UC San Diego

Scripps Institution of Oceanography Technical Report

Title
Improving earthquake source spectrum estimation using multitaper techniques

Permalink
https://escholarship.org/uc/item/6v67h185

Author
Prieto, German A

Publication Date
2007-05-04

eScholarship.org Powered by the California Digital Library

University of California
 UNIVERSITY OF CALIFORNIA, SAN DIEGO

Improving Earthquake Source Spectrum Estimation

using Multitaper Techniques

A dissertation submitted in partial satisfaction of the

requirements for the degree Doctor of Philosophy
in
Earth Sciences

by

Germán A. Prieto

Committee in charge:

Professor Frank L. Vernon, Co-Chair

Professor Peter M. Shearer Co-Chair
Professor Joel Conte
Professor Bruce Cornuelle
Professor Robert L. Parker
Professor David J. Thomson

2007

c 2007
Germán A. Prieto,
All rights reserved.
The dissertation of Germán A. Prieto is approved, and it is
acceptable in quality and form for publication on microfilm:

Co-Chair

Co-Chair

University of California, San Diego

2007

iii
To my beloved wife and best friend, Carolina

iv
 TABLE OF CONTENTS

Signature Page . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

Table of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v

List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ix

Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x

Vita, Publications, and Fields of Study . . . . . . . . . . . . . . . . . . . . . . xii

Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiii

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1. Earthquake physics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1. Static and dynamic earthquake parameters . . . . . . . . . . . . . . . . 2
2. Scaling of earthquakes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
3. How do we estimate source parameters? . . . . . . . . . . . . . . . . . . . 11
4. Earthquake source parameters . . . . . . . . . . . . . . . . . . . . . . . . . 14
5. Spectrum estimation of seismic signals . . . . . . . . . . . . . . . . . . . . 15
6. Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2 Confidence intervals for earthquake source parameters . . . . . . . . . . . . . 21

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2. The Jackknife Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
1. Jackknife in Regression Problems . . . . . . . . . . . . . . . . . . . . . 25
3. Multitaper Spectrum estimates . . . . . . . . . . . . . . . . . . . . . . . . 26
4. Source parameter jackknife . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
1. Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2. Confidence intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3. Seismic Moment, Source Radius and Stress Drop . . . . . . . . . . . . . 33
5. Application to Cajon Pass Data . . . . . . . . . . . . . . . . . . . . . . . . 34
1. Extension to multiple stations . . . . . . . . . . . . . . . . . . . . . . . 38
6. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3 Earthquake source scaling and self-similarity estimation from stacking P and

S spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2. Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3. Implications of Self-Similarity . . . . . . . . . . . . . . . . . . . . . . . . . 51
4. Source Parameter Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . 54
5. Calibration to absolute moment and energy . . . . . . . . . . . . . . . . . 56

v
 6. Results for corner frequency and apparent stress . . . . . . . . . . . . . . . 59
7. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

4 Uncertainties in earthquake source spectrum estimation using empirical Green

functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
2. Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3. The Combined Empirical Green Function . . . . . . . . . . . . . . . . . . . 71
4. Results for radiated seismic energy . . . . . . . . . . . . . . . . . . . . . . 75
5. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

5 Quadratic Multitaper Spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 79

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2. Spectrum Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
3. Multitaper Spectrum estimates . . . . . . . . . . . . . . . . . . . . . . . . 84
1. Properties of Slepian sequences and functions . . . . . . . . . . . . . . 86
2. The multitaper algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 88
4. Estimating the derivatives of the spectrum . . . . . . . . . . . . . . . . . . 89
5. Quadratic Multitaper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
1. Random signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2. Periodic Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
3. Resolution test and the choice of multitaper parameters . . . . . . . . . 98
4. Synthetic earthquake signal . . . . . . . . . . . . . . . . . . . . . . . . . 102
5. Bathymetry profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
7. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
1. Main Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
2. Future Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . 112

A Propagation of errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114

B Quadratic mean-square error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

vi
 LIST OF FIGURES

Figure 1.1: Relation between near-field and far-field displacement and veloc-
ity, and corresponding far-field amplitude spectra . . . . . . . . . 4
Figure 1.2: Variation of stress at a point on a fault as a function of slip based
on the slip-weakening model . . . . . . . . . . . . . . . . . . . . . 6
Figure 1.3: Comparison of static and dynamic source parameters . . . . . . . 8
Figure 1.4: Models of earthquake rupture for large and small earthquakes . . 10
Figure 1.5: Illustration of the relation between the spectrum of the earthquake
source and static and dynamic parameters . . . . . . . . . . . . . 15
Figure 1.6: The scaling of the energy-seismic moment ratio and apparent stress 18

Figure 2.1: Map of Cajon Pass study area . . . . . . . . . . . . . . . . . . . . 24

Figure 2.2: Slepian tapers and Slepian functions . . . . . . . . . . . . . . . . . 29
Figure 2.3: Example seismogram of ML 3.5 earthquake recorded at the 2.9 km
borehole station . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Figure 2.4: Delete-one spectrum of ML 3.5 earthquake . . . . . . . . . . . . . . 36
Figure 2.5: Spectral modeling of P waves for two different sized earthquakes . 37
Figure 2.6: Source parameters and 5-95% confidence intervals . . . . . . . . . 39

Figure 3.1: Map showing the cluster of over 400 earthquakes and Anza stations 45
Figure 3.2: Example of computed spectra from the largest magnitude earth-
quake in the study area (ML = 3.4) . . . . . . . . . . . . . . . . . 46
Figure 3.3: Cartoon explaining how spectral stacking is used to obtain the
earthquake term, as in Warren and Shearer (2002) . . . . . . . . . 48
Figure 3.4: Examples of path-station terms for P and S waves . . . . . . . . . 49
Figure 3.5: Relative source spectral shapes for some selected bins . . . . . . . 50
Figure 3.6: An illustration of the effects of self-similarity when an earthquake
is increased in size . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Figure 3.7: EGF corrected stacked spectra for bins of different source moment,
showing the self-similarity of the spectra when shifted along an
ω −3 line. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Figure 3.8: EGF corrected stacked spectra and best-fitting source models for
P and S-waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Figure 3.9: A comparison between ML as measured by the southern California
Seismic Network (SCSN) and ML as estimated from our relative
moment measures using an empirical scaling factor. . . . . . . . . 58
Figure 3.10: A comparison between P and S-wave corner frequencies . . . . . . 61
Figure 3.11: Radiated S-wave energy EsS versus P -wave energy EsP . . . . . . . 61
Figure 3.12: P and S corner frequencies versus moment . . . . . . . . . . . . . 62
Figure 3.13: Apparent stress σa versus moment . . . . . . . . . . . . . . . . . . 63

Figure 4.1: Maps of the study area around the M 5.1 Anza earthquake . . . . 69
Figure 4.2: Estimates of the spectrum for ground motion of the M 5.1 earth-
quake and largest aftershock M 2.9 at station AGA and corre-
sponding spectral ratio . . . . . . . . . . . . . . . . . . . . . . . . 71

vii
Figure 4.3: Selected bins of spectral ratios at station AGA . . . . . . . . . . . 72
Figure 4.4: Mean source spectrum over 8 stations with 95% confidence inter-
vals for the M5.1 target earthquake . . . . . . . . . . . . . . . . . 76

Figure 5.1: Selected Slepian sequences and corresponding Slepian functions . . 86

Figure 5.2: Orthonormal version of the Slepian functions in Figure 5.1, Vk (f )
in the inner domain . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Figure 5.3: Comparison between basis functions used in Thomson (1990) and
Chebyshev polynomials used in this study . . . . . . . . . . . . . . 91
Figure 5.4: Comparison between the different basis matrices for zero, first,
and second order coefficients . . . . . . . . . . . . . . . . . . . . . 92
Figure 5.5: Spectrum estimation of pseudo-random number signal . . . . . . . 95
Figure 5.6: Spectrum estimation for a high dynamic range signal with two
periodic components at 0.05Hz and 0.3Hz. . . . . . . . . . . . . . 96
Figure 5.7: Spectrum estimation of a normally distributed random signal with
two sinusoids at 0.05 Hz and 0.3 Hz . . . . . . . . . . . . . . . . . 98
Figure 5.8: Resolution test comparison between Thomson’s and the Quadratic
multitaper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Figure 5.9: Spectrum analysis of synthetic earthquake model . . . . . . . . . . 103
Figure 5.10: Mean derivative of the spectrum from 100 random realizations of
two different earthquake models . . . . . . . . . . . . . . . . . . . 103
Figure 5.11: Location of the study area, where bethymetry profiles were taken
from . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Figure 5.12: Spectral Analysis of 5 selected profiles of bathymetric data . . . . 106

viii
 LIST OF TABLES

Table 2.1: Hypocentral parameters for the earthquakes recorded at Cajon

Pass used in this study. . . . . . . . . . . . . . . . . . . . . . . . . 24
Table 2.2: Source parameters and confidence intervals obtained by spectral
fitting and jackknife analysis. . . . . . . . . . . . . . . . . . . . . . 39

Table 5.1: Essential notation and mathematical symbols used in this chapter. 81
Table 5.2: Comparison of smoothness for the multitaper methods . . . . . . 95
Table 5.3: Comparison between multitaper methods by their 3-dB and 9-dB
bandwidths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100

ix
 ACKNOWLEDGMENTS

I would like to start by thanking Carolina for her constant support throughout
this process. Her love and friendship have been fundamental and have helped me become
a better person. Even though nobody believes it, she is the one who has to listen to
me talking when we are by ourselves. Thank you for being such a patient editor. I will
always love you.
I thank my two advisors, Frank Vernon and Peter Shearer. They were always
supportive, guided me through my research and my life as a graduate student. Discus-
sions included seismology and time series analysis, but also my future in academia, what
good science is, tennis, etc. They gave me enough independence to develop my research
which helped me in becoming a better scientist. Thank you for being there to answer
all my questions.
I was very lucky to have the opportunity to interact with Robert L. Parker and
David J. Thomson and to have them as part of my committee. They are among the best
scientists in inverse theory and data analysis and I have learned a lot from each one of
them. Given their different ways to solve problems and points of view, discussion with
them gave me a very wide view of problem solving techniques, something I could never
have had anywhere else. Joel Conte and Bruce Cornuelle were also very supportive of my
work and provided significant comments, ideas and points of view on where the project
could be improved.
I would also like to thank the large number of scientists with whom I had the
honor to talk to and even get their reviews for some of my papers. Special thanks to
Greg Beroza, Kevin Mayeda and Bill Walters, Rachel Abercrombie, Ralph Archuleta,
Jim Brune, Hiroo Kanamori and Luis Rivera.
For two quarters, I was a teaching assistant for Gabi Laske’s Natural Disasters
class. Her class allowed me to experience teaching and gave me the assurance to continue
in academia.
I am grateful to my family. My parents showed me love for science and education
and were always supportive of my decisions. My sister Helena and my brother-in-law
helped us start a life in San Diego and have been great family and friends.
Special thanks to Debi Kilb for her trust and for triggering my interest in

x
earthquake source physics with her willingness to discuss crustal seismology. I also
thank my colleagues and classmates at IGPP for lunch hours, coffee and chat time, and
our good friend, Mónica Pachón for all her energy.
Some of the chapters are reformatted version of papers that appeared in the geo-
physical literature. Chapter 2 (G.A. Prieto, D.J. Thomson, F.L. Vernon, P.M. Shearer
and R.L. Parker., 2007, Confidence intervals of earthquake source parameters., Geophys.
J. Int., 168, p. 1227–1234., doi:10.1111/j.1365-246X.2006.03257.x.) is reprinted by per-
mission of the Geophysical Journal International editorial staff and Blackwell Publishing.
Chapters 3 (G.A. Prieto, P.M. Shearer, F.L. Vernon, and D. Kilb., 2004, Earthquake
source scaling and self-similarity estimation from stacking P and S spectra., J. Geophys.
Res., 109, B08310, doi:10.1029/2004JB003084) and Chapter 4 (G.A. Prieto, R.L. Parker,
F.L. Vernon, P.M. Shearer and D.J. Thomson., 2006, Uncertainties in earthquake source
spectrum estimation using empirical Green functions., In Earthquakes: Radiated Energy
and the Physics of Faulting. Abercrombie, McGarr, Kanamori, and di Toro eds. AGU
Geophys. Monograph 170. pp 69–74) are reprinted with permission of the American
Geophysical Union. In all cases I was senior author with guidance provided by the
co-authors.
Financial support during my years at UCSD was provided by the NSF Grant
number EAR0417983. I was also partly supported by a Lawrence Livermore Mini-Grant.

xi
 VITA
1979 Born, Jülich, Germany
2002 B.S. Geosciences, Universidad Nacional de Colombia
2004 M.S., University of California, San Diego
2005–2006 Teaching Assistant, Department of Earth Sciences
University of California, San Diego
2002–2007 Research Assistant, University of California, San Diego
2007 PhD., University of California, San Diego

PUBLICATIONS
G. A. Prieto, D. J. Thomson, F. L. Vernon, P. M. Shearer and R. L. Parker. (2006)
Confidence intervals of earthquake source parameters. Geophys. J. Int., 168, 1227–1234,
doi:10.1111/j.1365-246X.2006.03257.x
G. A. Prieto, R. L. Parker, F. L. Vernon, P. M. Shearer and D. J. Thomson. (2006)
Uncertainties in earthquake source spectrum estimation using empirical Green functions.
In Earthquakes: Radiated Energy and the Physics of Faulting., Abercrombie, McGarr,
Kanamori, and di Toro eds. AGU Geophys. Monograph 170. pp 69–74
P. M. Shearer, G. A. Prieto, E. Hauksson. (2006) Comprehensive Analysis of Earth-
quake Source Spectra in Southern California. J. Geophys. Res. 111, B06303, doi:
10.1029/2005JB003979.
G. A. Prieto, F. L. Vernon, T. G. Masters, and D. J. Thomson. (2005) Multitaper
Wigner-Ville Spectrum for Detecting Dispersive Signals from Earthquake Records. Pro-
ceedings of the Thirty-Ninth Asilomar Conference on Signals, Systems, and Computers,
938–941, Pacific Grove, CA.
G. A. Prieto, P. M. Shearer, F. L. Vernon, and D. Kilb. (2004) Earthquake source
scaling and self-similarity estimation from stacking P and S spectra. J. Geophys. Res.,
109, B08310, doi:10.1029/2004JB003084.

FIELDS OF STUDY
Major Field: Seismology
Studies in Earthquake Source Physics.
Professors Frank L. Vernon and Peter M. Shearer
Studies in Earthquake Parameter Uncertainties.
Professors Frank L. Vernon, Peter M. Shearer, Robert L. Parker and David J.
Thomson
Studies in Time Series Analysis.
Professors David J. Thomson and Robert L. Parker

xii
 ABSTRACT OF THE DISSERTATION

Improving Earthquake Source Spectrum Estimation

using Multitaper Techniques

by

Germán A. Prieto
Doctor of Philosophy in Earth Sciences
University of California, San Diego, 2007
Professor Frank L. Vernon and
Professor Peter M. Shearer, Chairs

Understanding the physics of the earthquake rupture mechanism is essential, given that
earthquakes are among the most harmful natural disasters. Some earthquake source
parameters such as radiated seismic energy and stress drop can be used to investigate
the properties and dynamics of faulting. Estimates of these parameters have large un-
certainties, leading to discrepancies among different studies, particularly investigations
of the scaling relations of earthquakes.
In order to understand the physics of earthquakes and their behavior as a
function of magnitude, it is necessary to have an idea of the uncertainties of the estimated
parameters (e.g., when comparing two earthquakes). We have developed a method to
estimate the uncertainties of the source parameters as measured from the seismic wave
spectra. The large uncertainties expected require improving the methodologies used to
obtain the source parameters. We present two methods that take advantage of the large
amounts of seismic data available.
In the first method we attempt to separate the effects of anelastic attenuation
from the earthquake source spectrum characteristics. Analyzing the latter we are able
to obtain source parameters with significantly reduced scatter and which indicate that
the earthquake rupture is self-similar in the magnitude range 1.8 to 3.4. In the second
method we perform a weighted average of spectral ratios using 160 small earthquakes
as empirical Green functions to obtain estimates of the source spectrum of the 2001
M5.1 Anza earthquake. The averaging scheme significantly reduces the uncertainties

xiii
and allows us to estimate the radiated seismic energy for this earthquake with greater
confidence than is otherwise possible.
Given that in the methods discussed above the seismic parameters were esti-
mated from the spectrum of the seismic waves, we present a new multitaper algorithm
that has significant bias reduction compared to standard multitaper techniques and at
the same time reducing the roughness of the estimated spectrum. We show that the
method has the ability to estimate both the spectrum and its slope, thus increasing the
degrees of freedom if parameters are to be estimated.

xiv
1

Introduction

Every day there are an immense number of earthquakes occurring somewhere

on Earth. Some of these events may be strong enough to be felt, and some, such as
the recent Sumatra earthquake in December 2004 in combination with its associated
tsunami, generate considerable losses in both infrastructure and human life.
Given that earthquakes are among the most harmful and costly natural disas-
ters, it is essential to have an understanding of the physical processes that lead to their
occurrence as well as a deep comprehension of the actual rupture process. Seismology
– the study of the structure of the Earth and the physics of earthquakes – has thus a
substantial role in mitigating the damaging effects of large earthquakes on our society.
Earthquakes can rupture along just a few meters or along hundreds of kilo-
meters. This study focuses on understanding earthquake ruptures, and to what extent
small earthquakes (such as the commonly occurring minor earthquakes in seismically ac-
tive regions) and large earthquakes (such as the Sumatra earthquake) are generated by
similar physical processes or if they are fundamentally different. The observable features
of earthquake rupture need to be quantified for comparison between these events, and
I investigate how novel time-series analysis tools and inverse problem solving methods
can be brought to answer relevant questions in this field.

1
 2

1.1 Earthquake physics

Earth’s tectonic plates slide past each other, in some cases being accommodated
by gradual sliding, in other cases by earthquake rupture that accommodates this motion
by sudden slip on a fault plane. Seismologists try to understand this sudden behavior
by looking at quantifiable features that can be extracted from records of the radiated
elastic waves at seismic stations on or near the surface of the Earth.
An earthquake is a failure in Earth’s crust. Due to plate motion there is a
certain amount of potential energy (gravitational and strain energy) available within
a certain region S. We may assume (for a short-term process) that the accumulated
strain energy is released in the region S by the earthquake rupture. During the failure
process, some energy is radiated as seismic waves (radiated energy ES ) and some energy
is dissipated mechanically (fracture energy EG ) and thermally (thermal or frictional
heating energy EF ). I will discuss and describe these and other terms in the following
section.

1.1.1 Static and dynamic earthquake parameters

In order to understand the physics of earthquakes, it is important to quantify

the behavior of some seismic parameters that describe the earthquake rupture process.
I will discuss both static as well as dynamic source parameters.

Seismic Moment

Consider a point source in which a displacement offset D between the two

sides of the fault occurs. It can be shown that a double-couple force can produce a
displacement field equivalent to a point dislocation. The scalar seismic moment M0 of
such a double-couple source is given by (e.g., Kanamori and Anderson, 1975; Shearer,
1999):
M0 = µDA (1.1)

where µ is the shear modulus of the material surrounding the fault, D is the fault
displacement, and A is the rupture area. The dimensions of M0 are force × length =
energy, and usually the unit Nm is used.
 3

Since the rupture in the solid Earth is irreversible, the displacement that occurs
between the two sides of the fault is permanent. This displacement also occurs over some
finite duration, leading to a ramp-like near-field (very close to the fault) displacement as
a function of time.
The far-field displacement, on the other hand, is not permanent and is propor-
tional to the time derivative of the near-field displacement (see Figure 1.1) . Assuming
a seismic station is in the far-field (and there is no attenuation or scattering) the scalar
seismic moment could be represented by the area under the displacement pulse (e.g.,
Madariaga, 1976; Shearer, 1999):
M0 = µD̄A (1.2)

where D̄ is the average displacement across the fault.

Seismic moment is believed to be the most useful and easily measured quantifi-
cation of the size of an earthquake. Unlike other magnitude estimates (local magnitude
or surface wave magnitude), M0 does not saturate for large earthquakes. However, as
discussed above, M0 is a static measure of the size of an earthquake and does not provide
any information on the dynamic properties of the source.

Stress Drop

As suggested above, a certain amount of stress is released by an earthquake rup-

ture. The stress drop is defined as the average difference between the stress (we usually
consider shear stress) on a fault before an earthquake and the stress after the earth-
quake (Kanamori and Anderson, 1975; Kanamori and Brodsky, 2004; Shearer, 1999).
For a point source the stress drop is:

∆σ = σ0 − σ1 (1.3)

where σ0 and σ1 represent the stress on the point before and after the earthquake,
respectively. Since the stress drop can actually be highly variable in certain regions of
the fault plane, we prefer the stress drop averaged over the entire fault plane (Kanamori
and Anderson, 1975; Shearer, 1999):
Z
1
∆σ = [σ0 − σ1 ] dS (1.4)
A
S
 4

Near-field Far-field Far-field

time domain time domain spectral domain

fc
M0
S(f)

D FFT

log amplidtude
displacement displacement displacement

amplidtude
amplidtude

|2πfS(f)|
FFT velocity
velocity
velocity

time time log frequency

Figure 1.1: Relation between near-field and far-field displacement and velocity, and
corresponding far-field amplitude spectra. The seismic moment M0 is proportional to the
shaded area under the far-field displacement curve. The two plots on the right represent
the amplitude spectrum. Another source parameter seen in the near-field signal is the
displacement D. The pulse duration τ and the corner frequency fc are also represented
in time and frequency domains respectively.

where the integral is performed over the surface of the fault.

The limited resolution in seismological methods does not allow determining the
displacement D everywhere on the fault, and forces us to use the approximation

D̄
∆σ = Cµ (1.5)
L̃

where µ is the shear modulus and comes from the relation between stress and strain
(σ = 2µ), L̃ is a characteristic rupture dimension, and C is a non-dimensional constant
that depends on the geometry of the fault plane.
For the particular case of a circular fault of radius r (Eshelby, 1957; Brune,
1970; Madariaga, 1976), it can be shown (plug 1.2 into 1.5) that the stress drop is
related to the seismic moment M0 by:

7M0
∆σ = (1.6)
16r3

Note that the stress drop is inversely proportional to the cube of the fault dimension, and
thus any uncertainty in the radius of the fault will propagate into a large uncertainty in
the stress drop. Once again, the stress drop is a static parameter of the seismic source.
 5

Corner frequency and source dimension

As discussed above, the stress drop can be obtained if we know the seismic
moment (e.g., from the area under the displacement pulse) and the source dimension; in
this case, assuming a circular fault, the source radius. But the source radius r cannot
be measured directly from seismological data and further assumptions are needed.
As suggested by Madariaga (1976), assume a circular fault with radius r with
a rupture starting in the center and moving radially outside with rupture velocity vr =
0.9β, where β is the S-wave speed. The rupture duration time is then
r
τ= (1.7)
vr
where τ can effectively be extracted from the pulse duration in the displacement record
(Figure 1.1).
One can also study the rupture duration by looking at the spectra of the dis-
placement records. Figure 1.1 plots the spectrum of the far-field displacement. Note that
the spectrum remains constant until it reaches what is known as the corner frequency
fc and then the amplitudes decrease rapidly. The corner frequency is clearly related to
the pulse duration (τ ∝ fc−1 ), and once plugged into Equation 1.7 can also be related to
the radius of the fault
kβ
r= (1.8)
fc
where k is a nondimensional factor (0.32 for P and 0.21 for S-wave, Brune, 1970;
Madariaga, 1976), depending on whether we are measuring P or S-wave corner fre-
quencies.

Seismic Energy

When an earthquake occurs, some fraction of the total energy is radiated as

seismic waves, while the rest is released as thermal and fracture energies, which together
represent the energy dissipation. I now discuss the energy budget involved in the rupture
process in order to provide a general understanding of the observational and the physical
processes involved. I will assume that the slip weakening model (Ida, 1972; Palmer and
Rice, 1973) is valid. An illustration of the energies related to fractures is shown in Figure
1.2.
 6

σp

σ0

Δσ

Stress
σf
EG ES
σ1

EF

0 Slip D

Figure 1.2: Variation of stress at a point on a fault as a function of slip based on the
slip-weakening model. This model explains the partition of energy during rupture and
the relation between radiated energy ES and stress drop ∆σ. The frictional stress σf (s)
is shown as the thick curve. Note that this figure represents a unit fault, the stress
behavior might be different in various regions on the fault plane.

As explained by Kanamori and Rivera (2006), in an expanding crack the initial

stress σ0 increases to the peak stress σp (also known as yield stress) at the onset of
rupture (see Figure 1.2) and then drops following the curve σf (s) as a function of slip.
The behavior of this curve σf (s), the frictional stress, shows the particular state of
stress during the rupture process. The actual variation σf (s) may be very complex and
seismological estimates are most likely smoothed versions of the real behavior.
The energy that is dissipated (EG + EF ) is represented by the area under σf .
This energy includes the energy used in creating new crack surface, energy breaking the
surrounding rock (creating fault gauge, etc) and energy released as heat, due to friction
between the two sides of the fault sliding past each other. Other types of energies
(such as latent heat) may be present if, for example, phase transitions due to heating are
present (Tinti et al., 2005). Even though Figure 1.2 shows a clear separation between the
fracture energy EG and the thermal energy EF , this is not always clear and is dependent
on many assumptions (Abercrombie and Rice, 2005).
The seismically radiated energy is given by the difference between the entire
colored area and the dissipated energy. We can only measure directly from seismograms
 7

the energy that is radiated as seismic waves ES .

Figure 1.2 also shows the relation between the radiated seismic energy and the
stress drop. As explained before, the stress drop ∆σ is the difference between the initial
stress on the fault σ0 and the final stress σ1 . For example, assume that the stress drops
instantaneously to σ1 , which would mean there is no fracture energy dissipation and
much more energy is radiated seismically. On the other hand, if stress drops quasi-
statically (e.g., creep, slow earthquakes) there is minimal or no seismic wave radiation.
An intermediate case is the one shown in the example of Figure 1.2.
As suggested by many studies (Mayeda and Walter, 1996; Kanamori and Rivera,
2004; Venkataraman et al., 2006, and many others) the radiated seismic energy is a
dynamic measure of the size of an earthquake. Note that both seismic moment M0 and
stress drop ∆σ are static source parameters, while ES is a dynamic one. The seismic
moment depends on the area of the fault rupture and the average displacement, the stress
drop is a function of the difference between initial and final stress states (see equation
1.2 and 1.3), while the seismic energy is a function of the behavior of the stress (the
function σf ) during earthquake rupture.
While the stress drop and seismic moment remain fixed, the radiated seismic
energy ES may be extremely variable if, for example, σf decreases below the final stress
σ1 and then increases back before rupture stops. Figure 1.3 shows two cases where
the stress drop remains constant while the radiated seismic energy varies, depending on
the rupture history and not only on the initial and final states. Similarly (Kanamori
and Rivera, 2004), we can have a far-field displacement pulse from which we obtain the
seismic moment and may estimate the radiated seismic energy by differentiating. One
can think of a displacement pulse with a very complicated structure, but still keeping
the area under the curve fixed, which would have a very large ES . Again, the seismic
energy is related to the rupture characteristics and is a dynamic parameter.

