Download ebooks file Signal Processing for Neuroscientists A Companion Volume Advanced Topics Nonlinear Techniques and Multi Channel Analysis Elsevier Insights 1st Edition Drongelen all chapters

Download the full version of the ebook now at ebookultra.
com
Signal Processing for Neuroscientists A

Companion Volume Advanced Topics Nonlinear
Techniques and Multi Channel Analysis
Elsevier Insights 1st Edition Drongelen
https://ebookultra.com/download/signal-processing-
for-neuroscientists-a-companion-volume-advanced-
topics-nonlinear-techniques-and-multi-channel-
analysis-elsevier-insights-1st-edition-drongelen/
Explore and download more ebook at https://ebookultra.com

Recommended digital products (PDF, EPUB, MOBI) that
you can download immediately if you are interested.
Signal Processing for Neuroscientists 1st Edition Wim Van

Drongelen
https://ebookultra.com/download/signal-processing-for-
neuroscientists-1st-edition-wim-van-drongelen/
ebookultra.com
Digital Signal Processing and Applications with the

TMS320C6713 and TMS320C6416 DSK Topics in Digital Signal
Processing 2nd Edition Rulph Chassaing
https://ebookultra.com/download/digital-signal-processing-and-
applications-with-the-tms320c6713-and-tms320c6416-dsk-topics-in-
digital-signal-processing-2nd-edition-rulph-chassaing/
ebookultra.com
Statistical techniques for neuroscientists 1st Edition

Young K. Truong
https://ebookultra.com/download/statistical-techniques-for-
neuroscientists-1st-edition-young-k-truong/
ebookultra.com
Time Frequency Signal Analysis and Processing A

Comprehensive Reference 1st Edition Boualem Boashash
https://ebookultra.com/download/time-frequency-signal-analysis-and-
processing-a-comprehensive-reference-1st-edition-boualem-boashash/
ebookultra.com
Digital Signal Processing Techniques and Applications in
Radar Image Processing 1st Edition Bu-Chin Wang
https://ebookultra.com/download/digital-signal-processing-techniques-
and-applications-in-radar-image-processing-1st-edition-bu-chin-wang/
ebookultra.com
Wavelet Analysis and Transient Signal Processing

Applications for Power Systems 1st Edition Zhengyou He
https://ebookultra.com/download/wavelet-analysis-and-transient-signal-
processing-applications-for-power-systems-1st-edition-zhengyou-he/
ebookultra.com
Signal Processing for Digital Communications Artech House

Signal Processing Library George J. Miao
https://ebookultra.com/download/signal-processing-for-digital-
communications-artech-house-signal-processing-library-george-j-miao/
ebookultra.com
Multi Platinum Pro Tools Advanced Editing Pocketing and

Autotuning Techniques Nathan Adam
https://ebookultra.com/download/multi-platinum-pro-tools-advanced-
editing-pocketing-and-autotuning-techniques-nathan-adam/
ebookultra.com
Multiple Input Multiple Output Channel Models Theory and

Practice Adaptive and Learning Systems for Signal
Processing Communications and Control Series 1st Edition
Nelson Costa
https://ebookultra.com/download/multiple-input-multiple-output-
channel-models-theory-and-practice-adaptive-and-learning-systems-for-
signal-processing-communications-and-control-series-1st-edition-
nelson-costa/
ebookultra.com
Signal Processing for Neuroscientists A Companion
Volume Advanced Topics Nonlinear Techniques and
Multi Channel Analysis Elsevier Insights 1st Edition
Drongelen Digital Instant Download
Author(s): Drongelen, Wim van
ISBN(s): 9781282770294, 1282770292
Edition: 1
File Details: PDF, 6.64 MB
Year: 2010
Language: english
Signal Processing for Neuroscientists,
A Companion Volume
Ae jωt = A[cos(ωt) + j sin(ωt)] = A cos(ωt) + jA sin(ωt)
Av = λv
P(a, b, f, tn)= a cos(2πftn) + b sin(2πftn)
dy
x = RC— + y
dt
y = f (α) + 1– (x – α) f '(α) + 1– (x – α)2 f "(α) + –1 (x – α)3f "'(α) +...

1! 2! 3!
(τ1, τ2) x(t – τ1) x(t – τ2) dτ1dτ2

Signal Processing
for Neuroscientists,
A Companion Volume
Advanced Topics, Nonlinear Techniques
and Multi-Channel Analysis
Wim van Drongelen
AMSTERDAM BOSTON HEIDELBERG LONDON NEW YORK OXFORD

G G G G G
PARIS SAN DIEGO SAN FRANCISCO SINGAPORE SYDNEY TOKYO

G G G G G
Elsevier
32 Jamestown Road London NW1 7BY
30 Corporate Drive, Suite 400, Burlington, MA 01803, USA
First edition 2010
Copyright r 2010 Elsevier Inc. All rights reserved
No part of this publication may be reproduced or transmitted in any form or by any means, electronic or
mechanical, including photocopying, recording, or any information storage and retrieval system, without
permission in writing from the publisher. Details on how to seek permission, further information about
the Publisher’s permissions policies and our arrangement with organizations such as the Copyright
Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/
permissions
This book and the individual contributions contained in it are protected under copyright by the Publisher
(other than as may be noted herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and experience
broaden our understanding, changes in research methods, professional practices, or medical treatment
may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in evaluating and
using any information, methods, compounds, or experiments described herein. In using such information
or methods they should be mindful of their own safety and the safety of others, including parties for
whom they have a professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume
any liability for any injury and/or damage to persons or property as a matter of products liability,
negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas
contained in the material herein.
British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library
Library of Congress Cataloging-in-Publication Data
A catalog record for this book is available from the Library of Congress
ISBN: 978-0-12-384915-1
For information on all Elsevier publications

visit our website at elsevierdirect.com
This book has been manufactured using Print On Demand technology. Each copy is produced to order
and is limited to black ink. The online version of this book will show color figures where appropriate.
Preface
This text is based on a course I teach at the University of Chicago for students
in Computational Neuroscience. It is a continuation of the previously published
text Signal Processing for Neuroscientists: An Introduction to the Analysis of
Physiological Signals and includes some of the more advanced topics of linear and
nonlinear systems analysis and multichannel analysis. In the following, it is assumed
that the reader is familiar with the basic concepts that are covered in the introductory
text and, to help the student, multiple references to the basics are included.
The popularity of signal processing in neuroscience is increasing, and with the
current availability and development of computer hardware and software it may be
anticipated that the current growth will continue. Because electrode fabrication has
improved and measurement equipment is getting less expensive, electrophysiologi-
cal measurements with large numbers of channels are now very common. In addi-
tion, neuroscience has entered the age of light, and fluorescence measurements are
fully integrated into the researcher’s toolkit. Because each image in a movie con-
tains multiple pixels, these measurements are multichannel by nature. Furthermore,
the availability of both generic and specialized software packages for data analysis
has altered the neuroscientist’s attitude toward some of the more complex analysis
techniques. Interestingly, the increased accessibility of hardware and software may
lead to a rediscovery of analysis procedures that were initially described decades
ago. At the time when these procedures were developed, only few researchers had
access to the required instrumentation, but now most scientists can access both the
necessary equipment and modern computer hardware and software to perform com-
plex experiments and analyses.
The considerations given above have provided a strong motivation for the devel-
opment of this text, where we discuss several advanced techniques, rediscover
methods to describe nonlinear systems, and examine the analysis of multichannel
recordings. The first chapter describes two very specialized algorithms: Lomb’s
algorithm to analyze unevenly sampled data sets and the Hilbert transform to detect
instantaneous phase and amplitude of a signal. The remainder of the text can be
divided into two main components: (I) modeling systems (Chapter 2) and the anal-
ysis of nonlinear systems with the Volterra and Wiener series (Chapters 3 5) and
(II) the analysis of multichannel measurements using a statistical approach
(Chapter 6) and examination of causal relationships (Chapter 7). Throughout this
text, we adopt an informal approach to the development of algorithms and we
include practical examples implemented in MATLAB. (All the MATLAB scripts
used in this text can be obtained via http://www.elsevierdirect.com/companions/
9780123849151)
viii Preface
It is a pleasure to acknowledge those who have assisted (directly and indirectly) in

the preparation of this text: Drs. V.L. Towle, P.S. Ulinski, D. Margoliash, H.C. Lee,
M.H. Kohrman, P. Adret, and N. Hatsopoulos. I also thank the teaching assistants for
their help in the course and in the development of the material in this text: thanks,
Matt Green, Peter Kruskal, Chris Rishel, and Jared Ostmeyer. There is a strong cou-
pling between my teaching efforts and research interests. Therefore, I am indebted to
the Dr. Ralph and Marian Falk Medical Research Trust for supporting my research
and to the graduate and undergraduate students in my laboratory: Jen Dwyer, Marc
Benayoun, Amber Martell, Mukta Vaidya, and Valeriya Talovikova. They provided
useful feedback, tested some of the algorithms, and collected several example data
sets. Special thanks to the group of students in the 2010 winter class who helped me
with reviewing this material: Matt Best, Kevin Brown, Jonathan Jui, Matt Kearney,
Lane McIntosh, Jillian McKee, Leo Olmedo, Alex Rajan, Alex Sadovsky, Honi
Sanders, Valeriya Talovikova, Kelsey Tupper, and Richard Williams. Their multiple
suggestions and critical review helped to significantly improve the text and some of
the figures. At Elsevier I want to thank Lisa Tickner, Clare Caruana, Lisa Jones, Mani
Prabakaran, and Johannes Menzel for their help and advice. Last but not least, thanks
to my wife Ingrid for everything and supporting the multiple vacation days used for
writing.
1 Lomb’s Algorithm and the
Hilbert Transform
1.1 Introduction
This first chapter describes two of the more advanced techniques in signal proces-
sing: Lomb’s algorithm and the Hilbert transform. Throughout this chapter (and the
remainder of this text) we assume that you have a basic understanding of signal
processing procedures; for those needing to refresh these skills, we include multiple
references to van Drongelen (2007).
In the 1970s, the astrophysicist Lomb developed an algorithm for spectral analy-
sis to deal with signals consisting of unevenly sampled data. You might comment
that in astrophysics considering uneven sampling is highly relevant (you cannot
observe the stars on a cloudy day), but in neuroscience data are always evenly sam-
pled. Although this is true, one can consider the action potential (or its extracellular
recorded equivalent, the spike) or neuronal burst as events that represent or sample
an underlying continuous process. Since these events occur unevenly, the sampling
of the underlying process is also uneven. In this context we will explore how to
obtain spectral information from unevenly distributed events.
The second part of this chapter introduces the Hilbert transform that allows one
to compute the instantaneous phase and amplitude of a signal. The fact that one
can determine these two metrics in an instantaneous fashion is unique because usu-
ally this type of parameter can only be associated with an interval of the signal. For
example, in spectral analysis the spectrum is computed for an epoch and the spec-
tral resolution is determined by epoch length. Being able to determine parameters
such as the phase instantaneously is especially useful if one wants to determine
relationships between multiple signals generated within a neuronal network.
1.2 Unevenly Sampled Data

In most measurements we have evenly sampled data—for instance, the interval Δt
between the sample points of the time series is constant, pixels in a picture have uni-
form interdistance, and so forth. Usually this is the case, but there are instances
when uneven sampling cannot be avoided. Spike trains (chapter 14, van Drongelen,
Signal Processing for Neuroscientists, A Companion Volume. DOI: 10.1016/B978-0-12-384915-1.00001-2

r 2010 Elsevier Inc. All rights reserved.
2 Signal Processing for Neuroscientists, A Companion Volume
2007) or time series representing heart rate (van Drongelen et al., 2009) are two
such examples; in these cases one may consider the spike or the heartbeat to repre-
sent events that sample an underlying process that is invisible to the experimenter
(Fig. 1.1A).
The heart rate signal is usually determined by measuring the intervals between
peaks in the QRS complexes. The inverse value of the interval between pairs of
subsequent QRS complexes can be considered a measure of the instantaneous rate
(Fig. 1.1B). This rate value can be positioned in a time series at the instant of either
the first or second QRS complex of the pair and, because the heartbeats do occur at
slightly irregular intervals, the time series is sampled unevenly. This example for
the heartbeat could be repeated, in a similar fashion, for determining the firing rate
associated with a spike train.
When a signal is unevenly sampled, many algorithms that are based on a fixed
sample interval (such as the direct Fourier transform [DFT] or fast Fourier transform
[FFT]) cannot be applied. In principle there are several solutions to this problem:
(1) An evenly sampled time series can be constructed from the unevenly sampled one by
using interpolation. In this approach the original signal is resampled at evenly spaced
intervals. The interpolation technique (e.g., linear, cubic, spline) may vary with the
application. In MATLAB resampling may be accomplished with the interp1 command
or any of the other related functions. After resampling the time series one can use stan-
dard Fourier analysis methods. The disadvantage is that the interpolation algorithm may
introduce frequency components that are not related to the underlying process.
(2) The measurements can be represented as the number of events in a binned trace; now
our time series is a sequence of numbers, with one number for each bin. Since the bins
are equally spaced, the standard DFT/FFT can be applied. In case of low-frequency
activity, the bins must be relatively wide to avoid an overrepresentation of empty bins.
Figure 1.1 The QRS complexes in the ECG or extracellularly recorded spike trains can be
considered as a series of events such as shown in (A). The rate of events can be depicted as
the inverse of the interval between the events (B); here the inverse of the interval between
each pair of events is plotted at the instant of the second event of the pair. The signal in (B)
is unevenly sampled because the rate measure is available only at the occurrence of the
events; the dashed line is a linear interpolation between these measures.
Lomb’s Algorithm and the Hilbert Transform 3
The disadvantage of this is that wide bins are associated with a low sample rate and thus
a low Nyquist frequency, which limits the bandwidth of the spectral analysis.
(3) The most elegant solution is to use Lomb’s algorithm for estimating the spectrum. This
algorithm is specially designed to deal with unevenly sampled time series directly with-
out the assumptions demanded by interpolation and resampling techniques (Lomb, 1976;
Press et al., 1992; Scargle, 1982; van Drongelen et al., 2009). The background and appli-
cation of this algorithm will be further described in Sections 1.2.1 and 1.2.2.
1.2.1 Lomb’s Algorithm

The idea of Lomb’s algorithm is similar to the development of the Fourier series,
namely, to represent a signal by a sum of sinusoidal waves (see chapter 5 in van
Drongelen, 2007). Lomb’s procedure is to fit a demeaned time series x that may be
sampled unevenly to a weighted pair of cosine and sine waves, where the cosine is
weighted by coefficient a and the sine by coefficient b. The fitting procedure is per-
formed over N samples of x(n) obtained at times tn and repeated for each frequency f.
Pða; b; f; tn Þ 5 a cosð2πftn Þ 1 b sinð2πftn Þ ð1:1Þ
Coefficients a and b are unknown and must be obtained from the fitting procedure.
For example, we can fit P to signal x by minimizing the squared difference between
P
N21
them over all samples: that is, minimize ε2 5 ½P 2 XðnÞ2 : We repeat this
n50
minimization for each frequency f. To accomplish this, we follow the same proce-
dure for developing the Fourier series (chapter 5 in van Drongelen, 2007) and set
the partial derivative for each coefficient to zero to find the minimum of the error,
that is:
@ε2 =@a 5 0 ð1:2aÞ
and
@ε2 =@b 5 0 ð1:2bÞ
For convenience, in the following we use a shorthand notation in addition to the full
notation. In the shorthand notation: C 5 cosð2πftn Þ; S 5 sinð2πftn Þ; and X 5 xðnÞ:
For the condition in Equation (1.2a) we get:
X @½P 2 xðnÞ X
@ε2 =@a 5 2½P 2 xðnÞ 5 2ðaC 1 bS 2 XÞC
@a
2 3
X
N 21
5 24a cosð2πftn Þ 1 b sinð2πftn Þ 2 xðnÞ5 cosð2πftn Þ 5 0
n50
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl}
P @½P 2 xðnÞ=@a
This and a similar expression obtained from the condition in Equation (1.2b) results
in the following two equations:
X X X
XC 5 a C2 1 b CS
X
N 21 X
N 21 X
N 21 ð1:3aÞ
XðnÞcosð2πftn Þ 5 a cos2 ð2πftn Þ 1 b cosð2πftn Þsinð2πftn Þ
n50 n50 n50
and
X X X
XS 5 a CS 1 b S2
X
N 21 X
N 21 X
N 21 ð1:3bÞ
XðnÞsinð2πftn Þ 5 a cosð2πftn Þsinð2πftn Þ 1 b sin2 ð2πftn Þ
n50 n50 n50
Thus far the procedure is similar to the standard Fourier analysis described in chapter 5
in van Drongelen (2007). The special feature in Lomb’s algorithm is that for each
frequency f, the sample times tn are now shifted by an amount τ (Fig. 1.2). Thus,
in Equations (1.3a) and (1.3b), tn becomes tn2τ. The critical step is that for each
frequency
P f, we select an optimal time shift τ so that the cosinesine cross-terms
( CS) disappear, that is:
X X
N 21
CS 5 cosð2πf ðtn 2 τÞÞsinð2πf ðtn 2 τÞÞ 5 0 ð1:4Þ
n50
Using the trigonometric identity cosðAÞsinðBÞ 5 12½sinðA 2 BÞ 2 sinðA 1 BÞ; this can
be simplified into:
2 3
X
N 21 X
N 21
14 5
sinð0Þ 2 sinð4πf ðt 2 τÞÞ 5 0- sinð4πf ðtn 2 τÞÞ 5 0
2
n50
|fflfflffl{zfflfflffl} n
n50
0
To separate the expressions for tn and τ, we use the trigonometric relationship

sinðA 2 BÞ 5 sinðAÞcosðBÞ 2 cosðAÞsinðBÞ to get the following expression:
Figure 1.2 The Lomb algorithm

fits sinusoidal signals to time
series that may be unevenly
sampled, as in the example in
(A). The fit procedure (B) is
Shifted sinewave optimized by shifting the
sinusoidal signals by an amount τ.
X
N21 X
N21
sinð4πftn Þcosð4πf τÞ 2 cosð4πftn Þsinð4πf τÞ
n50 n50
X
N 21 X
N 21
5 cosð4πf τÞ sinð4πftn Þ 2 sinð4πf τÞ cosð4πftn Þ 5 0
n50 n50
This can be further simplified into:
X
N 21 NX
21
sinð4πf τÞ=cosð4πf τÞ 5 tanð4πf τÞ 5 sinð4πftn Þ cosð4πftn Þ
n50 n50
Hence, condition (1.4) is satisfied if:

" X #
X
N 21 N 21
21
τ 5 tan sinð4πftn Þ cosð4πftn Þ 4πf ð1:5Þ
n50 n50
The value of variable τ as a function of frequency f can be found with Equation

