Todorovska MI, Hao T-Y (2003). Information granulation and dimensionality reduction of seismic vibration monitoring data using orthonormal discrete wavelet transform for possible application to data mining, Report CE 03-02, Dept. of Civil Engrg., Univ. of Southern California, Los Angeles, California, pp.171.

UNIVERSITY OF SOUTHERN CALIFORNIA
Department of Civil Engineering
INFORMATION GRANULATION AND DIMENSIONALITY REDUCTION OF

SEISMIC VIBRATION MONITORING DATA USING ORTHONORMAL
DISCRETE WAVELET TRANSFORM FOR POSSIBLE APPLICATION TO
DATA MINING
by
M.I. Todorovska and T.-Y. Hao
Report CE 03-02
December, 2003
Los Angeles, California
www.usc.edu/dept/civil_eng/Earthquake_eng/
ABSTRACT
This report explores the advantages of expansion in orthonormal wavelet bases—as a

preprocessing tool—in analysis of large sets of seismic vibration monitoring data, of ground or
structural response, for possible application in data mining. The focus is on the insight that can
be gained from the wavelet domain representation, convenience in estimation of energy and
correlation, estimation of derivatives, efficiency of representation (data compression property),
and dimensionality reduction. Local and global aggregates and distributions of energy and
related quantities (e.g. power, power spectrum density, Fourier amplitude, cross-energy, cross-
power, cross power spectrum density, etc.), computed directly in the wavelet domain, are
introduced and interpreted as information granules, representative of a frequency interval, of a
partition of the phase plane, or of the entire record. The concepts explored are illustrated on a
mini database of strong motion records from the 1994 Northridge in a 7-story reinforced concrete
building located in the Los Angeles area. Nodal spectra and time-frequency distribution of
energy and power spectrum density (as well as of cross-energy and cross-power spectrum
density) are shown for the ground floor and absolute and relative roof responses. Dimensionality
reduction by thresholding and lower resolution approximation is illustrated and compared with
sub-sampling. The errors associated with compression are illustrates for a small database of
ground response and roof records from six earthquakes recorded in the same building. The
results show that the error is very small even for high compression ratios. It is concluded that
expansion in orthonormal wavelet series is potentially a very useful preprocessing tool in mining
large data sets of ground and structural response vibration data under earthquake excitation. One
drawback of the orthonormal wavelet transform is poor resolution at high frequencies, which can
be eliminated by using expansion in orthonormal wavelet packets, to which all of the presented
concepts are directly applicable, except the multiresolution structure. The theory and concepts
presented in this report apply directly to datasets of any time series data.
i
ii
ACKNOWLEDGEMENTS
This work was partially supported by a 2002/2003 grant from the University of Southern
California Women in Science and Engineering (WISE) Program. Strong motion data for the
earthquakes after 1971 was provided by the California Geological Survey (formerly California
Division of Mines and Geology), either in digital form (for the larger earthquakes) or as photo-
copies (for the smaller earthquakes), which were digitized at USC.
iii
iv
TABLE OF CONTENTS
ABSTRACT..................................................................................................................................... i
ACKNOWLEDGEMENTS........................................................................................................... iii
TABLE OF CONTENTS................................................................................................................ v
1. INTRODUCTION .................................................................................................................... 1
1.3 General Introduction ............................................................................................................ 1
1.2 Motivation for and Objectives of this Work ........................................................................ 2
1.3 Organization of this Report.................................................................................................. 5
2. SIGNAL PROCESSING PRELIMINARIES ........................................................................... 7

2.1 Inner Product and Norm....................................................................................................... 7
2.2 Expansion in a Basis ............................................................................................................ 8
2.3 Continuous-Time Fourier Transform.................................................................................. 9
2.4 Discrete-Time Fourier Transform...................................................................................... 10
2.5 z-Transform........................................................................................................................ 11
2.6 Continuous-Time Fourier Series....................................................................................... 11
2.7 Discrete-Time Fourier Series............................................................................................. 13
2.8 Discrete Fourier Transform................................................................................................ 13
2.9 Sampling of Continuous-time Functions ........................................................................... 14
2.10 Convolution...................................................................................................................... 17
3. DISCRETE WAVELET TRANSFORM AND EXPANSION IN WAVELET BASES ....... 19

3.1 Wavelet Theory and Time-Scale Analysis ........................................................................ 19
3.2 Wavelet Transforms........................................................................................................... 21
2.3 Multiresolution Analysis.................................................................................................... 27
3.4 Subband Decomposition and Its Relation to Wavelet Basis Expansion............................ 29
3.5 Subband Central Frequencies and Bandwidth for Ideal Prototype Filters......................... 32
3.6 Perfect Reconstruction Orthonormal Filter Banks............................................................. 33
3.7 Aliasisng Cancellation ....................................................................................................... 35
3.8 The Pyramid Algorithm ..................................................................................................... 36
3.9 Examples of Orthonormal Wavelet Bases ......................................................................... 39
3.9.1 The sinc Basis ............................................................................................................. 39
2.9.2 The Haar Basis............................................................................................................ 41
3.9.3 Daubechies Compactly Supported Wavelets .............................................................. 42
3.9.4 Properties of Wavelet Families................................................................................... 47
3.10 Parseval’s Equality for the Orthonormal Discrete Wavelet Transform........................... 47
v
4. THE DISCRETE WAVELET TRANSFORM AS A DATA MINING TOOL ..................... 49
4.1 Some Basic Concepts in Data Mining and Knowledge Discovery.................................... 49
4.2 Wavelet Transformed Database......................................................................................... 52
4.3 Reducing Data Dimensionality by Compression............................................................... 55
4.3.1 Data Compression by Lower Resolution Approximation........................................... 58
4.3.2 Data Compression by Thresholding............................................................................ 59
4.4 Estimation of Energy and Power— Global and Local Aggregates and Averages ............ 61
4.4.1 Single Value Estimates ............................................................................................... 61
4.4.2 Nodal Spectra.............................................................................................................. 61
4.4.3 Nodal Time-Frequency Energy and Power Distribution ............................................ 63
4.4.4 Nodal Time Distributions ........................................................................................... 64
4.5 Estimation of Fourier Spectrum and Power Spectrum Density ......................................... 64
4.5.1 Estimation of Fourier Spectrum for Ideal Filters........................................................ 65
4.5.2 Estimation of Fourier Spectrum for Non-Ideal Filters................................................ 66
4.5.3 Estimation of Power Spectrum Density...................................................................... 67
4.6 Estimation of Correlation — Global and Local Aggregates and Averages....................... 70
4.6.1 Single Value Estimates ............................................................................................... 72
4.6.2 Nodal Cross-Energy Spectrum ................................................................................... 72
4.6.3 Nodal Time-Frequency Distribution of Cross-Correlation......................................... 74
4.6.4 Nodal Cross-Power Spectrum Density ....................................................................... 75
4.7 Estimation of Time Derivatives ........................................................................................ 76
4.8 Estimation of Continuous-Time Signals............................................................................ 79
4.8.1 Energy, Power and Root Mean Square Value ............................................................ 79
4.8.2 Fourier Transform....................................................................................................... 81
4.8.3 Power Spectrum Density............................................................................................. 81
4.8.4 Correlation, Cross-Power, and Power Cross-Spectrum Density ................................ 82
5. ILLUSTRATIONS OF CONCEPTS...................................................................................... 83
5.1 Brief Description of the Building and a Simplified Model................................................ 83
5.2 DWT Signature and Multiresolution Decomposition of Acceleration .............................. 92
5.2.1 Absolute Acceleration Responses............................................................................... 93
5.2.2 Relative Roof Response........................................................................................... 100
5.3 Multiresolution Approximations of Acceleration and Reduction of Dimensionality...... 104
5.4 Reduction of Dimensionality by Shrinkage of the Smallest Coefficients ....................... 104
5.5 Energy, Power, Nodal Energy Spectra and Average Fourier Spectra of Signals ............ 114
5.6 Estimation of Spectra of Integrals and Derivatives of Signals ........................................ 119
5.7 Energy and PSD Time-Frequency Distributions ............................................................. 122
5.8 Estimation of Cross-Correlation of Motions across the Building Height - Global and
Subband Aggregates and Averages ........................................................................................ 130
vi
5.9 Estimation of Cross-Correlation of Motions across the Building Length - Global and
Subband Aggregates and Averages ........................................................................................ 136
5.10 Time-Frequency Distribution of Correlation across the Building Height ..................... 141
5.11 Time-Frequency Distribution of Correlation across the Building Length..................... 145
6. ANALYSIS OF COMPRESSION ERROR FOR A MINI DATABASE ............................ 150
7. DISCUSSION, SUMMARY AND CONCLUSIONS.......................................................... 162

7.1 Discussion ........................................................................................................................ 162
7.2 Summary and Conclusions .............................................................................................. 165
REFERENCES ........................................................................................................................... 167
vii
viii
1. INTRODUCTION
1.3 General Introduction
Since first emerged as a consistent theory in the 1980s from the work of French
geophysicists Morlet and Grossman (Morlet et al., 1982a,b; Goupillaud et al., 1984/85;
Grossmann and Morlet, 1984), wavelet analysis has become a very popular tool for analysis of
signals and images in many fields of science and engineering. The wavelet transform is
particularly suitable for analysis of transient signals and of time varying systems, because it is
localized both in time and frequency. Its widespread use is also due to the existence of
orthogonal and bi-orthogonal wavelet bases, the efficiency of representing transient signals in
such bases (data compression), and the existence of fast and accurate computational algorithms
for signal/image transformation and reconstruction (asymptotically even faster than the fast
Fourier transform). The major milestones in the development of wavelet theory, as well as its
origins (before it emerged in a full grown theory), will be discussed in Chapter 3.
The most wide spread use of wavelets is in digital signal and image processing, in particular
for compression and subband coding. Wavelet analysis is also used in artificial intelligence, for
edge detection and defining contours of objects in bitmap images (from larger scale
reconstructions from wavelet bases expansion), and for pattern recognition (Tang et al., 1999).
Wavelets also enable selective noise removal, and are used in statistics for nonparametric
function estimation, and for detection of trends and surprises in data (Antoniades and
Oppenheim, 1995). Wavelets are a convenient tool for exploratory visual data analysis, and are
used extensively in analyses of medical signals. As a diagnostics tool, wavelets can be used to
detect and localize various singularities (discontinuities) in a signal. The intuitive visual
observations can be formalized into algorithms that can be applied automatically to a large set of
time series. Wavelets have also bee used in solving general partial differential equations, e.g. by
Amaratunga and Williams (1993), and Williams and Amaratunga (1995).
The use of wavelets in analyses of mechanical vibrations was first introduced by Newland
(1993a,b; 1994a,b). He proposed the use of wavelet maps (time-frequency distributions) as a
diagnostics tool for time-varying systems, in particular in detecting and localizing in time hidden
details and small perturbations that are practically invisible in the time representation, as well as
in providing insight into local correlation of two signals. He also introduced the use of the
harmonic wavelet, which is a complex extension of the Shannon wavelets. The use of the
wavelet transform to detect faults in machines and cracks in a structure has been proposed and
demonstrated for numerically simulated data, e.g. by Staszewski and Tomlinson (1994), Deng
and Wang (1998), Wang and Deng (1999), Hou et al. (2000), and Gentille and Messina (2003),
and its use for compression and feature selection in machine vibrations has been proposed by
Staszewski (1998a). The continuous wavelet transform has been applied to determination of
1
instantaneous frequency, e.g. by Staszewski (1998b) and Todorovska (2001). In structural
vibration, expansion in wavelet bases has been use to transform the equation of motion and the
exciting force, by Basu and Gupta (1997a,b; 1999) in stochastic analysis of response of linear
systems subjected to seismic excitation, and by Ghanem and Romeo (2000) and Pettit et al.
(2000) in parametric identification of nonlinear and linear time varying dynamical systems.
Wavelets have also been used in representation of random fields by Zeldin and Spanos (1996),
and Spanos and Rao (2001), and in earthquake engineering for representation of the earthquake
energy input into structures by Iyama and Kuwamura (1999). Wavelet analysis is slowly but
steadily gaining popularity in earthquake engineering (Ventura and Rezai, 1999; Gurley and
Kareem, 1999; Mukherjee and Gupta, 2002; Pan and Lee, 2002; Sun and Chag, 2002; Ching and
Glaser, 2003).
1.2 Motivation for and Objectives of this Work
The recent advances in sensor, computer, and communication technologies have enabled and
encouraged collection of large volumes of scientific data that needs to be efficiently stored,
managed and analyzed. These needs have stimulated many advances in the fields of data
engineering and digital signal and image processing, which have enabled respectively efficient
organization, retrieval and classification of diverse types of data, and efficient storage, transfer,
and analysis. In strong motion seismology, the volume of data is increasing rapidly not as much
due to the number of new sensors deployed (e.g., Kyoslin-net, Kinoshita, 1998; Taiwan Strong
Motion Network, Lee et al., 2001; and Southern California Integrated Seismic Network,
Haukson et al., 2001) as due to the increasing sensor sensitivity and recorder dynamic range
(currently approaching 26 bits or 156 dB), which makes it possible to record, with strong motion
instruments, ground and structural response to very small and distant earthquakes as well as to
ambient noise (Trifunac and Todorovska, 2001).
The increasing capability of recording, combined with the dramatic reduction of the cost of
digital storage media, lead to lowering the triggering level of recording and even to continuous
recording by a separate channel at some ground stations (Hauksson et al., 2001). As a part of the
Advanced National Seismic System (ANSS) initiative launched by the U.S. Geological Survey
(Benz et al., 2001; COSMOS, 2001), a large number of instruments will be installed in structures
as well, which will further increase the number of recordings. Potential benefits of continuous
recording of vibration of soil-structure systems have already been recognized, and it is likely to
see it implemented in the near future. One important benefit is better understanding of the soil-
structure interaction, in particular, of the degree to which the changes in the soil due to
environmental influences affect the system response. Such environmental influences are, e.g., the
water content in the soil, which depends on the weather, strong and weak shaking from
earthquakes and their aftershocks, and which affects the degree of consolidation of the soil.
2
Another important benefit is development and refinement of methods for structural health
monitoring and detection of structural deterioration and damage. Vibration monitoring data is
normally recorded at a sampling rate of 100 per second. Use of the data for studying wave
propagation within a structure (Todorovska et al, 2001) or within a small array of close by
stations would likely require recording at a sampling rate higher than 100 per second.
Continuous recording in structures and higher sampling rate will further augment the size of the
strong motion archives.
The traditional way of archiving seismic and other mechanical vibration data is by storing
the waveform data in the time domain (e.g., Hefner and Clayton, 2001). However, retrieval of
such time series data that is of high dimension (i.e. consisting of many data points) from
permanent archives on a remote server is slow, especially when, for the specific application,
lowering the dimensionality of the data (e.g. by reducing the resolution of the representation, by
data compression, or by representing values in intervals by local averages) is permissible or even
desirable (e.g. in pattern recognition).
A novel approach to archive time series or spatial data would be to expand the data in
wavelet series and archive the coefficients of the expansion. Its possible advantages are as
follows.
1. A wavelet-transformed database contains already preprocessed data, enabling an

insight that cannot be gained from the time domain representation without further
processing. One important feature of the wavelet transform is its multi-resolution
hierarchical structure, which is convenient for studying and extracting features of very
different temporal scales (e.g. sharp changes in the system due to damage, versus long
term changes, such as daily or seasonal changes due to environmental influences), or of
very different spatial scales (e.g. in analysis of data recorded by an array of sensors).
Another important feature of the wavelet transform is that it is a time-frequency
distribution (nodal in the case of the discrete wavelet transform), and hence enables
tracking changes in a system simultaneously in frequency and in time, or only in
frequency, after computing the corresponding marginal distribution.
2. The original time series representation of the data, if needed, can be reconstructed very
fast from the transform. For a large class of wavelets—compactly supported wavelets,
an algorithm for expansion and reconstruction exists—the pyramid algorithm (of
complexity O ( N ) ), which enables perfect reconstruction (up to the machine precision),
and which is asymptotically even faster than the fast Fourier transform (FFT) algorithm
(of complexity O ( N log 2 N ) ). For all other wavelets, the forward and inverse
transform can be performed also very fast in the Fourier transform domain using FFT.
3
3. Lower resolution approximations of the data, required for quick browsing and
visualization, or for studying large scale features, can be reconstructed on a remote
server by retrieval from the database and transfer over the network only of a small
fraction of the coefficients.
4. Lower dimensional approximations of the time series data, such that preserve the high
resolution features that are significant, can also be created fast on a remote server by
retrieval from the database and transfer over the network only of a fraction of the
coefficients stored. This fact is due to the data compression property of the discrete
wavelet transform, which is particularly significant for transient data. This property
refers to the ability to represent a large fraction of the energy of the signal by only a
small fraction of the wavelet coefficients. As it will be shown later in this report, for
strong motion data, high compression rates can be achieved with very small loss of
accuracy (e.g. 3% of the wavelet coefficients, or 67 real numbers of the original 2048,
account for 90% of the total energy of the signal). The data compression property is
particularly significant for storage of spatial data (images).
5. A wavelet-transformed database is the natural means of storing compressed data. While
data from rare events should always be stored without any loss of information, data
compression may be an attractive future solution for archiving very small amplitude
motions, such as ground and structural responses to ambient noise or to teleseismic
events), optimizing the costs and benefits of maintaining archives of such.
6. Data mining of a wavelet-transformed database can be done directly at database level,
without having to transfer large amounts of data from a remote server (Shahabi et al.,
2000, 2001a,b).
7. As there is no perfect wavelet for all applications to construct an all purpose permanent
wavelet transformed database, a working (possibly compact) wavelet transformed
database can be constructed on the remote server where the original data has been
stored, and used for data mining and feature extraction, also on the remote server, thus
avoiding transfer over the network of large amounts of data.
8. This vision applies equally to large databases of experimental data.
Data mining a wavelet-transformed database of strong and weak ground motion or structural
response caused by earthquakes, actuator forced or ambient noise, as well as of similar type of
data created by laboratory experimental, is a novel concept not yet explored. The first step to be
taken in pursuing this vision further, and ultimately implementing it practically for a wide-spread
use by the professional community, would be to demonstrate the usefulness of the insight gained
in the wavelet transform domain (especially for the discrete transform), in view of the fact that
the use of the wavelet transform in earthquake engineering is still in its infancy. The second step
4
would be to lay down the foundations for the use of the discrete wavelet transform as a data-
mining tool. This report addresses both steps.
The objectives of this report are to:
(a) Review the theory of the discrete wavelet transform and expansion in orthonormal
wavelet series, viewed as a special case of wavelet frames, and its relation to subband
decomposition. It is intended that this review is both rigorous as well as accessible even
for uninitiated readers from the earthquake engineering and from the engineering
mechanics communities.
(b) Present an interpretation of the discrete wavelet transform and related nodal marginal
distributions as granules of information of general time series data, for possible use in
feature extraction and in data mining of large sets of time series data. Here, information
granulation, refers to reducing the dimensionality of the data by encapsulation of
numeric data, contained e.g. in an interval, into a single conceptual entity (Cios et al.,
1998). Of particular interest to the authors of this report is the use in data mining of
large data sets of strong and weak ground motion or structural vibration data caused by
earthquakes, actuator forced or ambient noise, as well as of similar type of data created
by laboratory experiments. To the knowledge of the authors of this report, this
interpretation is original, especially for mechanical vibration data. Data mining refers to
extracting information and knowledge discovery from large sets of data, usually done at
database level.
(c) Illustrate the introduced granules of information and the insight they provide on a mini
database of strong earthquake motion records in a building in the Los Angeles area
recorded during the Northridge, California, earthquake of January 17, 1994.
(d) Illustrate on a mini database of strong earthquake motion records in the same building
during a series of earthquakes, the data compression property of the orthonormal discrete
wavelet transform and its efficiency to represent compactly strong motion records.
(e) Critically examine the usefulness and shortcomings of information granulation based on
orthonormal wavelet series expansion, and suggest possible extensions.
1.3 Organization of this Report
This report is organized as follows. Chapter 2 presents a review of some elementary

concepts in functional analysis and digital signal processing necessary to follow the material
presented in Chapters 3 and 4. Chapter 3 presents a review of the theory of the wavelet
transform and subband decomposition. Chapter 4 presented a review of the basic concepts in
data mining, and the theoretical basis for the interpretation of the discrete wavelet transform as a
tool for dimensionality reduction and information granulation. The methods considered for
5
dimensionality reduction are compression by thresholding (soft or hard), or by construction of a
lower resolution approximation. The information granules considered are local and global
aggregates and averages of related quantities, such as e.g. energy, power, spectral density, and
correlation. In this context, nodal spectra and spectral densities are introduced, as well as nodal
instantaneous spectra. Chapter 5 presents illustrations of the concepts introduced in Chapter 4
using strong motion data recorded in a 7-story reinforced concrete building in the city of Van
Nuys in the Los Angeles metropolitan area during the Northridge, California earthquake of
January 17, 1994. Chapter 6 presents results for various measures of error as a function of the
compression ratio for a variety of records recorded in the same building during several
earthquakes, and a discussion of how the nature of the earthquake shaking influences the
compression error. Chapter 7 presents a summary and the conclusions.
6
2. SIGNAL PROCESSING PRELIMINARIES
This section reviews the definitions of some basic mathematical concepts from vector
spaces, such as inner product, norm and expansion in a basis, as well as the Fourier transform
pairs and the Parseval’s equality for the Fourier transform, which is one of the fundamental
relations in signal processing. The reason for including this in a separate chapter is to avoid
repetition in the following chapters, and to clarify up front the conventions followed in this
report. The latter particularly applies to the Fourier Transform pairs for which there are different
definitions used in literature, which differ only by a constant, but are a source of confusion when
formulae and results from different publications are being compared. In this report, we follow
the convention used in Vetteri and Kovacević (1995). More detailed review of signal processing,
as related to the theory of wavelets, can be found in any of the many books on wavelets (e.g.,
Vetteri and Kovacević, 1995; and Strang and Nguyen, 1997).
2.1 Inner Product and Norm
Let f ( t ) , t ∈ be a real or complex valued function of a Hilbert space of square

integrable functions, and let f [ n] , n ∈ be a real or complex valued sequence also of a
Hilbert space of square integrable sequences, where is the real line and is the set of
integers, and a Hilbert space is a complete inner product vector space. Let us first define inner
product and norm in these Hilbert spaces.
The inner product of two functions f ( t ) and g ( t ) is by definition
∞
f ,g ∫ f ( t ) g ( t ) dt
−∞
(2.1.1)
and the L2 norm of a function, ⋅ L2

, consistent with this inner product definition is
1/ 2
∞ 
f , f =  ∫ f ( t ) dt 
2
f L2
= (2.1.2)
 −∞ 
Similarly, the inner product of two sequences f [ n] and g [ n ] is by definition
∞
f ,g ∑ f [ n] g [ n]
n =−∞
(2.1.3)
and the L2 norm of a sequence, consistent with this inner product is
7
1/ 2
 ∞ 2
f L2
= f, f  ∑ f [ n]  (2.1.4)
 n =−∞ 
In equations (2.1.1) and (2.1.3), the bar indicates complex conjugate. From now on, for
simplicity, the subscript referring to L2 will be omitted, and unless otherwise specified, ⋅ and
“norm” will refer to the L2 norm.
2.2 Expansion in a Basis
Let { xi }i∈ be a set of orthonormal vectors in a Hilbert space, e.g. the space of square
integrable functions defined on the real line or of square integrable sequences defined on the set
of integers, representing a basis. Then every vector f of the Hilbert space can be represented
uniquely as a linear combination of the basis vectors
f = ∑ α i xi (2.2.1)
i
with the coefficients of the expansion α i —the Fourier coefficients—being the inner product of
the basis vectors and the function
α i = xi , f (2.2.2)
Then the norm of the function can also be computed directly from the Fourier coefficients
f = ∑ αi
2
i
(2.2.3)
= ∑ xi , f
2
as well as the inner product of two vectors f and g of the Hilbert space, having respectively
Fourier coefficients α i and βi ,
f , g = ∑ α i βi
i
(2.2.4)
= ∑ f , xi xi , g
i
which is known as the Generalized Parseval’s equality.
8
2.3 Continuous-Time Fourier Transform
The Fourier Transform of a function f ( t ) , t ∈ is defined as
∞
fˆ ( Ω ) ∫ f (t ) e
− i Ωt
dt , Ω ∈
−∞ (2.3.1)
iΩt
= e ,f
This transform contains all the information about the function, and the function can be recovered
from its transform by applying the inverse transform formula
∞
1
f (t ) ∫ fˆ ( Ω ) e
iΩt
dΩ (2.3.2)
2π −∞
where Ω is circular frequency in rad/s if t is time in seconds. Here we use the symbol Ω
following the convention in Oppenheim and Shaffer (1999), and will use the symbol ω for the
circular frequency in rad/sample, to avoid confusion between these two frequencies. The first
equation is known as the analysis formula and the second one—as the synthesis formula. A
condition that the transform exists is that f ( t ) is square integrable, i.e.
∫ f (t )
2
dt < ∞ (2.3.3)
−∞
and the transform pair holds in the sense that f ( t ) and its reconstruction from its Fourier
Transform, f rec ( t ) , are equal to each other in the L2 sense, i.e. in the sense that the norm of the
error is zero
f ( t ) − f rec ( t ) = 0 (2.3.4)
while the two functions are not necessarily equal at every point t in the domain. This is
particularly the case for points of discontinuity at which the reconstructed version equals the
average of the left and right limits of the function at the discontinuity. Hence, for f ( t )
piecewise continuous, with a finite number of discontinuities, one can say that f ( t ) and f rec ( t )
are equal to each other almost everywhere, i.e. everywhere except at the points of discontinuity,
when at these points the value of the function is not the average of its left and right limits. The
Fourier Transform is a linear transform and is self-dual.
9
The Fourier Transform is an orthogonal transform and hence energy conserving, which is
formally described by the Parseval’s equality. For two continuous-time functions f ( t ) and
g ( t ) , the Parseval’s equality is
∞ ∞
1
f,g = ∫ f ( t ) g ( t ) dt = ∫ fˆ ( Ω ) gˆ ( Ω ) d Ω (2.3.5)
−∞
2π −∞
and when f ( t ) = g ( t )
∞ ∞
1 2
f ( t ) dt = fˆ ( Ω ) d Ω
2
∫ ∫
2
f = (2.3.6)
−∞
2π −∞
2.4 Discrete-Time Fourier Transform
The Fourier Transform pair for a sequence { f [ n]}n∈] is by definition
∞
fˆ ( eiω ) ∑ f [ n] e − iω n
n =−∞ (2.4.1)
iω n
= e , fn
π
1
f [ n] ∫π fˆ ( e ) e dω
ω ω i i n
(2.4.2)
2π −
where ω is circular frequency in rad/sample. The transform exists, i.e. the series in the right-
hand side of eqn. (2.4.1) converges, if the sequence { f [ n]}n∈] is square summable, in which case
the convergence is in the mean square sense and not necessarily uniform (for the partial sums of
the series in the right-hand-side of eqn (2.4.1), the average energy of the error goes to zero as
n → ∞ ). If { f [ n]} n∈]
is also absolutely summable, then the convergence is uniform (the
absolute value of the error at every point decreases uniformly as n → ∞ ). The Discrete-time
Fourier Transform is also linear. An important property of the Discrete-Time Fourier Transform
is that it is a periodic function of ω , with period 2π , in contrast to the Continuous-Time Fourier
Transform. This implies that the largest frequency a discrete sequence can have is ω = π . This
is particularly important in sampling continuous time functions and using the sampled version to
represent the continuous-time function. The discrete-time Fourier transform is dual of the
Fourier series expansion, reviewed in the proceeding text.
10
For two discrete-time sequences { f [ n]}n∈ and { g [ n]}n∈ , the Parseval’s equality is
∞ π
1
∑ f [ n] g [ n] = ∫π fˆ ( e ) gˆ ( e ) dω
ω ω
f,g = i i
(2.4.3)
n =−∞ 2π −
and when f [ n] = g [ n] , n ∈
∞ π
1
fˆ ( eiω ) dω
2
∑ f [ n]
2
∫
2
f = = (2.4.4)
n =−∞ 2π −π
2.5 z-Transform
While the Fourier transform is defined on the real line, the z-transform is defined on the
complex plane, and can be considered a generalization of the discrete-time Fourier transform.
For a sequence { f [ n]}n∈ , the z-transform is by definition
∞
F (z) = ∑ f [n] z
n =−∞
−n
, z∈ (2.5.1)
where is the complex plane. The domain where F ( z ) exists is its region of convergence
(ROC). The inverse transform involves contour integration in the complex plane. The discrete-
time Fourier transform is a special case of the z-transform when z is on the unit circle, i.e.
z = eiω ⇔ z = 1 , and their relationship is
fˆ ( eiω ) = F ( z ) z =eiω (2.5.2)
provided that the unit circle is inside the region of convergence of F ( z ) . The z-transform is
extensively used in discrete-time signal processing. One advantage of the z-transform is that it
may exist on some domain in the complex plane for signals for which the Fourier transform does
not exist (in this case the ROC excludes the unit circle).
2.6 Continuous-Time Fourier Series
This is the expansion of a periodic function f ( t ) , of period T , in the basis of complex

2π
, k ∈ , orthogonal on the interval [ −T / 2, T / 2] , which has the form
ik t
exponentials e T
11
∞ 2π
− in
∑ F [ n] e
t
f (t ) T
, t∈ (2.6.1)
n =−∞
where the Fourier coefficients are
T 2π
1
F [ n]
in
( ) T dt , n ∈
T −∫T
f t e (2.6.2)
The equality in eqn (2.6.1) is true in the L2 sense, and not necessarily point-wise. For the series
to converge, f ( t ) needs to be square integrable, and the convergence is in the mean square
sense, and not necessarily uniform. While the norm of the error between the function and the
partial sums is zero in the limit when infinitely many terms are considered, the error is not
necessarily zero at each point t . If f ( t ) is continuous, the convergence is uniform. If it is
piecewise continuous, with finite number of discontinuities, then the convergence is uniform
inside the intervals of continuity, and at the points of discontinuity, the series converges to the
average of the left and right side limits. Hence, the series is equal to the function almost
everywhere. The continuous-time Fourier series are the dual of the discrete-time Fourier
transform, as for T = π equations (2.6.1) and (2.6.2) are equivalent to eqns (2.4.1) and (2.4.2). It
should be noted that functions that are not periodic but are defined on a finite domain can also be
expanded in Fourier series (with an arbitrary period grater than the length of the domain), if they
are first extended periodically on the entire real line, and then their periodic extension is
expanded in Fourier series. The reconstruction formula, however, should be restricted to the
original domain of the function. Finally, the Parseval’s equality for two functions f ( t ) and
g ( t ) having Fourier series coefficients F [ n] and G [ n] is
T /2 ∞
f ,g  T T
− 2 , 2 
 
= ∫ f ( t ) g ( t ) dt = T ∑ F [ k ]G [ k ]
n =−∞
(2.6.3)
−T / 2
and for f ( t ) = g ( t )
T /2 ∞
f ( t ) dt = T ∑ F [k ]
2
∫
2 2
f  T T = (2.6.4)
− 2 , 2 
  −T / 2 n =−∞
12
2.7 Discrete-Time Fourier Series
Periodic sequences can also be expanded in Fourier series of orthogonal complex

exponential sequences, as follows. Let { f [ n]}n=0 be a periodic sequence with period N . Then
N −1
its Fourier series expansion and Fourier coefficients are
N −1 2π
1 i nk
f [ n] = ∑ F [k ] e N
, n∈ (2.7.1)
N k =0
N −1 2π
−i nk
F [ k ] = ∑ f [ n] e N
, k∈ (2.7.2)
n =0
2π
i nk
Here, the basis vectors e N
are also periodic sequences with period N , and there are only N
{ f [ n]}
N −1
distinct basis vectors. The Parseval’s equality for two periodic sequences n=0
and
{ g [ n]}
N −1
n =0
both with period N and having coefficients F [ k ] and G [ k ] of their discrete-time
Fourier series is
N −1 N −1
1
f,g [0, N −1]
= ∑ f [ n] g [ n] = ∑ F [k ]G [k ] (2.7.3)
n =0 N k =0
and when f [ n ] = g [ n ]
N −1 N −1
1
= ∑ f [ n] = ∑ F [k ]
2 2 2
f [0, N −1]
(2.7.4)
n =0 N k =0
The symmetry of eqns (2.7.1) and (2.7.2) imply that the discrete-time Fourier series is a self-
dual.
2.8 Discrete Fourier Transform
This transform is a generalization of the discrete-time Fourier series so that they can apply to
{ f [ n]}
M −1
a non-periodic but finite sequence. Let n =0
be such a sequence of length M . Then this
sequence can be expanded in discrete-time Fourier series of period N such that N ≥ M , first by
padding zeros to achieve length N , then by periodically extending the padded sequence on the
entire set of integers, and finally by expanding in discrete time Fourier series the periodic
extension. The analysis and synthesis equations are same as for the discrete time Fourier series
13
expansion of truly periodic sequences, given by eqns (2.7.1) and (2.7.2), except that the synthesis
equation is restricted to n ∈ [ 0, N − 1] . This transform is very important for numeric
computations, because for N equal to a power of 2, it can be computed by the Fast Fourier
Transform (FFT) algorithm.
Table 2.8.1 summarizes the different forms of Fourier transform, along with the
characteristics of the function in the time domain and in the transformed domain.
Table 2.8.1 Summary of Fourier Transforms
Transform Condition on f Characteristics of Dual

fˆ
(Continuous time) Fourier transform (CTFT) continuous continuous Self-dual
Discrete time Fourier transform (DTFT) discrete continuous and CTFS

periodic
(Continuous time) Fourier series (CTFS) continuous and discrete DTFT

periodic
Discrete time Fourier series (DTFS) discrete and discrete and self-dual
periodic periodic
2.9 Sampling of Continuous-time Functions
Although digital computers can perform some symbolic operations on continuous functions,
numerical computations using digital computers can be performed only on finite discrete
sequences. Hence, continuous-time functions need to be discretized, and infinite sequences and
sums need to be truncated. This implies that the results need to be interpreted with caution. This
particularly applies to the Fourier Transform, as it has quite different nature for continuous-time
functions than for sequences. For example, the Fourier transform of continuous time signal is
non-periodic and is defined, in general, on the entire range of frequencies, while the Fourier
transform of sequences is periodic and there is a limit on the maximum frequency of the signal.
This section will relate some fundamental concepts defined for continuous-time signals with the
related concepts for its sampled version, which are the ones that are computed numerically.
Let f [ n] , n ∈ be a sampled version of f ( t ) , t ∈ at equally spaced intervals ∆t .

