z) has a value dependent on position;
iii. Using the well-known trigonometric relations, one can introduce the
quantities Dn , ϕ n such that An = Dn sin ϕ n , Bn = Dn cos ϕ n and write the
Fourier series in the form
∞
f (t) =
A0
2
+
n=1
∑
Dn sin (ω nt + ϕ n ). (B.3)
At this point, without being too much concerned about mathematical

rigour, we can start from the complex form of the series and determine the
C-coefficients by (a) multiplying both sides by exp ( − iω mt ), (b) integrating
in dt over a period (say, from 0 to T , but any interval from t1 to t2 will do as
long as t2 − t1 = T ), (c) assuming that the series and the integral signs can be
interchanged and (d) taking into account the (easy-to-check) relation
T

∫e
0
i nω1t
e − i mω1t dt = T δ n m,(B.4)
where δ n m is the Kronecker delta (δ n m = 1 for n = m and δ n m = 0 for n ≠ m).

The final result of steps (a)–(d) is the formal expression of the Fourier coef-
ficients as
T T T
1 1 1
Cn =
T ∫
0
f (t)e − iω nt dt , C0 =
T ∫
0
f (t) dt , C− n =
T ∫ f (t)e
0
iω nt
dt ,
(B.5)
which, in turn, can be substituted into Equations B.2 to give the original
‘trigonometric’ coefficients as
T T
2 2
An =
T ∫
0
f (t) cos (ω n t ) dt , Bn =
T ∫ f (t)sin (ω t ) dt (B.6)
0
n
T
for n = 0,1,2,, where, in particular – since A0 = 2 T −1
∫ 0
f (t) dt – the term
A0 2 of Equation B.11 is the average value of f (t). Also, note that Equations
B.5 and B.6 show that f (t) must be integrable on its periodicity interval for
the coefficients to exist.
Appendix B 291
If now we ask about the relation between the mean-square value of f (t)
and its Fourier coefficients, we must first recall that the mean-square value
is defined as
T T
1 1
∫ ∫ f (t)
2 2 2
f (t) ≡ f (t) dt = dt , (B.7)
T T
0 0
when the integral on the r.h.s. exists. Then, we can substitute the series
expansion of Equation B.12 into the r.h.s. of Equation B.7, exchange the
integral and series signs, and use Equation B.4 again to get
∫ (∑ C e )(∑ C e ) dt
T T
1 1
f 2 (t) =
T ∫
0
f 2 (t) dt =
T
0
n
n
i ωn t
m
∗ −i ωm t
m
T ∞
CnCm∗
= ∑∑ n m T ∫ e i ω n t e − i ω m t dt = ∑∑ n m
CnCm∗ δ nm = ∑C
n=−∞
n
2
,
0
(B.8)
where the absence of cross-terms of the form CnCm∗ (with n ≠ m) in the final
result is worthy of notice because it means that each Fourier component
Cn makes its own contribution (to the mean-square value of f (t)), indepen-
dently of all the other components. When proved on a rigorous mathemati-
cal basis, this result is known as Parseval’s relation, and the reader is invited
to check that its ‘trigonometric version’ reads
∞
f 2 (t) =
A02 1
4 2
+ ∑(A n=1
2
n )
+ Bn2 . (B.9)
On physical grounds, it should be noted that Equations B.8 and B.9 are
particularly important because the squared value of a function (or ‘signal’,
as it is sometimes called in applications-oriented literature) is related to
key physical quantities such as energy or power. In this respect, in fact, an
important mathematical result is that for every square-integrable function
T
∫
2
on [0, T ] – that is functions such that f (t) dt is finite – Parseval’s rela-
0
tion holds and its Fourier series converges to f (t) in the mean-square sense,
which means that we have
T
∫ f (t) − S
2
lim N dt = 0 , (B.10)
N →∞
0
292 Appendix B
∑
N
where SN = Cn e i ω nt is the sequence of partial sums of the series.
n=− N
However, since mean-square convergence is a ‘global’ (on the interval [0, T ])
type of convergence that does not imply pointwise convergence – which, in
general, is the type of convergence of most interest in applications – a result
in this direction is as follows (Boas (1983)).
PROPOSITION B.1 (DIRICHLET)
If f (t) is periodic of period T and in its interval of periodicity is single-

valued, has a finite number of maximum and minimum values, and a finite
T
number of discontinuities, and if
∫0
f (t) dt is finite, then its Fourier series
converges to f (t) at all points where f (t) is continuous; at jumps, the series
converges to the midpoint of the jump (this includes jumps that may occur
at the endpoints 0, T ).
Remark B.2
i. More precisely, if t0 ∈[0, T ] is a point where f (t) has a jump with left-
( ) ( )
hand and right-hand limits f t0− , f t0+ , respectively, then the series
( ) ( )
converges to the value  f t0+ + f t0−  2;
ii. Proposition B.1 is useful because in applications, one often encounters
periodic functions that are integrable on any finite interval and that
have a finite number of jumps and/or corners in that interval. In these
cases, the theorem tells us that we do not need to test the convergence
of the Fourier series, because once we have calculated its coefficients,
the series will converge as stated;
iii. An important property of Fourier coefficients of an integrable func-
tion is that they tend to zero as n → ∞ (Riemann–Lebesgue lemma). In
practice, this means that in approximate computations, we can trun-
cate the series and calculate only a limited number of coefficients.
Example B.1
Since on its interval of definition (−π, π), the rectangular pulse
 −1 −π < t < 0
f (t) =  (B.11)
 +1 0<t <π
satisfies the Dirichlet conditions of proposition B.1, it has a Fourier

series that converges to it. In order to determine the coefficients, it is
Appendix B 293
more convenient in this case to use the trigonometric version by observ-

ing that f (t) is an odd function of t (i.e. f (−t) = − f (t)), and therefore, all
the cosine coefficients An must be zero. Then, for the sine coefficients,
some easy calculations show that Equation B.62 gives Bn = 0 for n even
and Bn = 4 nπ for n odd. Consequently, the Fourier expansion is
∞
sin ( 2n + 1) t 
f (t) =
4
π
1
3
1
5
 4
 sin t + sin3t + sin 5t +  =
π ∑
n =0
2n + 1
, (B.12)
 ( ) ( )
which, as expected, converges to  f 0+ + f 0−  2 = 0 at the point of

discontinuity t = 0. The same occurs also at the endpoints t = ±π if we
consider the function f (t) to be extended by periodicity over the entire
real line R. In this case, in fact, the extension leads to the left and
( ) ( )
right limits f −π − = 1, f −π + = −1 at t = −π and f π − = 1, f π + = −1 ( ) ( )
at t = π.
Example B.2
On the interval [ −π, π ], the function f (t) = t is continuous, satisfies
f (−π) = f (π) at its endpoints and is even (i.e. f (−t) = f (t)). Since this last
property suggests that also in this case, the trigonometric form is more
convenient for the calculation of the Fourier coefficients (because the
sine coefficients Bn are all zero), the reader is invited to determine the
cosine coefficients and show that (a) A0 = π and (b) only the An for n odd
are nonzero, with A1 = − 4 π , A3 = − 4 9π , A5 = − 4 25π , etc. The series
expansion, therefore, can be equivalently written in the two forms
( −1)n − 1 cos nt = π − 4 cos [(2n − 1) t ]

∞ ∞
f (t) =
π 2
+
2 π ∑
n =1
n2 2 π ∑
n =1
(2n − 1)2
, (B.13)
which, it should be noticed, has better (i.e. faster) convergence prop-

erties than the series (B.12) because its coefficients go to zero as 1 n2,
while the coefficients in B.12 go to zero as 1 n. The reason is that the
function t – being continuous with a piecewise continuous first
derivative on [ −π, π ] – is ‘better behaved’ than the function B.11, which
is only piecewise continuous.
Remark B.3
i. As a matter of fact, it turns out that the series B.13 converges uni-
formly because of the regularity properties of t . In this respect, more-
over, it can be shown (see, e.g., Howell (2001) or Sagan (1989)) that
these same properties allow the termwise differentiation of the series
B.13. By so doing, in fact, we get the coefficients of the series B.12,
294 Appendix B
and the result makes sense because the function B.11 is the derivative
of t . On the other hand, if we differentiate term-by-term the coef-
ficients of the series B.12 – which does not converge uniformly – we
get nonsense (and surely not the derivative of the pulse B.11). And this
is because piecewise continuity of f (t) is not sufficient for termwise
differentiation;
ii. Having pointed out in the previous remark that term-by-term dif-
ferentiation of a Fourier series requires some care, it turns out that
termwise integration is less problematic and, in general, the piecewise
continuity of the function suffices. An example is precisely the pulse
B.11, whose integral is t . By integrating term-by-term its Fourier
coefficients, in fact, the result is that we obtain the coefficients of the
series B.13;
iii. Another aspect worthy of mention is the so-called Gibbs phenom-
enon, which consists in the fact that near a jump discontinuity the
Fourier series overshoots (or undershoots) the function by approxi-
mately 9% of the jump. So, for instance, if we drew a graph of the
partial sum SN of the series on the r.h.s. of B.12 for a few different
values of N (and the reader is invited to do so), we would observe the
presence of ‘ripples’ (or ‘wiggles’) in the vicinity of the discontinuity
points of the function f (t). These ‘ripples’ are due to the non-uniform
convergence of the Fourier series and persists no matter how many
terms of the series are employed (even though, for increasing N, they
get confined to a steadily narrower region near the discontinuity).
B.2 FOURIER TRANSFORM
Intuitively, if we consider a non-periodic function f (t) as a periodic function

with an infinite period, we may expect to find a ‘Fourier representation’ of
f (t) by starting from Equation B.1 and passing to the limit T → ∞. In order
to do so, let us define ∆ω = ω n+1 − ω n; then ∆ω = 2π T and the coefficient Cn
of Equation B.5 can be rewritten as
T 2
∆ω
Cn =
2π ∫ f (s) e
−T 2
− iω n s
ds ,
where s is a dummy variable of integration. Substituting this expression into