1.2 Scaling of earthquakes

These are some of the most relevant source parameters used to describe earth-
quake properties. Now, let us consider how they relate to each other, given that these
 8

σ0

Δσ

Stress
σf
EG ES EG ES
σ1

EF EF σf

0 Slip D 0 Slip D

Figure 1.3: Comparison of static and dynamic source parameters. For simplicity I
assume peak stress being equal to the initial stress (σp = σ0 ). Note that while the
stress drop only depends on initial and final stresses, the radiated seismic energy ES is a
function of the frictional stress σf (s) (thick black curve) throughout the rupture process.

relations may be useful in constraining the processes involved in or around the fault.
In addition, I will discuss the behavior of these parameters as a function of
the earthquake size. In other words, is the physics of the faulting mechanism associated
with an M8.0 earthquake different from that of an M2.0 earthquake? The differences (or
lack thereof) could potentially provide means for rapid determination of the size of an
earthquake to use in early warning systems (Kanamori, 2005). Aki (1967) suggested a
scale invariance of the rupture process, consistent with observations that many geological
processes are similar over a wide range of scales (Abercrombie, 1995).

Seismic moment and corner frequency

The relation between the static parameters seismic moment M0 and the length
scale or characteristic rupture dimension L̃ (see equation 1.5) has been widely used in the
literature. As seen from equation (1.8), under certain assumptions the corner frequency
fc or the source duration τ can be used as a proxy for the source dimension.
From compilations of a variety of studies (Brune, 1970; Abercrombie and Leary,
1993; Prieto et al., 2004), a very common scaling relation is M0 ∝ fc−3 . This would
suggest that if this scaling holds (assuming a constant rupture velocity (see Kanamori
and Rivera, 2004)), the scaling of stress drop with seismic moment would be M0 ∝ ∆σ,
 9

meaning that ∆σ is independent of earthquake size. Other studies suggest (Abercrombie,

1995; Kanamori and Rivera, 2004; Izutani and Kanamori, 2001) that in fact the scaling
−(3+)
should be M0 ∝ fc , where  6 1 and ∆σ could be scale dependent.

Seismic moment and radiated energy

The radiated energy is of considerable interest because it has relevant informa-

tion about the dynamics of rupture during an earthquake, and it can be measured with
seismological methods.
A very useful dynamic parameter associated with the radiated seismic energy
ES is the apparent stress, as introduced by Wyss and Brune (1968):

ES
σa = µ (1.9)
M0

which describes the dynamic properties of an earthquake. Replacing the seismic moment
M0 (Equation 1.2) we have
ES
σa = (1.10)
D̄A
and can be interpreted as the radiated seismic energy per unit area per unit displacement.
The behavior of apparent stress as a function of earthquake magnitude, the
scaling of σa ∝ M0 , is of key importance. Does the earthquake apparent stress change
with magnitude? Does the seismic energy ES (seismic waves that leave the source region)
increase proportionally as the seismic moment M0 (the fault area A and the slip D̄)
increases?
In a very simple way, in Figure 1.4 imagine a unit fault for a small and a large
earthquake (the large earthquake would be composed of a large number of these unit
faults) and their corresponding stress behavior as a function of slip. Clearly, the stress
and slip behavior is not necessarily uniform throughout the fault plane, and we must
deal with the averages during rupture for each event. The focus here is to investigate
how the source parameters (frictional stress, seismic energy, stress drop, etc.) vary over
a wide range of earthquake magnitudes, represented by the slip D̄. The slip can range
from millimeters for small earthquakes to meters for very large ones, and here we take
the averages for individual earthquakes as representative of the relation between stress
and slip.
 10

Small Large
a
σ0

Δσ
Stress

EG ES EG ES
σ1

EF EF

0 Slip D 0 Slip D

b
σ0

Δσ
Stress

EG ES EG ES
σ1

EF EF

0 Slip D 0 Slip D

c
σ0

Δσ
Stress

Δσ
EG ES EG ES
σ1

EF EF

0 Slip D 0 Slip D

Figure 1.4: Models of earthquake rupture for large and small earthquakes. Peak stress
σp is neglected for simplicity and the frictional stress σf (s) is represented by the thick
curves. Due to rupture, the initial stress state σ0 falls to a final value σ1 as the slip
increases to the value D. a) In the self-similar model, ES increases proportionally as a
function of slip D; b) the fracture energy varies with increasing slip D, ES does not scale
proportionally as a function of D and thus apparent stress σa will be scale-dependent; c)
the fault lubrication model of Kanamori and Heaton (2000) in which friction decreases
as in the self-similar model, but after a certain slip it decreases even further, generating
a larger stress drop and radiating more energy ES .
 11

Figure 1.4 shows a small set of models that can be used to describe the earth-
quake rupture process. The self-similar model (Aki, 1967; Prieto et al., 2004) in Figure
1.4a assumes that the fracture energy is constant (at least for a particular region) gen-
erating an ES that scales proportionally to the final slip D. In this case, both the stress
drop ∆σ and the apparent stress σa will be constant as a function of magnitude.
If the fracture energy scales with size (Figure 1.4b) and is scale dependent, then
the σa will not be constant. In this case, the stress drop ∆σ is scale independent, and
the size of the earthquake is related to the fracture energy, that is, the fracture energy is
a property of the fault zone and the ultimate size of the earthquake is in part governed
by the fault zone properties.
A final case is shown in Figure 1.4c, which is known as the fault lubrication
model (Kanamori and Heaton, 2000). In this model the rupture process may behave
like the self-similar model, until it reaches a certain amount of slip, at which point an
additional drop in the frictional stress occurs. This would clearly suggest very different
physical processes during rupture between large and small earthquakes. The additional
weakening mechanism has been explained in various ways, thermal weakening by pore
fluids (Lachenbruch, 1980), elastohydrodynamic lubrication (Brodsky and Kanamori,
2001), and normal stress variations and interface separation during slip (Brune et al.,
1993).
If the source parameters ∆σ and σa are constant within the uncertainties over
a wide magnitude range, the model in Figure 1.4a is possible, and the rupture process
should be considered self-similar. If, on the other hand either or both ∆σ and σa are
scale dependent, then other models must be considered. As in Figure 1.4b, the properties
of the fault zone might be controlling the rupture, or as in Figure 1.4c, it might be that
the rupture for very large earthquakes (large slip D) follows very different physics than
the smaller earthquakes do.

1.3 How do we estimate source parameters?

So far, we have discussed some of the source parameters used for describing
earthquake rupture. What is recorded at a seismic station (using mainly velocity sensor
 12

and/or accelerometers) is the ground motion associated with the earthquake rupture and
the radiated seismic energy.
Seismic waves travel inside the Earth’s crust and mantle and are affected by
attenuation, velocity and density variations, near surface effects and scattering until they
finally arrive at the seismic station. These disturbances of the original radiated waves
need to be accounted for. In this section I will discuss some of the methods used in the
literature (some used in this thesis) to correct for the propagation effects in order to be
able to investigate the source parameters.
There are also source effects that may be present and need to be corrected or
accounted for to obtain a reliable estimate. These include the radiation pattern and the
directivity effects.
The radiation pattern may alter the amplitudes of the far-field signals recorded,
depending on the azimuthal direction between the earthquake source and the receiver.
The directivity effect is a result of the Doppler shift or Doppler effect, where high-
frequencies are expected to be radiated in the direction of rupture while lower frequencies
will be present in the opposite direction. It has also been argued that the directivity effect
focuses the energy radiation in the direction of rupture (Venkataraman and Kanamori,
2004). A more complete discussion about the Doppler effect and directivity in seismic
sources can be found in Douglas et al. (1988). It is essential to have a good azimuthal
coverage of stations to properly take into account these effects.
Correcting waveforms or their spectra for propagation effects is one of the chal-
lenging aspects in determining the earthquake source properties. Different methods used
for correction of the propagation effects can introduce significant variability in the cal-
culated source parameters as pointed out by Sonley and Abercrombie (2006), even when
applied to the same data (e.g., Prejean and Ellsworth, 2001; Ide et al., 2003).
Many researchers use the amplitude spectra of the seismic waves (Abercrombie,
1995; Ide et al., 2003; Prieto et al., 2004) to calculate radiated seismic energy, stress drop
and other source parameters. A similar result should be obtained if working on the time-
domain signals (Kanamori et al., 1993; Mori et al., 2003), but corrections for attenuation
and deconvolutions are in many cases easier in the spectral domain. As we will discuss
in the next section, in the spectral domain there are state-of-the-art methods to analyze
 13

the frequency content of complicated signals and study their statistical reliability, which
is more difficult in the time domain.
From this point on, it is assumed that the instrument response has been cor-
rected from the recorded signals and that the effects of any incorrect instrument response
correction are negligible.

Attenuation correction

To obtain earthquake source parameters, there is a need to correct for attenua-

tion, which may vary from study to study. The basic idea is that the source spectrum is
attenuated by the anelastic crust through which it travels. It is customary in earthquake
physics to represent the attenuation with the inverse of the quality factor Q.
It is possible to represent the quality factor by a constant Q (Abercrombie,
1995; Prieto et al., 2006) or as a frequency-dependent function Q(f ) = Q0 f b , where
both Q0 and b are constants (Ide et al., 2003; Sonley and Abercrombie, 2006). The
calculation of the Q(f ) is not trivial and it is in many cases done simultaneously with
some source parameters, and some trade-offs between these parameters are inevitable.
Even in deep borehole stations it has been shown that a frequency dependent quality
factor may be necessary (Ide et al., 2003).

Empirical Green Functions

The Empirical Green Function (EGF) method (Mueller, 1985; Hartzell, 1978;
Hough, 1997) takes advantage of the records of a smaller earthquake that is collocated
with a larger one. We can assume that up to a certain frequency, the smaller earthquake
can be approximated by a point source in time and space. This means that the ground
motion recorded at a particular station for that earthquake is approximately the impulse
response of the path between the source and the receiver.
This impulse response is then deconvolved from the larger earthquake, in this
way accounting for attenuation, scattering, near-source and other effects. In a sense,
the EGF method provides a more accurate account of path effects than the attenuation
correction. Nevertheless, it also has shortcomings, namely that only events that are
collocated can potentially be used, the signal-to-noise ratio for the smaller earthquakes
 14

degrades at both low and high frequencies and that other characteristics, such as similar
focal mechanisms for the two earthquakes, are needed.
In principle, this deconvolution can be performed in the time domain (e.g., Mori
et al., 2003) or in the frequency domain (Ide et al., 2003; Abercrombie and Rice, 2005;
Prieto et al., 2004). As we will show in this thesis, the frequency domain methods used
provide an advantage in terms of uncertainty estimation.

1.4 Earthquake source parameters

In this thesis we will focus on the study of earthquake source parameters es-
timated from the spectrum of the seismic signals. Assuming all the propagation effects
have been accounted for, all the parameters can be estimated from the spectra of the
seismic waves.
The seismic moment M0 and the corner frequency fc can be fitted from the
displacement spectra using the Brune (1970) model:
M0
S(f ) = (1.11)
1 + (f /fc )2
or the Boatwright (1980) model:
M0
S(f ) = h i0.5 (1.12)
1 + (f /fc )4

As shown in Figure 1.5, the seismic moment M0 is proportional to the zero frequency
amplitude of the displacement spectrum, equivalent to the area under the displacement
pulse in the time domain (see Figure 1.1). The corner frequency is given by the strong
change in slope of the amplitude spectrum, which in the time domain is related to the
width of the source pulse, and can be thought as a measure of the rupture duration. In
general, some kind of non-linear fitting algorithm needs to be used.
The radiated seismic energy ES can also be obtained by converting (rotating)
the spectrum to velocity, squaring and integrating:
Z∞
ES = C1 [2πf S(f )]2 df (1.13)
0

where the constant C1 has additional parameters such as the wave speed and density of
the material surrounding the fault area to obtain the correct units.
 15

fc fc
S(f) |2πf S(f)|2

log amplidtude
M0
rotate to
velocity ES

log frequency log frequency

Figure 1.5: Illustration of the relation between the spectrum of the earthquake source
(after correcting for all propagation and other source effects) and static and dynamic
parameters. The seismic moment M0 and corner frequency fc can be related to the
displacement source spectrum (left). The energy is related to the area under the velocity
spectrum squared (left). See text for explanation.

1.5 Spectrum estimation of seismic signals

There are many applications in geophysics where relevant information contained

in a given signal may be extracted from the frequency content of the spectrum. In some
cases the scientist may be interested in periodic components usually immersed in some
background noise (e.g., normal mode seismology (Gilbert, 1970), climate time series
(Chappellaz et al., 1990), etc.), in a general continuous spectrum to be estimated from
a short time series (e.g., earthquake source spectra (Brune, 1970; Prieto et al., 2004))
or in comparing two signals and investigating where the similarities or differences are
(in seismology for example (Vernon, 1989; Hough and Field, 1996), transfer functions in
electromagnetism (Constable and Constable, 2004), elastic thickness of the lithosphere
(Daly et al., 2004), etc.). In each of these cases, it is desirable to be able to obtain a
reasonable spectrum with little or no bias and small uncertainties.
As suggested above, all source parameters we have discussed above can be
obtained by analyzing the spectrum of the seismic signals, after other effects (attenuation,
directivity, etc.) have been accounted for. But the spectrum estimation of seismic signals
poses many difficulties. First, all the seismic waves we are interested in are transient
and, in the case of small earthquakes, the seismic phases (e.g., body waves) are contained
in a short segment within the record. Second, the signals have very high dynamic range,
which might lead to severely biased estimates due to spectral leakage, where frequency
 16

information with high amplitudes (e.g., close to the corner frequency) leaks into frequency
regions with low amplitudes. A final difficulty is that the signal is non-stationary; that
is, the statistical character of the data changes with position in the record.
It is standard practice to use the discrete Fourier transform (DFT) to estimate
the spectrum of a particular series. However, simply using the DFT and squaring to
obtain the spectrum, in other words using the periodogram, is a poor choice and should
never be done (for discussion about the choice of the periodogram, see Harris, 1978;
Thomson, 1982, and references therein). In general, it is much better practice to window
the time series with a taper before performing the DFT and squaring to reduce spectral
leakage.
Conventional tapers (in the time domain) used for spectrum estimation have
a bell-shaped curve (sometimes with a flat top) and tend to zero at the edges. This
approach has a major limitation, in that by applying a taper we are effectively discarding
significant statistical information in a given time series. The data points at the edges of
the record are down-weighted, while the data in the center is emphasized, which causes
the variance of the spectral estimate to be greater than that of the periodogram.
In Thomson (1982) a different method, called the multitaper spectral analysis,
was introduced. As its name suggests, the idea of the method is to use multiple orthog-
onal tapers to window the time series and reduce spectral leakage, and by applying the
DFT and squaring, obtain almost independent estimates of the spectrum, called eigen-
spectra. As shown in Thomson (1982) and in many other studies (Vernon, 1989; Park
et al., 1987b; Riedel and Sidorenko, 1995) as long as only one taper is used, there will
be a trade-off between the resistance to spectral leakage and the variance of a spectral
estimate.
The tapers are constructed to be leakage-resistant, and sample the time series
in different ways. The information that is discarded by the first taper (whose shape is
very similar to the conventional tapers) is partially recovered by the second taper, and
the information down-weighted by these two tapers is recovered by the third, fourth, etc.
As explained by Park et al. (1987b), single-taper spectral estimates have rel-
atively large variance (increasing as a larger fraction of the data is discarded and the
bias of the estimate is reduced) and are inconsistent estimates (i.e., the variance of the
 17

estimate does not drop as one increases the number of data points). In the case of single
taper methods a smooth estimate can be achieved by applying a moving-average, thus
reducing the variance, while at the same time reducing the frequency resolution and
increasing the bias of the estimate.
In the multitaper algorithm only a few tapers are used to construct the spectral
estimate. A weighted sum of the eigenspectra is formed, leading to a smooth estimate of
the spectrum with variance reduced due to the averaging process. By using the multiple
tapers, the estimator is also consistent. It has been shown in multiple cases that the
multitaper algorithm outperforms single-taper smoothed spectral estimates (Park et al.,
1987b; Bronez, 1992; Riedel and Sidorenko, 1995).
In terms of non-stationarity, the multitaper estimation is also a better choice.
Because single taper estimates weight the data in the center of the signal more, the
information present at the end of the signal is not used, which may lead to the misrepre-
sentation of the spectrum. Multitaper estimates discard much less data and are sensitive
to information from almost the entire signal.
Another very important feature of multitaper spectral estimates is the possibil-
ity of obtaining measures of uncertainties and confidence intervals from the data. Given
that we have almost independent eigenspectra, it is possible to obtain error estimates of
the spectrum and associated parameters (Thomson and Chave, 1991).

1.6 Objectives

The primary goal of this thesis is to develop methods to obtain better, more
reliable estimates of the seismic source parameters used to investigate the physics of
earthquakes. One of the issues in the physics of earthquakes is the scaling of static and
dynamic parameters and what this tells us about the rupture process. As mentioned
earlier, all these parameters and their scaling can be determined from the spectrum of
the seismic waves, and I will focus on using and improving the state-of-the-art multitaper
spectrum algorithm to obtain the most reliable parameter estimations.
The basic problem in earthquake source scaling can be observed in Figure 1.6.
The figure shows a compilation of different studies on the radiated energy and apparent
 18

Magnitude MW
1 2 3 4 5 6 7 8

10

σa (MPa)
1

0.1

0.01

10-3
10 12 14 16 18 20 22
log M0 (Nm)
Kanamori et al. [1993] Mayeda and Walter [1996] Prieto et al. [2004]

Abercrombie [1995] Pérez-Campos and Beroza [2001] Prieto et al. [2006b]

Figure 1.6: The scaling controversy on energy-seismic moment ratio and apparent stress
as first compiled by Ide and Beroza (2001). Each symbol denotes a different data set as
shown. In red, I added results from studies in the Anza region from Chapter 3 and the
M5.1 Anza earthquake from Chapter 4 with 95% confidence bounds.

stress scaling, ranging from very small earthquakes (M0.0) to very large ones (M8.0)
(Abercrombie, 1995; Mayeda and Walter, 1996; Kanamori et al., 1993; Pérez-Campos
and Beroza, 2001; Prieto et al., 2004; Shearer et al., 2006). The question is whether the
dynamic parameter is constant over this wide range of magnitudes or if there is a change
in behavior as the magnitude increases. Depending on the data you focus on, there could
be a linear increasing trend, but with the large scatter of the data, it is also possible to
have a constant scaling of the dynamics of the earthquake rupture.
I have not organized this thesis by the temporal evolution of my research, but
rather by topic. The first question that may arise from looking at Figure 1.6 is what
points can be believed? What are the error bars of each of these points? Is the large
scatter seen in these data points real, meaning that the rupture process may have very
different dynamic behavior even if the magnitude of the earthquakes is similar, or is this
scatter due to large uncertainties in the individual estimates?
Data shown in Figure 1.6 come from different regions (although most of them
are in California), the signal processing methods are quite different (e.g., time domain
in Kanamori et al. (1993); multitaper spectral analysis in Abercrombie (1995); and coda
wave envelopes in Mayeda and Walter (1996)) and earthquakes include strike-slip, thrust
and normal faulting, making them very difficult to compare.
 19

Many of the studies shown in the figure and many others published in the
literature lack the analysis of the uncertainties associated with the methodologies used
in the analysis of the seismic parameters. In order to compare two or more estimates, it
is of key importance to have an understanding of the uncertainties and assumptions of
the methods used.
Since this, to me, represents a major shortcoming of the present literature, in
Chapter 2 we present a method to obtain confidence intervals on earthquake source
parameters using the multitaper algorithm from single-station measurements. We discuss
the use of the jackknife variance applied to source parameter estimation, and show that
large uncertainties are expected even in the ideal conditions of deep borehole records. An
extension to multiple station measurements is also discussed. Chapter 2 has appeared
in Geophysical Journal International under the title ”Confidence intervals for earthquake
source parameters” (Prieto et al., 2007a). I participated as primary author in all phases
of the development of this research, including the programing of the computer algorithms.
Given that large uncertainties are expected (even in ideal conditions), we pre-
sent in Chapter 3 and Chapter 4 two different methods to improve the estimation of
the source parameters by taking advantage of the large number of earthquakes that occur
in southern California using the Anza Seismic Network in Southern California. Chap-
ter 3 and Chapter 4 are reformatted versions of the papers ”Earthquake source scaling
and self-similarity estimation from stacking P and S spectra” (in Journal of Geophysical
Research (Prieto et al., 2004)) and ”Uncertainties in earthquake source spectrum esti-
mation using empirical Green functions” (appeared in the AGU Monograph on Radiated
Energy and the Physics of Earthquake Faulting (Prieto et al., 2006)) respectively. In
both cases I was senior author and under the supervision of my co-authors developed
the research and computer codes that form the basis of these chapters.
In Chapter 3, we study the self-similarity and scaling relations of a cluster of
400 earthquakes in the range M 0.5 to M 3.4 by iteratively stacking spectra of P and S
waves. The iterative approach is aimed at separating the propagation effects from the
source spectra and, hence, be able to study the scaling properties. As I discussed above,
we study the scaling by looking at source parameters (e.g., M0 , fc , ES , etc.) and their
relation, but also introduce a test on self-similarity, which is independent of the choice
 20

of parameterization of the earthquake spectra.

In Chapter 4, we investigate the problem of using Empirical Green functions
(EGF) to obtain the source spectrum and associated source parameters. We show that
the spectrum obtained via the EGF method has large uncertainties and introduce a
method to reduce the uncertainties by using multiple EGFs instead of a single one. The
method is applied to the 2001 M 5.1 Anza earthquake in southern California.
The multitaper algorithm presented by Thomson (1982) reduces spectral leak-
age effectively, but suffers from local bias. By performing the weighted averaging of the
different eigenspectra as proposed by Thomson (1982), it is assumed that the spectrum
is white within a certain frequency band. As shown by Riedel and Sidorenko (1995),
there is considerable bias present due to the curvature of the spectrum. In Chapter 5,
we present an extension to the multitaper, which reduces the curvature bias and provides
additional information about the derivative or slope of the spectrum of a particular sig-
nal. We show that this Quadratic multitaper method provides improved estimates of the
spectrum, especially where strong structure is present, for example around the corner
frequency fc . Chapter 5 is under consideration for publication in Geophysical Journal
International as ”Quadratic Multitaper Spectrum” (Prieto et al., 2007b).
Chapter 6 summarizes the main conclusions of the thesis.

Acknowledgments

I would like to thank Greg Beroza and Luis Rivera and Luciana Astiz for
comments on early versions of this chapter. Funding for this research was provided by
NSF Grant number EAR0417983.
2

Confidence intervals for

earthquake source parameters

We develop a method to obtain confidence intervals of earthquake source pa-

rameters, such as stress drop, seismic moment and corner frequency, from single sta-
tion measurements. We use the idea of jackknife variance combined with a multitaper
spectrum estimation to obtain the confidence regions. The approximately independent
spectral estimates provide an ideal case to perform jackknife analysis. Given the partic-
ular properties of the problem to solve for source parameters, including high dynamic
range, non-negativity, nonlinearity, etc., a log transformation is necessary before per-
forming the jackknife analysis. We use a Student-t distribution after transformation to
obtain accurate confidence intervals. Even without the distribution assumption, we can
generate typical standard deviation confidence regions. We apply this approach to four
earthquakes recorded at 1.5 km and 2.9 km depth at Cajon Pass, California. It is nec-
essary to propagate the errors from all unknowns to obtain reliable confidence regions.
From the example, it is shown that a 50% error in stress drop is not unrealistic, and
even higher errors are expected if velocity structure and location errors are present. An
extension to multiple station measurement is discussed.

21
 22

2.1 Introduction

There is a long-standing controversy on whether stress drop increases with

earthquake magnitude or remains constant over a wide range of earthquake sizes (Aki,
1967; Archuleta et al., 1982; Kanamori et al., 1993; Abercrombie, 1995; Mayeda and
Walter, 1996; Ide and Beroza, 2001). The behavior of source parameters including stress
drop, corner frequency, radiated seismic energy and apparent stress are of key importance
in understanding the physics of earthquakes. However, it is difficult to estimate stress
drop reliably from seismograms since it is dependent on the cube of the corner frequency
fc and in turn, fc is dependent on an accurate account of seismic attenuation, path
effects, etc. These factors lead to considerable uncertainty in estimates of stress drop
and other source parameters.
Abercrombie (1995) used records from a 2.5 km deep borehole in Cajon Pass and
showed that the data supported a constant stress drop, but also an increasing apparent
stress with earthquake magnitude. Also from deep borehole data, Prejean and Ellsworth
(2001) reported a similar result. A magnitude dependency has been supported by some
studies (e.g., Kanamori et al., 1993; Mayeda and Walter, 1996; Mori et al., 2003). while
other studies have suggested scale independence (e.g., McGarr, 1999; Ide and Beroza,
2001; Ide et al., 2003), finding no evidence of increasing stress drop or apparent stress
with magnitude.
More recently Abercrombie and Rice (2005) revisited some of the Cajon Pass
seismograms and using both spectral fitting and Empirical Green Functions (EGF) con-
cluded that both apparent stress and stress drop may increase with increasing earthquake
size, but noted that the uncertainties were still large and scale independence could not
be entirely discarded.
So, what are the uncertainties of the estimated source parameters? Error analy-
sis for source parameters has been attempted before (e.g., Archuleta et al., 1982; Fletcher
et al., 1984) but seems to have been neglected more recently. Recently Prieto et al. (2006)
developed an approach to obtain uncertainties in earthquake source spectrum using EGF
and applied it to obtain confidence intervals of radiated seismic energy. As pointed out
by Tukey (1960):
“Probably the greatest ultimate importance, among all types of statistical
 23

procedures we now know, belongs to confidence procedures which, by making

interval estimates, attempt to reach as strong conclusions as are reasonable
by pointing out, not single likely values, but rather whole classes (intervals,
regions, etc.) of possible values, so chosen that there can be high confidence
that the ‘true’ value is somewhere among them. Such procedures are clearly
quantitative conclusion procedures. ... ”

In this paper we use the idea of the jackknife variance (Tukey, 1958) and follow
a similar recipe to the one applied for spectra (Vernon, 1989; Thomson and Chave, 1991)
to construct confidence intervals for earthquake source parameters. This is applied to
single station seismograms but can easily be extended to multiple station and spectral
ratios and EGF techniques. The confidence intervals are of paramount importance to
obtain meaningful scaling relations when different studies, regions, etc. are compared.
We present an example from data recorded at the Cajon Pass Borehole Ex-
periment Phase II with some records also used in Abercrombie (1997) and show the
resultant confidence intervals for stress drop and other source parameters for four small
earthquakes (Table 2.1) that were also recorded by the Southern California Seismic Net-
work (SCSN). Figure (2.1) shows a map with the relocated earthquakes and the borehole
location.