Then f [ n] and f ( t ) are related by
14
f [ n] = f ( n ⋅ ∆t ) , n ∈ (2.9.1)
and the frequencies ω in rad/sample is related to the frequency Ω in rad/s by
ω
Ω= (2.9.2)
∆t
We saw earlier that the largest frequency a sequence can represent is ω = π , which correspond
to frequency
π
Ω max = = Nyquist frequency = Ωsampling 2 (2.9.3)
∆t
of the continuous time signal, f ( t ) , where
Ωsampling = 2π ∆t (2.9.4)
is the sampling frequency. On the contrary, for continuous-time signal there is no upper bound
for the largest frequency. Another major difference between the Fourier transform of
continuous-time signals and that of their sampled versions is that the latter is periodic with
period ω = 2π , which corresponds to period Ω = ω ∆t = 2π ∆t = Ωsampling . An important
consequence of this is that the entire energy of the continuous time signal is mapped into the
interval ω ∈ [ −π , π ] , which corresponds to Ω ∈  −Ωsampling 2 , Ωsampling 2  =  −Ω Nyquist , Ω Nyquist  ,
which would make it impossible to extract the Fourier transform of f ( t ) from that of its
sampled version. This is formally described by the following relationship between the two
Fourier transforms
1 ∞  2π 
fˆ∆t ( Ω ) = ∑
∆t k =−∞
fˆ  Ω − k


∆t 
(2.9.5)
where
∞
f ∆t ( t ) = ∑ f ( k ∆t ) δ ( t − k ∆t )
k =−∞
(2.9.6)
is the sampled version of f ( t ) and δ ( t ) is the Dirac delta-function (the sampled version is a
comb of delta-functions, with weights equal to the values at the sample points. Equation (2.9.5)
states that, if the frequency axis is divided into intervals of length Ωsampling , e.g. in such away that
15
the first interval is  −Ωsampling 2 , Ωsampling 2  , then the Fourier transform of the sampled version
on the interval  −Ωsampling 2 , Ωsampling 2  is obtained by shifting the Fourier transform defined
on the intervals so that they overlap with  −Ωsampling 2 , Ωsampling 2  , and adding the transforms
for all the intervals, and then mapping the result periodically on the entire frequency axis. This
phenomenon is called aliasing. From eqn (2.9.5) it follows that the only way to make a one-to-
one correspondence between the Fourier transform of the f ( t ) and that of its sampled version is
to leave only the first term of the sum on the right-hand-side of the equation, which can be done
by setting to zero the Fourier transform of f ( t ) outside he interval  −Ωsampling 2 , Ωsampling 2  ,
i.e. by low-pass filtering it below its Nyquist frequency i.e. so that the maximum frequency in the
signal is Ω max ≤ Ω Nyquist = Ωsampling 2 = 2π ∆t , in which case
π
fˆ ( Ω ) = ∆t fˆ∆t ( Ω ) = fˆ ( ei Ω ∆t ) , Ω ≤ Ω Nyquist = (2.9.7)
∆t
If the physical application imposes a restrictions on the maximum frequency of the signal, which
has to be set to some value Ω max , then the sampling interval should be chosen sufficiently small,
i.e. so that
π
∆t ≤ ∆tcrit = (2.9.8)
Ω max
There is another important consequence of the one-to-one correspondence between the

Fourier transform of the continuous-time signal and that of its sampled version, which is that, for
band-pass signals, there is no loss of information if they are appropriately sampled. This means
that the continuous-time signal can be exactly recovered from its samples, which is formally
stated by the Nyquist sampling theorem (Oppenheim and Schafer, 1999). The reconstruction
formula (sue to Shannon, Strang and Nguyen, 1997) is
∞
f (t ) = ∑ f ( n ∆t ) s inc ( t − n ∆t )
n =−∞
∆t (2.9.9)
where
sin (π t / ∆t )
s inc ∆t ( t ) = (2.9.10)
π t / ∆t
and its shifts serve as interpolation functions, interpolating the signal between the samples.
16
2.10 Convolution
This operator is reviewed here because it reveals the relationship between coefficients of
series expansion, defined as inner products, and filters, which is necessary for the theory of
wavelets presented in the subsequent sections. In discrete time, the convolution of two
sequences x [ n ] and h [ n ] , n ∈ is a sequence y [ n ] such that
∞ ∞
y [ n] = x [ n] ∗ y [ n] = ∑ x [k ] h[n − k ] = ∑ x [ n − k ] h [ k ], n∈ (2.10.1)
k =−∞ k =−∞
or in the frequency domain
yˆ ( eiω ) = xˆ ( eiω ) hˆ ( eiω ) (2.10.2)
The convolution operator is often used in linear systems theory to describe the action of a linear
system on an input signal, and x [ n ] can be thought of as the input signal, h [ n ] —as the impulse
response function of the system, and hˆ ( eiω ) — as the transfer-function of the system.
Without loss of generality, let us assume the two sequences are real. Then Equation (2.10.1)
can also be written as
∞ ∞
y [ n] = ∑ x [k ] h[n − k ] = ∑ x [k ] h [k ] = h, x (2.10.3)
k =−∞ k =−∞
where
h [k ] = h [n − k ] (2.10.4)
i.e. for each n , y [ n ] is equal to the inner product of x [ n ] with a time reversed and shifted
version of h [ n ] . Equation (2.10.3) can be written in matrix form as
y=H x (2.10.5)
where H is a matrix with entries
H [ n, k ] = h [ n − k ] (2.10.6)
17
Vice versa, an inner product of two real sequences can be represented as a convolution of one of
the signals with an operator with impulse response which is a time reversed and shifted version
of the signal. Hence, computing Fourier coefficients of an orthonormal series expansion of a
signal is equivalent to passing the signal through an appropriate filter.
18
3. DISCRETE WAVELET TRANSFORM AND EXPANSION IN WAVELET BASES
This chapter reviews the general theory behind the discrete wavelet transform and expansion
in orthogonal wavelet bases, which are of concern in this report, while Chapter 4 deals with the
theory behind the application of orthonormal wavelet bases as a tool for dimensionality reduction
and information granulation of time series signals. In this chapter, the discrete wavelet transform
is introduced as a special case of the continuous wavelet transform, consisting of its samples on a
particular discrete time-scale grid, and orthonormal wavelet bases are introduced as a special
case of wavelet frames, which are set of functions, usually over-determined, that can be used to
transform and reconstruct back a function. This is followed by the review of: multi-resolution
analysis—the general framework for construction of wavelet bases; the connection between
wavelets and subband decomposition, including perfect reconstruction filter banks, quadrature
mirror filters and the pyramid algorithm. Finally, three examples of orthonormal wavelet bases
are presented—of Haar and sinc wavelets and of the s8 compactly supported wavelets, and the
Parseval’s equality for the orthonormal discrete wavelet bases expansion, much of the material
presented in Chapter 4 is based on. Main sources for this review are Vetterli and Kovacević
(1995), Bruce and Gao (1996), and Daubechies (1992).
3.1 Wavelet Theory and Time-Scale Analysis
The modern wavelet theory emerged as a coherent and full-grown theory in the early and
mid 1980s from the work of French exploration geophysicists Grossman and Morlet (Morlet et
al., 1982a,b; Grossman and Morlet, 1984; Goupillaud et al., 1984/85), and was motivated by the
need for high resolution methods for analysis of seismic reflection signals. The origins of the
wavelet theory, however, appear much earlier, in separate efforts of several mathematicians, e.g.,
of Haar in 1910, of Littlewood and Palley in the 1930s, of Lusin in 1930s, and of Calderon in
1960s (Meyer, 1992). A very good review of the early historical developments and rediscoveries
of the 1980s can be found in the published lectures on wavelet analysis by Meyer (1992), and by
Daubechies (1992). A major milestone in the development of the modern wavelet theory is the
work of Mallat who saw the connection between wavelet theory and subband decomposition,
established a general theoretical framework for construction of wavelet bases—multiresolution
analysis, and proposed a very fast algorithm for computation of wavelet expansion and
synthesis—the pyramid algorithm (Mallat, 1989). Another major milestone, which triggered
numerous applications, is the work of Daubechies who, building on the framework proposed by
Mallat, proposed design of wavelet bases with compact support and pre-selected regularity
properties (Daubechies, 1988). These developments will be discussed in more detail as they
appear in the review of wavelet theory in the remaining part of this chapter.
19
According to the modern wavelet theory, a wavelet is a real or complex zero mean wiggle
on the real line that is localized both in time and in frequency. For the zero mean condition (also
called admissibility condition) to be satisfied, it must be oscillatory—hence the name wavelet
(Vetterli and Kovacević, 1995; Daubechies, 1992). By elementary operations consisting of shifts
in time and change of scale (i.e. dilation or contraction), a family of wavelets
1 t −b 
ψ a,b (t ) = ψ  , a > 0, b ∈ \, t ∈ \ (3.1.1)
a  a 
can be constructed from a prototype wavelet, ψ (t ) ∈ L2 ( R) (called the “mother wavelet”). Here a
is the scale variable and b is the time shift. The wavelet is dilated if a > 1 and it is contracted
for a < 1 . The normalizing constant 1/ a is such that all wavelets have same L2 norm, usually
set to unity
∞ 2
ψ L2
= ∫ ψ a ,b ( t ) dt = 1 (3.1.2)
−∞
An important property of wavelets in a family is their localization, described by their central

time and central frequency, and spread in time and in frequency. If the central time and
frequency of the mother wavelet are ω0 and t0 , the central time of all wavelets with time shift b,
and the central frequency of all wavelets at scale a are
t ( b ) = t0 + b (3.1.3)
and
ω ( a ) = ω0 / a (3.1.4)
If the spread in time and frequency of the mother wavelet are σ t , 0 and σ ω , 0 , the spread in time
and the spread in frequency of all wavelets with scale a will be
σ t ( a ) = a σ t ,0 (3.1.5)
and
σ
σ ω ( a ) = ωa,0 (3.1.6)
20
Figure 3.1.1 illustrates the localization bins in the time-frequency plane for three wavelets,
with a = 1 , a < 1 and a > 1 . It can be seen that for frequencies ω > ω0 ( a < 1 ) the bean is
stretched along the ω -axis, but compressed along the t-axis, and the opposite is true for
ω < ω0 ( a > 1 ). This means that the lower frequency components are better localized in
frequency than higher frequency components. The relative spread however σ ω / ω is const.
This is the major difference between time-scale and time-frequency analysis (e.g., the short time
Fourier transform, also called moving window Fourier analysis, or Gabor transform—when the
moving window has Gaussian shape), in which absolute spreads in time and frequency are
constant throughout the entire time-frequency plane.
ω (=ω 0/a)
2 σ t,0 a
∆ω σ ω,0 /a σ ω,0
= = =const.
ω ω 0/a ω0
ω>ω 0 2 σ ω,0 /a
2 σ t,0
a<1
a=1 ω0 2 σ ω,0
2 σ t,0 a
a>1
ω<ω 0 2 σω ,0 /a
t
t0 t0+b
Fig. 3.1.1 Time-frequency resolution for three wavelets in a family.
3.2 Wavelet Transforms
The most general transform involving wavelets is the Continuous Wavelet Transform
(CWT). For a signal s (t ) ∈ L2 ( R) , the CWT is defined as its inner product with a family of
admissible waveletsψ a ,b (t )
21
Ws (a, b) =< ψ a ,b , s > L2
∞
= ∫ s(t ) ψ a,b (t )dt (3.2.1)
−∞
∞
1  t −b 
=
a
∫ s(t )ψ  a 
 dt , a > 0, b ∈
−∞
where the bar indicates complex conjugate. From the nature of the inner product of two
functions, it follows that the CWT at scale a and time b is a projection of the signal onto the
wavelet with scale a and time shift b, and hence it shows how similar it is to that particular
wavelet. From the Parseval’s equality, the wavelet transform can also be computed from the
Fourier transforms of the wavelet and of the signal
Ws (a, b) =< ψ a ,b , s > L2

∞
1
=
2π ∫ sˆ(Ω) ψˆ a,b (Ω) d Ω (3.2.2)
−∞
∞
1 i Ωb
=
2π
a ∫ sˆ(Ω) ψˆ (a Ω) e d Ω, a > 0, b ∈
−∞
and also is a measure how similar the Fourier transform of signal is to the one of the
corresponding wavelet. From the last line of eqn (3.2.2) it follows that, up to a constant, the
wavelet transform is the inverse Fourier transform of sˆ(Ω) ψˆ (a Ω) and can be computed using
the Fast Fourier Transform.
The CWT is a time-scale distribution and shows the local properties of the signal at a
particular time and a particular scale. As scale and frequency are closely related and can be used
interchangeably (see eqn (3.1.4), it is also a time-frequency distribution. The localization
properties (central time and frequency and spread) of Ws (a, b) are same as those of the wavelet
ψ a,b (t ) (see Fig. 3.1.1.).
The continuous wavelet transform is highly redundant, as a function defined on is

+
transformed into a function defined on × (where indicates the real line representing
+
time, and indicates the positive real line representing scale). It contains all the information
about the signal, which can be reconstructed from its transform. The reconstruction formula is
known as the resolution of identity. The reconstruction formula, in general, requires to integrate
the transform over the full range of scales ( a ∈ , a ≠ 0 ), even though negative scales are
physically meaningless, similarly as the inverse Fourier transform requires integration also over
22
negative frequencies which have no physical meaning. For real wavelets, however, which are of
concern in this report, it is necessary to integrate only on the half-range of scales ( a ∈ + ), due
to the symmetry of the magnitude of their Fourier transform. The half-range reconstruction
formula is
∞ ∞
1  t − b  dadb
s (t ) =
Cψ ∫ ∫ Ws (a, b)ψ 
 a  a2
 (3.2.3)
0 −∞
where
∞ 0
|ψˆ (Ω) |2 | ψˆ (Ω) |2
Cψ = ∫ dΩ = ∫ dΩ < ∞ (3.2.4)
0
Ω −∞
Ω
with
∫ ψ (t )e
− iΩt
ψˆ (Ω) = dt (3.2.5)
−∞
being the Fourier Transform of the mother wavelet, and the full-range formula is
∞ ∞
1  t − b  dadb
s (t ) =
Cψ ∫ ∫ Ws (a, b)ψ  
a  a2
(3.2.6)
−∞ −∞
where
∞
|ψˆ (Ω) |2
Cψ = ∫ Ω dΩ < ∞ (3.2.7)
−∞
We should mention here that the function expanded should be zero mean for the reconstruction
formula to be valid strictly speaking, because the wavelets are zero mean, and hence cannot
represent a constant function. This however is not a restriction, as any function that is not zero
mean can be represented as sum of its average value and the difference, which is zero mean.
Both the representation of a function by its wavelet transform and the inversion formula
(resolution of identity) were proposed by French exploration geophysicists Grossman and
Morlet, first intuitively and with numerical examples, and then within a rigorous mathematical
framework (Morlet et al., 1982a,b; Grossman and Morlet, 1984; Goupillaud et al., 1984/85), not
being aware of the similar work of the mathematician Calderon in 1960 (Meyer, 1992).
23
The Discrete Wavelet Transform (DWT) is a special case of the CWT when the scale and
time are sampled on discrete grids of the form
a = a0m , b = nb0 a0m , m, n ∈ , a0 ≠ 1, b0 ≠ 0 (3.2.8)
where is the set of integers, in which case the transform becomes
Ws (m, n) = ψ m,n , s
L2
∞ (3.2.9)
= a0− m / 2 ∫ ( )
s (t )ψ a0− mt − nb0 dt , m, n ∈ , a0 ≠ 1, b0 ≠ 0
−∞
It also can be evaluated in the frequency domain, using eqn (3.2.2) and substituting for a and b
from eqn (3.2.8).
The DWT is also a time-frequency distribution. Along the frequency axis, the neighboring
nodes on the grid differ by a factor of a0 . The nodes on the grid are denser at smaller
frequencies and become progressively more sparse as the frequency increases, but are equally
spaced on a logarithmic scale (at distance log a0 ). In time, the nodes are equally spaces at a
particular frequency, but the spacing changes as function of frequency, being larger for the
smaller frequencies, i.e. when the wavelets are more spread in time. This is natural as the
spacing is proportional to the width of the time-frequency bins for the corresponding wavelets,
illustrated in Fig. 3.1.1. Most widespread is the DWT with a0 = 2 and b0 = 1 , which is referred
to as dyadic wavelet transform. However, using a0 < 2 may be advantageous for some
applications, as in that case the sampling in frequency is denser.
( )
If the sampling is sufficiently dense (i.e. for appropriately chosen pair a0 , b0 ), the signal
can be reconstructed as
s (t ) = ∑∑ ψ m,n , s ψ m,n (t ) (3.2.10)

n n
where, in general, functions ψ m,n ≠ ψ m,n . Appropriately sampled wavelet families (i.e. such that
the transform is continuous and invertible) are a special case of families of functions called
frames. The following is a brief review of frames, which will lead to the definition of
orthonormal wavelet bases.
24
A series of functions {ψ m,n } ∈ H , where H is a Hilbert space, constitutes a frame if for
m ,n∈
all signals s ∈ H there exist constants 0 < A ≤ B < ∞ such that
2
∑∑
2 2
A s ≤ ψ m ,n , s ≤B s (3.2.11)
m∈ n∈
where ⋅ indicates L2 norm (Daubechies, 1992). The constants A and B are the frame bounds
( B < ∞ guarantees that the transform is continuous, and A > 0 guarantees that it is invertible).
In general, the functions constituting a frame are linearly dependent (there are more functions
than necessary to reconstruct the signal), and hence the expansion is redundant. The degree of
redundancy is described by the redundancy ratio B / A ≥ 1 . If A = B , the frame is called tight
and the reconstruction formula is simple
1
s (t ) = ∑
A m∈
∑ ψ m ,n , s ψ m ,n ( t ) (3.2.12)
n∈
If A ≈ B , the frame is called snug, in which case there is a simple approximate reconstruction
formula
1  ε 
s (t ) =
( A + B) 2 ∑∑ ψ m , n , s ψ m ,n ( t ) + O  s 
 2+ε 
(3.2.13)
m∈ n∈
where
B
ε= −1 (3.2.14)
A
An exact reconstruction formula exists even for frames that are neither tight nor snug but is
more complex and can be found, e.g. in Abbate et al. (2002) and Vetterli and Kovacević (1995).
A special case of a tight wavelet frame is when A = B = 1 , in which case the wavelet family
is linearly independent and constitutes a basis (there are exactly as many functions as necessary
to reconstruct the signal), and the expansion is nonredundant. It is very convenient if the wavelet
basis is also orthogonal. Without loss of generality, we assume that an orthogonal wavelet basis
is also orthonormal (the basis functions have unit norm)
1, m = m ' and n = n '

ψ m,n ,ψ m ',n ' =  (3.2.15)
0, otherwise
in which case the reconstruction formula is
25
s (t ) = ∑∑ ψ m,n , s ψ m,n ( t )
m n
(3.2.16)
= ∑∑ W ( m, n ) ψ m,n ( t )
m n
A wavelet basis that is not orthogonal is said to be bi-orthogonal, as one can always find a
{
dual basis ψ m* ,n } m,n∈
∈ H such that
1, m = m ' and n = n '

ψ m,n ,ψ m* ',n ' =  (3.2.17)
0, otherwise
and in which case the bi-orthogonal basis expansion is
s (t ) = ∑∑ ψ m* ,n , s ψ m,n ( t )
m n
(3.2.18)
= ∑∑ W * ( m, n ) ψ m,n ( t )
m n
Orthonormal expansions are convenient because orthogonality simplifies the analysis.

However, it also imposes an additional restriction on the wavelets, which may result in lack of
some desirable properties, such as smoothness and linear phase. For example, the relaxation of
the orthogonality condition makes it possible to construct bi-orthogonal bases of wavelet that are
both smooth and of linear phase. Also, there are applications, e.g. noise removal, for which
over-complete expansions are more desirable than the expansions in a basis, because the over
sampling in time helps reduce artifacts. This study focuses on expansion in real orthonormal
bases, for which there exists a fast algorithm for expansion and synthesis.
The first known orthogonal wavelet basis expansion was proposed by the mathematician
Haar in 1910 (Daubechies, 1992; Vetterli and Kovacević, 1995) as an alternative to the Fourier
Series expansion such that can preserve the time localization. The Haar wavelet is a localized up
and down rectangular pulse, and hence is discontinuous, which is a property not desirable in
expansion of continuous functions. A curiosity at the time it was first introduced, and mostly of
historical significance in wavelet analysis, the Haar basis is still used in analyses of signals with
very abrupt variations, such as e.g. EKG signals in medicine, and in applications that require
very good time localization, as it is the most compact and the only symmetric orthogonal basis of
compactly supported wavelets. Its total opposite is the orthogonal basis of sinc functions (known
in electrical engineering literature as the Shannon wavelet and in mathematics as the Littlewood-
Paley wavelet; Vetterli and Kovacević, 1995), which is a square pulse in the Fourier domain —
hence perfectly localized in frequency, but which has infinite support in the time domain —
26
hence poor localization in time. Another orthogonal basis, constructed by Fourier techniques, is
the one of Meyer wavelets, which are “softer” versions of the sinc wavelets, such that are better
localized in the time domain, though still having infinite support, at the expense of non-perfect
localization in the frequency domain (Vetterli and Kovacević, 1995). Other popular wavelets,
also with infinite support in the time domain, are the Morlet wavelet, proposed by the French
geophysical engineer Morlet (Morlet et al., 1982a,b), which is a complex exponential modulated
by a Gaussian envelope, and the Mexican hat wavelet, which itself is the second derivative of the
Gaussian function (Daubechies, 1992; Vetterli and Kovacević, 1995). Another orthogonal
basis—that of harmonic wavelets—was proposed by Newland (1993) for analysis of mechanical
vibration signals. These wavelets represent a complex generalization of the Shannon wavelets.
All of these wavelets, except for the Haar wavelet, have infinite support in the time domain.
Orthogonal bases of compactly supported wavelets (i.e. wavelets with finite support in the time
domain) such that are more “regular” than the Haar basis (regularity here refers to continuity of
the function itself and possibly of a certain number of its derivatives) were constructed for the
first time by Daubechies in the late 1980s (Daubechies, 1988). Several classes of such wavelets
all constructed by Daubechies are the doublets (named after Daubechies), symlets (named for
their higher degree of symmetry compared to the daublets), and coiflets (named in honor of R.
Coifman), a detailed review of which can be found in Daubechies (1992). The current
widespread use of wavelets in many fields of science and engineering is, to a large degree, due to
the existence of such orthogonal wavelet bases. The expansion and synthesis for these wavelets
can be implemented via filter banks that are perfect reconstruction and using very fast
algorithms.
3.3 Multiresolution Analysis
The general framework for construction of wavelet bases—multiresolution analysis—was

developed in 1986 by Mallat and Meyer (Daubechies, 1992), who saw the connection between
wavelet analysis and subband theory (Mallat, 1989; Meyer, 1990). Multiresolution analysis
consists of a sequence of embedded closed approximation subspaces
V2 ⊂ V1 ⊂ V0 ⊂ V−1 ⊂ V−2 ⊂ (3.3.1)
with the following properties:
a) lim Vm =
m→−∞
∪ Vm = L2 ( ) (3.3.2)
m∈Z
b) lim Vm =
m→∞
∩ Vm = {∅} (3.3.3)
m∈Z
27
c) s (t ) ∈ Vm ⇔ s (2m t ) ∈ V0 (3.3.4)
d) s (t ) ∈ V0 ⇔ s (t − n) ∈ V0 , n ∈ Z (3.3.5)
e) There exists ϕ ∈ V0 (called a scaling function) such that {ϕ ( t − n )}n∈Z is an orthonormal

basis for V0 .
Spaces Vm are approximation spaces that approximate a function at different resolutions,

with larger m corresponding to lower resolution approximation, and the limiting cases V−∞ and
V∞ corresponding respectively to the highest resolution (the one containing all the detail) and the
lowest resolution (containing no detail) approximation, as stated by properties a) and b).
Property c) is the scale invariance property (the approximation subspaces are scaled versions of
each other) and d) is the shift invariance property (all time shifts belong to the same
{(
approximation subspace). Properties c), d) and e) imply that ϕ 2m t − n )} n∈Z
is a basis for V− m .
The following are derived properties from the five axioms of multiresolution analysis. Let
Wm be the orthogonal complement of approximation subspace Vm in approximation subspace
Vm−1 . Then, Vm−1 is a direct sum of Vm and Wm
Vm−1 = Wm ⊕ Vm (3.3.6)
Because Vm is a lower resolution approximation subspace than Vm−1 , then Wm contains the
detail that is removed from the higher resolution approximation to obtain the lower resolution
approximation. Hence, the subspaces Wm , m ∈ Z are called detail subspaces. The detail spaces
also have the scaling and shift properties as the approximation subspaces, i.e.
s (t ) ∈ Wm ⇔ s (2m t − n) ∈ W0 , n ∈ Z (3.3.7)
In contrast to subspaces Vm , m ∈ Z , which are embedded, subspaces Wm , m ∈ Z are

disjoint and hence orthogonal to each other. By applying eqn (3.3.6) recursively, one gets
Vm = Wm+1 ⊕ Vm+1
= Wm+1 ⊕ Wm+ 2 ⊕ Vm+ 2
(3.3.8)
= Wm+1 ⊕ Wm+ 2 ⊕ Wm+3 ⊕ Vm+3
28
which together with eqn (3.3.2) implies that the entire space is a direct sum of the detail
subspaces
L2 ( ) = ⊕ Wm (3.3.9)
m∈Z
There is theorem that guarantees the existence of an orthonormal wavelet basis
{ψ m,n ( t )}m,n∈Z = {2−m / 2ψ ( 2m t − n )}m,n∈Z (3.3.10)
such that spans L2 ( ) , and such that for given m ∈ Z , {ψ m,n ( t )} is a basis for Wm .
n∈Z
Multiresolution analysis can be used to construct wavelet bases. The simplest wavelet bases that
satisfy the axioms of multiresolution analysis are the Haar and the sinc bases, which are duals of
each other.
Important relationships in multireslution analysis are the two-scale equations
ϕ ( t ) = 2 ∑ g0 [ n ] ϕ ( 2t − n ) (3.3.11)
n∈Z
ψ ( t ) = 2 ∑ g1 [ n ] ϕ ( 2t − n ) (3.3.12)
n∈Z
which follow from the fact that every function in V0 and W0 , including their basis functions, can
be expanded in a series of the basis functions of V−1 , which is the next higher resolution
approximation space containing both V0 and W0 . The coefficients of the expansion, g 0 [ n ] and
g1 [ n ] , constitute filters called synthesis filters.
3.4 Subband Decomposition and Its Relation to Wavelet Basis Expansion
Subband theory is an area of digital signal processing, which is based on the design of a set
of prototype filters such that divide the signal frequency band in equal parts (subbands). For
example, a two-channel filter bank consists of a high-pass and low-pass filter. These filters are
applied recursively to the signal, which leads to its division in subbands. The filters applied at
each stage have identical properties except for scale. The set of filters is referred to as filter bank.
The prototype filters are used to construct the wavelet and the scaling function. Hence, subband
decomposition becomes equivalent to basis decomposition for discrete time signals. The
29
following example shows how multiresolution approximations of a signal can be obtained using
subband decomposition.
All discrete time signals are band-limited, and the highest frequency they can contain is the
Nyquist frequency, f max = 1/ ( 2∆t ) Hz (where ∆t is the sampling period in seconds), which
corresponds to circular frequency ωmax = π radians per sample (this frequency is equal to the
circular frequency in radians per second if ∆t = 1 s). Hence, all discrete signals belong to the
space of band-limited functions on ω ∈ [ −π , π ] , which we will call V0 . Let the prototype filters
be ideal low- and high-pass filters, as shown in Fig. 3.4.1 (top), with impulse responses
respectively h0 [ n ] and h1 [ n ] , and with Fourier transforms H 0 (ω ) and H1 (ω ) . Filters H 0 (ω )
and H1 (ω ) split the signal into two components, S1 and D1 , such that the former is a smooth
approximation of the signal and the latter is the remainder, containing the detail. These two
components are respectively band-limited on ω ∈ [ −π / 2, π / 2] and ω ∈ [ −π , −π / 2] ∪ [π / 2, π ] .
Next, the smooth component, S1 , is split into a low-pass (smooth) and high-pass (detail)
components S 2 and D2 , by applying filters H 0 ( 2ω ) and H1 ( 2ω ) . If this procedure is repeated
recursively, after J steps we have
s ( t ) = D1 ( t ) + S1 ( t )
= D1 ( t ) + D2 ( t ) + S2 ( t )
= D1 ( t ) + D2 ( t ) + D3 ( t ) + S3 ( t )
(3.4.1)
= D1 ( t ) + D2 ( t ) + + DJ ( t ) + S J ( t )
J
= ∑ D j (t ) + SJ (t )
j =1
This leads to division of the space V0 into a sequence of detail subspaces W j , j = 1,… J and a
smooth subspace VJ
V0 = W1 ⊕ W2 ⊕ W3 ⊕ ⊕ WJ ⊕ VJ (3.4.2)
shown in Fig. 3.4.1 (bottom) for the positive half-band only. Recalling that D j ∈ W j , j = 1,… J
and S J ∈ VJ , and that each Wm has basis {ψ m,n ( t )} and VJ has basis {ϕ J ,n ( t )} , it follows
n∈Z n∈Z
that the subband components can be expanded in series of the basis functions as follows
30
N /2j
D j (t ) = ∑ d j ,kψ j ,k (t ) (3.4.3)
k =1
and
N / 2J
S J (t ) = ∑ s J ,kϕ J ,k (t ) (3.4.4)
k =1
which gives the relationship between subbands and the wavelet bases. The coefficients of the
expansion represent the discrete wavelet transform (DWT) of the signal
d j ,k = ψ j ,k , s ( t ) (3.4.5)
s J ,k = ϕ J ,k , s ( t ) (3.4.6)
We recall that in eqns (3.4.3) through (3.4.6)
(2 )
−
−j
ψ j ,k (t ) = 2 2ψ t−k (3.4.7)
| H 0 (ω) | | H 1 (ω) |
ω ω
π π π π
−π − 0 π −π − 0 π
2 2 2 2
V J-1
V2
V1
V0
VJ W J W2 W1
ω
π π π π π
0
2 J 2 J-1 22 2
Fig. 3.4.1 Ideal low and high pass filters as prototype filters for a two-channel filter bank (top), and an
illustration of a J-level division of the space of discrete time signals, V0, by application of the filter bank.
31
j
( )
−
ϕ j ,k (t ) = 2 2 ϕ 2− j t − k (3.4.8)
where ψ (t ) and ϕ (t ) are respectively the prototype wavelet and scaling functions for the basis.
3.5 Subband Central Frequencies and Bandwidth for Ideal Prototype Filters
For ideal low- and high-pass prototype filters and discretely sampled signals, the division of
the signal bandwidth into subbands, and the central frequencies ω and bandwidths ∆ω of the
subbands are illustrated in Fig. 3.5.1. It can be seen that for some intermediate stage of filtering
j, the high and low pass channels have central frequency and bandwidth
3π π
Dj : ω = j +1
, ∆ω = j
2 2 (3.5.1)
π π
Sj : ω = , ∆ω =
2 j +1 2j
For a J level decomposition, there are J+1 subbands with central frequency and bandwidth
3π π
D1 : ω = , ∆ω =
4 2
3π π
D2 : ω = , ∆ω =
8 4
(3.5.2)
3π π
DJ : ω = J +1
, ∆ω = J
2 2
π π
SJ : ω = J +1
, ∆ω =
2 2J
We recall that ω is the frequency in radians per sample. For signal sampled at time intervals
equally spaced at ∆t seconds, the frequency in Hz is
ω
f = ∆t (3.5.3)
2π
which implies
32
3 1
Dj : f = j+2
∆t Hz, ∆f = ∆t Hz
2 2 j +1 (3.5.4)
1 1
SJ : f = ∆t Hz, ∆f = j +1 ∆t Hz
2 j+2 2
| H 0 (ω) | | H 1 (ω) |
ω ω
π π π π
−π − 0 π −π − 0 π
2 2 2 2
π π π π
J J
2 2 4 2
V0
VJ W J W2 W1
ω
π π π π π
0
2 J 2 J-1 4 2
π 3π 3π 3π
2 J+1 2 J+1 8 4
Fig. 3.5.1 Subband central frequencies and bandwidth for ideal filters.
3.6 Perfect Reconstruction Orthonormal Filter Banks
Section 3.4 showed conceptually the link between subband decomposition and
multiresolution analysis and wavelet bases decomposition. This section deals with the details of
this relationship, in particular with the characteristics of and the relation between the different
filters involved, and with how these filters are related to the wavelets. The analysis will be
restricted to real and orthogonal wavelet bases and discrete signals, which are of interest for the
analysis in this report.
Section 2.10 showed the relationship between inner product and convolution, which is the
basis for the relationship between subband decomposition and wavelet bases expansion. Let us
recall the analysis and synthesis formulae for an orthonormal basis expansion
33
ck = ϕk , s , k∈ (3.6.1)
s [ n ] = ∑ ckϕk [ n ] (3.6.2)
k
The analysis formula can be written as a convolution with a filter
ck = ϕk , s = s ∗ hk (3.6.3)
where the filter is a time-reversal of the k-th basis function
hk [ n ] = ϕk [ − n ] (3.6.4)
Hence, the analysis filters are time reversals of the basis functions. Let c be a vector of the
Fourier coefficients and s be a vector representing the signal. Then, eqn (3.6.3) can be written
in matrix form as
c=Hs (3.6.5)
where the analysis matrix H has entries
H k ,n = ϕ [ k − n ] (3.6.6)
Let us consider this basis expansion as the level 1 expansion in a wavelet basis. As it turns
out, the basis functions corresponding to even k are shifts of the scaling functions, denoted here
by ϕ0 [ n ] , and those for odd k are shifts of the wavelets, denoted here by ϕ1 [ n ] , and the filters
hk for even k are shifts of the low-pass prototype filter h0 [ n ] and those for odd k are shifts of
the high-pass prototype filter h1 [ n ] . This implies that the low-pass filters are time reversals of
the scaling functions and the high-pass filters are a time reversal of the wavelets, which for zero
shift can be written as
h0 [ n ] = ϕ0 [ − n ] (3.6.7)
h1 [ n ] = ϕ1 [ − n ] (3.6.8)
We note that convolution evaluated at even indices is equivalent to applying the filter and
then down sampling by 2 (disregarding every other sample), which is the main idea behind the
analysis past of the pyramid algorithm described in the following sections.
34
The synthesis formula (3.6.2) can also be implemented by filtering with low-pass and high-
pass filters, called the synthesis filters, such that are equal to the basis functions. This defines the
prototype low-pass and high-pass, g 0 [ n ] and g1 [ n ] , as
g 0 [ n ] = ϕ0 [ n ] (3.6.9)
g1 [ n ] = ϕ1 [ n ] (3.6.10)
These filters are applied respectively on the even and odd subsequences of the sequence of
the Fourier coefficients, and the output sequences of these two convolutions are summed up.
Convolution with every other sample of the input sequence is equivalent to up-sampling the
input sequence of the Fourier coefficients by 2 (by inserting zeros between the samples) and then
performing the convolution, which is the main idea behind the synthesis part of the pyramid
algorithm.
The analysis and synthesis filters form a filter bank. Those filter banks that implement a
basis expansion are called perfect reconstruction filter banks. In filter banks analysis, perfect
reconstruction filter banks are such that, up to a shift and scale, the impulse responses of the
analysis filters and of the synthesis filters form a bi-orthogonal basis (Vetterli and Kovacević,
1995). Hence, for a general basis expansion (not necessarily orthogonal), these filter banks are
bi-orthogonal, and for an orthonormal basis expansions—they are orthogonal. Equations (3.6.7)
through (3.6.10), imply the following relationship between the analysis and synthesis filters
( )
gi [ n ] = hi [ − n ] ⇔ Gi ( z ) = H i z −1 , i = 1, 2 (3.6.11)
which is the perfect reconstruction condition for orthonormal filter banks. Then, for
orthonormal perfect reconstruction filter banks, the synthesis filters are completely specified by
the analysis filters, and are in fact their time reversals.
3.7 Aliasisng Cancellation
The ideal low- and high-pass prototype filters have infinite impulse responses (IIR filters)
that decay slowly in time (as 1/n), which implies poor time localization. To implement them in
a digital computer, they have to be truncated. Truncation, however, changes their Fourier
transforms, in that it is not any more a perfect box, but has ripples both in the pass and in the stop
bands, which cause leakage of energy in the stop band and aliasing. The ripple effect can be
abated by tapering in the Fourier domain, but the leakage of energy in the stop band and the
aliasing distortions will remain. The reconstruction process also introduce similar errors, which
would further magnify the distortion. These distortions will be further magnified when the
35
signal is being reconstructed using the same filters (up to a time reversal). While the aliasing
cannot be eliminated, with a special design of the filters, it is possible to make the distortions
from the synthesis such that they cancel the distortion from the analysis, so that the
reconstruction is indeed perfect, even in the numerical implementation. One such design
solution is to use quadrature mirror filters (QMF), which in the Fourier domain are symmetric
with respect to ω = π . For orthonormal filters, the perfect reconstruction requirement already
imposes symmetry between the low-pass synthesis and low-pass analysis filters, and between the
high-pass synthesis and high-pass analysis filters, and the only new quadrature-mirror
requirement is the symmetry between the low and high pass filters, which up to a time shift has
the form
(
G1 ( z ) = − z −2 K +1G0 − z −1 ) ⇔ g1 [ n ] = ( −1) g 0 [ 2 K − 1 − n ]
n
(3.7.1)
Then, for a given low pass filter, the high pass filter is obtained by a time reversal and by
flipping the sigh of every other entry. Hence, only one of the filters needs to be specified, and
the other three filters can be obtained through equations (3.6.11) and (3.7.1).
3.8 The Pyramid Algorithm
For expansion of discrete time signals in orthogonal wavelet bases such that their prototype
filters are finite impulse response (FIR) filters, there exists a fast algorithm for computation of
the coefficients of the expansion, which is based on the theory of multiresolution analysis and
uses a two-channel filter bank. For large N—the length of the signal, it has complexity O(N),
and is efficient because the coefficients at each stage are computed using results from the
previous stage. It will be referred to as Fast Wavelet Transform (FWT). It is called the pyramid
algorithm because of its structure, or Mallat’s algorithm (Vetterli and Kovacević, 1995). It
consists of splitting the signal into a high- and a low-frequency component, followed by sub-
sampling by a factor of two, keeping the high-frequency output, while operating further on the
low-frequency output by splitting and sub-sampling it in the same fashion. The splitting and
sub-sampling of the low-frequency output is applied iteratively in several stages. The final
results consist of the outputs of the high frequency channels and the low frequency output of the
last channel. The maximum number of iteration stages, called levels of the decomposition,
depends on the length of the signal and on the length of the impulse responses of the digital
filters used to spit the signal (it is larger for a longer signal and for a shorter filter). The output
sequences contain the coefficients of the Discrete Wavelet Transform of the signal. The
algorithm is illustrated in Fig. 3.7.1 (top), where H0 and H1 are the low and high-pass prototype
filters, d j ,k , k = 1,..2 j and s j ,k , k = 1,..2 j are the sequences of the “detail” coefficients and
“smooth” coefficients for level j of the expansion, with j = 1,..., J = total number of levels. As a
36
result of the repeated sub-sampling, these sequences are progressively shorter for the higher
levels of the decomposition, i.e. for the low frequency components of the signal. Then, the
forward FWT is equivalent to splitting the signal into frequency bands, and keeping less samples
for the lower frequency bands. In each subband, the signal can be reconstructed from the
samples using appropriate wavelets as interpolating functions, much like a band-limited
continuous signal can be reconstructed from its samples (if sampled at a sufficiently high
sampling rate) using a basis of translated sinc functions as interpolators, as stated by the
sampling theorem. This can also be done using the inverse FWT (also of complexity O(N)) by
setting to 0 the coefficients for all the other subbands. The entire signal is reconstructed as
shown in Fig. 3.7.1 (bottom).
The FWT is nonredundant due to the down sampling. If the signal consists of N samples,
then after each stage the total number of coefficients is also N. For example, the number of
coefficients for J level decomposition is
N / 2 + N / 2 2 + N / 23 + + N / 2 J −1 + N / 2 J −1 = N (3.8.1)
d1,k d 2,k d3,k d J ,k s j ,k
It should be noted that the FWT and IFWT do not require explicit knowledge of the
wavelets themselves, but only of the prototype filters, which can be used to construct the
prototype wavelet and scaling function (using the two-scale equations (3.3.11) and (3.3.12)) if
that is necessary. We recall here that, for expansion of discrete signals, on the integers, the
wavelets are equal to synthesis high-pass filters, and the scaling functions are equal to the
synthesis low-pass filters (between the integers, they can be obtained by interpolation), which is
true for every stage of filtering. Let G0( j ) ( z ) and G1( j ) ( z ) be the equivalent iterated low and
high pass filters that correspond to stage j of filtering. These filters are products of all of the
previously applied filters, after an appropriate scaling to account for the up-sampling by 2. In the
z-domain, these filters are given by
 ( 2k ) 
j −1
G0( j ) ( z ) = ∏ G0  z 
k =0  
G1 ( z ) , j =1 (3.8.2)