B.12 leads to
∞  T 2  ∞
f (t) =
1
∑ ∫ 
2π n=−∞ 
 −T 2
∑
f (s) e iω n (t − s) ds ∆ω =

1
2π n=−∞
g (ω n )∆ω , (B.14)
Appendix B 295
where in the last expression we called g (ω n ) the integral within parenthe-

ses. Now, if ∆ω → 0 as T → ∞, then
∞ T 2 ∞
g (ω n ) g (ω n )
∞
∑
n=−∞
2π
∆ω →
∫ 2π
dω ,
∫ f (s) e i ω n (t − s)
ds →
∫ f (s) e i ω (t − s)
ds
−∞ −T 2 −∞
and Equation B.14 becomes (with all integrals from −∞ to ∞)
1 1
f (t) =
2π ∫ g(ω ) dω = 2π ∫∫ f (s) e i ω (t − s)
ds dω
(B.15)
 1 
=
∫ 
2π ∫ f (s) e −i ω s
ds e i ω t dω =
 ∫ F(ω ) e iω t
dω ,
where in the last equality we have defined

∞
1
F(ω ) =
2π ∫ f (t) e
−∞
−iω t
dt (B.16)
and we could return to the original variable t because s was just a dummy
variable of integration. But then, since Equation B.15 gives
∞
f (t) =
∫ F(ω ) e
−∞
iω t
dω , (B.17)
these last two relations show that when the integrals exist, the functions
f (t) and F(ω ) form a pair. This is called a Fourier transform pair and usu-
ally one calls F(ω ) the (forward) Fourier transform of f (t), while f (t) is the
inverse Fourier transform of F(ω ). In accordance with these names, one can
formally introduce the forward and inverse Fourier transform ‘operators’
F, F-1 and conveniently rewrite Equations B.16 and B.17 in the more concise
notation
F(ω ) = F [ f (t)] , f (t) = F−1 [ F(ω )]. (B.18)
Remark B.4
i. Although our notation (with the multiplying factor 1 2π in the for-

ward transform) is quite common, other authors may adopt different
conventions and it is possible to find the factor 1 2π in the inverse
transform, or a multiplying factor 1 2π in both integrals. It is also
296 Appendix B
quite common to find the transform pair expressed in terms of the

ordinary frequency ν = ω 2π as
∞ ∞
F(ν ) =

∫
−∞
f (t) e − i 2πνt dt , f (t) =
∫ F(ν) e
−∞
i 2 πνt
dν . (B.19)
From a strictly mathematical point of view, these differences of nota-

tion are not really important, but care must be exercised in practical
cases, especially when using tables of Fourier transforms;
ii. Equations B.18 suggest that the operators F, F−1 are the inverses of
each other, so that, for instance, one can use F to transform f (t) into
F(ω ) and then recover f (t) by means of F-1. As a matter of fact, this
is generally the case for ‘reasonably nice’ functions – such as when
both f (t) and F(ω ) are piecewise continuous and absolutely integrable
on the real line. However, since there is definitely more than this and
since our rather ‘free manipulations’ certainly do not provide a mathe-
matical proof, the interested reader should take Equations B.18 with a
grain of salt and refer, for instance, to Appel (2007) or Howell (2001)
for a rigorous account of these aspects.
Example B.3
Given the rectangular pulse of unit area (sometimes called the boxcar
function)
 1 2 −1 ≤ t ≤ 1
f (t) =  , (B.20)
 0 otherwise
a quick calculation gives the Fourier transform
1
sin ω
F(ω ) =
1
4π ∫e −iω t
dt =
1
4i πω
(
ei ω − e− i ω =)2πω
, (B.21)
−1
where F(ω ) is real because f (t) is even. If, on the other hand, we con-
sider the time-shifted version of the boxcar function defined as f (t) = 1 2
for 0 ≤ t ≤ 2 (and zero otherwise), we get the complex transform
F(ω ) = ( 2πω ) e − i ω sin ω , which has the same magnitude of B.21 but a
−1
different phase. Also, if we want the transform Fˆ (ν ) of the boxcar func-

tion in terms of ordinary frequency, we can observe that Fˆ (ν )dν = F(ω )dω
gives Fˆ (ν ) = 2π F(ω ); consequently, Fˆ (ν ) = ( 2πν ) sin 2πν . Needless to
−1
say, this is the same result that we obtain by using the transform B.19.
If now we continue with our heuristic manipulations, we can obtain a

non-periodic version of Parseval’s relation (Equation B.8 in the periodic
case). With the understanding that all integrals extend from −∞ to ∞ and
Appendix B 297
that an asterisk denotes complex conjugation, let us consider the double

integral I = (2π)−1
∫∫ F (ω )f (t) dω dt. By assuming that the order of inte-
∗
gration is immaterial (we recall that a result known as Fubini’s theorem

establishes the conditions under which this is the case), we can write I in
two different forms. In the first, we use Equation B.16 to obtain
 1 
∫ F (ω )  2π ∫ f (t)e ∫ F (ω )F(ω ) dω = ∫ F(ω )
∗ −iω t ∗ 2
I= dt  dω = dω

(B.22a)
while in the second form, we use the complex conjugate of Equation

B.17 to get
1  1 1
dω  dt =
∫ f (t)  ∫ F (ω )e ∫ f (t)f (t) dt = 2π ∫ f (t)
∗ −iω t ∗ 2
I= dt
2π  2π
(B.22b)
Putting these two last equations together leads to Parseval’s relation in

the form
∫ f (t) ∫ F(ω )
2 2
dt = 2π dω . (B.23)
B.2.1 Main properties of Fourier transforms

Let f (t), g(t) be two Fourier-transformable functions with transforms
F(ω ), G(ω ), respectively, and let a, b be two constants. Then, the function
af (t) + bg(t) is itself Fourier-transformable and linearity is immediate, that is
F [ af (t) + bg(t)] = aF(ω ) + bG(ω ). (B.24)
Four other properties (which the reader is invited to check) are
F  f ∗ (t)  = F ∗ (−ω ), F [ f (t − a)] = e − i ω a F(ω )

(B.25)
F [ f (at)] = a F (ω a ) ,
−1
F  f (t) e i at  = F(ω − a),
where a is assumed to be nonzero in the last relation.

Turning to the derivative f ′(t) = df dt , assuming it to be Fourier-
transformable implies that f (t) → 0 as t → ±∞. Then, we can use integration
by parts to get
∞ ∞ ∞
1 f e− i ω t iω
F [ f ′(t)] =
2π ∫
−∞
f ′(t) e − i ω t dt =
2π −∞
+
2π ∫ f (t) e
−∞
−i ω t
dt = i ω F(ω ), (B.26a)
298 Appendix B
which, denoting by f (k) (t) the kth-order derivative of f (t) and assuming it to
be Fourier-transformable in its own right, is just a special case of the more
general relation
F  f (k) (t)  = (i ω )k F(ω ). (B.26b)
t
Next, if we consider the function I (t) =
∫ t0
f (s) ds – that is the ‘anti-derivative’
of f (t) – and assume it to be Fourier-transformable, Equation B.26a gives
(iω )−1 F[ I '(t)] = F[ I (t)]. But since I '(t) = f (t), this means (iω )−1 F [ f (t)] = F [ I (t)]
and therefore
1
F [ I (t)] = F(ω ). (B.27)
iω
On the other hand, for the derivatives F (k) (ω ) = d kF dω of F(ω ), the result is
that if the function t k f (t) is transformable for k = 1,2,, n, then F(ω ) can be
differentiated n times and we have
F t k f (t)  = i kF (k) (ω ) (k = 1,..., n ). (B.28)