(1.5), and by applying the appropriate shift tn - (tn2τ), the cross-terms in
Equations (1.3a) and (1.3b) become zero. Now we can determine the a and b co-
efficients for each frequency from the simplified expressions obtained from
Equations (1.3a) and (1.3b) without the cross-terms:
X X
XC 5 a C2
X
N 21 X
N 21
XðnÞcosð2πf ðtn 2 τÞÞ 5 a cos2 ð2πf ðtn 2 τÞÞ
n50 n50 ð1:6aÞ
X X X
N 21 NX
21
- a5 XC= C2 5 xðnÞcosð2πf ðtn 2 τÞÞ cos2 ð2πf ðtn 2 τÞÞ
n50 n50
and
X X
XS 5 b S2
X
N 21 X
N 21
xðnÞsinð2πf ðtn 2 τÞÞ 5 b sin2 ð2πf ðtn 2 τÞÞ
n50 n50 ð1:6bÞ
X X X
N 21 X
N 21
- b5 XS= S2 5 xðnÞsinð2πf ðtn 2 τÞÞ sin2 ð2πf ðtn 2 τÞÞ
n50 n50
Now we can compute the sum of P2 ða; b; f; tn Þ—that is, the sum of squares of the
sinusoidal signal in Equation (1.1) for all tn—in order to obtain an expression that
is proportional with the power spectrum S of x(n) as a function of f:
cross-terms
X
N 21 X X zfflfflffl}|fflfflffl{
Sðf; a; bÞ 5 P ða; b; f; tn Þ 5
2
ðaC 1 bSÞ 5 2
a C 1 b S 1 2abCS
2 2 2 2
n50
2 3
X
N 21
5 4a cos ð2πf ðtn 2 τÞÞ 1 b sin ð2πf ðtn 2 τÞÞ 1 cross-terms 5
2 2 2 2
|fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
n50 0
ð1:7Þ
Since we shift by τ, all cross-terms vanish and by substitution of the expressions

for the a and b coefficients in Equation (1.7) we get:
X 2 X 2
XC X XS X
Sðf Þ 5 X 2 C 2 1 X 2 S2
2 2
C S
" #2
X
N 21
xðnÞcosð2πf ðtn 2 τÞÞ N21
n50
X
5 " #2 cos2 ð2πf ðtn 2 τÞÞ
X
N 21
n50
cos2 ð2πf ðtn 2 τÞÞ
"n50 #2
X
N21
xðnÞsinð2πf ðtn 2 τÞÞ
n50
X
N 21
1 " #2 sin2 ð2πf ðtn 2 τÞÞ
X
N 21
n50
sin2 ð2πf ðtn 2 τÞÞ
n50
This can be further simplified into:
X 2 X 2
XC XS
Sðf Þ 5 X 1 X
C2 S2
" #2 " #2
X
N 21 X
N 21
ð1:8Þ
xðnÞcosð2πf ðtn 2 τÞÞ xðnÞsinð2πf ðtn 2 τÞÞ
n50 n50
5 1
X
N 21 X
N 21
cos2 ð2πf ðtn 2 τÞÞ sin2 ð2πf ðtn 2 τÞÞ
n50 n50
The expression for the power spectrum in Equation (1.8) is sometimes divided by 2
(to make it equal to the standard power spectrum based on the Fourier transform;
see Appendix 1.1), or by 2σ2 (σ2—variance of x) for the determination of the statis-
tical significance of spectral peaks. (Some of the background for this normalization
is described in Appendix 1.1; for more details, see Scargle, 1982.) By applying the
normalization we finally get:
8" #2 " #2 9
>
> X
N 21 X
N 21 >
>
>
> ðt 2 τÞÞ ðt 2 τÞÞ >
>
>
< xðnÞcosð2πf n xðnÞsinð2πf n >
=
1 n50 n50
Sðf Þ 5 2 1
2σ >
> X
N 21 X
N 21 >
>
>
> cos2 ð2πf ðtn 2 τÞÞ sin2 ð2πf ðtn 2 τÞÞ > >
>
: >
;
n50 n50
ð1:9Þ
From the above derivation, we can see that Lomb’s procedure allows (but does
not require) unevenly sampled data. Note that in Equations (1.7) and (1.8) we did not
compute power as the square of the cosine and sine coefficients, a and b, as we
would do in the standard Fourier transform; this is because in Lomb’s approach the
sinusoidal signals are not required to have a complete period within the epoch deter-
mined by the samples x(n). Because we do not have this requirement, the frequency f
is essentially a continuous variable and the spectral estimate we obtain by this
approach is therefore not limited by frequency resolution (in the DFT/FFT, the fre-
quency resolution is determined by the total epoch of the sampled data) and range
(in the DFT/FFT, the maximum frequency is determined by the Nyquist frequency).
However, to avoid misinterpretation, it is common practice to limit the bandwidth of
the Lomb spectrum to less than or equal to half the average sample rate. Similarly,
the commonly employed frequency resolution is the inverse of the signal’s epoch.
1.2.2 A MATLAB Example

To test Lomb’s algorithm we apply it to a signal that consists of two sinusoidal sig-
nals (50 and 130 Hz) plus a random noise component (this is the same example used
in fig. 7.2A in van Drongelen, 2007). In this example (implemented in MATLAB
script Pr1_1.m ), we sample the signal with randomly distributed intervals (2000
points) and specify a frequency scale (f in the script) up to 500 Hz. Subsequently we
use Equations (1.5), (1.8), and (1.9) to compute τ (tau in the script) and the unscaled
and scaled versions of power spectrum S(f) (Pxx in the script) of input x(n) (x in
the script). This script is available on http://www.elsevierdirect.com/companions/
9780123849151.
The following script ( Pr1_1.m ) uses the Lomb algorithm to compute the spectrum
from an unevenly sampled signal. The output of the script is a plot of the input (an
unevenly sampled time domain) signal and its associated Lomb spectrum.
% Pr1_1.m
% Application of Lomb Spectrum
clear;
t=rand(2000,1); t=sort(t); % An array of 2000 random sample intervals
f=[1:500]; % The desired frequency scale
% frequencies same as pr7_1.m in van Drongelen (2007)
f1=50;
f2=130;
% data plus noise as in pr7_1.m in van Drongelen (2007)
x=sin(2*pi*f1*t)+sin(2*pi*f2*t);
x=x+randn(length(t),1);
var=(std(x))^2; % The signal’s variance
% Main Loop
for i=1:length(f)
h1=4*pi*f(i)*t;
%Equation (1.5)
tau=atan2(sum(sin(h1)), sum(cos(h1)))/(4*pi*f(i));
h2=2*pi*f(i)*(t-tau);
%Equation (1.8)
Pxx(i)=(sum(x.*cos(h2)).^2)/sum(cos(h2).^2)+. . .
(sum(x.*sin(h2)).^2)/sum(sin(h2).^2);
end;
% Normalize; Equation (1.9)
Pxx=Pxx/(2*var);
% Plot the Results
figure;
subplot(2,1,1), plot(t,x,‘.2’)
title(‘Irregularly Sampled Signal (USE ZOOM TO INSPECT UNEVEN
SAMPLING)’)
xlabel(‘Time (s)’);ylabel(‘Amplitude’)
subplot(2,1,2),plot(f,Pxx);
title(‘Lomb Spectrum’)
xlabel(‘Frequency (Hz)’);ylabel(‘Normalized Power’)
1.3 The Hilbert Transform

One of the current frontiers in neuroscience is marked by our lack of understanding
of neuronal network function. A first step in unraveling network activities is to
record from multiple neurons and/or networks simultaneously. A question that
often arises in this context is which signals lead or lag; the underlying thought here
is that the signals that lead cause the signals that lag. Although this approach is not
foolproof, since one can only make reasonable inferences about causality if all
connections between and activities of the neuronal elements are established, it is a

first step in analyzing network function. Multiple techniques to measure lead and
lag can be used. The simplest ones are cross-correlation and coherence (for an
overview of these techniques, see chapter 8 in van Drongelen, 2007). A rather
direct method to examine lead and lag is to determine the phase of simultaneously
recorded signals. If the phase difference between two signals is not too big, one
considers signal 1 to lead signal 2 if the phase of signal 1 (φ1) is less than the phase
of signal 2 (φ2): φ1 , φ2. Of course this procedure should be considered as a heu-
ristic approach to describe a causal sequence between the components in the net-
work activity since there is no guarantee that a phase difference reflects a causal
relationship between neural element 1 (generating signal 1) and neural element 2
(generating signal 2). In this example one could easily imagine alternatives where
neural elements 1 and 2 are both connected to a common source causing both sig-
nals, or where element 2 is connected to element 1 via a significant number of
relays; in both alternatives the condition φ1 , φ2 might be satisfied without a direct
causal relationship from element 1 to element 2. A frequently used technique to
compute a signal’s phase is the Hilbert transform, which will be described in the
remainder of this chapter. An alternative approach to study causality in multichan-
nel data is discussed in Chapter 7.
The Hilbert transform is a useful tool to determine the amplitude and instanta-
neous phase of a signal. We will first define the transform before demonstrating
the underlying mathematics. An easy way of introducing the application of the
Hilbert transform is by considering Euler’s equation multiplied with a constant A:
Aejωt 5 A½cosðωtÞ 1 j sinðωtÞ 5 A cosðωtÞ 1 jA sinðωtÞ ð1:10Þ
In this example we consider the first term in Equation (1.10), f ðtÞ 5 A cosðωtÞ; as
the signal under investigation. This signal is ideal to demonstrate the Hilbert trans-
form application because in this example we can see that the amplitude of f ðtÞ is A,
and its instantaneous phase φ is ωt. The terminology for the Hilbert transform is as
follows: the imaginary component, the second term, in Equation (1.10)
f~ðtÞ 5 A sinðωtÞ is defined as the Hilbert transform of f ðtÞ (we will discuss further
details in Sections 1.3.1 and 1.3.2 below), and the sum of both the signal and its
Hilbert transform multiplied by j generates a complex signal:
fa ðtÞ 5 A e jωt 5 A cosðωtÞ 1 j A sinðωtÞ 5 f ðtÞ 1 j f~ðtÞ
in which fa ðtÞ is defined as the analytic signal.

To summarize, the real part of the analytic signal is the signal under investiga-
tion f ðtÞ and its imaginary component is the Hilbert transform f~ðtÞ of the signal.
The analysis procedure is summarized in Fig. 1.3. As can be seen in Fig. 1.3A
and B, we can use the analytic signal A ejωt to determine amplitude A and instanta-
neous phase ωt of any point, such as the one indicated by *. The amplitude is:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
A5 real component2 1 imaginary component2
(A) Analytic signal (B) Time domain

f (t) = A cos(ωt)
fa(t) = A e jωt = A cos(ωt) + A sin(ωt)
Imaginary axis
Hilbert transform
ωt
Real axis
~
fa(t) = A sin(ωt)
Instantaneous phase = ωt
Fa(ω) = 2Aπ δ (ω −ω0) Aπ δ (ω − ω0) Aπ δ (ω − ω0)
A j π δ (ω +ω0) − j sgn(ω)
− ω0 ω0 ω
Fourier transform of the j

analytic signal A j π δ (ω −ω0)
− ω0 ω0
Frequency domain
Figure 1.3 The signal amplitude A and instantaneous phase ωt of point * of the cosine
function (B, f ðtÞ; blue) can be determined with the so-called analytic signal (A). The analytic
signal consists of a real part equal to the signal under investigation (the cosine) and an
imaginary component (the sine). The imaginary component (red) is defined as the Hilbert
transform f~ðtÞ of the signal f ðtÞ: The frequency domain equivalents of the cosine wave, the
sine wave, the Hilbert transform procedure, and the analytic signal are shown in (C). See
text for further explanation.
and the phase is:

21 imaginary component
φ 5 tan
real component
Again, in this example we did not need the analytic signal to determine phase and
amplitude for our simple cosine signal, but our finding may be generalized to other
signals where such a determination is not trivial.
1.3.1 The Hilbert Transform in the Frequency Domain

As can be seen in the earlier example (depicted in Fig. 1.3), the Hilbert transform
can be considered as a phase shift operation on f ðtÞ to generate f~ðtÞ: In our example
the signal cosðωtÞ is shifted by 2π/2 rad (or 290 ) to generate its Hilbert trans-
form cosðωt 2 π=2Þ 5 sinðωtÞ: We may generalize this property and define a
(A)
~
f (t ) Hilbert Transformer f (t )
Phase Shift by -π /2 rad
(B) (C)
Imaginary Axis
Imaginary Axis
z = a + jb
*
-π/2
Real Axis Real Axis
*
-π /2 *
e − j (π / 2 ) = − j * ~
z = b − ja
Figure 1.4 (A) The Hilbert transform can be represented as the operation of an LTI system
(the Hilbert transformer). Input f ðtÞ is transformed into f~ðtÞ by shifting it by 2π/2 rad
(290 ). (B) The Hilbert transform operation in the frequency domain can be represented as
a multiplication with e2jðπ=2Þ 5 2j (orange arrow). (C) Example of the Hilbert transform in
the frequency domain—that is, multiplication of a complex number z 5 a 1 jb with 2j. The
result is z~ 5 b 2 ja: As can be seen, the result is a 290 rotation. Note in this panel that 90
angles are indicated by M and that the angles indicated by and * add up to 90 .G
Hilbert transformer as a phase-shifting (linear time invariant, LTI) system that gen-
erates the Hilbert transform of its input (Fig. 1.4A). The generalization of this prop-
erty associated with the cosine is not too far of a stretch if you recall that, with the
real Fourier series, any periodic signal can be written as the sum of sinusoidal sig-
nals (the cosine and sine waves in equation (5.1) in van Drongelen, 2007) and that
our above results can be applied to each of these sinusoidal components.
To further define the shifting property of the Hilbert transformer (see Fig.
1.4A), we begin to explore this operation in the frequency domain, because here
the procedure of shifting the phase of a signal by 2π/2 rad is relatively easy
to define as a multiplication by e2jðπ=2Þ 5 2j (Fig. 1.4B). If this is not obvious
to you, consider the effect of this multiplication for any complex number
z 5 ejφ (representing phase φ) that can also be written as the sum of its real and
imaginary parts z 5 a 1 jb: Multiplication by 2j gives its Hilbert transform
z~ 5 2jða 1 jbÞ 5 b 2 ja; indeed corresponding to a 290 rotation of z (see
Fig. 1.4C). Although the multiplication with 2j is correct for the positive frequen-
cies, a 290 shift for the negative frequencies in the Fourier transform (due to the
negative values of ω) corresponds to multiplication with ejðπ=2Þ 5 j: Therefore, the
operation of the Hilbert transform in the frequency domain can be summarized as:
multiplication by 2 j sgnðωÞ ð1:11Þ

Here we use the so-called signum function sgn (Appendix 1.2, Fig. A2.1)
defined as:
8
< 21 for ω , 0
sgnðωÞ 0 for ω 5 0 ð1:12Þ
:
1 for ω . 0
Let us go back to our phase-shifting system depicted in Fig. 1.4A and define its
unit impulse response as h(t) and its associated frequency response as H(ω). Within
this approach, the Hilbert transform is the convolution of input f(t) with h(t). Using
our knowledge about convolution (if you need to review this, see section 8.3.2 in
van Drongelen, 2007), we can also represent the Hilbert transform in the frequency
domain as the product of F(ω)—the Fourier transform of f(t)—and H(ω). This is
very convenient because we just determined above that the Hilbert transform in the
frequency domain corresponds to a multiplication with 2j sgnðωÞ: To summarize,
we now have the following three Fourier transform pairs:
System’s input3Fourier transform: f ðtÞ 3FðωÞ

System’s unit impulse response3Fourier transform: hðtÞ3HðωÞ
Hilbert transform3Fourier transform: f ðtÞ hðtÞ3FðωÞHðωÞ
ð1:13Þ
Using these relationships and Equation (1.11), we may state that the Fourier trans-
form of the unit impulse response (i.e., the frequency response) of the Hilbert trans-
former is:
HðωÞ 5 2jsgnðωÞ ð1:14Þ
We can use the expression we found for H(ω) to examine the above example of
Euler’s equation
Analytical signal
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
Ae jω0 t
5 A cosðω0 tÞ 1 j A sinðω0 tÞ
|fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl} |fflfflfflfflfflfflffl{zfflfflfflfflfflfflffl}
Signal Hilbert transform
in the frequency domain. The Fourier transform of the cosine term (using equation
(6.13) in van Drongelen, 2007) is:
Aπ½δðω 1 ω0 Þ 1 δðω 2 ω0 Þ ð1:15Þ
Now, according to Equation (1.13), the Fourier transform of the cosine’s Hilbert
transform is the product of the Fourier transform of the input signal (the cosine)
and the frequency response of the Hilbert transformer H(ω), that is:
Aπ½δðω 1 ω0 Þ 1 δðω 2 ω0 Þ 2j sgnðωÞ

ð1:16Þ
Aπ½δðω 1 ω0 Þð2j sgnðωÞÞ 1 δðω 2 ω0 Þð2j sgnðωÞÞ
Because δðω 1 ω0 Þ is only nonzero for ω 5 2ω0 and δðω 2 ω0 Þ is only nonzero for
ω 5 ω0 ; we may rewrite the 2j sgn(ω) factors in Equation (1.16) and we get:
Aπ½δðω 1 ω0 Þð2 j sgnð2ω0 ÞÞ 1 δðω 2 ω0 Þð2j sgnðω0 ÞÞ
Now we use the definition of sgn, sgnð2ω0 Þ 5 21 and sgnðω0 Þ 5 1 (Equation

(1.12)), and simplify the expression to:
Aπ½δðω 1 ω0 ÞðjÞ 1 δðω 2 ω0 Þð 2 jÞ 5 Ajπ½δðω 1 ω0 Þ 2 δðω 2 ω0 Þ ð1:17Þ
As expected, Equation (1.17) is the Fourier transform of A sinðωtÞ (see equation

(6.14) in van Drongelen, 2007), which is indeed the Hilbert transform f~ðtÞ of
f ðtÞ 5 A cosðωtÞ:
Combining the above results, we can find the Fourier transform of the analytic
signal fa ðtÞ 5 A cosðωtÞ 1 jA sinðωtÞ 5 f ðtÞ 1 j f~ðtÞ: If we define the following pairs:
fa ðtÞ3Fa ðωÞ
f ðtÞ 3FðωÞ
f~ðtÞ 3FðωÞ
~
the above expressions can be combined in the following Fourier transform pair:
fa ðtÞ 5 f ðtÞ 1 j f~ðtÞ3Fa ðωÞ 5 FðωÞ 1 jFðωÞ