2.2 The Jackknife Method

The jackknife was first introduced by Quenouille (1949) and then named and
extended by Tukey (1958) to estimate variances. It is one of many resampling methods
used for statistical inference. One of the great advantages of the jackknife is that one
does not need to know the statistical distribution of the parameter in question and that
it works on complicated processes reliably (a detailed proof is given in Reeds, 1978). In
this paper we will use the so-called delete-one jackknife, which we will refer to simply as
the jackknife. A good review can be found in Miller (1974) and Efron (1982).
Assume X1 , X2 , . . . , XK are K independent random observations taken from
an unknown probability distribution characterized by a parameter θ which is to be
estimated. The estimate of θ using all observations is:

θ̂ = θ̂[X1 , X2 , . . . , XK ] (2.1)
Table 2.1: Hypocentral parameters for the earthquakes recorded at Cajon Pass used in this study. The relocations by Shearer et al.
(2005), model SHLK 1.02, are given for SCSN earthquakes.
ID Year Month Day Hour Min Sec Lat (◦ ) Lon (◦ ) Depth (km) ML Dist (km) CuspID
01 1994 01 01 14 56 42.087 34.2139 −117.4194 14.98 1.1 12.37 3138796
02 1994 01 01 17 47 31.257 34.3863 −117.0185 11.04 3.5 42.86 3138805
03 1994 08 22 21 27 51.366 34.3522 −117.6339 11.03 2.3 14.98 3181641
04 1994 09 21 04 16 51.762 34.2316 −117.4623 13.63 2.3 9.32 3185485
ML is the local magnitude given by the SHLK catalog, Dist is the hypocentral distance from
the borehole station, CuspID is the SCSN ID for the earthquakes.

-117.8 -117.6 -117.4 -117.2 -117.0 -116.8

0
12.5
25

34.4 km

03 02
Cajon Pass Borehole

34.2 04
01

34.0

Figure 2.1: Map showing the Cajon Pass Borehole location and relocated earthquakes considered in this study. Relocations from
Shearer, Hauksson and Lin (2005) model SHLK 1.02
24
 25

Let
θ̂_
ı = θ̂[X1 , . . . , Xi−1 , Xi+1 , . . . , XK ] (2.2)

be the delete-one estimate of θ, where the ith observation Xi is not used to estimate

ı . The data are thus subdivided in K groups of size (K − 1) by deleting each entry
θ̂ _
in turn.
An important application of the jackknife was suggested by Tukey (1958), and
is the jackknife estimate of the variance of θ̂
K
K − 1 Xh i2
var{θ̂} = θ̂_
ı − θ̂_
· (2.3)
K
i=1

where
K
1 X
θ̂_
· = θ̂_
ı (2.4)
K
i=1

is the mean of the delete-one estimates (2.2). Although it has been proposed (Wu,
1986) that deleting an arbitrary number of observations might have better convergence
properties, we use throughout the paper the delete-one jackknife because of its simplicity,
efficiency and independence of an arbitrary chosen subdivision for the groups.
As suggested by Miller (1974) and applied in spectrum estimation (Vernon,
1989; Thomson and Chave, 1991) it is sometimes necessary to use a transformation that
stabilizes the variance, especially when the statistic being investigated is bounded or its
distribution is strongly non-gaussian. This can be important when estimating errors in
stress drop ∆τ , seismic moment M0 , and corner frequency fc , all with a range [0, ∞).

2.2.1 Jackknife in Regression Problems

Consider the regression problem for a basic model

Y = Aβ + e (2.5)

where Y, e are m sized vectors of the data and the errors, A is a m × p matrix from the
model, and β is a p size vector of the parameters we wish to find.
Miller (1974) examined the traditional jackknife approach by deleting rows of
both Y and A simultaneously and showed the asymptotic normality of the jackknife
 26

solution vector and its variance under general conditions. The delete-one estimate is
given by solving
Y_ ı β̂_
ı = A_ ı (2.6)

where Y_
ı and A_
ı have the ith row removed. As will be clear in the subsequent

sections, the problem to solve for source parameters is non-linear and the model (2.5) is
not appropriate. Instead we have

yi = gi (β) + ei (2.7)

where gi is a nonlinear smooth function of the parameters in β (Fox et al., 1980; Wu,
1986). In an analogous way, we want to obtain the delete-one estimates β̂_
ı that satisfy

(2.7) by means of one of many non-linear parameter estimation techniques (non-linear

least-squares, grid search, etc).

2.3 Multitaper Spectrum estimates

The multitaper spectrum algorithm was introduced by Thomson (1982) and

has been widely used in the geophysical community (e.g., Park et al., 1987b; Vernon,
1989; Chappellaz et al., 1990; Lees and Park, 1995; Abercrombie, 1995). The method
takes advantage of a family of orthogonal tapers which are resistant to spectral leakage.
Given a time series x(t) with N contiguous data samples and assuming unit
sampling, we multiply the time series by a sequence a(t) called a taper and apply a DFT
N
X −1 N
X −1
Y (f ) = x(t)a(t)e−2πif t with |a(t)|2 = 1 (2.8)
t=0 t=0

to obtain a direct estimate of the true spectrum S(f ) of the signal

Ŝ(f ) = |Y (f )|2 (2.9)

The question is then what taper to use? Is there a reason to prefer one taper over the
other?
Spectral leakage is the bias introduced by energy leaking from frequencies dif-
ferent from the frequency f for S(f ). Now the question becomes: what taper a(t) has
 27

the greatest concentration of energy in its Fourier transform? Spectral properties of the
taper can be studied from its DFT
N
X −1
A(f ) = a(t)e−2πif t (2.10)
t=0

The function |A(f )| for conventional tapers has a broad main lobe and a succession of
smaller sidelobes. The larger the sidelobes, the more spectral leakage is biasing Ŝ(f ).
We can express the estimate in equation (2.9) as a convolution of the taper
transform (2.10) and the true spectrum S(f ) (see Thomson, 1982; Park et al., 1987b, for
derivation):
Z1/2
2
Ŝ(f ) = |A(f − f 0 )| S(f 0 ) df 0 (2.11)
−1/2

The interpretation of this equation is as a convolution describing the smearing of the true
spectrum as a consequence of the discrete sampling. A good taper will have a spectral
window with low amplitudes whenever |f − f 0 | gets large and large amplitudes whenever
|f − f 0 | is small.
Slepian (1978) suggested choosing a frequency W , where 0 < |W | ≤ 1/2 (unit
sampling) and maximizing the fraction of energy of A at frequencies from (−W, W ). In
mathematical form this is equivalent to:

ZW
|A(f )|2 df
−W
λ(N, W ) = (2.12)
Z1/2
|A(f )|2 df
−1/2

Since no finite time series can be completely band-limited, λ < 1. The spectral leakage
comes from the sidelobes of A(f ) convolved with the spectrum outside the band (f −
W, f + W ). One can think of λ(N, W ) as the amount of spectral energy at Ŝ(f ) that
comes from (f − W, f + W ) and 1 − λ as the amount that comes from outside the band
or as the bias from outside the band.
We wish to maximize the value of λ by choosing A(f ) appropriately. Substitute
(2.10) into (2.12) and represent a(t) by an N-vector of coefficients a; taking the gradient
 28

of λ with respect to a and setting to zero leads to the matrix eigenvalue problem:

D · a − λa = 0 (2.13)

where D is a symmetric matrix

sin2πW (t − t0 )
D(t, t0 ) = (2.14)
π(t − t0 )

with eigenvalues 1 > λ0 > λ1 > · · · > λN −1 > 0 and associated eigenvectors vk (t; N, W )
called the Slepian sequences (Slepian, 1978). From now on we will drop the explicit
dependence on N and W .
The eigenvector with the largest eigenvalue is the best possible taper for the
suppression of spectral leakage, and in practice we find λ0 is usually extraordinarily close
to one. But in fact it can be proved that the first 2N W −1 eigenvalues are also very close
to one, leading to a whole family of excellent tapers. The multitaper method exploits this
fact by using all of these tapers rather than merely the first one. Because the eigentapers
are orthogonal (both in time and frequency domains), the estimates based on them are
statistically independent of each other and can therefore be combined together to yield
a more reliable overall estimate as we will explain.
In practice we choose a bandwidth W over which the spectrum is to be smoo-
thed, thus fixing N W , which is called the time-bandwidth product of the system under
study. For practical problems we always choose N W > 1, because we cannot expect to
obtain good concentration into a frequency band narrower than fR = 1/N , the Rayleigh
resolution.
Figure (2.2) shows the Slepian sequences and their Fourier transforms with cor-
responding eigenvalues for a time series with N = 100 samples, N W = 4, and W = 0.04
for unit sampling. The horizontal axis is shown in Rayleigh units, basically equivalent
to the frequency sampling. Figure (2.2) suggests that N W = 4 is equivalent to saying
that the smoothing will take place over N W Rayleigh bins around the frequency of in-
terest. Note that here we have assumed unit sampling – if that is not the case, then the
time-bandwidth product is actually ∆tN W = 4, where ∆t is the sampling rate, in order
to maintain the proper units of W .
Turning back to the spectral estimation problem, given a particular bandwidth
 29

4π Slepian tapers
0.2

0.1
Amplitude

-0.1

-0.2
0 20 40 60 80 100
Sample
4π taper transforms
1 4
10 3 λ0 = 0.99999
2 λ1 = 0.99999
1 λ2 = 0.99999
0 λ3 = 0.99996
Transform Amplitude

-1 λ4 = 0.99942
10

-3
10

-5
10

0 4 8 12 16
Frequency (Rayleigh units)

Figure 2.2: The five lowest order 4π Slepian tapers (top panel) and corresponding
Fourier transform amplitudes (lower panel). Solid lines correspond to the zero order
sequence v0 (t), higher order tapers are plotted with dashed lines. We have used N = 100
and N W = 4. Estimated eigenvalues are also provided.
 30

W , we compute DFTs of the tapered data Yk (f ), called the eigencomponents,

N
X −1
Yk (f ) = x(t)vk (t)e−2πif t (2.15)
t=0

We generally use k = 1, . . . , K, where K = 2N W − 1. As expressed above, the corre-

sponding eigenvalues are λk ≈ 1 with good leakage properties.
As suggested by Thomson (1982) we use the adaptive weighting procedure

Yk (f ) = dk (f )Yk (f ) (2.16)

and the corresponding adaptive spectral estimate

K−1
X
|Yk (f )|2
k=0
Ŝ(f ) = K−1
(2.17)
X 2
|dk (f )|
k=0

where the weights dk (f ) are chosen to reduce bias from spectral leakage. The frequency
dependent weights are useful in the analysis of high dynamic range spectral processes.
The weights work as follows. At frequencies where the spectrum is reasonably flat, the
weights dk (f ) ≈ 1, thus reducing the variance of the spectral estimate by averaging over
all the eigencomponents Yk . At frequencies where the spectrum has a large dynamic
range the higher order eigencomponents might be biased and the weights reduce the
contributions from these components.
The optimal weights dk (f ) can be found by minimizing the misfit between the
estimated spectrum and the true spectrum S(f ). The approximate optimum weights are
√
λk S(f )
dk (f ) ≈ (2.18)
λk S(f ) + (1 − λk )σ 2

where σ 2 represents the variance of the time series. The term (1 − λk )σ 2 represents
an approximation to the bias from spectral leakage. Since we do not know the true
spectrum, we replace S(f ) by an estimate Ŝ(f ).
We find the weights and estimated spectrum Ŝ(f ) by iteration. As an initial
estimate of S(f ) we take the arithmetic average of the first two squared eigencomponents
|Y0 (f )|2 and |Y1 (f )|2 and substitute in (2.18) to obtain estimates of dk (f ). The weights
are then used in (2.16 - 2.17) to obtain a new spectral estimate Ŝ(f ) and this process is
 31

repeated. Convergence is rapid and only a few cycles are necessary. Note that both the
tapers and weights are normalized in order to keep the spectrum in physical units.
The kth eigenspectrum is

2
Ŝk (f ) = |Ŷk (f )| (2.19)

For the jackknife approach, we will use the Ŝk as the K independent estimates of the
spectrum. At each frequency f the multitaper estimate of the log spectrum is given by
K
" #
1 X
ln Ŝ = ln Ŝk (2.20)
K
k=1

and we also define the delete-one spectrum

 
K
1 X
ln Ŝ_
ı = ln  Ŝk  (2.21)
K −1
k=1,k6=i

The logarithmic transformation of the spectrum is suggested in Thomson and Chave

(1991), providing a more symmetric distribution than the standard χ2 for spectral esti-
mates.

2.4 Source parameter jackknife

A general source model of the displacement spectra of both P and S waves

(e.g., Abercrombie, 1995) is:

Ω0 e−(πf t/Q)
u(f ) =  1/γ (2.22)
1 + (f /fc )nγ

where Ω0 is the long period amplitude, f is the frequency, fc is the corner frequency, n
the high frequency fall-off rate, γ is a constant, t is the travel time, and Q a frequency
independent quality factor. Modified versions of spectral shapes proposed by Brune
(1970) and Boatwright (1980) can be obtained by changing γ. Based on previous studies
of data from the Cajon Pass Borehole (Abercrombie, 1995, 1997) a value γ = 2 and a
variable fall-off n fits the spectra reasonably well. In this paper we will use

Ω0 e−(πf t/Q)
u(f ) =  0.5 (2.23)
1 + (f /fc )2n
 32

Following Ide et al. (2003) take the logarithm

ln u(f ) = g(f ; β) (2.24)

πf t
= ln Ω0 − 0.5 ln(1 + (f /fc )2n ) − (2.25)
Q

where β is a vector of three components given by the parameters we are searching

for, namely Ω0 , fc , Q. The function g(f, β) is clearly non-linear and some kind of non-
linear inversion is necessary. It is not the aim of this paper to discuss the difficulties
encountered in solving this problem, and we suggest reading Bard (1974) on nonlinear
parameter estimation, as applied to source physics (see Abercrombie, 1995; Ide et al.,
2003, and references therein).
The estimate β̂ of β using all observations (as in equation 2.1) is given by
the solution of (2.24) using the multitaper spectrum estimate (2.20). The delete-one
β̂_
ı parameter instead uses the delete-one spectrum (2.21). The result is K delete-one

estimates of the long period amplitude, corner frequency and quality factor, denoted
respectively Ω0, _
ı , fc, _
ı , Q_
ı .

2.4.1 Transformations

The use of transformations before performing the jackknife is in some cases

necessary. In terms of source parameters a logarithmic transformation should provide
more stable estimates of variance. Some reasons for this are as follows:

• The three seismic parameters are non-negative. If a simple Gaussian distribution

is assumed, the tails give a nonzero probability that a parameter is negative, which
is not physical. The jackknife does not constrain variances to be positive.

• As suggested by (Archuleta et al., 1982), if no transformation is performed, the

arithmetical average will be biased to the larger values, while taking a transformed
parameter gives equal weight to all independent estimates.

• A closer to normal distribution of the errors is achieved by such a transformation.

In this respect we will perform the jackknife on θ = ln Ω0 rather than Ω0 .

 33

2.4.2 Confidence intervals

As a result of section 4.1 we obtain an estimate θ̂ = ln β̂ and the variance of

the transformed variable

σ̃ 2 = var{ln β̂} (2.26)

K
K −1 X 2
= ı − ln β̂_
[ln β̂_ · ] (2.27)
K
i=1

where β̂_
ı is any one of Ω̂0, _
ˆ ı , and Q̂_
ı , fc, _ ı .

ı − ln β̂_
Tukey (1958) suggested that (ln β̂_ · )/σ̃ is nearly distributed as Stu-

dent’s t with K − 1 degrees of freedom for small samples. Hinkley (1977) on the other
hand stated that if the data have strongly nonnormal distributions, the Student t approx-
imation can lead to substantial errors. However, if the transformation performed leads to
more nearly normal distributions, the approximation is reasonably accurate (Davidson
and Hinkley, 1997). Note that this distribution is very close to the Gaussian distribution
and for 30 or more degrees of freedom they are almost indistinguishable. With this in
mind, the double-sided 1 − α confidence interval of the long period amplitude is

Ω̂0 e−tK−1 (1−α/2)σ̃ < Ω̂0 ≤ Ω̂0 etK−1 (1−α/2)σ̃ (2.28)

and similar for fˆc and Q̂. If the Student t approximation is not appropriate for the
particular data, one can always simply plot the ±σ̃ bounds by adjusting (2.28). Note
that because of the transformation, the lower limit is never negative.

2.4.3 Seismic Moment, Source Radius and Stress Drop

Other important source parameters estimated from the spectrum are the seismic
moment (M0 ), the source radius (r) and the stress drop (∆τ ) and are often calculated
assuming a circular fault (Brune, 1970; Madariaga, 1976), in which case
4πρc3 RΩ̂0
M0 = (2.29)
Uθφ
kβ
r = (2.30)
fˆc
where a constant rupture velocity is assumed. From the mean estimates of the previous
two equations,
7M0
∆τ = (2.31)
16r3
 34

where ρ, c, R, Uθφ , β are density, wave velocity, hypocentral distance, the mean radiation
pattern (0.52 and 0.63 for P and S waves) and the shear wave velocity at the source. k
is 0.32 and 0.21 for P waves and S waves respectively, assuming the rupture velocity is
0.9β (Madariaga, 1976).
We will assume that the parameters not associated with the source (shear wave
speed, density, etc.) are known exactly, that is, do not contribute to the uncertainties of
seismic moment, source radius and stress drop. We will use the idea of propagation of
errors (Taylor, 1997) to obtain confidence limits of these parameters.
We perform the propagation of errors in the log domain, since it is where we have
2 and σ 2 . The idea is to
variance estimates of ln Ω0 and ln fc , denoted respectively σΩ 0 fc
2 = var{ln M }, source radius σ 2 = var{ln r},
obtain the variance of seismic moment σM 0 0 r
2
and stress drop σ∆τ = var{ln ∆τ }. After this, equation (2.28) can be used to obtain
confidence intervals. Some rules of propagation of errors are shown in the appendix.
The relation of errors between the source and spectral parameters are

2 2
σM 0
= σΩ 0
(2.32)

σr2 = σf2c (2.33)

and a more complicated relation is obtained for the stress drop, since it depends on two
variables

2 2
σ∆τ = σM 0
+ 9 σr2
2
= σΩ 0
+ 9 σf2c (2.34)

where it is assumed that the covariance of Ω0 and fc is negligible. This relation was
used by Fletcher et al. (1984) to estimate uncertainties of stress drop using a multiplica-
tive error. Again, the bounds (either ±σ or confidence intervals using the Student t
approximation) can be transformed back to the linear domain using (2.28).

2.5 Application to Cajon Pass Data

Because attenuation can also cause fall-off at high-frequencies it is important

to correct observed spectra for Q effects. Full consideration of these effects is beyond our
 35

ML 3.5 Distance = 43.12 km

H2

Displacement (microns)
8
H1
4

Z
0

-4
8 9 10 11 12 13
Time (seconds)

Figure 2.3: Example seismogram of the largest event used in this study, ML 3.5 recorded
at the deepest borehole sensor 2.9 km. Seismograms have been corrected for instrument
response and are flat to displacement between 2 and 300 Hz. Horizontal bars show the
choice of noise and P wave window.

focus here, therefore as a demonstration of the jackknife procedure to obtain variance

and confidence intervals we choose seismograms recorded at two different depths (1.5 and
2.9 km) at the Cajon Pass Borehole, where attenuation effects are relatively small and
have been previously modeled by Abercrombie (1995). The seismometers that recorded
this data set are 10-Hz L-15LA high temperature geophones, with sample rates of 1000
samples/sec.
Figure 2.3 shows displacement seismograms recorded at the 2.9 km depth sensor
for a ML 3.5 earthquake 43 km away (ID 02 in Table 2.1). The spectrum is computed
for a 1 second window, starting 0.15 seconds before the P pick at the station, similar to
windows used in previous work (e.g., Abercrombie, 1995; Prieto et al., 2004; Abercrombie
and Rice, 2005) for small earthquakes.
For each of the three components (Z, H1, H2) we estimate the amplitude spec-
trum using a time-bandwidth product N W = 4 and work with K = 7 tapers. This means
we also compute for each component 7 delete-one spectra as in equation (2.21). The fi-
nal amplitude spectrum is then computed by vector summation of the three component
spectra, and similarly for the delete-one spectra.
Figure 2.4 shows the complete set of delete-one spectra for the ML 3.5 earth-
quake recorded at the 2.9 km sensor. The spectra have been shifted for comparison.
In general the spectral shapes are very similar and only slight differences at very low
 36

0
10
ML3.5

Displacement (microns/Hz)
-2
10

-4 fc, i
10
10.99
10.75
10.71
11.39
-6 10.86
10 10.72
10.98

10 100
Frequency (Hz)

Figure 2.4: The delete-one spectrum for the ML 3.5 earthquake, from signal in figure
2.3 at the 2.9 km sensor. The spectra have been shifted for comparison purposes. All
spectra show a similar behavior and slight differences are seen. For each spectrum, the
delete-one corner frequency estimates are listed.

frequencies and roughness at higher frequencies are visible. Note that in Figure 2.4 the
delete-one spectra are plotted, which are not independent estimates. Only Ŝk (f ) are
treated as independent, given the orthogonality properties of the Slepian tapers.
Here we have a good example of the properties of the multitaper algorithm.
The data from the Cajon Pass have a very strong 60 Hz signal. In this work, we have
N = 1000 samples, dt = 0.001, we chose N W = 4, the band W = 4 Hz and we use
K = 7 tapers. This means that the 60 Hz peak will be smoothed over the band between
56 − 64 Hz. In Figure 2.4 we can see a very sharp discontinuity at 64 Hz due to the fact
that outside the band, very little energy is leaked to frequencies f > 64 Hz.
Following Abercrombie (1995) and Abercrombie and Rice (2005) we use Q =
1000 and correct the spectra before performing spectral fitting. We vary the fitting
bandwidth to obtain optimal fits, but the same bandwidth is used for all delete-one
spectra and the average spectra. An example fit to the model (2.25) is shown in Figure
2.5 for the largest and smallest earthquake in this study. Note however that the 1.5 km
sensor data are used for the small earthquake, due to complicated resonances present at
 37

0
10

Displacement (microns/Hz)
ML3.5

-2
10

ML1.1

-4
10

10 100
Frequency (Hz)

Figure 2.5: Spectral modeling of P waves for two different sized earthquakes. The
bandwidth use to fit the source model varies depending on event size. Attenuation
correction was previously performed, using Q = 1000 as suggested by Abercrombie
(1995).

the 2.9 km sensor spectra that may have affected the results.
As in Abercrombie and Rice (2005) we use density ρ = 2700 kg/m3 , α = 6000
√
m/s and β = α/ 3 and using equations (2.29 to 2.31) we estimate the source parameters
M0 , fc , r, ∆τ from the average spectra and the jackknife parameters (e.g., M0, _
ı ) from

the delete-one spectra to get the jackknife variance and confidence intervals. Table 2.2
shows the source parameters and 5-95% confidence limits. Figure 2.6 shows plots of
seismic moment and corner frequencies and seismic moment and stress drop for the data
used in this paper.
It is important to note the assumptions and unknowns in the calculations. For
example, we have assumed that the wave speeds (α, β) are known exactly. If there are
errors (and certainly there are) associated with the wave speed, errors will propagate to
M0 , r and subsequently to the stress drop. Assuming a 5% error in the S wave speed,
thus affecting the radius uncertainties (and rupture speed), the confidence region for the
stress drop for the ML 3.5 earthquake recorded at 2.9 km sensor would be (47, 108), a
change of about 10%. Other sources of errors for this example include the attenuation
correction, the constant Q assumption used, earthquake location errors, radiation pattern
 38

and directivity, etc. Perhaps most importantly, we assume the validity of the source
model; our method provides an estimate of the errors in ∆τ with respect to random
fluctuations in the data but is not a test of the validity of the model itself.

2.5.1 Extension to multiple stations

A generalization of the jackknife to multiple stations is desirable. A major

source of uncertainty would be directivity, because as expected from directivity, the
pulse width of the source time functions is narrower in the direction of rupture and
broader in the opposite direction, also changing the corner frequency (e.g., McGuire,
2004).
One approach is to treat the different station estimates of corner frequencies fc
and seismic moment M0 as independent, and, after suitable transformations, compute
confidence intervals as explained in sections 2 and 4. If the earthquake source spectrum is
simple (no directivity effects, radiation pattern correctly accounted for, etc.) all stations
would return a similar estimate, thus having small uncertainties, while if there is strong
directivity, certain stations will have considerably different corner frequencies, increasing
uncertainties. A recipe for a multiple station jackknife of fc is as follows:

(1). Compute fc for every station that recorded the earthquake.

(2). Use the log transformation.

(3). Compute a mean ln fˆc (eq. 2.1).

(4). Compute delete-one ln fˆc, _

ı (eq. 2.2).

(5). Compute variance var{ln fc } = σf2c (eq. 2.3).

(6). Obtain confidence intervals (eq. 2.28).

A similar approach could be used for other parameters such as long period
amplitude Ω0 , Q, etc. Propagation of errors (eqs 2.32 to 2.34) is necessary to obtain
confidence intervals on seismic moment, source radii and stress drop.
 Table 2.2: Source parameters and confidence intervals obtained by spectral fitting and jackknife analysis.
ID ML Sensor M0 (Nm) M0 conf fc (Hz) fc conf n r(m) ∆τ (MPa) ∆τ conf
01 1.1 1.5 0.96e+11 (0.78 1.17)e+11 43.80 (33.29 57.63) 1.72 25 2.60 ( 1.11 6.07)
02 3.5 1.5 1.73e+14 (1.11 2.68)e+14 6.92 ( 4.77 10.03) 1.72 160 18.34 ( 5.75 58.47)
02 3.5 2.9 1.72e+14 (1.35 2.19)e+14 10.91 ( 9.91 12.00) 2.09 101 71.55 (52.78 96.98)
03 2.3 2.9 1.53e+12 (1.31 1.79)e+12 20.74 (17.46 24.63) 1.77 53 4.55 ( 2.72 7.64)
04 2.3 2.9 1.36e+12 (1.00 1.85)e+12 21.66 (18.74 25.02) 2.03 51 4.66 ( 2.90 7.50)
M0 conf, fc conf and ∆τ conf are the 5-95% confidence intervals for the seismic moment, corner frequency and stress
drop respectively. n is the high-frequency fall-off for the P waves. Q = 1000 is assumed. The column named Sensor
represents the depth in km. of the sensor recording the earthquake.

Mw
1 1.5 2 2.5 3 3.5
100

10

fc (Hz)
1

100

10

Stress Drop (MPa)

10 11 12 13 14 15
10 10 10 10 10 10
Seismic Moment (Nm)

Figure 2.6: Source parameters and 5-95% confidence intervals using a Student t approximation. Symbols with dot at the center
have parameters determined at the 1.5 km deep sensor. Note that the confidence regions may vary from event to event.
39
 40

2.6 Conclusions

In estimating source parameters from the seismic spectrum, it is important

not only to obtain a measure of the source parameters but also to obtain a measure
of the uncertainties, by means of confidence intervals. The jackknife is a reliable way
of estimating the variance of source parameters, and, given suitable transformations,
confidence intervals. It should be easy to extend this approach to multiple station studies,
where other sources of error include the radiation pattern and directivity effects which
might generate different corner frequencies and radiated energy.
We calculate source parameters and confidence intervals for four small earth-
quakes as an example of the use of the jackknife approach. The error analysis is necessary
if the data are to be used to constrain rupture models (Abercrombie and Rice, 2005),
examine scaling relations and the size dependence of earthquake parameters, in order to
conclude, within a reasonable reliability, something about the physics of earthquakes.
From Figure (2.6) there appears to be a slight increase of stress drop with
earthquake magnitude, which, unless the errors are kept small, would pass unnoticed.
Note also that the larger uncertainties are associated with the 1.5 km sensor, compared to
at 2.9 km. If some of the assumptions such as radiation pattern and earthquake location
contribute to the errors, the stress drop scaling would be less apparent, suggesting the
need to find ways of reducing uncertainties.
The M3.5 earthquake (ID02) was recorded at two different depths and the
corner frequency confidence regions barely overlap. This variation is likely explained
by other sources of error such as near site effects at the shallower station. This also
shows that even stations close to each other may have very different estimates of source
parameters and uncertainties are needed to address the significance of these estimates.