G1( j ) ( z ) =   ( 2 j −1 )  j −2  ( 2k ) 
 1z
G  ∏ G0  z , j >1
   k =0  
and in the Fourier domain by
37
The Pyramid Algorithm for Fast Wavelet Transform
H1 2 d 1,k
s[k] H1 2 d 2,k
H0 2 s 1,k H1 2 d 3,k
H0 2 s 2,k
j=1
H0 2 s 3k
j=2 H1 2
d J,k
j=3
Analysis - DWT H0 2
s J,k
j=J
38
d 1,k 2 G1
d 2,k 2 G1 + s[k]
d 3,k 2 G1 + s 1,k 2 G0
+ s 2,k 2 G0
j=1
s 3k 2 G0
2 G1 j=2
d J,k
+ j=3
s J,k 2 G0 Synthesis - Inverse DWT
j=J
Fig. 3.7.1 The forward (top) and inverse (bottom) pyramid algorithms for Fast Wavelet Transform.
( )
j −1
( )
G0( j ) eiω = ∏ G0 e −iω 2
k
k =0

( )
G1 eiω , j = 1 (3.8.3)
( )
G1( j ) eiω =
( ) ( )
j −2
G1 e −iω 2 ∏ G0 e −iω 2 ,
j −1 k
j >1
 k =0
These iterated filters define all the wavelets and scaling functions, up to the shifts in time, which
are equal to 2 j ( k − 1) , k = 1,...N / 2 j . All of these filters have norm 2π , which by the
Parseval’s equality agrees with the fact that all the wavelets have unit norm.
The following are some remarks about the two-channel filter banks. The two-channel filter
banks are called constant Q filtering because for each subband the relative bandwidth
(bandwidth/central frequency) is the same. These are also called logarithmic filter banks because
the channels have equal bandwidth on a logarithmic scale. These are also called octave band
filter banks because each successive high pass output contains an octave of the input bandwidth.
In English language dictionary, octave is: (a) series of eight notes occupying the interval between
and including two notes, one having half/twice the frequency of the other; (b) the interval
between the two notes; (c) each of the two notes at the end of the interval; (d) the two notes
sounding together.
3.9 Examples of Orthonormal Wavelet Bases
3.9.1 The sinc Basis
If the prototype filters are ideal low- and high-pass filters
  π π
 2, ω ∈  − , 
( )
G0 e iω
=  2 2
0, otherwise

(3.9.1.1)
 − iω  π  π 
 2 e , ω ∈  −π , −  ∪  , π 
( )
G1 eiω =  2 2 
0, otherwise

then the wavelet basis is the sinc basis, which has prototype scaling function
sin (π t )
ϕ ( t ) = sinc1 = (3.9.1.2)
πt
39
(as it could be anticipated recalling the reconstruction formula for a band-limited continuous
signals from their discrete samples) and a prototype wavelet
sin (π t / 2 )
ψ (t ) = cos ( 3π t / 2 ) (3.9.1.3)
πt / 2
(Vetterli and Kovacević, 1995). This wavelet is known as the sinc1 wavelet or the Shannon
wavelet, and in mathematical literature as the Littlewood-Paley wavelet. The corresponding
wavelet basis is
{ψ m,n ( t )}m,n∈Z = 2−m / 2ψ  2m t − n − 12  (3.9.1.4)

 m,n ∈Z
where {ψ m,n ( t )} is a basis for functions band-limited on

m ,n∈Z
ω ∈  −2− m+1π , −2− m π  ∪  2− m π , 2− m+1π  (3.9.1.5)
Figure 3.9.1.1 illustrates the prototype wavelet and scaling function (computed using the S-
plus wavelet toolbox; Bruce and Gao, 1996). What is characteristic about this basis is that the
sinc Wavelet
low-pass high-pass
1.2 1.2
0.8 0.8
0.4 0.4
G 0(e iω) G 1(e iω)
0.0 0.0
0 π/2 ω π 0 π/2 ω π
2 φ 2 ψ
0 0
-2 -2
-10 -5 0 5 t 10 -10 -5 0 5 t 10
Fig. 3.9.1.1 Prototype filters, scaling function and wavelet for the sinc wavelet basis.
40
wavelets are very well localized in frequency but have infinite support in time and decay slowly
as t → ±∞ , just the opposite of what is the case for its dual—the Haar basis. Also, because the
prototype filters are not FIR, the sinc basis cannot be constructed using the pyramid algorithm
but using other, e.g. Fourier techniques.
3.9.2 The Haar Basis
The Haar basis, constructed by the mathematician Haar in 1910 (Vetterli and Kovacević,
1995), is a dual of the sinc basis. While the sinc basis functions are box functions in frequency
and have infinite support in the time domain, the Haar basis functions are square waves (i.e. an
up and down box function) in time and have infinite support in frequency. For the Haar basis,
the prototype scaling and wavelet functions are
1, 0 ≤ t < 1
ϕ (t ) =  (3.9.2.1)
0, otherwise
 1
 1, 0≤t <
2

 1
ψ ( t ) = −1, ≤ t <1 (3.9.2.2)
 2
 0, otherwise


The prototype low- and high-pass filters in the frequency domain are
( )
G0 eiω =
1
2
(
1 + e −iω )
(3.9.2.3)
( )
G1 e iω
=
1
2
(
1 − e −iω )
and in the time domain are the two-point moving average and moving difference filters. The
Haar basis, besides its dual—the sinc basis, is the simplest basis that satisfies the axioms of the
multiresolution analysis and is the simplest basis of compactly supported wavelets (hence
associated with FIR filters) that can be constructed using the pyramid algorithm. Figure 3.9.2.1
illustrates the prototype filters in the Fourier and time domain and the prototype wavelet and
scaling function function (computed using the S-plus wavelet toolbox; Bruce and Gao, 1996).
An obvious feature of the Haar wavelet and scaling function is that they are both discontinuous.
41
Haar Wavelet
low-pass high-pass
1.0 1.0
g 0[n] g 1[n]
0.0 0.0
-1.0 -1.0
0 2 4 6 n 0 2 4 6 n
1.2 1.2
0.8 0.8
0.4 0.4
G 0(e iω) G 1(e iω)
0.0 0.0
0 π/2 ω π 0 π/2 ω π
2 φ 2 ψ
0 0
-2 -2
-10 -5 0 5 t 10 -10 -5 0 5 t 10
Fig. 3.9.2.1 Prototype filters, scaling function and wavelet for the Haar wavelet basis.
3.9.3 Daubechies Compactly Supported Wavelets
Some of the most widely used orthogonal wavelet families are those constructed by Ingrid
Daubechies, who introduced the concept of maximally flat compactly supported wavelets. She
first constructed the doublets, which were the first continuous wavelets with compact support
(the Haar wavelet has compact support but is discontinuous). The doublets come in different
lengths and have different degree of smoothness. Some examples of doublets are: d4, d6, d8,
s10, d12, d14, d16, d18 and d20 wavelets (all implemented in the S-Plus wavelet toolbox, Bruce
and Gao, 1996). These wavelets were named after their discoverer, Ingrid Daubechies, and the
number next to the letter “d” is the length of the prototype (FIR) filters. The larger this number,
the longer and smoother mother wavelet and scaling function. The doublets are quite
42
asymmetric. Daubechies later constructed the symmlets (e.g., s4, s6, s8, s10, s12, s14, s16, s18
and s20), which are as symmetric as possible, considering the other constraints (such as compact
support and orthogonality). It should be noted here that there are no orthogonal compactly
supported wavelets such that are both symmetric and smooth, and that the only symmetric of the
orthogonal compactly supported wavelets are those of the Haar basis, which are discontinuous.
Daubechies also constructed the coiflets (e.g., c6, c12, c24 and c30), named after Ronald
Coifman—one of the contributors to the theory of wavelet analysis, such that, besides being
orthogonal and compactly supported, are also nearly symmetric and have certain number of
vanishing moments. Figure 3.9.3.1 (produced using the S-plus wavelet toolbox; Bruce and Gao,
1996)) shows the prototype filters, in the time and Fourier domains, and the prototype wavelet
and scaling function for the s8 wavelet, which will be used in the illustrations of expansion
S8 Wavelet
low-pass high-pass
1.0 1.0
g 0[n] g 1[n]
0.0 0.0
-1.0 -1.0
0 2 4 6 n 0 2 4 6 n
1.2 1.2
0.8 0.8
0.4 0.4
G 0(e iω ) G 1(e iω)
0.0 0.0
0 π/2 ω π 0 π/2 ω π
2 φ 2 ψ
0 0
-2 -2
-10 -5 0 5 t 10 -10 -5 0 5 t 10
Fig. 3.9.3.1 Prototype filters, scaling function and wavelet for the s8 basis.
43
of strong motion signals in the next chapter. Figure 3.9.3.2 (produced by a Fortran program)
shows the equivalent iterated filters (the wavelet filters are time shifts of these filters) for the s8-
wavelet. Plotted on a logarithmic frequency scale. The filter for subband 5 is shown by a thicker
line to emphasize the ripples.
S8 wavelet and J=7

G 0(7) (e iω)
j=7
10
j=6
G 1(j) (e iω)
j=5
5
j=4
j=3
j=2
j=1
0
0.001 0.01 0.1 ω /π 1
Fig. 3.9.3.2 Wavelet filters for the s8 wavelet basis and J=7 level decomposition.
It can be seen from Fig. 3.9.3.2 that the detail subbands have equal bandwidth on a log scale
and that their central frequencies are equally spaced on a log scale. For filters that are not perfect
low- and high-pass, the central frequency, ω , and bandwidth, ∆ω , can be estimated using the
first and second moments of the magnitude of the low-pass prototype filter as follows. The
central frequency (for the positive half band only) and the variance (for the full band) are
π
( )
2
iω
∫ ω G0 e dω
ωl . p. = 0
π
(3.9.3.1)
∫ G0 ( e )
2
iω
dω
0
44
π
( )
2
∫ω
2
G0 eiω dω
σ 2l . p. = 0
π
(3.9.3.2)
∫ G0 ( e )
2
iω
dω
0
where
  
( )
2
G0 eiω =  ∑ h0 [ n ] e −iω n   ∑ h0 [ m ] e −iω m 
 n  m  (3.9.3.3)
N −1 N −1
− iω ( n − m )
= ∑ ∑ h0 [ n] h0 [ m]e
n =0 m =0
and N is the length of h0 [ n ] . Substituting eqns (3.9.3.3) in (3.9.3.1) gives
N −1 N −1 π
∑ ∑ h0 [ n] h0 [ m]∫ ω e−iω( n−m) dω
n =0 m =0
ωl . p. = 0
π
(3.9.3.4)
N −1 N −1
− iω ( n − m )
∑ ∑ h0 [ n] h0 [ m]∫ e dω
n =0 m = 0 0
where the integrals in the numerator and denominator can be evaluated analytically
π
π , n = m
− iω ( n − m ) 
∫e dω =  ( −1)n− m − 1 (3.9.3.5)
 −i n − m , n ≠ m
 ( )
0
π 2
π
 , n=m
− iω ( n − m )  2
∫ ωe dω =  n−m
( −1)n−m − 1 , n ≠ m
(3.9.3.6)
0  π ( −1) −
 −i ( n − m )  −i ( n − m )  2
  
Similarly, substituting eqns (3.9.3.3) in (3.9.3.2) gives
45
N −1 N −1 π
− iω ( n − m )
∑ ∑ h0 [ n] h0 [ m]∫ ω 2 e dω
n =0 m =0
σ 2l . p. = 0
π
(3.9.3.7)
N −1 N −1
− iω ( n − m )
∑ ∑ h0 [ n] h0 [ m]∫ e dω
n =0 m =0 0
where
π 3
 , n=m
π
2 − iω ( n − m )  2
∫ ω e d ω =  2
π ( −1 ) n−m
2 π ( −1) n−m 2 ( −1)
n−m
− 1
(3.9.3.8)
0  − −  , n≠m
 −i ( n − m ) 2 3
  −i ( n − m )   −i ( n − m ) 
From symmetry, for the high pass prototype filter
ωh. p. = π − ωl . p.
(3.9.3.9)
σ 2 h. p . = σ 2 l . p .
Then, approximating ∆ω by σ l . p. and using the scaling property for the wavelets/subband
ωh . p . σ h. p .
Dj : ω = , ∆ω =
2 j −1 2 j −1 (3.9.3.10)
ωl . p. σ l . p.
Sj : ω = , ∆ω =
2 j −1 2 j −1
It should be noted that ω is very close to that for ideal filters. For example, for the s8
wavelet, ωl . p. = 0.26π ≈ 0.25π , which is the central frequency for an ideal low-pass filter. The
spread, measured via fitting a distribution, is also similar— σ l . p. = 0.31π for the s8 wavelets
versus σ l . p. = 0.29π for the ideal low-pass filter. However, the effective bandwidth, i.e. the
bandwidth containing most of the band-passed energy, remains the same as that of the ideal low
pass filter, i.e. ∆ωeffective = 0.5π .
The wavelets and scaling functions have the same central frequency and bandwidth as the
corresponding subband. Because the mother wavelet and scaling functions belong to the bases
for W0 and V0 , they have the following properties
46
ψ ( t ) : ω = 2ωh. p.
(3.9.3.11)
ϕ ( t ) : ω = 2ωl . p.
The equations in this section are not found in standard books on wavelets.
3.9.4 Properties of Wavelet Families
Important properties of wavelet families are: orthogonality, compactness of support,

smoothness, degree of symmetry and number of vanishing moments. The following discusses
briefly the significance of these properties.
The wavelets with more compact support are better localized in time/space but more poorly
in frequency (as per the Heisenberg principle). Wavelets that are smoother (i.e. continuous and
also differentiable certain number of times) represent/approximate better (with less artifacts)
smooth functions. For example, the Haar wavelet is not continuous and not differentiable. The
d8 wavelet is continuous but not differentiable, and the d12 wavelet is continuous and twice
differentiable. The wavelets that are more symmetric avoid phase shift artifacts, such as
“drifting” of the coefficients relative to the signal (Bruce and Gao, 1996). A wavelet with M
vanishing moments is such that
∫ t ψ ( t ) dt = 0,
m
m = 1,..., M − 1 (3.9.4.1)
−∞
Hence, the inner product of such a wavelet with signals that are polynomials of degree < M − 1 is
equal to zero. Such wavelets are better (more efficient) for representing signals that are higher
degree polynomials, and are more efficient in detecting singularities in signals.
As mentioned in the section 3.9.3, the only symmetric wavelet of an orthogonal and
compactly supported family is the Haar wavelet, which is discontinuous, and the degree of
symmetry and smoothness of such families is inversely proportional to the length of the support.
However, there is no such constraint for families that are not orthogonal. For example, there are
smooth compactly supported wavelets in bi-orthogonal wavelet families.
3.10 Parseval’s Equality for the Orthonormal Discrete Wavelet Transform
The Parseval’s equality for the Discrete Orthogonal Wavelet Transform and its inverse—
expansion in orthogonal wavelet bases—is obtained by substitution in the generalized Parseval’s
equality in eqn (2.2.4) for the appropriate basis functions and Fourier coefficients. The
transform pair, consisting of the synthesis and analysis transforms, is
47
J N /2j N / 2J
s [ n] = ∑ ∑ d j ,kψ j ,k [ n] + ∑ s J ,k ϕ J ,k [ n ] (3.10.1)
j =1 k =1 k =1
d j ,k = ψ j ,k , s , s J ,k = ϕ J ,k , s (3.10.2)
and is very similar to the Fourier series pair, except for the time localization of the basis
functions. Let s [ n ] and s [ n ] be two signals, and d (j s,k) and s J( s,k) , and d (j ,gk) and s J( g,k) ,
respectively, be their coefficients of expansion in orthonormal wavelet basis. The Parseval’s
equality for these two signals is
∞ J N /2j N / 2J
s, g = ∑ s [ n] g [ n] = ∑ ∑ d (j s,k) d (j ,gk) + ∑ s J( s,k) sJ( g,k) (3.10.1)
n =−∞ j =1 k =1 k =1
If the two sequences are equal
∞ J N /2j 2 N / 2J
2
∑ s [ n] =∑ ∑ ∑
2 2
s = d j ,k + sJ ,k (3.10.2)
n =−∞ j =1 k =1 k =1
2
where s is also the energy of the signal.
48
4. THE DISCRETE WAVELET TRANSFORM AS A DATA MINING TOOL
This chapter deals with the theoretical aspects of dimensionality reduction and information
granulation of time series data, using representation in orthonormal wavelet bases, and for
possible application to data mining. Of particular interest to the authors of this report are time
series data of ground and structural vibration response to strong earthquake shaking, as well as to
ambient noise, or to forced vibration. There apparent suitability of the wavelet domain for this
purpose is mainly due to the following properties: (1) inherent hierarchical structure, which
enables automatically different resolution views of the data, i.e. to zoom in and view the detail
(i.e. the trees) or to zoom out and see a coarser view (i.e. the forest); (b) sparse representation
compared to that in the time domain of non-stationary and/or band-limited time series data,
which leads to high data compression rates with little loss of information; (c) discretely sampled
representation, convenient for automated analysis; and (d) complete representation, which
enables exact reconstruct of the time domain if all the information granules are used.
Illustrations of these concepts for strong earthquake motion records are presented in Chapter 5.
This chapter first reviews some basic concepts in data mining and knowledge discovery,
which is followed by a review of dimensionality reduction of time series data by compression.
The simplest forms of compression are considered, which are thresholding and lower resolution
approximation. This is followed by estimation of various global and local aggregates and
averages computed directly in the wavelet domain—such as energy, power, power spectrum
density, Fourier transform amplitude, correlation, cross-power and cross-power spectrum
density—and their interpretation as granules of information, representative of the entire signal, of
the subbands (which are intervals on the frequency axis), or of the tiles partitioning the time-
frequency plane. The estimation of the local aggregates and averages leads to nodal spectra and
nodal time-frequency distributions (or instantaneous spectra). Next, it is shown how the wavelet
coefficients and the same local and global aggregates and averages computed in the wavelet
domain can be estimated for derivatives and integrals of the signal. The estimation of all of these
quantities is presented for discrete signals, representing sampled continuous processes. Finally,
estimation of the same quantities for the related continuous time signal is presented. These
estimates for the global and local aggregates and averages represent preprocessed data, which
can be further analyzed by data mining methods, e.g. statistical regression, cluster analysis,
classification, etc.
4.1 Some Basic Concepts in Data Mining and Knowledge Discovery
Sensors collect data about a process or an object, but data represent only raw information,
which is not useful per se. What is useful is the knowledge revealed from the information
contained in the data, which enables one to draw opinions based on the data and take actions
accordingly. The purpose of these opinions may be to understand the nature of the physical
49
process that generated the data (i.e. understand the past), or further to predict future events, or to
serve as a basis for decision-making (i.e. shape the future). This section reviews some basic
concepts in data mining and knowledge discovery, with the purpose of providing a context for
the application of the discrete wavelet transform, which follows. The definitions of these
concepts have been taken exactly or closely following Cios et al. (1998).
Data mining methods are tools used in knowledge discovery to reveal new pieces of
knowledge from large data sets. The terms knowledge discovery and data mining first appeared
in the late 1980s. According to Cios et al. (1998), the former appeared first around 1989, is
attributed to Frawley, Piatetsky-Shapiro and Matheus (1991), and is defined as “the nontrivial
process of identifying valid, novel, potentially useful, and ultimately understandable patterns in
data.” Here, a pattern is an entity representing (characterizing, describing) an abstract concept or
a physical object, and may describe relationships, correlations, trends, etc. It is a collection of
individual attributes (features) of the data, and hence is a vector quantity. The number of
attributes defines the dimensionality of a pattern in the pattern space. One may search for
patterns in the data and classify the data into classes according to the nature of the patterns
discovered. A class can be thought of as a state of nature that governs the pattern generation.
For example, in data mining of structural vibration data, the concept of interest can be damage,
and a pattern characterizing damage can be a collection of attributes such as visible fractures and
cracks in the structure, reduction of the natural frequency, specific changes of the mode-shapes
of vibration, reduced wave travel times across a fractured structural element, etc. According to
this pattern, the data can be classified into two groups (classes), one indicating possible damage
and the other one—no damage in the structure. In data mining of ground vibration data, the
concept of interest can be hazardous ground shaking, with respect to some event such as
structural failure of a particular structure at the site where the ground motion was observed, or
liquefaction at the site which may initiate foundation failure, which can be characterized by
different patterns, depending on the criticality condition adopted. For example, a pattern
indicating hazardous ground shaking can be the peak acceleration or the spectral acceleration
exceeding some specified (design) level, or the total energy of ground shaking exceeding some
level (which depends on a pattern of velocity and duration of shaking, or on a pattern of the
Fourier amplitude spectrum). Then according to these patterns, the ground shaking can be
classified as hazardous or not. It should be noted here that the classification of the data could be
hard, i.e. with deterministically defined boundaries and exclusive membership in a class, or soft,
i.e. with fuzzy boundaries between the classes and membership defined by a probability
distribution function.
There are different levels of data mining, e.g., (a) undirected or pure data mining, (b)
directed data mining, and (c) hypothesis testing and refinement. The purpose of undirected data
mining is to find totally unexpected patterns in the data, and it has no strict constrains. The
50
directed data mining is more constrained—the user looks for patterns in subsets of the data, and
draws conclusions about the subset by induction. The hypothesis testing and refinement is even
more specific—the user has a specific hypothesis (model), and uses the data to validate, reject or
refine the hypothesis. The common architecture of all of these is the closed loop consisting of a
database, search for patterns, knowledge discovery, and a feedback loop of refinements at each
stage, and some exit criterion to terminate the loop. The refinements at the database level may
consist of cleaning the data, adding new data by modifying the experiment based on the
knowledge derived from the previous loop, and adding new attributes to the data (e.g. supporting
information about the site conditions where the time series have been recorded, or about the type
of structure the response of which has been recorded). At the search for patterns level, the
definition of the patterns themselves or of the search algorithms may be refined. Similarly, at the
knowledge discovery level, refinements may be introduced in the post-processing of the outcome
of the search for patterns, e.g. some specific visualization of the outcome, which may improve
the knowledge discovery.
These examples mentioned earlier in this section already indicate that the patterns
characterizing some concept or object may be more naturally defined in some transformed
domain, rather than in the domain where the data was recorded. In the case of mechanical
vibrations, the data is typically recorded in the time domain, but is commonly Fourier
transformed and analyzed in the frequency domain. In applications that require preserving the
time localization, the data is transformed by some time-frequency distribution, e.g. by the
wavelet transform or the Gabor transform. Transformation of the data is part of the
preprocessing stage.
Another important element of preprocessing is information granulation, which refers to

reducing the dimensionality of the data by encapsulating numeric data, e.g. in an interval, into a
single conceptual entity. Information granulation is necessary for several reasons. Firstly, while
data is usually recorded as some long sequence of numbers, it is the nature of the human mind to
process information by reducing this longer sequence of numbers into fewer groups or intervals,
and associating to each interval some concept, which represents some higher level of abstraction
but which is more naturally related to the phenomenon of interest. For example, the sequence of
amplitudes of recorded vibrations can be granulated by classification into intervals, which, at a
higher level of abstraction, are viewed as representing small, moderate or large response.
Information granulation also often consists of aggregating the information by computing local
and global averages, for example, representative of a segment of the domain or of the entire
domain. For example, one such aggregate granule of information may be the average amplitude
of the signal, the root mean square amplitude, which is related to the total energy of the signal, or
energy of the signal contained in different frequency subbands. Information granulation is also a
practical necessity to avoid a combinatorial explosion in analysis of large sets of data, but may
51
be also convenient in comparing different patterns, i.e. in defining distance between patterns.
This report explores the use of the discrete wavelet transform as a means for information
granulation and definition of patterns in data mining through large set of mechanical vibration
data, such as earthquake response data. As it will be seen later in this chapter, the orthonormal
discrete wavelet transform appears to be a very convenient tool for representing the signal
energy, both on a local and on a global scale.
Data can be characterized (explained) in different ways by a variety of models. Examples of

such models are: clustering, regression, classification, summarization, link analysis and sequence
analysis. Cluster analysis is concerned with finding groups in data, and hence, reveals a
structure in a database. It is a form of unsupervised learning, and is a part of the field of pattern
recognition. Regression analysis is concerned with constructing functional models that explain
the data. The functional forms have to be assumed a priori, while the unknown parameters of the
model are determined by solving an optimization problem (usually by minimization of the mean
square error). Hence, regression analysis is a form of supervised learning. Its applied part is
known as system identification. Classification analysis is concerned with classification of data
in predetermined categories. It can be viewed as a special case of regression analysis such that
the function that explains the data has only a finite number of values. Summarization is
concerned with characterization of the data by a small number attributes, such as aggregates or
averages. Because it does not involve an a priori assumption of a model, it is a form of
unsupervised learning. Link analysis is concerned with revealing relationships (dependencies)
among the data, such as correlations. Finally, sequence analysis deals with modeling sequential
data, and uses methods such as, e.g., time series analysis and temporal neural networks. The
techniques used to analyze these models have been developed in different branches of science
and engineering, such as mathematical statistics, systems theory, signal processing, and artificial
intelligence. A crucial requirement, however, for the success of the modeling is knowledge of
the domain of the data and a clear understanding of the application.
Finally, data mining and knowledge discovery are highly interdisciplinary, and involve a
fusion of a variety of methods traditionally developed in many different research fields, such as:
databases, statistics, rule-based systems, visualization, high-performance computing, fuzzy sets,
knowledge representation, machine learning, rough sets, evolutionary computing, and neural
networks.
4.2 Wavelet Transformed Database
The following sections describe how the discrete wavelet transform of a time series data can
be used as a tool in data mining, in particular for reducing the dimensionality of the data, as well
as for particular characterization of the data (feature selection) and pattern recognition. We
assume that the database has been transformed, and are interested, in particular, in features that
52
can be obtained directly in the wavelet domain, by simple manipulation of the wavelet
coefficients, such that can be done at a database level. The record stored in the database for each
discrete time series s [ n] , n = 1,..., N is the set of N coefficients of orthonormal wavelet
expansion, {d j ,k }k =1,..., N / 2 j and {s J ,k }