~
In the above equation we substitute the expressions for FðωÞ from Equation (1.15)
~
and FðωÞ from Equation (1.17) and get:
~
Fa ðωÞ 5 FðωÞ 1 jFðωÞ
Fourier transform of analytical signal
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
5 Aπ½δðω 1 ω0 Þ 1 δðω 2 ω0 Þ 1 j Ajπ½δðω 1 ω0 Þ 2δðω 2 ω0 Þ
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
Fourier transform of signal Fourier transform of Hilbert transform
With a bit of algebra we obtain:
~
Fa ðωÞ 5 FðωÞ 1 jFðωÞ 5 2πAδðω 2 ω0 Þ ð1:18Þ
This interesting finding shows that the Fourier transform of the analytic signal has
zero energy at negative frequencies and only a peak at 1 ω0 : The peak’s amplitude
at 1 ω0 is double the size of the corresponding peak in FðωÞ (Fig. 1.3C). This find-
ing may be generalized as: “The Fourier transform of the analytic signal Fa ðωÞ has
no energy at negative frequencies 2ω0 ; it only has energy at positive frequencies

1ω0 and its amplitude is double the amplitude at 1 ω0 in FðωÞ.”
1.3.2 The Hilbert Transform in the Time Domain

From the frequency response presented in Equation (1.14) and the relationship
between convolution in the time and frequency domains (section 8.3.2, van
Drongelen, 2007), we know that the unit impulse response h(t) of the Hilbert trans-
former (Fig. 1.4A) is the inverse Fourier transform of 2j sgnðωÞ. You can find
details of sgnðtÞ and its Fourier transform in Appendix 1.2; using the signum’s
Fourier transform, we can apply the duality property (section 6.2.1, van Drongelen,
2007) to determine the inverse Fourier transform for 2j sgnðωÞ: For convenience
we restate the duality property as:
if f ðtÞ3FðωÞ; then FðtÞ32πf ð2ωÞ ð1:19aÞ
Applying this to our signum function (see also Appendix 1.2), we can define the
inverse Fourier transform of sgnðωÞ:
2 2
sgnðtÞ3 ; therefore 32π sgnð2ωÞ ð1:19bÞ
jω jt |fflfflfflfflffl{zfflfflfflfflffl}
2sgnðωÞ
Note that we can substitute 2sgnðωÞ 5 sgnð2ωÞ because the signum function
(Fig. A2.1) has odd symmetry (defined in Appendix 5.2 in van Drongelen, 2007).
Using the result from applying the duality property in Equation (1.19b), we can
determine the inverse Fourier transform for the frequency response of the Hilbert
transformer HðωÞ 5 2j sgnðωÞ and find the corresponding unit impulse response
hðtÞ: Because 2π and j are both constants, we can multiply both sides with j and
divide by 2π; this generates the following Fourier transform pair:
1
hðtÞ 5 3Hð jωÞ 5 2j sgnðωÞ ð1:20Þ
πt
In Equation (1.14) we found that the frequency response of the Hilbert transformer
is 2j sgnðωÞ: Because we know that multiplication in the frequency domain is
equivalent to convolution in the time domain (chapter 8 in van Drongelen, 2007),
we can use the result in Equation (1.20) to define the Hilbert transform f~ðtÞ of sig-
nal f ðtÞ in both the time and frequency domains. We define the following Fourier
transform pairs:
the input: f ðtÞ3 FðωÞ

the Hilbert transform of the input: f~ðtÞ3 FðωÞ
~
the unit impulse response of the Hilbert transformer: hðtÞ3HðωÞ
Using the above pairs and Equation (1.20), the Hilbert transform and its frequency
domain equivalent are:
ð
N
1 f ðtÞ
f~ðtÞ 5 f ðtÞ hðtÞ 5 ~
dτ3FðωÞ 5 FðωÞHðωÞ ð1:21Þ
|ffl{zffl} π t2τ
1 2N
πt
There is, however, a problem with our finding for the Hilbert transform expression
in Equation (1.21), which is that there is a pole for f ðtÞ=ðt 2 τÞ within the integration
limits at t 5 τ: The solution to this problem is to define the Hilbert transform as:
ð
N
1 f ðtÞ
f~ðtÞ 5 CPV dτ ð1:22Þ
π t2τ
2N
in which CPV indicates the Cauchy principal value of the integral. The CPV is a
mathematical tool to evaluate integrals that include poles within the integration
limits. An example of such an application is given in Appendix 1.3. For those inter-
ested in the CPV procedure, we refer to a general mathematics text such as Boas
(1966).
1.3.3 Examples
The Hilbert transform is available in MATLAB via the hilbert command. Note
that this command produces the analytic signal f ðtÞ 1 j f~ðtÞ and not the Hilbert
transform itself; the Hilbert transform is the imaginary component of the output.
You can evaluate the example from Equation (1.10) by computing the Hilbert
transform for the cosine and plot the amplitude and phase. Type the following in
the MATLAB command window:
step=0.00001; % step size=1/sample rate

t=0:step:1; % timebase
x=cos(2*pi*4*t); % 4 Hz signal
xa=hilbert(x); % compute the analytic signal
Amplitude=abs(xa); % amplitude of the signal
Phase=atan2(imag(xa),real(xa)); % instantaneous phase
Ohmega=diff(Phase)/(2*pi*step); % instantaneous frequency in Hz
figure;plot(t,x,‘k’);hold;
plot(t,Amplitude,‘r’);
plot(t,Phase,‘g’);
plot(t(1:length(t)21),Ohmega,‘m.’)
axis([0 1 -5 5])
You will obtain a graph of a 4-Hz cosine function with an indication of its amplitude
(a constant) in red, its instantaneous phase in green (note that we use the atan2
MATLAB command in the above example because we want to obtain phase angles
between 2π and 1π), and the frequency as the derivative of the phase in magenta.
You can now check both frequency characteristics we discussed by computing
the Fourier transforms and plotting these in the same graph.
X=fft(x); % Fourier transform of the signal

XA=fft(xa); % Fourier transform of the analytic signal
figure;plot(abs(X),‘k’);hold;plot(abs(XA))
If you use the zoom function of the graph to study the peaks in the plot you will
see that the peaks for the positive frequencies (far-left part of the graph) show a
difference of a factor two between the Fourier transform of the analytic signal and
the Fourier transform of the signal. The negative component (in the discrete version
of the Fourier transform this is the far-right part of the graph) shows only a peak in
the Fourier transform of the signal. Both observations are as expected from the the-
oretical considerations in Section 1.3.1.
Another property to look at is the phase shift between the signal and its Hilbert
transform. This can be accomplished by typing the following lines:
figure; hold;
plot(t,imag(xa),‘r’); % the imaginary part of the analytic signal=
% the Hilbert transform
plot(t,x,‘k.’) % the signal
plot(t,real(xa),‘y’) % real part of the analytic signal=signal
Now you will get a figure with the signal (4-Hz cosine wave) in both black (thick
line) and yellow (thin line); the Hilbert transform (the 4-Hz sine wave) is plotted
in red.
Finally we will apply these techniques to an example in which we have two neu-
ral signals, one signal generated by a single neuron and one signal generated by the
network in which the neuron is embedded. Our question here is how the phases of
these two signals relate. First, the raw extracellular trace is rectified and sent
through a low-pass filter with a 50-ms time constant (this technique of using the
analytic signal to find the instantaneous phase usually works better with signals
composed of a small band of frequencies and, in our case, we are only interested in
the low-frequency behavior; see Pikovsky et al., 2001, for more details). For the
cellular activity, we create a raster plot of the spike times and send it through the
same low-pass filter. We now have two signals representing the low-pass-filtered
spiking behavior of the cell and network (see Fig. 1.5Aiii). We can use the Hilbert
(A) Pre-processing of the signal

(i) Raw data (iii) Low-pass filtered
(ii) Spike raster
Cell activity
Network
activity
(B) Using the Hilbert transform

(i) Instantaneous phase (iii) Histogram of phase
(ii) Phase difference
differences
π
π
–π
–π
π
–π –π π
Figure 1.5 (A) Processing of a cellular and network activity (i) into a low-frequency index
of spiking activity (iii) (see text for details). (B) The low-pass-filtered signals of (A) were
transformed using the analytic signal technique to find the instantaneous phase over time (i).
The relationship between the two signals was investigated by finding the difference between the
phases over time (ii) and plotting these phase differences in a histogram (iii). In this example we
observe that the overall effect is that the network activity leads and the cell activity lags—that is,
the histogram (iii) of network activity phase minus cell activity phase is predominantly positive.
(From A. Martell, unpublished results, with permission.)
transform technique to find the instantaneous phase of each signal (Fig. 1.5Bi). For
our case, we are interested in how the phases of the cellular and network signals
are related. To find this relationship, we calculate the difference between the two
instantaneous phase signals at each point in time and then use this information to
generate a histogram (see Fig. 1.5Biiiii). This method has been used to compare
how the phases of cellular and network signals are related for different types of cel-
lular behavior (Martell et al., 2008).
Appendix 1.1
In the case of the standard power spectrum we have S 5 XX*/N (equation (7.1)
in van Drongelen, 2007). The normalization by 1/N ensures that Parseval’s conser-
vation of energy theorem is satisfied (this theorem states that the sum of squares
in the time domain and the sum of all elements in the power spectrum are equal; see
Table 7.1 and Appendix 7.1 in van Drongelen, 2007). In the case of Lomb’s algo-
rithm we compute the sum of squares for each frequency by using the expression in
Power spectrum
based on
Lomb’s algorithm Power spectrum
based on the
Fourier transform
Figure A1.1 Spectral analysis of a 1-s epoch of a 50-Hz signal sampled at 1000 Hz. The
graph depicts the superimposed results from a standard power spectrum (red) based on the
Fourier transform and the power spectrum obtained with Lomb’s algorithm (dark blue). Note
that the total energy in both cases is identical. This figure can be created with Pr1_2.m .
Equation (1.8), which is based on Equation (1.7). Our expectation is therefore that
Lomb’s spectrum will also satisfy Parseval’s theorem. However, there is a slight dif-
ference. In the standard Fourier transform the positive and negative frequencies
each contain half the energy. Basically, this is due to the fact that the Fourier trans-
form is based on the complex Fourier series, which includes negative frequencies.
In contrast, if we compute the Lomb spectrum only up to the Nyquist frequency, we
have all energy in the positive frequencies, and therefore its values are twice as large
as compared to the standard power spectrum. An example for a single frequency is
shown in Figure A1.1. This figure is based on a standard power spectrum and Lomb
spectrum computed for the same input, a sine wave of 50 Hz. Thus, if we want the
Lomb spectrum to have the same amplitudes as the standard power spectrum, we
need to divide by two. Furthermore, if we want to normalize by the total power, we
can divide by the variance σ2. This normalization by 2σ2 is exactly the normalization
commonly applied for Lomb’s spectrum (see Equation (1.9) and Pr1_1.m ).
Appendix 1.2
This appendix describes the signum function sgnðtÞ; its derivative, and Fourier
transform. The signum function is 1 for positive t and 21 for negative t
(Fig. A2.1). Similar to the derivative of the unit step function U(t) (section 2.2.2,
fig. 2.4A in van Drongelen, 2007), the derivative of this function is only nonzero at
t 5 0. The only difference is that for sgnðtÞ the function increases by 2 units (from
21 to 1) instead of 1 unit (from 0 to 1) in U(t). Since the derivative of the unit step
is δðtÞ; the derivative of the signum function would be twice as large, that is:
d½sgnðtÞ
5 2δðtÞ ðA1:2:1Þ
dt
The Fourier transform of the derivative of a function is equal to the Fourier trans-
form of that function multiplied with jω. This property is similar to the relationship
of the Laplace transform of a derivative of a function and the Laplace transform of
the function itself (see section 9.3, equation (9.3) in van Drongelen, 2007). If we
Figure A2.1 The signum function

and its derivative, the unit impulse
– function with an amplitude of two.
d
δ
d
now use this property and define the Fourier transform of sgnðtÞ as S(ω), we can
apply the Fourier transform to both sides of Equation (A1.2.1):
jωSðωÞ 5 2 ðA1:2:2Þ
Recall that the Fourier transform of the unit impulse is 1 (see section 6.2.1, equa-
tion (6.9) in van Drongelen, 2007). Therefore, the Fourier transform pair for the
signum function is:
2
sgnðtÞ3SðωÞ 5 ðA1:2:3Þ
jω
Appendix 1.3
In Equation (1.22) we use the Cauchy principal value, CPV. This technique is used
to approach integration of a function that includes a pole within the integration lim-
its. We will not go into the mathematical details (for more on this subject please
see a mathematics textbook such as Boas, 1966), but we will give an example to
Ðd
show the principle. For example, consider the integral ð1=xÞdx: The function 1=x
2d
in this integral has a pole (is unbounded) at x 5 0: The Cauchy principal value tech-
nique approximates the integral as the sum of two separate integral:
ðd ð
2ε ðd
1 1 1
dx dx 1 dx
x x x
2d 2d ε
where ε is a small positive value approaching zero. In this case the two integrals
Ðd
cancel and approach ð1=xÞdx: Our final result can be summarized as:
2d
2 2ε 3
ðd ð ðd
1 1 1 5
CPV dx 5 lim 4 dx 1 dx 5 0
x ε-0 x x
2d 2d ε
Here the Cauchy principal value is indicated by CPV; in other texts you may also
find PV or P.
2 Modeling
2.1 Introduction
Signal analysis is frequently used to characterize systems. In van Drongelen
(2007), chapter 8, we described linear systems and associated techniques that allow
us determine system characteristics. In the last chapter of van Drongelen (2007)
(section 17.3) we showed how these linear methods, such as cross-correlation, fail
to characterize signals with a nonlinear component. To address this shortcoming,
we used metrics such as correlation dimension, the Lyapunov exponent, or
Kolmogorov entropy to characterize nonlinear signal properties.
The goal of this chapter is to introduce basics for modeling systems, with an
emphasis on techniques used to characterize nonlinear systems and their signals. In
this context, this chapter will also provide an introduction to the application of the
Volterra series, which forms the basis for the identification of dynamical nonlinear
systems, and which we will go over in more detail in Chapter 3. The systems that
we will introduce in this chapter are considered higher-order systems, since they
include operators beyond the (linear) first-order one. Useful references on the char-
acterization of nonlinear systems are the seminal text by Marmarelis and
Marmarelis (1978) and the reprint edition of a text from the 1980s by Schetzen
(2006). For more recent overviews, see Westwick and Kearney (2003) and
Marmarelis (2004).
2.2 Different Types of Models

Before going into mathematical detail, it is useful to summarize some of the types
of models that one may encounter in neuroscience. Attenuators and amplifiers are
both examples of linear systems, since output is simply the product of input and a
constant (e.g., y 5 3x). Alternatively, expressions that characterize nonlinear sys-
tems include higher-order terms: these systems, as we will see in Chapters 35, do
not obey the scaling and superposition rules of linear models (to review these rules
see section 8.3.1.1 in van Drongelen, 2007). Examples of nonlinear higher-order
systems are y 5 x2 (second-order system) and y 5 5 1 x 1 3x3 (third-order system).
At this point it is important to note that an expression such as y 5 a 1 bx 1 cx3 can
still be considered linear, but with respect to its parameters a, b, and c. This is a

property that we will use when developing the regression procedure in

Section 2.4.1.
All of the examples earlier are static models (systems without memory), mean-
ing that their output depends only on present input. In neuroscience we usually
must deal with dynamical models, in which output depends on present and past
input (but not on future input); these systems are also called causal. Static models
are represented by algebraic equations (such as the ones in the previous paragraph),
whereas dynamical systems are modeled by differential equations (for continuous
time models) or difference equations (for discrete time models). General examples
of linear dynamical systems with input x and output y are:
dn yðtÞ dn21 yðtÞ

An 1 A n21 1 ? 1 A0 yðtÞ
dtn dtn21
dm xðtÞ dm21 xðtÞ
5 Bm m
1 Bm21 1 ? 1 B0 xðtÞ
dt dtm21
for continuous time systems and:
An yðk 2 nÞ 1 An21 yðk 2 n 1 1Þ 1 ? 1 A0 yðkÞ

5 Bm xðk 2 mÞ 1 Bm21 xðk 2 m 1 1Þ 1 ? 1 B0 xðkÞ
for discrete time systems (for details see chapter 8 in van Drongelen, 2007). If one
of the terms in a differential or difference equation is of a higher order, we have a
nonlinear dynamical system. For example, y 2 4ðdy=dtÞ2 5 2x represents a second-
order dynamical system.
Time invariance is a critical condition for the development of the convolution
formalism (see section 8.3.1.1 in van Drongelen, 2007). This property allows us to
state that a system’s response to identical stimuli at different points in time is
always the same (provided that the system is in the same state, of course). Just as
we have linear time invariant systems, we also have nonlinear time invariant sys-
tems (usually abbreviated as LTI or NLTI systems).
Models of real systems can be generated according to two major methodological
approaches. One might follow a deductive path and start from (physical) assump-
tions about the system, generating a hypothetical model to create predictions that
can be empirically tested. These empirical measurements can be used to establish
the parameters of the hypothetical model, and therefore this type of representation
is often called a parametric model. An alternative to this procedure, the inductive
path, is followed if one starts from the measurements of a system’s input and out-
put. This data-driven method uses measurements, rather than assumptions about the
system, to mathematically relate input and output. Here, we can consider the sys-
tem as a black box, modeled by a mathematical relationship that transforms input
into output. This type of model is often referred to as nonparametric (note, how-
ever, that nonparametric does not refer to the absence of parameters; in many cases,
Modeling 23
these models will have more parameters than parametric models). The method of
induction is appropriate when dealing with complex systems that resist a reduction
to a simpler parametric model. It can also be a starting point in which a system is
first characterized as a black box and in subsequent steps parts with physical mean-
ing replace pieces of the black box. In this combined approach, parts of the model
may still be part of the black box, whereas other parts may be associated with a
physical interpretation. In this case, the distinction between parametric and non-
parametric models may become a bit fuzzy.
2.3 Examples of Parametric and Nonparametric Models

A parametric model usually has relatively few parameters. A simple example of a
parametric model of a dynamical LTI system is the ordinary differential equation
(ODE) for a filter. For example, x 5 RCðdy=dtÞ 1 y describes input x and output y of
a simple RC circuit (Fig. 2.1A). The only parameters in this case are the values of
the resistor R and the capacitor C in the equation. Subsequently, the value for these
parameters can be determined experimentally from observing the system’s
behavior.
Note: See van Drongelen (2007) for further details about determining these
parameters from measurements: in section 11.2.1 it is shown how RC can be
obtained from the filter’s unit step response (equation (11.8)), and in section
12.3, RC is determined from the 3 dB point of the filter’s frequency charac-
teristic (equation (12.5)).
A very famous parametric model in neuroscience is the Hodgkin and Huxley