Acknowledgments

We thank R. E. Abercrombie for help on using the Cajon Pass data and two
anonymous reviewers for their constructive comments. Funding for this research was
provided by NSF Grant number EAR0417983. This research was also supported by the
Southern California Earthquake Center. SCEC is funded by NSF Cooperative Agree-
 41

ment EAR-0106924 and USGS Cooperative Agreement 02HQAG0008. The SCEC con-
tribution number for this paper is 979. We also thank IRIS and the staff at the Data
Management Center for access to the Cajon Pass data.
3

Earthquake source scaling and self-similarity

estimation from stacking P and S spectra

We study the scaling relationships of source parameters and the self-similarity

of earthquake spectra by analyzing a cluster of over 400 small earthquakes (ML = 0.5
to 3.4) recorded by the ANZA seismic network in southern California. We compute P ,
S and pre-event noise spectra from each seismogram using a multitaper technique and
approximate source and receiver terms by iteratively stacking the spectra. To estimate
scaling relationships, we average the spectra in size bins based on their relative moment.
We correct for attenuation by using the smallest moment bin as an empirical Green’s
function (EGF) for the stacked spectra in the larger moment bins. The shapes of the
log spectra agree within their estimated uncertainties after shifting along the ω −3 line
expected for self-similarity of the source spectra. We also estimate corner frequencies
and radiated energy from the relative source spectra using a simple source model. The
ratio between radiated seismic energy and seismic moment (proportional to apparent
stress) is nearly constant with increasing moment over the magnitude range of our EGF
corrected data (ML = 1.8 to 3.4). Corner frequencies vary inversely as the cube root of
moment, as expected from the observed self-similarity in the spectra. The ratio between
P and S corner frequencies is observed to be 1.6 ± 0.2. We obtain values for absolute
moment and energy by calibrating our results to local magnitudes for these earthquakes.
This yields a S to P energy ratio of 9 ± 1.5 and a value of apparent stress of about 1
MPa.

42
 43

3.1 Introduction

A major question in seismology is whether the faulting mechanism of large

and small earthquakes involves different physics. That is, is a M = 8 earthquake just
a M = 2 earthquake scaled upward by a large factor or is something fundamentally
different occurring? Aki (1967) proposed scale invariance of the rupture process, con-
sistent with observations that many geological processes are similar over a wide range
of scales (Abercrombie, 1995). There is currently a debate regarding whether earth-
quakes are truly self-similar over their entire size range or if systematic departures from
self-similarity are observed (see, for example Abercrombie, 1995; Ide and Beroza, 2001).
Thus, although many mechanisms have been proposed for differences in the physics of
larger earthquakes, including shear melting (Jeffreys, 1942; Kanamori and Heaton, 2000),
acoustic fluidization (Melosh, 1979), rough fault sliding-induced normal stress reduction
(Brune et al., 1993), fluid pressurization (Sibson, 1973), and elastohydrodynamic lubri-
cation (Brodsky and Kanamori, 2001), the need for different mechanisms is not yet firmly
established.
Studies of earthquake scaling generally involve comparisons between static mea-
sures of size (e.g., moment) and dynamic measures of size (e.g., energy). Both measures
are typically derived from spectra of seismograms recorded at some distance from the
earthquakes. Because moment is obtained from the low frequency part of the spectra, it
is usually measured much more reliably than energy or corner frequency measurements,
which require the high-frequency part of the spectra where correcting for attenuation
and other path effects can be difficult. Current estimates of seismic moment made inde-
pendently from local, regional and teleseismic data usually agree within about a factor
of two. In contrast, estimates of seismically radiated energy by different investigators for
the same earthquake often differ by more than an order of magnitude (e.g., Singh and
Ordaz, 1994; Mayeda and Walter, 1996).
This uncertainty in seismic energy leads to different interpretations of the en-
ergy density of earthquakes, as measured by the Energy/Moment ratio, which is often
scaled by rigidity to represent the apparent stress. Several authors find evidence that
apparent stress increases with magnitude (Kanamori et al., 1993; Abercrombie, 1995;
 44

Mayeda and Walter, 1996; Izutani and Kanamori, 2001; Mori et al., 2003) while others
argue that apparent stress is approximately constant (Choy and Boatwright, 1995; Mc-
Garr, 1999; Ide and Beroza, 2001; Ide et al., 2003). Constant apparent stress implies
similar physics for small and large earthquakes, while increasing apparent stress with
magnitude implies that large earthquakes are more efficient radiators of seismic energy
than small ones.
Our approach here is to improve the reliability and stability of source spectra by
stacking and averaging thousands of records from the Anza seismic network in southern
California. We use a simple method (Warren and Shearer, 2000, 2002) to isolate the
relative source spectra from the path and site effects by stacking the computed log
spectra after subtracting the appropriate path-site terms. This approach is also similar
to that used by Andrews (1986) to analyze spectra of the 1980 Mammoth Lakes California
earthquake sequence. Rather than obtaining an absolute measure of individual source
spectra, we obtain relative shapes of spectra with respect to other earthquakes. We then
stack the spectra in bins of similar moment to obtain average spectra (and estimated
uncertainties) as a function of earthquake size and apply attenuation corrections using the
smallest earthquakes as empirical Green’s functions (EGF) (e.g., Mueller, 1985; Hough,
1997). The resulting spectra are sufficiently smooth that direct tests of the self-similarity
hypothesis are possible, as well as measurements of corner frequency and energy. All of
our results indicate self-similarity is closely obeyed over the ML = 1.8 to 3.4 size range
of our EGF corrected data.

3.2 Data Processing

We used records from the Anza seismic network (Berger et al., 1984; Vernon,
1989) [Berger et al., 1984; Vernon, 1989], 9 high-quality, three-component stations lo-
cated on hard rock sites near an active part of the Clark Lake segment of the San Jacinto
fault in southern California (Figure 3.1). We began this study by selecting about 800
earthquakes located in a tightly clustered volume (4.5 km sided area, with most of the
events between 5 and 12 km depth) near the Toro Peak station (TRO) and 50 km from the
most distant station (RDM). In this region, the database is complete to about ML ≥ 0.5
 45

0 10 20 km

33.8

KNW
Sa
nJ
ac
int
oF
au
lt
RDM
33.6 PFO

CRY WMC
SND TRO
BZN
FRD

33.4
-117 -116.8 -116.6 -116.4 -116.2

Figure 3.1: Map showing the cluster of over 400 earthquakes (small black dots) and
ANZA stations (solid triangles) used in this study. The inset shows the location of the
study area (rectangle) in the state of California.

with generally good signal-to-noise ratio records. The earthquakes occurred from 1983
to 1993, at which time the network recorded at 250 samples per second with Geospace
HS–10 2–Hz seismometers. We selected a relatively compact group of earthquakes so
that the path to each station would be similar between different earthquakes, permitting
the use of simple corrections for attenuation and other path effects.
We use both P and S waves and select time windows for P on the vertical
component and time windows for S on all three components. Both windows start 0.5
seconds before the analyst pick of the arrival, with a total window length of 1.28 seconds.
We also select a noise window of the same length, with the last data point just before
the P -wave window. The velocity spectrum is estimated using the multi-taper algorithm
(Park et al., 1987b) and then corrected for the instrument response function. The S wave
spectrum is calculated as the vector summation of spectra from all three components.
Figure 3.2 shows an example of this process for a vertical-component record from station
F RD.
We apply a signal-to-noise ratio cutoff, where we use spectra only when the
mean ratio is greater than 5.0 in the 0–80 Hz frequency band and the ratio is greater
 46

FRD 3.24 ML
x 105
3
a

Velocity (nm/sec)
2
1
0
-1
-2
-3
2 4 6 8 10 12 14
Time (sec)

5
10
b

Amplitude (nm/sec)

Noise
P wave
S wave
0
10 0 1 2
10 10 10
Frequency (Hz)

Figure 3.2: Example of computed spectra from the largest magnitude earthquake in
the study area (ML = 3.4) recorded at station FRD, vertical component. (a) The time
series, with horizontal bars showing the noise, P , and S windows used to compute the
spectra; in this case the S-wave shows up more clearly on the horizontal components.
(b) Spectra for the windows shown in (a), computed using a multitaper method. Note
the rapid decrease in signal to noise ratio at the higher frequencies.

than 3.0 at 80 Hz. At higher frequencies, the signal to noise ratio decreases very rapidly
(see Figure 3.2), so we limit our analysis to frequencies below 80 Hz. After applying
the cutoff, we have 2735 records (including both P and S waves) from 470 earthquakes.
Because of their larger amplitudes, the S waves have generally higher signal-to-noise
ratios than the P waves; thus our signal-to-noise cutoff excludes P waves from the
smallest earthquakes in our data set, which are represented only by S wave spectra.
One possible concern is that P -wave coda may be contaminating the S-wave
window. This potential source of bias is likely to have its largest effect on the closest
stations where the S − P time is the smallest. To test what effect this may be having
on our results, we repeated our analyses using subsets of the data where we removed the
closest, the two closest and the four closest stations from the source region. Although
there was some increase in the variability of the stacked spectra as we reduced the number
 47

of data in the stacks, there were no systematic changes in the S spectral shapes. Thus
it does not appear that P contamination of S is a significant factor in our analyses.
Since multiple stations record every earthquake and many earthquakes are re-
corded at each station, we can isolate the source and receiver contributions to the spectra.
Because our source region is relatively compact, the receiver contributions will also in-
clude most of the path effects. Following the method described by Warren and Shearer
(2002) it is possible to isolate the relative source spectrum (Figure 3.3) if we assume
that the observed spectrum Dij (f ) from each source and receiver (denoted Si for the
ith earthquake and Rj for the jth station) is a product of source effects and path-site
effects. We iteratively stack all log spectra from each earthquake, after removing the
appropriate station term, to obtain the earthquake term:
n
1X
log(Si ) = [log(Dij ) − log(Rj )] (3.1)
n
j=1

and we also stack all log spectra from each station, after removing the earthquake term,
to obtain the path-station term:
m
1 X
log(Rj ) = [log(Dij ) − log(Si )] (3.2)
m
i=0

where Dij is the computed spectrum, Si the earthquake term for the ith earthquake and
Rj the path-station term for the jth station. Since the earthquake term and the path-
station term are dependent upon each other, we solve the set of equations iteratively
until we reach a stable result, where the fractional change in either the source or path-
station terms is less than 10−4 . We normalize the average log source spectra for all our
earthquakes to unity, as a starting point for the iteration process. In practice we are
mapping the deviations of the source spectra from this reference flat spectrum.
After source and path-station terms are separated we obtain 470 relative source
spectra and 9 path-station spectra (separately for P and S waves). Figure 3.4 shows the
P and S path-site spectra for nine different stations. Because we have not yet assumed a
source model (e.g., ω −2 , etc.), the shape of each of these spectra will include both source
and attenuation contributions. The information is contained in the differences between
these curves, which are significant because all of the stations recorded the same set of
earthquakes. Variations in attenuation among the stations can be seen in the position at
 48

a
A B C

b Computed Stacked Path-

Log Spectrum Station Term

A - =

+
B - =

+
C - =

Stacked
Earthquake
Term

Figure 3.3: Cartoon explaining how spectral stacking is used to obtain the earthquake
term, as in Warren and Shearer (2002). If a given earthquake (star) is recorded at
stations A, B, and C (Figure 3a), the earthquake term is computed by stacking the log
spectrum from earthquake 1 computed for stations A, B and C after removing the path-
station terms for these stations (Figure 3b). An analogous procedure is used to compute
the station terms.

which the spectra begin to falloff at high frequencies. For example, it is clear that station
SN D, located within 100 meters of the surface trace of the San Jacinto fault, records a
more attenuating path than station F RD, despite the fact that F RD is located slightly
closer to the earthquake cluster. In general, there is no clear distance dependence to the
observed path-station spectra, suggesting that local site effects beneath each station are
dominating the spectral differences among the stations.
Each of the 470 relative source spectra represents the average log spectra of all
stations recording the earthquake, after correcting for differences among the path-station
terms. To study how these source spectra vary as a function of earthquake size, we divide
our data into 20 bins in relative moment, which is estimated from the low-frequency
 49

BZN (298) CRY (275) FRD (433)

5
10
P wave S wave

Norm. Amplitude
3
10

1
10

KNW (373) PFO (355) RDM (182)

5
Norm. Amplitude 10

3
10

1
10

SND (384) TRO (106) WMC (329)

5
10
Norm. Amplitude

3
10

1
10

0 1 2 0 1 2 0 1 2
10 10 10 10 10 10 10 10 10
Freq (Hz) Freq (Hz) Freq (Hz)

Figure 3.4: Examples of path-station terms for P (solid line) and S (dashed line)
waves. In parenthesis is the number of earthquakes recorded at each particular station.
Note that the path-station term for the S-waves is always larger, reflecting the higher
amplitude of S compared to P .

spectral amplitude. Because our S-wave data span a larger total moment range than
the P -wave data, the moment range within each S-wave bin is larger than in the P -wave
bins. At this stage in our analysis, we do not compare P and S amplitudes directly;
rather we process the P and S spectra separately and obtain independent results for
each phase (later comparisons between P and S corner frequencies will jointly consider
the data).
Selected S wave results are plotted in Figure 3.5a. Each binned source spectra
is the result of averaging between 1 and 86 earthquake spectra (each of which is itself a
stack of spectra from different stations recording the earthquake). The resulting binned
source spectra are much smoother than the individual spectra that go into the stacks.
There are generally many more earthquakes in the bins at smaller moments because of
the much larger number of smaller earthquakes in the data set. The relative moments
among the bins can be seen in the low frequency limit of the spectra (i.e., at about
 50

S spectra
3 4
10 10
1 a b

2 2 3
10 1 10
2
2
2
Spectral Ratio
8
1 2
10 10 10
24
17
32
52
0 1
10 49 10
86
74
62
26
0
13 10
1 10 80 1 10 80
Freq (Hz) Freq (Hz)

Figure 3.5: Relative source spectral shapes for some selected bins. In Figure 3.5a the
shapes are relative to the average spectra, which was forced to have a constant log am-
plitude of 1; numbers to the left of each spectra indicate the number of earthquakes in
that particular bin. Figure 3.5b shows the corrected source shapes, after applying the
smallest magnitude bin as an empirical Green’s function (EGF). To avoid passing un-
wanted line components present in the smallest bin, we smoothed the reference spectrum
with a 20 point moving average before subtracting it from the data.

1 Hz). These moments are not evenly spaced in Figure 3.5 because the moments of
the earthquakes within each bin are not always evenly distributed. We use a bootstrap
technique that randomly resamples the earthquakes within each source spectral bin in
order to estimate uncertainties on the binned spectra, and later to estimate uncertainties
on properties, such as corner frequency and energy, that we compute from these spectra.
As in the case of the path-site terms, the absolute shape of the spectra plotted
in Figure 3.5a is unconstrained (owing to the intrinsic tradeoff between the average
source spectrum and the average path-site response function). We resolve this tradeoff
in our iterative method by forcing the average source spectrum to unity. This is why the
spectra for the small earthquake bins curve upward at high-frequency. This indicates
that, as expected, these earthquakes have a shallower falloff at high frequencies compared
 51

to larger earthquakes. To obtain an estimate of the true spectral shapes of the sources,
we use the smallest moment bin as an empirical Green’s function (EGF) for all the other
bins.
Figure 3.5b shows the results of subtracting the log spectra of the smallest bin
from the others. These EGF corrected spectra have the features expected for source
spectra—a flat response out to a corner frequency that increases with decreasing earth-
quake size, and a rapid falloff beyond the corner frequency. As we will discuss later this
falloff closely agrees with the Brune ω −2 source model (Brune, 1970). We plot all of
the EGF corrected spectra in Figure 3.5, but our later analyses will focus only on those
bins at least one order of magnitude larger in moment than the EGF reference bin. For
comparison, Mori et al. (2003) used a M ∼ 1.5 smaller EGF and Frankel et al. (1986)
used earthquakes with M ≤ 2.1 as EGF of M ∼ 3 earthquakes. As previously noted, due
to signal-to-noise limitations, we do not use P -wave data from the smallest earthquakes
so the smallest P -wave EGF bin represents the same bin as the third smallest S-wave
EGF bin.

3.3 Implications of Self-Similarity

Before further analysis, it is instructive to consider the predicted effects of

earthquake self-similarity on recorded spectra (e.g., Aki, 1967). Figure 3.6 illustrates
the expected change in the pulse shape and spectrum for an earthquake rupture that is
increased in size by a factor b. Assuming the dimensions of the larger rupture are scaled
proportionally, then the fault area, A, will increase by a factor b2 , the displacement,
D, will increase by b, and the moment, M0 = µDA, will increase by a factor of b3 .
Figure 3.6b shows the resulting change in a displacement pulse, u(t), recorded in the far
field, assuming identical source and receiver locations and no attenuation. The exact
form of the shape of this pulse depends upon details of the source, but, assuming simple
scaling between the two earthquakes, the pulse shape will change in a predictable way. In
particular, assuming the rupture speed is constant between the earthquakes (as simple
self-similarity predicts), the pulse length will increase by a factor of b and the pulse
height will increase by a factor of b2 . This is necessary in order for the moment, which
 52

is proportional to the integrated area under the pulse, to increase by b3 .

It follows that the displacement pulse, u∗ , recorded by the second earthquake
can be expressed as
u∗ (t) = b2 u(t/b) (3.3)

where u(t) is the recorded displacement pulse of the first earthquake. The seismic energy,
Es , in the recorded pulse will be proportional to u̇2 (t) dt (the integrated square of the
R

slope of the pulse), so the second pulse will contain a factor b3 more energy than the
first pulse. Thus the energy density (Es /M0 ) remains constant.
Using the similarity theorem for the Fourier transform, it follows that the spec-
trum of the second earthquake is given by

u∗ (ω) = b3 u(bω) (3.4)

where u(ω) is the spectrum of the first earthquake. This relationship predicts that the
shape of all spectra on a log-log plot will be identical, but offset along a line of ω −3
(Figure 3.6c).
This provides a possible test of self-similarity that does not depend upon any
assumptions regarding which source model is most appropriate (ω −2 , ω −3 , etc.). We
perform this test (Figure 3.7) by shifting the EGF corrected spectra along an ω −3 line and
find that the shapes are in agreement within their estimated uncertainties. Furthermore,
there is no systematic dependence with moment exhibited in the alignment of the binned
spectra (Figures 3.7c and 3.7d). The P -wave spectra do not align as closely as the S-wave
spectra at low frequencies (≤ 1 Hz) because the individual P -wave stacks are not flat at
low frequencies (Figure 3.7a). Although we do not fully understand the reason for this
behavior, the shapes of the P spectra are nonetheless similar within their uncertainties.
It is likely that this anomaly in the P -wave spectra is related to decreasing signal-to-noise
ratios at low frequencies, which could bias the EGF reference stack because it is derived
from the smallest earthquakes.
The S-wave spectra are noticeably smoother and provide our most reliable
constraints on the similarity of the spectra as a function of moment. This is the most
fundamental result in our study and suggests the self-similarity hypothesis is valid for
our data set. The great advantage of this analysis is that we can check if spectral shapes
 53

(a)

x b
A = area A* = b2 A
M0 = μDA M0*= b3M0

(b)
u*(t)
u(t)
x b2

x b

(c)
ω−3
u*(ω)
log(u(ω))

3 log(b)

u(ω)

log(b)

log(ω)

Figure 3.6: An illustration of the effects of self-similarity when an earthquake is in-

creased in size by a factor of b. (a) The rupture area increases by b2 , the displacement
by b and the moment and energy by b3 . (b) A recorded far-field displacement pulse will
increase in length by b and in height by b2 . (c) Log-log plots of the spectra will have
identical shapes, but shifted along an ω −3 line.
 54

are self-similar or if there are systematic differences in the shapes as magnitude increases,
without assuming a particular model of corner frequency and high-frequency falloff. In
contrast, conventional methods for making inferences about source scaling are heavily
focused on parametric data derived from the spectra rather than the spectra themselves.
These parameters do, however, provide further insight regarding source properties.
Implicit in our spectral comparisons is that the focal mechanisms and rupture
directions do not vary systematically between smaller and larger earthquakes because
this could bias the results obtained at particular stations. We have not examined the
focal mechanisms for our earthquakes but have no evidence that this is the case. Such
problems are likely to be minimized in our analysis because we are averaging results from
many stations at different azimuths and distances from the earthquakes. Furthermore
it seems unlikely that these possible biases would have the effect of producing appar-
ent self-similarity in our measured spectra without self-similarity being present in the
earthquakes themselves.

3.4 Source Parameter Modeling

The source parameters seismic moment (M0 ), corner frequency (fc ) and radi-
ated energy (Es ) can be estimated from the source spectra and are important in the
understanding of the physics of the earthquake source, as well as for computing appar-
ent stress (σa ), defined as µEs /M0 , where µ is the rigidity. If self-similarity holds, as
tested in the previous section, this ratio should remain constant over the same range of
magnitudes.
We initially fit our P and S displacement spectral stacks with a general source
model (e.g., Abercrombie, 1995):

Ω0
u(f ) = h i1/γ (3.5)
f
γn
1+ fc

where Ω0 is the long-period amplitude (relative seismic moment), f is the frequency,

fc is the corner frequency, n is the high-frequency falloff rate and γ is a constant. We
allowed the values of n to vary while using γ=1 as well as γ=2, that is we experimented
with both the Brune (1970) and Boatwright (1980) models, allowing the falloff term to
 55

4 5
10 10
P spectra a P spectra c

ω
-3
3 4
10 10
Spectral Ratio

2 3
10 10

1 2
10 10

0 1
10 10
4 5
10 10
S spectra b S spectra d
ω
-3

3 4
10 10
Spectral Ratio

2 3
10 10

1 2
10 10

0 1
10 10
1 10 80 1 10 80
Freq (Hz) Freq (Hz)

Figure 3.7: EGF corrected stacked spectra for bins of different source moment, showing
the self-similarity of the spectra when shifted along an ω −3 line. (a) P -wave spectra and
(b) S-wave spectra, with 1-σ error bars estimated using a bootstrap resampling method.
The spectra shifted along an ω −3 line (dashed lines at left) for (c) P -waves and (d)
S-waves. The spectra agree in shape within their estimated errors, consistent with the
earthquake self-similarity hypothesis.
 56

vary as well as corner frequency and relative seismic moment. We used a grid search
technique to find the best-fitting set parameters (Ω0 , fc , n, and γ). We restricted this
procedure to those size bins that have relative moments ten times larger than that of
the EGF bin (see Figure 3.5).
In general we found that a simple ω −n model (i.e., γ = 1) worked reasonably
well with values of n ranging from 1.8 to 2.2 (i.e., very close to the Brune ω −2 model), and
that allowing additional free parameters did not significantly improve the fit. Predictions
obtained using γ = 2 yielded spectra with sharper corners than are seen in our stacked
spectra. It is possible that individual events have spectra with these sharp corners, but,
given some variability in the positions of the corners, the corner is smoothed and widened
in the stacks over many events so that the Brune model gives the best fit. We therefore
used the model
Ω0
u(f ) = (3.6)
1 + (f /fc )n
and solved for the best-fitting Ω0 , fc and n for the results presented here (see Figure 3.8
for examples of the resulting fits to the stacked spectra).
The radiated seismic energy is proportional to u̇2 (t) dt, the integrated square
R

of the measured velocity. We perform this calculation in the frequency domain by con-
verting the displacement spectra to velocity, squaring and integrating (e.g., following Ide
and Beroza, 2001), being careful to extrapolate to very high frequencies using the model
falloff rate. In this study we compute energy from the best fitting model rather than
directly from the data, i.e., we use
Z∞  2
2πf Ω0
I= df (3.7)
1 + (f /fc )n
0

where I is the relative seismic energy. Because we are integrating the model predicted
spectrum rather than the data, we can extend the upper integration limit to a sufficiently
high frequency to avoid any underestimation of the energy.

3.5 Calibration to absolute moment and energy

Our results described so far involve only relative estimates of moment and
seismic energy. To obtain absolute measures of these parameters directly from our data,
 57

4
10
3 a b
10

3
10

Spectral Ratio
2
10
2
10

1
10 1
10

1 10 80 1 10 80
Freq (Hz) Freq (Hz)

Figure 3.8: EGF corrected stacked spectra and best-fitting source models for (a) P -
waves and (b) S-waves. For clarity only some of the moment bins are plotted.

we would need to apply corrections for geometrical spreading, radiation pattern, free-
surface, and source-receiver impedance contrast effects. Because these corrections are
often difficult to estimate precisely, this will introduce considerable uncertainty into
our results. However, because our earthquakes are in a single compact region, these
correction factors are likely to be highly correlated, implying that relative measures of
moment and energy among our earthquakes are determined more accurately than their
absolute level. Thus, our most precise results involve relative measurements among
our earthquakes. However, for comparisons to other studies it is useful to have some
measure of absolute moment and energy. Our approach to this problem is to exploit
the fact that these earthquakes were also recorded by the Southern California Seismic
Network (SCSN), which provides well-calibrated local magnitude estimates (moment is
not routinely computed for earthquakes this small).
Assuming that ML ≈ MW , we can estimate moment M0 using the Kanamori
(1977) relation
MW = (2/3) log10 M0 − 10.7 (3.8)

In this way, we can compute a scaling factor to relate our relative moment estimates Ω0
to local magnitude and to true moment M0 . A comparison between SCSN mean catalog
ML versus our estimated ML (Figure 3.9) shows a linear relationship with a slope close
to unity, as expected if the 2/3 factor in (3.8) is accurate. Previous studies have shown
 58

3.5

Catalogue local magnitude

3

2.5

1.5

1
1 1.5 2 2.5 3 3.5 4
Estimated local magnitude

Figure 3.9: A comparison between ML as measured by the southern California Seismic

Network (SCSN) and ML as estimated from our relative moment measures using an
empirical scaling factor.

that in the magnitude range of our data set, a linear relation between log(M0 ) and MW
fits the data in Southern California, although with some variations on the 2/3 factor (see
Hanks and Boore, 1984; Abercrombie, 1996, for more detailed discussion). A change in
the scaling factor would change the absolute moment magnitude after calibration, but
the relative moment between the different earthquake bins should remain constant.
Now consider the theoretical relationships for M0 and Es for a double-couple
source in the far field in a uniform wholespace. The standard formula (e.g., Aki and
Richards, 1980; Kanamori and Rivera, 2004) for the moment in this case is

−1
M0 = 4πρc3 r c Uφθ Ω0 (3.9)

where ρ is the density, c is the seismic velocity (either α for P wave or β for S wave),
r is the source-receiver distance, c Uφθ is the radiation pattern, and Ω0 is the observed
long-period amplitude. Now assume that we know M0 , ρ and c independently. We can
then rewrite (3.9) as
c Uφθ Ω0
= 4πρc3 (3.10)
r M0
Note that 1/r represents a geometrical spreading term that could be generalized to a
more complicated model.
 59

For the same whole-space double-couple model, the radiated seismic energy
may be expressed as (e.g., Boatwright and Fletcher, 1984)
hc Uφθ 2 i
Esc = 4πρcr2 2 I (3.11)
c Uφθ

where hc Uφθ 2 i is the mean over the focal sphere of (c Uφθ )2 (= 4/15 for P waves and 2/5
for S waves) and I is the measured relative energy (i.e., the integrated velocity squared).
Because this equation involves the ratio of c Uφθ and r we can use (3.10) to obtain
hc Uθφ 2 i 2 I
Esc = M (3.12)
4πρc5 0 Ω20
which is independent of the geometrical spreading and radiation pattern terms. This
equation remains accurate if free-surface corrections are applied or if the instrument
gain is incorrectly known, provided M0 is determined independently. In the case where
ρ and c vary between source and receiver, carrying through the impedance correction
terms shows that (3.12) is still valid provided ρ and c are taken at the source.
Because the estimated energy varies inversely as c5 , the results are very sensitive
to errors in velocity at the source. A 15% error in c will produce about a factor of two
√
error in Es . In this study we use α = 6.0 km/s, β = α/ 3, and ρ = 2.7 kg/m3 , which
leads to the value µ = 3.24 × 1010 Pa. The values of velocity are very close to those from
a 3-D seismic velocity inversion (Scott et al., 1994) for the source region. We estimate
the uncertainty in our source velocity estimates to be less than 5%.
The total radiated seismic energy is obtained by adding the energy for P and
S waves
Es = EsP + EsS (3.13)

Finally it is important to recognize that absolute energy estimates are also very sensi-
tive to attenuation corrections. We assume here that the EGF approach has correctly
removed attenuation effects, but this remains another possible source of uncertainty in
our results.