N / 2J
, of the signal in a chosen orthonormal wavelet basis
k =1
j =1,..., J
{ψ j ,k [ n]}k =1,...,N / 2 ∪ {ϕ J ,k [ n]}k =1,..., N / 2

j J for agreed level of expansion J , such that
j =1,..., J
J N /2j N / 2J
s [ n] = ∑ ∑ d j ,kψ j,k [ n] + ∑ sJ ,kϕ J ,k [ n ], j = 1,..., J
j =1 k =1 k =1
(4.2.1)
It is understood that the discrete time series s [ n] , n = 1,..., N is a sampled version of a

continuous time signal s ( t ) at sampling interval ∆t . Each coefficient of the expansion is the
orthogonal projection of the signal onto the corresponding basis function, and hence is a measure
of how similar the signal is to that basis function. The basis functions are localized both in time
and in frequency, effectively within a tile of the time-frequency plane. Figure 4.2.1 shows a
multiresolution partition of the time-frequency plane in non-overlapping tiles. In reality, and
even theoretically, the time-frequency distributions of the basis functions extend beyond the
boundaries of the corresponding tiles, but most of their energy is within the tile. The leakage of
energy in the neighboring tiles depends on the nature of the basis. Extreme cases are the Haar
basis, for which there is no leakage of energy in the neighboring tiles along the time axis but
there is leakage along the frequency axis, and the sinc basis, for which there is no leakage of
energy in the neighboring tiles along the frequency axis but there is leakage in the neighboring
tiles along the time axis. The compactly supported wavelets, such as the s8 wavelet used in the
examples in this report, are localized reasonably well both in time and in frequency.
It can be seen from Fig. 4.2.1 that, for smaller frequencies, the tiles become wider and
shorter, while the area of each tile remains always constant and equal to π . For the detail
coefficient d j ,k , the tile is a rectangle with width ∆t = 2 j and height ∆ω = π / 2 j , and centered
at time n = 2 j ( k + 1/ 2 ) and at frequency ω = 3π / 2 j +1 . For the smooth coefficient s J ,k , the tile
is also a rectangle with width ∆t = 2 J and height ∆ω = π / 2 J , and centered at time
n = 2 J ( k + 1/ 2 ) and at frequency ω = π / 2 J +1 .
53
Multiresolution partition of the time-frequency plane
ω for a J=4 level expansion
D1
π
2
D2
π
22
d 3,5 D3
π
23 D4
π
24 s 4,2 S4
0
t
Tile for coefficient d j,k : Tile for coefficient s J,k :
2j
2J
π 3π π π
j J
2 2 j+1
2 2 j+1
1 J 1
j
2 (k + ) 2 (k + )
2 2
Fig. 4.2.1. A multiresolution division of the time-frequency plane. Note that the half plane is sufficient
to represent the signal.
54
The N coefficients of orthonormal wavelet expansion represent all the information about
the signal, and the signal can be exactly reconstructed (i.e. up to the precision of the machine)
using the inverse wavelet transform, which can be practically implemented for compactly
supported wavelets by the pyramid algorithm. Hence, the set of all of the wavelet coefficients
represents the signal at the finest information granulation level.
While the coefficients for fixed j represent a sub-sampled version of the subband
components, the subband components as functions of time can be reconstructed at the initial
sampling rate of the signal by synthesis of the corresponding coefficients
N /2j
D j [ n] = ∑ d j ,kψ j ,k [ n], j = 1,..., J
k =1
(4.2.2)
N / 2J
S j [ n] = ∑ s j ,k ϕ j , k [ n ]
k =1
which can be also accomplished by the inverse pyramid algorithm, by setting to zero all other
coefficients.
4.3 Reducing Data Dimensionality by Compression
Data compression is reduction of the size of a signal while preserving its significant features
or most of the energy. The wavelet basis expansion and its relative—the wavelet packet
expansion are very efficient for compression of non-stationary time series, such as strong motion
earthquake records. As it will be seen in the illustrations in the next chapters of this report,
especially for structural response records, dramatic compression rates are achieved (i.e. reduction
of size) by preserving most of the energy.
Data compression is accomplished by expanding the signal in a wavelet basis, and dropping
the coefficients of the expansion that are not significant, based on an adopted rule, depending on
the specific application. One rule may consist of creating a lower resolution approximation of the
signal, by dropping the detail coefficients that correspond to the smaller scales, considering them
as noise. This may be desirable in applications in which the features at the coarser scales are
considered as those defining the signal, while the details are considered not so important or
contaminated with noise. This rule is applied, e.g. in creating thumbnails for images. Another
rule is the one based on thresholding, which consists of defining some application specific or
universal threshold level, and dropping the coefficients that have magnitude smaller than that
level. This rule may be desirable when the signal may have some significant higher resolution
features (e.g. some singularities) that need to be preserved. Then this method will preserve the
high-resolution features associated with a large wavelet coefficient, while still dropping the less
significant features. The threshold can be a specified as a single value (hard thresholding) or as a
55
range, such that below that range, all coefficients are dropped, while within the range, the
coefficients are gradually “shrunk” (soft thresholding), which may reduce some artifacts
resulting from the compression. Depending on specific applications, other thresholding schemes
can be defined, e.g. such that have subband-specific threshold levels.
Data compression is closely related to nonparametric estimation and noise removal using
wavelets (Bruce and Gao, 1996), as both are based on the same principle of “shrinking” the less
significant coefficients. These were pioneered by Donoho at Johnston at Stanford University
(Donoho and Johnstone, 1994; Donoho, 1995).
For discrete sequences, data compression or nonparametric estimation can be formally

defined as creating and approximation s [ n] of the signal s [ n] such that
J 2j 2j
s [ n ] = ∑∑ d j ,kψ j ,k [ n ] + ∑ sJ ,k ϕ J ,k [ n ] (4.3.1)
j =1 k =1 k =1
where the new coefficients of the expansion d j ,k and s J ,k are
( )
d j ,k = δ d j ,k
(4.3.2)
s J ,k = δ ( s J ,k )
where the adopted threshold function δ ( x ) depends on the adopted threshold scheme. The
remaining part of this section will show examples of threshold functions δ ( x ) .
The degree of compression can be measured by the compression rate, herein defined as
N
Compression rate = 1 − (4.3.3)
N
where N is the number of real numbers used to represent the approximation (i.e. the coefficients
in the wavelet expansion that are different from zero) and the number of real numbers that
represents the discrete signal exactly, i.e. the length of the signal, N. In digital signal and image
processing, the compression ratio is defined via the ratio of the number of bits used to represent
the original signal and the number of bits used to represent the compressed and quantized and
coded signal. Quantization consists of dividing the range of possible values of the coefficients in
discrete levels, and assigning the coefficient to one of these discrete levels, and coding consists
of assigning each level a unique string of bits, with length that it is shorter for the intervals
occurring most often. Quantization and coding are out of the scope of the analysis in this report,
56
and here the compression is achieved only by describing a signal in the DWT domain by less real
numbers than the original length, while preserving most of its significant features.
The error associated with compression is the vector
Error [ n] = s [ n ] − s [ n] (4.3.4)
and can be measured using its norm, based on different Minkowski distances between the
original signal and its approximation, e.g. the Tschebyshev distance
Error ∞
= max sn − sn (4.3.5)
n
the Eucledian distance
∑ ( sn − sn )
2
Error 2
= (4.3.6)
n
or the Hamming distance
Error 1 = ∑ sn − sn (4.3.6)
n
where sn and sn are the n-th samples of the original signal and of its approximation. The
Tschebyshev distance gives the peak error. Based on the Eucledian norm and the Hamming
norm, one can define mean square error and mean absolute error
Mean Square Error = ∑ sn − sn

2
N
n (4.3.7)
( )
2
= Error 2
N
Mean Absolute Error = ∑ sn − sn N

n (4.3.8)
= Error 1 N
The error can also be expressed in terms of some measure of the variability of the signal itself,
by an appropriate normalization. As normalization factors, one can use the Tschebyshev norm of
the signal
s ∞
= max sn (4.3.9)
n
57
its Eucledian norm
∑ ( sn )
2
s 2
=
n
its Hamming norm
s 1 = ∑ sn (4.3.10)
n
or its standard deviation
∑ ( si − s ) / ( N − 1)
2
σs = (4.3.11)
i
where
s = ∑ sn N (4.3.12)
n
is the mean value of the signal. For example, in the following chapter, we will consider the
following normalized measures of the error:

Error ∞
,
Error ∞
,
( Mean Square Error )1/ 2 , and
σs s ∞
σs
Error 2
.
s 2
4.3.1 Data Compression by Lower Resolution Approximation
In data compression by lower resolution (coarser) approximation, the threshold function is
0, if j ≤ J
δ ( d j ,k ) =  (4.3.1.1)
d j ,k , if j > J
where J is the level of the desired smooth approximation of the signal. The compressed signal
is then the level J smooth approximation, represented by the corresponding smooth coefficients
sJ ,k , k = 1,..., N / 2 J , and reconstructed as
58
N / 2J
s [ n] = ∑ s J ,k ϕ J , k [ n ] (4.3.1.2)
k =1
If the signal has already been expanded in a level J > J expansion, the compressed signal is
represented by the higher-level detail coefficients d j ,k , k = 1,..., N / 2 j ; j = J + 1,..., J and by the
highest-level smooth coefficients sJ ,k , k = 1,..., N / 2 J
J N /2j N / 2J
s [n] = ∑ ∑ d j ,kψ j ,k [ n ] + ∑ s J ,k ϕ J ,k [ n ] (4.3.1.3)
j = J +1 k =1 k =1
Such a compression scheme would reduce the number of real numbers representing the signal
from N to N / 2 J , and the compression rate would be
N / 2J
Compression rate = 1 −
N (4.3.1.4)
1
= 1−
2J
and reduction of size by a factor of 2 for a level 1 approximation, by a factor of 4 for a level 2
approximation, by a factor of 8 for a level 3 approximation, etc. The amount of energy of the
signal that is lost by the compression is
J N /2j 2
Energy lost by the compression = ∑∑ d j ,k
j =1 k =1
4.3.2 Data Compression by Thresholding
In data compression by hard thresholding, the threshold function is
0, if x ≤ λ
δ ( x) =  (4.3.2.1)
 x, if x > λ
and for semi-soft thresholding it is
59
0, if x ≤ λ1

 λ2 ( x − λ1 )
δ ( x ) = sgn ( x ) , if λ1 < x ≤ λ2 (4.3.2.2)
 λ 2 − λ1
 x, if x > λ2

where λ , λ1 and λ2 are adopted threshold levels (Bruce and Gao, 1996). Semi-soft
thresholding may be desirable when the hard thresholding creates artifacts in the approximation
of the signal. These threshold functions are illustrated in Fig. 4.3.2.1. For hard threshoding,
e.g., this results in preserving only the first ntop of the coefficients, ordered in decreasing
magnitude, such that have magnitude larger than the threshold level, and the compression rate is
ntop
Compression rate = 1 − (4.3.2.3)
N
The energy lost by the compression is the sum of the square of the coefficients below the first
ntop coefficients.
Hard δ (x) Semi-soft δ (x)

thresholding thresholding
−λ −λ 2 −λ 1
λ x λ1 λ2 x
Fig. 4.3.2.1 Graphical representation of the threshold functions defined in eqns (4.3.2.1) and (4.3.2.2).
60
4.4 Estimation of Energy and Power— Global and Local Aggregates and Averages
This section deals with estimation of energy and power at different granulation levels, all in
the wavelet domain.
4.4.1 Single Value Estimates
On a global scale, i.e. at the coarsest granulation level, the single value aggregate
representing the signal is its total energy, which can be used to compute the average power over
the entire length of the signal. For orthonormal expansions, by the Parseval equality (eqn
(3.10.2)), the energy can be computed entirely in the wavelet domain by
∞ J N /2j 2 N / 2J
2
∑ s [ n]
2
Total Energy = E = =∑ ∑ d j ,k + ∑ s J ,k (4.4.1.1)
n =−∞ j =1 k =1 k =1
and the average power (energy per unit time) by
E 1  J N /2 2
j J
2 N /2
Average Power = =  ∑ ∑ d j ,k + ∑ sJ ,k  (4.4.1.2)
N N  j =1 k =1 k =1 
Another related single value quantity is the root-mean-square value of the signal, which is the
value such that its square multiplied by the number of points equals the total energy of the signal,
and which can be computed as follows
r.m.s. value = E / N (4.4.1.3)
4.4.2 Nodal Spectra
At a finer granulation level, the energy of the signal can be represented by its nodal energy
spectrum, which is a discrete spectrum consisting of impulses with amplitudes equal to the
energy of the subband components, and with nodes at the central frequencies of the subbands.
Hence, the nodal energy spectrum is the collection of ( J + 1) real numbers
 N /2j 2
 ED , j = ∑ d j , k , j = 1,..., J
 k =1
Nodal Energy Spectrum =  (4.4.2.1)
N / 2J
 2
 ES , J = ∑ s J , k
 k =1
61
where ED j is the energy in detail subband at level j , and ES J is the energy in the last smooth
subband J . This nodal spectrum, also called discrete wavelet spectrum, has nodes equally
spaced on a logarithmic frequency axis, as illustrated in Figure 4.4.2.1.
Nodal Energy Spectrum
Energy
ω
0 ω
SJ
ω
DJ
... ω
D3
ω
D2
ω
D1
Fig. 4.4.2.1 Illustration of a Discrete wavelet spectrum of a signal.
The total energy equals the sum of the nodal energy spectrum
J
E = ∑ ED , j + ES , J (4.4.2.2)
j =1
which implies that, even for non-ideal filters for which there is a “fuzzy” boundary between the
subbands, the total energy of the signal is exactly partitioned among the subbands, which may be
an advantage in some applications.
The nodal energy spectrum, normalized by the length of the signal, gives the average power
within each of the subbands, and is named here nodal power spectrum
 N /2j 2
E ∑ d j ,k
 D j = k =1 , j = 1,..., J
 N N
Nodal Power Spectrum =  (4.4.2.3)
N / 2J
 2
E ∑ sJ ,k
 S J = k =1
 N N
62
4.4.3 Nodal Time-Frequency Energy and Power Distribution
At the finest granulation level, the signal energy is represented by its nodal time-frequency
energy distribution, which is the N -dimensional collection
E 2
= d j ,k , j = 1,... J ; k = 1,... N / 2 j
Nodal Time-Frequency  D ( j , k )
= (4.4.3.1)
Energy Distribution 2
 ES = s j ,k , k = 1,... N / 2 J
 ( J , k )
of the energies contained within the tiles of the time-frequency plane (see Fig. 4.2.1). This
collection represents a discrete two-dimensional distribution, with nodes at the centers of gravity
of the tiles, and can be visualized as a collection of finite amplitude impulses distributed
accordingly in the time-frequency plane. It can also be viewed as a nodal instantaneous energy
spectrum where the instants are granules of time encapsulating a time interval, the length of
which depends on the frequency.
The sum of the nodal instantaneous energy spectrum over all nodes/tiles equals exactly the
total energy of the signal
J N /2j N / 2J
E=∑ ∑ ED( j , k ) + ∑ E S( J , k ) (4.4.3.2)
j =1 k =1 k =1
and the sum over all tiles in a subband equals exactly the energy of the subband
N /2j
ED j = ∑ ED( j , k )
k =1
(4.4.3.3)
N / 2J
ES J = ∑ E S( J , k )
k =1
It should be noted here that, while their sum equals exactly the total energy of the signal, and the
tiles are non-overlapping and cover the entire time-frequency plane (over positive half-band),
strictly speaking, the nodal energies represent an aggregate of energy in the time frequency plane
that spreads beyond the boundaries of the corresponding tile, but this spread is small, and one
can say that they represent the energy effectively contained within the tile.
Next, by dividing the nodal time-frequency energy distribution by the width ∆n of the
corresponding tiles, one can get the average power of the signal within each tile, which is named
here nodal time-frequency power distribution
63
 ED d j ,k
2
 ( j ,k )
= , j = 1,... J ; k = 1,... N / 2 j
Nodal Time-Frequency  ∆n 2 j
= (4.4.3.4)
Power Distribution 2
 ES( J ,k ) s j ,k
 = , k = 1,... N / 2 J
 ∆ n 2 j
This distribution can also be interpreted as an instantaneous power spectrum.
4.4.4 Nodal Time Distributions
Finally, by computing the energy and power for lower resolution approximations of the
signal, one can define distributions versus time of the signal energy and power for different
granulation levels, i.e. represented by fewer granules the coarser the signal approximations. For
example, for a level J approximation, one can define a distribution of the aggregate energies in
each of the tiles of the J -th smooth subband, named here nodal energy time distribution, as the
following collection of N / 2 J positive real numbers
Nodal Energy Time Distribution J = ES { ( J ,k )

2
= sJ ,k , k = 1,..., N / 2 J (4.4.4.1)
where the subscript J indicates the level of the approximation. By dividing the energy within
each tile by the width of the tile, one can obtain the average power within the tile, and hence a
nodal time distribution of power, which can also be interpreted as instantaneous power versus
time
 ES s J ,k
2
 ( J ,k )
Nodal Instantaneous Power Distribution J =  = , k = 1,... N / 2 J (4.4.4.2)
 ∆n
J
2

For increasing level of approximation, J , the number of granules decreases while the length of
the tiles (i.e. the time intervals that are encapsulated) increases, by a factor of two for each
consecutive level of approximation. The finest granulation level corresponds to the J = 0 level
approximation of the signal, which is the time series of the original signal.
4.5 Estimation of Fourier Spectrum and Power Spectrum Density
The frequency content of a signal is most commonly expressed via its Fourier spectrum,
and, hence, it is of interest to find out how the Fourier spectrum can be estimated in the discrete
64
wavelet transform domain. In electrical engineering signal processing literature, Fourier
spectrum is defined as the square of the modulus of the Fourier transform, while in earthquake
engineering literature it is just the modulus of the Fourier transform. To avoid confusion,
wherever this distinction matters, we refer directly to the Fourier transform of the signal, sˆ eiω . ( )
The following describes how the Fourier spectrum is related to the discrete wavelet spectrum, i.e.
to the nodal energy spectrum, first for ideal low and high pass filters, and later for other filters.
4.5.1 Estimation of Fourier Spectrum for Ideal Filters
By the Parseval equality for the discrete Fourier transform, for ideal filters, the energy of the
signal in the j-th detail subband is
∫ ( )
1 2
ED j = sˆ eiω dω =
2π ID j
1
( )
2
= 2 ∫ sˆ eiω dω (4.5.1.1)
2π I +
Dj
1 
( )
2
= 2  sˆ eiω  ∆ωD+ j
2π  av. on I D+ j
where
I D j =  −π / 2 j −1 , −π / 2 j  ∪ π / 2 j , π / 2 j −1  (4.5.1.2)
is the entire support of the subband band-pass filter and
I D+ j = π / 2 j , π / 2 j −1  (4.5.1.3)
is the support in the positive half-band only, ∆ωD+ j = π / 2 j is the length of I D+ j , and

( ) ( )
2 2
 sˆ e
iω
 is the average of sˆ eiω on I +j . Further, substituting for ∆ω +j = π / 2 j in
 av. on I D+ j
the last line of eqn (4.5.1.1) gives
1 
( )
2
ED j = j 
sˆ eiω  (4.5.1.4)
2  av. on I D+ j
Similarly, for the J-th smooth subband we have
65
1 
( )
2
ES J = 2  sˆ eiω  ∆ωS+J
2π  av. on I S+J
(4.5.1.5)
1 
( )
2
= J  sˆ eiω 
2  av.on I S+J
where I S+j = 0, π / 2 J  .
Then, from eqns (4.5.1.4) and (4.5.1.5), one can compute the average Fourier spectrum
within each subband, and define a nodal Fourier spectrum of these averages as the collection

( )
2
iω
 sˆ e  , j = 1,..., J
 av.on I D+ j
Nodal Fourier Spectrum = 
 sˆ eiω
( )
2
 
av.on I S+J
 (4.5.1.6)
 j N /2j 2
2 ∑ d j ,k , j = 1,..., J
 k =1
= J
 J N /2 2
 2 ∑ s J ,k
 k =1
This granulated version of Fourier spectrum consists only of J + 1 values, as opposed to

infinitely many values on the real line, and hence would be a convenient representation for
regression analyses.
4.5.2 Estimation of Fourier Spectrum for Non-Ideal Filters
For non-ideal filters, the relationship between the subband energies and the Fourier spectrum
( )
2
is identical to the one for ideal filters, except that the average of sˆ eiω on the corresponding
half-band interval is substituted by the weighed average on ω ∈ [ 0, π ] . The weighing functions
( ) ( )
2 2
are G1( ) eiω / π , j = 1,..., J for the detail subbands, and G0( ) eiω
j J
/ π for the last smooth
approximation subband, where G1( ) eiω

j
( ) and G( ) ( e ω ) are the iterated filters defined by eqns
0
J i
(2.8.3), which are windows in the frequency domain. Also, for non-ideal filters, ∆ωD+ j = π / 2 j
66
and ∆ωS+J = π / 2 J are the effective bandwidths of the corresponding subbands. The proof for
this general case is as follows.
The energy for subband D j is an integral of the Fourier spectrum multiplied by a

corresponding window
2
π j
( )
G1( ) eiω
∫ sˆ ( e )
1 iω
2
ED j = dω (4.5.2.1)
2π
( 2)
j
−π
where the window function

j
( )
G1( ) eiω
is an iterated filter from true low and high-pass filters
( 2)
j
( )
G0 eiω / 2 ( )
and G1 eiω / 2 , i.e. such that the ( )
G0 eiω / 2 1 as ω → 0 and
G (e ω ) /
1
i
2 1 as ω → π . By rearranging the terms and from the symmetry of the Fourier
transform of real signals, eqn (4.5.2.1) can be rewritten as
( )
2
G1( ) eiω
j
π
π
1
( )
2
ED j = 2 ∫ sˆ eiω dω
2π 0 π 2j (4.5.2.2)
1 
( )
2
= 2  sˆ eiω  ∆ωD+ j ,eff.
2π   w. av. on [0,π ]
( )
2
G1( ) eiω
j
where is such that

π
( )
2
G1( ) eiω
j
π
∫ π
dω = 1 (4.5.2.3)
0
and hence is a weighing function on [ 0, π ] .
4.5.3 Estimation of Power Spectrum Density
Energy or Power Spectrum Density (PSD) was first introduced in the theory of random
processes to signify the Fourier transform of the autocorrelation sequence of a stationary or a
67
wide-sense stationary random process, but was later extended to signify the magnitude square of
the Fourier transform of any signal. The term “power” and “energy” in the name comes from its
relation to the total energy and power of the signal through the Parseval equality. Although
energy and power are often used interchangeably, here we will make a distinction between the
two, and define power spectrum density of a signal s [ n] of length N to be the quantity such that
π
E 1
Average Power = =
N 2π ∫ PSD (ω ) dω
−π
(4.5.3.1)
from where
sˆ ( eiω )
2
PSD (ω ) = (4.5.3.2)
N
Then, similarly as for the Fourier spectrum, one can define a nodal PSD as the collection of the
( J + 1) values, each representing the average PSD within each subband
 PSDD j , j = 1,..., J
Nodal Power Spectrum Density = 
 PSDS J
1
 N sˆ ( e ) av.on I D+ j , j = 1,..., J
iω
2
(4.5.3.3)
=
 1 sˆ ( eiω ) 2
 N av.on I S+
J
which can be computed entirely in the DWT domain. Substituting for the nodal Fourier spectrum
from eqn (4.5.1.6) implies
 N /2
j
 ∑ d j ,k
2
 k =1 , j = 1,..., J
 N / 2j
Nodal Power Spectrum Density =  J (4.5.3.4)
N /2

∑
2
sJ , k
k =1

 N / 2J
68
Recalling that N / 2 j is the total number of coefficients/tiles in detail subband D j , the nodal
2
PSD for this subband is equal to the average d j , k over all shifts k in this subband. Further
extending this reasoning to smooth subband S J , one can write
average d 2 in subband D , j = 1,..., J

 j ,k j
Nodal Power Spectrum Density =  2
(4.5.3.5)
average sJ , k in subband SJ
Equation (4.5.3.5) suggests that one can also define a nodal time-frequency distribution of
PSD, which will give the average PSD within each tile, and which can also be interpreted as
instantaneous PSD within a frequency band. This can again be done through the Parseval
equality written for each tile, according to which
2
d j ,k 1
Average Power in Tile D( j , k ) = = ∫ PSD (ω ) d ω
2π
j D( j ,k )
2 ID j (4.5.3.6)
1 π
= ( PSD )av. on tile D( j ,k ) 2 j
2π 2
where PSDD( j ,k ) is the PSD of the signal multiplied by the time window of tile ( j, k ) , which
implies
( PSD )av.on tile D(

2
= d j ,k (4.5.3.7)
j ,k )
After extending the same reasoning to the tiles of the last smooth subband, one can readily define
nodal time-frequency PSD distribution as the collection of the average values of the PSD of the
signal in each tile
( PSD )av.on tile D , j = 1,..., J ; k = 1,..., N / 2 j

Nodal Time-Frequency  ( j ,k )
= (4.5.3.8)
( PSD )av.on tile S( J ,k ) , k = 1,..., N / 2
PSD Distribution J
 d 2 , j = 1,..., J ; k = 1,..., N / 2 j
 j ,k
= 2
 sJ , k , k = 1,..., N / 2 J
69
This distribution can also be interpreted as instantaneous PSD, specified on a multiresolution
grid of nodes.
4.6 Estimation of Correlation — Global and Local Aggregates and Averages
The definition of correlation of two signals, originally introduced as a concept in analyses of

random signals, can be extended to deterministic signals, which are of interest in this report. Let
s ( ) [n] and s ( ) [ n] be two real discrete signals. Their correlation (or cross-correlation) is by
1 2
definition their inner product (Newland, 1993)
Cross-Correlation s ( ) , s (( 1 2)
)= s( ) , s(
1 2)
∞ (4.6.1)
∑ s( ) [ n] s( ) [ n]
1 1
=
n =−∞
and as such, it is equal to the projection of one of the signals onto the other times the norm of the
other signal, and is a measure of how similar the two signals are. Two signals differ most if they
are orthogonal, in which case their correlation is zero. Positive correlation means that, on the
average, they have the same phase, and negative correlation means that, on the average, they
have opposite phase.
The cross-correlation of a signal with itself is equal to its energy
( )
2
Cross-Correlation s ( ) , s ( ) = s ( )
1 1 1
(4.6.2)
= Energy s ( ) ( ) 1
Hence, the cross-correlation of two signals is also called cross-energy, while the energy of a
signal is called auto-correlation, and we introduce notation
E(
1,2 )
≡ Cross-Energy s ( ) , s ( )( 1 1
) (4.6.3)
From the Cauchy-Schwarz inequality (Vetterli and Kovacevic, 1995)
s( ) , s(
2)
≤ s( ) s(
1 1 2)
(4.6.4)
which implies that the upper bound for the absolute value of the cross-correlation is the product
of the norms of the two signals
70
(
Cross-Correlation s ( ) , s ( ) ≤ s ( ) s (
1 1
) 1 2)
(4.6.5)
which is achieved when the signals are perfectly “aligned.” Viewing the two sequences as
vectors in an abstract Hilbert space ( l2 —the space of square integrable sequences), one can
compute the “angle” α between the two sequences, such that
s( ) , s(
2)
= s( ) s(
2)
cos α
1 1
(4.6.6)
from where
s( ) , s(
1 2)
cos α = (4.6.7)
s( ) s(
1 2)
or using the “energy” notation
E(
1,2 )
cos α = (4.6.8)
E( )E(
1 2)
For two signals that are perfectly aligned (have same phase), α = 0 ; for signals that have
opposite phase, α = −π ; and for signals that are orthogonal, α = π / 2 or α = −π / 2 .
By the Parseval’s relation, the correlation can be computed in the frequency domain as
π
Cross-Correlation s ( ) , s ( ) = ( 1 1
) 1
2π −
∫π sˆ ( e ) sˆ ( e ) dω
() 1ω ( ) ω i 2 i
(4.6.9)
where the product of the Fourier transforms of the two signals is the cross-energy spectrum
density
1 1 1 2
(
Cross-Energy Spectrum Density s ( ) , s ( ) = sˆ( ) ( eiω ) sˆ( ) ( eiω ) ) (4.6.10)
Further, by dividing the energy cross-spectrum density by the length of the signals, one can
define cross-power spectrum density
Cross-Power Spectrum Density s ( ) , s ( ) = ( 1 1

) 1 (1) iω ( 2) iω
N
sˆ ( e ) sˆ ( e ) (4.6.11)
71
4.6.1 Single Value Estimates
The cross-correlation of two signals expanded in an orthonormal wavelet basis can be

conveniently computed in the wavelet domain as
(
Cross-Energy s ( ) , s (
1 2)
) ≡ E( 1,2 )
s( ) , s(
1 2)
J N /2j N / 2J (4.6.1.1)
= ∑ ∑ d j ,k d j ,k + (1) ( 2)
∑ sJ( ,)k sJ( ,k)
1 2
j =1 k =1 k =1
One can also define average cross-correlation of two signals per unit length (analogous to
power of a signal), or average cross-power as
Average Cross-Power s ( ) , s ( ( 1 2)
) = N1 s( ) , s(
1 2)
(4.6.1.2)
1  J N /2 1 2 N /2 1 2 
j J
=  ∑ ∑ d (j , k) d (j , k) + ∑ sJ( ,)k sJ( ,k) 

N  j =1 k =1 k =1 
4.6.2 Nodal Cross-Energy Spectrum
The terms on the right-hand-side of eqn (4.6.1.1) can be rearranged so that
( ) ≡ E(
J
Cross-Energy s ( ) , s ( = ∑ ED( , j ) + ES( , J )
1 2) 1,2 ) 1,2 1,2
(4.6.2.1)
j =1
where the terms on the right hand side define a nodal cross-energy spectrum
 (1,2 ) N / 2 j (1) ( 2 )
 ED , j = ∑ d j ,k d j , k , j = 1,..., J
1 2 
Nodal Cross-Energy Spectrum s( ) , s( ) =  (
 (1,2 )
k =1
N / 2J
(1) ( 2 )
) (4.6.2.2)
 ES , J = ∑ s J , k s J , k
 k =1
Hence, the sum of the nodal values of the cross-energy spectrum is equal to the cross-correlation
of the two signals. The physical interpretation of the nodal cross-spectrum follows from the fact
that the nodal values are also equal to the inner product of the corresponding subband
components of the two signals
72
 (1,2 ) ∞
E
 D, j = ∑ D(j1) [ n] D(j2) [ n], j = 1,..., J
1 2 
Nodal Cross-Energy Spectrum s ( ) , s ( ) = 
 E (1,2 ) =
∞ (
n =−∞
) (4.6.2.3)
∑ S J( ) [ n] S J( ) [ n]
1 2
 S ,J
 n =−∞
Hence, the nodal values of the cross-spectrum represent the correlation of the corresponding
subband components of the two signals.
Viewing the subband components as signals themselves, one can define the “angle” between
the corresponding subband components of two signals, using eqn (4.6.5). Then for the detail
subbands, the angle α D , j is such that
ED( , j )
1,2
cosα D,j =
ED( ,) j ED( ,)j
1 2
N /2j
(4.6.2.4)
∑ d (j ,k) d (j ,k)
1 2
k =1
= , j = 1,..., J
∑ ( d (j ,k) ) ∑ ( d (j ,k) )
j j
N /2 2 N /2 2
1 2
k =1 k =1
and for the J -the approximation
ES( , J )
1,2
cosα S , J =
ES( ,)J ES( , J)
1 2
N / 2J
s J( ,)k s J( ,k) (4.6.2.5)
∑
1 2
k =1
=
∑( ) ∑( )
J J
N /2 2 N /2 2
(1)
sJ( ,k)
2
sJ ,k
k =1 k =1
The average of the nodal energy cross-spectrum per unit length of the signal defines the
nodal cross-power spectrum
73
 (1,2 ) 1 N /2j
∑ d (j ,k) d (j ,k) ,
1 2
 ED , j = j = 1,..., J
( )
(1) ( 2 )  N k =1
Nodal Cross-Power Spectrum s , s = (4.6.2.6)
N / 2J
 (1,2 ) 1
∑ sJ( ,)k sJ( ,k)
1 2
 ES , J = N
 k =1
The sum of the nodal values of the power cross-spectrum is equal to the average correlation per
unit length of the two signals.
To simplify the notation, further on, it will be understood that all the quantities related to
correlation refer to signals s ( ) and s ( ) , and the argument s ( ) , s (
1 2
( 1 2)
) will be omitted.
4.6.3 Nodal Time-Frequency Distribution of Cross-Correlation
Each term in the sum on the right-hand-side of eqn (4.6.2.2), which comprises a nodal value
of the energy cross-spectrum, represents the cross-correlation of the corresponding subband
components of the two signals within a time window, and the collection of the cross-correlations
for the time windows for all the subbands comprises a nodal time-frequency distribution of
cross-correlation
 E (1,2 ) = d (1) d ( 2) , j = 1,... J ; k = 1,... N / 2 j

Nodal Time-Frequency  D( j ,k ) j ,k j ,k
= (4.6.3.1)
Distribution of Cross-Energy
 ES J ,k = sJ( ,)k sJ( ,k) , k = 1,... N / 2 J
1 2
 ( )
Similarly as in the case of the nodal time-frequency distribution of the energy of a signal, each
nodal value of the time-frequency distribution of cross-correlation of two signals represents the
cross-correlation of the components of the signals within the corresponding tile of the time-
frequency plane. This distribution can also be viewed as instantaneous cross-energy spectrum,
defined on a multi-resolution grid.
One can also define nodal instantaneous cross-power spectrum by dividing the cross-
correlation in each tile by the length of the time window corresponding to that tile
 (1,2 ) 1 1
ED = j d (j ,k) d (j ,k) ,
2
 j = 1,... J ; k = 1,... N / 2 j
Nodal Instantaneous  ( ) 2 j , k
= (4.6.3.2)
Cross-Power Spectrum 1 (1) ( 2)
E J
S( J , k ) = J s J , k sJ ,k , k = 1,... N / 2
 2
74
4.6.4 Nodal Cross-Power Spectrum Density
The cross-power spectrum density, defined by eqn (4.6.5), can also be granulized into a
nodal cross-power spectrum density by substituting for the Fourier transform amplitudes of the
two signals their nodal Fourier transform amplitudes, equal to their average on the corresponding
subband
 PCSDD j , j = 1,..., J
Nodal Cross-Power Spectrum Density = 
 PCSDS J
 1 (1) iω
 N sˆ ( e ) av.on I D+ sˆ ( e ) av.on I D+ , j = 1,..., J
( 2 ) iω (4.6.4.1)
=
j j
 1 sˆ(1) ( eiω ) sˆ( ) ( eiω )

2
 N +
av.on I S
J
av.on I S+
J
Further, substituting in eqn (4.6.3.1) for the nodal Fourier spectra from eqn (4.5.1.6) gives
 N / 2 (1) ( 2)
j
 ∑ d j ,k d j ,k
 k =1 , j = 1,..., J
 N / 2j
Nodal Cross-Power Spectrum Density =  J (4.6.4.2)
N /2

∑ sJ( ,)k sJ( ,k)
1 2
k =1

 N / 2J
Recalling that N / 2 j is the total number of coefficients (and also tiles) in detail subband D j , the
nodal value of the power cross-spectrum density for this subband is equal to the average of the
product d (j , k) d (j ,k) over all shifts k in this subband. Similarly, the nodal value for the smooth
1 2
subband S J is the average of sJ( ,)k sJ( , k) over all shifts k in this subband, and one can write
1 2
average d (j1,k) d (j 2, k) in subband D j , j = 1,..., J

Nodal Cross-Power Spectrum Density =  (4.6.4.3)
(1) ( 2 )
average sJ , k sJ ,k in subband SJ
This interpretation suggests defining instantaneous nodal cross-power spectrum distribution
75
( PCSD )av.on tile D , j = 1,..., J ; k = 1,..., N / 2 j
Instantaneous Nodal  ( j ,k )
= (4.6.4.4)
( PCSD )av.on tile S( J ,k ) , k = 1,..., N / 2
Cross-PowerSpectrum Distribution J
d (j1, k) d (j 2,k) , j = 1,..., J ; k = 1,..., N / 2 j