(1952) formalism using four variables to describe the action potential generated in
the squid’s giant axon: the membrane potential V and three other variables m, h,
and n describe the membrane potential-dependent characteristics of sodium and
potassium conductivity. In the following it is assumed you are somewhat familiar
with the Hodgkin and Huxley model; if you need to review the details, chapter 2 in
Izhikevich (2007) provides an excellent overview.
Initially in the 1950s, the formalism was entirely hypothetical, and it was not
until after the molecular basis for Na1 and K1 ion channels was elucidated that a
physical interpretation of large parts of the model could be made. The gating vari-
able m characterizes the depolarization process caused by increased conductance of
sodium ions (causing an influx of positively charged sodium ions) that occurs dur-
ing the action potential generation. The variables h and n are recovery variables
that represent the inactivation of sodium ion conductance (reduced Na1 influx) and
the activation of potassium conductance (causing an outward flux of positively
charged potassium ions).
d
d
IC + INa + IK + IL = 0
RL
IC INa
RNA
IK
RK
IL –
Na
EK EL
Figure 2.1 (A) Example of a parametric model of a dynamical linear system (a low-pass
filter) and its input and output (x and y respectively). (B) The black box, nonparametric
equivalent of the same system is the white curve representing the (sampled) unit impulse
response (UIR). Both models permit us to predict the output resulting from an arbitrary input
such as the unit step function. The parametric model has two parameters (R and C) with
physical meaning. The nonparametric model consists of many parameters (the samples
making up the UIR) without a direct physical meaning. (C) Hodgkin and Huxley’s electronic
equivalent circuit for the biomembrane. The model consists of the membrane capacitance
(C) and three parallel ion channels: one for sodium, one for potassium, and a leakage
channel. According to Kirchhoff’s first law the sum of all currents at the node (arrow) must
be zero. (D) Model for gating variable m in the Hodgkin and Huxley formalism.
Hodgkin and Huxley’s model relates all these variables in an equivalent circuit
of the excitable biomembrane (Fig. 2.1C) by setting the sum of all membrane cur-
rents equal to zero according to Kirchhoff’s first law (see appendix 1.1 in van
Drongelen, 2007). By applying this law to the node indicated with the arrow in the
membrane model in Fig. 2.1C we obtain:
dV V 2 ENa V 2 EK
C 1 1 1 IL 5 0 ð2:1Þ
dt R RK |{z}
|fflffl{zfflffl} |fflfflfflfflffl{zfflfflfflfflffl}
Na
|fflfflfflfflffl{zfflfflfflffl
ffl}
Leak
Capacitive Sodium Potassium
current
current current current
5 IC 5 INa 5 IK
In this expression we have several parameters: membrane capacitance C; the resis-

tance values for sodium and potassium ions, RNa and RK, respectively; ENa and EK
Modeling 25
are the equilibrium potentials for sodium and potassium ions computed with the
Nernst equation (appendix 1.1 in van Drongelen, 2007); and IL is a constant leak-
age current attributed to Cl2 ions. The sodium and potassium currents are deter-
mined with Ohm’s law (appendix 1.1 in van Drongelen, 2007): each ion species
experiences a potential drop equal to the difference between the membrane poten-
tial V and its equilibrium potential (e.g., for sodium: V 2 ENa ), and this potential
drop divided by the resistance is the ion current (e.g., for sodium the current is
ðV 2 ENa Þ=RNa ). In addition to Equation (2.1), Hodgkin and Huxley (1952)
described the dynamics for RNa and RK with the nonlinear relationships
m3 h and gK 5 1=RK 5 gK n4 ; where gNa
gNa 5 1=RNa 5 gNa and gK are the maximum
conductivity values for sodium and potassium. Furthermore, the gating variable m
is modeled by a reversible process between the open (m) and closed (1m) states
(Fig. 2.1D), which can be represented by the following ODE:
dm
5 αm ðVÞð1 2 mÞ 2 β m ðVÞm ð2:2Þ
dt
The rate parameters αm and β m that govern this process depend on the membrane
potential V in a nonlinear fashion. The two other gating variables h and n follow
the same formalism with membrane potential-dependent rate constants αh, β h, αn,
and β n. Hodgkin and Huxley determined these nonlinear relationships between the
rate parameters and membrane potential from voltage clamp experiments.
Over time, other ion channels were introduced into the model using the same
formalism as for the sodium and potassium channels. Since the development of
computer technology, the Hodgkin and Huxley formalism has been widely used in
simulations of neural systems ranging from very detailed models of single neurons
(e.g., De Schutter and Bower, 1994a, b) to large-scale networks of neocortex (e.g.,
Traub et al., 2005; van Drongelen et al., 2006).
Although Hodgkin and Huxley’s model only contains four variables (V, m, h, n),
it is still too complex to approach analytically. Several authors solved this problem
by reducing the 4D model to a 2D one; the FitzhughNagumo model (Fitzhugh,
1961) is an example of such a reduction. In these models, the gating variable m of
the Hodgkin and Huxley model is removed by considering sodium activation to be
instantaneous; subsequently, h and n are combined into a single recovery variable w.
Fitzhugh used the following pair of coupled differential equations:
dV dw
5 Vða 2 VÞðV 2 1Þ 2 w 1 I and 5 bV 2 cw ð2:3Þ
dt dt
in which a, b, and c are parameters; I is a term representing injected current. The

equations are coupled because w occurs in the expression for dV/dt and V in the
expression for dw/dt. The remaining two variables in these models are the mem-
brane potential V and a single recovery variable w, generating a 2D model that is
amenable to mathematical analysis (for an excellent discussion of simplified ver-
sions of the Hodgkin and Huxley model, see Izhikevich, 2007).
Nonparametric models describe a system’s inputoutput relationship, usually

by using a large number of parameters, and these parameters do not necessarily
have a physical interpretation. Generally speaking, a nonparametric model is gener-
ated from a procedure in which we relate a system’s input x(t) and output y(t). Just
as we can relate two variables with a function, we can link two time series with an
operator. An example of such a nonparametric model would be the characteriza-
tion of an LTI dynamical system with its (sampled) unit impulse response (UIR)
(Fig. 2.1B). The operator in this case would be convolution, since convolution of
the input time series x(t) with the system’s UIR h(t) generates the system’s output
time series y(t): yðtÞ 5 hðtÞ xðtÞ (see section 8.3.1.1 in van Drongelen, 2007).
Although one might point out that such a nonparametric description does not neces-
sarily provide direct insight into the system’s components or the mechanisms
underlying the system’s operation, the curve of the UIR permits us to predict the
system’s response to any input, such as the unit step function (Fig. 2.1B).
2.4 Polynomials
For static systems, both linear and nonlinear, one can use algebraic expressions to
describe their inputoutput characteristic, and polynomials are often used for this
purpose. Polynomials are sums of monomials, which are expressions that consist of
a constant multiplied by one or more variables; the exponent of the variable is its
degree. For example, zðtÞ 5 axðtÞ4 yðtÞ3 is a monomial with a constant (parameter) a
and a degree of 4 for x and 3 for y. We can see that this expression represents a static
process because at any time t, output z depends only on the present values of inputs
x and y. It is important to note that although the relationship between z and x, y is
nonlinear, the expression can be considered a linear function of the parameter a.
2.4.1 Describing Discrete Time Data Sets

Applying the above to the characterization of nonlinear systems, we could describe
the relationship between input x(t) and output y(t) of a static nonlinearity (a nonlin-
ear system without memory) with a polynomial such as the following power series:
X
N
yðtÞ 5 a0 1 a1 xðtÞ 1 a2 xðtÞ2 1 a3 xðtÞ3 1 ? 1 ai xðtÞi 1 ? 5 ai xðtÞi ð2:4Þ
i50
In principle, power series are infinite; however, in our applications they will always
consist of a finite number of monomials. The fact that Equation (2.4) is linear with
respect to its parameters ai can be used to fit the series by using a technique
called least squares minimization. Using this approach of fitting polynomials to
recorded data sets is often called regression analysis. This procedure works as fol-
lows. Suppose we have two sets of N measurements: a system’s input xn and asso-
ciated output yn. If we model our system as a second-order static system, we can
Modeling 27
truncate the expression in Equation (2.4) above the second power and estimate the
output y as a0 1 a1 xn 1 a2 x2n : Subsequently we can define the error of our estimate
ε2 as:
X
N
ε2 5 ½yn 2 ða0 1 a1 xn 1 a2 x2n Þ 2 ð2:5Þ
n51
By following the same approach we used to find the coefficients in Lomb’s algo-
rithm (Section 1.2.1), we can find the minimum associated with the best choice for
parameters a0, a1, and a2 by setting the partial derivatives of ε2 (with respect to
these three parameters a0, a1, and a2 ) equal to zero:
@ε2 X N @½yn 2 ða0 1 a1 xn 1 a2 x2n Þ

5 2 yn 2 ða0 1 a1 xn 1 a2 x2n Þ 50
@ai n51 @ai ð2:6aÞ
for i 5 0; 1; 2
and we get what are called the normal equations:
@ε2 XN
5 2 2 ½yn 2 a0 2 a1 xn 2 a2 x2n 5 0
@a0 n51
X
N X
N X
N
N 1 a1
-a0 |{z} x n 1 a2 x2n 5 yn
n51 n51 n51
X
N
1
n51
@ε2 XN
5 2 2 ½yn 2 a0 2 a1 xn 2 a2 x2n xn 5 0
@a1 n51
X
N X
N X
N X
N
-a0 xn 1 a1 x2n 1 a2 x3n 5 yn xn
n51 n51 n51 n51
@ε 2 X
N
5 2 2 ½yn 2 a0 2 a1 xn 2 a2 x2n x2n 5 0
@a2 n51
X
N X
N X
N X
N
-a0 x2n 1 a1 x3n 1 a2 x4n 5 yn x2n ð2:6bÞ
n51 n51 n51 n51
Note that in Equation (2.6b) all summation (Σ) expressions are numbers that can
be computed from the observations; therefore, there are three linear equations with
three unknown parameters a0, a1, and a2 to compute (this should be no problem
provided, of course, that the set of equations can be solved). Note that if we had
truncated Equation (2.4) at a1, the normal equations that we would have obtained
would have been the well-known equations for linear regression.
It is a bit tedious to solve the three equations in (2.6b); therefore, one might prefer
to solve the coefficients by using the matrix notation XA 5 Y for the three equations:
2 3 2 3
X
N XN X N
6 N xn 2
xn 7 2 3 6 yn 7
6 7 a0 6 n51 7
6 n51 n51 7 6 7 6 7
6X X X 7 6 7
6 7 6X
N N N N
6 37 6 7 7
6 xn xn2
xn 7 6 7a 1 5 6 y n xn 7 ð2:6cÞ
6 n51 7 4 5 6 7
6 n51 n51 7 6 n51 7
6X N X N XN 7 6X N 7
4 2 3 45 a2 4 25
xn xn xn |fflfflffl{zfflfflffl} y x
n n
n51 n51 n51 A n51
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl} |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
X Y
The coefficients can be found by solving (Equation (2.6c)): that is, A 5 X 21 Y: In

MATLAB we can use the \ operator to obtain this result: A 5 X\Y . An example
( Pr2_1.m ) for approximating an exponential function y 5 ex is available on http://
www.elsevierdirect.com/companions/9780123849151.
2.4.2 Describing Analytic Functions

The previous example works with discrete time data such as a set of digitized
recordings of a system’s input and output. In other applications, one might deal with
a parametric model and consequently have access to analytic functions that describe
some nonlinear system under investigation (recall that an analytic function is
smooth and differentiable and that this is not the same as the analytic signal we
introduced for the Hilbert transform in Chapter 1). In this case, the so-called
Maclaurin or Taylor series approaches, which will be explained in Sections 2.4.2.1
and 2.4.2.2, may be applied to convert the function into a power series. Such a power
series approach can also be helpful for creating a linear approximation of a nonlinear
function in the neighborhood of a point of interest. Because linear relationships are
easier to analyze than nonlinear ones, this technique of linearization of nonlinear
functions can help us understand the behavior of complex nonlinear processes.
Like the polynomials discussed in the previous section, the Maclaurin and Taylor
series describe static systems. To describe dynamical systems, we can use the
Volterra series, which is discussed in detail in Chapter 3. In Section 2.5, we will
show that the Taylor series can be considered the static version of a Volterra series.
2.4.2.1 Maclaurin Series

A famous power series describing a function about the origin is the Maclaurin
series. Let us consider an example with the exponential function and use the power
series approach in Equation (2.4) to represent this function:
X
N
f ðtÞ 5 et 5 a0 1 a1 t 1 a2 t2 1 a3 t3 1 ? 1 ai ti 1 ? 5 ai t i ð2:7Þ
i50
Modeling 29
The task at hand is to determine the values of the coefficients ai for function et. We
can use the following approach to perform this task. First we determine the deriva-
tives of f.
f ðtÞ 5 et 5 a0 1 a1 t 1 a2 t2 1 a3 t3 1 ? 1 ai ti 1 ?
df ðtÞ
5 et 5 a1 1 2a2 t 1 3a3 t2 1 ? 1 iai ti21 1 ?
dt
d2 f ðtÞ ð2:8Þ
5 et 5 2a2 1 ð2 3 3Þa3 t 1 ? 1 ði 3 ði 2 1ÞÞai ti22 1 ?
dt2
d3 f ðtÞ
5 et 5 ð2 3 3Þa3 1 ? 1 ði 3 ði 2 1Þ 3 ði 2 2ÞÞai ti23 1 ?
dt3
^
The second step is to consider f(t) 5 et about the origin. As we approach the origin
(i.e., t becomes 0), Equation (2.8) simplifies to:
f ð0Þ 5 e0 5 15 ½a0 1 a1 t 1a2 t2 1a3 t3 1? 1 ai ti 1 ?t50 5 a0

df ð0Þ
5 e0 5 15 ½a1 1 2a2 t 1 3a3 t2 1 ?1 iai ti21 1 ?t50 5 a1
dt
d2 f ð0Þ
5 e0 5 15 ½2a2 1ð23 3Þa3 t 1?1 ði3 ði21ÞÞai ti22 1 ?t50 5 2a2 ð2:9Þ
dt2
d3 f ð0Þ
5 e0 5 15 ½ð2 3 3Þa3 1? 1 ði3 ði21Þ 3 ði22ÞÞai ti23 1 ?t50
dt3
5 ð23 3Þa3
^
With the results obtained in Equation (2.9), we can see that for the function, the
values for the coefficients ai are:
1
ai 5 ð2:10Þ
i!
Combining this result in Equation (2.10) with Equation (2.7), we have found the
well-known power series expansion of the exponential function:
1 1 1 1 XN
1 i
f ðtÞ 5 et 5 1 1 t 1 t2 1 t3 1 ? 1 ti 1 ? 5 t ð2:11Þ
1! 2! 3! i! i50
i!
P
N
In the last expression ð1=i!Þ ti ; we use the definition 0! 1. Note that by using
i50
this approach, we include only values of t—there are no previous or future values
(t6τ) included in the power series; therefore, this approach is static (or memory-
less). An example of this approximation is implemented in MATLAB script Pr2_1.m
(http://www.elsevierdirect.com/companions/9780123849151).
In the above example we used the exponential exp(t) for f(t); if we consider the
development of Equation (2.4) for any function that can be differentiated, we get:
f ð0Þ 5 ½a0 1 a1 t 1 a2 t2 1 a3 t3 1 ? 1 ai ti 1 ?t50 5 a0 -a0 5 f ð0Þ

df ð0Þ
5 ½a1 1 2a2 t 1 3a3 t2 1 ? 1 iai ti21 1 ?t50 5 a1 -a1 5 f 0 ð0Þ
dt
d2 f ð0Þ f 00 ð0Þ
5 ½2a 2 1 ð2 3 3Þa3 t 1 ? 1 ði 3 ði21ÞÞai t i22
1 ? t50 5 2a 2 -a 2 5
dt2 2
d3 f ð0Þ
5 ½ð2 3 3Þa3 1 ? 1 ði 3 ði21Þ 3 ði22ÞÞai ti23 1 ?t50
dt3
f 000 ð0Þ
5 ð2 3 3Þa3 -a3 5
ð2 3 3Þ
^
ð2:12Þ
Here the notation f 0 ð0Þ; f 00 ð0Þ; f 000 ð0Þ; . . . are not functions but represent the numbers
computed as the value of the 1st, 2nd, 3rd, . . . derivatives of f at t 5 0. From this more
general notation we obtain the expression for the so-called Maclaurin series of f(t):
1 0 1 1
f ðtÞ 5 f ð0Þ 1 t f ð0Þ 1 t2 f 00 ð0Þ 1 t3 f 000 ð0Þ 1 ? ð2:13Þ
1! 2! 3!
2.4.2.2 Taylor Series

In the above example, we developed the power series for a function about the ori-
gin. The development of the Taylor series follows a similar approach but now
about any point α. For a power series of power N this becomes:
f ðtÞ 5 a0 1 a1 ðt 2 αÞ 1 a2 ðt2αÞ2 1 a3 ðt2αÞ3 1 ? 1 ai ðt2αÞi 1 ?