Finally, if we let f (t), g(t) be two Fourier-transformable functions (with
transforms F(ω ), G(ω )), an important property concerns the convolution
w(t) of the two functions, defined as
∞
w(t) =
∫ f (t − τ )g(τ )dτ = ( f ∗ g )(t), (B.29)
−∞
where the ‘ ∗ ’ symbol in the rightmost expression is a standard nota-

tion for convolution. If the product f (t)g(t) is absolutely integrable, that
∞
is if
∫ −∞
f (t)g(t) dt < ∞, we can write the chain of relations (all integrals
are from −∞ to ∞)
1 1
F[ w(t)] =
2π ∫∫ f (t − τ )g(τ ) e − iω t
dτ dt =
2π ∫ g(τ )  ∫ f (t − τ ) e − iω t
dt  dτ

 1 
=
∫ g(τ ) e − iω τ
 2π ∫ f (s) e − iω s
∫
ds  dτ = F(ω ) g(τ ) e − i ω τ dτ = 2π F(ω )G(ω ),

(B.30)
where in the third equality we made the change of variable s = t − τ in the

integral within square brackets (or, which is the same, used property B.252).
Equation B.30 shows that the convolution ( f ∗ g ) (t) is transformed into 2π
Appendix B 299
times the product of the transforms F(ω ) and G(ω ). By a ‘symmetric’ argu-
ment, the reader is invited to show that F−1 [(F ∗ G)(ω )] = f (t)g(t), which, in
turn, suggests that we also have
F [ f (t)g(t)] = ( F ∗ G ) (ω ). (B.31)
However, in order to show how Equation B.31 can be obtained directly, it is

convenient to make use of the Dirac delta function introduced in Section B.3.
Remark B.5
i. The factor 2π in Equation B.30 depends on the convention adopted

to define the Fourier transform and on the definition of con-
volution. Note, in fact, that one sometimes finds the definition
∞
( f ∗ g ) (t) ≡ A ∫ f (t − τ )g(τ ) dτ where, depending on the author, we
−∞
may have A = 1 2π or A = 1 2π;
ii. Omitting the argument t for brevity of notation, it can be shown that
the convolution product satisfies the following basic properties: (a)
f ∗ g = g ∗ f , (b) f ∗ ( g + h) = ( f ∗ g ) + ( f ∗ h) and (c) f ∗ ( g ∗ h) = ( f ∗ g ) ∗ h.
B.2.2 The uncertainty principle

In its most celebrated form, the uncertainty principle is the quantum
mechanical relation ∆x∆p ≥  2 given by the physicist W. Heisenberg during
the 1920s. In just a few words, the principle states that the product of the
uncertainties of position (∆x ) and momentum (∆p ) of a particle is always
greater than  2, where  = 1.05 × 10−34 Js is Planck’s constant. In more gen-
eral terms, it turns out that any pair of conjugate variables – of which posi-
tion and momentum in quantum mechanics are a typical example – must
obey some form of uncertainty relation. For our purposes, the conjugate
variables are time t and frequency ν and the uncertainty principle in the
field of signal analysis relates the time duration of a signal f (t) to the range
of frequencies that are present in its Fourier transform. So, if ∆t is the dura-
tion of the signal f (t) and ∆ν is the range of frequencies spanned by F(ν ),
the principle reads
∆t ∆ν ≅ 1, (B.32)
which, in this form, is often called bandwidth theorem. The approxi-

mate sign means that in most cases, the product generally lies in the range
0.5–3.0 (the reader is invited to try with the boxcar function of Example
B.2, where ∆ν can be approximately taken as the width at the basis of the
300 Appendix B
central peak), but the precise value of the number on the r.h.s. of Equation
B.32 is not really important; the point of the theorem is that two members
of a Fourier transform pair – each one in its appropriate domain – cannot
be both ‘narrow’. The implications of this fact pervade the whole subject
of signal analysis and have important consequences in both theory and
practice.
In applications, it is quite common to encounter situations that confirm
the principle. For example, when a lightly damped structure in free vibration
oscillates for a relatively long time (large ∆ t ) at its natural frequency ω n , this
implies that a graph of the Fourier transform of this vibration signal will be
strongly peaked at ω = ω n (i.e. with a small ∆ω ); by contrast, if we want to
excite many modes of vibration of a structure in a relatively large band of fre-
quencies (large ∆ω ), we can do so by a sudden blow with a short time duration
(small ∆ t ). So, the point is that, as far as we know, the uncertainty principle
represents an inescapable law of nature with which we must come to terms.
B.3 A short digression on the Dirac delta ‘function’

The Dirac delta – denoted by δ (t) or δ (x), depending on the independent
variable of interest – is a mathematical entity frequently used in applications
to represent some finite physical phenomenon achieved in an arbitrarily
small interval of time or space, where by ‘arbitrarily small’ we mean highly
localised in time or space, and in any case much smaller than the typical
intervals (of time or space) that characterise the problem under study. For
our purposes, two illustrative examples can be (a) an impulsive force that
lasts for a very short interval of time or (b) an action/excitation (e.g. a force)
applied at a given point of a structure.
For any reasonably well-behaved function f (t), the standard ‘definition’
of the Dirac delta is given by means of the property
∞ ∞

∫
−∞
f (t)δ (t) dt = f (0),
∫ f (τ )δ (t − τ ) dτ = f (t),(B.33)
−∞
(where the second relation is a translated version of the first), which, in the
light of the fact that the l.h.s. of Equation B.332 is the convolution product
(f ∗ δ )(t), tells us that the Dirac delta is the unit element for the operation of
convolution, that is
(f ∗ δ )(t) = f (t).(B.34)
Remark B.6
The reason for the above quotation marks in ‘function’ and ‘definition’ is
that the Dirac delta is not an ordinary function and Equation B.33 is not a
Appendix B 301
proper definition. In fact, when one considers that Equation B.331 is trying
to tell us is that the Dirac delta must∞simultaneously have the two proper-
ties: (a) δ (t) = 0 for all t ≠ 0, and (b)
∫
δ (t) dt = 1, the weakness of the argu-
−∞
ment is evident; unless δ (t) is something other than an ordinary function,
the integral of a function that is zero everywhere except at one point is
necessarily zero no matter what definition of integral is used. The rigorous
mathematical justification of these facts was given in the 1940s by Laurent
Schwartz with his theory of distributions, where he showed that the Dirac
delta is a so-called distribution (or generalised function). Since, however,
our main interest lies in the many ways in which the Dirac delta is used in
applications, for these rigorous aspects we refer the more mathematically
oriented reader to Chapters 7 and 8 of Appel (2007), Chapter 6 of Debnath
and Mikusinski (1999) or to the book of Vladimirov (1981).
Now consider the definitions of Fourier transform of a function f (t) and
its inverse (Equations B.16 and B.17). If we substitute one into the other, we
get (all integrals are from −∞ to ∞)
 1   1 
f (t) =
∫ F(ω )e iω t
dω =
∫  2π ∫ f (r)e − iω r
dr  e i ω t dω =
 ∫ f (r)  2π ∫ e i ω (t − r )
dω  dr ,

which, owing to Equation B.332 , shows that the last integral within brack-
ets can be interpreted as an integral representation of the delta function and
that we can write
∞ ∞
1 1
δ (t − r) =
2π ∫
−∞
e i ω (t − r) dω , δ (t) =
2π ∫e
−∞
iω t
dω , (B.35)
where in the second expression, we set r = 0. This last relation, moreover,

suggests that δ (t) can be interpreted as the inverse Fourier transform of the
constant 1 2π , and consequently that the Fourier transform of δ (t) is 1 2π .
These considerations imply that, at least formally, we have
δ (t) = F−1 [1 2π ] , F [δ (t)] = 1 2π. (B.36)
Owing to Equations B.35, it turns out that the Dirac delta can be conve-
niently used in many cases. As a first example, we can now directly obtain
Equation B.31. In fact, starting from the chain of relations (again, all inte-
grals are from −∞ to ∞)
 ds e i r t dr
∫
f (t)g(t) = f (t) G(r) e i r t dr =
∫ G(r)  ∫ F(s)e i st


 ds dr ,
=
∫ G(r)  ∫ F(s)e i (s + r ) t

302 Appendix B
we can make the change of variable z = s + r (so that s = z − r and dz = ds ) in

the integral within brackets to get
 dz  dr =  

∫ G(r)  ∫ F(z − r)e izt
 ∫  ∫ F(z − r)G(r) dr  e izt
dz ,
where the integral within brackets is the convolution (F ∗ G)(z). So, if now
for present convenience we denote this convolution by W (z), the result thus
obtained is

∫
f (t) g(t) = W (z) e i z t dz (B.37)
and we can apply the Fourier transform operator on both sides of Equation
B.37 to get, as desired, Equation B.31, that is
1   1 
dz  e − i ω t dt =