= 1 2
() ( )
 sJ ,k sJ , k , k = 1,..., N / 2
J
4.7 Estimation of Time Derivatives
Because of the narrow band nature of the wavelets, it is possible to estimate approximately
the wavelet transform of derivatives of a signal from the wavelet transform of the signal itself, as
follows. Let Ws ( a, b ) be the wavelet transform of a signal s ( t ) , and Ws ( a, b ) be the wavelet
transform of its derivative s ( t ) , where a is the scale variable and b is the shift. From the
definition of wavelet transform and the Parseval equality
∞
1
Ws ( a, b ) = ψ a ,b , s = ∫ s ( Ω )ψˆ ( Ω ) d Ω
*
(4.7.1)
2π
a ,b
−∞
and
∞
1
Ws ( a, b ) = ψ a ,b , s = ∫ sˆ ( Ω )ψˆ ( Ω ) d Ω
*
(4.7.2)
2π
a ,b
−∞
Recalling that
sˆ ( Ω ) = i Ω sˆ ( Ω ) (4.7.3)
and that ψˆ a ,b ( Ω ) is narrow band, and has significant amplitudes only near the central frequency
of the wavelet Ω a ,b , the integrand in the right-hand-side of eqn (4.6.2) also has significant
amplitudes only near Ω a ,b , and the integral can be approximated by
76
∞
1
Ws ( a, b ) = ∫ iΩ sˆ ( Ω )ψˆ ( Ω ) d Ω
*
2π
a ,b
−∞
∞
(4.7.4)
1
≈ iΩ a , b ∫ sˆ ( Ω )ψˆ ( Ω ) d Ω
*
2π
a ,b
−∞
which implies
Ws ( a, b ) ≈ iΩa ,bWs ( a, b ) (4.7.5)
Similarly
Ws ( a, b ) ≈ Ωa2 ,bWs ( a, b ) (4.7.6)
Although this relationship is approximate, it is extremely convenient as it reduces the amount of

computation to estimate the wavelet transform of the derivatives only to multiplication by a
scalar, provided the wavelet transform of the signal is already available. One use of this
relationship is in estimation of nodal spectra (energy or Fourier) of velocity and displacement
from the corresponding nodal spectrum of absolute acceleration, which is the quantity usually
recorded by strong motion instruments.
Equations (4.4.2.1) and (4.7.5) and (4.7.6) imply that the nodal energy spectra of s and of s
are related to the nodal energy spectrum of s by
ωD2 j ED j , j = 1,..., J
Nodal Energy Spectrum s =  2
ωS J EDJ (4.7.7)
and
ωD4 j ED j , j = 1,..., J
Nodal Energy Spectrum s =  4 (4.7.8)
ωS J EDJ
Similarly, eqns (4.5.1.6) and (4.7.5) and (4.7.6) imply that he nodal Fourier spectra (defined as
the square of the magnitude of the Fourier transform) of the derivatives of the signal are
77
ωD2 j 2 j ED j , j = 1,..., J
Nodal Fourier Spectrum s =  2 J (4.7.9)
ωSJ 2 EDJ
and
ωD4 j 2 j ED j , j = 1,..., J
Nodal Fourier Spectrum s =  4 J (4.7.10)
ωSJ 2 EDJ
From eqns (4.5.3.4) and (4.7.5) and (4.7.6), the Nodal Power Spectrum Density of the derivatives
is
ωD2 j ED j 2 j / N , j = 1,..., J
Nodal Power Spectrum Density s =  2 (4.7.11)
ωS J EDJ 2 / N
J
and
ωD4 j ED j 2 j / N , j = 1,..., J
Nodal Power Spectrum Density s =  4 (4.7.12)
ωSJ EDJ 2 / N
J
Finally, the nodal time-frequency distribution of power spectrum density of the derivatives is
ω 2 d 2 , j = 1,..., J ; k = 1,..., N / 2 j
 D j j ,k
Nodal Time-Frequency PSD Distribution s =  2
(4.7.13)
ωS2J sJ ,k , k = 1,..., N / 2 J
and
ω 4 d 2 , j = 1,..., J ; k = 1,..., N / 2 j
 D j j ,k
Nodal Time-Frequency PSD Distribution s =  2
(4.7.14)
ωS4J sJ ,k , k = 1,..., N / 2 J
Finally, one can define a nodal transfer-function by dividing the nodal spectrum of the output by
the nodal spectrum of the input to the system of interest.
78
4.8 Estimation of Continuous-Time Signals
Discrete-time signals are often representations of continuous time processes, which can be
recorded and analyzed by digital computers only as finite length discrete-time signals. While the
computations for such signals are carried out in discrete time, ultimately of interest is the
characterization of the corresponding continuous time signal and the physical process it
represents. The following derives the relationships between the continuous time and discrete
time estimates of energy, power, Fourier transform and power spectrum density, as well as of
cross-energy, cross-power and cross power spectrum density. These relationships are derived for
the global aggregates and averages, but apply also to the corresponding spectral and time-
frequency estimates.
Let s ( t ) be a continuous-time signal, band-limited below its Nyquist frequency

Ω Nyquist = π / ∆t , and s∆t [ n ] be its sampled version at equally spaced time intervals ∆t . The two
are related by
s∆t [ n] = s ( n ∆t ) (4.8.1)
Let E , ŝ ( Ω ) and PSD ( Ω ) be the energy, Fourier transform and power spectrum density of the
continuous-time signal, and E∆t , sˆ∆t ( Ω ) and PSD∆t ( Ω ) be the corresponding quantities for its
sampled version. Consistently with the rest of this report, Ω is circular frequency in rad/s and
ω is circular frequency in rad/sample, and the two are related by
Ω = ω / ∆t (4.8.2)
4.8.1 Energy, Power and Root Mean Square Value
By definition
∫ s (t )
2
E= dt (4.8.1.1)
−∞
and
∑ s [ n]
2
E∆t = (4.8.1.2)
n =−∞
79
After representing the integral in eqn (4.7.1.1) as a sum of integrals on segments of length ∆t ,
and approximating s ( t ) on each segment by the samples s [ n] , it follows
∫ s (t )
2
E= dt
−∞
∞ ( n +1) ∆t
∑ ∫ s ( t ) dt
2
=
n =−∞ n∆t (4.8.1.3)
∞
∑ s ( n ∆t )
2
≈ ∆t
n =−∞
∞
∑ s [ n]
2
= ∆t = E∆t ∆t
n =−∞
Hence
E ≈ E∆t ∆t (4.8.1.4)
This relationship can be extended to the nodal energy spectrum and to the nodal energy time-
frequency distribution.
By definition, power is energy per unit time. In continuous and in discrete time, the average
power of the signal is
E
Average Power = (4.8.1.5)
N ∆t
and
E∆t
Average Power∆t = (4.8.1.6)
N
By substituting in eqn (4.8.1.5) for E from eqn (4.8.1.4) it follows
Average Power ≈ Average Power ∆t (4.8.1.7)
The same relationship between continuous-time and discrete-time average power holds also for
the average power in the tiles of the time-frequency plane, and power in general.
80
The root-mean-square value of the signal is essentially the square root of the average power
(see eqn (4.4.1.3)). Hence, from eqn (4.8.1.7)
r.m.s. value ≈ r.m.s. value ∆t (4.8.1.8)
4.8.2 Fourier Transform
By definition
∞
sˆ ( Ω ) = ∫ s (t ) e
− iΩt
dt (4.8.2.1)
−∞
and
∞
( ) ∑ s [n] e
sˆ∆t eiω = − iω n
(4.8.2.2)
n =−∞
By representing the integral in eqn (4.8.2.1) by a sum of integrals on the segments, and
approximating s ( t ) on the segments by its sampled version and e − iΩt by e − i Ω n∆t , and recalling
eqn (4.8.2)
( )
sˆ ( Ω ) ≈ sˆ∆t eiω ∆t (4.8.2.3)
and
( ) ( ∆t )
2
sˆ ( Ω ) ≈ sˆ∆t eiω
2 2
(4.8.2.4)
The relationship in eqn (4.8.2.4) also applies to the nodal Fourier spectra.
4.8.3 Power Spectrum Density
By definition
PSD ( Ω ) = s ( Ω )
2
(4.8.3.1)
and
81
( )
2
PSD∆t (ω ) = s∆t eiω (4.8.3.2)
Recalling eqn (4.8.2.4) it follows
PSD ( Ω ) ≈ PSD∆t (ω )( ∆t )
2
(4.8.3.3)
which is also valid for the nodal power spectrum densities.
4.8.4 Correlation, Cross-Power, and Power Cross-Spectrum Density
By the analogy of their definitions, the relations between continuous-time and discrete-time
correlation, cross-power, and power cross-spectrum density are the same as those for energy,
power, and power spectrum density. Hence
Cross-Correlation ≈ Cross-Correlation ∆t ∆t (4.8.4.1)
Average Cross-Power ≈ Average Cross-Power ∆t (4.8.4.2)
and
PCSD ( Ω ) ≈ PCSD∆t (ω )( ∆t )
2
(4.8.4.3)
82
5. ILLUSTRATIONS OF CONCEPTS
This chapter illustrates the method described in Chapter 3 and concepts introduced in
Chapter 4 on strong motion acceleration data recorded in a seven-story reinforced concrete
building in Van Nuys, California. The concepts are illustrated on the horizontal components of
the ground floor and roof response records of the January 17, 1994, Northridge earthquake
(Shakal et al., 1994). The building and the data are first briefly described, which is followed by
an illustration of the following concepts: DWT signature, multiresolution decomposition and
multiresolution approximation of the absolute acceleration at ground level and at the roof, and of
the relative acceleration of the roof with respect to the ground floor; total energy and average
power and their nodal spectra and time-frequency distributions; and efficiency of representation
and data compression. In all the illustrations, DWT refers to dyadic orthonormal wavelet
transform, and unless specified otherwise, the analyzing wavelet is the s8 wavelet. Most of the
results presented in this chapter were computed using S-plus (Bruce and Gao, 1996), and
checked against results by a Fortran code based on the wavelets subroutine from Numerical
Recipes (Press et al., 1994). The authors hope these illustrations will motivate more detailed
specialized analyses, which are out of the scope of this report.
The seven-story reinforced concrete building in Van Nuys, California has been the subject
of many studies (Blume and Assoc., 1973; Freeman and Honda, 1973; Islam, 1996; Ivanovic et
al., 1999; Trifunac et al., 2001a,b,c,d; 2003; De la Liera et al., 2001; Todorovska et al., 2001a,b),
and is an interesting case study because it was severely damaged by the Northridge earthquake
(ML=6.4, R=1.5 km) and its aftershocks, and because it recorded other earthquakes since 1968,
including the San Fernando earthquake of 1971, which also damaged the building.
5.1 Brief Description of the Building and a Simplified Model
The Van Nuys seven-story hotel is located in central San Fernando Valley of the Los
Angeles metropolitan area (at 34.221o N and 118.471o W), north-west from downtown Los
Angeles. Figure 5.1.1 shows San Fernando Valley and the building location, relative to the
major freeways and to the horizontal projections of the fault planes of 1971 San Fernando and
1994 Northridge earthquakes (Trifunac, 1974, Wald and Heaton, 1996), which both damaged the
building and were recorded in the building, This figure also shows the epicenters of two
Northridge aftershocks (shown by solid stars) and the directions and epicentral distances to seven
other earthquakes also recorded in the building (Trifunac et al., 1999a).
83
34 o 25 ' N
epicenter
S. Fernando, 1971:
1
Northridge, 1994: Northridge, 1994:
6 Dec. aftershock 2 sec
main event
Newhall 3
5 4
5 sec
5
6
4 6
3 7
8 sec
9
118 2 sec
o
o 22
-42 210
62 o v=2 km/s
o
1 , 1991 44 km
Sierra Madre
-90
Landers, 1992 186 km

epicenter SITE Big Bear , 1992
0 149 km
Canoga Park Pasa
Ch. 1 13 dena
,1 988
Reseda Wh 32 km
Northridge, 1994: M
on
i t t ie
r, 1 Burbank
km
20 March aftershock 101 te 98

be 7
35
llo 41
,1 km
98
89
34 09 ' N
o
9
, 19
Encino 0 1 2 mi 34
km
l ibu
118 39 ' W
o
405 1 2 km 118 o 18 ' W
Ma
Fig. 5.1.1 Central San Fernando Valley, California, and the site of the instrumented seven-story
reinforced concrete building. The dashed lines represent the horizontal projections of the faults of the
1971 San Fernando and 1994 Northridge earthquakes. The duration of faulting during these two
earthquakes was respectively 9 and 6 s. Directions and distances to seven other earthquakes recorded in
the building are shown by arrows. The epicenters of two Northridge aftershocks are shown by solid stars.
84
The building was designed in 1965, constructed in 1966 (Blume et al., 1973), and served as
a hotel at the time of the 1994 Northridge earthquake. Figure 5.1.2 shows (a) plan view of a
typical floor, (b) plan view of the foundation layout, (c) side view of the building frame, and (d)
typical soil-boring log data at the building site. The building is 18.9 m × 45.7 m in plan. The
typical framing consists of columns spaced at 6.1 m centers in the transverse direction and 5.8 m
centers in the longitudinal direction. Spandrel beams surround the perimeter of the structure.
Lateral forces in the EW (lateral) direction are resisted by interior column-slab frames and
exterior column spandrel beam frames. The added stiffness in the exterior frames associated with
the spandrel beams creates exterior frames that are roughly twice as stiff as interior frames. The
floor system is reinforced concrete flat slab, 25.4 cm thick at the second floor, 21.6 cm thick at
the third to seventh floors and 20.3 cm thick at the roof.
The building is situated on undifferentiated Holocene alluvium, uncemented and

unconsolidated, with thickness < 30 m, and age < 10,000 years. The average shear wave
velocity in the top 30 m of soil is 300 m/s. The soil-boring log in Fig. 5.1.2d shows that the
underlying soil consists primarily of fine sandy silts and silty fine sands. The foundation system
(Fig. 5.1.2b) consists of 96.5 cm deep pile caps, supported by groups of two to four poured-in-
place 61 cm diameter reinforced concrete friction piles. These are centered under the main
building columns. All the pile caps are connected by a grid of beams. Each pile is roughly 12.2
m long and has design capacity of over 444.82 × 103 N vertical load and up to 88.96 × 103 N
lateral load. The structure is constructed of normal weight reinforced concrete (Blume and
Assoc., 1973).
The ML = 6.6 San Fernando earthquake of February 9, 1971 (Fig. 5.1.1) (Trifunac, 1974)
caused minor structural damage (Blume and Assoc., 1973). Epoxy was used to repair spalled
concrete of the second floor beam column joints on the North side and East end of the building.
The recorded peak accelerations in the building were: 0.13g (L), 0.24g (T) and 0.18g (V) at the
base, and 0.32g (L), 0.39g (T) and 0.22g (V) at the roof, along the longitudinal (L), transverse
(T) and vertical (V) axes of symmetry.
The ML = 6.4 Northridge earthquake of 17 January 1994 (Fig. 5.1.1) severely damaged the
building, and it was declared unsafe by the Los Angeles Housing Authorities. The structural
damage was extensive in the exterior north (D) and south (A) frames, designed to take most of
the lateral load in the longitudinal direction. Severe shear cracks occurred at the middle columns
of frame A, near the contact with the spandrel beam of the 5th floor. Those cracks significantly
decreased the axial, moment, and shear capacity of the columns. The shear cracks which
appeared in the north (D) frame on the 3rd and 4th floors, and the damage of columns D2, D3 and
D4 on the 1st floor caused minor to moderate changes in the capacity of these structural elements.
No major damage of the interior longitudinal (B and C) frames was noticed. There was no visible
85
a)
1 2 3 4 5 6 7 8 9
D
C N
19.1 m
B
8 bays @ 5.72 m = 45.72 m
b)
1 2 3 4 5 6 7 8 9
19.1 m
B
N
8 bays @ 5.72 m = 45.72 m
Fig. 5.1.2 VN7SH building: (a) typical floor plan, and (b) foundation plan.
86
N Depth in m Elev. 243.2 m
0
Silty fine sand - brown,
6% - 102 3 SM
D C B A some gravel; dry
1.5
roof 20.3 cm slab
6% - 96 2 Fine sandy silt - light
ML
3.0 brown; slightly damp
2 Becoming more damp;
2.64 m
7th 21.6 cm slab
includes small pores
4.6
Silty fine sand - buff;
6% - 105 3 SM
6th 21.6 cm slab damp
6.1
Fine sandy silt - buff,
3 ML
some gravel and clay
7.6
5th 21.6 cm slab Silty very fine sand -
7% - 98 10* SM
light buff
9.1
8* Very fine sandy silt -
4th 21.6 cm slab ML
brown; slightly moist
20.03 m
10.7
5 x 2.65 m
11% - 103 8* Silty fine sand -
SM
3rd 21.6 cm slab
87
buff; damp
12.2
10*
2nd 25.4 cm slab 13.7

No water, no caving
Boring completed - 7/19/65
1st 10.2 cm slab 4.1 m LEGEND :

A Field moisture expressed as a percentage
of the dry weight soil.
B Dry density expressed in pounds per cubic foot.
6.35 m 6.35 m 6.35 m
C Blows per foot of penetration using a
19.1 m 3
17.8 x 10 N hammer dropping 30.5 cm
3
* - a 17.8 x 10 N hammer was used.
c) d) Depth at which undesturbed sample was extracted.
Fig. 5.1.2 (cont.) VN7SH building: (c) typical transverse section, and (d) soil boring data from 7/17/1965.
damage in the slabs and around the foundation. The nonstructural damage was significant. The
recorded peak accelerations in the building were: 0.46g (L), 0.40g (T) and 0.28g (V) at the base,
and 0.59g (L) and 0.58g (T) at the roof, along the longitudinal (L), transverse (T) and vertical
(V) axes of symmetry (there were no sensors installed on the roof to measure vertical motions).
Photographs and detailed description of the damage from the Northridge earthquake can be
found in Trifunac et al. (1999a). Analysis of the relationship between the observed damage and
the change in equivalent vertical shear wave velocity in the building can be found in Ivanović et
al. (1999). A discussion on the extent to which this damage has contributed to the changes in the
apparent period of the soil-structure system can be found in Trifunac et al. (2001c,d).
Two ambient vibration surveys of this building were conducted after the 1994 Northridge
earthquake of January 17, 1994 (ML=6.4, R=1.5 km) and its early aftershocks, while the building
was still damaged and was not in use. Table 5.1.1 shows the first few mode-shapes and
corresponding frequencies identified from these surveys. A detailed description of these surveys
can be found in Ivanovic et al. (1999). Detailed surveys of the damage were conducted at the
time of the experiments and are reported in Trifunac et al. (1999a) along with a summary of the
recorded strong motion response.
Table 5.1.1 Frequencies of the longitudinal (EW) and transverse (NS) modes of vibration.
f - Hz f - Hz
Mode shapes Mode shapes
EW Expl. I Expl. II ∆f - % NS Expl. I Expl. II ∆f - %
Feb. 94 Apr. 94 Feb. 94 Apr. 94
1.0 1.1 10 1.4 1.4 0
3.5 3.7 6 1.6 1.6 0
5.7 5.7 0 3.9 4.2 10
8.1 8.5 5 4.9 4.9 0
Finally, the Northridge earthquake and other earthquakes recorded in the building following
the 1971 San Fernando earthquake were recorded on film by a 13-channel CR-1 recording
system and a stand-alone 3-channel SMA-1 accelerograph, synchronized with the CR-1 recorder.
The 1971 San Fernando earthquake was recorded by a system consisting of three stand alone but
synchronized AR-240 accelerographs. Figure 5.1.3 shows a sketch of the building with the
88
2 N 3
9
4
13.2 m 5.3
10
1.8
6 8.0 m
5
2.7 11
7 8
4.1 2.7
1 12
15 4.1
16 14
13 8.1
40.0 m
45.7 m
ground=1st
Fig. 5.1.3 A sketch of the building with the location of the sensors and their orientation indicated by
arrows. Top: Sensor location of the CR-1 system and SMA-1 accelerograph (channels 14, 15 and 16)
installed after the 1971 San Fernando earthquake. Bottom: Sensor location for the AR-240
accelerographs operating during the 1971San Fernando earthquake.
89
location of the sensors and their orientation indicated by arrows for the CR-1 system and SMA-1
accelerograph (top), and for the earlier AR-240 system (bottom).
In this report, only the data of longitudinal (EW) and transverse (NS) motions recorded at
the ground floor and at the roof are used for 6 earthquakes: 1971 San Fernando, 1989 Malibu,
1989 Montebello, 1992 Landers, 1994 Northridge, and the 1994 Northridge aftershock of March
20, 1994. The records of the 1992 Landers, 1994 Northridge, and 1994 Northridge aftershock
were digitized and processed by the Strong Motion Instrumentation Program of the California
Division of Mines and Geology (CDMG) using an automatic digitization system. The 1971 San
Fernando earthquake records were digitized manually at Caltech following the earthquake, at
sampling rate grater than 50 per second (Trifunac et al., 1973), and the other records were
digitized and processed at USC from photo copies provided by CDMG, using an automatic
digitization system. In this report, acceleration data is used corrected for the instrument response
and band-pass filtered by Ormsby filters (0.1−0.2 and 23−25 Hz) to remove the high frequency
digitization and low frequency baseline errors, using the LaBatch software system (Lee and
Trifunac, 1990). The processed data was made available subsampled at 50 samples per second
(sampling period ∆t=0.02 s), which implies Nyquist frequency of 25 Hz. These and other
earthquake records in the building, not used in this report are described in Trifunac et al. (1999a).
Fig. 5.1.4 shows a simple two-dimensional soil-structure interaction model for in-plane
response of the building, appropriate to model its EW motion, as well as its average NS
translational motion, i.e. the average of the motions at the east and west ends, which differ
because of the torsional response. In this figure, ugnd fl and uroof are the absolute (horizontal)
displacements at the ground floor and at the roof, which have been recorded, urel is the relative
displacement of the roof with respect to the ground floor, φ is the rocking angle of the
foundation (assumed to be rigid), urock is the horizontal displacement of the roof due to rocking
of the foundation, udef is the horizontal displacement of the roof due to deformation of the
building, and H is the building height. As it can be seen from the sketch in Fig. 5.1.4, the
displacement on the roof due to rocking of the foundation is
urock = φ H (5.1.1)
The rocking of the foundation, and this contribution to the absolute roof motion, cannot be
determined from the strong motion instrumentation shown in Fig. 5.1.3, which has only one
sensor on the ground floor that records vertical motions. What can be determined from the
recorded motions is the relative displacement of the roof with respect to the ground floor
90
u roof
u gnd fl u rel
u rock u def
roof
4 H
1
φ
Fig. 5.1.4 A simple model for in-plane soil-structure interaction of the building.
urel = uroof − ugnd fl (5.1.2)
which is the sum of the motion due to rigid body rocking and due to deformation of the building
urel = φ H + udef (5.1.3)
which one has to bare in mind when interpreting the recorded motions. The instrumentation is
also inadequate to separate the free-field motion from the foundation motion. The free-field
91
motion is the motion on the ground surface unaffected by the building and other structures, and
represents the exciting motion. What has been recorded is the motion at the ground floor, which
is affected by the presence of the building and of the building foundation (pile foundation in this
case). The effect of the building vibration is most significant near the first frequency of the soil-
structure system, where the coupling of the foundation and building motions is the strongest, and
for higher frequencies shorter wavelengths compared with the dimension of the foundation,
which are filtered out by the foundation, which is usually stiffer than the surrounding soil. The
foundation motion (in this case the first floor motion) can be represented as a sum of the free-
field motion, uff (the true input), and the motion of the foundation relative to the free-field
motion, u0 , i.e.
ugnd fl = uff + u0 (5.1.4)
The rotation of the foundation, φ , is also a sum of the (point) rotation of the free-field motion,
φff , and the rotation of the foundation relative to the free-field motion, φ0 , i.e.
φ = φff + φ0 (5.1.5)
5.2 DWT Signature and Multiresolution Decomposition of Acceleration
This section shows plots of the discrete wavelet transform and of the multiresolution
components of the recorded absolute accelerations, in Section 5.2.1, and of the relative
acceleration and relative torsion of the roof with respect to the ground level, in Section 5.2.2,
obtained in the wavelet domain from the wavelet transforms of the corresponding absolute
accelerations, based on the linearity property of the wavelet transform. The discrete wavelet
transform plots consist of the detail coefficients d j ,k , k = 1,..., N / 2 j , j = 1,..., J and smooth
coefficients sJ ,k , k = 1,..., N / 2 J of the expansion of the signal in an orthonormal wavelet basis
(see eqn (4.2.1)), plotted for each subband as impulses positioned along the time axis at the
central time of the tile in the time-frequency plane within which they represent the signal (see
Fig. 4.2.1). Hence, there are progressively fewer coefficients for increasing level j . The
multiresolution components are the sequences D j [ n ] , j = 1,..., J and S J [ n ] , computed from the
coefficients by eqn (4.2.2), and represent the subband components sampled at same rate as the
original signal. For each subband, the sequence of coefficients of wavelet expansion and the
corresponding subband component are two equivalent representations of the projection of the
signal onto that subband, one in the wavelet transform domain and the other one in the time
domain. All the results in this section are for a level J = 7 expansion.
92
5.2.1 Absolute Acceleration Responses
Figures 5.2.1.1, 5.2.1.2 and 5.2.1.3 show respectively results for the EW accelerations, and
for the NS accelerations recorded at the East and West sides of the building. Parts a and b show
respectively the wavelet coefficients and the subband components. The plots in the left and right
hand sides of each figure correspond to the motions recorded at ground level and at the roof, all
plotted on same scale for easier comparison. In each plot, the top curve is the original signal. For
convenience in the interpretation of these plots, on the right hand side, the central frequencies of
the subbands (in radians per sample and in Hz) are listed. The x-axis shows the time coordinate
in samples and in seconds.
A comparison of the wavelet coefficients in different subbands (parts a of these figures)

shows that, for the ground floor records, the coefficients are the largest for detail subbands 3 and
4, indicating that this subband contained the largest fraction of the energy of the input ground
motion. For the roof records, the wavelet coefficients are the largest in detail subband 6.
A comparison of the wavelet coefficients for the ground floor with those for the roof
response indicates that the higher frequency components of the input ground motion were
attenuated while propagating from the ground towards the roof (the wavelet coefficients for
detail subbands 1, 2 and 3 are smaller for the roof record), while the lower frequency
components contained in detail subbands 5, 6 and 7 were amplified by the building. The former
is to be expected, as the shorter waves are more prone to scattering and attenuation along the
wave path, while the latter indicates that the fundamental frequency of the soil-structure system
during the shaking from this earthquake was contained somewhere within these subbands. A
curious case is detail subband 4, the motion in which is somewhat smaller at the roof for the EW
but is larger at the roof for the NS component of acceleration. This subband (with central
frequency 2.34 Hz) contains the frequency of the air conditioning system which apparently
operated during this earthquake, and which was also detected in the data from the ambient
vibration tests of this building (Ivanovic et al., 1999). Hence, the larger amplitudes in this
subband at the roof could be a result of a 2 Hz vibration source at the roof rather than an intrinsic
characteristic of the building-soil system.
Comparison of the wavelet coefficients in Figs 5.2.1.2 and 5.2.1.3, both corresponding to
NS accelerations but recorded at the opposite ends of the building, shows that the NS responses
at the East and West ends are similar but not identical, even at the ground floor. For the ground
floor responses, the difference is mostly due to the finiteness of the fault (the energy from the
different asperities on the fault arrives at the West and East sides of the building from different
angles, and with some time delay at the East side relative to the West side, which is closer to the
fault surface (see Fig. 5.1.1). These effects are magnified by the softer soil conditions at the
building site. It also can be noticed that the difference is progressively smaller for the lower
93
EW acceleration Ground Roof
s s
ω/π f
[rad/sample] [Hz]
d1 d1
3/4 18.75
d2 d2
3/8 9.37
d3 d3
3/16 4.69
d4 d4
3/32 2.34
1000
1000
d5 d5
94
0
0
3/64 1.17
cm/s 2
cm/s 2
d6 d6
3/128 0.59
d7 d7
3/256 0.29
s7 s7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.1.1a Discrete wavelet transform of the EW acceleration recorded at the ground floor and at the roof.
s s
ω/π f
[rad/sample] [Hz]
D1 D1
3/4 18.75
D2 D2
3/8 9.37
D3 D3
3/16 4.69
D4 D4
3/32 2.34
400
400
D5 D5
95
0
0
3/64 1.17
cm/s 2
cm/s 2
D6 D6
3/128 0.59
D7 D7
3/256 0.29
S7 S7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.1.1b Multiresolution components of the EW acceleration recorded at the ground floor and on the roof.
NS acceleration at East side Ground Roof
s s
ω/π f
[rad/sample] [Hz]
d1 d1
3/4 18.75
d2 d2
3/8 9.37
d3 d3
3/16 4.69
d4 d4
3/32 2.34
1000
1000
d5 d5
96
0
0
3/64 1.17
cm/s 2
cm/s 2
d6 d6
3/128 0.59
d7 d7
3/256 0.29
s7 s7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.1.2a Discrete wavelet transform of the NS acceleration recorded on the ground floor and on the roof at the East side of the
building.
NS acceleration at East side Ground Roof
s s
ω/π f
[rad/sample] [Hz]
D1 D1
3/4 18.75
D2 D2
3/8 9.37
D3 D3
3/16 4.69
D4 D4
3/32 2.34
400
400
D5 D5
97
0
0
3/64 1.17
cm/s 2
cm/s 2
D6 D6
3/128 0.59
D7 D7
3/256 0.29
S7 S7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.1.2b Multiresolution components of the NS acceleration recorded on the ground floor and on the roof at the East side of the
building.
NS acceleration at West side Ground Roof
s s
ω/π f
[rad/sample] [Hz]
d1 d1
3/4 18.75
d2 d2
3/8 9.37
d3 d3
3/16 4.69
d4 d4
3/32 2.34
1000
1000
d5 d5
98
0
0
3/64 1.17
cm/s 2
cm/s 2
d6 d6
3/128 0.59
d7 d7
3/256 0.29
s7 s7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.1.3a Discrete wavelet transform of the NS acceleration recorded on the ground floor and on the roof at the West side of the
building.
NS acceleration at West side Ground Roof
s s
ω/π f
[rad/sample] [Hz]
D1 D1
3/4 18.75
D2 D2
3/8 9.37
D3 D3
3/16 4.69
D4 D4
3/32 2.34
400
400
D5 D5
99
0
0
3/64 1.17
cm/s 2
cm/s 2
D6 D6
3/128 0.59
D7 D7
3/256 0.29
S7 S7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.1.3b Multiresolution components of the NS acceleration recorded on the ground floor and on the roof at the West side of the
building.
frequency subbands, which can be explained by the fact that the longer waves are less sensitive
to the propagation path. For the roof responses, the difference is due mostly to the torsional
response recorded by these sensors, which is anti-symmetric and hence has opposite phase at the
East and West ends of the building. Contrary to the ground floor responses, which are almost
identical in the 6th subband, the roof responses are quite different in this subband, indicating that
the first torsional mode frequency of the soil-structure system is within this subband.
The above conclusions can be reached also from an analysis of the subband components of
these signals, shown in parts b of these figures, which represent the same information but in the
time domain.
5.2.2 Relative Roof Response
Figures 5.2.2.1 and 5.2.2.2 show the wavelet transform (left) and the subband components
(right) of the EW and average NS translational acceleration responses on the roof relative to the
response of the ground floor. The NS response shown is the average of the responses at the East
and West ends of the building, and represents approximately the NS translation at the center of
the building, not affected by the torsional response. On the other hand, Figure 5.2.2.3 shows the
same for the difference between the NS responses at the East and West ends, which is equal to
the relative angle of torsion of the roof with respect to the ground floor times the normal distance
between the sensors (approximately equal to the length of the building), and hence is a measure
of the relative torsional response. It should be noted here that these measures of the NS
translational and torsional responses were computed from the recorded translational responses at
the East and West walls assuming rigid floor slabs and that the center of torsion for all the floors
coincides with the geometric centers of the floors, which is not true strictly speaking, because the
building was severely damaged during this response. Hence, the computed NS translational
response is contaminated with torsional response, and vice versa. It is also noted that the wavelet
transforms of the relative acceleration responses were computed entirely in the wavelet domain,
from those for the absolute responses by appropriate additions and subtractions, due to the
linearity of the wavelet transform, which required a much smaller number of floating point
operations, avoiding performing three subband decompositions.
A comparison of the coefficients in different subbands shows that, for the relative
translations, the wavelet coefficients are the largest in subbands 5 and 6, and, for the relative
torsion, they are the largest in subband 6. Because the subbands have different bandwidth, we
restrain here from making conclusions about the frequencies of the first longitudinal and
transverse translation and the first torsion frequencies of the soil-structure system until Section
3.6 which deals with the estimation of nodal Fourier spectra of acceleration and of its integrals—
velocity and displacement.
100
Relative EW acceleration at roof
s s
ω/π f
[rad/sample] [Hz]
d1 D1
3/4 18.75
d2 D2
3/8 9.37
d3 D3
3/16 4.69
d4 D4
3/32 2.34
400
1000
d5 D5
101
0
0
3/64 1.17
cm/s 2
cm/s 2
d6 D6
3/128 0.59
d7 D7
3/256 0.29
s7 S7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.2.1 Discrete wavelet transform and multiresolution components of the relative EW acceleration at the roof with respect to
the ground floor.
Relative NS acceleration at roof (average of acceleration at East and West sides)
s s
ω/π f
[rad/sample] [Hz]
d1 D1
3/4 18.75
d2 D2
3/8 9.37
d3 D3
3/16 4.69
d4 D4
3/32 2.34
400
1000
d5 D5
0
0
102
3/64 1.17
cm/s 2
cm/s 2
d6 D6
3/128 0.59
d7 D7
3/256 0.29
s7 S7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.2.2 Discrete wavelet transform and multiresolution components of the relative NS acceleration at the roof with respect to
the ground floor.
Relative torsional acceleration at roof (difference of acceleration at East and West sides)
s s
ω/π f
[rad/sample] [Hz]
d1 D1
3/4 18.75
d2 D2
3/8 9.37
d3 D3
3/16 4.69
d4 D4
3/32 2.34
400
1000
d5 D5
103
0
0
3/64 1.17
cm/s 2
cm/s 2
d6 D6
3/128 0.59
d7 D7
3/256 0.29
s7 S7
1/256 0.10
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.2.2.3 Discrete wavelet transform and multiresolution components of the relative NS acceleration due to torsion at the roof with
respect to the ground floor (estimated by the difference between the relative accelerations at the East and West walls. .
5.3 Multiresolution Approximations of Acceleration and Reduction of Dimensionality
Figures 5.3.1 through 5.3.3 show successive lower resolution approximations of the EW, NS
at east wall and NS at west wall recorded absolute accelerations, at ground level (left) and at the
roof (right). The top signals on each plot are the original time series, sampled at ∆t = 0.05 s and
of length N = 2048 , and below are smooth approximations for levels of expansion J = 1,..., 7 ,
also sampled at ∆t = 0.05 s . The successive approximations have been obtained by successive
removal of detailed information (see Figs. 3.4.1 and 3.7.1), and are in fact a synthesis of the
corresponding level smooth coefficients (see Section 4.3.1), the total number of which is N / 2 J .
Hence, the smooth approximations have smaller dimension than the original signal, and represent
a compressed version of the signal, with some loss, which is the removed detail information.
The numbers to the right in each of the figures represent the corresponding number of smooth
coefficients (i.e. the true dimension of the approximations) and the maximum frequency (in Hz)
of the approximation for ideal filters. The progressive loss of high frequencies is apparent from
these figures. However, to a necked eye, the difference between the original signal and the
approximations up to level J = 3 is barely visible while the reduction of dimensionality is very
significant (up to 1/8 of the original size, i.e. 256 compared to 2048). For browsing through a
database, even J = 4 for the ground response and J = 5 for the roof response records appear to
be satisfactory, which translates respectively to dimension of 128 and 64. Such a reduction of
dimensionality would significantly reduce the time to view a record, especially if the data is on a
remote server.
5.4 Reduction of Dimensionality by Shrinkage of the Smallest Coefficients
The following illustrates the efficiency of the discrete wavelet transform to represent strong
motion records, which are nonstationary signals, with efficiency referring to the ability to
represent a large fraction of the total energy of the signal only by a small number of the largest
wavelet coefficients.
We first recall eqn (4.4.1.1) by which the total energy is the sum of the squares of the
wavelet coefficients, and order the coefficients by decreasing magnitude. Then we compute the
energy that is contained in the first few terms of the wavelet series expansion that have the
largest magnitude coefficients by summing the squares of the corresponding top coefficients, and
plot in Fig. 5.4.1 the fraction of the total energy represented by the top coefficients versus their
number for the roof and for the ground floor EW acceleration records. The analyzing wavelet is
the s8 wavelet and the number of levels is J = 7. It can be seen that, for the roof record, only
3.3% of the largest terms of the expansion (67 of 2048) contain 90% of the total energy. For the
ground floor record, this number is slightly larger—5% of the largest terms (114 of 2048)
104
Dimension fmax
[fraction of N] [Hz]
s s
1 25.00
S1 S1
1/2 12.50
S2 S2
1/4 6.25
S3 S3
1/8 3.12
400
400
S4 S4
0
0
1/16 1.56
105
cm/s 2
cm/s 2
S5 S5
1/32 0.78
S6 S6
1/64 0.39
S7 S7
1/128 0.20
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.3.2 Successive smooth approximations of the NS acceleration recorded on the ground floor and on the roof at the East side of
the building.
NS acceleration at East side Ground Roof Dimension fmax
s s
1 25.00
S1 S1
1/2 12.50
S2 S2
1/4 6.25
S3 S3
1/8 3.12
400
400
S4 S4
0
0
1/16 1.56
106
cm/s 2
cm/s 2
S5 S5
1/32 0.78
S6 S6
1/64 0.39
S7 S7
1/128 0.20
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.3.1 Successive smooth approximations of the NS acceleration recorded on the ground floor and on the roof at the East side of
the building.
NS acceleration at West side Ground Roof Dimension fmax
s s
1 25.00
S1 S1
1/2 12.50
S2 S2
1/4 6.25
S3 S3
1/8 3.12
400
400
S4 S4
0
0
1/16 1.56
107
cm/s 2
cm/s 2
S5 S5
1/32 0.78
S6 S6
1/64 0.39
S7 S7
1/128 0.20
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.3.1 Successive smooth approximations of the NS acceleration recorded on the ground floor and on the roof at the West side of
the building.
representing 90% of the total energy of the signal. It can also be seen that as much as 99% of the
total energy is represented by the top 11.5% coefficients for the roof record and in the top 20%
of the coefficients for the ground floor record. The associated errors of the approximation are
very small. For example, for the ground floor record, the L2 norm of the error, normalized by the
L2 norm of the corresponding signal, is 0.7% if only the top 25% coefficients are used to
reconstruct the signal (75% compression), and it is 1.6% if the top 12.5% coefficients are used to
reconstruct the signal (256 of 2048 coefficients). These results indicate that the discrete wavelet
transform is extremely efficient in representing strong earthquake ground motion records.
EW accel.
roof
1.0
ground floor
0.8
represented by the top coefficients
Fraction of total energy
0.6
0.4
0.2
0.0
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Number of top coefficients
Fig. 5.4.1 Fraction of the total energy represented by the top coefficients versus their number for the EW
acceleration records at the ground floor and at the roof.
It is important to note that the reduction of size by shrinking the less significant wavelet
coefficients is not associated with loss of bandwidth as in the case of subsampling (to avoid
aliasing, subsampling needs to be preceded by low-pass filtering). This is the case because the
wavelet shrinkage preserves the high frequency features in the segments where they are
significant. To illustrate this, in Fig. 5.4.2 we plot reconstructions of these two signals from the
top coefficients that represent 90% of the total energy, and contrast these with uniformly
subsampled versions of the signals that have approximately the same length as the dimension of
108
EW acceleration
original signal
Ground floor: top 114 coef. used (5.5% of all coef.,90% total energy)
compressed signal
downsampled signal Roof: top 67 coef. used (3.27% of all coef., 90% total energy)
roof
0
cm/s 2
ground floor
109
1,000
0 200 400 600 800 1000 1200 1400 1600 1800 2000
0 20 40 s
Fig. 5.4.2 Comparison of results of reduction of dimensionality of a signal in the wavelet domain—by shrinkage of the less
significant wavelet coefficients—and in the time domain by sub-sampling, for the EW accelerations recorded at the ground floor
(bottom) and at the roof (top). The shrinkage is such that the remaining top coefficients represent 90% of the total energy, and the
sub-sampling is at equal time intervals such that the new length is approximately equal to the number of top coefficients used in the
reconstruction of the other compressed version.
the compressed signals. The dimension of the approximations is 114 of 2048 for the ground
floor record and 67 of 2048 for the roof record, i.e. 5.5% and 3.3% of the original dimensions.
Clearly, even for such high compression rates, the wavelet shrinkage preserves the significant
high frequency features of the signals.
The efficiency of representation depends on the nature of the analyzing wavelet and on the
number of levels in the expansion. Figure 5.4.3 illustrates these differences for the EW
EW acceleration
1.0 J=7 1.0 wavelet series
s8 wavelet
J=2 J=7
Normalized Cummulative Energy
J=2
0.8 0.8
0.6 0.6
time series
0.4 0.4 time series
0.2 0.2
Ground Roof
0.0 0.0
1 10 100 n 1000 1 10 100 n 1000
EW acceleration
1.0 J=7 1.0
Normalized Cummulative Energy
s8 s8
0.8 Haar 0.8 Haar
0.6 0.6
time series time series