X
N ð2:14Þ
5 ai ðt2αÞi
i50
We will now use a similar approach for the development of this series about α as
we used in the Maclaurin series about the origin—except in this case we set t 5 α
Modeling 31
(instead of t 5 0) so that all terms in Equation (2.14) with ðt2αÞi vanish. By fol-
lowing this procedure we get:
f ðαÞ 5½a0 1 a1 ðt2αÞ 1 a2 ðt2αÞ2 1a3 ðt2αÞ3 1? 1 ai ðt2αÞi 1 ?t 5 α
5 a0 -a0 5 f ðαÞ
df ðαÞ
5 ½a1 1 2a2 ðt2αÞ 1 3a3 ðt2αÞ2 1? 1 iai ðt2αÞi21 1 ?t 5α
dt
5 a1 -a1 5 f 0 ðαÞ
d f ðαÞ
2
5 ½2a2 1 ð2 33Þa3 ðt2αÞ 1 ?1ði3 ði21ÞÞai ðt2αÞi22 1 ?t 5 α
dt2 ð2:15Þ
f 00 ðαÞ
5 2a2 -a2 5
2
d3 f ðαÞ
5 ½ð23 3Þa3 1?1 ði3 ði21Þ 3 ði22ÞÞai ðt2αÞi23 1 ?t 5 0
dt3
f 000 ðαÞ
5 ð2 3 3Þa3 -a3 5
ð23 3Þ
^
Similar to the notation used in the previous section, the notation f 0 ðαÞ; f 00 ðαÞ;
f 000 ðαÞ; . . . does not refer to functions, but represents the numbers computed as the
value of the 1st, 2nd, 3rd, . . . derivatives of f at t 5 α. Substituting the findings in
Equation (2.15) into Equation (2.14) we obtain the Taylor series about t 5 α:
1 1 1
f ðtÞ 5 f ðαÞ 1 ðt 2 αÞf 0 ðαÞ 1 ðt2αÞ2 f 00 ðαÞ 1 ðt2αÞ3 f 000 ðαÞ 1 ? ð2:16Þ
1! 2! 3!
Comparing Equations (2.13) and (2.16), we can establish that the Maclaurin series is
the same as a Taylor series computed about the origin (i.e., α 5 0). This approach can
be extended to higher-dimensional systems with multiple inputs; see Appendix 2.1
for examples of the 2D case. It must be noted that in many texts the distinction
between Maclaurin and Taylor series is not always made and it is not uncommon to
use the term Taylor series for both, a habit we will adopt in the following.
The number of terms in a Taylor series may be infinite. However, if we evaluate
a system close to an equilibrium at the origin or α, the value of t or (t2α) is a
small number {1; therefore, one can ignore higher-order terms in the power series
tn or (t2α)n because they become increasingly smaller. Thus, in general we can
approximate any function close to α with a linear expression obtained from a
Taylor series in which higher-order terms are ignored f ðtÞ f ðαÞ 1 ðt 2 αÞf 0 ðαÞ; or,
in the case where we evaluate the expression about the origin, we obtain the
approximation f ðtÞ f ð0Þ 1 tf 0 ð0Þ: This technique of linearizing a nonlinear func-
tion plays an important role in the analysis of nonlinear systems. A system’s
behavior in a restricted part of its domain can be understood and approximated by

a linear version of its characteristic. Sometimes, with the more complex systems, a
piecewise approximation with linear functions is the best option for their analysis.
For example, if we wanted to evaluate sin(t) around the origin, we can apply
Equation (2.13) and find the series:
t2 t3
sinðtÞ 5 sinð0Þ 1 t cosð0Þ 2 sinð0Þ 2 cosð0Þ 1 ?
|fflfflffl{zfflfflffl} |fflfflffl{zfflfflffl} 2 |fflfflffl{zfflfflffl} 6 |fflfflffl{zfflfflffl}
0 1 0 1
For small values of t (around 0) we may ignore all higher-order terms and we find
that sinðtÞ t: In general, such an approach may be useful if one studies a system
close to an equilibrium. For example, if one examines a neuron’s subthreshold
behavior, one must describe the membrane potential close to the resting potential; in
this case it makes sense to linearize the nonlinear equations that govern the cell’s
electrical activity around resting potential. An example of this approach, where the
nonlinear Hodgkin and Huxley equations are linearized, can be found in Chapter 10
in Koch (1999).
When fitting a truncated power series to an analytic function, one could truncate
the Taylor series at the desired order. However, due to the error introduced by trunca-
tion, one may actually obtain a better fit by using a linear regression approach. An
example is if one wants to approximate et with a second-order power function over a
limited interval. The truncated Taylor series (see Equation (2.11)) is 1 1 t 1 0:5t2 ; but
with a regression approach over the interval [1,1] one obtains a better fit with
0:9963 1 1:1037t 1 0:5368t2 : This can be seen by running MATLAB script Pr2_1
(available on http://www.elsevierdirect.com/companions/9780123849151) where the
original exponential function (red), the Taylor series (blue), and the regression result
(black) are superimposed. The regression approach for obtaining a power series
approximation is also a valid solution if the Taylor series cannot be applied, as in the
case of a function that is nonanalytic, such as y 5 jxj (no [unique] derivative at x 5 0).
2.5 Nonlinear Systems with Memory

In the above examples, the output y(t) of the nonlinear systems could be described
with a polynomial of x(t) because there was a direct relationship between x and y;
that is, in these examples there was no memory in the system. However, nonlinear
systems with memory do exist, and for these systems we must describe how the
output y(t) depends on both the present and the past input: x(t2τ) with τ $ 0.
In the following chapter, we will consider the details of the so-called Volterra
series for the characterization of dynamical nonlinear systems (nonlinear systems
that do have a memory). Here we will demonstrate the similarities between the
Volterra and Taylor series. With the Taylor series we can link output value y 5 f(x)
to input value x in the following manner:
1 1 1
y 5 f ðαÞ 1 ðx 2 αÞf 0 ðαÞ 1 ðx2αÞ2 f 00 ðαÞ 1 ðx2αÞ3 f 000 ðαÞ 1 ? ð2:17Þ
1! 2! 3!
Modeling 33
In the example below, we will approximate a nonlinearity with a series truncated at

the second order:
yðtÞ 5 a0 1 a1 xðtÞ 1 a2 xðtÞ2 ð2:18Þ
Before we introduce the Volterra series, we generalize the procedure in which

we relate two values x and y into a slightly altered procedure in which we relate a pair
of time series x(t) and y(t). Just as we can relate two values x and y with a function f:
y 5 f ðxÞ ð2:19aÞ
we can link two time series x(t) and y(t) with an operator F:
yðtÞ 5 FfxðtÞg ð2:19bÞ
Note: In some texts on Volterra series F will be called a functional. Because

F connects two functions x(t) and y(t), it is better to use the term “operator”
because strictly speaking, a functional maps a function onto a value, whereas
an operator maps one function to another function.
A Volterra series can perform such an operation:

ð
N
yðtÞ 5 h0 1 h1 ðτ 1 Þxðt 2 τ 1 Þdτ 1

|{z}
0th order term 2N
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
1st order term
ð N
N ð
1 h2 ðτ 1 ; τ 2 Þxðt 2 τ 1 Þxðt 2 τ 2 Þdτ 1 dτ 2 1 ?
2N 2N
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
2nd order term
N ð
ð N ð
N
1 ? hn ðτ 1 ; τ 2 ; ?; τ n Þxðt2τ 1 Þxðt2τ 2 Þ?xðt2τ n Þdτ 1 dτ 2 ?dτ n ð2:20Þ

2N2N 2N
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl
ffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
nth order term
Do not be intimidated by this first appearance of the expression for the Volterra
series. In the following text and Chapter 3 we will discuss and explain the different
components of this representation. At this point it is worthwhile to mention that the
Volterra series is essentially the convolution integral extended to nonlinear sys-
tems. We could simplify the notation in Equation (2.20) with the commonly used
symbol for convolution (chapter 8 in van Drongelen, 2007), and we get:
y 5 h0 1 h1 x 1 h2 x x 1 ? 1 hn x ffl{zfflfflfflfflfflfflfflfflffl
|fflfflfflfflfflfflfflfflffl ? xffl}
n copies of x
In the Volterra series (Equation 2.20), input function x(t) determines the output
function y(t). The expression is analogous to the Taylor series except that the dif-
ferentials of the Taylor series are replaced by integrals. The symbols h0, h1, h2, and
hn represent the so-called Volterra kernels. The term “kernel” is uniquely defined
for this type of series and should not be confused with the use of this term in com-
puter science or other areas in mathematics. Note that the first-order component
Ð
N
h1 ðτ 1 Þxðt 2 τ 1 Þdτ 1 in the Volterra series is the convolution integral (see section
2N
8.3.1.1 in van Drongelen, 2007) and the higher-order components in Equation
(2.20) are convolution-like integrals. Thus for a linear system, kernel h1 is the UIR.
Representations that utilize Volterra series are usually nonparametric—that is, one
can predict system output when the input is known, but one cannot necessarily
intuit the system’s components or underlying mechanisms. In the following we will
examine examples of the relationship between Volterra and Taylor series. See also
Chapter 3 for further details on the Volterra series.
Despite the similarities between the Taylor series in Equation (2.17) and the
Volterra series in Equation (2.20) discussed above, it may not be immediately obvi-
ous that they are related. Therefore, we will discuss the similarities for a simple
dynamical nonlinear system, which we will subsequently transform into a static
nonlinear one. Let us consider a dynamical second-order system that consists of
a cascade of a dynamical linear component and a static nonlinear module
(Fig. 2.2A). Such a cascade approach with the dynamics in the linear component
combined with static nonlinearities is frequently applied in dynamical nonlinear
system analysis. In this example, we have the linear component’s UIR h(t) and the
static second-order nonlinear component a0 1 a1 y 1 a2 y2 (Equation (2.18)). From
Fig. 2.2A we can establish that the output y of the linear module can be obtained
from the convolution of the input x and the linear module’s UIR h:
ð
N
yðtÞ 5 hðτÞxðt 2 τÞdτ ð2:21Þ

2N
Second-order dynamical nonlinear system

Figure 2.2 (A) Diagram of a
second-order dynamical nonlinear
system consisting of a cascade of a
dynamical LTI system and a
Second-order static second-order static nonlinearity.
nonlinearity (B) A similar system for which the
dynamical linear component is
Second-order static nonlinear system replaced by a static one.
Second-order static
nonlinearity
Modeling 35
The cascade’s final output z can be obtained from the static nonlinearity character-
istic by substituting the output of the linear component (Equation (2.21)) into the
input of the static nonlinearity (Equation (2.18)):
2 32
ð
N ð
N
zðtÞ 5 a0 1 a1 hðτÞxðt 2 τÞdτ 1 a2 4 hðτÞxðt2τÞdτ5 ð2:22Þ

2N 2N
This can be rewritten as:
ð
N
zðtÞ 5 a0 1 a1 hðτÞxðt 2 τÞdτ

20
2N 10 N 13
ð
N ð
1 a2 4@ hðτ 1 Þxðt 2 τ 1 Þdτ 1 A@ hðτ 2 Þxðt 2 τ 2 Þdτ 2 A5 ð2:23Þ
2N 2N
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
Nð N ð
hðτ 1 Þhðτ 2 Þxðt 2 τ 1 Þxðt 2 τ 2 Þdτ 1 dτ 2
2N 2N
This expression can be rearranged in the form of the Volterra series shown in
Equation (2.20):
1st order term

zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
0th order term ð
N
z}|{
zðtÞ 5 a0 1 a1 hðτÞ xðt 2 τÞdτ
|{z} |fflfflffl{zfflfflffl}
h0 2N h1 ðτÞ
2nd order term

zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
ð N
N ð
1 a2 hðτ 1 Þhðτ 2 Þ xðt 2 τ 1 Þxðt 2 τ 2 Þdτ 1 dτ 2 ð2:24Þ
|fflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflffl}
2N 2N h2 ðτ 1 ;τ 2 Þ
Equation (2.24) shows that the system in Fig. 2.2A can be characterized by a
Volterra series for a second-order system with Volterra kernels h0, h1, and h2.
To demonstrate that the Taylor series is the static equivalent of the Volterra
series, we show the equivalence of Equation (2.24) to the power series in Equation
(2.18). To accomplish this, we consider the case where our dynamical component
in the cascade becomes static; the linear component is now replaced by the static
function yðtÞ 5 xðtÞ: In other words, the linear module’s UIR is the unit impulse δ
itself, indicating that for this linear component output equals input (Fig. 2.2B).
Therefore, we can substitute δ(t) for h(t) in Equation (2.24):
ð
N
zðtÞ 5 a0 1 a1 δðτÞxðt 2 τÞdτ

2N
|fflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
xðtÞ
ð N
N ð
1 a2 δðτ 1 Þδðτ 2 Þxðt 2 τ 1 Þxðt 2 τ 2 Þdτ 1 dτ 2 ð2:25Þ
2N 2N
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
xðtÞxðtÞ 5 xðtÞ2
5 a0 1 a1 xðtÞ 1 a2 xðtÞ2
Thus, in the static case, we can use the Volterra series to recover
zðtÞ 5 a0 1 a1 xðtÞ 1 a2 xðtÞ2 ; which is the original expression of the power series in
Equation (2.18).
Note: The integrals in Equation (2.25) are evaluated using the property of the
Ð
N
unit impulse δ (see section 2.2.2 in van Drongelen, 2007): xðτÞδðτÞdτ 5
NÐ 2N
xð0Þ and accordingly xðt 2 τÞδðτÞdτ 5 xðtÞ:
2N
Appendix 2.1
Taylor Series for a 2D Function
We can extend the Taylor series in Equation (2.16) to a function f ðτ; σÞ of two
variables τ and σ. In the case where we can subdivide the function into two sepa-
rate ones (e.g., f ðτ; σÞ 5 f ðτÞ 1 f ðσÞ or f ðτ; σÞ 5 f ðτÞf ðσÞ), we can compute the
Taylor series for each function f ðτÞ and f ðσÞ and add or multiply the individual
series to obtain the expression for f ðτ; σÞ: Such an approach would work if, for
example, f ðτ; σÞ 5 eτ sinðσÞ:
Alternatively, one can approach the development of a 2D Taylor series more
generally, and consider f about point α, β.
f ðτ; σÞ 5 a00 1 a10 ðτ 2 αÞ 1 a01 ðσ 2 βÞ 1 a20 ðτ2αÞ2 1 a11 ðτ 2 αÞðσ 2 βÞ
1 a02 ðσ2βÞ2 1 a30 ðτ2αÞ3 1 a21 ðτ2αÞ2 ðσ 2 βÞ
1 a12 ðτ 2 αÞðσ2βÞ2 1 a03 ðσ2βÞ3 1 a40 ðτ2αÞ4 1 ?

ðA2:1:1Þ
Modeling 37
Using a similar approach as the one for the single-variable Taylor series, we set τ
and σ to α and β and find f ðα; βÞ 5 a00 : To find the other coefficients we use par-
tial differentiation of f at point α, β:
@f ðα; βÞ @f ðα; βÞ @2 f ðα; βÞ

5 a10 ; 5 a01 ; 5 2a20 ;
@τ @σ @τ 2
ðA2:1:2Þ
@2 f ðα; βÞ @2 f ðα; βÞ
5 a11 ; 5 2a02
@τ @σ @σ2
This technique can be used to obtain the full power series of f. In most applications
we are interested in the linear approximation of the 2D series:
@f ðα; βÞ @f ðα; βÞ
f ðτ; σÞ f ðα; βÞ 1 ðτ 2 αÞ 1 ðσ 2 βÞ ðA2:1:3aÞ
@τ @σ
The higher-order nonlinear terms are often not considered because we assume that
we only look at f closely around point α, β; therefore, τ2α and σ2β are very
small numbers, and higher powers of these small contributions are even smaller. In
other words, when f is in the neighborhood of point α, β, the function can be
approximated with the linear terms in Equation (A2.13a). In many cases, especially
in physics literature, you may encounter an alternative notation for the linear
approximation of a nonlinear system. The small fluctuations τ2α and σ2β around
α, β are indicated as perturbations δτ and δσ; and the notation for f ðα; βÞ;
ð@f ðα; βÞÞ=@τ; and ð@f ðα; βÞÞ=@σ is changed to ½f α;β ; ½@f=@τα;β ; and ½@f=@σα;β :
2 3 2 3
@f @f
f ðτ; σÞ ½f α;β 1 4 5 δτ 1 4 5 δσ ðA2:1:3bÞ
@τ @σ
α;β α;β
Again, recall that in this notation ½f α;β ; ½@f=@τα;β ; and ½@ f=@σα;β represent the
coefficients in the equation. They are numbers and not functions, since these repre-
sent the function and its derivatives when evaluated at point α, β. An example of
an application that linearizes the nonlinear Hodgkin and Huxley equations can be
found in Chapter 10 of Koch (1999).
3 Volterra Series
3.1 Introduction
Most physiological systems cannot be modeled successfully as linear systems. At
best, a linear model can be considered an approximation of physiological activity
in cases where the output of a physiological system behaves (almost) linearly over
a limited range of the input. In the following, we extend the convolution integral
that describes the behavior of linear devices to the convolution-like Volterra series,
which can be used to represent nonlinear systems. Because the expressions for
higher-order nonlinear terms require significant computational resources and
become very complex to deal with, we will demonstrate the general principles for
second-order systems. See Schetzen (2006) if you are interested in details of higher-
order systems.
In a linear time invariant (LTI) system, the convolution integral links output y(t)
and input x(t) by means of its weighting function h(t) (Fig. 3.1) (Chapter 8 in van
Drongelen, 2007):
ð
N
yðtÞ 5 hðtÞ xðtÞ 5 hðτÞ xðt 2 τÞdτ ð3:1Þ

2N
Here symbolizes the convolution operation and the system’s weighting func-
tion h(t) is its unit impulse response (UIR). This role of h(t) can be verified by
using a unit impulse δ(t) as the input. In this case we obtain (using the sifting
property):
ð
N
hðτÞδðt 2 τÞdτ 5 hðtÞ ð3:2Þ

2N

Note: In the following we will use the sifting property of the unit impulse
repeatedly (for a discussion, see section 2.2.2 in van Drongelen, 2007). The sift-
ing property is defined as:
ð
N ð
N
xðtÞ 5 xðτÞδðτ 2 tÞdτ 5 xðτÞδðt 2 τÞdτ

2N 2N
The unit impulse δ has properties of a function with even symmetry; there-
fore, the evaluation of the integral above is the same for δðt 2 τÞ and δðτ 2 tÞ:
You can also see that this must be the case since the outcome is δð0Þ for t 5 τ
in both cases, δðt 2 τÞ and δðτ 2 tÞ:
Such an LTI system shows superposition and scaling properties. For instance, if we
introduce a scaled delta function Cδ(t) (C—constant) at the input, we get a scaled
UIR function Ch(t) at the output:
ð
N ð
N
hðτÞCδðt 2 τÞdτ 5 C hðτÞδðt 2 τÞdτ 5 ChðtÞ ð3:3Þ

2N 2N
h(t) d
x(t) h2(τ1,τ2) h2(τ1,τ2)x(t − τ1)x(t − τ2) dτ1 dτ2
h2(τ1,τ2)
h1(τ1)x(t − τ1) dτ1
h2(τ1,τ2)x(t − τ1)x(t − τ2) dτ1 dτ2 + τ ...

h2(τ1,τ2,τ3, ... ,τn )
h2(τ1,τ2, ...,τn )x(t − τ1)x(t − τ2)...x(t − τn) dτ1 dτ2 ...dτn
Figure 3.1 Example of LTI and NLTI systems. (A) A linear system. (B) A second-order
system and (C) a combined nth order system. The output of the first-order components is
determined by the convolution integral, and the output of higher-order components is
obtained from convolution-like integrals. The output of the nth order system is represented
by a Volterra series consisting of the sum of the individual components, each determined by
a convolution-like expression.
Volterra Series 41
Now consider a system that is governed by an equation that is convolution-like:
ð
N ð
N
yðtÞ 5 h2 ðτ 1 ; τ 2 Þxðt 2 τ 1 Þxðt 2 τ 2 Þdτ 1 dτ 2 ð3:4Þ