F [ f (t)g(t)] =
2π ∫  ∫ W (z) e izt
 ∫ W (z)  2π ∫ e i (z −ω ) t
dt  dz

=
∫ W (z)δ (z − ω ) dz = W (ω ) = (F ∗ G)(ω ), (B.38)
where we have used the integral representation B.351 in the third equality
and the defining property B.331 in the fourth.
A second example is Parseval’s relation B.23, which can be obtained by
writing the chain of relations
1  dt   f ∗ (r) e i ω r dr  dω
∫ F(ω ) ∫ F(ω )F (ω ) dω = (2π) ∫  ∫ f (t) e ∫
2 ∗ −i ω t
dω = 2  
1  1 
=
2π ∫∫ f (t)f (r)  2π ∫ e
∗ i ω (r − t )
dω  drdt

1 1 1
f (t)  f ∗ (r)δ (r − t) dr  dt =
∫ ∫ ∫ ∫
2
= f (t)f ∗ (t) dt = f (t) dt
2π   2π 2π
(B.39)
At this point, one may ask about the integral of the Dirac delta. So, by
introducing the so-called Heaviside (or unit step) function
 0 t <0
θ (t) =  ,(B.40)
 1 t ≥0
we notice that we can formally write the relations

Appendix B 303
t
dθ (t)
δ (t) =
dt
, θ (t) =
∫ δ (τ ) dτ ,(B.41)
−∞
which, in turn, lead to the question about the derivative δ ′(t) of δ (t). For
this, however, we must take an indirect approach because δ ′(t) cannot be
directly determined from the definition of δ (t); by letting f (t) be a well-
behaved function and assuming the usual rule of integration by parts to
hold, we get
∫ f (t)δ ′(t) dt = f (t)δ (t) ∫ f ′(t)δ (t) dt = − f ′(0), (B.42)

∞
−∞
−
thus showing that δ ′(t), like δ (t), vanishes for all t ≠ 0 but ‘operates’ (under
the integral sign) on the derivative f ′(t) rather than on f (t). Then, generalis-
ing the relation B.42, the kth derivative δ (k) (t) is such that

∫ f (t)δ (k)
(t) dt = (−1)k f (k) (0).(B.43)
Finally, in many applications, it may be convenient to see δ (t) as the limit of

a sequence of functions wε (t), where ε is a small parameter that goes to zero
and the sequence is such that
∫ w (t) dt = 1 for all values of ε . In this light, we
ε
can write δ (t) = lim wε (t) and two illustrative examples of such sequences are
ε →0
sin(t ε )
wε (t) =
1
ε π
exp − t 2 ε 2 , ( ) wε (t) =
πt
, (B.44)
which, in the limit, satisfy the defining property B.331. With the Gaussian
functions B.441, in fact, for any well-behaved function f (t) that vanishes
fast enough at infinity to ensure the convergence of any integral in which it
occurs, we get
1 1 f (0) − s2
lim
e→0 ε π ∫ f (t) e −t 2 ε 2
dt = lim
e→0 π ∫ f (ε s) e − s2
ds =
π ∫e ds = f (0),
where we made the change of variable s = t ε in the first equality and the
last relation holds because
∫e − s2
ds = π . Also, it may be worth noticing
that by defining A = 1 ε in the functions B.442 , we can once again obtain
the integral representation of Equation B.352; in fact, we have
A ∞
sin At 1 1
δ (t) = lim
A→∞ πt
= lim
A→∞ 2π ∫
−A
e i ω t dω =
2π ∫e
−∞
iω t
dω ,
304 Appendix B
where in the second equality we used the relation

A
( π t )−1 sin(At) = (2π )−1 ∫ e i ω t dω .
−A
B.4 LAPLACE TRANSFORM
From a general mathematical viewpoint, the Fourier transform is a special

case of integral transformation, that is a correspondence between two func-
tions f (t) and F(u) such that
F(u) = T [ f (t)] =
∫ K(t, u)f (t) dt ,(B.45)
a
where the function K(t , u) is called the kernel of the transformation and
F(u) – often symbolically denoted by T[ f (t)] – is the transform of f (t) with
respect to the kernel. The various transformations differ (and hence have
different names) depending on the kernel and on the integration limits a, b.
So, for example, we have seen that the choice K(t , u) = (1 2π ) e − i u t together
with a = −∞, b = ∞ gives the Fourier transform, but other frequently encoun-
tered types are called the Laplace, Hankel and Mellin transforms, just to
name a few. Together with the Fourier transform, the Laplace transform is
probably the most popular integral transformation and is defined as
∞
F(s) ≡ L [ f (t)] =
∫ f (t) e
0
− st
dt ,(B.46)
where s is a complex variable. The form of Equation B.46 suggests that

the Laplace transform can be particularly useful when we are interested
in functions f (t) such that (a) f (t) = 0 for t ≤ 0, and (b) f (t) is not integrable
on the interval [0, ∞). In this case, the ‘bad behaviour’ of f (t) at infinity
can often be ‘fixed’ by multiplying it by a factor exp ( −ct ), where c is a real
number larger than some value α (α is sometimes called the convergence
abscissa, and its exact value depends on the function to be transformed).
By so doing, we can Fourier transform the function g(t) = f (t) e − ct and
∞
obtain
∫ 0
f (t) e − (c+ i ω ) t dt = F(c + iω ), which, introducing the complex vari-
able s = c + iω , is exactly of the form of Equation B.46 and is such that the
integral exists for all c > α , that is in the right-hand region of the complex
s-plane where c = Re(s) > α . So, for example, if b is a real positive constant
and we consider the function defined as f (t) = exp(bt) for t ≥ 0 and zero oth-
erwise, then its Fourier transform does not exist but the Fourier transform
of ebt e − ct does exist for c > b and we get
Appendix B 305
1
F(s) = L ebt  =
s−b
(c = Re(s) > b ). (B.47)
Remark B.7
i. In the above Fourier transform, we have intentionally omitted the

multiplying factor 1 2π because the Laplace transform is almost
universally defined as in Equation B.46. This factor, as we will see
shortly, will appear in the inverse transform;
ii. In regard to terminology, one sometimes calls original any Laplace-
transformable function f (t), while its transform F(s) is called the image
(of the original).

In the light of definition B.46, it is natural to ask for the inverse
transform of an image F(s). So, even if in most cases it is common to
make use of (widely available) tables of integral transforms, it can be
shown (see, for example, Boas (1983), Mathews and Walker (1970)
or Sidorov et al. (1985)) that the inversion formula (also known as
Bromwich integral) is
c+ i ∞
1
−1
L [ F(s)] =
2iπ ∫ F(s) e
c− i ∞
st
ds, (B.48)
where the notation in the term on the r.h.s. means that we integrate along
the vertical straight line Re(s) = c (c > α ) in the complex plane. Then, the
integral converges to f (t), where f (t) is continuous, while at jumps, it con-
verges to the mid-value of the jump; in particular, for t = 0, the integral
( )
converges to f 0+ 2. Also, it may be worth mentioning that the integral of
Equation B.48 can be evaluated as a contour integral by means of the so-
called theorem of residues, one of the key results in the theory of functions
of a complex variable.
B.4.1 L aplace transform: Basic properties

and some examples
The basic properties of the Laplace transform are not very dissimilar from
the properties of the Fourier transform. First of all, the transformation is lin-
ear, and for any two constants a, b, we have L [ af (t) + bg(t)] = a F(s) + bG(s)
whenever f (t), g(t) are Laplace-transformable with transforms F(s), G(s).
Also, it is not difficult to show that
L [ f (t − a)] = e − as F(s), L  f (t) e − at  = F(s + a).(B.49)