3.6 Results for corner frequency and apparent stress

To compare P and S corner frequencies, we performed a separate analysis in

which the relative moment of each earthquake was estimated from both the P and S
 60

spectra so that the same earthquakes would be contained in each moment bin. As
discussed in the previous section, this relative moment will later be calibrated with esti-
mates of moment from local magnitude determinations. We find that the P wave corner
frequencies determined here are systematically higher than those estimated for S waves
from the same earthquakes (Figure 3.10). The ratio fc (P )/fc (S) is about 1.6 (individual
measurements range from 1.3 to 2.0), consistent with the model of Madariaga (1976) and
very close to values determined using borehole recordings at 2.5-km depth in the Cajon
Pass, California by Abercrombie (1995). This ratio is likely to correspond principally to
source effects since attenuation and other path-site effects have been removed.
As analyzed by Abercrombie (1995) the ratio of S to P wave energy (known as
q) is also very important. From Boatwright and Fletcher (1984) we have

3 α 5 fc (S) 3
   
q= (3.14)
2 β fc (P )

where fc (S) and fc (P ) are the corner frequencies for S and P (which are assumed to
have the same falloff rate at high frequencies). Note that q = 23.4 for a Poisson solid if
the corner frequencies are identical. In our study, fc (P ) is about 1.6 times larger than
fc (S), reducing the predicted value of q to about 6.
Our estimated P and S energies (Figure 3.11), calculated using equation (3.12)
for the different moment bins, which yield q = 9 ± 1.5, the difference from the predicted
value (q = 6) resulting from the fact that our models permit the falloff exponent to vary
slightly between P and S waves. Previous studies have found considerable variation in q
estimates, as they are highly dependent upon corner frequency shifts, but our results are
in reasonable agreement with, for example, Boatwright and Fletcher (1984) (q = 13.7
± 7.3) and Abercrombie (1995) (q = 14.31 with values from 4.43 to 46.26). We did not
directly obtain P -wave energy for the two smallest spectral bins because their relative
moment was not ten times larger than the P -wave EGF. To obtain P energy for these
bins, we divided the S-wave energy by the q = 9 scaling parameter estimated from the
other bins.
Another commonly applied test of self-similarity (e.g., Abercrombie, 1995; Ide
et al., 2003; Kanamori and Rivera, 2004) is to plot corner frequency versus seismic
moment. As previously discussed, self-similarity predicts that M0 ∝ fc−3 . We determined
 61

100

P spectra corner frequency (Hz)

10

2.0
f cS 1.5
/
f cP 1.0

1
1 10 100
S spectra corner frequency (Hz)

Figure 3.10: A comparison between P and S-wave corner frequencies as measured for
the different moment bins. Dashed lines represent different scaling factors. The data
suggest fc (P ) ≈ 1.6fc (S).

11
10

10
10
S Wave Energy (J)

9
10

8
10

7
10

ES = 9.1 E P
6
10 5 6 7 8 9 10
10 10 10 10 10 10
P Wave Energy (J)

Figure 3.11: Radiated S-wave energy EsS versus P -wave energy EsP estimated from the
different moment bins. The best fitting line is for EsS = 9EsP .
 62

Mw
2 1.5 2 2.5 3 3.5
10
a

fc (Hz)
1
10

P waves
0
10
2
10
b
fc (Hz)

1
10

S waves
0
10 11 12 13 14 15
10 10 10 10 10
Seismic Moment (N m)

Figure 3.12: P and S corner frequencies versus moment, derived from the stacked
spectra for the different moment bins. The results indicate M0 ∝ fc−3 as shown by the
dashed lines.

the relative seismic moment and corner frequencies for P and S waves independently
(see Figure 3.12). The relative moment is scaled to obtain an approximation of the
absolute seismic moment (see section 3.5). Corner frequencies follow the cube root scaling
expected from self-similarity, as previously observed by Ide et al. (2003). Of course this is
not surprising, given that the spectra themselves obey self-similarity scaling (see Figure
3.7 and prior discussion). Due to the corner frequency shift for P and S waves, we plot
this relationship independently. It is possible that the 40 Hz and higher corner frequencies
for the smaller moment bins are constrained less accurately than the corner frequencies
for the larger earthquakes because our analysis extends only to 80 Hz. However, there are
more earthquakes in the smaller moment bins, resulting in smoother stacked spectra (see
Figure 3.7 and 3.8), which likely improves the reliability of the corner frequency estimates
even when less of the spectrum is available. It is clear from Figures 3.7 and 3.12 that
the available part of the spectra are consistent with the self-similarity hypothesis.
The relationship between seismic moment and radiated seismic energy is also
important and has been a focus of many previous studies. This relationship is commonly
expressed in terms of apparent stress, defined as σa = µEs /M0 . Figure 3.13 shows
apparent stress plotted as a function of moment for our spectral stacks. The Es /M0
 63

Mw
1.5 2 2.5 3 3.5
10

σa (MPa)
1

0.1 11 12 13 14 15
10 10 10 10 10
Seismic Moment (N m)

Figure 3.13: Apparent stress σa versus moment for the different moment bins. Appar-
ent stress is nearly constant over MW = 1.8 to 3.4, with an average value of about 1
MPa.

ratio is approximately constant as moment increases, as predicted if self-similarity is

obeyed and apparent stress is constant as a function of earthquake size. Given the
scatter in our data, a small degree of scaling is possible. A weighted least squares fit to
the points in Figure 3.13 results in Es /M0 ∝ M00.08±0.10 , providing relatively tight error
bounds that include the zero exponent result expected from self-similarity.
The average apparent stress is about 1 MPa, but for the reasons discussed above,
this value is less well constrained than the relative σa between our different earthquake
bins. The two largest sources of error in our absolute σa estimate are likely to be: (1) our
calibration factor between relative moment and MW , and (2) the assumed S velocity at
the source. Any calibration factor error will scale directly as M0 . From (3.8) we see that
if, for example, our MW estimates (assumed equal to the SCSN ML values) are 0.2 units
too large, this will result in σa estimates that are about two times too large. A 5% error
in our assumed source S velocity will yield about 30% uncertainty in σa . Given these
uncertainties and the scatter shown in Figure 3.13, a reasonable range for the possible
values of the average apparent stress is 0.3 to 3.0 MPa.

3.7 Discussion

Our study indicates that self-similarity of the earthquake source is consistent

with data from over 400 small earthquakes in our study region, as shown by the scaling of
 64

source parameters such as corner frequencies and apparent stress as well as the similar-
ity in the shapes of the source spectra themselves, independent of any particular source
model. This conclusion is based on stacks of earthquake spectra in bins of similar seismic
moment, a process that averages the properties of earthquakes in these bins. Spectra
of individual earthquakes may also be obtained using our technique; these show much
greater variability in corner frequency and apparent stress but their average properties
are consistent with the results presented here. Although we do not take into account
possible biasing effects, such as systematic changes in focal mechanism or rupture direc-
tivity, it is likely that these effects are minimized by averaging over stations at different
distances and azimuths from the source region.
Our study supports models in which the average apparent stress is constant as a
function of earthquake size, as suggested by Ide et al. (2003) and others. Our results are
limited by the small magnitude range spanned by our earthquakes (1.8 ≤ M ≤ 3.4 for
the EGF corrected data), but have sufficiently low scatter that fairly tight constraints
can be placed on any possible moment dependence of apparent stress. Mayeda and
1/4
Walter (1996) proposed that Es /M0 is proportional to M0 over the magnitude range
3.3 ≤ M ≤ 7.3, consistent with the suggestion of Abercrombie (1995) that apparent
stress appears to increase gradually with moment over a magnitude range from 0 to 7.
Such a strong dependence on M0 is not supported by our results over the limited size
range of our data (our best fitting scaling is M00.08±0.10 ). Comparisons with other studies
can extend the applicability of our results. Our estimated average apparent stress of 1
MPa is above most of the estimates of Abercrombie (1995) for similar size earthquakes
(i.e., MW = 1.8 to 3.4) and is consistent with the suggestion of Ide and Beroza (2001)
that apparent stress has a nearly constant value of 1 MPa over the entire observed range
of earthquake sizes.
A large number of studies have suggested that the source spectra might have
more complex behavior than that estimated from simple corner frequency models (e.g.,
Singh and Ordaz, 1994; Mayeda and Walter, 1996) and should include intermediate
falloffs. Differences in the results obtained in different studies might be due to model
assumptions that depend upon parametric data derived from the spectra rather than
the spectra themselves. An advantage of our approach is that we can directly use the
 65

shapes of the spectra to test for self-similarity without any source model assumptions.
A source of concern for our parametric analysis is whether the maximum frequency of 80
Hz that we use in our study is affecting our results, especially for estimates of the corner
frequency for the smallest earthquakes. This does not appear to be a problem because
we observe no saturation of the corner frequencies for the small events (see Figure 3.12).
The values of apparent stress that we obtain have much less scatter than those
seen in most previous studies, probably because of the averaging that we perform within
each moment bin. Thus, although our study spans a quite limited magnitude range, our
nearly constant values of apparent stress place fairly tight constraints on the amount
of any scaling with moment that could be present within our data. Recently Mayeda
et al. (2004) have argued that a potential problem exists in comparing apparent stress
for events over a broad region because the regional scatter of the estimates could make
resolving scaling variations problematic. Also, some of the trends of previous studies
might be masking (or exposing) the true trend, because of the large range of apparent
stress uncertainties. Our study has the advantage of being restricted to a specific source
region and of averaging over a large number of earthquakes, reducing the scatter and
likely biases in our apparent stress estimates.
Our results are limited to the cluster of earthquakes in our study region but
the spectral stacking method should readily be applicable to other data sets. In par-
ticular, it would be useful to study clusters or aftershock sequences that contain larger
earthquakes to extend the magnitude range. There are a number of possible candidates
in southern California for such an analysis, including the Northridge and Landers after-
shock sequences. In addition, studies of large numbers of distributed earthquakes, as
recorded by local and regional seismic networks, might reveal spatial patterns in source
properties. In this case, corrections for attenuation effects will be more complicated than
when the earthquakes are restricted to a single cluster, but in principle attenuation and
source effects can be still be separated using a spectral stacking approach.

Acknowledgments

We thank the personnel of IGPP and the Anza Seismic group at UCSD who
recorded, picked and archived the seismograms. Rachel Abercrombie, Greg Beroza, Jack
 66

Boatwright and Arthur McGarr provided useful comments and reviews. Funding for this
research was provided by NEHRP/USGS grants 03HQPA0001 and 01HQAG0021. This
research was also supported by the Southern California Earthquake Center. SCEC is
funded by NSF Cooperative Agreement EAR-0106924 and USGS Cooperative Agreement
02HQAG0008. The SCEC contribution number for this paper is 766.
4

Uncertainties in earthquake
source spectrum estimation using
empirical Green functions

We analyze the problem of reliably estimating uncertainties of the earthquake

source spectrum and related source parameters using Empirical Green Functions (EGF).
We take advantage of the large dataset available from 10 seismic stations at hypocentral
distances (10 km < d < 50 km) to average spectral ratios of the 2001 M5.1 Anza earth-
quake and 160 nearby aftershocks. We estimate the uncertainty of the average source
spectrum of the M5.1 target earthquake by performing propagation of errors, which, due
to the large number of EGFs used, is significantly smaller than that obtained using a
single EGF. Our approach provides estimates of both the earthquake source spectrum
and its uncertainties, plus confidence intervals on related source parameters such as radi-
ated seismic energy or apparent stress, allowing the assessment of statistical significance.
This is of paramount importance when comparing different sized earthquakes and ana-
lyzing source scaling of the earthquake rupture process. Our best estimate of radiated
energy for the target earthquake is 1.24 × 1011 Joules with 95% confidence intervals
(0.73 × 1011 , 2.28 × 1011 ). The estimated apparent stress of 0.33 (0.19, 0.59) MPa is
relatively low compared to previous estimates from smaller earthquakes (1MPa) in the
same region.

67
 68

4.1 Introduction

A fundamental problem in seismology is accurate estimation of the radiated

seismic energy of an earthquake. When an earthquake occurs, some fraction of the
total energy is radiated as seismic waves, providing important information about the
earthquake rupture process. Determining the amount of radiated seismic energy can
be difficult and large differences in these estimates have been found among different
techniques and groups (Pérez-Campos et al., 2003).
Partly the difficulty lies in the need to estimate radiated seismic energy over
a large dynamic range and a wide frequency band, from the low frequencies needed
to define the seismic moment to well beyond the corner frequency. Another difficulty
comes from the need for path and site corrections; these corrections have much larger
uncertainties at higher frequencies. Finally, the earthquake source might have more
complicated features than first expected, including directivity effects that need to be
taken into account in order to avoid biasing the estimation.
Over the last 15 years there have been many studies of the radiated seismic
energy of different sized earthquakes in different regions of the Earth (e.g., Kanamori
et al., 1993; Abercrombie, 1995; Choy and Boatwright, 1995; Mayeda and Walter, 1996;
Ide and Beroza, 2001; Venkataraman et al., 2002; Mori et al., 2003; Prieto et al., 2004,
and many others). What is often missing in these studies is a measure of the uncertainty
of each individual estimate. The question of the uncertainty of estimates is of key
importance for describing the significance of one measurement compared to another.
In particular, how do the errors in the assumed attenuation model propagate into the
uncertainty of the source spectrum?
The major unknown in the system is the transfer function between source and
receiver. We will focus in this paper on developing a technique to use Empirical Green
Functions and estimating and reducing the uncertainties by averaging over a set of suit-
able aftershocks. By using many EGFs we effectively are randomly sampling the errors
in the path effects and averaging over the propagation space. As an example we will
present results for the October 31 2001 M5.1 Anza earthquake, using data from local
stations.
 69

33.6

Target EQ
34.0 DEV M5.1

PLC MGE

PFO AGA
DGR 33.5

33.5 DNR

PLM
BOR
0
12.5
25
0
2.5
5

km
km

JCS
33.0 33.4
-117.0 -116.5 -116.0 -116.6 -116.5 -116.4

Figure 4.1: Maps of the study area. The left map covers the entire study region,
showing the seismic stations (black triangles), the M 5.1 target Anza earthquake (black
circle) and aftershocks M2.9 and lower (gray circles). The inset shows the location of
the study area and the state of California. A close-up region (dashed box) is shown in
the right hand map. The target earthquake is shown as an open gray circle of radius
proportional to a 1MPa stress drop event. Aftershocks used in this study (black circles)
and general aftershock seismicity of the region (gray circles) are shown.

4.2 Data Processing

A ML 5.1 earthquake occurred on 31 October 2001 in the Anza region in south-

ern California (hypocenter 33.5081◦ N, 116.5143◦ W, depth 15.2km, MW 4.7) The earth-
quake exhibited thrust motion on a vertical fault striking N35◦ E (Hauksson et al., 2002).
The data comes from Anza broadband velocity sensors and IDA strong motion sensors
at station PFO and both broadband and strong motion sensors from TriNet.
All stations are within 10 to 50 km from the source region. The choice of using
local stations to obtain the source spectrum is motivated by the need to obtain reliable
estimates of the spectrum at high frequencies and also the need to have good signal-to-
noise ratios for the smaller aftershocks. The epicentral distance between the mainshock
and all the aftershocks used is less than 2 km (see map in Figure 4.1)
Since S-P times for the closer stations are small, we take a 12 second win-
dow that includes both P and S waves, starting 1 second before the P wave pick (e.g.,
Venkataraman et al., 2002). We also select a noise window of the same length, immedi-
 70

ately preceding the signal window. A signal-to-noise ratio (SNR) is obtained by the ratio
of the signal and noise spectra. The spectra of the waveforms are estimated using the
multitaper technique (Thomson, 1982), which not only allows a variance reduction of the
spectra within a certain bandwidth, but also provides an estimate of the uncertainties
at each frequency bin using a jackknife approach (Vernon, 1989; Thomson and Chave,
1991). For each waveform at each station we obtain the spectrum and the 5 − 95%
confidence interval.
Since we are interested in the source spectrum of the M 5.1 earthquake, we
perform spectral division between the mainshock and all the aftershocks recorded at
each station. Following the theory of propagation of errors (Taylor, 1997), dividing two
random variables results in a new random variable Ri (f ) (the spectral ratio), where the
i term represents the ith aftershock used for deconvolution, with uncertainties being a
function of the uncertainties of the spectra of the mainshock and aftershock. As shown in
Figure 4.2, the relative uncertainties in R(f ) are larger than the individual components
in X(f ) and Y (f ), due to the instability of the deconvolution process, or simply because
we are dividing two noisy spectra. The gray shaded area is larger in Figure 4.2 for the
spectral ratio.
As noted by Aki and Richards (1980) (Volume II, Chapter 11.5.5) simple spec-
tral density estimates have only two degrees of freedom and are distributed as χ22 ; con-
sequently, their ratio is distributed statistically as F2,2 , a distribution so broad it has
infinite variance. We improve on this situation by making use of multitapers: each spec-
tral estimate is made by weighted average over 7 tapers (and time-bandwidth product
4). Then the number of degrees of freedom increases to about 10 (less than the expected
14 because of the weighting).
We take the log of the spectral ratios and stack our data in 15 bins divided by the
magnitude of the EGF used (∆ML = 0.2). Given that variations in spectra are generally
observed to be log-normally distributed, throughout the paper we will average and stack
spectra in the log domain, which is equivalent to using a geometric average rather than
the arithmetic mean. This also ensures that low amplitude spectra contribute equally to
average spectral shapes compared to high amplitude spectra. Figure 4.3 shows a set of
typical spectral ratios Ri (f ) for EGFs with a range of earthquake magnitudes. Note the
 71

12

10
R(f)
8

log Amplitude
6
Y(f)
4

2
X(f)
0

0.1 1 10
Freq (Hz)

Figure 4.2: Estimates of the spectrum for ground motion associated with the M5.1
earthquake Y (f ) at station AGA and the spectrum of the largest aftershock M 2.9 X(f )
at the same station and corresponding spectral ratio R(f ) = Y (f )/X(f ) (offset for
comparison purposes). The gray area represents the 95% confidence interval estimated
using the jackknife approach and propagation of errors. Note how uncertainties grow on
R(f ) after spectral division.

effect of the corner frequency of the larger aftershocks and the effect of the low SNR at
low frequencies for the smaller aftershocks. The uncertainties (not shown in Figure 4.3)
for each particular bin are again estimated using the method of propagation of errors
(Taylor, 1997). For each station a scaling factor for the spectral ratios is determined
over a narrow band where the shapes are consistent (gray area in Figure 4.3).

4.3 The Combined Empirical Green Function

In order to take advantage of the large wealth of data available from the local
stations, including very small aftershocks (M 0–0.5), we will average together the dif-
 72

0.0 < M < 0.2

4 Station AGA

Log Amplitude
2

1
2.8 < M < 3.0

0.1 1 10
Freq (Hz)

Figure 4.3: Selected bins of spectral ratios (∆ML = 0.2) at station AGA. The effect of
the corner frequency of the 2.8 < M < 3.0 earthquakes (lower line) flattens the spectral
ratio at high frequencies, while the low SNR affects lower frequencies of the smaller
EGF s, but gives reasonable spectral fall-off at high frequencies. The gray shaded area
shows the band used to estimate a scaling factor for combining the spectral ratios.

ferent estimates of spectral ratios. This will enable us to bring down the variance or
uncertainties of the source spectrum.
The spectral ratio R(f ) contains two sources of error, (a) the variance due to the
intrinsic spectral estimation process, which we take from the multitaper algorithm and
(b) systematic errors (which we will call bias) due to the effect of the corner frequency of
the EGF. As described by Ide et al. (2003), the spectral ratio R(f ) of two earthquakes
located close to each other, assuming the same focal mechanisms and path effects, can
be expressed as
Y (f ) S(f )
R(f ) = = (4.1)
X(f ) G(f )
where Y (f ), X(f ) represent the spectrum of the ground motions of the mainshock and
aftershock and S(f ), G(f ) represent the source spectrum of the mainshock and the
 73

aftershock respectively.
An approximate form of the source spectrum (Brune, 1970) is
M0
G(f ) = (4.2)
1 + (f /fc )2
where M0 is the seismic moment and fc is the corner frequency of the EGF. An ideal
case for an EGF would be to use an earthquake whose corner frequency fc was very
large, giving then in the log domain

log R(f ) = log S(f ) − log M0 (4.3)

resulting in a scaled version of the source spectrum, without changing its shape.
The bias of the log spectrum that is created by a finite corner frequency is then

b(f ) = log(1 + (f /fc )2 ) (4.4)

which clearly shows that as the frequency f grows and approaches the corner frequency
of the EGF, the bias of the source spectrum increases.
We do not know the true corner frequency of the EGF and the relatively low
sampling rate of the stations in the network (100 sps) is not enough to estimate it from
spectral ratios (as in Hough, 1997; Ide et al., 2003). Instead, we use a simple scaling
relation to get an approximate corner frequency. We assume that fc can be approximated
(Venkataraman et al., 2002) from fc = 0.49β(∆τ /M0 )1/3 where ∆τ is stress drop, and
we use a value of 1 MPa. This choice of stress drop is rather arbitrary but certainly
within the average in southern California and the study region (Vernon, 1989; Shearer
et al., 2006). We chose a rather low stress drop, as a conservative value to obtain small
fc ’s for the EGFs and limit the potential bias at high frequencies. We estimate the
seismic moment following the procedure from Prieto et al. (2004), that is, we assume
local magnitude ML = Mw for the small earthquakes and use the Kanamori (1977)
relation to obtain M0 . One could argue that this approximation is not accurate, but
as it turns out, even allowing the corner frequency of the EGF to change by 20 − 30%
does not substantially affect the results, changing the radiated energy estimate of the
mainshock by less than 3% in our example.
A common technique for dealing with estimates that contain variance and bias
as sources of error is the mean-square error (MSE) (Rice, 1995). At a given station we
 74

construct the source spectrum of the target earthquake S(f ) from a linear combination
of the spectral ratios
N
X
log S(f ) = wi (f ) log Ri (f ) (4.5)
i=1

where the index i in the sum runs over the events, wi are the weights for each spectral
ratio, and N is the number of EGF available at a particular station. As explained earlier,
the idea is to create a weighted average of the spectral ratios.
The MSE is the sum of the variance and the bias squared of the estimate.
Applying this idea to our linear combination of spectra

N N
!2 N
X X X
mse2 = wi2 σi2 + wi bi −λ wi (4.6)
i=1 i=1 i=1

where the first term represents the variance (σi2 ), the second term is the bias squared
and we added a normalization constraint as a Lagrange multiplier λ.
Taking the derivative of (4.6) with respect to the unknowns wi and λ and
minimizing, we obtain two sets of linear equations
N
X  2 
σi δij + bi bj wj − λ = 0 (4.7)
j=1

N
X
wj = 1 (4.8)
j=1

We solve the linear set of equations for each individual frequency f with a non-
negative least squares approach (see Lawson and Hanson, 1974), to obtain only positive
weights. We find that requiring a SNR of 5 or larger for a particular Ri (f ) leads to
better results. This approach will discard most of the spectra below 0.8 Hz due to low
SNR for the smaller magnitude EGFs, and will only use the set of larger EGFs. At
higher frequencies more spectral ratios have good SNR, but the larger earthquakes have
either larger bias terms or the variance is much larger than for the smaller EGF (since
the network records many small aftershocks, thus the variance of a particular spectral
ratio bin is decreased), so that the weights prefer the smaller EGF. This means we keep
the part of each spectrum that has good SNR and discard only the frequency points with
low SNR.
 75

Figure 4.3 shows Ri (f ) for a range of EGF earthquake magnitudes at station

AGA. We scale the spectral ratios before combining the estimates in (4.5), since the
absolute amplitudes depend on the seismic moment of the EGF used. The scaling factor
is obtained from a narrow band (shaded area in Figure 4.3) where the spectral ratios are
consistent. We tested the effect of the scaling factor by changing the band width used for
shifting the spectra. Because the resultant energy estimate does not vary significantly,
we choose to ignore this effect.
Solving for the weights in (4.6) and applying the result in (4.5) we obtain a
relative source spectrum for each station in the network. Following Prieto et al. (2004)
we calibrate the relative source spectrum at each station to the seismic moment of the
target earthquake. We express the mean source spectrum by averaging over the different
seismic stations.
K
1 X
log S(f ) = log Sk (f ) (4.9)
K
k=1
where K are the number of stations and Sk is the outcome of (4.5) for station k. The
processing and averaging is done one frequency at a time, at individual stations, and the
mean source spectrum S represents a station average.
Figure 4.4 shows the source spectrum S(f ) for our target event, with 5 − 95%
confidence intervals at each frequency point. Note how for the lower frequencies the
uncertainties are larger, since the weights are non-zero only for the larger aftershocks,
while at higher frequencies many events are available, an average over many independent
estimates, increasing the number of degrees of freedom and decreasing the variance.