2N 2N
Unlike the convolution in Equation (3.1), this system works on two copies of input
x(t) instead of only one. As we discussed in Section 2.5, such a system is an exam-
ple of a so-called second-order Volterra system. Note that the double integral in
Equation (3.4) is identical to the last term in the expression in Equation (2.24). If
we determine the UIR for the system in Equation (3.4), we get:
ð
N ð
N
h2 ðt; tÞ 5 h2 ðτ 1 ; τ 2 Þδðt 2 τ 1 Þδðt 2 τ 2 Þdτ 1 dτ 2 ð3:5Þ

2N 2N
Here we applied the sifting property twice: once for each of the delays τ 1 and τ 2.
The result h2(t,t) is the diagonal of kernel h2.
Note: You can see that in a second-order Volterra system, the UIR h2(t,t) does
not fully characterize the system (unlike the situation in an LTI system).
Instead it only characterizes the 2D function h2 ðτ 1 ; τ 2 Þ along the diagonal
τ 1 5 τ 2 in the τ 1 ; τ 2 plane. As we will see in Section 3.3 we need sets of
paired impulses to fully characterize h2.
The system in Equation (3.4) is nonlinear because scaling does not hold. For exam-
ple, the response to a scaled delta function Cδ(t) at the input is:
ð
N ð
N
h2 ðτ 1 ; τ 2 ÞCδðt 2 τ 1 ÞCδðt 2 τ 2 Þdτ 1 dτ 2 5

2N 2N
ð N
N ð ð3:6Þ
C 2
h2 ðτ 1 ; τ 2 Þδðt 2 τ 1 Þδðt 2 τ 2 Þdτ 1 dτ 2 5 C h2 ðt; tÞ
2
2N 2N
By comparing Equations (3.5) and (3.6) we can see that in this system the UIR
h2(t,t) scales with C2 instead of C. As we will show in Section 3.2.1, superposition
does not hold for this system either, but showing that scaling does not hold is suffi-
cient to negate linearity of the system. In the remainder of this chapter we will con-
tinue our introduction of Section 2.5 by studying the properties of the Volterra
series and applying it for the characterization of higher-order systems.
3.2 Volterra Series

The mathematician Vito Volterra used series of convolution-like expressions to
define the inputoutput relationship of NLTI systems:
1st order term
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
ð
N
yðtÞ 5 h1 ðτ 1 Þxðt 2 τ 1 Þdτ 1

2N
2nd order term
zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{
ð N
N ð
1 h2 ðτ 1 ; τ 2 Þxðt 2 τ 1 Þxðt 2 τ 2 Þdτ 1 dτ 2 1 ?
2N 2N
ð
N ð
N ð
N
1 ? hn ðτ 1 ; τ 2 ; . . .; τ n Þxðt 2 τ 1 Þxðt 2 τ 2 Þ. . .xðt 2 τ n Þdτ 1 dτ 2 . . . dτ n

2N 2N 2N
|fflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflfflffl}
nth order term
ð3:7Þ
The output y(t) of an nth order system depends on multiple copies of the input and
is the sum of the 1st, 2nd, . . ., nth order convolution-like expressions. The func-
tions h1, h2, . . ., hn are called the 1st, 2nd, . . ., nth order Volterra kernels. In some
texts, a zero-order kernel (h0) representing a DC term, or offset, is added to
Equation (3.7). Just as in an LTI system, y(t) is the UIR if the input x(t) is a unit
impulse δ(t). In higher-order systems, the contribution of the nth order Volterra
kernel to the UIR is a so-called diagonal slice through the kernel, that is, a section
through the kernel with all delays τ 1, τ 2, . . ., τ n equal. An example for a second-
order system (n 5 2) is shown in Equation (3.5).
Note: We can refer to h1 as the UIR only if we deal with a first-order Volterra
system without a DC term—that is, a (linear) system where h1 is the only
term of y(t). In all other cases, the UIR is determined by the contributions of
all of the system’s Volterra kernels and not just by h1.
If we represent the 1st, 2nd, . . ., nth order convolution-like terms in Equation

(3.7) as H1, H2, . . ., Hn we get an alternative, simplified notation:
yðtÞ 5 H1 ½xðtÞ 1 H2 ½xðtÞ 1 ? 1 Hn ½xðtÞ ð3:8Þ
Equation (3.8) can be generalized for an nth order system:
X
N
yðtÞ 5 Hn ½xðtÞ ð3:9aÞ
n51
Other documents randomly have
different content
"H—m," chuckled Hora, glancing at a print hung upon the wall
opposite him, "that Morland would have been a cheap investment,
even if it had been a fake. As it is——"
"As it is," laughed Guy, "your capital has returned to you more
than a thousandfold. Still I can't help marvelling at your wonderful
eye for detail. You could not have been in Flurscheim's house more
than an hour, and yet I found every wire, every lock, every catch,
exactly where you told me I should find them. Some of the doors and
windows you could never have seen? How could you know?"
"It was not through any capacity for seeing through brick walls,"
said Hora drily. "But merely a deduction from what Flurscheim
himself did not tell me. He was very proud of a system of alarm
designed by himself for the protection of his treasures. He told me
that it was impossible for a window sash to be lifted or an outer door
to be unlatched without setting off the alarm—I observed from
outside that the attics were fitted with swing casements and I drew
my own conclusions."
"You omitted to inform me that the servants slept in those attics,"
remarked Guy. "I nearly stepped on the bed of one of them when I
entered the window."
"So that is why you left by the front door, was it?" enquired Hora.
"It was a little bit risky, wasn't it?"
"No," said the young man. "I calculated that I should get a
minute's start, and thirty seconds was quite enough. As a matter of
fact, I had a clear minute. I looked out into the street from a window
and saw that the coast was clear and the brougham was waiting.
There were two or three parties just leaving Lady Greyston's and I
calculated upon being able to join them without exciting observation.
The street was very much in shadow, and just between lights, after a
dance, you know, no one observes very clearly."
"Still it was a risk," observed Hora. "I should have returned by
the way I had gone."
"I still think I took the lesser evil," replied Guy. "Besides the
chance of finding Flurscheim's servants awake, there was the
possibility of being seen from the street as I passed along the
parapet back to the window of the Greyston's house. Then suppose I
had met someone on the stairs at the Greyston's. The function was
practically over. There was every likelihood that some of the servants
would be going to their quarters—it would have been deuced
unpleasant to have had to explain what I was doing there."
"At all events," remarked Hora, "you ought to have cut off the
alarm. Did you forget how to do it?"
A smile flickered across the young man's face.
"No," he said, "I left it in position on purpose. I thought I should
like to give Flurscheim a sporting chance of getting his own back.
There were just two flights of stairs and a bedroom door between us.
I thought that if that were not sufficient to enable me to get away I
should deserve to be captured."
Myra, who had been listening to the conversation in silence, half
uttered an exclamation. But she checked it so that only Hora's keen
ears heard. He smiled, but said nothing. Guy continued lazily: "You
see that I did not misjudge the conditions. I am here." Then he
repeated the words he had used a few hours previously. "You must
set me a more difficult task next time, Commandatore."
"What an enthusiast you are," remarked Hora. "If you go on at
this rate, there will be nothing left for you to do."
"I hate being idle," remarked the young man.
"Never fear, never fear," said Hora, "I have no doubt you will
manage to amuse yourself. You did so last night, did you not?"
Although the question was asked carelessly, the young man
flushed slightly as he answered: "Tolerably well."
"Only tolerably well?" asked Hora, "and yet you postponed your
enterprise until almost too late, for only 'tolerable' amusement."
"Admitted, Commandatore," answered Guy gaily, "the adverb is
not sufficient. To tell the truth, I met some very pleasant people, and
the time passed swiftly."
Myra sprung the next question.
"Who were they, Guy? Anyone I have met?"
"No," he answered. "A Captain and Mrs. Marven and——"
He did not get to the end of his sentence. Lynton Hora had risen
from his chair and interrupted him: "Who did you say? Say the name
again," he cried hoarsely.
Both Myra and Guy looked at him in amazement. Hora was not
given to showing emotion, and there could be no doubt but that he
was deeply moved. His lips were drawn closely together, beads of
perspiration broke out on his forehead, every line in his face
deepened.
"What's the matter, Commandatore—father?" cried Guy in
alarm, as he sprung to Hora's side and laid his hand on the elder
man's arm.
Hora shook off the touch roughly.
"Say the name again," he repeated.
"Marven," repeated Guy, "Captain Marven."
Gradually Hora regained control of himself. His features
resumed their normal air of petulant acquiescence with the world, but
there was a gleam in his eyes which revealed a very different spirit
within him. Presently he spoke.
"You are surprised to see me so much moved by the mention of
a name. You would not be if you knew what reason I have to hate
the possessor of it. So you found Captain Marven very good
company, eh, Guy?" He laughed sardonically.
"Why, yes," replied the young man.
"I wonder," he mused, "if you would have thought him as
entertaining if you knew the part he has played in my life."
"In your life?" queried Myra and Guy in the same breath.
"In my life," repeated Hora with deliberation. Then he continued
in accents which showed how deeply memories of the past rankled:
"That is the man, Guy, to whose actions my quarrel with the world is
due. Owing to him I found every man's hand raised against me.
Owing to him I was compelled in self-defence to raise my hand
against every man. Owing to him I became another Ishmael—thrust
out into the world, branded, a mark for every man's scorn and every
woman's jeers. Oh, I have taken my revenge upon the scorners," he
laughed harshly, "but not upon him—not upon him—yet."
He paused, and once more, it was only with an effort that he
regained control of himself. He did not again trust himself to speech.
He turned on his heel abruptly. At the door he paused.
"You have given me much to think about, Guy," he said. "At
present I am unable to think calmly. Some other time I will discuss
the matter with you."
He left the room swiftly and the firm step of his sound leg and
the following shuffle as he dragged the other foot along after it was
the only sound to be heard until the closing of another door told Myra
and Guy that he had shut himself in his own apartment.
CHAPTER III
THE MAKING OF A CRIMINAL
The philosophy of Lynton Hora had for once given way under
the stress of a deep emotion. There could be no doubt about that,
and no doubt either that the emotion which had strained the
philosophy to breaking point was the emotion of hate.
Never before had Guy seen him so wrought upon. Often he had
regretted that the man he called father should have been of so calm
a temperament—regretted even while he admired. Himself of an
impulsive, even ardent nature, he had longed to express his feelings
to the one being who had been his sole companion from infancy,
who had treated him with unfailing and unvarying kindness, but who
chilled, with what appeared to be temperamental coldness, any
expression of affection.
Guy was thrilled with the discovery that a deep sea of passion
underlay Hora's cold exterior. If Hora hated, of necessity he must
love.
He must love him, Guy Hora, his son. Did not every action in his
life show it?
The thought awakened Guy's memory actively. His earliest
memories were of the Commandatore. He had no knowledge of a
mother, or but shadowy recollections, and those might merely be the
offspring of his own imagination. Lynton Hora had been father and
mother both. Guy could recall Hora's face bending over his bed in
the days of his babyhood. He had one vivid recollection of being
parted from his father when he himself was about seven years old.
He had been left in the charge of some dark-haired, swarthy-faced
people, and they had neglected him—had beaten him. How he had
cried for his father, and when his father had returned, he
remembered running to him and sobbing out his tale of misery. He
remembered how Hora had told him that men never cried when they
were hurt, and that he, stricken with shame, had answered that it
was not the beating but the loneliness which had brought the tears to
his eyes.
Hora had smiled and had left him alone for a few minutes. He
had smiled still more when he had returned. Guy remembered
seeing the man who had beaten him later that same day with a
bruised face and an arm hanging helpless in a sling from his neck.
But that was not his most vivid memory of Hora's return. Chiefly
it was a conversation that took place when Hora had taken the boy's
hand and led him up into the mountains. Often the boy had recalled
the words which had been spoken to him. He could never see a pine
tree without their being fresh spoken to his ear, for they had been
uttered beneath the pine woods, on the edge of a translucent
mountain lake, which mirrored the snowy peaks above it so perfectly
that it seemed strange that the pebbles at the bottom could not be
counted.
Hora had taken the boy's tears as his text.
"Women weep when they are hurt," he said. "Men strike back.
Remember that, Guy; remember too that if you cannot strike with the
arm, there are other ways of driving the blow home."
Though Guy had understood the meaning of Hora's words but
dimly then, he had remembered them, and later he understood. Hora
had often given him practical illustration of his precepts. He never
forgot an injury or a slight, and Guy was often allowed to see how
Hora avenged either. Memory has no chronological exactitude, and
as Guy allowed his thoughts to drift, an instance occurred to him
which had happened some years later. They had been travelling in
France together and had been hurrying on to Italy. The one other
traveller in the same compartment had been a blusterous
Englishman of the most unpleasantly self-assertive type. Hora had
attempted to engage him in conversation and had met with a surly
repulse. When the frontier was reached, the assertive person was
asleep. Hora had dexterously possessed himself of the man's watch
and when the custom's official made his appearance had transferred,
with equal dexterity, the watch to his pocket, leaving a portion of the
chain visible. When awakened, the Englishman discovered his loss
almost immediately. The official was before him asking him in a
language he did not comprehend, whether he had any dutiable
articles to declare. The visible piece of chain caught the eye of the
excited passenger. He made a grab at the presumed thief. The
official, thinking he was being attacked by a madman, made a wild
dive for the door and reached the platform. The Englishman followed
in pursuit and captured his man. There was a wild melée, from which
the victim did not emerge victorious. When the train moved on, Hora
was gratified by seeing their late companion ineffectually struggling
in the grasp of half-a-dozen stalwart carabinieri.
Guy was fifteen years old when this event had happened, and
long before then he had imbibed from his father ideas of morality
which were directly at variance with those generally accepted. Guy
could never remember a time when Hora had bade him restrain any
desire. How well he recalled a day, he could not have been more
than six, when they had passed a shop wherein a basket of golden
oranges were displayed. "Buy me one," he had cried. Hora had
stopped. There was no one in the shop. "I'll teach you a new game,"
he said. "Go and fetch a couple, Guy. Mind you choose the best," he
said.
Guy had obeyed and Hora had praised him. As Guy ate the
oranges he thought the game the best he had ever heard of. Next
day they had passed the shop and Guy was about to repeat the
foray, but Hora had restrained him.
"Look, Guy," he said. "There is somebody there now; when you
want oranges or anything else without paying you must be quite sure
there is no one about, or you will lose the game."
Guy remembered the precept and acted upon it. It was a
delightful new game for anyone to play, if you were only clever
enough to play it properly. He used to beg Hora to take him out for a
day's stealing, and sometimes, as a reward for perseverance in his
studies, Hora would accede to the boy's request. He had no notion
that he was doing anything wrong, though he had been taught that
there were things he must not do. He knew that he must not tell his
father a lie; he knew too that he was to be silent when bidden.
Of course a time had eventually arrived when he had become
conscious that there was some lack of harmony between the life he
and his father led and the lives of those upon whom they preyed.
Hora had taken the boy to see a big penal establishment and his
curiosity had been stirred as to the reason of this gathering of men in
mud-coloured garb, marked all over with broad arrows. "Why are
they all dressed alike? Why do their masters carry guns?" he asked.
Hora had silenced him with a sign at the time, but later, when
they were alone, he had explained.
"They are all men who have been trying to play the game of
stealing and have lost," he said. "If you were to get caught, you
would be taken away and shut up at night in a cell all alone, and
dressed in ugly clothes, and when you went out men with guns
would be set to watch you so that they should shoot you if you tried
to run away."
"Have you ever been caught, father?" Guy asked.
Hora had never replied to that question. His face had grown so
dark that Guy had forborne to press for an answer, and the memory
of the singular expression which had passed over his countenance
had been sufficient to prevent Guy ever repeating the enquiry.
After the visit to the convict establishment, Guy had been
timorous at playing his new game, but Hora had chaffed him,
advised him, stood beside him, protected him, until he became
exceedingly dexterous in a variety of forms of petty larceny. He was
never allowed at this time to mix with other boys. Hora had him
always under his own eye, educating him according to a system
which was a fair sample of the average boy's education as regards
matter, but differing vastly from the average boy's education as
regards the application of the knowledge imparted to him.
Cæsar was never to him a mere handbook by which the
intricacies of a dead language were revealed, but a wonderful history
of a man who played the game of stealing in a great way. Hora made
quite clear to the boy's mind that there was only a difference in
degree between the stealing of oranges and the stealing of
kingdoms, but that if one wanted to steal kingdoms it was just as well
to begin early and learn the principles of the art by stealing oranges.
He explained, too, that the world looked with very different eyes
upon the theft of a crown and the theft of an orange or an apple. The
man who annexed an empire was an emperor whom men acclaimed
and set on a throne in a garb of purple, while the man who stole a
loaf of bread to assuage his hunger was a petty thief at whom the
world hurled opprobrium and thrust into a prison, garbed in mud-
coloured clothes and covered over with broad arrows.
Guy began to comprehend what Hora intended him to
comprehend, that there was something mean about petty theft, and
he no longer found pleasure in his game, but turned instead to the
weaving of romances of magnificent depredations.
Even the fiction which was supplied by Hora for the boy's
amusement was insidiously utilised for the inculcation of the same
perverted morality. With Robinson Crusoe, for instance, it was easy
for a man of Hora's equipment to make fun of Crusoe's naïve
dependence upon Providence and his exhibition of piety in moments
of stress. Hora pointed out that Crusoe's prayers were mere
expression of the terror of an uneducated mind when confronted with
personal danger—of a mind which had been trained in youth to rely
upon supernatural agencies for relief and comfort. He pointed out
that Crusoe really secured his own safety through the exercise of his
own constructive and observatory powers, and through no other
agency.
As Guy grew older, Hora sedulously built upon the foundation of
disbelief which he laid down as the basis of the boy's education. Guy
was taught that religion was merely the means by which a
priestarchy levied toll upon the body corporate by playing upon
inherited superstitions—while history supplied him with plenty of
illustrations. History supplied him, too, with plenty of examples to
point the arguments with which he supported what was in effect a
complete criminal philosophy. Guy was not taught only that atheism
was the hope of humanity. Hora had read much of Nietzsche, and he
skilfully adopted the Nietzschean philosophy to his purpose. A
particular appeal to Guy's mind was to be found in Hora's definition
of virtue, as a thirst for danger and courage for the forbidden. As
translated by Hora, both in precept and in practice, the highest virtue
was to be found in the breaking of laws. He imbibed the doctrines
with avidity, for Hora had a persuasive tongue. He learned at the
same time to keep them to himself, for, as Hora explained, if sheep
knew as much as men, men would have no mutton.
Until eighteen, Guy's education progressed under his father's
tuition, and then, feeling sure of him, Hora thought it time to launch
him on the world. Guy went to Oxbridge to make acquaintance with
his fellows, to survey the flock of sheep which were to supply him
with mutton in the future. The time then passed pleasantly enough,
and plenty of active exercise supplied him with a vent for his
energies. He did not shear any of the sheep, for Hora had bidden
him stay his hand. A blameless university career would, he knew, be
of great value in the future.
When Guy came down from the University it was with the
reputation of being one of its wildest spirits. Great things were
predicted of him. Others might excel him in individual efforts in the
field and the schools, but none could excel him in fearlessness of
demeanour. Besides, Hora's education had supplied him with a
serene belief in himself, which had been communicated to those with
whom he came in contact. He had been the leader of a set, the
model for the freshman, the autocrat of his time. Like most autocrats,
he cherished a profound contempt for those who bowed down before
him. He was to them as his father was to him, something so much
greater than they that their tribute became merely a thing of no
account. He understood why his father had no affection for him. How
could anyone love the thing beneath; the moth could love the star,
but the star could not love the moth—and——
Guy awoke from the reverie into which he had been betrayed by
his father's emotion on hearing the name of Captain Marven
mentioned. He was quite alone. Myra had left the room after vainly
trying to engage his attention. His hand unconsciously sought his
pocket, and, when he drew it out, he held in his palm the snuff-box
he had reserved for himself from the booty he had brought home on
the previous night. He gazed earnestly at the miniature set in the lid.
"So Captain Marven is father's enemy," he muttered, "and this—
this must be a portrait of Captain Marven's daughter."
His face grew troubled. His brow puckered. He thrust the box
back into his pocket and rose impatiently from his seat.
"Bah!" he said, "what says the Commandatore? Man is trained
for war, and woman for the relaxation of the warrior; all else is folly."
CHAPTER IV
THE REFLECTIONS OF LYNTON HORA
There was undoubted reason why the name of Marven should