306 Appendix B
Passing to the first and second derivatives f ′(t), f ′′(t) – which we assume to
be originals in the sense of Remark B.7 (ii) – a single and a double integra-
tion by parts, respectively, lead to
( )
L[ f ′(t)] = s F(s) − f 0+ , ( ) ( )
L[ f ′′(t)] = s2F(s) − sf 0+ − f ′ 0+ , (B.50a)
( ) ( )
where f 0+ , f ′ 0+ are the limits as t approaches zero from the positive side.
More generally, if we denote by f (k) (t) the kth derivative of f (t), Equation
B.50a are special cases of the relation
k
L  f (k) (t)  = skF(s) − ∑s

j =1
k− j ( j −1)
f (0 ). (B.50b)
+
t
If f (t) is an original with image F(s) and if the function I (t) =
∫ 0
f (u) du is
also an original, then we can apply Equation B.491 to I (t) and take into
( )
account that I 0+ = 0. Then
F(s)
L[ I (t)] = . (B.51)
s
A final property we consider here is the counterpart of Equation B.30. If, in
fact, we let F(s), G(s) be the images of the two originals f (t), g(t) with conver-
gence abscissas α , β , respectively, then the convolution theorem for Laplace
transforms reads
L [(f ∗ g)(t)] = F(s)G(s) (B.52)
for Re(s) > max(α , β ).
Example B.4
With reference to inverse Laplace transforms, in applications one often
finds functions that can be written as the ratio of two polynomials in
s, that is F(s) = P(s) Q(s), where Q(s) is of higher degree than P(s). Then,
calling f (t) the inverse transform of F(s), an easy recipe in these cases is
provided by the two following rules:
a. To a simple zero q of Q(s) there corresponds in f (t) the term
P(q) q t
e (B.53)
Q′(q)
Appendix B 307
b. To two complex conjugate simple zeros q ± iω of Q(s) there cor-

responds in f (t) a term of the form ( A cos ω t − B sin ω t ) eq t , where
the real numbers A, B are given by the relation
 P(s) 
A + iB = 2   . (B.54)
 Q′(s)  s=q + i ω
As a specific example, consider the function F(s) = (s + 1) [ s(s + 2)],

where the two simple zeros of the denominator are q1 = 0, q2 = −2.
Then, Q′(s) = 2s + 2, so that Q′ ( q1 ) = 2, Q′ ( q2 ) = −2 . For the numer-
ator, on the other hand, we have P ( q1 ) = 1, P ( q2 ) = −1, thus implying,
by rule (a), that f (t) is given by the sum of the two terms
P ( q1 ) P ( q2 ) 1 1 −2 t
f (t) = L−1 [ F(s)] = eq1 t + eq2 t = + e .
Q′ ( q1 ) Q′ ( q2 ) 2 2
As a second example, consider the function
s
F(s) = ,
(
(s + 3) s2 + 4s + 5 )
whose denominator has the simple root q1 = −3 and the complex
conjugate pair of zeros q ± iω = −2 ± i. For the simple root, rule (a)
leads to the term −(3 2) e −3 t, while, on the other hand, rule (b) tells
us that the term associated with the complex conjugate zeros is
e −2 t (A cos t − B sin t). Then, using Equation B.54, we leave to the
reader the calculations that lead to A = 3 2, B = 1 2. Finally, putting
the pieces back together, we can write the desired result as
3 −3 t  3 1 
f (t) = L−1 [ F(s)] = − e +  cos t − sin t  e −2 t .
2 2 2 
Example B.5
Equations B.50 show that Laplace transforms ‘automatically’ provide
for the initial conditions at t = 0. In this example, we exploit this prop-
erty by considering the ordinary (homogeneous) differential equation
d 2 f (t)
+ a2 f (t) = 0, (B.55)
dt 2
where a is a constant and we are given the initial conditions f (0) = f0

and f ′(0) = f0′. Laplace transformation of both sides of Equation B.55
gives  s2F(s) − sf0 − f0′  + a2F(s) = 0, so that the solution in the s-domain
is easily obtained as
308 Appendix B
sf0 f0′
F(s) = + .
s 2 + a2 s 2 + a2
Then, since from a table of Laplace transforms, we get

{ (
L−1 s s2 + a2 )} = cos at and L−1 {(s2 + a2 )-1 } = a−1 sin at, the solution is
f0′
f (t) = f0 cos at + sin at, (B.56)
a
which is exactly the result that can be obtained by standard methods

(with the appropriate modifications, in fact, see Equation 3.51 and the
first two of Equation 3.6).
On the other hand, if we consider the non-homogeneous differential
equation
d 2 f (t)
+ a2 f (t) = g(t), (B.57)
dt 2
where the forcing function on the r.h.s. is of the sinusoidal form

g(t) = gˆ cos ω t and the initial conditions are as above, Laplace transfor-
mation of both sides leads to
ˆ
gs sf f′
F(s) = + 2 0 2 + 2 0 2
( )(
s + a s2 + ω 2
2 2
s +a) ( s +a ) ( )
( )
because L[cos ω t ] = s s2 + ω 2 . In order to return to the time domain,
we already know the inverse transform of the last two terms, while for
the first term, we can use the convolution theorem and write
 s   1  −1  s  −1  1 
L−1  2 2 2 2 = L  2 2 ∗L  2
 s + ω   s + a    s + ω   s + a 
2
(
= (cos ω t) ∗ a−1 sin at =) 1
a ∫ cos(ω τ )sin a (t − τ ) dτ ,
0
(B.58)
from which it follows that the time-domain solution of Equation B.57 is
t
f0′ gˆ
f (t) = f0 cos at +
a
sin at +
a ∫ cos(ω τ )sin a(t − τ ) dτ ,
0
(B.59)
which, again, is the same result that we can obtain by the standard
methods.
Appendix B 309
Example B.6 – Initial value problem for an

infinitely long flexible string.
Partial differential equations can also be solved with the aid of Laplace
transforms, the effect of the transformation being the reduction of inde-
pendent variables by one. Also, Laplace and Fourier transforms can be
used together, as in the following example. It is shown in Chapter 5
that the equation of motion for the small oscillations of a vibrating
string is the one-dimensional wave equation
∂2 y 1 ∂2 y
= , (B.60)
∂ x2 c2 ∂t 2
where y = y(x, t), and here, we assume that the initial conditions are
given by the two functions
∂ y(x, t)
y(x,0) = u(x), = w(x). (B.61)
∂t t =0
If now we call Y (x, s) the Laplace transform of y(x, t) relative to the time
variable and transform both sides of Equation B.60, we get
∂ 2 Y (x, s) 1 2
= 2  s Y (x, s) − su(x) − w(x)  , (B.62)
∂ x2 c
which, in turn, can be Fourier-transformed with respect to the space

variable x to give
1 2
−k2 Ψ(k, s) =  s Ψ(k, s) − s U(k) − W (k) , (B.63)
c2 
where k, the so-called wavenumber, is the (Fourier) conjugate variable
of x and we define
Ψ(k, s) = F[Y (x, s)], U(k) = F[ u(x)], W (k) = F [ w(x)].
Then, some easy algebraic manipulations of Equation B.63 give

sU(k) + W (k)
Ψ(k, s) = , (B.64)
s 2 + k2 c 2
and we can use the tabulated inverse Laplace transforms used in
Example B.5 to get the inverse Laplace transform of Equation B.64 as
 s   1 
χ (k, t) ≡ L−1 [ Ψ(k, s)] = U(k) L−1  2 + W (k) L−1  2
 s + k c 
2 2 2 2 
s + k c 
(B.65)
W (k)
= U(k) cos(kct) + sin(kct),
kc
310 Appendix B
from which we can obtain the desired result by inverse Fourier trans-
formation. For the first term on the r.h.s. of Equation B.65, we have
∞
1
F −1 [U(k) cos kct ] =
2 ∫ U(k) e
−∞
i kc t
+ e − ikct  e i k x dk
∞ ∞
1 1
=
2 ∫
−∞
U(k) e i k (x+c t ) dk +
2 ∫ U(k) e
−∞
i k (x−c t )
dk
1
=
2
[ u(x + ct) + u(x − ct)] . (B.66a)
For the second term, on the other hand, we first note that
 sin(kct)   t

F −1 W (k)
 kc  

 0
∫
= F −1 W (k) cos(kcr) dr  ,


where r is a dummy variable of integration; then, it just takes a small

effort to obtain (hint: Use Euler’s relations)
x + ct
 sin(kct)  1
F −1 W (k)

=
kc  2c
x − ct
∫
w(ξ ) d ξ . (B.66b)
Finally, putting Equations B.66a and B.66b back together yields the
solution y(x, t) of Equation B.60 as
x + ct
1
y(x, t) = F −1 [ χ (k, t)] =
2
[ u(x + ct) + u(x − ct)] + 21c ∫ w(ξ) dξ , (B.67)
x − ct
which is the so-called d’Alembert solution (of Equation B.60 with ini-
tial conditions B.61). As an incidental remark, we note for a string
of finite length L, we can proceed from Equation B.62 by expanding
Y (x, s), u(x) and w(x) in terms of Fourier series rather than taking their
Fourier transforms.
References and further reading
Adhikari, S. (1999). Rates of change of eigenvalues and eigenvectors in damped