0.4 0.4
0.2 0.2
Ground Roof
0.0 0.0
1 10 100 n 1000 1 10 100 n 1000
Fig. 5.4.3 Top: Illustration of how the total number of levels of the expansion and the type of wavelet
used affect the efficiency in representation by discrete wavelet transform, for the EW acceleration records
at the ground floor (left) and at the roof (right). Top: normalized energy versus number of top
coefficients for level 7 (solid line) and for level 2 (dashed and dotted line) wavelet expansions. Bottom:
normalized energy for level 7 expansions in s8 (solid line) and in Haar (dashed and dotted line) bases.
The dotted lines in all the plots show the energy contained in the largest amplitude samples of the time
series representations of the same signals.
110
acceleration records at the ground floor (left) and at the roof (right). The top plots show the
normalized energy versus number of top coefficients for level 7 (solid line) and for level 2
(dashed and dotted line) wavelet expansions. The number of coefficients is plotted on a log scale
to emphasize the differences. Similarly, the bottom plots show the normalized energy for level 7
expansions in s8 (solid line) and in Haar (dashed and dotted line) bases. The dotted lines in all
the plots show the energy contained in the largest amplitude samples of the time series
representations of the same signals. It can be seen that the wavelet representation is much more
efficient than the time series representation, more so for the roof record which is more narrow
band and which contains more energy in the lower frequency bands, which have less wavelet
coefficients. The top graphs show that level 7 expansions in s8 basis are more efficient than
level 2 expansions, but the difference is mall for the ground floor record. The bottom graphs
show that, for these signals, level 7 expansion in s8 wavelet basis is more efficient than level 7
Haar basis expansion.
Finally, the error associated with the compression is illustrated for the EW component of
acceleration recorded at the ground floor and at the roof, for the s8 wavelet, and in terms of the
measures outlined in Section 4.3. Figure 5.4.4 shows normalized errors
Error max si − sî

∞ i
= (5.4.1)
σs σs

∞ i
= (5.4.2)
s ∞
max si
i
∑ ( si − sî )
2
/N
( Mean Square Error ) 1/ 2
= i
(5.4.3)
σs σs
and
∑ ( si − sî )
2
Error 2 i
= (5.4.4)
s 2 ∑ si 2
i
versus compression rate
111
n top
Compression rate = 1 − (5.4.5)
N
The latter two measures of normalized error, defined by (eqns) (5.4.3), and (5.4.3) show as
identical curves because the signal is zero mean. It can be seen that the maximum error
normalized by sigma of the signal, Error ∞ σ s , is the largest and can be used as a conservative
estimate of the error, while the maximum error normalized by the maximum of the signal,
Error ∞ s ∞ , and the L2 norm of the error normalized by the L2 norm of the signal,
Error 2
s 2
, are similar, with the latter being somewhat larger for compression ratios grater
than about 10%. As it can be expected, the error is larger for the ground floor record than for
the roof one, and is larger for larger compression rates. For example, for 10% compression,
≅ ( Mean Square Error )
1/ 2
Error 2
s 2
σ s is 0.07% for the ground record and 0.02% for the roof
record, for 50% compression—it is less than 2% for the ground record and 0.50.6% for the roof
record, and for 80% compression—it is 10% for the ground record and 4% for the roof record.
These figures indicate that the mean square error is small even for compression ratio as large as
80%. The maximum error normalized by the standard deviation of the signal is larger and is as
follows: for 10% compression, Error ∞ σ s is 0.4% for the ground record and 0.1−0.2% for the
roof record, for 50% compression—it is 6% for the ground record and 2−3% for the roof record,
and for 80% compression—it is 40% for the ground record and 20% for the roof record. Yet,
normalized by the peak value of the signal, this error is small even for 80% compression, for
which it is 5−6% for the ground floor record and 2−3% for the roof record.
It is interesting to compare the maximum error to the precision of the digitization of the film
records. For 600 dots per inch (236 dots per cm) resolution of scanning, and for transducer
sensitivity of 1.7 cm/g, where g = 981 cm/s2 is the acceleration due to gravity, 1 dot = 0.42
cm/s2. This value and σ s = 69.1 cm/s2 for the ground floor record and σ s = 93.4 cm/s2 for the
roof record imply that the peak error for 50% compression is les than 9.5 pixels for the ground
floor record and 4.4−6.7 pixels for the roof record. The digitization error is larger than the
precision of the scanner and also depends on the width of the trace. The width of the trace
depends on the adopted threshold level, and for chosen threshold level changes depending on the
trace amplitude in the sense that the trace is thinner where the amplitudes are larger (Trifunac
and Lee, 1979). At the peaks, however, the width also depends on the frequency of the signal
and appears thicker if the frequency is larger (Trifunac et al., 1999b). To estimate the order of
magnitude of the error compared to the width of the trace, let us assume that the thickness of the
trace at the largest peak is 3 pixels. This implies that for 50% compression, the maximum error
is of the same order of magnitude as the thickness of the trace. Whether such a small error
would be acceptable would depend on the application.
112
3
10
Ground h
Error
2 100
10 σ
s
h
10
1 Error Error
2
100 100
Normalized error - %
s s
0 2
10
Error
2
100
10
-1 σ
s
-2
10
-3
10
-4
10
1 10 100
3
10
Roof h
Error
2 100
10 σ
s
h
10
1 Error Error
2
100 100
s s
0 2
10
Error
2
100
10
-1 σ
s
-2
10
-3
10
-4
10
1 10 100
Compression rate - %
Fig. 5.4.4 Several measures of normalized error versus compression rate for the EW components of
acceleration recorded on the ground floor and on the roof during the 1994 Northridge earthquake.
113
5.5 Energy, Power, Nodal Energy Spectra and Average Fourier Spectra of Signals
This section illustrates the spectral characteristics of a signal for the recorded accelerations
at the ground floor and on the roof of the building during the Northridge earthquake. These
were computed from a level 8 wavelet series expansion. Because all the signals are high-pass
filtered (by Ormsby filters with ramp between 0.1 and 0.2 Hz) to eliminate low frequency noise
due to baseline distortions, the nodal values for the smooth subband ( S8 ) are not shown. The
last detailed subband, D8 , is essentially the transition band (i.e. the ramp) of the high-pass filter,
which requires a caution in the interpretation of the results, in this and in the subsequent sections
of this chapter. All spectra are plotted on a logarithmic frequency axis in which case the
frequencies of the nodes (i.e. the central frequencies of the subbands) are equally spaced. All
spectra were computed from the first 2048 points of the recorded accelerations, which
corresponds to duration of about 41 s.
Figure 5.5.1 shows the nodal energy spectra, and Fig. 5.5.2—the average Fourier transform
amplitude spectra (called Fourier spectra hereafter), for the EW (top), NS at the East side
(middle) and NS at the West side (bottom) accelerations recorded at the ground floor (left) and at
the roof (right). Each nodal value of the energy spectra, being an aggregate quantity, is plotted
as a vertical bar at the corresponding subband central frequency. The values for the average
Fourier spectra in the subbands are plotted as dots at the nodes. The Fourier spectrum is a
quantity distributed on the frequency axis. To emphasize this, between the nodes, the Fourier
spectra are interpolated linearly (the broken solid line), which is physically meaningful when the
average is computed by placing a much larger weight on the frequencies at and near the nodes
and when there is leakage of energy in the neighboring subbands (true for very short wavelet
filters, which are well localized in time but poorly in frequency). The other extreme is to
interpolate so that the Fourier spectrum is constant throughout the entire subband and
discontinuous at the subband boundaries, which is illustrated only on the top-left plot in Fig.
5.5.2, and which is physically meaningful for ideal wavelet filters, which are perfectly localized
in frequency but poorly in time. For wavelet filters of some “intermediate” length, the
appropriate interpolation function should be between these two.
In the corner of each nodal energy spectrum in Fig. 5.5.1, the total signal energy, average
power for the duration considered, and the corresponding root mean square (r.m.s.) value are
shown. It can be seen that the roof response records have considerably larger total energy and
power than the ground response ones, and are more narrow-band. Both the ground and roof
response records have small energy in the highest and lowest frequency bands, the latter being
ultimately a consequence of the data processing. For the ground floor accelerations, most of the
energy is contained in subband 3 and 4, while for the roof accelerations—in subband 6.
114
Absolute Acceleration - Nodal Energy Spectra
a) 16
EW acceleration
16
Ground Roof
E = 19.6 m 2/s 3 E = 35.7 m 2 /s 3

12 12
P = 0.48 m 2/s 4 P = 0.87 m 2/s 4
r.m.s. = 0.69 m/ s2 r.m.s. = 0.93 m/ s 2
E - m 2 /s 3
8 8
4 4
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d8
d7
d4
d3
d1
d8
d7
d4
d3
d1
b) 16
NS acceleration at East side
16
Ground Roof
E = 19.8 m 2/s 3 E = 32.9 m 2 /s 3

12 12
P = 0.48 m 2/s 4 P = 0.80 m 2/s 4
r.m.s. = 0.69 m/ s 2 r.m.s. = 0.90 m/ s 2
E - m 2 /s 3
8 8
4 4
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d6
d5
d2
d8
d6
d5
d2
d7
d4
d3
d1
d7
d4
d3
d1
c) NS acceleration at West side
16 16
Ground Roof
E = 20.8 m 2 /s 3 E = 33.8 m 2/s 3
12 P = 0.51 m 2 /s 4 12 P = 0.83 m 2 /s 4
r.m.s. = 0.71 m/ s2 r.m.s. = 0.91 m/ s 2
E - m 2 /s 3
8 8
4 4
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d4
d6
d5
d2
d7
d6
d5
d3
d2
d1
d8
d4
d7
d3
d1
Fig. 5.5.1 Nodal energy spectra of the EW acceleration (top), NS acceleration at the West side (middle)
and NS acceleration at the East side (bottom), recorded at the ground floor (left) and on the roof (right)
during the 1994 Northridge earthquake, plotted versus the central frequency of the subbands.
115
Absolute Acceleration - Average Fourier Spectra
a) 2
EW acceleration
4
Ground Roof
3
|FT| - m/s
1 2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d8
d7
d4
d8
d7
d4
d3
d1
d3
d1
b) 2
NS acceleration at East side
4
Ground Roof
3
|FT| - m/s
1 2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d8
d7
d4
d3
d1
d8
d7
d4
d3
d1
c) NS acceleration at West side
2 4
Ground Roof
3
|FT| - m/s
1 2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d4
d6
d2
d7
d6
d3
d2
d1
d8
d5
d4
d5
d7
d3
d1
Fig. 5.5.2 Average subband Fourier transform amplitude versus the subband central frequency (dots) of
the EW acceleration (top), NS acceleration at the West side (middle) and NS acceleration at the East side
(bottom), recorded at the ground floor (left) and on the roof (right) during the 1994 Northridge
earthquake. The broken lines represent an interpretation of interpolated values between the nodes for
non-ideal filters, and the horizontal segments in the upper left plot—for ideal filters.
116
Relative Acceleration - Nodal Energy and Average Fourier Spectra
a) 24
Relative EW acceleration at the roof
5
20
E = 59.1 m 2/s 3 4
16 P = 1.44 m 2/s 4
r.m.s. = 1.20 m/s 2 3
|FT| - m/s
E - m 2 /s 3
12
2
8
4 1
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d6
d5
d2
d5
d2
d7
d4
d3
d1
d7
d4
d3
d1
d8
d8
b) 24
Relative NS acceleration at the roof *
5
20 E = 45.6 m 2 /s 3 4
P = 1.11 m 2 /s 4
16
r.m.s. = 1.06 m/s 2 3
|FT| - m/s
E - m 2 /s 3
12
2
8
4 1
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d2
d7
d4
d5
d4
d3
d1
d7
d3
d1
d8
d8
c) 24
Relative Torsional acceleration at the roof **
5
20 E = 36.4 m 2 /s 3 4
P = 0.89 m 2 /s 4
16
r.m.s. = 0.94 m/s 2
3
|FT| - m/s
E - m 2/s 3
12
2
8
4 1
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d7
d4
d5
d4
d2
d3
d1
d7
d3
d1
d8
d8
* Average of relative NS accelerations at East and West sides ; **Difference of relative NS accelerations at East and West sides
Fig. 5.5.3 Nodal energy spectra (left) and average Fourier spectra (right) versus subband central
frequency of the relative EW acceleration (top), average NS relative acceleration (middle), and relative
“torsional” acceleration during the 1994 Northridge earthquake.
117
The shape of the Fourier spectra in Fig. 5.5.2 is different from that of the nodal energy
spectra in Fig. 5.5.1 because the subbands have different width. For the ground floor
acceleration records, the Fourier spectra are the largest in the 6th subband, but are also large in
subbands 3, 4 and 5. For the roof acceleration records, the Fourier spectra are by far the largest
in the 6th subband, indicating that the predominant (first) frequency of the soil-structure system is
within subband 6.
Fig. 5.5.3 shows the nodal energy spectra (left) and the corresponding average Fourier
spectra (right) for the relative roof responses: EW translation (top), NS translation (the average
of the relative responses at the East and West walls), and “torsion” (the difference of the relative
responses at the East and West walls). The total energy, average power and root mean square
values for these relative responses are shown in the corner of the plots of the nodal energy
spectra. It can be seen that, for the relative acceleration responses, the total energy is larger than
that for the absolute roof responses. The shapes of the Fourier spectra of these relative
acceleration responses are similar to those for the roof responses, and are the largest for the 6th
subband. This indicates that the predominant system frequencies for longitudinal (EW) and
transverse (NS) translations and torsion are within subband 6. Most accurately, this would be
obtained from the shape of the average Fourier spectra of displacement, which can be estimated
from those of acceleration, and even more accurately from the corresponding transfer-functions
between the spectra of the relative displacements and those of the ground floor accelerations.
However before attempting to do that, we asses how close the shape of the average Fourier
spectra is to the shape of the actual Fourier spectra, and how good are the estimated average
Fourier spectra of displacement from those of acceleration.
Figure 5.5.4 shows a comparison of (weighted) average Fourier spectra of acceleration with
the actual spectra for the EW ground floor (left) and roof (right) records, plotted on a logarithmic
frequency and linear amplitude scale. The actual spectra were computed on a dense grid of
frequencies equally spaced on a log scale, using eqns (2.4.1) and (4.7.2.3). It can be seen that the
actual spectra have large fluctuations (on a linear scale), which are well smoothed by the wavelet
transform, but the number of subbands (nodes) is too small for the roof record to estimate
accurately the spectral shape. Also, the leakage of energy in the neighboring subbands is
obvious, especially for the roof record. Both problems can be solved, the former by using
related transforms, e.g. wavelet packet transform or wavelet transform that is not dyadic and has
more subbands within the frequency interval of interest for strong motion studies, and the latter
by using wavelets that are better localized in frequency, and consequently—more spread in time
and hence described by longer prototype filters (the s8 wavelet is an 8 point filter). Because of
the intrinsically large fluctuations of the amplitudes of the Fourier spectra of earthquake strong
motion records, the average spectra represent well the total energy in the subband, but not the
extreme values.
118
Absolute Acceleration - Average and Actual Fourier Spectra
EW acceleration
4 10
Ground Roof
8
|FT| - m/s
6
2
4
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d6
d8
d5
d2
d8
d5
d2
d7
d4
d3
d1
d7
d4
d3
d1
Fig. 5.5.4 Comparison of the actual Fourier transform amplitudes with the weighted average values for
the subbands plotted at the nodes (dots), for the EW acceleration records of the 1994 Northridge
earthquake at the ground floor and at the roof.
5.6 Estimation of Spectra of Integrals and Derivatives of Signals
In Section 4.7, we showed how the wavelet transform and nodal energy and average Fourier
spectra of derivatives (or integrals) of a signal can be estimated directly from the wavelet
transform of the signal itself (see eqns (4.7.5) and (4.7.6)). Here we examine how good these
estimates are for velocity and displacement obtained from recorded acceleration, band-pass
filtered to eliminate long period errors due to uncertainty in the baseline position. In the
standard data processing of strong motion records, the velocity and displacement integrated from
the acceleration time history are also band-pass filtered after integration (Trifunac and Lee,
1973), which is not accounted for by eqns (4.7.5) and (4.7.6). As we will see, this will influence
considerably the estimated spectra in the subband which contains the high-pass filter transition
band, as well as to some degree also the neighboring subbands, depending on the frequency
localization of the (non-ideal) wavelet filters. Figure 5.6.1 compares nodal energy spectra (top)
and average Fourier spectra bottom) of EW velocity at the ground floor (left) and at the roof
(right), computed from the velocity time histories (open circles) and estimated from the
corresponding spectra of acceleration (triangular symbols). The estimated nodal energy spectra
have been shifted slightly towards higher frequencies to avoid clutter. Similarly, Figure 5.6.2
compares nodal energy spectra (top) and average Fourier spectra (bottom) of EW displacement
at the ground floor (left) and at the roof (right), computed from the displacement time histories
(open circles) and estimated from the corresponding spectra of acceleration (triangular symbols).
It can be seen that the agreement between the actual and estimated spectra is good for the higher
frequency subbands and not so good for the lower frequency subbands, especially for the
displacement, due to the fact that the actual spectra have been obtained from high-pass filtered
119
Comparison of Estimated and Actual VelocitySpectra
a) 0.2
EW motion
2
Ground Roof
E = 0.27 m 2 /s E = 2.14 m 2 /s
P = 0.0065 m 2/s 2 P = 0.052 m 2/s 2
r.m.s. = 0.081 m/ s r.m.s. = 0.23 m/ s
E - m 2 /s
0.1 1
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d8
d7
d4
d3
d1
d8
d7
d4
d3
d1
Estimated from velocity time series Estimated from DWT of acceleration
b) 0.6
EW motion
1.4
Ground Roof
1.2
1.0
0.4
|FT| - m
0.8
0.6
0.2
0.4
0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d6
d2
d8
d5
d4
d8
d5
d4
d7
d3
d1
d7
d3
d1
Fig. 5.6.1 Comparison of nodal energy spectra (top) and average Fourier spectra (bottom) of velocity
computed from velocity time history, obtained by integration of recorded acceleration, and high-pass
filtered afterwards to eliminate long period noise (open circles) with the corresponding spectra estimated
from the discrete wavelet transform of acceleration, for the EW motion at the ground floor (left) and at
the roof (right) during the 1994 Northridge earthquake.
data, and due to the leakage of energy in neighboring subbands. We go even further by
comparing in Fig. 5.6.3 the actual displacement time histories (thin lines) with the estimated time
histories obtained by inverting the estimated wavelet transform of displacement from that of
acceleration (thick lines), for the EW motions at the ground floor (top) and at the roof (bottom).
It can be seen that there is a very good qualitative agreement of the actual and estimated time
histories, which is at first sight counter intuitive considering the large discrepancies in the
corresponding spectra, not only in the lowest frequency subband—due to the poor frequency
localization of the wavelet basis, but which actually can be explained by the poor frequency
120
localization, as poor frequency localization implies very good time localization—apparently the
reason for the very good qualitative agreement. Better agreement of the actual and estimated
spectra of derivatives and integrals of the signal can be achieved by using a wavelet basis that is
well localized in frequency (i.e. with longer wavelet filters) and related transforms that lead to
more subbands within the frequency interval of interest.
Comparison of Estimated and Actual Displacement Spectra

a) 0.02
EW motion
0.2
Ground Roof
E = 0.027 m 2 s E = 0.219 m 2 s
P = 0.00065 m2 P = 0.0053 m 2
r.m.s. = 0.026 m r.m.s. = 0.073 m
E - m2 s
0.01 0.1
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d6
d2
d8
d5
d4
d8
d5
d4
d7
d3
d1
d7
d3
d1
Estimated from velocity time series Estimated from DWT of acceleration
b) 0.3
EW motion
0.8
Ground Roof
0.6
0.2
|FT| - m s
0.4
0.1
0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d6
d2
d8
d5
d4
d8
d5
d4
d7
d3
d1
d7
d3
d1
Fig. 5.6.2 Comparison of nodal energy spectra (top) and average Fourier spectra (bottom) of displacement
computed from displacement time history, obtained by two times integration of recorded acceleration, and
high-pass filtered afterwards to eliminate long period noise (open circles) with the corresponding spectra
estimated from the discrete wavelet transform of acceleration, for the EW motion at the ground floor (left)
and at the roof (right) during the 1994 Northridge earthquake.
121
EW displacement
Estimated from DWT of acceleration
Actual
Ground
0 10
Roof
0
30 cm
0 200 400 600 800 1000 1200 1400 1600 1800 2000
0 20 40 s
Fig. 5.6.3 Comparison of displacement time history, obtained by two times integration of recorded
acceleration, and high-pass filtered to eliminate long period noise (thin line) with the displacement time
history estimated from the discrete wavelet transform of acceleration, for the EW motion at the ground
floor (top) and at the roof (bottom) during the 1994 Northridge earthquake.
5.7 Energy and PSD Time-Frequency Distributions
I this section, we illustrate nodal time-frequency distributions of energy, computed from eqn
(4.4.3.1), and time-frequency distributions of power spectrum density (PSD)—more precisely, of
average PSD in the tiles of the time-frequency plane, computed by eqn (4.5.3.7). These are
illustrated on the EW components of acceleration at the ground floor and at the roof, and on the
EW displacement of the roof relative to that of the ground floor. As it can be seen from eqns
(4.4.3.1) and (4.5.3.7), the numerical values of the nodal energy distribution (aggregate energy in
a tile) and of the average PSD in the tiles have same numerical value (a coincidence of the fact
that all the tiles, regardless of the difference in shape, have area equal to ) but different
physical units.
Figure 5.7.1 shows nodal energy spectra for the tiles of each subband for the acceleration at
the ground floor (left) and that at the roof (right) plotted as vertical bars at the central frequency
of the subband. It can be seen that the nodal energies are very small for all but subbands 2, 4, 5
and 6 for the ground floor acceleration, and in subbands 3, 4, 5, 6 and 7 for the roof acceleration.
It can be seen that the propagation through the building amplified the motions mostly in subband
6 but also in subbands 5 and 7, and attenuated the acceleration in subbands 1 through 4. The
122
Energy Nodal Time-Frequency Distribution
400
400
s s
ω/π
0
0
f
[rad/sample] [Hz]
cm/s 2
cm/s 2
|d 1k |2 |d 1k |2
3/4 18.75
|d 2k |2 |d 2k |2
3/8 9.37
|d 3k |2 |d 3k |2
3/16 4.69
|d 4k |2 |d 4k |2
3/32 2.34
123
|d 5k |2 |d 5k |2
0
3/64 1.17
0
m 2/s 3
m 2/s 3
|d 6k |2 |d 6k |2
3/128 0.59
|d 7k |2 |d 7k |2
3/256 0.29
|d 8k |2 |d 8k |2
3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.7.1 Nodal time-frequency distribution of energy of the EW accelerations recorded at the ground floor and on the roof
during the 1994 Northridge earthquake.
degree of amplification depends on time as well. For example, during the first 10 seconds of
motion, the accelerations in subband 5 are amplified the most, and afterwards, those in subbands
6 and 7 are amplified the most, which indicates a change in the system approximately about that
time. Figure 5.7.2 shows the nodal energy distribution for the ground floor acceleration (the
input to the system) and for the displacement at the roof relative to the ground floor. It can be
seen from this plot that the energy of the relative displacement is contained mostly in subbands 5
and to a smaller degree in subbands 5 and 7, and that even more obviously, the building
amplified the motion in subband 5 within the first 10 s of motion, and in subbands 6 and 7
afterwards, reaffirming the hypothesis about a system change around the 10th second of recorded
response. The energy of the relative displacement in all the other subbands is very small and
cannot be distinguished from the baseline in these two plots.
Figures 5.7.3 through 5.7.5 show the normalized (to unit amplitude in each subband) PSD
distribution in the tiles, respectively for the ground floor and roof (absolute) accelerations, for the
ground floor acceleration and the roof relative displacement, and for the ground floor and roof
absolute displacement. A plot of the predominant system frequency versus time is shown at the
bottom, determined from the ridge of the continuous wavelet transform of the relative roof
displacement, computed using the complex Morlet wavelet Todorovska (2001) under the
assumption that the relative displacement response is essentially a chirp signal. These two plots
make it possible to observe the time variations in PSD for all of the subbands, and possibly relate
some of the variations to the changes in the predominant system frequency, which varied during
the shaking and was significantly smaller than the value determined from small amplitude
(ambient noise) tests (see Table 5.1.1). In the brief discussion of these plots, we focus on the
patterns in the higher frequency subbands. Sharp spikes in the PSD versus time can be are
noticed in the highest frequency subbands, indicating sudden changes (possibly discontinuities)
in the amplitudes of these signals.
Before we attempt to interpret these spikes in the PSD plots in Figs. 5.7.3 through 5.7.5 we
show in Fig. 5.7.6 normalized PSD in the three highest frequency subbands for the relative
displacement signal s [ n ] modified so that an increment ∆s is added to it only in the interval
n ∈ [1000,1500] , hence artificially creating a “jump” ∆s at n = 1000 and at n = 1500 . Figure
5.7.6 shows PSD for ∆s = 1 , 0.5, 0.1 and 0.01 cm, which is small fraction of the amplitude of the
signal within this interval (~15 cm). This figure shows that even such small “jumps” result in
large spikes in the PSD. These spikes are very well localized in time (especially in the higher
frequency subband), and precisely point out to the time of these rapid variations in amplitude.
The amplitude of the spikes becomes smaller for progressively smaller ∆s . Even for the
smallest jump considered, ∆s = 0.05 cm, the spikes in the PSD are at least an order of magnitude
larger than all the values for n > 600 , but are comparable to the spikes in the PSD for n < 500 .
Such rapid changes in the amplitude of response of the building can be caused by a sudden loss
124
Energy Nodal Time-Frequency Distribution
Relative EW displacement
30
400
s s
ω/π
0
0
f
[rad/sample] [Hz]
cm
cm/s 2
|d 1k |2 |d 1k |2
3/4 18.75
|d 2k |2 |d 2k |2
3/8 9.37
|d 3k |2 |d 3k |2
3/16 4.69
|d 4k |2 |d 4k |2
3/32 2.34
125
0.02
|d 5k |2 |d 5k |2
3/64 1.17
0
0
m 2s
m 2/s 3
|d 6k |2 |d 6k |2
3/128 0.59
|d 7k |2 |d 7k |2
3/256 0.29
|d 8k |2 |d 8k |2
3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.7.2 Nodal time-frequency distribution of energy of the EW accelerations recorded at the ground floor and of the relative
displacement of the roof during the 1994 Northridge earthquake.
Normalized PSD Distribution |d jk |2 / max |d jk |2
k
400
400
s s
ω/π
0
0
f
[rad/sample] [Hz]
subband D 1 subband D 1
cm/s 2
cm/s 2
3/4 18.75
D2 D2
3/8 9.37
D3 D3
3/16 4.69
D4 D4
3/32 2.34
126
D5 D5
3/64 1.17
D6 D6
3/128 0.59
D7 D7
3/256 0.29
0.8 Predominant system frequency - Hz 0.8 Predominant system frequency - Hz
0.6 0.6
0.4 0.4
0 500 1000 1500 2000 0 500 1000 1500 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.7.3 Time-frequency distribution of average power spectrum density (PSD), normalized to unit amplitude for each
subband, of the EW accelerations recorded at the ground floor and on the roof during the 1994 Northridge earthquake.
k
30
400
EW acceleration Ground Relative EW displacement Roof
s s
ω/π
0
0
f
[rad/sample] [Hz]
cm
cm/s 2
3/4 18.75
D2 D2
3/8 9.37
D3 D3
3/16 4.69
D4 D4
3/32 2.34
127
D5 D5
3/64 1.17
D6 D6
3/128 0.59
D7 D7
3/256 0.29
0.6 0.6
0.4 0.4
0 500 1000 1500 2000 0 500 1000 1500 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
subband, of the EW accelerations recorded at the ground floor and on the roof during the 1994 Northridge earthquake.
k
30
30
EW displacement Ground Roof
s s
ω/π
0
0
f
[rad/sample] [Hz]
cm
cm
3/4 18.75
D2 D2
3/8 9.37
D3 D3
3/16 4.69
D4 D4
3/32 2.34
128
D5 D5
3/64 1.17
D6 D6
3/128 0.59
D7 D7
3/256 0.29
0.6 0.6
0.4 0.4
0 500 1000 1500 2000 0 500 1000 1500 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
subband, of the EW absolute displacement at the ground floor and on the roof during the 1994 Northridge earthquake.
k
s Relative EW displacement
0
∆ s = 1 cm ∆ s = 0.5 cm ω/π
subband f
[rad/sample] [Hz]
D1 D1
3/4 18.75
D2 D2
3/8 9.37
D3 D3
3/16 4.69
129
0 500 1000 1500 2000 0 500 1000 1500 2000
∆ s = 0.05 cm
∆ s = 0.1 cm
D1
D1 3/4 18.75
D2 D2 3/8 9.37
D3 D3 3/16 4.69
0 500 1000 1500 2000 0 500 1000 1500 2000
Fig. 5.7.6 Influence on the PSD time distribution of discontinuities in the signal, artificially created at n=1000 and 1500 by
adding increment ∆s to the signal.
of stiffness due to cracking or a brittle failure of some of its structural elements. Hence, such
patterns in the PSD could be used as a tool for damage detection. The PSD time frequency
distribution can be used to localize the onset of damage in time, and such distributions for many
points in the structure could be used to localize the damage spatially. In Fig. 5.7.5, spikes in the
PSD of the absolute roof response that could be indicative of sudden loss of stiffness are
identified by arrows (at ~4.7 s, 8.28 s and 8.9 s). The use of the discrete wavelet transform
along these lines to identify damage in a structure has been pointed out earlier by Rezai et al.
(1994). One difficulty with this approach is to distinguish spikes due to damage from spikes due
to high frequency motions that have been propagated through the structure from the ground, and
other high frequency “noise” in the response. Hence, the presence of such spikes in the PSD can
be considered sufficient but not necessary condition to identify damage.
5.8 Estimation of Cross-Correlation of Motions across the Building Height - Global and
Subband Aggregates and Averages
This section illustrates the global and spectral aggregates and averages of cross-correlation
across the building height, computed for the EW components of motion, in particular, between
the ground floor and roof motions.
Figures 5.8.1 through 5.8.3 show: a) time series of the two signals, b) their nodal cross-
energy spectrum, c) their nodal cross-power spectrum density, d) their maximum cross-energy
spectrum, and e) the cosine of the angle between them, respectively for the EW accelerations at
the roof and at ground level, for the EW displacements at the roof and at ground level, and for
the EW relative roof displacement and acceleration at ground level. The nodal cross-energy
{ }
spectrum, ED( , j ) , j = 1,..., J ; ES( , J ) , computed by eqn (4.6.2.2), represents the distribution across
1,2 1,2
the different subbands of the total cross-energy, and represents a measure of the similarity of the
corresponding subband components of the two signals. The nodal cross-power spectrum density
(PSD), computed by eqn (4.6.4.2) also represents a measure of the similarity of the
corresponding subband components of the two signals, but its nodal values, representing subband
averages, do not depend on the width of the subbands, while the nodal values of the cross-energy
spectrum, representing subband aggregates, do depend on the width of the subbands. The value
of the total cross-energy, ED( , j ) , computed by eqn (4.6.1.1), is written in the upper right corner of
1,2
the plot, along with the average cross-power. The maximum nodal cross-energy spectrum, equal
to the product of the norms of the corresponding subband components of the two signals, is equal
to the nodal cross-energy spectrum if the subband components of the two signals were perfectly
aligned, i.e. if cos α computed by eqns (4.6.2.4) and (4.6.2.5) is equal to zero. The numerical
values in all the figures represent estimates for the continuous time signals.
130
The plots of cos α in parts e) of Figs 5.8.1 and 5.8.2 show that both the absolute
accelerations and the absolute displacements at the ground floor and at the roof are positively
correlated (have same sign) far from the (first) system frequency, f1 , while in the neighborhood
of the system, where the absolute motion on the roof is dominated by the contribution from the
relative roof response with respect to that of the ground floor, the correlation changes sign. Near
the subband that contains the system frequency, the correlation is relatively large in magnitude,
and positive for f < f1 and negative for f > f1 . Within the subband that contains f1 , however,
cancellation occurs from the (large) positive and negative contributions from the frequencies
very close but less than f1 , and from those very close but grater than f1 , so that the resultant
nodal correlation is small in magnitude, and positive or negative depending on the contribution
that prevails. For high frequencies, the degree of correlation, as measured by cos α , is smaller
for the accelerations than for the displacements, especially for the two highest frequency
subbands. In absolute value, as it can be seen form the plots of the cross-energy (part a), the
correlation is small for the highest frequency subbands, especially for the displacement, as the
displacements of the signals themselves have little energy in these subbands, as it can be seen
from the plots of the maximum cross-energy (part d). The same is true for the lowest frequency
subband, D8 , which overlaps with the transition band for the high pass filter applied to the
preprocessed data to remove the long period noise.
Figure 5.8.3 shows that the relative roof displacement is negatively correlated with the
ground floor acceleration for f < f1 , and positively correlated for f > f1 . The degree of
correlation is small near f ≈ f1 , where the correlation changes sign, for the same reasons
explained in the previous paragraph. The absolute value of the correlation is small for the
highest frequency subbands.
The nature of the correlation between the motion at the ground floor and at the roof (both
relative and absolute) near the (first) system frequency, f1 , can be explained approximately by
the differential equation of a single degree-of-freedom oscillator which can be used to
approximate the deflection of multi-story buildings. Then, the displacement of the roof due to
deflection of the columns, udef (see the model in Fig. 5.1.4), satisfies
udef + 2ω1ζ udef + ω12udef = −ugnd fl (5.8.1)
where ugnd fl is the absolute displacement of the oscillator support (the ground floor in this case),
ω = 2π f is the circular frequency, and ζ is the damping ratio. For lightly damped systems
131
(a)
EW acceleration Ground
400
cm/s 2
0
Roof
400
cm/s 2
0
0 500 1000 1500 2000
0 10 20 30 t 40 s
(b) (c)
2
2
/s 3
/s 2
2
1
2
Cross - Energy - m
0
Cross - PSD - m
0
-2
-1
E (1,2) = − 1.9 m 2 /s 3
-2
-4
-3
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
max E (1,2) = 22.1 m 2/s 3 (d) (e)
6
1.0
/s 3
5
2
0.5
Max. Cross - Energy - m
4
cos α
3
0.0
2
-0.5
cos α = −0.09
1
α = 0.53 π
-1.0
0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
Fig. 5.8.1 Estimation of correlation between the EW acceleration at the ground floor and at the roof
recorded during the 1994 Northridge earthquake. (a) Time series of the two signals; (b) nodal cross-
energy spectrum, (c) nodal cross-power spectrum density; (d) maximum nodal cross-energy; and (e) the
angle between the corresponding subband components of the two signals.
132
(a)
EW displacement Ground
20
cm
0
Roof
20
cm
0
0 500 1000 1500 2000
0 10 20 30 t 40 s
(b) (c)
0.03
0.06
2
s
E (1,2) = 0.03 m 2 s
2
s
0.02
Cross - Energy - m
0.04
Cross - PSD - m
0.01
0.02
0.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
(d) (e)
0.04
1.0
s
max E (1,2) = 0.075 m 2s