move Lynton Hora to emotion. It swept him back over the thirty years
which bridged him from his youth. He would not have answered to
the name of Hora in those days; the days when he and Richard
Marven—"Gay" Marven—had been subalterns together in the same
cavalry regiment. But the name he had borne then was buried and
forgotten long since, and the young man who had borne it was dead
to the knowledge of the world, though his virtues and his sins, his
memories and hatreds—most certainly his hatreds—lived actively in
the recluse connoisseur and antiquarian, Lynton Hora.
He had had good reason for burying his earlier self—an all-
sufficing excuse for blotting out his existence from his regimental
companions, from the friends of his youth, from the parents who had
wept over his downfall perhaps even more than they had mourned
the presumed death which had followed his punishment.
The name of Marven had brought vividly before his mind a
picture of the bitterest moment of his life. Never could the memory of
that moment lose the poignancy of its sting. The hollow square in the
barrack yard, the epaulets he had once worn on his shoulder lying
on the ground, the look of scorn on the faces of his brother officers
and reflected on the faces of the men who had been till then beneath
him, for the convicted thief, he saw these things clearly at the
mention of Captain Marven's name.
He had always held that Marven was responsible for his
dishonour, Marven who had everything which he, Hora, had desired
and which fate had denied him. On the day he had first met him envy
had entered into his heart. The contented smile on Marven's face,
the expression which declared that everything is the best possible in
the best possible world, had irritated him. Hora had not shown his
irritation! Early in his youth he had learned to control the expression
of his feelings. But companionship had deepened the irritation day
by day. Gay Marven was the most popular man in the mess, Hora
the least. Marven was wealthy, a credit to a smart cavalry regiment;
Hora's allowance barely sufficed to meet the bare necessary
expenditure, and so he was debarred from indulging in the
extravagances which his comrade affected.
Some of them sneered at him and Hora attributed the sneers to
Marven's influence, though wrongfully, and his irritation became
anger.
Later, a greater cause of jealousy arose through the
interposition of the essential feminine element in all drama. Hora had
fallen hopelessly in love, and he had reason to think that the
affection he had bestowed would be returned. Then Marven
appeared on the scene, and Hora's hopes had vanished. Marven
had only to be natural to dazzle the eyes of all beholders with the
rays of his sunshiny disposition. Hora's temperament was of an
intellectual coldness more likely to provoke esteem than love. The
attack of aerotitis which affected both of them accentuated in each
his natural characteristics. Marven became more brilliant than ever,
Hora more passionately reserved. Then Hora, his natural judgment
in suspension, had imagined that it would be possible to out-dazzle
Marven. Reckless of consequences, he joined in all the pursuits from
which he had hitherto stood aloof. His useful charger had been
replaced by two magnificent mounts. His tailor had been made
temporarily happy by a swiftly swelling account. He had begun to
entertain lavishly—a year's income, apart from his pay, would not
have met one single week's expenditure. He had known that the
pace could not last, but fate had been kind to him at the outset. He
had speculated on the turf and had won. From the card table, too, he
rarely rose a loser and the play that went on in the card room of the
mess, when the Colonel was not there, would have genuinely
shocked the commanding officer, had he been aware of the amount
of the stakes at issue. Hora's comrades thought he had come in for a
legacy, and he was no longer deemed a discredit to their ranks.
Though delayed, the day when the inevitable reckoning was to
be met could not be averted forever. When fortune frowned, instead
of smiling upon his turf speculations, he was forced to visit the Jews.
There he could obtain but trifling accommodation, for he had never
had any expectations, and was heir to nothing but an unstained
name. Even the five hundred pounds he had ultimately raised was
only advanced at ruinous interest on a three months' bill. He had
plunged more wildly than ever. He had lost. He had become short of
cash to meet his daily out-of-pocket expenses. Even then, he might
have been saved from utter extinction had not he imagined that he
had succeeded in putting his rival in the shade. He had staked
everything upon one last hazard. There had been under his control
certain regimental funds. He had made use of them, knowing full well
that so soon as his anticipated engagement was announced he
would have no difficulty in obtaining a further loan, since the object of
rivalry between Marven and himself was wealthy, as well as
beautiful. With that loan he had counted on being able to replace the
money he had embezzled. But Marven was before him, and the day
Marven's engagement to Beatrice Challys was announced an
unexpected investigation of Hora's accounts by the Colonel of the
regiment disclosed the defalcation. Hora was placed under arrest,
and the torrent rushed over him.
When he had reached his own apartment, Lynton Hora spared
himself not a single pang of bitterness of the memories of what had
followed. The weary days under arrest, the long-drawn-out inquiries,
the court martial, the day of the promulgation of the sentence, when
he was drummed out of the regiment, and had walked out of the
barrack yard into the hands of the civil police, who were awaiting him
to bring him to further trial.
He had been spared nothing. There was no influence which
could have been exerted to save him from any one of the ignominies
which he had incurred. He had supplied an excellent example for
exhibiting the impartiality of the law. No private in the ranks should
be able to say that he received harsher treatment than the officer of
a crack cavalry regiment.
He had faced his punishment bravely, indeed, he had welcomed
the solitude of the cell when he had eventually exchanged his
cavalry dress for another of H. M.'s uniforms. There he had not to
meet the scorn of men's eyes.
One by one he recalled the incidents. They had never ceased to
pain him, even though he tried to laugh at his weakness in imagining
that the wound to his pride still rankled. But he would not have been
without that smarting sore. He took the same fierce satisfaction in
the pain with which the martyrs of Smithfield solaced themselves as
they thrust their arms into the fire. He told himself always that the
mental suffering, the intolerable scorn he had faced, had shown him
the world as it is, and not as it pretends to be. He postulated a
deceitful, hypocritical world with a smile on its face for the man of
wealth, and a frown and a brick for the poor devil who had the will to
enjoy and not the means to gratify his longings.
Before his disgrace he had hated only one man—afterwards he
hated all men, and at least one woman—she who preferred Gay
Marven, fortune's favourite, to himself, fortune's scapegoat. But in
addition to enabling him to appreciate the smiles and frowns of the
world at their proper worth he told himself that his experience had
made a man of him. It certainly left him a purposeful, resourceful,
scrupleless being, with a definite object in existence.
That object was revenge. Revenge on the world which had
scorned him, revenge on the world which had labelled him criminal,
revenge above all upon Marven.
He had made all his plans long before his sentence had expired.
He saw that he must die to the world if the future was to have any
promise at all, for a past, such as his, would have been an incubus
no man might carry for long. So, when his term of imprisonment was
over, he disappeared in the broad light of day. At least the ex-convict
disappeared from English eyes when he sculled out to sea in a fair-
weather craft from a south coast watering-place. A day or two later
the overturned boat was picked up with the ex-convict's coat still
entangled in the seat, and with his ticket-of-leave still in the pocket.
There was nothing to connect the Lynton Hora who a few weeks
later landed from an English tramp steamer at an Italian port with the
missing man.
Hora had not found existence present many difficulties. He had
buried his scruples with his identity, and a man of brains, with
courage and no scruples, need never look very far for the means of
subsistence. For a while he preyed on British tourists. They were of
his own race, and, therefore, his chiefest enemies, and, besides, he
knew that, since he would need a place where he might build a
reputation and a new identity if his purposes were to be fulfilled, it
would be unwise to prey upon the inhabitants of the selected spot.
Italy appealed to him. He became to all intents and purposes Italian.
An English soap-maker's wife, "seeing" the Eternal City, supplied
him, unwittingly, with funds to purchase a vineyard in Tuscany. He
stocked his farm and furnished a house with the contents of a
duchess' jewel casket. The capital necessary for pursuing his
agricultural operations was provided indirectly by the Casino
authorities at Monte Carlo. Hora had ventured no stake at the tables.
He had merely relieved a successful gambler of his winnings. Thus
provided with a home, he had paid a visit to England. When he
returned, six months later, he brought his reputed son with him, a
child of three; and away in England, his old comrade, now Captain
Marven, together with Mrs. Marven, had mourned beside an empty
cot in their nursery.
Hora had succeeded in the initial step towards the
accomplishment of his revenge. But this had been only the first step.
His appetite was not to be sated with one simple meal of vengeance.
His rival, like himself, should never be allowed to forget his loss. So
punctually every year, on the anniversary of the stolen boy's birthday,
Marven had received a brief type-written note stating that the child
was alive and well—nothing more. Hora would gladly have signed
the note with his forgotten name, but that thereby he might have
incurred danger to himself and the overthrow of his whole scheme of
revenge. When the appointed time came, when the child was full
grown, when by his own acts the child should be damned beyond all
redemption—then the woman who had refused his offer of marriage
should have her son restored to her, the rival who had won that
woman's love from him should have the paternity of the criminal
thrust upon him, and the whole world should be made aware of
Guy's real parentage. That was the complete scheme of revenge
Hora contemplated; to consummate which he had instilled into the
baby ears the subtle poison of his perverted morality, had skilfully
taken advantage of the boy's adventurous nature to interest him in
the romantic possibilities of a criminal career, had laboured and
watched the unfolding of a mind with the patience of a Japanese
gardener producing a dwarfed and twisted miniature of a fair tree of
the forest.
He had been discreet in his work. He had no intention of making
of his pupil a rod for his own scourging. His conception of the great
criminal he desired to make of Guy had nothing in common with the
average conception of a person given to indulgence in all the
commonplace vices of humanity. Self-control he had early realised
was of more importance to the man who was waging single-handed
warfare with the world, than to the units of the community with whom
he was at issue. His own predilections, too, were instinctively
refined. The grosser forms of self-indulgence had never appealed to
him. He was an epicure of life, and had no desire to spoil his palate
with a surfeit of coarse pleasures. Clean living himself, he demanded
cleanliness of life in those about him. To what happened outside of
his own household he was cynically indifferent.
Guy had proved a credit to his training. He was healthy in body,
full of the enthusiasm of living, and possessed of a fine rapture for
the profession to which he had served his apprenticeship. Almost the
time was ripe for the consummation of Hora's revenge, when chance
had brought Guy into contact with his real parents. This was a
contingency Hora had not foreseen, and it needed careful
consideration. He did not fear that the relationship would be
disclosed. Guy himself had no suspicion of the facts. He knew no
parent but Hora, though he believed that he remembered the mother
whom Hora had invented for his benefit, whose portrait hung on the
wall of his bedroom, and of whom Hora had spoken to him on many
occasions. Yes, Guy Marven's real identity was sufficiently sunken in
that of Guy Hora to ensure him against discovery, even though
physical likeness should lead to comment.
Yet, Hora's first emotion, on learning that his foster-son had met
his father and mother, was one he thought he had banished forever.
A sensation of fear had passed over him, a dread lest the natural
inclination of son to mother should manifest itself, lest the blood
which pulsed eagerly in the son's arteries should cry out to the blood
which ran more sluggishly in his father's veins, and, his own mock
relationship disestablished, there be destroyed the living instrument
for his revenge he had spent so many years in fashioning. Nor had
his only fear been for the loss of his whole scheme of revenge. He
realised, for the first time, that his interest in Guy was more than that
of the artist in his artistry. Guy had always looked to him, had repaid
him for his attention with all the warmth of an affectionate nature. He
was the one being in whom, save Myra, Hora had taken a personal
interest. Suppose someone else were to take his, Hora's, place in
the young man's thoughts? The dread was in his mind, though he
would not acknowledge it—though he denied its existence. That
would be a piece of sentimentalism utterly foreign to his whole
nature. He told himself that he had no affection for the child of his
adoption, save that of the master craftsman in his tool. Of course he
would regret the necessity, when it arose, of giving the tool to
destruction, but he would admit to himself no warmer interest in
Guy's fate than that.
Self-persuaded on the point, he considered whether the
meeting, of which he had been apprised, might not be utilised for the
furtherance of his plans. Nor was it long before he became
persuaded that Fate was playing into his hands. Supposing that the
acquaintance developed into intimacy. A thousand vague
possibilities floated, shadow-like, before Hora's eyes. He determined
that the acquaintance should be continued, but still fearful, he
determined also that Guy should be plunged more deeply into the
vortex of crime than hitherto, so that, struggle and strive as he might,
he should find it impossible to escape. Fortunately for his purpose,
Guy had expressed himself as hungering for further adventure. Well,
Hora was fertile of plans, and he saw very good reasons why Guy's
desires should be humoured. His household saw nothing more of the
Commandatore that day. He remained alone with his thoughts.
CHAPTER V
THE COMMANDATORE MAKES A DEDUCTION
"We are getting near the end of our resources, Guy," remarked
Hora quietly, as he held a glass of port up to the light, sipped the
wine, nodded his head approvingly, and set the glass down gently.
It was the evening of the second day after Hora's exhibition of
emotion upon hearing the name of Marven. He had not referred
again to the object of his hatred, and neither Myra nor Guy, who sat
with him at the table, had prompted his memory.
Guy looked round the room before he answered. He had been
well trained in the observance of caution. But the servants had
retired, the door was closed. The three were alone.
"All London offers replenishment of our empty coffers," he
answered light-heartedly. "Who is to have the honour?" He turned to
Myra. "Shall I peel a peach for you?" he asked.
The woman seemed not to hear the question. She was looking
at Hora, with an appeal in her glance.
Hora answered her glance. "Myra is tired of London," he
remarked. "What do you say, Guy? Shall we finish the campaign
now, strike our tents and retire like contented bourgeoisie to our
vineyard to watch the grapes ripen?"
Guy's eyebrows arched in surprise. "Retire empty-handed?" he
asked incredulously. "Why, what has come upon you,
Commandatore?"
"Myra is tired," he answered briefly.
Guy looked, smilingly, at her. She flushed slightly. "Not a bit of
it," he answered. "I am quite sure she does not desire to exchange
the delights of a London season, even for the dolce far niente of an
Italian summer."
"I should not mind," she answered. "London is a beastly place.
The Commandatore is right. I am sick of the sight and sound of
people, and of the perpetual menace of our life—I——"
Hora checked her speech with a gesture. The door opened and
a servant entered with coffee, and while he was present the
conversation passed lightly over topics of the day.
"I don't like that man," said Guy, as the servant withdrew. "I
caught him prying about amongst my belongings the other day when
I returned to the flat unexpectedly."
"All servants do that," murmured Hora indifferently. "Curiosity is
the mental badge of servitude. The servant is never happy until he
has surprised one of his master's secrets. It would be just as well,
Guy, if you were to supply him with a few facts to exercise his
imagination upon. Get some girl to write you a few love letters and
hide them where he can find them. He will never be at a loss then to
supply a reason for any erratic movement of yours."
Guy laughed. "Not a bad suggestion," he agreed. "Do you adopt
the same plan to protect yourself?"
Hora shrugged his shoulders. "I carefully built up my own
reputation in advance," he remarked. "Haven't I told you? I suppose
not, for you were both too young when I first located myself here."
He looked round the pleasant dining-room complacently. "I've had
the place for ten years now, and for one's name to be for ten years in
the London directory, at the same address, is a certificate of
respectability which is not easily discredited."
"Still I wonder that you did not seek greater privacy," remarked
Guy, as he lit a cigarette.
Hora smiled. "A decision for privacy always awakens suspicion,
and thus in our profession privacy de facto is perhaps the one luxury
we cannot afford. Nevertheless a greater degree of privacy is
possible in the midst of a crowd than would be possible anywhere
else in the wide world. This is not such a paradoxical statement as it
sounds. In the crowd no one is intent on the doings of his
neighbours. Put a ring-fence round a man, and every eye would be
fixed upon him. Thus you see my reason for selecting a residential
flat for my London residence. The servants are not mine. Each of
them has half a dozen other objects of curiosity. When they have
attended to our requirements they disappear."
"But, nevertheless, they must be curious concerning the
contents of the art gallery?"
The allusion was to a portion of the abode into which the
servants were not supposed to enter. Though situated on the eighth
story, Hora's flat at Westminster Mansions was not the ultimate
achievement of the builder. Above were attics to which a narrow
staircase gave entrance. The stairs were shut off by a door, and the
door was always locked.
"When I see any signs of curiosity I always take an early
opportunity of gratifying it," said Hora. "Every one of the servants
who has ever waited upon me has had the privilege of inspecting
that chamber, and not one of them has ever been sufficiently
interested to enter it a second time, except at my especial request.
You see they are all aware why I took possession of the attic. They
think it is the fad of a nervous invalid. Those attics were entered from
another staircase when first I took the flat, and some of the servants
slept there. I complained of the noise, continually. Half a dozen of the
poor devils must have been dismissed at one time or another for
purely imaginary offences in consequence. Then I declared I could
stay no longer, and I gave notice to leave. The agent for the landlord
was apologetic, and asked if there was no way in which he would not
be able to meet me. I offered to rent the place, saying that I would
make it into a storeroom for the books and trifles which I am
continually accumulating. He jumped at the offer I made, and I know
he thought me a fool." Hora chuckled. "How surprised he would be to
learn that the proceeds of many a rich haul have been stored there
for months. But I have drifted away from my original point. I was
telling you of the manner in which I built up my original reputation for
eccentricity, the safest cloak a man may wear. It was a simple
matter. I merely answered for myself the references I gave to my
landlord. I described myself as an unfavourable tenant from every
point of view, but the pecuniary one. My habits I described as
irregular, my requirements exacting to a degree, my manner brusque
and overbearing, and my disposition faddy and changeable, and
further said I was given to making continual requests for structural
alterations in any dwelling place that I occupied in order to make
accommodation for any new collecting craze which seized me."
"I wonder any landlord ventured to accept you," laughed Myra.
"The London landlord has a high opinion of his capability for
withstanding the demands of his tenants," said Hora drily. "He is a
man lavish of promises, but meagre of fulfilments, and possessed of
a genius for extracting the uttermost farthing of his rent. Moreover,
he would take Satan himself as a tenant if he offered to pay six
months' rent in advance. Naturally I proved acceptable, and not
turning out to be the terror I depicted myself I am now looked upon
as the best tenant in the whole building. I am free to do as I like. My
treasure-house ceases to excite curiosity, and I believe if I were to
place the crown jewels upon one of the tables up there they would
be undisturbed, so long as my rent was paid regularly, until they
were hidden beneath the accumulated dust of ages."
The allusion gave Guy an idea.
"Do you contemplate an imitation of Colonel Blood's exploit for
the replenishment of our empty exchequer?" he said, smiling.
"I have often envied Blood's opportunities," answered Hora
thoughtfully, "but at the present day there are much greater
difficulties in the way than Blood had to contend with. Some day,
perhaps, but just now I have another scheme in my mind." He rose
from the table. "I have something to tell you," he remarked. "You will
excuse me for a minute."
He left the room. As the door closed on Hora, Myra turned
eagerly to her companion. She felt that, despite her promise to Hora,
she must give utterance to the fears which once again possessed
her mind.
"Guy," she said, "I wish you would persuade the Commandatore
to leave London for a while. He would listen to any wish of yours."
"Do you think so?" he asked. "I don't think that any expression
of mine would turn him from any purpose he has in view."
"But can you not try?" she persisted. "For my sake, Guy."
"Why, whatever is the matter with you, Myra?" asked the young
man, his attention captured by the obvious anxiety in her voice.
"Surely you are not becoming afraid?"
"Becoming afraid?" she repeated after him mechanically. "No, I
am not becoming afraid. I learned what fear was long ago, when first
I ventured to put my own desires in opposition to the will of the
Commandatore. I have always been afraid since then." She fell to
silence.
"There's no reason to fear the Commandatore," answered Guy
cheerfully. "You are growing morbid, Myra."
She paid no heed to his comment. "It is not fear now, or at least
not what is generally understood by fear. There is an oppression in
the air, the weight of something unseen and unknown presses on
me."
"But there is nothing for you to fear. Whatever were to happen
you would be quite safe," argued Guy.
"Myself? It is not myself I am thinking about," she cried
passionately. "Whatever impends does not threaten me. It is you,
Guy, I fear for. Ever since the night of Lady Greyston's dance I have
felt it. I thought you would never return that evening, but you came
back, and for a while I could laugh at my fears. But, now the
Commandatore has some other proposal to make, my dread has
returned. I shall not have a moment's rest."
"Why this is sheer hysteria, if not madness," said Guy in great
concern.
"Call it what you like," she replied earnestly, "but listen to what I
say—promise me!"
She heard Hora's lagging footstep in the passage outside, and
she ceased speaking suddenly. "Not a word of this to the
Commandatore," she said hastily, as the door opened, and Hora re-
entered.
If the elder man observed that his re-entry had broken in upon a
confidence from which he was excluded, he gave no signs of having
done so. Myra breathed more freely when he seated himself again at
the table, and spread out a newspaper he had brought with him on
the table.
"There are three items of news in this evening's paper," he
remarked quietly, "which supply the data from which may be
deduced the means whereby an enterprising man may build a
fortune."
Guy was all attention on the instant, and Myra, viewing his keen
face, let her head droop upon her hand.
"Those items are?" queried the young man, as Hora paused.
"You will find the first in the Court News," was the reply. "The Rt.
Hon. Sir Gadsby Dimbleby, who is the minister in attendance upon
His Majesty the King, arrived at Sandringham last night."
"The Minister of Foreign Affairs, is he not?" asked Guy.
Hora nodded, and turned to another page from which he read:
"Just before the close all markets sagged badly on selling orders
from Berlin. A variety of rumours were afloat as to the reason, but no
definite information which would supply justification for a bear raid on
the market was forthcoming from any well-informed quarter. In the
street, afterwards, prices were put up again generally, though
fluctuations were considerable."
"Yes," said Guy, beginning to look puzzled.
"The third item is a mere addition to the Reuter's telegram from
Australia, giving particulars of the cricket match between the English
and Australian teams. The result is placed in the space left for late
news, and over it are the words "delayed in transmission.""
Hora ceased speaking.
"If the rehabilitation of our fortunes depends upon translating
that puzzle we shall end our days in the workhouse yet," said Guy.
"Yet, there is much wealth for the man who can piece together
those scraps of information, and will act promptly on the knowledge,"
answered Hora.
"How? By speculation on the Stock Exchange?" asked Guy. "I
thought, Commandatore, that you eschewed all forms of gambling."
"I do," said Hora drily. "But to buy and sell on a certainty has
nothing of the gambling element about it. I feel inclined to make
either the bulls or the bears contribute to our maintenance. But
action must be prompt if it is to succeed. There is work for you to do
to-night, Guy, if you care about it."
"Care about it?" The young man sprang to his feet, every fibre of
his frame quivering for action.
Hora laughed good-naturedly. "There—there, Guy, take matters
a bit easily. There's plenty of time before you yet, if you decide to go
on with the job. It's more risky than the last."
"The greater the risk, the better I shall be pleased," exclaimed
Guy, as he dropped again into his chair, "though how you are going
to evolve anything of a risky nature from those paragraphs you have
just read, I entirely fail to understand."
"You'll understand soon enough," remarked Hora quietly, "and
you will then be surprised that the meaning of these three items of
news should have conveyed so little to you. Let me reverse the order
and read into these three facts my own conclusions. What can be
the reason for the delay in the transmission of the cable containing
the cricket result? Either the cable had broken down, or it was
monopolised for more important work. The former theory is
untenable, for if you take the trouble to compare the time of the
insertion of the news with the time when it should have been
inserted, you will find a delay of three or four hours only has to be
accounted for. Thus I arrive at the decision that the cable was fully
occupied by someone with a prior claim for its use. Who could that
be? Here again the choice is between two possibilities. Either some
big financier or body of financiers or the Government. Again the
indications point to one conclusion. The City was merely uneasy by
reason of German selling, which could not be accounted for, and not
because of information which had come over the wires. Therefore,
the wires must have been occupied by important despatches to the
British Government. I think," said Hora, "that if the knowledge of
what has passed over the cable is in my possession by to-morrow
morning, we shall be in a position to spoil the Egyptians of
Throgmorton Street to some purpose."
Guy looked at Hora with admiration. Some idea of his
companion's purpose dawned upon him—but only faintly. He asked
eagerly for further guidance.
"As to the nature of the despatch which has been received at
the Foreign Office, I have no more idea than yourself," he continued,
"though it probably affects Germany, and it is hardly worth while
troubling to guess. I am only concerned with times, places, and
people. As I calculate, the cable was not clear for ordinary business
until close upon six o'clock. Six would, therefore, be very near the
time when the end of the message was delivered at the Foreign
Office. Of course it would have been cabled in the official cypher. By
the time the message would be de-coded there is only one train by
which a special messenger could take the de-coded despatch to his
chief, who happens to be the minister in attendance upon His
Majesty at Sandringham."
Hora looked up at the clock. "That train starts from St. Pancras
at 9.50. It proceeds as far as Lynn, where the messenger carrying
the despatch will probably be met by a motor-car. It is just nine
o'clock now, Guy, so there is plenty of time for you to decide whether
it is worth while making an effort to obtain the information which will
be in his despatch box."
Guy's eyes sparkled. "It's worth while trying any way,
Commandatore." He turned to the young woman. "Wish me luck,
Myra," he said.
CHAPTER VI
WHEREIN A KING'S MESSENGER IS DESPOILED
OF HIS DESPATCHES
Guy had ten minutes to spare as he stepped out of the cab at