4.4 Results for radiated seismic energy

From the mean source spectrum S(f ) and the uncertainties, it is possible now
to obtain an estimate of the radiated seismic energy (Vossiliou and Kanamori, 1982) and
its uncertainties for our target earthquake:
  Z∞
4π 1 1 2
ER = + f 2 |S(f )| df (4.10)
5ρ 3α5 2β 5
0

where ρ is the density, and α and β are the P and S wave velocity at the source. We
set ρ = 2700 kg/m3 , and the velocities are taken from Scott et al. (1994), with α = 6000
 76

10

Log Relative Amplitude

9

0.1 1 10
Freq (Hz)

Figure 4.4: Mean source spectrum over 8 stations with 95% confidence intervals (gray
area) for the M5.1 target earthquake. Compared to Figure 2, the uncertainties have been
decreased due to averaging over different stations and by the combined EGF, especially at
high frequencies where many small aftershocks can be used to perform spectral division.

m/s and β = 3465 m/s. The apparent stress is

τa = µER /M0 (4.11)

where we use µ = 3.24 × 104 MPa.

Since the S wave contains more than 90% of the energy we only use the second
term of (4.10). The radiated energy is calculated by integrating (numerically) the mean
source spectrum up to the highest frequency possible (30 Hz). We also extrapolate at the
higher frequencies assuming a fall-off rate of ω −1 in velocity, which contains less than 5%
of the total energy. The 95% bounds are integrated to obtain energy uncertainties. The
value of radiated energy obtained is 1.24 × 1011 (0.73 × 1011 , 2.28 × 1011 ) Joules, with the
95% confidence interval in parenthesis. We also compute a value of the apparent stress
(proportional to ER /M0 ) and obtain 0.33 (0.19, 0.59) MPa, with the 95% confidence
interval.
The apparent stress obtained is within the values of previous results for similar
 77

sized earthquakes, and a little lower than smaller events in the same region (Prieto
et al., 2004). But note in this case we have not only obtained the measure of radiated
energy, but also the uncertainties, which because we have used over 100 EGFs, have
been reduced significantly compared to what one would obtain using just one EGF (i.e.,
compare Figures 4.2 and 4.4). Our analysis focuses on random variations in our input
data (spectral ratios) rather than uncertainties in model parameters. Therefore we do not
attempt to propagate errors associated with parameters such as wave speed or density,
even though these errors might be important. From equation (4.10) it would be straight
forward to propagate such errors, if known, following Taylor (1997).

4.5 Discussion

As explained before, using spectral ratios will increase the variance and a larger
number of source models (corner frequencies, fall-off rates) are to be allowed within the
uncertainties. Since stress drop varies as fc3 , the uncertainties of the stress drop as esti-
mated from fc will grow considerably, making the inference of scaling features between
different earthquakes more difficult. Our goal is to show a consistent way of estimating
the uncertainties of the earthquake source spectrum and some source parameters using
the method of propagation of errors and the variance of the spectrum estimation pro-
cedure. A source spectrum is then constructed from a weighted set of spectral ratios in
order to reduce the variance.
As pointed out by Sonley and Abercrombie (2006) it is good practice to check
whether a realistic source pulse is obtained after deconvolution, by inverse FFT of the
multitaper eigencomponents. We compute source pulses for our target M5.1 earthquake
and obtain results similar to those of McGuire [pers. com., 2005] for the same event.
However it is not possible to check all our EGFs since SNR limitations at low frequencies
affect the very small EGFs and it is not always possible to recover source pulses.
In our data set we find that there is no single EGF with good SNR and appro-
priate bias reduction on the entire frequency band of interest (about 0.1 - 30 Hz) and
it is necessary to use multiple EGFs. Even if an ideal EGF is found, the uncertainties
of the radiated energy estimate would be considerably larger than presented here and
 78

should be taken into account when comparing different results or looking for scaling of
energy with earthquake magnitude.
A possible major source of bias is if many EGFs used in this study (especially
the small ones) have consistently different focal mechanisms that are not accounted for.
We believe that by using many EGFs we are sampling a wide variety of focal mechanisms.
For the target event we obtain realistic source pulses, suggesting similar mechanisms. The
aim of this study is to show a way of estimating and reducing uncertainties of source
spectra based on the target and EGF spectra. It is not intended to completely remove
all possible biases associated with the choice of events and model parameterization.

Acknowledgments

We thank Glen Offield for field operations and the Anza group at UCSD who
picked and archived the seismograms. We thank M. Hellweg, V. Oye and R. Abercrombie
(editor) for thoughtful comments. Funding for this research was provided by NSF Grant
number EAR0417983. This research was also supported by the Southern California
Earthquake Center. SCEC is funded by NSF Cooperative Agreement EAR-0106924 and
USGS Cooperative Agreement 02HQAG0008. The SCEC contribution number for this
paper is 933.
5

Quadratic Multitaper Spectrum

The power spectral density of geophysical signals provides relevant information

about the processes that generated these particular signals. We present a new method
to optimally use the multitaper spectral analysis method. The method is an extension
of the algorithm by Thomson (1982) with a reduction of bias due to the curvature close
to the frequency of interest. A comparison of the original and the new Quadratic mul-
titaper with the same resolution bandwidth demonstrates the reduction of bias in areas
where the signal has significant quadratic structure without the introduction of addi-
tional sidelobe leakage. In addition, the methodology provides independent estimation
of the derivatives of the spectrum (e.g., the slope of the spectrum). The extra information
can be implemented for parameter estimation or in comparing different signals.

79
 80

5.1 Introduction

There are many applications in geophysics where relevant information contained

in a given signal may be extracted from the frequency content of the spectrum. In
some cases, the scientist may be interested in periodic components usually immersed
in some background noise (e.g., normal mode seismology (Gilbert, 1970), climate-time
series (Chappellaz et al., 1990), etc.), in a general continuous spectrum to be estimated
from a short time series (e.g., earthquake source spectra (Brune, 1970), bathymetry data
(Goff and Jordan, 1988), etc.) or in comparing two signals and investigating where the
similarities or differences are (in seismology for example (Vernon, 1989; Hough and Field,
1996); transfer functions in electromagnetism (Constable and Constable, 2004), elastic
thickness of the lithosphere (Daly et al., 2004), etc.). In each of these cases, it is desirable
to be able to obtain a reasonable spectrum, with little or no bias and small uncertainties.
In Thomson (1982) the multitaper spectral analysis method was introduced.
The original algorithm has been widely used in geophysical applications and has been
shown in multiple cases to outperform the single-tapered, smoothed periodogram (Park
et al., 1987b; Bronez, 1992; Riedel et al., 1993). In the latter, a multitaper estimate that
is subsequently smoothed is preferred.
Single taper estimates have a major limitation, in the sense that by using one
taper a significant portion of the signal is discarded. The data points at the extremes
are down-weighted, causing the variance of the direct spectral estimate to be greater
than that of the periodogram. In the multitaper algorithm, the statistical information
discarded by one taper is partially recovered by the others. The tapers are constructed to
optimize resistance to spectral leakage and only a few of them are computed. The multi-
taper spectrum is constructed by a weighted sum of these single tapered periodograms.
The weighting function is defined to generate a smooth estimate with less variance than
single taper methods and at the same time to have reduced bias from spectral leakage.
The weighting proposed (Thomson, 1982) is ideal for spectral leakage but suffers
from local bias. What we mean by local bias is that by averaging over a number of
tapered spectra, the estimate will have a broad response around the center frequency,
whose width is that of the frequency resolution chosen to reduce spectral leakage.
Table 5.1: Essential notation and mathematical symbols used in this chapter. In the text, θ̂ means estimate of the variable θ; E[·]
is the expected value and ∗ represents the complex conjugate.

Symbol Description
x(t) The time series to be analyzed (assumed with time unit intervals)
N The number of data points of the time series
t The time variable (t = 0, 1, 2, . . . , N − 1)
f ,f 0 Continuous frequency variables
vk (t) Slepian sequences. Are a function of both N and a bandwidth W
W Bandwidth of the windows vk , which define the inner domain (f − W, f + W )
λk Eigenvalues of the Slepian sequences. Also give the fraction of energy in the inner band (−W, W ).
K Number of tapers or windows to be used. Number of Slepian sequences used
dZ(f ) Orthogonal incremental process from Cramér Spectral representation. Also known as generalized Fourier transform.
Vk The Slepian functions, Fourier transform of vk ’s. Functions are orthogonal in the principal domain (− 21 , 12 )
Vk Orthonormal version of Slepian functions in inner domain (−W, W ). Defined as Vk = Vk /λk
Yk (f ) kth eigencomponent, the discrete Fourier transform of x(t)vk (t) in the domain (− 12 , 12 ).
Yk (f ) Idealized kth eigencomponent in the domain (−W, W ). Theoretically contain information from the inner domain.
Sk (f ) The kth eigenspectra. A direct spectral estimate using vk as a taper.
Cjk (f ) Covariance matrix of the idealized eigencomponents. Cjk = Yj Yk∗ .
S(f ) The power spectral density (PSD) of the time series.
Tn (f ) The nth Chebyshev polynomial function.
αn The nth Chebyshev coefficient for the expansion of the spectrum S(f )
dk (f ) Multitaper weights used to obtain estimates of the bandlimited coefficients Yk (f )
81
 82

Riedel and Sidorenko (1995) suggested a different weighting function to mini-

mize the local bias. They take a different set of orthogonal tapers and, assuming qua-
dratic structure of the spectrum, reduce the bias introduced by the curvature of the
spectrum. Because the chosen tapers minimize local bias, there is little spectral leak-
age reduction and this method does not perform as well with very high dynamic range
signals.
In this paper we present an improved multitaper spectral estimate using Thom-
son’s algorithm modified to reduce curvature bias. The second derivative of the spectrum
is estimated by means of an expansion of the spectrum via the Chebyshev Polynomials.
Because we use the Slepian tapers and the original weighting functions, the estimate is
resistant to spectral leakage, yet reduces bias due to spectral curvature. In addition,
we can estimate the derivative of the spectrum (slope), which can be used in parameter
estimation or as a discriminant for comparing different signals.
This paper is organized as follows. In the next section, we outline the problem
associated with the estimation of the power spectral density (PSD). After this, we provide
a brief review of the multitaper spectrum estimation algorithm and discuss some of its
properties. Then, we introduce the Quadratic multitaper, first explaining in section
5.4 the methodology to estimate the derivatives of the spectrum and in section 5.5
the estimation of the unbiased spectral estimate. Finally, there is a fairly extensive
section in which we give a number of examples to compare Thomson’s multitaper and
the new algorithm. There is an application to bathymetric data and the application of
the derivatives of the spectrum for comparing signals.

5.2 Spectrum Estimation

In time-series analysis it is often useful to describe the power spectral density

(PSD) of the signal, given that it may have information of the background noise, periodic
components, and transients. These pieces of information are fundamental in geophysics
(Thomson, 1990).
To start, we need to consider a stationary stochastic process x(t), a zero mean
discrete time series consisting of N contiguous samples and assume that the sampling
 83

rate is always unity, so that t = 0, 1, 2, . . . , N − 1. Define the Fourier transform of the

observations
N −1
X 1 1
Y (f ) = x(t) e−2πif t , − 6f 6 . (5.1)
2 2
t=0

We assume the signal is a harmonizable process, thus has a Cramér spectral representa-
tion (Cramér, 1940)
1
Z2
x(t) = e2πif t dZ(f ) (5.2)
− 12

for all t, where dZ(·) is an orthogonal incremental process (Doob, 1952).

The random orthogonal measure dZ(f ) has its first moment

E [dZ(f )] = 0 (5.3)

and second moment (of relevance for our purposes)

h i
E |dZ(f )|2 = S(f )df (5.4)

where S(f ) is defined as the power spectral density function of the process and E[·] is
the expected value. Note here that the frequency variable f is continuous and so we are
in fact trying to find a function S(f ) from the finite series x(t).
Plugging the Cramér spectral representation (5.2) into the discrete Fourier
transform (5.1), we arrive at the basic integral equation (Thomson, 1982, 1990)
1 1
Z2 Z2
sin N π(f − v)
Y (f ) = dZ(v) = D(f, v)dZ(v) (5.5)
sin π(f − v)
− 21 − 12

where D(f, v) is the Dirichlet kernel. The basic integral equation is a convolution that
can be interpreted as the smearing of the true dZ(f ) projected into Y (f ), due to the
finite duration of the time series x(t).
This paper is about obtaining an approximate solution to this equation and
based on that solution (via Equation 5.4) obtaining reliable estimates of the power spec-
tral density of the signal.
 84

5.3 Multitaper Spectrum estimates

We give a brief review of the standard theory for multitapers. Proof of various
assertions can be found in (Thomson, 1982, 1990; Park et al., 1987b; Percival and Walden,
1993). Note that given the properties of (5.5) and the smoothing of the Dirichlet kernel
there is no unique solution to this problem. The multitaper spectrum estimate is an
approximate least-squares solution to equation (5.5) using an eigenfunction expansion.
The choice of this type of solution will be explained next.
The most obvious first guess for the spectrum is the squared Fourier transform.
Slightly rewriting the discrete Fourier transform (5.1) and squaring
2
−1

NX 
2 −2πif t 
Ŝ(f ) = |Y (f )| =  x(t)a(t) e (5.6)


 
t=0

where the sequence a(t) is called a taper. In the case of (5.1) the taper a(t) = 1 is a
boxcar function. To maintain total power correctly a(t) needs to be normalized:
N
X −1
|a(t)|2 = 1 (5.7)
t=0

In the frequency domain, the properties of the taper are deduced from its Fourier trans-
form
N
X −1
A(f ) = a(t) e−2πif t (5.8)
t=0
We call the function A(f ) the spectral window associated with a. For conventional
tapers, |A(f )| has a broad main lobe and a succession of smaller sidelobes (see Figure
5.1).
The choice of the taper can have a significant effect on the resultant spectrum
estimate. One can observe this by expressing equation (5.6) as a convolution of the taper
transform (5.8) and the true spectrum
1
h i Z2
2
E Ŝ(f ) = |A(f 0 )| S(f − f 0 )df 0 (5.9)
− 12

Here, as in (5.5), there is smearing or smoothing of the true spectrum. This means
that the choice of window is important. A good taper will have a spectral window with
low amplitudes whenever |f − f 0 | is large, leading to an estimate Ŝ(f ) based primarily
 85

on information close to the frequency f of interest. The objective of the taper a(t) is
to prevent energy at distant frequencies from biasing the estimate at the frequency of
interest. This bias is known as spectral leakage. We wish to minimize the leakage at
frequency f from frequencies f 0 6= f .
In practice, it is not sensible to be concerned about |f 0 − f | 6 1/N , since this
is the lowest frequency accessible from a record of length N . A bandwidth W is chosen,
where 1/N < W ≤ 1/2, and the fraction of energy of A in the interval (−W, W ) is given
by:
ZW
|A(f )|2 df
−W
λ(N, W ) = 1
(5.10)
Z2
|A(f )|2 df
− 12

where λ is a measure of spectral concentration. It is clear that no choice of W can make

λ greater than unity. Our task is to maximize λ.
To maximize λ substitute (5.8) in (5.10), take the gradient of λ with respect to
the vector a = [a(0), a(1), . . . , a(N − 1)] and set to zero to obtain a matrix eigenvalue
problem:
D · a − λa = 0 (5.11)

where the matrix D has components

sin2πW (t − t0 )
Dt,t0 = , t, t0 = 0, 1, . . . , N − 1 (5.12)
π(t − t0 )

and is symmetric.
The solution of (5.11) has (dropping dependence on N and W ) eigenvalues
1 > λ0 > λ1 > · · · > λN −1 > 0 and associated eigenvectors vk (t), called the Slepian
sequences (Slepian, 1978). The first eigenvalue λ0 is extraordinarily close to unity, thus
making the choice a(t) = v0 (t) the taper with the best possible suppression of spectral
leakage for the particular choice of bandwidth W . In fact, the first 2N W − 1 eigenvalues
are also very close to one, leading to a family of very good tapers for bias reduction.
The multitaper algorithm exploits the fact that a number of tapers have good spectral
leakage reduction, and uses all of them rather than only one.
 86

0.2
v k (t)
0.1

Amplitude
0.0

-.1

k=1 k=4 k=7

-.2
0 20 40 60 80 100
Sample
10 0
Vk (f)
Transfrm Amplitude

-2
10

10-4

10-6

10-8
-5W -3W -W W 3W 5W
Frequency

Figure 5.1: Selected Slepian sequences and corresponding Slepian functions for N = 100
samples and a choice of N W = 4. Sometimes called 4π Slepian sequences.

5.3.1 Properties of Slepian sequences and functions

The Slepian sequences are solutions of the symmetric matrix eigenvalue problem
(5.11) - (5.12). The eigenvectors with associated eigenvalues λk are real and orthogonal
as usual
N
X −1
vj (t)vk (t) = δjk (5.13)
t=0

These vectors will be used as tapers in (5.6). We define the Slepian functions as the
spectral windows, the Fourier transforms of the sequences
N
X −1
Vk (f ) = vk (t)e−2πif t (5.14)
t=0

Note that the Vk ’s are complex functions of frequency. Figure (5.1) shows three Slepian
sequences and their corresponding Slepian functions.
Orthogonality conditions also hold in the frequency domain, making the choice
 87

1 Imaginary Vk (f)
Real Vk (f) k=1
k=4

Normalized Amplitude
k=7

k=1
k=7 k=4
-1
-W 0 W -W 0 W
Frequency Frequency

Figure 5.2: Orthonormal version of the Slepian functions in Figure 5.1, Vk (f ) in the
inner domain. Only three functions are shown in the example with their real and imagi-
nary amplitudes normalized. Number and thin lines show the index of the corresponding
Slepian function plotted. The symmetry and amplitude of the functions are of interest.

of the Slepian sequences so interesting:

1
Z2
Vj (f )Vk∗ (f )df = δjk (5.15)
− 12

ZW
Vj (f )Vk∗ (f )df = λk δjk (5.16)
−W

It is convenient to define an orthonormal version of the Vk ’s on the inner domain (−W, W )

Vk (f )
Vk (f ) = √ (5.17)
λk

with the obvious property

ZW
Vj (f )Vk∗ (f )df = δjk (5.18)
−W

This last property will be exploited in the sections to come. Figure (5.2) shows a selection
of the Vk (f ) functions in the inner interval. Note how the real part of the functions is
always even and the imaginary part is odd. See also how the amplitude of the functions
increases outward as the Slepian function index increases and are more sensitive to
structure further from the center frequency.
 88

5.3.2 The multitaper algorithm

The objective of this algorithm is to estimate the spectrum S(f ) by using K of

the Slepian sequences to obtain the k eigencomponents:
N
X −1
Yk (f ) = x(t)vk (t)e−2πif t (5.19)
t=0

and a set of K eigenspectra as in (5.6):

Ŝk (f ) = |Yk (f )|2 (5.20)

from which we can form the mean spectrum

K
1 X
S̄(f ) = Ŝk (f ) (5.21)
K
k=1
The idea of taking an average is to reduce the variance in the spectral estimate. As will be
shown below, the mean spectrum is not an ideal estimate and we prefer a weighted aver-
age instead, one that minimizes some measure of discrepancy. While the spectral leakage
properties of the Ŝ0 eigenspectrum are very good, since the eigenvalues are close to unity
when K < 2N W − 1, the leakage characteristics of the succesive estimates degrade. It
is clear that by using Ŝ = |Y0 |2 , the least amount of spectral leakage is achieved. Never-
theless, including the other eigencomponents (Y1 , Y2 , . . . , YK ), while increasing spectral
leakage, reduces the variance of the spectral estimate and is thus preferred.
In order to estimate the discrepancy of the different eigencomponents Yk , we
combine (5.19) with (5.2)
N
X −1
Yk (f ) = x(t)vk (t)e−2πif t
t=0
1
N −1 Z2
0
X
= dZ(f 0 )e2πif t vk (t)e−2πif t
t=0
− 21
1
Z2 NX
−1
0
= vk (t)e−2πi(f −f )t dZ(f 0 )
t=0
− 12

and using the definition of the Fourier transform of the taper (5.14) we obtain:
1
Z2
Yk (f ) = Vk (f − f 0 ) dZ(f 0 ) (5.22)
− 12
 89

containing information from the whole interval (− 12 , 21 ).

If the sequence x(t) were passed by a perfect bandpass filter from f − W to
f + W before truncation to the sample size with N data points, we would obtain the
idealized eigencomponents Yk (f ) that, though unobservable, would be represented by:
ZW ZW
Vk (f 0 )
Yk (f ) = √ dZ(f − f 0 ) = Vk (f 0 )dZ(f − f 0 ) (5.23)
λk
−W −W

Note that here we adopt the orthonormal functions Vk , in order to maintain the correct
normalization. The Yk takes only information over the inner interval (−W, W ).
In order to estimate Ŝ(f ), we find a set of frequency dependent weights dk (f ),
as proposed by (Thomson, 1982, 1990):
√
λk S(f )
dk (f ) = (5.24)
λk S(f ) + (1 − λk )σ 2
where σ 2 is the variance of the signal x(t). The multitaper spectrum is then obtained
K−1
X 2
d2k |Yk (f )|
k=0
Ŝ(f ) = K−1
(5.25)
X
d2k
k=0

Since we don’t know the spectrum S(f ) in (5.24), we are required to assume an initial
estimate of the spectrum (averaging the first two k eigenspectra S0 + S1 for example)
and find the weights dk iteratively. A complete derivation of the weights in (5.24) can
be found in Thomson (1982) or Percival and Walden (1993).

5.4 Estimating the derivatives of the spectrum

More information about the spectrum can be obtained by looking at the co-
variance matrix of the K components:

Cjk (f ) = E [dj Yj dk Yk∗ ] = E [Yj Yk∗ ] (5.26)

with j, k = 1, . . . , K. We use the orthogonal increment property (5.4) and substituting

(5.23):
ZW
Cjk (f ) = Gjk (f 0 ) S(f − f 0 )df 0 (5.27)
−W
 90

where Gjk (f ) = Vj (f )Vk∗ (f ). If the spectrum does not vary in the interval (−W, W ) then
the covariance matrix is diagonal

C(f ) = S(f ) I (5.28)

where I is the K × K identity matrix. Note that the multitaper spectrum in (5.25) is
equivalent to taking the trace of C(f ) and normalizing by the weights.
When the spectrum is constant in the interval (−W, W ), Thomson’s multitaper
is unbiased and provides an appropriate estimate of the spectrum. If, however, the
spectrum varies within the interval, the matrix C(f ) will not be diagonal and (5.25)
may be biased. Clearly, we rarely obtain perfectly diagonal covariance matrices and the
spectrum is not perfectly resolved.
Like (5.5) equation (5.27) is a Fredholm integral of the first kind and suffers
from similar non-uniqueness and smearing features, except in this case we have reduced
spectral leakage and are only concerned about the interval (−W, W ).
Thomson (1990) suggested taking a set of orthogonal eigenfunctions to expand
the spectrum to solve (5.27). Here we propose to employ the Chebyshev polynomials
for estimating the derivatives of the spectrum and using these derivatives to obtain an
improved solution of the spectrum S(f ). We prefer the Chebyshev polinomials (Mason
and Handscomb, 2003) because, as can be seen from Figure (5.3), these polynomials are
sensitive to structure at the edges of the interval, where the eigenfunctions proposed by
Thomson (1990) have very little energy.
We write the spectrum in the inner interval as:

f0 f0 f0
S(f − f 0 ) = α0 T0 ( ) + α1 T1 ( ) + α2 T2 ( ), (5.29)
W W W
0
−W 6 f 6 W

where Tn (x) is the Chebyshev polinomial of degree n (see Mason and Handscomb, 2003).
We show only the first three polynomials, since these are applied throughout the study,
and express the quadratic terms of the signal.
Returning to the inverse problem (5.27) in spectrum estimation and inserting
(5.29):
(0) (1) (2)
Cjk (f ) = α0 Hjk + α1 Hjk + α2 Hjk + O(f − f 0 )3 (5.30)
 91

Normalized Amplitude
0

-1 A B
-W 0 W -W 0 W
Frequency Frequency

Figure 5.3: Comparison between first 3 basis functions used in Thomson (1990) (A)
and Chebyshev polynomials (B) used in this study. Note that in (A), the basis functions
always tend to zero when getting close to either −W or W and are not sensitive to
structure at the boundaries. The Chebyshev polynomials in (B) are also very simple
approximations of a constant, slope, and quadratic terms.

where the matrices

ZW
(0)
Hjk = Gjk (f 0 ) T0 (f 0 ) df 0 = δjk
−W
ZW
(1)
Hjk = Gjk (f 0 ) T1 (f 0 ) df 0
−W
ZW
(2)
Hjk = Gjk (f 0 ) T2 (f 0 ) df 0
−W

describe the zero, first, and second derivative basis matrices. We can then obtain the
Chebyshev coefficients, α0 , α1 , α2 , by solving the least squares problem where we use the
observed values of dj dk Yj Yk∗ to approximate the left side.
The Chebyshev coefficients are estimates of the derivatives of the spectrum

α0 ≈ S(f )

α1 ≈ S 0 (f )

α2 ≈ S 00 (f )

around the center frequency on the interval (f − W, f + W ).

 92

H(0) H(1) H(2)

0

REAL
3

j index
6
0

IMAGINARY 1
3

60 3 60 3 60 3 6 0

k index

Figure 5.4: Comparison between the different basis matrices for zero (left), first (mid-
dle), and second (right) order coefficients. The absolute values are shown for simplicity.
Note that the completely white matrices show that the real part of the covariance ma-
trix is insensitive to slopes, while the imaginary part is insensitive to a constant and
quadratic structure of the spectrum.

(1) (2)
In practice, the calculation of the integrals in Hjk and Hjk is done numerically
using a trapezoidal quadrature. In Figure (5.4) we plot the case of the three matrices in
(5.30). The absolute values are plotted. Note that both the zero and second order terms
are only present in the real part of the covariance matrix, while the first order term is
present only in the imaginary part of the covariance matrix.
It is clear that the constant term will result in a diagonal covariance matrix,
(1)
while the effects of the first and second order terms are quite different. Hjk has no effect
on the diagonal terms, suggesting that this term does not bias the spectrum estimate
(2)
(5.25). In contrast, Hjk has an important contribution to the diagonal but is also
present in the off-diagonal terms, showing the dependence of the eigencomponents in
spectra that are highly variable. A slight correlation between the estimates of α0 and
α2 is present.

5.5 Quadratic Multitaper

Up to now, the literature (e.g., Thomson (1982); Park (1992); Percival and
Walden (1993), and many others) has assumed the spectrum varies slowly in this interval
 93

and can be taken out of the integral in equation (5.27). Now, within (−W, W ) we can try
to find further information on the structure of the spectrum and relax the assumption
of a constant spectrum inside the interval.
Assume the spectrum has a Taylor series expansion on the interval (f − W, f +
W ) of the form:

S(f ) = S(f 0 ) + (f − f 0 )S 0 (f 0 ) + 12 (f − f 0 )2 S 00 (f 0 ) + O(f − f 0 )3

We know that the estimate of the spectrum Ŝ(f ) at any frequency f is an average over
the interval (f − W, f + W ):

ZW
1
S(f 0 ) + 12 (f − f 0 )2 S 00 (f 0 ) df 0
 
E[Ŝ(f )] =
2W
−W
1
= S(f ) + W 2 S 00 (f )
6

where the term associated with the first derivative does not contribute to the integral
due to symmetry. Note that in Figure 5.4 the matrix H(1) , associated with the slope, is
zero in the main diagonal and does not bias E[Ŝ(f )].
We can obtain the Quadratic multitaper estimate of the spectrum S̃(f ) at
frequency f by applying the correction:

1
S̃(f ) = Ŝ(f ) − W 2 α̂2 (5.31)
6

where we assume α̂2 ≈ S 00 (f ) obtained by solving (5.30). Note that we apply the
correction to the multitaper estimate Ŝ(f ) in (5.25).
Applying the quadratic correction in (5.31) will increase the variance of the
overall estimate, because α̂2 is also uncertain. We propose to implement a mean-square
error criteria instead of directly applying (5.31) to avoid exacerbating the uncertainties
of the Quadratic multitaper:

1
S̃(f ) = Ŝ(f ) − µ W 2 α̂2 (5.32)
6

where µ is a weight:
α2 2
µ= (5.33)
(α2 2 + var{α2 })
 94

In Appendix B the derivation of the weight µ in (5.33) is explained as well as the

approximate estimation of the variance of α2 .
The Quadratic multitaper is an approximately unbiased estimate of the PSD of
the signal analyzed. As will be demonstrated in the next section with different examples,
the Quadratic multitaper provides a reduction of curvature bias while at the same time
generating smooth estimates.