2
0.03
0.5
cos α
0.02
0.0
0.01
-0.5
cos α = 0.39
α = 0.37 π
-1.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
Fig. 5.8.2 Estimation of correlation between the EW displacement at the ground floor and at the roof
caused by the 1994 Northridge earthquake. (a) Time series of the two signals; (b) nodal cross-energy
spectrum, (c) nodal cross-power spectrum density; (d) maximum nodal cross-energy; and (e) the angle
between the corresponding subband components of the two signals.
133
(a)
EW ground acceleration
400
cm/s 2
0
Relative roof displacement
20
cm
0
0 500 1000 1500 2000
0 10 20 30 t 40 s
(b) (c)
E (1,2) = 0.27 m 2/s
0.20
0.20
2
/s
2
Cross - Energy - m
Cross - PSD - m
0.10
0.10
0.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
(d) (e)
0.5
1.0
/s
max E (1,2) = 0.95 m 2/s

2
0.4
0.5
0.3
cos α
0.0
0.2
-0.5
cos α =0.28
0.1
α = 0.41 π
-1.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
Fig. 5.8.3 Estimation of correlation between the EW acceleration at the ground floor and the relative roof
displacement caused by the 1994 Northridge earthquake. (a) Time series of the two signals; (b) nodal
cross-energy spectrum, (c) nodal cross-power spectrum density; (d) maximum nodal cross-energy; and (e)
the angle between the corresponding subband components of the two signals.
134
(ζ 1 ), the second term on the left hand side of eqn (5.8.1) is much smaller than the other terms
and can be neglected. Further on, by substituting for udefl = −ω 2udefl in eqn (5.8.1) implies
(ω 2
− ω12 ) udefl = ugnd fl (5.8.2)
from where it follows that
udefl ∼ −ugnd fl , ω 2 < ω12

(5.8.3)
udefl ∼ ugnd fl , ω 2 > ω12
According to the simple soil-structure interaction model in Fig. 5.1.4, the relative displacement
of the roof with respect to the ground floor is
urel = φ H + udef (5.8.4)
where φ is the foundation rocking angle. According to linear analytical models of building-soil
interaction (Todorovska and Trifunac, 1990), near the first system frequency, where the
foundation rocking is the largest, the building deflection and displacement die to rigid body
rocking are approximately in phase, which along with eqn (5.8.3) explains the results in Fig.
5.8.3. This also explains the results in Fig. 5.8.1 and 5.8.2 near the firsts system frequency,
where the relative response is large, and hence is the dominant contributor to the absolute roof
displacement
uroof = ugnd fl + urel (5.8.5)
The explanation requires expressing urel in eqn (5.8.1) via eqn (5.8.5), and appropriate
cancellations, which implies
− (ω 2 − ω12 ) udefl = ugnd fl (5.8.6)
and
− (ω 2 − ω12 ) udefl = ugnd fl (5.8.7)
135
5.9 Estimation of Cross-Correlation of Motions across the Building Length - Global and
Subband Aggregates and Averages
This section illustrates the global and spectral aggregates and averages of cross-correlation
across the building length, computed from the NS motions recorded at the east and west ends of
the building, at the ground floor and at the roof.
Figures 5.9.1 through 5.9.4 show: a) time series of the two signals, b) their nodal cross-
energy spectrum, c) their nodal cross-power spectrum density, d) their maximum cross-energy
spectrum, and e) the cosine of the angle between them, respectively for the NS accelerations
recorded on the ground floor near the East and West walls, for the NS displacements recorded on
the ground floor near the East and West walls, for the NS accelerations recorded on the roof near
the East and West walls, for the NS displacements recorded on the roof near the East and West
walls.
Figures 5.9.1 and 5.9.2 shows that, the accelerations on the ground floor are positively
correlated to a very high degree (as measured by cos α 1 ) up to f = 1−2 Hz, when the
correlation starts to decrease, and changes sign in the highest frequency subband, while the
displacements are positively correlated to a much higher degree even in the high frequency
subbands. The change of sign for the correlation of acceleration in the highest frequency
subband may be related to the phase difference between the short wavelength incident waves
reaching the East and West ends of the foundation, which are at a different distance from the
source. Narrowing down the physical explanations for this effect is out of the scope of this
report. The cross-energy is the largest in subbands d3 and d 4 where both signals have most of
their energy, while the cross-PSD is the largest in subband d 6 . The cross-energy, as well as the
maximum cross-energy in the highest frequency subband, d1 , is very small.
Figures 5.9.3 and 5.9.4 shows that, at the roof, the accelerations are positively correlated for
all subbands except for the two highest frequency ones, when the cross-energy is very small and
were the correlation changes sign. The cross-energy, and cross-PSD, and maximum cross-
energy, all are maximum in subband d 6 , in which the Fourier spectrum of both the (average)
relative roof translation and the relative roof torsion is the maximum, indicating that both the
first system frequency of NS translation and the first system torsional frequency are in or near
this subband. The degree of correlation for the accelerations in subband d 6 , however, as
measured by cos α , has a local minimum but is still positive and large ( cos α > 0.5 ), which
indicates that the contribution to the roof response from the rigid body rocking plus deflection of
the first NS translational mode is much larger than the contribution from the first torsional mode.
Form Fig. 5.9.4, it can be seen that, for the roof displacements, the degree of correlation starts to
136
East side
(a)
400
NS ground acceleration
cm/s 2
0
400
West side
cm/s 2
0
0 500 1000 1500 2000
0 10 20 30 t 40 s
(b) (c)
3.0
E (1,2) = 12.8 m 2/s 3
/s 3
0.20
/s 2
2
2
Cross - Energy - m
2.0
Cross - PSD - m
0.10
1.0
0.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
(d) (e)
1.0
8
max E (1,2) = 20.2 m 2/s 3

/s 3
2
0.5
6
cos α
0.0
4
cos α = 0.63
-0.5
2
α = 0.28 π
-1.0
0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
Fig. 5.9.1 Estimation of correlation between the NS acceleration recorded on the ground floor near the
East and West walls during the 1994 Northridge earthquake. (a) Time series of the two signals; (b) nodal
cross-energy spectrum, (c) nodal cross-power spectrum density; (d) maximum nodal cross-energy; and (e)
the angle between the corresponding subband components of the two signals.
137
NS ground displacement East side
(a)
20
cm
0
20
West side
cm
0
0 500 1000 1500 2000
0 10 20 30 t 40 s
(b) (c)
0.020
s
0.04
2
2
s
Cross - Energy - m
Cross - PSD - m
0.010
0.02
E (1,2) = 0.04 m 2 s
0.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
(d) (e)
0.02
1.0
2
s
max E (1,2) = 0.041 m 2s

0.5
cos α
0.01
0.0
-0.5
cos α =0.985
α = 0.055 π
-1.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
Fig. 5.9.2 Estimation of correlation between the NS displacements of the ground floor near the East and
West walls during the 1994 Northridge earthquake. (a) Time series of the two signals; (b) nodal cross-
138
NS roof acceleration East side
(a)
400
cm/s 2
0
400
West side
cm/s 2
0
0 500 1000 1500 2000
0 10 20 30 t 40 s
(b) (c)
/s 3
E (1,2) = 20.98 m 2 /s 3
12
12
/s 2
2
2
Cross - Energy - m
Cross - PSD - m
8
8
4
4
0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
(d) (e)
1.0
/s 3
12
2
0.5
max E (1,2) = 32.6 m 2/s 3

8
cos α
0.0
4
-0.5
cos α = 0.64
α = 0.28 π
-1.0
0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
Fig. 5.9.3 Estimation of correlation between the NS acceleration recorded on the roof near the East and
West walls during the 1994 Northridge earthquake. (a) Time series of the two signals; (b) nodal cross-
139
NS roof displacement East side
(a)
20
cm
0
20
West side
cm
0
0 500 1000 1500 2000
0 10 20 30 t 40 s
(b) (c)
0.2
0.2
E (1,2) = 0.18 m 2 s
2
s
2
s
Cross - Energy - m
Cross - PSD - m
0.1
0.1
0.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
(e)
(d)
0.2
1.0
max E (1,2) = 0.21 m 2s

2
s
0.5
cos α
0.1
0.0
-0.5
cos α = 0.85
α = 0.18 π
-1.0
0.0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d8
d7
d6
d5
d4
d3
d2
d1
d8
d7
d6
d5
d4
d3
d2
d1
Fig. 5.9.4 Estimation of correlation between the NS displacements of the roof near the East and West
walls during the 1994 Northridge earthquake. (a) Time series of the two signals; (b) nodal cross-energy
spectrum, (c) nodal cross-power spectrum density; (d) maximum nodal cross-energy; and (e) the angle
between the corresponding subband components of the two signals.
140
decrease with frequency in subband d 6 , but has a local minimum in subband d5 and remains
small also in subbands d 4 and d3 , and then increases in subbands d 2 and d1 . The energy of the
signals, however, in the four highest frequency subband, d1 through d 4 , is very small.
5.10 Time-Frequency Distribution of Correlation across the Building Height
We conclude this chapter by presenting time frequency distributions of cross-correlation, in

this section for EW motions across the building height, and in the following section— for NS
motions across the building length. Figures 5.10.1 through 5.10.3 show time-frequency
distribution of correlation respectively between the EW accelerations at the ground floor and at
the roof, EW displacements at the ground floor and at the roof, and EW acceleration at the
ground floor and relative displacement at the roof. On the left hand side of each figure, the nodal
values of cross-energy are shown by vertical bars for the detail subbands 1 through 8, computed
by eqn (4.6.3.1), and then converted to continuous time estimates. On the right hand side, the
nodal values of the cross-power spectrum density are shown, plotted as horizontal levels over the
length of the tiles of the time-frequency plane (this way emphasizing that they represent
distributions), and with amplitude normalized by the maximum value in the subband, so that
their variation with time can be seen also in the low energy subbands. For a discrete time signal,
the cross-energy and cross-power spectrum density, computed by eqns (4.6.3.1) and (4.6.4.4),
have same numerical values, equal to the product of the corresponding wavelet coefficients.
Hence, the plots in the left hand side of Figs 5.10.1 through 5.10.3 are same as those in the right-
hand side, except for the amplitude normalization and style.
Each nodal value of the cross-energy time-frequency distribution equals the product of the
corresponding wavelet coefficients, which are projections of the signals onto the corresponding
tile of the phase plane. Hence, the amplitude of a nodal cross-energy is a measure of the energy
both signals have in that tile, while the sign indicates whether, on the overall, the signals are in
phase of out of phase within that tile. As the sums of the nodal cross-energies along the subbands
comprise the nodal energy spectrum, change of sign of the nodal cross-energies along a subband
leads to cancellations and a small spectral value for that subband even when the individual
signals have significant energy in that subband. Such cancellations occur within the subbands
the frequency of which is near a zero-crossing of the (interpolated) nodal energy spectrum, or of
cos α for that spectrum, as can be seen from a comparison of Figs 5.10.1 through 5.10.3 with
Figs 5.8.1 through 5.8.3. The physical explanation of these zero-crossings is included in Section
5.8. Cancellations of the nodal cross-energies occur also within subbands that are not near a
zero-crossing of the corresponding cross-energy spectrum, e.g. within the three highest
frequency subbands of the cross-energy between the acceleration at the ground floor and the
141
Correlation of EW Acceleration at Ground Floor and Roof
Ground floor Ground floor ω/π f

Roof Roof [rad/sample] [Hz]
d1 d1 3/4 18.75
d2 d2 3/8 9.37
d3 d3 3/16 4.69
d4 d4 3/32 2.34
142
d5 d5 3/64 1.17
0
m 2/s 3
d6 d6 3/128 0.59
d7 d7 3/256 0.29
d8 d8 3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.10.1 Time-frequency distribution of correlation between EW accelerations recorded at the ground floor and on the roof during
the 1994 Northridge earthquake. Left: cross-energy. Right: cross-PSD, normalized to unit amplitude for each subband.
Correlation of EW Displacement at Ground Floor and Roof
Ground floor Ground floor ω/π f

Roof Roof [Hz]
[rad/sample]
d1 d1 3/4 18.75
d2 d2 3/8 9.37
d3 d3 3/16 4.69
d4 d4 3/32 2.34
143
0.005
d5 d5 3/64 1.17
0
m 2s
d6 d6 3/128 0.59
d7 d7 3/256 0.29
d8 d8 3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.10.2 Time-frequency distribution of correlation between EW displacements at the ground floor and at the roof during the 1994
Northridge earthquake. Left: cross-energy. Right: cross-PSD, normalized to unit amplitude for each subband.
Correlation of EW Ground Floor Acceleration and Roof Relative Displacement
Ground floor accel. Ground floor accel. ω/π f

Roof rel. displ. Roof rel. displ. [rad/sample] [Hz]
d1 d1 3/4 18.75
d2 d2 3/8 9.37
d3 d3 3/16 4.69
d4 d4 3/32 2.34
0.05
144
d5 d5 3/64 1.17
0
m 2/s
d6 d6 3/128 0.59
d7 d7 3/256 0.29
d8 d8 3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.10.3 Time-frequency distribution of correlation between EW accelerations recorded at the ground floor and the relative roof
displacement during the 1994 Northridge earthquake. Left: cross-energy. Right: normalized cross-PSD.
acceleration at the roof, and is a result of “lack of coherence.” A more detailed analysis of the
cross-energy time-frequency distributions is out of the scope of this report.
5.11 Time-Frequency Distribution of Correlation across the Building Length
Finally, we show nodal time-frequency distributions of correlation between NS motions

recorded at the opposite ends of the building. Figures 5.11.1 and 5.11.2 correspond respectively
to acceleration and displacement recorded on the ground floor, and Figs 5.11.3 and 5.11.4—to
acceleration and displacement recorded on the roof. These figure show that, except for the
acceleration in the two highest frequency subbands, there is a very high degree of correlation
between the motions at both ends of the building, even at the roof, in spite of the contributions
from torsional vibrations, which are anti-symmetric at the two corners. The prevailing and
significant positive correlation of these motions indicates that most of the roof NS response is
due to foundation rocking and first NS mode relative deflection.
145
Correlation of NS Acceleration at Ground Floor
East side East side ω/π f

West side West side [rad/sample] [Hz]
d1 d1 3/4 18.75
d2 d2 3/8 9.37
d3 d3 3/16 4.69
d4 d4 3/32 2.34
146
d5 d5 3/64 1.17
0
m 2/s 3
d6 d6 3/128 0.59
d7 d7 3/256 0.29
d8 d8 3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.11.1 Time-frequency distribution of correlation between NS accelerations recorded on the ground floor at the East and West
ends of the building during the 1994 Northridge earthquake. Left: cross-energy. Right: normalized cross-PSD.
Correlation of NS Displacement at Ground Floor

d1 d1 3/4 18.75
d2 d2 3/8 9.37
d3 d3 3/16 4.69
d4 d4 3/32 2.34
0.004
147
d5 d5 3/64 1.17
0
m 2s
d6 d6 3/128 0.59
d7 d7 3/256 0.29
d8 d8 3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.11.2 Time-frequency distribution of correlation between NS displacements recorded on the ground floor at the East and West
ends of the building during the 1994 Northridge earthquake. Left: cross-energy. Right: normalized cross-PSD.
Correlation of NS Acceleration at Roof

d1 d1 3/4 18.75
d2 d2 3/8 9.37
d3 d3 3/16 4.69
d4 d4 3/32 2.34
148
d5 d5
3/64 1.17
0
m 2/s 3
d6 d6 3/128 0.59
d7 d7 3/256 0.29
d8 d8 3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.11.3 Time-frequency distribution of correlation between NS accelerations recorded on the roof at the East and West ends of the
building during the 1994 Northridge earthquake. Left: cross-energy. Right: normalized cross-PSD.
Correlation of NS Displacement at Roof

West side West side
[rad/sample] [Hz]
d1 d1 3/4 18.75
d2 d2 3/8 9.37
d3 d3 3/16 4.69
d4 d4 3/32 2.34
0.02
149
d5 d5 3/64 1.17
0
m 2s
d6 d6 3/128 0.59
d7 d7 3/256 0.29
d8 d8 3/512 0.15
0 500 1000 1500 n 2000 0 500 1000 1500 n 2000

0 10 20 30 t 40 s 0 10 20 30 t 40 s
Fig. 5.11.4 Time-frequency distribution of correlation between NS displacements recorded on the roof at the East and West ends of
the building during the 1994 Northridge earthquake. Left: cross-energy. Right: normalized cross-PSD.
6. ANALYSIS OF COMPRESSION ERROR FOR A MINI DATABASE
This chapter presents an analysis of the error associated with shrinkage of the smallest
wavelet coefficients as a function of the compression rate for a mini “database” of acceleration
recorded at the ground floor and at the roof of the 7-story hotel during six earthquakes,
representing a variety of excitations. The earthquakes are listed in chronological order in Table
6.1. Column 1 through 10 show respectively the order number, name, date, earthquake
magnitude, epicentral distance, number of samples in the time series considered, the frequencies
of the transition bands of Ormsby high- and low-pass filters used to band-pass filter the records,
peak acceleration of the EW and NS ground floor records and root-mean-square values of the
EW and NS ground floor records. The listed values for the NS motions are those for the motions
recorded at the East wall, except for the 1971 earthquake for which there was only one
instrument at the roof. Figure 6.1 shows the acceleration time series of this mini database,
plotted with same time and amplitude scales. It can be seen that these records represent a variety
of signals with different frequency contents and different degrees of nonstationarity. Figures 6.2
through 6.5 show the nodal energy spectra of the signals in Fig. 6.1, which will be used in the
discussion explaining the variations in error due to compression. The total number of points,
total energy, average power, and root mean square value are also listed for each signal.
Table 6.1 A list of earthquakes recorded in the building for which the error versus compression rate was
computed.
Earthquake Date M R N* f EW** NS** EW** NS**

km [Hz] amax amax arms arms
cm/s2 cm/s2 cm/s2 cm/s2
1 San 02/09/1971 6.6 22 2,920 0.05-0.10 129.3 247.9 26.6 36.6

Fernando to 25-27
2 Malibu 01/19//1989 5.0 36 1,254 0.10-0.20 21.6 16.1 2.9 3.4
to 25-27
3 Montebello 06/12/1989 4.1 34 1,200 0.10-0.20 21.6 12.1 2.4 1.4
to 25-27
4 Landers 06/28/1992 7.5 186 4,000 0.20-0.40 40.7 40.8 8.9 9.2
to 25-27
5 Northridge 01/17/1994 6.5 1.5 3,000 0.10-0.20 444.5 393.8 57.2 57.5
to 23-25
6 Northridge 03/20/1994 5.2 1.2 1,317 0.10-0.20 143.6 261.7 20.1 23.1
aft. to 23-25
* The sampling interval is ∆t = 0.02 s for all the records.

** Refer to the corresponding ground floor records.
150
E-W acceleration N-S acceleration at east wall
San Fernando (1971)

Base
Roof
Base 400
Malibu (1989)
Roof 200
-200
Base
Montebello (1989) -400
Roof (cm/s 2)
Landers (1991)
Base
Roof
Northridge (1994)
Base
Roof
Northridge March Aft. (1994)

Base
Roof
0 20 40 s 60 0 20 40 s 60
Fig. 6.1 Acceleration time histories used in the comparative study of compression error.
151
EW acceleration, Ground floor
Energy in subbands / Total energy 0.6 San Fernnado, 1971 0.6 Landers, 1992
N = 2,920 N = 4,000
E = 420 x10 2 cm 2/s 3 E = 60 x10 2 cm 2/s 3
0.4 0.4
P = 708 c m 2/s 4 P = 79 c m 2/s 4
r.m.s. = 26.6 cm/ s 2 r.m.s. = 8.9 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d7
d4
d3
d1
d7
d4
d3
d1
d8
d8
Energy in subbands / Total energy
0.6 Malibu, 1989 0.6 Northridge, 1994
N = 1,254 N = 3,000
0.4 E = 2 x10 2 cm 2 /s 3 0.4 E = 1960 x10 2 cm 2 /s 3

P = 8.3 c m 2/s 4 P = 3,268 c m 2/s 4
r.m.s. = 2.9 cm/ s 2 r.m.s. = 57.2 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d6
d5
d2
d7
d3
d1
d7
d4
d3
d1
d8
d8
0.6 Montebello, 1989 0.6 Northridge March aftershock, 1994
N = 1,200 N = 1317
E = 1.4 x10 2 cm 2/s 3 E = 100 x10 2 cm 2/s 3
0.4 0.4
P = 5.8 c m 2 /s 4 P = 403 c m 2/s 4
r.m.s. = 2.4 cm/ s 2 r.m.s. = 20.1 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d7
d3
d1
d6
d5
d2
d7
d4
d8
d3
d1
d8
Fig. 6.2 Nodal energy spectra of the EW acceleration recorded at the ground floor during six earthquakes,
shown in Fig. 6.1.
152
EW acceleration, Roof

0.6 San Fernnado, 1971 0.6 Landers, 1992
N = 2920 N = 4,000
E = 1,760 x10 2 cm 2/s 3 E = 620 x10 2 cm 2/s 3
0.4 P = 3,003 c m 2/s 4 0.4 P = 769 c m 2/s 4
r.m.s. = 54.8 cm/ s 2 r.m.s. = 27.7 cm/ s2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d7
d4
d7
d4
d3
d1
d3
d1
d8
d8
N = 1,254 N = 3,000
2 2 cm 2/s 3
E = 6.8 x10 cm 2/s 3 E = 3,600 x10
0.4 P = 26.8 c m 2/s 4 0.4 P = 5,987 c m 2 /s 4
r.m.s. = 5.2 cm/ s 2 r.m.s. = 77.4 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d6
d5
d2
d7
d3
d1
d7
d4
d3
d1
d8
d8
N = 1,200 N = 1317
E = 2 x10 2 cm 2/s 3 E = 80 x10 2 cm 2 /s 3
0.4 0.4
P = 8.5 c m 2/s 4 P = 322 c m 2/s 4
r.m.s. = 2.9 cm/ s 2 r.m.s. = 17.9 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d7
d3
d1
d6
d5
d2
d7
d4
d3
d1
d8
d8
Fig. 6.3 Nodal energy spectra of the EW acceleration recorded at the roof during six earthquakes, shown
in Fig. 6.1.
153
NS acceleration at East wall, Ground floor
N = 2,920 N = 4,000
2 cm 2/s 3 2
E = 780 x10 E = 60 x10 cm 2/s 3
0.4 0.4
P = 1,338 c m 2/s 4 P = 84 c m 2 /s 4
r.m.s. = 36.6 cm/ s2 r.m.s. = 9.2 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d7
d4
d3
d1
d7
d4
d3
d1
d8
d8
N = 1,254 N = 3,000
2 cm 2/s 3 2
E = 2.8 x10 E = 1980 x10 cm 2 /s 3
0.4 0.4
P = 11.5 c m 2/s 4 P = 3,305 c m 2/s 4
r.m.s. = 3.4 cm/ s 2 r.m.s. = 57.5 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d6
d5
d2
d7
d3
d1
d7
d4
d3
d1
d8
d8
N = 1,200 N = 1317
E = 0.5 x10 2 cm 2/s 3 E = 140 x10 2 cm 2 /s 3
0.4 0.4
P = 2.0 c m 2/s 4 P = 532 c m 2/s 4
r.m.s. = 1.4 cm/ s 2 r.m.s. = 23.1 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d7
d3
d1
d6
d5
d2
d7
d4
d8
d3
d1
d8
Fig. 6.4 Nodal energy spectra of the NS acceleration recorded at the ground floor (near the East wall)
during six earthquakes, shown in Fig. 6.1.
154
NS acceleration at East wall, Roof
N = 2,920 N = 4,000
E = 2,660 x10 2 cm 2/s 3 E = 300 x10 2 cm 2/s 3
0.4 0.4
P = 4,556 c m 2 /s 4 P = 386 c m 2 /s 4
r.m.s. = 67.5 cm/ s 2 r.m.s. = 19.7 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d5
d2
d6
d5
d2
d7
d4
d7
d4
d3
d1
d3
d1
d8
d8
N = 1,254 N = 3,000
2 cm 2 /s 3
E = 8 x10 2 cm 2/s 3 E = 3,320 x10
0.4 P = 31.8 c m 2/s 4 0.4 P = 5,539 c m 2/s 4
r.m.s. = 5.6 cm/ s2 r.m.s. = 74.4 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d6
d5
d2
d7
d3
d1
d7
d4
d3
d1
d8
d8
N = 1,200 N = 1317
E = 1 x10 2 cm 2/s 3 E = 94 x10 2 cm 2/s 3
0.4 0.4
P = 4.4 c m 2/s 4 P = 354 c m 2/s 4
r.m.s. = 2.1 cm/ s 2 r.m.s. = 18.8 cm/ s 2
0.2 0.2
0 0
0.1 1 f - Hz 10 0.1 1 f - Hz 10
d6
d2
d5
d4
d7
d3
d1
d6
d5
d2
d7
d4
d3
d1
d8
d8
Fig. 6.5 Nodal energy spectra of the NS acceleration recorded at the roof (near the East wall) during six
earthquakes, shown in Fig. 6.1.
155
The error associated with the compression is illustrated in terms of the following normalized
measures of errors:

∞ i
= (6.1)
σs σs
and
∑ ( si − sî )
2
/N
Root Mean Square Error i
= (6.2)
σs σs
which are plotted versus the compression rate, defined here as
n top
Compression rate = 1 − (6.3)
N
In Chapter 5, Section 5.4, we also showed two other measures of error

∞ i
= (6.4)
s ∞
max si
i
and
∑ ( si − sî )
2
Error 2 i
= (6.5)
s 2 ∑ si 2
i
for the EW ground floor and roof acceleration during the Northridge earthquake only, and
concluded that Error 2 / s 2 = Root Mean Square Error / σ s because the signals are zero mean, and
that Error ∞
/ s ∞
≈ Error 2 / s 2
(hence, we consider them practically redundant). Figures 6.6
and 6.7 show normalized error Error ∞
/ σ s respectively for the EW and NS accelerations
recorded at the ground floor (top) and at the roof (bottom), and Figs 6.8 and 6.9 show normalized
error Root Mean Square Error / σ s for the same motions. The entire available digitized length of
the signals was considered, and the signals were zero-padded so that the total number of points
was equal to the nearest power of 2, so that the fast (pyramid) algorithm could be applied. The
compression rate was computed by dividing the number of top coefficients by the original
number of points (i.e. by the number of non-trivial samples of the transformed time series).
156
3
10
EW acceleration: Ground
2
10 Error
100
8
σ
s
1
10
0
10
Montebello, 1989
-1 Malibu, 1989
10
Northridge Mar. Aft.
Landers, 1991 1994
-2
10 Northridge, 1994
San Fernando, 1971
-3
10
-4
10
1 10 100
3
10
EW acceleration: Roof
2 Error
10
100
8
σ
s
1
10
0
10
-1 Montebello, 1989
10 Malibu, 1989
Northridge Mar. Aft., 1994
-2
10 Landers, 1991
Northridge, 1994
-3 San Fernando, 1971
10
-4
10
1 10 100
Fig. 6.6 Normalized peak error by the standard deviation of the signal, for the EW accelerations recorded
on the ground floor and on the roof during six earthquakes.
157
3
10
NS acceleration: Ground
2
10 Error
100
8
σ
s
1
10
0
10
Montebello, 1989
-1 Malibu, 1989
10
Landers, 1991
-2 1994
10 Northridge, 1994
San Fernando, 1971
-3
10
-4
10
1 10 100
3
10
NS acceleration: Roof
2 Error
10
100
8
σ
s
1
10
0
10
Montebello, 1989
-1 Malibu, 1989
10
1994
-2 Landers, 1991
10 Northridge, 1994
San Fernando, 1971

-3
10
-4
10
1 10 100
Fig. 6.7 Normalized peak error by the standard deviation of the signal, for the NS accelerations recorded
on the ground floor and on the roof (near the east wall) during six earthquakes.
158
3
10 EW acceleration: Ground
2
10 Root Mean Square Error
100
σ
s
1
10
0
10
-1 Landers, 1991
10
Northridge
Mar. Aft., 1994
-2
10 Montebello, 1989
Malibu, 1989
-3 Northridge, 1994
10
San Fernando, 1971
-4
10
1 10 100
3
10 EW acceleration: Roof
10
100
σ
s
1
10
0
10
-1
10
Montebello, 1989
-2 Malibu, 1989
10
1994
-3 Landers, 1991
10 Northridge, 1994
San Fernando, 1971
-4
10
1 10 100
Fig. 6.8 Normalized root mean square error by the standard deviation of the signal, for the EW
accelerations recorded on the ground floor and on the roof during six earthquakes.
159
3
10 NS acceleration: Ground
2
100
σ
s
1
10
0
10
-1 Landers, 1991
10
Northridge
Mar. Aft., 1994
-2
10 Montebello, 1989
Malibu, 1989
-3 Northridge, 1994
10
San Fernando, 1971
-4
10
1 10 100
3
10 NS acceleration: Roof
10
100
σ
s
1
10
0
10
-1
10
Montebello, 1989
-2 Malibu, 1989
10
Landers, 1991 1994
-3
10 Northridge, 1994
San Fernando, 1971
-4
10
1 10 100
Fig. 6.9 Normalized root mean square error by the standard deviation of the signal, for the NS
accelerations recorded on the ground floor and on the roof (near the East wall) during six earthquakes.
160
As it can be seen from the plots of the signals in Fig. 6.1, the set of 6 earthquakes considered
represents a good variety, in terms of amplitudes, frequency content, and “degree” of being
transient in nature. The 1971 San Fernando and 1994 Northridge earthquakes signals are
mutually similar by their large amplitudes, relatively high frequency content, and transient
nature, for which, as it can be expected, the normalized error is very similar and is the smallest
compared to the error for the other earthquakes. Both of these events were of moderate size but
occurred very close to the building, and caused the largest amplitude responses, which damaged
the building, and caused nonlinear response of the local soil. Other two mutually similar signals
are those from the 1989 Malibu and Montebello earthquakes, for which the normalized error is
also similar, and is the largest compared to the other earthquakes. Both of these events had
smaller magnitude and occurred relatively close to the building site (see Table 6.1), so that the
acceleration records in the building have more energy in the higher frequency subbands. The
motions from the 1992 Landers and 1994 Northridge aftershock are very different in nature. The
former (1992 Landers) is a large magnitude but distant event, hence with more of the ground
motion energy at the building site in the longer periods, and causing moderate level of response
at the roof, also with more energy in the longer periods. In contrast, the latter (1994 Northridge
aftershock) is a relatively small event but very close to the building site, hence the ground
motion is rich in high frequencies. Interestingly, the normalized error is very similar for the
ground floor records of these two earthquakes, even though these two signals have very different
nature. Intuitively, the normalized error is expected to be smaller for the 1992 Landers
earthquake because of the longer period content. However, this record is also more “stationary”
in nature compared to the 1994 Northridge aftershock record, which has lead to larger error.
Apparently, the more transient nature of the Northridge aftershock records at ground level has
compensated for their high frequency content effect on the error—hence the error curves are
similar to those for the Landers earthquake. The error for the roof response records of these two
earthquakes is significantly different. It is significantly smaller for the 1992 Landers earthquake,
which can be explained by the fact that the roof records for this earthquake are more transient in
nature than the ground response records. A comparison of the error for the EW and NS recorded
responses shows that there is no apparent significant difference depending on the component of
motion considered. A comparison of the errors for the roof records with those for the ground
floor records shows that the error is smaller for the roof records, especially for larger
compression rates, as it can be expected because of the more narrow-band and longer period
nature of the roof response signals.
161
7. DISCUSSION, SUMMARY AND CONCLUSIONS
7.1 Discussion
As a time-frequency distribution, the wavelet transform provides insight simultaneously into

the frequency and temporal characteristics of a signal, and hence is convenient for analyses of
transients and time varying systems. For a critically sampled discrete wavelet transform, this
time-frequency distribution is nodal. This report illustrated the insight that can be gained from
nodal time-frequency distributions of strong motion data recorded by a structural array in a 7-
story reinforced building in the Los Angeles area. Besides the distribution of the wavelet
coefficients, this report showed nodal time-frequency distributions of energy, power, and power
spectrum density for the motions at ground level and at the roof, as well as of correlation
between motions recorded by sensors distributed horizontally and vertically along the structure.
It also illustrated how discontinuities in the response (possibly due to sudden loss of stiffness,
e.g. from structural damage) produce large amplitude energy pulses in the high frequency
subbands, providing a possible explanation for such pulses observed for the data of the 1994
Northridge earthquake, which damaged the building. Such applications of the discrete wavelet
transform was first introduced in analyses of mechanical vibrations by Newland (1993a,b;
1994a,b). Its use for detection of cracks in a structure was later studied and demonstrated for
numerically simulated response by a number of investigators, e.g., Staszewski and Tomlinson
(1994), Deng and Wang (1998), Wang and Deng (1999), Hou et al. (2000), and Gentille and
Messina (2003).
As shown in this report, the expansion of strong motion data in orthonormal wavelet series
can be used to estimate nodal spectra of energy, power, power density, and Fourier transform
amplitude or a particular signal, as well as nodal spectra of cross-energy, cross-power, cross-
power density, and the angle between two signals, showing the frequency nature of their
correlation. These spectra are marginal distributions of the corresponding nodal time-frequency
distributions, mentioned in the preceding paragraph, and in turn can be used to produce single
value estimates of these quantities (e.g., total energy and average power of a signal, and
correlation and angle between two signals).
The nodal time-frequency distribution, spectrum and single value estimate of the signal
energy were interpreted as aggregates of the energy within a tile of the phase plane, a frequency
band of the phase plane, and of the entire phase plane. Similarly, the time-frequency and
spectral distributions of power density were interpreted as local averages of the power density
within a tile and a frequency band of the phase plane, and the single value estimate of power—as
the global average of the power over the entire length of the signal. Hence, all of these nodal
distributions and global estimates can be interpreted as information granules, encapsulating
information about a property of the signal over an interval of the frequency axis ( ) or a patch
162
of the Cartesian product of the frequency axis and the time axis ( × ). All of these estimates
of energy, power, and power density were obtained as a linear combination of the modulus
square of the wavelet coefficients. As each wavelet coefficient represents an orthogonal
projection of the signal onto a particular tile of the phase plane, the wavelet coefficients
themselves can be viewed as information granules. Further, for two signals, the products of the
corresponding wavelet coefficients and their linear combinations can be used to produce various
information granules of the local and global correlation and degree of correlation. Such
interpretation of the discrete orthonormal wavelet transform, and estimation of spectra and
spectral densities of signals and of their derivatives and integrals—all in the wavelet domain, are
novelties in analyses of mechanical vibrations.
The multi-resolution structure of the wavelet transform is such that the ratio of the subband
width to the central frequency is equal for all the subbands, resulting in equal subband bandwidth
on a logarithmic frequency scale. Hence, wavelet analysis is most appropriate for processes that
are analyzed on a logarithmic frequency axis, and that benefit from better frequency resolution
for the longer period components. It is most appropriate for analysis of ground motion records,
which is usually done on a logarithmic frequency axis. For structural response records, it is an
appropriate tool for detection and time localization of sudden changes in the system (possibly
indicative of structural degradation), which is done from the highest frequency components, for
which the time resolution is very good. The examples presented in this report showed that the
nodal Fourier and power spectrum density spectra do indicate the location of the system
frequencies, but the density of nodes (in frequency), at least for the dyadic (or octave band)
critically sampled discrete wavelet transform, is too low for accurate detection of the system
frequencies and their temporal changes. The density of nodes can be increased while preserving
the multi-resolution hierarchy by using discrete wavelet transforms that are not dyadic (i.e. with
a0 ≠ 2 , see eqns (3.2.8) and (3.2.9)). An even better tool to use is wavelet packet expansions,
which can subdivide further some or all of the subbands of the dyadic wavelet transform. While
departing from the multi-resolution structure of the wavelet transform, the wavelet packet
expansions do preserve the orthogonal structure and nodal nature, while enabling a high degree
of flexibility in choosing the position of the nodes (i.e. in the partition of the phase plane into
tiles). The theory presented in Chapter 4, for the estimation of various global and local
aggregates and averages, is directly extensible to wavelet packets, which is a natural continuation
of the work presented in this report.
The best choice of a wavelet depends on the particular application. Because of the
Heisenberg uncertainty principle, there is no wavelet that is perfectly localized both in time and
in frequency, and choosing a particular wavelet is a compromise. Wavelets with more compact
support (i.e. described by shorter filters) are better localized in time and worse in frequency, and
vice versa. For the s8 wavelet, e.g., there is a significant leakage of energy in the neighboring
163
subbands, and if spectral analysis is of primary importance, longer wavelets have to be used (e.g.
the s20 wavelet).
The discrete nodal structure of the orthonormal wavelet series and wavelet packet
expansions and of the associated marginal distributions would be very convenient for automated
analyses of large sets of data (possibly at database level, Shahabi et al., 2001), e.g. by cluster
analysis, statistical regressions, and pattern recognition techniques. The orthogonality property
is particularly convenient for the estimation of energy—of ground motion that excites structures,
and of structural response—previously recognized by Iyama and Kuwamura (1999), who used
the discrete wavelet transform to study the cumulative energy and the rate of energy input for
ground motions of the 1995 Hyogoken-Nanbu earthquake in Japan.
This report also showed that the wavelet transform is a very efficient tool for compression of
strong motion records at a very low loss of accuracy. The presented results showed that, for a
particular earthquake, the error due to compression was smaller for the structural response
records than for ground motion records, in particular for high compression ratios. This is
explained by (a) the more narrow-band nature of the structural response records, and (b) more of
the energy in the longer period subbands (which are represented exactly by less coefficients than
the higher frequency subbands). Considering the fact that the first mode contributes most to the
total response, the compression error is expected to be smaller for structural response records of
longer period structures than for shorter period structures, and for smaller damping ratios. For
ground motion records, the compression error is smaller for motions that are more transient in
nature, and that have more energy in the longer periods. Hence, for given acceptable error,
higher compression ratios can be expected for larger magnitude earthquakes, for larger epicentral
distances, and for type of travel paths and local site conditions such that attenuate more the
higher frequencies of ground motion and amplify the longer period components (e.g. “soft”
sites). For records of ambient noise response, the compression rates are expected to be much
smaller, due to the stationary nature of these motions, and due to their significant high frequency
content.
The data compression property of the discrete wavelet transform could be useful for
reducing data storage requirements for vibration data for which loss of accuracy is permissible,
e.g. for excitation by ambient noise, or by very small local or very distant earthquakes, that
otherwise might not have been permanently preserved.
A more significant and universal use of the data compression property is to reduce the
dimensionality of time series data, for the purpose of facilitating a particular analysis.
Dimensionality reduction can be achieved by eliminating the small detail that is not necessary for
the analysis (i.e. by creating a lower resolution approximation), or by eliminating the less
significant features (i.e. by setting to zero the wavelet coefficients that are smaller than an
164
adopted threshold level). Both of these procedures perform denoising, where “noise” may also
include features that can be explained but are not of interest for the analysis. Wavelet based
compression and feature selection has been proposed and demonstrated for vibration of
machinery by Staszewsky (1998a). In addition to thresholding, Staszewski (1998a) also
considered optimal wavelet compression, obtained using a genetic algorithm, taking into
consideration the position of the wavelet coefficients in addition to their amplitude, which
resulted in more accurate time localization of the preserved features in the compressed data.
Denoising (compression) by thresholding has been known to produce artifacts due to the sparse
distribution of the wavelet coefficients in a critically sampled discrete wavelet transform. These
artifacts can also be avoided by using oversampled expansions (Bruce and Gao, 1996). An
important application of pattern recognition in the wavelet domain is structural health
monitoring.
7.2 Summary and Conclusions
The objective of this report was to examine the usefulness of the orthonormal discrete
wavelet transform as a preprocessing tool in data mining of large sets of ground and structural
response vibration data under earthquake, ambient noise, or forced excitation. The report
reviews the basics theory of the transform and how it can be computed—in the frequency
domain using FFT and by the pyramid algorithm using quadrature mirror filters. It also reviews
the theory behind its use for data compression—by thresholding and by lower resolution
approximation, and the theory behind the interpretation of the wavelet coefficients themselves
and their square values for a particular signal, of their products for two signals, and the
corresponding sums as information granules representative of the properties of the signal(s) on
an interval of or on a rectangle of × . Illustrations are presented for seismic monitoring
data recorded by a structural array in a building in the Los Angeles area. The concepts are
illustrated on data from the 1994 Northridge earthquake, and compression ratios and associated
error are illustrated and compared for data from six earthquakes representing a variety of motions
(by amplitude, frequency content, and “degree” of nonstationarity).
From the theory and illustrations presented in this report, it can be concluded that expansion
in orthonormal wavelet series is potentially a very useful preprocessing tool in mining large data
sets of ground and structural response vibration data under earthquake excitation. The beneficial
properties of the transform are: linearity (enabling computation of the transform of linear
combination of signals from the transforms of the individual signals), orthogonality (simplifying
analyses due to various cancellations, in particular useful in estimation of energy and
correlation), time-frequency localization (convenient for analysis of time varying systems; also
enables estimation of the transform of derivatives and integrals of the signal from the transform
of the signal), information granulation (i.e. encapsulation of information about an interval into a
165
single value, hence providing more robust estimation than single values), data compression
(enabling a very high degree of dimensionality reduction while preserving the significant
features at all scales), multi-resolution structure (enabling separation of features that appear at
different scales, as well as dimensionality reduction by lower resolution approximation), and
discrete structure (convenient for automated analysis in data mining). One drawback of the
orthonormal wavelet transform is poor resolution at high frequencies. This drawback can be
eliminated in principle by further expansion of the subbands that require better frequency
resolution in wavelet packets, which will be demonstrated in our future work.
The theory and concepts presented apply directly to datasets of any time series data. Similar
conclusions as those reached in this report are expected for ground and structural response
vibration monitoring data caused by other natural forces (e.g. ambient noise and wind), as well as
by man made forces (e.g. created by explosions or actuators). An important application of such
data is in structural health monitoring.
Similar conclusions are also expected for vibration data from laboratory experiments, a large
amount of which will be generated by the experimental facilities of the Network for Earthquake
Engineering Simulation. The errors associated with data compression, however, will be much
larger for time series data that are more stationary in nature.
166
REFERENCES
1. Amaratunga, K., and J.R. Williams (1993). “Wavelet based Green’s functions approach to 2D PDEs”,
Engineering Computations, 10(4), 349-367.
2. Abbate, A., C.M. DeCusatis, and P.K. Das (2002). “Wavelets and Subbands,” Birkhäuser, Boston,
Massachusetts.
3. Antoniadis, A., and G. Oppenheim (1995). “Wavelets and Statistics”, Springer-Verlag, New York,
NY.
4. Basu, B., and V.K. Gupta (1997a). “On wavelet-analyzed seismic response of SDOF systems”,
ASME, Proc. Design Engineering Technical Conference, Sept. 14-17, 1997, Sacramento CA.
5. Basu, B. and V.K. Gupta (1997b). “Non-stationary seismic response of MDOF systems by wavelet
transform”, Earthquake Engineering & Structural Dynamics, 26(12), 1243-1258.
6. Basu, B. and Gupta, V. K. (1999). “Wavelet-based analysis of the non-stationary response of a
slipping foundation”, Journal of Sound and Vibration, 222 (4), 547-563.
7. Benz, H., R. Buland, J. Filson, A. Frankel, and K. Sheldok (2001). “Advanced National Seismic
System,” Seism. Res. Letters, 72(1), 70-75.
8. Blume, J.A. and Assoc. (1973). “Holiday inn,” Chapter 29 in “San Fernando, California earthquake
of February 9, 1971,” Volume I, Part A, U.S. Dept. of Commerce, National Oceanic and Atmospheric
Administration, Washington, D.C.
9. Bruce, A., and H.-Y. Gao (1996). “Applied Wavelet Analysis with S-Plus”, Springer-Verlag, New
York, NY.
10. Ching, J., and S.D. Glaser (2003). “Tracking rapidly changing dynamical systems using a non-
parametric statistical method based on wavelets,” Earthquake Engineering and Structural Dynamics,
32, 2377-2406.
11. Cios, K.J., W. Pedrycz, and R. Swiniarski (1998). “Data mining methods for knowledge discovery,”
Kluwer Academic Publishers, Boston, Massachusetts.
12. Consortium of Organizations for Strong-Motion Observations Systems (COSMOS) (2001). “Strong-
Motion Instrumentation of Buildings,” Proceedings, Invited Workshop on Strong-Motion
Instrumentation of Buildings, November 14-15, 2001, Emeryville, California, COSMOS Publication
No. CP-2001/04.
13. Daubechies, I. (1988). “Orthonormal Bases of Compactly Supported Wavelets”, Comm. Pure and
Appl. Maths, Vol. XLI, 909-996.
14. Daubechies, I. (1992). “Ten lectures on wavelets,” Society for Industrial Application of Mathematics
(SIAM), Philadelphia, Pennsylvania.
15. De la Llera, J.C., Chopra, A.K., and Almazan, J.L. (2001). “Three-dimensional inelastic response of
an RC building during the Northridge earthquake,” J. of Struct. Eng., ASCE, 127(5), 482-489.
16. Deng, X., and Q. Wang (1998). “Crack detection using spatial measurements and wavelet analysis,”
International J. of Fracture, 91, L23-L28.
17. Donoho, L.D. (1995). “Denoising by soft thresholding,” IEEE Transactions on Information Theory,
41(3), 613-627.
18. Donoho, L.D., and I.M. Johnstone (1994). “Ideal spatial adaptation via wavelet shinkage,”
Biometrika, 81, 425-455.
167
19. Freeman, S.A. and K.K. Honda (1973). Response of two identical seven-story structures to the San
Fernando earthquake of February 9, 1971, J.A. Blume Assoc. Report, submitted to Nevada Operations
Office, U.S. Atomic Energy Commission.
20. Frawley, W.J., G. Piatetsky-Shapiro, and C.J. Matheus (1991). “Knowledge discovery in databases:
an overview.” In G. Piatetsky-Shapiro and W.J. Frawley (eds.) “Knowledge discovery in databases,”
AAAI/MIT Press, 1-27.
21. Gentile, A., and A. Messina (2003). “On the continuous wavelet transform applied to discrete
vibrational data for detecting open cracks in damaged beams,” International J. of Solids and
Structures, 40, 295-315.
22. Ghanem, R., and F. Romeo (2000). “A Wavelet-Based Approach for the Identification of Linear
Time-Varying Dynamical Systems”, J. of Sound and Vibration, 234(4), 555-576.
23. Grossmann, A., and J. Morlet (1984). “Decomposition of Hardy Functions into Square Integrable
Wavelets of Constant Shape”, SIAM J. of Math. Anal., 15(4), 723-736.
24. Goupillaud, P., A. Grossmann and J. Morlet (1984/85). “Cysle-Octave and Related Transforms in
Seismic Signal Analysis”, Geoexploration, 23, 85-102.
25. Gurley, K., and A. Kareem (1999). “Applications of wavelet transform in earthquake, wind and
ocean engineering,” Engineering Structures, 21, 149-167.
26. Hauksson, E., P. Small, K. Hafner, R. Busby, R. Clayton, J. Goltz, T. Heaton, K. Hutton, H.
Kanamori, J. Polet, D. Given, L.M. Jones, and D. Wald (2001). “Southern California Seismic
network: Caltech/USGS Element of TriNet 1997-2001,” Seism. Res. Letters, 72(6), 690-704.
27. Hefner, K., and R. Clayton (2001). “The Southern California Earthquake Data Center (SCEDC),”
Seism. Res. Letters, 72(6), 705-711.
28. Hou, Z., M. Noori, and R. St. Amand (2000). “Wavelet based approach for structural damage
detection,” J. of Engineering Mechanics, ASCE, 126(7), 677-683.
29. Hudson, D.E., M.D. Trifunac, A.G. Brady and A. Vijayaraghavan (1971). Strong-motion earthquake
accelerograms, II, corrected accelerograms and integrated velocity and displacement curves, Report
EERL 71-51, Earthquake Eng. Res. Lab., Calif. Inst. of Tech., Pasadena, California.
30. Islam, M.S. (1996). Analysis of the response of an instrumented 7-story nonductile concrete frame
building damaged during the Northridge earthquake, Professional paper 96-9, Los Angeles Tall
Building Structural Design Council, 1996 Annual Meeting.
31. Ivanović, S.S., M.D. Trifunac, E.I. Novikova, A.A. Gladkov, and M.I. Todorovska (1999).
“Instrumented 7-storey reinforced concrete building in Van Nuys, California: ambient vibration
surveys following the damage from the 1994 Northridge earthquake,” Report CE 99-03, Department
of Civil Engineering, University of Southern California, Los Angeles, California.
32. Iyama, J., and H. Kuwamura (1999). “Application of wavelets to analysis and simulation of
earthquake motions,” Earthquake Engineering and Structural Dynamics, 28, 255-272.
33. Kinoshita, S. (1998). “Kyoshin net (K-Net),” Seism. Res. Letters, 69(4), 309-332.
34. Lee, V.W. and M.D. Trifunac (1987). “Strong Earthquake Ground Motion Data in EQUINFOS: Part
I,” Dept. of Civil Engineering Report CE 87-01, University of Southern California, Los Angeles.
35. Lee, V.W. and M.D. Trifunac (1990). “Automatic digitization and processing of accelerograms using
PC,” Report No. 90-03, Dept. of Civil Eng. U. of So. California, Los Angeles, CA.
36. Lee, W.H.K., T.C. Shin, K.W. Kuo, K.C. Chen, and C.F. Wu (2001). “CWB free-field strong motion
data from the 21 September Chi-Chi, Taiwan, earthquake,” Bull. Seism. Soc. Am., 91(5), 1370-1376.
168
37. Mallat, S.G. (1989). “Multiresolution approximations and wavelet orthonormal bases of L2(R),”
Trans. Amer. Math. Soc, 315, 69-87.
38. Mallat, S. and Z. Zhang (1993). “Matching Pursuit with Time Frequency Dictionaries”, IEEE Trans.
on Signal Processing, 41(12), 3397-3415.
39. Meyer, Y. (1990). “Ondelettes et opérateurs,” Vol. I and II, Hemann, Paris.
40. Meyer, Y. (1993). “Wavelets: algorithms and applications,” Society for Industrial and Applied
Mathematics (SIAM), Philadelphia.
41. Morlet, J., G. Arens, E. Fourgeau and D. Giard (1982a). “Wave Propagation and Sampling Theory-
Part I: Complex Signal and Scattering in Multilayred Media”, Geophysics, 47(2), 203-221.
42. Morlet, J., G. Arens, E. Fourgeau and D. Giard (1982b). “Wave Propagation and Sampling Theory-
Part II: Sampling Theory and Complex Waves”, Geophysics, 47(2), 222-236.
43. Mukherjee, S., and V.K. Gupta (2002). “Wavelet based characterization of design ground motions,”
Earthquake Engineering and Structural Dynamics, 31, 1173-1190.
44. Newland, D.E (1993a). “Random Vibrations, Spectral and Wavelet Analysis”, 3rd ed., Longman
Scientific & Technical, New York.
45. Newland, D.E (1993b). “Harmonic wavelet analysis,” Proc. Royal Soc. London A Mat, 443, 203-225.
46. Newland, D.E (1994a). “Wavelet Analysis of Vibration, Part I: Theory”, ASME, J. of Vibration and
Acoustics, 116, 409-416.
47. Newland, D.E (1994b). “Wavelet Analysis of Vibration, Part II: Wavelet Maps”, ASME, J. of
Vibration and Acoustics, 116, 417-425.
48. Oppenheim, A.V. and R.W. Schafer (1999). “Discrete-time signal processing,” 2nd edition, Prentice
Hall, Upper Saddle River, New Jersey.
49. Pan, T.-C., and C.L. Lee (2002). “Application of wavelet theory to identify yielding in seismic
response of bi-linear structures,” Earthquake Engineering and Structural Dynamics, 31, 379-398.
50. Pettit, C.L., N. Jones and R. Ghanem (2000). “Wavelet-Based Detection and Classification of Roof-
Corner Pressure Transients”, Wind and Structures, 3(3), 159-175.
51. Rezai, M., P. Rahmatian and C. E. Ventura (1996). “Seismic data analysis of a seven-storey building
using frequency response function and wavelet transform,” Proc. NEHRP Conf. And Workshop on
Research on the Northridge, California Earthquake of January 17, 1994, Vol. III, 421-428.
52. Shahabi, C, X. Tian and W. Zhao (2000). “TSA-tree: A Wavelet-Based Approach to Improve the
Efficiency of Multi-Level Surprise and Trend Queries on Time-Series Data,” The 12th International
Conference on Scientific and Statistical Database Management (SSDBM 2000), Berlin, Germany,
July, 2000.
53. Shahabi, C., S. Chung, M. Safar and G. Hajj (2001a). “2D TSA-tree: A Wavelet-Based Approach to
Improve the Efficiency of Multi-Level Spatial Data Mining,” The 13th International Conference on
Scientific and Statistical Database Management (SSDBM 2001), Fairfax, Virginia, July 2001.
54. Shahabi, C., S. Chung and M. Safar (2001b). “A Wavelet-Based Approach to Improve the Efficiency
of Multi-Level Surprise Mining,” PAKDD International Workshop on Mining Spatial and Temporal
Data 2001, Hong Kong, May 2001.
55. Shakal, A., M. Huang, R. Darragh, T. Cao, R. Sherburne, P. Malhotra, C. Cramer, R. Syndov, V.
Graizer, G. Maldonado, C. Peterson and J. Wimpole (1994). CSMIP strong motion records from the
169
Northridge, California, Earthquake of 17 January 1994, Report No. OSMS 94-07, Calif. Dept. of
Conservation, Div. of Mines and Geology, Sacramento, California.
56. Spanos, P.D., and V.R.S. Rao (2001). “Random field representation in biorthogonal wavelet basis,” J.
of Engineering Mech., ASCE, 127(2), 194-205.
57. Staszewski, W.J. (1998a). “Wavelet based compression and feature selection for vibration analysis,”
J. of Sound and Vibration, 211(5), 735-760.
58. Staszewski, W.J. (1998b). Identification of non-linear systems using multiscale ridges and skeletons
of the wavelet transform, J. of Sound and Vibration, 214(4), 639-658.
59. Staszewski, W.J., and G.R. Tomlinson (1994). “Application of the wavelet transform to fault
detection in a spur gear,” Mechanical Systems and Signal Processing, 8, 289-307.
60. Strang, G. and T. Nguyen (1997). “Wavelets and filter banks,” Wellesley-Cambridge Press, Wessley,
Massachusetts.
61. Sun, Z., and C.C. Chang (2002). “Structural damage assessment based on wavelet packet transform,”
J. of Structural Engineering, ASCE, 128(10), 1354-1361.
62. Tang, Y.Y., J.M. Liu, H. Ma and B.F. Li (1999). “Wavelet Orthonormal Decomposition for
Extracting Features in pattern Recognition,” Int. J. of Pattern Recognition and Artificial Intelligence,
13(6), 803-831.
63. Todorovska,M.I. (2001). “Estimation of instantaneous frequency of signals using the continuous
wavelet transform,” Report CE 01-07, Dept. of Civil Engrg., Univ. of Southern California.
64. Todorovska, M.I., and M.D. Trifunac (1990). “Analytical model for in-plane building-foundation-soil
interaction: incident P-, S- and Rayleigh waves,” Dept. of Civil Eng., Report CE Report CE 90-01,
Dept. of Civil Engrg., Univ. of Southern California.
65. Todorovska, M.I., S.S. Ivanović and M.D. Trifunac (2001a). “Wave propagation in a seven-story
reinforced concrete building: I. theoretical models,” Soil Dynam. and Earthquake Engrg, 21, 211-
223.
66. Todorovska, M.I., S.S. Ivanović & M.D. Trifunac (2001b). “Wave propagation in a seven-story
reinforced concrete building, Part II: observed wavenumbers,” Soil Dynam. and Earthquake Engrg,
21(3), 224-236.
67. Trifunac, M.D. (1974). A three-dimensional dislocation model for the San Fernando, California,
Earthquake of February 9, 1971, Bull. Seism. Soc. Amer., 64, 149-172.
68. Trifunac, M.D. and V.W. Lee (1973). Routine computer processing of strong motion accelerograms,
Earthquake Eng. Res. Lab., Report EERL 73-03, Calif. Inst. of Tech., Pasadena, California.
69. Trifunac, M.D. and V.W. Lee (1979). Automatic digitization and processing of strong-motion
accelerograms, Parts I and II, Dept. of Civil Eng., Report CE 79-15, Univ. of Southern California,
Los Angeles, California.
70. Trifunac, M.D., and M.I. Todorovska (2001). “Evolution of accelerographs, data processing, strong
motion arrays and amplitude and spatial resolution in recording strong earthquake motion,” Soil
Dynamics and Earthquake Engineering, 21, 537-555.
71. Trifunac, M.D., A.G. Brady and D.E. Hudson (1973). Strong motion earthquake accelerograms,
digitized and plotted data, Volume II, Part C, Earthquake Eng. Res. Lab., Report EERL-72-51, Calif.
Institute of Tech., Pasadena, California.
72. Trifunac, M.D., S.S. Ivanović & M.I. Todorovska (1999a). “Instrumented 7-storey reinforced
concrete building in Van Nuys, California: description of damage from the 1994 Northridge
170
earthquake and strong motion data,” Dept. of Civil Eng., Report CE 99-02, Univ. of Southern
California, Los Angeles, California.
73. Trifunac, M.D., V.W. Lee & M.I. Todorovska (1999b). “Common problems in automatic digitization
of accelerograms,” Soil Dynamics & Earthquake Engrg, 18, 519-530.
74. Trifunac, M.D., T.Y. Hao and M.I. Todorovska (2001a). “On energy flow in earthquake response,”
Report CE 01-03, Dept. of Civil Engrg., Univ. of Southern California.
75. Trifunac, M.D., T.Y. Hao and M.I. Todorovska (2001b). “Energy of earthquake response as a design
tool,” Proc. 13th Mexican National Conf. on Earthquake Eng., Guadalajara, Mexico.
76. Trifunac, M.D., S.S. Ivanović and M.I. Todorovska, (2001c). Apparent periods of a building. I:
Fourier analysis, ASCE, J. of Struct. Engrg, 127(5), 517-526.
77. Trifunac, M.D., S.S. Ivanović and M.I. Todorovska, (2001d). Apparent periods of a building. II: time-
frequency analysis:, ASCE, J. of Struct. Engrg, 127(5), 527-537.
78. Trifunac, M.D., S.S. Ivanović and M.I. Todorovska, (2003). Wave propagation in a seven-story
reinforced concrete building: III. Damage detection via changes in wavenumbers, Soil Dynam. and
Earthquake Engrg, 23(1), 65-75.
79. Ventura, C.E., and M. Rezai (1999). “Wavelet transform of ground motions recorded at two sites
during Loma Prieta earthquake,” ISET J. of Earthquake Technology, 36(2-4), 85-106.
80. Vetterli, M., and J. Kovacević (1995). “Wavelets and Subband Coding”, Prentice Hall PTR, Upper
Saddle River, New Jersey.
81. Wald, D.J., Heaton, T.H. and Hudnut. K.W. (1996). The slip history of the 1994 Northridge,
California, earthquake determined from strong motion, teleseismic, GPS and leveling data, Bull.
Seism. Soc. Amer., 86(1B), S49-S70.
82. Wang, Q., and X. Deng (1999). “Damage detection with spatial wavelets,” International J. of Solids
and Structures, 36, 3443-3468.
83. Williams, J.R., and K. Amaratunga (1995). “A Multiscale Wavelet Solver with O(n) Complexity”, J.
of Computational Physics, 122, 30-38.
84. Zeldin, B., and P.D. Spanos (1996). “Random field representation and synthesis using wavelet
bases,” J. of Applied Mech., ASME, 63, 946-952.
171

Uploaded by

Document Information

Original Description:

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

Uploaded by

Copyright:

Available Formats

UNIVERSITY OF SOUTHERN CALIFORNIA

Department of Civil Engineering

INFORMATION GRANULATION AND DIMENSIONALITY REDUCTION OF

M.I. Todorovska and T.-Y. Hao

This report explores the advantages of expansion in orthonormal wavelet bases—as a

2. SIGNAL PROCESSING PRELIMINARIES ........................................................................... 7

3. DISCRETE WAVELET TRANSFORM AND EXPANSION IN WAVELET BASES ....... 19

6. ANALYSIS OF COMPRESSION ERROR FOR A MINI DATABASE ............................ 150

7. DISCUSSION, SUMMARY AND CONCLUSIONS.......................................................... 162

REFERENCES ........................................................................................................................... 167

1.3 General Introduction

1.2 Motivation for and Objectives of this Work

1. A wavelet-transformed database contains already preprocessed data, enabling an

The objectives of this report are to:

1.3 Organization of this Report

This report is organized as follows. Chapter 2 presents a review of some elementary

2.1 Inner Product and Norm

Let f ( t ) , t ∈ be a real or complex valued function of a Hilbert space of square

The inner product of two functions f ( t ) and g ( t ) is by definition

and the L2 norm of a function, ⋅ L2

Similarly, the inner product of two sequences f [ n] and g [ n ] is by definition

and the L2 norm of a sequence, consistent with this inner product is

2.2 Expansion in a Basis

which is known as the Generalized Parseval’s equality.

The Fourier Transform of a function f ( t ) , t ∈ is defined as

2.4 Discrete-Time Fourier Transform

The Fourier Transform pair for a sequence { f [ n]}n∈] is by definition

fˆ ( eiω ) = F ( z ) z =eiω (2.5.2)

2.6 Continuous-Time Fourier Series

This is the expansion of a periodic function f ( t ) , of period T , in the basis of complex

where the Fourier coefficients are

Periodic sequences can also be expanded in Fourier series of orthogonal complex

its Fourier series expansion and Fourier coefficients are

2.8 Discrete Fourier Transform

Table 2.8.1 Summary of Fourier Transforms

Transform Condition on f Characteristics of Dual

(Continuous time) Fourier transform (CTFT) continuous continuous Self-dual

Discrete time Fourier transform (DTFT) discrete continuous and CTFS

(Continuous time) Fourier series (CTFS) continuous and discrete DTFT

2.9 Sampling of Continuous-time Functions

Let f [ n] , n ∈ be a sampled version of f ( t ) , t ∈ at equally spaced intervals ∆t .

and the frequencies ω in rad/sample is related to the frequency Ω in rad/s by

of the continuous time signal, f ( t ) , where

There is another important consequence of the one-to-one correspondence between the

or in the frequency domain

yˆ ( eiω ) = xˆ ( eiω ) hˆ ( eiω ) (2.10.2)

where H is a matrix with entries

3.1 Wavelet Theory and Time-Scale Analysis

An important property of wavelets in a family is their localization, described by their central

Fig. 3.1.1 Time-frequency resolution for three wavelets in a family.

3.2 Wavelet Transforms

Ws (a, b) =< ψ a ,b , s > L2

The continuous wavelet transform is highly redundant, as a function defined on is

a = a0m , b = nb0 a0m , m, n ∈ , a0 ≠ 1, b0 ≠ 0 (3.2.8)

where is the set of integers, in which case the transform becomes

s (t ) = ∑∑ ψ m,n , s ψ m,n (t ) (3.2.10)

all signals s ∈ H there exist constants 0 < A ≤ B < ∞ such that

1, m = m ' and n = n '

in which case the reconstruction formula is

1, m = m ' and n = n '

and in which case the bi-orthogonal basis expansion is

Orthonormal expansions are convenient because orthogonality simplifies the analysis.

3.3 Multiresolution Analysis

The general framework for construction of wavelet bases—multiresolution analysis—was

V2 ⊂ V1 ⊂ V0 ⊂ V−1 ⊂ V−2 ⊂ (3.3.1)