St. Pancras Station, and, handing a bag to a porter, made his way
leisurely to the booking office and took a ticket for Lynn. He would
have been easily recognised by any of his acquaintance, for he had
made no effort to disguise himself. Hora professed no liking for
physical or material disguises, contending, indeed, that such were
merely the clumsy devices of incompetence. "The man who
anticipates being seen in any affair in which he is not prepared to
meet scrutiny is anticipating failure," he was accustomed to declare.
"In all other cases," he had told Guy, "your own identity will be your
most certain protection. A bourgeois reputation for wealth and half a
dozen society acquaintances will provoke apologies from inquisitive
persons when a mere hint that you had dyed your moustache would
mean instant arrest." Guy was disposed to agree with him. He
disliked the theatrical, as much as he was enamoured of the
dramatic, aspect of his profession.
Yet on this occasion he had not been long on the platform
before he was betrayed into the wish that for once he had assumed
some disguise. The porter to whom he had entrusted his bag had
deposited it in the corner of a first-class smoking compartment, and
Guy strolled along the waiting train, glancing into each compartment
in turn, in order to locate the messenger who, if Hora's deductions
were correct, was to travel by it. He had made no plans as to the
means he was to take to obtain a knowledge of the despatches. Nor
had Hora made any suggestions.
"I leave the matter entirely in your hands, Guy," he had said. "If
you see a favourable opportunity, take it. Opportunity only passes by
the unobservant. Still, such may be lacking, and if none presents
itself do not strive to make one. One failure in our profession is the
end of a career. I don't want to see your career cut short yet. It is a
difficult task you have undertaken, and I shall not be surprised if no
opportunity presents itself. Don't think I shall blame you."
The words were in his mind, as he passed coach after coach,
and saw no one who seemed at all likely to be the messenger
bearing the despatches. He returned to the door of the compartment
where his bag lay. A porter came to the door, glanced in, then
opened the door of the next compartment, and pasted a little label
"reserved" on the window. Guy, standing on the platform, looked at
the entrance a dozen yards distant. He became conscious of a little
thrill of excitement in the thought that Hora's deductions were likely
to be proved accurate. It was the next instant that he wished that he
had concealed his identity. A passenger made his appearance and
the moment their eyes met Guy recognised him, even as he knew he
was himself recognised. It was too late to seek retirement in the
train. He stood still while the newcomer advanced with outstretched
hand.
"Guy Hora, if I'm not mistaken, by all that's lucky. Going
anywhere on this train, eh?"
Mechanically Guy took the offered hand. There was no need for
him to attempt to infuse any surprise into his voice. He could only
ejaculate "Captain Marven!"
"Himself," said Marven, "and as much surprised to find a friend
here as you are." He glanced into the compartment where Guy's bag
remained.
"Unless you want to sleep undisturbed, I should propose that we
travel together," he remarked.
"If I might intrude," begun Guy hesitatingly.
"Intrude? You don't know how delighted I shall be," said Marven
heartily. He beckoned the porter who stood holding open the door of
the reserved compartment. "Put this gentleman's bag in here," he
said. "Now, jump in," he continued to Guy.
The young man obeyed. The porter closed the door, touched his
cap, and turned away. Captain Marven tossed a bundle of
magazines on the seat and settled himself in a corner. "Better to look
upon than magazines is the face of a friend on a five hours' journey,"
he said laughingly.
"Much better," replied Guy. His first surprise had been
succeeded by a second much greater. Captain Marven's coat had
swung back as he entered the compartment, and Guy had observed
beneath it a despatch case. It flashed upon him that Captain Marven
must be the carrier of the despatches. There seemed to be
something uncanny in Hora's prescience. Fate had given him the
opportunity which the Commandatore had declared awaited the
observant, and it remained for Guy to take advantage of it. But there
was a bitter thought commingled with the realisation of the fact that
the opportunity he desired had been given to him. This man had
seemed so friendly disposed towards him from the moment they first
met. There seemed to Guy to be something particularly mean in an
act of treachery towards him. But he had no opportunity to brood
upon the thought, even had he desired to do so, and he did not
desire to contemplate it. Besides, not to take full advantage of the
opportunity would be treachery to the Commandatore, who was
depending upon his success. Guy put aside his new-born scruples,
and concentrated his attention on the business in hand.
He perceived that he would have to make use of all his faculties
if he were to succeed in his object. The man opposite him would not
be easily cozened. But the effort must be made. He began to turn
over in his brain a variety of plans, while he chatted with his
companion.
The whistle sounded, doors banged, the train moved out into the
night. There was no escape now. But how was he to gain possession
of the case in Captain Marven's possession without exciting
suspicion? As a stranger, there would have been greater chance of
effecting the coup by some sudden stratagem. But, positioned as he
now found himself, he was practically helpless unless Marven should
sleep. Guy looked up at the Captain's face. It was an alert face, not
the face of a man who sleeps while on duty. It would be difficult to
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade
Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.
Let us accompany you on the journey of exploring knowledge and

personal growth!
ebookultra.com

Download ebooks file Signal Processing for Neuroscientists A Companion Volume Advanced Topics Nonlinear Techniques and Multi Channel Analysis Elsevier Insights 1st Edition Drongelen all chapters

Uploaded by

Download ebooks file Signal Processing for Neuroscientists A Companion Volume Advanced Topics Nonlinear Techniques and Multi Channel Analysis Elsevier Insights 1st Edition Drongelen all chapters

Uploaded by

Download the full version of the ebook now at ebookultra.

Signal Processing for Neuroscientists A

Explore and download more ebook at https://ebookultra.com

Signal Processing for Neuroscientists 1st Edition Wim Van

Digital Signal Processing and Applications with the

Statistical techniques for neuroscientists 1st Edition

Time Frequency Signal Analysis and Processing A

Wavelet Analysis and Transient Signal Processing

Signal Processing for Digital Communications Artech House

Multi Platinum Pro Tools Advanced Editing Pocketing and

Multiple Input Multiple Output Channel Models Theory and

y = f (α) + 1– (x – α) f '(α) + 1– (x – α)2 f "(α) + –1 (x – α)3f "'(α) +...

(τ1, τ2) x(t – τ1) x(t – τ2) dτ1dτ2

Wim van Drongelen

AMSTERDAM BOSTON HEIDELBERG LONDON NEW YORK OXFORD

PARIS SAN DIEGO SAN FRANCISCO SINGAPORE SYDNEY TOKYO

First edition 2010

Copyright r 2010 Elsevier Inc. All rights reserved

British Library Cataloguing-in-Publication Data

For information on all Elsevier publications

It is a pleasure to acknowledge those who have assisted (directly and indirectly) in

1.2 Unevenly Sampled Data

Signal Processing for Neuroscientists, A Companion Volume. DOI: 10.1016/B978-0-12-384915-1.00001-2

1.2.1 Lomb’s Algorithm

Pða; b; f; tn Þ 5 a cosð2πftn Þ 1 b sinð2πftn Þ ð1:1Þ

@ε2 =@a 5 0 ð1:2aÞ

@ε2 =@b 5 0 ð1:2bÞ

To separate the expressions for tn and τ, we use the trigonometric relationship

Figure 1.2 The Lomb algorithm

This can be further simplified into:

Hence, condition (1.4) is satisfied if:

The value of variable τ as a function of frequency f can be found with Equation

Since we shift by τ, all cross-terms vanish and by substitution of the expressions

This can be further simplified into:

1.2.2 A MATLAB Example

1.3 The Hilbert Transform

connections between and activities of the neuronal elements are established, it is a

Aejωt 5 A½cosðωtÞ 1 j sinðωtÞ 5 A cosðωtÞ 1 jA sinðωtÞ ð1:10Þ

fa ðtÞ 5 A e jωt 5 A cosðωtÞ 1 j A sinðωtÞ 5 f ðtÞ 1 j f~ðtÞ

in which fa ðtÞ is defined as the analytic signal.

(A) Analytic signal (B) Time domain

Fa(ω) = 2Aπ δ (ω −ω0) Aπ δ (ω − ω0) Aπ δ (ω − ω0)

Fourier transform of the j

and the phase is:

1.3.1 The Hilbert Transform in the Frequency Domain

multiplication by 2 j sgnðωÞ ð1:11Þ

System’s input3Fourier transform: f ðtÞ 3FðωÞ

HðωÞ 5 2jsgnðωÞ ð1:14Þ

Aπ½δðω 1 ω0 Þ 1 δðω 2 ω0 Þ ð1:15Þ

Aπ½δðω 1 ω0 Þ 1 δðω 2 ω0 Þ 2j sgnðωÞ

Aπ½δðω 1 ω0 Þð2 j sgnð2ω0 ÞÞ 1 δðω 2 ω0 Þð2j sgnðω0 ÞÞ

Now we use the definition of sgn, sgnð2ω0 Þ 5 21 and sgnðω0 Þ 5 1 (Equation

Aπ½δðω 1 ω0 ÞðjÞ 1 δðω 2 ω0 Þð 2 jÞ 5 Ajπ½δðω 1 ω0 Þ 2 δðω 2 ω0 Þ ð1:17Þ

As expected, Equation (1.17) is the Fourier transform of A sinðωtÞ (see equation

fa ðtÞ 5 f ðtÞ 1 j f~ðtÞ3Fa ðωÞ 5 FðωÞ 1 jFðωÞ

With a bit of algebra we obtain:

no energy at negative frequencies 2ω0 ; it only has energy at positive frequencies

1.3.2 The Hilbert Transform in the Time Domain

if f ðtÞ3FðωÞ; then FðtÞ32πf ð2ωÞ ð1:19aÞ

the input: f ðtÞ3 FðωÞ

step=0.00001; % step size=1/sample rate

X=fft(x); % Fourier transform of the signal

(A) Pre-processing of the signal

(ii) Spike raster

(B) Using the Hilbert transform

Figure A2.1 The signum function

2.2 Different Types of Models

Signal Processing for Neuroscientists, A Companion Volume. DOI: 10.1016/B978-0-12-384915-1.00002-4

property that we will use when developing the regression procedure in

dn yðtÞ dn21 yðtÞ

for continuous time systems and:

An yðk 2 nÞ 1 An21 yðk 2 n 1 1Þ 1 ? 1 A0 yðkÞ

2.3 Examples of Parametric and Nonparametric Models

A very famous parametric model in neuroscience is the Hodgkin and Huxley

In this expression we have several parameters: membrane capacitance C; the resis-

in which a, b, and c are parameters; I is a term representing injected current. The