dynamic systems, AIAA Journal, 37(11): 1452–1458.
Adhikari, S. (2014a). Structural Dynamic Analysis with Generalized Damping
Models: Analysis, London: ISTE; Hoboken, NJ: John Wiley & Sons.
Adhikari, S. (2014b). Structural Dynamic Analysis with Generalized Damping
Models: Identification, London: ISTE; Hoboken, NJ: John Wiley & Sons.
Appel, W. (2007). Mathematics for Physics and Physicists, Princeton, NJ: Princeton
University Press.
Baruh, H. (1999). Analytical Dynamics, Singapore: WCB/McGraw-Hill.
Baruh, H. (2015). Applied Dynamics, Boca Raton, FL: CRC Press.
Bathe, K.J. (1996). Finite Element Procedures, Englewood Cliffs, NJ: Prentice-Hall.
Billingham, J. and King, A.C. (2000). Wave Motion, Cambridge, UK: Cambridge
University Press.
Bisplinghoff, R.L., Mar, J.W. and Pian, T.H.H. (1965). Statics of Deformable
Solids, New York: Dover Publications.
Boas, M.L. (1983). Mathematical Methods in the Physical Sciences, New York:
John Wiley & Sons.
Boyce, W.E. and DiPrima, R.C. (2005). Elementary Differential Equations and
Boundary Value Problems, 8th ed., Hoboken, NJ: John Wiley & Sons.
Brandt, A. (2011). Noise and Vibration Analysis: Signal Analysis and Experimental
Procedures, Chichester, UK: John Wiley & Sons.
Byron Jr., F.W and Fuller, R.W. (1992). Mathematics of Classical and Quantum
Physics, New York: Dover Publications.
Caughey, T.K. and O’Kelly, M.E.J. (1965). Classical normal modes in damped lin-
ear dynamic systems, ASME Journal of Applied Mechanics, 32: 583–588.
Chakraverty, S. (2009). Vibration of Plates, Boca Raton, FL: CRC Press
Clough, R.W. and Penzien, J. (1975). Dynamics of Structures, Singapore:
McGraw-Hill.
Collins, P.J. (2006). Differential and Integral Equations, Oxford: Oxford University
Press.
Corben, H.C. and Stehle, P. (1994). Classical Mechanics, 2nd ed., New York:
Dover Publications.
Davis, H.T. and Thomson, K.T. (2000). Linear Algebra and Linear Operators in
Engineering, San Diego, CA: Academic Press.
311
312 References and further reading
Daya Reddy, B. (1998). Introductory Functional Analysis, with Applications to

Boundary Value Problems and Finite Elements, New York: Springer-Verlag.
Debnath, L. and Mikusinski, P. (1999). Introduction to Hilbert Spaces with
Applications, London: Academic Press.
Dyke, P.P.G. (2001). An Introduction to Laplace Transforms and Fourier Series,
London: Springer.
Elishakoff, I. (1999). Probabilistic Theory of Structures, 2nd ed., Mineola, NY:
Dover Publications.
Elmore, W.C. and Heald, M.A. (1985). Physics of Waves, New York: Dover
Publications.
Ewins, D.J. (2000). Modal Testing: Theory, Practice and Applications, 2nd ed.,
Baldock: Research Studies Press.
Gatti, P.L. (2005). Probability Theory and Mathematical Statistics for Engineers,
Abingdon: Spon Press.
Gatti, P.L. (2014). Applied Structural and Mechanical Vibrations, Theory and
Methods, 2nd ed., Boca Raton, FL: CRC Press.
Gelfand, I.M. and Fomin, S.V. (2000). Calculus of Variations, Mineola, NY: Dover
Publications.
Géradin, M. and Rixen, D.J. (2015) Mechanical Vibrations: Theory and Applications
to Structural Dynamics, 3rd ed., Chichester, UK: John Wiley & Sons.
Ginsberg, J. (2008). Engineering Dynamics, Cambridge: Cambridge University
Press.
Goldstein, H. (1980). Classical Mechanics, 2nd ed., Reading, MA: Addison-Wesley.
Gould, S.H. (1995). Variational Methods for Eigenvalue Problems: An Introduction
to the Methods of Rayleigh, Ritz, Weinstein, and Aronszajn, New York:
Dover Publications.
Graff, K.F. (1991). Wave Motion in Elastic Solids, New York: Dover Publications.
Greenwood, D.T. (1977). Classical Dynamics, Mineola, NY: Dover Publications.
Greenwood, D.T. (2003). Advanced Dynamics, Cambridge: Cambridge University
Press.
Guenther, R.B. and Lee, J.W. (2019). Sturm-Liouville Problems, Theory and
Numerical Implementation, Boca Raton, FL: CRC Press.
Hagedorn, P. and DasGupta, A. (2007). Vibration and Waves in Continuous
Mechanical Systems, Chichester, UK: John Wiley & Sons.
Halmos, P.R. (2017). Finite-dimensional Vector Spaces, 2nd ed., Mineola, NY:
Dover Publications.
Hamming, R.W. (1998). Digital Filters, 3rd ed., Mineola, NY: Dover Publications.
Hildebrand, F.B. (1992). Methods of Applied Mathematics, New York: Dover
Publications.
Horn, R.A. and Johnson, C.R. (1993). Matrix Analysis, Cambridge: Cambridge
University Press.
Howell, K.B. (2001). Principles of Fourier Analysis, Boca Raton, FL: Chapman &
Hall/CRC.
Humar, J.L. (2012). Dynamics of Structures, 3rd ed., Boca Raton, FL: CRC
Press.
Inman, D.J. (1994). Engineering Vibration, Englewood Cliffs, NJ: Prentice-Hall.
Inman, D.J. (2006), Vibration with Control, Chichester, UK: John Wiley & Sons.
Ivchenko, G.I. and Medvedev, Yu.I. (1990) Mathematical Statistics, Moscow: Mir
Publishers.
Jazar, R.N. (2013). Advanced Vibrations: A Modern Approach, New York:
Springer.
Junkins, J.L. and Kim, Y. (1993). Introduction to Dynamics and Control of Flexible
Structures, Washington DC: AIAA (American Institute of Aeronautics and
Astronautics).
Kelly, S.G. (2007) Advanced Vibration Analysis, Boca Raton, FL: CRC Press.
Komzsik, L. (2009). Applied Calculus of Variations for Engineers, Boca Raton,
FL: CRC Press.
Köylüoglu, H.U. (1995). Stochastic response and reliability analyses of struc-
tures with random properties subject to stationary random excitation, PhD
Dissertation, Princeton University, Princeton, NJ.
Lanczos, C. (1970). The Variational Principles in Mechanics, New York: Dover
Publications.
Landau, L.D. and Lifshitz E.M. (1982). Meccanica, Roma: Editori Riuniti.
Laub, A.J. (2005). Matrix Analysis for Scientists and Engineers, Philadelphia, PA:
SIAM (Society for Industrial and Applied Mathematics).
Leissa, A.W. (1969). Vibration of Plates, NASA SP-160, Washington DC, US
Government Printing Office.
Leissa, A.W. (1973). The free vibration of rectangular plates, Journal of Sound and
Vibration, 31: 257–293.
Leissa, A.W. and Qatu, M.S. (2011). Vibrations of Continuous Systems, New York:
McGraw-Hill.
Lurie, A.I. (2002). Analytical Mechanics, Berlin Heidelberg: Springer.
Lutes, L.D. and Sarkani, S. (1997). Stochastic Analysis of Structural and Mechanical
Vibrations, Upper Saddle River, NJ: Prentice-Hall.
Ma, F., Caughey, T.K. (1995), Analysis of linear nonconservative vibrations, ASME
Journal of Applied Mechanics, 62: 685–691.
Ma, F., Imam, A., Morzfeld, M. (2009). The decoupling of damped linear systems
in oscillatory free vibration, Journal of Sound and Vibration, 324: 408–428.
Ma, F., Morzfeld, M., Imam, A. (2010). The decoupling of damped linear systems
in free or forced vibration, Journal of Sound and Vibration, 329: 3182–3202.
Maia, N.M.M. and Silva, J.M.M. (eds.). (1997) Theoretical and Experimental
Modal Analysis, Taunton: Research Studies Press.
Mathews, J. and Walker, R.L. (1970). Mathematical Methods of Physics, 2nd ed.,
Redwood City, CA: Addison-Wesley.
McConnell, K.G. (1995). Vibration Testing, Theory and Practice, New York: John
Wiley & Sons.
Meirovitch, L. (1980). Computational Methods in Structural Dynamics, Alphen
aan den Rijn: Sijthoff & Noordhoff.
Meirovitch, L. (1997). Principles and Techniques of Vibrations, Upper Saddle
River, NJ: Prentice-Hall
Moon, F.C. (2004). Chaotic Vibrations: An Introduction for Applied Scientists
and Engineers, Hoboken, NJ: John Wiley & Sons.
Naylor, A.W. and Sell, G.R. (1982). Linear Operator Theory in Engineering and
Science, New York: Springer-Verlag.
314 References and further reading
Newland, D.E. (1989). Mechanical Vibration Analysis and Computation, Harlow,