5.6 Examples

As a demonstration of the benefits of the Quadratic spectrum algorithm (QMT)

we show a number of synthetic examples. The main features we would like to concentrate
on are the resolution close to significant structure in the spectrum, smoothness of the
resultant estimates and the overall spectral leakage properties.

5.6.1 Random signal

The first signal to be analyzed is a simple pseudo-random number r(t) with a

normal distribution and standard deviation σ=1. The number of data points for this
example is N =1000, For all plots in this paper we compute 6 tapers (K=6) with 3.5 as
the time-bandwidth product.
A visual inspection of the results of the two different methods in Figure 5.5
shows that the algorithm presented here generates a smoother spectrum than the original
Thomson method (TMT). A more quantitative comparison is provided in Table 5.2,
where 10 random realizations of a 1000 sample long random time series were analyzed
and two different measures were used to assess the smoothness of the resultant spectrum:
the norm of the second (numerical) derivative of the spectrum and a count of the number
of maxima in each spectra.
Results of these measures of smoothness in Table 5.2 demonstrate that the QMT
generates smoother estimates of the spectrum. In all 10 realizations, both measures of
smoothness where lower when using the Quadratic algorithm.
 95

Table 5.2: Comparison of smoothness for the multitaper methods

Thomson Quadratic
Number of realizations 10 10
Second derivative norm 216.8 49.9
Standard deviation 22.2 5.99
Count of maxima 123.3 67.3
Standard deviation 5.07 3.62

-3
0 200 400 600 800 1000
Time (seconds)

5 Thomson MT 5 Quadratic MT
PSD

PSD

3 3

1 1

0 0.2 0.5 0 0.2 0.5

3 2W 3 2W
PSD

PSD

2 2

1 1

0.05 0.1 0.05 0.1

Frequency (Hz) Frequency (Hz)

Figure 5.5: Spectrum estimation of pseudo-random number signal. The signal is a

normally distributed random vector, with N =1000 samples. The top panel shows the
time-series random signal. The middle panels show the TMT (left) and the QMT (right)
estimates of the spectrum. Bottom panels show a detailed view of the estimates between
0.01 and 0.1 Hz and the 2W bandwidth for reference. The QMT generates a smoother
spectral estimate.
 96

1e11
1e11

1e09
8e10
-W W
1e07

Log PSD
Linear PSD
-W W
6e10

1e05
4e10

1e03
2e10

0 1e01
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Frequency (Hz) Frequency (Hz)

Figure 5.6: Spectrum estimation for a high dynamic range signal with two periodic
components at 0.05Hz and 0.3Hz. N =100 points, time-bandwidth product N W =3.5
and K=6 . The linear scale spectrum (top panel) shows the improved performance
of the QMT in describing the periodic components, while the logarithmic scale (lower
panel) shows the similar spectral leakage properties of both algorithms.

5.6.2 Periodic Components

We test the effectiveness of the new algorithm with two signals; we examine a
random signal r(t) with σ=1.0 and a pair of periodic components. The number of data
points is reduced to N =100 in order to have a comparison of spectral leakage around
the linear components.
Our first test is to see whether this algorithm represents the periodic compo-
nents in the signal better, without introducing additional spectral leakage. For this, we
take the signal:
x(t) = A0 sin(2πf1 t) + A0 sin(2πf2 t) + r(t) (5.34)

where f1 = 0.05, f2 = 0.3, and the amplification factor A0 = 105 . Two questions arise
here. How well can we describe the periodic components; effectively, line features in the
spectrum, and is there any spectral leakage introduced due to the Quadratic algorithm?
Figure 5.6 shows the results of spectral analysis from both TMT (gray) and QMT, on
linear and log axes. The linear plot clearly shows the more accurate description of the
periodic components provided by the QMT, the logarithmic plot shows that no additional
spectral leakage is introduced. Note how both methods overlap at very low amplitudes.
The signal has 8 to 10 orders of magnitude dynamic range and both methods behave
similarly in terms of spectral leakage.
 97

The second test signal has a much smaller signal-to-noise ratio. In this case we
let the amplification factor be A0 = 1.0 and the standard deviation σ of r(t) remains
fixed. Figure 5.7 shows the result of spectral analysis on this signal. We would like
to stress two important features that can be seen from these results. First, the linear
components are better described by the QMT. Second, the information outside the range
of the periodic components is smoother, as shown in the random signal example (Table
5.2).
Additionally, we have also obtained extra information about the spectrum.
Figure (5.7) shows the estimate of the first and second derivatives of the spectral contents
of the signal. As expected, the first derivative should be very close to zero, when getting
close to the periodic component. Similarly, the second derivative of the spectrum should
have a large negative value, showing that the line represents a local maxima of the
spectrum.. These two features are clearly present in the estimates of the derivatives.
Given the randomness of the signal and also uncertainties due to the non-uniqueness
of the problem in our example, the second derivative around the 0.3Hz component is
not exactly the largest negative value, but is rather off by a frequency bin. This shows
that still some uncertainties remain in all estimates, including the spectrum and its
derivatives. Nevertheless, the extra information that is gained from the derivatives could
certainly be relevant.
Note that the estimates of the derivatives are not computed by a numerical
differentiation of the spectrum estimate Ŝ(f ), but rather by the steps described in the
previous sections.
A method for the detection of periodic components using the multitaper algo-
rithm was developed by Thomson (1982), known as the F-test for spectral lines. For a
complete description of the methodology the reader is refered to Thomson (1982) and
Percival and Walden (1993). The method can be applied for reshaping the spectrum near
spectral lines (e.g., Park et al., 1987a; Thomson, 1990; Denison et al., 1999) or even for
removal of these periodic signals embedded in a colored spectrum (Lees, 1995; Percival
and Walden, 1993, Chapter 10). For high signal-to-noise ratios as in Figure 5.6, F-test
or other line detection algorithms are preferred for harmonic analysis.
The general idea of spectral reshaping is to subtract the effect of the statistically
 98

A 4 B
2 Spectrum Derivatives

Amplitude
400
0
200
-2

-4 0
0 20 40 60 80 100
Time (s) -200
S‘
16 -400
Thomson MT Quadratic MT
400
12
200
-W W -W W
PSD

8 0

-200
S‘’
4
-400
0 0.1 0.2 0.3 0.4 0.5
Frequency (Hz)
0
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
Frequency (Hz) Frequency (Hz)

Figure 5.7: Spectrum estimation of a normally distributed random signal with two
sinusoids at 0.05 Hz and 0.3 Hz. Parameters as in Figure 5.6. Plots below the time series
contain the estimates of the spectrum using the TMT (left) and the QMT (right). The
figures in the bounded box (right) show the first and second derivatives as estimated from
the covariance matrix. Vertical gray lines represent the location of the line components.
Note how the derivative estimates help in pinning where the line components are located.

significant lines from the eigencomponents Yk (see equation 5.22). This subtraction is
done before the adaptive weighting in equations (5.24) and (5.25), meaning that it is
possible to perform either TMT or the QMT estimates on the remaining stochastic part
of the spectrum (without the deterministic periodic components, just removed), with
similar improvements as presented in the examples above, if the Quadratic algorithm is
applied.

5.6.3 Resolution test and the choice of multitaper parameters

One important question that remains unanswered in spectral analysis using

multitaper methods is, what is the optimal choice of the time-bandwidth product N W
(the averaging bandwidth) and the number of tapers K (the more tapers, the smoother
the estimate)? Or, having a chosen bandwidth W , what is the ideal number of tapers
that should be used?. Riedel and Sidorenko (1995) invented the sine multitaper method
to get around this problem by choosing an optimal number of tapers iteratively at each
frequency.
In the multitaper literature, it has been proposed that K = 2N W − 1 as
 99

an appropriate choice, since the eigenvalues λk are all close to unity. This choice is
essentially based on the leakage properties of the tapers, but does not take into account
the particular shape of the spectrum of the signal under analysis. In this subsection,
we present a comparison of the effect of the choices of N W and K on the resolution of
the spectra around a periodic component. We show how the QMT is less dependent on
these choices compared with TMT.
In Figure 5.8 and Table 5.3 we compare the resolution of TMT and the Qua-
dratic multitaper. A useful criterion is that of the width of the half-power points, also
known as the 3-dB bandwidth. This criterion reflects the fact that two equal-strength
periodic components separated by less than the 3-dB bandwidth will show in the spec-
trum as a single peak instead of two (Harris, 1978). We use the signal in the previous
section (Figure 5.6) and plot the spectrum centered on one of the periodic components
on a dB scale defined as:
dB = 10 log10 (S(f )/S(f0 )) (5.35)

where f0 is the frequency of the periodic component. We vary the time-bandwidth N W

and the number of tapers K to investigate the effect of these choices. We also present
in Table 5.3 the result of the 3-dB and 9-dB ( 18 th power) bandwidths for the different
choices of N W and K by applying a linear interpolation.
The QMT always outperforms Thomson’s algorithm given the same parameters
as shown in Table 5.3. Note that at the 9-dB line in Figure 5.8 (see Table 5.3 as well)
both methods provide similar results, with the method introduced here being slightly
better.
An important result obtained by conducting this test is the fact that the 3-dB
bandwidth is less sensitive to the choice of N W for the QMT (see Figure 5.8A). Once
we reach the 9-dB bandwidth, a larger value of N W decreases the resolving power. On
the other hand the choice of K is directly proportional to the resolution bandwidth for
both algorithms (this is also evident from Figure 5.2), with the Quadratic algorithm
having narrower 3-dB and 9-dB bandwidths in all cases. A final observation from Table
5.3 indicates that a comparatively better resolution is achieved with the QMT even if
one more taper K is used compared to TMT (compare also red and gray lines in Figure
5.8B), leading to smoother estimates due to the increased degrees of freedom.
Table 5.3: Comparison between multitaper methods by their 3-dB and 9-dB bandwidths in Rayleigh frequency fR units with
different choices of time-bandwidth product N W and number of tapers K for a periodic component in the spectrum in Figure 5.6.

3-dB 9-dB
K N W Thomson Quadratic Thomson Quadratic
7 3.0 3.07 2.85 3.22 3.16
7 3.5 3.11 2.96 3.32 3.22
7 4.0 3.33 2.84 4.01 3.65
7 4.5 3.65 2.75 4.16 4.03
7 5.0 4.03 2.81 4.28 4.10
Constant number of tapers, varying time-bandwidth

3-dB 9-dB
K N W Thomson Quadratic Thomson Quadratic
4 4.0 2.47 1.79 3.08 2.68
5 4.0 3.01 2.05 3.25 3.08
6 4.0 3.17 2.54 3.53 3.29
7 4.0 3.33 2.84 4.01 3.65
8 4.0 4.01 3.23 4.17 4.09
Varying number of tapers, constant time-bandwidth
100
 THOMSON MT QUADRATIC MT
A
0
NW=5 NW=5

-3

-6
NW=3 NW=3

dB [10 • log(S(f)/S(f0) ]
-9

B CONSTANT # TAPERS
0 K=4 K=8 K=5 K=8

-3

-6
K=5 K=4

dB [10 • log(S(f)/S(f0) ]
-9
CONSTANT BANDWIDTH

f0 - f f0 f0+f f0 - f f0 f0+f
Frequency Frequency

Figure 5.8: Resolution test comparison between TMT and the QMT with different choices of time-bandwidth product N W and
number of tapers K for a periodic component in the spectrum in Figure 5.6. In A we plot the spectral shape around a periodic
component at frequency f0 for two choices of time-bandwidth product N W with K = 7 tapers using the two methods. Two dashed
lines represent the 3-dB (half-power) and 9-dB lines. The QMT has a 3-dB bandwidth that is narrower than the equivalent multitaper
and is relatively independent of the choice of NW. In B we fix the time-bandwidth product N W = 4.0 and vary the number of tapers.
For the same choice of parameters the QMT outperforms the TMT with a narrower 3-dB bandwidth. We added the Quadratic
estimate using N W = 4.0 and K = 5 on both plots for reference (gray line), resulting in a similar resolution to that of the K = 4
TMT (red line on left plot) showing that the QMT can effectively provide smoother estimates (by using more tapers) without the
loss of resolution.
101
 102

5.6.4 Synthetic earthquake signal

In geophysical applications many signals have spectral shapes with large dy-
namic range (red spectra) but rarely with deterministic components (periodic signals).
The spectra are continuous, for example the Earth’s background seismic noise (Berger
et al., 2004), medium and small sized earthquake sources (e.g., Abercrombie, 1995; Prieto
et al., 2004), the crustal magnetic field (Korte et al., 2002), and many others.
Consider the spectrum of an earthquake, which follows the Brune (1970) model:

2πf M0
u̇(f ) = (5.36)
1 + (f /fc )2

where u̇(f ) is the velocity amplitude source spectrum associated with the earthquake,
M0 is the seismic moment (related to the size of the earthquake) and fc is the corner
frequency. The corner frequency represents the predominant frequency content of the
radiated seismic energy from the earthquake rupture. The spectrum from this model has
a triangular shape if plotted in log-log axes with a slope of two in power.
In this synthetic example, we generate a pseudo-random time series with 1000
samples whose spectra follow the source model in Equation (5.36). Even though TMT
possesses good spectral leakage reduction, the spectrum may be biased due to the quadra-
tic effects we discussed previously. This is especially true when the corner frequency gets
extremely close to the sampling frequency, so that the curvature around fc is described
by a small number of spectral points.
Figure (5.9) shows the spectral estimates from a realization of a synthetic source
model with fc = 0.005Hz using TMT and the QMT. The triangular shape that is ex-
pected from source spectra is better constraint using the new algorithm.
In addition to the standard spectrum, it is also possible to obtain an estimate
of the derivative of the spectrum. The derivative estimate of two different source mod-
els are shown in Figure (5.10), taken from an average of 100 random realizations and
corresponding standard errors. The two cases presented have corner frequencies close
to the Rayleigh frequency fR . Whenever the corner frequency is close to the sampling
frequency, its curvature is represented by few spectrum bins.
The uncertainties of the derivative estimate are, similar to the uncertainties of
the PSD, proportional to the amplitude of the spectrum. Using the derivative provides
 103

fc = 0.005 fc = 0.005
13
10

12
10
PSD

curvature
11 biased
10

Thomson MT Quadratic MT
10
10

-3 -2 -1 -3 -2 -1
10 10 10 10 10 10
Frequency (Hz) Frequency (Hz)

Figure 5.9: Spectrum analysis of synthetic earthquake model. The time series (not
shown) has 1000 samples (N = 1000) and we use time-bandwidth product of 3.5. The
corner frequency fc = 0.005Hz (shown by an arrow) is to be estimated from the com-
puted spectrum. Note how Thomson’s algorithm biases the lower frequency part of the
spectrum and does not resemble the triangular shape expected for these kind of signals.
The QMT reduces the bias at lower frequencies considerably and is a better description
of the shape of the spectrum around fc .

10 PSD Derivative
PSD Derivative (Normalized)

8 Estimated Derivative
2 std error bound
6
4
fc ~ 0.01
2
0
-2
fc ~ 0.005
-4

.001 .01 .1
Frequency (Hz)

Figure 5.10: Mean derivative of the spectrum from 100 random realizations and two
standard error bounds for two different earthquake models (see equation 5.36), with
corner frequencies fc = 0.005 and fc = 0.01Hz. Time series analyzed is 1000 samples
long. Note how the estimate closely resembles the model. In the case of the lower corner
frequency, there is considerable bias at the lower frequency band.
 104

additional degrees of freedom for estimating parameters from the spectrum.

5.6.5 Bathymetry profiles

A simple, isotropic, three-parameter model for the power spectrum of marine

topography has been proposed of the form (Goff and Jordan, 1988):

a4
S(k) = (5.37)
(1 + (k/kc )2 )µ

where a is the amplitude of the total root-mean-square roughness of the topography,

k = |k| is the wavenumber, µ > 1 is the slope in the roll-off in the short wavelength part
of the spectrum, and kc is the corner wavenumber. We decided to use bathymetry data,
given their more closely isotropic behavior (stationarity in terms of time series) and the
availability of very high quality data sets.
We used bathymetry data obtained in the Central Pacific region (See Figure
5.11), drawn from ship multibeam data (Macdonald et al., 1992, see also Marine Geo-
science Data System, http://www.marine-geo.org). From the available data we chose 5
profiles (east-west directions), four of them crossing the mid-ocean ridge, the other along
the transform fault.
The idea of this example is to show what extra information can be extracted
via the QMT. The bathymetry profiles have a sampling rate of 500 samples per deg and
a total of 1251 samples per profile. The location of the profiles is shown in Figure (5.11).
Figure (5.12) shows the QMT analysis of the selected profiles. The spectrum
is shown to have a large dynamic range (about 7 orders of magnitude) over the entire
frequency range. We focus our attention at the lower frequency range, where the corner
wavenumber is expected from the model in Equation (5.37). The spectral shapes are very
similar for all profiles and the different k0 ’s are hardly distinguishable. An independent
observation of the different behavior between profiles can be drawn from the derivative
estimates. See Figure (5.12) and caption for discussion.
From the methods described above, we can obtain estimates of the derivative
of the spectrum, and by normalizing by the spectrum,

S 0 (f ) d
= {log S(f )} (5.38)
S(f ) df
 105

Depth (m)
4500 3500 2500 1500

10.0º
Profile A

Profile B

Profile C
9.0º
Profile D

Profile E

8.0º

-105º -104º -103º

Figure 5.11: Location of the study area, where the profiles were taken from. Four of
these profiles run across the mid-ocean ridge, and one is paralel to the transform fault.
Location of the profiles is shown as thick black lines.

we have an estimate of the derivative of the log spectrum (Thomson, 1994). The deriva-
tives also provide the means for comparing the different profiles, and clearly show the
presence of two groups with particular spectral characteristics. This suggests that the
profiles sampling the transform fault posses a lower corner frequency than the profiles
sampling the mid-ocean ridge structure.

5.7 Conclusions

Multitaper spectral analysis can be obtained by multiplying the data by a set of

orthogonal (in time and frequency) sequences, all having good spectral leakage properties.
The sequences have the property to concentrate within a band 2W the frequency content
of the spectral estimate. A simple average of the eigenspectra Ŝk is not ideal, given the
large dynamic ranges of some signals, and an adaptive weighting function is necessary,
especially in regions where the spectrum has low amplitudes, and thus is prone to leakage
from frequencies that have much larger amplitudes.
As noted by Riedel and Sidorenko (1995) and confirmed in this study, in regions
where spectral leakage is not expected, corresponding to the regions of the spectrum with
large amplitudes, the local or quadratic bias can have an important effect on the shape
 106

105

103
PSD (m2 * deg)

101

10-1

10-3

100 101 102

Wavenumber (deg-1)

105 Derivative of ln(PSD)

2
PSD (m2 * deg)

104
d[log(PSD)] / df

103
-2

102 -4
100 101 100 101
Wavenumber (deg-1) Wavenumber (deg-1)

Figure 5.12: Spectral Analysis of 5 selected profiles of bathymetric data in the Central
Pacific Ocean around 9◦ N (see map in Figure 5.11). QMT (left plots) show very similar
behavior of the spectra. In addition to the standard spectra, the algorithm provides
an estimate of the derivative of the spectra (right plot). We show the scaled derivative
(approximately the derivative of ln(PSD). Note that the spectra can be grouped accord-
ing to the derivatives, with one profile having a quite different behavior (magenta line)
showing a lower corner frequency. This derivative corresponds to profile E in Figure 5.11,
which samples the transform fault; while the rest of the profiles sample the mid-ocean
ridge.
 107

of the spectrum.
We introduce the Quadratic multitaper method, which estimates the derivatives
of the spectrum, that is, the slope and curvature of the spectrum on the interval (−W, W ),
by solving a parameter estimation problem relating the derivatives of the spectrum and
the K eigencomponents.
With the estimation of the second derivative (the curvature of the spectrum)
we can apply a correction to the spectrum to obtain a new estimate that is unbiased to
quadratic structure. This algorithm reduces to the original Thomson (1982) multitaper
when the spectrum is locally flat in the interval (−W, W ).
We present a variety of examples that indicate that the Quadratic multitaper
provides a smoother, less biased spectral estimate of the data, in addition to independent
estimates of the derivatives of the spectrum. When the dynamic range of the signal is
very large, the improvements are not as striking, but the information contained in the
slope estimates can readily be applied in parameter estimation, or as an additional
discriminant to compare two signals. No additional spectral leakage was introduced in
the examples shown in this study.
We also discuss the effect of chosen multitaper parameters such as the time-
bandwidth product and the number of tapers to compute. Even though it is still a
user-defined set of parameters, we show that the Quadratic multitaper leads to increased
resolution compared to TMT and it is less dependent on the choice of the time-bandwidth
in the inner interval. It allows the use of more tapers without the loss of resolution power
compared to Thomson’s algorithm.
Finally, model parameters can be found by analysing the goodness-of-fit be-
tween a Quadratic spectral estimate plus the slope of the spectrum of a data set with
a theoretical model of the spectrum and its derivative. Another approach would be to
generate from the theoretical model a covariance matrix Cjk (as in Equation 5.27) and
find the model that best fits the data. In the later case, the information is not restricted
to curvature; rather all information from the theoretical models is used.
 108

Acknowledgments

We thank David Sandwell and J.J. Becker for providing the bathymetry data
used here. Funding for this research was provided by NSF Grant number EAR0417983.
6

Conclusions

6.1 Main Results

This thesis presents new methods to estimate and reduce the uncertainties and
biases that are unavoidable in the analysis of the earthquake rupture process. The meth-
ods include spectral stacking (Prieto et al., 2004, Chapter 3), empirical Green function
(EGF) analysis (Prieto et al., 2006, Chapter 4), source spectral fitting and uncertainty
estimation (Prieto et al., 2007a, Chapter 2), and improving analysis tools to reduce bias
and variance in spectrum estimates (Prieto et al., 2007b, Chapter 5). We have presented
three basic results:

1. Measuring the uncertainties in earthquake source parameter estimation.

2. Reducing these uncertainties by stacking or averaging spectral estimates.

3. Generating a multitaper spectral estimate with significant bias reduction, with

extra independent information from the slope of the PSD.