Essex: Longman Scientific & Technical.
Newland, D.E. (1993). An Introduction to Random Vibrations, Spectral and
Wavelet Analysis, 3rd ed., Harlow: Longman Scientific and Technical.
Palazzolo, A.B. (2016). Vibration Theory and Applications with Finite Element
and Active Vibration Control, Chichester, UK: John Wiley & Sons.
Papoulis, A. (1981). Signal Analysis, New York: McGraw-Hill.
Pars, L.A. (1965). A Treatise on Analytical Dynamics, London: Heinemann.
Petyt, M. (1990). Introduction to Finite Element Vibration Analysis, Cambridge:
Cambridge University Press.
Pfeiffer, F. and Schindler, T. (2015). Introduction to Dynamics, Heidelberg:
Springer.
Preumont, A. (1994). Random Vibration and Spectral Analysis, Dordrecht:
Springer Science + Business Media.
Preumont, A. (2013). Twelve Lectures on Structural Dynamics, Dordrecht:
Springer.
Rao, S.S. (2007). Vibration of Continuous Systems, Hoboken, NJ: John Wiley &
Sons.
Rektorys, K. (1980). Variational Methods in Mathematics, Science and Engineering,
Dordrecht-Holland: D. Reidel Publishing Company.
Rivin, E.I. (2003). Passive Vibration Isolation, New York: ASME Press.
Rosenberg, R.M. (1977). Analytical Dynamics of Discrete Systems, New York:
Plenum Press.
Sagan, H. (1989). Boundary and Eigenvalue Problems in Mathematical Physics,
New York: Dover Publications.
Savelyev, I.V. (1982). Fundamentals of Theoretical Physics, Vol. 1, Moscow: Mir
Publishers.
Schmidt, G. and Tondl, A. (1986). Non-linear Vibrations, Cambridge: Cambridge
University Press.
Shin, K. and Hammond, J.K. (2008). Fundamentals of Signal Processing for Sound
and Vibration Engineers, Chichester, UK: John Wiley & Sons.
Sidorov, Yu.V., Fedoryuk, M.V. and Shabunin, M.I. (1985). Lectures on the Theory
of Functions of a Complex Variable, Moscow: Mir Publishers.
Solnes, J. (1997). Stochastic Processes and Random Vibrations, Theory and
Practice, Chichester, UK: John Wiley & Sons.
Szilard, R. (2004). Theories and Applications of Plate Analysis: Classical,
Numerical and Engineering Methods, Hoboken, NJ: John Wiley & Sons.
Thomsen, J.J. (2003). Vibrations and Stability: Advanced Theory, Analysis and
Tools, 2nd ed., Heidelberg: Springer.
Vladimirov, V.S. (1981). Le Distribuzioni nella Fisica Matematica, Mosca:
Edizioni Mir.
Vladimirov, V.S. (1987). Equazioni della Fisica Matematica, Mosca: Edizioni Mir.
Vu, H.V. and Esfandiari, R.S. (1998). Dynamic Systems, Modeling and Analysis,
New York: McGraw-Hill.
Vujanovic, B.D. and Atanackovic, T.M. (2004). An Introduction to Modern
Variational Techniques in Mechanics and Engineering, New York: Springer
Science+Business Media.
Wilkinson, J.H. (1996). The Algebraic Eigenvalue Problem, Oxford: Clarendon

Press.
Williams, M. (2016). Structural Dynamics, Boca Raton, FL: CRC Press.
Yang, C.Y. (1986). Random Vibration of Structures, New York: John Wiley &
Sons.
Zachmanoglou, E.C. and Thoe, D.W. (1986). Introduction to Partial Differential
Equations with Applications, New York: Dover Publications.
Ziegler, H. (1977). Principles of Structural Stability, 2nd ed., Switzerland: Springer
Basel AG.
Index
accelerance 114, 138 constraint

acceleration 6 bilateral16
accelerometer 7 equations 15
action integral 34 forces 16, 33
amplitude 3 holonomic 15
complex 4 non-holonomic 15
peak 3 rheonomic 15
angular frequency 3 scleronomic 15
asymptotically stable system 64 unilateral 16
continuous system 2
convolution 101, 225, 298
base excitation 105–106 co-ordinates
basis 261 Cartesian 15
orthonormal 264–267 generalised 15
beams 162–167 ignorable (or cyclic) 53
axial-force effects 168–170 correlation function 209, 214–216
Euler-Bernoulli 163, 187, 197 correlation matrix 241
Timoshenko 163, 170–174 covariance function 209, 214–216
Rayleigh 173 critically damped system 62
shear 172
beats 5–6, 76 dashpot 9
biorthogonality 275 d’Alembert solution 141, 310
bode diagram 108 damping
boundary condition (BC) 40, 143 classical 87–90, 120–122
geometric (or imposed) 40 coefficient 9
natural (or force) 40 critical 61
boundary-value problems (BVP) 150 evaluation 116–117
broadband process 230–232 matrix 47
BVP see boundary-value problem non-classical 90–92, 126–133
light damping perturbation 80–81
Caughey series 89 ratio 61
Cholesky factorisation 83 decibel 7–8
Christoffel symbol 27 degree of freedom (DOF) 2, 16
Clapeyron’s law 30 deterministic vibration 2
complete orthonormal system 154 differential eigenvalue problem 186
condition number 287 dirac delta ‘function’ 300–304
configuration space 93 discrete system 2
317
318 Index
displacement 6 continuous systems 148–150,

Duhamel integral 101 156–167, 174–181, 188–190
DOF see degree of freedom frequency 3
dynamic coupling 77 equation 67, 144
dynamic magnification factor 107 fundamental 67, 289
dynamic potential 27 natural 60
of damped oscillation 62
eccentric excitation 112–114 ratio 102
eigenfunction 145, 151 frequency response function (FRF)
eigenvalue 67, 145, 151, 274 102–103, 114–116, 137–138,
degeneracy 68, 83, 278 199–201
sensitivity 78–81, 96–97 estimate 235
eigenvector 67, 274 estimate 235
left 130, 275 modal 121, 192, 244
right 130, 275 FRF see frequency response function
sensitivity 78–81 function
energy Bessel 159, 180
function 28 coherence 236
kinetic 18 complementary 99
potential 19 Green 201–205
strain 30 harmonic 3
ensemble 211 Heaviside 302
equation periodic 109, 289
Bessel 159, 180 weight 151
Helmholtz 157
ergodic random process 216–219 Gaussian (or normal) processes
Euler’s buckling load 169 239–241
Euler relations 3 generalised momenta 20
extended Hamilton’s principle 36 generalised potential 24
flexibility (or compliance) 9 generalised eigenvalue problem (GEP)
force 68, 92–96, 274
conservative 18 Gibbs’ phenomenon 294
elastic 29–30 Gram-Schmidt orthonormalisation
fictitious 50, 53 process 267–268
generalised 18 Green’s formula 151
gyroscopic 27
inertia 19 half-power bandwidth 116
monogenic 24 Hamilton function (or
nonconservative 19 Hamiltonian) 20
viscous 30–32 Hamilton equations 21
forced vibration 48 non-holonomic form 34
SDOF systems 100–103, 107–110 Hamilton’s principle 34–37
continuous systems 190–199 extended 36
Fourier harmonic excitation 107–108
coefficients 289 Hermitian form 282
series 289–294
transform 294–300 independent random variables 210
free vibration 48 impulse response function (IRF) 100
SDOF systems 59–65 modal 120, 191, 244
MDOF systems 67–68, 70–72 initial value problem (IVP) 14, 141
Index 319
inner product 265, 268–269 mean 210

IRF see impulse response function mean square value 210, 291
isomorphism 264 matrix 245
mechanical impedance 116
Lagrange’s equations 18 membrane free vibration 156–157
non-holonomic form 33 circular 158–162
standard form 19 rectangular 157–158
Lagrange function (or Lagrangian) 19 multiple inputs-multiple outputs
Lagrange identity 151 (MIMO) systems 241–243
Lagrange’s multipliers 32–34 mobility 114, 117
Lagrangian density 38, 157, 175 modal
Laplace transform 304–310 damping ratio 86
linear system 1 force vector 120
logarithmic decrement 65–66 matrix 70, 281
participation factor 120
mass 9 mode
apparent 116 acceleration method 122–124
coefficients 46 complex 81, 91–92
matrix 46 shape 67, 145
modal 69 truncation 122
matrix 255 moments (in probability) 209
defective 278 central 210
determinant 258 non-central 210
diagonal 257 multiple degrees of freedom (MDOF)
diagonalisable 277 systems
dynamic 93 multiplicity
exponential 282–283 algebraic 68, 275, 278
Hermitian (or self-adjoint) 256, geometric 83, 278
279–281
Hermitian adjoint 256 narrowband process 229–230
inverse 259 peak distribution 251–254
negative-definite 282 threshold crossing rates 249–251
negative-semidefinite 282 natural system 25, 28
nondefective 278 Newton’s laws 13
nonsingular 259 non-inertial frame of reference 48–53
normal 257 nonlinear system 1
orthogonal 257 non-natural system 25
positive-definite 282 norm
positive-semidefinite 282 matrix 285–288
rank 259 vector 265–266
similar 270, 277 normal (or modal) co-ordinates 72–73,
singular 259 120, 244
skew-hermitian 256 normalisation 68–70, 147, 274
skew-symmetric 256 mass 69
spectral 70 Nyquist plot 108, 117
symmetric 256, 279–281
trace 258 operator
transpose 256 beam 182–183
unitary 257 biharmonic 175
MDOF see multiple degree of freedom Laplace (also Laplacian) 157
320 Index
operator (Cont.) variable (r.v.) 207