As discussed in the introduction (see Figure 1.6), in order to compare different

studies of the earthquake rupture it is useful to have a measure of some static and
dynamic source parameters, as well as their associated uncertainties. The error analysis
is necessary if issues regarding scaling relations and source parameter size dependence
are to be discussed.
In Chapter 2 we use a multitaper spectrum algorithm combined with jackknife
statistical analysis to obtain, from the seismic spectrum, the source parameters and their

109
 110

confidence intervals. Analysis of the Cajon Pass Borehole finds a slight increase in stress
drop with earthquake magnitude, deviating from the constant stress drop expected if
earthquake ruptures are self-similar.
Two additional observations from the results are relevant. First, stations close
to each other (i.e., two borehole stations at 2.9 km and 1.5 km depths) may produce
very different estimates of the source parameters, indicating the need to address the
significance of individual estimates when comparing results. Second, given the assump-
tions used in the analysis (radiation pattern, earthquake locations, velocity model, etc.),
it is likely that the confidence intervals presented here represent a lower bound. This
suggests we need to find ways of reducing uncertainties, and the subsequent chapters
seek to address this point.
In Chapter 3, we study the scaling relationships of source parameters and
the self-similarity seen in 400 small earthquakes located in a compact region near the
Anza Seismic Network in southern California. By iteratively stacking P and S spectra,
we are able to separate source and receiver contributions and, after an EGF correction,
obtain a relative earthquake source spectrum. Due to the large number of events and
multiple stations, the stacked source spectra are smooth compared to individual earth-
quake estimates. We show using standard scaling relationships that for the magnitude
range ML = 1.8 to 3.4 the earthquake rupture is self-similar. The static scaling relation
between the size of the earthquake – given by the seismic moment M0 – and the length
scale of the rupture measured from the corner frequency fc is M0 ∝ fc−3 . The dynamic
parameter apparent stress σa shows an average value of 1MPa and is constant as a func-
tion of M0 . This is confirmed by directly testing for self-similarity with the spectral
shapes at various seismic moments. We applied a similar methodology to a comprehen-
sive analysis of over 60,000 earthquakes in southern California in Shearer, Prieto, and
Hauksson (2006).
The empirical Green function, where a small earthquake is taken as an approx-
imate impulse response between a source and a receiver, has been shown in multiple
cases to be a very useful method. It has a number of drawbacks, however, with one of
them being that the uncertainties after performing the EGF deconvolution (in time or
frequency domains) tend to be very large.
 111

In Chapter 4 we take advantage of the large data set available in southern

California to average spectral ratios of the 2001 M 5.1 Anza earthquake and 160 after-
shocks. Due to the large number of EGFs used, we can use a weighted averaging of the
spectral ratios and are able to control the trade-off between the variance reduction and
the bias inherent in the EGF approximation by using a mean-square error criteria. Using
propagation of errors, we also obtain the confidence intervals of the source spectrum and
dynamic source parameters (radiated seismic energy ES and apparent stress σa ).
The results of this study show that in order to reduce uncertainties, it is nec-
essary to use multiple EGFs; otherwise, even if an ideal EGF is available, the large
uncertainties would make the result less significant. Note in Figure 1.6 that the error
bars of the σa estimate for this particular study are almost invisible in the given scale,
while this is not so for the majority of the other data points in the same plot.
In Chapter 5 we present an improvement to the multitaper spectral analysis
(TMT) tools introduced by Thomson (1982), which we have used throughout this thesis.
As discussed in this chapter, the TMT algorithm suffers from local bias, i.e., at frequencies
close to spectral lines the spectrum estimate is approximately constant. The Quadratic
multitaper (QMT) introduced can be used to estimate the derivatives of the spectrum
as a function of frequency. The estimate of the second derivative is used to apply a
correction and obtain a spectrum that is unbiased with respect to quadratic structure.
In addition, we present the use of the first derivative – the slope – of the spectrum for
comparing signals from marine topography profiles.
In order to use the multitaper algorithm (either TMT or QMT), it is necessary
to define two basic parameters: a frequency bandwidth (W ) over which the user wishes to
average the spectrum and the number (K) of tapers to compute. If the user chooses a very
narrow bandwidth or very few tapers, the spectral estimate will have large uncertainties
and will not be smooth. A considerable advantage of the QMT presented in Chapter 5
is that for the same choice of parameters (W and K), the spectrum has higher resolution
and is smoother than TMT estimates. We also show that the user may choose to compute
one more taper without reducing the resolution of the estimate compared to TMT and
since one more taper is used, an even smoother estimate is expected.
 112

6.2 Future Research Directions

The methodologies developed during this research should provide considerable

advantages in studying the physics of earthquake rupture from seismic recordings. They
provide improved estimation of some earthquake source parameters such as stress drop
and radiated seismic energy. The methods proposed to reduce uncertainties can easily
be applied to different data sets in which either lots of earthquakes or lots of stations are
available. In the applications we presented mainly small and medium sized earthquakes
(M < 6.0) analyzed at either local or regional distances.
The stacking or averaging methods are not useful for large teleseismic earth-
quakes, given that at teleseismic distances it is rare to record many small earthquakes
near a large one. Another difficulty at teleseismic distances is that large earthquakes rup-
ture over tens of kilometers, requiring different EGFs for different patches of the fault,
and complicating any averaging between these patches. Nevertheless, EGF method-
ologies have been adapted for prediction or simulation of ground motion from large
earthquakes using small earthquakes as EGFs (e.g., Wössner et al., 2002).
One particularly interesting approach is the analysis of earthquake data recor-
ded by small aperture arrays. The idea is to look at the consistency of spectral analysis
across the array. If two earthquakes – one large and one small that can be used as an
EGF – are recorded by the stations of the small aperture array, we can test the stability
of the EGF deconvolution. It is assumed that all stations are sampling an identical region
of the focal sphere and that the EGF is removing the effects of attenuation, near-site
effects, scattering, etc., so the result should be identical for every station. We can then
test the uncertainties and resolving power of the stations in the array, given that the
Quadratic multitaper (QMT) will provide the least variable, least biased estimate, which
will then be used in the spectral division.
With the advent of high-quality borehole seismic networks, it is now possible to
investigate the rupture properties of smaller earthquakes. The increasingly available data
from the Japanese Hi-net and F-net networks (with over 700 stations), the EarthScope
PBO Borehole Strainmeter and Seismometer network with 7 stations near the Anza
Seismic Gap, and the SAFOD array, with seismic sensors along the borehole and some
 113

sensors a few hundred meters from known seismic sources in the San Andreas Fault can
be exploited using the methodologies described in this thesis.
Using EGF methods, McGuire (2004) showed that it was possible to investigate
the fault plane, rupture length, and directivity of M5 earthquakes by using a network
of surface stations. Having the borehole data described above, which are expected to
have better signal-to-noise ratios and record signals at higher frequencies than surface
instruments, we have the opportunity to investigate the rupture properties of even smaller
earthquakes (M2 - M4). The network of stations is needed to analyze the variability of
source properties as a function of the take-off angle (as a function of the focal sphere
sampled at a given station), helpful in solving the fault plane ambiguity for a double-
couple source and providing information about directivity of small earthquakes.
Another important application is the use of the methodologies described here
for the analysis of attenuation structure, which we have in part neglected. From the
stacking procedures or from EGF analysis, one can potentially obtain information about
the properties of the medium through which the seismic waves have traveled, i.e., Q
structure (attenuation) or near-site effects (see for example Warren and Shearer, 2000;
Tsuda et al., 2006; Hauksson and Shearer, 2006). Variations in Q structure may be
caused by cracks, chemical composition and temperature variations, which can then be
modeled from the spectral analysis.
There are some limitations to using the Quadratic multitaper. As explained
in Chapter 5, QMT provides an estimate of the power spectrum and we lose the phase
information. This clearly limits our ability to investigate the source time function (STF)
and the behavior of the radiated seismic energy in time. For example Mori et al. (2003)
use the STF to investigate the dynamic stress drops or average frictional stress ave{σf }
(see Figure 1.2) by looking at the initial slope of the deconvolved source time functions.
Nevertheless, estimation of radiated seismic energy in the spectral domain might be
better constrained by using multitaper methods, where uncertainties are easily quanti-
fied, and even STF can potentially be extracted if Thomson’s multitaper is used (for an
example deconvolution applied to receiver functions see: Park and Levin, 2000).
A

Propagation of errors

Some rules of propagation of errors (Taylor, 1997) are listed here. Assume u, v
are random variables with associated variance σu2 , σv2 , and covariance σuv
2 . Constants a, b

do not contribute to uncertainties. Then we have:

x = au ± bv σx2 = a2 σu2 + b2 σv2 + 2abσuv

2

σx2 σu2 σv2 2

σuv
x = auv = + + 2
x2 u2 v2 uv

au σx2 σu2 σv2 2

σuv
x= = + − 2
v x2 u2 v2 uv
σx σu
x = au±b = b
x u
σx
x = ae±bu = bσu
x
σu
x = a ln(±bu) σx = a
u

114
B

Quadratic mean-square error

In Equation (5.32), a correction for the curvature or quadratic bias is applied

to the multitaper estimate. In section 5.5, we defined the expected value of the spectrum
as an average over the inner interval (−W, W ):
h i 1
E Ŝ(f ) = S(f ) + W 2 S 00 (f ) (B.1)
6

and by applying the correction in 5.32, we have the expected value of the Quadratic
multitaper
h i 1 1
E S̃(f ) = S(f ) + W 2 S 00 (f ) − µ W 2 S 00 (f ) (B.2)
6 6
where we assume that E[α̂2 ] = S 00 (f ).
The bias of the Quadratic multitaper is then
h i 1
E [β] = E S̃(f ) − S(f ) = W 2 S 00 (f )(1 − µ) (B.3)
6

and the variance, using the rules of propagation of errors (Taylor, 1997) in Equation
5.32,
W4
var{S̃} = var{Ŝ} + µ2 var{S 00 } (B.4)
36
and we can now define the mean square error (bias2 + variance):
2
W 2 00 W4

L = (1 − µ) S + var{Ŝ} + µ2 var{S 00 } (B.5)
6 36

where the first term is the bias squared and the two on the right represent the variance.
It is assumed that the covariance is insignificant. Taking the derivative with respect to

115
 116

µ and setting to zero

∂L  2
= (1 − µ) S 00 + µ var{S 00 } = 0 (B.6)
∂µ
and rearranging yields
2
[S 00 ]
µ= 2 (B.7)
([S 00 ] + var{S 00 })
which is the solution shown in 5.33, using α2 as estimates of S 00 .
To obtain the variance of the estimates α̂0 , α̂1 , and α̂2 in the least squares
problem (5.30), we compute the covariance matrix of the coefficients, following Lawson
and Hanson (1974).
Bibliography

Abercrombie, R. E. (1995). Earthquake source scaling relationships from −1 to 5 ML

using seismograms recorded at 2.5-km depth. J. Geophys. Res., 100, 24015–24036.
Abercrombie, R. E. (1996). The magnitude-frequency distribution of earthquakes recor-
ded with deep seismometers at Cajon Pass, southern California. Tectonophysics, 261,
1–7.
Abercrombie, R. E. (1997). Near-surface attenuation and site effects from comparison
of surface and deep borehole recordings. Bull. Seism. Soc. Am., 87(3), 731–744.
Abercrombie, R. E. and Leary, P. (1993). Source parameters of small earthquakes re-
corded at 2.5 km depth, Cajon Pass, southern California: Implications for earthquake
scaling. Geophysics Research Letters, 20, 1511–1514.
Abercrombie, R. E. and Rice, J. R. (2005). Can observations of earthquake scaling
constrain slip weakening? Geophys. J. Int., 162, 406–424.
Aki, K. (1967). Scaling law of seismic spectrum. J. Geophys. Res., 72, 1217–1231.
Aki, K. and Richards, P. (1980). Quantitative Seismology. W. H. Freeman, New York.
Andrews, D. J. (1986). Objective determination of source parameters and similarity of
earthquakes of different size. In S. Das, J. Boatwright, and C. H. Scholz, editors, Earth-
quake Source Mechanics, Geophys. Monogr. Ser., pages 259–267. AGU, Washington,
DC.
Archuleta, R. J., Cranswick, E., Mueller, C., and Spudich, P. (1982). Source parameters
of the 1980 mammoth lakes, california, earthquake sequence. J. Geophys. Res., 87(B6),
4595–4607.
Bard, Y. (1974). Nonlinear parameter estimation. Academic Press, New York.
Berger, J., Davis, P., and Ekström, G. (2004). Ambient earth noise: A survey of the
global seismographic network. J. Geophys. Res., 109(B11307).
Berger, J. L., Baker, L., Brune, J., Fletcher, J., Hanks, T., and Vernon, F. L. (1984).
The Anza array: A high dynamic range, broadband, digitally radiometered seismic
array. Bull. Seism. Soc. Am., 74, 1469–1481.
Boatwright, J. A. (1980). A spectral theory for circular seismic sources: Simple estimates
of source dimension, dynamic stress drop and radiated energy. Bull. Seism. Soc. Am.,
70, 1–27.

117
 118

Boatwright, J. A. and Fletcher, J. B. (1984). The partition of radiated energy between

P and S waves. Bull. Seism. Soc. Am., 74, 361–376.

Brodsky, E. E. and Kanamori, H. (2001). Elastohydrodynamic lubrication of faults. J.

Geophys. Res., 106, 16357–16374.

Bronez, T. P. (1992). On the performance advantage of multitaper spectral analysis.

IEEE Trans. Sig. Proc., 40(12), 2941–2946.

Brune, J., Brown, S., and Johnson, P. (1993). Rupture mechanics and interface sep-
aration in foam rubber models of earthquakes: a possible solution to the heat flow
paradox and the paradox of large overthrusts. Tectonophysics, 218, 59–67.

Brune, J. N. (1970). Tectonic stress and seismic shear waves from earthquakes. J.
Geophys. Res., 75, 4997–5009.

Chappellaz, J., Barnola, J. M., Raynaud, D., Korotkevich, Y. S., and Lorius, C. (1990).
Ice-core record of atmospheric methane over the past 160,000 years. Nature, 345,
127–131.

Choy, G. L. and Boatwright, J. L. (1995). Global patterns of radiated seismic energy

and apparent stress. J. Geophys. Res., 100, 18205–18228.

Constable, S. and Constable, C. (2004). Observing geomagnetic induction in magnetic

satellite measurements and associated implications for mantle conductivity. Geochem.
Geophys. Geosyst., 5, Q01006.

Cramér, H. (1940). On the theory of stationary random processes. Ann. of Math., 41,
215–230.

Daly, E., Brown, C., Stark, C. P., and Ebinger, C. J. (2004). Wavelet and multita-
per coherence methods for assesing the elastic tichkness of the Irish Atlantic margin.
Geophys. J. Int., 159, 445–459.

Davidson, A. C. and Hinkley, D. V. (1997). Bootstrap Methods and their Application.

Cambridge University Press, New York.

Denison, D. G. T., Walden, A. T., Balogh, A., and Forsyth, J. (1999). Multitaper testing
of spectral lines and the detection of the solar rotation frequency and its harmonics.
Appl. Statist., 48, 427–439.

Doob, J. L. (1952). Stochastic processes. John Wiley and Sons, New York.

Douglas, A., Hudson, J. A., and Pearce, R. G. (1988). Directivity and the Doppler effect.
Bull. Seism. Soc. Am., 78, 1367–1372.

Efron, B. (1982). The Jackknife, the Bootstrap, and other resampling plans. SIAM,
Philadelphia.

Eshelby, J. D. (1957). The determination of the elastic field of an ellipsoidal inclusion

and related problems. Proc. Roy. Soc. London Series A, 241, 376–396.
 119

Fletcher, J., Boatwright, J., Haar, Hanks, and McGarr, A. (1984). Source parameters
for aftershocks of the oroville california earthquake. Bull. Seism. Soc. Am., 74(4),
1101–1123.
Fox, T., Hinkley, D. V., and Larntz, K. (1980). Jackknifing in nonlinear regresion.
Technometrics, 22, 29–33.
Frankel, A., Fletcher, J., Vernon, F., Haar, L., Berger, J., and Hanks, T. (1986). Rupture
characteristics and tomographic source imaging of ML ∼ 3 earthquakes near Anza,
Southern California. J. Geophys. Res., 91, 12633 – 12650.
Gilbert, F. (1970). Excitation of normal modes of the Earth by earthquake sources.
Geophys. J. Royal Astr. Soc., 22, 223–226.
Goff, J. A. and Jordan, T. H. (1988). Stochastic modeling of seafloor morphology:
inversion of sea beam data for second-order statistics. J. Geophys. Res., 93, 13589–
13608.
Hanks, T. C. and Boore, D. M. (1984). Moment-magnitude relations in theory and in
practice. J. Geophys. Res., 89, 6229 – 6235.
Harris, F. (1978). On the use of windows for harmonic analysis with the discrete Fourier
transform. Proceedings of the IEEE , 66, 51–83.
Hartzell, S. H. (1978). Earthquake Aftershocks as Greens Functions. Geophysics Research
Letters, 5, 1–4.
Hauksson, E. and Shearer, P. M. (2006). Attenuation models (QP andQS ) in three
dimensions of the southern California crust: Inferred fluid saturation at seismogenic
depths. J. Geophys. Res., 111, B05302.
Hauksson, E., Jones, L. M., Perry, S., and Hutton, K. (2002). Emerging from the Stress
shadow of the 1992 Mw7.3 Landers southern California earthquake? A preliminary
Assessment. Seismol. Res. Let., 73, 33–38.
Hinkley, D. V. (1977). Jackknife confidence limits using Student t approximations.
Biometrika, 64, 21–28.
Hough, S. E. (1997). Empirical Green’s function analysis: Taking the next step. J.
Geophys. Res., 102, 5369 – 5384.
Hough, S. E. and Field, E. H. (1996). On the coherence of ground motion in the San
Fernando Valley. Bull. Seism. Soc. Am., 86(6), 1724–1736.
Ida, Y. (1972). Cohesive force across the tip of a longitudinal-shear crack and Griffiths
specific surface energy. J. Geophys. Res., 77, 3796–3805.
Ide, S. and Beroza, G. C. (2001). Does apparent stress vary with earthquake size?
Geophysics Research Letters, 28, 3349–3352.
Ide, S., Beroza, G. C., Prejean, S. G., and Ellsworth, W. L. (2003). Apparent break
in earthquake scaling due to path and site effects on deep borehole recordings. J.
Geophys. Res., 108(B5 2271).
 120

Izutani, Y. and Kanamori, H. (2001). Scale dependence of seismic energy-to-moment

ratio for strike-slip earthquakes in Japan. Geophysics Research Letters, 28, 4007–
4010.

Jeffreys, H. (1942). On the mechanics of faulting. Geol. Mag., 79, 291.

Kanamori, H. (1977). The energy release in great earthquakes. J. Geophys. Res., 82,
2981–2987.

Kanamori, H. (2005). Real-time seismology and earthquake damage mitigation. Annu.

Rev. Earth Planet. Sci., 33, 195–214.

Kanamori, H. and Anderson, D. L. (1975). Theoretical basis of some empirical relations

in seismology. Bull. Seism. Soc. Am., 65(5), 1073–1095.

Kanamori, H. and Brodsky, E. E. (2004). The physics of earthquakes. Reports on

Progress in Physics, 67, 1429–1496.

Kanamori, H. and Heaton, T. H. (2000). Microscopic and macroscopic physics of earth-

quakes. In J. Rundle, D. L. Turcotte, and W. Kein, editors, Geocomplexity and the
Physics of Earthquakes, pages 147–155, Washington, D. C. American Geophysical
Union. Monograph 120.

Kanamori, H. and Rivera, L. (2004). Static and dynamic scaling relations for earthquakes
and their implications for rupture speed and stress drop. Bull. Seism. Soc. Am., 94,
314–319.

Kanamori, H. and Rivera, L. (2006). Energy partitioning during an earthquake. In

Abercrombie, McGarr, Kanamori, and diToro, editors, Earthquakes: Radiated Energy
and the Physics of Faulting, pages 3–13. AGU Geophys. Monograph 170.

Kanamori, H., Mori, J., Hauksson, E., Heaton, E., Hutton, T., and Jones, K. (1993).
Determination of earthquake energy release and ML using terrascope. Bull. Seism.
Soc. Am., 83(2), 330–346.

Korte, M., Constable, C. G., and Parker, R. L. (2002). Revised magnetic power spectrum
of the oceanic crust. J. Geophys. Res., 107(B9), 2205.

Lachenbruch, A. H. (1980). Frictional heating, fluid pressure and the resistance to fault
motion. J. Geophys. Res., 85, 6097–6112.

Lawson, C. L. and Hanson, R. J. (1974). Solving Least Squares Problems. Prentice Hall,
Englewood Cliffs, NJ.

Lees, J. (1995). Reshaping spectrum estimates by removing periodic noise: Application

to seismic spectral ratios. Geophysics Research Letters, 22(4), 513–516.

Lees, J. and Park, J. (1995). Multiple-taper spectral analysis: a stand-alone C-

subroutine. Compt. Geosci., 21(2), 199–236.
 121

Macdonald, K. C., Fox, P., Miller, S., Carbotte, S., Edwards, M., Eisen, M., Fornari, D.,
Perram, L., Pockalny, R., Scheirer, D., Tighe, S., Weiland, C., and Wilson, D. (1992).
The East Pacific Rise and its flanks 8-18N: History of segmentation, propagation and
spreading direction based on SeaMARC II and SeaBeam Studies. Marine Geophys.
Researches, 14, 299–344.
Madariaga, R. (1976). Dynamics of an expanding circular fault. Bull. Seism. Soc. Am.,
66, 639–666.
Mason, J. and Handscomb, D. (2003). Chebyshev Polynomials. Chapman & Hall/CRC
Press, Boca Raton, FL.
Mayeda, K. and Walter, W. R. (1996). Moment, energy, stress drop, and source spectra
of Western United States earthquakes from regional coda envelopes. J. Geophys. Res.,
101, 11195–11208.
Mayeda, K., Dreger, D. S., Walter, W. R., and Tajima, F. (2004). Bdsn calibration for
northern california earthquakes from coda-derived source spectra: Moment magnitude
and radiated energy (abstract). Seism. Res. Let., 75, 278.
McGarr, A. (1999). On relating apparent stress to the stress causing earthquake fault
slip. J. Geophys. Res., 104, 3001–3003.
McGuire, J. J. (2004). Estimating the finite source properties of small earthquake rup-
tures. Bull. Seism. Soc. Am., 94, 377–393.
Melosh, J. (1979). Acoustic fluidization: a new geologic process? J. Geophys. Res., 84,
7512–7520.
Miller, R. G. (1974). The jackknife – a review. Biometrika, 61, 1–15.
Mori, J., Kanamori, H., and Abercrombie, R. E. (2003). Stress drops and radiated
energies of aftershocks of the 1994 Northridge, California, earthquake. J. Geophys.
Res., 108(B11 2545).
Mueller, C. S. (1985). Source pulse enhancement by deconvolution of an empirical Green’s
function. J. Geophys. Res., 12(1), 33–36.
Palmer, A. C. and Rice, J. R. (1973). The growth of slip surfaces in the progressive
failure of overconsolidated clay. Proc. R. Soc. Lond., A, 332, 527–548.
Park, J. and Levin, V. (2000). Receiver functions from multiple-taper spectral correlation
estimates. Bull. Seism. Soc. Am., 90, 1507–1520.
Park, J., Lindberg, C. R., and Thomson, D. J. (1987a). Multiple-taper spectral analysis
of terrestial free oscillations: part I. Geophys. J. Royal Astr. Soc., 91, 755–794.
Park, J., Lindberg, C. R., and Vernon, F. L. (1987b). Multitaper spectral analysis of
high-frequency seismograms. J. Geophys. Res., 92, 12675–12684.
Park, J. J. (1992). Envelope estimation for quasi-periodic geophysical signals in noise:
a multitaper approach. In A. T. Walden and P. Guttorp, editors, Statistics in the
Enviroment & Earth Sciences, pages 189–219. Edward Arnold, London.
 122

Percival, D. B. and Walden, A. T. (1993). Spectral analysis for physical applications:

Multitaper and Conventional univariate techniques. Cambridge Univ. Press.

Pérez-Campos, X. and Beroza, G. C. (2001). An apparent mechanism dependence of

radiated seismic energy. J. Geophys. Res., 106(B6), 11127–11136.

Pérez-Campos, X., Singh, S. K., and Beroza, G. C. (2003). Reconciling teleseismic and
regional estimates of seismic energy. Bull. Seism. Soc. Am., 93(2), 2123–2130.

Prejean, S. G. and Ellsworth, W. L. (2001). Observations of earthquake source param-

eters and attenuation at 2 km depth in the Long Valley Caldera, eastern California.
Bull. Seism. Soc. Am., 95, 165–177.

Prieto, G. A., Shearer, P. M., Vernon, F. L., and Kilb, D. (2004). Earthquake source
scaling and self-similarity estimation from stacking P and S spectra. J. Geophys. Res.,
109(B08310).

Prieto, G. A., Parker, R. L., Vernon, F. L., Shearer, P. M., and Thomson, D. J. (2006).
Uncertainties in earthquake source spectrum estimation using empirical Green func-
tions. In Abercrombie, McGarr, Kanamori, and diToro, editors, Earthquakes: Radiated
Energy and the Physics of Faulting, pages 69–74. AGU Geophys. Monograph 170.

Prieto, G. A., Thomson, D. J., Vernon, F. L., Shearer, P. M., and Parker, R. L. (2007a).
Confidence intervals for earthquake source parameters. Geophys. J. Int., 168, 1227–
1234.

Prieto, G. A., Parker, R. L., Thomson, D. J., Vernon, F. L., and Graham, R. L. (2007b).
Quadratic multitaper spectrum. Submitted to Geophys. J. Int.

Quenouille, M. (1949). Approximate tests of correlation in time series. J. R. Stat. Soc.,

B11, 18–84.

Rice, J. A. (1995). Mathematical Statistics and Data Analysis. Duxbury Press, Belmont,
CA, 2nd. edition.

Riedel, K. S. and Sidorenko, A. (1995). Minimum bias multiple taper spectral estimation.
IEEE Trans. on Signal Processing, 43, 188–195.

Riedel, K. S., Sidorenko, A., and Thomson, D. J. (1993). Spectral estimation of plasma
fluctuations i. comparison of methods. Phys. Plasma, 1(3), 485–500.

Scott, J. S., Masters, T. G., and Vernon, F. L. (1994). 3-D velocity structure of the San
Jacinto fault zone near Anza, California—I. P waves. Geophys. J. Int., 119, 611–626.

Shearer, P. M. (1999). Introduction to Seismology. Cambridge University Press, Cam-

bridge.

Shearer, P. M., Hauksson, E., and Lin, G. (2005). Southern California hypocenter
relocation with waveform cross-correlation, Part 2: results using sourcespecific station
terms and cluster analysis. Bull. Seism. Soc. Am., 95, 904–915.
 123

Shearer, P. M., Prieto, G. A., and Hauksson, E. (2006). Comprehensive analysis of

earthquake source spectra in southern California. J. Geophys. Res., 111(B06303).

Sibson, R. (1973). Interactions between temperature and pore-fluid pressure during

earthquake faulting and a mechanism for partial or total stress relief. Nature, 243,
66–68.

Singh, S. K. and Ordaz, M. (1994). Seismic energy release in mexican subduction zone
earthquakes. Bull. Seism. Soc. Am., 84, 1533–1550.

Slepian, D. (1978). Prolate spheroidal wavefunctions, Fourier analysis, and uncertainty

V: the discrete case. Bell System Tech. J., 57, 1371–1429.

Sonley, E. and Abercrombie, R. E. (2006). Variability introduced by methods of attenu-

ation correction: the effect on source parameter determination. In Abercrombie, Mc-
Garr, Kanamori, and diToro, editors, Earthquakes: Radiated Energy and the Physics
of Faulting. AGU Geophys. Monograph 170.

Taylor, J. R. (1997). An introduction to Error Analysis. University Science Books,

Sausalito, CA.

Thomson, D. J. (1982). Spectrum estimation and harmonic analysis. In Proceedings of

the IEEE , volume 70, pages 1055–1096.

Thomson, D. J. (1990). Quadratic-Inverse spectrum estimates: applications to paleocli-

matology. Phys. Trans. R. Soc. London A, 332, 539–597.

Thomson, D. J. (1994). An overview of multiple-window and quadratic-inverse spectrum

estimation methods. In ICASSP - 94 , volume VI, pages 185–194.

Thomson, D. J. and Chave, A. D. (1991). Jackknife error estimates for spectra, coher-
ences, and transfer functions. In S. Haykin, editor, Advances in Spectrum Analysis
and Array Processing, volume 1 chapter 2, pages 58–113. Prentice Hall.

Tinti, E., Spudich, P., and Cocco, M. (2005). Earthquake fracture energy inferred from
kinematic rupture models on extended faults. J. Geophys. Res., 110, B12303.

Tsuda, K., Archuleta, R. J., and Koketsu, K. (2006). Quantifying the spatial distribution
of site response by use of the Yokohama High-Density Strong-Motion Network. Bull.
Seism. Soc. Am., 96(3), 926–942.

Tukey, J. W. (1958). Bias and confidence in not-quite large samples. Ann. Math. Stat.,
29, 614.

Tukey, J. W. (1960). Conclusions vs decisions. Technometrics, 2(4), 423–433.

Venkataraman, A. and Kanamori, H. (2004). Effect of directivity on estimates of radiated

seismic energy. J. Geophys. Res., 109(B04301).

Venkataraman, A., Kanamori, H., and Rivera, L. (2002). Radiated energy from the 16
October 1999 Hector Mine Earthquake: Regional and Teleseismic Estimates. Bull.
Seism. Soc. Am., 92, 1256–1265.
 124

Venkataraman, A., Boatwright, J., and Beroza, G. C. (2006). A brief review of techniques
used to estimate radiated seismic energy. In Abercrombie, McGarr, Kanamori, and
diToro, editors, Earthquakes: Radiated Energy and the Physics of Faulting, pages –.
AGU Geophys. Monograph 170.

Vernon, F. L. (1989). Analysis of data recorded on the ANZA seismic network . Ph.D.
thesis, University of California, San Diego.

Vossiliou, M. S. and Kanamori, H. (1982). The energy release in earthquakes. Bull.

Seism. Soc. Am., 72, 371–387.

Warren, L. M. and Shearer, P. M. (2000). Investigating the frequency dependance of

mantle Q by stacking P and P P spectra. J. Geophys. Res., 105, 25391–25402.

Warren, L. M. and Shearer, P. M. (2002). Mapping lateral variations in upper mantle

attenuation by stacking P and P P spectra. J. Geophys. Res., 107(B12 2342).

Wössner, J., Treml, M., and Wenzel, F. (2002). Simulation of MW = 6.0 earthquakes
in the Upper Rhinegraben using empirical Green functions. Geophys. J. Int., 151(2),
487–500.

Wu, C. F. J. (1986). Jackknife, bootstrap and other resampling methods in regression

analysis (with discussion). Ann. Stat., 14, 1261–1350.

Wyss, M. and Brune, J. N. (1968). Seismic moment, stress, and source dimensions for
earthquakes in the california- nevada region. J. Geophys. Res., 73, 4681–4694.