linear 263–267 vibration 2, 207
mass 186 Rayleigh
plate 183–185 damping 88
stiffness 186 dissipation function 30–32
Sturm-Liouville 150 distribution 251, 253
self-adjoint 152 quotient 154
orthogonality 68–70, 82, 92 receptance 114
mass 69 reciprocity 122, 205
of beam eigenfunctions 167–168 resonance 108, 110
of plate eigenfunctions 181–182 response
stiffness 69 indicial 104
overdamped system 62 pseudo-static 123
resonant 109
Parseval’s relation 291, 296–297 spectrum 117–119
pendulum transient237–239
compound (or physical) 44 Riemann-Lebesgue lemma 292
double 23, 44 rigid-body modes 83–87, 125, 149
simple 21–23, 56–57 rod vibration 148–150
period 3 root mean square 210
periodic excitation 109–110 rotary (or rotatory) inertia 170, 173,
Phase angle 3 174
phase space 93 Routh function (or Routhian) 53–55
phasor 4
plates 174–176 sample function 211
circular 180–181 SDOF see single degree of freedom
flexural stiffness 175 shaft vibration 148–150
Kirchhoff theory 174 shear deformation 163, 170, 172, 174,
Mindlin theory 174 187
rectangular 176–180 single degree of freedom (SDOF)
power spectral density (PSD) 223, systems 92, 233, 236–239,
227–228 244, 251
matrix 241 singular value decomposition 283–285
principle single input-single output (SISO)
d’Alembert 17 systems 233–236
of least action 35 spectral density see power spectral
of virtual work 17 density
superposition 1 spectral matrix 70
uncertainty 299 spring 8
probability equivalent 9–11
density function (pdf) 208 stable system 64
distribution function (PDF) 208 standard deviation 210
progressive wave 141 standard eigenvalue problem (SEP) 68,
PSD see power spectral density 273–276
standing wave 142
quadratic eigenvalue problem (QEP) stationary random process 212–214,
80, 92–96 223–227
quadratic form 282 weakly stationary (WS) 214
state space 93–96, 126–133
random stiffness 8
process 207, 219–223 coefficients 45
Index 321
coupling 77 uncorrelated random variables 211

dynamic 116 underdamped system 62
matrix 46 unstable system 64
modal 69
stochastic process see random process
string vibrations 140–148 variance 210
Sturm-Liouville problem (SLp) 146, varied path 35
150–156 vector space 260–263
regular 150 dimension 261–262
singular 161 subspace 261
velocity 6
theorem group 162, 173
bandwidth 299–300 phase 162, 173
Bauer-Ficke 288 vibration isolation 110–112
Betti’s 182 virtual displacement 16
Dirichlet 292 viscous damper 8
Euler’s 29
Murnaghan-Wintner 284 wave equation 140, 148
Schur’s decomposition 283–284 wavelength 142
spectral for normal matrices 281 wavenumber 142, 309
transmissibility Weibull distribution 253
force 111 white noise 231
motion 110 band-limited 231
transfer function (TF) 102 Wiener-Khintchine relations 224
Taylor & Francis eBooks
www.taylorfrancis.com
A single destination for eBooks from Taylor & Francis

with increased functionality and an improved user
experience to meet the needs of our customers.
90,000+ eBooks of award-winning academic content in

Humanities, Social Science, Science, Technology, Engineering,
and Medical written by a global network of editors and authors.
TA YLOR & FRANCIS EBOOKS OFFERS:
Improved
A streamlined A single point search and
experience for of discovery discovery of
our library for all of our content at both
customers eBook content book and
chapter level
REQUEST A FREE TRIAL

[email protected]

Advanced Mechanical Vibrations - Physics, Mathematics and Applications

Uploaded by

Copyright:

Available Formats

Advanced Mechanical Vibrations - Physics, Mathematics and Applications

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Advanced Mechanical Vibrations - Physics, Mathematics and Applications

Uploaded by

Copyright:

Available Formats

What is the book about?

What is the book about?

What topics are covered in the book?

What topics are covered in the book?

Advanced Mechanical

Paolo Luciano Gatti

and by CRC Press

© 2021 Paolo Luciano Gatti

CRC Press is an imprint of Informa UK Limited

British Library Cataloguing-in-Publication Data

Library of Congress Cataloging‑in‑Publication Data

ISBN: 978-1-138-54228-0 (hbk)

1 A few preliminary fundamentals 1

2 Formulating the equations of motion 13

2.4.5 More co-ordinates than DOFs:

3 Finite DOFs systems: Free vibration 59

4 Finite-DOFs systems: Response to external excitation 99

4.2.1 Excitation due to base displacement,

5 Vibrations of continuous systems 139

5.9.1 Self-adjointness and positive-definiteness

6 Random vibrations 207

A few preliminary fundamentals

1.2 MODELLING VIBRATIONS AND

In order to make sense of the multifarious complexities of real-life physical

application of the excitations f1 , f2 is x1 + x2, where x1 , x2 are the system’s

i. The well-known finite element method – probably the most widely

Third, in studying the response of a system to an external excitation,

1.3 SOME BASIC CONCEPTS

Since the motion of a point particle necessarily occurs in time, it is natural

x(t) = X cos(ω t − θ ) (1.1)

X is the maximum or peak amplitude (in the appropriate units)

Harmonic motion can be conveniently represented by a vector x of mag-

where i is the imaginary unit ( i ≡ −1 ) and e = 2.71828 is the basis of

x(t) = C e i ω t = Xe i (ωt −θ ) (1.4)

where C = Xe − iθ is the complex amplitude, which – as well-known

x(t) = a cos(ω t) + b sin(ω t) = X cos(ω t − θ ) = Ce − iω t = Xe − i (ω t −θ ) (1.5)

1.3.1 The phenomenon of beats

x1(t) + x2 (t) = Xe i ω1 t + Xe i ω 2 t = Xe i (ω1 +ω 2 ) t 2 e i (ω 2 −ω1 ) t 2 + e − i (ω 2 −ω1 ) t 2 

whose real part – defining ω avg = (ω 1 + ω 2 ) 2 and ω = (ω 2 − ω 1 ) 2 – is

2X cos(ω t) cos (ω avg t ) (1.7)

which can be seen as an oscillation of frequency ω avg and a time-dependent

tends to give more weight to high-frequency components. This, in turn, means

thus implying that a typical general-purpose accelerometer with a sen-

L (dB) = 20 log10 ( y /y0 ) (1.9)

where, in vibrations, y can be displacement, velocity or acceleration.

acoustics – universally accepted. It is not always so in the field of vibrations,

1.4 SPRINGS, DAMPERS AND MASSES

Figure 1.2 Ideal massless spring.

Figure 1.3 Ideal massless dashpot.

where c is the coefficient of viscous damping, with units Ns/m.

As well-known from basic physics, in the SI system of units, the mass is

In the first case, in fact, we have F = k1x1 , F = k2 x2 because the force is

total elongation is x = x1 + x2 and the equivalent spring must be such that

∑ (1 ki ) and keq = ∑ i=1 ki ,

• Figure 1.5b: fixed-fixed bar of length L with transverse localised force

In order to describe the motion of a physical object or system, we must spec-

i. In accordance with the knowledge of his time, Newton regarded the

assumption is only an approximation. It is, however, an excellent

2.2 SYSTEMS OF MATERIAL PARTICLES

In most problems, it is convenient to ideally separate a ‘system’ of N mutu-

1.2 MODELLING VIBRATIONS AND

1.3 SOME BASIC CONCEPTS

1.3.1 The phenomenon of beats

1.4 SPRINGS, DAMPERS AND MASSES

Figure 1.2 Ideal massless spring.

Figure 1.3 Ideal massless dashpot.

2.2 SYSTEMS OF MATERIAL PARTICLES

2.2.1 Generalised co-ordinates, constraints

2.3 VIRTUAL WORK AND D’ALEMBERT’S PRINCIPLES –

2.3.1 Hamilton’s equations (HEs)

Figure 2.1 Simple pendulum.

Figure 2.2 Double pendulum.

2.4 ON THE PROPERTIES AND STRUCTURE

2.4.1 Invariance in the form of LEs

2.4.2 The structure of the kinetic energy

ql + fl ( q, q , t ) = 0 (l = 1,… , n ) (2.36)

2.4.3 The energy function and the

2.4.4 Elastic forces, viscous forces and

2.4.5 More co-ordinates than DOFs: